A previous commit added a trylock for getting the SQPOLL thread info via
fdinfo, but this introduced a regression where we often fail to get it if
the thread is busy. For that case, we end up not printing the current CPU
and PID info.
Rather than rely on this lock, just print the pid we already stored in
the io_sq_data struct, and ensure we update the current CPU every time
we've slept or potentially rescheduled. The latter won't potentially be
100% accurate, but that wasn't the case before either as the task can
get migrated at any time unless it has been pinned at creation time.
We retain keeping the io_sq_data dereference inside the ctx->uring_lock,
as it has always been, as destruction of the thread and data happen below
that. We could make this RCU safe, but there's little point in doing that.
With this, we always print the last valid information we had, rather than
have spurious outputs with missing information.
Fixes: 7644b1a1c9 ("io_uring/fdinfo: lock SQ thread while retrieving thread cpu/pid")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When doing a multishot read, the code path reuses the old read
paths. However this breaks an assumption built into those paths,
namely that struct io_rw::len is available for reuse by __io_import_iovec.
For multishot this results in len being set for the first receive
call, and then subsequent calls are clamped to that buffer length
incorrectly.
Instead keep len as zero after recycling buffers, to reuse the full
buffer size of the next selected buffer.
Fixes: fc68fcda04 ("io_uring/rw: add support for IORING_OP_READ_MULTISHOT")
Signed-off-by: Dylan Yudaken <dyudaken@gmail.com>
Link: https://lore.kernel.org/r/20231106203909.197089-4-dyudaken@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For addr: this field is not used, since buffer select is forced.
But by forcing it to be zero it leaves open future uses of the field.
len is actually usable, you could imagine that you want to receive
multishot up to a certain length.
However right now this is not how it is implemented, and it seems
safer to force this to be zero.
Fixes: fc68fcda04 ("io_uring/rw: add support for IORING_OP_READ_MULTISHOT")
Signed-off-by: Dylan Yudaken <dyudaken@gmail.com>
Link: https://lore.kernel.org/r/20231106203909.197089-3-dyudaken@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than sprinkle opcode checks in the generic read/write prep handler,
have a separate prep handler for the vectored readv/writev operation.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Rather than sprinkle opcode checks in the generic read/write prep handler,
have a separate prep handler for the vectored readv/writev operation.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring does non-blocking connection attempts, which can yield some
unexpected results if a connect request is re-attempted by an an
application. This is equivalent to the following sync syscall sequence:
sock = socket(AF_INET, SOCK_STREAM | SOCK_NONBLOCK, IPPROTO_TCP);
connect(sock, &addr, sizeof(addr);
ret == -1 and errno == EINPROGRESS expected here. Now poll for POLLOUT
on sock, and when that returns, we expect the socket to be connected.
But if we follow that procedure with:
connect(sock, &addr, sizeof(addr));
you'd expect ret == -1 and errno == EISCONN here, but you actually get
ret == 0. If we attempt the connection one more time, then we get EISCON
as expected.
io_uring used to do this, but turns out that bluetooth fails with EBADFD
if you attempt to re-connect. Also looks like EISCONN _could_ occur with
this sequence.
Retain the ->in_progress logic, but work-around a potential EISCONN or
EBADFD error and only in those cases look at the sock_error(). This
should work in general and avoid the odd sequence of a repeated connect
request returning success when the socket is already connected.
This is all a side effect of the socket state being in a CONNECTING
state when we get EINPROGRESS, and only a re-connect or other related
operation will turn that into CONNECTED.
Cc: stable@vger.kernel.org
Fixes: 3fb1bd6881 ("io_uring/net: handle -EINPROGRESS correct for IORING_OP_CONNECT")
Link: https://github.com/axboe/liburing/issues/980
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The new read multishot method doesn't need to allocate async data ever,
as it doesn't do vectored IO and it must only be used with provided
buffers. While it doesn't have ->prep_async() set, it also sets
->async_size to 0, which is different from any other read/write type we
otherwise support.
If it's used on a file type that isn't pollable, we do try and allocate
this async data, and then try and use that data. But since we passed in
a size of 0 for the data, we get a NULL back on data allocation. We then
proceed to dereference that to copy state, and that obviously won't end
well.
Add a check in io_setup_async_rw() for this condition, and avoid copying
state. Also add a check for whether or not buffer selection is specified
in prep while at it.
Fixes: fc68fcda04 ("io_uring/rw: add support for IORING_OP_READ_MULTISHOT")
Link: https://bugzilla.kernel.org/show_bug.cgi?id=218101
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull io_uring futex support from Jens Axboe:
"This adds support for using futexes through io_uring - first futex
wake and wait, and then the vectored variant of waiting, futex waitv.
