Commit a94e065726 upstream.
The current approach is a bit naive, and hence calls the time querying
way too often. Only start the "doing work" timer when there's actual
work to do, and then use that information to terminate (and account) the
work time once done. This greatly reduces the frequency of these calls,
when they cannot have changed anyway.
Running a basic random reader that is setup to use SQPOLL, a profile
before this change shows these as the top cycle consumers:
+ 32.60% iou-sqp-1074 [kernel.kallsyms] [k] thread_group_cputime_adjusted
+ 19.97% iou-sqp-1074 [kernel.kallsyms] [k] thread_group_cputime
+ 12.20% io_uring io_uring [.] submitter_uring_fn
+ 4.13% iou-sqp-1074 [kernel.kallsyms] [k] getrusage
+ 2.45% iou-sqp-1074 [kernel.kallsyms] [k] io_submit_sqes
+ 2.18% iou-sqp-1074 [kernel.kallsyms] [k] __pi_memset_generic
+ 2.09% iou-sqp-1074 [kernel.kallsyms] [k] cputime_adjust
and after this change, top of profile looks as follows:
+ 36.23% io_uring io_uring [.] submitter_uring_fn
+ 23.26% iou-sqp-819 [kernel.kallsyms] [k] io_sq_thread
+ 10.14% iou-sqp-819 [kernel.kallsyms] [k] io_sq_tw
+ 6.52% iou-sqp-819 [kernel.kallsyms] [k] tctx_task_work_run
+ 4.82% iou-sqp-819 [kernel.kallsyms] [k] nvme_submit_cmds.part.0
+ 2.91% iou-sqp-819 [kernel.kallsyms] [k] io_submit_sqes
[...]
0.02% iou-sqp-819 [kernel.kallsyms] [k] cputime_adjust
where it's spending the cycles on things that actually matter.
Reported-by: Fengnan Chang <changfengnan@bytedance.com>
Cc: stable@vger.kernel.org
Fixes: 3fcb9d1720 ("io_uring/sqpoll: statistics of the true utilization of sq threads")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 8ac9b0d33e upstream.
getrusage() does a lot more than what the SQPOLL accounting needs, the
latter only cares about (and uses) the stime. Rather than do a full
RUSAGE_SELF summation, just query the used stime instead.
Cc: stable@vger.kernel.org
Fixes: 3fcb9d1720 ("io_uring/sqpoll: statistics of the true utilization of sq threads")
Reviewed-by: Gabriel Krisman Bertazi <krisman@suse.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit c5efc6a0b3 ]
The __must_hold annotation references &req->ctx->uring_lock, but req
is not in scope in io_install_fixed_file. This change updates the
annotation to reference the correct ctx->uring_lock.
improving code clarity.
Fixes: f110ed8498 ("io_uring: split out fixed file installation and removal")
Signed-off-by: Alok Tiwari <alok.a.tiwari@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
Commit 927069c4ac upstream.
This reverts commit 90bfb28d5f.
Kevin reports that this commit causes an issue for him with LVM
snapshots, most likely because of turning off NOWAIT support while a
snapshot is being created. This makes -EOPNOTSUPP bubble back through
the completion handler, where io_uring read/write handling should just
retry it.
Reinstate the previous check removed by the referenced commit.
Cc: stable@vger.kernel.org
Fixes: 90bfb28d5f ("io_uring/rw: drop -EOPNOTSUPP check in __io_complete_rw_common()")
Reported-by: Salvatore Bonaccorso <carnil@debian.org>
Reported-by: Kevin Lumik <kevin@xf.ee>
Link: https://lore.kernel.org/io-uring/cceb723c-051b-4de2-9a4c-4aa82e1619ee@kernel.dk/
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 2c139a47ef ]
In io_link_skb function, there is a bug where prev_notif is incorrectly
assigned using 'nd' instead of 'prev_nd'. This causes the context
validation check to compare the current notification with itself instead
of comparing it with the previous notification.
