sched: prevent divide by zero error in cpu_avg_load_per_task
authorSteven Rostedt <rostedt@goodmis.org>
Thu, 27 Nov 2008 02:04:24 +0000 (21:04 -0500)
committerIngo Molnar <mingo@elte.hu>
Thu, 27 Nov 2008 09:29:52 +0000 (10:29 +0100)
Impact: fix divide by zero crash in scheduler rebalance irq

While testing the branch profiler, I hit this crash:

divide error: 0000 [#1] PREEMPT SMP
[...]
RIP: 0010:[<ffffffff8024a008>]  [<ffffffff8024a008>] cpu_avg_load_per_task+0x50/0x7f
[...]
Call Trace:
 <IRQ> <0> [<ffffffff8024fd43>] find_busiest_group+0x3e5/0xcaa
 [<ffffffff8025da75>] rebalance_domains+0x2da/0xa21
 [<ffffffff80478769>] ? find_next_bit+0x1b2/0x1e6
 [<ffffffff8025e2ce>] run_rebalance_domains+0x112/0x19f
 [<ffffffff8026d7c2>] __do_softirq+0xa8/0x232
 [<ffffffff8020ea7c>] call_softirq+0x1c/0x3e
 [<ffffffff8021047a>] do_softirq+0x94/0x1cd
 [<ffffffff8026d5eb>] irq_exit+0x6b/0x10e
 [<ffffffff8022e6ec>] smp_apic_timer_interrupt+0xd3/0xff
 [<ffffffff8020e4b3>] apic_timer_interrupt+0x13/0x20

The code for cpu_avg_load_per_task has:

if (rq->nr_running)
rq->avg_load_per_task = rq->load.weight / rq->nr_running;

The runqueue lock is not held here, and there is nothing that prevents
the rq->nr_running from going to zero after it passes the if condition.

The branch profiler simply made the race window bigger.

This patch saves off the rq->nr_running to a local variable and uses that
for both the condition and the division.

Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched.c

index 9b1e79371c207b37c1617d3f7c0460709a3cc39b..700aa9a1413fc783028170bcfebc014b17a38120 100644 (file)
@@ -1453,9 +1453,10 @@ static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
 static unsigned long cpu_avg_load_per_task(int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
+       unsigned long nr_running = rq->nr_running;
 
-       if (rq->nr_running)
-               rq->avg_load_per_task = rq->load.weight / rq->nr_running;
+       if (nr_running)
+               rq->avg_load_per_task = rq->load.weight / nr_running;
        else
                rq->avg_load_per_task = 0;