For both wait/wake/waitv, we support the bitset variant, as the
'normal' variants can be easily implemented on top of that.
PI and requeue are not supported through io_uring, just the above
mentioned parts. This may change in the future, but in the spirit of
keeping this small (and based on what people have been asking for),
this is what we currently have.
Wake support is pretty straight forward, most of the thought has gone
into the wait side to avoid needing to offload wait operations to a
blocking context. Instead, we rely on the usual callbacks to retry and
post a completion event, when appropriate.
As far as I can recall, the first request for futex support with
io_uring came from Andres Freund, working on postgres. His aio rework
of postgres was one of the early adopters of io_uring, and futex
support was a natural extension for that. This is relevant from both a
usability point of view, as well as for effiency and performance. In
Andres's words, for the former:
Futex wait support in io_uring makes it a lot easier to avoid
deadlocks in concurrent programs that have their own buffer pool:
Obviously pages in the application buffer pool have to be locked
during IO. If the initiator of IO A needs to wait for a held lock
B, the holder of lock B might wait for the IO A to complete. The
ability to wait for a lock and IO completions at the same time
provides an efficient way to avoid such deadlocks
and in terms of effiency, even without unlocking the full potential
yet, Andres says:
Futex wake support in io_uring is useful because it allows for more
efficient directed wakeups. For some "locks" postgres has queues
implemented in userspace, with wakeup logic that cannot easily be
implemented with FUTEX_WAKE_BITSET on a single "futex word"
(imagine waiting for journal flushes to have completed up to a
certain point).
Thus a "lock release" sometimes need to wake up many processes in a
row. A quick-and-dirty conversion to doing these wakeups via
io_uring lead to a 3% throughput increase, with 12% fewer context
switches, albeit in a fairly extreme workload"
* tag 'io_uring-futex-2023-10-30' of git://git.kernel.dk/linux:
io_uring: add support for vectored futex waits
futex: make the vectored futex operations available
futex: make futex_parse_waitv() available as a helper
futex: add wake_data to struct futex_q
io_uring: add support for futex wake and wait
futex: abstract out a __futex_wake_mark() helper
futex: factor out the futex wake handling
futex: move FUTEX2_VALID_MASK to futex.h
Pull io_uring {get,set}sockopt support from Jens Axboe:
"This adds support for using getsockopt and setsockopt via io_uring.
The main use cases for this is to enable use of direct descriptors,
rather than first instantiating a normal file descriptor, doing the
option tweaking needed, then turning it into a direct descriptor. With
this support, we can avoid needing a regular file descriptor
completely.
The net and bpf bits have been signed off on their side"
* tag 'for-6.7/io_uring-sockopt-2023-10-30' of git://git.kernel.dk/linux:
selftests/bpf/sockopt: Add io_uring support
io_uring/cmd: Introduce SOCKET_URING_OP_SETSOCKOPT
io_uring/cmd: Introduce SOCKET_URING_OP_GETSOCKOPT
io_uring/cmd: return -EOPNOTSUPP if net is disabled
selftests/net: Extract uring helpers to be reusable
tools headers: Grab copy of io_uring.h
io_uring/cmd: Pass compat mode in issue_flags
net/socket: Break down __sys_getsockopt
net/socket: Break down __sys_setsockopt
bpf: Add sockptr support for setsockopt
bpf: Add sockptr support for getsockopt
Pull io_uring updates from Jens Axboe:
"This contains the core io_uring updates, of which there are not many,
and adds support for using WAITID through io_uring and hence not
needing to block on these kinds of events.
Outside of that, tweaks to the legacy provided buffer handling and
some cleanups related to cancelations for uring_cmd support"
* tag 'for-6.7/io_uring-2023-10-30' of git://git.kernel.dk/linux:
io_uring/poll: use IOU_F_TWQ_LAZY_WAKE for wakeups
io_uring/kbuf: Use slab for struct io_buffer objects
io_uring/kbuf: Allow the full buffer id space for provided buffers
io_uring/kbuf: Fix check of BID wrapping in provided buffers
io_uring/rsrc: cleanup io_pin_pages()
io_uring: cancelable uring_cmd
io_uring: retain top 8bits of uring_cmd flags for kernel internal use
io_uring: add IORING_OP_WAITID support
exit: add internal include file with helpers
exit: add kernel_waitid_prepare() helper
exit: move core of do_wait() into helper
exit: abstract out should_wake helper for child_wait_callback()
io_uring/rw: add support for IORING_OP_READ_MULTISHOT
io_uring/rw: mark readv/writev as vectored in the opcode definition
io_uring/rw: split io_read() into a helper
Pull misc vfs updates from Christian Brauner:
"This contains the usual miscellaneous features, cleanups, and fixes
for vfs and individual fses.
