We checked for 'force_nonblock' higher up, so it's definitely false
at this point. Kill the check, it's a remnant of when we tried to do
inline splice without always punting to async context.
Fixes: 2fb3e82284 ("io_uring: punt splice async because of inode mutex")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add IORING_OP_TEE implementing tee(2) support. Almost identical to
splice bits, but without offsets.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
req->flags stores all sqe->flags. After checking that sqe->flags are
valid set if IOSQE* flags, no need to double check it, just forward them
all.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_file_put() deals with flushing state's file refs, adding "state" to
its name makes it a bit clearer. Also, avoid double check of
state->file in __io_file_get() in some cases.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A submission is "async" IIF it's done by SQPOLL thread. Instead of
passing @async flag into io_submit_sqes(), deduce it from ctx->flags.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We only need apoll in the one section, do the juggling with the work
restoration there. This removes a special case further down as well.
No functional changes in this patch.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As for other not inlined requests, alloc req->io for FORCE_ASYNC reqs,
so they can be prepared properly.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If req->io is not NULL, it's already prepared. Don't do it again,
it's dangerous.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There's no point in using list_del_init() on entries that are going
away, and the associated lock is always used in process context so
let's not use the IRQ disabling+saving variant of the spinlock.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This new flag should be set/clear from the application to
disable/enable eventfd notifications when a request is completed
and queued to the CQ ring.
Before this patch, notifications were always sent if an eventfd is
registered, so IORING_CQ_EVENTFD_DISABLED is not set during the
initialization.
It will be up to the application to set the flag after initialization
if no notifications are required at the beginning.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This patch adds the new 'cq_flags' field that should be written by
the application and read by the kernel.
This new field is available to the userspace application through
'cq_off.flags'.
We are using 4-bytes previously reserved and set to zero. This means
that if the application finds this field to zero, then the new
functionality is not supported.
In the next patch we will introduce the first flag available.
Signed-off-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some file descriptors use separate waitqueues for their f_ops->poll()
handler, most commonly one for read and one for write. The io_uring
poll implementation doesn't work with that, as the 2nd poll_wait()
call will cause the io_uring poll request to -EINVAL.
This affects (at least) tty devices and /dev/random as well. This is a
big problem for event loops where some file descriptors work, and others
don't.
With this fix, io_uring handles multiple waitqueues.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently embed and queue a work item per fixed_file_ref_node that
we update, but if the workload does a lot of these, then the associated
kworker-events overhead can become quite noticeable.
Since we rarely need to wait on these, batch them at 1 second intervals
instead. If we do need to wait for them, we just flush the pending
delayed work.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We used to have three completions, now we just have two. With the two,
let's not allocate them dynamically, just embed then in the ctx and
name them appropriately.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When we changed the file registration handling, it became important to
iterate the bulk request freeing list for fixed files as well, or we
miss dropping the fixed file reference. If not, we're leaking references,
and we'll get a kworker stuck waiting for file references to disappear.
This also means we can remove the special casing of fixed vs non-fixed
files, we need to iterate for both and we can just rely on
__io_req_aux_free() doing io_put_file() instead of doing it manually.
Fixes: 0558955373 ("io_uring: refactor file register/unregister/update handling")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Remove duplicate semicolon at the end of line in io_file_from_index()
Signed-off-by: Xiaoming Ni <nixiaoming@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
do_splice() doesn't expect len to be 0. Just always return 0 in this
case as splice(2) does.
Fixes: 7d67af2c01 ("io_uring: add splice(2) support")
Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The "struct io_submit_state *state" parameter is not used, remove it.
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The attempt protecting us from closing the ring itself wasn't really
complete, and we actually don't need it. The referencing of requests
themselve, and the references they hold on the ring, ensures that the
life time of the ring is sane. With the check removed, we can also
remove the need to have the close operation fget() the file.
Reported-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We currently make some guesses as when to open this fd, but in reality
we have no business (or need) to do so at all. In fact, it makes certain
things fail, like O_PATH.
Remove the fd lookup from these opcodes, we're just passing the 'fd' to
generic helpers anyway. With that, we can also remove the special casing
of fd values in io_req_needs_file(), and the 'fd_non_neg' check that
we have. And we can ensure that we only read sqe->fd once.
This fixes O_PATH usage with openat/openat2, and ditto statx path side
oddities.
Cc: stable@vger.kernel.org: # v5.6
Reported-by: Max Kellermann <mk@cm4all.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If copy_to_user() in io_uring_setup() failed, we'll leak many kernel
resources, which will be recycled until process terminates. This bug
can be reproduced by using mprotect to set params to PROT_READ. To fix
this issue, refactor io_uring_create() a bit to add a new 'struct
io_uring_params __user *params' parameter and move the copy_to_user()
in io_uring_setup() to io_uring_setup(), if copy_to_user() failed,
we can free kernel resource properly.
Suggested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The prepare_to_wait() and finish_wait() calls in io_uring_cancel_files()
are mismatched. Currently I don't see any issues related this bug, just
find it by learning codes.