Fix by using the correct prev_nd parameter when obtaining prev_notif.
Signed-off-by: Yang Xiuwei <yangxiuwei@kylinos.cn>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Fixes: 6fe4220912 ("io_uring/notif: implement notification stacking")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
Partially based on commit 98b6fa62c8 upstream.
This can be triggered by userspace, so just drop it. The condition
is appropriately handled.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit df8922afc3 upstream.
A recent commit:
fc582cd26e ("io_uring/msg_ring: ensure io_kiocb freeing is deferred for RCU")
fixed an issue with not deferring freeing of io_kiocb structs that
msg_ring allocates to after the current RCU grace period. But this only
covers requests that don't end up in the allocation cache. If a request
goes into the alloc cache, it can get reused before it is sane to do so.
A recent syzbot report would seem to indicate that there's something
there, however it may very well just be because of the KASAN poisoning
that the alloc_cache handles manually.
Rather than attempt to make the alloc_cache sane for that use case, just
drop the usage of the alloc_cache for msg_ring request payload data.
Fixes: 50cf5f3842 ("io_uring/msg_ring: add an alloc cache for io_kiocb entries")
Link: https://lore.kernel.org/io-uring/68cc2687.050a0220.139b6.0005.GAE@google.com/
Reported-by: syzbot+baa2e0f4e02df602583e@syzkaller.appspotmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 3539b1467e upstream.
When running task_work for an exiting task, rather than perform the
issue retry attempt, the task_work is canceled. However, this isn't
done for a ring that has been closed. This can lead to requests being
successfully completed post the ring being closed, which is somewhat
confusing and surprising to an application.
Rather than just check the task exit state, also include the ring
ref state in deciding whether or not to terminate a given request when
run from task_work.
Cc: stable@vger.kernel.org # 6.1+
Link: https://github.com/axboe/liburing/discussions/1459
Reported-by: Benedek Thaler <thaler@thaler.hu>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Parts of commit b6f58a3f4a upstream.
Backport io_should_terminate_tw() helper to judge whether task_work
should be run or terminated.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit df3b8ca604 upstream.
When the taks that submitted a request is dying, a task work for that
request might get run by a kernel thread or even worse by a half
dismantled task. We can't just cancel the task work without running the
callback as the cmd might need to do some clean up, so pass a flag
instead. If set, it's not safe to access any task resources and the
callback is expected to cancel the cmd ASAP.
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: David Sterba <dsterba@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit fc582cd26e upstream.
syzbot reports that defer/local task_work adding via msg_ring can hit
a request that has been freed:
CPU: 1 UID: 0 PID: 19356 Comm: iou-wrk-19354 Not tainted 6.16.0-rc4-syzkaller-00108-g17bbde2e1716 #0 PREEMPT(full)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 05/07/2025
Call Trace:
<TASK>
dump_stack_lvl+0x189/0x250 lib/dump_stack.c:120
print_address_description mm/kasan/report.c:408 [inline]
print_report+0xd2/0x2b0 mm/kasan/report.c:521
kasan_report+0x118/0x150 mm/kasan/report.c:634
io_req_local_work_add io_uring/io_uring.c:1184 [inline]
__io_req_task_work_add+0x589/0x950 io_uring/io_uring.c:1252
io_msg_remote_post io_uring/msg_ring.c:103 [inline]
io_msg_data_remote io_uring/msg_ring.c:133 [inline]
__io_msg_ring_data+0x820/0xaa0 io_uring/msg_ring.c:151
io_msg_ring_data io_uring/msg_ring.c:173 [inline]
io_msg_ring+0x134/0xa00 io_uring/msg_ring.c:314
__io_issue_sqe+0x17e/0x4b0 io_uring/io_uring.c:1739
io_issue_sqe+0x165/0xfd0 io_uring/io_uring.c:1762
io_wq_submit_work+0x6e9/0xb90 io_uring/io_uring.c:1874
io_worker_handle_work+0x7cd/0x1180 io_uring/io-wq.c:642
io_wq_worker+0x42f/0xeb0 io_uring/io-wq.c:696
ret_from_fork+0x3fc/0x770 arch/x86/kernel/process.c:148
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
</TASK>
which is supposed to be safe with how requests are allocated. But msg
ring requests alloc and free on their own, and hence must defer freeing
to a sane time.
