The memset() used is measurably slower in targeted benchmarks, wasting
about 1% of the total runtime, or 50% of the (later) hot path cached
bio alloc. Get rid of it and fill in the bio manually.
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_prep_linked_timeout() grew too heavy and compiler now refuse to
inline the function. Help it by splitting in two and annotating with
inline.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/560636717a32e9513724f09b9ecaace942dde4d4.1628705069.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

As submission references are gone, there is only one initial reference
left. Instead of actually doing atomic refcounting, add a flag
indicating whether we're going to take more refs or doing any other sync
magic. The flag should be set before the request may get used in
parallel.

Together with the previous patch it saves 2 refcount atomics per request
for IOPOLL and IRQ completions, and 1 atomic per req for inline
completions, with some exceptions. In particular, currently, there are
three cases, when the refcounting have to be enabled:
- Polling, including apoll. Because double poll entries takes a ref.
  Might get relaxed in the near future.
- Link timeouts, enabled for both, the timeout and the request it's
  bound to, because they work in-parallel and we need to synchronise
  to cancel one of them on completion.
- When a request gets in io-wq, because it doesn't hold uring_lock and
  we need guarantees of submission references.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/8b204b6c5f6643062270a1913d6d3a7f8f795fd9.1628705069.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Requests are by default given with two references, submission and
completion. Completion references are straightforward, they represent
request ownership and are put when a request is completed or so.
Submission references are a bit more trickier. They're needed when
io_issue_sqe() followed deep into the submission stack (e.g. in fs,
block, drivers, etc.), request may have given away for concurrent
execution or already completed, and the code unwinding back to
io_issue_sqe() may be accessing some pieces of our requests, e.g.
file or iov.

Now, we prevent such async/in-depth completions by pushing requests
through task_work. Punting to io-wq is also done through task_works,
apart from a couple of cases with a pretty well known context. So,
there're two cases:
1) io_issue_sqe() from the task context and protected by ->uring_lock.
Either requests return back to io_uring or handed to task_work, which
won't be executed because we're currently controlling that task. So,
we can be sure that requests are staying alive all the time and we don't
need submission references to pin them.

2) io_issue_sqe() from io-wq, which doesn't hold the mutex. The role of
submission reference is played by io-wq reference, which is put by
io_wq_submit_work(). Hence, it should be fine.

Considering that, we can carefully kill the submission reference.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/6b68f1c763229a590f2a27148aee77767a8d7750.1628705069.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Soon, we won't need to put several references at once, remove
req_ref_sub_and_test() and @nr argument from io_put_req_deferred(),
and put the rest of the references by hand.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1868c7554108bff9194fb5757e77be23fadf7fc0.1628705069.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Move all request refcount helpers to avoid forward declarations in the
future.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/89fd36f6f3fe5b733dfe4546c24725eee40df605.1628705069.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

We have no hard/soft IRQ users of this lock left, remove any IRQ
disabling/saving and restoring when grabbing this lock.

This is straight forward with no users entering with IRQs disabled
anymore, the only thing to look out for is the waitqueue poll head
lock which nests inside the completion lock. That needs IRQs disabled,
and hence we have to do that now instead of relying on the outer lock
doing so.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This is in preparation to making the completion lock work outside of
hard/soft IRQ context.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This is in preparation to making the completion lock work outside of
hard/soft IRQ context.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This is in preparation to making the completion lock work outside of
hard/soft IRQ context.

Add a timeout_lock to handle the ordering of timeout completions or
cancelations with the timeouts actually triggering.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

For requests with non-fixed files, instead of grabbing just one
reference, we get by the number of left requests, so the following
requests using the same file can take it without atomics.

However, it's not all win. If there is one request in the middle
not using files or having a fixed file, we'll need to put back the left
references. Even worse if an application submits requests dealing with
different files, it will do a put for each new request, so doubling the
number of atomics needed. Also, even if not used, it's still takes some
cycles in the submission path.

