Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes:
> On Sat, Sep 25, 2021 at 1:32 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> - io-wq core dump exit fix (me)
>
> Hmm.
>
> That one strikes me as odd.
>
> I get the feeling that if the io_uring thread needs to have that
> signal_group_exit() test, something is wrong in signal-land.
>
> It's basically a "fatal signal has been sent to another thread", and I
> really get the feeling that "fatal_signal_pending()" should just be
> modified to handle that case too.
>
> Because what about a number of other situations where we have that
> "killable" logic (ie "stop waiting for locks or IO if you're just
> going to get killed anyway" - things like lock_page_killable() and
> friends)
>
> Adding Eric, Oleg and Al to the participants, so that somebody else can pipe up.
>
> That piping up may quite possibly be to just tell me I'm being stupid,
> and that this is just a result of some io_uring thread thing, and
> nobody else has this problem.
>
> It's commit 87c169665578 ("io-wq: ensure we exit if thread group is
> exiting") in my tree.
>
> Comments?
I agree it smells.
It smells that there needs to be a test after get_signal returns with a
signal from an io_uring thread.
As I recall io-wq threads block all signals except for SIGSTOP and
SIGKILL. The signal SIGSTOP is never returned from get_signal. So at
a first glance every return from get_signal should be SIGKILL and thus
fatal.
So I don't understand why there needs to be a test at all after
get_signal.
Further the SIGKILL should have been delivered by get_signal so it
should not be pending so I am not certain when fatal_signal_pending.
Can someone please explain commit 15e20db2e0ce ("io-wq: only exit on
fatal signals") to me?
What signals is get_signal returning to be delivered to userspace that
are not fatal and that are ok to ignore?
I think I am missing something.
Eric
