Move the check for async request after check for the request being already
finished and done with.
Reported-by: syzbot+ae0bb7aae3de6b4594e2@syzkaller.appspotmail.com
Fixes: d49937749f ("fuse: stop copying args to fuse_req")
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Fix sparse warning:
fs/fuse/dev.c:468:6: warning: symbol 'fuse_args_to_req' was not declared. Should it be static?
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Fixes: 68583165f9 ("fuse: add pages to fuse_args")
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
This function has been made static, which now causes a compile-time
warning:
WARNING: "fuse_put_request" [vmlinux] is a static EXPORT_SYMBOL_GPL
Remove the unneeded export.
Fixes: 66abc3599c ("fuse: unexport request ops")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
As of now fuse_dev_alloc() both allocates a fuse device and installs it in
fuse_conn list. fuse_dev_alloc() can fail if fuse_device allocation fails.
virtio-fs needs to initialize multiple fuse devices (one per virtio queue).
It initializes one fuse device as part of call to fuse_fill_super_common()
and rest of the devices are allocated and installed after that.
But, we can't afford to fail after calling fuse_fill_super_common() as we
don't have a way to undo all the actions done by fuse_fill_super_common().
So to avoid failures after the call to fuse_fill_super_common(),
pre-allocate all fuse devices early and install them into fuse connection
later.
This patch provides two separate helpers for fuse device allocation and
fuse device installation in fuse_conn.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
The /dev/fuse device uses fiq->waitq and fasync to signal that requests are
available. These mechanisms do not apply to virtio-fs. This patch
introduces callbacks so alternative behavior can be used.
Note that queue_interrupt() changes along these lines:
spin_lock(&fiq->waitq.lock);
wake_up_locked(&fiq->waitq);
+ kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
spin_unlock(&fiq->waitq.lock);
- kill_fasync(&fiq->fasync, SIGIO, POLL_IN);
Since queue_request() and queue_forget() also call kill_fasync() inside
the spinlock this should be safe.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
File systems like virtio-fs need to do not have to play directly with
forget list data structures. There is a helper function use that instead.
Rename dequeue_forget() to fuse_dequeue_forget() and export it so that
stacked filesystems can use it.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
virtio-fs will need unique IDs for FORGET requests from outside
fs/fuse/dev.c. Make the symbol visible.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
virtio-fs will need to query the length of fuse_arg lists. Make the symbol
visible.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
virtio-fs will need to complete requests from outside fs/fuse/dev.c. Make
the symbol visible.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
The page array pointers are also duplicated across fuse_args_pages and
fuse_req. Get rid of the fuse_req ones.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
No need to duplicate the argument arrays in fuse_req, so just dereference
req->args instead of copying to the fuse_req internal ones.
This allows further cleanup of the fuse_req structure.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
All requests are now sent with one of the fuse_simple_... helpers. Get rid
of the old api from the fuse internal header.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Rename fuse_request_send_notify_reply() to fuse_simple_notify_reply() and
convert to passing fuse_args instead of fuse_req.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Derive fuse_writepage_args from fuse_io_args.
Sending the request is tricky since it was done with fi->lock held, hence
we must either use atomic allocation or release the lock. Both are
possible so try atomic first and if it fails, release the lock and do the
regular allocation with GFP_NOFS and __GFP_NOFAIL. Both flags are
necessary for correct operation.
Move the page realloc function from dev.c to file.c and convert to using
fuse_writepage_args.
The last caller of fuse_write_fill() is gone, so get rid of it.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Create a helper named fuse_simple_background() that is similar to
fuse_simple_request(). Unlike the latter, it returns immediately and calls
the supplied 'end' callback when the reply is received.
The supplied 'args' pointer is stored in 'fuse_req' which allows the
callback to interpret the output arguments decoded from the reply.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
fuse_simple_request() is converted to return length of last (instead of
single) out arg, since FUSE_IOCTL_OUT has two out args, the second of which
is variable length.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
fuse_req_pages_alloc() is moved to file.c, since its internal use by the
device code will eventually be removed.
