Searched +hist:7 +hist:f7f25e8 (Results 1 - 4 of 4) sorted by relevance

/linux-master/fs/
H A Dfile_table.cdiff 7cb537b6 Sat Nov 25 19:08:34 MST 2023 Al Viro <viro@zeniv.linux.org.uk> file: massage cleanup of files that failed to open

A file that has never gotten FMODE_OPENED will never have RCU-accessed
references, its final fput() is equivalent to file_free() and if it
doesn't have FMODE_BACKING either, it can be done from any context and
won't need task_work treatment.

Now that we have SLAB_TYPESAFE_BY_RCU we can simplify this and have
other callers benefit. All of that can be achieved easier is to make
fput() recoginze that case and call file_free() directly.

No need to introduce a special primitive for that. It also allowed
things like failing dentry_open() could benefit from that as well.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
[Christian Brauner <brauner@kernel.org>: massage commit message]
Link: https://lore.kernel.org/r/20231126020834.GC38156@ZenIV
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 021a160a Tue Aug 08 11:26:35 MDT 2023 Linus Torvalds <torvalds@linux-foundation.org> fs: use __fput_sync in close(2)

close(2) is a special case which guarantees a shallow kernel stack,
making delegation to task_work machinery unnecessary. Said delegation is
problematic as it involves atomic ops and interrupt masking trips, none
of which are cheap on x86-64. Forcing close(2) to do it looks like an
oversight in the original work.

Moreover presence of CONFIG_RSEQ adds an additional overhead as fput()
-> task_work_add(..., TWA_RESUME) -> set_notify_resume() makes the
thread returning to userspace land in resume_user_mode_work(), where
rseq_handle_notify_resume takes a SMAP round-trip if rseq is enabled for
the thread (and it is by default with contemporary glibc).

Sample result when benchmarking open1_processes -t 1 from will-it-scale
(that's an open + close loop) + tmpfs on /tmp, running on the Sapphire
Rapid CPU (ops/s):
stock+RSEQ: 1329857
stock-RSEQ: 1421667 (+7%)
patched: 1523521 (+14.5% / +7%) (with / without rseq)

Patched result is the same regardless of rseq as the codepath is avoided.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 021a160a Tue Aug 08 11:26:35 MDT 2023 Linus Torvalds <torvalds@linux-foundation.org> fs: use __fput_sync in close(2)

close(2) is a special case which guarantees a shallow kernel stack,
making delegation to task_work machinery unnecessary. Said delegation is
problematic as it involves atomic ops and interrupt masking trips, none
of which are cheap on x86-64. Forcing close(2) to do it looks like an
oversight in the original work.

Moreover presence of CONFIG_RSEQ adds an additional overhead as fput()
-> task_work_add(..., TWA_RESUME) -> set_notify_resume() makes the
thread returning to userspace land in resume_user_mode_work(), where
rseq_handle_notify_resume takes a SMAP round-trip if rseq is enabled for
the thread (and it is by default with contemporary glibc).

Sample result when benchmarking open1_processes -t 1 from will-it-scale
(that's an open + close loop) + tmpfs on /tmp, running on the Sapphire
Rapid CPU (ops/s):
stock+RSEQ: 1329857
stock-RSEQ: 1421667 (+7%)
patched: 1523521 (+14.5% / +7%) (with / without rseq)

Patched result is the same regardless of rseq as the codepath is avoided.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff e9c15bad Mon Jun 15 06:13:58 MDT 2020 Mel Gorman <mgorman@techsingularity.net> fs: Do not check if there is a fsnotify watcher on pseudo inodes

The kernel uses internal mounts created by kern_mount() and populated
with files with no lookup path by alloc_file_pseudo() for a variety of
reasons. An example of such a mount is for anonymous pipes. For pipes,
every vfs_write() regardless of filesystem, calls fsnotify_modify()
to notify of any changes which incurs a small amount of overhead in
fsnotify even when there are no watchers. It can also trigger for reads
and readv and writev, it was simply vfs_write() that was noticed first.

A patch is pending that reduces, but does not eliminate, the overhead of
fsnotify but for files that cannot be looked up via a path, even that
small overhead is unnecessary. The user API for all notification
subsystems (inotify, fanotify, ...) is based on the pathname and a dirfd
and proc entries appear to be the only visible representation of the
files. Proc does not have the same pathname as the internal entry and
the proc inode is not the same as the internal inode so even if fanotify
is used on a file under /proc/XX/fd, no useful events are notified.

This patch changes alloc_file_pseudo() to always opt out of fsnotify by
setting FMODE_NONOTIFY flag so that no check is made for fsnotify
watchers on pseudo files. This should be safe as the underlying helper
for the dentry is d_alloc_pseudo() which explicitly states that no
lookups are ever performed meaning that fanotify should have nothing
useful to attach to.

The test motivating this was "perf bench sched messaging --pipe". On
a single-socket machine using threads the difference of the patch was
as follows.

5.7.0 5.7.0
vanilla nofsnotify-v1r1
Amean 1 1.3837 ( 0.00%) 1.3547 ( 2.10%)
Amean 3 3.7360 ( 0.00%) 3.6543 ( 2.19%)
Amean 5 5.8130 ( 0.00%) 5.7233 * 1.54%*
Amean 7 8.1490 ( 0.00%) 7.9730 * 2.16%*
Amean 12 14.6843 ( 0.00%) 14.1820 ( 3.42%)
Amean 18 21.8840 ( 0.00%) 21.7460 ( 0.63%)
Amean 24 28.8697 ( 0.00%) 29.1680 ( -1.03%)
Amean 30 36.0787 ( 0.00%) 35.2640 * 2.26%*
Amean 32 38.0527 ( 0.00%) 38.1223 ( -0.18%)

The difference is small but in some cases it's outside the noise so
while marginal, there is still some small benefit to ignoring fsnotify
for files allocated via alloc_file_pseudo() in some cases.

Link: https://lore.kernel.org/r/20200615121358.GF3183@techsingularity.net
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Reviewed-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
diff 7f7f25e8 Tue Feb 11 15:49:24 MST 2014 Al Viro <viro@zeniv.linux.org.uk> replace checking for ->read/->aio_read presence with check in ->f_mode

Since we are about to introduce new methods (read_iter/write_iter), the
tests in a bunch of places would have to grow inconveniently. Check
once (at open() time) and store results in ->f_mode as FMODE_CAN_READ
and FMODE_CAN_WRITE resp. It might end up being a temporary measure -
once everything switches from ->aio_{read,write} to ->{read,write}_iter
it might make sense to return to open-coded checks. We'll see...

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 7f7f25e8 Tue Feb 11 15:49:24 MST 2014 Al Viro <viro@zeniv.linux.org.uk> replace checking for ->read/->aio_read presence with check in ->f_mode

Since we are about to introduce new methods (read_iter/write_iter), the
tests in a bunch of places would have to grow inconveniently. Check
once (at open() time) and store results in ->f_mode as FMODE_CAN_READ
and FMODE_CAN_WRITE resp. It might end up being a temporary measure -
once everything switches from ->aio_{read,write} to ->{read,write}_iter
it might make sense to return to open-coded checks. We'll see...

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 7f4b36f9 Thu Mar 13 22:45:29 MDT 2014 Al Viro <viro@zeniv.linux.org.uk> get rid of files_defer_init()

the only thing it's doing these days is calculation of
upper limit for fs.nr_open sysctl and that can be done
statically

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 9c225f26 Mon Mar 03 10:36:58 MST 2014 Linus Torvalds <torvalds@linux-foundation.org> vfs: atomic f_pos accesses as per POSIX

Our write() system call has always been atomic in the sense that you get
the expected thread-safe contiguous write, but we haven't actually
guaranteed that concurrent writes are serialized wrt f_pos accesses, so
threads (or processes) that share a file descriptor and use "write()"
concurrently would quite likely overwrite each others data.

This violates POSIX.1-2008/SUSv4 Section XSI 2.9.7 that says:

"2.9.7 Thread Interactions with Regular File Operations

All of the following functions shall be atomic with respect to each
other in the effects specified in POSIX.1-2008 when they operate on
regular files or symbolic links: [...]"

and one of the effects is the file position update.

This unprotected file position behavior is not new behavior, and nobody
has ever cared. Until now. Yongzhi Pan reported unexpected behavior to
Michael Kerrisk that was due to this.

This resolves the issue with a f_pos-specific lock that is taken by
read/write/lseek on file descriptors that may be shared across threads
or processes.

Reported-by: Yongzhi Pan <panyongzhi@gmail.com>
Reported-by: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 9c225f26 Mon Mar 03 10:36:58 MST 2014 Linus Torvalds <torvalds@linux-foundation.org> vfs: atomic f_pos accesses as per POSIX

Our write() system call has always been atomic in the sense that you get
the expected thread-safe contiguous write, but we haven't actually
guaranteed that concurrent writes are serialized wrt f_pos accesses, so
threads (or processes) that share a file descriptor and use "write()"
concurrently would quite likely overwrite each others data.

