Re: [RFC][PATCH] signals: don't copy siginfo_t on dequeue

From: Oleg Nesterov
Date: Thu Feb 26 2009 - 15:40:07 EST


On 02/26, Vegard Nossum wrote:
>
> 2009/2/26 Oleg Nesterov <oleg@xxxxxxxxxx>:
> > So. dequeue_signal() returns NULL if there is no siginfo queued. In that
> > case we assume that the signal is not pending.
> >
> > But this is not right. Think about SEND_SIG_FORCED, or __sigqueue_alloc()
> > failure when the signal is sent. Or look at zap_other_threads() for example,
> > it just sets the bit in ->pending but doesn't queue siginfo.
>
> I will investigate.

Cough. Well, I must admit I am a bit skeptical about this patch ;) Because
I suspect it will add more complications to the code. And _I think_ avoiding
copy_siginfo() does not buy too much. I will be happy if I am wrong, though.

But. If you are going to do another version, then please note there is another
problem with this patch, SIGQUEUE_PREALLOC.

If collect_signal() returns SIGQUEUE_PREALLOC info, we can not drop ->siglock.
I mean, once we drop ->siglock, this info can be freed, so for example

spin_unlock(&tsk->sighand->siglock);
- do_schedule_next_timer(info);
+ do_schedule_next_timer(&signal->info);

even this part is not safe.

Also. The patch uses __sigqueue_free() to free the delivered siginfo, but
this is not safe without ->siglock, we can race with sigqueue_free().

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/