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Nonblocking do_splice() still may wait for some time on an inode mutex.
Let's play safe and always punt it async.
Reported-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_req_defer() do double-checked locking. Use proper helpers for that,
i.e. list_empty_careful().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
While working on to make io_uring sqpoll mode support syscalls that need
struct files_struct, I got cpu soft lockup in io_ring_ctx_wait_and_kill(),
while (ctx->sqo_thread && !wq_has_sleeper(&ctx->sqo_wait))
cpu_relax();
above loop never has an chance to exit, it's because preempt isn't enabled
in the kernel, and the context calling io_ring_ctx_wait_and_kill() and
io_sq_thread() run in the same cpu, if io_sq_thread calls a cond_resched()
yield cpu and another context enters above loop, then io_sq_thread() will
always in runqueue and never exit.
Use cond_resched() can fix this issue.
Reported-by: syzbot+66243bb7126c410cefe6@syzkaller.appspotmail.com
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use ctx->fallback_req address for test_and_set_bit_lock() and
clear_bit_unlock().
Signed-off-by: Bijan Mottahedeh <bijan.mottahedeh@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We do blocking retry from our poll handler, if the file supports polled
notifications. Only mark the request as needing an async worker if we
can't poll for it.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can have files like eventfd where it's perfectly fine to do poll
based retry on them, right now io_file_supports_async() doesn't take
that into account.
Pass in data direction and check the f_op instead of just always needing
an async worker.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Clay reports that OP_STATX fails for a test case with a valid fd
and empty path:
-- Test 0: statx:fd 3: SUCCEED, file mode 100755
-- Test 1: statx:path ./uring_statx: SUCCEED, file mode 100755
-- Test 2: io_uring_statx:fd 3: FAIL, errno 9: Bad file descriptor
-- Test 3: io_uring_statx:path ./uring_statx: SUCCEED, file mode 100755
This is due to statx not grabbing the process file table, hence we can't
lookup the fd in async context. If the fd is valid, ensure that we grab
the file table so we can grab the file from async context.
Cc: stable@vger.kernel.org # v5.6
Reported-by: Clay Harris <bugs@claycon.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When testing io_uring IORING_FEAT_FAST_POLL feature, I got below panic:
BUG: kernel NULL pointer dereference, address: 0000000000000030
PGD 0 P4D 0
Oops: 0000 [#1] SMP PTI
CPU: 5 PID: 2154 Comm: io_uring_echo_s Not tainted 5.6.0+ #359
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996),
BIOS rel-1.11.1-0-g0551a4be2c-prebuilt.qemu-project.org 04/01/2014
RIP: 0010:io_wq_submit_work+0xf/0xa0
Code: ff ff ff be 02 00 00 00 e8 ae c9 19 00 e9 58 ff ff ff 66 0f 1f
84 00 00 00 00 00 0f 1f 44 00 00 41 54 49 89 fc 55 53 48 8b 2f <8b>
45 30 48 8d 9d 48 ff ff ff 25 01 01 00 00 83 f8 01 75 07 eb 2a
RSP: 0018:ffffbef543e93d58 EFLAGS: 00010286
RAX: ffffffff84364f50 RBX: ffffa3eb50f046b8 RCX: 0000000000000000
RDX: ffffa3eb0efc1840 RSI: 0000000000000006 RDI: ffffa3eb50f046b8
RBP: 0000000000000000 R08: 00000000fffd070d R09: 0000000000000000
R10: 0000000000000000 R11: 0000000000000000 R12: ffffa3eb50f046b8
R13: ffffa3eb0efc2088 R14: ffffffff85b69be0 R15: ffffa3eb0effa4b8
FS: 00007fe9f69cc4c0(0000) GS:ffffa3eb5ef40000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000030 CR3: 0000000020410000 CR4: 00000000000006e0
Call Trace:
task_work_run+0x6d/0xa0
do_exit+0x39a/0xb80
? get_signal+0xfe/0xbc0
do_group_exit+0x47/0xb0
get_signal+0x14b/0xbc0
? __x64_sys_io_uring_enter+0x1b7/0x450
do_signal+0x2c/0x260
? __x64_sys_io_uring_enter+0x228/0x450
exit_to_usermode_loop+0x87/0xf0
do_syscall_64+0x209/0x230
entry_SYSCALL_64_after_hwframe+0x49/0xb3
RIP: 0033:0x7fe9f64f8df9
Code: Bad RIP value.
task_work_run calls io_wq_submit_work unexpectedly, it's obvious that
struct callback_head's func member has been changed. After looking into
codes, I found this issue is still due to the union definition:
union {
/*
* Only commands that never go async can use the below fields,
* obviously. Right now only IORING_OP_POLL_ADD uses them, and
* async armed poll handlers for regular commands. The latter
* restore the work, if needed.
