On Wed, Mar 21, 2018 at 08:32:30AM +0100, Christoph Hellwig wrote: > The upcoming aio poll support would like to be able to complete the > iocb inline from the cancellation context, but that would cause > a lock order reversal. Add support for optionally moving the cancelation > outside the context lock to avoid this reversal. Ouch... Seeing that you've just taken out cmpxchg loop out of kiocb_cancel() with "serialized on ->ctx_lock" for explanation of safety... Let me check the aio_poll side of it; this commit might be better off in the poll series, *if* it is actually correct. What's to prevent double completions there? Suppose we have iocb sitting on the wait queue; cancellation callback set, so's "delayed cancel" flag. Now, somebody tries to cancel the fucker on CPU1. With ctx->lock held the sucker is found on the list and, just as we mark it "cancelled", driver sends a wakeup, executing (on CPU2) aio_poll_wake(), calling aio_complete_poll() (without ctx->lock, so no exclusion with io_cancel(2) on CPU1), which checks AIO_IOCB_CANCELLED and does not notice the flag being set on CPU1, then proceeds to __aio_complete_poll() and fput() in there. In the meanwhile, CPU1 has taken the sucker off the list, dropped the lock and called kiocb_cancel() on it. Now we get aio_poll_cancel() and __aio_complete_poll() on CPU1, with *another* fput(). What am I missing here that would prevent such a race?
