Pavel Begunkov [Mon, 4 Oct 2021 19:02:57 +0000 (20:02 +0100)]
io_uring: remove struct io_completion
We keep struct io_completion only as a temporal storage of cflags, Place
it in io_kiocb, it's cleaner, removes extra bits and even might be used
for future optimisations.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:56 +0000 (20:02 +0100)]
io_uring: control ->async_data with a REQ_F flag
->async_data is a slow path, so it won't matter much if we do the clean
up inside io_clean_op(). Moreover, in many cases it's allocated together
with setting one or more of IO_REQ_CLEAN_FLAGS flags, so it'd go through
io_clean_op() anyway.
Control ->async_data allocation with a new flag REQ_F_ASYNC_DATA, so we
can do all the maintainence under io_req_needs_clean() fast check.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:55 +0000 (20:02 +0100)]
io_uring: optimise io_free_batch_list()
Delay reading the next node in io_free_batch_list(), allows the compiler
to load the value a bit later improving register spilling in some cases.
With gcc 11.1 it helped to move @task_refs variable from the stack to a
register and optimises out a couple of per request instructions.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:53 +0000 (20:02 +0100)]
io_uring: optimise ctx referencing by requests
Currenlty, we allocate one ctx reference per request at submission time
and put them at free. It's batched and not so expensive but it still
bloats the kernel, adds 2 function calls for rcu and adds some overhead
for request counting in io_free_batch_list().
Always keep one reference with a request, even when it's freed and in
io_uring request caches. There is extra work at ring exit / quiesce
paths, which now need to put all cached requests. io_ring_exit_work() is
already looping, so it's not a problem. Add hybrid-busy waiting to
io_ctx_quiesce() as well for now.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:52 +0000 (20:02 +0100)]
io_uring: merge CQ and poll waitqueues
->cq_wait and ->poll_wait and waken up in the same manner, use a single
waitqueue for both of them. CQ waiters are queued exclusively, so wake
up should first go over all pollers and that's what we need.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:51 +0000 (20:02 +0100)]
io_uring: don't wake sqpoll in io_cqring_ev_posted
io_cqring_ev_posted() doesn't need to wake SQPOLL, it's either done by
userspace or with task_work, but no action is required on request
completion. Rip off bits waking it up in io_cqring_ev_posted().
Pavel Begunkov [Mon, 4 Oct 2021 19:02:50 +0000 (20:02 +0100)]
io_uring: optimise INIT_WQ_LIST
The invariant of io_wq_work_list is that it's empty IFF ->first is NULL,
so no need to initially set ->last. With now having more users of the
list it may play a role, i.e. used in each tw iteration and on every
completion flushing.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:49 +0000 (20:02 +0100)]
io_uring: optimise request allocation
Even after fully inlining io_alloc_req() my compiler does a NULL check
in the path of successful allocation, no hacks like an empty dereference
help it. Restructure io_alloc_req() by splitting out refilling part, so
the compiler generate a slightly better binary.
Pavel Begunkov [Mon, 4 Oct 2021 19:02:48 +0000 (20:02 +0100)]
io_uring: delay req queueing into compl-batch list
io_req_complete_state() is inlined and used in lots of places, so we
want to keep it concise. Move adding a request into a completion batch
list from io_req_complete_state() into the consumer, i.e.
__io_queue_sqe().
before vs after
text data bss dec hex filename
91894 14002 8 105904 19db0 ./fs/io_uring.o
91046 14002 8 105056 19a60 ./fs/io_uring.o
Pavel Begunkov [Mon, 4 Oct 2021 19:02:46 +0000 (20:02 +0100)]
io_uring: optimise kiocb layout
We want ->comp_list in the second cacheline, which is hotter comparing
to the 3rd. Swap the field with ->link, which is not as hot and
controlled by flags and so not accessed unless there is a link.
By the way add a couple of comments for io_kiocb fields.
Pavel Begunkov [Sat, 2 Oct 2021 18:36:14 +0000 (19:36 +0100)]
io_uring: add flag to not fail link after timeout
For some reason non-off IORING_OP_TIMEOUT always fails links, it's
pretty inconvenient and unnecessary limits chaining after it to hard
linking, which is far from ideal, e.g. doesn't pair well with timeout
cancellation. Add a flag forcing it to not fail links on -ETIME.
Pavel Begunkov [Fri, 1 Oct 2021 17:07:03 +0000 (18:07 +0100)]
io_uring: clean up buffer select
Hiding a pointer to a struct io_buffer in rw.addr is error prone. We
have some place in io_kiocb, so keep kbuf's in a separate field
without aliasing and risks of it being misused.
