When using SQPOLL, the submission queue polling thread calls
task_work_run() to run queued work. However, when work is added with
TWA_SIGNAL - as done by io_uring itself - the TIF_NOTIFY_SIGNAL remains
set afterwards and is never cleared.
Consequently, when the submission queue polling thread checks whether
signal_pending(), it may always find a pending signal, if
task_work_add() was ever called before.
The impact of this bug might be different on different kernel versions.
It appears that on 5.14 it would only cause unnecessary calculation and
prevent the polling thread from sleeping. On 5.13, where the bug was
found, it stops the polling thread from finding newly submitted work.
Instead of task_work_run(), use tracehook_notify_signal() that clears
TIF_NOTIFY_SIGNAL. Test for TIF_NOTIFY_SIGNAL in addition to
current->task_works to avoid a race in which task_works is cleared but
the TIF_NOTIFY_SIGNAL is set.
Fixes: 685fe7feed ("io-wq: eliminate the need for a manager thread")
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Nadav Amit <namit@vmware.com>
Link: https://lore.kernel.org/r/20210808001342.964634-2-namit@vmware.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For pure poll requests, it doesn't remove the second poll wait entry
when it's done, neither after vfs_poll() or in the poll completion
handler. We should remove the second poll wait entry.
And we use io_poll_remove_double() rather than io_poll_remove_waitqs()
since the latter has some redundant logic.
Fixes: 88e41cf928 ("io_uring: add multishot mode for IORING_OP_POLL_ADD")
Cc: stable@vger.kernel.org # 5.13+
Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
Link: https://lore.kernel.org/r/20210728030322.12307-1-haoxu@linux.alibaba.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some setups, like SCSI, can throw spurious -EAGAIN off the softirq
completion path. Normally we expect this to happen inline as part
of submission, but apparently SCSI has a weird corner case where it
can happen as part of normal completions.
This should be solved by having the -EAGAIN bubble back up the stack
as part of submission, but previous attempts at this failed and we're
not just quite there yet. Instead we currently use REQ_F_REISSUE to
handle this case.
For now, catch it in io_rw_should_reissue() and prevent a reissue
from a bogus path.
Cc: stable@vger.kernel.org
Reported-by: Fabian Ebner <f.ebner@proxmox.com>
Tested-by: Fabian Ebner <f.ebner@proxmox.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As a safeguard, if we're going to queue async work, do it from task_work
from the original task. This ensures that we can always sanely create
threads, regards of what the reissue context may be.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use a bit to manage if we need to add the shared task_work, but
a list + lock for the pending work. Before aborting a current run
of the task_work we check if the list is empty, but we do so without
grabbing the lock that protects it. This can lead to races where
we think we have nothing left to run, where in practice we could be
racing with a task adding new work to the list. If we do hit that
race condition, we could be left with work items that need processing,
but the shared task_work is not active.
Ensure that we grab the lock before checking if the list is empty,
so we know if it's safe to exit the run or not.
Link: https://lore.kernel.org/io-uring/c6bd5987-e9ae-cd02-49d0-1b3ac1ef65b1@tnonline.net/
Cc: stable@vger.kernel.org # 5.11+
Reported-by: Forza <forza@tnonline.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Catch an illegal case to queue async from an unrelated task that got
the ring fd passed to it. This should not be possible to hit, but
better be proactive and catch it explicitly. io-wq is extended to
check for early IO_WQ_WORK_CANCEL being set on a work item as well,
so it can run the request through the normal cancelation path.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are two reasons why this shouldn't be done:
1) Ring is exiting, and we're canceling requests anyway. Any request
should be canceled anyway. In theory, this could iterate for a
number of times if someone else is also driving the target block
queue into request starvation, however the likelihood of this
happening is miniscule.
2) If the original task decided to pass the ring to another task, then
we don't want to be reissuing from this context as it may be an
unrelated task or context. No assumptions should be made about
the context in which ->release() is run. This can only happen for pure
read/write, and we'll get -EFAULT on them anyway.