*/
struct {
struct callback_head task_work;
struct hlist_node hash_node;
struct async_poll *apoll;
};
struct io_wq_work work;
};
When task_work_run has multiple work to execute, the work that calls
io_poll_remove_all() will do req->work restore for non-poll request
always, but indeed if a non-poll request has been added to a new
callback_head, subsequent callback will call io_async_task_func() to
handle this request, that means we should not do the restore work
for such non-poll request. Meanwhile in io_async_task_func(), we should
drop submit ref when req has been canceled.
Fix both issues.
Fixes: b1f573bd15 ("io_uring: restore req->work when canceling poll request")
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Use io_double_put_req()
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When checking for draining with __req_need_defer(), it tries to match
how many requests were sent before a current one with number of already
completed. Dropped SQEs are included in req->sequence, and they won't
ever appear in CQ. To compensate for that, __req_need_defer() substracts
ctx->cached_sq_dropped.
However, what it should really use is number of SQEs dropped __before__
the current one. In other words, any submitted request shouldn't
shouldn't affect dequeueing from the drain queue of previously submitted
ones.
Instead of saving proper ctx->cached_sq_dropped in each request,
substract from req->sequence it at initialisation, so it includes number
of properly submitted requests.
note: it also changes behaviour of timeouts, but
1. it's already diverge from the description because of using SQ
2. the description is ambiguous regarding dropped SQEs
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
req->timeout.count and req->io->timeout.seq_offset store the same value,
which is sqe->off. Kill the second one
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_timeout() can be executed asynchronously by a worker and without
holding ctx->uring_lock
1. using ctx->cached_sq_head there is racy there
2. it should count events from a moment of timeout's submission, but
not execution
Use req->sequence.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the request has been marked as canceled, don't try and issue it.
Instead just fill a canceled event and finish the request.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We added this for just the regular poll requests in commit a6ba632d2c
("io_uring: retry poll if we got woken with non-matching mask"), we
should do the same for the poll handler used pollable async requests.
Move the re-wait check and arm into a helper, and call it from
io_async_task_func() as well.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The splice file punt check uses file->f_mode to check for O_NONBLOCK,
but it should be checking file->f_flags. This leads to punting even
for files that have O_NONBLOCK set, which isn't necessary. This equates
to checking for FMODE_PATH, which will never be set on the fd in
question.
Fixes: 7d67af2c01 ("io_uring: add splice(2) support")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Requests initialisation is scattered across several functions, namely
io_init_req(), io_submit_sqes(), io_submit_sqe(). Put it
in io_init_req() for better data locality and code clarity.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It's a good idea to not read sqe->flags twice, as it's prone to security
bugs. Instead of passing it around, embeed them in req->flags. It's
already so except for IOSQE_IO_LINK.
1. rename former REQ_F_LINK -> REQ_F_LINK_HEAD
2. introduce and copy REQ_F_LINK, which mimics IO_IOSQE_LINK
And leave req_set_fail_links() using new REQ_F_LINK, because it's more
sensible.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Having only one place for cleaning up a request after a link assembly/
submission failure will play handy in the future. At least it allows
to remove duplicated cleanup sequence.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As a preparation for extracting request init bits, remove self-coded mm
tracking from io_submit_sqes(), but rely on current->mm. It's more
convenient, than passing this piece of state in other functions.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If io_submit_sqes() can't grab an mm, it fails and exits right away.
There is no need to track the fact of the failure. Remove @mm_fault.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can't reliably wait in io_ring_ctx_wait_and_kill(), since the
task_works list isn't ordered (in fact it's LIFO ordered). We could
either fix this with a separate task_works list for io_uring work, or
just punt the wait-and-free to async context. This ensures that
task_work that comes in while we're shutting down is processed
correctly. If we don't go async, we could have work past the fput()
work for the ring that depends on work that won't be executed until
after we're done with the wait-and-free. But as this operation is
blocking, it'll never get a chance to run.
This was reproduced with hundreds of thousands of sockets running
memcached, haven't been able to reproduce this synthetically.
Reported-by: Dan Melnic <dmm@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If completion queue overflow occurs, __io_cqring_fill_event() will
update req->cflags, which is in a union with req->work and happens to
be aliased to req->work.fs. Following io_free_req() ->
io_req_work_drop_env() may get a bunch of different problems (miscount
fs->users, segfault, etc) on cleaning @fs.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Don't re-read userspace-shared sqe->flags, it can be exploited.
sqe->flags are copied into req->flags in io_submit_sqe(), check them
there instead.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_get_req() do two different things: io_kiocb allocation and
initialisation. Move init part out of it and rename into
io_alloc_req(). It's simpler this way and also have better data
locality.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
As io_get_sqe() split into 2 stage get/consume, get an sqe before
allocating io_kiocb, so no free_req*() for a failure case is needed,
and inline back __io_req_do_free(), which has only 1 user.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Make io_get_sqring() care only about sqes themselves, not initialising
the io_kiocb. Also, split it into get + consume, that will be helpful in
the future.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In io_read_prep() or io_write_prep(), io_req_map_rw() takes
struct io_async_rw's fast_iov as argument to call io_import_iovec(),
and if io_import_iovec() uses struct io_async_rw's fast_iov as
valid iovec array, later indeed io_req_map_rw() does not need
to do the memcpy operation, because they are same pointers.