This violates POSIX.1-2008/SUSv4 Section XSI 2.9.7 that says:

"2.9.7 Thread Interactions with Regular File Operations

All of the following functions shall be atomic with respect to each
other in the effects specified in POSIX.1-2008 when they operate on
regular files or symbolic links: [...]"

and one of the effects is the file position update.

This unprotected file position behavior is not new behavior, and nobody
has ever cared. Until now. Yongzhi Pan reported unexpected behavior to
Michael Kerrisk that was due to this.

This resolves the issue with a f_pos-specific lock that is taken by
read/write/lseek on file descriptors that may be shared across threads
or processes.

Reported-by: Yongzhi Pan <panyongzhi@gmail.com>
Reported-by: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 7e360c38 Tue Oct 05 01:32:55 MDT 2010 Eric Dumazet <eric.dumazet@gmail.com> fs: allow for more than 2^31 files

Andrew,

Could you please review this patch, you probably are the right guy to
take it, because it crosses fs and net trees.

Note : /proc/sys/fs/file-nr is a read-only file, so this patch doesnt
depend on previous patch (sysctl: fix min/max handling in
__do_proc_doulongvec_minmax())

Thanks !

[PATCH V4] fs: allow for more than 2^31 files

Robin Holt tried to boot a 16TB system and found af_unix was overflowing
a 32bit value :

<quote>

We were seeing a failure which prevented boot. The kernel was incapable
of creating either a named pipe or unix domain socket. This comes down
to a common kernel function called unix_create1() which does:

atomic_inc(&unix_nr_socks);
if (atomic_read(&unix_nr_socks) > 2 * get_max_files())
goto out;

The function get_max_files() is a simple return of files_stat.max_files.
files_stat.max_files is a signed integer and is computed in
fs/file_table.c's files_init().

n = (mempages * (PAGE_SIZE / 1024)) / 10;
files_stat.max_files = n;

In our case, mempages (total_ram_pages) is approx 3,758,096,384
(0xe0000000). That leaves max_files at approximately 1,503,238,553.
This causes 2 * get_max_files() to integer overflow.

</quote>

Fix is to let /proc/sys/fs/file-nr & /proc/sys/fs/file-max use long
integers, and change af_unix to use an atomic_long_t instead of
atomic_t.

get_max_files() is changed to return an unsigned long.
get_nr_files() is changed to return a long.

unix_nr_socks is changed from atomic_t to atomic_long_t, while not
strictly needed to address Robin problem.

Before patch (on a 64bit kernel) :
# echo 2147483648 >/proc/sys/fs/file-max
# cat /proc/sys/fs/file-max
-18446744071562067968

After patch:
# echo 2147483648 >/proc/sys/fs/file-max
# cat /proc/sys/fs/file-max
2147483648
# cat /proc/sys/fs/file-nr
704 0 2147483648

Reported-by: Robin Holt <holt@sgi.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: David Miller <davem@davemloft.net>
Reviewed-by: Robin Holt <holt@sgi.com>
Tested-by: Robin Holt <holt@sgi.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
H A Dread_write.cdiff bdeb77bc Sat Jul 16 22:37:10 MDT 2022 Andrei Vagin <avagin@gmail.com> fs: sendfile handles O_NONBLOCK of out_fd

sendfile has to return EAGAIN if out_fd is nonblocking and the write into
it would block.

Here is a small reproducer for the problem:

#define _GNU_SOURCE /* See feature_test_macros(7) */
#include <fcntl.h>
#include <stdio.h>
#include <unistd.h>
#include <errno.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/sendfile.h>


#define FILE_SIZE (1UL << 30)
int main(int argc, char **argv) {
int p[2], fd;

if (pipe2(p, O_NONBLOCK))
return 1;

fd = open(argv[1], O_RDWR | O_TMPFILE, 0666);
if (fd < 0)
return 1;
ftruncate(fd, FILE_SIZE);

if (sendfile(p[1], fd, 0, FILE_SIZE) == -1) {
fprintf(stderr, "FAIL\n");
}
if (sendfile(p[1], fd, 0, FILE_SIZE) != -1 || errno != EAGAIN) {
fprintf(stderr, "FAIL\n");
}
return 0;
}

It worked before b964bf53e540, it is stuck after b964bf53e540, and it
works again with this fix.

This regression occurred because do_splice_direct() calls pipe_write
that handles O_NONBLOCK. Here is a trace log from the reproducer:

1) | __x64_sys_sendfile64() {
1) | do_sendfile() {
1) | __fdget()
1) | rw_verify_area()
1) | __fdget()
1) | rw_verify_area()
1) | do_splice_direct() {
1) | rw_verify_area()
1) | splice_direct_to_actor() {
1) | do_splice_to() {
1) | rw_verify_area()
1) | generic_file_splice_read()
1) + 74.153 us | }
1) | direct_splice_actor() {
1) | iter_file_splice_write() {
1) | __kmalloc()
1) 0.148 us | pipe_lock();
1) 0.153 us | splice_from_pipe_next.part.0();
1) 0.162 us | page_cache_pipe_buf_confirm();
... 16 times
1) 0.159 us | page_cache_pipe_buf_confirm();
1) | vfs_iter_write() {
1) | do_iter_write() {
1) | rw_verify_area()
1) | do_iter_readv_writev() {
1) | pipe_write() {
1) | mutex_lock()
1) 0.153 us | mutex_unlock();
1) 1.368 us | }
1) 1.686 us | }
1) 5.798 us | }
1) 6.084 us | }
1) 0.174 us | kfree();
1) 0.152 us | pipe_unlock();
1) + 14.461 us | }
1) + 14.783 us | }
1) 0.164 us | page_cache_pipe_buf_release();
... 16 times
1) 0.161 us | page_cache_pipe_buf_release();
1) | touch_atime()
1) + 95.854 us | }
1) + 99.784 us | }
1) ! 107.393 us | }
1) ! 107.699 us | }

Link: https://lkml.kernel.org/r/20220415005015.525191-1-avagin@gmail.com
Fixes: b964bf53e540 ("teach sendfile(2) to handle send-to-pipe directly")
Signed-off-by: Andrei Vagin <avagin@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
diff 7b84b665 Fri Oct 02 20:55:23 MDT 2020 Matthew Wilcox (Oracle) <willy@infradead.org> fs: Allow a NULL pos pointer to __kernel_read

Match the behaviour of new_sync_read() and __kernel_write().

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 10dce8af Tue Mar 26 16:20:43 MDT 2019 Kirill Smelkov <kirr@nexedi.com> fs: stream_open - opener for stream-like files so that read and write can run simultaneously without deadlock

Commit 9c225f2655e3 ("vfs: atomic f_pos accesses as per POSIX") added
locking for file.f_pos access and in particular made concurrent read and
write not possible - now both those functions take f_pos lock for the
whole run, and so if e.g. a read is blocked waiting for data, write will
deadlock waiting for that read to complete.

This caused regression for stream-like files where previously read and
write could run simultaneously, but after that patch could not do so
anymore. See e.g. commit 581d21a2d02a ("xenbus: fix deadlock on writes
to /proc/xen/xenbus") which fixes such regression for particular case of
/proc/xen/xenbus.

The patch that added f_pos lock in 2014 did so to guarantee POSIX thread
safety for read/write/lseek and added the locking to file descriptors of
all regular files. In 2014 that thread-safety problem was not new as it
was already discussed earlier in 2006.

However even though 2006'th version of Linus's patch was adding f_pos
locking "only for files that are marked seekable with FMODE_LSEEK (thus
avoiding the stream-like objects like pipes and sockets)", the 2014
version - the one that actually made it into the tree as 9c225f2655e3 -
is doing so irregardless of whether a file is seekable or not.

See

https://lore.kernel.org/lkml/53022DB1.4070805@gmail.com/
https://lwn.net/Articles/180387
https://lwn.net/Articles/180396

for historic context.

The reason that it did so is, probably, that there are many files that
are marked non-seekable, but e.g. their read implementation actually
depends on knowing current position to correctly handle the read. Some
examples:

kernel/power/user.c snapshot_read
fs/debugfs/file.c u32_array_read
fs/fuse/control.c fuse_conn_waiting_read + ...
drivers/hwmon/asus_atk0110.c atk_debugfs_ggrp_read
arch/s390/hypfs/inode.c hypfs_read_iter
...

Despite that, many nonseekable_open users implement read and write with
pure stream semantics - they don't depend on passed ppos at all. And for
those cases where read could wait for something inside, it creates a
situation similar to xenbus - the write could be never made to go until
read is done, and read is waiting for some, potentially external, event,
for potentially unbounded time -> deadlock.