Pavel Begunkov [Fri, 1 Oct 2021 17:07:00 +0000 (18:07 +0100)]
io_uring: extra a helper for drain init
Add a helper io_init_req_drain for initialising requests with
IOSQE_DRAIN set. Also move bits from preambule of io_drain_req() in
there, because we already modify all the bits needed inside the helper.
Pavel Begunkov [Fri, 24 Sep 2021 21:00:04 +0000 (22:00 +0100)]
io_uring: disable draining earlier
Clear ->drain_active in two more cases where we check for a need of
draining. It's not a bug, but still may lead to some extra requests
being punted to io-wq, and that may be not desirable.
Pavel Begunkov [Fri, 24 Sep 2021 21:00:02 +0000 (22:00 +0100)]
io_uring: kill off ->inflight_entry field
->inflight_entry is not used anymore after converting everything to
single linked lists, remove it. Also adjust io_kiocb layout, so all hot
bits are in first 3 cachelines.
Put an explicit check for number of requests to submit. First,
we can turn while into do-while and it generates better code, and second
that if can be cheaper, e.g. by using CPU flags after sub in
io_sqring_entries().
Pavel Begunkov [Fri, 24 Sep 2021 21:00:00 +0000 (22:00 +0100)]
io_uring: reshuffle queue_sqe completion handling
If a request completed inline the result should only be zero, it's a
grave error otherwise. So, when we see REQ_F_COMPLETE_INLINE it's not
even necessary to check the return code, and the flag check can be moved
earlier.
It's one "if" less for inline completions, and same two checks for it
normally completing (ret == 0). Those are two cases we care about the
most.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:59 +0000 (21:59 +0100)]
io_uring: inline hot path of __io_queue_sqe()
Extract slow paths from __io_queue_sqe() into a function and inline the
hot path. With that we have everything completely inlined on the
submission path up until io_issue_sqe().
Pavel Begunkov [Fri, 24 Sep 2021 20:59:57 +0000 (21:59 +0100)]
io_uring: remove drain_active check from hot path
req->ctx->active_drain is a bit too expensive, partially because of two
dereferences. Do a trick, if we see it set in io_init_req(), set
REQ_F_FORCE_ASYNC and it automatically goes through a slower path where
we can catch it. It's nearly free to do in io_init_req() because there
is already ->restricted check and it's in the same byte of a bitmask.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:56 +0000 (21:59 +0100)]
io_uring: deduplicate io_queue_sqe() call sites
There are two call sites of io_queue_sqe() in io_submit_sqe(), combine
them into one, because io_queue_sqe() is inline and we don't want to
bloat binary, and will become even bigger
text data bss dec hex filename
92126 13986 8 106120 19e88 ./fs/io_uring.o
91966 13986 8 105960 19de8 ./fs/io_uring.o
Pavel Begunkov [Fri, 24 Sep 2021 20:59:54 +0000 (21:59 +0100)]
io_uring: don't pass tail into io_free_batch_list
io_free_batch_list() iterates all requests in the passed in list,
so we don't really need to know the tail but can keep iterating until
meet NULL. Just pass the first node into it and it will be enough.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:52 +0000 (21:59 +0100)]
io_uring: optimise batch completion
First, convert rest of iopoll bits to single linked lists, and also
replace per-request list_add_tail() with splicing a part of slist.
With that, use io_free_batch_list() to put/free requests. The main
advantage of it is that it's now the only user of struct req_batch and
friends, and so they can be inlined. The main overhead there was
per-request call to not-inlined io_req_free_batch(), which is expensive
enough.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:51 +0000 (21:59 +0100)]
io_uring: convert iopoll_completed to store_release
Convert explicit barrier around iopoll_completed to smp_load_acquire()
and smp_store_release(). Similar on the callback side, but replaces a
single smp_rmb() with per-request smp_load_acquire(), neither imply any
extra CPU ordering for x86. Use READ_ONCE as usual where it doesn't
matter.
Use it to move filling CQEs by iopoll earlier, that will be necessary
to avoid traversing the list one extra time in the future.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:48 +0000 (21:59 +0100)]
io_uring: split iopoll loop
The main loop of io_do_iopoll() iterates and does ->iopoll() until it
meets a first completed request, then it continues from that position
and splices requests to pass them through io_iopoll_complete().
