[PATCH] posix-timers: fix posix_cpu_timer_set() vs run_posix_cpu_timers() race
Oleg Nesterov [Mon, 24 Oct 2005 14:29:58 +0000 (18:29 +0400)]
This might be harmless, but looks like a race from code inspection (I
was unable to trigger it).  I must admit, I don't understand why we
can't return TIMER_RETRY after 'spin_unlock(&p->sighand->siglock)'
without doing bump_cpu_timer(), but this is what original code does.

posix_cpu_timer_set:

read_lock(&tasklist_lock);

spin_lock(&p->sighand->siglock);
list_del_init(&timer->it.cpu.entry);
spin_unlock(&p->sighand->siglock);

We are probaly deleting the timer from run_posix_cpu_timers's 'firing'
local list_head while run_posix_cpu_timers() does list_for_each_safe.

Various bad things can happen, for example we can just delete this timer
so that list_for_each() will not notice it and run_posix_cpu_timers()
will not reset '->firing' flag. In that case,

....

if (timer->it.cpu.firing) {
read_unlock(&tasklist_lock);
timer->it.cpu.firing = -1;
return TIMER_RETRY;
}

sys_timer_settime() goes to 'retry:', calls posix_cpu_timer_set() again,
it returns TIMER_RETRY ...

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

kernel/posix-cpu-timers.c

index 92a0380..b15462b 100644 (file)
@@ -730,9 +730,15 @@ int posix_cpu_timer_set(struct k_itimer *timer, int flags,
         * Disarm any old timer after extracting its expiry time.
         */
        BUG_ON(!irqs_disabled());
+
+       ret = 0;
        spin_lock(&p->sighand->siglock);
        old_expires = timer->it.cpu.expires;
-       list_del_init(&timer->it.cpu.entry);
+       if (unlikely(timer->it.cpu.firing)) {
+               timer->it.cpu.firing = -1;
+               ret = TIMER_RETRY;
+       } else
+               list_del_init(&timer->it.cpu.entry);
        spin_unlock(&p->sighand->siglock);
 
        /*
@@ -780,7 +786,7 @@ int posix_cpu_timer_set(struct k_itimer *timer, int flags,
                }
        }
 
-       if (unlikely(timer->it.cpu.firing)) {
+       if (unlikely(ret)) {
                /*
                 * We are colliding with the timer actually firing.
                 * Punt after filling in the timer's old value, and
@@ -788,8 +794,6 @@ int posix_cpu_timer_set(struct k_itimer *timer, int flags,
                 * it as an overrun (thanks to bump_cpu_timer above).
                 */
                read_unlock(&tasklist_lock);
-               timer->it.cpu.firing = -1;
-               ret = TIMER_RETRY;
                goto out;
        }