The debug l3_ick/rate is not displaying the actual rate of the clock in
hardware. This is because, the core dpll set_rate function doesn't update the
clk.rate. After fixing, the l3_ick/rate is displaying proper values.
Signed-off-by: Shweta Gulati <shweta.gulati@ti.com>
Signed-off-by: Avinash.H.M <avinashhm@ti.com>
Cc: Rajendra Nayak <rnayak@ti.com>
Cc: Paul Wamsley <paul@pwsan.com>
Acked-by: Paul Walmsley <paul@pwsan.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
On rev <= 1.1 tables, the offset is absolute,
on newer tables, it's relative.
Fixes:
https://bugzilla.redhat.com/show_bug.cgi?id=700326
Signed-off-by: Alex Deucher <alexdeucher@gmail.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Cc: stable@kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
The location of MC_ARB_RAMCFG changed on fusion.
I've diffed all the other regs in evergreend.h and this
is the only other reg that changed.
Signed-off-by: Alex Deucher <alexdeucher@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
alg_data is already a pointer which must be passed directly.
Reported-by: Dieter Ripp <ripp@systecnet.com>
Signed-off-by: Wolfram Sang <w.sang@pengutronix.de>
Cc: Russell King <linux@arm.linux.org.uk>
Cc: Ben Dooks <ben-i2c@fluff.org>
Signed-off-by: Ben Dooks <ben-linux@fluff.org>
Since commit e9df2e8fd8 (Use appropriate sock tclass setting for
routing lookup) we lost ability to properly add ECN codemarks to ipv6
TCP frames.
It seems like TCP_ECN_send() calls INET_ECN_xmit(), which only sets the
ECN bit in the IPv4 ToS field (inet_sk(sk)->tos), but after the patch,
what's checked is inet6_sk(sk)->tclass, which is a completely different
field.
Close bug https://bugzilla.kernel.org/show_bug.cgi?id=34322
[Eric Dumazet] : added the INET_ECN_dontxmit() fix and replace macros
by inline functions for clarity.
Signed-off-by: Steinar H. Gunderson <sgunderson@bigfoot.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
kptr_restrict has been triggering bugs in apps such as perf, and it also makes
the system less useful by default, so turn it off by default.
This is how we generally handle security features that remove functionality,
such as firewall code or SELinux - they have to be configured and activated
from user-space.
Distributions can turn kptr_restrict on again via this line in
/etc/sysctrl.conf:
kernel.kptr_restrict = 1
( Also mark the variable __read_mostly while at it, as it's typically modified
only once per bootup, or not at all. )
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
When p9pdu_readf() is called with "s" attribute, it allocates a pointer that
will store a string. In p9dirent_read(), this pointer is not being released,
leading to out of memory errors.
This patch releases this pointer after string is copyed to dirent->d_name.
Signed-off-by: Pedro Scarapicchia Junior <pedro.scarapiccha@br.flextronics.com>
Signed-off-by: Eric Van Hensbergen <ericvh@gmail.com>
This is a fix for the SGI Altix-UV Broadcast Assist Unit code,
which is used for TLB flushing.
Certain hardware configurations (that customers are ordering)
cause nasids (numa address space id's) to be non-consecutive.
Specifically, once you have more than 4 blades in a IRU
(Individual Rack Unit - or 1/2 rack) but less than the maximum
of 16, the nasid numbering becomes non-consecutive. This
currently results in a 'catastrophic error' (CATERR) detected by
the firmware during OS boot. The BAU is generating an 'INTD'
request that is targeting a non-existent nasid value. Such
configurations may also occur when a blade is configured off
because of hardware errors. (There is one UV hub per blade.)
This patch is required to support such configurations.
The problem with the tlb_uv.c code is that is using the
consecutive hub numbers as indices to the BAU distribution bit
map. These are simply the ordinal position of the hub or blade
within its partition. It should be using physical node numbers
(pnodes), which correspond to the physical nasid values. Use of
the hub number only works as long as the nasids in the partition
are consecutive and increase with a stride of 1.
This patch changes the index to be the pnode number, thus
allowing nasids to be non-consecutive.
