The series prevents submissions from SQPOLL thread when sqo_task is getting killed (or exec()ed). That should solve races introduced by a patch that allowed task_works during cancellation. Jens, IIRC you said that io_rw_reissue() is called only during submission, right? 4/4 depends on that, so 1/4 should help to catch any misbehaviour. I reduced nr_requests but wasn't able to trigger io_resubmit_prep() for iopoll or not. What was the trick again? 4/4 is actually fairly simple, but safety measures and comments make it to bloat. The overhead is pretty negligible, and it allows to kill more, but that's for-next. Pavel Begunkov (4): io_uring: io_rw_reissue lockdep annotations io_uring: inline io_uring_attempt_task_drop() io_uring: add warn_once for io_uring_flush() io_uring: stop SQPOLL submit on creator's death fs/io_uring.c | 97 +++++++++++++++++++++++++++++++++++++-------------- 1 file changed, 71 insertions(+), 26 deletions(-) -- 2.24.0
