sched: fix wait_for_completion_timeout() spurious failure under heavy load
authorRoland Dreier <rdreier@cisco.com>
Thu, 19 Jun 2008 22:04:07 +0000 (15:04 -0700)
committerIngo Molnar <mingo@elte.hu>
Fri, 20 Jun 2008 11:19:32 +0000 (13:19 +0200)
It seems that the current implementaton of wait_for_completion_timeout()
has a small problem under very high load for the common pattern:

if (!wait_for_completion_timeout(&done, timeout))
/* handle failure */

because the implementation very roughly does (lots of code deleted to
show the basic flow):

static inline long __sched
do_wait_for_common(struct completion *x, long timeout, int state)
{
if (x->done)
return timeout;

do {
timeout = schedule_timeout(timeout);

if (!timeout)
return timeout;

} while (!x->done);

return timeout;
}

so if the system is very busy and x->done is not set when
do_wait_for_common() is entered, it is possible that the first call to
schedule_timeout() returns 0 because the task doing wait_for_completion
doesn't get rescheduled for a long time, even if it is woken up early
enough.

In this case, wait_for_completion_timeout() returns 0 without even
checking x->done again, and the code above falls into its failure case
purely for scheduler reasons, even if the hardware event or whatever was
being waited for happened early enough.

It would make sense to add an extra test to do_wait_for() in the timeout
case and return 1 if x->done is actually set.

A quick audit (not exhaustive) of wait_for_completion_timeout() callers
seems to indicate that no one actually cares about the return value in
the success case -- they just test for 0 (timed out) versus non-zero
(wait succeeded).

Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched.c

index 4a3cb0614158b5e99fb24ccbab906d641e8d56d5..577f160131bd5922da713fb69415fbc88c5aa71d 100644 (file)
@@ -4405,6 +4405,16 @@ do_wait_for_common(struct completion *x, long timeout, int state)
                        spin_unlock_irq(&x->wait.lock);
                        timeout = schedule_timeout(timeout);
                        spin_lock_irq(&x->wait.lock);
+
+                       /*
+                        * If the completion has arrived meanwhile
+                        * then return 1 jiffy time left:
+                        */
+                       if (x->done && !timeout) {
+                               timeout = 1;
+                               break;
+                       }
+
                        if (!timeout) {
                                __remove_wait_queue(&x->wait, &wait);
                                return timeout;