Rename to fuse_pages_alloc() to signify that it's not only usable for
fuse_req page array.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Derive fuse_args_pages from fuse_args. This is used to handle requests
which use pages for input or output. The related flags are added to
fuse_args.
New FR_ALLOC_PAGES flags is added to indicate whether the page arrays in
fuse_req need to be freed by fuse_put_request() or not.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
In some cases it makes no sense to set pid/uid/gid fields in the request
header. Allow fuse_simple_background() to omit these. This is only
required in the "force" case, so for now just WARN if set otherwise.
Fold fuse_get_req_nofail_nopages() into its only caller. Comment is
obsolete anyway.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
This will be used by fuse_force_forget().
We can expand fuse_request_send() into fuse_simple_request(). The
FR_WAITING bit has already been set, no need to check.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Add 'force' to fuse_args and use fuse_get_req_nofail_nopages() to allocate
the request in that case.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Instead of complex games with a reserved request, just use __GFP_NOFAIL.
Both calers (flush, readdir) guarantee that connection was already
initialized, so no need to wait for fc->initialized.
Also remove unneeded clearing of FR_BACKGROUND flag.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
...to make future expansion simpler. The hiearachical structure is a
historical thing that does not serve any practical purpose.
The generated code is excatly the same before and after the patch.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
When IOCB_CMD_POLL is used on the FUSE device, aio_poll() disables IRQs
and takes kioctx::ctx_lock, then fuse_iqueue::waitq.lock.
This may have to wait for fuse_iqueue::waitq.lock to be released by one
of many places that take it with IRQs enabled. Since the IRQ handler
may take kioctx::ctx_lock, lockdep reports that a deadlock is possible.
Fix it by protecting the state of struct fuse_iqueue with a separate
spinlock, and only accessing fuse_iqueue::waitq using the versions of
the waitqueue functions which do IRQ-safe locking internally.
Reproducer:
#include <fcntl.h>
#include <stdio.h>
#include <sys/mount.h>
#include <sys/stat.h>
#include <sys/syscall.h>
#include <unistd.h>
#include <linux/aio_abi.h>
int main()
{
char opts[128];
int fd = open("/dev/fuse", O_RDWR);
aio_context_t ctx = 0;
struct iocb cb = { .aio_lio_opcode = IOCB_CMD_POLL, .aio_fildes = fd };
struct iocb *cbp = &cb;
sprintf(opts, "fd=%d,rootmode=040000,user_id=0,group_id=0", fd);
mkdir("mnt", 0700);
mount("foo", "mnt", "fuse", 0, opts);
syscall(__NR_io_setup, 1, &ctx);
syscall(__NR_io_submit, ctx, 1, &cbp);
}
Beginning of lockdep output:
=====================================================
WARNING: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected
5.3.0-rc5 #9 Not tainted
-----------------------------------------------------
syz_fuse/135 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
000000003590ceda (&fiq->waitq){+.+.}, at: spin_lock include/linux/spinlock.h:338 [inline]
000000003590ceda (&fiq->waitq){+.+.}, at: aio_poll fs/aio.c:1751 [inline]
000000003590ceda (&fiq->waitq){+.+.}, at: __io_submit_one.constprop.0+0x203/0x5b0 fs/aio.c:1825
and this task is already holding:
0000000075037284 (&(&ctx->ctx_lock)->rlock){..-.}, at: spin_lock_irq include/linux/spinlock.h:363 [inline]
0000000075037284 (&(&ctx->ctx_lock)->rlock){..-.}, at: aio_poll fs/aio.c:1749 [inline]
0000000075037284 (&(&ctx->ctx_lock)->rlock){..-.}, at: __io_submit_one.constprop.0+0x1f4/0x5b0 fs/aio.c:1825
which would create a new lock dependency:
(&(&ctx->ctx_lock)->rlock){..-.} -> (&fiq->waitq){+.+.}
but this new dependency connects a SOFTIRQ-irq-safe lock:
(&(&ctx->ctx_lock)->rlock){..-.}
[...]
