On 19/01/2021 13:12, Joseph Qi wrote: > On 1/19/21 7:45 PM, Pavel Begunkov wrote: >> On 19/01/2021 08:00, Joseph Qi wrote: >>> On 1/19/21 10:38 AM, Pavel Begunkov wrote: >>>> On 19/01/2021 01:58, Joseph Qi wrote: >>>>>> Hmm, I hastened, for files we need IO_WQ_WORK_FILES, >>>>>> +IO_WQ_WORK_BLKCG for same reasons. needs_file would make >>>>>> it to grab a struct file, that is wrong. >>>>>> Probably worked out because it just grabbed fd=0/stdin. >>>>>> >>>>> >>>>> I think IO_WQ_WORK_FILES can work since it will acquire >>>>> files when initialize async cancel request. >>>> >>>> That the one controlling files in the first place, need_file >>>> just happened to grab them submission. >>>> >>>>> Don't quite understand why we should have IO_WQ_WORK_BLKCG. >>>> >>>> Because it's set for IORING_OP_CLOSE, and similar situation >>>> may happen but with async_cancel from io-wq. >>>> >>> So how about do switch and restore in io_run_cancel(), seems it can >>> take care of direct request, sqthread and io-wq cases. >> >> It will get ugly pretty quickly, + this nesting of io-wq handlers >> async_handler() -> io_close() is not great... >> >> I'm more inclined to skip them in io_wqe_cancel_pending_work() >> to not execute inline. That may need to do some waiting on the >> async_cancel side though to not change the semantics. Can you >> try out this direction? >> > Sure, I'll try this way and send v2. There may be a much better way, that's to remove IO_WQ_WORK_NO_CANCEL and move -EAGAIN section of io_close() before close_fd_get_file(), so not splitting it in 2 and not keeping it half-done. IIRC, it was done this way because of historical reasons when we didn't put more stuff around files, but may be wrong. Jens, do you remember what it was? -- Pavel Begunkov
