Double fixed files for splice/tee are done in a nasty way, it takes 2
ref_node refs, and during the second time it blindly overrides
req->fixed_file_refs hoping that it haven't changed. That works because
all that is done under iouring_lock in a single go but is error-prone.

Bind everything explicitly to a single ref_node and take only one ref,
with current ref_node ordering it's guaranteed to keep all files valid
awhile the request is inflight.

That's mainly a cleanup + preparation for generic resource handling,
but also saves pcpu_ref get/put for splice/tee with 2 fixed files.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

As tasks now cancel only theirs requests, and inflight_wait is awaited
only in io_uring_cancel_files(), which should be called with ->in_idle
set, instead of keeping a separate inflight_wait use tctx->wait.

That will add some spurious wakeups but actually is safer from point of
not hanging the task.

e.g.
task1                   | IRQ
                        | *start* io_complete_rw_common(link)
                        |        link: req1 -> req2 -> req3(with files)
*cancel_files()         |
io_wq_cancel(), etc.    |
                        | put_req(link), adds to io-wq req2
schedule()              |

So, task1 will never try to cancel req2 or req3. If req2 is
long-standing (e.g. read(empty_pipe)), this may hang.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We don't even allow not plain data msg_control, which is disallowed in
__sys_{send,revb}msg_sock(). So no need in fs for IORING_OP_SENDMSG and
IORING_OP_RECVMSG. fs->lock is less contanged not as much as before, but
there are cases that can be, e.g. IOSQE_ASYNC.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If IORING_SETUP_SQPOLL is enabled, sqes are either handled in sq thread
task context or in io worker task context. If current task context is sq
thread, we don't need to check whether should wake up sq thread.

io_iopoll_req_issued() calls wq_has_sleeper(), which has smp_mb() memory
barrier, before this patch, perf shows obvious overhead:
  Samples: 481K of event 'cycles', Event count (approx.): 299807382878
  Overhead  Comma  Shared Object     Symbol
     3.69%  :9630  [kernel.vmlinux]  [k] io_issue_sqe

With this patch, perf shows:
  Samples: 482K of event 'cycles', Event count (approx.): 299929547283
  Overhead  Comma  Shared Object     Symbol
     0.70%  :4015  [kernel.vmlinux]  [k] io_issue_sqe

It shows some obvious improvements.
Signed-off-by: NXiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Both IOPOLL and sqes handling need to acquire uring_lock, combine
them together, then we just need to acquire uring_lock once.
Signed-off-by: NXiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Some static checker reports below warning:
    fs/io_uring.c:6939 io_sq_thread()
    error: uninitialized symbol 'timeout'.

This is a false positive, but let's just initialize 'timeout' to make
sure we don't trip over this.
Reported-by: NDan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NXiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Reviewed-by: NStefano Garzarella <sgarzare@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

There are some issues about current io_sq_thread() implementation:
  1. The prepare_to_wait() usage in __io_sq_thread() is weird. If
multiple ctxs share one same poll thread, one ctx will put poll thread
in TASK_INTERRUPTIBLE, but if other ctxs have work to do, we don't
need to change task's stat at all. I think only if all ctxs don't have
work to do, we can do it.
  2. We use round-robin strategy to make multiple ctxs share one same
poll thread, but there are various condition in __io_sq_thread(), which
seems complicated and may affect round-robin strategy.

To improve above issues, I take below actions:
  1. If multiple ctxs share one same poll thread, only if all all ctxs
don't have work to do, we can call prepare_to_wait() and schedule() to
make poll thread enter sleep state.
  2. To make round-robin strategy more straight, I simplify
__io_sq_thread() a bit, it just does io poll and sqes submit work once,
does not check various condition.
  3. For multiple ctxs share one same poll thread, we choose the biggest
sq_thread_idle among these ctxs as timeout condition, and will update
it when ctx is in or out.
  4. Not need to check EBUSY especially, if io_submit_sqes() returns
EBUSY, IORING_SQ_CQ_OVERFLOW should be set, helper in liburing should
be aware of cq overflow and enters kernel to flush work.
Signed-off-by: NXiaoguang Wang <xiaoguang.wang@linux.alibaba.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Instead of iterating over each request and cancelling it individually in
io_uring_cancel_files(), try to cancel all matching requests and use
->inflight_list only to check if there anything left.