Split the loop in two for clearness, iopolling and reaping completed
requests from the list.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:46 +0000 (21:59 +0100)]
io-wq: add io_wq_work_node based stack
Apart from just using lists (i.e. io_wq_work_list), we also want to have
stacks, which are a bit faster, and have some interoperability between
them. Add a stack implementation based on io_wq_work_node and some
helpers.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:44 +0000 (21:59 +0100)]
io_uring: use slist for completion batching
Currently we collect requests for completion batching in an array.
Replace them with a singly linked list. It's as fast as arrays but
doesn't take some much space in ctx, and will be used in future patches.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:42 +0000 (21:59 +0100)]
io_uring: force_nonspin
We don't really need to pass the number of requests to complete into
io_do_iopoll(), a flag whether to enforce non-spin mode is enough.
Should be straightforward, maybe except io_iopoll_check(). We pass !min
there, because we do never enter with the number of already reaped
requests is larger than the specified @min, apart from the first
iteration, where nr_events is 0 and so the final check should be
identical.
Pavel Begunkov [Fri, 24 Sep 2021 20:59:41 +0000 (21:59 +0100)]
io_uring: mark having different creds unlikely
Hint the compiler that it's not as likely to have creds different from
current attached to a request. The current code generation is far from
ideal, hopefully it can help to some compilers to remove duplicated jump
tables and so.
Pavel Begunkov [Wed, 15 Sep 2021 11:03:38 +0000 (12:03 +0100)]
io_uring: optimise io_req_init() sqe flags checks
IOSQE_IO_DRAIN is quite marginal and we don't care too much about
IOSQE_BUFFER_SELECT. Save to ifs and hide both of them under
SQE_VALID_FLAGS check. Now we first check whether it uses a "safe"
subset, i.e. without DRAIN and BUFFER_SELECT, and only if it's not
true we test the rest of the flags.
Pavel Begunkov [Wed, 15 Sep 2021 11:04:20 +0000 (12:04 +0100)]
io_uring: remove ctx referencing from complete_post
Now completions are done from task context, that means that it's either
the task itself, task_work or io-wq worker. In all those cases the ctx
will be staying alive by mutexing, explicit referencing or req references
by iowq. Remove extra ctx pinning from io_req_complete_post().
Developers may need some uring info to help themselves debug and address
issues in production. This includes sqring/cqring head/tail and the
detailed sqe/cqe info, which is very useful when an application is hung
on a ring.
Pavel Begunkov [Wed, 8 Sep 2021 15:40:52 +0000 (16:40 +0100)]
io_uring: dedup CQE flushing non-empty checks
We don't do io_submit_flush_completions() when there is no requests
enqueued, and every single caller checks for it. Hide that check into
the function not forgetting about inlining. That will make it much
easier for changing the empty check condition in the future.
Pavel Begunkov [Wed, 8 Sep 2021 15:40:49 +0000 (16:40 +0100)]
io_uring: kill off ios_left
->ios_left is only used to decide whether to plug or not, kill it to
avoid this extra accounting, just use the initial submission number.
There is no much difference in regards of enabling plugging, where this
one does it in a few more cases, but all major ones should be covered
well.
I recently had to look at a production problem where a request ended
up getting the dreaded -EINVAL error on submit. The most used and
hence useless of error codes, as it just tells you that something
was wrong with your request, but not more than that.
Let's dump the full sqe contents if we run into an issue failure,
that'll allow easier diagnosing of a wide variety of issues.
Jens Axboe [Tue, 19 Oct 2021 02:54:39 +0000 (20:54 -0600)]
block: fix too broad elevator check in blk_mq_free_request()
We added RQF_ELV to tell whether there's an IO scheduler attached, and
RQF_ELVPRIV tells us whether there's an IO scheduler with private data
attached. Don't check RQF_ELV in blk_mq_free_request(), what we care
about here is just if we have scheduler private data attached.
This fixes a boot crash
Fixes: 2ff0682da6e0 ("block: store elevator state in request") Reported-by: Yi Zhang <yi.zhang@redhat.com> Reported-by: syzbot+eb8104072aeab6cc1195@syzkaller.appspotmail.com Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Mon, 18 Oct 2021 14:45:39 +0000 (08:45 -0600)]
nvme: wire up completion batching for the IRQ path
Trivial to do now, just need our own io_comp_batch on the stack and pass
that in to the usual command completion handling.