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
OPENAT2 correctly sets O_LARGEFILE if it has to, but that escaped the
OPENAT opcode. Dmitry reports that his test case that compares openat()
and IORING_OP_OPENAT sees failures on large files:
*** sync openat
openat succeeded
sync write at offset 0
write succeeded
sync write at offset 4294967296
write succeeded
*** sync openat
openat succeeded
io_uring write at offset 0
write succeeded
io_uring write at offset 4294967296
write succeeded
*** io_uring openat
openat succeeded
sync write at offset 0
write succeeded
sync write at offset 4294967296
write failed: File too large
*** io_uring openat
openat succeeded
io_uring write at offset 0
write succeeded
io_uring write at offset 4294967296
write failed: File too large
Ensure we set O_LARGEFILE, if force_o_largefile() is true.
Cc: stable@vger.kernel.org # v5.6
Fixes: 15b71abe7b ("io_uring: add support for IORING_OP_OPENAT")
Reported-by: Dmitry Kadashev <dkadashev@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
An earlier commit "io_uring: remove @nxt from handlers" removed the
setting of pointer nxt and now it is always null, hence the non-null
check and call to io_wq_assign_next is redundant and can be removed.
Addresses-Coverity: ("'Constant' variable guard")
Reviewed-by: Chaitanya Kulkarni <chaitanya.kulkarni@wdc.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If io_get_req() fails, it drops a ref. Then, awhile keeping @submitted
unmodified, io_submit_sqes() breaks the loop and puts @nr - @submitted
refs. For each submitted req a ref is dropped in io_put_req() and
friends. So, for @nr taken refs there will be
(@nr - @submitted + @submitted + 1) dropped.
Remove ctx refcounting from io_get_req(), that at the same time makes
it clearer.
Fixes: 2b85edfc0c ("io_uring: batch getting pcpu references")
Cc: stable@vger.kernel.org # v5.6
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A request that completes with an -EAGAIN result after it has been added
to the poll list, will not be removed from that list in io_do_iopoll()
because the f_op->iopoll() will not succeed for that request.
Maintain a retryable local list similar to the done list, and explicity
reissue requests with an -EAGAIN result.
Signed-off-by: Bijan Mottahedeh <bijan.mottahedeh@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We already checked this limit when the file was opened, and we keep it
open in the file table. Hence when we added unit_inflight to the count
we want to register, we're doubly accounting these files. This results
in -EMFILE for file registration, if we're at half the limit.
Cc: stable@vger.kernel.org # v5.1+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If the original task is (or has) exited, then the task work will not get
queued properly. Allow for using the io-wq manager task to queue this
work for execution, and ensure that the io-wq manager notices and runs
this work if woken up (or exiting).
Reported-by: Dan Melnic <dmm@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We can have a task exit if it's not the owner of the ring. Be safe and
grab an actual reference to it, to avoid a potential use-after-free.
Reported-by: Dan Melnic <dmm@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If we get woken and the poll doesn't match our mask, re-add the task
to the poll waitqueue and try again instead of completing the request
with a mask of 0.
Reported-by: Dan Melnic <dmm@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add it to pair with prepare_to_wait() in an attempt to avoid
anything weird in the field.
Fixes: b41e98524e ("io_uring: add per-task callback handler")
Reported-by: syzbot+0c3370f235b74b3cfd97@syzkaller.appspotmail.com
Signed-off-by: Hillf Danton <hdanton@sina.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
While diving into io_uring fileset register/unregister/update codes, we
found one bug in the fileset update handling. io_uring fileset update
use a percpu_ref variable to check whether we can put the previously
registered file, only when the refcnt of the perfcpu_ref variable
reaches zero, can we safely put these files. But this doesn't work so
well. If applications always issue requests continually, this
perfcpu_ref will never have an chance to reach zero, and it'll always be
in atomic mode, also will defeat the gains introduced by fileset
register/unresiger/update feature, which are used to reduce the atomic
operation overhead of fput/fget.
To fix this issue, while applications do IORING_REGISTER_FILES or
IORING_REGISTER_FILES_UPDATE operations, we allocate a new percpu_ref
and kill the old percpu_ref, new requests will use the new percpu_ref.
Once all previous old requests complete, old percpu_refs will be dropped
and registered files will be put safely.
Link: https://lore.kernel.org/io-uring/5a8dac33-4ca2-4847-b091-f7dcd3ad0ff3@linux.alibaba.com/T/#t
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----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=bYLJ
-----END PGP SIGNATURE-----
Merge tag 'for-5.7/io_uring-2020-03-29' of git://git.kernel.dk/linux-block
Pull io_uring updates from Jens Axboe:
"Here are the io_uring changes for this merge window. Light on new
features this time around (just splice + buffer selection), lots of
cleanups, fixes, and improvements to existing support. In particular,
this contains:
- Cleanup fixed file update handling for stack fallback (Hillf)
- Re-work of how pollable async IO is handled, we no longer require
thread offload to handle that. Instead we rely using poll to drive
this, with task_work execution.