It also provides a table in local memory for each cpu to
translate target cpu number to target pnode and nasid.
And it improves naming to properly reflect 'node' and 'uvhub'
versus 'nasid'.
Signed-off-by: Cliff Wickman <cpw@sgi.com>
Cc: <stable@kernel.org>
Link: http://lkml.kernel.org/r/E1QJmxX-0002Mz-Fk@eag09.americas.sgi.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Changeset dcd39c9029 ("ne-h8300: convert to
net_device_ops") broke ne-h8300 by adding 8390.o to the link. That
meant that lib8390.c was included twice, once in ne-h8300.c and once in
8390.c, subject to different macros. This patch reverts that by
avoiding the wrappers in 8390.c.
Fix based on commits 217cbfa856 ("mac8390:
fix regression caused during net_device_ops conversion") and
4e0168fa48 ("mac8390: fix build with
NET_POLL_CONTROLLER").
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: stable@kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
Changeset 5618f0d119 ("hydra: convert to
net_device_ops") broke hydra by adding 8390.o to the link. That
meant that lib8390.c was included twice, once in hydra.c and once in
8390.c, subject to different macros. This patch reverts that by
avoiding the wrappers in 8390.c.
Fix based on commits 217cbfa856 ("mac8390:
fix regression caused during net_device_ops conversion") and
4e0168fa48 ("mac8390: fix build with
NET_POLL_CONTROLLER").
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: stable@kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
Changeset b6114794a1 ("zorro8390: convert to
net_device_ops") broke zorro8390 by adding 8390.o to the link. That
meant that lib8390.c was included twice, once in zorro8390.c and once in
8390.c, subject to different macros. This patch reverts that by
avoiding the wrappers in 8390.c.
Fix based on commits 217cbfa856 ("mac8390:
fix regression caused during net_device_ops conversion") and
4e0168fa48 ("mac8390: fix build with
NET_POLL_CONTROLLER").
Reported-by: Christian T. Steigies <cts@debian.org>
Suggested-by: Finn Thain <fthain@telegraphics.com.au>
Signed-off-by: Geert Uytterhoeven <geert@linux-m68k.org>
Tested-by: Christian T. Steigies <cts@debian.org>
Cc: stable@kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
The filename_trans rule processing has some printk(KERN_ERR ) messages
which were intended as debug aids in creating the code but weren't removed
before it was submitted. Remove them.
Reported-by: Paul Bolle <pebolle@tiscali.nl>
Signed-off-by: Eric Paris <eparis@redhat.com>
This reverts commit 49183b2818.
Quoth Franz Melchior:
"This patch introduces a bug on my infamous "Acer Travelmate
5735Z-452G32Mnss": when KMS takes over, the frame buffer contents get
completely garbled up on screen, with colored stripes and unreadable
text (photo on request). Only when X11 is started, the screen gets
restored again. Closing and re-opening the lid partly cures the
mess, too: it makes the font readable, though horizontally stretched."
Acked-by: Keith Packard <keithp@keithp.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
With CONFIG_DEBUG_SECTION_MISMATCH=y I see these warnings in next-20110415:
LD vmlinux.o
MODPOST vmlinux.o
WARNING: vmlinux.o(.text+0x1ba48): Section mismatch in reference from the function native_pagetable_reserve() to the function .init.text:memblock_x86_reserve_range()
The function native_pagetable_reserve() references
the function __init memblock_x86_reserve_range().
This is often because native_pagetable_reserve lacks a __init
annotation or the annotation of memblock_x86_reserve_range is wrong.
This patch fixes the issue.
Thanks to pipacs from PaX project for help on IRC.
Acked-by: "H. Peter Anvin" <hpa@zytor.com>
Signed-off-by: Sedat Dilek <sedat.dilek@gmail.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Introduce a new x86_init hook called pagetable_reserve that at the end
of init_memory_mapping is used to reserve a range of memory addresses for
the kernel pagetable pages we used and free the other ones.
On native it just calls memblock_x86_reserve_range while on xen it also
takes care of setting the spare memory previously allocated
for kernel pagetable pages from RO to RW, so that it can be used for
other purposes.
