On Thu, Mar 07, 2019 at 12:03:10AM +0000, Al Viro wrote: > From: Al Viro <viro@xxxxxxxxxxxxxxxxxx> > > In case of early wakeups, aio_poll() assumes that aio_poll_complete() > has either already happened or is imminent. In that case we do not > want to put iocb on the list of cancellables. However, ignored > wakeups need to be treated as if wakeup has not happened at all. > Trivially fixed by having aio_poll_wake() set ->woken only after > it's committed to taking iocb out of the waitqueue. > > Spotted-by: zhengbin <zhengbin13@xxxxxxxxxx> > Signed-off-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx> ... and unfortunately it's worse than just that - what both of us have missed is that one could have non-specific wakep + schedule_work + aio_poll_complete_work() rechecking ->poll(), seeing nothing of interest and reinserting into queue. All before vfs_poll() manages to return into aio_poll(). The window is harder to hit, but it's still there, with exact same "failed to add to cancel list" kind of bug if we do hit it ;-/