Besides xenbus, there are 14 such places in the kernel that I've found
with semantic patch (see below):

drivers/xen/evtchn.c:667:8-24: ERROR: evtchn_fops: .read() can deadlock .write()
drivers/isdn/capi/capi.c:963:8-24: ERROR: capi_fops: .read() can deadlock .write()
drivers/input/evdev.c:527:1-17: ERROR: evdev_fops: .read() can deadlock .write()
drivers/char/pcmcia/cm4000_cs.c:1685:7-23: ERROR: cm4000_fops: .read() can deadlock .write()
net/rfkill/core.c:1146:8-24: ERROR: rfkill_fops: .read() can deadlock .write()
drivers/s390/char/fs3270.c:488:1-17: ERROR: fs3270_fops: .read() can deadlock .write()
drivers/usb/misc/ldusb.c:310:1-17: ERROR: ld_usb_fops: .read() can deadlock .write()
drivers/hid/uhid.c:635:1-17: ERROR: uhid_fops: .read() can deadlock .write()
net/batman-adv/icmp_socket.c:80:1-17: ERROR: batadv_fops: .read() can deadlock .write()
drivers/media/rc/lirc_dev.c:198:1-17: ERROR: lirc_fops: .read() can deadlock .write()
drivers/leds/uleds.c:77:1-17: ERROR: uleds_fops: .read() can deadlock .write()
drivers/input/misc/uinput.c:400:1-17: ERROR: uinput_fops: .read() can deadlock .write()
drivers/infiniband/core/user_mad.c:985:7-23: ERROR: umad_fops: .read() can deadlock .write()
drivers/gnss/core.c:45:1-17: ERROR: gnss_fops: .read() can deadlock .write()

In addition to the cases above another regression caused by f_pos
locking is that now FUSE filesystems that implement open with
FOPEN_NONSEEKABLE flag, can no longer implement bidirectional
stream-like files - for the same reason as above e.g. read can deadlock
write locking on file.f_pos in the kernel.

FUSE's FOPEN_NONSEEKABLE was added in 2008 in a7c1b990f715 ("fuse:
implement nonseekable open") to support OSSPD. OSSPD implements /dev/dsp
in userspace with FOPEN_NONSEEKABLE flag, with corresponding read and
write routines not depending on current position at all, and with both
read and write being potentially blocking operations:

See

https://github.com/libfuse/osspd
https://lwn.net/Articles/308445

https://github.com/libfuse/osspd/blob/14a9cff0/osspd.c#L1406
https://github.com/libfuse/osspd/blob/14a9cff0/osspd.c#L1438-L1477
https://github.com/libfuse/osspd/blob/14a9cff0/osspd.c#L1479-L1510

Corresponding libfuse example/test also describes FOPEN_NONSEEKABLE as
"somewhat pipe-like files ..." with read handler not using offset.
However that test implements only read without write and cannot exercise
the deadlock scenario:

https://github.com/libfuse/libfuse/blob/fuse-3.4.2-3-ga1bff7d/example/poll.c#L124-L131
https://github.com/libfuse/libfuse/blob/fuse-3.4.2-3-ga1bff7d/example/poll.c#L146-L163
https://github.com/libfuse/libfuse/blob/fuse-3.4.2-3-ga1bff7d/example/poll.c#L209-L216

I've actually hit the read vs write deadlock for real while implementing
my FUSE filesystem where there is /head/watch file, for which open
creates separate bidirectional socket-like stream in between filesystem
and its user with both read and write being later performed
simultaneously. And there it is semantically not easy to split the
stream into two separate read-only and write-only channels:

https://lab.nexedi.com/kirr/wendelin.core/blob/f13aa600/wcfs/wcfs.go#L88-169

Let's fix this regression. The plan is:

1. We can't change nonseekable_open to include &~FMODE_ATOMIC_POS -
doing so would break many in-kernel nonseekable_open users which
actually use ppos in read/write handlers.

2. Add stream_open() to kernel to open stream-like non-seekable file
descriptors. Read and write on such file descriptors would never use
nor change ppos. And with that property on stream-like files read and
write will be running without taking f_pos lock - i.e. read and write
could be running simultaneously.

3. With semantic patch search and convert to stream_open all in-kernel
nonseekable_open users for which read and write actually do not
depend on ppos and where there is no other methods in file_operations
which assume @offset access.

4. Add FOPEN_STREAM to fs/fuse/ and open in-kernel file-descriptors via
steam_open if that bit is present in filesystem open reply.

It was tempting to change fs/fuse/ open handler to use stream_open
instead of nonseekable_open on just FOPEN_NONSEEKABLE flags, but
grepping through Debian codesearch shows users of FOPEN_NONSEEKABLE,
and in particular GVFS which actually uses offset in its read and
write handlers

https://codesearch.debian.net/search?q=-%3Enonseekable+%3D
https://gitlab.gnome.org/GNOME/gvfs/blob/1.40.0-6-gcbc54396/client/gvfsfusedaemon.c#L1080
https://gitlab.gnome.org/GNOME/gvfs/blob/1.40.0-6-gcbc54396/client/gvfsfusedaemon.c#L1247-1346
https://gitlab.gnome.org/GNOME/gvfs/blob/1.40.0-6-gcbc54396/client/gvfsfusedaemon.c#L1399-1481

so if we would do such a change it will break a real user.

5. Add stream_open and FOPEN_STREAM handling to stable kernels starting
from v3.14+ (the kernel where 9c225f2655 first appeared).

This will allow to patch OSSPD and other FUSE filesystems that
provide stream-like files to return FOPEN_STREAM | FOPEN_NONSEEKABLE
in their open handler and this way avoid the deadlock on all kernel
versions. This should work because fs/fuse/ ignores unknown open
flags returned from a filesystem and so passing FOPEN_STREAM to a
kernel that is not aware of this flag cannot hurt. In turn the kernel
that is not aware of FOPEN_STREAM will be < v3.14 where just
FOPEN_NONSEEKABLE is sufficient to implement streams without read vs
write deadlock.

This patch adds stream_open, converts /proc/xen/xenbus to it and adds
semantic patch to automatically locate in-kernel places that are either
required to be converted due to read vs write deadlock, or that are just
safe to be converted because read and write do not use ppos and there
are no other funky methods in file_operations.

Regarding semantic patch I've verified each generated change manually -
that it is correct to convert - and each other nonseekable_open instance
left - that it is either not correct to convert there, or that it is not
converted due to current stream_open.cocci limitations.

The script also does not convert files that should be valid to convert,
but that currently have .llseek = noop_llseek or generic_file_llseek for
unknown reason despite file being opened with nonseekable_open (e.g.
drivers/input/mousedev.c)

Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Yongzhi Pan <panyongzhi@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Tejun Heo <tj@kernel.org>
Cc: Kirill Tkhai <ktkhai@virtuozzo.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: Nikolaus Rath <Nikolaus@rath.org>
Cc: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff 10dce8af Tue Mar 26 16:20:43 MDT 2019 Kirill Smelkov <kirr@nexedi.com> fs: stream_open - opener for stream-like files so that read and write can run simultaneously without deadlock

Commit 9c225f2655e3 ("vfs: atomic f_pos accesses as per POSIX") added
locking for file.f_pos access and in particular made concurrent read and
write not possible - now both those functions take f_pos lock for the
whole run, and so if e.g. a read is blocked waiting for data, write will
deadlock waiting for that read to complete.

This caused regression for stream-like files where previously read and
write could run simultaneously, but after that patch could not do so
anymore. See e.g. commit 581d21a2d02a ("xenbus: fix deadlock on writes
to /proc/xen/xenbus") which fixes such regression for particular case of
/proc/xen/xenbus.

The patch that added f_pos lock in 2014 did so to guarantee POSIX thread
safety for read/write/lseek and added the locking to file descriptors of
all regular files. In 2014 that thread-safety problem was not new as it
was already discussed earlier in 2006.

However even though 2006'th version of Linus's patch was adding f_pos
locking "only for files that are marked seekable with FMODE_LSEEK (thus
avoiding the stream-like objects like pipes and sockets)", the 2014
version - the one that actually made it into the tree as 9c225f2655e3 -
is doing so irregardless of whether a file is seekable or not.

See

https://lore.kernel.org/lkml/53022DB1.4070805@gmail.com/
https://lwn.net/Articles/180387
https://lwn.net/Articles/180396

for historic context.

The reason that it did so is, probably, that there are many files that
are marked non-seekable, but e.g. their read implementation actually
depends on knowing current position to correctly handle the read. Some
examples:

kernel/power/user.c snapshot_read
fs/debugfs/file.c u32_array_read
fs/fuse/control.c fuse_conn_waiting_read + ...
drivers/hwmon/asus_atk0110.c atk_debugfs_ggrp_read
arch/s390/hypfs/inode.c hypfs_read_iter
...

Despite that, many nonseekable_open users implement read and write with
pure stream semantics - they don't depend on passed ppos at all. And for
those cases where read could wait for something inside, it creates a
situation similar to xenbus - the write could be never made to go until
read is done, and read is waiting for some, potentially external, event,
for potentially unbounded time -> deadlock.