Link: https://lore.kernel.org/io-uring/YPr4OaHv0iv0KTOc@zeniv-ca.linux.org.uk/
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A previous commit shuffled some code around, and inadvertently used
struct file after fdput() had been called on it. As we can't touch
the file post fdput() dropping our reference, move the fdput() to
after that has been done.
Cc: Pavel Begunkov <asml.silence@gmail.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/io-uring/YPnqM0fY3nM5RdRI@zeniv-ca.linux.org.uk/
Fixes: f2a48dd09b ("io_uring: refactor io_sq_offload_create()")
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
__io_queue_proc() can enqueue both poll entries and still fail
afterwards, so the callers trying to cancel it should also try to remove
the second poll entry (if any).
For example, it may leave the request alive referencing a io_uring
context but not accessible for cancellation:
[ 282.599913][ T1620] task:iou-sqp-23145 state:D stack:28720 pid:23155 ppid: 8844 flags:0x00004004
[ 282.609927][ T1620] Call Trace:
[ 282.613711][ T1620] __schedule+0x93a/0x26f0
[ 282.634647][ T1620] schedule+0xd3/0x270
[ 282.638874][ T1620] io_uring_cancel_generic+0x54d/0x890
[ 282.660346][ T1620] io_sq_thread+0xaac/0x1250
[ 282.696394][ T1620] ret_from_fork+0x1f/0x30
Cc: stable@vger.kernel.org
Fixes: 18bceab101 ("io_uring: allow POLL_ADD with double poll_wait() users")
Reported-and-tested-by: syzbot+ac957324022b7132accf@syzkaller.appspotmail.com
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/0ec1228fc5eda4cb524eeda857da8efdc43c331c.1626774457.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If __io_queue_proc() fails to add a second poll entry, e.g. kmalloc()
failed, but it goes on with a third waitqueue, it may succeed and
overwrite the error status. Count the number of poll entries we added,
so we can set pt->error to zero at the beginning and find out when the
mentioned scenario happens.
Cc: stable@vger.kernel.org
Fixes: 18bceab101 ("io_uring: allow POLL_ADD with double poll_wait() users")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/9d6b9e561f88bcc0163623b74a76c39f712151c3.1626774457.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When we use delayed_work for fallback execution of requests, current
will be not of the submitter task, and so checks in io_req_task_submit()
may not behave as expected. Currently, it leaves inline completions not
flushed, so making io_ring_exit_work() to hang. Use the submitter task
for all those checks.
Cc: stable@vger.kernel.org
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/cb413c715bed0bc9c98b169059ea9c8a2c770715.1625881431.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Colin reports that Coverity complains about checking for poll being
non-zero after having dereferenced it multiple times. This is a valid
complaint, and actually a leftover from back when this code was based
on the aio poll code.
Kill the redundant check.
Link: https://lore.kernel.org/io-uring/fe70c532-e2a7-3722-58a1-0fa4e5c5ff2c@canonical.com/
Reported-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have requests like IORING_OP_FILES_UPDATE that don't go through
->iopoll_list but get completed in place under ->uring_lock, and so
after dropping the lock io_iopoll_check() should expect that some CQEs
might have get completed in a meanwhile.
Currently such events won't be accounted in @nr_events, and the loop
will continue to poll even if there is enough of CQEs. It shouldn't be a
problem as it's not likely to happen and so, but not nice either. Just
return earlier in this case, it should be enough.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/66ef932cc66a34e3771bbae04b2953a8058e9d05.1625747741.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If one entered io_req_task_work_add() not seeing PF_EXITING, it will set
a ->task_state bit and try task_work_add(), which may fail by that
moment. If that happens the function would try to cancel the request.
However, in a meanwhile there might come other io_req_task_work_add()
callers, which will see the bit set and leave their requests in the
list, which will never be executed.
Don't propagate an error, but clear the bit first and then fallback
all requests that we can splice from the list. The callback functions
have to be able to deal with PF_EXITING, so poll and apoll was modified
via changing io_poll_rewait().
