Re: [PATCH 5/5] coredump: ignore non-fatal signals when core dumpingto a pipe

[Mandeep Singh Baines]

On Sat, Feb 16, 2013 at 11:46 AM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> On 02/16, Oleg Nesterov wrote:
>>
>> On 02/16, Mandeep Singh Baines wrote:
>> >
>> > +static int sigkill_pending(struct task_struct *tsk)
>> > +{
>> > +   return  signal_pending(tsk) &&
>> > +           (sigismember(&tsk->pending.signal, SIGKILL) ||
>> > +            sigismember(&tsk->signal->shared_pending.signal, SIGKILL));
>> > +}
>>
>> Why? __fatal_signal_pending() is enough, you do not need to check
>> ->shared_pending. And once again, ignoring the freezer problems I
>> do not think we need this check at all.
>>

The problem is that the kill signal remains in shared pending since
it'll never get dequeued.

localhost ~ # kill -KILL $!
localhost ~ # cat /proc/$!/status | grep -A4 SigPnd
SigPnd: 0000000000000000
ShdPnd: 0000000000000100
SigBlk: 0000000000000000
SigIgn: 0000000000000000
SigCgt: 0000000000000000

Normally a fatal signal will get propagated to the whole group but
that doesn't happen here because GROUP_EXIT is set:

        /*
         * Found a killable thread.  If the signal will be fatal,
         * then start taking the whole group down immediately.
         */
        if (sig_fatal(p, sig) &&
            !(signal->flags & (SIGNAL_UNKILLABLE | SIGNAL_GROUP_EXIT)) &&
            !sigismember(&t->real_blocked, sig) &&
            (sig == SIGKILL || !t->ptrace)) {
                /*
                 * This signal will be fatal to the whole group.
                 */


GROUP_EXIT is set in do_coredump->coredump_wait->zap_threads->zap_process.

What if complete_signal was changed to propagate KILL even if
SIGNAL_GROUP_EXIT is set?

Regards,
Mandeep

>> IOW. Yes, we will probably need to do this change but only to be
>> freezer-friendly.
>
> And, forgot to mention, this logic is not right in the multi-
> threaded case. I mean, you can't assume that 'kill -9 dumpingtask"
> will wake the coredumping thread up. So this sigkill_pending() or
> __fatal_signal_pending() check can only work in the single-threaded
> case.
>
>> --- x/fs/coredump.c
>> +++ x/fs/coredump.c
>> @@ -416,17 +416,17 @@ static void wait_for_dump_helpers(struct
>>       pipe_lock(pipe);
>>       pipe->readers++;
>>       pipe->writers--;
>> +     // TODO: wake_up_interruptible_sync_poll ?
>> +     wake_up_interruptible_sync(&pipe->wait);
>> +     kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
>> +     pipe_unlock(pipe);
>>
>> -     while ((pipe->readers > 1) && (!signal_pending(current))) {
>> -             wake_up_interruptible_sync(&pipe->wait);
>> -             kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
>> -             pipe_wait(pipe);
>> -     }
>> +     wait_event_freezekillable(&pipe->wait, pipe->readers == 1);
>
> I tried to check (but didn't even try to test). I think this should
> work. Assuming that we teach SIGKILL to actually kill the dumper, but
> we need this in any case.
>
> But. Then we need to change pipe_release() to use wake_up_sync_poll()
> (which we do not have). Probably we can do this... but otoh if we protect
> the dumping thread from the non-fatal signals (and again, we need this
> anyway ;) then we can simply do wait_event_freezable().
>
> Damn. I need to think more.
>
> Oleg.
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/