Besides xenbus, there are 14 such places in the kernel that I've found
with semantic patch (see below):

drivers/xen/evtchn.c:667:8-24: ERROR: evtchn_fops: .read() can deadlock .write()
drivers/isdn/capi/capi.c:963:8-24: ERROR: capi_fops: .read() can deadlock .write()
drivers/input/evdev.c:527:1-17: ERROR: evdev_fops: .read() can deadlock .write()
drivers/char/pcmcia/cm4000_cs.c:1685:7-23: ERROR: cm4000_fops: .read() can deadlock .write()
net/rfkill/core.c:1146:8-24: ERROR: rfkill_fops: .read() can deadlock .write()
drivers/s390/char/fs3270.c:488:1-17: ERROR: fs3270_fops: .read() can deadlock .write()
drivers/usb/misc/ldusb.c:310:1-17: ERROR: ld_usb_fops: .read() can deadlock .write()
drivers/hid/uhid.c:635:1-17: ERROR: uhid_fops: .read() can deadlock .write()
net/batman-adv/icmp_socket.c:80:1-17: ERROR: batadv_fops: .read() can deadlock .write()
drivers/media/rc/lirc_dev.c:198:1-17: ERROR: lirc_fops: .read() can deadlock .write()
drivers/leds/uleds.c:77:1-17: ERROR: uleds_fops: .read() can deadlock .write()
drivers/input/misc/uinput.c:400:1-17: ERROR: uinput_fops: .read() can deadlock .write()
drivers/infiniband/core/user_mad.c:985:7-23: ERROR: umad_fops: .read() can deadlock .write()
drivers/gnss/core.c:45:1-17: ERROR: gnss_fops: .read() can deadlock .write()

In addition to the cases above another regression caused by f_pos
locking is that now FUSE filesystems that implement open with
FOPEN_NONSEEKABLE flag, can no longer implement bidirectional
stream-like files - for the same reason as above e.g. read can deadlock
write locking on file.f_pos in the kernel.

FUSE's FOPEN_NONSEEKABLE was added in 2008 in a7c1b990f715 ("fuse:
implement nonseekable open") to support OSSPD. OSSPD implements /dev/dsp
in userspace with FOPEN_NONSEEKABLE flag, with corresponding read and
write routines not depending on current position at all, and with both
read and write being potentially blocking operations:

See

https://github.com/libfuse/osspd
https://lwn.net/Articles/308445

https://github.com/libfuse/osspd/blob/14a9cff0/osspd.c#L1406
https://github.com/libfuse/osspd/blob/14a9cff0/osspd.c#L1438-L1477
https://github.com/libfuse/osspd/blob/14a9cff0/osspd.c#L1479-L1510

Corresponding libfuse example/test also describes FOPEN_NONSEEKABLE as
"somewhat pipe-like files ..." with read handler not using offset.
However that test implements only read without write and cannot exercise
the deadlock scenario:

https://github.com/libfuse/libfuse/blob/fuse-3.4.2-3-ga1bff7d/example/poll.c#L124-L131
https://github.com/libfuse/libfuse/blob/fuse-3.4.2-3-ga1bff7d/example/poll.c#L146-L163
https://github.com/libfuse/libfuse/blob/fuse-3.4.2-3-ga1bff7d/example/poll.c#L209-L216

I've actually hit the read vs write deadlock for real while implementing
my FUSE filesystem where there is /head/watch file, for which open
creates separate bidirectional socket-like stream in between filesystem
and its user with both read and write being later performed
simultaneously. And there it is semantically not easy to split the
stream into two separate read-only and write-only channels:

https://lab.nexedi.com/kirr/wendelin.core/blob/f13aa600/wcfs/wcfs.go#L88-169

Let's fix this regression. The plan is:

1. We can't change nonseekable_open to include &~FMODE_ATOMIC_POS -
doing so would break many in-kernel nonseekable_open users which
actually use ppos in read/write handlers.

2. Add stream_open() to kernel to open stream-like non-seekable file
descriptors. Read and write on such file descriptors would never use
nor change ppos. And with that property on stream-like files read and
write will be running without taking f_pos lock - i.e. read and write
could be running simultaneously.

3. With semantic patch search and convert to stream_open all in-kernel
nonseekable_open users for which read and write actually do not
depend on ppos and where there is no other methods in file_operations
which assume @offset access.

4. Add FOPEN_STREAM to fs/fuse/ and open in-kernel file-descriptors via
steam_open if that bit is present in filesystem open reply.

It was tempting to change fs/fuse/ open handler to use stream_open
instead of nonseekable_open on just FOPEN_NONSEEKABLE flags, but
grepping through Debian codesearch shows users of FOPEN_NONSEEKABLE,
and in particular GVFS which actually uses offset in its read and
write handlers

https://codesearch.debian.net/search?q=-%3Enonseekable+%3D
https://gitlab.gnome.org/GNOME/gvfs/blob/1.40.0-6-gcbc54396/client/gvfsfusedaemon.c#L1080
https://gitlab.gnome.org/GNOME/gvfs/blob/1.40.0-6-gcbc54396/client/gvfsfusedaemon.c#L1247-1346
https://gitlab.gnome.org/GNOME/gvfs/blob/1.40.0-6-gcbc54396/client/gvfsfusedaemon.c#L1399-1481

so if we would do such a change it will break a real user.

5. Add stream_open and FOPEN_STREAM handling to stable kernels starting
from v3.14+ (the kernel where 9c225f2655 first appeared).

This will allow to patch OSSPD and other FUSE filesystems that
provide stream-like files to return FOPEN_STREAM | FOPEN_NONSEEKABLE
in their open handler and this way avoid the deadlock on all kernel
versions. This should work because fs/fuse/ ignores unknown open
flags returned from a filesystem and so passing FOPEN_STREAM to a
kernel that is not aware of this flag cannot hurt. In turn the kernel
that is not aware of FOPEN_STREAM will be < v3.14 where just
FOPEN_NONSEEKABLE is sufficient to implement streams without read vs
write deadlock.

This patch adds stream_open, converts /proc/xen/xenbus to it and adds
semantic patch to automatically locate in-kernel places that are either
required to be converted due to read vs write deadlock, or that are just
safe to be converted because read and write do not use ppos and there
are no other funky methods in file_operations.

Regarding semantic patch I've verified each generated change manually -
that it is correct to convert - and each other nonseekable_open instance
left - that it is either not correct to convert there, or that it is not
converted due to current stream_open.cocci limitations.

The script also does not convert files that should be valid to convert,
but that currently have .llseek = noop_llseek or generic_file_llseek for
unknown reason despite file being opened with nonseekable_open (e.g.
drivers/input/mousedev.c)

Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Yongzhi Pan <panyongzhi@gmail.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: David Vrabel <david.vrabel@citrix.com>
Cc: Juergen Gross <jgross@suse.com>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Tejun Heo <tj@kernel.org>
Cc: Kirill Tkhai <ktkhai@virtuozzo.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Julia Lawall <Julia.Lawall@lip6.fr>
Cc: Nikolaus Rath <Nikolaus@rath.org>
Cc: Han-Wen Nienhuys <hanwen@google.com>
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff 7c0f6ba6 Sat Dec 24 12:46:01 MST 2016 Linus Torvalds <torvalds@linux-foundation.org> Replace <asm/uaccess.h> with <linux/uaccess.h> globally

This was entirely automated, using the script by Al:

PATT='^[[:blank:]]*#[[:blank:]]*include[[:blank:]]*<asm/uaccess.h>'
sed -i -e "s!$PATT!#include <linux/uaccess.h>!" \
$(git grep -l "$PATT"|grep -v ^include/linux/uaccess.h)

to do the replacement at the end of the merge window.

Requested-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
diff 7f7f25e8 Tue Feb 11 15:49:24 MST 2014 Al Viro <viro@zeniv.linux.org.uk> replace checking for ->read/->aio_read presence with check in ->f_mode

Since we are about to introduce new methods (read_iter/write_iter), the
tests in a bunch of places would have to grow inconveniently. Check
once (at open() time) and store results in ->f_mode as FMODE_CAN_READ
and FMODE_CAN_WRITE resp. It might end up being a temporary measure -
once everything switches from ->aio_{read,write} to ->{read,write}_iter
it might make sense to return to open-coded checks. We'll see...

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 7f7f25e8 Tue Feb 11 15:49:24 MST 2014 Al Viro <viro@zeniv.linux.org.uk> replace checking for ->read/->aio_read presence with check in ->f_mode

Since we are about to introduce new methods (read_iter/write_iter), the
tests in a bunch of places would have to grow inconveniently. Check
once (at open() time) and store results in ->f_mode as FMODE_CAN_READ
and FMODE_CAN_WRITE resp. It might end up being a temporary measure -
once everything switches from ->aio_{read,write} to ->{read,write}_iter
it might make sense to return to open-coded checks. We'll see...

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 9c225f26 Mon Mar 03 10:36:58 MST 2014 Linus Torvalds <torvalds@linux-foundation.org> vfs: atomic f_pos accesses as per POSIX

Our write() system call has always been atomic in the sense that you get
the expected thread-safe contiguous write, but we haven't actually
guaranteed that concurrent writes are serialized wrt f_pos accesses, so
threads (or processes) that share a file descriptor and use "write()"
concurrently would quite likely overwrite each others data.

This violates POSIX.1-2008/SUSv4 Section XSI 2.9.7 that says:

"2.9.7 Thread Interactions with Regular File Operations

All of the following functions shall be atomic with respect to each
other in the effects specified in POSIX.1-2008 when they operate on
regular files or symbolic links: [...]"

and one of the effects is the file position update.