Reported-by: syzbot+af05535bb79520f95431@syzkaller.appspotmail.com
Reported-by: syzbot+d86c4426a01f60feddc7@syzkaller.appspotmail.com
Fixes: bfe4037e72 ("aio: implement IOCB_CMD_POLL")
Cc: <stable@vger.kernel.org> # v4.19+
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
[ This retries commit d4b13963f2 ("fuse: require /dev/fuse reads to have
enough buffer capacity"), which was reverted. In this version we require
only `sizeof(fuse_in_header) + sizeof(fuse_write_in)` instead of 4K for
FUSE request header room, because, contrary to libfuse and kernel client
behaviour, GlusterFS actually provides only so much room for request
header. ]
A FUSE filesystem server queues /dev/fuse sys_read calls to get filesystem
requests to handle. It does not know in advance what would be that request
as it can be anything that client issues - LOOKUP, READ, WRITE, ... Many
requests are short and retrieve data from the filesystem. However WRITE and
NOTIFY_REPLY write data into filesystem.
Before getting into operation phase, FUSE filesystem server and kernel
client negotiate what should be the maximum write size the client will ever
issue. After negotiation the contract in between server/client is that the
filesystem server then should queue /dev/fuse sys_read calls with enough
buffer capacity to receive any client request - WRITE in particular, while
FUSE client should not, in particular, send WRITE requests with >
negotiated max_write payload. FUSE client in kernel and libfuse
historically reserve 4K for request header. However an existing filesystem
server - GlusterFS - was found which reserves only 80 bytes for header room
(= `sizeof(fuse_in_header) + sizeof(fuse_write_in)`).
Since
`sizeof(fuse_in_header) + sizeof(fuse_write_in)` ==
`sizeof(fuse_in_header) + sizeof(fuse_read_in)` ==
`sizeof(fuse_in_header) + sizeof(fuse_notify_retrieve_in)`
is the absolute minimum any sane filesystem should be using for header
room, the contract is that filesystem server should queue sys_reads with
`sizeof(fuse_in_header) + sizeof(fuse_write_in)` + max_write buffer.
If the filesystem server does not follow this contract, what can happen
is that fuse_dev_do_read will see that request size is > buffer size,
and then it will return EIO to client who issued the request but won't
indicate in any way that there is a problem to filesystem server.
This can be hard to diagnose because for some requests, e.g. for
NOTIFY_REPLY which mimics WRITE, there is no client thread that is
waiting for request completion and that EIO goes nowhere, while on
filesystem server side things look like the kernel is not replying back
after successful NOTIFY_RETRIEVE request made by the server.
We can make the problem easy to diagnose if we indicate via error return to
filesystem server when it is violating the contract. This should not
practically cause problems because if a filesystem server is using shorter
buffer, writes to it were already very likely to cause EIO, and if the
filesystem is read-only it should be too following FUSE_MIN_READ_BUFFER
minimum buffer size.
Please see [1] for context where the problem of stuck filesystem was hit
for real (because kernel client was incorrectly sending more than
max_write data with NOTIFY_REPLY; see also previous patch), how the
situation was traced and for more involving patch that did not make it
into the tree.
[1] https://marc.info/?l=linux-fsdevel&m=155057023600853&w=2
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Tested-by: Sander Eikelenboom <linux@eikelenboom.it>
Cc: Han-Wen Nienhuys <hanwen@google.com>
Cc: Jakob Unterwurzacher <jakobunt@gmail.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
This reverts commit d4b13963f2.
The commit introduced a regression in glusterfs-fuse.
Reported-by: Sander Eikelenboom <linux@eikelenboom.it>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
A FUSE filesystem server queues /dev/fuse sys_read calls to get
filesystem requests to handle. It does not know in advance what would be
that request as it can be anything that client issues - LOOKUP, READ,
WRITE, ... Many requests are short and retrieve data from the
filesystem. However WRITE and NOTIFY_REPLY write data into filesystem.
