signals not reliably interrupting io_uring_enter anymore

[Andres Freund <andres@anarazel.de>]

Hi,

I haven't yet fully analyzed the problem, but after updating to
cdd3bb54332f82295ed90cd0c09c78cd0c0ee822 io_uring using postgres does
not work reliably anymore.

The symptom is that io_uring_enter(IORING_ENTER_GETEVENTS) isn't
interrupted by signals anymore. The signal handler executes, but
afterwards the syscall is restarted. Previously io_uring_enter reliably
returned EINTR in that case.

Currently postgres relies on signals interrupting io_uring_enter(). We
probably can find a way to not do so, but it'd not be entirely trivial.

I suspect the issue is

commit ce593a6c480a22acba08795be313c0c6d49dd35d (tag: io_uring-5.8-2020-07-01, linux-block/io_uring-5.8)
Author: Jens Axboe <axboe@kernel.dk>
Date:   2020-06-30 12:39:05 -0600

    io_uring: use signal based task_work running

as that appears to have changed the error returned by
io_uring_enter(GETEVENTS) after having been interrupted by a signal from
EINTR to ERESTARTSYS.


I'll check to make sure that the issue doesn't exist before the above
commit.


Greetings,

Andres Freund