This unprotected file position behavior is not new behavior, and nobody
has ever cared. Until now. Yongzhi Pan reported unexpected behavior to
Michael Kerrisk that was due to this.

This resolves the issue with a f_pos-specific lock that is taken by
read/write/lseek on file descriptors that may be shared across threads
or processes.

Reported-by: Yongzhi Pan <panyongzhi@gmail.com>
Reported-by: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 9c225f26 Mon Mar 03 10:36:58 MST 2014 Linus Torvalds <torvalds@linux-foundation.org> vfs: atomic f_pos accesses as per POSIX

Our write() system call has always been atomic in the sense that you get
the expected thread-safe contiguous write, but we haven't actually
guaranteed that concurrent writes are serialized wrt f_pos accesses, so
threads (or processes) that share a file descriptor and use "write()"
concurrently would quite likely overwrite each others data.

This violates POSIX.1-2008/SUSv4 Section XSI 2.9.7 that says:

"2.9.7 Thread Interactions with Regular File Operations

All of the following functions shall be atomic with respect to each
other in the effects specified in POSIX.1-2008 when they operate on
regular files or symbolic links: [...]"

and one of the effects is the file position update.

This unprotected file position behavior is not new behavior, and nobody
has ever cared. Until now. Yongzhi Pan reported unexpected behavior to
Michael Kerrisk that was due to this.

This resolves the issue with a f_pos-specific lock that is taken by
read/write/lseek on file descriptors that may be shared across threads
or processes.

Reported-by: Yongzhi Pan <panyongzhi@gmail.com>
Reported-by: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
diff 7b8e8924 Tue May 08 01:24:13 MDT 2007 Chris Snook <csnook@redhat.com> use symbolic constants in generic lseek code

Convert magic numbers to SEEK_* values from fs.h

Signed-off-by: Chris Snook <csnook@redhat.com>
Acked-by: David Howells <dhowells@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
H A Dopen.cdiff 021a160a Tue Aug 08 11:26:35 MDT 2023 Linus Torvalds <torvalds@linux-foundation.org> fs: use __fput_sync in close(2)

close(2) is a special case which guarantees a shallow kernel stack,
making delegation to task_work machinery unnecessary. Said delegation is
problematic as it involves atomic ops and interrupt masking trips, none
of which are cheap on x86-64. Forcing close(2) to do it looks like an
oversight in the original work.

Moreover presence of CONFIG_RSEQ adds an additional overhead as fput()
-> task_work_add(..., TWA_RESUME) -> set_notify_resume() makes the
thread returning to userspace land in resume_user_mode_work(), where
rseq_handle_notify_resume takes a SMAP round-trip if rseq is enabled for
the thread (and it is by default with contemporary glibc).

Sample result when benchmarking open1_processes -t 1 from will-it-scale
(that's an open + close loop) + tmpfs on /tmp, running on the Sapphire
Rapid CPU (ops/s):
stock+RSEQ: 1329857
stock-RSEQ: 1421667 (+7%)
patched: 1523521 (+14.5% / +7%) (with / without rseq)

Patched result is the same regardless of rseq as the codepath is avoided.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 021a160a Tue Aug 08 11:26:35 MDT 2023 Linus Torvalds <torvalds@linux-foundation.org> fs: use __fput_sync in close(2)

close(2) is a special case which guarantees a shallow kernel stack,
making delegation to task_work machinery unnecessary. Said delegation is
problematic as it involves atomic ops and interrupt masking trips, none
of which are cheap on x86-64. Forcing close(2) to do it looks like an
oversight in the original work.

Moreover presence of CONFIG_RSEQ adds an additional overhead as fput()
-> task_work_add(..., TWA_RESUME) -> set_notify_resume() makes the
thread returning to userspace land in resume_user_mode_work(), where
rseq_handle_notify_resume takes a SMAP round-trip if rseq is enabled for
the thread (and it is by default with contemporary glibc).

Sample result when benchmarking open1_processes -t 1 from will-it-scale
(that's an open + close loop) + tmpfs on /tmp, running on the Sapphire
Rapid CPU (ops/s):
stock+RSEQ: 1329857
stock-RSEQ: 1421667 (+7%)
patched: 1523521 (+14.5% / +7%) (with / without rseq)

Patched result is the same regardless of rseq as the codepath is avoided.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff a0fc452a Sat Aug 05 10:11:58 MDT 2023 Aleksa Sarai <cyphar@cyphar.com> open: make RESOLVE_CACHED correctly test for O_TMPFILE

O_TMPFILE is actually __O_TMPFILE|O_DIRECTORY. This means that the old
fast-path check for RESOLVE_CACHED would reject all users passing
O_DIRECTORY with -EAGAIN, when in fact the intended test was to check
for __O_TMPFILE.

Cc: stable@vger.kernel.org # v5.12+
Fixes: 99668f618062 ("fs: expose LOOKUP_CACHED through openat2() RESOLVE_CACHED")
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Message-Id: <20230806-resolve_cached-o_tmpfile-v1-1-7ba16308465e@cyphar.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 7b8c9d7b Sun Jun 11 06:24:29 MDT 2023 Amir Goldstein <amir73il@gmail.com> fsnotify: move fsnotify_open() hook into do_dentry_open()

fsnotify_open() hook is called only from high level system calls
context and not called for the very many helpers to open files.

This may makes sense for many of the special file open cases, but it is
inconsistent with fsnotify_close() hook that is called for every last
fput() of on a file object with FMODE_OPENED.

As a result, it is possible to observe ACCESS, MODIFY and CLOSE events
without ever observing an OPEN event.

Fix this inconsistency by replacing all the fsnotify_open() hooks with
a single hook inside do_dentry_open().

If there are special cases that would like to opt-out of the possible
overhead of fsnotify() call in fsnotify_open(), they would probably also
want to avoid the overhead of fsnotify() call in the rest of the fsnotify
hooks, so they should be opening that file with the __FMODE_NONOTIFY flag.

However, in the majority of those cases, the s_fsnotify_connectors
optimization in fsnotify_parent() would be sufficient to avoid the
overhead of fsnotify() call anyway.

Signed-off-by: Amir Goldstein <amir73il@gmail.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Message-Id: <20230611122429.1499617-1-amir73il@gmail.com>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 43b45063 Tue Mar 21 02:18:07 MDT 2023 Christian Brauner <brauner@kernel.org> open: return EINVAL for O_DIRECTORY | O_CREAT

After a couple of years and multiple LTS releases we received a report
that the behavior of O_DIRECTORY | O_CREAT changed starting with v5.7.

On kernels prior to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
had the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: create regular file
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: create regular file
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

On kernels since to v5.7 combinations of O_DIRECTORY, O_CREAT, O_EXCL
have the following semantics:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: EISDIR

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: ENOTDIR (create regular file)
* d exists and is a regular file: EEXIST
* d exists and is a directory: EEXIST

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

This is a fairly substantial semantic change that userspace didn't
notice until Pedro took the time to deliberately figure out corner
cases. Since no one noticed this breakage we can somewhat safely assume
that O_DIRECTORY | O_CREAT combinations are likely unused.

The v5.7 breakage is especially weird because while ENOTDIR is returned
indicating failure a regular file is actually created. This doesn't make
a lot of sense.

Time was spent finding potential users of this combination. Searching on
codesearch.debian.net showed that codebases often express semantical
expectations about O_DIRECTORY | O_CREAT which are completely contrary
to what our code has done and currently does.

The expectation often is that this particular combination would create
and open a directory. This suggests users who tried to use that
combination would stumble upon the counterintuitive behavior no matter
if pre-v5.7 or post v5.7 and quickly realize neither semantics give them
what they want. For some examples see the code examples in [1] to [3]
and the discussion in [4].

There are various ways to address this issue. The lazy/simple option
would be to restore the pre-v5.7 behavior and to just live with that bug
forever. But since there's a real chance that the O_DIRECTORY | O_CREAT
quirk isn't relied upon we should try to get away with murder(ing bad
semantics) first. If we need to Frankenstein pre-v5.7 behavior later so
be it.

So let's simply return EINVAL categorically for O_DIRECTORY | O_CREAT
combinations. In addition to cleaning up the old bug this also opens up
the possiblity to make that flag combination do something more intuitive
in the future.

Starting with this commit the following semantics apply:

(1) open("/tmp/d", O_DIRECTORY | O_CREAT)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(2) open("/tmp/d", O_DIRECTORY | O_CREAT | O_EXCL)
* d doesn't exist: EINVAL
* d exists and is a regular file: EINVAL
* d exists and is a directory: EINVAL

(3) open("/tmp/d", O_DIRECTORY | O_EXCL)
* d doesn't exist: ENOENT
* d exists and is a regular file: ENOTDIR
* d exists and is a directory: open directory

One additional note, O_TMPFILE is implemented as:

#define __O_TMPFILE 020000000
#define O_TMPFILE (__O_TMPFILE | O_DIRECTORY)
#define O_TMPFILE_MASK (__O_TMPFILE | O_DIRECTORY | O_CREAT)

For older kernels it was important to return an explicit error when
O_TMPFILE wasn't supported. So O_TMPFILE requires that O_DIRECTORY is
raised alongside __O_TMPFILE. It also enforced that O_CREAT wasn't
specified. Since O_DIRECTORY | O_CREAT could be used to create a regular
allowing that combination together with __O_TMPFILE would've meant that
false positives were possible, i.e., that a regular file was created
instead of a O_TMPFILE. This could've been used to trick userspace into
thinking it operated on a O_TMPFILE when it wasn't.

