We set empty registered buffers to dummy_ubuf as an optimisation.
Currently, we allocate the dummy entry for each ring, whenever we can
simply have one global instance.
We're casting out const on assignment, it's fine as we're not going to
change the content of the dummy, the constness gives us an extra layer
of protection if sth ever goes wrong.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/e4a96dda35ab755914bc43f6781bba0df97ac489.1691757663.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now all callers of io_aux_cqe() set allow_overflow to false, remove the
parameter and not allow overflowing auxilary multishot cqes.
When CQ is full the function callers and all multishot requests in
general are expected to complete the request. That prevents indefinite
in-background grows of the overflow list and let's the userspace to
handle the backlog at its own pace.
Resubmitting a request should also be faster than accounting a bunch of
overflows, so it should be better for perf when it happens, but a well
behaving userspace should be trying to avoid overflows in any case.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/bb20d14d708ea174721e58bb53786b0521e4dd6d.1691757663.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We never use io_move_task_work_from_local() before it's defined in the
file anyway, so kill the forward declaration.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We return 0 for success, or -error when there's an error. Move the 'ret'
variable into the loop where we are actually using it, to make it
clearer that we don't carry this variable forward for return outside of
the loop.
While at it, also move the need_resched() break condition out of the
while check itself, keeping it with the signal pending check.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_req_local_work_add() peeks into the work list, which can be executed
in the meanwhile. It's completely fine without KASAN as we're in an RCU
read section and it's SLAB_TYPESAFE_BY_RCU. With KASAN though it may
trigger a false positive warning because internal io_uring caches are
sanitised.
Remove sanitisation from the io_uring request cache for now.
Cc: stable@vger.kernel.org
Fixes: 8751d15426a31 ("io_uring: reduce scheduling due to tw")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/c6fbf7a82a341e66a0007c76eefd9d57f2d3ba51.1691541473.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
cq_extra is protected by ->completion_lock, which io_get_sqe() misses.
The bug is harmless as it doesn't happen in real life, requires invalid
SQ index array and racing with submission, and only messes up the
userspace, i.e. stall requests execution but will be cleaned up on
ring destruction.
Fixes: 15641e427070f ("io_uring: don't cache number of dropped SQEs")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/66096d54651b1a60534bb2023f2947f09f50ef73.1691538547.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
I observed poor performance of io_uring compared to synchronous IO. That
turns out to be caused by deeper CPU idle states entered with io_uring,
due to io_uring using plain schedule(), whereas synchronous IO uses
io_schedule().
The losses due to this are substantial. On my cascade lake workstation,
t/io_uring from the fio repository e.g. yields regressions between 20%
and 40% with the following command:
./t/io_uring -r 5 -X0 -d 1 -s 1 -c 1 -p 0 -S$use_sync -R 0 /mnt/t2/fio/write.0.0
This is repeatable with different filesystems, using raw block devices
and using different block devices.
Use io_schedule_prepare() / io_schedule_finish() in
io_cqring_wait_schedule() to address the difference.
After that using io_uring is on par or surpassing synchronous IO (using
registered files etc makes it reliably win, but arguably is a less fair
comparison).
There are other calls to schedule() in io_uring/, but none immediately
jump out to be similarly situated, so I did not touch them. Similarly,
it's possible that mutex_lock_io() should be used, but it's not clear if
there are cases where that matters.
Cc: stable@vger.kernel.org # 5.10+
Cc: Pavel Begunkov <asml.silence@gmail.com>
Cc: io-uring@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Andres Freund <andres@anarazel.de>
Link: https://lore.kernel.org/r/20230707162007.194068-1-andres@anarazel.de
[axboe: minor style fixup]
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring offloads task_work for cancelation purposes when the task is
exiting. This is conceptually fine, but we should be nicer and actually
wait for that work to complete before returning.
Add an argument to io_fallback_tw() telling it to flush the deferred
work when it's all queued up, and have it flush a ctx behind whenever
the ctx changes.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is no reason not to use __io_cq_unlock_post_flush for intermediate
aux CQE flushing, all ->task_complete should apply there, i.e. if set it
should be the submitter task. Combine them, get rid of of
__io_cq_unlock_post() and rename the left function.
This place was also taking a couple percents of CPU according to
profiles for max throughput net benchmarks due to multishot recv
flooding it with completions.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/bbed60734cbec2e833d9c7bdcf9741aada5d8aab.1687518903.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We're abusing ->completion_lock helpers. io_cq_unlock() neither
locking conditionally nor doing CQE flushing, which means that callers
must have some side reason of taking the lock and should do it directly.
