Re: 3.14-rc2 XFS backtrace because irqs_disabled.

[Oleg Nesterov <oleg@redhat.com>]

On 02/15, Al Viro wrote:
>
> On Sat, Feb 15, 2014 at 03:27:00PM +0100, Oleg Nesterov wrote:
>
> > 1. info->q can be already freed if SIGQUEUE_PREALLOC.
> >
> >    Once get_signal_to_deliver() or any other caller drops ->siglock
> >    another thread can do sys_timer_delete()->sigqueue_free().
>
> How the devil would it find the sucker?

It simply frees the SIGQUEUE_PREALLOC sigqueue, k_itimer->sigq.

> It's off the list already.

Exactly, list_empty(q->list) == T. So release_posix_timer()->sigqueue_free()
assumes we can safely free it.

> > 2. We need to move do_schedule_next_timer() from dequeue_signal()
> >    here.
> >
> >    Otherwise ->q can be reused/overwritten by the next send_sigqueue()
> >    right affter ->siglock is dropped.
>
> Ditto.  We rip them out of queue on collect_signal();

Yes, and dequeue_signal()->do_schedule_next_timer() can trigger another
send_sigqueue() which uses the same SIGQUEUE_PREALLOC sigqueue once we
drop ->siglock.

This is not that bad, but at least ->si_overrun can be overwritten
before __setup_rt_frame()->copy_siginfo_to_user().

Oleg.