Now that we block O_DIRECTORY | O_CREAT completely the check for O_CREAT
in the __O_TMPFILE branch via if ((flags & O_TMPFILE_MASK) != O_TMPFILE)
can be dropped. Instead we can simply check verify that O_DIRECTORY is
raised via if (!(flags & O_DIRECTORY)) and explain this in two comments.

As Aleksa pointed out O_PATH is unaffected by this change since it
always returned EINVAL if O_CREAT was specified - with or without
O_DIRECTORY.

Link: https://lore.kernel.org/lkml/20230320071442.172228-1-pedro.falcato@gmail.com
Link: https://sources.debian.org/src/flatpak/1.14.4-1/subprojects/libglnx/glnx-dirfd.c/?hl=324#L324 [1]
Link: https://sources.debian.org/src/flatpak-builder/1.2.3-1/subprojects/libglnx/glnx-shutil.c/?hl=251#L251 [2]
Link: https://sources.debian.org/src/ostree/2022.7-2/libglnx/glnx-dirfd.c/?hl=324#L324 [3]
Link: https://www.openwall.com/lists/oss-security/2014/11/26/14 [4]
Reported-by: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
/linux-master/include/linux/
H A Dfs.hdiff 70dfe3f0 Tue Feb 20 22:14:08 MST 2024 Gabriel Krisman Bertazi <krisman@suse.de> libfs: Add helper to choose dentry operations at mount-time

In preparation to drop the similar helper that sets d_op at lookup time,
add a version to set the right d_op filesystem-wide, through sb->s_d_op.
The operations structures are shared across filesystems supporting
fscrypt and/or casefolding, therefore we can keep it in common libfs
code.

Reviewed-by: Eric Biggers <ebiggers@google.com>
Link: https://lore.kernel.org/r/20240221171412.10710-7-krisman@suse.de
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
diff 2824083d Tue Feb 20 22:14:03 MST 2024 Gabriel Krisman Bertazi <krisman@suse.de> ovl: Always reject mounting over case-insensitive directories

overlayfs relies on the filesystem setting DCACHE_OP_HASH or
DCACHE_OP_COMPARE to reject mounting over case-insensitive directories.

Since commit bb9cd9106b22 ("fscrypt: Have filesystems handle their
d_ops"), we set ->d_op through a hook in ->d_lookup, which
means the root dentry won't have them, causing the mount to accidentally
succeed.

In v6.7-rc7, the following sequence will succeed to mount, but any
dentry other than the root dentry will be a "weird" dentry to ovl and
fail with EREMOTE.

mkfs.ext4 -O casefold lower.img
mount -O loop lower.img lower
mount -t overlay -o lowerdir=lower,upperdir=upper,workdir=work ovl /mnt

Mounting on a subdirectory fails, as expected, because DCACHE_OP_HASH
and DCACHE_OP_COMPARE are properly set by ->lookup.

Fix by explicitly rejecting superblocks that allow case-insensitive
dentries. Yes, this will be solved when we move d_op configuration back
to ->s_d_op. Yet, we better have an explicit fix to avoid messing up
again.

While there, re-sort the entries to have more descriptive error messages
first.

Fixes: bb9cd9106b22 ("fscrypt: Have filesystems handle their d_ops")
Acked-by: Amir Goldstein <amir73il@gmail.com>
Reviewed-by: Eric Biggers <ebiggers@google.com>
Link: https://lore.kernel.org/r/20240221171412.10710-2-krisman@suse.de
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
diff d3b1a9a7 Fri Feb 02 01:33:04 MST 2024 JonasZhou <JonasZhou@zhaoxin.com> fs/address_space: move i_mmap_rwsem to mitigate a false sharing with i_mmap.

In the struct address_space, there is a 32-byte gap between i_mmap
and i_mmap_rwsem. Due to the alignment of struct address_space
variables to 8 bytes, in certain situations, i_mmap and i_mmap_rwsem
may end up in the same CACHE line.

While running Unixbench/execl, we observe high false sharing issues
when accessing i_mmap against i_mmap_rwsem. We move i_mmap_rwsem
after i_private_list, ensuring a 64-byte gap between i_mmap and
i_mmap_rwsem.

For Intel Silver machines (2 sockets) using kernel v6.8 rc-2, the score
of Unixbench/execl improves by ~3.94%, and the score of Unixbench/shell
improves by ~3.26%.

Baseline:
-------------------------------------------------------------
162 546 748 11374 21 0xffff92e266af90c0
-------------------------------------------------------------
46.89% 44.65% 0.00% 0.00% 0x0 1 1 0xffffffff86d5fb96 460 258 271 1069 32 [k] __handle_mm_fault [kernel.vmlinux] memory.c:2940 0 1
4.21% 4.41% 0.00% 0.00% 0x4 1 1 0xffffffff86d0ed54 473 311 288 95 28 [k] filemap_read [kernel.vmlinux] atomic.h:23 0 1
0.00% 0.00% 0.04% 4.76% 0x8 1 1 0xffffffff86d4bcf1 0 0 0 5 4 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:204 0 1
6.41% 6.02% 0.00% 0.00% 0x8 1 1 0xffffffff86d4ba85 411 271 339 210 32 [k] vma_interval_tree_insert [kernel.vmlinux] interval_tree.c:23 0 1
0.00% 0.00% 0.47% 95.24% 0x10 1 1 0xffffffff86d4bd34 0 0 0 74 32 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:339 0 1
0.37% 0.13% 0.00% 0.00% 0x10 1 1 0xffffffff86d4bb4f 328 212 380 7 5 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:338 0 1
5.13% 5.08% 0.00% 0.00% 0x10 1 1 0xffffffff86d4bb4b 416 255 357 197 32 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:338 0 1
1.10% 0.53% 0.00% 0.00% 0x28 1 1 0xffffffff86e06eb8 395 228 351 24 14 [k] do_dentry_open [kernel.vmlinux] open.c:966 0 1
1.10% 2.14% 57.07% 0.00% 0x38 1 1 0xffffffff878c9225 1364 792 462 7003 32 [k] down_write [kernel.vmlinux] atomic64_64.h:109 0 1
0.00% 0.00% 0.01% 0.00% 0x38 1 1 0xffffffff878c8e75 0 0 252 3 2 [k] rwsem_down_write_slowpath [kernel.vmlinux] atomic64_64.h:109 0 1
0.00% 0.13% 0.00% 0.00% 0x38 1 1 0xffffffff878c8e23 0 596 63 2 2 [k] rwsem_down_write_slowpath [kernel.vmlinux] atomic64_64.h:15 0 1
2.38% 2.94% 6.53% 0.00% 0x38 1 1 0xffffffff878c8ccb 1150 818 570 1197 32 [k] rwsem_down_write_slowpath [kernel.vmlinux] atomic64_64.h:109 0 1
30.59% 32.22% 0.00% 0.00% 0x38 1 1 0xffffffff878c8cb4 423 251 380 648 32 [k] rwsem_down_write_slowpath [kernel.vmlinux] atomic64_64.h:15 0 1
1.83% 1.74% 35.88% 0.00% 0x38 1 1 0xffffffff86b4f833 1217 1112 565 4586 32 [k] up_write [kernel.vmlinux] atomic64_64.h:91 0 1

with this change:
-------------------------------------------------------------
360 12 300 57 35 0xffff982cdae76400
-------------------------------------------------------------
50.00% 59.67% 0.00% 0.00% 0x0 1 1 0xffffffff8215fb86 352 200 191 558 32 [k] __handle_mm_fault [kernel.vmlinux] memory.c:2940 0 1
8.33% 5.00% 0.00% 0.00% 0x4 1 1 0xffffffff8210ed44 370 284 263 42 24 [k] filemap_read [kernel.vmlinux] atomic.h:23 0 1
0.00% 0.00% 5.26% 2.86% 0x8 1 1 0xffffffff8214bce1 0 0 0 4 4 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:204 0 1
33.33% 14.33% 0.00% 0.00% 0x8 1 1 0xffffffff8214ba75 344 186 219 140 32 [k] vma_interval_tree_insert [kernel.vmlinux] interval_tree.c:23 0 1
0.00% 0.00% 94.74% 97.14% 0x10 1 1 0xffffffff8214bd24 0 0 0 88 29 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:339 0 1
8.33% 20.00% 0.00% 0.00% 0x10 1 1 0xffffffff8214bb3b 296 209 226 167 31 [k] vma_interval_tree_remove [kernel.vmlinux] rbtree_augmented.h:338 0 1
0.00% 0.67% 0.00% 0.00% 0x28 1 1 0xffffffff82206f45 0 140 334 4 3 [k] do_dentry_open [kernel.vmlinux] open.c:966 0 1
0.00% 0.33% 0.00% 0.00% 0x38 1 1 0xffffffff8250a6c4 0 286 126 5 5 [k] errseq_sample [kernel.vmlinux] errseq.c:125 0

Signed-off-by: JonasZhou <JonasZhou@zhaoxin.com>
Link: https://lore.kernel.org/r/20240202083304.10995-1-JonasZhou-oc@zhaoxin.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff af7628d6 Fri Nov 17 09:14:47 MST 2023 Matthew Wilcox (Oracle) <willy@infradead.org> fs: convert error_remove_page to error_remove_folio

There were already assertions that we were not passing a tail page to
error_remove_page(), so make the compiler enforce that by converting
everything to pass and use a folio.