Before getting into operation phase, FUSE filesystem server and kernel
client negotiate what should be the maximum write size the client will
ever issue. After negotiation the contract in between server/client is
that the filesystem server then should queue /dev/fuse sys_read calls with
enough buffer capacity to receive any client request - WRITE in
particular, while FUSE client should not, in particular, send WRITE
requests with > negotiated max_write payload. FUSE client in kernel and
libfuse historically reserve 4K for request header. This way the
contract is that filesystem server should queue sys_reads with
4K+max_write buffer.
If the filesystem server does not follow this contract, what can happen
is that fuse_dev_do_read will see that request size is > buffer size,
and then it will return EIO to client who issued the request but won't
indicate in any way that there is a problem to filesystem server.
This can be hard to diagnose because for some requests, e.g. for
NOTIFY_REPLY which mimics WRITE, there is no client thread that is
waiting for request completion and that EIO goes nowhere, while on
filesystem server side things look like the kernel is not replying back
after successful NOTIFY_RETRIEVE request made by the server.
We can make the problem easy to diagnose if we indicate via error return to
filesystem server when it is violating the contract. This should not
practically cause problems because if a filesystem server is using shorter
buffer, writes to it were already very likely to cause EIO, and if the
filesystem is read-only it should be too following FUSE_MIN_READ_BUFFER
minimum buffer size.
Please see [1] for context where the problem of stuck filesystem was hit
for real (because kernel client was incorrectly sending more than
max_write data with NOTIFY_REPLY; see also previous patch), how the
situation was traced and for more involving patch that did not make it
into the tree.
[1] https://marc.info/?l=linux-fsdevel&m=155057023600853&w=2
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Cc: Han-Wen Nienhuys <hanwen@google.com>
Cc: Jakob Unterwurzacher <jakobunt@gmail.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
FUSE filesystem server and kernel client negotiate during initialization
phase, what should be the maximum write size the client will ever issue.
Correspondingly the filesystem server then queues sys_read calls to read
requests with buffer capacity large enough to carry request header + that
max_write bytes. A filesystem server is free to set its max_write in
anywhere in the range between [1*page, fc->max_pages*page]. In particular
go-fuse[2] sets max_write by default as 64K, wheres default fc->max_pages
corresponds to 128K. Libfuse also allows users to configure max_write, but
by default presets it to possible maximum.
If max_write is < fc->max_pages*page, and in NOTIFY_RETRIEVE handler we
allow to retrieve more than max_write bytes, corresponding prepared
NOTIFY_REPLY will be thrown away by fuse_dev_do_read, because the
filesystem server, in full correspondence with server/client contract, will
be only queuing sys_read with ~max_write buffer capacity, and
fuse_dev_do_read throws away requests that cannot fit into server request
buffer. In turn the filesystem server could get stuck waiting indefinitely
for NOTIFY_REPLY since NOTIFY_RETRIEVE handler returned OK which is
understood by clients as that NOTIFY_REPLY was queued and will be sent
back.
Cap requested size to negotiate max_write to avoid the problem. This
aligns with the way NOTIFY_RETRIEVE handler works, which already
unconditionally caps requested retrieve size to fuse_conn->max_pages. This
way it should not hurt NOTIFY_RETRIEVE semantic if we return less data than
was originally requested.
Please see [1] for context where the problem of stuck filesystem was hit
for real, how the situation was traced and for more involving patch that
did not make it into the tree.
[1] https://marc.info/?l=linux-fsdevel&m=155057023600853&w=2
[2] https://github.com/hanwen/go-fuse
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Cc: Han-Wen Nienhuys <hanwen@google.com>
Cc: Jakob Unterwurzacher <jakobunt@gmail.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Functions, like pr_err, are a more modern variant of printing compared to
printk. They could be used to denoise sources by using needed level in
the print function name, and by automatically inserting per-driver /
function / ... print prefix as defined by pr_fmt macro. pr_* are also
said to be used in Documentation/process/coding-style.rst and more
recent code - for example overlayfs - uses them instead of printk.
Convert CUSE and FUSE to use the new pr_* functions.