A detailed explanation of the reason why this hook is needed follows.
As a consequence of the commit:
commit 4b239f458c
Author: Yinghai Lu <yinghai@kernel.org>
Date: Fri Dec 17 16:58:28 2010 -0800
x86-64, mm: Put early page table high
at some point init_memory_mapping is going to reach the pagetable pages
area and map those pages too (mapping them as normal memory that falls
in the range of addresses passed to init_memory_mapping as argument).
Some of those pages are already pagetable pages (they are in the range
pgt_buf_start-pgt_buf_end) therefore they are going to be mapped RO and
everything is fine.
Some of these pages are not pagetable pages yet (they fall in the range
pgt_buf_end-pgt_buf_top; for example the page at pgt_buf_end) so they
are going to be mapped RW. When these pages become pagetable pages and
are hooked into the pagetable, xen will find that the guest has already
a RW mapping of them somewhere and fail the operation.
The reason Xen requires pagetables to be RO is that the hypervisor needs
to verify that the pagetables are valid before using them. The validation
operations are called "pinning" (more details in arch/x86/xen/mmu.c).
In order to fix the issue we mark all the pages in the entire range
pgt_buf_start-pgt_buf_top as RO, however when the pagetable allocation
is completed only the range pgt_buf_start-pgt_buf_end is reserved by
init_memory_mapping. Hence the kernel is going to crash as soon as one
of the pages in the range pgt_buf_end-pgt_buf_top is reused (b/c those
ranges are RO).
For this reason we need a hook to reserve the kernel pagetable pages we
used and free the other ones so that they can be reused for other
purposes.
On native it just means calling memblock_x86_reserve_range, on Xen it
also means marking RW the pagetable pages that we allocated before but
that haven't been used before.
Another way to fix this is without using the hook is by adding a 'if
(xen_pv_domain)' in the 'init_memory_mapping' code and calling the Xen
counterpart, but that is just nasty.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Acked-by: Yinghai Lu <yinghai@kernel.org>
Acked-by: H. Peter Anvin <hpa@zytor.com>
Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
* 'fixes' of master.kernel.org:/home/rmk/linux-2.6-arm:
ARM: 6870/1: The mandatory barrier rmb() must be a dsb() in for device accesses
ARM: 6892/1: handle ptrace requests to change PC during interrupted system calls
ARM: 6890/1: memmap: only free allocated memmap entries when using SPARSEMEM
ARM: zImage: the page table memory must be considered before relocation
ARM: zImage: make sure not to relocate on top of the relocation code
ARM: zImage: Fix bad SP address after relocating kernel
ARM: zImage: make sure the stack is 64-bit aligned
ARM: RiscPC: acornfb: fix section mismatches
ARM: RiscPC: etherh: fix section mismatches
read/write/ioctl on a fbcon file descriptor has traditionally used the
fbcon not when it was opened, but as it was at the time of the call.
That makes no sense, but the lack of sense is much more obvious now that
we properly ref-count the usage - it means that the ref-counting doesn't
actually protect operations we do on the frame buffer.
This changes it to look at the fb_info that we got at open time, but in
order to avoid using a frame buffer long after it has been unregistered,
we do verify that it is still current, and return -ENODEV if not.
Acked-by: Tim Gardner <tim.gardner@canonical.com>
Tested-by: Daniel J Blueman <daniel.blueman@gmail.com>
Tested-by: Anca Emanuel <anca.emanuel@gmail.com>
Cc: Bruno Prémont <bonbons@linux-vserver.org>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Dave Airlie <airlied@redhat.com>
Cc: Andy Whitcroft <andy.whitcroft@canonical.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This just adds the refcount and the new registration lock logic. It
does not (for example) actually change the read/write/ioctl routines to
actually use the frame buffer that was opened: those function still end
up alway susing whatever the current frame buffer is at the time of the
call.
Without this, if something holds the frame buffer open over a
framebuffer switch, the close() operation after the switch will access a
fb_info that has been free'd by the unregistering of the old frame
buffer.