Link: https://lkml.kernel.org/r/20231117161447.2461643-7-willy@infradead.org
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Naoya Horiguchi <naoya.horiguchi@nec.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
diff 3e327154 Sat Aug 05 01:25:01 MDT 2023 Linus Torvalds <torvalds@linux-foundation.org> vfs: get rid of old '->iterate' directory operation

All users now just use '->iterate_shared()', which only takes the
directory inode lock for reading.

Filesystems that never got convered to shared mode now instead use a
wrapper that drops the lock, re-takes it in write mode, calls the old
function, and then downgrades the lock back to read mode.

This way the VFS layer and other callers no longer need to care about
filesystems that never got converted to the modern era.

The filesystems that use the new wrapper are ceph, coda, exfat, jfs,
ntfs, ocfs2, overlayfs, and vboxsf.

Honestly, several of them look like they really could just iterate their
directories in shared mode and skip the wrapper entirely, but the point
of this change is to not change semantics or fix filesystems that
haven't been fixed in the last 7+ years, but to finally get rid of the
dual iterators.

Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff c402a9a9 Thu Jun 01 08:58:58 MDT 2023 Christoph Hellwig <hch@lst.de> filemap: add a kiocb_invalidate_post_direct_write helper

Add a helper to invalidate page cache after a dio write.

Link: https://lkml.kernel.org/r/20230601145904.1385409-7-hch@lst.de
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Damien Le Moal <dlemoal@kernel.org>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Acked-by: Darrick J. Wong <djwong@kernel.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Andreas Gruenbacher <agruenba@redhat.com>
Cc: Anna Schumaker <anna@kernel.org>
Cc: Chao Yu <chao@kernel.org>
Cc: Christian Brauner <brauner@kernel.org>
Cc: Ilya Dryomov <idryomov@gmail.com>
Cc: Jaegeuk Kim <jaegeuk@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Thumshirn <johannes.thumshirn@wdc.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Miklos Szeredi <miklos@szeredi.hu>
Cc: Miklos Szeredi <mszeredi@redhat.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Trond Myklebust <trond.myklebust@hammerspace.com>
Cc: Xiubo Li <xiubli@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
diff a7bc2e8d Thu Jun 01 03:24:00 MDT 2023 chenzhiyin <zhiyin.chen@intel.com> fs.h: Optimize file struct to prevent false sharing

In the syscall test of UnixBench, performance regression occurred due
to false sharing.

The lock and atomic members, including file::f_lock, file::f_count and
file::f_pos_lock are highly contended and frequently updated in the
high-concurrency test scenarios. perf c2c indentified one affected
read access, file::f_op.
To prevent false sharing, the layout of file struct is changed as
following
(A) f_lock, f_count and f_pos_lock are put together to share the same
cache line.
(B) The read mostly members, including f_path, f_inode, f_op are put
into a separate cache line.
(C) f_mode is put together with f_count, since they are used frequently
at the same time.
Due to '__randomize_layout' attribute of file struct, the updated layout
only can be effective when CONFIG_RANDSTRUCT_NONE is 'y'.

The optimization has been validated in the syscall test of UnixBench.
performance gain is 30~50%. Furthermore, to confirm the optimization
effectiveness on the other codes path, the results of fsdisk, fsbuffer
and fstime are also shown.

Here are the detailed test results of unixbench.

Command: numactl -C 3-18 ./Run -c 16 syscall fsbuffer fstime fsdisk

Without Patch
------------------------------------------------------------------------
File Copy 1024 bufsize 2000 maxblocks 875052.1 KBps (30.0 s, 2 samples)
File Copy 256 bufsize 500 maxblocks 235484.0 KBps (30.0 s, 2 samples)
File Copy 4096 bufsize 8000 maxblocks 2815153.5 KBps (30.0 s, 2 samples)
System Call Overhead 5772268.3 lps (10.0 s, 7 samples)

System Benchmarks Partial Index BASELINE RESULT INDEX
File Copy 1024 bufsize 2000 maxblocks 3960.0 875052.1 2209.7
File Copy 256 bufsize 500 maxblocks 1655.0 235484.0 1422.9
File Copy 4096 bufsize 8000 maxblocks 5800.0 2815153.5 4853.7
System Call Overhead 15000.0 5772268.3 3848.2
========
System Benchmarks Index Score (Partial Only) 2768.3

With Patch
------------------------------------------------------------------------
File Copy 1024 bufsize 2000 maxblocks 1009977.2 KBps (30.0 s, 2 samples)
File Copy 256 bufsize 500 maxblocks 264765.9 KBps (30.0 s, 2 samples)
File Copy 4096 bufsize 8000 maxblocks 3052236.0 KBps (30.0 s, 2 samples)
System Call Overhead 8237404.4 lps (10.0 s, 7 samples)

System Benchmarks Partial Index BASELINE RESULT INDEX
File Copy 1024 bufsize 2000 maxblocks 3960.0 1009977.2 2550.4
File Copy 256 bufsize 500 maxblocks 1655.0 264765.9 1599.8
File Copy 4096 bufsize 8000 maxblocks 5800.0 3052236.0 5262.5
System Call Overhead 15000.0 8237404.4 5491.6
========
System Benchmarks Index Score (Partial Only) 3295.3

Signed-off-by: chenzhiyin <zhiyin.chen@intel.com>
Message-Id: <20230601092400.27162-1-zhiyin.chen@intel.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff a7bc2e8d Thu Jun 01 03:24:00 MDT 2023 chenzhiyin <zhiyin.chen@intel.com> fs.h: Optimize file struct to prevent false sharing

In the syscall test of UnixBench, performance regression occurred due
to false sharing.

The lock and atomic members, including file::f_lock, file::f_count and
file::f_pos_lock are highly contended and frequently updated in the
high-concurrency test scenarios. perf c2c indentified one affected
read access, file::f_op.
To prevent false sharing, the layout of file struct is changed as
following
(A) f_lock, f_count and f_pos_lock are put together to share the same
cache line.
(B) The read mostly members, including f_path, f_inode, f_op are put
into a separate cache line.
(C) f_mode is put together with f_count, since they are used frequently
at the same time.
Due to '__randomize_layout' attribute of file struct, the updated layout
only can be effective when CONFIG_RANDSTRUCT_NONE is 'y'.

The optimization has been validated in the syscall test of UnixBench.
performance gain is 30~50%. Furthermore, to confirm the optimization
effectiveness on the other codes path, the results of fsdisk, fsbuffer
and fstime are also shown.

Here are the detailed test results of unixbench.

Command: numactl -C 3-18 ./Run -c 16 syscall fsbuffer fstime fsdisk

Without Patch
------------------------------------------------------------------------
File Copy 1024 bufsize 2000 maxblocks 875052.1 KBps (30.0 s, 2 samples)
File Copy 256 bufsize 500 maxblocks 235484.0 KBps (30.0 s, 2 samples)
File Copy 4096 bufsize 8000 maxblocks 2815153.5 KBps (30.0 s, 2 samples)
System Call Overhead 5772268.3 lps (10.0 s, 7 samples)

System Benchmarks Partial Index BASELINE RESULT INDEX
File Copy 1024 bufsize 2000 maxblocks 3960.0 875052.1 2209.7
File Copy 256 bufsize 500 maxblocks 1655.0 235484.0 1422.9
File Copy 4096 bufsize 8000 maxblocks 5800.0 2815153.5 4853.7
System Call Overhead 15000.0 5772268.3 3848.2
========
System Benchmarks Index Score (Partial Only) 2768.3

With Patch
------------------------------------------------------------------------
File Copy 1024 bufsize 2000 maxblocks 1009977.2 KBps (30.0 s, 2 samples)
File Copy 256 bufsize 500 maxblocks 264765.9 KBps (30.0 s, 2 samples)
File Copy 4096 bufsize 8000 maxblocks 3052236.0 KBps (30.0 s, 2 samples)
System Call Overhead 8237404.4 lps (10.0 s, 7 samples)

System Benchmarks Partial Index BASELINE RESULT INDEX
File Copy 1024 bufsize 2000 maxblocks 3960.0 1009977.2 2550.4
File Copy 256 bufsize 500 maxblocks 1655.0 264765.9 1599.8
File Copy 4096 bufsize 8000 maxblocks 5800.0 3052236.0 5262.5
System Call Overhead 15000.0 8237404.4 5491.6
========
System Benchmarks Index Score (Partial Only) 3295.3

Signed-off-by: chenzhiyin <zhiyin.chen@intel.com>
Message-Id: <20230601092400.27162-1-zhiyin.chen@intel.com>
Signed-off-by: Christian Brauner <brauner@kernel.org>
diff 17e81022 Fri Dec 30 14:52:52 MST 2022 Yu Zhao <yuzhao@google.com> mm: support POSIX_FADV_NOREUSE

This patch adds POSIX_FADV_NOREUSE to vma_has_recency() so that the LRU
algorithm can ignore access to mapped files marked by this flag.