CUSE output stays completely unchanged, while FUSE output is amended a
bit for "trying to steal weird page" warning - the second line now comes
also with "fuse:" prefix. I hope it is ok.
Suggested-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Reviewed-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Merge page ref overflow branch.
Jann Horn reported that he can overflow the page ref count with
sufficient memory (and a filesystem that is intentionally extremely
slow).
Admittedly it's not exactly easy. To have more than four billion
references to a page requires a minimum of 32GB of kernel memory just
for the pointers to the pages, much less any metadata to keep track of
those pointers. Jann needed a total of 140GB of memory and a specially
crafted filesystem that leaves all reads pending (in order to not ever
free the page references and just keep adding more).
Still, we have a fairly straightforward way to limit the two obvious
user-controllable sources of page references: direct-IO like page
references gotten through get_user_pages(), and the splice pipe page
duplication. So let's just do that.
* branch page-refs:
fs: prevent page refcount overflow in pipe_buf_get
mm: prevent get_user_pages() from overflowing page refcount
mm: add 'try_get_page()' helper function
mm: make page ref count overflow check tighter and more explicit
Change pipe_buf_get() to return a bool indicating whether it succeeded
in raising the refcount of the page (if the thing in the pipe is a page).
This removes another mechanism for overflowing the page refcount. All
callers converted to handle a failure.
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Cc: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The only caller that needs fc->aborted set is fuse_conn_abort_write().
Setting fc->aborted is now racy (fuse_abort_conn() may already be in
progress or finished) but there's no reason to care.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
This is rather natural action after previous patches, and it just decreases
load of fc->lock.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
When queue_interrupt() is called from fuse_dev_do_write(), it came from
userspace directly. Userspace may pass any request id, even the request's
we have not interrupted (or even background's request). This patch adds
sanity check to make kernel safe against that.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Currently, we wait on req->waitq in request_wait_answer() function only,
and it's never used for background requests. Since wake_up() is not a
light-weight macros, instead of this, it unfolds in really called function,
which makes locking operations taking some cpu cycles, let's avoid its call
for the case we definitely know it's completely useless.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
We take global fiq->waitq.lock every time, when we are in this function,
but interrupted requests are just small subset of all requests. This patch
optimizes request_end() and makes it to take the lock when it's really
needed.
queue_interrupt() needs small change for that. After req is linked to
interrupt list, we do smp_mb() and check for FR_FINISHED again. In case of
FR_FINISHED bit has appeared, we remove req and leave the function:
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
We should sent signal only in case of interrupt is really queued. Not a
real problem, but this makes the code clearer and intuitive.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Some of the pipe_buf_release() handlers seem to assume that the pipe is
locked - in particular, anon_pipe_buf_release() accesses pipe->tmp_page
without taking any extra locks. From a glance through the callers of
pipe_buf_release(), it looks like FUSE is the only one that calls
pipe_buf_release() without having the pipe locked.
This bug should only lead to a memory leak, nothing terrible.
Fixes: dd3bb14f44 ("fuse: support splice() writing to fuse device")
Cc: stable@vger.kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
In current fuse_drop_waiting() implementation it's possible that
fuse_wait_aborted() will not be woken up in the unlikely case that
fuse_abort_conn() + fuse_wait_aborted() runs in between checking
fc->connected and calling atomic_dec(&fc->num_waiting).
Do the atomic_dec_and_test() unconditionally, which also provides the
necessary barrier against reordering with the fc->connected check.
The explicit smp_mb() in fuse_wait_aborted() is not actually needed, since
the spin_unlock() in fuse_abort_conn() provides the necessary RELEASE
barrier after resetting fc->connected. However, this is not a performance
sensitive path, and adding the explicit barrier makes it easier to
document.
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Fixes: b8f95e5d13 ("fuse: umount should wait for all requests")
Cc: <stable@vger.kernel.org> #v4.19
fuse_request_send_notify_reply() may fail if the connection was reset for
some reason (e.g. fs was unmounted). Don't leak request reference in this
case. Besides leaking memory, this resulted in fc->num_waiting not being
decremented and hence fuse_wait_aborted() left in a hanging and unkillable
state.
