On 9/7/23 16:36, Jens Axboe wrote:
On 9/1/23 9:13 AM, Jens Axboe wrote:
On Fri, 01 Sep 2023 21:49:16 +0800, Ming Lei wrote:
io_wq_put_and_exit() is called from do_exit(), but all FIXED_FILE requests
in io_wq aren't canceled in io_uring_cancel_generic() called from do_exit().
Meantime io_wq IO code path may share resource with normal iopoll code
So if any HIPRI request is submittd via io_wq, this request may not get resouce
for moving on, given iopoll isn't possible in io_wq_put_and_exit().
[1/1] io_uring: fix IO hang in io_wq_put_and_exit from do_exit()
This causes a regression with the test/thread-exit.t test case, as it's
canceling requests from other tasks as well. I will drop this patch for
And this one has never hit my mailbox... Anyway, I'm confused with
1) request tracking is done only for io_uring polling io_uring, which
shouldn't be the case for t/io_uring, so it's probably unrelated?
2) In case of iopoll, io-wq only submits a request but doesn't wait/poll
for it. If io_issue_sqe() returned -EAGAIN or an error, the request is
considered not yet submitted to block and can be just cancelled normally
without any dancing like io_iopoll_try_reap_events().
3) If we condense the code it sounds like it effectively will be
void io_wq_exit_start(struct io_wq *wq)
We set the flag, interrupt workers (TIF_NOTIFY_SIGNAL + wake up), and
wait for them. Workers are woken up (or just return), see
the flag and return. At least that's how it was intended to work.
What's missing? Racing for IO_WQ_BIT_EXIT? Not breaking on IO_WQ_BIT_EXIT
correctly? Not honouring / clearing TIF_NOTIFY_SIGNAL?
I'll try to repro later