Add an rcu_head and use kfree_rcu() in both spots where requests are
freed. Only the one in io_msg_tw_complete() is strictly required as it
has been visible on the other ring, but use it consistently in the other
spot as well.
This should not cause any other issues outside of KASAN rightfully
complaining about it.
Link: https://lore.kernel.org/io-uring/686cd2ea.a00a0220.338033.0007.GAE@google.com/
Reported-by: syzbot+54cbbfb4db9145d26fc2@syzkaller.appspotmail.com
Cc: stable@vger.kernel.org
Fixes: 0617bb500b ("io_uring/msg_ring: improve handling of target CQE posting")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
(cherry picked from commit fc582cd26e)
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 508c1314b3 upstream.
The io_futex_data is allocated upfront and assigned to the io_kiocb
async_data field, but the request isn't marked with REQ_F_ASYNC_DATA
at that point. Those two should always go together, as the flag tells
io_uring whether the field is valid or not.
Additionally, on failure cleanup, the futex handler frees the data but
does not clear ->async_data. Clear the data and the flag in the error
path as well.
Thanks to Trend Micro Zero Day Initiative and particularly ReDress for
reporting this.
Cc: stable@vger.kernel.org
Fixes: 194bb58c60 ("io_uring: add support for futex wake and wait")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 41b70df5b3 upstream.
Ring provided buffers are potentially only valid within the single
execution context in which they were acquired. io_uring deals with this
and invalidates them on retry. But on the networking side, if
MSG_WAITALL is set, or if the socket is of the streaming type and too
little was processed, then it will hang on to the buffer rather than
recycle or commit it. This is problematic for two reasons:
1) If someone unregisters the provided buffer ring before a later retry,
then the req->buf_list will no longer be valid.
2) If multiple sockers are using the same buffer group, then multiple
receives can consume the same memory. This can cause data corruption
in the application, as either receive could land in the same
userspace buffer.
Fix this by disallowing partial retries from pinning a provided buffer
across multiple executions, if ring provided buffers are used.
Cc: stable@vger.kernel.org
Reported-by: pt x <superman.xpt@gmail.com>
Fixes: c56e022c0a ("io_uring: add support for user mapped provided buffer ring")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 825aea662b upstream.
kernel test robot reports that a recent change of the sqe->rw_flags
field throws a sparse warning on 32-bit archs:
>> io_uring/rw.c:291:19: sparse: sparse: incorrect type in assignment (different base types) @@ expected restricted __kernel_rwf_t [usertype] flags @@ got unsigned int @@
io_uring/rw.c:291:19: sparse: expected restricted __kernel_rwf_t [usertype] flags
io_uring/rw.c:291:19: sparse: got unsigned int
Force cast it to rwf_t to silence that new sparse warning.
Fixes: cf73d9970e ("io_uring: don't use int for ABI")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202507032211.PwSNPNSP-lkp@intel.com/
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit c7cafd5b81 upstream.
8c8492ca64 ("io_uring/net: don't retry connect operation on EPOLLERR")
is a little dirty hack that
1) wrongfully assumes that POLLERR equals to a failed request, which
breaks all POLLERR users, e.g. all error queue recv interfaces.
2) deviates the connection request behaviour from connect(2), and
3) racy and solved at a wrong level.
Nothing can be done with 2) now, and 3) is beyond the scope of the
patch. At least solve 1) by moving the hack out of generic poll handling
into io_connect().