Fixes: 7cbf1722d5 ("io_uring: provide FIFO ordering for task_work")
Reported-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/060002f19f1fdbd130ba24aef818ea4d3080819b.1625142209.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since we don't really use req->task_work anymore, get rid of it together
with the nasty ->func aliasing between ->io_task_work and ->task_work,
and hide ->fallback_node inside of io_task_work.
Also, as task_work is gone now, replace the callback type from
task_work_func_t to a function taking io_kiocb to avoid casting and
simplify code.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When task_work_add() fails, we use ->exit_task_work to queue the work.
That will be run only in the cancellation path, which happens either
when the ctx is dying or one of tasks with inflight requests is exiting
or executing. There is a good chance that such a request would just get
stuck in the list potentially hodling a file, all io_uring rsrc
recycling or some other resources. Nothing terrible, it'll go away at
some point, but we don't want to lock them up for longer than needed.
Replace that hand made ->exit_task_work with delayed_work + llist
inspired by fput_many().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently spin in iopoll() when requests to be iopolled are for
same file(device), while one device may have multiple hardware queues.
given an example:
hw_queue_0 | hw_queue_1
req(30us) req(10us)
If we first spin on iopolling for the hw_queue_0. the avg latency would
be (30us + 30us) / 2 = 30us. While if we do round robin, the avg
latency would be (30us + 10us) / 2 = 20us since we reap the request in
hw_queue_1 in time. So it's better to do spinning only when requests
are in same hardware queue.
Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Most of requests are allocated from an internal cache, so it's waste of
time fully initialising them every time. Instead, let's pre-init some of
the fields we can during initial allocation (e.g. kmalloc(), see
io_alloc_req()) and keep them valid on request recycling. There are four
of them in this patch:
->ctx is always stays the same
->link is NULL on free, it's an invariant
->result is not even needed to init, just a precaution
->async_data we now clean in io_dismantle_req() as it's likely to
never be allocated.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/892ba0e71309bba9fe9e0142472330bbf9d8f05d.1624739600.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since cancellation got moved before exit_signals(), there is no one left
who can call io_run_task_work() with PF_EXIING set, so remove the check.
Note that __io_req_task_submit() still needs a similar check.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/f7f305ececb1e6044ea649fb983ca754805bb884.1624739600.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
gcc 11 goes a weird path and duplicates most of io_arm_poll_handler()
for READ and WRITE cases. Help it and move all pollin vs pollout
specific bits under a single if-else, so there is no temptation for this
kind of unfolding.
before vs after:
text data bss dec hex filename
85362 12650 8 98020 17ee4 ./fs/io_uring.o
85186 12650 8 97844 17e34 ./fs/io_uring.o
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1deea0037293a922a0358e2958384b2e42437885.1624739600.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It is quite frequent that when an operation fails and returns EAGAIN,
the data becomes available between that failure and the call to
vfs_poll() done by io_arm_poll_handler().
Detecting the situation and reissuing the operation is much faster
than going ahead and push the operation to the io-wq.
Performance improvement testing has been performed with:
Single thread, 1 TCP connection receiving a 5 Mbps stream, no sqpoll.
4 measurements have been taken:
1. The time it takes to process a read request when data is already available
2. The time it takes to process by calling twice io_issue_sqe() after vfs_poll() indicated that data was available
3. The time it takes to execute io_queue_async_work()
4. The time it takes to complete a read request asynchronously
2.25% of all the read operations did use the new path.
ready data (baseline)
avg 3657.94182918628
min 580
max 20098
stddev 1213.15975908162
reissue completion
average 7882.67567567568
min 2316
max 28811
stddev 1982.79172973284
insert io-wq time
average 8983.82276995305
min 3324
max 87816
stddev 2551.60056552038
async time completion
average 24670.4758861127
min 10758
max 102612
stddev 3483.92416873804
Conclusion:
On average reissuing the sqe with the patch code is 1.1uSec faster and
in the worse case scenario 59uSec faster than placing the request on
io-wq
On average completion time by reissuing the sqe with the patch code is
16.79uSec faster and in the worse case scenario 73.8uSec faster than
async completion.
