io_uring: poll rework

It's not possible to go forward with the current state of io_uring
polling, we need a more straightforward and easier synchronisation.
There are a lot of problems with how it is at the moment, including
missing events on rewait.

The main idea here is to introduce a notion of request ownership while
polling, no one but the owner can modify any part but ->poll_refs of
struct io_kiocb, that grants us protection against all sorts of races.

Main users of such exclusivity are poll task_work handler, so before
queueing a tw one should have/acquire ownership, which will be handed
off to the tw handler.
The other user is __io_arm_poll_handler() do initial poll arming. It
starts taking the ownership, so tw handlers won't be run until it's
released later in the function after vfs_poll. note: also prevents
races in __io_queue_proc().
Poll wake/etc. may not be able to get ownership, then they need to
increase the poll refcount and the task_work should notice it and retry
if necessary, see io_poll_check_events().
There is also IO_POLL_CANCEL_FLAG flag to notify that we want to kill
request.

It makes cancellations more reliable, enables double multishot polling,
fixes double poll rewait, fixes missing poll events and fixes another
bunch of races.

Even though it adds some overhead for new refcounting, and there are a
couple of nice performance wins:
- no req->refs refcounting for poll requests anymore
- if the data is already there (once measured for some test to be 1-2%
  of all apoll requests), it removes it doesn't add atomics and removes
  spin_lock/unlock pair.
- works well with multishots, we don't do remove from queue / add to
  queue for each new poll event.

Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
Link: https://lore.kernel.org/r/6b652927c77ed9580ea4330ac5612f0e0848c946.1639605189.git.asml.silence@gmail.com


Signed-off-by: Jens Axboe <axboe@kernel.dk>