Features:
- Rename and export helpers that get write access to a mount. They
are used in overlayfs to get write access to the upper mount.
- Print the pretty name of the root device on boot failure. This
helps in scenarios where we would usually only print
"unknown-block(1,2)".
- Add an internal SB_I_NOUMASK flag. This is another part in the
endless POSIX ACL saga in a way.
When POSIX ACLs are enabled via SB_POSIXACL the vfs cannot strip
the umask because if the relevant inode has POSIX ACLs set it might
take the umask from there. But if the inode doesn't have any POSIX
ACLs set then we apply the umask in the filesytem itself. So we end
up with:
(1) no SB_POSIXACL -> strip umask in vfs
(2) SB_POSIXACL -> strip umask in filesystem
The umask semantics associated with SB_POSIXACL allowed filesystems
that don't even support POSIX ACLs at all to raise SB_POSIXACL
purely to avoid umask stripping. That specifically means NFS v4 and
Overlayfs. NFS v4 does it because it delegates this to the server
and Overlayfs because it needs to delegate umask stripping to the
upper filesystem, i.e., the filesystem used as the writable layer.
This went so far that SB_POSIXACL is raised eve on kernels that
don't even have POSIX ACL support at all.
Stop this blatant abuse and add SB_I_NOUMASK which is an internal
superblock flag that filesystems can raise to opt out of umask
handling. That should really only be the two mentioned above. It's
not that we want any filesystems to do this. Ideally we have all
umask handling always in the vfs.
- Make overlayfs use SB_I_NOUMASK too.
- Now that we have SB_I_NOUMASK, stop checking for SB_POSIXACL in
IS_POSIXACL() if the kernel doesn't have support for it. This is a
very old patch but it's only possible to do this now with the wider
cleanup that was done.
- Follow-up work on fake path handling from last cycle. Citing mostly
from Amir:
When overlayfs was first merged, overlayfs files of regular files
and directories, the ones that are installed in file table, had a
"fake" path, namely, f_path is the overlayfs path and f_inode is
the "real" inode on the underlying filesystem.
In v6.5, we took another small step by introducing of the
backing_file container and the file_real_path() helper. This change
allowed vfs and filesystem code to get the "real" path of an
overlayfs backing file. With this change, we were able to make
fsnotify work correctly and report events on the "real" filesystem
objects that were accessed via overlayfs.
This method works fine, but it still leaves the vfs vulnerable to
new code that is not aware of files with fake path. A recent
example is commit db1d1e8b98 ("IMA: use vfs_getattr_nosec to get
the i_version"). This commit uses direct referencing to f_path in
IMA code that otherwise uses file_inode() and file_dentry() to
reference the filesystem objects that it is measuring.
This contains work to switch things around: instead of having
filesystem code opt-in to get the "real" path, have generic code
opt-in for the "fake" path in the few places that it is needed.
Is it far more likely that new filesystems code that does not use
the file_dentry() and file_real_path() helpers will end up causing
crashes or averting LSM/audit rules if we keep the "fake" path
exposed by default.
This change already makes file_dentry() moot, but for now we did
not change this helper just added a WARN_ON() in ovl_d_real() to
catch if we have made any wrong assumptions.
After the dust settles on this change, we can make file_dentry() a
plain accessor and we can drop the inode argument to ->d_real().
- Switch struct file to SLAB_TYPESAFE_BY_RCU. This looks like a small
change but it really isn't and I would like to see everyone on
their tippie toes for any possible bugs from this work.
Essentially we've been doing most of what SLAB_TYPESAFE_BY_RCU for
files since a very long time because of the nasty interactions
between the SCM_RIGHTS file descriptor garbage collection. So
extending it makes a lot of sense but it is a subtle change. There
are almost no places that fiddle with file rcu semantics directly
and the ones that did mess around with struct file internal under
rcu have been made to stop doing that because it really was always
dodgy.
I forgot to put in the link tag for this change and the discussion
in the commit so adding it into the merge message:
https://lore.kernel.org/r/20230926162228.68666-1-mjguzik@gmail.com
Cleanups:
- Various smaller pipe cleanups including the removal of a spin lock
that was only used to protect against writes without pipe_lock()
from O_NOTIFICATION_PIPE aka watch queues. As that was never
implemented remove the additional locking from pipe_write().
- Annotate struct watch_filter with the new __counted_by attribute.