In many cases it should be faster, and we can reuse a lot of code from
task cancellation.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Make io_poll_remove_all() and io_kill_timeouts() to match against files
as well. A preparation patch, effectively not used by now.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_uring_cancel_files() guarantees to cancel all matching requests,
that's not necessary to do that in a loop. Move it up in the callchain
into io_uring_cancel_task_requests().
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_uring_cancel_files() cancels all request that match files regardless
of task. There is no real need in that, cancel only requests of the
specified task. That also handles SQPOLL case as it already changes task
to it.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Add io_match_task() that matches both task and files.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If IORING_SETUP_SQPOLL is set all requests belong to the corresponding
SQPOLL task, so skip task checking in that case and always match.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Inline io_import_iovec() and leave only its former __io_import_iovec()
renamed to the original name. That makes it more obious what is reused in
io_read/write().
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

io_size and iov_count in io_read() and io_write() hold the same value,
kill the last one.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This is the only code that relies on import_iovec() returning
iter.count on success.
This allows a better interface to import_iovec().
Signed-off-by: NDavid Laight <david.laight@aculab.com>
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Reviewed-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

SQPOLL task may find sqo_task->files == NULL and
__io_sq_thread_acquire_files() would leave it unset, so following
fget_many() and others try to dereference NULL and fault. Propagate
an error files are missing.

[  118.962785] BUG: kernel NULL pointer dereference, address:
	0000000000000020
[  118.963812] #PF: supervisor read access in kernel mode
[  118.964534] #PF: error_code(0x0000) - not-present page
[  118.969029] RIP: 0010:__fget_files+0xb/0x80
[  119.005409] Call Trace:
[  119.005651]  fget_many+0x2b/0x30
[  119.005964]  io_file_get+0xcf/0x180
[  119.006315]  io_submit_sqes+0x3a4/0x950
[  119.007481]  io_sq_thread+0x1de/0x6a0
[  119.007828]  kthread+0x114/0x150
[  119.008963]  ret_from_fork+0x22/0x30
Reported-by: NJosef Grieb <josef.grieb@gmail.com>
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Now users who want to get woken when waiting for events should submit a
timeout command first. It is not safe for applications that split SQ and
CQ handling between two threads, such as mysql. Users should synchronize
the two threads explicitly to protect SQ and that will impact the
performance.

This patch adds support for timeout to existing io_uring_enter(). To
avoid overloading arguments, it introduces a new parameter structure
which contains sigmask and timeout.

I have tested the workloads with one thread submiting nop requests
while the other reaping the cqe with timeout. It shows 1.8~2x faster
when the iodepth is 16.
Signed-off-by: NJiufei Xue <jiufei.xue@linux.alibaba.com>
Signed-off-by: NHao Xu <haoxu@linux.alibaba.com>
[axboe: various cleanups/fixes, and name change to SIG_IS_DATA]
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We unconditionally call blk_start_plug() when starting the IO
submission, but we only really should do that if we have more than 1
request to submit AND we're potentially dealing with block based storage
underneath. For any other type of request, it's just a waste of time to
do so.

Add a ->plug bit to io_op_def and set it for read/write requests. We
could make this more precise and check the file itself as well, but it
doesn't matter that much and would quickly become more expensive.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

We've got extra 8 bytes in the 2nd cacheline, put ->fixed_file_refs
there, so inline execution path mostly doesn't touch the 3rd cacheline
for fixed_file requests as well.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Singly linked list for keeping linked requests is enough, because we
almost always operate on the head and traverse forward with the
exception of linked timeouts going 1 hop backwards.

Replace ->link_list with a handmade singly linked list. Also kill
REQ_F_LINK_HEAD in favour of checking a newly added ->list for NULL
directly.

That saves 8B in io_kiocb, is not as heavy as list fixup, makes better
use of cache by not touching a previous request (i.e. last request of
the link) each time on list modification and optimises cache use further
in the following patch, and actually makes travesal easier removing in
the end some lines. Also, keeping invariant in ->list instead of having
REQ_F_LINK_HEAD is less error-prone.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

In preparation for converting singly linked lists for chaining requests,
make linked timeouts save requests that they're responsible for and not
count on doubly linked list for back referencing.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Explicitly save not only a link's head in io_submit_sqe[s]() but the
tail as well. That's in preparation for keeping linked requests in a
singly linked list.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Don't use a single struct for polls and poll remove requests, they have
totally different layouts.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

IORING_OP_UNLINKAT behaves like unlinkat(2) and takes the same flags
and arguments.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

IORING_OP_RENAMEAT behaves like renameat2(), and takes the same flags
etc.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Now that SQPOLL supports non-registered files and grabs the file table,
we can relax the restriction on open/close/accept/connect and allow
them on a ring that is setup with IORING_SETUP_SQPOLL.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

The restriction of needing fixed files for SQPOLL is problematic, and
prevents/inhibits several valid uses cases. With the referenced
files_struct that we have now, it's trivially supportable.