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/9e8441419bb1b8f3c3fcc607b2713efecdef2136.1624364038.git.olivier@trillion01.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can't support IOPOLL with non-pollable request types, and we should
check for unused/reserved fields like we do for other request types.
Fixes: 14a1143b68 ("io_uring: add support for IORING_OP_UNLINKAT")
Cc: stable@vger.kernel.org
Reported-by: Dmitry Kadashev <dkadashev@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can't support IOPOLL with non-pollable request types, and we should
check for unused/reserved fields like we do for other request types.
Fixes: 80a261fd00 ("io_uring: add support for IORING_OP_RENAMEAT")
Cc: stable@vger.kernel.org
Reported-by: Dmitry Kadashev <dkadashev@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The magic number used to cap the number of entries extracted from an
io_uring instance SQ before moving to the other instances is an
interesting parameter to experiment with.
A define has been created to make it easy to change its value from a
single location.
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/b401640063e77ad3e9f921e09c9b3ac10a8bb923.1624473200.git.olivier@trillion01.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If task_state is cleared, io_req_task_work_add() will go the slow path
adding a task_work, setting the task_state, waking up the task and so
on. Not to mention it's expensive. tctx_task_work() first clears the
state and then executes all the work items queued, so if any of them
resubmits or adds new task_work items, it would unnecessarily go through
the slow path of io_req_task_work_add().
Let's clear the ->task_state at the end. We still have to check
->task_list for emptiness afterward to synchronise with
io_req_task_work_add(), do that, and set the state back if we're going
to retry, because clearing not-ours task_state on the next iteration
would be buggy.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1ef72cdac7022adf0cd7ce4bfe3bb5c82a62eb93.1623949695.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Entering tctx_task_work() with empty task_list is a strange scenario,
that can happen only on rare occasion during task exit, so let's not
check for task_list emptiness in advance and do it do-while style. The
code still correct for the empty case, just would do extra work about
which we don't care.
Do extra step and do the check before cond_resched(), so we don't
resched if have nothing to execute.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/c4173e288e69793d03c7d7ce826f9d28afba718a.1623949695.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We don't need a full copy of tctx->task_list in tctx_task_work(), but
only a first one, so just assign node directly.
Taking into account that task_works are run in a context of a task,
it's very unlikely to first see non-empty tctx->task_list and then
splice it empty, can only happen with task_work cancellations that is
not-normal slow path anyway. Hence, get rid of the check in the end,
it's there not for validity but "performance" purposes.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d076c83fedb8253baf43acb23b8fafd7c5da1714.1623949695.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
tctx_task_work() tries to fetch a next batch of requests, but before it
would flush completions from the previous batch that may be sub-optimal.
E.g. io_req_task_queue() executes a head of the link where all the
linked may be enqueued through the same io_req_task_queue(). And there
are more cases for that.
Do the flushing at the end, so it can cache completions of several waves
of a single tctx_task_work(), and do the flush at the very end.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/3cac83934e4fbce520ff8025c3524398b3ae0270.1623949695.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, if req->creds is not NULL, then there are creds assigned.
Track the invariant with a new flag in req->flags. No need to clear the
field at init, and also cleanup can be efficiently moved into
io_clean_op().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/5f8baeb8d3b909487f555542350e2eac97005556.1623949695.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io-wq defaults to per-node masks for IO workers. This works fine by
default, but isn't particularly handy for workloads that prefer more
specific affinities, for either performance or isolation reasons.
This adds IORING_REGISTER_IOWQ_AFF that allows the user to pass in a CPU
mask that is then applied to IO thread workers, and an
IORING_UNREGISTER_IOWQ_AFF that simply resets the masks back to the
default of per-node.
Note that no care is given to existing IO threads, they will need to go
through a reschedule before the affinity is correct if they are already
running or sleeping.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
