Re: [PATCH] coredump: Retry writes where appropriate

[Oleg Nesterov <oleg@redhat.com>]

On 06/03, Roland McGrath wrote:
>
> > But since the coredumping task is not freezable anyway, perhaps we should
> > change fake_signal_wake_up() to ignore SIGNAL_GROUP_DUMPING task.
>
> That could be a long delay and a lot of i/o before suspending.
>
> > Or we should make the coredumping freezable. This means dump_write/seek
> > and exit_mm() should do try_to_freeze().
>
> Yes, I think this is the thing to do for that issue.

Fortunately, this doesn't look hard. Whatever we do, we should modify
dump_write/seek to check fatal_signal_pending() anyway. Because we can't
know if f_ops->write() pays attention to signals. This means we can just
add try_to_freeze().

As for exit_mm(), we can use freezer_do_not_count() + freezer_count()
around the "for (;;)" loop.

> > In any case, the coredumping is special. If ->write() returns -ERESTART/EINTR
> > it assumes the return to ths user-space, this is not true for the coredump.
> > This means that handling the spurious signals in coredump_file_write() is
> > not so bad if we can't avoid this.
>
> I am not so confident.  It seems far too easy to wind up with some other
> way that TIF_SIGPENDING gets continually set and this loops, for example.
> (This could be some day in the future when fs, driver or pipe-io code
> changes somehow completely obscure.)  It's far better to have confidence
> just in the signals code itself: the only things that set TIF_SIGPENDING
> interlock with the logic of the only things that are expected to clear it.

Looks like, if we introduce a difference between "really killed" tasks and
exiting/execing/coredumping tasks (as discussed in another thread), we get
this all for free.

Oleg.