- In conjunction with the above, allow expendable buffer selection,
so that poll+recv (for example) no longer has to be a split
operation.
- Make sure we honor RLIMIT_FSIZE for buffered writes
- Add support for splice (Pavel)
- Linked work inheritance fixes and optimizations (Pavel)
- Async work fixes and cleanups (Pavel)
- Improve io-wq locking (Pavel)
- Hashed link write improvements (Pavel)
- SETUP_IOPOLL|SETUP_SQPOLL improvements (Xiaoguang)"
* tag 'for-5.7/io_uring-2020-03-29' of git://git.kernel.dk/linux-block: (54 commits)
io_uring: cleanup io_alloc_async_ctx()
io_uring: fix missing 'return' in comment
io-wq: handle hashed writes in chains
io-uring: drop 'free_pfile' in struct io_file_put
io-uring: drop completion when removing file
io_uring: Fix ->data corruption on re-enqueue
io-wq: close cancel gap for hashed linked work
io_uring: make spdxcheck.py happy
io_uring: honor original task RLIMIT_FSIZE
io-wq: hash dependent work
io-wq: split hashing and enqueueing
io-wq: don't resched if there is no work
io-wq: remove duplicated cancel code
io_uring: fix truncated async read/readv and write/writev retry
io_uring: dual license io_uring.h uapi header
io_uring: io_uring_enter(2) don't poll while SETUP_IOPOLL|SETUP_SQPOLL enabled
io_uring: Fix unused function warnings
io_uring: add end-of-bits marker and build time verify it
io_uring: provide means of removing buffers
io_uring: add IOSQE_BUFFER_SELECT support for IORING_OP_RECVMSG
...
Cleanup io_alloc_async_ctx() a bit, add a new __io_alloc_async_ctx(),
so io_setup_async_rw() won't need to check whether async_ctx is true
or false again.
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The missing 'return' work may make it hard for other developers to
understand it.
Signed-off-by: Chucheng Luo <luochucheng@vivo.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Sync removal of file is only used in case of a GFP_KERNEL kmalloc
failure at the cost of io_file_put::done and work flush, while a
glich like it can be handled at the call site without too much pain.
That said, what is proposed is to drop sync removing of file, and
the kink in neck as well.
Signed-off-by: Hillf Danton <hdanton@sina.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A case of task hung was reported by syzbot,
INFO: task syz-executor975:9880 blocked for more than 143 seconds.
Not tainted 5.6.0-rc6-syzkaller #0
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
syz-executor975 D27576 9880 9878 0x80004000
Call Trace:
schedule+0xd0/0x2a0 kernel/sched/core.c:4154
schedule_timeout+0x6db/0xba0 kernel/time/timer.c:1871
do_wait_for_common kernel/sched/completion.c:83 [inline]
__wait_for_common kernel/sched/completion.c:104 [inline]
wait_for_common kernel/sched/completion.c:115 [inline]
wait_for_completion+0x26a/0x3c0 kernel/sched/completion.c:136
io_queue_file_removal+0x1af/0x1e0 fs/io_uring.c:5826
__io_sqe_files_update.isra.0+0x3a1/0xb00 fs/io_uring.c:5867
io_sqe_files_update fs/io_uring.c:5918 [inline]
__io_uring_register+0x377/0x2c00 fs/io_uring.c:7131
__do_sys_io_uring_register fs/io_uring.c:7202 [inline]
__se_sys_io_uring_register fs/io_uring.c:7184 [inline]
__x64_sys_io_uring_register+0x192/0x560 fs/io_uring.c:7184
do_syscall_64+0xf6/0x7d0 arch/x86/entry/common.c:294
entry_SYSCALL_64_after_hwframe+0x49/0xbe
and bisect pointed to 05f3fb3c53 ("io_uring: avoid ring quiesce for
fixed file set unregister and update").
It is down to the order that we wait for work done before flushing it
while nobody is likely going to wake us up.
We can drop that completion on stack as flushing work itself is a sync
operation we need and no more is left behind it.
To that end, io_file_put::done is re-used for indicating if it can be
freed in the workqueue worker context.
Reported-and-Inspired-by: syzbot <syzbot+538d1957ce178382a394@syzkaller.appspotmail.com>
Signed-off-by: Hillf Danton <hdanton@sina.com>
Rename ->done to ->free_pfile
Signed-off-by: Jens Axboe <axboe@kernel.dk>
work->data and work->list are shared in union. io_wq_assign_next() sets
->data if a req having a linked_timeout, but then io-wq may want to use
work->list, e.g. to do re-enqueue of a request, so corrupting ->data.
->data is not necessary, just remove it and extract linked_timeout
through @link_list.
Fixes: 60cf46ae60 ("io-wq: hash dependent work")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With the previous fixes for number of files open checking, I added some
debug code to see if we had other spots where we're checking rlimit()
against the async io-wq workers. The only one I found was file size
checking, which we should also honor.