If a file used many times, it rather makes sense to pre-register it, if
not, we may fall in the described pitfall. So, this optimisation is a
matter of use case. Go with the simpliest code-wise way, remove it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

After recent fixes, tctx_task_work() always does proper spinlocking
before looking into ->task_list, so now we don't need atomics for
->task_state, replace it with non-atomic task_running using the critical
section.

Tide it up, combine two separate block with spinlocking, and always try
to splice in there, so we do less locking when new requests are arriving
during the function execution.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
[axboe: fix missing ->task_running reset on task_work_add() failure]
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Inline io_poll_remove_waitqs() into its only user and clean it up.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/2f1a91a19ffcd591531dc4c61e2f11c64a2d6a6d.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Unlike __io_cqring_overflow_flush(), nobody does forced flushing with
io_cqring_overflow_flush(), so removed the argument from it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/7594f869ca41b7cfb5a35a3c7c2d402242834e9e.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Inline struct io_comp_state into struct io_submit_state. They are
already coupled tightly, together with mixed responsibilities it
only brings confusion having them separately.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/e55bba77426b399e3a2e54e3c6c267c6a0fc4b57.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

req->compl.list is used to cache freed requests, and so can't overlap in
time with req->inflight_entry. So, use inflight_entry to link requests
and remove compl.list.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/e430e79d22d70a190d718831bda7bfed1daf8976.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

We don't use @tsk argument of io_req_cache_free(), remove it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/6a28b4a58ee0aaf0db98e2179b9c9f06f9b0cca1.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Don't kfree requests in __io_free_req() but put them back into the
internal request cache. That makes allocations more sustainable and will
be used for refcounting optimisations.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/9f4950fbe7771c8d41799366d0a3a08ac3040236.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Move io_fallback_req_func() to kill yet another forward declaration.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/d0a8f9d9a0057ed761d6237167d51c9378798d2d.1628536684.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

We cache all the reference to task + tctx, so if io_put_task() is
called by the corresponding task itself, we can save on atomics and
return the refs right back into the cache.

It's beneficial for all inline completions, and also iopolling, when
polling and submissions are done by the same task, including
SQPOLL|IOPOLL.

Note: io_uring_cancel_generic() can return refs to the cache as well,
so those should be flushed in the loop for tctx_inflight() to work
right.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/6fe9646b3cb70e46aca1f58426776e368c8926b3.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

In case of on-exec io_uring cancellations, tasks already wait for all
submitted requests to get completed/cancelled, so we don't need to check
for ->in_execve separately.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/be8707049f10df9d20ca03dc4ca3316239b5e8e0.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

IO_IOPOLL_BATCH is not used, delete it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/b2bdf19dbee2c9fc8865bbab9412135a14e24a64.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

If io_ring_exit_work() can't get it done in 5 minutes, something is
going very wrong, don't keep spinning at HZ / 20 rate, it doesn't help
and it may take much of CPU time if there is a lot of workers stuck as
such.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/9e2d1ca81d569f6bc628af1a42ff6663bff7ce9c.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Move IORING_SETUP_IOPOLL check into __io_openat_prep(), so both openat
and openat2 reuse it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/9a73ce83e4ee60d011180ef177eecef8e87ff2a2.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Inline io_free_req_deferred(), there is no reason to keep it separated.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/ce04b7180d4eac0d69dd00677b227eefe80c2cc5.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Move the function together with io_rsrc_node_ref_zero() in the source
file as it is to get rid of forward declarations.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/4d81f6f833e7d017860b24463a9a68b14a8a5ed2.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Move the function in the source file as it is to get rid of forward
declarations.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/33d917d69e4206557c75a5b98fe22bcdf77ce47d.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Refactor __io_uring_register() by extracting a helper responsible for
ctx queisce. Looks better and will make it easier to add more
optimisations.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/0339e0027504176be09237eefa7945bf9a6f153d.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Turns out we always init struct io_wait_queue in io_cqring_wait(), even
if it's not used after, i.e. there are already enough of CQEs. And often
it's exactly what happens, for instance, requests may have been
completed inline, or in case of io_uring_enter(submit=N, wait=1).

