linux/fs/namei.c

5388 lines
143 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 22:07:57 +08:00
// SPDX-License-Identifier: GPL-2.0
/*
* linux/fs/namei.c
*
* Copyright (C) 1991, 1992 Linus Torvalds
*/
/*
* Some corrections by tytso.
*/
/* [Feb 1997 T. Schoebel-Theuer] Complete rewrite of the pathname
* lookup logic.
*/
/* [Feb-Apr 2000, AV] Rewrite to the new namespace architecture.
*/
#include <linux/init.h>
#include <linux/export.h>
#include <linux/slab.h>
#include <linux/wordpart.h>
#include <linux/fs.h>
#include <linux/filelock.h>
#include <linux/namei.h>
#include <linux/pagemap.h>
#include <linux/sched/mm.h>
#include <linux/fsnotify.h>
#include <linux/personality.h>
#include <linux/security.h>
#include <linux/syscalls.h>
#include <linux/mount.h>
#include <linux/audit.h>
#include <linux/capability.h>
#include <linux/file.h>
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
#include <linux/fcntl.h>
cgroups: implement device whitelist Implement a cgroup to track and enforce open and mknod restrictions on device files. A device cgroup associates a device access whitelist with each cgroup. A whitelist entry has 4 fields. 'type' is a (all), c (char), or b (block). 'all' means it applies to all types and all major and minor numbers. Major and minor are either an integer or * for all. Access is a composition of r (read), w (write), and m (mknod). The root device cgroup starts with rwm to 'all'. A child devcg gets a copy of the parent. Admins can then remove devices from the whitelist or add new entries. A child cgroup can never receive a device access which is denied its parent. However when a device access is removed from a parent it will not also be removed from the child(ren). An entry is added using devices.allow, and removed using devices.deny. For instance echo 'c 1:3 mr' > /cgroups/1/devices.allow allows cgroup 1 to read and mknod the device usually known as /dev/null. Doing echo a > /cgroups/1/devices.deny will remove the default 'a *:* mrw' entry. CAP_SYS_ADMIN is needed to change permissions or move another task to a new cgroup. A cgroup may not be granted more permissions than the cgroup's parent has. Any task can move itself between cgroups. This won't be sufficient, but we can decide the best way to adequately restrict movement later. [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: fix may-be-used-uninitialized warning] Signed-off-by: Serge E. Hallyn <serue@us.ibm.com> Acked-by: James Morris <jmorris@namei.org> Looks-good-to: Pavel Emelyanov <xemul@openvz.org> Cc: Daniel Hokka Zakrisson <daniel@hozac.com> Cc: Li Zefan <lizf@cn.fujitsu.com> Cc: Paul Menage <menage@google.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-29 16:00:10 +08:00
#include <linux/device_cgroup.h>
#include <linux/fs_struct.h>
#include <linux/posix_acl.h>
vfs: fix bad hashing of dentries Josef Bacik found a performance regression between 3.2 and 3.10 and narrowed it down to commit bfcfaa77bdf0 ("vfs: use 'unsigned long' accesses for dcache name comparison and hashing"). He reports: "The test case is essentially for (i = 0; i < 1000000; i++) mkdir("a$i"); On xfs on a fio card this goes at about 20k dir/sec with 3.2, and 12k dir/sec with 3.10. This is because we spend waaaaay more time in __d_lookup on 3.10 than in 3.2. The new hashing function for strings is suboptimal for < sizeof(unsigned long) string names (and hell even > sizeof(unsigned long) string names that I've tested). I broke out the old hashing function and the new one into a userspace helper to get real numbers and this is what I'm getting: Old hash table had 1000000 entries, 0 dupes, 0 max dupes New hash table had 12628 entries, 987372 dupes, 900 max dupes We had 11400 buckets with a p50 of 30 dupes, p90 of 240 dupes, p99 of 567 dupes for the new hash My test does the hash, and then does the d_hash into a integer pointer array the same size as the dentry hash table on my system, and then just increments the value at the address we got to see how many entries we overlap with. As you can see the old hash function ended up with all 1 million entries in their own bucket, whereas the new one they are only distributed among ~12.5k buckets, which is why we're using so much more CPU in __d_lookup". The reason for this hash regression is two-fold: - On 64-bit architectures the down-mixing of the original 64-bit word-at-a-time hash into the final 32-bit hash value is very simplistic and suboptimal, and just adds the two 32-bit parts together. In particular, because there is no bit shuffling and the mixing boundary is also a byte boundary, similar character patterns in the low and high word easily end up just canceling each other out. - the old byte-at-a-time hash mixed each byte into the final hash as it hashed the path component name, resulting in the low bits of the hash generally being a good source of hash data. That is not true for the word-at-a-time case, and the hash data is distributed among all the bits. The fix is the same in both cases: do a better job of mixing the bits up and using as much of the hash data as possible. We already have the "hash_32|64()" functions to do that. Reported-by: Josef Bacik <jbacik@fb.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christoph Hellwig <hch@infradead.org> Cc: Chris Mason <clm@fb.com> Cc: linux-fsdevel@vger.kernel.org Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-09-14 02:30:10 +08:00
#include <linux/hash.h>
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
#include <linux/bitops.h>
#include <linux/init_task.h>
#include <linux/uaccess.h>
#include "internal.h"
#include "mount.h"
/* [Feb-1997 T. Schoebel-Theuer]
* Fundamental changes in the pathname lookup mechanisms (namei)
* were necessary because of omirr. The reason is that omirr needs
* to know the _real_ pathname, not the user-supplied one, in case
* of symlinks (and also when transname replacements occur).
*
* The new code replaces the old recursive symlink resolution with
* an iterative one (in case of non-nested symlink chains). It does
* this with calls to <fs>_follow_link().
* As a side effect, dir_namei(), _namei() and follow_link() are now
* replaced with a single function lookup_dentry() that can handle all
* the special cases of the former code.
*
* With the new dcache, the pathname is stored at each inode, at least as
* long as the refcount of the inode is positive. As a side effect, the
* size of the dcache depends on the inode cache and thus is dynamic.
*
* [29-Apr-1998 C. Scott Ananian] Updated above description of symlink
* resolution to correspond with current state of the code.
*
* Note that the symlink resolution is not *completely* iterative.
* There is still a significant amount of tail- and mid- recursion in
* the algorithm. Also, note that <fs>_readlink() is not used in
* lookup_dentry(): lookup_dentry() on the result of <fs>_readlink()
* may return different results than <fs>_follow_link(). Many virtual
* filesystems (including /proc) exhibit this behavior.
*/
/* [24-Feb-97 T. Schoebel-Theuer] Side effects caused by new implementation:
* New symlink semantics: when open() is called with flags O_CREAT | O_EXCL
* and the name already exists in form of a symlink, try to create the new
* name indicated by the symlink. The old code always complained that the
* name already exists, due to not following the symlink even if its target
* is nonexistent. The new semantics affects also mknod() and link() when
* the name is a symlink pointing to a non-existent name.
*
* I don't know which semantics is the right one, since I have no access
* to standards. But I found by trial that HP-UX 9.0 has the full "new"
* semantics implemented, while SunOS 4.1.1 and Solaris (SunOS 5.4) have the
* "old" one. Personally, I think the new semantics is much more logical.
* Note that "ln old new" where "new" is a symlink pointing to a non-existing
* file does succeed in both HP-UX and SunOs, but not in Solaris
* and in the old Linux semantics.
*/
/* [16-Dec-97 Kevin Buhr] For security reasons, we change some symlink
* semantics. See the comments in "open_namei" and "do_link" below.
*
* [10-Sep-98 Alan Modra] Another symlink change.
*/
/* [Feb-Apr 2000 AV] Complete rewrite. Rules for symlinks:
* inside the path - always follow.
* in the last component in creation/removal/renaming - never follow.
* if LOOKUP_FOLLOW passed - follow.
* if the pathname has trailing slashes - follow.
* otherwise - don't follow.
* (applied in that order).
*
* [Jun 2000 AV] Inconsistent behaviour of open() in case if flags==O_CREAT
* restored for 2.4. This is the last surviving part of old 4.2BSD bug.
* During the 2.4 we need to fix the userland stuff depending on it -
* hopefully we will be able to get rid of that wart in 2.5. So far only
* XEmacs seems to be relying on it...
*/
/*
* [Sep 2001 AV] Single-semaphore locking scheme (kudos to David Holland)
* implemented. Let's see if raised priority of ->s_vfs_rename_mutex gives
* any extra contention...
*/
/* In order to reduce some races, while at the same time doing additional
* checking and hopefully speeding things up, we copy filenames to the
* kernel data space before using them..
*
* POSIX.1 2.4: an empty pathname is invalid (ENOENT).
* PATH_MAX includes the nul terminator --RR.
*/
#define EMBEDDED_NAME_MAX (PATH_MAX - offsetof(struct filename, iname))
syscalls: implement execveat() system call This patchset adds execveat(2) for x86, and is derived from Meredydd Luff's patch from Sept 2012 (https://lkml.org/lkml/2012/9/11/528). The primary aim of adding an execveat syscall is to allow an implementation of fexecve(3) that does not rely on the /proc filesystem, at least for executables (rather than scripts). The current glibc version of fexecve(3) is implemented via /proc, which causes problems in sandboxed or otherwise restricted environments. Given the desire for a /proc-free fexecve() implementation, HPA suggested (https://lkml.org/lkml/2006/7/11/556) that an execveat(2) syscall would be an appropriate generalization. Also, having a new syscall means that it can take a flags argument without back-compatibility concerns. The current implementation just defines the AT_EMPTY_PATH and AT_SYMLINK_NOFOLLOW flags, but other flags could be added in future -- for example, flags for new namespaces (as suggested at https://lkml.org/lkml/2006/7/11/474). Related history: - https://lkml.org/lkml/2006/12/27/123 is an example of someone realizing that fexecve() is likely to fail in a chroot environment. - http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=514043 covered documenting the /proc requirement of fexecve(3) in its manpage, to "prevent other people from wasting their time". - https://bugzilla.redhat.com/show_bug.cgi?id=241609 described a problem where a process that did setuid() could not fexecve() because it no longer had access to /proc/self/fd; this has since been fixed. This patch (of 4): Add a new execveat(2) system call. execveat() is to execve() as openat() is to open(): it takes a file descriptor that refers to a directory, and resolves the filename relative to that. In addition, if the filename is empty and AT_EMPTY_PATH is specified, execveat() executes the file to which the file descriptor refers. This replicates the functionality of fexecve(), which is a system call in other UNIXen, but in Linux glibc it depends on opening "/proc/self/fd/<fd>" (and so relies on /proc being mounted). The filename fed to the executed program as argv[0] (or the name of the script fed to a script interpreter) will be of the form "/dev/fd/<fd>" (for an empty filename) or "/dev/fd/<fd>/<filename>", effectively reflecting how the executable was found. This does however mean that execution of a script in a /proc-less environment won't work; also, script execution via an O_CLOEXEC file descriptor fails (as the file will not be accessible after exec). Based on patches by Meredydd Luff. Signed-off-by: David Drysdale <drysdale@google.com> Cc: Meredydd Luff <meredydd@senatehouse.org> Cc: Shuah Khan <shuah.kh@samsung.com> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Kees Cook <keescook@chromium.org> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Rich Felker <dalias@aerifal.cx> Cc: Christoph Hellwig <hch@infradead.org> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2014-12-13 08:57:29 +08:00
struct filename *
getname_flags(const char __user *filename, int flags)
{
struct filename *result;
char *kname;
int len;
result = audit_reusename(filename);
if (result)
return result;
result = __getname();
if (unlikely(!result))
return ERR_PTR(-ENOMEM);
/*
* First, try to embed the struct filename inside the names_cache
* allocation
*/
kname = (char *)result->iname;
result->name = kname;
len = strncpy_from_user(kname, filename, EMBEDDED_NAME_MAX);
/*
* Handle both empty path and copy failure in one go.
*/
if (unlikely(len <= 0)) {
if (unlikely(len < 0)) {
__putname(result);
return ERR_PTR(len);
}
/* The empty path is special. */
if (!(flags & LOOKUP_EMPTY)) {
__putname(result);
return ERR_PTR(-ENOENT);
}
}
/*
* Uh-oh. We have a name that's approaching PATH_MAX. Allocate a
* separate struct filename so we can dedicate the entire
* names_cache allocation for the pathname, and re-do the copy from
* userland.
*/
if (unlikely(len == EMBEDDED_NAME_MAX)) {
const size_t size = offsetof(struct filename, iname[1]);
kname = (char *)result;
/*
* size is chosen that way we to guarantee that
* result->iname[0] is within the same object and that
* kname can't be equal to result->iname, no matter what.
*/
result = kzalloc(size, GFP_KERNEL);
if (unlikely(!result)) {
__putname(kname);
return ERR_PTR(-ENOMEM);
}
result->name = kname;
len = strncpy_from_user(kname, filename, PATH_MAX);
if (unlikely(len < 0)) {
__putname(kname);
kfree(result);
return ERR_PTR(len);
}
/* The empty path is special. */
if (unlikely(!len) && !(flags & LOOKUP_EMPTY)) {
__putname(kname);
kfree(result);
return ERR_PTR(-ENOENT);
}
if (unlikely(len == PATH_MAX)) {
__putname(kname);
kfree(result);
return ERR_PTR(-ENAMETOOLONG);
}
}
audit,io_uring: io_uring openat triggers audit reference count underflow An io_uring openat operation can update an audit reference count from multiple threads resulting in the call trace below. A call to io_uring_submit() with a single openat op with a flag of IOSQE_ASYNC results in the following reference count updates. These first part of the system call performs two increments that do not race. do_syscall_64() __do_sys_io_uring_enter() io_submit_sqes() io_openat_prep() __io_openat_prep() getname() getname_flags() /* update 1 (increment) */ __audit_getname() /* update 2 (increment) */ The openat op is queued to an io_uring worker thread which starts the opportunity for a race. The system call exit performs one decrement. do_syscall_64() syscall_exit_to_user_mode() syscall_exit_to_user_mode_prepare() __audit_syscall_exit() audit_reset_context() putname() /* update 3 (decrement) */ The io_uring worker thread performs one increment and two decrements. These updates can race with the system call decrement. io_wqe_worker() io_worker_handle_work() io_wq_submit_work() io_issue_sqe() io_openat() io_openat2() do_filp_open() path_openat() __audit_inode() /* update 4 (increment) */ putname() /* update 5 (decrement) */ __audit_uring_exit() audit_reset_context() putname() /* update 6 (decrement) */ The fix is to change the refcnt member of struct audit_names from int to atomic_t. kernel BUG at fs/namei.c:262! Call Trace: ... ? putname+0x68/0x70 audit_reset_context.part.0.constprop.0+0xe1/0x300 __audit_uring_exit+0xda/0x1c0 io_issue_sqe+0x1f3/0x450 ? lock_timer_base+0x3b/0xd0 io_wq_submit_work+0x8d/0x2b0 ? __try_to_del_timer_sync+0x67/0xa0 io_worker_handle_work+0x17c/0x2b0 io_wqe_worker+0x10a/0x350 Cc: stable@vger.kernel.org Link: https://lore.kernel.org/lkml/MW2PR2101MB1033FFF044A258F84AEAA584F1C9A@MW2PR2101MB1033.namprd21.prod.outlook.com/ Fixes: 5bd2182d58e9 ("audit,io_uring,io-wq: add some basic audit support to io_uring") Signed-off-by: Dan Clash <daclash@linux.microsoft.com> Link: https://lore.kernel.org/r/20231012215518.GA4048@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net Reviewed-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Christian Brauner <brauner@kernel.org>
2023-10-13 05:55:18 +08:00
atomic_set(&result->refcnt, 1);
result->uptr = filename;
result->aname = NULL;
audit_getname(result);
return result;
}
struct filename *
getname_uflags(const char __user *filename, int uflags)
{
int flags = (uflags & AT_EMPTY_PATH) ? LOOKUP_EMPTY : 0;
return getname_flags(filename, flags);
}
struct filename *
getname(const char __user * filename)
{
return getname_flags(filename, 0);
}
struct filename *
getname_kernel(const char * filename)
{
struct filename *result;
int len = strlen(filename) + 1;
result = __getname();
if (unlikely(!result))
return ERR_PTR(-ENOMEM);
if (len <= EMBEDDED_NAME_MAX) {
result->name = (char *)result->iname;
} else if (len <= PATH_MAX) {
const size_t size = offsetof(struct filename, iname[1]);
struct filename *tmp;
tmp = kmalloc(size, GFP_KERNEL);
if (unlikely(!tmp)) {
__putname(result);
return ERR_PTR(-ENOMEM);
}
tmp->name = (char *)result;
result = tmp;
} else {
__putname(result);
return ERR_PTR(-ENAMETOOLONG);
}
memcpy((char *)result->name, filename, len);
result->uptr = NULL;
result->aname = NULL;
audit,io_uring: io_uring openat triggers audit reference count underflow An io_uring openat operation can update an audit reference count from multiple threads resulting in the call trace below. A call to io_uring_submit() with a single openat op with a flag of IOSQE_ASYNC results in the following reference count updates. These first part of the system call performs two increments that do not race. do_syscall_64() __do_sys_io_uring_enter() io_submit_sqes() io_openat_prep() __io_openat_prep() getname() getname_flags() /* update 1 (increment) */ __audit_getname() /* update 2 (increment) */ The openat op is queued to an io_uring worker thread which starts the opportunity for a race. The system call exit performs one decrement. do_syscall_64() syscall_exit_to_user_mode() syscall_exit_to_user_mode_prepare() __audit_syscall_exit() audit_reset_context() putname() /* update 3 (decrement) */ The io_uring worker thread performs one increment and two decrements. These updates can race with the system call decrement. io_wqe_worker() io_worker_handle_work() io_wq_submit_work() io_issue_sqe() io_openat() io_openat2() do_filp_open() path_openat() __audit_inode() /* update 4 (increment) */ putname() /* update 5 (decrement) */ __audit_uring_exit() audit_reset_context() putname() /* update 6 (decrement) */ The fix is to change the refcnt member of struct audit_names from int to atomic_t. kernel BUG at fs/namei.c:262! Call Trace: ... ? putname+0x68/0x70 audit_reset_context.part.0.constprop.0+0xe1/0x300 __audit_uring_exit+0xda/0x1c0 io_issue_sqe+0x1f3/0x450 ? lock_timer_base+0x3b/0xd0 io_wq_submit_work+0x8d/0x2b0 ? __try_to_del_timer_sync+0x67/0xa0 io_worker_handle_work+0x17c/0x2b0 io_wqe_worker+0x10a/0x350 Cc: stable@vger.kernel.org Link: https://lore.kernel.org/lkml/MW2PR2101MB1033FFF044A258F84AEAA584F1C9A@MW2PR2101MB1033.namprd21.prod.outlook.com/ Fixes: 5bd2182d58e9 ("audit,io_uring,io-wq: add some basic audit support to io_uring") Signed-off-by: Dan Clash <daclash@linux.microsoft.com> Link: https://lore.kernel.org/r/20231012215518.GA4048@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net Reviewed-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Christian Brauner <brauner@kernel.org>
2023-10-13 05:55:18 +08:00
atomic_set(&result->refcnt, 1);
audit_getname(result);
return result;
}
EXPORT_SYMBOL(getname_kernel);
void putname(struct filename *name)
{
if (IS_ERR(name))
return;
audit,io_uring: io_uring openat triggers audit reference count underflow An io_uring openat operation can update an audit reference count from multiple threads resulting in the call trace below. A call to io_uring_submit() with a single openat op with a flag of IOSQE_ASYNC results in the following reference count updates. These first part of the system call performs two increments that do not race. do_syscall_64() __do_sys_io_uring_enter() io_submit_sqes() io_openat_prep() __io_openat_prep() getname() getname_flags() /* update 1 (increment) */ __audit_getname() /* update 2 (increment) */ The openat op is queued to an io_uring worker thread which starts the opportunity for a race. The system call exit performs one decrement. do_syscall_64() syscall_exit_to_user_mode() syscall_exit_to_user_mode_prepare() __audit_syscall_exit() audit_reset_context() putname() /* update 3 (decrement) */ The io_uring worker thread performs one increment and two decrements. These updates can race with the system call decrement. io_wqe_worker() io_worker_handle_work() io_wq_submit_work() io_issue_sqe() io_openat() io_openat2() do_filp_open() path_openat() __audit_inode() /* update 4 (increment) */ putname() /* update 5 (decrement) */ __audit_uring_exit() audit_reset_context() putname() /* update 6 (decrement) */ The fix is to change the refcnt member of struct audit_names from int to atomic_t. kernel BUG at fs/namei.c:262! Call Trace: ... ? putname+0x68/0x70 audit_reset_context.part.0.constprop.0+0xe1/0x300 __audit_uring_exit+0xda/0x1c0 io_issue_sqe+0x1f3/0x450 ? lock_timer_base+0x3b/0xd0 io_wq_submit_work+0x8d/0x2b0 ? __try_to_del_timer_sync+0x67/0xa0 io_worker_handle_work+0x17c/0x2b0 io_wqe_worker+0x10a/0x350 Cc: stable@vger.kernel.org Link: https://lore.kernel.org/lkml/MW2PR2101MB1033FFF044A258F84AEAA584F1C9A@MW2PR2101MB1033.namprd21.prod.outlook.com/ Fixes: 5bd2182d58e9 ("audit,io_uring,io-wq: add some basic audit support to io_uring") Signed-off-by: Dan Clash <daclash@linux.microsoft.com> Link: https://lore.kernel.org/r/20231012215518.GA4048@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net Reviewed-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Christian Brauner <brauner@kernel.org>
2023-10-13 05:55:18 +08:00
if (WARN_ON_ONCE(!atomic_read(&name->refcnt)))
return;
audit,io_uring: io_uring openat triggers audit reference count underflow An io_uring openat operation can update an audit reference count from multiple threads resulting in the call trace below. A call to io_uring_submit() with a single openat op with a flag of IOSQE_ASYNC results in the following reference count updates. These first part of the system call performs two increments that do not race. do_syscall_64() __do_sys_io_uring_enter() io_submit_sqes() io_openat_prep() __io_openat_prep() getname() getname_flags() /* update 1 (increment) */ __audit_getname() /* update 2 (increment) */ The openat op is queued to an io_uring worker thread which starts the opportunity for a race. The system call exit performs one decrement. do_syscall_64() syscall_exit_to_user_mode() syscall_exit_to_user_mode_prepare() __audit_syscall_exit() audit_reset_context() putname() /* update 3 (decrement) */ The io_uring worker thread performs one increment and two decrements. These updates can race with the system call decrement. io_wqe_worker() io_worker_handle_work() io_wq_submit_work() io_issue_sqe() io_openat() io_openat2() do_filp_open() path_openat() __audit_inode() /* update 4 (increment) */ putname() /* update 5 (decrement) */ __audit_uring_exit() audit_reset_context() putname() /* update 6 (decrement) */ The fix is to change the refcnt member of struct audit_names from int to atomic_t. kernel BUG at fs/namei.c:262! Call Trace: ... ? putname+0x68/0x70 audit_reset_context.part.0.constprop.0+0xe1/0x300 __audit_uring_exit+0xda/0x1c0 io_issue_sqe+0x1f3/0x450 ? lock_timer_base+0x3b/0xd0 io_wq_submit_work+0x8d/0x2b0 ? __try_to_del_timer_sync+0x67/0xa0 io_worker_handle_work+0x17c/0x2b0 io_wqe_worker+0x10a/0x350 Cc: stable@vger.kernel.org Link: https://lore.kernel.org/lkml/MW2PR2101MB1033FFF044A258F84AEAA584F1C9A@MW2PR2101MB1033.namprd21.prod.outlook.com/ Fixes: 5bd2182d58e9 ("audit,io_uring,io-wq: add some basic audit support to io_uring") Signed-off-by: Dan Clash <daclash@linux.microsoft.com> Link: https://lore.kernel.org/r/20231012215518.GA4048@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net Reviewed-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Christian Brauner <brauner@kernel.org>
2023-10-13 05:55:18 +08:00
if (!atomic_dec_and_test(&name->refcnt))
return;
if (name->name != name->iname) {
__putname(name->name);
kfree(name);
} else
__putname(name);
}
EXPORT_SYMBOL(putname);
/**
* check_acl - perform ACL permission checking
* @idmap: idmap of the mount the inode was found from
* @inode: inode to check permissions on
* @mask: right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC ...)
*
* This function performs the ACL permission checking. Since this function
* retrieve POSIX acls it needs to know whether it is called from a blocking or
* non-blocking context and thus cares about the MAY_NOT_BLOCK bit.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
static int check_acl(struct mnt_idmap *idmap,
struct inode *inode, int mask)
{
#ifdef CONFIG_FS_POSIX_ACL
struct posix_acl *acl;
if (mask & MAY_NOT_BLOCK) {
acl = get_cached_acl_rcu(inode, ACL_TYPE_ACCESS);
if (!acl)
return -EAGAIN;
fs: rename current get acl method The current way of setting and getting posix acls through the generic xattr interface is error prone and type unsafe. The vfs needs to interpret and fixup posix acls before storing or reporting it to userspace. Various hacks exist to make this work. The code is hard to understand and difficult to maintain in it's current form. Instead of making this work by hacking posix acls through xattr handlers we are building a dedicated posix acl api around the get and set inode operations. This removes a lot of hackiness and makes the codepaths easier to maintain. A lot of background can be found in [1]. The current inode operation for getting posix acls takes an inode argument but various filesystems (e.g., 9p, cifs, overlayfs) need access to the dentry. In contrast to the ->set_acl() inode operation we cannot simply extend ->get_acl() to take a dentry argument. The ->get_acl() inode operation is called from: acl_permission_check() -> check_acl() -> get_acl() which is part of generic_permission() which in turn is part of inode_permission(). Both generic_permission() and inode_permission() are called in the ->permission() handler of various filesystems (e.g., overlayfs). So simply passing a dentry argument to ->get_acl() would amount to also having to pass a dentry argument to ->permission(). We should avoid this unnecessary change. So instead of extending the existing inode operation rename it from ->get_acl() to ->get_inode_acl() and add a ->get_acl() method later that passes a dentry argument and which filesystems that need access to the dentry can implement instead of ->get_inode_acl(). Filesystems like cifs which allow setting and getting posix acls but not using them for permission checking during lookup can simply not implement ->get_inode_acl(). This is intended to be a non-functional change. Link: https://lore.kernel.org/all/20220801145520.1532837-1-brauner@kernel.org [1] Suggested-by/Inspired-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-09-22 23:17:00 +08:00
/* no ->get_inode_acl() calls in RCU mode... */
posix_acl: Inode acl caching fixes When get_acl() is called for an inode whose ACL is not cached yet, the get_acl inode operation is called to fetch the ACL from the filesystem. The inode operation is responsible for updating the cached acl with set_cached_acl(). This is done without locking at the VFS level, so another task can call set_cached_acl() or forget_cached_acl() before the get_acl inode operation gets to calling set_cached_acl(), and then get_acl's call to set_cached_acl() results in caching an outdate ACL. Prevent this from happening by setting the cached ACL pointer to a task-specific sentinel value before calling the get_acl inode operation. Move the responsibility for updating the cached ACL from the get_acl inode operations to get_acl(). There, only set the cached ACL if the sentinel value hasn't changed. The sentinel values are chosen to have odd values. Likewise, the value of ACL_NOT_CACHED is odd. In contrast, ACL object pointers always have an even value (ACLs are aligned in memory). This allows to distinguish uncached ACLs values from ACL objects. In addition, switch from guarding inode->i_acl and inode->i_default_acl upates by the inode->i_lock spinlock to using xchg() and cmpxchg(). Filesystems that do not want ACLs returned from their get_acl inode operations to be cached must call forget_cached_acl() to prevent the VFS from doing so. (Patch written by Al Viro and Andreas Gruenbacher.) Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-03-24 21:38:37 +08:00
if (is_uncached_acl(acl))
return -ECHILD;
return posix_acl_permission(idmap, inode, acl, mask);
}
fs: rename current get acl method The current way of setting and getting posix acls through the generic xattr interface is error prone and type unsafe. The vfs needs to interpret and fixup posix acls before storing or reporting it to userspace. Various hacks exist to make this work. The code is hard to understand and difficult to maintain in it's current form. Instead of making this work by hacking posix acls through xattr handlers we are building a dedicated posix acl api around the get and set inode operations. This removes a lot of hackiness and makes the codepaths easier to maintain. A lot of background can be found in [1]. The current inode operation for getting posix acls takes an inode argument but various filesystems (e.g., 9p, cifs, overlayfs) need access to the dentry. In contrast to the ->set_acl() inode operation we cannot simply extend ->get_acl() to take a dentry argument. The ->get_acl() inode operation is called from: acl_permission_check() -> check_acl() -> get_acl() which is part of generic_permission() which in turn is part of inode_permission(). Both generic_permission() and inode_permission() are called in the ->permission() handler of various filesystems (e.g., overlayfs). So simply passing a dentry argument to ->get_acl() would amount to also having to pass a dentry argument to ->permission(). We should avoid this unnecessary change. So instead of extending the existing inode operation rename it from ->get_acl() to ->get_inode_acl() and add a ->get_acl() method later that passes a dentry argument and which filesystems that need access to the dentry can implement instead of ->get_inode_acl(). Filesystems like cifs which allow setting and getting posix acls but not using them for permission checking during lookup can simply not implement ->get_inode_acl(). This is intended to be a non-functional change. Link: https://lore.kernel.org/all/20220801145520.1532837-1-brauner@kernel.org [1] Suggested-by/Inspired-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-09-22 23:17:00 +08:00
acl = get_inode_acl(inode, ACL_TYPE_ACCESS);
if (IS_ERR(acl))
return PTR_ERR(acl);
if (acl) {
int error = posix_acl_permission(idmap, inode, acl, mask);
posix_acl_release(acl);
return error;
}
#endif
return -EAGAIN;
}
/**
* acl_permission_check - perform basic UNIX permission checking
* @idmap: idmap of the mount the inode was found from
* @inode: inode to check permissions on
* @mask: right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC ...)
*
* This function performs the basic UNIX permission checking. Since this
* function may retrieve POSIX acls it needs to know whether it is called from a
* blocking or non-blocking context and thus cares about the MAY_NOT_BLOCK bit.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
static int acl_permission_check(struct mnt_idmap *idmap,
struct inode *inode, int mask)
{
unsigned int mode = inode->i_mode;
vfsuid_t vfsuid;
/* Are we the owner? If so, ACL's don't matter */
vfsuid = i_uid_into_vfsuid(idmap, inode);
if (likely(vfsuid_eq_kuid(vfsuid, current_fsuid()))) {
mask &= 7;
mode >>= 6;
return (mask & ~mode) ? -EACCES : 0;
}
/* Do we have ACL's? */
if (IS_POSIXACL(inode) && (mode & S_IRWXG)) {
int error = check_acl(idmap, inode, mask);
if (error != -EAGAIN)
return error;
}
/* Only RWX matters for group/other mode bits */
mask &= 7;
/*
* Are the group permissions different from
* the other permissions in the bits we care
* about? Need to check group ownership if so.
*/
if (mask & (mode ^ (mode >> 3))) {
vfsgid_t vfsgid = i_gid_into_vfsgid(idmap, inode);
if (vfsgid_in_group_p(vfsgid))
mode >>= 3;
}
/* Bits in 'mode' clear that we require? */
return (mask & ~mode) ? -EACCES : 0;
}
/**
* generic_permission - check for access rights on a Posix-like filesystem
* @idmap: idmap of the mount the inode was found from
* @inode: inode to check access rights for
* @mask: right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC,
* %MAY_NOT_BLOCK ...)
*
* Used to check for read/write/execute permissions on a file.
* We use "fsuid" for this, letting us set arbitrary permissions
* for filesystem access without changing the "normal" uids which
* are used for other things.
*
* generic_permission is rcu-walk aware. It returns -ECHILD in case an rcu-walk
* request cannot be satisfied (eg. requires blocking or too much complexity).
* It would then be called again in ref-walk mode.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int generic_permission(struct mnt_idmap *idmap, struct inode *inode,
int mask)
{
int ret;
/*
* Do the basic permission checks.
*/
ret = acl_permission_check(idmap, inode, mask);
if (ret != -EACCES)
return ret;
if (S_ISDIR(inode->i_mode)) {
/* DACs are overridable for directories */
if (!(mask & MAY_WRITE))
if (capable_wrt_inode_uidgid(idmap, inode,
CAP_DAC_READ_SEARCH))
return 0;
if (capable_wrt_inode_uidgid(idmap, inode,
CAP_DAC_OVERRIDE))
return 0;
return -EACCES;
}
/*
* Searching includes executable on directories, else just read.
*/
mask &= MAY_READ | MAY_WRITE | MAY_EXEC;
if (mask == MAY_READ)
if (capable_wrt_inode_uidgid(idmap, inode,
CAP_DAC_READ_SEARCH))
return 0;
/*
* Read/write DACs are always overridable.
* Executable DACs are overridable when there is
* at least one exec bit set.
*/
if (!(mask & MAY_EXEC) || (inode->i_mode & S_IXUGO))
if (capable_wrt_inode_uidgid(idmap, inode,
CAP_DAC_OVERRIDE))
return 0;
return -EACCES;
}
EXPORT_SYMBOL(generic_permission);
/**
* do_inode_permission - UNIX permission checking
* @idmap: idmap of the mount the inode was found from
* @inode: inode to check permissions on
* @mask: right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC ...)
*
* We _really_ want to just do "generic_permission()" without
* even looking at the inode->i_op values. So we keep a cache
* flag in inode->i_opflags, that says "this has not special
* permission function, use the fast case".
*/
static inline int do_inode_permission(struct mnt_idmap *idmap,
struct inode *inode, int mask)
{
if (unlikely(!(inode->i_opflags & IOP_FASTPERM))) {
if (likely(inode->i_op->permission))
return inode->i_op->permission(idmap, inode, mask);
/* This gets set once for the inode lifetime */
spin_lock(&inode->i_lock);
inode->i_opflags |= IOP_FASTPERM;
spin_unlock(&inode->i_lock);
}
return generic_permission(idmap, inode, mask);
}
/**
* sb_permission - Check superblock-level permissions
* @sb: Superblock of inode to check permission on
* @inode: Inode to check permission on
* @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
*
* Separate out file-system wide checks from inode-specific permission checks.
*/
static int sb_permission(struct super_block *sb, struct inode *inode, int mask)
{
if (unlikely(mask & MAY_WRITE)) {
umode_t mode = inode->i_mode;
/* Nobody gets write access to a read-only fs. */
if (sb_rdonly(sb) && (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode)))
return -EROFS;
}
return 0;
}
/**
* inode_permission - Check for access rights to a given inode
* @idmap: idmap of the mount the inode was found from
* @inode: Inode to check permission on
* @mask: Right to check for (%MAY_READ, %MAY_WRITE, %MAY_EXEC)
*
* Check for read/write/execute permissions on an inode. We use fs[ug]id for
* this, letting us set arbitrary permissions for filesystem access without
* changing the "normal" UIDs which are used for other things.
*
* When checking for MAY_APPEND, MAY_WRITE must also be set in @mask.
*/
int inode_permission(struct mnt_idmap *idmap,
struct inode *inode, int mask)
{
int retval;
retval = sb_permission(inode->i_sb, inode, mask);
if (retval)
return retval;
if (unlikely(mask & MAY_WRITE)) {
/*
* Nobody gets write access to an immutable file.
*/
if (IS_IMMUTABLE(inode))
return -EPERM;
/*
* Updating mtime will likely cause i_uid and i_gid to be
* written back improperly if their true value is unknown
* to the vfs.
*/
if (HAS_UNMAPPED_ID(idmap, inode))
return -EACCES;
}
retval = do_inode_permission(idmap, inode, mask);
if (retval)
return retval;
retval = devcgroup_inode_permission(inode, mask);
if (retval)
return retval;
return security_inode_permission(inode, mask);
}
EXPORT_SYMBOL(inode_permission);
/**
* path_get - get a reference to a path
* @path: path to get the reference to
*
* Given a path increment the reference count to the dentry and the vfsmount.
*/
void path_get(const struct path *path)
{
mntget(path->mnt);
dget(path->dentry);
}
EXPORT_SYMBOL(path_get);
/**
* path_put - put a reference to a path
* @path: path to put the reference to
*
* Given a path decrement the reference count to the dentry and the vfsmount.
*/
void path_put(const struct path *path)
{
dput(path->dentry);
mntput(path->mnt);
}
EXPORT_SYMBOL(path_put);
#define EMBEDDED_LEVELS 2
struct nameidata {
struct path path;
struct qstr last;
struct path root;
struct inode *inode; /* path.dentry.d_inode */
unsigned int flags, state;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
unsigned seq, next_seq, m_seq, r_seq;
int last_type;
unsigned depth;
int total_link_count;
struct saved {
struct path link;
struct delayed_call done;
const char *name;
unsigned seq;
} *stack, internal[EMBEDDED_LEVELS];
struct filename *name;
struct nameidata *saved;
unsigned root_seq;
int dfd;
vfsuid_t dir_vfsuid;
umode_t dir_mode;
} __randomize_layout;
#define ND_ROOT_PRESET 1
#define ND_ROOT_GRABBED 2
#define ND_JUMPED 4
static void __set_nameidata(struct nameidata *p, int dfd, struct filename *name)
{
struct nameidata *old = current->nameidata;
p->stack = p->internal;
p->depth = 0;
p->dfd = dfd;
p->name = name;
p->path.mnt = NULL;
p->path.dentry = NULL;
p->total_link_count = old ? old->total_link_count : 0;
p->saved = old;
current->nameidata = p;
}
static inline void set_nameidata(struct nameidata *p, int dfd, struct filename *name,
const struct path *root)
{
__set_nameidata(p, dfd, name);
p->state = 0;
if (unlikely(root)) {
p->state = ND_ROOT_PRESET;
p->root = *root;
}
}
static void restore_nameidata(void)
{
struct nameidata *now = current->nameidata, *old = now->saved;
current->nameidata = old;
if (old)
old->total_link_count = now->total_link_count;
if (now->stack != now->internal)
kfree(now->stack);
}
static bool nd_alloc_stack(struct nameidata *nd)
{
struct saved *p;
p= kmalloc_array(MAXSYMLINKS, sizeof(struct saved),
nd->flags & LOOKUP_RCU ? GFP_ATOMIC : GFP_KERNEL);
if (unlikely(!p))
return false;
memcpy(p, nd->internal, sizeof(nd->internal));
nd->stack = p;
return true;
}
/**
* path_connected - Verify that a dentry is below mnt.mnt_root
* @mnt: The mountpoint to check.
* @dentry: The dentry to check.
*
* Rename can sometimes move a file or directory outside of a bind
* mount, path_connected allows those cases to be detected.
*/
static bool path_connected(struct vfsmount *mnt, struct dentry *dentry)
{
fs: Teach path_connected to handle nfs filesystems with multiple roots. On nfsv2 and nfsv3 the nfs server can export subsets of the same filesystem and report the same filesystem identifier, so that the nfs client can know they are the same filesystem. The subsets can be from disjoint directory trees. The nfsv2 and nfsv3 filesystems provides no way to find the common root of all directory trees exported form the server with the same filesystem identifier. The practical result is that in struct super s_root for nfs s_root is not necessarily the root of the filesystem. The nfs mount code sets s_root to the root of the first subset of the nfs filesystem that the kernel mounts. This effects the dcache invalidation code in generic_shutdown_super currently called shrunk_dcache_for_umount and that code for years has gone through an additional list of dentries that might be dentry trees that need to be freed to accomodate nfs. When I wrote path_connected I did not realize nfs was so special, and it's hueristic for avoiding calling is_subdir can fail. The practical case where this fails is when there is a move of a directory from the subtree exposed by one nfs mount to the subtree exposed by another nfs mount. This move can happen either locally or remotely. With the remote case requiring that the move directory be cached before the move and that after the move someone walks the path to where the move directory now exists and in so doing causes the already cached directory to be moved in the dcache through the magic of d_splice_alias. If someone whose working directory is in the move directory or a subdirectory and now starts calling .. from the initial mount of nfs (where s_root == mnt_root), then path_connected as a heuristic will not bother with the is_subdir check. As s_root really is not the root of the nfs filesystem this heuristic is wrong, and the path may actually not be connected and path_connected can fail. The is_subdir function might be cheap enough that we can call it unconditionally. Verifying that will take some benchmarking and the result may not be the same on all kernels this fix needs to be backported to. So I am avoiding that for now. Filesystems with snapshots such as nilfs and btrfs do something similar. But as the directory tree of the snapshots are disjoint from one another and from the main directory tree rename won't move things between them and this problem will not occur. Cc: stable@vger.kernel.org Reported-by: Al Viro <viro@ZenIV.linux.org.uk> Fixes: 397d425dc26d ("vfs: Test for and handle paths that are unreachable from their mnt_root") Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-03-15 07:20:29 +08:00
struct super_block *sb = mnt->mnt_sb;
/* Bind mounts can have disconnected paths */
if (mnt->mnt_root == sb->s_root)
return true;
return is_subdir(dentry, mnt->mnt_root);
}
static void drop_links(struct nameidata *nd)
{
int i = nd->depth;
while (i--) {
struct saved *last = nd->stack + i;
do_delayed_call(&last->done);
clear_delayed_call(&last->done);
}
}
static void leave_rcu(struct nameidata *nd)
{
nd->flags &= ~LOOKUP_RCU;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->seq = nd->next_seq = 0;
rcu_read_unlock();
}
static void terminate_walk(struct nameidata *nd)
{
drop_links(nd);
if (!(nd->flags & LOOKUP_RCU)) {
int i;
path_put(&nd->path);
for (i = 0; i < nd->depth; i++)
path_put(&nd->stack[i].link);
if (nd->state & ND_ROOT_GRABBED) {
path_put(&nd->root);
nd->state &= ~ND_ROOT_GRABBED;
}
} else {
leave_rcu(nd);
}
nd->depth = 0;
nd->path.mnt = NULL;
nd->path.dentry = NULL;
}
/* path_put is needed afterwards regardless of success or failure */
static bool __legitimize_path(struct path *path, unsigned seq, unsigned mseq)
{
int res = __legitimize_mnt(path->mnt, mseq);
if (unlikely(res)) {
if (res > 0)
path->mnt = NULL;
path->dentry = NULL;
return false;
}
if (unlikely(!lockref_get_not_dead(&path->dentry->d_lockref))) {
path->dentry = NULL;
return false;
}
return !read_seqcount_retry(&path->dentry->d_seq, seq);
}
static inline bool legitimize_path(struct nameidata *nd,
struct path *path, unsigned seq)
{
return __legitimize_path(path, seq, nd->m_seq);
}
static bool legitimize_links(struct nameidata *nd)
{
int i;
if (unlikely(nd->flags & LOOKUP_CACHED)) {
drop_links(nd);
nd->depth = 0;
return false;
}
for (i = 0; i < nd->depth; i++) {
struct saved *last = nd->stack + i;
if (unlikely(!legitimize_path(nd, &last->link, last->seq))) {
drop_links(nd);
nd->depth = i + 1;
return false;
}
}
return true;
}
static bool legitimize_root(struct nameidata *nd)
{
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
/* Nothing to do if nd->root is zero or is managed by the VFS user. */
if (!nd->root.mnt || (nd->state & ND_ROOT_PRESET))
return true;
nd->state |= ND_ROOT_GRABBED;
return legitimize_path(nd, &nd->root, nd->root_seq);
}
/*
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
* Path walking has 2 modes, rcu-walk and ref-walk (see
* Documentation/filesystems/path-lookup.txt). In situations when we can't
* continue in RCU mode, we attempt to drop out of rcu-walk mode and grab
* normal reference counts on dentries and vfsmounts to transition to ref-walk
* mode. Refcounts are grabbed at the last known good point before rcu-walk
* got stuck, so ref-walk may continue from there. If this is not successful
* (eg. a seqcount has changed), then failure is returned and it's up to caller
* to restart the path walk from the beginning in ref-walk mode.
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
*/
/**
* try_to_unlazy - try to switch to ref-walk mode.
* @nd: nameidata pathwalk data
* Returns: true on success, false on failure
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
*
* try_to_unlazy attempts to legitimize the current nd->path and nd->root
* for ref-walk mode.
* Must be called from rcu-walk context.
* Nothing should touch nameidata between try_to_unlazy() failure and
* terminate_walk().
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
*/
static bool try_to_unlazy(struct nameidata *nd)
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
{
struct dentry *parent = nd->path.dentry;
BUG_ON(!(nd->flags & LOOKUP_RCU));
vfs: fix dentry RCU to refcounting possibly sleeping dput() This is the fix that the last two commits indirectly led up to - making sure that we don't call dput() in a bad context on the dentries we've looked up in RCU mode after the sequence count validation fails. This basically expands d_rcu_to_refcount() into the callers, and then fixes the callers to delay the dput() in the failure case until _after_ we've dropped all locks and are no longer in an RCU-locked region. The case of 'complete_walk()' was trivial, since its failure case did the unlock_rcu_walk() directly after the call to d_rcu_to_refcount(), and as such that is just a pure expansion of the function with a trivial movement of the resulting dput() to after 'unlock_rcu_walk()'. In contrast, the unlazy_walk() case was much more complicated, because not only does convert two different dentries from RCU to be reference counted, but it used to not call unlock_rcu_walk() at all, and instead just returned an error and let the caller clean everything up in "terminate_walk()". Happily, one of the dentries in question (called "parent" inside unlazy_walk()) is the dentry of "nd->path", which terminate_walk() wants a refcount to anyway for the non-RCU case. So what the new and improved unlazy_walk() does is to first turn that dentry into a refcounted one, and once that is set up, the error cases can continue to use the terminate_walk() helper for cleanup, but for the non-RCU case. Which makes it possible to drop out of RCU mode if we actually hit the sequence number failure case. Acked-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-09 09:13:49 +08:00
if (unlikely(!legitimize_links(nd)))
goto out1;
if (unlikely(!legitimize_path(nd, &nd->path, nd->seq)))
goto out;
if (unlikely(!legitimize_root(nd)))
goto out;
leave_rcu(nd);
BUG_ON(nd->inode != parent->d_inode);
return true;
out1:
nd->path.mnt = NULL;
nd->path.dentry = NULL;
out:
leave_rcu(nd);
return false;
}
/**
* try_to_unlazy_next - try to switch to ref-walk mode.
* @nd: nameidata pathwalk data
* @dentry: next dentry to step into
* Returns: true on success, false on failure
*
* Similar to try_to_unlazy(), but here we have the next dentry already
* picked by rcu-walk and want to legitimize that in addition to the current
* nd->path and nd->root for ref-walk mode. Must be called from rcu-walk context.
* Nothing should touch nameidata between try_to_unlazy_next() failure and
* terminate_walk().
*/
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
static bool try_to_unlazy_next(struct nameidata *nd, struct dentry *dentry)
{
int res;
BUG_ON(!(nd->flags & LOOKUP_RCU));
if (unlikely(!legitimize_links(nd)))
goto out2;
res = __legitimize_mnt(nd->path.mnt, nd->m_seq);
if (unlikely(res)) {
if (res > 0)
goto out2;
goto out1;
}
if (unlikely(!lockref_get_not_dead(&nd->path.dentry->d_lockref)))
goto out1;
RCU'd vfsmounts * RCU-delayed freeing of vfsmounts * vfsmount_lock replaced with a seqlock (mount_lock) * sequence number from mount_lock is stored in nameidata->m_seq and used when we exit RCU mode * new vfsmount flag - MNT_SYNC_UMOUNT. Set by umount_tree() when its caller knows that vfsmount will have no surviving references. * synchronize_rcu() done between unlocking namespace_sem in namespace_unlock() and doing pending mntput(). * new helper: legitimize_mnt(mnt, seq). Checks the mount_lock sequence number against seq, then grabs reference to mnt. Then it rechecks mount_lock again to close the race and either returns success or drops the reference it has acquired. The subtle point is that in case of MNT_SYNC_UMOUNT we can simply decrement the refcount and sod off - aforementioned synchronize_rcu() makes sure that final mntput() won't come until we leave RCU mode. We need that, since we don't want to end up with some lazy pathwalk racing with umount() and stealing the final mntput() from it - caller of umount() may expect it to return only once the fs is shut down and we don't want to break that. In other cases (i.e. with MNT_SYNC_UMOUNT absent) we have to do full-blown mntput() in case of mount_lock sequence number mismatch happening just as we'd grabbed the reference, but in those cases we won't be stealing the final mntput() from anything that would care. * mntput_no_expire() doesn't lock anything on the fast path now. Incidentally, SMP and UP cases are handled the same way - no ifdefs there. * normal pathname resolution does *not* do any writes to mount_lock. It does, of course, bump the refcounts of vfsmount and dentry in the very end, but that's it. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-09-30 10:06:07 +08:00
/*
* We need to move both the parent and the dentry from the RCU domain
* to be properly refcounted. And the sequence number in the dentry
* validates *both* dentry counters, since we checked the sequence
* number of the parent after we got the child sequence number. So we
* know the parent must still be valid if the child sequence number is
*/
if (unlikely(!lockref_get_not_dead(&dentry->d_lockref)))
goto out;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
if (read_seqcount_retry(&dentry->d_seq, nd->next_seq))
goto out_dput;
vfs: fix dentry RCU to refcounting possibly sleeping dput() This is the fix that the last two commits indirectly led up to - making sure that we don't call dput() in a bad context on the dentries we've looked up in RCU mode after the sequence count validation fails. This basically expands d_rcu_to_refcount() into the callers, and then fixes the callers to delay the dput() in the failure case until _after_ we've dropped all locks and are no longer in an RCU-locked region. The case of 'complete_walk()' was trivial, since its failure case did the unlock_rcu_walk() directly after the call to d_rcu_to_refcount(), and as such that is just a pure expansion of the function with a trivial movement of the resulting dput() to after 'unlock_rcu_walk()'. In contrast, the unlazy_walk() case was much more complicated, because not only does convert two different dentries from RCU to be reference counted, but it used to not call unlock_rcu_walk() at all, and instead just returned an error and let the caller clean everything up in "terminate_walk()". Happily, one of the dentries in question (called "parent" inside unlazy_walk()) is the dentry of "nd->path", which terminate_walk() wants a refcount to anyway for the non-RCU case. So what the new and improved unlazy_walk() does is to first turn that dentry into a refcounted one, and once that is set up, the error cases can continue to use the terminate_walk() helper for cleanup, but for the non-RCU case. Which makes it possible to drop out of RCU mode if we actually hit the sequence number failure case. Acked-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-09 09:13:49 +08:00
/*
* Sequence counts matched. Now make sure that the root is
* still valid and get it if required.