Fixes: 2d45ba381a ("fuse: add retrieve request")
Fixes: b8f95e5d13 ("fuse: umount should wait for all requests")
Reported-and-tested-by: syzbot+6339eda9cb4ebbc4c37b@syzkaller.appspotmail.com
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Cc: <stable@vger.kernel.org> #v2.6.36
Writeback caching currently allocates requests with the maximum number of
possible pages, while the actual number of pages per request depends on a
couple of factors that cannot be determined when the request is allocated
(whether page is already under writeback, whether page is contiguous with
previous pages already added to a request).
This patch allows such requests to start with no page allocation (all pages
inline) and grow the page array on demand.
If the max_pages tunable remains the default value, then this will mean
just one allocation that is the same size as before. If the tunable is
larger, then this adds at most 3 additional memory allocations (which is
generously compensated by the improved performance from the larger
request).
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Replace FUSE_MAX_PAGES_PER_REQ with the configurable parameter max_pages to
improve performance.
Old RFC with detailed description of the problem and many fixes by Mitsuo
Hayasaka (mitsuo.hayasaka.hu@hitachi.com):
- https://lkml.org/lkml/2012/7/5/136
We've encountered performance degradation and fixed it on a big and complex
virtual environment.
Environment to reproduce degradation and improvement:
1. Add lag to user mode FUSE
Add nanosleep(&(struct timespec){ 0, 1000 }, NULL); to xmp_write_buf in
passthrough_fh.c
2. patch UM fuse with configurable max_pages parameter. The patch will be
provided latter.
3. run test script and perform test on tmpfs
fuse_test()
{
cd /tmp
mkdir -p fusemnt
passthrough_fh -o max_pages=$1 /tmp/fusemnt
grep fuse /proc/self/mounts
dd conv=fdatasync oflag=dsync if=/dev/zero of=fusemnt/tmp/tmp \
count=1K bs=1M 2>&1 | grep -v records
rm fusemnt/tmp/tmp
killall passthrough_fh
}
Test results:
passthrough_fh /tmp/fusemnt fuse.passthrough_fh \
rw,nosuid,nodev,relatime,user_id=0,group_id=0 0 0
1073741824 bytes (1.1 GB) copied, 1.73867 s, 618 MB/s
passthrough_fh /tmp/fusemnt fuse.passthrough_fh \
rw,nosuid,nodev,relatime,user_id=0,group_id=0,max_pages=256 0 0
1073741824 bytes (1.1 GB) copied, 1.15643 s, 928 MB/s
Obviously with bigger lag the difference between 'before' and 'after'
will be more significant.
Mitsuo Hayasaka, in 2012 (https://lkml.org/lkml/2012/7/5/136),
observed improvement from 400-550 to 520-740.
Signed-off-by: Constantine Shulyupin <const@MakeLinux.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
When allocating page array for a request the array for the page pointers
and the array for page descriptors are allocated by two separate kmalloc()
calls. Merge these into one allocation.
Also instead of initializing the request and the page arrays with memset(),
use the zeroing allocation variants.
Reserved requests never carry pages (page array size is zero). Make that
explicit by initializing the page array pointers to NULL and make sure the
assumption remains true by adding a WARN_ON().
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
We noticed the performance bottleneck in FUSE running our Virtuozzo storage
over rdma. On some types of workload we observe 20% of times spent in
request_find() in profiler. This function is iterating over long requests
list, and it scales bad.
The patch introduces hash table to reduce the number of iterations, we do
in this function. Hash generating algorithm is taken from hash_add()
function, while 256 lines table is used to store pending requests. This
fixes problem and improves the performance.
Reported-by: Alexey Kuznetsov <kuznet@virtuozzo.com>
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
This field is not needed after the previous patch, since we can easily
convert request ID to interrupt request ID and vice versa.
Signed-off-by: Kirill Tkhai <ktkhai@virtuozzo.com>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