It shows up in my profiler, so optimise it by delaying the struct init.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/6f1b81c60b947d165583dc333947869c3d85d037.1628471125.git.asml.silence@gmail.com
[axboe: fixed up for new cqring wait]
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Add more annotations for submission path functions holding ->uring_lock.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/128ec4185e26fbd661dd3a424aa66108ee8ff951.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

IOPOLL users should care more about getting completions for requests
they submitted, but not in "device did/completed something". Currently,
io_do_iopoll() may return a positive number, which will instruct
io_iopoll_check() to break the loop and end the syscall, even if there
is not enough CQEs or none at all.

Don't return positive numbers, so io_iopoll_check() exits only when it
gets an actual error, need reschedule or got enough CQEs.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/641a88f751623b6758303b3171f0a4141f06726e.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Replace the main if of io_flush_cached_reqs() with inverted condition +
goto, so all the cases are handled in the same way. And also extract
io_preinit_req() to make it cleaner and easier to refer to.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/1abcba1f7b55dc53bf1dbe95036e345ffb1d5b01.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Prepare nodes that we're going to add before actually linking them, it's
always safer and costs us nothing.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/f7e53f0c84c02ed6748c488ed0789b98f8cc6185.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

We prefer nornal task_works even if it would fail requests inside. Kill
a PF_EXITING check in io_req_task_work_add(), task_work_add() handles
well dying tasks, i.e. return error when can't enqueue due to late
stages of do_exit().
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/fc14297e8441cd8f5d1743a2488cf0df09bf48ac.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Move io-wq callbacks closer to each other, so it's easier to work with
them, and rename io_free_work() into io_wq_free_work() for consistency.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/851bbc7f0f86f206d8c1333efee8bcb9c26e419f.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

If we use fixed files, we can be sure (almost) that REQ_F_ISREG is set.
However, for non-reg files io_prep_rw() still will look into inode to
double check, and that's expensive and can be avoided.

The only caveat is that it only currently works with 64+ bit
architectures, see FFS_ISREG, so we should consider that.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/0a62780c491ca2522cd52db4ae3f16e03aafed0f.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

io_file_supports_async() checks whether a file supports nowait
operations, so "async" in the name is misleading. Rename it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/33d55b5ce43aa1884c637c1957f1e30d30dc3bec.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Optimise io_file_get() with registered files, which is in a hot path,
by inlining parts of the function. Saves a function call, and
inefficiencies of passing arguments, e.g. evaluating
(sqe_flags & IOSQE_FIXED_FILE).

It couldn't have been done before as compilers were refusing to inline
it because of the function size.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/52115cd6ce28f33bd0923149c0e6cb611084a0b1.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Instead of hand-coded two-level tables for registered files, allocate
them with kvmalloc(). In many cases small enough tables are enough, and
so can be kmalloc()'ed removing an extra memory load and a bunch of bit
logic instructions from the hot path. If the table is larger, we trade
off all the pros with a TLB-assisted memory lookup.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/280421d3b48775dabab773006bb5588c7b2dabc0.1628471125.git.asml.silence@gmail.comSigned-off-by: NJens Axboe <axboe@kernel.dk>

Currently we only wake the first waiter, even if we have enough entries
posted to satisfy multiple waiters. Improve that situation so that
every waiter knows how much the CQ tail has to advance before they can
be safely woken up.

With this change, if we have N waiters each asking for 1 event and we get
4 completions, then we wake up 4 waiters. If we have N waiters asking
for 2 completions and we get 4 completions, then we wake up the first
two. Previously, only the first waiter would've been woken up.
Signed-off-by: NJens Axboe <axboe@kernel.dk>