During write and fallocate prep, store the max file size and override
that for the current ask if we're in io-wq worker context.
Cc: stable@vger.kernel.org # 5.1+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Just like commit 4022e7af86, this fixes the fact that
IORING_OP_ACCEPT ends up using get_unused_fd_flags(), which checks
current->signal->rlim[] for limits.
Add an extra argument to __sys_accept4_file() that allows us to pass
in the proper nofile limit, and grab it at request prep time.
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Dmitry reports that a test case shows that io_uring isn't honoring a
modified rlimit nofile setting. get_unused_fd_flags() checks the task
signal->rlimi[] for the limits. As this isn't easily inheritable,
provide a __get_unused_fd_flags() that takes the value instead. Then we
can grab it when the request is prepared (from the original task), and
pass that in when we do the async part part of the open.
Reported-by: Dmitry Kadashev <dkadashev@gmail.com>
Tested-by: Dmitry Kadashev <dkadashev@gmail.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It's a preparation patch removing io_wq_enqueue_hashed(), which
now should be done by io_wq_hash_work() + io_wq_enqueue().
Also, set hash value for dependant works, and do it as late as possible,
because req->file can be unavailable before. This hash will be ignored
by io-wq.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Processing links, io_submit_sqe() prepares requests, drops sqes, and
passes them with sqe=NULL to io_queue_sqe(). There IOSQE_DRAIN and/or
IOSQE_ASYNC requests will go through the same prep, which doesn't expect
sqe=NULL and fail with NULL pointer deference.
Always do full prepare including io_alloc_async_ctx() for linked
requests, and then it can skip the second preparation.
Cc: stable@vger.kernel.org # 5.5
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Ensure we keep the truncated value, if we did truncate it. If not, we
might read/write more than the registered buffer size.
Also for retry, ensure that we return the truncated mapped value for
the vectorized versions of the read/write commands.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When SETUP_IOPOLL and SETUP_SQPOLL are both enabled, applications don't need
to do io completion events polling again, they can rely on io_sq_thread to do
polling work, which can reduce cpu usage and uring_lock contention.
I modify fio io_uring engine codes a bit to evaluate the performance:
static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
continue;
}
- if (!o->sqpoll_thread) {
+ if (o->sqpoll_thread && o->hipri) {
r = io_uring_enter(ld, 0, actual_min,
IORING_ENTER_GETEVENTS);
if (r < 0) {
and use "fio -name=fiotest -filename=/dev/nvme0n1 -iodepth=$depth -thread
-rw=read -ioengine=io_uring -hipri=1 -sqthread_poll=1 -direct=1 -bs=4k
-size=10G -numjobs=1 -time_based -runtime=120"
original codes
--------------------------------------------------------------------
iodepth | 4 | 8 | 16 | 32 | 64
bw | 1133MB/s | 1519MB/s | 2090MB/s | 2710MB/s | 3012MB/s
fio cpu usage | 100% | 100% | 100% | 100% | 100%
--------------------------------------------------------------------
with patch
--------------------------------------------------------------------
iodepth | 4 | 8 | 16 | 32 | 64
bw | 1196MB/s | 1721MB/s | 2351MB/s | 2977MB/s | 3357MB/s
fio cpu usage | 63.8% | 74.4%% | 81.1% | 83.7% | 82.4%
--------------------------------------------------------------------
bw improve | 5.5% | 13.2% | 12.3% | 9.8% | 11.5%
--------------------------------------------------------------------
From above test results, we can see that bw has above 5.5%~13%
improvement, and fio process's cpu usage also drops much. Note this
won't improve io_sq_thread's cpu usage when SETUP_IOPOLL|SETUP_SQPOLL
are both enabled, in this case, io_sq_thread always has 100% cpu usage.
I think this patch will be friendly to applications which will often use
io_uring_wait_cqe() or similar from liburing.
Signed-off-by: Xiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If CONFIG_NET is not set, gcc warns:
fs/io_uring.c:3110:12: warning: io_setup_async_msg defined but not used [-Wunused-function]
static int io_setup_async_msg(struct io_kiocb *req,
^~~~~~~~~~~~~~~~~~
There are many funcions wraped by CONFIG_NET, move them
together to simplify code, also fix this warning.
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Minor tweaks.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Not easy to tell if we're going over the size of bits we can shove
in req->flags, so add an end-of-bits marker and a BUILD_BUG_ON()
check for it.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We have IORING_OP_PROVIDE_BUFFERS, but the only way to remove buffers
is to trigger IO on them. The usual case of shrinking a buffer pool
would be to just not replenish the buffers when IO completes, and
instead just free it. But it may be nice to have a way to manually
remove a number of buffers from a given group, and
IORING_OP_REMOVE_BUFFERS provides that functionality.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This adds support for the vectored read. This is limited to supporting
just 1 segment in the iov, and is provided just for convenience for
applications that use IORING_OP_READV already.