*/
if (unlikely(!legitimize_root(nd)))
goto out_dput;
leave_rcu(nd);
return true;
out2:
nd->path.mnt = NULL;
out1:
nd->path.dentry = NULL;
vfs: fix dentry RCU to refcounting possibly sleeping dput() This is the fix that the last two commits indirectly led up to - making sure that we don't call dput() in a bad context on the dentries we've looked up in RCU mode after the sequence count validation fails. This basically expands d_rcu_to_refcount() into the callers, and then fixes the callers to delay the dput() in the failure case until _after_ we've dropped all locks and are no longer in an RCU-locked region. The case of 'complete_walk()' was trivial, since its failure case did the unlock_rcu_walk() directly after the call to d_rcu_to_refcount(), and as such that is just a pure expansion of the function with a trivial movement of the resulting dput() to after 'unlock_rcu_walk()'. In contrast, the unlazy_walk() case was much more complicated, because not only does convert two different dentries from RCU to be reference counted, but it used to not call unlock_rcu_walk() at all, and instead just returned an error and let the caller clean everything up in "terminate_walk()". Happily, one of the dentries in question (called "parent" inside unlazy_walk()) is the dentry of "nd->path", which terminate_walk() wants a refcount to anyway for the non-RCU case. So what the new and improved unlazy_walk() does is to first turn that dentry into a refcounted one, and once that is set up, the error cases can continue to use the terminate_walk() helper for cleanup, but for the non-RCU case. Which makes it possible to drop out of RCU mode if we actually hit the sequence number failure case. Acked-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-09 09:13:49 +08:00
out:
leave_rcu(nd);
return false;
out_dput:
leave_rcu(nd);
dput(dentry);
return false;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
}
static inline int d_revalidate(struct dentry *dentry, unsigned int flags)
{
if (unlikely(dentry->d_flags & DCACHE_OP_REVALIDATE))
return dentry->d_op->d_revalidate(dentry, flags);
else
return 1;
}
/**
* complete_walk - successful completion of path walk
* @nd: pointer nameidata
*
* If we had been in RCU mode, drop out of it and legitimize nd->path.
* Revalidate the final result, unless we'd already done that during
* the path walk or the filesystem doesn't ask for it. Return 0 on
* success, -error on failure. In case of failure caller does not
* need to drop nd->path.
*/
static int complete_walk(struct nameidata *nd)
{
struct dentry *dentry = nd->path.dentry;
int status;
if (nd->flags & LOOKUP_RCU) {
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
/*
* We don't want to zero nd->root for scoped-lookups or
* externally-managed nd->root.
*/
if (!(nd->state & ND_ROOT_PRESET))
if (!(nd->flags & LOOKUP_IS_SCOPED))
nd->root.mnt = NULL;
nd->flags &= ~LOOKUP_CACHED;
if (!try_to_unlazy(nd))
return -ECHILD;
}
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
if (unlikely(nd->flags & LOOKUP_IS_SCOPED)) {
/*
* While the guarantee of LOOKUP_IS_SCOPED is (roughly) "don't
* ever step outside the root during lookup" and should already
* be guaranteed by the rest of namei, we want to avoid a namei
* BUG resulting in userspace being given a path that was not
* scoped within the root at some point during the lookup.
*
* So, do a final sanity-check to make sure that in the
* worst-case scenario (a complete bypass of LOOKUP_IS_SCOPED)
* we won't silently return an fd completely outside of the
* requested root to userspace.
*
* Userspace could move the path outside the root after this
* check, but as discussed elsewhere this is not a concern (the
* resolved file was inside the root at some point).
*/
if (!path_is_under(&nd->path, &nd->root))
return -EXDEV;
}
if (likely(!(nd->state & ND_JUMPED)))
return 0;
vfs: kill FS_REVAL_DOT by adding a d_weak_revalidate dentry op The following set of operations on a NFS client and server will cause server# mkdir a client# cd a server# mv a a.bak client# sleep 30 # (or whatever the dir attrcache timeout is) client# stat . stat: cannot stat `.': Stale NFS file handle Obviously, we should not be getting an ESTALE error back there since the inode still exists on the server. The problem is that the lookup code will call d_revalidate on the dentry that "." refers to, because NFS has FS_REVAL_DOT set. nfs_lookup_revalidate will see that the parent directory has changed and will try to reverify the dentry by redoing a LOOKUP. That of course fails, so the lookup code returns ESTALE. The problem here is that d_revalidate is really a bad fit for this case. What we really want to know at this point is whether the inode is still good or not, but we don't really care what name it goes by or whether the dcache is still valid. Add a new d_op->d_weak_revalidate operation and have complete_walk call that instead of d_revalidate. The intent there is to allow for a "weaker" d_revalidate that just checks to see whether the inode is still good. This is also gives us an opportunity to kill off the FS_REVAL_DOT special casing. [AV: changed method name, added note in porting, fixed confusion re having it possibly called from RCU mode (it won't be)] Cc: NeilBrown <neilb@suse.de> Signed-off-by: Jeff Layton <jlayton@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-02-21 00:19:05 +08:00
if (likely(!(dentry->d_flags & DCACHE_OP_WEAK_REVALIDATE)))
return 0;
vfs: kill FS_REVAL_DOT by adding a d_weak_revalidate dentry op The following set of operations on a NFS client and server will cause server# mkdir a client# cd a server# mv a a.bak client# sleep 30 # (or whatever the dir attrcache timeout is) client# stat . stat: cannot stat `.': Stale NFS file handle Obviously, we should not be getting an ESTALE error back there since the inode still exists on the server. The problem is that the lookup code will call d_revalidate on the dentry that "." refers to, because NFS has FS_REVAL_DOT set. nfs_lookup_revalidate will see that the parent directory has changed and will try to reverify the dentry by redoing a LOOKUP. That of course fails, so the lookup code returns ESTALE. The problem here is that d_revalidate is really a bad fit for this case. What we really want to know at this point is whether the inode is still good or not, but we don't really care what name it goes by or whether the dcache is still valid. Add a new d_op->d_weak_revalidate operation and have complete_walk call that instead of d_revalidate. The intent there is to allow for a "weaker" d_revalidate that just checks to see whether the inode is still good. This is also gives us an opportunity to kill off the FS_REVAL_DOT special casing. [AV: changed method name, added note in porting, fixed confusion re having it possibly called from RCU mode (it won't be)] Cc: NeilBrown <neilb@suse.de> Signed-off-by: Jeff Layton <jlayton@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-02-21 00:19:05 +08:00
status = dentry->d_op->d_weak_revalidate(dentry, nd->flags);
if (status > 0)
return 0;
if (!status)
status = -ESTALE;
return status;
}
static int set_root(struct nameidata *nd)
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
{
struct fs_struct *fs = current->fs;
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
/*
* Jumping to the real root in a scoped-lookup is a BUG in namei, but we
* still have to ensure it doesn't happen because it will cause a breakout
* from the dirfd.
*/
if (WARN_ON(nd->flags & LOOKUP_IS_SCOPED))
return -ENOTRECOVERABLE;
if (nd->flags & LOOKUP_RCU) {
unsigned seq;
do {
seq = read_seqcount_begin(&fs->seq);
nd->root = fs->root;
nd->root_seq = __read_seqcount_begin(&nd->root.dentry->d_seq);
} while (read_seqcount_retry(&fs->seq, seq));
} else {
get_fs_root(fs, &nd->root);
nd->state |= ND_ROOT_GRABBED;
}
return 0;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
}
static int nd_jump_root(struct nameidata *nd)
{
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
if (unlikely(nd->flags & LOOKUP_BENEATH))
return -EXDEV;
namei: LOOKUP_NO_XDEV: block mountpoint crossing /* Background. */ The need to contain path operations within a mountpoint has been a long-standing usecase that userspace has historically implemented manually with liberal usage of stat(). find, rsync, tar and many other programs implement these semantics -- but it'd be much simpler to have a fool-proof way of refusing to open a path if it crosses a mountpoint. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_XDEV will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_XDEV applies to all components of the path. With LOOKUP_NO_XDEV, any path component which crosses a mount-point during path resolution (including "..") will yield an -EXDEV. Absolute paths, absolute symlinks, and magic-links will only yield an -EXDEV if the jump involved changing mount-points. /* Testing. */ LOOKUP_NO_XDEV is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:32 +08:00
if (unlikely(nd->flags & LOOKUP_NO_XDEV)) {
/* Absolute path arguments to path_init() are allowed. */
if (nd->path.mnt != NULL && nd->path.mnt != nd->root.mnt)
return -EXDEV;
}
if (!nd->root.mnt) {
int error = set_root(nd);
if (error)
return error;
}
if (nd->flags & LOOKUP_RCU) {
struct dentry *d;
nd->path = nd->root;
d = nd->path.dentry;
nd->inode = d->d_inode;
nd->seq = nd->root_seq;
if (read_seqcount_retry(&d->d_seq, nd->seq))
return -ECHILD;
} else {
path_put(&nd->path);
nd->path = nd->root;
path_get(&nd->path);
nd->inode = nd->path.dentry->d_inode;
}
nd->state |= ND_JUMPED;
return 0;
}
/*
* Helper to directly jump to a known parsed path from ->get_link,
* caller must have taken a reference to path beforehand.
*/
int nd_jump_link(const struct path *path)
{
namei: LOOKUP_NO_MAGICLINKS: block magic-link resolution /* Background. */ There has always been a special class of symlink-like objects in procfs (and a few other pseudo-filesystems) which allow for non-lexical resolution of paths using nd_jump_link(). These "magic-links" do not follow traditional mount namespace boundaries, and have been used consistently in container escape attacks because they can be used to trick unsuspecting privileged processes into resolving unexpected paths. It is also non-trivial for userspace to unambiguously avoid resolving magic-links, because they do not have a reliable indication that they are a magic-link (in order to verify them you'd have to manually open the path given by readlink(2) and then verify that the two file descriptors reference the same underlying file, which is plagued with possible race conditions or supplementary attack scenarios). It would therefore be very helpful for userspace to be able to avoid these symlinks easily, thus hopefully removing a tool from attackers' toolboxes. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_MAGICLINKS will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_MAGICLINKS applies to all components of the path. With LOOKUP_NO_MAGICLINKS, any magic-link path component encountered during path resolution will yield -ELOOP. The handling of ~LOOKUP_FOLLOW for a trailing magic-link is identical to LOOKUP_NO_SYMLINKS. LOOKUP_NO_SYMLINKS implies LOOKUP_NO_MAGICLINKS. /* Testing. */ LOOKUP_NO_MAGICLINKS is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:31 +08:00
int error = -ELOOP;
struct nameidata *nd = current->nameidata;
namei: LOOKUP_NO_MAGICLINKS: block magic-link resolution /* Background. */ There has always been a special class of symlink-like objects in procfs (and a few other pseudo-filesystems) which allow for non-lexical resolution of paths using nd_jump_link(). These "magic-links" do not follow traditional mount namespace boundaries, and have been used consistently in container escape attacks because they can be used to trick unsuspecting privileged processes into resolving unexpected paths. It is also non-trivial for userspace to unambiguously avoid resolving magic-links, because they do not have a reliable indication that they are a magic-link (in order to verify them you'd have to manually open the path given by readlink(2) and then verify that the two file descriptors reference the same underlying file, which is plagued with possible race conditions or supplementary attack scenarios). It would therefore be very helpful for userspace to be able to avoid these symlinks easily, thus hopefully removing a tool from attackers' toolboxes. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_MAGICLINKS will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_MAGICLINKS applies to all components of the path. With LOOKUP_NO_MAGICLINKS, any magic-link path component encountered during path resolution will yield -ELOOP. The handling of ~LOOKUP_FOLLOW for a trailing magic-link is identical to LOOKUP_NO_SYMLINKS. LOOKUP_NO_SYMLINKS implies LOOKUP_NO_MAGICLINKS. /* Testing. */ LOOKUP_NO_MAGICLINKS is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:31 +08:00
if (unlikely(nd->flags & LOOKUP_NO_MAGICLINKS))
goto err;
namei: LOOKUP_NO_XDEV: block mountpoint crossing /* Background. */ The need to contain path operations within a mountpoint has been a long-standing usecase that userspace has historically implemented manually with liberal usage of stat(). find, rsync, tar and many other programs implement these semantics -- but it'd be much simpler to have a fool-proof way of refusing to open a path if it crosses a mountpoint. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_XDEV will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_XDEV applies to all components of the path. With LOOKUP_NO_XDEV, any path component which crosses a mount-point during path resolution (including "..") will yield an -EXDEV. Absolute paths, absolute symlinks, and magic-links will only yield an -EXDEV if the jump involved changing mount-points. /* Testing. */ LOOKUP_NO_XDEV is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:32 +08:00
error = -EXDEV;
if (unlikely(nd->flags & LOOKUP_NO_XDEV)) {
if (nd->path.mnt != path->mnt)
goto err;
}
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
/* Not currently safe for scoped-lookups. */
if (unlikely(nd->flags & LOOKUP_IS_SCOPED))
goto err;
namei: LOOKUP_NO_XDEV: block mountpoint crossing /* Background. */ The need to contain path operations within a mountpoint has been a long-standing usecase that userspace has historically implemented manually with liberal usage of stat(). find, rsync, tar and many other programs implement these semantics -- but it'd be much simpler to have a fool-proof way of refusing to open a path if it crosses a mountpoint. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_XDEV will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_XDEV applies to all components of the path. With LOOKUP_NO_XDEV, any path component which crosses a mount-point during path resolution (including "..") will yield an -EXDEV. Absolute paths, absolute symlinks, and magic-links will only yield an -EXDEV if the jump involved changing mount-points. /* Testing. */ LOOKUP_NO_XDEV is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:32 +08:00
namei: LOOKUP_NO_MAGICLINKS: block magic-link resolution /* Background. */ There has always been a special class of symlink-like objects in procfs (and a few other pseudo-filesystems) which allow for non-lexical resolution of paths using nd_jump_link(). These "magic-links" do not follow traditional mount namespace boundaries, and have been used consistently in container escape attacks because they can be used to trick unsuspecting privileged processes into resolving unexpected paths. It is also non-trivial for userspace to unambiguously avoid resolving magic-links, because they do not have a reliable indication that they are a magic-link (in order to verify them you'd have to manually open the path given by readlink(2) and then verify that the two file descriptors reference the same underlying file, which is plagued with possible race conditions or supplementary attack scenarios). It would therefore be very helpful for userspace to be able to avoid these symlinks easily, thus hopefully removing a tool from attackers' toolboxes. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_MAGICLINKS will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_MAGICLINKS applies to all components of the path. With LOOKUP_NO_MAGICLINKS, any magic-link path component encountered during path resolution will yield -ELOOP. The handling of ~LOOKUP_FOLLOW for a trailing magic-link is identical to LOOKUP_NO_SYMLINKS. LOOKUP_NO_SYMLINKS implies LOOKUP_NO_MAGICLINKS. /* Testing. */ LOOKUP_NO_MAGICLINKS is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:31 +08:00
path_put(&nd->path);
nd->path = *path;
nd->inode = nd->path.dentry->d_inode;
nd->state |= ND_JUMPED;
return 0;
namei: LOOKUP_NO_MAGICLINKS: block magic-link resolution /* Background. */ There has always been a special class of symlink-like objects in procfs (and a few other pseudo-filesystems) which allow for non-lexical resolution of paths using nd_jump_link(). These "magic-links" do not follow traditional mount namespace boundaries, and have been used consistently in container escape attacks because they can be used to trick unsuspecting privileged processes into resolving unexpected paths. It is also non-trivial for userspace to unambiguously avoid resolving magic-links, because they do not have a reliable indication that they are a magic-link (in order to verify them you'd have to manually open the path given by readlink(2) and then verify that the two file descriptors reference the same underlying file, which is plagued with possible race conditions or supplementary attack scenarios). It would therefore be very helpful for userspace to be able to avoid these symlinks easily, thus hopefully removing a tool from attackers' toolboxes. This is part of a refresh of Al's AT_NO_JUMPS patchset[1] (which was a variation on David Drysdale's O_BENEATH patchset[2], which in turn was based on the Capsicum project[3]). /* Userspace API. */ LOOKUP_NO_MAGICLINKS will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_NO_MAGICLINKS applies to all components of the path. With LOOKUP_NO_MAGICLINKS, any magic-link path component encountered during path resolution will yield -ELOOP. The handling of ~LOOKUP_FOLLOW for a trailing magic-link is identical to LOOKUP_NO_SYMLINKS. LOOKUP_NO_SYMLINKS implies LOOKUP_NO_MAGICLINKS. /* Testing. */ LOOKUP_NO_MAGICLINKS is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [2]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [3]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:31 +08:00
err:
path_put(path);
return error;
}
static inline void put_link(struct nameidata *nd)
{
struct saved *last = nd->stack + --nd->depth;
do_delayed_call(&last->done);
if (!(nd->flags & LOOKUP_RCU))
path_put(&last->link);
}
static int sysctl_protected_symlinks __read_mostly;
static int sysctl_protected_hardlinks __read_mostly;
static int sysctl_protected_fifos __read_mostly;
static int sysctl_protected_regular __read_mostly;
#ifdef CONFIG_SYSCTL
static struct ctl_table namei_sysctls[] = {
{
.procname = "protected_symlinks",
.data = &sysctl_protected_symlinks,
.maxlen = sizeof(int),
.mode = 0644,
.proc_handler = proc_dointvec_minmax,
.extra1 = SYSCTL_ZERO,
.extra2 = SYSCTL_ONE,
},
{
.procname = "protected_hardlinks",
.data = &sysctl_protected_hardlinks,
.maxlen = sizeof(int),
.mode = 0644,
.proc_handler = proc_dointvec_minmax,
.extra1 = SYSCTL_ZERO,
.extra2 = SYSCTL_ONE,
},
{
.procname = "protected_fifos",
.data = &sysctl_protected_fifos,
.maxlen = sizeof(int),
.mode = 0644,
.proc_handler = proc_dointvec_minmax,
.extra1 = SYSCTL_ZERO,
.extra2 = SYSCTL_TWO,
},
{
.procname = "protected_regular",
.data = &sysctl_protected_regular,
.maxlen = sizeof(int),
.mode = 0644,
.proc_handler = proc_dointvec_minmax,
.extra1 = SYSCTL_ZERO,
.extra2 = SYSCTL_TWO,
},
};
static int __init init_fs_namei_sysctls(void)
{
register_sysctl_init("fs", namei_sysctls);
return 0;
}
fs_initcall(init_fs_namei_sysctls);
#endif /* CONFIG_SYSCTL */
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
/**
* may_follow_link - Check symlink following for unsafe situations
* @nd: nameidata pathwalk data
* @inode: Used for idmapping.
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
*
* In the case of the sysctl_protected_symlinks sysctl being enabled,
* CAP_DAC_OVERRIDE needs to be specifically ignored if the symlink is
* in a sticky world-writable directory. This is to protect privileged
* processes from failing races against path names that may change out
* from under them by way of other users creating malicious symlinks.
* It will permit symlinks to be followed only when outside a sticky
* world-writable directory, or when the uid of the symlink and follower
* match, or when the directory owner matches the symlink's owner.
*
* Returns 0 if following the symlink is allowed, -ve on error.
*/
static inline int may_follow_link(struct nameidata *nd, const struct inode *inode)
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
{
struct mnt_idmap *idmap;
vfsuid_t vfsuid;
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
if (!sysctl_protected_symlinks)
return 0;
idmap = mnt_idmap(nd->path.mnt);
vfsuid = i_uid_into_vfsuid(idmap, inode);
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
/* Allowed if owner and follower match. */
if (vfsuid_eq_kuid(vfsuid, current_fsuid()))
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return 0;
/* Allowed if parent directory not sticky and world-writable. */
if ((nd->dir_mode & (S_ISVTX|S_IWOTH)) != (S_ISVTX|S_IWOTH))
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return 0;
/* Allowed if parent directory and link owner match. */
if (vfsuid_valid(nd->dir_vfsuid) && vfsuid_eq(nd->dir_vfsuid, vfsuid))
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return 0;
if (nd->flags & LOOKUP_RCU)
return -ECHILD;
audit_inode(nd->name, nd->stack[0].link.dentry, 0);
audit_log_path_denied(AUDIT_ANOM_LINK, "follow_link");
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return -EACCES;
}
/**
* safe_hardlink_source - Check for safe hardlink conditions
* @idmap: idmap of the mount the inode was found from
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
* @inode: the source inode to hardlink from
*
* Return false if at least one of the following conditions:
* - inode is not a regular file
* - inode is setuid
* - inode is setgid and group-exec
* - access failure for read and write
*
* Otherwise returns true.
*/
static bool safe_hardlink_source(struct mnt_idmap *idmap,
struct inode *inode)
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
{
umode_t mode = inode->i_mode;
/* Special files should not get pinned to the filesystem. */
if (!S_ISREG(mode))
return false;
/* Setuid files should not get pinned to the filesystem. */
if (mode & S_ISUID)
return false;
/* Executable setgid files should not get pinned to the filesystem. */
if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
return false;
/* Hardlinking to unreadable or unwritable sources is dangerous. */
if (inode_permission(idmap, inode, MAY_READ | MAY_WRITE))
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return false;
return true;
}
/**
* may_linkat - Check permissions for creating a hardlink
* @idmap: idmap of the mount the inode was found from
* @link: the source to hardlink from
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
*
* Block hardlink when all of:
* - sysctl_protected_hardlinks enabled
* - fsuid does not match inode
* - hardlink source is unsafe (see safe_hardlink_source() above)
* - not CAP_FOWNER in a namespace with the inode owner uid mapped
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
* Returns 0 if successful, -ve on error.
*/
int may_linkat(struct mnt_idmap *idmap, const struct path *link)
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
{
struct inode *inode = link->dentry->d_inode;
/* Inode writeback is not safe when the uid or gid are invalid. */
if (!vfsuid_valid(i_uid_into_vfsuid(idmap, inode)) ||
!vfsgid_valid(i_gid_into_vfsgid(idmap, inode)))
return -EOVERFLOW;
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
if (!sysctl_protected_hardlinks)
return 0;
/* Source inode owner (or CAP_FOWNER) can hardlink all they like,
* otherwise, it must be a safe source.
*/
if (safe_hardlink_source(idmap, inode) ||
inode_owner_or_capable(idmap, inode))
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return 0;
audit_log_path_denied(AUDIT_ANOM_LINK, "linkat");
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
return -EPERM;
}
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
/**
* may_create_in_sticky - Check whether an O_CREAT open in a sticky directory
* should be allowed, or not, on files that already
* exist.
* @idmap: idmap of the mount the inode was found from
namei: fix kernel-doc for struct renamedata and more Fix kernel-doc warnings in namei.c: ../fs/namei.c:1149: warning: Excess function parameter 'dir_mode' description in 'may_create_in_sticky' ../fs/namei.c:1149: warning: Excess function parameter 'dir_uid' description in 'may_create_in_sticky' ../fs/namei.c:3396: warning: Function parameter or member 'open_flag' not described in 'vfs_tmpfile' ../fs/namei.c:3396: warning: Excess function parameter 'open_flags' description in 'vfs_tmpfile' ../fs/namei.c:4460: warning: Function parameter or member 'rd' not described in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'old_mnt_userns' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'old_dir' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'old_dentry' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'new_mnt_userns' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'new_dir' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'new_dentry' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'delegated_inode' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'flags' description in 'vfs_rename' Link: https://lore.kernel.org/r/20210216042929.8931-3-rdunlap@infradead.org Fixes: 9fe61450972d ("namei: introduce struct renamedata") Signed-off-by: Randy Dunlap <rdunlap@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: linux-fsdevel@vger.kernel.org Cc: Christian Brauner <christian.brauner@ubuntu.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Christian Brauner <christian.brauner@ubuntu.com> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
2021-02-16 12:29:28 +08:00
* @nd: nameidata pathwalk data
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
* @inode: the inode of the file to open
*
* Block an O_CREAT open of a FIFO (or a regular file) when:
* - sysctl_protected_fifos (or sysctl_protected_regular) is enabled
* - the file already exists
* - we are in a sticky directory
* - we don't own the file
* - the owner of the directory doesn't own the file
* - the directory is world writable
* If the sysctl_protected_fifos (or sysctl_protected_regular) is set to 2
* the directory doesn't have to be world writable: being group writable will
* be enough.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
* Returns 0 if the open is allowed, -ve on error.
*/
static int may_create_in_sticky(struct mnt_idmap *idmap, struct nameidata *nd,
struct inode *const inode)
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
{
umode_t dir_mode = nd->dir_mode;
vfsuid_t dir_vfsuid = nd->dir_vfsuid, i_vfsuid;
if (likely(!(dir_mode & S_ISVTX)))
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
return 0;
if (S_ISREG(inode->i_mode) && !sysctl_protected_regular)
return 0;
if (S_ISFIFO(inode->i_mode) && !sysctl_protected_fifos)
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
return 0;
i_vfsuid = i_uid_into_vfsuid(idmap, inode);
if (vfsuid_eq(i_vfsuid, dir_vfsuid))
return 0;
if (vfsuid_eq_kuid(i_vfsuid, current_fsuid()))
return 0;
if (likely(dir_mode & 0002)) {
audit_log_path_denied(AUDIT_ANOM_CREAT, "sticky_create");
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
return -EACCES;
}
if (dir_mode & 0020) {
if (sysctl_protected_fifos >= 2 && S_ISFIFO(inode->i_mode)) {
audit_log_path_denied(AUDIT_ANOM_CREAT,
"sticky_create_fifo");
return -EACCES;
}
if (sysctl_protected_regular >= 2 && S_ISREG(inode->i_mode)) {
audit_log_path_denied(AUDIT_ANOM_CREAT,
"sticky_create_regular");
return -EACCES;
}
}
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
return 0;
}
/*
* follow_up - Find the mountpoint of path's vfsmount
*
* Given a path, find the mountpoint of its source file system.
* Replace @path with the path of the mountpoint in the parent mount.
* Up is towards /.
*
* Return 1 if we went up a level and 0 if we were already at the
* root.
*/
int follow_up(struct path *path)
{
struct mount *mnt = real_mount(path->mnt);
struct mount *parent;
struct dentry *mountpoint;
RCU'd vfsmounts * RCU-delayed freeing of vfsmounts * vfsmount_lock replaced with a seqlock (mount_lock) * sequence number from mount_lock is stored in nameidata->m_seq and used when we exit RCU mode * new vfsmount flag - MNT_SYNC_UMOUNT. Set by umount_tree() when its caller knows that vfsmount will have no surviving references. * synchronize_rcu() done between unlocking namespace_sem in namespace_unlock() and doing pending mntput(). * new helper: legitimize_mnt(mnt, seq). Checks the mount_lock sequence number against seq, then grabs reference to mnt. Then it rechecks mount_lock again to close the race and either returns success or drops the reference it has acquired. The subtle point is that in case of MNT_SYNC_UMOUNT we can simply decrement the refcount and sod off - aforementioned synchronize_rcu() makes sure that final mntput() won't come until we leave RCU mode. We need that, since we don't want to end up with some lazy pathwalk racing with umount() and stealing the final mntput() from it - caller of umount() may expect it to return only once the fs is shut down and we don't want to break that. In other cases (i.e. with MNT_SYNC_UMOUNT absent) we have to do full-blown mntput() in case of mount_lock sequence number mismatch happening just as we'd grabbed the reference, but in those cases we won't be stealing the final mntput() from anything that would care. * mntput_no_expire() doesn't lock anything on the fast path now. Incidentally, SMP and UP cases are handled the same way - no ifdefs there. * normal pathname resolution does *not* do any writes to mount_lock. It does, of course, bump the refcounts of vfsmount and dentry in the very end, but that's it. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-09-30 10:06:07 +08:00
read_seqlock_excl(&mount_lock);
parent = mnt->mnt_parent;
if (parent == mnt) {
RCU'd vfsmounts * RCU-delayed freeing of vfsmounts * vfsmount_lock replaced with a seqlock (mount_lock) * sequence number from mount_lock is stored in nameidata->m_seq and used when we exit RCU mode * new vfsmount flag - MNT_SYNC_UMOUNT. Set by umount_tree() when its caller knows that vfsmount will have no surviving references. * synchronize_rcu() done between unlocking namespace_sem in namespace_unlock() and doing pending mntput(). * new helper: legitimize_mnt(mnt, seq). Checks the mount_lock sequence number against seq, then grabs reference to mnt. Then it rechecks mount_lock again to close the race and either returns success or drops the reference it has acquired. The subtle point is that in case of MNT_SYNC_UMOUNT we can simply decrement the refcount and sod off - aforementioned synchronize_rcu() makes sure that final mntput() won't come until we leave RCU mode. We need that, since we don't want to end up with some lazy pathwalk racing with umount() and stealing the final mntput() from it - caller of umount() may expect it to return only once the fs is shut down and we don't want to break that. In other cases (i.e. with MNT_SYNC_UMOUNT absent) we have to do full-blown mntput() in case of mount_lock sequence number mismatch happening just as we'd grabbed the reference, but in those cases we won't be stealing the final mntput() from anything that would care. * mntput_no_expire() doesn't lock anything on the fast path now. Incidentally, SMP and UP cases are handled the same way - no ifdefs there. * normal pathname resolution does *not* do any writes to mount_lock. It does, of course, bump the refcounts of vfsmount and dentry in the very end, but that's it. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-09-30 10:06:07 +08:00
read_sequnlock_excl(&mount_lock);
return 0;
}
mntget(&parent->mnt);
mountpoint = dget(mnt->mnt_mountpoint);
RCU'd vfsmounts * RCU-delayed freeing of vfsmounts * vfsmount_lock replaced with a seqlock (mount_lock) * sequence number from mount_lock is stored in nameidata->m_seq and used when we exit RCU mode * new vfsmount flag - MNT_SYNC_UMOUNT. Set by umount_tree() when its caller knows that vfsmount will have no surviving references. * synchronize_rcu() done between unlocking namespace_sem in namespace_unlock() and doing pending mntput(). * new helper: legitimize_mnt(mnt, seq). Checks the mount_lock sequence number against seq, then grabs reference to mnt. Then it rechecks mount_lock again to close the race and either returns success or drops the reference it has acquired. The subtle point is that in case of MNT_SYNC_UMOUNT we can simply decrement the refcount and sod off - aforementioned synchronize_rcu() makes sure that final mntput() won't come until we leave RCU mode. We need that, since we don't want to end up with some lazy pathwalk racing with umount() and stealing the final mntput() from it - caller of umount() may expect it to return only once the fs is shut down and we don't want to break that. In other cases (i.e. with MNT_SYNC_UMOUNT absent) we have to do full-blown mntput() in case of mount_lock sequence number mismatch happening just as we'd grabbed the reference, but in those cases we won't be stealing the final mntput() from anything that would care. * mntput_no_expire() doesn't lock anything on the fast path now. Incidentally, SMP and UP cases are handled the same way - no ifdefs there. * normal pathname resolution does *not* do any writes to mount_lock. It does, of course, bump the refcounts of vfsmount and dentry in the very end, but that's it. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-09-30 10:06:07 +08:00
read_sequnlock_excl(&mount_lock);
dput(path->dentry);
path->dentry = mountpoint;
mntput(path->mnt);
path->mnt = &parent->mnt;
return 1;
}
EXPORT_SYMBOL(follow_up);
static bool choose_mountpoint_rcu(struct mount *m, const struct path *root,
struct path *path, unsigned *seqp)
{
while (mnt_has_parent(m)) {
struct dentry *mountpoint = m->mnt_mountpoint;
m = m->mnt_parent;
if (unlikely(root->dentry == mountpoint &&
root->mnt == &m->mnt))
break;
if (mountpoint != m->mnt.mnt_root) {
path->mnt = &m->mnt;
path->dentry = mountpoint;
*seqp = read_seqcount_begin(&mountpoint->d_seq);
return true;
}
}
return false;
}
static bool choose_mountpoint(struct mount *m, const struct path *root,
struct path *path)
{
bool found;
rcu_read_lock();
while (1) {
unsigned seq, mseq = read_seqbegin(&mount_lock);
found = choose_mountpoint_rcu(m, root, path, &seq);
if (unlikely(!found)) {
if (!read_seqretry(&mount_lock, mseq))
break;
} else {
if (likely(__legitimize_path(path, seq, mseq)))
break;
rcu_read_unlock();
path_put(path);
rcu_read_lock();
}
}
rcu_read_unlock();
return found;
}
/*
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
* Perform an automount
* - return -EISDIR to tell follow_managed() to stop and return the path we
* were called with.
*/
static int follow_automount(struct path *path, int *count, unsigned lookup_flags)
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
{
struct dentry *dentry = path->dentry;
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
/* We don't want to mount if someone's just doing a stat -
* unless they're stat'ing a directory and appended a '/' to
* the name.
*
* We do, however, want to mount if someone wants to open or
* create a file of any type under the mountpoint, wants to
* traverse through the mountpoint or wants to open the
* mounted directory. Also, autofs may mark negative dentries
* as being automount points. These will need the attentions
* of the daemon to instantiate them before they can be used.
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
*/
if (!(lookup_flags & (LOOKUP_PARENT | LOOKUP_DIRECTORY |
LOOKUP_OPEN | LOOKUP_CREATE | LOOKUP_AUTOMOUNT)) &&
dentry->d_inode)
return -EISDIR;
if (count && (*count)++ >= MAXSYMLINKS)
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
return -ELOOP;
return finish_automount(dentry->d_op->d_automount(path), path);
}
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
/*
* mount traversal - out-of-line part. One note on ->d_flags accesses -
* dentries are pinned but not locked here, so negative dentry can go
* positive right under us. Use of smp_load_acquire() provides a barrier
* sufficient for ->d_inode and ->d_flags consistency.
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
*/
static int __traverse_mounts(struct path *path, unsigned flags, bool *jumped,
int *count, unsigned lookup_flags)
{
struct vfsmount *mnt = path->mnt;
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
bool need_mntput = false;
VFS: Fix vfsmount overput on simultaneous automount [Kudos to dhowells for tracking that crap down] If two processes attempt to cause automounting on the same mountpoint at the same time, the vfsmount holding the mountpoint will be left with one too few references on it, causing a BUG when the kernel tries to clean up. The problem is that lock_mount() drops the caller's reference to the mountpoint's vfsmount in the case where it finds something already mounted on the mountpoint as it transits to the mounted filesystem and replaces path->mnt with the new mountpoint vfsmount. During a pathwalk, however, we don't take a reference on the vfsmount if it is the same as the one in the nameidata struct, but do_add_mount() doesn't know this. The fix is to make sure we have a ref on the vfsmount of the mountpoint before calling do_add_mount(). However, if lock_mount() doesn't transit, we're then left with an extra ref on the mountpoint vfsmount which needs releasing. We can handle that in follow_managed() by not making assumptions about what we can and what we cannot get from lookup_mnt() as the current code does. The callers of follow_managed() expect that reference to path->mnt will be grabbed iff path->mnt has been changed. follow_managed() and follow_automount() keep track of whether such reference has been grabbed and assume that it'll happen in those and only those cases that'll have us return with changed path->mnt. That assumption is almost correct - it breaks in case of racing automounts and in even harder to hit race between following a mountpoint and a couple of mount --move. The thing is, we don't need to make that assumption at all - after the end of loop in follow_manage() we can check if path->mnt has ended up unchanged and do mntput() if needed. The BUG can be reproduced with the following test program: #include <stdio.h> #include <sys/types.h> #include <sys/stat.h> #include <unistd.h> #include <sys/wait.h> int main(int argc, char **argv) { int pid, ws; struct stat buf; pid = fork(); stat(argv[1], &buf); if (pid > 0) wait(&ws); return 0; } and the following procedure: (1) Mount an NFS volume that on the server has something else mounted on a subdirectory. For instance, I can mount / from my server: mount warthog:/ /mnt -t nfs4 -r On the server /data has another filesystem mounted on it, so NFS will see a change in FSID as it walks down the path, and will mark /mnt/data as being a mountpoint. This will cause the automount code to be triggered. !!! Do not look inside the mounted fs at this point !!! (2) Run the above program on a file within the submount to generate two simultaneous automount requests: /tmp/forkstat /mnt/data/testfile (3) Unmount the automounted submount: umount /mnt/data (4) Unmount the original mount: umount /mnt At this point the kernel should throw a BUG with something like the following: BUG: Dentry ffff880032e3c5c0{i=2,n=} still in use (1) [unmount of nfs4 0:12] Note that the bug appears on the root dentry of the original mount, not the mountpoint and not the submount because sys_umount() hasn't got to its final mntput_no_expire() yet, but this isn't so obvious from the call trace: [<ffffffff8117cd82>] shrink_dcache_for_umount+0x69/0x82 [<ffffffff8116160e>] generic_shutdown_super+0x37/0x15b [<ffffffffa00fae56>] ? nfs_super_return_all_delegations+0x2e/0x1b1 [nfs] [<ffffffff811617f3>] kill_anon_super+0x1d/0x7e [<ffffffffa00d0be1>] nfs4_kill_super+0x60/0xb6 [nfs] [<ffffffff81161c17>] deactivate_locked_super+0x34/0x83 [<ffffffff811629ff>] deactivate_super+0x6f/0x7b [<ffffffff81186261>] mntput_no_expire+0x18d/0x199 [<ffffffff811862a8>] mntput+0x3b/0x44 [<ffffffff81186d87>] release_mounts+0xa2/0xbf [<ffffffff811876af>] sys_umount+0x47a/0x4ba [<ffffffff8109e1ca>] ? trace_hardirqs_on_caller+0x1fd/0x22f [<ffffffff816ea86b>] system_call_fastpath+0x16/0x1b as do_umount() is inlined. However, you can see release_mounts() in there. Note also that it may be necessary to have multiple CPU cores to be able to trigger this bug. Tested-by: Jeff Layton <jlayton@redhat.com> Tested-by: Ian Kent <raven@themaw.net> Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-06-16 22:10:06 +08:00
int ret = 0;
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
while (flags & DCACHE_MANAGED_DENTRY) {
Add a dentry op to allow processes to be held during pathwalk transit Add a dentry op (d_manage) to permit a filesystem to hold a process and make it sleep when it tries to transit away from one of that filesystem's directories during a pathwalk. The operation is keyed off a new dentry flag (DCACHE_MANAGE_TRANSIT). The filesystem is allowed to be selective about which processes it holds and which it permits to continue on or prohibits from transiting from each flagged directory. This will allow autofs to hold up client processes whilst letting its userspace daemon through to maintain the directory or the stuff behind it or mounted upon it. The ->d_manage() dentry operation: int (*d_manage)(struct path *path, bool mounting_here); takes a pointer to the directory about to be transited away from and a flag indicating whether the transit is undertaken by do_add_mount() or do_move_mount() skipping through a pile of filesystems mounted on a mountpoint. It should return 0 if successful and to let the process continue on its way; -EISDIR to prohibit the caller from skipping to overmounted filesystems or automounting, and to use this directory; or some other error code to return to the user. ->d_manage() is called with namespace_sem writelocked if mounting_here is true and no other locks held, so it may sleep. However, if mounting_here is true, it may not initiate or wait for a mount or unmount upon the parameter directory, even if the act is actually performed by userspace. Within fs/namei.c, follow_managed() is extended to check with d_manage() first on each managed directory, before transiting away from it or attempting to automount upon it. follow_down() is renamed follow_down_one() and should only be used where the filesystem deliberately intends to avoid management steps (e.g. autofs). A new follow_down() is added that incorporates the loop done by all other callers of follow_down() (do_add/move_mount(), autofs and NFSD; whilst AFS, NFS and CIFS do use it, their use is removed by converting them to use d_automount()). The new follow_down() calls d_manage() as appropriate. It also takes an extra parameter to indicate if it is being called from mount code (with namespace_sem writelocked) which it passes to d_manage(). follow_down() ignores automount points so that it can be used to mount on them. __follow_mount_rcu() is made to abort rcu-walk mode if it hits a directory with DCACHE_MANAGE_TRANSIT set on the basis that we're probably going to have to sleep. It would be possible to enter d_manage() in rcu-walk mode too, and have that determine whether to abort or not itself. That would allow the autofs daemon to continue on in rcu-walk mode. Note that DCACHE_MANAGE_TRANSIT on a directory should be cleared when it isn't required as every tranist from that directory will cause d_manage() to be invoked. It can always be set again when necessary. ========================== WHAT THIS MEANS FOR AUTOFS ========================== Autofs currently uses the lookup() inode op and the d_revalidate() dentry op to trigger the automounting of indirect mounts, and both of these can be called with i_mutex held. autofs knows that the i_mutex will be held by the caller in lookup(), and so can drop it before invoking the daemon - but this isn't so for d_revalidate(), since the lock is only held on _some_ of the code paths that call it. This means that autofs can't risk dropping i_mutex from its d_revalidate() function before it calls the daemon. The bug could manifest itself as, for example, a process that's trying to validate an automount dentry that gets made to wait because that dentry is expired and needs cleaning up: mkdir S ffffffff8014e05a 0 32580 24956 Call Trace: [<ffffffff885371fd>] :autofs4:autofs4_wait+0x674/0x897 [<ffffffff80127f7d>] avc_has_perm+0x46/0x58 [<ffffffff8009fdcf>] autoremove_wake_function+0x0/0x2e [<ffffffff88537be6>] :autofs4:autofs4_expire_wait+0x41/0x6b [<ffffffff88535cfc>] :autofs4:autofs4_revalidate+0x91/0x149 [<ffffffff80036d96>] __lookup_hash+0xa0/0x12f [<ffffffff80057a2f>] lookup_create+0x46/0x80 [<ffffffff800e6e31>] sys_mkdirat+0x56/0xe4 versus the automount daemon which wants to remove that dentry, but can't because the normal process is holding the i_mutex lock: automount D ffffffff8014e05a 0 32581 1 32561 Call Trace: [<ffffffff80063c3f>] __mutex_lock_slowpath+0x60/0x9b [<ffffffff8000ccf1>] do_path_lookup+0x2ca/0x2f1 [<ffffffff80063c89>] .text.lock.mutex+0xf/0x14 [<ffffffff800e6d55>] do_rmdir+0x77/0xde [<ffffffff8005d229>] tracesys+0x71/0xe0 [<ffffffff8005d28d>] tracesys+0xd5/0xe0 which means that the system is deadlocked. This patch allows autofs to hold up normal processes whilst the daemon goes ahead and does things to the dentry tree behind the automouter point without risking a deadlock as almost no locks are held in d_manage() and none in d_automount(). Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:26 +08:00
/* Allow the filesystem to manage the transit without i_mutex
* being held. */
if (flags & DCACHE_MANAGE_TRANSIT) {
ret = path->dentry->d_op->d_manage(path, false);
flags = smp_load_acquire(&path->dentry->d_flags);
Add a dentry op to allow processes to be held during pathwalk transit Add a dentry op (d_manage) to permit a filesystem to hold a process and make it sleep when it tries to transit away from one of that filesystem's directories during a pathwalk. The operation is keyed off a new dentry flag (DCACHE_MANAGE_TRANSIT). The filesystem is allowed to be selective about which processes it holds and which it permits to continue on or prohibits from transiting from each flagged directory. This will allow autofs to hold up client processes whilst letting its userspace daemon through to maintain the directory or the stuff behind it or mounted upon it. The ->d_manage() dentry operation: int (*d_manage)(struct path *path, bool mounting_here); takes a pointer to the directory about to be transited away from and a flag indicating whether the transit is undertaken by do_add_mount() or do_move_mount() skipping through a pile of filesystems mounted on a mountpoint. It should return 0 if successful and to let the process continue on its way; -EISDIR to prohibit the caller from skipping to overmounted filesystems or automounting, and to use this directory; or some other error code to return to the user. ->d_manage() is called with namespace_sem writelocked if mounting_here is true and no other locks held, so it may sleep. However, if mounting_here is true, it may not initiate or wait for a mount or unmount upon the parameter directory, even if the act is actually performed by userspace. Within fs/namei.c, follow_managed() is extended to check with d_manage() first on each managed directory, before transiting away from it or attempting to automount upon it. follow_down() is renamed follow_down_one() and should only be used where the filesystem deliberately intends to avoid management steps (e.g. autofs). A new follow_down() is added that incorporates the loop done by all other callers of follow_down() (do_add/move_mount(), autofs and NFSD; whilst AFS, NFS and CIFS do use it, their use is removed by converting them to use d_automount()). The new follow_down() calls d_manage() as appropriate. It also takes an extra parameter to indicate if it is being called from mount code (with namespace_sem writelocked) which it passes to d_manage(). follow_down() ignores automount points so that it can be used to mount on them. __follow_mount_rcu() is made to abort rcu-walk mode if it hits a directory with DCACHE_MANAGE_TRANSIT set on the basis that we're probably going to have to sleep. It would be possible to enter d_manage() in rcu-walk mode too, and have that determine whether to abort or not itself. That would allow the autofs daemon to continue on in rcu-walk mode. Note that DCACHE_MANAGE_TRANSIT on a directory should be cleared when it isn't required as every tranist from that directory will cause d_manage() to be invoked. It can always be set again when necessary. ========================== WHAT THIS MEANS FOR AUTOFS ========================== Autofs currently uses the lookup() inode op and the d_revalidate() dentry op to trigger the automounting of indirect mounts, and both of these can be called with i_mutex held. autofs knows that the i_mutex will be held by the caller in lookup(), and so can drop it before invoking the daemon - but this isn't so for d_revalidate(), since the lock is only held on _some_ of the code paths that call it. This means that autofs can't risk dropping i_mutex from its d_revalidate() function before it calls the daemon. The bug could manifest itself as, for example, a process that's trying to validate an automount dentry that gets made to wait because that dentry is expired and needs cleaning up: mkdir S ffffffff8014e05a 0 32580 24956 Call Trace: [<ffffffff885371fd>] :autofs4:autofs4_wait+0x674/0x897 [<ffffffff80127f7d>] avc_has_perm+0x46/0x58 [<ffffffff8009fdcf>] autoremove_wake_function+0x0/0x2e [<ffffffff88537be6>] :autofs4:autofs4_expire_wait+0x41/0x6b [<ffffffff88535cfc>] :autofs4:autofs4_revalidate+0x91/0x149 [<ffffffff80036d96>] __lookup_hash+0xa0/0x12f [<ffffffff80057a2f>] lookup_create+0x46/0x80 [<ffffffff800e6e31>] sys_mkdirat+0x56/0xe4 versus the automount daemon which wants to remove that dentry, but can't because the normal process is holding the i_mutex lock: automount D ffffffff8014e05a 0 32581 1 32561 Call Trace: [<ffffffff80063c3f>] __mutex_lock_slowpath+0x60/0x9b [<ffffffff8000ccf1>] do_path_lookup+0x2ca/0x2f1 [<ffffffff80063c89>] .text.lock.mutex+0xf/0x14 [<ffffffff800e6d55>] do_rmdir+0x77/0xde [<ffffffff8005d229>] tracesys+0x71/0xe0 [<ffffffff8005d28d>] tracesys+0xd5/0xe0 which means that the system is deadlocked. This patch allows autofs to hold up normal processes whilst the daemon goes ahead and does things to the dentry tree behind the automouter point without risking a deadlock as almost no locks are held in d_manage() and none in d_automount(). Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:26 +08:00
if (ret < 0)
VFS: Fix vfsmount overput on simultaneous automount [Kudos to dhowells for tracking that crap down] If two processes attempt to cause automounting on the same mountpoint at the same time, the vfsmount holding the mountpoint will be left with one too few references on it, causing a BUG when the kernel tries to clean up. The problem is that lock_mount() drops the caller's reference to the mountpoint's vfsmount in the case where it finds something already mounted on the mountpoint as it transits to the mounted filesystem and replaces path->mnt with the new mountpoint vfsmount. During a pathwalk, however, we don't take a reference on the vfsmount if it is the same as the one in the nameidata struct, but do_add_mount() doesn't know this. The fix is to make sure we have a ref on the vfsmount of the mountpoint before calling do_add_mount(). However, if lock_mount() doesn't transit, we're then left with an extra ref on the mountpoint vfsmount which needs releasing. We can handle that in follow_managed() by not making assumptions about what we can and what we cannot get from lookup_mnt() as the current code does. The callers of follow_managed() expect that reference to path->mnt will be grabbed iff path->mnt has been changed. follow_managed() and follow_automount() keep track of whether such reference has been grabbed and assume that it'll happen in those and only those cases that'll have us return with changed path->mnt. That assumption is almost correct - it breaks in case of racing automounts and in even harder to hit race between following a mountpoint and a couple of mount --move. The thing is, we don't need to make that assumption at all - after the end of loop in follow_manage() we can check if path->mnt has ended up unchanged and do mntput() if needed. The BUG can be reproduced with the following test program: #include <stdio.h> #include <sys/types.h> #include <sys/stat.h> #include <unistd.h> #include <sys/wait.h> int main(int argc, char **argv) { int pid, ws; struct stat buf; pid = fork(); stat(argv[1], &buf); if (pid > 0) wait(&ws); return 0; } and the following procedure: (1) Mount an NFS volume that on the server has something else mounted on a subdirectory. For instance, I can mount / from my server: mount warthog:/ /mnt -t nfs4 -r On the server /data has another filesystem mounted on it, so NFS will see a change in FSID as it walks down the path, and will mark /mnt/data as being a mountpoint. This will cause the automount code to be triggered. !!! Do not look inside the mounted fs at this point !!! (2) Run the above program on a file within the submount to generate two simultaneous automount requests: /tmp/forkstat /mnt/data/testfile (3) Unmount the automounted submount: umount /mnt/data (4) Unmount the original mount: umount /mnt At this point the kernel should throw a BUG with something like the following: BUG: Dentry ffff880032e3c5c0{i=2,n=} still in use (1) [unmount of nfs4 0:12] Note that the bug appears on the root dentry of the original mount, not the mountpoint and not the submount because sys_umount() hasn't got to its final mntput_no_expire() yet, but this isn't so obvious from the call trace: [<ffffffff8117cd82>] shrink_dcache_for_umount+0x69/0x82 [<ffffffff8116160e>] generic_shutdown_super+0x37/0x15b [<ffffffffa00fae56>] ? nfs_super_return_all_delegations+0x2e/0x1b1 [nfs] [<ffffffff811617f3>] kill_anon_super+0x1d/0x7e [<ffffffffa00d0be1>] nfs4_kill_super+0x60/0xb6 [nfs] [<ffffffff81161c17>] deactivate_locked_super+0x34/0x83 [<ffffffff811629ff>] deactivate_super+0x6f/0x7b [<ffffffff81186261>] mntput_no_expire+0x18d/0x199 [<ffffffff811862a8>] mntput+0x3b/0x44 [<ffffffff81186d87>] release_mounts+0xa2/0xbf [<ffffffff811876af>] sys_umount+0x47a/0x4ba [<ffffffff8109e1ca>] ? trace_hardirqs_on_caller+0x1fd/0x22f [<ffffffff816ea86b>] system_call_fastpath+0x16/0x1b as do_umount() is inlined. However, you can see release_mounts() in there. Note also that it may be necessary to have multiple CPU cores to be able to trigger this bug. Tested-by: Jeff Layton <jlayton@redhat.com> Tested-by: Ian Kent <raven@themaw.net> Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-06-16 22:10:06 +08:00
break;
Add a dentry op to allow processes to be held during pathwalk transit Add a dentry op (d_manage) to permit a filesystem to hold a process and make it sleep when it tries to transit away from one of that filesystem's directories during a pathwalk. The operation is keyed off a new dentry flag (DCACHE_MANAGE_TRANSIT). The filesystem is allowed to be selective about which processes it holds and which it permits to continue on or prohibits from transiting from each flagged directory. This will allow autofs to hold up client processes whilst letting its userspace daemon through to maintain the directory or the stuff behind it or mounted upon it. The ->d_manage() dentry operation: int (*d_manage)(struct path *path, bool mounting_here); takes a pointer to the directory about to be transited away from and a flag indicating whether the transit is undertaken by do_add_mount() or do_move_mount() skipping through a pile of filesystems mounted on a mountpoint. It should return 0 if successful and to let the process continue on its way; -EISDIR to prohibit the caller from skipping to overmounted filesystems or automounting, and to use this directory; or some other error code to return to the user. ->d_manage() is called with namespace_sem writelocked if mounting_here is true and no other locks held, so it may sleep. However, if mounting_here is true, it may not initiate or wait for a mount or unmount upon the parameter directory, even if the act is actually performed by userspace. Within fs/namei.c, follow_managed() is extended to check with d_manage() first on each managed directory, before transiting away from it or attempting to automount upon it. follow_down() is renamed follow_down_one() and should only be used where the filesystem deliberately intends to avoid management steps (e.g. autofs). A new follow_down() is added that incorporates the loop done by all other callers of follow_down() (do_add/move_mount(), autofs and NFSD; whilst AFS, NFS and CIFS do use it, their use is removed by converting them to use d_automount()). The new follow_down() calls d_manage() as appropriate. It also takes an extra parameter to indicate if it is being called from mount code (with namespace_sem writelocked) which it passes to d_manage(). follow_down() ignores automount points so that it can be used to mount on them. __follow_mount_rcu() is made to abort rcu-walk mode if it hits a directory with DCACHE_MANAGE_TRANSIT set on the basis that we're probably going to have to sleep. It would be possible to enter d_manage() in rcu-walk mode too, and have that determine whether to abort or not itself. That would allow the autofs daemon to continue on in rcu-walk mode. Note that DCACHE_MANAGE_TRANSIT on a directory should be cleared when it isn't required as every tranist from that directory will cause d_manage() to be invoked. It can always be set again when necessary. ========================== WHAT THIS MEANS FOR AUTOFS ========================== Autofs currently uses the lookup() inode op and the d_revalidate() dentry op to trigger the automounting of indirect mounts, and both of these can be called with i_mutex held. autofs knows that the i_mutex will be held by the caller in lookup(), and so can drop it before invoking the daemon - but this isn't so for d_revalidate(), since the lock is only held on _some_ of the code paths that call it. This means that autofs can't risk dropping i_mutex from its d_revalidate() function before it calls the daemon. The bug could manifest itself as, for example, a process that's trying to validate an automount dentry that gets made to wait because that dentry is expired and needs cleaning up: mkdir S ffffffff8014e05a 0 32580 24956 Call Trace: [<ffffffff885371fd>] :autofs4:autofs4_wait+0x674/0x897 [<ffffffff80127f7d>] avc_has_perm+0x46/0x58 [<ffffffff8009fdcf>] autoremove_wake_function+0x0/0x2e [<ffffffff88537be6>] :autofs4:autofs4_expire_wait+0x41/0x6b [<ffffffff88535cfc>] :autofs4:autofs4_revalidate+0x91/0x149 [<ffffffff80036d96>] __lookup_hash+0xa0/0x12f [<ffffffff80057a2f>] lookup_create+0x46/0x80 [<ffffffff800e6e31>] sys_mkdirat+0x56/0xe4 versus the automount daemon which wants to remove that dentry, but can't because the normal process is holding the i_mutex lock: automount D ffffffff8014e05a 0 32581 1 32561 Call Trace: [<ffffffff80063c3f>] __mutex_lock_slowpath+0x60/0x9b [<ffffffff8000ccf1>] do_path_lookup+0x2ca/0x2f1 [<ffffffff80063c89>] .text.lock.mutex+0xf/0x14 [<ffffffff800e6d55>] do_rmdir+0x77/0xde [<ffffffff8005d229>] tracesys+0x71/0xe0 [<ffffffff8005d28d>] tracesys+0xd5/0xe0 which means that the system is deadlocked. This patch allows autofs to hold up normal processes whilst the daemon goes ahead and does things to the dentry tree behind the automouter point without risking a deadlock as almost no locks are held in d_manage() and none in d_automount(). Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:26 +08:00
}
if (flags & DCACHE_MOUNTED) { // something's mounted on it..