(The read/write/ioctl operations will normally not cause problems,
because they will - illogically - pick up the new fbcon instead. But a
switch that happens just as one of those is going on might see problems
too, the window is just much smaller: one individual op rather than the
whole open-close sequence.)
This use-after-free is apparently fairly easily triggered by the Ubuntu
11.04 boot sequence.
Acked-by: Tim Gardner <tim.gardner@canonical.com>
Tested-by: Daniel J Blueman <daniel.blueman@gmail.com>
Tested-by: Anca Emanuel <anca.emanuel@gmail.com>
Cc: Bruno Prémont <bonbons@linux-vserver.org>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Cc: Paul Mundt <lethal@linux-sh.org>
Cc: Dave Airlie <airlied@redhat.com>
Cc: Andy Whitcroft <andy.whitcroft@canonical.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
We enabled write-combining for memory-mapped registers in commit
65f0b417de, but inhibited it for the
MCDI shared memory where this is not supported. However,
write-combining mappings also allow read-reordering, which may also
be a problem.
I found that when an SFC9000-family controller is connected to an
Intel 3000 chipset, and write-combining is enabled, the controller
stops responding to PCIe read requests during driver initialisation
while the driver is polling for completion of an MCDI command. This
results in an NMI and system hang. Adding read memory barriers
between all reads to the shared memory area appears to reduce but not
eliminate the probability of this.
We have not yet established whether this is a bug in our BIU or in the
PCIe bridge. For now, work around by mapping the shared memory area
separately.
Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Since mandatory barriers may be used (explicitly or implicitly via readl
etc.) to ensure the ordering between Device and Normal memory accesses,
a DMB is not enough. This patch converts it to a DSB.
Cc: Colin Cross <ccross@android.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
GDB's interrupt.exp test cases currenly fail on ARM. The problem is how do_signal
handled restarting interrupted system calls:
The entry.S assembler code determines that we come from a system call; and that
information is passed as "syscall" parameter to do_signal. That routine then
calls get_signal_to_deliver [*] and if a signal is to be delivered, calls into
handle_signal. If a system call is to be restarted either after the signal
handler returns, or if no handler is to be called in the first place, the PC
is updated after the get_signal_to_deliver call, either in handle_signal (if
we have a handler) or at the end of do_signal (otherwise).
Now the problem is that during [*], the call to get_signal_to_deliver, a ptrace
intercept may happen. During this intercept, the debugger may change registers,
including the PC. This is done by GDB if it wants to execute an "inferior call",
i.e. the execution of some code in the debugged program triggered by GDB.
To this purpose, GDB will save all registers, allocate a stack frame, set up
PC and arguments as appropriate for the call, and point the link register to
a dummy breakpoint instruction. Once the process is restarted, it will execute
the call and then trap back to the debugger, at which point GDB will restore
all registers and continue original execution.
This generally works fine. However, now consider what happens when GDB attempts
to do exactly that while the process was interrupted during execution of a to-be-
restarted system call: do_signal is called with the syscall flag set; it calls
get_signal_to_deliver, at which point the debugger takes over and changes the PC
to point to a completely different place. Now get_signal_to_deliver returns
without a signal to deliver; but now do_signal decides it should be restarting
a system call, and decrements the PC by 2 or 4 -- so it now points to 2 or 4
bytes before the function GDB wants to call -- which leads to a subsequent crash.
To fix this problem, two things need to be supported:
- do_signal must be able to recognize that get_signal_to_deliver changed the PC
to a different location, and skip the restart-syscall sequence
- once the debugger has restored all registers at the end of the inferior call
sequence, do_signal must recognize that *now* it needs to restart the pending
system call, even though it was now entered from a breakpoint instead of an
actual svc instruction
This set of issues is solved on other platforms, usually by one of two
mechanisms:
- The status information "do_signal is handling a system call that may need
restarting" is itself carried in some register that can be accessed via
ptrace. This is e.g. on Intel the "orig_eax" register; on Sparc the kernel
defines a magic extra bit in the flags register for this purpose.
This allows GDB to manage that state: reset it when doing an inferior call,
and restore it after the call is finished.