- Clarify do_unlinkat() cleanup so that it doesn't look like an extra
iput() is done that would cause issues.
- Simplify file cleanup when the file has never been opened.
- Use module helper instead of open-coding it.
- Predict error unlikely for stale retry.
- Use WRITE_ONCE() for mount expiry field instead of just commenting
that one hopes the compiler doesn't get smart.
Fixes:
- Fix readahead on block devices.
- Fix writeback when layztime is enabled and inodes whose timestamp
is the only thing that changed reside on wb->b_dirty_time. This
caused excessively large zombie memory cgroup when lazytime was
enabled as such inodes weren't handled fast enough.
- Convert BUG_ON() to WARN_ON_ONCE() in open_last_lookups()"
* tag 'vfs-6.7.misc' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs: (26 commits)
file, i915: fix file reference for mmap_singleton()
vfs: Convert BUG_ON to WARN_ON_ONCE in open_last_lookups
writeback, cgroup: switch inodes with dirty timestamps to release dying cgwbs
chardev: Simplify usage of try_module_get()
ovl: rely on SB_I_NOUMASK
fs: fix umask on NFS with CONFIG_FS_POSIX_ACL=n
fs: store real path instead of fake path in backing file f_path
fs: create helper file_user_path() for user displayed mapped file path
fs: get mnt_writers count for an open backing file's real path
vfs: stop counting on gcc not messing with mnt_expiry_mark if not asked
vfs: predict the error in retry_estale as unlikely
backing file: free directly
vfs: fix readahead(2) on block devices
io_uring: use files_lookup_fd_locked()
file: convert to SLAB_TYPESAFE_BY_RCU
vfs: shave work on failed file open
fs: simplify misleading code to remove ambiguity regarding ihold()/iput()
watch_queue: Annotate struct watch_filter with __counted_by
fs/pipe: use spinlock in pipe_read() only if there is a watch_queue
fs/pipe: remove unnecessary spinlock from pipe_write()
...
Pull misc filesystem fixes from Al Viro:
"Assorted fixes all over the place: literally nothing in common, could
have been three separate pull requests.
All are simple regression fixes, but not for anything from this cycle"
* tag 'pull-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
ceph_wait_on_conflict_unlink(): grab reference before dropping ->d_lock
io_uring: kiocb_done() should *not* trust ->ki_pos if ->{read,write}_iter() failed
sparc32: fix a braino in fault handling in csum_and_copy_..._user()
->ki_pos value is unreliable in such cases. For an obvious example,
consider O_DSYNC write - we feed the data to page cache and start IO,
then we make sure it's completed. Update of ->ki_pos is dealt with
by the first part; failure in the second ends up with negative value
returned _and_ ->ki_pos left advanced as if sync had been successful.
In the same situation write(2) does not advance the file position
at all.
Reviewed-by: Christian Brauner <brauner@kernel.org>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
If an application does O_DIRECT writes with io_uring and the file system
supports IOCB_DIO_CALLER_COMP, then completions of the dio write side is
done from the task_work that will post the completion event for said
write as well.
Whenever a dio write is done against a file, the inode i_dio_count is
elevated. This enables other callers to use inode_dio_wait() to wait for
previous writes to complete. If we defer the full dio completion to
task_work, we are dependent on that task_work being run before the
inode i_dio_count can be decremented.
If the same task that issues io_uring dio writes with
IOCB_DIO_CALLER_COMP performs a synchronous system call that calls
inode_dio_wait(), then we can deadlock as we're blocked sleeping on
the event to become true, but not processing the completions that will
result in the inode i_dio_count being decremented.
Until we can guarantee that this is the case, then disable the deferred
caller completions.
Fixes: 099ada2c87 ("io_uring/rw: add write support for IOCB_DIO_CALLER_COMP")
Reported-by: Andres Freund <andres@anarazel.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We could race with SQ thread exit, and if we do, we'll hit a NULL pointer
dereference when the thread is cleared. Grab the SQPOLL data lock before
attempting to get the task cpu and pid for fdinfo, this ensures we have a
stable view of it.
Cc: stable@vger.kernel.org
Link: https://bugzilla.kernel.org/show_bug.cgi?id=218032
Reviewed-by: Gabriel Krisman Bertazi <krisman@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add initial support for SOCKET_URING_OP_SETSOCKOPT. This new command is
similar to setsockopt. This implementation leverages the function
do_sock_setsockopt(), which is shared with the setsockopt() system call
path.
Important to say that userspace needs to keep the pointer's memory alive
until the operation is completed. I.e, the memory could not be
deallocated before the CQE is returned to userspace.