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
struct vfsmount *mounted = lookup_mnt(path);
if (mounted) { // ... in our namespace
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
dput(path->dentry);
if (need_mntput)
mntput(path->mnt);
path->mnt = mounted;
path->dentry = dget(mounted->mnt_root);
// here we know it's positive
flags = path->dentry->d_flags;
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
need_mntput = true;
continue;
}
}
if (!(flags & DCACHE_NEED_AUTOMOUNT))
break;
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
// uncovered automount point
ret = follow_automount(path, count, lookup_flags);
flags = smp_load_acquire(&path->dentry->d_flags);
if (ret < 0)
break;
}
VFS: Fix vfsmount overput on simultaneous automount [Kudos to dhowells for tracking that crap down] If two processes attempt to cause automounting on the same mountpoint at the same time, the vfsmount holding the mountpoint will be left with one too few references on it, causing a BUG when the kernel tries to clean up. The problem is that lock_mount() drops the caller's reference to the mountpoint's vfsmount in the case where it finds something already mounted on the mountpoint as it transits to the mounted filesystem and replaces path->mnt with the new mountpoint vfsmount. During a pathwalk, however, we don't take a reference on the vfsmount if it is the same as the one in the nameidata struct, but do_add_mount() doesn't know this. The fix is to make sure we have a ref on the vfsmount of the mountpoint before calling do_add_mount(). However, if lock_mount() doesn't transit, we're then left with an extra ref on the mountpoint vfsmount which needs releasing. We can handle that in follow_managed() by not making assumptions about what we can and what we cannot get from lookup_mnt() as the current code does. The callers of follow_managed() expect that reference to path->mnt will be grabbed iff path->mnt has been changed. follow_managed() and follow_automount() keep track of whether such reference has been grabbed and assume that it'll happen in those and only those cases that'll have us return with changed path->mnt. That assumption is almost correct - it breaks in case of racing automounts and in even harder to hit race between following a mountpoint and a couple of mount --move. The thing is, we don't need to make that assumption at all - after the end of loop in follow_manage() we can check if path->mnt has ended up unchanged and do mntput() if needed. The BUG can be reproduced with the following test program: #include <stdio.h> #include <sys/types.h> #include <sys/stat.h> #include <unistd.h> #include <sys/wait.h> int main(int argc, char **argv) { int pid, ws; struct stat buf; pid = fork(); stat(argv[1], &buf); if (pid > 0) wait(&ws); return 0; } and the following procedure: (1) Mount an NFS volume that on the server has something else mounted on a subdirectory. For instance, I can mount / from my server: mount warthog:/ /mnt -t nfs4 -r On the server /data has another filesystem mounted on it, so NFS will see a change in FSID as it walks down the path, and will mark /mnt/data as being a mountpoint. This will cause the automount code to be triggered. !!! Do not look inside the mounted fs at this point !!! (2) Run the above program on a file within the submount to generate two simultaneous automount requests: /tmp/forkstat /mnt/data/testfile (3) Unmount the automounted submount: umount /mnt/data (4) Unmount the original mount: umount /mnt At this point the kernel should throw a BUG with something like the following: BUG: Dentry ffff880032e3c5c0{i=2,n=} still in use (1) [unmount of nfs4 0:12] Note that the bug appears on the root dentry of the original mount, not the mountpoint and not the submount because sys_umount() hasn't got to its final mntput_no_expire() yet, but this isn't so obvious from the call trace: [<ffffffff8117cd82>] shrink_dcache_for_umount+0x69/0x82 [<ffffffff8116160e>] generic_shutdown_super+0x37/0x15b [<ffffffffa00fae56>] ? nfs_super_return_all_delegations+0x2e/0x1b1 [nfs] [<ffffffff811617f3>] kill_anon_super+0x1d/0x7e [<ffffffffa00d0be1>] nfs4_kill_super+0x60/0xb6 [nfs] [<ffffffff81161c17>] deactivate_locked_super+0x34/0x83 [<ffffffff811629ff>] deactivate_super+0x6f/0x7b [<ffffffff81186261>] mntput_no_expire+0x18d/0x199 [<ffffffff811862a8>] mntput+0x3b/0x44 [<ffffffff81186d87>] release_mounts+0xa2/0xbf [<ffffffff811876af>] sys_umount+0x47a/0x4ba [<ffffffff8109e1ca>] ? trace_hardirqs_on_caller+0x1fd/0x22f [<ffffffff816ea86b>] system_call_fastpath+0x16/0x1b as do_umount() is inlined. However, you can see release_mounts() in there. Note also that it may be necessary to have multiple CPU cores to be able to trigger this bug. Tested-by: Jeff Layton <jlayton@redhat.com> Tested-by: Ian Kent <raven@themaw.net> Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-06-16 22:10:06 +08:00
if (ret == -EISDIR)
ret = 0;
// possible if you race with several mount --move
if (need_mntput && path->mnt == mnt)
mntput(path->mnt);
if (!ret && unlikely(d_flags_negative(flags)))
ret = -ENOENT;
*jumped = need_mntput;
return ret;
}
static inline int traverse_mounts(struct path *path, bool *jumped,
int *count, unsigned lookup_flags)
{
unsigned flags = smp_load_acquire(&path->dentry->d_flags);
/* fastpath */
if (likely(!(flags & DCACHE_MANAGED_DENTRY))) {
*jumped = false;
if (unlikely(d_flags_negative(flags)))
return -ENOENT;
return 0;
}
return __traverse_mounts(path, flags, jumped, count, lookup_flags);
}
Add a dentry op to allow processes to be held during pathwalk transit Add a dentry op (d_manage) to permit a filesystem to hold a process and make it sleep when it tries to transit away from one of that filesystem's directories during a pathwalk. The operation is keyed off a new dentry flag (DCACHE_MANAGE_TRANSIT). The filesystem is allowed to be selective about which processes it holds and which it permits to continue on or prohibits from transiting from each flagged directory. This will allow autofs to hold up client processes whilst letting its userspace daemon through to maintain the directory or the stuff behind it or mounted upon it. The ->d_manage() dentry operation: int (*d_manage)(struct path *path, bool mounting_here); takes a pointer to the directory about to be transited away from and a flag indicating whether the transit is undertaken by do_add_mount() or do_move_mount() skipping through a pile of filesystems mounted on a mountpoint. It should return 0 if successful and to let the process continue on its way; -EISDIR to prohibit the caller from skipping to overmounted filesystems or automounting, and to use this directory; or some other error code to return to the user. ->d_manage() is called with namespace_sem writelocked if mounting_here is true and no other locks held, so it may sleep. However, if mounting_here is true, it may not initiate or wait for a mount or unmount upon the parameter directory, even if the act is actually performed by userspace. Within fs/namei.c, follow_managed() is extended to check with d_manage() first on each managed directory, before transiting away from it or attempting to automount upon it. follow_down() is renamed follow_down_one() and should only be used where the filesystem deliberately intends to avoid management steps (e.g. autofs). A new follow_down() is added that incorporates the loop done by all other callers of follow_down() (do_add/move_mount(), autofs and NFSD; whilst AFS, NFS and CIFS do use it, their use is removed by converting them to use d_automount()). The new follow_down() calls d_manage() as appropriate. It also takes an extra parameter to indicate if it is being called from mount code (with namespace_sem writelocked) which it passes to d_manage(). follow_down() ignores automount points so that it can be used to mount on them. __follow_mount_rcu() is made to abort rcu-walk mode if it hits a directory with DCACHE_MANAGE_TRANSIT set on the basis that we're probably going to have to sleep. It would be possible to enter d_manage() in rcu-walk mode too, and have that determine whether to abort or not itself. That would allow the autofs daemon to continue on in rcu-walk mode. Note that DCACHE_MANAGE_TRANSIT on a directory should be cleared when it isn't required as every tranist from that directory will cause d_manage() to be invoked. It can always be set again when necessary. ========================== WHAT THIS MEANS FOR AUTOFS ========================== Autofs currently uses the lookup() inode op and the d_revalidate() dentry op to trigger the automounting of indirect mounts, and both of these can be called with i_mutex held. autofs knows that the i_mutex will be held by the caller in lookup(), and so can drop it before invoking the daemon - but this isn't so for d_revalidate(), since the lock is only held on _some_ of the code paths that call it. This means that autofs can't risk dropping i_mutex from its d_revalidate() function before it calls the daemon. The bug could manifest itself as, for example, a process that's trying to validate an automount dentry that gets made to wait because that dentry is expired and needs cleaning up: mkdir S ffffffff8014e05a 0 32580 24956 Call Trace: [<ffffffff885371fd>] :autofs4:autofs4_wait+0x674/0x897 [<ffffffff80127f7d>] avc_has_perm+0x46/0x58 [<ffffffff8009fdcf>] autoremove_wake_function+0x0/0x2e [<ffffffff88537be6>] :autofs4:autofs4_expire_wait+0x41/0x6b [<ffffffff88535cfc>] :autofs4:autofs4_revalidate+0x91/0x149 [<ffffffff80036d96>] __lookup_hash+0xa0/0x12f [<ffffffff80057a2f>] lookup_create+0x46/0x80 [<ffffffff800e6e31>] sys_mkdirat+0x56/0xe4 versus the automount daemon which wants to remove that dentry, but can't because the normal process is holding the i_mutex lock: automount D ffffffff8014e05a 0 32581 1 32561 Call Trace: [<ffffffff80063c3f>] __mutex_lock_slowpath+0x60/0x9b [<ffffffff8000ccf1>] do_path_lookup+0x2ca/0x2f1 [<ffffffff80063c89>] .text.lock.mutex+0xf/0x14 [<ffffffff800e6d55>] do_rmdir+0x77/0xde [<ffffffff8005d229>] tracesys+0x71/0xe0 [<ffffffff8005d28d>] tracesys+0xd5/0xe0 which means that the system is deadlocked. This patch allows autofs to hold up normal processes whilst the daemon goes ahead and does things to the dentry tree behind the automouter point without risking a deadlock as almost no locks are held in d_manage() and none in d_automount(). Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:26 +08:00
int follow_down_one(struct path *path)
{
struct vfsmount *mounted;
mounted = lookup_mnt(path);
if (mounted) {
dput(path->dentry);
mntput(path->mnt);
path->mnt = mounted;
path->dentry = dget(mounted->mnt_root);
return 1;
}
return 0;
}
EXPORT_SYMBOL(follow_down_one);
/*
* Follow down to the covering mount currently visible to userspace. At each
* point, the filesystem owning that dentry may be queried as to whether the
* caller is permitted to proceed or not.
*/
int follow_down(struct path *path, unsigned int flags)
{
struct vfsmount *mnt = path->mnt;
bool jumped;
int ret = traverse_mounts(path, &jumped, NULL, flags);
if (path->mnt != mnt)
mntput(mnt);
return ret;
}
EXPORT_SYMBOL(follow_down);
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
/*
* Try to skip to top of mountpoint pile in rcuwalk mode. Fail if
* we meet a managed dentry that would need blocking.
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
*/
static bool __follow_mount_rcu(struct nameidata *nd, struct path *path)
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
{
struct dentry *dentry = path->dentry;
unsigned int flags = dentry->d_flags;
if (likely(!(flags & DCACHE_MANAGED_DENTRY)))
return true;
if (unlikely(nd->flags & LOOKUP_NO_XDEV))
return false;
for (;;) {
/*
* Don't forget we might have a non-mountpoint managed dentry
* that wants to block transit.
*/
if (unlikely(flags & DCACHE_MANAGE_TRANSIT)) {
int res = dentry->d_op->d_manage(path, true);
if (res)
return res == -EISDIR;
flags = dentry->d_flags;
}
if (flags & DCACHE_MOUNTED) {
struct mount *mounted = __lookup_mnt(path->mnt, dentry);
if (mounted) {
path->mnt = &mounted->mnt;
dentry = path->dentry = mounted->mnt.mnt_root;
nd->state |= ND_JUMPED;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->next_seq = read_seqcount_begin(&dentry->d_seq);
flags = dentry->d_flags;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
// makes sure that non-RCU pathwalk could reach
// this state.
if (read_seqretry(&mount_lock, nd->m_seq))
return false;
continue;
}
if (read_seqretry(&mount_lock, nd->m_seq))
return false;
}
return !(flags & DCACHE_NEED_AUTOMOUNT);
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
}
}
static inline int handle_mounts(struct nameidata *nd, struct dentry *dentry,
struct path *path)
{
bool jumped;
int ret;
path->mnt = nd->path.mnt;
path->dentry = dentry;
if (nd->flags & LOOKUP_RCU) {
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
unsigned int seq = nd->next_seq;
if (likely(__follow_mount_rcu(nd, path)))
return 0;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
// *path and nd->next_seq might've been clobbered
path->mnt = nd->path.mnt;
path->dentry = dentry;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->next_seq = seq;
if (!try_to_unlazy_next(nd, dentry))
return -ECHILD;
}
ret = traverse_mounts(path, &jumped, &nd->total_link_count, nd->flags);
if (jumped) {
if (unlikely(nd->flags & LOOKUP_NO_XDEV))
ret = -EXDEV;
else
nd->state |= ND_JUMPED;
}
if (unlikely(ret)) {
dput(path->dentry);
if (path->mnt != nd->path.mnt)
mntput(path->mnt);
}
return ret;
}
/*
* This looks up the name in dcache and possibly revalidates the found dentry.
* NULL is returned if the dentry does not exist in the cache.
*/
static struct dentry *lookup_dcache(const struct qstr *name,
struct dentry *dir,
unsigned int flags)
{
struct dentry *dentry = d_lookup(dir, name);
if (dentry) {
int error = d_revalidate(dentry, flags);
if (unlikely(error <= 0)) {
if (!error)
d_invalidate(dentry);
dput(dentry);
return ERR_PTR(error);
}
}
return dentry;
}
/*
* Parent directory has inode locked exclusive. This is one
* and only case when ->lookup() gets called on non in-lookup
* dentries - as the matter of fact, this only gets called
* when directory is guaranteed to have no in-lookup children
* at all.
*/
struct dentry *lookup_one_qstr_excl(const struct qstr *name,
struct dentry *base,
unsigned int flags)
{
struct dentry *dentry = lookup_dcache(name, base, flags);
struct dentry *old;
struct inode *dir = base->d_inode;
if (dentry)
return dentry;
/* Don't create child dentry for a dead directory. */
if (unlikely(IS_DEADDIR(dir)))
return ERR_PTR(-ENOENT);
dentry = d_alloc(base, name);
if (unlikely(!dentry))
return ERR_PTR(-ENOMEM);
old = dir->i_op->lookup(dir, dentry, flags);
if (unlikely(old)) {
dput(dentry);
dentry = old;
}
return dentry;
}
EXPORT_SYMBOL(lookup_one_qstr_excl);
/**
* lookup_fast - do fast lockless (but racy) lookup of a dentry
* @nd: current nameidata
*
* Do a fast, but racy lookup in the dcache for the given dentry, and
* revalidate it. Returns a valid dentry pointer or NULL if one wasn't
* found. On error, an ERR_PTR will be returned.
*
* If this function returns a valid dentry and the walk is no longer
* lazy, the dentry will carry a reference that must later be put. If
* RCU mode is still in force, then this is not the case and the dentry
* must be legitimized before use. If this returns NULL, then the walk
* will no longer be in RCU mode.
*/
static struct dentry *lookup_fast(struct nameidata *nd)
{
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
struct dentry *dentry, *parent = nd->path.dentry;
int status = 1;
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
/*
* Rename seqlock is not required here because in the off chance
* of a false negative due to a concurrent rename, the caller is
* going to fall back to non-racy lookup.
*/
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
if (nd->flags & LOOKUP_RCU) {
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
dentry = __d_lookup_rcu(parent, &nd->last, &nd->next_seq);
if (unlikely(!dentry)) {
if (!try_to_unlazy(nd))
return ERR_PTR(-ECHILD);
return NULL;
}
vfs: clean up __d_lookup_rcu() and dentry_cmp() interfaces The calling conventions for __d_lookup_rcu() and dentry_cmp() are annoying in different ways, and there is actually one single underlying reason for both of the annoyances. The fundamental reason is that we do the returned dentry sequence number check inside __d_lookup_rcu() instead of doing it in the caller. This results in two annoyances: - __d_lookup_rcu() now not only needs to return the dentry and the sequence number that goes along with the lookup, it also needs to return the inode pointer that was validated by that sequence number check. - and because we did the sequence number check early (to validate the name pointer and length) we also couldn't just pass the dentry itself to dentry_cmp(), we had to pass the counted string that contained the name. So that sequence number decision caused two separate ugly calling conventions. Both of these problems would be solved if we just did the sequence number check in the caller instead. There's only one caller, and that caller already has to do the sequence number check for the parent anyway, so just do that. That allows us to stop returning the dentry->d_inode in that in-out argument (pointer-to-pointer-to-inode), so we can make the inode argument just a regular input inode pointer. The caller can just load the inode from dentry->d_inode, and then do the sequence number check after that to make sure that it's synchronized with the name we looked up. And it allows us to just pass in the dentry to dentry_cmp(), which is what all the callers really wanted. Sure, dentry_cmp() has to be a bit careful about the dentry (which is not stable during RCU lookup), but that's actually very simple. And now that dentry_cmp() can clearly see that the first string argument is a dentry, we can use the direct word access for that, instead of the careful unaligned zero-padding. The dentry name is always properly aligned, since it is a single path component that is either embedded into the dentry itself, or was allocated with kmalloc() (see __d_alloc). Finally, this also uninlines the nasty slow-case for dentry comparisons: that one *does* need to do a sequence number check, since it will call in to the low-level filesystems, and we want to give those a stable inode pointer and path component length/start arguments. Doing an extra sequence check for that slow case is not a problem, though. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-05 05:59:14 +08:00
/*
* This sequence count validates that the parent had no
* changes while we did the lookup of the dentry above.
*/
if (read_seqcount_retry(&parent->d_seq, nd->seq))
return ERR_PTR(-ECHILD);
status = d_revalidate(dentry, nd->flags);
if (likely(status > 0))
return dentry;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
if (!try_to_unlazy_next(nd, dentry))
return ERR_PTR(-ECHILD);
if (status == -ECHILD)
/* we'd been told to redo it in non-rcu mode */
status = d_revalidate(dentry, nd->flags);
} else {
dentry = __d_lookup(parent, &nd->last);
if (unlikely(!dentry))
return NULL;
status = d_revalidate(dentry, nd->flags);
Add a dentry op to handle automounting rather than abusing follow_link() Add a dentry op (d_automount) to handle automounting directories rather than abusing the follow_link() inode operation. The operation is keyed off a new dentry flag (DCACHE_NEED_AUTOMOUNT). This also makes it easier to add an AT_ flag to suppress terminal segment automount during pathwalk and removes the need for the kludge code in the pathwalk algorithm to handle directories with follow_link() semantics. The ->d_automount() dentry operation: struct vfsmount *(*d_automount)(struct path *mountpoint); takes a pointer to the directory to be mounted upon, which is expected to provide sufficient data to determine what should be mounted. If successful, it should return the vfsmount struct it creates (which it should also have added to the namespace using do_add_mount() or similar). If there's a collision with another automount attempt, NULL should be returned. If the directory specified by the parameter should be used directly rather than being mounted upon, -EISDIR should be returned. In any other case, an error code should be returned. The ->d_automount() operation is called with no locks held and may sleep. At this point the pathwalk algorithm will be in ref-walk mode. Within fs/namei.c itself, a new pathwalk subroutine (follow_automount()) is added to handle mountpoints. It will return -EREMOTE if the automount flag was set, but no d_automount() op was supplied, -ELOOP if we've encountered too many symlinks or mountpoints, -EISDIR if the walk point should be used without mounting and 0 if successful. The path will be updated to point to the mounted filesystem if a successful automount took place. __follow_mount() is replaced by follow_managed() which is more generic (especially with the patch that adds ->d_manage()). This handles transits from directories during pathwalk, including automounting and skipping over mountpoints (and holding processes with the next patch). __follow_mount_rcu() will jump out of RCU-walk mode if it encounters an automount point with nothing mounted on it. follow_dotdot*() does not handle automounts as you don't want to trigger them whilst following "..". I've also extracted the mount/don't-mount logic from autofs4 and included it here. It makes the mount go ahead anyway if someone calls open() or creat(), tries to traverse the directory, tries to chdir/chroot/etc. into the directory, or sticks a '/' on the end of the pathname. If they do a stat(), however, they'll only trigger the automount if they didn't also say O_NOFOLLOW. I've also added an inode flag (S_AUTOMOUNT) so that filesystems can mark their inodes as automount points. This flag is automatically propagated to the dentry as DCACHE_NEED_AUTOMOUNT by __d_instantiate(). This saves NFS and could save AFS a private flag bit apiece, but is not strictly necessary. It would be preferable to do the propagation in d_set_d_op(), but that doesn't normally have access to the inode. [AV: fixed breakage in case if __follow_mount_rcu() fails and nameidata_drop_rcu() succeeds in RCU case of do_lookup(); we need to fall through to non-RCU case after that, rather than just returning with ungrabbed *path] Signed-off-by: David Howells <dhowells@redhat.com> Was-Acked-by: Ian Kent <raven@themaw.net> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-01-15 02:45:21 +08:00
}
if (unlikely(status <= 0)) {
if (!status)
d_invalidate(dentry);
dput(dentry);
return ERR_PTR(status);
}
return dentry;
}
/* Fast lookup failed, do it the slow way */
static struct dentry *__lookup_slow(const struct qstr *name,
struct dentry *dir,
unsigned int flags)
{
struct dentry *dentry, *old;
struct inode *inode = dir->d_inode;
DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
/* Don't go there if it's already dead */
parallel lookups machinery, part 3 We will need to be able to check if there is an in-lookup dentry with matching parent/name. Right now it's impossible, but as soon as start locking directories shared such beasts will appear. Add a secondary hash for locating those. Hash chains go through the same space where d_alias will be once it's not in-lookup anymore. Search is done under the same bitlock we use for modifications - with the primary hash we can rely on d_rehash() into the wrong chain being the worst that could happen, but here the pointers are buggered once it's removed from the chain. On the other hand, the chains are not going to be long and normally we'll end up adding to the chain anyway. That allows us to avoid bothering with ->d_lock when doing the comparisons - everything is stable until removed from chain. New helper: d_alloc_parallel(). Right now it allocates, verifies that no hashed and in-lookup matches exist and adds to in-lookup hash. Returns ERR_PTR() for error, hashed match (in the unlikely case it's been found) or new dentry. In-lookup matches trigger BUG() for now; that will change in the next commit when we introduce waiting for ongoing lookup to finish. Note that in-lookup matches won't be possible until we actually go for shared locking. lookup_slow() switched to use of d_alloc_parallel(). Again, these commits are separated only for making it easier to review. All this machinery will start doing something useful only when we go for shared locking; it's just that the combination is too large for my taste. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 14:42:04 +08:00
if (unlikely(IS_DEADDIR(inode)))
return ERR_PTR(-ENOENT);
parallel lookups machinery, part 3 We will need to be able to check if there is an in-lookup dentry with matching parent/name. Right now it's impossible, but as soon as start locking directories shared such beasts will appear. Add a secondary hash for locating those. Hash chains go through the same space where d_alias will be once it's not in-lookup anymore. Search is done under the same bitlock we use for modifications - with the primary hash we can rely on d_rehash() into the wrong chain being the worst that could happen, but here the pointers are buggered once it's removed from the chain. On the other hand, the chains are not going to be long and normally we'll end up adding to the chain anyway. That allows us to avoid bothering with ->d_lock when doing the comparisons - everything is stable until removed from chain. New helper: d_alloc_parallel(). Right now it allocates, verifies that no hashed and in-lookup matches exist and adds to in-lookup hash. Returns ERR_PTR() for error, hashed match (in the unlikely case it's been found) or new dentry. In-lookup matches trigger BUG() for now; that will change in the next commit when we introduce waiting for ongoing lookup to finish. Note that in-lookup matches won't be possible until we actually go for shared locking. lookup_slow() switched to use of d_alloc_parallel(). Again, these commits are separated only for making it easier to review. All this machinery will start doing something useful only when we go for shared locking; it's just that the combination is too large for my taste. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 14:42:04 +08:00
again:
dentry = d_alloc_parallel(dir, name, &wq);
parallel lookups machinery, part 3 We will need to be able to check if there is an in-lookup dentry with matching parent/name. Right now it's impossible, but as soon as start locking directories shared such beasts will appear. Add a secondary hash for locating those. Hash chains go through the same space where d_alias will be once it's not in-lookup anymore. Search is done under the same bitlock we use for modifications - with the primary hash we can rely on d_rehash() into the wrong chain being the worst that could happen, but here the pointers are buggered once it's removed from the chain. On the other hand, the chains are not going to be long and normally we'll end up adding to the chain anyway. That allows us to avoid bothering with ->d_lock when doing the comparisons - everything is stable until removed from chain. New helper: d_alloc_parallel(). Right now it allocates, verifies that no hashed and in-lookup matches exist and adds to in-lookup hash. Returns ERR_PTR() for error, hashed match (in the unlikely case it's been found) or new dentry. In-lookup matches trigger BUG() for now; that will change in the next commit when we introduce waiting for ongoing lookup to finish. Note that in-lookup matches won't be possible until we actually go for shared locking. lookup_slow() switched to use of d_alloc_parallel(). Again, these commits are separated only for making it easier to review. All this machinery will start doing something useful only when we go for shared locking; it's just that the combination is too large for my taste. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 14:42:04 +08:00
if (IS_ERR(dentry))
return dentry;
parallel lookups machinery, part 3 We will need to be able to check if there is an in-lookup dentry with matching parent/name. Right now it's impossible, but as soon as start locking directories shared such beasts will appear. Add a secondary hash for locating those. Hash chains go through the same space where d_alias will be once it's not in-lookup anymore. Search is done under the same bitlock we use for modifications - with the primary hash we can rely on d_rehash() into the wrong chain being the worst that could happen, but here the pointers are buggered once it's removed from the chain. On the other hand, the chains are not going to be long and normally we'll end up adding to the chain anyway. That allows us to avoid bothering with ->d_lock when doing the comparisons - everything is stable until removed from chain. New helper: d_alloc_parallel(). Right now it allocates, verifies that no hashed and in-lookup matches exist and adds to in-lookup hash. Returns ERR_PTR() for error, hashed match (in the unlikely case it's been found) or new dentry. In-lookup matches trigger BUG() for now; that will change in the next commit when we introduce waiting for ongoing lookup to finish. Note that in-lookup matches won't be possible until we actually go for shared locking. lookup_slow() switched to use of d_alloc_parallel(). Again, these commits are separated only for making it easier to review. All this machinery will start doing something useful only when we go for shared locking; it's just that the combination is too large for my taste. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 14:42:04 +08:00
if (unlikely(!d_in_lookup(dentry))) {
int error = d_revalidate(dentry, flags);
if (unlikely(error <= 0)) {
if (!error) {
d_invalidate(dentry);
dput(dentry);
goto again;
}
dput(dentry);
dentry = ERR_PTR(error);
}
parallel lookups machinery, part 3 We will need to be able to check if there is an in-lookup dentry with matching parent/name. Right now it's impossible, but as soon as start locking directories shared such beasts will appear. Add a secondary hash for locating those. Hash chains go through the same space where d_alias will be once it's not in-lookup anymore. Search is done under the same bitlock we use for modifications - with the primary hash we can rely on d_rehash() into the wrong chain being the worst that could happen, but here the pointers are buggered once it's removed from the chain. On the other hand, the chains are not going to be long and normally we'll end up adding to the chain anyway. That allows us to avoid bothering with ->d_lock when doing the comparisons - everything is stable until removed from chain. New helper: d_alloc_parallel(). Right now it allocates, verifies that no hashed and in-lookup matches exist and adds to in-lookup hash. Returns ERR_PTR() for error, hashed match (in the unlikely case it's been found) or new dentry. In-lookup matches trigger BUG() for now; that will change in the next commit when we introduce waiting for ongoing lookup to finish. Note that in-lookup matches won't be possible until we actually go for shared locking. lookup_slow() switched to use of d_alloc_parallel(). Again, these commits are separated only for making it easier to review. All this machinery will start doing something useful only when we go for shared locking; it's just that the combination is too large for my taste. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-04-15 14:42:04 +08:00
} else {
old = inode->i_op->lookup(inode, dentry, flags);
d_lookup_done(dentry);
if (unlikely(old)) {
dput(dentry);
dentry = old;
}
}
return dentry;
}
static struct dentry *lookup_slow(const struct qstr *name,
struct dentry *dir,
unsigned int flags)
{
struct inode *inode = dir->d_inode;
struct dentry *res;
inode_lock_shared(inode);
res = __lookup_slow(name, dir, flags);
inode_unlock_shared(inode);
return res;
}
static inline int may_lookup(struct mnt_idmap *idmap,
struct nameidata *restrict nd)
{
int err, mask;
mask = nd->flags & LOOKUP_RCU ? MAY_NOT_BLOCK : 0;
err = inode_permission(idmap, nd->inode, mask | MAY_EXEC);
if (likely(!err))
return 0;
// If we failed, and we weren't in LOOKUP_RCU, it's final
if (!(nd->flags & LOOKUP_RCU))
return err;
// Drop out of RCU mode to make sure it wasn't transient
if (!try_to_unlazy(nd))
return -ECHILD; // redo it all non-lazy
if (err != -ECHILD) // hard error
return err;
return inode_permission(idmap, nd->inode, MAY_EXEC);
}
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
static int reserve_stack(struct nameidata *nd, struct path *link)
{
if (unlikely(nd->total_link_count++ >= MAXSYMLINKS))
return -ELOOP;
if (likely(nd->depth != EMBEDDED_LEVELS))
return 0;
if (likely(nd->stack != nd->internal))
return 0;
if (likely(nd_alloc_stack(nd)))
return 0;
if (nd->flags & LOOKUP_RCU) {
// we need to grab link before we do unlazy. And we can't skip
// unlazy even if we fail to grab the link - cleanup needs it
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
bool grabbed_link = legitimize_path(nd, link, nd->next_seq);
if (!try_to_unlazy(nd) || !grabbed_link)
return -ECHILD;
if (nd_alloc_stack(nd))
return 0;
}
return -ENOMEM;
}
enum {WALK_TRAILING = 1, WALK_MORE = 2, WALK_NOFOLLOW = 4};
static const char *pick_link(struct nameidata *nd, struct path *link,
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
struct inode *inode, int flags)
{
struct saved *last;
const char *res;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
int error = reserve_stack(nd, link);
if (unlikely(error)) {
if (!(nd->flags & LOOKUP_RCU))
path_put(link);
return ERR_PTR(error);
}
last = nd->stack + nd->depth++;
last->link = *link;
clear_delayed_call(&last->done);
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
last->seq = nd->next_seq;
if (flags & WALK_TRAILING) {
error = may_follow_link(nd, inode);
if (unlikely(error))
return ERR_PTR(error);
}
if (unlikely(nd->flags & LOOKUP_NO_SYMLINKS) ||
unlikely(link->mnt->mnt_flags & MNT_NOSYMFOLLOW))
return ERR_PTR(-ELOOP);
if (!(nd->flags & LOOKUP_RCU)) {
touch_atime(&last->link);
cond_resched();
} else if (atime_needs_update(&last->link, inode)) {
if (!try_to_unlazy(nd))
return ERR_PTR(-ECHILD);
touch_atime(&last->link);
}
error = security_inode_follow_link(link->dentry, inode,
nd->flags & LOOKUP_RCU);
if (unlikely(error))
return ERR_PTR(error);
res = READ_ONCE(inode->i_link);
if (!res) {
const char * (*get)(struct dentry *, struct inode *,
struct delayed_call *);
get = inode->i_op->get_link;
if (nd->flags & LOOKUP_RCU) {
res = get(NULL, inode, &last->done);
if (res == ERR_PTR(-ECHILD) && try_to_unlazy(nd))
res = get(link->dentry, inode, &last->done);
} else {
res = get(link->dentry, inode, &last->done);
}
if (!res)
goto all_done;
if (IS_ERR(res))
return res;
}
if (*res == '/') {
error = nd_jump_root(nd);
if (unlikely(error))
return ERR_PTR(error);
while (unlikely(*++res == '/'))
;
}
if (*res)
return res;
all_done: // pure jump
put_link(nd);
return NULL;
}
/*
* Do we need to follow links? We _really_ want to be able
* to do this check without having to look at inode->i_op,
* so we keep a cache of "no, this doesn't need follow_link"
* for the common case.
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
*
* NOTE: dentry must be what nd->next_seq had been sampled from.
*/
static const char *step_into(struct nameidata *nd, int flags,
struct dentry *dentry)
{
struct path path;
struct inode *inode;
int err = handle_mounts(nd, dentry, &path);
if (err < 0)
return ERR_PTR(err);
inode = path.dentry->d_inode;
if (likely(!d_is_symlink(path.dentry)) ||
((flags & WALK_TRAILING) && !(nd->flags & LOOKUP_FOLLOW)) ||
(flags & WALK_NOFOLLOW)) {
/* not a symlink or should not follow */
if (nd->flags & LOOKUP_RCU) {
if (read_seqcount_retry(&path.dentry->d_seq, nd->next_seq))
return ERR_PTR(-ECHILD);
if (unlikely(!inode))
return ERR_PTR(-ENOENT);
} else {
dput(nd->path.dentry);
if (nd->path.mnt != path.mnt)
mntput(nd->path.mnt);
}
nd->path = path;
nd->inode = inode;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->seq = nd->next_seq;
return NULL;
}
if (nd->flags & LOOKUP_RCU) {
/* make sure that d_is_symlink above matches inode */
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
if (read_seqcount_retry(&path.dentry->d_seq, nd->next_seq))
return ERR_PTR(-ECHILD);
} else {
if (path.mnt == nd->path.mnt)
mntget(path.mnt);
}
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
return pick_link(nd, &path, inode, flags);
}
static struct dentry *follow_dotdot_rcu(struct nameidata *nd)
{
struct dentry *parent, *old;
if (path_equal(&nd->path, &nd->root))
goto in_root;
if (unlikely(nd->path.dentry == nd->path.mnt->mnt_root)) {
struct path path;
unsigned seq;
if (!choose_mountpoint_rcu(real_mount(nd->path.mnt),
&nd->root, &path, &seq))
goto in_root;
if (unlikely(nd->flags & LOOKUP_NO_XDEV))
return ERR_PTR(-ECHILD);
nd->path = path;
nd->inode = path.dentry->d_inode;
nd->seq = seq;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
// makes sure that non-RCU pathwalk could reach this state
if (read_seqretry(&mount_lock, nd->m_seq))
return ERR_PTR(-ECHILD);
/* we know that mountpoint was pinned */
}
old = nd->path.dentry;
parent = old->d_parent;
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->next_seq = read_seqcount_begin(&parent->d_seq);
// makes sure that non-RCU pathwalk could reach this state
if (read_seqcount_retry(&old->d_seq, nd->seq))
return ERR_PTR(-ECHILD);
if (unlikely(!path_connected(nd->path.mnt, parent)))
return ERR_PTR(-ECHILD);
return parent;
in_root:
if (read_seqretry(&mount_lock, nd->m_seq))
return ERR_PTR(-ECHILD);
if (unlikely(nd->flags & LOOKUP_BENEATH))
return ERR_PTR(-ECHILD);
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->next_seq = nd->seq;
return nd->path.dentry;
}
static struct dentry *follow_dotdot(struct nameidata *nd)
{
struct dentry *parent;
if (path_equal(&nd->path, &nd->root))
goto in_root;
if (unlikely(nd->path.dentry == nd->path.mnt->mnt_root)) {
struct path path;
if (!choose_mountpoint(real_mount(nd->path.mnt),
&nd->root, &path))
goto in_root;
path_put(&nd->path);
nd->path = path;
nd->inode = path.dentry->d_inode;
if (unlikely(nd->flags & LOOKUP_NO_XDEV))
return ERR_PTR(-EXDEV);
}
/* rare case of legitimate dget_parent()... */
parent = dget_parent(nd->path.dentry);
if (unlikely(!path_connected(nd->path.mnt, parent))) {
dput(parent);
return ERR_PTR(-ENOENT);
}
return parent;
in_root:
if (unlikely(nd->flags & LOOKUP_BENEATH))
return ERR_PTR(-EXDEV);
return dget(nd->path.dentry);
}
static const char *handle_dots(struct nameidata *nd, int type)
{
if (type == LAST_DOTDOT) {
const char *error = NULL;
struct dentry *parent;
if (!nd->root.mnt) {
error = ERR_PTR(set_root(nd));
if (error)
return error;
}
if (nd->flags & LOOKUP_RCU)
parent = follow_dotdot_rcu(nd);
else
parent = follow_dotdot(nd);
if (IS_ERR(parent))
return ERR_CAST(parent);
error = step_into(nd, WALK_NOFOLLOW, parent);
if (unlikely(error))
return error;
if (unlikely(nd->flags & LOOKUP_IS_SCOPED)) {
/*
* If there was a racing rename or mount along our
* path, then we can't be sure that ".." hasn't jumped
* above nd->root (and so userspace should retry or use
* some fallback).
*/
smp_rmb();
if (__read_seqcount_retry(&mount_lock.seqcount, nd->m_seq))
return ERR_PTR(-EAGAIN);
if (__read_seqcount_retry(&rename_lock.seqcount, nd->r_seq))
return ERR_PTR(-EAGAIN);
}
}
return NULL;
}
static const char *walk_component(struct nameidata *nd, int flags)
{
struct dentry *dentry;
/*
* "." and ".." are special - ".." especially so because it has
* to be able to know about the current root directory and
* parent relationships.
*/
if (unlikely(nd->last_type != LAST_NORM)) {
if (!(flags & WALK_MORE) && nd->depth)
put_link(nd);
return handle_dots(nd, nd->last_type);
}
dentry = lookup_fast(nd);
if (IS_ERR(dentry))
return ERR_CAST(dentry);
if (unlikely(!dentry)) {
dentry = lookup_slow(&nd->last, nd->path.dentry, nd->flags);
if (IS_ERR(dentry))
return ERR_CAST(dentry);
}
if (!(flags & WALK_MORE) && nd->depth)
put_link(nd);
return step_into(nd, flags, dentry);
}
/*
* We can do the critical dentry name comparison and hashing
* operations one word at a time, but we are limited to:
*
* - Architectures with fast unaligned word accesses. We could
* do a "get_unaligned()" if this helps and is sufficiently
* fast.
*
* - non-CONFIG_DEBUG_PAGEALLOC configurations (so that we
* do not trap on the (extremely unlikely) case of a page
* crossing operation.
*
* - Furthermore, we need an efficient 64-bit compile for the
* 64-bit case in order to generate the "number of bytes in
* the final mask". Again, that could be replaced with a
* efficient population count instruction or similar.
*/
#ifdef CONFIG_DCACHE_WORD_ACCESS
#include <asm/word-at-a-time.h>
#ifdef HASH_MIX
/* Architecture provides HASH_MIX and fold_hash() in <asm/hash.h> */
#elif defined(CONFIG_64BIT)
/*
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
* Register pressure in the mixing function is an issue, particularly
* on 32-bit x86, but almost any function requires one state value and
* one temporary. Instead, use a function designed for two state values
* and no temporaries.
*
* This function cannot create a collision in only two iterations, so
* we have two iterations to achieve avalanche. In those two iterations,
* we have six layers of mixing, which is enough to spread one bit's
* influence out to 2^6 = 64 state bits.
*
* Rotate constants are scored by considering either 64 one-bit input
* deltas or 64*63/2 = 2016 two-bit input deltas, and finding the
* probability of that delta causing a change to each of the 128 output
* bits, using a sample of random initial states.
*
* The Shannon entropy of the computed probabilities is then summed
* to produce a score. Ideally, any input change has a 50% chance of
* toggling any given output bit.
*
* Mixing scores (in bits) for (12,45):
* Input delta: 1-bit 2-bit
* 1 round: 713.3 42542.6
* 2 rounds: 2753.7 140389.8
* 3 rounds: 5954.1 233458.2
* 4 rounds: 7862.6 256672.2
* Perfect: 8192 258048
* (64*128) (64*63/2 * 128)
*/
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
#define HASH_MIX(x, y, a) \
( x ^= (a), \
y ^= x, x = rol64(x,12),\
x += y, y = rol64(y,45),\
y *= 9 )
/*
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
* Fold two longs into one 32-bit hash value. This must be fast, but
* latency isn't quite as critical, as there is a fair bit of additional
* work done before the hash value is used.
*/
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
static inline unsigned int fold_hash(unsigned long x, unsigned long y)
{
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
y ^= x * GOLDEN_RATIO_64;
y *= GOLDEN_RATIO_64;
return y >> 32;
}
#else /* 32-bit case */
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
/*
* Mixing scores (in bits) for (7,20):
* Input delta: 1-bit 2-bit
* 1 round: 330.3 9201.6
* 2 rounds: 1246.4 25475.4
* 3 rounds: 1907.1 31295.1
* 4 rounds: 2042.3 31718.6
* Perfect: 2048 31744
* (32*64) (32*31/2 * 64)
*/
#define HASH_MIX(x, y, a) \
( x ^= (a), \
y ^= x, x = rol32(x, 7),\
x += y, y = rol32(y,20),\
y *= 9 )
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
static inline unsigned int fold_hash(unsigned long x, unsigned long y)
{
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
/* Use arch-optimized multiply if one exists */
return __hash_32(y ^ __hash_32(x));
}
#endif
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
/*
* Return the hash of a string of known length. This is carfully
* designed to match hash_name(), which is the more critical function.
* In particular, we must end by hashing a final word containing 0..7
* payload bytes, to match the way that hash_name() iterates until it
* finds the delimiter after the name.