Cc: stable@vger.kernel.org
Fixes: 8c8492ca64 ("io_uring/net: don't retry connect operation on EPOLLERR")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/3dc89036388d602ebd84c28e5042e457bdfc952b.1752682444.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
A previous commit aborted mapping more for a non-incremental ring for
bundle peeking, but depending on where in the process this peeking
happened, it would not necessarily prevent a retry by the user. That can
create gaps in the received/read data.
Add struct buf_sel_arg->partial_map, which can pass this information
back. The networking side can then map that to internal state and use it
to gate retry as well.
Since this necessitates a new flag, change io_sr_msg->retry to a
retry_flags member, and store both the retry and partial map condition
in there.
Cc: stable@vger.kernel.org
Fixes: 26ec15e4b0 ("io_uring/kbuf: don't truncate end buffer for multiple buffer peeks")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
(cherry picked from commit 178b8ff66f)
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 9a709b7e98 upstream.
A bigger array of vecs could've been allocated, but
io_ring_buffers_peek() still decided to cap the mapped range depending
on how much data was available. Hence don't rely on the segment count
to know if the request should be marked as needing cleanup, always
check upfront if the iov array is different than the fast_iov array.
Fixes: 26ec15e4b0 ("io_uring/kbuf: don't truncate end buffer for multiple buffer peeks")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
A previous fix corrected the retry condition for when to continue a
current bundle, but it missed that the current (not the total) transfer
count also applies to the buffer put. If not, then for incrementally
consumed buffer rings repeated completions on the same request may end
up over consuming.
Reported-by: Roy Tang (ErgoniaTrading) <royonia@ergonia.io>
Cc: stable@vger.kernel.org
Fixes: 3a08988123 ("io_uring/net: only retry recv bundle for a full transfer")
Link: https://github.com/axboe/liburing/issues/1423
Signed-off-by: Jens Axboe <axboe@kernel.dk>
(cherry picked from commit 51a4598ad5)
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 2c7f023219 upstream.
Currently retry and general validity of msg_inq is gated on it being
larger than zero, but it's entirely possible for this to be slightly
inaccurate. In particular, if FIN is received, it'll return 1.
Just use larger than 1 as the check. This covers both the FIN case, and
at the same time, it doesn't make much sense to retry a recv immediately
if there's even just a single 1 byte of valid data in the socket.
Leave the SOCK_NONEMPTY flagging when larger than 0 still, as an app may
use that for the final receive.
Cc: stable@vger.kernel.org
Reported-by: Christian Mazakas <christian.mazakas@gmail.com>
Fixes: 7c71a0af81 ("io_uring/net: improve recv bundles")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 3a08988123 upstream.
If a shorter than assumed transfer was seen, a partial buffer will have
been filled. For that case it isn't sane to attempt to fill more into
the bundle before posting a completion, as that will cause a gap in
the received data.
Check if the iterator has hit zero and only allow to continue a bundle
operation if that is the case.
Also ensure that for putting finished buffers, only the current transfer
is accounted. Otherwise too many buffers may be put for a short transfer.
Link: https://github.com/axboe/liburing/issues/1409
Cc: stable@vger.kernel.org
Fixes: 7c71a0af81 ("io_uring/net: improve recv bundles")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Commit 7c71a0af81 upstream.
Current recv bundles are only supported for multishot receives, and
additionally they also always post at least 2 CQEs if more data is
available than what a buffer will hold. This happens because the initial
bundle recv will do a single buffer, and then do the rest of what is in
the socket as a followup receive. As shown in a test program, if 1k
buffers are available and 32k is available to receive in the socket,
you'd get the following completions:
bundle=1, mshot=0
cqe res 1024
cqe res 1024
[...]
cqe res 1024
bundle=1, mshot=1
cqe res 1024
cqe res 31744
where bundle=1 && mshot=0 will post 32 1k completions, and bundle=1 &&
mshot=1 will post a 1k completion and then a 31k completion.