- On s390, do_signal transparently handles this problem without requiring
GDB interaction, by performing system call restarting in the following
way: first, adjust the PC as necessary for restarting the call. Then,
call get_signal_to_deliver; and finally just continue execution at the
PC. This way, if GDB does not change the PC, everything is as before.
If GDB *does* change the PC, execution will simply continue there --
and once GDB restores the PC it saved at that point, it will automatically
point to the *restarted* system call. (There is the minor twist how to
handle system calls that do *not* need restarting -- do_signal will undo
the PC change in this case, after get_signal_to_deliver has returned, and
only if ptrace did not change the PC during that call.)
Because there does not appear to be any obvious register to carry the
syscall-restart information on ARM, we'd either have to introduce a new
artificial ptrace register just for that purpose, or else handle the issue
transparently like on s390. The patch below implements the second option;
using this patch makes the interrupt.exp test cases pass on ARM, with no
regression in the GDB test suite otherwise.
Cc: patches@linaro.org
Signed-off-by: Ulrich Weigand <ulrich.weigand@linaro.org>
Signed-off-by: Arnd Bergmann <arnd.bergmann@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The SPARSEMEM code allocates memmap entries only for sections which are
present (i.e. those which contain some valid memory). The membank checks
in free_unused_memmap do not take this into account and can incorrectly
attempt to free memory which is not allocated, resulting in a BUG() in
the bootmem code.
However, if memory is configured as follows:
|<----section---->|<----hole---->|<----section---->|
+--------+--------+--------------+--------+--------+
| bank 0 | unused | | bank 1 | unused |
+--------+--------+--------------+--------+--------+
where a bank only occupies part of a section, the memmap allocated for
the remainder of the section *can* be freed.
This patch modifies the checks in free_unused_memmap so that only valid
memmap entries are considered for removal.
Acked-by: Catalin Marinas <catalin.marinas@arm.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
When we are in the label cc_dword_align, registers %o0 and %o1 have the same last 2 bits,
but it's not guaranteed one of them is zero. So we can get unaligned memory access
in label ccte. Example of parameters which lead to this:
%o0=0x7ff183e9, %o1=0x8e709e7d, %g1=3
With the parameters I had a memory corruption, when the additional 5 bytes were rewritten.
This patch corrects the error.
One comment to the patch. We don't care about the third bit in %o1, because cc_end_cruft
stores word or less.
Signed-off-by: Tkhai Kirill <tkhai@yandex.ru>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, writebacks may end up recursing back into the filesystem due to
GFP_KERNEL direct reclaims in the pnfs subsystem.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
ceph: do not use i_wrbuffer_ref as refcount for Fb cap
ceph: fix list_add in ceph_put_snap_realm
ceph: print debug message before put mds session
This reverts commit f21ca5fff6.
Quoth Gustavo F. Padovan:
"Commit f21ca5fff6 can cause a NULL
dereference if we call shutdown in a bluetooth SCO socket and doesn't
wait the shutdown completion to call close(). Please revert it. I
may have a fix for it soon, but we don't have time anymore, so revert
is the way to go. ;)"
Requested-by: Gustavo F. Padovan <padovan@profusion.mobi>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/gfp.h and include/trace/events/gfpflags.h are out of sync.
When tracing is enabled, certain flags are not recognised and the text
output is less useful as a result. Add the missing flags.
Signed-off-by: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Testing the shmem_swaplist replacements for igrab() revealed another bug:
writes to /dev/loop0 on a tmpfs file which fills its filesystem were
sometimes failing with "Buffer I/O error"s.
These came from ENOSPC failures of shmem_getpage(), when racing with
swapoff: the same could happen when racing with another shmem_getpage(),
pulling the page in from swap in between our find_lock_page() and our
taking the info->lock (though not in the single-threaded loop case).
This is unacceptable, and surprising that I've not noticed it before:
it dates back many years, but (presumably) was made a lot easier to
reproduce in 2.6.36, which sited a page preallocation in the race window.
