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>
The req pointer uniquely identify a specific request.
Having it in traces can provide valuable insights that is not possible
to have if the calling process is reusing the same user_data value.
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Olivier Langlois <olivier@trillion01.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
In most cases io_commit_cqring() is just an smp_store_release(), and
it's hot enough, especially for IRQ rw, to want it to save on a function
call. Mark it inline and extract a non-inlined slow path doing drain
and timeout flushing. The inlined part is pretty slim to not cause
binary bloating.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/7350f8b6b92caa50a48a80be39909f0d83eddd93.1623772051.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Place all drain_next logic into io_drain_req(), so it's never executed
if there was no drained requests before. The only thing we need is to
set ->drain_active if we see a request with IOSQE_IO_DRAIN, do that in
io_init_req() where flags are definitely in registers.
Also, all drain-related code is encapsulated in io_drain_req(), makes it
cleaner.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/68bf4f7395ddaafbf1a26bd97b57d57d45a9f900.1623772051.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
->drain_used is one way, which is not optimal if users use DRAIN but
very rarely. However, we can just clear it in io_drain_req() when all
drained before requests are gone. Also rename the flag to reflect the
change and be more clear about it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/7f37a240857546a94df6348507edddacab150460.1623772051.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
fs/io_uring.c: In function 'io_alloc_page_table':
include/linux/minmax.h:20:28: warning: comparison of distinct pointer
types lacks a cast
Cast everything to size_t using min_t.
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Fixes: 9123c8ffce ("io_uring: add helpers for 2 level table alloc")
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/50f420a956bca070a43810d4a805293ed54f39d8.1623759527.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The sqe_ptr argument has been gone since 709b302fad (io_uring:
simplify io_get_sqring, 2020-04-08), made the return value of the
function. Update the comment accordingly.
Signed-off-by: Fam Zheng <fam.zheng@bytedance.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20210604164256.12242-1-fam.zheng@bytedance.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Replace drain checks with one-way flag set upon seeing the first
IOSQE_IO_DRAIN request. There are several places where it cuts cycles
well:
1) It's much faster than the fast check with two
conditions in io_drain_req() including pretty complex
list_empty_careful().
2) We can mark io_queue_sqe() inline now, that's a huge win.
3) It replaces timeout and drain checks in io_commit_cqring() with a
single flags test. Also great not touching ->defer_list there without a
reason so limiting cache bouncing.
It adds a small amount of overhead to drain path, but it's negligible.
The main nuisance is that once it meets any DRAIN request in io_uring
instance lifetime it will _always_ go through a slower path, so
drain-less and offset-mode timeout less applications are preferable.
The overhead in that case would be not big, but it's worth to bear in
mind.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/98d2fff8c4da5144bb0d08499f591d4768128ea3.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
->uring_lock is prevalently used for submission, even though it protects
many other things like iopoll, registeration, selected bufs, and more.
And it's placed together with ->cq_wait poked on completion and CQ
waiting sides. Move them apart, ->uring_lock goes to the submission
data, and cq_wait to completion related chunk. The last one requires
some reshuffling so everything needed by io_cqring_ev_posted*() is in
one cacheline.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/dea5e845caee4c98aa0922b46d713154d81f7bd8.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We use several wait_queue_head's for different purposes, but namings are
confusing. First rename ctx->cq_wait into ctx->poll_wait, because this
one is used for polling an io_uring instance. Then rename ctx->wait into
ctx->cq_wait, which is responsible for CQE waiting.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/47b97a097780c86c67b20b6ccc4e077523dce682.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There are no users of ->sq_check_overflow, only ->cq_check_overflow is
used. Combine it and move out of completion related part of struct
io_ring_ctx.
A not so obvious benefit of it is fitting all completion side fields
into a single cacheline. It was taking 2 lines before with 56B padding,
and io_cqring_ev_posted*() were still touching both of them.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/25927394964df31d113e3c729416af573afff5f5.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
submit_state.link is used only to assemble a link and not used for
actual submission, so clear it before io_queue_sqe() in io_submit_sqe(),
awhile it's hot and in caches and queueing doesn't spoil it. May also
potentially help compiler with spilling or to do other optimisations.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1579939426f3ad6b55af3005b1389bbbed7d780d.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
io_commit_cqring() might be very hot and we definitely don't want to
touch ->timeout_list there, because 1) it's shared with the submission
side so might lead to cache bouncing and 2) may need to load an extra
cache line, especially for IRQ completions.
We're interested in it at the completion side only when there are
offset-mode timeouts, which are not so popular. Replace
list_empty(->timeout_list) hot path check with a new one-way flag, which
is set when we prepare the first offset-mode timeout.
note: the flag sits in the same line as briefly used after ->rings
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/e4892ec68b71a69f92ffbea4a1499be3ec0d463b.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Kill ->cached_sq_dropped and wire DRAIN sequence number correction via
->cq_extra, which is there exactly for that purpose. User visible
dropped counter will be populated by incrementing it instead of keeping
a copy, similarly as it was done not so long ago with cq_overflow.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/088aceb2707a534d531e2770267c4498e0507cc1.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since moving locked_free_* out of struct io_submit_state
ctx->submit_state is accessed on submission side only, so move it into
the submission section. Same goes for rsrc table pointers/nodes/etc.,
they must be taken and checked during submission because sync'ed by
uring_lock, so move them there as well.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/8a5899a50afc6ccca63249e716f580b246f3dec6.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
ctx->flags are heavily used by both, completion and submission sides, so
move it out from the ctx fields related to submissions. Instead, place
it together with ctx->refs, because it's already cacheline-aligned and
so pads lots of space, and both almost never change. Also, in most
occasions they are accessed together as refs are taken at submission
time and put back during completion.
Do same with ctx->rings, where the pointer itself is never modified
apart from ring init/free.
Note: in percpu mode, struct percpu_ref doesn't modify the struct itself
but takes indirection with ref->percpu_count_ptr.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/4c48c173e63d35591383ba2b87e8b8e8dfdbd23d.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
sq_array and sq_sqes are always used together, however they are in
different cachelines, where the borderline is right before
cq_overflow_list is rather rarely touched. Move the fields together so
it loads only one cacheline.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/3ef2411a94874da06492506a8897eff679244f49.1623709150.git.asml.silence@gmail.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Static analysis is warning that the sizeof being used is should be
of *data->tags[i] and not data->tags[i]. Although these are the same
size on 64 bit systems it is not a portable assumption to assume
this is true for all cases. Fix this by using a temporary pointer
tag_slot to make the code a clearer.
Addresses-Coverity: ("Sizeof not portable")
Fixes: d878c81610 ("io_uring: hide rsrc tag copy into generic helpers")
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/20210615130011.57387-1-colin.king@canonical.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Hao Xu