To support bundle recv without multishot, it's possible to simply retry
the recv immediately and post a single completion, rather than split it
into two completions. With the below patch, the same test looks as
follows:
bundle=1, mshot=0
cqe res 32768
bundle=1, mshot=1
cqe res 32768
where mshot=0 works fine for bundles, and both of them post just a
single 32k completion rather than split it into separate completions.
Posting fewer completions is always a nice win, and not needing
multishot for proper bundle efficiency is nice for cases that can't
necessarily use multishot.
Reported-by: Norman Maurer <norman_maurer@apple.com>
Link: https://lore.kernel.org/r/184f9f92-a682-4205-a15d-89e18f664502@kernel.dk
Fixes: 2f9c9515bd ("io_uring/net: support bundles for recv")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit f2320f1dd6 ]
A recent commit moved the error handling of sqpoll thread and tctx
failures into the thread itself, as part of fixing an issue. However, it
missed that tctx allocation may also fail, and that
io_sq_offload_create() does its own error handling for the task_struct
in that case.
Remove the manual task putting in io_sq_offload_create(), as
io_sq_thread() will notice that the tctx did not get setup and hence it
should put itself and exit.
Reported-by: syzbot+763e12bbf004fb1062e4@syzkaller.appspotmail.com
Fixes: ac0b8b327a ("io_uring: fix use-after-free of sq->thread in __io_uring_show_fdinfo()")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 26ec15e4b0 upstream.
If peeking a bunch of buffers, normally io_ring_buffers_peek() will
truncate the end buffer. This isn't optimal as presumably more data will
be arriving later, and hence it's better to stop with the last full
buffer rather than truncate the end buffer.
Cc: stable@vger.kernel.org
Fixes: 35c8711c8f ("io_uring/kbuf: add helpers for getting/peeking multiple buffers")
Reported-by: Christian Mazakas <christian.mazakas@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit c538f400fa ]
The sqpoll thread is dereferenced with rcu read protection in one place,
so it needs to be annotated as an __rcu type, and should consistently
use rcu helpers for access and assignment to make sparse happy.
Since most of the accesses occur under the sqd->lock, we can use
rcu_dereference_protected() without declaring an rcu read section.
Provide a simple helper to get the thread from a locked context.
Fixes: ac0b8b327a ("io_uring: fix use-after-free of sq->thread in __io_uring_show_fdinfo()")
Signed-off-by: Keith Busch <kbusch@kernel.org>
Link: https://lore.kernel.org/r/20250611205343.1821117-1-kbusch@meta.com
[axboe: fold in fix for register.c]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
Commit b53e523261 upstream.
There are a few spots where linked timeouts are armed, and not all of
them adhere to the pre-arm, attempt issue, post-arm pattern. This can
be problematic if the linked request returns that it will trigger a
callback later, and does so before the linked timeout is fully armed.
Consolidate all the linked timeout handling into __io_issue_sqe(),
rather than have it spread throughout the various issue entry points.
Cc: stable@vger.kernel.org
Link: https://github.com/axboe/liburing/issues/1390
Reported-by: Chase Hiltz <chase@path.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 92835cebab ]
Our QA team reported a 10%-23%, throughput reduction on an io_uring
sqpoll testcase doing IO to a null_blk, that I traced back to a
reduction of the device submission queue depth utilization. It turns out
that, after commit af5d68f889 ("io_uring/sqpoll: manage task_work
privately"), we capped the number of task_work entries that can be
completed from a single spin of sqpoll to only 8 entries, before the
sqpoll goes around to (potentially) sleep. While this cap doesn't drive
the submission side directly, it impacts the completion behavior, which
affects the number of IO queued by fio per sqpoll cycle on the
submission side, and io_uring ends up seeing less ios per sqpoll cycle.
As a result, block layer plugging is less effective, and we see more
time spent inside the block layer in profilings charts, and increased
submission latency measured by fio.
