[PATCH] Fix cpu timers exit deadlock and races
Roland McGrath [Thu, 20 Oct 2005 05:21:23 +0000 (22:21 -0700)]
Oleg Nesterov reported an SMP deadlock.  If there is a running timer
tracking a different process's CPU time clock when the process owning
the timer exits, we deadlock on tasklist_lock in posix_cpu_timer_del via
exit_itimers.

That code was using tasklist_lock to check for a race with __exit_signal
being called on the timer-target task and clearing its ->signal.
However, there is actually no such race.  __exit_signal will have called
posix_cpu_timers_exit and posix_cpu_timers_exit_group before it does
that.  Those will clear those k_itimer's association with the dying
task, so posix_cpu_timer_del will return early and never reach the code
in question.

In addition, posix_cpu_timer_del called from exit_itimers during execve
or directly from timer_delete in the process owning the timer can race
with an exiting timer-target task to cause a double put on timer-target
task struct.  Make sure we always access cpu_timers lists with sighand
lock held.

Signed-off-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Chris Wright <chrisw@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

kernel/posix-cpu-timers.c

index 7a51a55..b3f3edc 100644 (file)
@@ -387,25 +387,19 @@ int posix_cpu_timer_del(struct k_itimer *timer)
        if (unlikely(p == NULL))
                return 0;
 
+       spin_lock(&p->sighand->siglock);
        if (!list_empty(&timer->it.cpu.entry)) {
-               read_lock(&tasklist_lock);
-               if (unlikely(p->signal == NULL)) {
-                       /*
-                        * We raced with the reaping of the task.
-                        * The deletion should have cleared us off the list.
-                        */
-                       BUG_ON(!list_empty(&timer->it.cpu.entry));
-               } else {
-                       /*
-                        * Take us off the task's timer list.
-                        */
-                       spin_lock(&p->sighand->siglock);
-                       list_del(&timer->it.cpu.entry);
-                       spin_unlock(&p->sighand->siglock);
-               }
-               read_unlock(&tasklist_lock);
+               /*
+                * Take us off the task's timer list.  We don't need to
+                * take tasklist_lock and check for the task being reaped.
+                * If it was reaped, it already called posix_cpu_timers_exit
+                * and posix_cpu_timers_exit_group to clear all the timers
+                * that pointed to it.
+                */
+               list_del(&timer->it.cpu.entry);
+               put_task_struct(p);
        }
-       put_task_struct(p);
+       spin_unlock(&p->sighand->siglock);
 
        return 0;
 }