tree 4fa2f7e11cc757160ae5f492fdbe0da72d30cf26
parent 829841146878e082613a49581ae252c071057c23
author Oleg Nesterov <oleg@tv-sign.ru> 1128692779 +0400
committer Linus Torvalds <torvalds@g5.osdl.org> 1128808411 -0700

[PATCH] fix do_coredump() vs SIGSTOP race

Let's suppose we have 2 threads in thread group:
	A - does coredump
	B - has pending SIGSTOP

thread A						thread B

do_coredump:						get_signal_to_deliver:

  lock(->sighand)
  ->signal->flags = SIGNAL_GROUP_EXIT
  unlock(->sighand)

							lock(->sighand)
							signr = dequeue_signal()
								->signal->flags |= SIGNAL_STOP_DEQUEUED
								return SIGSTOP;

							do_signal_stop:
							    unlock(->sighand)

  coredump_wait:

      zap_threads:
          lock(tasklist_lock)
          send SIGKILL to B
              // signal_wake_up() does nothing
          unlock(tasklist_lock)

							    lock(tasklist_lock)
							    lock(->sighand)
							    re-check sig->flags & SIGNAL_STOP_DEQUEUED, yes
							    set_current_state(TASK_STOPPED);
							    finish_stop:
							        schedule();
							            // ->state == TASK_STOPPED

      wait_for_completion(&startup_done)
         // waits for complete() from B,
         // ->state == TASK_UNINTERRUPTIBLE

We can't wake up 'B' in any way:

	SIGCONT will be ignored because handle_stop_signal() sees
	->signal->flags & SIGNAL_GROUP_EXIT.

	sys_kill(SIGKILL)->__group_complete_signal() will choose
	uninterruptible 'A', so it can't help.

	sys_tkill(B, SIGKILL) will be ignored by specific_send_sig_info()
	because B already has pending SIGKILL.

This scenario is not possbile if 'A' does do_group_exit(), because
it sets sig->flags = SIGNAL_GROUP_EXIT and delivers SIGKILL to
subthreads atomically, holding both tasklist_lock and sighand->lock.
That means that do_signal_stop() will notice !SIGNAL_STOP_DEQUEUED
after re-locking ->sighand. And it is not possible to any other
thread to re-add SIGNAL_STOP_DEQUEUED later, because dequeue_signal()
can only return SIGKILL.

I think it is better to change do_coredump() to do sigaddset(SIGKILL)
and signal_wake_up() under sighand->lock, but this patch is much
simpler.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>