*/
unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
{
unsigned long a, x = 0, y = (unsigned long)salt;
for (;;) {
if (!len)
goto done;
a = load_unaligned_zeropad(name);
if (len < sizeof(unsigned long))
break;
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
HASH_MIX(x, y, a);
name += sizeof(unsigned long);
len -= sizeof(unsigned long);
}
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
x ^= a & bytemask_from_count(len);
done:
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
return fold_hash(x, y);
}
EXPORT_SYMBOL(full_name_hash);
/* Return the "hash_len" (hash and length) of a null-terminated string */
u64 hashlen_string(const void *salt, const char *name)
{
unsigned long a = 0, x = 0, y = (unsigned long)salt;
unsigned long adata, mask, len;
const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
len = 0;
goto inside;
do {
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
HASH_MIX(x, y, a);
len += sizeof(unsigned long);
inside:
a = load_unaligned_zeropad(name+len);
} while (!has_zero(a, &adata, &constants));
adata = prep_zero_mask(a, adata, &constants);
mask = create_zero_mask(adata);
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
x ^= a & zero_bytemask(mask);
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
return hashlen_create(fold_hash(x, y), len + find_zero(mask));
}
EXPORT_SYMBOL(hashlen_string);
/*
* Calculate the length and hash of the path component, and
* return the length as the result.
*/
static inline const char *hash_name(struct nameidata *nd,
const char *name,
unsigned long *lastword)
{
unsigned long a, b, x, y = (unsigned long)nd->path.dentry;
unsigned long adata, bdata, mask, len;
word-at-a-time: make the interfaces truly generic This changes the interfaces in <asm/word-at-a-time.h> to be a bit more complicated, but a lot more generic. In particular, it allows us to really do the operations efficiently on both little-endian and big-endian machines, pretty much regardless of machine details. For example, if you can rely on a fast population count instruction on your architecture, this will allow you to make your optimized <asm/word-at-a-time.h> file with that. NOTE! The "generic" version in include/asm-generic/word-at-a-time.h is not truly generic, it actually only works on big-endian. Why? Because on little-endian the generic algorithms are wasteful, since you can inevitably do better. The x86 implementation is an example of that. (The only truly non-generic part of the asm-generic implementation is the "find_zero()" function, and you could make a little-endian version of it. And if the Kbuild infrastructure allowed us to pick a particular header file, that would be lovely) The <asm/word-at-a-time.h> functions are as follows: - WORD_AT_A_TIME_CONSTANTS: specific constants that the algorithm uses. - has_zero(): take a word, and determine if it has a zero byte in it. It gets the word, the pointer to the constant pool, and a pointer to an intermediate "data" field it can set. This is the "quick-and-dirty" zero tester: it's what is run inside the hot loops. - "prep_zero_mask()": take the word, the data that has_zero() produced, and the constant pool, and generate an *exact* mask of which byte had the first zero. This is run directly *outside* the loop, and allows the "has_zero()" function to answer the "is there a zero byte" question without necessarily getting exactly *which* byte is the first one to contain a zero. If you do multiple byte lookups concurrently (eg "hash_name()", which looks for both NUL and '/' bytes), after you've done the prep_zero_mask() phase, the result of those can be or'ed together to get the "either or" case. - The result from "prep_zero_mask()" can then be fed into "find_zero()" (to find the byte offset of the first byte that was zero) or into "zero_bytemask()" (to find the bytemask of the bytes preceding the zero byte). The existence of zero_bytemask() is optional, and is not necessary for the normal string routines. But dentry name hashing needs it, so if you enable DENTRY_WORD_AT_A_TIME you need to expose it. This changes the generic strncpy_from_user() function and the dentry hashing functions to use these modified word-at-a-time interfaces. This gets us back to the optimized state of the x86 strncpy that we lost in the previous commit when moving over to the generic version. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-27 01:43:17 +08:00
const struct word_at_a_time constants = WORD_AT_A_TIME_CONSTANTS;
/*
* The first iteration is special, because it can result in
* '.' and '..' and has no mixing other than the final fold.
*/
a = load_unaligned_zeropad(name);
b = a ^ REPEAT_BYTE('/');
if (has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)) {
adata = prep_zero_mask(a, adata, &constants);
bdata = prep_zero_mask(b, bdata, &constants);
mask = create_zero_mask(adata | bdata);
a &= zero_bytemask(mask);
*lastword = a;
len = find_zero(mask);
nd->last.hash = fold_hash(a, y);
nd->last.len = len;
return name + len;
}
len = 0;
x = 0;
do {
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
HASH_MIX(x, y, a);
len += sizeof(unsigned long);
a = load_unaligned_zeropad(name+len);
word-at-a-time: make the interfaces truly generic This changes the interfaces in <asm/word-at-a-time.h> to be a bit more complicated, but a lot more generic. In particular, it allows us to really do the operations efficiently on both little-endian and big-endian machines, pretty much regardless of machine details. For example, if you can rely on a fast population count instruction on your architecture, this will allow you to make your optimized <asm/word-at-a-time.h> file with that. NOTE! The "generic" version in include/asm-generic/word-at-a-time.h is not truly generic, it actually only works on big-endian. Why? Because on little-endian the generic algorithms are wasteful, since you can inevitably do better. The x86 implementation is an example of that. (The only truly non-generic part of the asm-generic implementation is the "find_zero()" function, and you could make a little-endian version of it. And if the Kbuild infrastructure allowed us to pick a particular header file, that would be lovely) The <asm/word-at-a-time.h> functions are as follows: - WORD_AT_A_TIME_CONSTANTS: specific constants that the algorithm uses. - has_zero(): take a word, and determine if it has a zero byte in it. It gets the word, the pointer to the constant pool, and a pointer to an intermediate "data" field it can set. This is the "quick-and-dirty" zero tester: it's what is run inside the hot loops. - "prep_zero_mask()": take the word, the data that has_zero() produced, and the constant pool, and generate an *exact* mask of which byte had the first zero. This is run directly *outside* the loop, and allows the "has_zero()" function to answer the "is there a zero byte" question without necessarily getting exactly *which* byte is the first one to contain a zero. If you do multiple byte lookups concurrently (eg "hash_name()", which looks for both NUL and '/' bytes), after you've done the prep_zero_mask() phase, the result of those can be or'ed together to get the "either or" case. - The result from "prep_zero_mask()" can then be fed into "find_zero()" (to find the byte offset of the first byte that was zero) or into "zero_bytemask()" (to find the bytemask of the bytes preceding the zero byte). The existence of zero_bytemask() is optional, and is not necessary for the normal string routines. But dentry name hashing needs it, so if you enable DENTRY_WORD_AT_A_TIME you need to expose it. This changes the generic strncpy_from_user() function and the dentry hashing functions to use these modified word-at-a-time interfaces. This gets us back to the optimized state of the x86 strncpy that we lost in the previous commit when moving over to the generic version. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-27 01:43:17 +08:00
b = a ^ REPEAT_BYTE('/');
} while (!(has_zero(a, &adata, &constants) | has_zero(b, &bdata, &constants)));
adata = prep_zero_mask(a, adata, &constants);
bdata = prep_zero_mask(b, bdata, &constants);
mask = create_zero_mask(adata | bdata);
a &= zero_bytemask(mask);
x ^= a;
len += find_zero(mask);
*lastword = 0; // Multi-word components cannot be DOT or DOTDOT
word-at-a-time: make the interfaces truly generic This changes the interfaces in <asm/word-at-a-time.h> to be a bit more complicated, but a lot more generic. In particular, it allows us to really do the operations efficiently on both little-endian and big-endian machines, pretty much regardless of machine details. For example, if you can rely on a fast population count instruction on your architecture, this will allow you to make your optimized <asm/word-at-a-time.h> file with that. NOTE! The "generic" version in include/asm-generic/word-at-a-time.h is not truly generic, it actually only works on big-endian. Why? Because on little-endian the generic algorithms are wasteful, since you can inevitably do better. The x86 implementation is an example of that. (The only truly non-generic part of the asm-generic implementation is the "find_zero()" function, and you could make a little-endian version of it. And if the Kbuild infrastructure allowed us to pick a particular header file, that would be lovely) The <asm/word-at-a-time.h> functions are as follows: - WORD_AT_A_TIME_CONSTANTS: specific constants that the algorithm uses. - has_zero(): take a word, and determine if it has a zero byte in it. It gets the word, the pointer to the constant pool, and a pointer to an intermediate "data" field it can set. This is the "quick-and-dirty" zero tester: it's what is run inside the hot loops. - "prep_zero_mask()": take the word, the data that has_zero() produced, and the constant pool, and generate an *exact* mask of which byte had the first zero. This is run directly *outside* the loop, and allows the "has_zero()" function to answer the "is there a zero byte" question without necessarily getting exactly *which* byte is the first one to contain a zero. If you do multiple byte lookups concurrently (eg "hash_name()", which looks for both NUL and '/' bytes), after you've done the prep_zero_mask() phase, the result of those can be or'ed together to get the "either or" case. - The result from "prep_zero_mask()" can then be fed into "find_zero()" (to find the byte offset of the first byte that was zero) or into "zero_bytemask()" (to find the bytemask of the bytes preceding the zero byte). The existence of zero_bytemask() is optional, and is not necessary for the normal string routines. But dentry name hashing needs it, so if you enable DENTRY_WORD_AT_A_TIME you need to expose it. This changes the generic strncpy_from_user() function and the dentry hashing functions to use these modified word-at-a-time interfaces. This gets us back to the optimized state of the x86 strncpy that we lost in the previous commit when moving over to the generic version. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-05-27 01:43:17 +08:00
nd->last.hash = fold_hash(x, y);
nd->last.len = len;
return name + len;
}
/*
* Note that the 'last' word is always zero-masked, but
* was loaded as a possibly big-endian word.
*/
#ifdef __BIG_ENDIAN
#define LAST_WORD_IS_DOT (0x2eul << (BITS_PER_LONG-8))
#define LAST_WORD_IS_DOTDOT (0x2e2eul << (BITS_PER_LONG-16))
#endif
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
#else /* !CONFIG_DCACHE_WORD_ACCESS: Slow, byte-at-a-time version */
/* Return the hash of a string of known length */
unsigned int full_name_hash(const void *salt, const char *name, unsigned int len)
{
unsigned long hash = init_name_hash(salt);
while (len--)
hash = partial_name_hash((unsigned char)*name++, hash);
return end_name_hash(hash);
}
EXPORT_SYMBOL(full_name_hash);
/* Return the "hash_len" (hash and length) of a null-terminated string */
u64 hashlen_string(const void *salt, const char *name)
{
unsigned long hash = init_name_hash(salt);
unsigned long len = 0, c;
c = (unsigned char)*name;
while (c) {
len++;
hash = partial_name_hash(c, hash);
c = (unsigned char)name[len];
}
return hashlen_create(end_name_hash(hash), len);
}
EXPORT_SYMBOL(hashlen_string);
/*
* We know there's a real path component here of at least
* one character.
*/
static inline const char *hash_name(struct nameidata *nd, const char *name, unsigned long *lastword)
{
unsigned long hash = init_name_hash(nd->path.dentry);
unsigned long len = 0, c, last = 0;
c = (unsigned char)*name;
do {
last = (last << 8) + c;
len++;
hash = partial_name_hash(c, hash);
c = (unsigned char)name[len];
} while (c && c != '/');
// This is reliable for DOT or DOTDOT, since the component
// cannot contain NUL characters - top bits being zero means
// we cannot have had any other pathnames.
*lastword = last;
nd->last.hash = end_name_hash(hash);
nd->last.len = len;
return name + len;
}
#endif
#ifndef LAST_WORD_IS_DOT
#define LAST_WORD_IS_DOT 0x2e
#define LAST_WORD_IS_DOTDOT 0x2e2e
#endif
/*
* Name resolution.
* This is the basic name resolution function, turning a pathname into
* the final dentry. We expect 'base' to be positive and a directory.
*
* Returns 0 and nd will have valid dentry and mnt on success.
* Returns error and drops reference to input namei data on failure.
*/
static int link_path_walk(const char *name, struct nameidata *nd)
{
link_path_walk(): simplify stack handling We use nd->stack to store two things: pinning down the symlinks we are resolving and resuming the name traversal when a nested symlink is finished. Currently, nd->depth is used to keep track of both. It's 0 when we call link_path_walk() for the first time (for the pathname itself) and 1 on all subsequent calls (for trailing symlinks, if any). That's fine, as far as pinning symlinks goes - when handling a trailing symlink, the string we are interpreting is the body of symlink pinned down in nd->stack[0]. It's rather inconvenient with respect to handling nested symlinks, though - when we run out of a string we are currently interpreting, we need to decide whether it's a nested symlink (in which case we need to pick the string saved back when we started to interpret that nested symlink and resume its traversal) or not (in which case we are done with link_path_walk()). Current solution is a bit of a kludge - in handling of trailing symlink (in lookup_last() and open_last_lookups() we clear nd->stack[0].name. That allows link_path_walk() to use the following rules when running out of a string to interpret: * if nd->depth is zero, we are at the end of pathname itself. * if nd->depth is positive, check the saved string; for nested symlink it will be non-NULL, for trailing symlink - NULL. It works, but it's rather non-obvious. Note that we have two sets: the set of symlinks currently being traversed and the set of postponed pathname tails. The former is stored in nd->stack[0..nd->depth-1].link and it's valid throught the pathname resolution; the latter is valid only during an individual call of link_path_walk() and it occupies nd->stack[0..nd->depth-1].name for the first call of link_path_walk() and nd->stack[1..nd->depth-1].name for subsequent ones. The kludge is basically a way to recognize the second set becoming empty. The things get simpler if we keep track of the second set's size explicitly and always store it in nd->stack[0..depth-1].name. We access the second set only inside link_path_walk(), so its size can live in a local variable; that way the check becomes trivial without the need of that kludge. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-02-24 11:04:15 +08:00
int depth = 0; // depth <= nd->depth
int err;
sanitize handling of nd->last_type, kill LAST_BIND ->last_type values are set in 3 places: path_init() (sets to LAST_ROOT), link_path_walk (LAST_NORM/DOT/DOTDOT) and pick_link (LAST_BIND). The are checked in walk_component(), lookup_last() and do_last(). They also get copied to the caller by filename_parentat(). In the last 3 cases the value is what we had at the return from link_path_walk(). In case of walk_component() it's either directly downstream from assignment in link_path_walk() or, when called by lookup_last(), the value we have at the return from link_path_walk(). The value at the entry into link_path_walk() can survive to return only if the pathname contains nothing but slashes. Note that pick_link() never returns such - pure jumps are handled directly. So for the calls of link_path_walk() for trailing symlinks it does not matter what value had been there at the entry; the value at the return won't depend upon it. There are 3 call chains that might have pick_link() storing LAST_BIND: 1) pick_link() from step_into() from walk_component() from link_path_walk(). In that case we will either be parsing the next component immediately after return into link_path_walk(), which will overwrite the ->last_type before anyone has a chance to look at it, or we'll fail, in which case nobody will be looking at ->last_type at all. 2) pick_link() from step_into() from walk_component() from lookup_last(). The value is never looked at due to the above; it won't affect the value seen at return from any link_path_walk(). 3) pick_link() from step_into() from do_last(). Ditto. In other words, assignemnt in pick_link() is pointless, and so is LAST_BIND itself; nothing ever looks at that value. Kill it off. And make link_path_walk() _always_ assign ->last_type - in the only case when the value at the entry might survive to the return that value is always LAST_ROOT, inherited from path_init(). Move that assignment from path_init() into the beginning of link_path_walk(), to consolidate the things. Historical note: LAST_BIND used to be used for the kludge with trailing pure jump symlinks (extra iteration through the top-level loop). No point keeping it anymore... Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-01-20 00:44:51 +08:00
nd->last_type = LAST_ROOT;
nd->flags |= LOOKUP_PARENT;
if (IS_ERR(name))
return PTR_ERR(name);
while (*name=='/')
name++;
if (!*name) {
nd->dir_mode = 0; // short-circuit the 'hardening' idiocy
return 0;
}
/* At this point we know we have a real path component. */
for(;;) {
struct mnt_idmap *idmap;
const char *link;
unsigned long lastword;
idmap = mnt_idmap(nd->path.mnt);
err = may_lookup(idmap, nd);
fs/namei.c: Improve dcache hash function Patch 0fed3ac866 improved the hash mixing, but the function is slower than necessary; there's a 7-instruction dependency chain (10 on x86) each loop iteration. Word-at-a-time access is a very tight loop (which is good, because link_path_walk() is one of the hottest code paths in the entire kernel), and the hash mixing function must not have a longer latency to avoid slowing it down. There do not appear to be any published fast hash functions that: 1) Operate on the input a word at a time, and 2) Don't need to know the length of the input beforehand, and 3) Have a single iterated mixing function, not needing conditional branches or unrolling to distinguish different loop iterations. One of the algorithms which comes closest is Yann Collet's xxHash, but that's two dependent multiplies per word, which is too much. The key insights in this design are: 1) Barring expensive ops like multiplies, to diffuse one input bit across 64 bits of hash state takes at least log2(64) = 6 sequentially dependent instructions. That is more cycles than we'd like. 2) An operation like "hash ^= hash << 13" requires a second temporary register anyway, and on a 2-operand machine like x86, it's three instructions. 3) A better use of a second register is to hold a two-word hash state. With careful design, no temporaries are needed at all, so it doesn't increase register pressure. And this gets rid of register copying on 2-operand machines, so the code is smaller and faster. 4) Using two words of state weakens the requirement for one-round mixing; we now have two rounds of mixing before cancellation is possible. 5) A two-word hash state also allows operations on both halves to be done in parallel, so on a superscalar processor we get more mixing in fewer cycles. I ended up using a mixing function inspired by the ChaCha and Speck round functions. It is 6 simple instructions and 3 cycles per iteration (assuming multiply by 9 can be done by an "lea" instruction): x ^= *input++; y ^= x; x = ROL(x, K1); x += y; y = ROL(y, K2); y *= 9; Not only is this reversible, two consecutive rounds are reversible: if you are given the initial and final states, but not the intermediate state, it is possible to compute both input words. This means that at least 3 words of input are required to create a collision. (It also has the property, used by hash_name() to avoid a branch, that it hashes all-zero to all-zero.) The rotate constants K1 and K2 were found by experiment. The search took a sample of random initial states (I used 1023) and considered the effect of flipping each of the 64 input bits on each of the 128 output bits two rounds later. Each of the 8192 pairs can be considered a biased coin, and adding up the Shannon entropy of all of them produces a score. The best-scoring shifts also did well in other tests (flipping bits in y, trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits), so the choice was made with the additional constraint that the sum of the shifts is odd and not too close to the word size. The final state is then folded into a 32-bit hash value by a less carefully optimized multiply-based scheme. This also has to be fast, as pathname components tend to be short (the most common case is one iteration!), but there's some room for latency, as there is a fair bit of intervening logic before the hash value is used for anything. (Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs. I need a better benchmark; the numbers seem to show a slight dip in performance between 4.6.0 and this patch, but they're too noisy to quote.) Special thanks to Bruce fields for diligent testing which uncovered a nasty fencepost error in an earlier version of this patch. [checkpatch.pl formatting complaints noted and respectfully disagreed with.] Signed-off-by: George Spelvin <linux@sciencehorizons.net> Tested-by: J. Bruce Fields <bfields@redhat.com>
2016-05-23 19:43:58 +08:00
if (err)
return err;
nd->last.name = name;
name = hash_name(nd, name, &lastword);
switch(lastword) {
case LAST_WORD_IS_DOTDOT:
nd->last_type = LAST_DOTDOT;
nd->state |= ND_JUMPED;
break;
case LAST_WORD_IS_DOT:
nd->last_type = LAST_DOT;
break;
default:
nd->last_type = LAST_NORM;
nd->state &= ~ND_JUMPED;
struct dentry *parent = nd->path.dentry;
if (unlikely(parent->d_flags & DCACHE_OP_HASH)) {
err = parent->d_op->d_hash(parent, &nd->last);
if (err < 0)
return err;
}
}
if (!*name)
goto OK;
/*
* If it wasn't NUL, we know it was '/'. Skip that
* slash, and continue until no more slashes.
*/
do {
name++;
} while (unlikely(*name == '/'));
if (unlikely(!*name)) {
OK:
link_path_walk(): simplify stack handling We use nd->stack to store two things: pinning down the symlinks we are resolving and resuming the name traversal when a nested symlink is finished. Currently, nd->depth is used to keep track of both. It's 0 when we call link_path_walk() for the first time (for the pathname itself) and 1 on all subsequent calls (for trailing symlinks, if any). That's fine, as far as pinning symlinks goes - when handling a trailing symlink, the string we are interpreting is the body of symlink pinned down in nd->stack[0]. It's rather inconvenient with respect to handling nested symlinks, though - when we run out of a string we are currently interpreting, we need to decide whether it's a nested symlink (in which case we need to pick the string saved back when we started to interpret that nested symlink and resume its traversal) or not (in which case we are done with link_path_walk()). Current solution is a bit of a kludge - in handling of trailing symlink (in lookup_last() and open_last_lookups() we clear nd->stack[0].name. That allows link_path_walk() to use the following rules when running out of a string to interpret: * if nd->depth is zero, we are at the end of pathname itself. * if nd->depth is positive, check the saved string; for nested symlink it will be non-NULL, for trailing symlink - NULL. It works, but it's rather non-obvious. Note that we have two sets: the set of symlinks currently being traversed and the set of postponed pathname tails. The former is stored in nd->stack[0..nd->depth-1].link and it's valid throught the pathname resolution; the latter is valid only during an individual call of link_path_walk() and it occupies nd->stack[0..nd->depth-1].name for the first call of link_path_walk() and nd->stack[1..nd->depth-1].name for subsequent ones. The kludge is basically a way to recognize the second set becoming empty. The things get simpler if we keep track of the second set's size explicitly and always store it in nd->stack[0..depth-1].name. We access the second set only inside link_path_walk(), so its size can live in a local variable; that way the check becomes trivial without the need of that kludge. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-02-24 11:04:15 +08:00
/* pathname or trailing symlink, done */
if (!depth) {
nd->dir_vfsuid = i_uid_into_vfsuid(idmap, nd->inode);
nd->dir_mode = nd->inode->i_mode;
nd->flags &= ~LOOKUP_PARENT;
return 0;
}
/* last component of nested symlink */
link_path_walk(): simplify stack handling We use nd->stack to store two things: pinning down the symlinks we are resolving and resuming the name traversal when a nested symlink is finished. Currently, nd->depth is used to keep track of both. It's 0 when we call link_path_walk() for the first time (for the pathname itself) and 1 on all subsequent calls (for trailing symlinks, if any). That's fine, as far as pinning symlinks goes - when handling a trailing symlink, the string we are interpreting is the body of symlink pinned down in nd->stack[0]. It's rather inconvenient with respect to handling nested symlinks, though - when we run out of a string we are currently interpreting, we need to decide whether it's a nested symlink (in which case we need to pick the string saved back when we started to interpret that nested symlink and resume its traversal) or not (in which case we are done with link_path_walk()). Current solution is a bit of a kludge - in handling of trailing symlink (in lookup_last() and open_last_lookups() we clear nd->stack[0].name. That allows link_path_walk() to use the following rules when running out of a string to interpret: * if nd->depth is zero, we are at the end of pathname itself. * if nd->depth is positive, check the saved string; for nested symlink it will be non-NULL, for trailing symlink - NULL. It works, but it's rather non-obvious. Note that we have two sets: the set of symlinks currently being traversed and the set of postponed pathname tails. The former is stored in nd->stack[0..nd->depth-1].link and it's valid throught the pathname resolution; the latter is valid only during an individual call of link_path_walk() and it occupies nd->stack[0..nd->depth-1].name for the first call of link_path_walk() and nd->stack[1..nd->depth-1].name for subsequent ones. The kludge is basically a way to recognize the second set becoming empty. The things get simpler if we keep track of the second set's size explicitly and always store it in nd->stack[0..depth-1].name. We access the second set only inside link_path_walk(), so its size can live in a local variable; that way the check becomes trivial without the need of that kludge. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-02-24 11:04:15 +08:00
name = nd->stack[--depth].name;
link = walk_component(nd, 0);
} else {
/* not the last component */
link = walk_component(nd, WALK_MORE);
}
if (unlikely(link)) {
if (IS_ERR(link))
return PTR_ERR(link);
/* a symlink to follow */
link_path_walk(): simplify stack handling We use nd->stack to store two things: pinning down the symlinks we are resolving and resuming the name traversal when a nested symlink is finished. Currently, nd->depth is used to keep track of both. It's 0 when we call link_path_walk() for the first time (for the pathname itself) and 1 on all subsequent calls (for trailing symlinks, if any). That's fine, as far as pinning symlinks goes - when handling a trailing symlink, the string we are interpreting is the body of symlink pinned down in nd->stack[0]. It's rather inconvenient with respect to handling nested symlinks, though - when we run out of a string we are currently interpreting, we need to decide whether it's a nested symlink (in which case we need to pick the string saved back when we started to interpret that nested symlink and resume its traversal) or not (in which case we are done with link_path_walk()). Current solution is a bit of a kludge - in handling of trailing symlink (in lookup_last() and open_last_lookups() we clear nd->stack[0].name. That allows link_path_walk() to use the following rules when running out of a string to interpret: * if nd->depth is zero, we are at the end of pathname itself. * if nd->depth is positive, check the saved string; for nested symlink it will be non-NULL, for trailing symlink - NULL. It works, but it's rather non-obvious. Note that we have two sets: the set of symlinks currently being traversed and the set of postponed pathname tails. The former is stored in nd->stack[0..nd->depth-1].link and it's valid throught the pathname resolution; the latter is valid only during an individual call of link_path_walk() and it occupies nd->stack[0..nd->depth-1].name for the first call of link_path_walk() and nd->stack[1..nd->depth-1].name for subsequent ones. The kludge is basically a way to recognize the second set becoming empty. The things get simpler if we keep track of the second set's size explicitly and always store it in nd->stack[0..depth-1].name. We access the second set only inside link_path_walk(), so its size can live in a local variable; that way the check becomes trivial without the need of that kludge. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-02-24 11:04:15 +08:00
nd->stack[depth++].name = name;
name = link;
continue;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
}
if (unlikely(!d_can_lookup(nd->path.dentry))) {
if (nd->flags & LOOKUP_RCU) {
if (!try_to_unlazy(nd))
return -ECHILD;
}
return -ENOTDIR;
}
}
}
/* must be paired with terminate_walk() */
static const char *path_init(struct nameidata *nd, unsigned flags)
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
{
int error;
const char *s = nd->name->name;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
/* LOOKUP_CACHED requires RCU, ask caller to retry */
if ((flags & (LOOKUP_RCU | LOOKUP_CACHED)) == LOOKUP_CACHED)
return ERR_PTR(-EAGAIN);
if (!*s)
flags &= ~LOOKUP_RCU;
if (flags & LOOKUP_RCU)
rcu_read_lock();
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
else
nd->seq = nd->next_seq = 0;
nd->flags = flags;
nd->state |= ND_JUMPED;
namei: LOOKUP_{IN_ROOT,BENEATH}: permit limited ".." resolution Allow LOOKUP_BENEATH and LOOKUP_IN_ROOT to safely permit ".." resolution (in the case of LOOKUP_BENEATH the resolution will still fail if ".." resolution would resolve a path outside of the root -- while LOOKUP_IN_ROOT will chroot(2)-style scope it). Magic-link jumps are still disallowed entirely[*]. As Jann explains[1,2], the need for this patch (and the original no-".." restriction) is explained by observing there is a fairly easy-to-exploit race condition with chroot(2) (and thus by extension LOOKUP_IN_ROOT and LOOKUP_BENEATH if ".." is allowed) where a rename(2) of a path can be used to "skip over" nd->root and thus escape to the filesystem above nd->root. thread1 [attacker]: for (;;) renameat2(AT_FDCWD, "/a/b/c", AT_FDCWD, "/a/d", RENAME_EXCHANGE); thread2 [victim]: for (;;) openat2(dirb, "b/c/../../etc/shadow", { .flags = O_PATH, .resolve = RESOLVE_IN_ROOT } ); With fairly significant regularity, thread2 will resolve to "/etc/shadow" rather than "/a/b/etc/shadow". There is also a similar (though somewhat more privileged) attack using MS_MOVE. With this patch, such cases will be detected *during* ".." resolution and will return -EAGAIN for userspace to decide to either retry or abort the lookup. It should be noted that ".." is the weak point of chroot(2) -- walking *into* a subdirectory tautologically cannot result in you walking *outside* nd->root (except through a bind-mount or magic-link). There is also no other way for a directory's parent to change (which is the primary worry with ".." resolution here) other than a rename or MS_MOVE. The primary reason for deferring to userspace with -EAGAIN is that an in-kernel retry loop (or doing a path_is_under() check after re-taking the relevant seqlocks) can become unreasonably expensive on machines with lots of VFS activity (nfsd can cause lots of rename_lock updates). Thus it should be up to userspace how many times they wish to retry the lookup -- the selftests for this attack indicate that there is a ~35% chance of the lookup succeeding on the first try even with an attacker thrashing rename_lock. A variant of the above attack is included in the selftests for openat2(2) later in this patch series. I've run this test on several machines for several days and no instances of a breakout were detected. While this is not concrete proof that this is safe, when combined with the above argument it should lend some trustworthiness to this construction. [*] It may be acceptable in the future to do a path_is_under() check for magic-links after they are resolved. However this seems unlikely to be a feature that people *really* need -- it can be added later if it turns out a lot of people want it. [1]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ [2]: https://lore.kernel.org/lkml/CAG48ez30WJhbsro2HOc_DR7V91M+hNFzBP5ogRMZaxbAORvqzg@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: Jann Horn <jannh@google.com> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:35 +08:00
nd->m_seq = __read_seqcount_begin(&mount_lock.seqcount);
nd->r_seq = __read_seqcount_begin(&rename_lock.seqcount);
smp_rmb();
if (nd->state & ND_ROOT_PRESET) {
struct dentry *root = nd->root.dentry;
struct inode *inode = root->d_inode;
if (*s && unlikely(!d_can_lookup(root)))
return ERR_PTR(-ENOTDIR);
nd->path = nd->root;
nd->inode = inode;
if (flags & LOOKUP_RCU) {
namei: LOOKUP_{IN_ROOT,BENEATH}: permit limited ".." resolution Allow LOOKUP_BENEATH and LOOKUP_IN_ROOT to safely permit ".." resolution (in the case of LOOKUP_BENEATH the resolution will still fail if ".." resolution would resolve a path outside of the root -- while LOOKUP_IN_ROOT will chroot(2)-style scope it). Magic-link jumps are still disallowed entirely[*]. As Jann explains[1,2], the need for this patch (and the original no-".." restriction) is explained by observing there is a fairly easy-to-exploit race condition with chroot(2) (and thus by extension LOOKUP_IN_ROOT and LOOKUP_BENEATH if ".." is allowed) where a rename(2) of a path can be used to "skip over" nd->root and thus escape to the filesystem above nd->root. thread1 [attacker]: for (;;) renameat2(AT_FDCWD, "/a/b/c", AT_FDCWD, "/a/d", RENAME_EXCHANGE); thread2 [victim]: for (;;) openat2(dirb, "b/c/../../etc/shadow", { .flags = O_PATH, .resolve = RESOLVE_IN_ROOT } ); With fairly significant regularity, thread2 will resolve to "/etc/shadow" rather than "/a/b/etc/shadow". There is also a similar (though somewhat more privileged) attack using MS_MOVE. With this patch, such cases will be detected *during* ".." resolution and will return -EAGAIN for userspace to decide to either retry or abort the lookup. It should be noted that ".." is the weak point of chroot(2) -- walking *into* a subdirectory tautologically cannot result in you walking *outside* nd->root (except through a bind-mount or magic-link). There is also no other way for a directory's parent to change (which is the primary worry with ".." resolution here) other than a rename or MS_MOVE. The primary reason for deferring to userspace with -EAGAIN is that an in-kernel retry loop (or doing a path_is_under() check after re-taking the relevant seqlocks) can become unreasonably expensive on machines with lots of VFS activity (nfsd can cause lots of rename_lock updates). Thus it should be up to userspace how many times they wish to retry the lookup -- the selftests for this attack indicate that there is a ~35% chance of the lookup succeeding on the first try even with an attacker thrashing rename_lock. A variant of the above attack is included in the selftests for openat2(2) later in this patch series. I've run this test on several machines for several days and no instances of a breakout were detected. While this is not concrete proof that this is safe, when combined with the above argument it should lend some trustworthiness to this construction. [*] It may be acceptable in the future to do a path_is_under() check for magic-links after they are resolved. However this seems unlikely to be a feature that people *really* need -- it can be added later if it turns out a lot of people want it. [1]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ [2]: https://lore.kernel.org/lkml/CAG48ez30WJhbsro2HOc_DR7V91M+hNFzBP5ogRMZaxbAORvqzg@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: Jann Horn <jannh@google.com> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:35 +08:00
nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
nd->root_seq = nd->seq;
} else {
path_get(&nd->path);
}
return s;
}
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
nd->root.mnt = NULL;
namei: LOOKUP_IN_ROOT: chroot-like scoped resolution /* Background. */ Container runtimes or other administrative management processes will often interact with root filesystems while in the host mount namespace, because the cost of doing a chroot(2) on every operation is too prohibitive (especially in Go, which cannot safely use vfork). However, a malicious program can trick the management process into doing operations on files outside of the root filesystem through careful crafting of symlinks. Most programs that need this feature have attempted to make this process safe, by doing all of the path resolution in userspace (with symlinks being scoped to the root of the malicious root filesystem). Unfortunately, this method is prone to foot-guns and usually such implementations have subtle security bugs. Thus, what userspace needs is a way to resolve a path as though it were in a chroot(2) -- with all absolute symlinks being resolved relative to the dirfd root (and ".." components being stuck under the dirfd root). It is much simpler and more straight-forward to provide this functionality in-kernel (because it can be done far more cheaply and correctly). More classical applications that also have this problem (which have their own potentially buggy userspace path sanitisation code) include web servers, archive extraction tools, network file servers, and so on. /* Userspace API. */ LOOKUP_IN_ROOT will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_IN_ROOT applies to all components of the path. With LOOKUP_IN_ROOT, any path component which attempts to cross the starting point of the pathname lookup (the dirfd passed to openat) will remain at the starting point. Thus, all absolute paths and symlinks will be scoped within the starting point. There is a slight change in behaviour regarding pathnames -- if the pathname is absolute then the dirfd is still used as the root of resolution of LOOKUP_IN_ROOT is specified (this is to avoid obvious foot-guns, at the cost of a minor API inconsistency). As with LOOKUP_BENEATH, Jann's security concern about ".."[1] applies to LOOKUP_IN_ROOT -- therefore ".." resolution is blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_IN_ROOT is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:34 +08:00
/* Absolute pathname -- fetch the root (LOOKUP_IN_ROOT uses nd->dfd). */
if (*s == '/' && !(flags & LOOKUP_IN_ROOT)) {
error = nd_jump_root(nd);
if (unlikely(error))
return ERR_PTR(error);
return s;
namei: LOOKUP_IN_ROOT: chroot-like scoped resolution /* Background. */ Container runtimes or other administrative management processes will often interact with root filesystems while in the host mount namespace, because the cost of doing a chroot(2) on every operation is too prohibitive (especially in Go, which cannot safely use vfork). However, a malicious program can trick the management process into doing operations on files outside of the root filesystem through careful crafting of symlinks. Most programs that need this feature have attempted to make this process safe, by doing all of the path resolution in userspace (with symlinks being scoped to the root of the malicious root filesystem). Unfortunately, this method is prone to foot-guns and usually such implementations have subtle security bugs. Thus, what userspace needs is a way to resolve a path as though it were in a chroot(2) -- with all absolute symlinks being resolved relative to the dirfd root (and ".." components being stuck under the dirfd root). It is much simpler and more straight-forward to provide this functionality in-kernel (because it can be done far more cheaply and correctly). More classical applications that also have this problem (which have their own potentially buggy userspace path sanitisation code) include web servers, archive extraction tools, network file servers, and so on. /* Userspace API. */ LOOKUP_IN_ROOT will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_IN_ROOT applies to all components of the path. With LOOKUP_IN_ROOT, any path component which attempts to cross the starting point of the pathname lookup (the dirfd passed to openat) will remain at the starting point. Thus, all absolute paths and symlinks will be scoped within the starting point. There is a slight change in behaviour regarding pathnames -- if the pathname is absolute then the dirfd is still used as the root of resolution of LOOKUP_IN_ROOT is specified (this is to avoid obvious foot-guns, at the cost of a minor API inconsistency). As with LOOKUP_BENEATH, Jann's security concern about ".."[1] applies to LOOKUP_IN_ROOT -- therefore ".." resolution is blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_IN_ROOT is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:34 +08:00
}
/* Relative pathname -- get the starting-point it is relative to. */
if (nd->dfd == AT_FDCWD) {
if (flags & LOOKUP_RCU) {
struct fs_struct *fs = current->fs;
unsigned seq;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
do {
seq = read_seqcount_begin(&fs->seq);
nd->path = fs->pwd;
nd->inode = nd->path.dentry->d_inode;
nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
} while (read_seqcount_retry(&fs->seq, seq));
} else {
get_fs_pwd(current->fs, &nd->path);
nd->inode = nd->path.dentry->d_inode;
}
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
} else {
/* Caller must check execute permissions on the starting path component */
struct fd f = fdget_raw(nd->dfd);
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
struct dentry *dentry;
if (!f.file)
return ERR_PTR(-EBADF);
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
vfs: relax linkat() AT_EMPTY_PATH - aka flink() - requirements "The definition of insanity is doing the same thing over and over again and expecting different results” We've tried to do this before, most recently with commit bb2314b47996 ("fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink") about a decade ago. But the effort goes back even further than that, eg this thread back from 1998 that is so old that we don't even have it archived in lore: https://lkml.org/lkml/1998/3/10/108 which also points out some of the reasons why it's dangerous. Or, how about then in 2003: https://lkml.org/lkml/2003/4/6/112 where we went through some of the same arguments, just wirh different people involved. In particular, having access to a file descriptor does not necessarily mean that you have access to the path that was used for lookup, and there may be very good reasons why you absolutely must not have access to a path to said file. For example, if we were passed a file descriptor from the outside into some limited environment (think chroot, but also user namespaces etc) a 'flink()' system call could now make that file visible inside a context where it's not supposed to be visible. In the process the user may also be able to re-open it with permissions that the original file descriptor did not have (eg a read-only file descriptor may be associated with an underlying file that is writable). Another variation on this is if somebody else (typically root) opens a file in a directory that is not accessible to others, and passes the file descriptor on as a read-only file. Again, the access to the file descriptor does not imply that you should have access to a path to the file in the filesystem. So while we have tried this several times in the past, it never works. The last time we did this, that commit bb2314b47996 quickly got reverted again in commit f0cc6ffb8ce8 (Revert "fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink"), with a note saying "We may re-do this once the whole discussion about the interface is done". Well, the discussion is long done, and didn't come to any resolution. There's no question that 'flink()' would be a useful operation, but it's a dangerous one. However, it does turn out that since 2008 (commit d76b0d9b2d87: "CRED: Use creds in file structs") we have had a fairly straightforward way to check whether the file descriptor was opened by the same credentials as the credentials of the flink(). That allows the most common patterns that people want to use, which tend to be to either open the source carefully (ie using the openat2() RESOLVE_xyz flags, and/or checking ownership with fstat() before linking), or to use O_TMPFILE and fill in the file contents before it's exposed to the world with linkat(). But it also means that if the file descriptor was opened by somebody else, or we've gone through a credentials change since, the operation no longer works (unless we have CAP_DAC_READ_SEARCH capabilities in the opener's user namespace, as before). Note that the credential equality check is done by using pointer equality, which means that it's not enough that you have effectively the same user - they have to be literally identical, since our credentials are using copy-on-write semantics. So you can't change your credentials to something else and try to change it back to the same ones between the open() and the linkat(). This is not meant to be some kind of generic permission check, this is literally meant as a "the open and link calls are 'atomic' wrt user credentials" check. It also means that you can't just move things between namespaces, because the credentials aren't just a list of uid's and gid's: they includes the pointer to the user_ns that the capabilities are relative to. So let's try this one more time and see if maybe this approach ends up being workable after all. Cc: Andrew Lutomirski <luto@kernel.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Cc: Peter Anvin <hpa@zytor.com> Cc: Jan Kara <jack@suse.cz> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://lore.kernel.org/r/20240411001012.12513-1-torvalds@linux-foundation.org [brauner: relax capability check to opener of the file] Link: https://lore.kernel.org/all/20231113-undenkbar-gediegen-efde5f1c34bc@brauner Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-04-11 08:10:12 +08:00
if (flags & LOOKUP_LINKAT_EMPTY) {
if (f.file->f_cred != current_cred() &&
!ns_capable(f.file->f_cred->user_ns, CAP_DAC_READ_SEARCH)) {
fdput(f);
return ERR_PTR(-ENOENT);
}
}
dentry = f.file->f_path.dentry;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
if (*s && unlikely(!d_can_lookup(dentry))) {
fdput(f);
return ERR_PTR(-ENOTDIR);
}
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
nd->path = f.file->f_path;
if (flags & LOOKUP_RCU) {
nd->inode = nd->path.dentry->d_inode;
nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
} else {
path_get(&nd->path);
nd->inode = nd->path.dentry->d_inode;
}
fdput(f);
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
}
namei: LOOKUP_IN_ROOT: chroot-like scoped resolution /* Background. */ Container runtimes or other administrative management processes will often interact with root filesystems while in the host mount namespace, because the cost of doing a chroot(2) on every operation is too prohibitive (especially in Go, which cannot safely use vfork). However, a malicious program can trick the management process into doing operations on files outside of the root filesystem through careful crafting of symlinks. Most programs that need this feature have attempted to make this process safe, by doing all of the path resolution in userspace (with symlinks being scoped to the root of the malicious root filesystem). Unfortunately, this method is prone to foot-guns and usually such implementations have subtle security bugs. Thus, what userspace needs is a way to resolve a path as though it were in a chroot(2) -- with all absolute symlinks being resolved relative to the dirfd root (and ".." components being stuck under the dirfd root). It is much simpler and more straight-forward to provide this functionality in-kernel (because it can be done far more cheaply and correctly). More classical applications that also have this problem (which have their own potentially buggy userspace path sanitisation code) include web servers, archive extraction tools, network file servers, and so on. /* Userspace API. */ LOOKUP_IN_ROOT will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_IN_ROOT applies to all components of the path. With LOOKUP_IN_ROOT, any path component which attempts to cross the starting point of the pathname lookup (the dirfd passed to openat) will remain at the starting point. Thus, all absolute paths and symlinks will be scoped within the starting point. There is a slight change in behaviour regarding pathnames -- if the pathname is absolute then the dirfd is still used as the root of resolution of LOOKUP_IN_ROOT is specified (this is to avoid obvious foot-guns, at the cost of a minor API inconsistency). As with LOOKUP_BENEATH, Jann's security concern about ".."[1] applies to LOOKUP_IN_ROOT -- therefore ".." resolution is blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_IN_ROOT is tested as part of the openat2(2) selftests. [1]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:34 +08:00
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
/* For scoped-lookups we need to set the root to the dirfd as well. */
if (flags & LOOKUP_IS_SCOPED) {
nd->root = nd->path;
if (flags & LOOKUP_RCU) {
nd->root_seq = nd->seq;
} else {
path_get(&nd->root);
nd->state |= ND_ROOT_GRABBED;
namei: LOOKUP_BENEATH: O_BENEATH-like scoped resolution /* Background. */ There are many circumstances when userspace wants to resolve a path and ensure that it doesn't go outside of a particular root directory during resolution. Obvious examples include archive extraction tools, as well as other security-conscious userspace programs. FreeBSD spun out O_BENEATH from their Capsicum project[1,2], so it also seems reasonable to implement similar functionality for Linux. This is part of a refresh of Al's AT_NO_JUMPS patchset[3] (which was a variation on David Drysdale's O_BENEATH patchset[4], which in turn was based on the Capsicum project[5]). /* Userspace API. */ LOOKUP_BENEATH will be exposed to userspace through openat2(2). /* Semantics. */ Unlike most other LOOKUP flags (most notably LOOKUP_FOLLOW), LOOKUP_BENEATH applies to all components of the path. With LOOKUP_BENEATH, any path component which attempts to "escape" the starting point of the filesystem lookup (the dirfd passed to openat) will yield -EXDEV. Thus, all absolute paths and symlinks are disallowed. Due to a security concern brought up by Jann[6], any ".." path components are also blocked. This restriction will be lifted in a future patch, but requires more work to ensure that permitting ".." is done safely. Magic-link jumps are also blocked, because they can beam the path lookup across the starting point. It would be possible to detect and block only the "bad" crossings with path_is_under() checks, but it's unclear whether it makes sense to permit magic-links at all. However, userspace is recommended to pass LOOKUP_NO_MAGICLINKS if they want to ensure that magic-link crossing is entirely disabled. /* Testing. */ LOOKUP_BENEATH is tested as part of the openat2(2) selftests. [1]: https://reviews.freebsd.org/D2808 [2]: https://reviews.freebsd.org/D17547 [3]: https://lore.kernel.org/lkml/20170429220414.GT29622@ZenIV.linux.org.uk/ [4]: https://lore.kernel.org/lkml/1415094884-18349-1-git-send-email-drysdale@google.com/ [5]: https://lore.kernel.org/lkml/1404124096-21445-1-git-send-email-drysdale@google.com/ [6]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/ Cc: Christian Brauner <christian.brauner@ubuntu.com> Suggested-by: David Drysdale <drysdale@google.com> Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Suggested-by: Andy Lutomirski <luto@kernel.org> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Aleksa Sarai <cyphar@cyphar.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2019-12-06 22:13:33 +08:00
}
}
return s;
}
static inline const char *lookup_last(struct nameidata *nd)
{
if (nd->last_type == LAST_NORM && nd->last.name[nd->last.len])
nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
return walk_component(nd, WALK_TRAILING);
}
static int handle_lookup_down(struct nameidata *nd)
{
if (!(nd->flags & LOOKUP_RCU))
dget(nd->path.dentry);
namei: stash the sampled ->d_seq into nameidata New field: nd->next_seq. Set to 0 outside of RCU mode, holds the sampled value for the next dentry to be considered. Used instead of an arseload of local variables, arguments, etc. step_into() has lost seq argument; nd->next_seq is used, so dentry passed to it must be the one ->next_seq is about. There are two requirements for RCU pathwalk: 1) it should not give a hard failure (other than -ECHILD) unless non-RCU pathwalk might fail that way given suitable timings. 2) it should not succeed unless non-RCU pathwalk might succeed with the same end location given suitable timings. The use of seq numbers is the way we achieve that. Invariant we want to maintain is: if RCU pathwalk can reach the state with given nd->path, nd->inode and nd->seq after having traversed some part of pathname, it must be possible for non-RCU pathwalk to reach the same nd->path and nd->inode after having traversed the same part of pathname, and observe the nd->path.dentry->d_seq equal to what RCU pathwalk has in nd->seq For transition from parent to child, we sample child's ->d_seq and verify that parent's ->d_seq remains unchanged. Anything that disrupts parent-child relationship would've bumped ->d_seq on both. For transitions from child to parent we sample parent's ->d_seq and verify that child's ->d_seq has not changed. Same reasoning as for the previous case applies. For transition from mountpoint to root of mounted we sample the ->d_seq of root and verify that nobody has touched mount_lock since the beginning of pathwalk. That guarantees that mount we'd found had been there all along, with these mountpoint and root of the mounted. It would be possible for a non-RCU pathwalk to reach the previous state, find the same mount and observe its root at the moment we'd sampled ->d_seq of that For transitions from root of mounted to mountpoint we sample ->d_seq of mountpoint and verify that mount_lock had not been touched since the beginning of pathwalk. The same reasoning as in the previous case applies. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2022-07-05 06:12:39 +08:00
nd->next_seq = nd->seq;
return PTR_ERR(step_into(nd, WALK_NOFOLLOW, nd->path.dentry));
}
/* Returns 0 and nd will be valid on success; Returns error, otherwise. */
static int path_lookupat(struct nameidata *nd, unsigned flags, struct path *path)
{
const char *s = path_init(nd, flags);
int err;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
if (unlikely(flags & LOOKUP_DOWN) && !IS_ERR(s)) {
err = handle_lookup_down(nd);
if (unlikely(err < 0))
s = ERR_PTR(err);
}
while (!(err = link_path_walk(s, nd)) &&
(s = lookup_last(nd)) != NULL)
;
if (!err && unlikely(nd->flags & LOOKUP_MOUNTPOINT)) {
err = handle_lookup_down(nd);
nd->state &= ~ND_JUMPED; // no d_weak_revalidate(), please...