There are other places that have increased overhead once sqpoll sleeps
more often, such as the sqpoll utilization calculation. But, in this
microbenchmark, those were not representative enough in perf charts, and
their removal didn't yield measurable changes in throughput. The major
overhead comes from the fact we plug less, and less often, when submitting
to the block layer.
My benchmark is:
fio --ioengine=io_uring --direct=1 --iodepth=128 --runtime=300 --bs=4k \
--invalidate=1 --time_based --ramp_time=10 --group_reporting=1 \
--filename=/dev/nullb0 --name=RandomReads-direct-nullb-sqpoll-4k-1 \
--rw=randread --numjobs=1 --sqthread_poll
In one machine, tested on top of Linux 6.15-rc1, we have the following
baseline:
READ: bw=4994MiB/s (5236MB/s), 4994MiB/s-4994MiB/s (5236MB/s-5236MB/s), io=439GiB (471GB), run=90001-90001msec
With this patch:
READ: bw=5762MiB/s (6042MB/s), 5762MiB/s-5762MiB/s (6042MB/s-6042MB/s), io=506GiB (544GB), run=90001-90001msec
which is a 15% improvement in measured bandwidth. The average
submission latency is noticeably lowered too. As measured by
fio:
Baseline:
lat (usec): min=20, max=241, avg=99.81, stdev=3.38
Patched:
lat (usec): min=26, max=226, avg=86.48, stdev=4.82
If we look at blktrace, we can also see the plugging behavior is
improved. In the baseline, we end up limited to plugging 8 requests in
the block layer regardless of the device queue depth size, while after
patching we can drive more io, and we manage to utilize the full device
queue.
In the baseline, after a stabilization phase, an ordinary submission
looks like:
254,0 1 49942 0.016028795 5977 U N [iou-sqp-5976] 7
After patching, I see consistently more requests per unplug.
254,0 1 4996 0.001432872 3145 U N [iou-sqp-3144] 32
Ideally, the cap size would at least be the deep enough to fill the
device queue, but we can't predict that behavior, or assume all IO goes
to a single device, and thus can't guess the ideal batch size. We also
don't want to let the tw run unbounded, though I'm not sure it would
really be a problem. Instead, let's just give it a more sensible value
that will allow for more efficient batching. I've tested with different
cap values, and initially proposed to increase the cap to 1024. Jens
argued it is too big of a bump and I observed that, with 32, I'm no
longer able to observe this bottleneck in any of my machines.
Fixes: af5d68f889 ("io_uring/sqpoll: manage task_work privately")
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://lore.kernel.org/r/20250508181203.3785544-1-krisman@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 687b2bae0e upstream.
Multishot normally uses io_req_post_cqe() to post completions, but when
stopping it, it may finish up with a deferred completion. This is fine,
except if another multishot event triggers before the deferred completions
get flushed. If this occurs, then CQEs may get reordered in the CQ ring,
as new multishot completions get posted before the deferred ones are
flushed. This can cause confusion on the application side, if strict
ordering is required for the use case.
When multishot posting via io_req_post_cqe(), flush any pending deferred
completions first, if any.
Cc: stable@vger.kernel.org # 6.1+
Reported-by: Norman Maurer <norman_maurer@apple.com>
Reported-by: Christian Mazakas <christian.mazakas@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 390513642e ]
io_uring always switches requests to atomic refcounting for iowq
execution before there is any parallilism by setting REQ_F_REFCOUNT,
and the flag is not cleared until the request completes. That should be
fine as long as the compiler doesn't make up a non existing value for
the flags, however KCSAN still complains when the request owner changes
oter flag bits:
BUG: KCSAN: data-race in io_req_task_cancel / io_wq_free_work
...
read to 0xffff888117207448 of 8 bytes by task 3871 on cpu 0:
req_ref_put_and_test io_uring/refs.h:22 [inline]
Skip REQ_F_REFCOUNT checks for iowq, we know it's set.