The iov helpers will be used for IORING_OP_RECVMSG as well.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If a server process has tons of pending socket connections, generally
it uses epoll to wait for activity. When the socket is ready for reading
(or writing), the task can select a buffer and issue a recv/send on the
given fd.
Now that we have fast (non-async thread) support, a task can have tons
of pending reads or writes pending. But that means they need buffers to
back that data, and if the number of connections is high enough, having
them preallocated for all possible connections is unfeasible.
With IORING_OP_PROVIDE_BUFFERS, an application can register buffers to
use for any request. The request then sets IOSQE_BUFFER_SELECT in the
sqe, and a given group ID in sqe->buf_group. When the fd becomes ready,
a free buffer from the specified group is selected. If none are
available, the request is terminated with -ENOBUFS. If successful, the
CQE on completion will contain the buffer ID chosen in the cqe->flags
member, encoded as:
(buffer_id << IORING_CQE_BUFFER_SHIFT) | IORING_CQE_F_BUFFER;
Once a buffer has been consumed by a request, it is no longer available
and must be registered again with IORING_OP_PROVIDE_BUFFERS.
Requests need to support this feature. For now, IORING_OP_READ and
IORING_OP_RECV support it. This is checked on SQE submission, a CQE with
res == -EOPNOTSUPP will be posted if attempted on unsupported requests.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
IORING_OP_PROVIDE_BUFFERS uses the buffer registration infrastructure to
support passing in an addr/len that is associated with a buffer ID and
buffer group ID. The group ID is used to index and lookup the buffers,
while the buffer ID can be used to notify the application which buffer
in the group was used. The addr passed in is the starting buffer address,
and length is each buffer length. A number of buffers to add with can be
specified, in which case addr is incremented by length for each addition,
and each buffer increments the buffer ID specified.
No validation is done of the buffer ID. If the application provides
buffers within the same group with identical buffer IDs, then it'll have
a hard time telling which buffer ID was used. The only restriction is
that the buffer ID can be a max of 16-bits in size, so USHRT_MAX is the
maximum ID that can be used.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
After more careful studying, Paul informs me that we cannot rely on
ordering of RCU callbacks in the way that the the tagged commit did.
The current construct looks like this:
void C(struct rcu_head *rhp)
{
do_something(rhp);
call_rcu(&p->rh, B);
}
call_rcu(&p->rh, A);
call_rcu(&p->rh, C);
and we're relying on ordering between A and B, which isn't guaranteed.
Make this explicit instead, and have a work item issue the rcu_barrier()
to ensure that A has run before we manually execute B.
While thorough testing never showed this issue, it's dependent on the
per-cpu load in terms of RCU callbacks. The updated method simplifies
the code as well, and eliminates the need to maintain an rcu_head in
the fileset data.
Fixes: c1e2148f8e ("io_uring: free fixed_file_data after RCU grace period")
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is a recipe to deadlock the kernel: submit a timeout sqe with a
linked_timeout (e.g. test_single_link_timeout_ception() from liburing),
and SIGKILL the process.
Then, io_kill_timeouts() takes @ctx->completion_lock, but the timeout
isn't flagged with REQ_F_COMP_LOCKED, and will try to double grab it
during io_put_free() to cancel the linked timeout. Probably, the same
can happen with another io_kill_timeout() call site, that is
io_commit_cqring().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The percpu refcount protects this structure, and we can have an atomic
switch in progress when exiting. This makes it unsafe to just free the
struct normally, and can trigger the following KASAN warning:
BUG: KASAN: use-after-free in percpu_ref_switch_to_atomic_rcu+0xfa/0x1b0
Read of size 1 at addr ffff888181a19a30 by task swapper/0/0
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.6.0-rc4+ #5747
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1ubuntu1 04/01/2014
Call Trace:
<IRQ>
dump_stack+0x76/0xa0
print_address_description.constprop.0+0x3b/0x60
? percpu_ref_switch_to_atomic_rcu+0xfa/0x1b0
? percpu_ref_switch_to_atomic_rcu+0xfa/0x1b0
__kasan_report.cold+0x1a/0x3d
? percpu_ref_switch_to_atomic_rcu+0xfa/0x1b0
percpu_ref_switch_to_atomic_rcu+0xfa/0x1b0
rcu_core+0x370/0x830
? percpu_ref_exit+0x50/0x50
? rcu_note_context_switch+0x7b0/0x7b0
? run_rebalance_domains+0x11d/0x140
__do_softirq+0x10a/0x3e9
irq_exit+0xd5/0xe0
smp_apic_timer_interrupt+0x86/0x200
apic_timer_interrupt+0xf/0x20
</IRQ>
RIP: 0010:default_idle+0x26/0x1f0
Fix this by punting the final exit and free of the struct to RCU, then
we know that it's safe to do so. Jann suggested the approach of using a
double rcu callback to achieve this. It's important that we do a nested
call_rcu() callback, as otherwise the free could be ordered before the
atomic switch, even if the latter was already queued.