Open code io_cq_unlock() into io_cqring_overflow_kill() and clean it up.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/7dabb36856db2b562e78780480396c52c29b2bf4.1687518903.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Extract a function for non-local task_work_add, and use it directly from
io_move_task_work_from_local(). Now we don't use IOU_F_TWQ_FORCE_NORMAL
and it can be killed.
As a small positive side effect we don't grab task->io_uring in
io_req_normal_work_add anymore, which is not needed for
io_req_local_work_add().
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/2e55571e8ff2927ae3cc12da606d204e2485525b.1687518903.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We're trying to batch io_put_task() in io_free_batch_list(), but
considering that the hot path is a simple inc, it's most cerainly and
probably faster to just do io_put_task() instead of task tracking.
We don't care about io_put_task_remote() as it's only for IOPOLL
where polling/waiting is done by not the submitter task.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/4a7ef7dce845fe2bd35507bf389d6bd2d5c1edf0.1687518903.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Request completion is a very hot path in general, but there are 3 places
that can be doing it: io_free_batch_list(), io_req_complete_post() and
io_free_req_tw().
io_free_req_tw() is used rather marginally and we don't care about it.
Killing it can help to clean up and optimise the left two, do that by
replacing it with io_req_task_complete().
There are two things to consider:
1) io_free_req() is called when all refs are put, so we need to reinit
references. The easiest way to do that is to clear REQ_F_REFCOUNT.
2) We also don't need a cqe from it, so silence it with REQ_F_CQE_SKIP.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/434a2be8f33d474ad888ce1c17fe5ea7bbcb2a55.1687518903.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Remove all the open coded magic on slot->file_ptr by introducing two
helpers that return the file pointer and the flags instead.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20230620113235.920399-9-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Two of the three callers want them, so return the more usual format,
and shift into the FFS_ form only for the fixed file table.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20230620113235.920399-6-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The SCM inflight mechanism has nothing to do with the fact that a file
might be a regular file or not and if it supports non-blocking
operations.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20230620113235.920399-4-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now that this only checks O_NONBLOCK and FMODE_NOWAIT, the helper is
complete overkilļ, and the comments are confusing bordering to wrong.
Just inline the check into the caller.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20230620113235.920399-2-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
WHen the ring exits, cleanup is done and the final cancelation and
waiting on completions is done by io_ring_exit_work. That function is
invoked by kworker, which doesn't take any signals. Because of that, it
doesn't really matter if we wait for completions in TASK_INTERRUPTIBLE
or TASK_UNINTERRUPTIBLE state. However, it does matter to the hung task
detection checker!
Normally we expect cancelations and completions to happen rather
quickly. Some test cases, however, will exit the ring and park the
owning task stopped (eg via SIGSTOP). If the owning task needs to run
task_work to complete requests, then io_ring_exit_work won't make any
progress until the task is runnable again. Hence io_ring_exit_work can
trigger the hung task detection, which is particularly problematic if
panic-on-hung-task is enabled.
As the ring exit doesn't take signals to begin with, have it wait
interruptibly rather than uninterruptibly. io_uring has a separate
stuck-exit warning that triggers independently anyway, so we're not
really missing anything by making this switch.
Cc: stable@vger.kernel.org # 5.10+
Link: https://lore.kernel.org/r/b0e4aaef-7088-56ce-244c-976edeac0e66@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Everybody is passing in the request, so get rid of the io_ring_ctx and
explicit user_data pass-in. Both the ctx and user_data can be deduced
from the request at hand.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use task_work for a variety of reasons, but doing completions or
triggering rety after poll are by far the hottest two. Use the indirect
funtion call wrappers to avoid the indirect function call if
CONFIG_RETPOLINE is set.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use lockless lists for the local and deferred task_work, which means
that when we queue up events for processing, we ultimately process them
in reverse order to how they were received. This usually doesn't matter,
but for some cases, it does seem to make a big difference. Do the right
thing and reverse the list before processing it, so that we know it's
processed in the same order in which it was received.