I pondered making this dependent on how many entries we had to process,
but even for a single entry there's no discernable difference in
performance or latency. Running a sync workload over io_uring:
Jens Axboe [Tue, 12 Oct 2021 15:28:46 +0000 (09:28 -0600)]
io_uring: utilize the io batching infrastructure for more efficient polled IO
Wire up using an io_comp_batch for f_op->iopoll(). If the lower stack
supports it, we can handle high rates of polled IO more efficiently.
This raises the single core efficiency on my system from ~6.1M IOPS to
~6.6M IOPS running a random read workload at depth 128 on two gen2
Optane drives.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Fri, 8 Oct 2021 11:59:37 +0000 (05:59 -0600)]
nvme: add support for batched completion of polled IO
Take advantage of struct io_comp_batch, if passed in to the nvme poll
handler. If it's set, rather than complete each request individually
inline, store them in the io_comp_batch list. We only do so for requests
that will complete successfully, anything else will be completed inline as
before.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Fri, 8 Oct 2021 11:50:46 +0000 (05:50 -0600)]
block: add support for blk_mq_end_request_batch()
Instead of calling blk_mq_end_request() on a single request, add a helper
that takes the new struct io_comp_batch and completes any request stored
in there.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Mon, 18 Oct 2021 14:53:19 +0000 (08:53 -0600)]
block: remove some blk_mq_hw_ctx debugfs entries
Just like the blk_mq_ctx counterparts, we've got a bunch of counters
in here that are only for debugfs and are of questionnable value. They
are:
- dispatched, index of how many requests were dispatched in one go
- poll_{considered,invoked,success}, which track poll sucess rates. We're
confident in the iopoll implementation at this point, don't bother
tracking these.
As a bonus, this shrinks each hardware queue from 576 bytes to 512 bytes,
dropping a whole cacheline.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
These were added as part of early days debugging for blk-mq, and they
are not really useful anymore. Rather than spend cycles updating them,
just get rid of them.
As a bonus, this shrinks the per-cpu software queue size from 256b
to 192b. That's a whole cacheline less.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pavel Begunkov [Mon, 18 Oct 2021 20:37:27 +0000 (21:37 +0100)]
block: skip elevator fields init for non-elv queue
Don't init rq->hash and rq->rb_node in blk_mq_rq_ctx_init() if there is
no elevator. Also, move some other initialisers that imply barriers to
the end, so the compiler is free to rearrange and optimise other the
rest of them.
note: fold in a change from Jens leaving queue_list unconditional, as
it might lead to problems otherwise.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Fri, 15 Oct 2021 15:44:38 +0000 (09:44 -0600)]
block: store elevator state in request
Add an rq private RQF_ELV flag, which tells the block layer that this
request was initialized on a queue that has an IO scheduler attached.
This allows for faster checking in the fast path, rather than having to
deference rq->q later on.
Elevator switching does full quiesce of the queue before detaching an
IO scheduler, so it's safe to cache this in the request itself.
Jens Axboe [Sat, 16 Oct 2021 02:06:18 +0000 (20:06 -0600)]
block: only mark bio as tracked if it really is tracked
We set BIO_TRACKED unconditionally when rq_qos_throttle() is called, even
though we may not even have an rq_qos handler. Only mark it as TRACKED if
it really is potentially tracked.
This saves considerable time for the case where the bio isn't tracked:
2.64% -1.65% [kernel.vmlinux] [k] bio_endio
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Fri, 15 Oct 2021 21:03:52 +0000 (15:03 -0600)]
block: improve layout of struct request
It's been a while since this was analyzed, move some members around to
better flow with the use case. Initial state up top, and queued state
after that. This improves my peak case by about 1.5%, from 7750K to
7900K IOPS.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Thu, 14 Oct 2021 15:17:01 +0000 (09:17 -0600)]
block: move update request helpers into blk-mq.c
For some reason we still have them in blk-core, with the rest of the
request completion being in blk-mq. That causes and out-of-line call
for each completion.
Move them into blk-mq.c instead, where they belong.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Jens Axboe [Wed, 13 Oct 2021 15:01:43 +0000 (09:01 -0600)]
block: don't bother iter advancing a fully done bio
If we're completing nbytes and nbytes is the size of the bio, don't bother
with calling into the iterator increment helpers. Just clear the bio
size and we're done.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pavel Begunkov [Thu, 14 Oct 2021 14:03:26 +0000 (15:03 +0100)]
block: cache request queue in bdev
There are tons of places where we need to get a request_queue only
having bdev, which turns into bdev->bd_disk->queue. There are probably a
hundred of such places considering inline helpers, and enough of them
are in hot paths.