Reported-by: syzbot+e017e49c39ab484ac87a@syzkaller.appspotmail.com
Suggested-by: Jann Horn <jannh@google.com>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This just prepares the ring for having lists of buffers associated with
it, that the application can provide for SQEs to consume instead of
providing their own.
The buffers are organized by group ID.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
First it changes io-wq interfaces. It replaces {get,put}_work() with
free_work(), which guaranteed to be called exactly once. It also enforces
free_work() callback to be non-NULL.
io_uring follows the changes and instead of putting a submission reference
in io_put_req_async_completion(), it will be done in io_free_work(). As
removes io_get_work() with corresponding refcount_inc(), the ref balance
is maintained.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
If after dropping the submission reference req->refs == 1, the request
is done, because this one is for io_put_work() and will be dropped
synchronously shortly after. In this case it's safe to steal a next
work from the request.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There will be no use for @nxt in the handlers, and it's doesn't work
anyway, so purge it
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The rule is simple, any async handler gets a submission ref and should
put it at the end. Make them all follow it, and so more consistent.
This is a preparation patch, and as io_wq_assign_next() currently won't
ever work, this doesn't care to use io_put_req_find_next() instead of
io_put_req().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
refcount_inc_not_zero() -> refcount_inc() fix.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Clang warns:
fs/io_uring.c:4178:6: warning: variable 'mask' is used uninitialized
whenever 'if' condition is false [-Wsometimes-uninitialized]
if (def->pollin)
^~~~~~~~~~~
fs/io_uring.c:4182:2: note: uninitialized use occurs here
mask |= POLLERR | POLLPRI;
^~~~
fs/io_uring.c:4178:2: note: remove the 'if' if its condition is always
true
if (def->pollin)
^~~~~~~~~~~~~~~~
fs/io_uring.c:4154:15: note: initialize the variable 'mask' to silence
this warning
__poll_t mask, ret;
^
= 0
1 warning generated.
io_op_defs has many definitions where pollin is not set so mask indeed
might be uninitialized. Initialize it to zero and change the next
assignment to |=, in case further masks are added in the future to avoid
missing changing the assignment then.
Fixes: d7718a9d25 ("io_uring: use poll driven retry for files that support it")
Link: https://github.com/ClangBuiltLinux/linux/issues/916
Signed-off-by: Nathan Chancellor <natechancellor@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io-wq cares about IO_WQ_WORK_UNBOUND flag only while enqueueing, so
it's useless setting it for a next req of a link. Thus, removed it
from io_prep_linked_timeout(), and inline the function.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
After __io_queue_sqe() ended up in io_queue_async_work(), it's already
known that there is no @nxt req, so skip the check and return from the
function.
Also, @nxt initialisation now can be done just before
io_put_req_find_next(), as there is no jumping until it's checked.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently io_uring tries any request in a non-blocking manner, if it can,
and then retries from a worker thread if we get -EAGAIN. Now that we have
a new and fancy poll based retry backend, use that to retry requests if
the file supports it.
This means that, for example, an IORING_OP_RECVMSG on a socket no longer
requires an async thread to complete the IO. If we get -EAGAIN reading
from the socket in a non-blocking manner, we arm a poll handler for
notification on when the socket becomes readable. When it does, the
pending read is executed directly by the task again, through the io_uring
task work handlers. Not only is this faster and more efficient, it also
means we're not generating potentially tons of async threads that just
sit and block, waiting for the IO to complete.
The feature is marked with IORING_FEAT_FAST_POLL, meaning that async
pollable IO is fast, and that poll<link>other_op is fast as well.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Add a pollin/pollout field to the request table, and have commands that
we can safely poll for properly marked.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
For poll requests, it's not uncommon to link a read (or write) after
the poll to execute immediately after the file is marked as ready.
Since the poll completion is called inside the waitqueue wake up handler,
we have to punt that linked request to async context. This slows down
the processing, and actually means it's faster to not use a link for this
use case.
We also run into problems if the completion_lock is contended, as we're
doing a different lock ordering than the issue side is. Hence we have
to do trylock for completion, and if that fails, go async. Poll removal
needs to go async as well, for the same reason.
eventfd notification needs special case as well, to avoid stack blowing
recursion or deadlocks.
These are all deficiencies that were inherited from the aio poll
implementation, but I think we can do better. When a poll completes,
simply queue it up in the task poll list. When the task completes the
list, we can run dependent links inline as well. This means we never
have to go async, and we can remove a bunch of code associated with
that, and optimizations to try and make that run faster. The diffstat
speaks for itself.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Store the io_kiocb in the private field instead of the poll entry, this
is in preparation for allowing multiple waitqueues.
No functional changes in this patch.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
IO_WQ_WORK_CB is used only for linked timeouts, which will be armed
before the work setup (i.e. mm, override creds, etc). The setup
shouldn't take long, so it's ok to arm it a bit later and get rid
of IO_WQ_WORK_CB.
Make io-wq call work->func() only once, callbacks will handle the rest.
i.e. the linked timeout handler will do the actual issue. And as a
bonus, it removes an extra indirect call.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