This makes a rather big difference for some medium load network tests,
where consistency of performance was a bit all over the place. Here's
a case that has 4 connections each doing two sends and receives:
io_uring port=10002: rps:161.13k Bps: 1.45M idle=256ms
io_uring port=10002: rps:107.27k Bps: 0.97M idle=413ms
io_uring port=10002: rps:136.98k Bps: 1.23M idle=321ms
io_uring port=10002: rps:155.58k Bps: 1.40M idle=268ms
and after the change:
io_uring port=10002: rps:205.48k Bps: 1.85M idle=140ms user=40ms
io_uring port=10002: rps:203.57k Bps: 1.83M idle=139ms user=20ms
io_uring port=10002: rps:218.79k Bps: 1.97M idle=106ms user=30ms
io_uring port=10002: rps:217.88k Bps: 1.96M idle=110ms user=20ms
io_uring port=10002: rps:222.31k Bps: 2.00M idle=101ms user=0ms
io_uring port=10002: rps:218.74k Bps: 1.97M idle=102ms user=20ms
io_uring port=10002: rps:208.43k Bps: 1.88M idle=125ms user=40ms
using more of the time to actually process work rather than sitting
idle.
No effects have been observed at the peak end of the spectrum, where
performance is still the same even with deep batch depths (and hence
more items to sort).
Signed-off-by: Jens Axboe <axboe@kernel.dk>
With IORING_REGISTER_USE_REGISTERED_RING, an application can register
the ring fd and use it via registered index rather than installed fd.
This allows using a registered ring for everything *except* the initial
mmap.
With IORING_SETUP_NO_MMAP, io_uring_setup uses buffers allocated by the
user, rather than requiring a subsequent mmap.
The combination of the two allows a user to operate *entirely* via a
registered ring fd, making it unnecessary to ever install the fd in the
first place. So, add a flag IORING_SETUP_REGISTERED_FD_ONLY to make
io_uring_setup register the fd and return a registered index, without
installing the fd.
This allows an application to avoid touching the fd table at all, and
allows a library to never even momentarily install a file descriptor.
This splits out an io_ring_add_registered_file helper from
io_ring_add_registered_fd, for use by io_uring_setup.
Signed-off-by: Josh Triplett <josh@joshtriplett.org>
Link: https://lore.kernel.org/r/bc8f431bada371c183b95a83399628b605e978a3.1682699803.git.josh@joshtriplett.org
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently io_uring applications must call mmap(2) twice to map the rings
themselves, and the sqes array. This works fine, but it does not support
using huge pages to back the rings/sqes.
Provide a way for the application to pass in pre-allocated memory for
the rings/sqes, which can then suitably be allocated from shmfs or
via mmap to get huge page support.
Particularly for larger rings, this reduces the TLBs needed.
If an application wishes to take advantage of that, it must pre-allocate
the memory needed for the sq/cq ring, and the sqes. The former must
be passed in via the io_uring_params->cq_off.user_data field, while the
latter is passed in via the io_uring_params->sq_off.user_data field. Then
it must set IORING_SETUP_NO_MMAP in the io_uring_params->flags field,
and io_uring will then map the existing memory into the kernel for shared
use. The application must not call mmap(2) to map rings as it otherwise
would have, that will now fail with -EINVAL if this setup flag was used.
The pages used for the rings and sqes must be contigious. The intent here
is clearly that huge pages should be used, otherwise the normal setup
procedure works fine as-is. The application may use one huge page for
both the rings and sqes.
Outside of those initialization changes, everything works like it did
before.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We do rings and sqes separately, move them into a helper that does both
the freeing and clearing of the memory.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In preparation for having more than one time of ring allocator, make the
existing one return valid/error-pointer rather than just NULL.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We only have two reserved members we're not clearing, do so manually
instead. This is in preparation for using one of these members for
a new feature.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Now that we have both sockets and block devices setting FMODE_NOWAIT
appropriately, we can get rid of all the odd special casing in
__io_file_supports_nowait() and rely soley on FMODE_NOWAIT and
O_NONBLOCK rather than special case sockets and (in particular) bdevs.