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20231016134750.1381153-11-leitao@debian.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add support for getsockopt command (SOCKET_URING_OP_GETSOCKOPT), where
level is SOL_SOCKET. This is leveraging the sockptr_t infrastructure,
where a sockptr_t is either userspace or kernel space, and handled as
such.
Differently from the getsockopt(2), the optlen field is not a userspace
pointers. In getsockopt(2), userspace provides optlen pointer, which is
overwritten by the kernel. In this implementation, userspace passes a
u32, and the new value is returned in cqe->res. I.e., optlen is not a
pointer.
Important to say that userspace needs to keep the pointer alive until
the CQE is completed.
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20231016134750.1381153-10-leitao@debian.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Protect io_uring_cmd_sock() to be called if CONFIG_NET is not set. If
network is not enabled, but io_uring is, then we want to return
-EOPNOTSUPP for any possible socket operation.
This is helpful because io_uring_cmd_sock() can now call functions that
only exits if CONFIG_NET is enabled without having #ifdef CONFIG_NET
inside the function itself.
Signed-off-by: Breno Leitao <leitao@debian.org>
Link: https://lore.kernel.org/r/20231016134750.1381153-9-leitao@debian.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With poll triggered retries, each event trigger will cause a task_work
item to be added for processing. If the ring is setup with
IORING_SETUP_DEFER_TASKRUN and a task is waiting on multiple events to
complete, any task_work addition will wake the task for processing these
items. This can cause more context switches than we would like, if the
application is deliberately waiting on multiple items to increase
efficiency.
For example, if an application has receive multishot armed for sockets
and wants to wait for N to complete within M usec of time, we should not
be waking up and processing these items until we have all the events we
asked for. By switching the poll trigger to lazy wake, we'll process
them when they are all ready, in one swoop, rather than wake multiple
times only to process one and then go back to sleep.
At some point we probably want to look at just making the lazy wake
the default, but for now, let's just selectively enable it where it
makes sense.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we specify a valid CQ ring address but an invalid SQ ring address,
we'll correctly spot this and free the allocated pages and clear them
to NULL. However, we don't clear the ring page count, and hence will
attempt to free the pages again. We've already cleared the address of
the page array when freeing them, but we don't check for that. This
causes the following crash:
Unable to handle kernel NULL pointer dereference at virtual address 0000000000000000
Oops [#1]
Modules linked in:
CPU: 0 PID: 20 Comm: kworker/u2:1 Not tainted 6.6.0-rc5-dirty #56
Hardware name: ucbbar,riscvemu-bare (DT)
Workqueue: events_unbound io_ring_exit_work
epc : io_pages_free+0x2a/0x58
ra : io_rings_free+0x3a/0x50
epc : ffffffff808811a2 ra : ffffffff80881406 sp : ffff8f80000c3cd0
status: 0000000200000121 badaddr: 0000000000000000 cause: 000000000000000d
[<ffffffff808811a2>] io_pages_free+0x2a/0x58
[<ffffffff80881406>] io_rings_free+0x3a/0x50
[<ffffffff80882176>] io_ring_exit_work+0x37e/0x424
[<ffffffff80027234>] process_one_work+0x10c/0x1f4
[<ffffffff8002756e>] worker_thread+0x252/0x31c
[<ffffffff8002f5e4>] kthread+0xc4/0xe0
[<ffffffff8000332a>] ret_from_fork+0xa/0x1c
Check for a NULL array in io_pages_free(), but also clear the page counts
when we free them to be on the safer side.
Reported-by: rtm@csail.mit.edu
Fixes: 03d89a2de2 ("io_uring: support for user allocated memory for rings/sqes")
Cc: stable@vger.kernel.org
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The allocation of struct io_buffer for metadata of provided buffers is
done through a custom allocator that directly gets pages and
fragments them. But, slab would do just fine, as this is not a hot path
(in fact, it is a deprecated feature) and, by keeping a custom allocator
implementation we lose benefits like tracking, poisoning,
sanitizers. Finally, the custom code is more complex and requires
keeping the list of pages in struct ctx for no good reason. This patch
cleans this path up and just uses slab.
I microbenchmarked it by forcing the allocation of a large number of
objects with the least number of io_uring commands possible (keeping
nbufs=USHRT_MAX), with and without the patch. There is a slight
increase in time spent in the allocation with slab, of course, but even
when allocating to system resources exhaustion, which is not very
realistic and happened around 1/2 billion provided buffers for me, it
wasn't a significant hit in system time. Specially if we think of a
real-world scenario, an application doing register/unregister of
provided buffers will hit ctx->io_buffers_cache more often than actually
going to slab.
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20231005000531.30800-4-krisman@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
