[PATCHSET v2] Improve io_uring cancellations

[Jens Axboe <axboe@kernel.dk>]

Hot on the heels of v1, here's v2 of this patchset. Due to some
brainfart on my end, I forgot that we have the worker ref count
to prevent workers from going away. So we don't need the ->exiting
change from v1. The other two items were totally valid, however.
To summarize, they are:

- While the wqe->lock protects the stability of the worker->cur_work
  pointer, it does NOT stabilize the actual work item. For cancelling
  specific work, io_uring needs to dereference it to see if it matches.

- We're not consistent in comparing worker->cur_work and sending a
  signal to cancel it.

These three patches fix the above issues. The first is a prep patch that
adds referencing of work items, which makes it safe to dereference
->cur_work if we ensure that ->cur_work itself is stable. The second
closes the hole that currently exists with signaling specific work for
cancellation, and the last one closes a tiny hole where we might miss a
worker during list iteration.

Changes since v1:
- Get rid of ->exiting, we don't need it
- Add third patch on closing list iteration hole

 fs/io-wq.c    | 113 +++++++++++++++++++++++++++-----------------------
 fs/io-wq.h    |   7 +++-
 fs/io_uring.c |  17 +++++++-
 3 files changed, 84 insertions(+), 53 deletions(-)

-- 
Jens Axboe