Link: https://lore.kernel.org/r/20230509151910.183637-4-axboe@kernel.dk
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----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=Erfk
-----END PGP SIGNATURE-----
Merge tag 'for-6.4/io_uring-2023-04-21' of git://git.kernel.dk/linux
Pull io_uring updates from Jens Axboe:
- Cleanup of the io-wq per-node mapping, notably getting rid of it so
we just have a single io_wq entry per ring (Breno)
- Followup to the above, move accounting to io_wq as well and
completely drop struct io_wqe (Gabriel)
- Enable KASAN for the internal io_uring caches (Breno)
- Add support for multishot timeouts. Some applications use timeouts to
wake someone waiting on completion entries, and this makes it a bit
easier to just have a recurring timer rather than needing to rearm it
every time (David)
- Support archs that have shared cache coloring between userspace and
the kernel, and hence have strict address requirements for mmap'ing
the ring into userspace. This should only be parisc/hppa. (Helge, me)
- XFS has supported O_DIRECT writes without needing to lock the inode
exclusively for a long time, and ext4 now supports it as well. This
is true for the common cases of not extending the file size. Flag the
fs as having that feature, and utilize that to avoid serializing
those writes in io_uring (me)
- Enable completion batching for uring commands (me)
- Revert patch adding io_uring restriction to what can be GUP mapped or
not. This does not belong in io_uring, as io_uring isn't really
special in this regard. Since this is also getting in the way of
cleanups and improvements to the GUP code, get rid of if (me)
- A few series greatly reducing the complexity of registered resources,
like buffers or files. Not only does this clean up the code a lot,
the simplified code is also a LOT more efficient (Pavel)
- Series optimizing how we wait for events and run task_work related to
it (Pavel)
- Fixes for file/buffer unregistration with DEFER_TASKRUN (Pavel)
- Misc cleanups and improvements (Pavel, me)
* tag 'for-6.4/io_uring-2023-04-21' of git://git.kernel.dk/linux: (71 commits)
Revert "io_uring/rsrc: disallow multi-source reg buffers"
io_uring: add support for multishot timeouts
io_uring/rsrc: disassociate nodes and rsrc_data
io_uring/rsrc: devirtualise rsrc put callbacks
io_uring/rsrc: pass node to io_rsrc_put_work()
io_uring/rsrc: inline io_rsrc_put_work()
io_uring/rsrc: add empty flag in rsrc_node
io_uring/rsrc: merge nodes and io_rsrc_put
io_uring/rsrc: infer node from ctx on io_queue_rsrc_removal
io_uring/rsrc: remove unused io_rsrc_node::llist
io_uring/rsrc: refactor io_queue_rsrc_removal
io_uring/rsrc: simplify single file node switching
io_uring/rsrc: clean up __io_sqe_buffers_update()
io_uring/rsrc: inline switch_start fast path
io_uring/rsrc: remove rsrc_data refs
io_uring/rsrc: fix DEFER_TASKRUN rsrc quiesce
io_uring/rsrc: use wq for quiescing
io_uring/rsrc: refactor io_rsrc_ref_quiesce
io_uring/rsrc: remove io_rsrc_node::done
io_uring/rsrc: use nospec'ed indexes
...
Replace completions with waitqueues for rsrc data quiesce, the main
wakeup condition is when data refs hit zero. Note that data refs are
only changes under ->uring_lock, so we prepare before mutex_unlock()
reacquire it after taking the lock back. This change will be needed
in the next patch.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1d0dbc74b3b4fd67c8f01819e680c5e0da252956.1681395792.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
So far io_req_complete_post() only covers DEFER_TASKRUN by completing
request via task work when the request is completed from IOWQ.
However, uring command could be completed from any context, and if io
uring is setup with DEFER_TASKRUN, the command is required to be
completed from current context, otherwise wait on IORING_ENTER_GETEVENTS
can't be wakeup, and may hang forever.
The issue can be observed on removing ublk device, but turns out it is
one generic issue for uring command & DEFER_TASKRUN, so solve it in
io_uring core code.
Fixes: e6aeb2721d3b ("io_uring: complete all requests in task context")
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/linux-block/b3fc9991-4c53-9218-a8cc-5b4dd3952108@kernel.dk/
Reported-by: Jens Axboe <axboe@kernel.dk>
Cc: Kanchan Joshi <joshi.k@samsung.com>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use io_rsrc_node_switch() coupled with io_rsrc_node_switch_start()
for a bunch of cases including initialising ctx->rsrc_node, i.e. by
passing NULL instead of rsrc_data. Leave it to only deal with actual
node changing.
For that, first remove it from io_uring_create() and add a function
allocating the first node. Then also remove all calls to
io_rsrc_node_switch() from files/buffers register as we already have a
node installed and it does essentially nothing.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d146fe306ff98b1a5a60c997c252534f03d423d7.1681210788.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_uring/io_uring.c:432 io_prep_async_work() error: we previously
assumed 'req->file' could be null (see line 425).
Even though it's a false positive as there will not be REQ_F_ISREG set
without a file, let's add a simple check to make the kernel test robot
happy. We don't care about performance here, but assumingly it'll be
optimised out by the compiler.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/a6cfbe92c74b789c0b4f046f7f98d19b1ca2e5b7.1681210788.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