Fix it by rechecking the page cache before settling on an ENOSPC error.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The use of igrab() in swapoff's shmem_unuse_inode() is just as vulnerable
to umount as that in shmem_writepage().
Fix this instance by extending the protection of shmem_swaplist_mutex
right across shmem_unuse_inode(): while it's on the list, the inode cannot
be evicted (and the filesystem cannot be unmounted) without
shmem_evict_inode() taking that mutex to remove it from the list.
But since shmem_writepage() might take that mutex, we should avoid making
memory allocations or memcg charges while holding it: prepare them at the
outer level in shmem_unuse(). When mem_cgroup_cache_charge() was
originally placed, we didn't know until that point that the page from swap
was actually a shmem page; but nowadays it's noted in the swap_map, so
we're safe to charge upfront. For the radix_tree, do as is done in
shmem_getpage(): preload upfront, but don't pin to the cpu; so we make a
habit of refreshing the node pool, but might dip into GFP_NOWAIT reserves
on occasion if subsequently preempted.
With the allocation and charge moved out from shmem_unuse_inode(),
we can also hold index map and info->lock over from finding the entry.
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Konstanin Khlebnikov reports that a dangerous race between umount and
shmem_writepage can be reproduced by this script:
for i in {1..300} ; do
mkdir $i
while true ; do
mount -t tmpfs none $i
dd if=/dev/zero of=$i/test bs=1M count=$(($RANDOM % 100))
umount $i
done &
done
on a 6xCPU node with 8Gb RAM: kernel very unstable after this accident. =)
Kernel log:
VFS: Busy inodes after unmount of tmpfs.
Self-destruct in 5 seconds. Have a nice day...
WARNING: at lib/list_debug.c:53 __list_del_entry+0x8d/0x98()
list_del corruption. prev->next should be ffff880222fdaac8, but was (null)
Pid: 11222, comm: mount.tmpfs Not tainted 2.6.39-rc2+ #4
Call Trace:
warn_slowpath_common+0x80/0x98
warn_slowpath_fmt+0x41/0x43
__list_del_entry+0x8d/0x98
evict+0x50/0x113
iput+0x138/0x141
...
BUG: unable to handle kernel paging request at ffffffffffffffff
IP: shmem_free_blocks+0x18/0x4c
Pid: 10422, comm: dd Tainted: G W 2.6.39-rc2+ #4
Call Trace:
shmem_recalc_inode+0x61/0x66
shmem_writepage+0xba/0x1dc
pageout+0x13c/0x24c
shrink_page_list+0x28e/0x4be
shrink_inactive_list+0x21f/0x382
...
shmem_writepage() calls igrab() on the inode for the page which came from
page reclaim, to add it later into shmem_swaplist for swapoff operation.
This igrab() can race with super-block deactivating process:
shrink_inactive_list() deactivate_super()
pageout() tmpfs_fs_type->kill_sb()
shmem_writepage() kill_litter_super()
generic_shutdown_super()
evict_inodes()
igrab()
atomic_read(&inode->i_count)
skip-inode
iput()
if (!list_empty(&sb->s_inodes))
printk("VFS: Busy inodes after...
This igrap-iput pair was added in commit 1b1b32f2c6 "tmpfs: fix
shmem_swaplist races" based on incorrect assumptions: igrab() protects the
inode from concurrent eviction by deletion, but it does nothing to protect
it from concurrent unmounting, which goes ahead despite the raised
i_count.
So this use of igrab() was wrong all along, but the race made much worse
in 2.6.37 when commit 63997e98a3 "split invalidate_inodes()" replaced
two attempts at invalidate_inodes() by a single evict_inodes().
Konstantin posted a plausible patch, raising sb->s_active too: I'm unsure
whether it was correct or not; but burnt once by igrab(), I am sure that
we don't want to rely more deeply upon externals here.
Fix it by adding the inode to shmem_swaplist earlier, while the page lock
on page in page cache still secures the inode against eviction, without
artifically raising i_count. It was originally added later because
shmem_unuse_inode() is liable to remove an inode from the list while it's
unswapped; but we can guard against that by taking spinlock before
dropping mutex.
Reported-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Tested-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