The advantages of POSIX_FADV_NOREUSE are:
1. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not alter the
default readahead behavior.
2. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not split VMAs and
therefore does not take mmap_lock.
3. Unlike MADV_COLD, setting it has a negligible cost, regardless of
how many pages it affects.

Its limitations are:
1. Like POSIX_FADV_RANDOM and POSIX_FADV_SEQUENTIAL, it currently does
not support range. IOW, its scope is the entire file.
2. It currently does not ignore access through file descriptors.
Specifically, for the active/inactive LRU, given a file page shared
by two users and one of them having set POSIX_FADV_NOREUSE on the
file, this page will be activated upon the second user accessing
it. This corner case can be covered by checking POSIX_FADV_NOREUSE
before calling folio_mark_accessed() on the read path. But it is
considered not worth the effort.

There have been a few attempts to support POSIX_FADV_NOREUSE, e.g., [1].
This time the goal is to fill a niche: a few desktop applications, e.g.,
large file transferring and video encoding/decoding, want fast file
streaming with mmap() rather than direct IO. Among those applications, an
SVT-AV1 regression was reported when running with MGLRU [2]. The
following test can reproduce that regression.

kb=$(awk '/MemTotal/ { print $2 }' /proc/meminfo)
kb=$((kb - 8*1024*1024))

modprobe brd rd_nr=1 rd_size=$kb
dd if=/dev/zero of=/dev/ram0 bs=1M

mkfs.ext4 /dev/ram0
mount /dev/ram0 /mnt/
swapoff -a

fallocate -l 8G /mnt/swapfile
mkswap /mnt/swapfile
swapon /mnt/swapfile

wget http://ultravideo.cs.tut.fi/video/Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
7z e -o/mnt/ Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
SvtAv1EncApp --preset 12 -w 3840 -h 2160 \
-i /mnt/Bosphorus_3840x2160.y4m

For MGLRU, the following change showed a [9-11]% increase in FPS,
which makes it on par with the active/inactive LRU.

patch Source/App/EncApp/EbAppMain.c <<EOF
31a32
> #include <fcntl.h>
35d35
< #include <fcntl.h> /* _O_BINARY */
117a118
> posix_fadvise(config->mmap.fd, 0, 0, POSIX_FADV_NOREUSE);
EOF

[1] https://lore.kernel.org/r/1308923350-7932-1-git-send-email-andrea@betterlinux.com/
[2] https://openbenchmarking.org/result/2209259-PTS-MGLRU8GB57

Link: https://lkml.kernel.org/r/20221230215252.2628425-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Righi <andrea.righi@canonical.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
diff 17e81022 Fri Dec 30 14:52:52 MST 2022 Yu Zhao <yuzhao@google.com> mm: support POSIX_FADV_NOREUSE

This patch adds POSIX_FADV_NOREUSE to vma_has_recency() so that the LRU
algorithm can ignore access to mapped files marked by this flag.

The advantages of POSIX_FADV_NOREUSE are:
1. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not alter the
default readahead behavior.
2. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not split VMAs and
therefore does not take mmap_lock.
3. Unlike MADV_COLD, setting it has a negligible cost, regardless of
how many pages it affects.

Its limitations are:
1. Like POSIX_FADV_RANDOM and POSIX_FADV_SEQUENTIAL, it currently does
not support range. IOW, its scope is the entire file.
2. It currently does not ignore access through file descriptors.
Specifically, for the active/inactive LRU, given a file page shared
by two users and one of them having set POSIX_FADV_NOREUSE on the
file, this page will be activated upon the second user accessing
it. This corner case can be covered by checking POSIX_FADV_NOREUSE
before calling folio_mark_accessed() on the read path. But it is
considered not worth the effort.

There have been a few attempts to support POSIX_FADV_NOREUSE, e.g., [1].
This time the goal is to fill a niche: a few desktop applications, e.g.,
large file transferring and video encoding/decoding, want fast file
streaming with mmap() rather than direct IO. Among those applications, an
SVT-AV1 regression was reported when running with MGLRU [2]. The
following test can reproduce that regression.

kb=$(awk '/MemTotal/ { print $2 }' /proc/meminfo)
kb=$((kb - 8*1024*1024))

modprobe brd rd_nr=1 rd_size=$kb
dd if=/dev/zero of=/dev/ram0 bs=1M

mkfs.ext4 /dev/ram0
mount /dev/ram0 /mnt/
swapoff -a

fallocate -l 8G /mnt/swapfile
mkswap /mnt/swapfile
swapon /mnt/swapfile

wget http://ultravideo.cs.tut.fi/video/Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
7z e -o/mnt/ Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
SvtAv1EncApp --preset 12 -w 3840 -h 2160 \
-i /mnt/Bosphorus_3840x2160.y4m

For MGLRU, the following change showed a [9-11]% increase in FPS,
which makes it on par with the active/inactive LRU.

patch Source/App/EncApp/EbAppMain.c <<EOF
31a32
> #include <fcntl.h>
35d35
< #include <fcntl.h> /* _O_BINARY */
117a118
> posix_fadvise(config->mmap.fd, 0, 0, POSIX_FADV_NOREUSE);
EOF

[1] https://lore.kernel.org/r/1308923350-7932-1-git-send-email-andrea@betterlinux.com/
[2] https://openbenchmarking.org/result/2209259-PTS-MGLRU8GB57

Link: https://lkml.kernel.org/r/20221230215252.2628425-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Righi <andrea.righi@canonical.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
diff 17e81022 Fri Dec 30 14:52:52 MST 2022 Yu Zhao <yuzhao@google.com> mm: support POSIX_FADV_NOREUSE

This patch adds POSIX_FADV_NOREUSE to vma_has_recency() so that the LRU
algorithm can ignore access to mapped files marked by this flag.

The advantages of POSIX_FADV_NOREUSE are:
1. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not alter the
default readahead behavior.
2. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not split VMAs and
therefore does not take mmap_lock.
3. Unlike MADV_COLD, setting it has a negligible cost, regardless of
how many pages it affects.

Its limitations are:
1. Like POSIX_FADV_RANDOM and POSIX_FADV_SEQUENTIAL, it currently does
not support range. IOW, its scope is the entire file.
2. It currently does not ignore access through file descriptors.
Specifically, for the active/inactive LRU, given a file page shared
by two users and one of them having set POSIX_FADV_NOREUSE on the
file, this page will be activated upon the second user accessing
it. This corner case can be covered by checking POSIX_FADV_NOREUSE
before calling folio_mark_accessed() on the read path. But it is
considered not worth the effort.

There have been a few attempts to support POSIX_FADV_NOREUSE, e.g., [1].
This time the goal is to fill a niche: a few desktop applications, e.g.,
large file transferring and video encoding/decoding, want fast file
streaming with mmap() rather than direct IO. Among those applications, an
SVT-AV1 regression was reported when running with MGLRU [2]. The
following test can reproduce that regression.

kb=$(awk '/MemTotal/ { print $2 }' /proc/meminfo)
kb=$((kb - 8*1024*1024))

modprobe brd rd_nr=1 rd_size=$kb
dd if=/dev/zero of=/dev/ram0 bs=1M

mkfs.ext4 /dev/ram0
mount /dev/ram0 /mnt/
swapoff -a

fallocate -l 8G /mnt/swapfile
mkswap /mnt/swapfile
swapon /mnt/swapfile

wget http://ultravideo.cs.tut.fi/video/Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
7z e -o/mnt/ Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
SvtAv1EncApp --preset 12 -w 3840 -h 2160 \
-i /mnt/Bosphorus_3840x2160.y4m

For MGLRU, the following change showed a [9-11]% increase in FPS,
which makes it on par with the active/inactive LRU.

patch Source/App/EncApp/EbAppMain.c <<EOF
31a32
> #include <fcntl.h>
35d35
< #include <fcntl.h> /* _O_BINARY */
117a118
> posix_fadvise(config->mmap.fd, 0, 0, POSIX_FADV_NOREUSE);
EOF

[1] https://lore.kernel.org/r/1308923350-7932-1-git-send-email-andrea@betterlinux.com/
[2] https://openbenchmarking.org/result/2209259-PTS-MGLRU8GB57

Link: https://lkml.kernel.org/r/20221230215252.2628425-2-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andrea Righi <andrea.righi@canonical.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

Completed in 1076 milliseconds