}
if (!err)
err = complete_walk(nd);
if (!err && nd->flags & LOOKUP_DIRECTORY)
if (!d_can_lookup(nd->path.dentry))
err = -ENOTDIR;
if (!err) {
*path = nd->path;
nd->path.mnt = NULL;
nd->path.dentry = NULL;
}
terminate_walk(nd);
return err;
}
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
int filename_lookup(int dfd, struct filename *name, unsigned flags,
vfs: Add configuration parser helpers Because the new API passes in key,value parameters, match_token() cannot be used with it. Instead, provide three new helpers to aid with parsing: (1) fs_parse(). This takes a parameter and a simple static description of all the parameters and maps the key name to an ID. It returns 1 on a match, 0 on no match if unknowns should be ignored and some other negative error code on a parse error. The parameter description includes a list of key names to IDs, desired parameter types and a list of enumeration name -> ID mappings. [!] Note that for the moment I've required that the key->ID mapping array is expected to be sorted and unterminated. The size of the array is noted in the fsconfig_parser struct. This allows me to use bsearch(), but I'm not sure any performance gain is worth the hassle of requiring people to keep the array sorted. The parameter type array is sized according to the number of parameter IDs and is indexed directly. The optional enum mapping array is an unterminated, unsorted list and the size goes into the fsconfig_parser struct. The function can do some additional things: (a) If it's not ambiguous and no value is given, the prefix "no" on a key name is permitted to indicate that the parameter should be considered negatory. (b) If the desired type is a single simple integer, it will perform an appropriate conversion and store the result in a union in the parse result. (c) If the desired type is an enumeration, {key ID, name} will be looked up in the enumeration list and the matching value will be stored in the parse result union. (d) Optionally generate an error if the key is unrecognised. This is called something like: enum rdt_param { Opt_cdp, Opt_cdpl2, Opt_mba_mpbs, nr__rdt_params }; const struct fs_parameter_spec rdt_param_specs[nr__rdt_params] = { [Opt_cdp] = { fs_param_is_bool }, [Opt_cdpl2] = { fs_param_is_bool }, [Opt_mba_mpbs] = { fs_param_is_bool }, }; const const char *const rdt_param_keys[nr__rdt_params] = { [Opt_cdp] = "cdp", [Opt_cdpl2] = "cdpl2", [Opt_mba_mpbs] = "mba_mbps", }; const struct fs_parameter_description rdt_parser = { .name = "rdt", .nr_params = nr__rdt_params, .keys = rdt_param_keys, .specs = rdt_param_specs, .no_source = true, }; int rdt_parse_param(struct fs_context *fc, struct fs_parameter *param) { struct fs_parse_result parse; struct rdt_fs_context *ctx = rdt_fc2context(fc); int ret; ret = fs_parse(fc, &rdt_parser, param, &parse); if (ret < 0) return ret; switch (parse.key) { case Opt_cdp: ctx->enable_cdpl3 = true; return 0; case Opt_cdpl2: ctx->enable_cdpl2 = true; return 0; case Opt_mba_mpbs: ctx->enable_mba_mbps = true; return 0; } return -EINVAL; } (2) fs_lookup_param(). This takes a { dirfd, path, LOOKUP_EMPTY? } or string value and performs an appropriate path lookup to convert it into a path object, which it will then return. If the desired type was a blockdev, the type of the looked up inode will be checked to make sure it is one. This can be used like: enum foo_param { Opt_source, nr__foo_params }; const struct fs_parameter_spec foo_param_specs[nr__foo_params] = { [Opt_source] = { fs_param_is_blockdev }, }; const char *char foo_param_keys[nr__foo_params] = { [Opt_source] = "source", }; const struct constant_table foo_param_alt_keys[] = { { "device", Opt_source }, }; const struct fs_parameter_description foo_parser = { .name = "foo", .nr_params = nr__foo_params, .nr_alt_keys = ARRAY_SIZE(foo_param_alt_keys), .keys = foo_param_keys, .alt_keys = foo_param_alt_keys, .specs = foo_param_specs, }; int foo_parse_param(struct fs_context *fc, struct fs_parameter *param) { struct fs_parse_result parse; struct foo_fs_context *ctx = foo_fc2context(fc); int ret; ret = fs_parse(fc, &foo_parser, param, &parse); if (ret < 0) return ret; switch (parse.key) { case Opt_source: return fs_lookup_param(fc, &foo_parser, param, &parse, &ctx->source); default: return -EINVAL; } } (3) lookup_constant(). This takes a table of named constants and looks up the given name within it. The table is expected to be sorted such that bsearch() be used upon it. Possibly I should require the table be terminated and just use a for-loop to scan it instead of using bsearch() to reduce hassle. Tables look something like: static const struct constant_table bool_names[] = { { "0", false }, { "1", true }, { "false", false }, { "no", false }, { "true", true }, { "yes", true }, }; and a lookup is done with something like: b = lookup_constant(bool_names, param->string, -1); Additionally, optional validation routines for the parameter description are provided that can be enabled at compile time. A later patch will invoke these when a filesystem is registered. Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-11-02 07:07:24 +08:00
struct path *path, struct path *root)
{
int retval;
struct nameidata nd;
if (IS_ERR(name))
return PTR_ERR(name);
set_nameidata(&nd, dfd, name, root);
retval = path_lookupat(&nd, flags | LOOKUP_RCU, path);
if (unlikely(retval == -ECHILD))
retval = path_lookupat(&nd, flags, path);
if (unlikely(retval == -ESTALE))
retval = path_lookupat(&nd, flags | LOOKUP_REVAL, path);
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
if (likely(!retval))
audit_inode(name, path->dentry,
flags & LOOKUP_MOUNTPOINT ? AUDIT_INODE_NOEVAL : 0);
restore_nameidata();
return retval;
}
/* Returns 0 and nd will be valid on success; Returns error, otherwise. */
static int path_parentat(struct nameidata *nd, unsigned flags,
struct path *parent)
{
const char *s = path_init(nd, flags);
int err = link_path_walk(s, nd);
if (!err)
err = complete_walk(nd);
if (!err) {
*parent = nd->path;
nd->path.mnt = NULL;
nd->path.dentry = NULL;
}
terminate_walk(nd);
return err;
}
namei: Fix use after free in kern_path_locked In 0ee50b47532a ("namei: change filename_parentat() calling conventions"), filename_parentat() was made to always call putname() on the filename before returning, and kern_path_locked() was migrated to this calling convention. However, kern_path_locked() uses the "last" parameter to lookup and potentially create a new dentry. The last parameter contains the last component of the path and points within the filename, which was recently freed at the end of filename_parentat(). Thus, when kern_path_locked() calls __lookup_hash(), it is using the filename after it has already been freed. In other words, these calling conventions had been wrong for the only remaining caller of filename_parentat(). Everything else is using __filename_parentat(), which does not drop the reference; so should kern_path_locked(). Switch kern_path_locked() to use of __filename_parentat() and move getting/dropping struct filename into wrapper. Remove filename_parentat(), now that we have no remaining callers. Fixes: 0ee50b47532a ("namei: change filename_parentat() calling conventions") Link: https://lore.kernel.org/linux-fsdevel/YS9D4AlEsaCxLFV0@infradead.org/ Link: https://lore.kernel.org/linux-fsdevel/YS+csMTV2tTXKg3s@zeniv-ca.linux.org.uk/ Cc: Christoph Hellwig <hch@infradead.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Reported-by: syzbot+fb0d60a179096e8c2731@syzkaller.appspotmail.com Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com> Co-authored-by: Dmitry Kadashev <dkadashev@gmail.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2021-09-02 01:51:41 +08:00
/* Note: this does not consume "name" */
static int __filename_parentat(int dfd, struct filename *name,
unsigned int flags, struct path *parent,
struct qstr *last, int *type,
const struct path *root)
{
int retval;
struct nameidata nd;
if (IS_ERR(name))
return PTR_ERR(name);
set_nameidata(&nd, dfd, name, root);
retval = path_parentat(&nd, flags | LOOKUP_RCU, parent);
if (unlikely(retval == -ECHILD))
retval = path_parentat(&nd, flags, parent);
if (unlikely(retval == -ESTALE))
retval = path_parentat(&nd, flags | LOOKUP_REVAL, parent);
if (likely(!retval)) {
*last = nd.last;
*type = nd.last_type;
audit_inode(name, parent->dentry, AUDIT_INODE_PARENT);
}
restore_nameidata();
return retval;
}
static int filename_parentat(int dfd, struct filename *name,
unsigned int flags, struct path *parent,
struct qstr *last, int *type)
{
return __filename_parentat(dfd, name, flags, parent, last, type, NULL);
}
/* does lookup, returns the object with parent locked */
static struct dentry *__kern_path_locked(int dfd, struct filename *name, struct path *path)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
struct dentry *d;
struct qstr last;
int type, error;
error = filename_parentat(dfd, name, 0, path, &last, &type);
if (error)
return ERR_PTR(error);
if (unlikely(type != LAST_NORM)) {
path_put(path);
return ERR_PTR(-EINVAL);
}
inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
d = lookup_one_qstr_excl(&last, path->dentry, 0);
if (IS_ERR(d)) {
inode_unlock(path->dentry->d_inode);
path_put(path);
}
return d;
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
namei: Fix use after free in kern_path_locked In 0ee50b47532a ("namei: change filename_parentat() calling conventions"), filename_parentat() was made to always call putname() on the filename before returning, and kern_path_locked() was migrated to this calling convention. However, kern_path_locked() uses the "last" parameter to lookup and potentially create a new dentry. The last parameter contains the last component of the path and points within the filename, which was recently freed at the end of filename_parentat(). Thus, when kern_path_locked() calls __lookup_hash(), it is using the filename after it has already been freed. In other words, these calling conventions had been wrong for the only remaining caller of filename_parentat(). Everything else is using __filename_parentat(), which does not drop the reference; so should kern_path_locked(). Switch kern_path_locked() to use of __filename_parentat() and move getting/dropping struct filename into wrapper. Remove filename_parentat(), now that we have no remaining callers. Fixes: 0ee50b47532a ("namei: change filename_parentat() calling conventions") Link: https://lore.kernel.org/linux-fsdevel/YS9D4AlEsaCxLFV0@infradead.org/ Link: https://lore.kernel.org/linux-fsdevel/YS+csMTV2tTXKg3s@zeniv-ca.linux.org.uk/ Cc: Christoph Hellwig <hch@infradead.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Reported-by: syzbot+fb0d60a179096e8c2731@syzkaller.appspotmail.com Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com> Co-authored-by: Dmitry Kadashev <dkadashev@gmail.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2021-09-02 01:51:41 +08:00
struct dentry *kern_path_locked(const char *name, struct path *path)
{
struct filename *filename = getname_kernel(name);
struct dentry *res = __kern_path_locked(AT_FDCWD, filename, path);
namei: Fix use after free in kern_path_locked In 0ee50b47532a ("namei: change filename_parentat() calling conventions"), filename_parentat() was made to always call putname() on the filename before returning, and kern_path_locked() was migrated to this calling convention. However, kern_path_locked() uses the "last" parameter to lookup and potentially create a new dentry. The last parameter contains the last component of the path and points within the filename, which was recently freed at the end of filename_parentat(). Thus, when kern_path_locked() calls __lookup_hash(), it is using the filename after it has already been freed. In other words, these calling conventions had been wrong for the only remaining caller of filename_parentat(). Everything else is using __filename_parentat(), which does not drop the reference; so should kern_path_locked(). Switch kern_path_locked() to use of __filename_parentat() and move getting/dropping struct filename into wrapper. Remove filename_parentat(), now that we have no remaining callers. Fixes: 0ee50b47532a ("namei: change filename_parentat() calling conventions") Link: https://lore.kernel.org/linux-fsdevel/YS9D4AlEsaCxLFV0@infradead.org/ Link: https://lore.kernel.org/linux-fsdevel/YS+csMTV2tTXKg3s@zeniv-ca.linux.org.uk/ Cc: Christoph Hellwig <hch@infradead.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Reported-by: syzbot+fb0d60a179096e8c2731@syzkaller.appspotmail.com Signed-off-by: Stephen Brennan <stephen.s.brennan@oracle.com> Co-authored-by: Dmitry Kadashev <dkadashev@gmail.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2021-09-02 01:51:41 +08:00
putname(filename);
return res;
}
struct dentry *user_path_locked_at(int dfd, const char __user *name, struct path *path)
{
struct filename *filename = getname(name);
struct dentry *res = __kern_path_locked(dfd, filename, path);
putname(filename);
return res;
}
EXPORT_SYMBOL(user_path_locked_at);
int kern_path(const char *name, unsigned int flags, struct path *path)
{
struct filename *filename = getname_kernel(name);
int ret = filename_lookup(AT_FDCWD, filename, flags, path, NULL);
putname(filename);
return ret;
}
EXPORT_SYMBOL(kern_path);
/**
* vfs_path_parent_lookup - lookup a parent path relative to a dentry-vfsmount pair
* @filename: filename structure
* @flags: lookup flags
* @parent: pointer to struct path to fill
* @last: last component
* @type: type of the last component
* @root: pointer to struct path of the base directory
*/
int vfs_path_parent_lookup(struct filename *filename, unsigned int flags,
struct path *parent, struct qstr *last, int *type,
const struct path *root)
{
return __filename_parentat(AT_FDCWD, filename, flags, parent, last,
type, root);
}
EXPORT_SYMBOL(vfs_path_parent_lookup);
fs: introduce vfs_path_lookup Stackable file systems, among others, frequently need to lookup paths or path components starting from an arbitrary point in the namespace (identified by a dentry and a vfsmount). Currently, such file systems use lookup_one_len, which is frowned upon [1] as it does not pass the lookup intent along; not passing a lookup intent, for example, can trigger BUG_ON's when stacking on top of NFSv4. The first patch introduces a new lookup function to allow lookup starting from an arbitrary point in the namespace. This approach has been suggested by Christoph Hellwig [2]. The second patch changes sunrpc to use vfs_path_lookup. The third patch changes nfsctl.c to use vfs_path_lookup. The fourth patch marks link_path_walk static. The fifth, and last patch, unexports path_walk because it is no longer unnecessary to call it directly, and using the new vfs_path_lookup is cleaner. For example, the following snippet of code, looks up "some/path/component" in a directory pointed to by parent_{dentry,vfsmnt}: err = vfs_path_lookup(parent_dentry, parent_vfsmnt, "some/path/component", 0, &nd); if (!err) { /* exits */ ... /* once done, release the references */ path_release(&nd); } else if (err == -ENOENT) { /* doesn't exist */ } else { /* other error */ } VFS functions such as lookup_create can be used on the nameidata structure to pass the create intent to the file system. Signed-off-by: Josef 'Jeff' Sipek <jsipek@cs.sunysb.edu> Cc: Al Viro <viro@zeniv.linux.org.uk> Acked-by: Christoph Hellwig <hch@lst.de> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Cc: Neil Brown <neilb@suse.de> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-19 16:48:18 +08:00
/**
* vfs_path_lookup - lookup a file path relative to a dentry-vfsmount pair
* @dentry: pointer to dentry of the base directory
* @mnt: pointer to vfs mount of the base directory
* @name: pointer to file name
* @flags: lookup flags
* @path: pointer to struct path to fill
fs: introduce vfs_path_lookup Stackable file systems, among others, frequently need to lookup paths or path components starting from an arbitrary point in the namespace (identified by a dentry and a vfsmount). Currently, such file systems use lookup_one_len, which is frowned upon [1] as it does not pass the lookup intent along; not passing a lookup intent, for example, can trigger BUG_ON's when stacking on top of NFSv4. The first patch introduces a new lookup function to allow lookup starting from an arbitrary point in the namespace. This approach has been suggested by Christoph Hellwig [2]. The second patch changes sunrpc to use vfs_path_lookup. The third patch changes nfsctl.c to use vfs_path_lookup. The fourth patch marks link_path_walk static. The fifth, and last patch, unexports path_walk because it is no longer unnecessary to call it directly, and using the new vfs_path_lookup is cleaner. For example, the following snippet of code, looks up "some/path/component" in a directory pointed to by parent_{dentry,vfsmnt}: err = vfs_path_lookup(parent_dentry, parent_vfsmnt, "some/path/component", 0, &nd); if (!err) { /* exits */ ... /* once done, release the references */ path_release(&nd); } else if (err == -ENOENT) { /* doesn't exist */ } else { /* other error */ } VFS functions such as lookup_create can be used on the nameidata structure to pass the create intent to the file system. Signed-off-by: Josef 'Jeff' Sipek <jsipek@cs.sunysb.edu> Cc: Al Viro <viro@zeniv.linux.org.uk> Acked-by: Christoph Hellwig <hch@lst.de> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Cc: Neil Brown <neilb@suse.de> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-19 16:48:18 +08:00
*/
int vfs_path_lookup(struct dentry *dentry, struct vfsmount *mnt,
const char *name, unsigned int flags,
struct path *path)
fs: introduce vfs_path_lookup Stackable file systems, among others, frequently need to lookup paths or path components starting from an arbitrary point in the namespace (identified by a dentry and a vfsmount). Currently, such file systems use lookup_one_len, which is frowned upon [1] as it does not pass the lookup intent along; not passing a lookup intent, for example, can trigger BUG_ON's when stacking on top of NFSv4. The first patch introduces a new lookup function to allow lookup starting from an arbitrary point in the namespace. This approach has been suggested by Christoph Hellwig [2]. The second patch changes sunrpc to use vfs_path_lookup. The third patch changes nfsctl.c to use vfs_path_lookup. The fourth patch marks link_path_walk static. The fifth, and last patch, unexports path_walk because it is no longer unnecessary to call it directly, and using the new vfs_path_lookup is cleaner. For example, the following snippet of code, looks up "some/path/component" in a directory pointed to by parent_{dentry,vfsmnt}: err = vfs_path_lookup(parent_dentry, parent_vfsmnt, "some/path/component", 0, &nd); if (!err) { /* exits */ ... /* once done, release the references */ path_release(&nd); } else if (err == -ENOENT) { /* doesn't exist */ } else { /* other error */ } VFS functions such as lookup_create can be used on the nameidata structure to pass the create intent to the file system. Signed-off-by: Josef 'Jeff' Sipek <jsipek@cs.sunysb.edu> Cc: Al Viro <viro@zeniv.linux.org.uk> Acked-by: Christoph Hellwig <hch@lst.de> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Cc: Neil Brown <neilb@suse.de> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-19 16:48:18 +08:00
{
struct filename *filename;
struct path root = {.mnt = mnt, .dentry = dentry};
int ret;
filename = getname_kernel(name);
/* the first argument of filename_lookup() is ignored with root */
ret = filename_lookup(AT_FDCWD, filename, flags, path, &root);
putname(filename);
return ret;
fs: introduce vfs_path_lookup Stackable file systems, among others, frequently need to lookup paths or path components starting from an arbitrary point in the namespace (identified by a dentry and a vfsmount). Currently, such file systems use lookup_one_len, which is frowned upon [1] as it does not pass the lookup intent along; not passing a lookup intent, for example, can trigger BUG_ON's when stacking on top of NFSv4. The first patch introduces a new lookup function to allow lookup starting from an arbitrary point in the namespace. This approach has been suggested by Christoph Hellwig [2]. The second patch changes sunrpc to use vfs_path_lookup. The third patch changes nfsctl.c to use vfs_path_lookup. The fourth patch marks link_path_walk static. The fifth, and last patch, unexports path_walk because it is no longer unnecessary to call it directly, and using the new vfs_path_lookup is cleaner. For example, the following snippet of code, looks up "some/path/component" in a directory pointed to by parent_{dentry,vfsmnt}: err = vfs_path_lookup(parent_dentry, parent_vfsmnt, "some/path/component", 0, &nd); if (!err) { /* exits */ ... /* once done, release the references */ path_release(&nd); } else if (err == -ENOENT) { /* doesn't exist */ } else { /* other error */ } VFS functions such as lookup_create can be used on the nameidata structure to pass the create intent to the file system. Signed-off-by: Josef 'Jeff' Sipek <jsipek@cs.sunysb.edu> Cc: Al Viro <viro@zeniv.linux.org.uk> Acked-by: Christoph Hellwig <hch@lst.de> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Cc: Neil Brown <neilb@suse.de> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-19 16:48:18 +08:00
}
EXPORT_SYMBOL(vfs_path_lookup);
fs: introduce vfs_path_lookup Stackable file systems, among others, frequently need to lookup paths or path components starting from an arbitrary point in the namespace (identified by a dentry and a vfsmount). Currently, such file systems use lookup_one_len, which is frowned upon [1] as it does not pass the lookup intent along; not passing a lookup intent, for example, can trigger BUG_ON's when stacking on top of NFSv4. The first patch introduces a new lookup function to allow lookup starting from an arbitrary point in the namespace. This approach has been suggested by Christoph Hellwig [2]. The second patch changes sunrpc to use vfs_path_lookup. The third patch changes nfsctl.c to use vfs_path_lookup. The fourth patch marks link_path_walk static. The fifth, and last patch, unexports path_walk because it is no longer unnecessary to call it directly, and using the new vfs_path_lookup is cleaner. For example, the following snippet of code, looks up "some/path/component" in a directory pointed to by parent_{dentry,vfsmnt}: err = vfs_path_lookup(parent_dentry, parent_vfsmnt, "some/path/component", 0, &nd); if (!err) { /* exits */ ... /* once done, release the references */ path_release(&nd); } else if (err == -ENOENT) { /* doesn't exist */ } else { /* other error */ } VFS functions such as lookup_create can be used on the nameidata structure to pass the create intent to the file system. Signed-off-by: Josef 'Jeff' Sipek <jsipek@cs.sunysb.edu> Cc: Al Viro <viro@zeniv.linux.org.uk> Acked-by: Christoph Hellwig <hch@lst.de> Cc: Trond Myklebust <trond.myklebust@fys.uio.no> Cc: Neil Brown <neilb@suse.de> Cc: Michael Halcrow <mhalcrow@us.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-07-19 16:48:18 +08:00
static int lookup_one_common(struct mnt_idmap *idmap,
const char *name, struct dentry *base, int len,
struct qstr *this)
{
this->name = name;
this->len = len;
this->hash = full_name_hash(base, name, len);
if (!len)
return -EACCES;
if (is_dot_dotdot(name, len))
return -EACCES;
while (len--) {
unsigned int c = *(const unsigned char *)name++;
if (c == '/' || c == '\0')
return -EACCES;
}
/*
* See if the low-level filesystem might want
* to use its own hash..
*/
if (base->d_flags & DCACHE_OP_HASH) {
int err = base->d_op->d_hash(base, this);
if (err < 0)
return err;
}
return inode_permission(idmap, base->d_inode, MAY_EXEC);
}
/**
* try_lookup_one_len - filesystem helper to lookup single pathname component
* @name: pathname component to lookup
* @base: base directory to lookup from
* @len: maximum length @len should be interpreted to
*
* Look up a dentry by name in the dcache, returning NULL if it does not
* currently exist. The function does not try to create a dentry.
*
* Note that this routine is purely a helper for filesystem usage and should
* not be called by generic code.
*
* The caller must hold base->i_mutex.
*/
struct dentry *try_lookup_one_len(const char *name, struct dentry *base, int len)
{
struct qstr this;
int err;
WARN_ON_ONCE(!inode_is_locked(base->d_inode));
err = lookup_one_common(&nop_mnt_idmap, name, base, len, &this);
if (err)
return ERR_PTR(err);
return lookup_dcache(&this, base, 0);
}
EXPORT_SYMBOL(try_lookup_one_len);
/**
* lookup_one_len - filesystem helper to lookup single pathname component
* @name: pathname component to lookup
* @base: base directory to lookup from
* @len: maximum length @len should be interpreted to
*
* Note that this routine is purely a helper for filesystem usage and should
* not be called by generic code.
*
* The caller must hold base->i_mutex.
*/
struct dentry *lookup_one_len(const char *name, struct dentry *base, int len)
{
struct dentry *dentry;
struct qstr this;
int err;
WARN_ON_ONCE(!inode_is_locked(base->d_inode));
err = lookup_one_common(&nop_mnt_idmap, name, base, len, &this);
if (err)
return ERR_PTR(err);
dentry = lookup_dcache(&this, base, 0);
return dentry ? dentry : __lookup_slow(&this, base, 0);
}
EXPORT_SYMBOL(lookup_one_len);
/**
* lookup_one - filesystem helper to lookup single pathname component
* @idmap: idmap of the mount the lookup is performed from
* @name: pathname component to lookup
* @base: base directory to lookup from
* @len: maximum length @len should be interpreted to
*
* Note that this routine is purely a helper for filesystem usage and should
* not be called by generic code.
*
* The caller must hold base->i_mutex.
*/
struct dentry *lookup_one(struct mnt_idmap *idmap, const char *name,
struct dentry *base, int len)
{
struct dentry *dentry;
struct qstr this;
int err;
WARN_ON_ONCE(!inode_is_locked(base->d_inode));
err = lookup_one_common(idmap, name, base, len, &this);
if (err)
return ERR_PTR(err);
dentry = lookup_dcache(&this, base, 0);
return dentry ? dentry : __lookup_slow(&this, base, 0);
}
EXPORT_SYMBOL(lookup_one);
/**
* lookup_one_unlocked - filesystem helper to lookup single pathname component
* @idmap: idmap of the mount the lookup is performed from
* @name: pathname component to lookup
* @base: base directory to lookup from
* @len: maximum length @len should be interpreted to
*
* Note that this routine is purely a helper for filesystem usage and should
* not be called by generic code.
*
* Unlike lookup_one_len, it should be called without the parent
* i_mutex held, and will take the i_mutex itself if necessary.
*/
struct dentry *lookup_one_unlocked(struct mnt_idmap *idmap,
const char *name, struct dentry *base,
int len)
{
struct qstr this;
int err;
struct dentry *ret;
err = lookup_one_common(idmap, name, base, len, &this);
if (err)
return ERR_PTR(err);
ret = lookup_dcache(&this, base, 0);
if (!ret)
ret = lookup_slow(&this, base, 0);
return ret;
}
EXPORT_SYMBOL(lookup_one_unlocked);
/**
* lookup_one_positive_unlocked - filesystem helper to lookup single
* pathname component
* @idmap: idmap of the mount the lookup is performed from
* @name: pathname component to lookup
* @base: base directory to lookup from
* @len: maximum length @len should be interpreted to
*
* This helper will yield ERR_PTR(-ENOENT) on negatives. The helper returns
* known positive or ERR_PTR(). This is what most of the users want.
*
* Note that pinned negative with unlocked parent _can_ become positive at any
* time, so callers of lookup_one_unlocked() need to be very careful; pinned
* positives have >d_inode stable, so this one avoids such problems.
*
* Note that this routine is purely a helper for filesystem usage and should
* not be called by generic code.
*
* The helper should be called without i_mutex held.
*/
struct dentry *lookup_one_positive_unlocked(struct mnt_idmap *idmap,
const char *name,
struct dentry *base, int len)
{
struct dentry *ret = lookup_one_unlocked(idmap, name, base, len);
if (!IS_ERR(ret) && d_flags_negative(smp_load_acquire(&ret->d_flags))) {
dput(ret);
ret = ERR_PTR(-ENOENT);
}
return ret;
}
EXPORT_SYMBOL(lookup_one_positive_unlocked);
/**
* lookup_one_len_unlocked - filesystem helper to lookup single pathname component
* @name: pathname component to lookup
* @base: base directory to lookup from
* @len: maximum length @len should be interpreted to
*
* Note that this routine is purely a helper for filesystem usage and should
* not be called by generic code.
*
* Unlike lookup_one_len, it should be called without the parent
* i_mutex held, and will take the i_mutex itself if necessary.
*/
struct dentry *lookup_one_len_unlocked(const char *name,
struct dentry *base, int len)
{
return lookup_one_unlocked(&nop_mnt_idmap, name, base, len);
}
EXPORT_SYMBOL(lookup_one_len_unlocked);
/*
* Like lookup_one_len_unlocked(), except that it yields ERR_PTR(-ENOENT)
* on negatives. Returns known positive or ERR_PTR(); that's what
* most of the users want. Note that pinned negative with unlocked parent
* _can_ become positive at any time, so callers of lookup_one_len_unlocked()
* need to be very careful; pinned positives have ->d_inode stable, so
* this one avoids such problems.
*/
struct dentry *lookup_positive_unlocked(const char *name,
struct dentry *base, int len)
{
return lookup_one_positive_unlocked(&nop_mnt_idmap, name, base, len);
}
EXPORT_SYMBOL(lookup_positive_unlocked);
devpts: Make each mount of devpts an independent filesystem. The /dev/ptmx device node is changed to lookup the directory entry "pts" in the same directory as the /dev/ptmx device node was opened in. If there is a "pts" entry and that entry is a devpts filesystem /dev/ptmx uses that filesystem. Otherwise the open of /dev/ptmx fails. The DEVPTS_MULTIPLE_INSTANCES configuration option is removed, so that userspace can now safely depend on each mount of devpts creating a new instance of the filesystem. Each mount of devpts is now a separate and equal filesystem. Reserved ttys are now available to all instances of devpts where the mounter is in the initial mount namespace. A new vfs helper path_pts is introduced that finds a directory entry named "pts" in the directory of the passed in path, and changes the passed in path to point to it. The helper path_pts uses a function path_parent_directory that was factored out of follow_dotdot. In the implementation of devpts: - devpts_mnt is killed as it is no longer meaningful if all mounts of devpts are equal. - pts_sb_from_inode is replaced by just inode->i_sb as all cached inodes in the tty layer are now from the devpts filesystem. - devpts_add_ref is rolled into the new function devpts_ptmx. And the unnecessary inode hold is removed. - devpts_del_ref is renamed devpts_release and reduced to just a deacrivate_super. - The newinstance mount option continues to be accepted but is now ignored. In devpts_fs.h definitions for when !CONFIG_UNIX98_PTYS are removed as they are never used. Documentation/filesystems/devices.txt is updated to describe the current situation. This has been verified to work properly on openwrt-15.05, centos5, centos6, centos7, debian-6.0.2, debian-7.9, debian-8.2, ubuntu-14.04.3, ubuntu-15.10, fedora23, magia-5, mint-17.3, opensuse-42.1, slackware-14.1, gentoo-20151225 (13.0?), archlinux-2015-12-01. With the caveat that on centos6 and on slackware-14.1 that there wind up being two instances of the devpts filesystem mounted on /dev/pts, the lower copy does not end up getting used. Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Greg KH <greg@kroah.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Peter Anvin <hpa@zytor.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Serge Hallyn <serge.hallyn@ubuntu.com> Cc: Willy Tarreau <w@1wt.eu> Cc: Aurelien Jarno <aurelien@aurel32.net> Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk> Cc: Jann Horn <jann@thejh.net> Cc: Jiri Slaby <jslaby@suse.com> Cc: Florian Weimer <fw@deneb.enyo.de> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-06-02 23:29:47 +08:00
#ifdef CONFIG_UNIX98_PTYS
int path_pts(struct path *path)
{
/* Find something mounted on "pts" in the same directory as
* the input path.
*/
struct dentry *parent = dget_parent(path->dentry);
struct dentry *child;
struct qstr this = QSTR_INIT("pts", 3);
devpts: Make each mount of devpts an independent filesystem. The /dev/ptmx device node is changed to lookup the directory entry "pts" in the same directory as the /dev/ptmx device node was opened in. If there is a "pts" entry and that entry is a devpts filesystem /dev/ptmx uses that filesystem. Otherwise the open of /dev/ptmx fails. The DEVPTS_MULTIPLE_INSTANCES configuration option is removed, so that userspace can now safely depend on each mount of devpts creating a new instance of the filesystem. Each mount of devpts is now a separate and equal filesystem. Reserved ttys are now available to all instances of devpts where the mounter is in the initial mount namespace. A new vfs helper path_pts is introduced that finds a directory entry named "pts" in the directory of the passed in path, and changes the passed in path to point to it. The helper path_pts uses a function path_parent_directory that was factored out of follow_dotdot. In the implementation of devpts: - devpts_mnt is killed as it is no longer meaningful if all mounts of devpts are equal. - pts_sb_from_inode is replaced by just inode->i_sb as all cached inodes in the tty layer are now from the devpts filesystem. - devpts_add_ref is rolled into the new function devpts_ptmx. And the unnecessary inode hold is removed. - devpts_del_ref is renamed devpts_release and reduced to just a deacrivate_super. - The newinstance mount option continues to be accepted but is now ignored. In devpts_fs.h definitions for when !CONFIG_UNIX98_PTYS are removed as they are never used. Documentation/filesystems/devices.txt is updated to describe the current situation. This has been verified to work properly on openwrt-15.05, centos5, centos6, centos7, debian-6.0.2, debian-7.9, debian-8.2, ubuntu-14.04.3, ubuntu-15.10, fedora23, magia-5, mint-17.3, opensuse-42.1, slackware-14.1, gentoo-20151225 (13.0?), archlinux-2015-12-01. With the caveat that on centos6 and on slackware-14.1 that there wind up being two instances of the devpts filesystem mounted on /dev/pts, the lower copy does not end up getting used. Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Greg KH <greg@kroah.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Peter Anvin <hpa@zytor.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Serge Hallyn <serge.hallyn@ubuntu.com> Cc: Willy Tarreau <w@1wt.eu> Cc: Aurelien Jarno <aurelien@aurel32.net> Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk> Cc: Jann Horn <jann@thejh.net> Cc: Jiri Slaby <jslaby@suse.com> Cc: Florian Weimer <fw@deneb.enyo.de> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-06-02 23:29:47 +08:00
if (unlikely(!path_connected(path->mnt, parent))) {
dput(parent);
return -ENOENT;
}
dput(path->dentry);
path->dentry = parent;
devpts: Make each mount of devpts an independent filesystem. The /dev/ptmx device node is changed to lookup the directory entry "pts" in the same directory as the /dev/ptmx device node was opened in. If there is a "pts" entry and that entry is a devpts filesystem /dev/ptmx uses that filesystem. Otherwise the open of /dev/ptmx fails. The DEVPTS_MULTIPLE_INSTANCES configuration option is removed, so that userspace can now safely depend on each mount of devpts creating a new instance of the filesystem. Each mount of devpts is now a separate and equal filesystem. Reserved ttys are now available to all instances of devpts where the mounter is in the initial mount namespace. A new vfs helper path_pts is introduced that finds a directory entry named "pts" in the directory of the passed in path, and changes the passed in path to point to it. The helper path_pts uses a function path_parent_directory that was factored out of follow_dotdot. In the implementation of devpts: - devpts_mnt is killed as it is no longer meaningful if all mounts of devpts are equal. - pts_sb_from_inode is replaced by just inode->i_sb as all cached inodes in the tty layer are now from the devpts filesystem. - devpts_add_ref is rolled into the new function devpts_ptmx. And the unnecessary inode hold is removed. - devpts_del_ref is renamed devpts_release and reduced to just a deacrivate_super. - The newinstance mount option continues to be accepted but is now ignored. In devpts_fs.h definitions for when !CONFIG_UNIX98_PTYS are removed as they are never used. Documentation/filesystems/devices.txt is updated to describe the current situation. This has been verified to work properly on openwrt-15.05, centos5, centos6, centos7, debian-6.0.2, debian-7.9, debian-8.2, ubuntu-14.04.3, ubuntu-15.10, fedora23, magia-5, mint-17.3, opensuse-42.1, slackware-14.1, gentoo-20151225 (13.0?), archlinux-2015-12-01. With the caveat that on centos6 and on slackware-14.1 that there wind up being two instances of the devpts filesystem mounted on /dev/pts, the lower copy does not end up getting used. Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Greg KH <greg@kroah.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Peter Anvin <hpa@zytor.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Serge Hallyn <serge.hallyn@ubuntu.com> Cc: Willy Tarreau <w@1wt.eu> Cc: Aurelien Jarno <aurelien@aurel32.net> Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk> Cc: Jann Horn <jann@thejh.net> Cc: Jiri Slaby <jslaby@suse.com> Cc: Florian Weimer <fw@deneb.enyo.de> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-06-02 23:29:47 +08:00
child = d_hash_and_lookup(parent, &this);
if (IS_ERR_OR_NULL(child))
devpts: Make each mount of devpts an independent filesystem. The /dev/ptmx device node is changed to lookup the directory entry "pts" in the same directory as the /dev/ptmx device node was opened in. If there is a "pts" entry and that entry is a devpts filesystem /dev/ptmx uses that filesystem. Otherwise the open of /dev/ptmx fails. The DEVPTS_MULTIPLE_INSTANCES configuration option is removed, so that userspace can now safely depend on each mount of devpts creating a new instance of the filesystem. Each mount of devpts is now a separate and equal filesystem. Reserved ttys are now available to all instances of devpts where the mounter is in the initial mount namespace. A new vfs helper path_pts is introduced that finds a directory entry named "pts" in the directory of the passed in path, and changes the passed in path to point to it. The helper path_pts uses a function path_parent_directory that was factored out of follow_dotdot. In the implementation of devpts: - devpts_mnt is killed as it is no longer meaningful if all mounts of devpts are equal. - pts_sb_from_inode is replaced by just inode->i_sb as all cached inodes in the tty layer are now from the devpts filesystem. - devpts_add_ref is rolled into the new function devpts_ptmx. And the unnecessary inode hold is removed. - devpts_del_ref is renamed devpts_release and reduced to just a deacrivate_super. - The newinstance mount option continues to be accepted but is now ignored. In devpts_fs.h definitions for when !CONFIG_UNIX98_PTYS are removed as they are never used. Documentation/filesystems/devices.txt is updated to describe the current situation. This has been verified to work properly on openwrt-15.05, centos5, centos6, centos7, debian-6.0.2, debian-7.9, debian-8.2, ubuntu-14.04.3, ubuntu-15.10, fedora23, magia-5, mint-17.3, opensuse-42.1, slackware-14.1, gentoo-20151225 (13.0?), archlinux-2015-12-01. With the caveat that on centos6 and on slackware-14.1 that there wind up being two instances of the devpts filesystem mounted on /dev/pts, the lower copy does not end up getting used. Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Greg KH <greg@kroah.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Peter Anvin <hpa@zytor.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Serge Hallyn <serge.hallyn@ubuntu.com> Cc: Willy Tarreau <w@1wt.eu> Cc: Aurelien Jarno <aurelien@aurel32.net> Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk> Cc: Jann Horn <jann@thejh.net> Cc: Jiri Slaby <jslaby@suse.com> Cc: Florian Weimer <fw@deneb.enyo.de> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-06-02 23:29:47 +08:00
return -ENOENT;
path->dentry = child;
dput(parent);
follow_down(path, 0);
devpts: Make each mount of devpts an independent filesystem. The /dev/ptmx device node is changed to lookup the directory entry "pts" in the same directory as the /dev/ptmx device node was opened in. If there is a "pts" entry and that entry is a devpts filesystem /dev/ptmx uses that filesystem. Otherwise the open of /dev/ptmx fails. The DEVPTS_MULTIPLE_INSTANCES configuration option is removed, so that userspace can now safely depend on each mount of devpts creating a new instance of the filesystem. Each mount of devpts is now a separate and equal filesystem. Reserved ttys are now available to all instances of devpts where the mounter is in the initial mount namespace. A new vfs helper path_pts is introduced that finds a directory entry named "pts" in the directory of the passed in path, and changes the passed in path to point to it. The helper path_pts uses a function path_parent_directory that was factored out of follow_dotdot. In the implementation of devpts: - devpts_mnt is killed as it is no longer meaningful if all mounts of devpts are equal. - pts_sb_from_inode is replaced by just inode->i_sb as all cached inodes in the tty layer are now from the devpts filesystem. - devpts_add_ref is rolled into the new function devpts_ptmx. And the unnecessary inode hold is removed. - devpts_del_ref is renamed devpts_release and reduced to just a deacrivate_super. - The newinstance mount option continues to be accepted but is now ignored. In devpts_fs.h definitions for when !CONFIG_UNIX98_PTYS are removed as they are never used. Documentation/filesystems/devices.txt is updated to describe the current situation. This has been verified to work properly on openwrt-15.05, centos5, centos6, centos7, debian-6.0.2, debian-7.9, debian-8.2, ubuntu-14.04.3, ubuntu-15.10, fedora23, magia-5, mint-17.3, opensuse-42.1, slackware-14.1, gentoo-20151225 (13.0?), archlinux-2015-12-01. With the caveat that on centos6 and on slackware-14.1 that there wind up being two instances of the devpts filesystem mounted on /dev/pts, the lower copy does not end up getting used. Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Greg KH <greg@kroah.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Peter Anvin <hpa@zytor.com> Cc: Andy Lutomirski <luto@amacapital.net> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Serge Hallyn <serge.hallyn@ubuntu.com> Cc: Willy Tarreau <w@1wt.eu> Cc: Aurelien Jarno <aurelien@aurel32.net> Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk> Cc: Jann Horn <jann@thejh.net> Cc: Jiri Slaby <jslaby@suse.com> Cc: Florian Weimer <fw@deneb.enyo.de> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-06-02 23:29:47 +08:00
return 0;
}
#endif
int user_path_at(int dfd, const char __user *name, unsigned flags,
struct path *path)
{
struct filename *filename = getname_flags(name, flags);
int ret = filename_lookup(dfd, filename, flags, path, NULL);
putname(filename);
return ret;
}
EXPORT_SYMBOL(user_path_at);
int __check_sticky(struct mnt_idmap *idmap, struct inode *dir,
struct inode *inode)
{
kuid_t fsuid = current_fsuid();
if (vfsuid_eq_kuid(i_uid_into_vfsuid(idmap, inode), fsuid))
return 0;
if (vfsuid_eq_kuid(i_uid_into_vfsuid(idmap, dir), fsuid))
return 0;
return !capable_wrt_inode_uidgid(idmap, inode, CAP_FOWNER);
}
EXPORT_SYMBOL(__check_sticky);
/*
* Check whether we can remove a link victim from directory dir, check
* whether the type of victim is right.
* 1. We can't do it if dir is read-only (done in permission())
* 2. We should have write and exec permissions on dir
* 3. We can't remove anything from append-only dir
* 4. We can't do anything with immutable dir (done in permission())
* 5. If the sticky bit on dir is set we should either
* a. be owner of dir, or
* b. be owner of victim, or
* c. have CAP_FOWNER capability
* 6. If the victim is append-only or immutable we can't do antyhing with
* links pointing to it.
vfs: Don't modify inodes with a uid or gid unknown to the vfs When a filesystem outside of init_user_ns is mounted it could have uids and gids stored in it that do not map to init_user_ns. The plan is to allow those filesystems to set i_uid to INVALID_UID and i_gid to INVALID_GID for unmapped uids and gids and then to handle that strange case in the vfs to ensure there is consistent robust handling of the weirdness. Upon a careful review of the vfs and filesystems about the only case where there is any possibility of confusion or trouble is when the inode is written back to disk. In that case filesystems typically read the inode->i_uid and inode->i_gid and write them to disk even when just an inode timestamp is being updated. Which leads to a rule that is very simple to implement and understand inodes whose i_uid or i_gid is not valid may not be written. In dealing with access times this means treat those inodes as if the inode flag S_NOATIME was set. Reads of the inodes appear safe and useful, but any write or modification is disallowed. The only inode write that is allowed is a chown that sets the uid and gid on the inode to valid values. After such a chown the inode is normal and may be treated as such. Denying all writes to inodes with uids or gids unknown to the vfs also prevents several oddball cases where corruption would have occurred because the vfs does not have complete information. One problem case that is prevented is attempting to use the gid of a directory for new inodes where the directories sgid bit is set but the directories gid is not mapped. Another problem case avoided is attempting to update the evm hash after setxattr, removexattr, and setattr. As the evm hash includeds the inode->i_uid or inode->i_gid not knowning the uid or gid prevents a correct evm hash from being computed. evm hash verification also fails when i_uid or i_gid is unknown but that is essentially harmless as it does not cause filesystem corruption. Acked-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-06-30 03:54:46 +08:00
* 7. If the victim has an unknown uid or gid we can't change the inode.
* 8. If we were asked to remove a directory and victim isn't one - ENOTDIR.
* 9. If we were asked to remove a non-directory and victim isn't one - EISDIR.
* 10. We can't remove a root or mountpoint.
* 11. We don't allow removal of NFS sillyrenamed files; it's handled by
* nfs_async_unlink().
*/
static int may_delete(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *victim, bool isdir)
{
struct inode *inode = d_backing_inode(victim);
int error;
if (d_is_negative(victim))
return -ENOENT;
BUG_ON(!inode);
BUG_ON(victim->d_parent->d_inode != dir);
/* Inode writeback is not safe when the uid or gid are invalid. */
if (!vfsuid_valid(i_uid_into_vfsuid(idmap, inode)) ||
!vfsgid_valid(i_gid_into_vfsgid(idmap, inode)))
return -EOVERFLOW;
audit_inode_child(dir, victim, AUDIT_TYPE_CHILD_DELETE);
error = inode_permission(idmap, dir, MAY_WRITE | MAY_EXEC);
if (error)
return error;
if (IS_APPEND(dir))
return -EPERM;
if (check_sticky(idmap, dir, inode) || IS_APPEND(inode) ||
IS_IMMUTABLE(inode) || IS_SWAPFILE(inode) ||
HAS_UNMAPPED_ID(idmap, inode))
return -EPERM;
if (isdir) {
if (!d_is_dir(victim))
return -ENOTDIR;
if (IS_ROOT(victim))
return -EBUSY;
} else if (d_is_dir(victim))
return -EISDIR;
if (IS_DEADDIR(dir))
return -ENOENT;
if (victim->d_flags & DCACHE_NFSFS_RENAMED)
return -EBUSY;
return 0;
}
/* Check whether we can create an object with dentry child in directory
* dir.
* 1. We can't do it if child already exists (open has special treatment for
* this case, but since we are inlined it's OK)
* 2. We can't do it if dir is read-only (done in permission())
* 3. We can't do it if the fs can't represent the fsuid or fsgid.
* 4. We should have write and exec permissions on dir
* 5. We can't do it if dir is immutable (done in permission())
*/
static inline int may_create(struct mnt_idmap *idmap,
struct inode *dir, struct dentry *child)
{
audit: add child record before the create to handle case where create fails Historically, when a syscall that creates a dentry fails, you get an audit record that looks something like this (when trying to create a file named "new" in "/tmp/tmp.SxiLnCcv63"): type=PATH msg=audit(1366128956.279:965): item=0 name="/tmp/tmp.SxiLnCcv63/new" inode=2138308 dev=fd:02 mode=040700 ouid=0 ogid=0 rdev=00:00 obj=staff_u:object_r:user_tmp_t:s15:c0.c1023 This record makes no sense since it's associating the inode information for "/tmp/tmp.SxiLnCcv63" with the path "/tmp/tmp.SxiLnCcv63/new". The recent patch I posted to fix the audit_inode call in do_last fixes this, by making it look more like this: type=PATH msg=audit(1366128765.989:13875): item=0 name="/tmp/tmp.DJ1O8V3e4f/" inode=141 dev=fd:02 mode=040700 ouid=0 ogid=0 rdev=00:00 obj=staff_u:object_r:user_tmp_t:s15:c0.c1023 While this is more correct, if the creation of the file fails, then we have no record of the filename that the user tried to create. This patch adds a call to audit_inode_child to may_create. This creates an AUDIT_TYPE_CHILD_CREATE record that will sit in place until the create succeeds. When and if the create does succeed, then this record will be updated with the correct inode info from the create. This fixes what was broken in commit bfcec708. Commit 79f6530c should also be backported to stable v3.7+. Signed-off-by: Jeff Layton <jlayton@redhat.com> Signed-off-by: Eric Paris <eparis@redhat.com> Signed-off-by: Richard Guy Briggs <rgb@redhat.com> Signed-off-by: Eric Paris <eparis@redhat.com>
2013-05-08 22:25:58 +08:00
audit_inode_child(dir, child, AUDIT_TYPE_CHILD_CREATE);
if (child->d_inode)
return -EEXIST;
if (IS_DEADDIR(dir))
return -ENOENT;
if (!fsuidgid_has_mapping(dir->i_sb, idmap))
return -EOVERFLOW;
return inode_permission(idmap, dir, MAY_WRITE | MAY_EXEC);
}
// p1 != p2, both are on the same filesystem, ->s_vfs_rename_mutex is held
static struct dentry *lock_two_directories(struct dentry *p1, struct dentry *p2)
{
struct dentry *p = p1, *q = p2, *r;
while ((r = p->d_parent) != p2 && r != p)
p = r;
if (r == p2) {
// p is a child of p2 and an ancestor of p1 or p1 itself
inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
inode_lock_nested(p1->d_inode, I_MUTEX_PARENT2);
return p;
}
// p is the root of connected component that contains p1
// p2 does not occur on the path from p to p1
while ((r = q->d_parent) != p1 && r != p && r != q)
q = r;
if (r == p1) {
// q is a child of p1 and an ancestor of p2 or p2 itself
inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
return q;
} else if (likely(r == p)) {
// both p2 and p1 are descendents of p
inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
inode_lock_nested(p2->d_inode, I_MUTEX_PARENT2);
return NULL;
} else { // no common ancestor at the time we'd been called
mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
return ERR_PTR(-EXDEV);
}
}
/*
* p1 and p2 should be directories on the same fs.
*/
struct dentry *lock_rename(struct dentry *p1, struct dentry *p2)
{
if (p1 == p2) {
inode_lock_nested(p1->d_inode, I_MUTEX_PARENT);
return NULL;
}
mutex_lock(&p1->d_sb->s_vfs_rename_mutex);
return lock_two_directories(p1, p2);
}
EXPORT_SYMBOL(lock_rename);
/*
* c1 and p2 should be on the same fs.
*/
struct dentry *lock_rename_child(struct dentry *c1, struct dentry *p2)
{
if (READ_ONCE(c1->d_parent) == p2) {
/*
* hopefully won't need to touch ->s_vfs_rename_mutex at all.
*/
inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
/*
* now that p2 is locked, nobody can move in or out of it,
* so the test below is safe.
*/
if (likely(c1->d_parent == p2))
return NULL;
/*
* c1 got moved out of p2 while we'd been taking locks;
* unlock and fall back to slow case.
*/
inode_unlock(p2->d_inode);
}
mutex_lock(&c1->d_sb->s_vfs_rename_mutex);
/*
* nobody can move out of any directories on this fs.