Cache queue pointer in struct block_device and make use of it in
bdev_get_queue().
Jens Axboe [Thu, 14 Oct 2021 17:17:43 +0000 (11:17 -0600)]
block: use flags instead of bit fields for blkdev_dio
This generates a lot better code for me, and bumps performance from
7650K IOPS to 7750K IOPS. Looking at profiles for the run and running
perf diff, it confirms that we're now sending a lot less time there:
Pavel Begunkov [Wed, 13 Oct 2021 08:57:11 +0000 (09:57 +0100)]
block: cache bdev in struct file for raw bdev IO
bdev = &BDEV_I(file->f_mapping->host)->bdev
Getting struct block_device from a file requires 2 memory dereferences
as illustrated above, that takes a toll on performance, so cache it in
yet unused file->private_data. That gives a noticeable peak performance
improvement.
block: don't allow writing to the poll queue attribute
The poll attribute is a historic artefact from before when we had
explicit poll queues that require driver specific configuration.
Just print a warning when writing to the attribute.
Signed-off-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Tested-by: Mark Wunderlich <mark.wunderlich@intel.com> Link: https://lore.kernel.org/r/20211012111226.760968-16-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
Replace the blk_poll interface that requires the caller to keep a queue
and cookie from the submissions with polling based on the bio.
Polling for the bio itself leads to a few advantages:
- the cookie construction can made entirely private in blk-mq.c
- the caller does not need to remember the request_queue and cookie
separately and thus sidesteps their lifetime issues
- keeping the device and the cookie inside the bio allows to trivially
support polling BIOs remapping by stacking drivers
- a lot of code to propagate the cookie back up the submission path can
be removed entirely.
Ming Lei [Tue, 12 Oct 2021 11:12:23 +0000 (13:12 +0200)]
block: define 'struct bvec_iter' as packed
'struct bvec_iter' is embedded into 'struct bio', define it as packed
so that we can get one extra 4bytes for other uses without expanding
bio.
'struct bvec_iter' is often allocated on stack, so making it packed
doesn't affect performance. Also I have run io_uring on both
nvme/null_blk, and not observe performance effect in this way.
Suggested-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Reviewed-by: Hannes Reinecke <hare@suse.de> Signed-off-by: Christoph Hellwig <hch@lst.de> Tested-by: Mark Wunderlich <mark.wunderlich@intel.com> Link: https://lore.kernel.org/r/20211012111226.760968-14-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk>
This flags ensures that the pages will not be reused for non-bio
allocations before the end of an RCU grace period. With that we can
safely use a RCU lookup for bio polling as long as we are fine with
occasionally polling the wrong device.
Unlike the RWF_HIPRI userspace ABI which is intentionally kept vague,
the bio flag is specific to the polling implementation, so rename and
document it properly.
block: replace the spin argument to blk_iopoll with a flags argument
Switch the boolean spin argument to blk_poll to passing a set of flags
instead. This will allow to control polling behavior in a more fine
grained way.
Signed-off-by: Christoph Hellwig <hch@lst.de> Tested-by: Mark Wunderlich <mark.wunderlich@intel.com> Link: https://lore.kernel.org/r/20211012111226.760968-10-hch@lst.de
[axboe: adapt to changed io_uring iopoll] Signed-off-by: Jens Axboe <axboe@kernel.dk>
The polling support in the legacy direct-io support is a little crufty.
It already doesn't support the asynchronous polling needed for io_uring
polling, and is hard to adopt to upcoming changes in the polling
interfaces. Given that all the major file systems already use the iomap
direct I/O code, just drop the polling support.
Jens Axboe [Thu, 14 Oct 2021 13:24:07 +0000 (07:24 -0600)]
block: only check previous entry for plug merge attempt
Currently we scan the entire plug list, which is potentially very
expensive. In an IOPS bound workload, we can drive about 5.6M IOPS with
merging enabled, and profiling shows that the plug merge check is the
(by far) most expensive thing we're doing:
Instead of browsing the whole list, just check the previously inserted
entry. That is enough for a naive merge check and will catch most cases,
and for devices that need full merging, the IO scheduler attached to
such devices will do that anyway. The plug merge is meant to be an
inexpensive check to avoid getting a request, but if we repeatedly
scan the list for every single insert, it is very much not a cheap
check.
With this patch, the workload instead runs at ~7.0M IOPS, providing
a 25% improvement. Disabling merging entirely yields another 5%
improvement.
Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
projects / thirdparty / kernel / linux.git / log
