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>

Like IORING_OP_READV, this is limited to supporting just a single
segment in the iovec passed in.
Signed-off-by: Jens Axboe <axboe@kernel.dk>

This splits it into two parts, one that imports the message, and one
that imports the iovec. This allows a caller to only do the first part,
and import the iovec manually afterwards.

No functional changes in this patch.
Acked-by: David Miller <davem@davemloft.net>
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>

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>

When executing non-linked hashed work, io_worker_handle_work()
will lock-unlock wqe->lock to update hash, and then immediately
lock-unlock to get next work. Optimise this case and do
lock/unlock only once.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

There are 2 optimisations:
- Now, io_worker_handler_work() do io_assign_current_work() twice per
request, and each one adds lock/unlock(worker->lock) pair. The first is
to reset worker->cur_work to NULL, and the second to set a real work
shortly after. If there is a dependant work, set it immediately, that
effectively removes the extra NULL'ing.

- And there is no use in taking wqe->lock for linked works, as they are
not hashed now. Optimise it out.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

This is a preparation patch, it adds some helpers and makes
the next patches cleaner.

- extract io_impersonate_work() and io_assign_current_work()
- replace @next label with nested do-while
- move put_work() right after NULL'ing cur_work.
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>

Don't abuse labels for plain and straightworward code.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
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: d7718a9d ("io_uring: use poll driven retry for files that support it")
Link: https://github.com/ClangBuiltLinux/linux/issues/916Signed-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>

As Peter pointed out, task_work() can avoid ->pi_lock and cmpxchg()
if task->task_works == NULL && !PF_EXITING.

And in fact the only reason why task_work_run() needs ->pi_lock is
the possible race with task_work_cancel(), we can optimize this code
and make the locking more clear.
Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

@hash_map is unsigned long, but BIT_ULL() is used for manipulations.
BIT() is a better match as it returns exactly unsigned long value.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
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>

IO_WQ_WORK_HAS_MM is set but never used, remove it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_recvmsg() and io_sendmsg() duplicate nonblock -EAGAIN finilising
part, so add helper for that.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Deduplicate call to io_cqring_fill_event(), plain and easy
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Add support for splice(2).

- output file is specified as sqe->fd, so it's handled by generic code
- hash_reg_file handled by generic code as well
- len is 32bit, but should be fine
- the fd_in is registered file, when SPLICE_F_FD_IN_FIXED is set, which
is a splice flag (i.e. sqe->splice_flags).
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Preparation without functional changes. Adds io_get_file(), that allows
to grab files not only into req->file.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Make do_splice(), so other kernel parts can reuse it
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

req->in_async is not really needed, it only prevents propagation of
@nxt for fast not-blocked submissions. Remove it.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_prep_async_worker() called io_wq_assign_next() do many useless checks:
io_req_work_grab_env() was already called during prep, and @do_hashed
is not ever used. Add io_prep_next_work() -- simplified version, that
can be called io-wq.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Many operations define custom work.func before getting into an io-wq.
There are several points against:
- it calls io_wq_assign_next() from outside io-wq, that may be confusing
- sync context would go unnecessary through io_req_cancelled()
- prototypes are quite different, so work!=old_work looks strange
- makes async/sync responsibilities fuzzy
- adds extra overhead

Don't call generic path and io-wq handlers from each other, but use
helpers instead
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Don't drop an early reference, hang on to it and let the caller drop
it. This makes it behave more like "regular" requests.
Signed-off-by: Jens Axboe <axboe@kernel.dk>

If the -EAGAIN happens because of a static condition, then a poll
or later retry won't fix it. We must call it again from blocking
condition. Play it safe and ensure that any -EAGAIN condition from read
or write must retry from async context.
Signed-off-by: Jens Axboe <axboe@kernel.dk>

io_wq_flush() is buggy, during cancelation of a flush, the associated
work may be passed to the caller's (i.e. io_uring) @match callback. That
callback is expecting it to be embedded in struct io_kiocb. Cancelation
of internal work probably doesn't make a lot of sense to begin with.

As the flush helper is no longer used, just delete it and the associated
work flag.
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>