Treat ->files like we do the mm for the SQPOLL thread - grab a reference
to it (and assign it), and drop it when we're done.

This feature is exposed as IORING_FEAT_SQPOLL_NONFIXED.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

This adds support for the shutdown(2) system call, which is useful for
dealing with sockets.

shutdown(2) may block, so we have to punt it to async context.
Suggested-by: NNorman Maurer <norman.maurer@googlemail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

CAP_SYS_ADMIN is too restrictive for a lot of uses cases, allow
CAP_SYS_NICE based on the premise that such users are already allowed
to raise the priority of tasks.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Don't recycle a refnode until we're done with all requests of nodes
ejected before.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Cc: stable@vger.kernel.org # v5.7+
Signed-off-by: NJens Axboe <axboe@kernel.dk>

An active ref_node always can be found in ctx->files_data, it's much
safer to get it this way instead of poking into files_data->ref_list.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Cc: stable@vger.kernel.org # v5.7+
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Zorro reports that an xfstest test case is failing, and it turns out that
for the reissue path we can potentially issue a double completion on the
request for the failure path. There's an issue around the retry as well,
but for now, at least just make sure that we handle the error path
correctly.

Cc: stable@vger.kernel.org
Fixes: b63534c4 ("io_uring: re-issue block requests that failed because of resources")
Reported-by: NZorro Lang <zlang@redhat.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Any attempt to do path resolution on /proc/self from an async worker will
yield -EOPNOTSUPP. We can safely do that resolution from the task itself,
and without blocking, so retry it from there.

Ideally io_uring would know this upfront and not have to go through the
worker thread to find out, but that doesn't currently seem feasible.
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If an application specifies IORING_SETUP_CQSIZE to set the CQ ring size
to a specific size, we ensure that the CQ size is at least that of the
SQ ring size. But in doing so, we compare the already rounded up to power
of two SQ size to the as-of yet unrounded CQ size. This means that if an
application passes in non power of two sizes, we can return -EINVAL when
the final value would've been fine. As an example, an application passing
in 100/100 for sq/cq size should end up with 128 for both. But since we
round the SQ size first, we compare the CQ size of 100 to 128, and return
-EINVAL as that is too small.

Cc: stable@vger.kernel.org
Fixes: 33a107f0 ("io_uring: allow application controlled CQ ring size")
Reported-by: NDan Melnic <dmm@fb.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Break this function into two helpers so that it's obvious that the
trylock versions return a value that must be checked, and the blocking
versions don't require that.  While we're at it, clean up the return
type mismatch.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

We can't just go over linked requests because it may race with linked
timeouts. Take ctx->completion_lock in that case.

Cc: stable@vger.kernel.org # v5.7+
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Previous commit changed how we index the registered credentials, but
neglected to update one spot that is used when the personalities are
iterated through ->show_fdinfo(). Ensure we use the right struct type
for the iteration.

Reported-by: syzbot+a6d494688cdb797bdfce@syzkaller.appspotmail.com
Fixes: 1e6fa521 ("io_uring: COW io_identity on mismatch")
Signed-off-by: NJens Axboe <axboe@kernel.dk>

If there is a long-standing request of one task locking up execution of
deferred requests, and the defer list contains requests of another task
(all files-less), then a potential execution of __io_uring_task_cancel()
by that another task will sleep until that first long-standing request
completion, and that may take long.

E.g.
tsk1: req1/read(empty_pipe) -> tsk2: req(DRAIN)
Then __io_uring_task_cancel(tsk2) waits for req1 completion.

It seems we even can manufacture a complicated case with many tasks
sharing many rings that can lock them forever.

Cancel deferred requests for __io_uring_task_cancel() as well.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>

Current io_match_files() check in io_cqring_overflow_flush() is useless
because requests drop ->files before going to the overflow list, however
linked to it request do not, and we don't check them.
Signed-off-by: NPavel Begunkov <asml.silence@gmail.com>
Signed-off-by: NJens Axboe <axboe@kernel.dk>