*/
if (likely(c1->d_parent != p2))
return lock_two_directories(c1->d_parent, p2);
/*
* c1 got moved into p2 while we were taking locks;
* we need p2 locked and ->s_vfs_rename_mutex unlocked,
* for consistency with lock_rename().
*/
inode_lock_nested(p2->d_inode, I_MUTEX_PARENT);
mutex_unlock(&c1->d_sb->s_vfs_rename_mutex);
return NULL;
}
EXPORT_SYMBOL(lock_rename_child);
void unlock_rename(struct dentry *p1, struct dentry *p2)
{
inode_unlock(p1->d_inode);
if (p1 != p2) {
inode_unlock(p2->d_inode);
mutex_unlock(&p1->d_sb->s_vfs_rename_mutex);
}
}
EXPORT_SYMBOL(unlock_rename);
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
/**
* vfs_prepare_mode - prepare the mode to be used for a new inode
* @idmap: idmap of the mount the inode was found from
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
* @dir: parent directory of the new inode
* @mode: mode of the new inode
* @mask_perms: allowed permission by the vfs
* @type: type of file to be created
*
* This helper consolidates and enforces vfs restrictions on the @mode of a new
* object to be created.
*
* Umask stripping depends on whether the filesystem supports POSIX ACLs (see
* the kernel documentation for mode_strip_umask()). Moving umask stripping
* after setgid stripping allows the same ordering for both non-POSIX ACL and
* POSIX ACL supporting filesystems.
*
* Note that it's currently valid for @type to be 0 if a directory is created.
* Filesystems raise that flag individually and we need to check whether each
* filesystem can deal with receiving S_IFDIR from the vfs before we enforce a
* non-zero type.
*
* Returns: mode to be passed to the filesystem
*/
static inline umode_t vfs_prepare_mode(struct mnt_idmap *idmap,
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
const struct inode *dir, umode_t mode,
umode_t mask_perms, umode_t type)
{
mode = mode_strip_sgid(idmap, dir, mode);
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
mode = mode_strip_umask(dir, mode);
/*
* Apply the vfs mandated allowed permission mask and set the type of
* file to be created before we call into the filesystem.
*/
mode &= (mask_perms & ~S_IFMT);
mode |= (type & S_IFMT);
return mode;
}
/**
* vfs_create - create new file
* @idmap: idmap of the mount the inode was found from
* @dir: inode of the parent directory
* @dentry: dentry of the child file
* @mode: mode of the child file
* @want_excl: whether the file must not yet exist
*
* Create a new file.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_create(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *dentry, umode_t mode, bool want_excl)
{
int error;
error = may_create(idmap, dir, dentry);
if (error)
return error;
if (!dir->i_op->create)
return -EACCES; /* shouldn't it be ENOSYS? */
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
mode = vfs_prepare_mode(idmap, dir, mode, S_IALLUGO, S_IFREG);
error = security_inode_create(dir, dentry, mode);
if (error)
return error;
error = dir->i_op->create(idmap, dir, dentry, mode, want_excl);
if (!error)
fsnotify_create(dir, dentry);
return error;
}
EXPORT_SYMBOL(vfs_create);
int vfs_mkobj(struct dentry *dentry, umode_t mode,
int (*f)(struct dentry *, umode_t, void *),
void *arg)
{
struct inode *dir = dentry->d_parent->d_inode;
int error = may_create(&nop_mnt_idmap, dir, dentry);
if (error)
return error;
mode &= S_IALLUGO;
mode |= S_IFREG;
error = security_inode_create(dir, dentry, mode);
if (error)
return error;
error = f(dentry, mode, arg);
if (!error)
fsnotify_create(dir, dentry);
return error;
}
EXPORT_SYMBOL(vfs_mkobj);
bool may_open_dev(const struct path *path)
{
return !(path->mnt->mnt_flags & MNT_NODEV) &&
!(path->mnt->mnt_sb->s_iflags & SB_I_NODEV);
}
static int may_open(struct mnt_idmap *idmap, const struct path *path,
int acc_mode, int flag)
{
struct dentry *dentry = path->dentry;
struct inode *inode = dentry->d_inode;
int error;
if (!inode)
return -ENOENT;
switch (inode->i_mode & S_IFMT) {
case S_IFLNK:
return -ELOOP;
case S_IFDIR:
if (acc_mode & MAY_WRITE)
return -EISDIR;
if (acc_mode & MAY_EXEC)
return -EACCES;
break;
case S_IFBLK:
case S_IFCHR:
if (!may_open_dev(path))
return -EACCES;
exec: move S_ISREG() check earlier The execve(2)/uselib(2) syscalls have always rejected non-regular files. Recently, it was noticed that a deadlock was introduced when trying to execute pipes, as the S_ISREG() test was happening too late. This was fixed in commit 73601ea5b7b1 ("fs/open.c: allow opening only regular files during execve()"), but it was added after inode_permission() had already run, which meant LSMs could see bogus attempts to execute non-regular files. Move the test into the other inode type checks (which already look for other pathological conditions[1]). Since there is no need to use FMODE_EXEC while we still have access to "acc_mode", also switch the test to MAY_EXEC. Also include a comment with the redundant S_ISREG() checks at the end of execve(2)/uselib(2) to note that they are present to avoid any mistakes. My notes on the call path, and related arguments, checks, etc: do_open_execat() struct open_flags open_exec_flags = { .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC, .acc_mode = MAY_EXEC, ... do_filp_open(dfd, filename, open_flags) path_openat(nameidata, open_flags, flags) file = alloc_empty_file(open_flags, current_cred()); do_open(nameidata, file, open_flags) may_open(path, acc_mode, open_flag) /* new location of MAY_EXEC vs S_ISREG() test */ inode_permission(inode, MAY_OPEN | acc_mode) security_inode_permission(inode, acc_mode) vfs_open(path, file) do_dentry_open(file, path->dentry->d_inode, open) /* old location of FMODE_EXEC vs S_ISREG() test */ security_file_open(f) open() [1] https://lore.kernel.org/lkml/202006041910.9EF0C602@keescook/ Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Aleksa Sarai <cyphar@cyphar.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <christian.brauner@ubuntu.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Eric Biggers <ebiggers3@gmail.com> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Link: http://lkml.kernel.org/r/20200605160013.3954297-3-keescook@chromium.org Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-12 09:36:26 +08:00
fallthrough;
case S_IFIFO:
case S_IFSOCK:
exec: move S_ISREG() check earlier The execve(2)/uselib(2) syscalls have always rejected non-regular files. Recently, it was noticed that a deadlock was introduced when trying to execute pipes, as the S_ISREG() test was happening too late. This was fixed in commit 73601ea5b7b1 ("fs/open.c: allow opening only regular files during execve()"), but it was added after inode_permission() had already run, which meant LSMs could see bogus attempts to execute non-regular files. Move the test into the other inode type checks (which already look for other pathological conditions[1]). Since there is no need to use FMODE_EXEC while we still have access to "acc_mode", also switch the test to MAY_EXEC. Also include a comment with the redundant S_ISREG() checks at the end of execve(2)/uselib(2) to note that they are present to avoid any mistakes. My notes on the call path, and related arguments, checks, etc: do_open_execat() struct open_flags open_exec_flags = { .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC, .acc_mode = MAY_EXEC, ... do_filp_open(dfd, filename, open_flags) path_openat(nameidata, open_flags, flags) file = alloc_empty_file(open_flags, current_cred()); do_open(nameidata, file, open_flags) may_open(path, acc_mode, open_flag) /* new location of MAY_EXEC vs S_ISREG() test */ inode_permission(inode, MAY_OPEN | acc_mode) security_inode_permission(inode, acc_mode) vfs_open(path, file) do_dentry_open(file, path->dentry->d_inode, open) /* old location of FMODE_EXEC vs S_ISREG() test */ security_file_open(f) open() [1] https://lore.kernel.org/lkml/202006041910.9EF0C602@keescook/ Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Aleksa Sarai <cyphar@cyphar.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <christian.brauner@ubuntu.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Eric Biggers <ebiggers3@gmail.com> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Link: http://lkml.kernel.org/r/20200605160013.3954297-3-keescook@chromium.org Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-12 09:36:26 +08:00
if (acc_mode & MAY_EXEC)
return -EACCES;
flag &= ~O_TRUNC;
break;
exec: move path_noexec() check earlier The path_noexec() check, like the regular file check, was happening too late, letting LSMs see impossible execve()s. Check it earlier as well in may_open() and collect the redundant fs/exec.c path_noexec() test under the same robustness comment as the S_ISREG() check. My notes on the call path, and related arguments, checks, etc: do_open_execat() struct open_flags open_exec_flags = { .open_flag = O_LARGEFILE | O_RDONLY | __FMODE_EXEC, .acc_mode = MAY_EXEC, ... do_filp_open(dfd, filename, open_flags) path_openat(nameidata, open_flags, flags) file = alloc_empty_file(open_flags, current_cred()); do_open(nameidata, file, open_flags) may_open(path, acc_mode, open_flag) /* new location of MAY_EXEC vs path_noexec() test */ inode_permission(inode, MAY_OPEN | acc_mode) security_inode_permission(inode, acc_mode) vfs_open(path, file) do_dentry_open(file, path->dentry->d_inode, open) security_file_open(f) open() /* old location of path_noexec() test */ Signed-off-by: Kees Cook <keescook@chromium.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Aleksa Sarai <cyphar@cyphar.com> Cc: Christian Brauner <christian.brauner@ubuntu.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Eric Biggers <ebiggers3@gmail.com> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Link: http://lkml.kernel.org/r/20200605160013.3954297-4-keescook@chromium.org Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-08-12 09:36:30 +08:00
case S_IFREG:
if ((acc_mode & MAY_EXEC) && path_noexec(path))
return -EACCES;
break;
}
error = inode_permission(idmap, inode, MAY_OPEN | acc_mode);
if (error)
return error;
/*
* An append-only file must be opened in append mode for writing.
*/
if (IS_APPEND(inode)) {
if ((flag & O_ACCMODE) != O_RDONLY && !(flag & O_APPEND))
return -EPERM;
if (flag & O_TRUNC)
return -EPERM;
}
/* O_NOATIME can only be set by the owner or superuser */
if (flag & O_NOATIME && !inode_owner_or_capable(idmap, inode))
return -EPERM;
return 0;
}
static int handle_truncate(struct mnt_idmap *idmap, struct file *filp)
{
const struct path *path = &filp->f_path;
struct inode *inode = path->dentry->d_inode;
int error = get_write_access(inode);
if (error)
return error;
error = security_file_truncate(filp);
if (!error) {
error = do_truncate(idmap, path->dentry, 0,
ATTR_MTIME|ATTR_CTIME|ATTR_OPEN,
filp);
}
put_write_access(inode);
return error;
}
static inline int open_to_namei_flags(int flag)
{
if ((flag & O_ACCMODE) == 3)
flag--;
return flag;
}
static int may_o_create(struct mnt_idmap *idmap,
const struct path *dir, struct dentry *dentry,
umode_t mode)
{
int error = security_path_mknod(dir, dentry, mode, 0);
if (error)
return error;
if (!fsuidgid_has_mapping(dir->dentry->d_sb, idmap))
return -EOVERFLOW;
error = inode_permission(idmap, dir->dentry->d_inode,
MAY_WRITE | MAY_EXEC);
if (error)
return error;
return security_inode_create(dir->dentry->d_inode, dentry, mode);
}
/*
* Attempt to atomically look up, create and open a file from a negative
* dentry.
*
* Returns 0 if successful. The file will have been created and attached to
* @file by the filesystem calling finish_open().
*
* If the file was looked up only or didn't need creating, FMODE_OPENED won't
* be set. The caller will need to perform the open themselves. @path will
* have been updated to point to the new dentry. This may be negative.
*
* Returns an error code otherwise.
*/
static struct dentry *atomic_open(struct nameidata *nd, struct dentry *dentry,
struct file *file,
int open_flag, umode_t mode)
{
struct dentry *const DENTRY_NOT_SET = (void *) -1UL;
struct inode *dir = nd->path.dentry->d_inode;
int error;
if (nd->flags & LOOKUP_DIRECTORY)
open_flag |= O_DIRECTORY;
file->f_path.dentry = DENTRY_NOT_SET;
file->f_path.mnt = nd->path.mnt;
error = dir->i_op->atomic_open(dir, dentry, file,
open_to_namei_flags(open_flag), mode);
d_lookup_done(dentry);
if (!error) {
if (file->f_mode & FMODE_OPENED) {
if (unlikely(dentry != file->f_path.dentry)) {
dput(dentry);
dentry = dget(file->f_path.dentry);
}
} else if (WARN_ON(file->f_path.dentry == DENTRY_NOT_SET)) {
error = -EIO;
} else {
if (file->f_path.dentry) {
dput(dentry);
dentry = file->f_path.dentry;
}
if (unlikely(d_is_negative(dentry)))
error = -ENOENT;
}
}
if (error) {
dput(dentry);
dentry = ERR_PTR(error);
}
return dentry;
}
/*
* Look up and maybe create and open the last component.
*
* Must be called with parent locked (exclusive in O_CREAT case).
*
* Returns 0 on success, that is, if
* the file was successfully atomically created (if necessary) and opened, or
* the file was not completely opened at this time, though lookups and
* creations were performed.
* These case are distinguished by presence of FMODE_OPENED on file->f_mode.
* In the latter case dentry returned in @path might be negative if O_CREAT
* hadn't been specified.
*
* An error code is returned on failure.
*/
static struct dentry *lookup_open(struct nameidata *nd, struct file *file,
const struct open_flags *op,
bool got_write)
{
struct mnt_idmap *idmap;
struct dentry *dir = nd->path.dentry;
struct inode *dir_inode = dir->d_inode;
int open_flag = op->open_flag;
struct dentry *dentry;
int error, create_error = 0;
umode_t mode = op->mode;
DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq);
if (unlikely(IS_DEADDIR(dir_inode)))
return ERR_PTR(-ENOENT);
file->f_mode &= ~FMODE_CREATED;
dentry = d_lookup(dir, &nd->last);
for (;;) {
if (!dentry) {
dentry = d_alloc_parallel(dir, &nd->last, &wq);
if (IS_ERR(dentry))
return dentry;
}
if (d_in_lookup(dentry))
break;
error = d_revalidate(dentry, nd->flags);
if (likely(error > 0))
break;
if (error)
goto out_dput;
d_invalidate(dentry);
dput(dentry);
dentry = NULL;
}
if (dentry->d_inode) {
/* Cached positive dentry: will open in f_op->open */
return dentry;
}
if (open_flag & O_CREAT)
audit_inode(nd->name, dir, AUDIT_INODE_PARENT);
/*
* Checking write permission is tricky, bacuse we don't know if we are
* going to actually need it: O_CREAT opens should work as long as the
* file exists. But checking existence breaks atomicity. The trick is
* to check access and if not granted clear O_CREAT from the flags.
*
* Another problem is returing the "right" error value (e.g. for an
* O_EXCL open we want to return EEXIST not EROFS).
*/
lookup_open(): don't bother with fallbacks to lookup+create We fall back to lookup+create (instead of atomic_open) in several cases: 1) we don't have write access to filesystem and O_TRUNC is present in the flags. It's not something we want ->atomic_open() to see - it just might go ahead and truncate the file. However, we can pass it the flags sans O_TRUNC - eventually do_open() will call handle_truncate() anyway. 2) we have O_CREAT | O_EXCL and we can't write to parent. That's going to be an error, of course, but we want to know _which_ error should that be - might be EEXIST (if file exists), might be EACCES or EROFS. Simply stripping O_CREAT (and checking if we see ENOENT) would suffice, if not for O_EXCL. However, we used to have ->atomic_open() fully responsible for rejecting O_CREAT | O_EXCL on existing file and just stripping O_CREAT would've disarmed those checks. With nothing downstream to catch the problem - FMODE_OPENED used to be "don't bother with EEXIST checks, ->atomic_open() has done those". Now EEXIST checks downstream are skipped only if FMODE_CREATED is set - FMODE_OPENED alone is not enough. That has eliminated the need to fall back onto lookup+create path in this case. 3) O_WRONLY or O_RDWR when we have no write access to filesystem, with nothing else objectionable. Fallback is (and had always been) pointless. IOW, we don't really need that fallback; all we need in such cases is to trim O_TRUNC and O_CREAT properly. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-03-13 02:07:27 +08:00
if (unlikely(!got_write))
open_flag &= ~O_TRUNC;
idmap = mnt_idmap(nd->path.mnt);
if (open_flag & O_CREAT) {
lookup_open(): don't bother with fallbacks to lookup+create We fall back to lookup+create (instead of atomic_open) in several cases: 1) we don't have write access to filesystem and O_TRUNC is present in the flags. It's not something we want ->atomic_open() to see - it just might go ahead and truncate the file. However, we can pass it the flags sans O_TRUNC - eventually do_open() will call handle_truncate() anyway. 2) we have O_CREAT | O_EXCL and we can't write to parent. That's going to be an error, of course, but we want to know _which_ error should that be - might be EEXIST (if file exists), might be EACCES or EROFS. Simply stripping O_CREAT (and checking if we see ENOENT) would suffice, if not for O_EXCL. However, we used to have ->atomic_open() fully responsible for rejecting O_CREAT | O_EXCL on existing file and just stripping O_CREAT would've disarmed those checks. With nothing downstream to catch the problem - FMODE_OPENED used to be "don't bother with EEXIST checks, ->atomic_open() has done those". Now EEXIST checks downstream are skipped only if FMODE_CREATED is set - FMODE_OPENED alone is not enough. That has eliminated the need to fall back onto lookup+create path in this case. 3) O_WRONLY or O_RDWR when we have no write access to filesystem, with nothing else objectionable. Fallback is (and had always been) pointless. IOW, we don't really need that fallback; all we need in such cases is to trim O_TRUNC and O_CREAT properly. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-03-13 02:07:27 +08:00
if (open_flag & O_EXCL)
open_flag &= ~O_TRUNC;
mode = vfs_prepare_mode(idmap, dir->d_inode, mode, mode, mode);
lookup_open(): don't bother with fallbacks to lookup+create We fall back to lookup+create (instead of atomic_open) in several cases: 1) we don't have write access to filesystem and O_TRUNC is present in the flags. It's not something we want ->atomic_open() to see - it just might go ahead and truncate the file. However, we can pass it the flags sans O_TRUNC - eventually do_open() will call handle_truncate() anyway. 2) we have O_CREAT | O_EXCL and we can't write to parent. That's going to be an error, of course, but we want to know _which_ error should that be - might be EEXIST (if file exists), might be EACCES or EROFS. Simply stripping O_CREAT (and checking if we see ENOENT) would suffice, if not for O_EXCL. However, we used to have ->atomic_open() fully responsible for rejecting O_CREAT | O_EXCL on existing file and just stripping O_CREAT would've disarmed those checks. With nothing downstream to catch the problem - FMODE_OPENED used to be "don't bother with EEXIST checks, ->atomic_open() has done those". Now EEXIST checks downstream are skipped only if FMODE_CREATED is set - FMODE_OPENED alone is not enough. That has eliminated the need to fall back onto lookup+create path in this case. 3) O_WRONLY or O_RDWR when we have no write access to filesystem, with nothing else objectionable. Fallback is (and had always been) pointless. IOW, we don't really need that fallback; all we need in such cases is to trim O_TRUNC and O_CREAT properly. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-03-13 02:07:27 +08:00
if (likely(got_write))
create_error = may_o_create(idmap, &nd->path,
dentry, mode);
lookup_open(): don't bother with fallbacks to lookup+create We fall back to lookup+create (instead of atomic_open) in several cases: 1) we don't have write access to filesystem and O_TRUNC is present in the flags. It's not something we want ->atomic_open() to see - it just might go ahead and truncate the file. However, we can pass it the flags sans O_TRUNC - eventually do_open() will call handle_truncate() anyway. 2) we have O_CREAT | O_EXCL and we can't write to parent. That's going to be an error, of course, but we want to know _which_ error should that be - might be EEXIST (if file exists), might be EACCES or EROFS. Simply stripping O_CREAT (and checking if we see ENOENT) would suffice, if not for O_EXCL. However, we used to have ->atomic_open() fully responsible for rejecting O_CREAT | O_EXCL on existing file and just stripping O_CREAT would've disarmed those checks. With nothing downstream to catch the problem - FMODE_OPENED used to be "don't bother with EEXIST checks, ->atomic_open() has done those". Now EEXIST checks downstream are skipped only if FMODE_CREATED is set - FMODE_OPENED alone is not enough. That has eliminated the need to fall back onto lookup+create path in this case. 3) O_WRONLY or O_RDWR when we have no write access to filesystem, with nothing else objectionable. Fallback is (and had always been) pointless. IOW, we don't really need that fallback; all we need in such cases is to trim O_TRUNC and O_CREAT properly. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-03-13 02:07:27 +08:00
else
create_error = -EROFS;
}
lookup_open(): don't bother with fallbacks to lookup+create We fall back to lookup+create (instead of atomic_open) in several cases: 1) we don't have write access to filesystem and O_TRUNC is present in the flags. It's not something we want ->atomic_open() to see - it just might go ahead and truncate the file. However, we can pass it the flags sans O_TRUNC - eventually do_open() will call handle_truncate() anyway. 2) we have O_CREAT | O_EXCL and we can't write to parent. That's going to be an error, of course, but we want to know _which_ error should that be - might be EEXIST (if file exists), might be EACCES or EROFS. Simply stripping O_CREAT (and checking if we see ENOENT) would suffice, if not for O_EXCL. However, we used to have ->atomic_open() fully responsible for rejecting O_CREAT | O_EXCL on existing file and just stripping O_CREAT would've disarmed those checks. With nothing downstream to catch the problem - FMODE_OPENED used to be "don't bother with EEXIST checks, ->atomic_open() has done those". Now EEXIST checks downstream are skipped only if FMODE_CREATED is set - FMODE_OPENED alone is not enough. That has eliminated the need to fall back onto lookup+create path in this case. 3) O_WRONLY or O_RDWR when we have no write access to filesystem, with nothing else objectionable. Fallback is (and had always been) pointless. IOW, we don't really need that fallback; all we need in such cases is to trim O_TRUNC and O_CREAT properly. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-03-13 02:07:27 +08:00
if (create_error)
open_flag &= ~O_CREAT;
if (dir_inode->i_op->atomic_open) {
dentry = atomic_open(nd, dentry, file, open_flag, mode);
if (unlikely(create_error) && dentry == ERR_PTR(-ENOENT))
dentry = ERR_PTR(create_error);
return dentry;
}
if (d_in_lookup(dentry)) {
struct dentry *res = dir_inode->i_op->lookup(dir_inode, dentry,
nd->flags);
d_lookup_done(dentry);
if (unlikely(res)) {
if (IS_ERR(res)) {
error = PTR_ERR(res);
goto out_dput;
}
dput(dentry);
dentry = res;
}
}
/* Negative dentry, just create the file */
if (!dentry->d_inode && (open_flag & O_CREAT)) {
file->f_mode |= FMODE_CREATED;
audit_inode_child(dir_inode, dentry, AUDIT_TYPE_CHILD_CREATE);
if (!dir_inode->i_op->create) {
error = -EACCES;
goto out_dput;
}
error = dir_inode->i_op->create(idmap, dir_inode, dentry,
mode, open_flag & O_EXCL);
if (error)
goto out_dput;
}
if (unlikely(create_error) && !dentry->d_inode) {
error = create_error;
goto out_dput;
}
return dentry;
out_dput:
dput(dentry);
return ERR_PTR(error);
}
static inline bool trailing_slashes(struct nameidata *nd)
{
return (bool)nd->last.name[nd->last.len];
}
static struct dentry *lookup_fast_for_open(struct nameidata *nd, int open_flag)
{
struct dentry *dentry;
if (open_flag & O_CREAT) {
if (trailing_slashes(nd))
return ERR_PTR(-EISDIR);
/* Don't bother on an O_EXCL create */
if (open_flag & O_EXCL)
return NULL;
}
if (trailing_slashes(nd))
nd->flags |= LOOKUP_FOLLOW | LOOKUP_DIRECTORY;
dentry = lookup_fast(nd);
if (IS_ERR_OR_NULL(dentry))
return dentry;
if (open_flag & O_CREAT) {
/* Discard negative dentries. Need inode_lock to do the create */
if (!dentry->d_inode) {
if (!(nd->flags & LOOKUP_RCU))
dput(dentry);
dentry = NULL;
}
}
return dentry;
}
static const char *open_last_lookups(struct nameidata *nd,
struct file *file, const struct open_flags *op)
{
struct dentry *dir = nd->path.dentry;
int open_flag = op->open_flag;
bool got_write = false;
struct dentry *dentry;
const char *res;
nd->flags |= op->intent;
if (nd->last_type != LAST_NORM) {
if (nd->depth)
put_link(nd);
return handle_dots(nd, nd->last_type);
}
/* We _can_ be in RCU mode here */
dentry = lookup_fast_for_open(nd, open_flag);
if (IS_ERR(dentry))
return ERR_CAST(dentry);
if (likely(dentry))
goto finish_lookup;
if (!(open_flag & O_CREAT)) {
if (WARN_ON_ONCE(nd->flags & LOOKUP_RCU))
return ERR_PTR(-ECHILD);
} else {
if (nd->flags & LOOKUP_RCU) {
if (!try_to_unlazy(nd))
return ERR_PTR(-ECHILD);
}
}
if (open_flag & (O_CREAT | O_TRUNC | O_WRONLY | O_RDWR)) {
got_write = !mnt_want_write(nd->path.mnt);
/*
* do _not_ fail yet - we might not need that or fail with
* a different error; let lookup_open() decide; we'll be
* dropping this one anyway.
*/
}
if (open_flag & O_CREAT)
inode_lock(dir->d_inode);
else
inode_lock_shared(dir->d_inode);
dentry = lookup_open(nd, file, op, got_write);
vfs: generate FS_CREATE before FS_OPEN when ->atomic_open used. When a file is opened and created with open(..., O_CREAT) we get both the CREATE and OPEN fsnotify events and would expect them in that order. For most filesystems we get them in that order because open_last_lookups() calls fsnofify_create() and then do_open() (from path_openat()) calls vfs_open()->do_dentry_open() which calls fsnotify_open(). However when ->atomic_open is used, the do_dentry_open() -> fsnotify_open() call happens from finish_open() which is called from the ->atomic_open handler in lookup_open() which is called *before* open_last_lookups() calls fsnotify_create. So we get the "open" notification before "create" - which is backwards. ltp testcase inotify02 tests this and reports the inconsistency. This patch lifts the fsnotify_open() call out of do_dentry_open() and places it higher up the call stack. There are three callers of do_dentry_open(). For vfs_open() and kernel_file_open() the fsnotify_open() is placed directly in that caller so there should be no behavioural change. For finish_open() there are two cases: - finish_open is used in ->atomic_open handlers. For these we add a call to fsnotify_open() at open_last_lookups() if FMODE_OPENED is set - which means do_dentry_open() has been called. - finish_open is used in ->tmpfile() handlers. For these a similar call to fsnotify_open() is added to vfs_tmpfile() With this patch NFSv3 is restored to its previous behaviour (before ->atomic_open support was added) of generating CREATE notifications before OPEN, and NFSv4 now has that same correct ordering that is has not had before. I haven't tested other filesystems. Fixes: 7c6c5249f061 ("NFS: add atomic_open for NFSv3 to handle O_TRUNC correctly.") Reported-by: James Clark <james.clark@arm.com> Closes: https://lore.kernel.org/all/01c3bf2e-eb1f-4b7f-a54f-d2a05dd3d8c8@arm.com Signed-off-by: NeilBrown <neilb@suse.de> Link: https://lore.kernel.org/r/171817619547.14261.975798725161704336@noble.neil.brown.name Fixes: 7b8c9d7bb457 ("fsnotify: move fsnotify_open() hook into do_dentry_open()") Tested-by: James Clark <james.clark@arm.com> Signed-off-by: Jan Kara <jack@suse.cz> Link: https://lore.kernel.org/r/20240617162303.1596-2-jack@suse.cz Reviewed-by: Amir Goldstein <amir73il@gmail.com> Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-06-18 00:23:01 +08:00
if (!IS_ERR(dentry)) {
if (file->f_mode & FMODE_CREATED)
fsnotify_create(dir->d_inode, dentry);
if (file->f_mode & FMODE_OPENED)
fsnotify_open(file);
}
if (open_flag & O_CREAT)
inode_unlock(dir->d_inode);
else
inode_unlock_shared(dir->d_inode);
if (got_write)
mnt_drop_write(nd->path.mnt);
if (IS_ERR(dentry))
return ERR_CAST(dentry);
if (file->f_mode & (FMODE_OPENED | FMODE_CREATED)) {
dput(nd->path.dentry);
nd->path.dentry = dentry;
return NULL;
}
finish_lookup:
if (nd->depth)
put_link(nd);
res = step_into(nd, WALK_TRAILING, dentry);
if (unlikely(res))
nd->flags &= ~(LOOKUP_OPEN|LOOKUP_CREATE|LOOKUP_EXCL);
return res;
}
/*
* Handle the last step of open()
*/
static int do_open(struct nameidata *nd,
struct file *file, const struct open_flags *op)
{
struct mnt_idmap *idmap;
int open_flag = op->open_flag;
bool do_truncate;
int acc_mode;
int error;
if (!(file->f_mode & (FMODE_OPENED | FMODE_CREATED))) {
error = complete_walk(nd);
if (error)
return error;
}
if (!(file->f_mode & FMODE_CREATED))
audit_inode(nd->name, nd->path.dentry, 0);
idmap = mnt_idmap(nd->path.mnt);
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
if (open_flag & O_CREAT) {
open_last_lookups(): lift O_EXCL|O_CREAT handling into do_open() Currently path_openat() has "EEXIST on O_EXCL|O_CREAT" checks done on one of the ways out of open_last_lookups(). There are 4 cases: 1) the last component is . or ..; check is not done. 2) we had FMODE_OPENED or FMODE_CREATED set while in lookup_open(); check is not done. 3) symlink to be traversed is found; check is not done (nor should it be) 4) everything else: check done (before complete_walk(), even). In case (1) O_EXCL|O_CREAT ends up failing with -EISDIR - that's open("/tmp/.", O_CREAT|O_EXCL, 0600) Note that in the same conditions open("/tmp", O_CREAT|O_EXCL, 0600) would have yielded EEXIST. Either error is allowed, switching to -EEXIST in these cases would've been more consistent. Case (2) is more subtle; first of all, if we have FMODE_CREATED set, the object hadn't existed prior to the call. The check should not be done in such a case. The rest is problematic, though - we have FMODE_OPENED set (i.e. it went through ->atomic_open() and got successfully opened there) FMODE_CREATED is *NOT* set O_CREAT and O_EXCL are both set. Any such case is a bug - either we failed to set FMODE_CREATED when we had, in fact, created an object (no such instances in the tree) or we have opened a pre-existing file despite having had both O_CREAT and O_EXCL passed. One of those was, in fact caught (and fixed) while sorting out this mess (gfs2 on cold dcache). And in such situations we should fail with EEXIST. Note that for (1) and (4) FMODE_CREATED is not set - for (1) there's nothing in handle_dots() to set it, for (4) we'd explicitly checked that. And (1), (2) and (4) are exactly the cases when we leave the loop in the caller, with do_open() called immediately after that loop. IOW, we can move the check over there, and make it If we have O_CREAT|O_EXCL and after successful pathname resolution FMODE_CREATED is *not* set, we must have run into a preexisting file and should fail with EEXIST. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2020-03-10 22:13:53 +08:00
if ((open_flag & O_EXCL) && !(file->f_mode & FMODE_CREATED))
return -EEXIST;
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
if (d_is_dir(nd->path.dentry))
return -EISDIR;
error = may_create_in_sticky(idmap, nd,
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
d_backing_inode(nd->path.dentry));
if (unlikely(error))
return error;
namei: allow restricted O_CREAT of FIFOs and regular files Disallows open of FIFOs or regular files not owned by the user in world writable sticky directories, unless the owner is the same as that of the directory or the file is opened without the O_CREAT flag. The purpose is to make data spoofing attacks harder. This protection can be turned on and off separately for FIFOs and regular files via sysctl, just like the symlinks/hardlinks protection. This patch is based on Openwall's "HARDEN_FIFO" feature by Solar Designer. This is a brief list of old vulnerabilities that could have been prevented by this feature, some of them even allow for privilege escalation: CVE-2000-1134 CVE-2007-3852 CVE-2008-0525 CVE-2009-0416 CVE-2011-4834 CVE-2015-1838 CVE-2015-7442 CVE-2016-7489 This list is not meant to be complete. It's difficult to track down all vulnerabilities of this kind because they were often reported without any mention of this particular attack vector. In fact, before hardlinks/symlinks restrictions, fifos/regular files weren't the favorite vehicle to exploit them. [s.mesoraca16@gmail.com: fix bug reported by Dan Carpenter] Link: https://lkml.kernel.org/r/20180426081456.GA7060@mwanda Link: http://lkml.kernel.org/r/1524829819-11275-1-git-send-email-s.mesoraca16@gmail.com [keescook@chromium.org: drop pr_warn_ratelimited() in favor of audit changes in the future] [keescook@chromium.org: adjust commit subjet] Link: http://lkml.kernel.org/r/20180416175918.GA13494@beast Signed-off-by: Salvatore Mesoraca <s.mesoraca16@gmail.com> Signed-off-by: Kees Cook <keescook@chromium.org> Suggested-by: Solar Designer <solar@openwall.com> Suggested-by: Kees Cook <keescook@chromium.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-24 08:00:35 +08:00
}
if ((nd->flags & LOOKUP_DIRECTORY) && !d_can_lookup(nd->path.dentry))
return -ENOTDIR;
do_truncate = false;
acc_mode = op->acc_mode;
if (file->f_mode & FMODE_CREATED) {
/* Don't check for write permission, don't truncate */
open_flag &= ~O_TRUNC;
acc_mode = 0;
} else if (d_is_reg(nd->path.dentry) && open_flag & O_TRUNC) {
error = mnt_want_write(nd->path.mnt);
if (error)
return error;
do_truncate = true;
}
error = may_open(idmap, &nd->path, acc_mode, open_flag);
if (!error && !(file->f_mode & FMODE_OPENED))
error = vfs_open(&nd->path, file);
if (!error)
error = security_file_post_open(file, op->acc_mode);
if (!error && do_truncate)
error = handle_truncate(idmap, file);
if (unlikely(error > 0)) {
WARN_ON(1);
error = -EINVAL;
}
if (do_truncate)
mnt_drop_write(nd->path.mnt);
return error;
}
/**
* vfs_tmpfile - create tmpfile
* @idmap: idmap of the mount the inode was found from
* @parentpath: pointer to the path of the base directory
* @file: file descriptor of the new tmpfile
* @mode: mode of the new tmpfile
*
* Create a temporary file.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_tmpfile(struct mnt_idmap *idmap,
const struct path *parentpath,
struct file *file, umode_t mode)
{
struct dentry *child;
struct inode *dir = d_inode(parentpath->dentry);
struct inode *inode;
int error;
int open_flag = file->f_flags;
/* we want directory to be writable */
error = inode_permission(idmap, dir, MAY_WRITE | MAY_EXEC);
if (error)
return error;
if (!dir->i_op->tmpfile)
return -EOPNOTSUPP;
child = d_alloc(parentpath->dentry, &slash_name);
if (unlikely(!child))
return -ENOMEM;
file->f_path.mnt = parentpath->mnt;
file->f_path.dentry = child;
mode = vfs_prepare_mode(idmap, dir, mode, mode, mode);
error = dir->i_op->tmpfile(idmap, dir, file, mode);
dput(child);
vfs: generate FS_CREATE before FS_OPEN when ->atomic_open used. When a file is opened and created with open(..., O_CREAT) we get both the CREATE and OPEN fsnotify events and would expect them in that order. For most filesystems we get them in that order because open_last_lookups() calls fsnofify_create() and then do_open() (from path_openat()) calls vfs_open()->do_dentry_open() which calls fsnotify_open(). However when ->atomic_open is used, the do_dentry_open() -> fsnotify_open() call happens from finish_open() which is called from the ->atomic_open handler in lookup_open() which is called *before* open_last_lookups() calls fsnotify_create. So we get the "open" notification before "create" - which is backwards. ltp testcase inotify02 tests this and reports the inconsistency. This patch lifts the fsnotify_open() call out of do_dentry_open() and places it higher up the call stack. There are three callers of do_dentry_open(). For vfs_open() and kernel_file_open() the fsnotify_open() is placed directly in that caller so there should be no behavioural change. For finish_open() there are two cases: - finish_open is used in ->atomic_open handlers. For these we add a call to fsnotify_open() at open_last_lookups() if FMODE_OPENED is set - which means do_dentry_open() has been called. - finish_open is used in ->tmpfile() handlers. For these a similar call to fsnotify_open() is added to vfs_tmpfile() With this patch NFSv3 is restored to its previous behaviour (before ->atomic_open support was added) of generating CREATE notifications before OPEN, and NFSv4 now has that same correct ordering that is has not had before. I haven't tested other filesystems. Fixes: 7c6c5249f061 ("NFS: add atomic_open for NFSv3 to handle O_TRUNC correctly.") Reported-by: James Clark <james.clark@arm.com> Closes: https://lore.kernel.org/all/01c3bf2e-eb1f-4b7f-a54f-d2a05dd3d8c8@arm.com Signed-off-by: NeilBrown <neilb@suse.de> Link: https://lore.kernel.org/r/171817619547.14261.975798725161704336@noble.neil.brown.name Fixes: 7b8c9d7bb457 ("fsnotify: move fsnotify_open() hook into do_dentry_open()") Tested-by: James Clark <james.clark@arm.com> Signed-off-by: Jan Kara <jack@suse.cz> Link: https://lore.kernel.org/r/20240617162303.1596-2-jack@suse.cz Reviewed-by: Amir Goldstein <amir73il@gmail.com> Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-06-18 00:23:01 +08:00
if (file->f_mode & FMODE_OPENED)
fsnotify_open(file);
if (error)
return error;
/* Don't check for other permissions, the inode was just created */
error = may_open(idmap, &file->f_path, 0, file->f_flags);
if (error)
return error;
inode = file_inode(file);
if (!(open_flag & O_EXCL)) {
spin_lock(&inode->i_lock);
inode->i_state |= I_LINKABLE;
spin_unlock(&inode->i_lock);
}
security_inode_post_create_tmpfile(idmap, inode);
return 0;
}
/**
* kernel_tmpfile_open - open a tmpfile for kernel internal use
* @idmap: idmap of the mount the inode was found from
* @parentpath: path of the base directory
* @mode: mode of the new tmpfile
* @open_flag: flags
* @cred: credentials for open
*
* Create and open a temporary file. The file is not accounted in nr_files,
* hence this is only for kernel internal use, and must not be installed into
* file tables or such.
*/
struct file *kernel_tmpfile_open(struct mnt_idmap *idmap,
const struct path *parentpath,
umode_t mode, int open_flag,
const struct cred *cred)
{
struct file *file;
int error;
file = alloc_empty_file_noaccount(open_flag, cred);
if (IS_ERR(file))
return file;
error = vfs_tmpfile(idmap, parentpath, file, mode);
if (error) {
fput(file);
file = ERR_PTR(error);
}
return file;
}
EXPORT_SYMBOL(kernel_tmpfile_open);
static int do_tmpfile(struct nameidata *nd, unsigned flags,
const struct open_flags *op,
struct file *file)
{
struct path path;
int error = path_lookupat(nd, flags | LOOKUP_DIRECTORY, &path);
if (unlikely(error))
return error;
error = mnt_want_write(path.mnt);
if (unlikely(error))
goto out;
error = vfs_tmpfile(mnt_idmap(path.mnt), &path, file, op->mode);
if (error)
goto out2;
audit_inode(nd->name, file->f_path.dentry, 0);
out2:
mnt_drop_write(path.mnt);
out:
path_put(&path);
return error;
}
static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
{
struct path path;
int error = path_lookupat(nd, flags, &path);
if (!error) {
audit_inode(nd->name, path.dentry, 0);
error = vfs_open(&path, file);
path_put(&path);
}
return error;
}
static struct file *path_openat(struct nameidata *nd,
const struct open_flags *op, unsigned flags)
{
struct file *file;
int error;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
file = alloc_empty_file(op->open_flag, current_cred());
if (IS_ERR(file))
return file;
fs: rcu-walk for path lookup Perform common cases of path lookups without any stores or locking in the ancestor dentry elements. This is called rcu-walk, as opposed to the current algorithm which is a refcount based walk, or ref-walk. This results in far fewer atomic operations on every path element, significantly improving path lookup performance. It also avoids cacheline bouncing on common dentries, significantly improving scalability. The overall design is like this: * LOOKUP_RCU is set in nd->flags, which distinguishes rcu-walk from ref-walk. * Take the RCU lock for the entire path walk, starting with the acquiring of the starting path (eg. root/cwd/fd-path). So now dentry refcounts are not required for dentry persistence. * synchronize_rcu is called when unregistering a filesystem, so we can access d_ops and i_ops during rcu-walk. * Similarly take the vfsmount lock for the entire path walk. So now mnt refcounts are not required for persistence. Also we are free to perform mount lookups, and to assume dentry mount points and mount roots are stable up and down the path. * Have a per-dentry seqlock to protect the dentry name, parent, and inode, so we can load this tuple atomically, and also check whether any of its members have changed. * Dentry lookups (based on parent, candidate string tuple) recheck the parent sequence after the child is found in case anything changed in the parent during the path walk. * inode is also RCU protected so we can load d_inode and use the inode for limited things. * i_mode, i_uid, i_gid can be tested for exec permissions during path walk. * i_op can be loaded. When we reach the destination dentry, we lock it, recheck lookup sequence, and increment its refcount and mountpoint refcount. RCU and vfsmount locks are dropped. This is termed "dropping rcu-walk". If the dentry refcount does not match, we can not drop rcu-walk gracefully at the current point in the lokup, so instead return -ECHILD (for want of a better errno). This signals the path walking code to re-do the entire lookup with a ref-walk. Aside from the final dentry, there are other situations that may be encounted where we cannot continue rcu-walk. In that case, we drop rcu-walk (ie. take a reference on the last good dentry) and continue with a ref-walk. Again, if we can drop rcu-walk gracefully, we return -ECHILD and do the whole lookup using ref-walk. But it is very important that we can continue with ref-walk for most cases, particularly to avoid the overhead of double lookups, and to gain the scalability advantages on common path elements (like cwd and root). The cases where rcu-walk cannot continue are: * NULL dentry (ie. any uncached path element) * parent with d_inode->i_op->permission or ACLs * dentries with d_revalidate * Following links In future patches, permission checks and d_revalidate become rcu-walk aware. It may be possible eventually to make following links rcu-walk aware. Uncached path elements will always require dropping to ref-walk mode, at the very least because i_mutex needs to be grabbed, and objects allocated. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 14:49:52 +08:00
if (unlikely(file->f_flags & __O_TMPFILE)) {
error = do_tmpfile(nd, flags, op, file);
} else if (unlikely(file->f_flags & O_PATH)) {
error = do_o_path(nd, flags, file);
} else {
const char *s = path_init(nd, flags);
while (!(error = link_path_walk(s, nd)) &&
(s = open_last_lookups(nd, file, op)) != NULL)
;
if (!error)
error = do_open(nd, file, op);
terminate_walk(nd);
}
if (likely(!error)) {
if (likely(file->f_mode & FMODE_OPENED))
return file;
WARN_ON(1);
error = -EINVAL;
}
fput(file);
if (error == -EOPENSTALE) {
if (flags & LOOKUP_RCU)
error = -ECHILD;
else
error = -ESTALE;
}
return ERR_PTR(error);
}
struct file *do_filp_open(int dfd, struct filename *pathname,
const struct open_flags *op)
{
struct nameidata nd;
int flags = op->lookup_flags;
struct file *filp;
set_nameidata(&nd, dfd, pathname, NULL);
filp = path_openat(&nd, op, flags | LOOKUP_RCU);
if (unlikely(filp == ERR_PTR(-ECHILD)))
filp = path_openat(&nd, op, flags);
if (unlikely(filp == ERR_PTR(-ESTALE)))
filp = path_openat(&nd, op, flags | LOOKUP_REVAL);
restore_nameidata();
return filp;
}
struct file *do_file_open_root(const struct path *root,
const char *name, const struct open_flags *op)
{
struct nameidata nd;
struct file *file;
struct filename *filename;
int flags = op->lookup_flags;
if (d_is_symlink(root->dentry) && op->intent & LOOKUP_OPEN)
return ERR_PTR(-ELOOP);
filename = getname_kernel(name);
if (IS_ERR(filename))
return ERR_CAST(filename);
set_nameidata(&nd, -1, filename, root);
file = path_openat(&nd, op, flags | LOOKUP_RCU);
if (unlikely(file == ERR_PTR(-ECHILD)))
file = path_openat(&nd, op, flags);
if (unlikely(file == ERR_PTR(-ESTALE)))
file = path_openat(&nd, op, flags | LOOKUP_REVAL);
restore_nameidata();
putname(filename);
return file;
}
static struct dentry *filename_create(int dfd, struct filename *name,
struct path *path, unsigned int lookup_flags)
{
struct dentry *dentry = ERR_PTR(-EEXIST);
struct qstr last;
VFS: filename_create(): fix incorrect intent. When asked to create a path ending '/', but which is not to be a directory (LOOKUP_DIRECTORY not set), filename_create() will never try to create the file. If it doesn't exist, -ENOENT is reported. However, it still passes LOOKUP_CREATE|LOOKUP_EXCL to the filesystems ->lookup() function, even though there is no intent to create. This is misleading and can cause incorrect behaviour. If you try ln -s foo /path/dir/ where 'dir' is a directory on an NFS filesystem which is not currently known in the dcache, this will fail with ENOENT. But as the name is not in the dcache, nfs_lookup gets called with LOOKUP_CREATE|LOOKUP_EXCL and so it returns NULL without performing any lookup, with the expectation that a subsequent call to create the target will be made, and the lookup can be combined with the creation. In the case with a trailing '/' and no LOOKUP_DIRECTORY, that call is never made. Instead filename_create() sees that the dentry is not (yet) positive and returns -ENOENT - even though the directory actually exists. So only set LOOKUP_CREATE|LOOKUP_EXCL if there really is an intent to create, and use the absence of these flags to decide if -ENOENT should be returned. Note that filename_parentat() is only interested in LOOKUP_REVAL, so we split that out and store it in 'reval_flag'. __lookup_hash() then gets reval_flag combined with whatever create flags were determined to be needed. Reviewed-by: David Disseldorp <ddiss@suse.de> Reviewed-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: NeilBrown <neilb@suse.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-04-14 11:57:35 +08:00
bool want_dir = lookup_flags & LOOKUP_DIRECTORY;
unsigned int reval_flag = lookup_flags & LOOKUP_REVAL;
unsigned int create_flags = LOOKUP_CREATE | LOOKUP_EXCL;
int type;
int err2;
int error;
VFS: filename_create(): fix incorrect intent. When asked to create a path ending '/', but which is not to be a directory (LOOKUP_DIRECTORY not set), filename_create() will never try to create the file. If it doesn't exist, -ENOENT is reported. However, it still passes LOOKUP_CREATE|LOOKUP_EXCL to the filesystems ->lookup() function, even though there is no intent to create. This is misleading and can cause incorrect behaviour. If you try ln -s foo /path/dir/ where 'dir' is a directory on an NFS filesystem which is not currently known in the dcache, this will fail with ENOENT. But as the name is not in the dcache, nfs_lookup gets called with LOOKUP_CREATE|LOOKUP_EXCL and so it returns NULL without performing any lookup, with the expectation that a subsequent call to create the target will be made, and the lookup can be combined with the creation. In the case with a trailing '/' and no LOOKUP_DIRECTORY, that call is never made. Instead filename_create() sees that the dentry is not (yet) positive and returns -ENOENT - even though the directory actually exists. So only set LOOKUP_CREATE|LOOKUP_EXCL if there really is an intent to create, and use the absence of these flags to decide if -ENOENT should be returned. Note that filename_parentat() is only interested in LOOKUP_REVAL, so we split that out and store it in 'reval_flag'. __lookup_hash() then gets reval_flag combined with whatever create flags were determined to be needed. Reviewed-by: David Disseldorp <ddiss@suse.de> Reviewed-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: NeilBrown <neilb@suse.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-04-14 11:57:35 +08:00
error = filename_parentat(dfd, name, reval_flag, path, &last, &type);
if (error)
return ERR_PTR(error);
/*
* Yucky last component or no last component at all?