Reported-by: syzbot+903a2ad71fb3f1e47cf5@syzkaller.appspotmail.com
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d880bc27fb8c3209b54641be4ff6ac02b0e5789a.1743679736.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit edd43f4d6f upstream.
A previous commit added a 'sync' parameter to io_fallback_tw(), which if
true, means the caller wants to wait on the fallback thread handling it.
But the logic is somewhat messed up, ensure that ctxs are swapped and
flushed appropriately.
Cc: stable@vger.kernel.org
Fixes: dfbe5561ae ("io_uring: flush offloaded and delayed task_work on exit")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 6889ae1b4d upstream.
[ 114.987980][ T5313] WARNING: CPU: 6 PID: 5313 at io_uring/io_uring.c:872 io_req_post_cqe+0x12e/0x4f0
[ 114.991597][ T5313] RIP: 0010:io_req_post_cqe+0x12e/0x4f0
[ 115.001880][ T5313] Call Trace:
[ 115.002222][ T5313] <TASK>
[ 115.007813][ T5313] io_send+0x4fe/0x10f0
[ 115.009317][ T5313] io_issue_sqe+0x1a6/0x1740
[ 115.012094][ T5313] io_wq_submit_work+0x38b/0xed0
[ 115.013223][ T5313] io_worker_handle_work+0x62a/0x1600
[ 115.013876][ T5313] io_wq_worker+0x34f/0xdf0
As the comment states, io_req_post_cqe() should only be used by
multishot requests, i.e. REQ_F_APOLL_MULTISHOT, which bundled sends are
not. Add a flag signifying whether a request wants to post multiple
CQEs. Eventually REQ_F_APOLL_MULTISHOT should imply the new flag, but
that's left out for simplicity.
Cc: stable@vger.kernel.org
Fixes: a05d1f625c ("io_uring/net: support bundles for send")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/8b611dbb54d1cd47a88681f5d38c84d0c02bc563.1743067183.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 67c007d6c1 upstream.
refcount_t: underflow; use-after-free.
WARNING: CPU: 0 PID: 5823 at lib/refcount.c:28 refcount_warn_saturate+0x15a/0x1d0 lib/refcount.c:28
RIP: 0010:refcount_warn_saturate+0x15a/0x1d0 lib/refcount.c:28
Call Trace:
<TASK>
io_notif_flush io_uring/notif.h:40 [inline]
io_send_zc_cleanup+0x121/0x170 io_uring/net.c:1222
io_clean_op+0x58c/0x9a0 io_uring/io_uring.c:406
io_free_batch_list io_uring/io_uring.c:1429 [inline]
__io_submit_flush_completions+0xc16/0xd20 io_uring/io_uring.c:1470
io_submit_flush_completions io_uring/io_uring.h:159 [inline]
Before the blamed commit, sendzc relied on io_req_msg_cleanup() to clear
REQ_F_NEED_CLEANUP, so after the following snippet the request will
never hit the core io_uring cleanup path.
io_notif_flush();
io_req_msg_cleanup();
The easiest fix is to null the notification. io_send_zc_cleanup() can
still be called after, but it's tolerated.
Reported-by: syzbot+cf285a028ffba71b2ef5@syzkaller.appspotmail.com
Tested-by: syzbot+cf285a028ffba71b2ef5@syzkaller.appspotmail.com
Fixes: cc34d8330e ("io_uring/net: don't clear REQ_F_NEED_CLEANUP unconditionally")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/e1306007458b8891c88c4f20c966a17595f766b0.1742643795.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit cc34d8330e upstream.
io_req_msg_cleanup() relies on the fact that io_netmsg_recycle() will
always fully recycle, but that may not be the case if the msg cache
was already full. To ensure that normal cleanup always gets run,
let io_netmsg_recycle() deal with clearing the relevant cleanup flags,
as it knows exactly when that should be done.
Cc: stable@vger.kernel.org
Reported-by: David Wei <dw@davidwei.uk>
Fixes: 7519134178 ("io_uring/net: add iovec recycling")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
