Re: [6.5-rc5 regression] core dump hangs (was Re: [Bug report] fstests generic/051 (on xfs) hang on latest linux v6.5-rc5+)

[Zorro Lang]

On Mon, Jun 12, 2023 at 03:45:12AM -0500, Eric W. Biederman wrote:
> Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes:
> 
> > On Sun, Jun 11, 2023 at 10:49 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> >>
> >> On Sun, Jun 11, 2023 at 10:34:29PM -0700, Linus Torvalds wrote:
> >> >
> >> > So that "!=" should obviously have been a "==".
> >>
> >> Same as without the condition - all the fsstress tasks hang in
> >> do_coredump().
> >
> > Ok, that at least makes sense. Your "it made things worse" made me go
> > "What?" until I noticed the stupid backwards test.
> >
> > I'm not seeing anything else that looks odd in that commit
> > f9010dbdce91 ("fork, vhost: Use CLONE_THREAD to fix freezer/ps
> > regression").
> >
> > Let's see if somebody else goes "Ahh" when they wake up tomorrow...
> 
> It feels like there have been about half a dozen bugs pointed out in
> that version of the patch.  I am going to have to sleep before I can get
> as far as "Ahh"
> 
> One thing that really stands out for me is.
> 
> if (test_if_loop_should_continue) {
> 	set_current_state(TASK_INTERRUPTIBLE);
>         schedule();
> }
> 
> /* elsewhere */
> llist_add(...);
> wake_up_process()
> 
> So it is possible that the code can sleep indefinitely waiting for a
> wake-up that has already come, because the order of set_current_state
> and the test are in the wrong order.
> 
> Unfortunately I don't see what would effect a coredump on a process that
> does not trigger the vhost_worker code.
> 
> 
> 
> About the only thing I can image is if io_uring is involved.  Some of
> the PF_IO_WORKER code was changed, and the test
> "((t->flags & (PF_USER_WORKER | PF_IO_WORKER)) != PF_USER_WORKER)"
> was sprinkled around.
> 
> That is the only code outside of vhost specific code that was changed.
> 
> 
> Is io_uring involved in the cases that hang?

Oh, right, I involved io_uring into in fstests' fsstress.c, and I built kernel
with CONFIG_IO_URING=y. If Darrick (said he didn't hit this issue) didn't enable
io_uring, that might mean it's io_uring related.

> 
> 
> Eric
>