* (foo/., foo/.., /////)
*/
if (unlikely(type != LAST_NORM))
goto out;
/* don't fail immediately if it's r/o, at least try to report other errors */
err2 = mnt_want_write(path->mnt);
/*
VFS: filename_create(): fix incorrect intent. When asked to create a path ending '/', but which is not to be a directory (LOOKUP_DIRECTORY not set), filename_create() will never try to create the file. If it doesn't exist, -ENOENT is reported. However, it still passes LOOKUP_CREATE|LOOKUP_EXCL to the filesystems ->lookup() function, even though there is no intent to create. This is misleading and can cause incorrect behaviour. If you try ln -s foo /path/dir/ where 'dir' is a directory on an NFS filesystem which is not currently known in the dcache, this will fail with ENOENT. But as the name is not in the dcache, nfs_lookup gets called with LOOKUP_CREATE|LOOKUP_EXCL and so it returns NULL without performing any lookup, with the expectation that a subsequent call to create the target will be made, and the lookup can be combined with the creation. In the case with a trailing '/' and no LOOKUP_DIRECTORY, that call is never made. Instead filename_create() sees that the dentry is not (yet) positive and returns -ENOENT - even though the directory actually exists. So only set LOOKUP_CREATE|LOOKUP_EXCL if there really is an intent to create, and use the absence of these flags to decide if -ENOENT should be returned. Note that filename_parentat() is only interested in LOOKUP_REVAL, so we split that out and store it in 'reval_flag'. __lookup_hash() then gets reval_flag combined with whatever create flags were determined to be needed. Reviewed-by: David Disseldorp <ddiss@suse.de> Reviewed-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: NeilBrown <neilb@suse.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-04-14 11:57:35 +08:00
* Do the final lookup. Suppress 'create' if there is a trailing
* '/', and a directory wasn't requested.
*/
VFS: filename_create(): fix incorrect intent. When asked to create a path ending '/', but which is not to be a directory (LOOKUP_DIRECTORY not set), filename_create() will never try to create the file. If it doesn't exist, -ENOENT is reported. However, it still passes LOOKUP_CREATE|LOOKUP_EXCL to the filesystems ->lookup() function, even though there is no intent to create. This is misleading and can cause incorrect behaviour. If you try ln -s foo /path/dir/ where 'dir' is a directory on an NFS filesystem which is not currently known in the dcache, this will fail with ENOENT. But as the name is not in the dcache, nfs_lookup gets called with LOOKUP_CREATE|LOOKUP_EXCL and so it returns NULL without performing any lookup, with the expectation that a subsequent call to create the target will be made, and the lookup can be combined with the creation. In the case with a trailing '/' and no LOOKUP_DIRECTORY, that call is never made. Instead filename_create() sees that the dentry is not (yet) positive and returns -ENOENT - even though the directory actually exists. So only set LOOKUP_CREATE|LOOKUP_EXCL if there really is an intent to create, and use the absence of these flags to decide if -ENOENT should be returned. Note that filename_parentat() is only interested in LOOKUP_REVAL, so we split that out and store it in 'reval_flag'. __lookup_hash() then gets reval_flag combined with whatever create flags were determined to be needed. Reviewed-by: David Disseldorp <ddiss@suse.de> Reviewed-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: NeilBrown <neilb@suse.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-04-14 11:57:35 +08:00
if (last.name[last.len] && !want_dir)
create_flags = 0;
inode_lock_nested(path->dentry->d_inode, I_MUTEX_PARENT);
dentry = lookup_one_qstr_excl(&last, path->dentry,
reval_flag | create_flags);
if (IS_ERR(dentry))
goto unlock;
error = -EEXIST;
if (d_is_positive(dentry))
goto fail;
/*
* Special case - lookup gave negative, but... we had foo/bar/
* From the vfs_mknod() POV we just have a negative dentry -
* all is fine. Let's be bastards - you had / on the end, you've
* been asking for (non-existent) directory. -ENOENT for you.
*/
VFS: filename_create(): fix incorrect intent. When asked to create a path ending '/', but which is not to be a directory (LOOKUP_DIRECTORY not set), filename_create() will never try to create the file. If it doesn't exist, -ENOENT is reported. However, it still passes LOOKUP_CREATE|LOOKUP_EXCL to the filesystems ->lookup() function, even though there is no intent to create. This is misleading and can cause incorrect behaviour. If you try ln -s foo /path/dir/ where 'dir' is a directory on an NFS filesystem which is not currently known in the dcache, this will fail with ENOENT. But as the name is not in the dcache, nfs_lookup gets called with LOOKUP_CREATE|LOOKUP_EXCL and so it returns NULL without performing any lookup, with the expectation that a subsequent call to create the target will be made, and the lookup can be combined with the creation. In the case with a trailing '/' and no LOOKUP_DIRECTORY, that call is never made. Instead filename_create() sees that the dentry is not (yet) positive and returns -ENOENT - even though the directory actually exists. So only set LOOKUP_CREATE|LOOKUP_EXCL if there really is an intent to create, and use the absence of these flags to decide if -ENOENT should be returned. Note that filename_parentat() is only interested in LOOKUP_REVAL, so we split that out and store it in 'reval_flag'. __lookup_hash() then gets reval_flag combined with whatever create flags were determined to be needed. Reviewed-by: David Disseldorp <ddiss@suse.de> Reviewed-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: NeilBrown <neilb@suse.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-04-14 11:57:35 +08:00
if (unlikely(!create_flags)) {
error = -ENOENT;
goto fail;
}
if (unlikely(err2)) {
error = err2;
goto fail;
}
return dentry;
fail:
dput(dentry);
dentry = ERR_PTR(error);
unlock:
inode_unlock(path->dentry->d_inode);
if (!err2)
mnt_drop_write(path->mnt);
out:
path_put(path);
return dentry;
}
struct dentry *kern_path_create(int dfd, const char *pathname,
struct path *path, unsigned int lookup_flags)
{
struct filename *filename = getname_kernel(pathname);
struct dentry *res = filename_create(dfd, filename, path, lookup_flags);
putname(filename);
return res;
}
EXPORT_SYMBOL(kern_path_create);
void done_path_create(struct path *path, struct dentry *dentry)
{
dput(dentry);
inode_unlock(path->dentry->d_inode);
mnt_drop_write(path->mnt);
path_put(path);
}
EXPORT_SYMBOL(done_path_create);
inline struct dentry *user_path_create(int dfd, const char __user *pathname,
struct path *path, unsigned int lookup_flags)
{
struct filename *filename = getname(pathname);
struct dentry *res = filename_create(dfd, filename, path, lookup_flags);
putname(filename);
return res;
}
EXPORT_SYMBOL(user_path_create);
/**
* vfs_mknod - create device node or file
* @idmap: idmap of the mount the inode was found from
* @dir: inode of the parent directory
* @dentry: dentry of the child device node
* @mode: mode of the child device node
* @dev: device number of device to create
*
* Create a device node or file.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_mknod(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *dentry, umode_t mode, dev_t dev)
{
bool is_whiteout = S_ISCHR(mode) && dev == WHITEOUT_DEV;
int error = may_create(idmap, dir, dentry);
if (error)
return error;
if ((S_ISCHR(mode) || S_ISBLK(mode)) && !is_whiteout &&
!capable(CAP_MKNOD))
return -EPERM;
if (!dir->i_op->mknod)
return -EPERM;
mode = vfs_prepare_mode(idmap, dir, mode, mode, mode);
cgroups: implement device whitelist Implement a cgroup to track and enforce open and mknod restrictions on device files. A device cgroup associates a device access whitelist with each cgroup. A whitelist entry has 4 fields. 'type' is a (all), c (char), or b (block). 'all' means it applies to all types and all major and minor numbers. Major and minor are either an integer or * for all. Access is a composition of r (read), w (write), and m (mknod). The root device cgroup starts with rwm to 'all'. A child devcg gets a copy of the parent. Admins can then remove devices from the whitelist or add new entries. A child cgroup can never receive a device access which is denied its parent. However when a device access is removed from a parent it will not also be removed from the child(ren). An entry is added using devices.allow, and removed using devices.deny. For instance echo 'c 1:3 mr' > /cgroups/1/devices.allow allows cgroup 1 to read and mknod the device usually known as /dev/null. Doing echo a > /cgroups/1/devices.deny will remove the default 'a *:* mrw' entry. CAP_SYS_ADMIN is needed to change permissions or move another task to a new cgroup. A cgroup may not be granted more permissions than the cgroup's parent has. Any task can move itself between cgroups. This won't be sufficient, but we can decide the best way to adequately restrict movement later. [akpm@linux-foundation.org: coding-style fixes] [akpm@linux-foundation.org: fix may-be-used-uninitialized warning] Signed-off-by: Serge E. Hallyn <serue@us.ibm.com> Acked-by: James Morris <jmorris@namei.org> Looks-good-to: Pavel Emelyanov <xemul@openvz.org> Cc: Daniel Hokka Zakrisson <daniel@hozac.com> Cc: Li Zefan <lizf@cn.fujitsu.com> Cc: Paul Menage <menage@google.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-04-29 16:00:10 +08:00
error = devcgroup_inode_mknod(mode, dev);
if (error)
return error;
error = security_inode_mknod(dir, dentry, mode, dev);
if (error)
return error;
error = dir->i_op->mknod(idmap, dir, dentry, mode, dev);
if (!error)
fsnotify_create(dir, dentry);
return error;
}
EXPORT_SYMBOL(vfs_mknod);
static int may_mknod(umode_t mode)
{
switch (mode & S_IFMT) {
case S_IFREG:
case S_IFCHR:
case S_IFBLK:
case S_IFIFO:
case S_IFSOCK:
case 0: /* zero mode translates to S_IFREG */
return 0;
case S_IFDIR:
return -EPERM;
default:
return -EINVAL;
}
}
static int do_mknodat(int dfd, struct filename *name, umode_t mode,
unsigned int dev)
{
struct mnt_idmap *idmap;
struct dentry *dentry;
struct path path;
int error;
unsigned int lookup_flags = 0;
error = may_mknod(mode);
if (error)
goto out1;
retry:
dentry = filename_create(dfd, name, &path, lookup_flags);
error = PTR_ERR(dentry);
if (IS_ERR(dentry))
goto out1;
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
error = security_path_mknod(&path, dentry,
mode_strip_umask(path.dentry->d_inode, mode), dev);
if (error)
goto out2;
idmap = mnt_idmap(path.mnt);
switch (mode & S_IFMT) {
case 0: case S_IFREG:
error = vfs_create(idmap, path.dentry->d_inode,
dentry, mode, true);
security: Place security_path_post_mknod() where the original IMA call was Commit 08abce60d63f ("security: Introduce path_post_mknod hook") introduced security_path_post_mknod(), to replace the IMA-specific call to ima_post_path_mknod(). For symmetry with security_path_mknod(), security_path_post_mknod() was called after a successful mknod operation, for any file type, rather than only for regular files at the time there was the IMA call. However, as reported by VFS maintainers, successful mknod operation does not mean that the dentry always has an inode attached to it (for example, not for FIFOs on a SAMBA mount). If that condition happens, the kernel crashes when security_path_post_mknod() attempts to verify if the inode associated to the dentry is private. Move security_path_post_mknod() where the ima_post_path_mknod() call was, which is obviously correct from IMA/EVM perspective. IMA/EVM are the only in-kernel users, and only need to inspect regular files. Reported-by: Steve French <smfrench@gmail.com> Closes: https://lore.kernel.org/linux-kernel/CAH2r5msAVzxCUHHG8VKrMPUKQHmBpE6K9_vjhgDa1uAvwx4ppw@mail.gmail.com/ Suggested-by: Al Viro <viro@zeniv.linux.org.uk> Fixes: 08abce60d63f ("security: Introduce path_post_mknod hook") Signed-off-by: Roberto Sassu <roberto.sassu@huawei.com> Reviewed-by: Christian Brauner <brauner@kernel.org> Reviewed-by: Mimi Zohar <zohar@linux.ibm.com> Acked-by: Paul Moore <paul@paul-moore.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2024-04-03 15:57:29 +08:00
if (!error)
security_path_post_mknod(idmap, dentry);
break;
case S_IFCHR: case S_IFBLK:
error = vfs_mknod(idmap, path.dentry->d_inode,
dentry, mode, new_decode_dev(dev));
break;
case S_IFIFO: case S_IFSOCK:
error = vfs_mknod(idmap, path.dentry->d_inode,
dentry, mode, 0);
break;
}
out2:
done_path_create(&path, dentry);
if (retry_estale(error, lookup_flags)) {
lookup_flags |= LOOKUP_REVAL;
goto retry;
}
out1:
putname(name);
return error;
}
SYSCALL_DEFINE4(mknodat, int, dfd, const char __user *, filename, umode_t, mode,
unsigned int, dev)
{
return do_mknodat(dfd, getname(filename), mode, dev);
}
SYSCALL_DEFINE3(mknod, const char __user *, filename, umode_t, mode, unsigned, dev)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_mknodat(AT_FDCWD, getname(filename), mode, dev);
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
/**
* vfs_mkdir - create directory
* @idmap: idmap of the mount the inode was found from
* @dir: inode of the parent directory
* @dentry: dentry of the child directory
* @mode: mode of the child directory
*
* Create a directory.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_mkdir(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *dentry, umode_t mode)
{
int error;
unsigned max_links = dir->i_sb->s_max_links;
error = may_create(idmap, dir, dentry);
if (error)
return error;
if (!dir->i_op->mkdir)
return -EPERM;
mode = vfs_prepare_mode(idmap, dir, mode, S_IRWXUGO | S_ISVTX, 0);
error = security_inode_mkdir(dir, dentry, mode);
if (error)
return error;
if (max_links && dir->i_nlink >= max_links)
return -EMLINK;
error = dir->i_op->mkdir(idmap, dir, dentry, mode);
if (!error)
fsnotify_mkdir(dir, dentry);
return error;
}
EXPORT_SYMBOL(vfs_mkdir);
int do_mkdirat(int dfd, struct filename *name, umode_t mode)
{
struct dentry *dentry;
struct path path;
int error;
unsigned int lookup_flags = LOOKUP_DIRECTORY;
retry:
dentry = filename_create(dfd, name, &path, lookup_flags);
error = PTR_ERR(dentry);
if (IS_ERR(dentry))
goto out_putname;
fs: move S_ISGID stripping into the vfs_*() helpers Move setgid handling out of individual filesystems and into the VFS itself to stop the proliferation of setgid inheritance bugs. Creating files that have both the S_IXGRP and S_ISGID bit raised in directories that themselves have the S_ISGID bit set requires additional privileges to avoid security issues. When a filesystem creates a new inode it needs to take care that the caller is either in the group of the newly created inode or they have CAP_FSETID in their current user namespace and are privileged over the parent directory of the new inode. If any of these two conditions is true then the S_ISGID bit can be raised for an S_IXGRP file and if not it needs to be stripped. However, there are several key issues with the current implementation: * S_ISGID stripping logic is entangled with umask stripping. If a filesystem doesn't support or enable POSIX ACLs then umask stripping is done directly in the vfs before calling into the filesystem. If the filesystem does support POSIX ACLs then unmask stripping may be done in the filesystem itself when calling posix_acl_create(). Since umask stripping has an effect on S_ISGID inheritance, e.g., by stripping the S_IXGRP bit from the file to be created and all relevant filesystems have to call posix_acl_create() before inode_init_owner() where we currently take care of S_ISGID handling S_ISGID handling is order dependent. IOW, whether or not you get a setgid bit depends on POSIX ACLs and umask and in what order they are called. Note that technically filesystems are free to impose their own ordering between posix_acl_create() and inode_init_owner() meaning that there's additional ordering issues that influence S_SIGID inheritance. * Filesystems that don't rely on inode_init_owner() don't get S_ISGID stripping logic. While that may be intentional (e.g. network filesystems might just defer setgid stripping to a server) it is often just a security issue. This is not just ugly it's unsustainably messy especially since we do still have bugs in this area years after the initial round of setgid bugfixes. So the current state is quite messy and while we won't be able to make it completely clean as posix_acl_create() is still a filesystem specific call we can improve the S_SIGD stripping situation quite a bit by hoisting it out of inode_init_owner() and into the vfs creation operations. This means we alleviate the burden for filesystems to handle S_ISGID stripping correctly and can standardize the ordering between S_ISGID and umask stripping in the vfs. We add a new helper vfs_prepare_mode() so S_ISGID handling is now done in the VFS before umask handling. This has S_ISGID handling is unaffected unaffected by whether umask stripping is done by the VFS itself (if no POSIX ACLs are supported or enabled) or in the filesystem in posix_acl_create() (if POSIX ACLs are supported). The vfs_prepare_mode() helper is called directly in vfs_*() helpers that create new filesystem objects. We need to move them into there to make sure that filesystems like overlayfs hat have callchains like: sys_mknod() -> do_mknodat(mode) -> .mknod = ovl_mknod(mode) -> ovl_create(mode) -> vfs_mknod(mode) get S_ISGID stripping done when calling into lower filesystems via vfs_*() creation helpers. Moving vfs_prepare_mode() into e.g. vfs_mknod() takes care of that. This is in any case semantically cleaner because S_ISGID stripping is VFS security requirement. Security hooks so far have seen the mode with the umask applied but without S_ISGID handling done. The relevant hooks are called outside of vfs_*() creation helpers so by calling vfs_prepare_mode() from vfs_*() helpers the security hooks would now see the mode without umask stripping applied. For now we fix this by passing the mode with umask settings applied to not risk any regressions for LSM hooks. IOW, nothing changes for LSM hooks. It is worth pointing out that security hooks never saw the mode that is seen by the filesystem when actually creating the file. They have always been completely misplaced for that to work. The following filesystems use inode_init_owner() and thus relied on S_ISGID stripping: spufs, 9p, bfs, btrfs, ext2, ext4, f2fs, hfsplus, hugetlbfs, jfs, minix, nilfs2, ntfs3, ocfs2, omfs, overlayfs, ramfs, reiserfs, sysv, ubifs, udf, ufs, xfs, zonefs, bpf, tmpfs. All of the above filesystems end up calling inode_init_owner() when new filesystem objects are created through the ->mkdir(), ->mknod(), ->create(), ->tmpfile(), ->rename() inode operations. Since directories always inherit the S_ISGID bit with the exception of xfs when irix_sgid_inherit mode is turned on S_ISGID stripping doesn't apply. The ->symlink() and ->link() inode operations trivially inherit the mode from the target and the ->rename() inode operation inherits the mode from the source inode. All other creation inode operations will get S_ISGID handling via vfs_prepare_mode() when called from their relevant vfs_*() helpers. In addition to this there are filesystems which allow the creation of filesystem objects through ioctl()s or - in the case of spufs - circumventing the vfs in other ways. If filesystem objects are created through ioctl()s the vfs doesn't know about it and can't apply regular permission checking including S_ISGID logic. Therfore, a filesystem relying on S_ISGID stripping in inode_init_owner() in their ioctl() callpath will be affected by moving this logic into the vfs. We audited those filesystems: * btrfs allows the creation of filesystem objects through various ioctls(). Snapshot creation literally takes a snapshot and so the mode is fully preserved and S_ISGID stripping doesn't apply. Creating a new subvolum relies on inode_init_owner() in btrfs_new_subvol_inode() but only creates directories and doesn't raise S_ISGID. * ocfs2 has a peculiar implementation of reflinks. In contrast to e.g. xfs and btrfs FICLONE/FICLONERANGE ioctl() that is only concerned with the actual extents ocfs2 uses a separate ioctl() that also creates the target file. Iow, ocfs2 circumvents the vfs entirely here and did indeed rely on inode_init_owner() to strip the S_ISGID bit. This is the only place where a filesystem needs to call mode_strip_sgid() directly but this is self-inflicted pain. * spufs doesn't go through the vfs at all and doesn't use ioctl()s either. Instead it has a dedicated system call spufs_create() which allows the creation of filesystem objects. But spufs only creates directories and doesn't allo S_SIGID bits, i.e. it specifically only allows 0777 bits. * bpf uses vfs_mkobj() but also doesn't allow S_ISGID bits to be created. The patch will have an effect on ext2 when the EXT2_MOUNT_GRPID mount option is used, on ext4 when the EXT4_MOUNT_GRPID mount option is used, and on xfs when the XFS_FEAT_GRPID mount option is used. When any of these filesystems are mounted with their respective GRPID option then newly created files inherit the parent directories group unconditionally. In these cases non of the filesystems call inode_init_owner() and thus did never strip the S_ISGID bit for newly created files. Moving this logic into the VFS means that they now get the S_ISGID bit stripped. This is a user visible change. If this leads to regressions we will either need to figure out a better way or we need to revert. However, given the various setgid bugs that we found just in the last two years this is a regression risk we should take. Associated with this change is a new set of fstests to enforce the semantics for all new filesystems. Link: https://lore.kernel.org/ceph-devel/20220427092201.wvsdjbnc7b4dttaw@wittgenstein [1] Link: e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes") [2] Link: 01ea173e103e ("xfs: fix up non-directory creation in SGID directories") [3] Link: fd84bfdddd16 ("ceph: fix up non-directory creation in SGID directories") [4] Link: https://lore.kernel.org/r/1657779088-2242-3-git-send-email-xuyang2018.jy@fujitsu.com Suggested-by: Dave Chinner <david@fromorbit.com> Suggested-by: Christian Brauner (Microsoft) <brauner@kernel.org> Reviewed-by: Darrick J. Wong <djwong@kernel.org> Reviewed-and-Tested-by: Jeff Layton <jlayton@kernel.org> Signed-off-by: Yang Xu <xuyang2018.jy@fujitsu.com> [<brauner@kernel.org>: rewrote commit message] Signed-off-by: Christian Brauner (Microsoft) <brauner@kernel.org>
2022-07-14 14:11:27 +08:00
error = security_path_mkdir(&path, dentry,
mode_strip_umask(path.dentry->d_inode, mode));
if (!error) {
error = vfs_mkdir(mnt_idmap(path.mnt), path.dentry->d_inode,
dentry, mode);
}
done_path_create(&path, dentry);
if (retry_estale(error, lookup_flags)) {
lookup_flags |= LOOKUP_REVAL;
goto retry;
}
out_putname:
putname(name);
return error;
}
SYSCALL_DEFINE3(mkdirat, int, dfd, const char __user *, pathname, umode_t, mode)
{
return do_mkdirat(dfd, getname(pathname), mode);
}
SYSCALL_DEFINE2(mkdir, const char __user *, pathname, umode_t, mode)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_mkdirat(AT_FDCWD, getname(pathname), mode);
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
/**
* vfs_rmdir - remove directory
* @idmap: idmap of the mount the inode was found from
* @dir: inode of the parent directory
* @dentry: dentry of the child directory
*
* Remove a directory.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_rmdir(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *dentry)
{
int error = may_delete(idmap, dir, dentry, 1);
if (error)
return error;
if (!dir->i_op->rmdir)
return -EPERM;
dget(dentry);
inode_lock(dentry->d_inode);
error = -EBUSY;
if (is_local_mountpoint(dentry) ||
(dentry->d_inode->i_flags & S_KERNEL_FILE))
goto out;
error = security_inode_rmdir(dir, dentry);
if (error)
goto out;
error = dir->i_op->rmdir(dir, dentry);
if (error)
goto out;
rmdir(),rename(): do shrink_dcache_parent() only on success Once upon a time ->rmdir() instances used to check if victim inode had more than one (in-core) reference and failed with -EBUSY if it had. The reason was race avoidance - emptiness check is worthless if somebody could just go and create new objects in the victim directory afterwards. With introduction of dcache the checks had been replaced with checking the refcount of dentry. However, since a cached negative lookup leaves a negative child dentry, such check had lead to false positives - with empty foo/ doing stat foo/bar before rmdir foo ended up with -EBUSY unless the negative dentry of foo/bar happened to be evicted by the time of rmdir(2). That had been fixed by doing shrink_dcache_parent() just before the refcount check. At the same time, ext2_rmdir() has grown a private solution that eliminated those -EBUSY - it did something (setting ->i_size to 0) which made any subsequent ext2_add_entry() fail. Unfortunately, even with shrink_dcache_parent() the check had been racy - after all, the victim itself could be found by dcache lookup just after we'd checked its refcount. That got fixed by a new helper (dentry_unhash()) that did shrink_dcache_parent() and unhashed the sucker if its refcount ended up equal to 1. That got called before ->rmdir(), turning the checks in ->rmdir() instances into "if not unhashed fail with -EBUSY". Which reduced the boilerplate nicely, but had an unpleasant side effect - now shrink_dcache_parent() had been done before the emptiness checks, leading to easily triggerable calls of shrink_dcache_parent() on arbitrary large subtrees, quite possibly nested into each other. Several years later the ext2-private trick had been generalized - (in-core) inodes of dead directories are flagged and calls of lookup, readdir and all directory-modifying methods were prevented in so marked directories. Remaining boilerplate in ->rmdir() instances became redundant and some instances got rid of it. In 2011 the call of dentry_unhash() got shifted into ->rmdir() instances and then killed off in all of them. That has lead to another problem, though - in case of successful rmdir we *want* any (negative) child dentries dropped and the victim itself made negative. There's no point keeping cached negative lookups in foo when we can get the negative lookup of foo itself cached. So shrink_dcache_parent() call had been restored; unfortunately, it went into the place where dentry_unhash() used to be, i.e. before the ->rmdir() call. Note that we don't unhash anymore, so any "is it busy" checks would be racy; fortunately, all of them are gone. We should've done that call right *after* successful ->rmdir(). That reduces contention caused by tree-walking in shrink_dcache_parent() and, especially, contention caused by evictions in two nested subtrees going on in parallel. The same goes for directory-overwriting rename() - the story there had been parallel to that of rmdir(). Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-05-28 04:23:51 +08:00
shrink_dcache_parent(dentry);
dentry->d_inode->i_flags |= S_DEAD;
dont_mount(dentry);
vfs: Lazily remove mounts on unlinked files and directories. With the introduction of mount namespaces and bind mounts it became possible to access files and directories that on some paths are mount points but are not mount points on other paths. It is very confusing when rm -rf somedir returns -EBUSY simply because somedir is mounted somewhere else. With the addition of user namespaces allowing unprivileged mounts this condition has gone from annoying to allowing a DOS attack on other users in the system. The possibility for mischief is removed by updating the vfs to support rename, unlink and rmdir on a dentry that is a mountpoint and by lazily unmounting mountpoints on deleted dentries. In particular this change allows rename, unlink and rmdir system calls on a dentry without a mountpoint in the current mount namespace to succeed, and it allows rename, unlink, and rmdir performed on a distributed filesystem to update the vfs cache even if when there is a mount in some namespace on the original dentry. There are two common patterns of maintaining mounts: Mounts on trusted paths with the parent directory of the mount point and all ancestory directories up to / owned by root and modifiable only by root (i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu, cpuacct, ...}, /usr, /usr/local). Mounts on unprivileged directories maintained by fusermount. In the case of mounts in trusted directories owned by root and modifiable only by root the current parent directory permissions are sufficient to ensure a mount point on a trusted path is not removed or renamed by anyone other than root, even if there is a context where the there are no mount points to prevent this. In the case of mounts in directories owned by less privileged users races with users modifying the path of a mount point are already a danger. fusermount already uses a combination of chdir, /proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races. The removable of global rename, unlink, and rmdir protection really adds nothing new to consider only a widening of the attack window, and fusermount is already safe against unprivileged users modifying the directory simultaneously. In principle for perfect userspace programs returning -EBUSY for unlink, rmdir, and rename of dentires that have mounts in the local namespace is actually unnecessary. Unfortunately not all userspace programs are perfect so retaining -EBUSY for unlink, rmdir and rename of dentries that have mounts in the current mount namespace plays an important role of maintaining consistency with historical behavior and making imperfect userspace applications hard to exploit. v2: Remove spurious old_dentry. v3: Optimized shrink_submounts_and_drop Removed unsued afs label v4: Simplified the changes to check_submounts_and_drop Do not rename check_submounts_and_drop shrink_submounts_and_drop Document what why we need atomicity in check_submounts_and_drop Rely on the parent inode mutex to make d_revalidate and d_invalidate an atomic unit. v5: Refcount the mountpoint to detach in case of simultaneous renames. Reviewed-by: Miklos Szeredi <miklos@szeredi.hu> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-10-02 09:33:48 +08:00
detach_mounts(dentry);
out:
inode_unlock(dentry->d_inode);
dput(dentry);
if (!error)
fsnotify: invalidate dcache before IN_DELETE event Apparently, there are some applications that use IN_DELETE event as an invalidation mechanism and expect that if they try to open a file with the name reported with the delete event, that it should not contain the content of the deleted file. Commit 49246466a989 ("fsnotify: move fsnotify_nameremove() hook out of d_delete()") moved the fsnotify delete hook before d_delete() so fsnotify will have access to a positive dentry. This allowed a race where opening the deleted file via cached dentry is now possible after receiving the IN_DELETE event. To fix the regression, create a new hook fsnotify_delete() that takes the unlinked inode as an argument and use a helper d_delete_notify() to pin the inode, so we can pass it to fsnotify_delete() after d_delete(). Backporting hint: this regression is from v5.3. Although patch will apply with only trivial conflicts to v5.4 and v5.10, it won't build, because fsnotify_delete() implementation is different in each of those versions (see fsnotify_link()). A follow up patch will fix the fsnotify_unlink/rmdir() calls in pseudo filesystem that do not need to call d_delete(). Link: https://lore.kernel.org/r/20220120215305.282577-1-amir73il@gmail.com Reported-by: Ivan Delalande <colona@arista.com> Link: https://lore.kernel.org/linux-fsdevel/YeNyzoDM5hP5LtGW@visor/ Fixes: 49246466a989 ("fsnotify: move fsnotify_nameremove() hook out of d_delete()") Cc: stable@vger.kernel.org # v5.3+ Signed-off-by: Amir Goldstein <amir73il@gmail.com> Signed-off-by: Jan Kara <jack@suse.cz>
2022-01-21 05:53:04 +08:00
d_delete_notify(dir, dentry);
return error;
}
EXPORT_SYMBOL(vfs_rmdir);
int do_rmdir(int dfd, struct filename *name)
{
int error;
struct dentry *dentry;
struct path path;
struct qstr last;
int type;
unsigned int lookup_flags = 0;
retry:
error = filename_parentat(dfd, name, lookup_flags, &path, &last, &type);
if (error)
goto exit1;
switch (type) {
case LAST_DOTDOT:
error = -ENOTEMPTY;
goto exit2;
case LAST_DOT:
error = -EINVAL;
goto exit2;
case LAST_ROOT:
error = -EBUSY;
goto exit2;
}
error = mnt_want_write(path.mnt);
if (error)
goto exit2;
inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
dentry = lookup_one_qstr_excl(&last, path.dentry, lookup_flags);
error = PTR_ERR(dentry);
if (IS_ERR(dentry))
goto exit3;
if (!dentry->d_inode) {
error = -ENOENT;
goto exit4;
}
error = security_path_rmdir(&path, dentry);
if (error)
goto exit4;
error = vfs_rmdir(mnt_idmap(path.mnt), path.dentry->d_inode, dentry);
exit4:
dput(dentry);
exit3:
inode_unlock(path.dentry->d_inode);
mnt_drop_write(path.mnt);
exit2:
path_put(&path);
if (retry_estale(error, lookup_flags)) {
lookup_flags |= LOOKUP_REVAL;
goto retry;
}
exit1:
putname(name);
return error;
}
SYSCALL_DEFINE1(rmdir, const char __user *, pathname)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_rmdir(AT_FDCWD, getname(pathname));
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
/**
* vfs_unlink - unlink a filesystem object
* @idmap: idmap of the mount the inode was found from
* @dir: parent directory
* @dentry: victim
* @delegated_inode: returns victim inode, if the inode is delegated.
*
* The caller must hold dir->i_mutex.
*
* If vfs_unlink discovers a delegation, it will return -EWOULDBLOCK and
* return a reference to the inode in delegated_inode. The caller
* should then break the delegation on that inode and retry. Because
* breaking a delegation may take a long time, the caller should drop
* dir->i_mutex before doing so.
*
* Alternatively, a caller may pass NULL for delegated_inode. This may
* be appropriate for callers that expect the underlying filesystem not
* to be NFS exported.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_unlink(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *dentry, struct inode **delegated_inode)
{
struct inode *target = dentry->d_inode;
int error = may_delete(idmap, dir, dentry, 0);
if (error)
return error;
if (!dir->i_op->unlink)
return -EPERM;
inode_lock(target);
if (IS_SWAPFILE(target))
error = -EPERM;
else if (is_local_mountpoint(dentry))
error = -EBUSY;
else {
error = security_inode_unlink(dir, dentry);
if (!error) {
error = try_break_deleg(target, delegated_inode);
if (error)
goto out;
error = dir->i_op->unlink(dir, dentry);
vfs: Lazily remove mounts on unlinked files and directories. With the introduction of mount namespaces and bind mounts it became possible to access files and directories that on some paths are mount points but are not mount points on other paths. It is very confusing when rm -rf somedir returns -EBUSY simply because somedir is mounted somewhere else. With the addition of user namespaces allowing unprivileged mounts this condition has gone from annoying to allowing a DOS attack on other users in the system. The possibility for mischief is removed by updating the vfs to support rename, unlink and rmdir on a dentry that is a mountpoint and by lazily unmounting mountpoints on deleted dentries. In particular this change allows rename, unlink and rmdir system calls on a dentry without a mountpoint in the current mount namespace to succeed, and it allows rename, unlink, and rmdir performed on a distributed filesystem to update the vfs cache even if when there is a mount in some namespace on the original dentry. There are two common patterns of maintaining mounts: Mounts on trusted paths with the parent directory of the mount point and all ancestory directories up to / owned by root and modifiable only by root (i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu, cpuacct, ...}, /usr, /usr/local). Mounts on unprivileged directories maintained by fusermount. In the case of mounts in trusted directories owned by root and modifiable only by root the current parent directory permissions are sufficient to ensure a mount point on a trusted path is not removed or renamed by anyone other than root, even if there is a context where the there are no mount points to prevent this. In the case of mounts in directories owned by less privileged users races with users modifying the path of a mount point are already a danger. fusermount already uses a combination of chdir, /proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races. The removable of global rename, unlink, and rmdir protection really adds nothing new to consider only a widening of the attack window, and fusermount is already safe against unprivileged users modifying the directory simultaneously. In principle for perfect userspace programs returning -EBUSY for unlink, rmdir, and rename of dentires that have mounts in the local namespace is actually unnecessary. Unfortunately not all userspace programs are perfect so retaining -EBUSY for unlink, rmdir and rename of dentries that have mounts in the current mount namespace plays an important role of maintaining consistency with historical behavior and making imperfect userspace applications hard to exploit. v2: Remove spurious old_dentry. v3: Optimized shrink_submounts_and_drop Removed unsued afs label v4: Simplified the changes to check_submounts_and_drop Do not rename check_submounts_and_drop shrink_submounts_and_drop Document what why we need atomicity in check_submounts_and_drop Rely on the parent inode mutex to make d_revalidate and d_invalidate an atomic unit. v5: Refcount the mountpoint to detach in case of simultaneous renames. Reviewed-by: Miklos Szeredi <miklos@szeredi.hu> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-10-02 09:33:48 +08:00
if (!error) {
dont_mount(dentry);
vfs: Lazily remove mounts on unlinked files and directories. With the introduction of mount namespaces and bind mounts it became possible to access files and directories that on some paths are mount points but are not mount points on other paths. It is very confusing when rm -rf somedir returns -EBUSY simply because somedir is mounted somewhere else. With the addition of user namespaces allowing unprivileged mounts this condition has gone from annoying to allowing a DOS attack on other users in the system. The possibility for mischief is removed by updating the vfs to support rename, unlink and rmdir on a dentry that is a mountpoint and by lazily unmounting mountpoints on deleted dentries. In particular this change allows rename, unlink and rmdir system calls on a dentry without a mountpoint in the current mount namespace to succeed, and it allows rename, unlink, and rmdir performed on a distributed filesystem to update the vfs cache even if when there is a mount in some namespace on the original dentry. There are two common patterns of maintaining mounts: Mounts on trusted paths with the parent directory of the mount point and all ancestory directories up to / owned by root and modifiable only by root (i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu, cpuacct, ...}, /usr, /usr/local). Mounts on unprivileged directories maintained by fusermount. In the case of mounts in trusted directories owned by root and modifiable only by root the current parent directory permissions are sufficient to ensure a mount point on a trusted path is not removed or renamed by anyone other than root, even if there is a context where the there are no mount points to prevent this. In the case of mounts in directories owned by less privileged users races with users modifying the path of a mount point are already a danger. fusermount already uses a combination of chdir, /proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races. The removable of global rename, unlink, and rmdir protection really adds nothing new to consider only a widening of the attack window, and fusermount is already safe against unprivileged users modifying the directory simultaneously. In principle for perfect userspace programs returning -EBUSY for unlink, rmdir, and rename of dentires that have mounts in the local namespace is actually unnecessary. Unfortunately not all userspace programs are perfect so retaining -EBUSY for unlink, rmdir and rename of dentries that have mounts in the current mount namespace plays an important role of maintaining consistency with historical behavior and making imperfect userspace applications hard to exploit. v2: Remove spurious old_dentry. v3: Optimized shrink_submounts_and_drop Removed unsued afs label v4: Simplified the changes to check_submounts_and_drop Do not rename check_submounts_and_drop shrink_submounts_and_drop Document what why we need atomicity in check_submounts_and_drop Rely on the parent inode mutex to make d_revalidate and d_invalidate an atomic unit. v5: Refcount the mountpoint to detach in case of simultaneous renames. Reviewed-by: Miklos Szeredi <miklos@szeredi.hu> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-10-02 09:33:48 +08:00
detach_mounts(dentry);
}
}
}
out:
inode_unlock(target);
/* We don't d_delete() NFS sillyrenamed files--they still exist. */
fsnotify: invalidate dcache before IN_DELETE event Apparently, there are some applications that use IN_DELETE event as an invalidation mechanism and expect that if they try to open a file with the name reported with the delete event, that it should not contain the content of the deleted file. Commit 49246466a989 ("fsnotify: move fsnotify_nameremove() hook out of d_delete()") moved the fsnotify delete hook before d_delete() so fsnotify will have access to a positive dentry. This allowed a race where opening the deleted file via cached dentry is now possible after receiving the IN_DELETE event. To fix the regression, create a new hook fsnotify_delete() that takes the unlinked inode as an argument and use a helper d_delete_notify() to pin the inode, so we can pass it to fsnotify_delete() after d_delete(). Backporting hint: this regression is from v5.3. Although patch will apply with only trivial conflicts to v5.4 and v5.10, it won't build, because fsnotify_delete() implementation is different in each of those versions (see fsnotify_link()). A follow up patch will fix the fsnotify_unlink/rmdir() calls in pseudo filesystem that do not need to call d_delete(). Link: https://lore.kernel.org/r/20220120215305.282577-1-amir73il@gmail.com Reported-by: Ivan Delalande <colona@arista.com> Link: https://lore.kernel.org/linux-fsdevel/YeNyzoDM5hP5LtGW@visor/ Fixes: 49246466a989 ("fsnotify: move fsnotify_nameremove() hook out of d_delete()") Cc: stable@vger.kernel.org # v5.3+ Signed-off-by: Amir Goldstein <amir73il@gmail.com> Signed-off-by: Jan Kara <jack@suse.cz>
2022-01-21 05:53:04 +08:00
if (!error && dentry->d_flags & DCACHE_NFSFS_RENAMED) {
fsnotify_unlink(dir, dentry);
} else if (!error) {
fsnotify_link_count(target);
fsnotify: invalidate dcache before IN_DELETE event Apparently, there are some applications that use IN_DELETE event as an invalidation mechanism and expect that if they try to open a file with the name reported with the delete event, that it should not contain the content of the deleted file. Commit 49246466a989 ("fsnotify: move fsnotify_nameremove() hook out of d_delete()") moved the fsnotify delete hook before d_delete() so fsnotify will have access to a positive dentry. This allowed a race where opening the deleted file via cached dentry is now possible after receiving the IN_DELETE event. To fix the regression, create a new hook fsnotify_delete() that takes the unlinked inode as an argument and use a helper d_delete_notify() to pin the inode, so we can pass it to fsnotify_delete() after d_delete(). Backporting hint: this regression is from v5.3. Although patch will apply with only trivial conflicts to v5.4 and v5.10, it won't build, because fsnotify_delete() implementation is different in each of those versions (see fsnotify_link()). A follow up patch will fix the fsnotify_unlink/rmdir() calls in pseudo filesystem that do not need to call d_delete(). Link: https://lore.kernel.org/r/20220120215305.282577-1-amir73il@gmail.com Reported-by: Ivan Delalande <colona@arista.com> Link: https://lore.kernel.org/linux-fsdevel/YeNyzoDM5hP5LtGW@visor/ Fixes: 49246466a989 ("fsnotify: move fsnotify_nameremove() hook out of d_delete()") Cc: stable@vger.kernel.org # v5.3+ Signed-off-by: Amir Goldstein <amir73il@gmail.com> Signed-off-by: Jan Kara <jack@suse.cz>
2022-01-21 05:53:04 +08:00
d_delete_notify(dir, dentry);
}
return error;
}
EXPORT_SYMBOL(vfs_unlink);
/*
* Make sure that the actual truncation of the file will occur outside its
* directory's i_mutex. Truncate can take a long time if there is a lot of
* writeout happening, and we don't want to prevent access to the directory
* while waiting on the I/O.
*/
int do_unlinkat(int dfd, struct filename *name)
{
int error;
struct dentry *dentry;
struct path path;
struct qstr last;
int type;
struct inode *inode = NULL;
struct inode *delegated_inode = NULL;
unsigned int lookup_flags = 0;
retry:
error = filename_parentat(dfd, name, lookup_flags, &path, &last, &type);
if (error)
goto exit1;
error = -EISDIR;
if (type != LAST_NORM)
goto exit2;
error = mnt_want_write(path.mnt);
if (error)
goto exit2;
retry_deleg:
inode_lock_nested(path.dentry->d_inode, I_MUTEX_PARENT);
dentry = lookup_one_qstr_excl(&last, path.dentry, lookup_flags);
error = PTR_ERR(dentry);
if (!IS_ERR(dentry)) {
/* Why not before? Because we want correct error value */
if (last.name[last.len] || d_is_negative(dentry))
fix wrong iput on d_inode introduced by e6bc45d65d Git bisection shows that commit e6bc45d65df8599fdbae73be9cec4ceed274db53 causes BUG_ONs under high I/O load: kernel BUG at fs/inode.c:1368! [ 2862.501007] Call Trace: [ 2862.501007] [<ffffffff811691d8>] d_kill+0xf8/0x140 [ 2862.501007] [<ffffffff81169c19>] dput+0xc9/0x190 [ 2862.501007] [<ffffffff8115577f>] fput+0x15f/0x210 [ 2862.501007] [<ffffffff81152171>] filp_close+0x61/0x90 [ 2862.501007] [<ffffffff81152251>] sys_close+0xb1/0x110 [ 2862.501007] [<ffffffff814c14fb>] system_call_fastpath+0x16/0x1b A reliable way to reproduce this bug is: Login to KDE, run 'rsnapshot sync', and apt-get install openjdk-6-jdk, and apt-get remove openjdk-6-jdk. The buggy part of the patch is this: struct inode *inode = NULL; ..... - if (nd.last.name[nd.last.len]) - goto slashes; inode = dentry->d_inode; - if (inode) - ihold(inode); + if (nd.last.name[nd.last.len] || !inode) + goto slashes; + ihold(inode) ... if (inode) iput(inode); /* truncate the inode here */ If nd.last.name[nd.last.len] is nonzero (and thus goto slashes branch is taken), and dentry->d_inode is non-NULL, then this code now does an additional iput on the inode, which is wrong. Fix this by only setting the inode variable if nd.last.name[nd.last.len] is 0. Reference: https://lkml.org/lkml/2011/6/15/50 Reported-by: Norbert Preining <preining@logic.at> Reported-by: Török Edwin <edwintorok@gmail.com> Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Török Edwin <edwintorok@gmail.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2011-06-16 05:06:14 +08:00
goto slashes;
inode = dentry->d_inode;
ihold(inode);
error = security_path_unlink(&path, dentry);
if (error)
goto exit3;
error = vfs_unlink(mnt_idmap(path.mnt), path.dentry->d_inode,
dentry, &delegated_inode);
exit3:
dput(dentry);
}
inode_unlock(path.dentry->d_inode);
if (inode)
iput(inode); /* truncate the inode here */
inode = NULL;
if (delegated_inode) {
error = break_deleg_wait(&delegated_inode);
if (!error)
goto retry_deleg;
}
mnt_drop_write(path.mnt);
exit2:
path_put(&path);
if (retry_estale(error, lookup_flags)) {
lookup_flags |= LOOKUP_REVAL;
inode = NULL;
goto retry;
}
exit1:
putname(name);
return error;
slashes:
if (d_is_negative(dentry))
error = -ENOENT;
else if (d_is_dir(dentry))
error = -EISDIR;
else
error = -ENOTDIR;
goto exit3;
}
SYSCALL_DEFINE3(unlinkat, int, dfd, const char __user *, pathname, int, flag)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
if ((flag & ~AT_REMOVEDIR) != 0)
return -EINVAL;
if (flag & AT_REMOVEDIR)
return do_rmdir(dfd, getname(pathname));
return do_unlinkat(dfd, getname(pathname));
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
SYSCALL_DEFINE1(unlink, const char __user *, pathname)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_unlinkat(AT_FDCWD, getname(pathname));
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
/**
* vfs_symlink - create symlink
* @idmap: idmap of the mount the inode was found from
* @dir: inode of the parent directory
* @dentry: dentry of the child symlink file
* @oldname: name of the file to link to
*
* Create a symlink.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_symlink(struct mnt_idmap *idmap, struct inode *dir,
struct dentry *dentry, const char *oldname)
{
int error;
error = may_create(idmap, dir, dentry);
if (error)
return error;
if (!dir->i_op->symlink)
return -EPERM;
error = security_inode_symlink(dir, dentry, oldname);
if (error)
return error;
error = dir->i_op->symlink(idmap, dir, dentry, oldname);
if (!error)
fsnotify_create(dir, dentry);
return error;
}
EXPORT_SYMBOL(vfs_symlink);
int do_symlinkat(struct filename *from, int newdfd, struct filename *to)
{
int error;
struct dentry *dentry;
struct path path;
unsigned int lookup_flags = 0;
if (IS_ERR(from)) {
error = PTR_ERR(from);
goto out_putnames;
}
retry:
dentry = filename_create(newdfd, to, &path, lookup_flags);
error = PTR_ERR(dentry);
if (IS_ERR(dentry))
goto out_putnames;
error = security_path_symlink(&path, dentry, from->name);
if (!error)
error = vfs_symlink(mnt_idmap(path.mnt), path.dentry->d_inode,
dentry, from->name);
done_path_create(&path, dentry);
if (retry_estale(error, lookup_flags)) {
lookup_flags |= LOOKUP_REVAL;
goto retry;
}
out_putnames:
putname(to);
putname(from);
return error;
}
SYSCALL_DEFINE3(symlinkat, const char __user *, oldname,
int, newdfd, const char __user *, newname)
{
return do_symlinkat(getname(oldname), newdfd, getname(newname));
}
SYSCALL_DEFINE2(symlink, const char __user *, oldname, const char __user *, newname)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_symlinkat(getname(oldname), AT_FDCWD, getname(newname));
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
/**
* vfs_link - create a new link
* @old_dentry: object to be linked
* @idmap: idmap of the mount
* @dir: new parent
* @new_dentry: where to create the new link
* @delegated_inode: returns inode needing a delegation break
*
* The caller must hold dir->i_mutex
*
* If vfs_link discovers a delegation on the to-be-linked file in need
* of breaking, it will return -EWOULDBLOCK and return a reference to the
* inode in delegated_inode. The caller should then break the delegation
* and retry. Because breaking a delegation may take a long time, the
* caller should drop the i_mutex before doing so.
*
* Alternatively, a caller may pass NULL for delegated_inode. This may
* be appropriate for callers that expect the underlying filesystem not
* to be NFS exported.
*
* If the inode has been found through an idmapped mount the idmap of
* the vfsmount must be passed through @idmap. This function will then take
* care to map the inode according to @idmap before checking permissions.
* On non-idmapped mounts or if permission checking is to be performed on the
* raw inode simply pass @nop_mnt_idmap.
*/
int vfs_link(struct dentry *old_dentry, struct mnt_idmap *idmap,
struct inode *dir, struct dentry *new_dentry,
struct inode **delegated_inode)
{
struct inode *inode = old_dentry->d_inode;
unsigned max_links = dir->i_sb->s_max_links;
int error;
if (!inode)
return -ENOENT;
error = may_create(idmap, dir, new_dentry);
if (error)
return error;
if (dir->i_sb != inode->i_sb)
return -EXDEV;
/*
* A link to an append-only or immutable file cannot be created.
*/
if (IS_APPEND(inode) || IS_IMMUTABLE(inode))
return -EPERM;
vfs: Don't modify inodes with a uid or gid unknown to the vfs When a filesystem outside of init_user_ns is mounted it could have uids and gids stored in it that do not map to init_user_ns. The plan is to allow those filesystems to set i_uid to INVALID_UID and i_gid to INVALID_GID for unmapped uids and gids and then to handle that strange case in the vfs to ensure there is consistent robust handling of the weirdness. Upon a careful review of the vfs and filesystems about the only case where there is any possibility of confusion or trouble is when the inode is written back to disk. In that case filesystems typically read the inode->i_uid and inode->i_gid and write them to disk even when just an inode timestamp is being updated. Which leads to a rule that is very simple to implement and understand inodes whose i_uid or i_gid is not valid may not be written. In dealing with access times this means treat those inodes as if the inode flag S_NOATIME was set. Reads of the inodes appear safe and useful, but any write or modification is disallowed. The only inode write that is allowed is a chown that sets the uid and gid on the inode to valid values. After such a chown the inode is normal and may be treated as such. Denying all writes to inodes with uids or gids unknown to the vfs also prevents several oddball cases where corruption would have occurred because the vfs does not have complete information. One problem case that is prevented is attempting to use the gid of a directory for new inodes where the directories sgid bit is set but the directories gid is not mapped. Another problem case avoided is attempting to update the evm hash after setxattr, removexattr, and setattr. As the evm hash includeds the inode->i_uid or inode->i_gid not knowning the uid or gid prevents a correct evm hash from being computed. evm hash verification also fails when i_uid or i_gid is unknown but that is essentially harmless as it does not cause filesystem corruption. Acked-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-06-30 03:54:46 +08:00
/*
* Updating the link count will likely cause i_uid and i_gid to
* be writen back improperly if their true value is unknown to
* the vfs.
*/
if (HAS_UNMAPPED_ID(idmap, inode))
vfs: Don't modify inodes with a uid or gid unknown to the vfs When a filesystem outside of init_user_ns is mounted it could have uids and gids stored in it that do not map to init_user_ns. The plan is to allow those filesystems to set i_uid to INVALID_UID and i_gid to INVALID_GID for unmapped uids and gids and then to handle that strange case in the vfs to ensure there is consistent robust handling of the weirdness. Upon a careful review of the vfs and filesystems about the only case where there is any possibility of confusion or trouble is when the inode is written back to disk. In that case filesystems typically read the inode->i_uid and inode->i_gid and write them to disk even when just an inode timestamp is being updated. Which leads to a rule that is very simple to implement and understand inodes whose i_uid or i_gid is not valid may not be written. In dealing with access times this means treat those inodes as if the inode flag S_NOATIME was set. Reads of the inodes appear safe and useful, but any write or modification is disallowed. The only inode write that is allowed is a chown that sets the uid and gid on the inode to valid values. After such a chown the inode is normal and may be treated as such. Denying all writes to inodes with uids or gids unknown to the vfs also prevents several oddball cases where corruption would have occurred because the vfs does not have complete information. One problem case that is prevented is attempting to use the gid of a directory for new inodes where the directories sgid bit is set but the directories gid is not mapped. Another problem case avoided is attempting to update the evm hash after setxattr, removexattr, and setattr. As the evm hash includeds the inode->i_uid or inode->i_gid not knowning the uid or gid prevents a correct evm hash from being computed. evm hash verification also fails when i_uid or i_gid is unknown but that is essentially harmless as it does not cause filesystem corruption. Acked-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-06-30 03:54:46 +08:00
return -EPERM;
if (!dir->i_op->link)
return -EPERM;
if (S_ISDIR(inode->i_mode))
return -EPERM;
error = security_inode_link(old_dentry, dir, new_dentry);
if (error)
return error;
inode_lock(inode);
/* Make sure we don't allow creating hardlink to an unlinked file */
if (inode->i_nlink == 0 && !(inode->i_state & I_LINKABLE))
error = -ENOENT;
else if (max_links && inode->i_nlink >= max_links)
error = -EMLINK;
else {
error = try_break_deleg(inode, delegated_inode);
if (!error)
error = dir->i_op->link(old_dentry, dir, new_dentry);
}
if (!error && (inode->i_state & I_LINKABLE)) {
spin_lock(&inode->i_lock);
inode->i_state &= ~I_LINKABLE;
spin_unlock(&inode->i_lock);
}
inode_unlock(inode);
if (!error)
fsnotify_link(dir, inode, new_dentry);
return error;
}
EXPORT_SYMBOL(vfs_link);
/*
* Hardlinks are often used in delicate situations. We avoid
* security-related surprises by not following symlinks on the
* newname. --KAB
*
* We don't follow them on the oldname either to be compatible
* with linux 2.0, and to avoid hard-linking to directories
* and other special files. --ADM
*/
int do_linkat(int olddfd, struct filename *old, int newdfd,
struct filename *new, int flags)
{
struct mnt_idmap *idmap;
struct dentry *new_dentry;
struct path old_path, new_path;
struct inode *delegated_inode = NULL;
int how = 0;
int error;
if ((flags & ~(AT_SYMLINK_FOLLOW | AT_EMPTY_PATH)) != 0) {
error = -EINVAL;
goto out_putnames;
}
/*
vfs: relax linkat() AT_EMPTY_PATH - aka flink() - requirements "The definition of insanity is doing the same thing over and over again and expecting different results” We've tried to do this before, most recently with commit bb2314b47996 ("fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink") about a decade ago. But the effort goes back even further than that, eg this thread back from 1998 that is so old that we don't even have it archived in lore: https://lkml.org/lkml/1998/3/10/108 which also points out some of the reasons why it's dangerous. Or, how about then in 2003: https://lkml.org/lkml/2003/4/6/112 where we went through some of the same arguments, just wirh different people involved. In particular, having access to a file descriptor does not necessarily mean that you have access to the path that was used for lookup, and there may be very good reasons why you absolutely must not have access to a path to said file. For example, if we were passed a file descriptor from the outside into some limited environment (think chroot, but also user namespaces etc) a 'flink()' system call could now make that file visible inside a context where it's not supposed to be visible. In the process the user may also be able to re-open it with permissions that the original file descriptor did not have (eg a read-only file descriptor may be associated with an underlying file that is writable). Another variation on this is if somebody else (typically root) opens a file in a directory that is not accessible to others, and passes the file descriptor on as a read-only file. Again, the access to the file descriptor does not imply that you should have access to a path to the file in the filesystem. So while we have tried this several times in the past, it never works. The last time we did this, that commit bb2314b47996 quickly got reverted again in commit f0cc6ffb8ce8 (Revert "fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink"), with a note saying "We may re-do this once the whole discussion about the interface is done". Well, the discussion is long done, and didn't come to any resolution. There's no question that 'flink()' would be a useful operation, but it's a dangerous one. However, it does turn out that since 2008 (commit d76b0d9b2d87: "CRED: Use creds in file structs") we have had a fairly straightforward way to check whether the file descriptor was opened by the same credentials as the credentials of the flink(). That allows the most common patterns that people want to use, which tend to be to either open the source carefully (ie using the openat2() RESOLVE_xyz flags, and/or checking ownership with fstat() before linking), or to use O_TMPFILE and fill in the file contents before it's exposed to the world with linkat(). But it also means that if the file descriptor was opened by somebody else, or we've gone through a credentials change since, the operation no longer works (unless we have CAP_DAC_READ_SEARCH capabilities in the opener's user namespace, as before). Note that the credential equality check is done by using pointer equality, which means that it's not enough that you have effectively the same user - they have to be literally identical, since our credentials are using copy-on-write semantics. So you can't change your credentials to something else and try to change it back to the same ones between the open() and the linkat(). This is not meant to be some kind of generic permission check, this is literally meant as a "the open and link calls are 'atomic' wrt user credentials" check. It also means that you can't just move things between namespaces, because the credentials aren't just a list of uid's and gid's: they includes the pointer to the user_ns that the capabilities are relative to. So let's try this one more time and see if maybe this approach ends up being workable after all. Cc: Andrew Lutomirski <luto@kernel.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Cc: Peter Anvin <hpa@zytor.com> Cc: Jan Kara <jack@suse.cz> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://lore.kernel.org/r/20240411001012.12513-1-torvalds@linux-foundation.org [brauner: relax capability check to opener of the file] Link: https://lore.kernel.org/all/20231113-undenkbar-gediegen-efde5f1c34bc@brauner Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-04-11 08:10:12 +08:00
* To use null names we require CAP_DAC_READ_SEARCH or
* that the open-time creds of the dfd matches current.
* This ensures that not everyone will be able to create
vfs: relax linkat() AT_EMPTY_PATH - aka flink() - requirements "The definition of insanity is doing the same thing over and over again and expecting different results” We've tried to do this before, most recently with commit bb2314b47996 ("fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink") about a decade ago. But the effort goes back even further than that, eg this thread back from 1998 that is so old that we don't even have it archived in lore: https://lkml.org/lkml/1998/3/10/108 which also points out some of the reasons why it's dangerous. Or, how about then in 2003: https://lkml.org/lkml/2003/4/6/112 where we went through some of the same arguments, just wirh different people involved. In particular, having access to a file descriptor does not necessarily mean that you have access to the path that was used for lookup, and there may be very good reasons why you absolutely must not have access to a path to said file. For example, if we were passed a file descriptor from the outside into some limited environment (think chroot, but also user namespaces etc) a 'flink()' system call could now make that file visible inside a context where it's not supposed to be visible. In the process the user may also be able to re-open it with permissions that the original file descriptor did not have (eg a read-only file descriptor may be associated with an underlying file that is writable). Another variation on this is if somebody else (typically root) opens a file in a directory that is not accessible to others, and passes the file descriptor on as a read-only file. Again, the access to the file descriptor does not imply that you should have access to a path to the file in the filesystem. So while we have tried this several times in the past, it never works. The last time we did this, that commit bb2314b47996 quickly got reverted again in commit f0cc6ffb8ce8 (Revert "fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink"), with a note saying "We may re-do this once the whole discussion about the interface is done". Well, the discussion is long done, and didn't come to any resolution. There's no question that 'flink()' would be a useful operation, but it's a dangerous one. However, it does turn out that since 2008 (commit d76b0d9b2d87: "CRED: Use creds in file structs") we have had a fairly straightforward way to check whether the file descriptor was opened by the same credentials as the credentials of the flink(). That allows the most common patterns that people want to use, which tend to be to either open the source carefully (ie using the openat2() RESOLVE_xyz flags, and/or checking ownership with fstat() before linking), or to use O_TMPFILE and fill in the file contents before it's exposed to the world with linkat(). But it also means that if the file descriptor was opened by somebody else, or we've gone through a credentials change since, the operation no longer works (unless we have CAP_DAC_READ_SEARCH capabilities in the opener's user namespace, as before). Note that the credential equality check is done by using pointer equality, which means that it's not enough that you have effectively the same user - they have to be literally identical, since our credentials are using copy-on-write semantics. So you can't change your credentials to something else and try to change it back to the same ones between the open() and the linkat(). This is not meant to be some kind of generic permission check, this is literally meant as a "the open and link calls are 'atomic' wrt user credentials" check. It also means that you can't just move things between namespaces, because the credentials aren't just a list of uid's and gid's: they includes the pointer to the user_ns that the capabilities are relative to. So let's try this one more time and see if maybe this approach ends up being workable after all. Cc: Andrew Lutomirski <luto@kernel.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Cc: Peter Anvin <hpa@zytor.com> Cc: Jan Kara <jack@suse.cz> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://lore.kernel.org/r/20240411001012.12513-1-torvalds@linux-foundation.org [brauner: relax capability check to opener of the file] Link: https://lore.kernel.org/all/20231113-undenkbar-gediegen-efde5f1c34bc@brauner Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-04-11 08:10:12 +08:00
* a hardlink using the passed file descriptor.
*/
vfs: relax linkat() AT_EMPTY_PATH - aka flink() - requirements "The definition of insanity is doing the same thing over and over again and expecting different results” We've tried to do this before, most recently with commit bb2314b47996 ("fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink") about a decade ago. But the effort goes back even further than that, eg this thread back from 1998 that is so old that we don't even have it archived in lore: https://lkml.org/lkml/1998/3/10/108 which also points out some of the reasons why it's dangerous. Or, how about then in 2003: https://lkml.org/lkml/2003/4/6/112 where we went through some of the same arguments, just wirh different people involved. In particular, having access to a file descriptor does not necessarily mean that you have access to the path that was used for lookup, and there may be very good reasons why you absolutely must not have access to a path to said file. For example, if we were passed a file descriptor from the outside into some limited environment (think chroot, but also user namespaces etc) a 'flink()' system call could now make that file visible inside a context where it's not supposed to be visible. In the process the user may also be able to re-open it with permissions that the original file descriptor did not have (eg a read-only file descriptor may be associated with an underlying file that is writable). Another variation on this is if somebody else (typically root) opens a file in a directory that is not accessible to others, and passes the file descriptor on as a read-only file. Again, the access to the file descriptor does not imply that you should have access to a path to the file in the filesystem. So while we have tried this several times in the past, it never works. The last time we did this, that commit bb2314b47996 quickly got reverted again in commit f0cc6ffb8ce8 (Revert "fs: Allow unprivileged linkat(..., AT_EMPTY_PATH) aka flink"), with a note saying "We may re-do this once the whole discussion about the interface is done". Well, the discussion is long done, and didn't come to any resolution. There's no question that 'flink()' would be a useful operation, but it's a dangerous one. However, it does turn out that since 2008 (commit d76b0d9b2d87: "CRED: Use creds in file structs") we have had a fairly straightforward way to check whether the file descriptor was opened by the same credentials as the credentials of the flink(). That allows the most common patterns that people want to use, which tend to be to either open the source carefully (ie using the openat2() RESOLVE_xyz flags, and/or checking ownership with fstat() before linking), or to use O_TMPFILE and fill in the file contents before it's exposed to the world with linkat(). But it also means that if the file descriptor was opened by somebody else, or we've gone through a credentials change since, the operation no longer works (unless we have CAP_DAC_READ_SEARCH capabilities in the opener's user namespace, as before). Note that the credential equality check is done by using pointer equality, which means that it's not enough that you have effectively the same user - they have to be literally identical, since our credentials are using copy-on-write semantics. So you can't change your credentials to something else and try to change it back to the same ones between the open() and the linkat(). This is not meant to be some kind of generic permission check, this is literally meant as a "the open and link calls are 'atomic' wrt user credentials" check. It also means that you can't just move things between namespaces, because the credentials aren't just a list of uid's and gid's: they includes the pointer to the user_ns that the capabilities are relative to. So let's try this one more time and see if maybe this approach ends up being workable after all. Cc: Andrew Lutomirski <luto@kernel.org> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Christian Brauner <brauner@kernel.org> Cc: Peter Anvin <hpa@zytor.com> Cc: Jan Kara <jack@suse.cz> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Link: https://lore.kernel.org/r/20240411001012.12513-1-torvalds@linux-foundation.org [brauner: relax capability check to opener of the file] Link: https://lore.kernel.org/all/20231113-undenkbar-gediegen-efde5f1c34bc@brauner Signed-off-by: Christian Brauner <brauner@kernel.org>
2024-04-11 08:10:12 +08:00
if (flags & AT_EMPTY_PATH)
how |= LOOKUP_LINKAT_EMPTY;
if (flags & AT_SYMLINK_FOLLOW)
how |= LOOKUP_FOLLOW;
retry:
error = filename_lookup(olddfd, old, how, &old_path, NULL);
if (error)
goto out_putnames;
new_dentry = filename_create(newdfd, new, &new_path,
(how & LOOKUP_REVAL));
error = PTR_ERR(new_dentry);
if (IS_ERR(new_dentry))
goto out_putpath;
error = -EXDEV;
if (old_path.mnt != new_path.mnt)
goto out_dput;
idmap = mnt_idmap(new_path.mnt);
error = may_linkat(idmap, &old_path);
fs: add link restrictions This adds symlink and hardlink restrictions to the Linux VFS. Symlinks: A long-standing class of security issues is the symlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given symlink (i.e. a root process follows a symlink belonging to another user). For a likely incomplete list of hundreds of examples across the years, please see: http://cve.mitre.org/cgi-bin/cvekey.cgi?keyword=/tmp The solution is to permit symlinks to only be followed when outside a sticky world-writable directory, or when the uid of the symlink and follower match, or when the directory owner matches the symlink's owner. Some pointers to the history of earlier discussion that I could find: 1996 Aug, Zygo Blaxell http://marc.info/?l=bugtraq&m=87602167419830&w=2 1996 Oct, Andrew Tridgell http://lkml.indiana.edu/hypermail/linux/kernel/9610.2/0086.html 1997 Dec, Albert D Cahalan http://lkml.org/lkml/1997/12/16/4 2005 Feb, Lorenzo Hernández García-Hierro http://lkml.indiana.edu/hypermail/linux/kernel/0502.0/1896.html 2010 May, Kees Cook https://lkml.org/lkml/2010/5/30/144 Past objections and rebuttals could be summarized as: - Violates POSIX. - POSIX didn't consider this situation and it's not useful to follow a broken specification at the cost of security. - Might break unknown applications that use this feature. - Applications that break because of the change are easy to spot and fix. Applications that are vulnerable to symlink ToCToU by not having the change aren't. Additionally, no applications have yet been found that rely on this behavior. - Applications should just use mkstemp() or O_CREATE|O_EXCL. - True, but applications are not perfect, and new software is written all the time that makes these mistakes; blocking this flaw at the kernel is a single solution to the entire class of vulnerability. - This should live in the core VFS. - This should live in an LSM. (https://lkml.org/lkml/2010/5/31/135) - This should live in an LSM. - This should live in the core VFS. (https://lkml.org/lkml/2010/8/2/188) Hardlinks: On systems that have user-writable directories on the same partition as system files, a long-standing class of security issues is the hardlink-based time-of-check-time-of-use race, most commonly seen in world-writable directories like /tmp. The common method of exploitation of this flaw is to cross privilege boundaries when following a given hardlink (i.e. a root process follows a hardlink created by another user). Additionally, an issue exists where users can "pin" a potentially vulnerable setuid/setgid file so that an administrator will not actually upgrade a system fully. The solution is to permit hardlinks to only be created when the user is already the existing file's owner, or if they already have read/write access to the existing file. Many Linux users are surprised when they learn they can link to files they have no access to, so this change appears to follow the doctrine of "least surprise". Additionally, this change does not violate POSIX, which states "the implementation may require that the calling process has permission to access the existing file"[1]. This change is known to break some implementations of the "at" daemon, though the version used by Fedora and Ubuntu has been fixed[2] for a while. Otherwise, the change has been undisruptive while in use in Ubuntu for the last 1.5 years. [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/linkat.html [2] http://anonscm.debian.org/gitweb/?p=collab-maint/at.git;a=commitdiff;h=f4114656c3a6c6f6070e315ffdf940a49eda3279 This patch is based on the patches in Openwall and grsecurity, along with suggestions from Al Viro. I have added a sysctl to enable the protected behavior, and documentation. Signed-off-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-07-26 08:29:07 +08:00
if (unlikely(error))
goto out_dput;
error = security_path_link(old_path.dentry, &new_path, new_dentry);
if (error)
goto out_dput;
error = vfs_link(old_path.dentry, idmap, new_path.dentry->d_inode,
new_dentry, &delegated_inode);
out_dput:
done_path_create(&new_path, new_dentry);
if (delegated_inode) {
error = break_deleg_wait(&delegated_inode);
if (!error) {
path_put(&old_path);
goto retry;
}
}
if (retry_estale(error, how)) {
path_put(&old_path);
how |= LOOKUP_REVAL;
goto retry;
}
out_putpath:
path_put(&old_path);
out_putnames:
putname(old);
putname(new);
return error;
}
SYSCALL_DEFINE5(linkat, int, olddfd, const char __user *, oldname,
int, newdfd, const char __user *, newname, int, flags)
{
return do_linkat(olddfd, getname_uflags(oldname, flags),
newdfd, getname(newname), flags);
}
SYSCALL_DEFINE2(link, const char __user *, oldname, const char __user *, newname)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_linkat(AT_FDCWD, getname(oldname), AT_FDCWD, getname(newname), 0);
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
/**
* vfs_rename - rename a filesystem object
namei: fix kernel-doc for struct renamedata and more Fix kernel-doc warnings in namei.c: ../fs/namei.c:1149: warning: Excess function parameter 'dir_mode' description in 'may_create_in_sticky' ../fs/namei.c:1149: warning: Excess function parameter 'dir_uid' description in 'may_create_in_sticky' ../fs/namei.c:3396: warning: Function parameter or member 'open_flag' not described in 'vfs_tmpfile' ../fs/namei.c:3396: warning: Excess function parameter 'open_flags' description in 'vfs_tmpfile' ../fs/namei.c:4460: warning: Function parameter or member 'rd' not described in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'old_mnt_userns' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'old_dir' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'old_dentry' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'new_mnt_userns' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'new_dir' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'new_dentry' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'delegated_inode' description in 'vfs_rename' ../fs/namei.c:4460: warning: Excess function parameter 'flags' description in 'vfs_rename' Link: https://lore.kernel.org/r/20210216042929.8931-3-rdunlap@infradead.org Fixes: 9fe61450972d ("namei: introduce struct renamedata") Signed-off-by: Randy Dunlap <rdunlap@infradead.org> Cc: David Howells <dhowells@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: linux-fsdevel@vger.kernel.org Cc: Christian Brauner <christian.brauner@ubuntu.com> Reviewed-by: Christoph Hellwig <hch@lst.de> Acked-by: Christian Brauner <christian.brauner@ubuntu.com> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
2021-02-16 12:29:28 +08:00
* @rd: pointer to &struct renamedata info
*
* The caller must hold multiple mutexes--see lock_rename()).
*
* If vfs_rename discovers a delegation in need of breaking at either
* the source or destination, it will return -EWOULDBLOCK and return a
* reference to the inode in delegated_inode. The caller should then
* break the delegation and retry. Because breaking a delegation may
* take a long time, the caller should drop all locks before doing
* so.
*
* Alternatively, a caller may pass NULL for delegated_inode. This may
* be appropriate for callers that expect the underlying filesystem not
* to be NFS exported.
*
* The worst of all namespace operations - renaming directory. "Perverted"
* doesn't even start to describe it. Somebody in UCB had a heck of a trip...
* Problems:
*
* a) we can get into loop creation.
* b) race potential - two innocent renames can create a loop together.
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
* That's where 4.4BSD screws up. Current fix: serialization on
* sb->s_vfs_rename_mutex. We might be more accurate, but that's another
* story.
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
* c) we may have to lock up to _four_ objects - parents and victim (if it exists),
* and source (if it's a non-directory or a subdirectory that moves to
* different parent).
* And that - after we got ->i_mutex on parents (until then we don't know
* whether the target exists). Solution: try to be smart with locking
* order for inodes. We rely on the fact that tree topology may change
* only under ->s_vfs_rename_mutex _and_ that parent of the object we
* move will be locked. Thus we can rank directories by the tree
* (ancestors first) and rank all non-directories after them.
* That works since everybody except rename does "lock parent, lookup,
* lock child" and rename is under ->s_vfs_rename_mutex.
* HOWEVER, it relies on the assumption that any object with ->lookup()
* has no more than 1 dentry. If "hybrid" objects will ever appear,
* we'd better make sure that there's no link(2) for them.
* d) conversion from fhandle to dentry may come in the wrong moment - when
* we are removing the target. Solution: we will have to grab ->i_mutex
* in the fhandle_to_dentry code. [FIXME - current nfsfh.c relies on
* ->i_mutex on parents, which works but leads to some truly excessive
* locking].
*/
int vfs_rename(struct renamedata *rd)
{
int error;
struct inode *old_dir = rd->old_dir, *new_dir = rd->new_dir;
struct dentry *old_dentry = rd->old_dentry;
struct dentry *new_dentry = rd->new_dentry;
struct inode **delegated_inode = rd->delegated_inode;
unsigned int flags = rd->flags;
bool is_dir = d_is_dir(old_dentry);
struct inode *source = old_dentry->d_inode;
struct inode *target = new_dentry->d_inode;
bool new_is_dir = false;
unsigned max_links = new_dir->i_sb->s_max_links;
struct name_snapshot old_name;
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
bool lock_old_subdir, lock_new_subdir;
if (source == target)
return 0;
error = may_delete(rd->old_mnt_idmap, old_dir, old_dentry, is_dir);
if (error)
return error;
if (!target) {
error = may_create(rd->new_mnt_idmap, new_dir, new_dentry);
} else {
new_is_dir = d_is_dir(new_dentry);
if (!(flags & RENAME_EXCHANGE))
error = may_delete(rd->new_mnt_idmap, new_dir,
new_dentry, is_dir);
else
error = may_delete(rd->new_mnt_idmap, new_dir,
new_dentry, new_is_dir);
}
if (error)
return error;
if (!old_dir->i_op->rename)
return -EPERM;
/*
* If we are going to change the parent - check write permissions,
* we'll need to flip '..'.
*/
if (new_dir != old_dir) {
if (is_dir) {
error = inode_permission(rd->old_mnt_idmap, source,
MAY_WRITE);
if (error)
return error;
}
if ((flags & RENAME_EXCHANGE) && new_is_dir) {
error = inode_permission(rd->new_mnt_idmap, target,
MAY_WRITE);
if (error)
return error;
}
}
error = security_inode_rename(old_dir, old_dentry, new_dir, new_dentry,
flags);
if (error)
return error;
take_dentry_name_snapshot(&old_name, old_dentry);
dget(new_dentry);
/*
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
* Lock children.
* The source subdirectory needs to be locked on cross-directory
* rename or cross-directory exchange since its parent changes.
* The target subdirectory needs to be locked on cross-directory
* exchange due to parent change and on any rename due to becoming
* a victim.
* Non-directories need locking in all cases (for NFS reasons);
* they get locked after any subdirectories (in inode address order).
*
* NOTE: WE ONLY LOCK UNRELATED DIRECTORIES IN CROSS-DIRECTORY CASE.
* NEVER, EVER DO THAT WITHOUT ->s_vfs_rename_mutex.
*/
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
lock_old_subdir = new_dir != old_dir;
lock_new_subdir = new_dir != old_dir || !(flags & RENAME_EXCHANGE);
if (is_dir) {
if (lock_old_subdir)
inode_lock_nested(source, I_MUTEX_CHILD);
if (target && (!new_is_dir || lock_new_subdir))
inode_lock(target);
} else if (new_is_dir) {
if (lock_new_subdir)
inode_lock_nested(target, I_MUTEX_CHILD);
inode_lock(source);
} else {
lock_two_nondirectories(source, target);
}
error = -EPERM;
if (IS_SWAPFILE(source) || (target && IS_SWAPFILE(target)))
goto out;
error = -EBUSY;
2013-10-05 10:15:13 +08:00
if (is_local_mountpoint(old_dentry) || is_local_mountpoint(new_dentry))
goto out;
if (max_links && new_dir != old_dir) {
error = -EMLINK;
if (is_dir && !new_is_dir && new_dir->i_nlink >= max_links)
goto out;
if ((flags & RENAME_EXCHANGE) && !is_dir && new_is_dir &&
old_dir->i_nlink >= max_links)
goto out;
}
if (!is_dir) {
error = try_break_deleg(source, delegated_inode);
if (error)
goto out;
}
if (target && !new_is_dir) {
error = try_break_deleg(target, delegated_inode);
if (error)
goto out;
}
error = old_dir->i_op->rename(rd->new_mnt_idmap, old_dir, old_dentry,
new_dir, new_dentry, flags);
if (error)
goto out;
if (!(flags & RENAME_EXCHANGE) && target) {
rmdir(),rename(): do shrink_dcache_parent() only on success Once upon a time ->rmdir() instances used to check if victim inode had more than one (in-core) reference and failed with -EBUSY if it had. The reason was race avoidance - emptiness check is worthless if somebody could just go and create new objects in the victim directory afterwards. With introduction of dcache the checks had been replaced with checking the refcount of dentry. However, since a cached negative lookup leaves a negative child dentry, such check had lead to false positives - with empty foo/ doing stat foo/bar before rmdir foo ended up with -EBUSY unless the negative dentry of foo/bar happened to be evicted by the time of rmdir(2). That had been fixed by doing shrink_dcache_parent() just before the refcount check. At the same time, ext2_rmdir() has grown a private solution that eliminated those -EBUSY - it did something (setting ->i_size to 0) which made any subsequent ext2_add_entry() fail. Unfortunately, even with shrink_dcache_parent() the check had been racy - after all, the victim itself could be found by dcache lookup just after we'd checked its refcount. That got fixed by a new helper (dentry_unhash()) that did shrink_dcache_parent() and unhashed the sucker if its refcount ended up equal to 1. That got called before ->rmdir(), turning the checks in ->rmdir() instances into "if not unhashed fail with -EBUSY". Which reduced the boilerplate nicely, but had an unpleasant side effect - now shrink_dcache_parent() had been done before the emptiness checks, leading to easily triggerable calls of shrink_dcache_parent() on arbitrary large subtrees, quite possibly nested into each other. Several years later the ext2-private trick had been generalized - (in-core) inodes of dead directories are flagged and calls of lookup, readdir and all directory-modifying methods were prevented in so marked directories. Remaining boilerplate in ->rmdir() instances became redundant and some instances got rid of it. In 2011 the call of dentry_unhash() got shifted into ->rmdir() instances and then killed off in all of them. That has lead to another problem, though - in case of successful rmdir we *want* any (negative) child dentries dropped and the victim itself made negative. There's no point keeping cached negative lookups in foo when we can get the negative lookup of foo itself cached. So shrink_dcache_parent() call had been restored; unfortunately, it went into the place where dentry_unhash() used to be, i.e. before the ->rmdir() call. Note that we don't unhash anymore, so any "is it busy" checks would be racy; fortunately, all of them are gone. We should've done that call right *after* successful ->rmdir(). That reduces contention caused by tree-walking in shrink_dcache_parent() and, especially, contention caused by evictions in two nested subtrees going on in parallel. The same goes for directory-overwriting rename() - the story there had been parallel to that of rmdir(). Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-05-28 04:23:51 +08:00
if (is_dir) {
shrink_dcache_parent(new_dentry);
target->i_flags |= S_DEAD;
rmdir(),rename(): do shrink_dcache_parent() only on success Once upon a time ->rmdir() instances used to check if victim inode had more than one (in-core) reference and failed with -EBUSY if it had. The reason was race avoidance - emptiness check is worthless if somebody could just go and create new objects in the victim directory afterwards. With introduction of dcache the checks had been replaced with checking the refcount of dentry. However, since a cached negative lookup leaves a negative child dentry, such check had lead to false positives - with empty foo/ doing stat foo/bar before rmdir foo ended up with -EBUSY unless the negative dentry of foo/bar happened to be evicted by the time of rmdir(2). That had been fixed by doing shrink_dcache_parent() just before the refcount check. At the same time, ext2_rmdir() has grown a private solution that eliminated those -EBUSY - it did something (setting ->i_size to 0) which made any subsequent ext2_add_entry() fail. Unfortunately, even with shrink_dcache_parent() the check had been racy - after all, the victim itself could be found by dcache lookup just after we'd checked its refcount. That got fixed by a new helper (dentry_unhash()) that did shrink_dcache_parent() and unhashed the sucker if its refcount ended up equal to 1. That got called before ->rmdir(), turning the checks in ->rmdir() instances into "if not unhashed fail with -EBUSY". Which reduced the boilerplate nicely, but had an unpleasant side effect - now shrink_dcache_parent() had been done before the emptiness checks, leading to easily triggerable calls of shrink_dcache_parent() on arbitrary large subtrees, quite possibly nested into each other. Several years later the ext2-private trick had been generalized - (in-core) inodes of dead directories are flagged and calls of lookup, readdir and all directory-modifying methods were prevented in so marked directories. Remaining boilerplate in ->rmdir() instances became redundant and some instances got rid of it. In 2011 the call of dentry_unhash() got shifted into ->rmdir() instances and then killed off in all of them. That has lead to another problem, though - in case of successful rmdir we *want* any (negative) child dentries dropped and the victim itself made negative. There's no point keeping cached negative lookups in foo when we can get the negative lookup of foo itself cached. So shrink_dcache_parent() call had been restored; unfortunately, it went into the place where dentry_unhash() used to be, i.e. before the ->rmdir() call. Note that we don't unhash anymore, so any "is it busy" checks would be racy; fortunately, all of them are gone. We should've done that call right *after* successful ->rmdir(). That reduces contention caused by tree-walking in shrink_dcache_parent() and, especially, contention caused by evictions in two nested subtrees going on in parallel. The same goes for directory-overwriting rename() - the story there had been parallel to that of rmdir(). Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2018-05-28 04:23:51 +08:00
}
dont_mount(new_dentry);
vfs: Lazily remove mounts on unlinked files and directories. With the introduction of mount namespaces and bind mounts it became possible to access files and directories that on some paths are mount points but are not mount points on other paths. It is very confusing when rm -rf somedir returns -EBUSY simply because somedir is mounted somewhere else. With the addition of user namespaces allowing unprivileged mounts this condition has gone from annoying to allowing a DOS attack on other users in the system. The possibility for mischief is removed by updating the vfs to support rename, unlink and rmdir on a dentry that is a mountpoint and by lazily unmounting mountpoints on deleted dentries. In particular this change allows rename, unlink and rmdir system calls on a dentry without a mountpoint in the current mount namespace to succeed, and it allows rename, unlink, and rmdir performed on a distributed filesystem to update the vfs cache even if when there is a mount in some namespace on the original dentry. There are two common patterns of maintaining mounts: Mounts on trusted paths with the parent directory of the mount point and all ancestory directories up to / owned by root and modifiable only by root (i.e. /media/xxx, /dev, /dev/pts, /proc, /sys, /sys/fs/cgroup/{cpu, cpuacct, ...}, /usr, /usr/local). Mounts on unprivileged directories maintained by fusermount. In the case of mounts in trusted directories owned by root and modifiable only by root the current parent directory permissions are sufficient to ensure a mount point on a trusted path is not removed or renamed by anyone other than root, even if there is a context where the there are no mount points to prevent this. In the case of mounts in directories owned by less privileged users races with users modifying the path of a mount point are already a danger. fusermount already uses a combination of chdir, /proc/<pid>/fd/NNN, and UMOUNT_NOFOLLOW to prevent these races. The removable of global rename, unlink, and rmdir protection really adds nothing new to consider only a widening of the attack window, and fusermount is already safe against unprivileged users modifying the directory simultaneously. In principle for perfect userspace programs returning -EBUSY for unlink, rmdir, and rename of dentires that have mounts in the local namespace is actually unnecessary. Unfortunately not all userspace programs are perfect so retaining -EBUSY for unlink, rmdir and rename of dentries that have mounts in the current mount namespace plays an important role of maintaining consistency with historical behavior and making imperfect userspace applications hard to exploit. v2: Remove spurious old_dentry. v3: Optimized shrink_submounts_and_drop Removed unsued afs label v4: Simplified the changes to check_submounts_and_drop Do not rename check_submounts_and_drop shrink_submounts_and_drop Document what why we need atomicity in check_submounts_and_drop Rely on the parent inode mutex to make d_revalidate and d_invalidate an atomic unit. v5: Refcount the mountpoint to detach in case of simultaneous renames. Reviewed-by: Miklos Szeredi <miklos@szeredi.hu> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-10-02 09:33:48 +08:00
detach_mounts(new_dentry);
}
if (!(old_dir->i_sb->s_type->fs_flags & FS_RENAME_DOES_D_MOVE)) {
if (!(flags & RENAME_EXCHANGE))
d_move(old_dentry, new_dentry);
else
d_exchange(old_dentry, new_dentry);
}
out:
rename(): fix the locking of subdirectories We should never lock two subdirectories without having taken ->s_vfs_rename_mutex; inode pointer order or not, the "order" proposed in 28eceeda130f "fs: Lock moved directories" is not transitive, with the usual consequences. The rationale for locking renamed subdirectory in all cases was the possibility of race between rename modifying .. in a subdirectory to reflect the new parent and another thread modifying the same subdirectory. For a lot of filesystems that's not a problem, but for some it can lead to trouble (e.g. the case when short directory contents is kept in the inode, but creating a file in it might push it across the size limit and copy its contents into separate data block(s)). However, we need that only in case when the parent does change - otherwise ->rename() doesn't need to do anything with .. entry in the first place. Some instances are lazy and do a tautological update anyway, but it's really not hard to avoid. Amended locking rules for rename(): find the parent(s) of source and target if source and target have the same parent lock the common parent else lock ->s_vfs_rename_mutex lock both parents, in ancestor-first order; if neither is an ancestor of another, lock the parent of source first. find the source and target. if source and target have the same parent if operation is an overwriting rename of a subdirectory lock the target subdirectory else if source is a subdirectory lock the source if target is a subdirectory lock the target lock non-directories involved, in inode pointer order if both source and target are such. That way we are guaranteed that parents are locked (for obvious reasons), that any renamed non-directory is locked (nfsd relies upon that), that any victim is locked (emptiness check needs that, among other things) and subdirectory that changes parent is locked (needed to protect the update of .. entries). We are also guaranteed that any operation locking more than one directory either takes ->s_vfs_rename_mutex or locks a parent followed by its child. Cc: stable@vger.kernel.org Fixes: 28eceeda130f "fs: Lock moved directories" Reviewed-by: Jan Kara <jack@suse.cz> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2023-11-20 09:25:58 +08:00
if (!is_dir || lock_old_subdir)
inode_unlock(source);
if (target && (!new_is_dir || lock_new_subdir))
inode_unlock(target);
dput(new_dentry);
if (!error) {
fsnotify_move(old_dir, new_dir, &old_name.name, is_dir,
!(flags & RENAME_EXCHANGE) ? target : NULL, old_dentry);
if (flags & RENAME_EXCHANGE) {
fsnotify_move(new_dir, old_dir, &old_dentry->d_name,
new_is_dir, NULL, new_dentry);
}
}
release_dentry_name_snapshot(&old_name);
return error;
}
EXPORT_SYMBOL(vfs_rename);
int do_renameat2(int olddfd, struct filename *from, int newdfd,
struct filename *to, unsigned int flags)
{
struct renamedata rd;
struct dentry *old_dentry, *new_dentry;
struct dentry *trap;
struct path old_path, new_path;
struct qstr old_last, new_last;
int old_type, new_type;
struct inode *delegated_inode = NULL;
unsigned int lookup_flags = 0, target_flags = LOOKUP_RENAME_TARGET;
bool should_retry = false;
int error = -EINVAL;
if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
goto put_names;
if ((flags & (RENAME_NOREPLACE | RENAME_WHITEOUT)) &&
(flags & RENAME_EXCHANGE))
goto put_names;
if (flags & RENAME_EXCHANGE)
target_flags = 0;
retry:
error = filename_parentat(olddfd, from, lookup_flags, &old_path,
&old_last, &old_type);
if (error)
goto put_names;
error = filename_parentat(newdfd, to, lookup_flags, &new_path, &new_last,
&new_type);
if (error)
goto exit1;
error = -EXDEV;
if (old_path.mnt != new_path.mnt)
goto exit2;
error = -EBUSY;
if (old_type != LAST_NORM)
goto exit2;
if (flags & RENAME_NOREPLACE)
error = -EEXIST;
if (new_type != LAST_NORM)
goto exit2;
error = mnt_want_write(old_path.mnt);
if (error)
goto exit2;
retry_deleg:
trap = lock_rename(new_path.dentry, old_path.dentry);
if (IS_ERR(trap)) {
error = PTR_ERR(trap);
goto exit_lock_rename;
}
old_dentry = lookup_one_qstr_excl(&old_last, old_path.dentry,
lookup_flags);
error = PTR_ERR(old_dentry);
if (IS_ERR(old_dentry))
goto exit3;
/* source must exist */
error = -ENOENT;
if (d_is_negative(old_dentry))
goto exit4;
new_dentry = lookup_one_qstr_excl(&new_last, new_path.dentry,
lookup_flags | target_flags);
error = PTR_ERR(new_dentry);
if (IS_ERR(new_dentry))
goto exit4;
error = -EEXIST;
if ((flags & RENAME_NOREPLACE) && d_is_positive(new_dentry))
goto exit5;
if (flags & RENAME_EXCHANGE) {
error = -ENOENT;
if (d_is_negative(new_dentry))
goto exit5;
if (!d_is_dir(new_dentry)) {
error = -ENOTDIR;
if (new_last.name[new_last.len])
goto exit5;
}
}
/* unless the source is a directory trailing slashes give -ENOTDIR */
if (!d_is_dir(old_dentry)) {
error = -ENOTDIR;
if (old_last.name[old_last.len])
goto exit5;
if (!(flags & RENAME_EXCHANGE) && new_last.name[new_last.len])
goto exit5;
}
/* source should not be ancestor of target */
error = -EINVAL;
if (old_dentry == trap)
goto exit5;
/* target should not be an ancestor of source */
if (!(flags & RENAME_EXCHANGE))
error = -ENOTEMPTY;
if (new_dentry == trap)
goto exit5;
error = security_path_rename(&old_path, old_dentry,
&new_path, new_dentry, flags);
if (error)
goto exit5;
rd.old_dir = old_path.dentry->d_inode;
rd.old_dentry = old_dentry;
rd.old_mnt_idmap = mnt_idmap(old_path.mnt);
rd.new_dir = new_path.dentry->d_inode;
rd.new_dentry = new_dentry;
rd.new_mnt_idmap = mnt_idmap(new_path.mnt);
rd.delegated_inode = &delegated_inode;
rd.flags = flags;
error = vfs_rename(&rd);
exit5:
dput(new_dentry);
exit4:
dput(old_dentry);
exit3:
unlock_rename(new_path.dentry, old_path.dentry);
exit_lock_rename:
if (delegated_inode) {
error = break_deleg_wait(&delegated_inode);
if (!error)
goto retry_deleg;
}
mnt_drop_write(old_path.mnt);
exit2:
if (retry_estale(error, lookup_flags))
should_retry = true;
path_put(&new_path);
exit1:
path_put(&old_path);
if (should_retry) {
should_retry = false;
lookup_flags |= LOOKUP_REVAL;
goto retry;
}
put_names:
putname(from);
putname(to);
return error;
}
SYSCALL_DEFINE5(renameat2, int, olddfd, const char __user *, oldname,
int, newdfd, const char __user *, newname, unsigned int, flags)
{
return do_renameat2(olddfd, getname(oldname), newdfd, getname(newname),
flags);
}
SYSCALL_DEFINE4(renameat, int, olddfd, const char __user *, oldname,
int, newdfd, const char __user *, newname)
{
return do_renameat2(olddfd, getname(oldname), newdfd, getname(newname),
0);
}
SYSCALL_DEFINE2(rename, const char __user *, oldname, const char __user *, newname)
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
{
return do_renameat2(AT_FDCWD, getname(oldname), AT_FDCWD,
getname(newname), 0);
[PATCH] vfs: *at functions: core Here is a series of patches which introduce in total 13 new system calls which take a file descriptor/filename pair instead of a single file name. These functions, openat etc, have been discussed on numerous occasions. They are needed to implement race-free filesystem traversal, they are necessary to implement a virtual per-thread current working directory (think multi-threaded backup software), etc. We have in glibc today implementations of the interfaces which use the /proc/self/fd magic. But this code is rather expensive. Here are some results (similar to what Jim Meyering posted before). The test creates a deep directory hierarchy on a tmpfs filesystem. Then rm -fr is used to remove all directories. Without syscall support I get this: real 0m31.921s user 0m0.688s sys 0m31.234s With syscall support the results are much better: real 0m20.699s user 0m0.536s sys 0m20.149s The interfaces are for obvious reasons currently not much used. But they'll be used. coreutils (and Jeff's posixutils) are already using them. Furthermore, code like ftw/fts in libc (maybe even glob) will also start using them. I expect a patch to make follow soon. Every program which is walking the filesystem tree will benefit. Signed-off-by: Ulrich Drepper <drepper@redhat.com> Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> Cc: Christoph Hellwig <hch@lst.de> Cc: Al Viro <viro@ftp.linux.org.uk> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Michael Kerrisk <mtk-manpages@gmx.net> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-01-19 09:43:53 +08:00
}
int readlink_copy(char __user *buffer, int buflen, const char *link)
{
int len = PTR_ERR(link);
if (IS_ERR(link))
goto out;
len = strlen(link);
if (len > (unsigned) buflen)
len = buflen;
if (copy_to_user(buffer, link, len))
len = -EFAULT;
out:
return len;
}
/**
* vfs_readlink - copy symlink body into userspace buffer
* @dentry: dentry on which to get symbolic link
* @buffer: user memory pointer
* @buflen: size of buffer
*
* Does not touch atime. That's up to the caller if necessary
*
* Does not call security hook.
*/
int vfs_readlink(struct dentry *dentry, char __user *buffer, int buflen)
{
struct inode *inode = d_inode(dentry);
DEFINE_DELAYED_CALL(done);
const char *link;
int res;
if (unlikely(!(inode->i_opflags & IOP_DEFAULT_READLINK))) {
if (unlikely(inode->i_op->readlink))
return inode->i_op->readlink(dentry, buffer, buflen);
if (!d_is_symlink(dentry))
return -EINVAL;
spin_lock(&inode->i_lock);
inode->i_opflags |= IOP_DEFAULT_READLINK;
spin_unlock(&inode->i_lock);
}
link = READ_ONCE(inode->i_link);
if (!link) {
link = inode->i_op->get_link(dentry, inode, &done);
if (IS_ERR(link))
return PTR_ERR(link);
}
res = readlink_copy(buffer, buflen, link);
do_delayed_call(&done);
return res;
}
EXPORT_SYMBOL(vfs_readlink);
/**
* vfs_get_link - get symlink body
* @dentry: dentry on which to get symbolic link
* @done: caller needs to free returned data with this
*
* Calls security hook and i_op->get_link() on the supplied inode.
*
* It does not touch atime. That's up to the caller if necessary.
*
* Does not work on "special" symlinks like /proc/$$/fd/N
*/
const char *vfs_get_link(struct dentry *dentry, struct delayed_call *done)
{
const char *res = ERR_PTR(-EINVAL);
struct inode *inode = d_inode(dentry);
if (d_is_symlink(dentry)) {
res = ERR_PTR(security_inode_readlink(dentry));
if (!res)
res = inode->i_op->get_link(dentry, inode, done);
}
return res;
}
EXPORT_SYMBOL(vfs_get_link);
/* get the link contents into pagecache */
const char *page_get_link(struct dentry *dentry, struct inode *inode,
struct delayed_call *callback)
{
char *kaddr;
struct page *page;
struct address_space *mapping = inode->i_mapping;
if (!dentry) {
page = find_get_page(mapping, 0);
if (!page)
return ERR_PTR(-ECHILD);
if (!PageUptodate(page)) {
put_page(page);
return ERR_PTR(-ECHILD);
}
} else {
page = read_mapping_page(mapping, 0, NULL);
if (IS_ERR(page))
return (char*)page;
}
set_delayed_call(callback, page_put_link, page);
BUG_ON(mapping_gfp_mask(mapping) & __GFP_HIGHMEM);
kaddr = page_address(page);
nd_terminate_link(kaddr, inode->i_size, PAGE_SIZE - 1);
return kaddr;
}
EXPORT_SYMBOL(page_get_link);
void page_put_link(void *arg)
{
put_page(arg);
}
EXPORT_SYMBOL(page_put_link);
int page_readlink(struct dentry *dentry, char __user *buffer, int buflen)
{
DEFINE_DELAYED_CALL(done);
int res = readlink_copy(buffer, buflen,
page_get_link(dentry, d_inode(dentry),
&done));
do_delayed_call(&done);
return res;
}
EXPORT_SYMBOL(page_readlink);
int page_symlink(struct inode *inode, const char *symname, int len)
{
struct address_space *mapping = inode->i_mapping;
const struct address_space_operations *aops = mapping->a_ops;
bool nofs = !mapping_gfp_constraint(mapping, __GFP_FS);
struct folio *folio;
mm: fs: initialize fsdata passed to write_begin/write_end interface Functions implementing the a_ops->write_end() interface accept the `void *fsdata` parameter that is supposed to be initialized by the corresponding a_ops->write_begin() (which accepts `void **fsdata`). However not all a_ops->write_begin() implementations initialize `fsdata` unconditionally, so it may get passed uninitialized to a_ops->write_end(), resulting in undefined behavior. Fix this by initializing fsdata with NULL before the call to write_begin(), rather than doing so in all possible a_ops implementations. This patch covers only the following cases found by running x86 KMSAN under syzkaller: - generic_perform_write() - cont_expand_zero() and generic_cont_expand_simple() - page_symlink() Other cases of passing uninitialized fsdata may persist in the codebase. Link: https://lkml.kernel.org/r/20220915150417.722975-43-glider@google.com Signed-off-by: Alexander Potapenko <glider@google.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Alexei Starovoitov <ast@kernel.org> Cc: Andrey Konovalov <andreyknvl@gmail.com> Cc: Andrey Konovalov <andreyknvl@google.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Christoph Hellwig <hch@lst.de> Cc: Christoph Lameter <cl@linux.com> Cc: David Rientjes <rientjes@google.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Eric Biggers <ebiggers@google.com> Cc: Eric Biggers <ebiggers@kernel.org> Cc: Eric Dumazet <edumazet@google.com> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Ilya Leoshkevich <iii@linux.ibm.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Jens Axboe <axboe@kernel.dk> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Kees Cook <keescook@chromium.org> Cc: Marco Elver <elver@google.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Matthew Wilcox <willy@infradead.org> Cc: Michael S. Tsirkin <mst@redhat.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Petr Mladek <pmladek@suse.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Vasily Gorbik <gor@linux.ibm.com> Cc: Vegard Nossum <vegard.nossum@oracle.com> Cc: Vlastimil Babka <vbabka@suse.cz> Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-15 23:04:16 +08:00
void *fsdata = NULL;
int err;
unsigned int flags;
retry:
if (nofs)
flags = memalloc_nofs_save();
err = aops->write_begin(NULL, mapping, 0, len-1, &folio, &fsdata);
if (nofs)
memalloc_nofs_restore(flags);
if (err)
goto fail;
memcpy(folio_address(folio), symname, len - 1);
err = aops->write_end(NULL, mapping, 0, len - 1, len - 1,
folio, fsdata);
if (err < 0)
goto fail;
if (err < len-1)
goto retry;
mark_inode_dirty(inode);
return 0;
fail:
return err;
}
EXPORT_SYMBOL(page_symlink);
const struct inode_operations page_symlink_inode_operations = {
.get_link = page_get_link,
};
EXPORT_SYMBOL(page_symlink_inode_operations);