Skip to contentSkip to conversation

sched: CPU hotplug race vs. set_cpus_allowed()

PATCH In review Kirill Korotaev 4 participants 4 messages Jun 23, 2006 – Jun 24, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 sched: CPU hotplug race vs. set_cpus_allowed()

Looks like there is a race between set_cpus_allowed() and move_task_off_dead_cpu().
__migrate_task() doesn't report any err code, so task can be left on its runqueue if its cpus_allowed mask changed so that dest_cpu is not longer a possible target. Also, chaning cpus_allowed mask requires rq->lock being held.

Kirill
P.S. against 2.6.17-mm1

kernel/sched.c +14 -5
@@ -4858,12 +4858,13 @@ EXPORT_SYMBOL_GPL(set_cpus_allowed);
  * So we race with normal scheduler movements, but that's OK, as long
  * as the task is no longer on this CPU.
  */
-static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
+static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
 {
 	runqueue_t *rq_dest, *rq_src;
+	int res = 0;
 
 	if (unlikely(cpu_is_offline(dest_cpu)))
-		return;
+		return 0;
 
 	rq_src = cpu_rq(src_cpu);
 	rq_dest = cpu_rq(dest_cpu);
@@ -4891,9 +4892,10 @@ static void __migrate_task(struct task_s
 		if (TASK_PREEMPTS_CURR(p, rq_dest))
 			resched_task(rq_dest->curr);
 	}
-
+	res = 1;
 out:
 	double_rq_unlock(rq_src, rq_dest);
+	return res;
 }
 
 /*
@@ -4964,10 +4966,13 @@ int sigstop_on_cpu_lost;
 /* Figure out where task on dead CPU should go, use force if neccessary. */
 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk)
 {
-	int dest_cpu;
+	runqueue_t *rq;
+	unsigned long flags;
+	int dest_cpu, res;
 	cpumask_t mask;
 	int force = 0;
 
+restart:
 	/* On same node? */
 	mask = node_to_cpumask(cpu_to_node(dead_cpu));
 	cpus_and(mask, mask, tsk->cpus_allowed);
@@ -4979,8 +4984,10 @@ static void move_task_off_dead_cpu(int d
 
 	/* No more Mr. Nice Guy. */
 	if (dest_cpu == NR_CPUS) {
+		rq = task_rq_lock(tsk, &flags);
 		cpus_setall(tsk->cpus_allowed);
 		dest_cpu = any_online_cpu(tsk->cpus_allowed);
+		task_rq_unlock(rq, &flags);
 
 		/*
 		 * Don't tell them about moving exiting tasks or
@@ -5000,7 +5007,9 @@ static void move_task_off_dead_cpu(int d
 		if (tsk->mm && sigstop_on_cpu_lost)
 			force = 1;
 	}
-	__migrate_task(tsk, dead_cpu, dest_cpu);
+	res = __migrate_task(tsk, dead_cpu, dest_cpu);
+	if (!res)
+		goto restart;
 
 	if (force)
 		force_sig_specific(SIGSTOP, tsk);
Acked-by on v1 1/1
#1 lore raw
Quoted 7 lines · Kirill Korotaev

Looks like there is a race between set_cpus_allowed() and move_task_off_dead_cpu(). __migrate_task() doesn't report any err code, so task can be left on its runqueue if its cpus_allowed mask changed so that dest_cpu is not longer a possible target. Also, chaning cpus_allowed mask requires rq->lock being held.

good one!

Ingo

#2 lore raw
In reply to Kirill Korotaev
Quoted 11 lines · Kirill Korotaev

Looks like there is a race between set_cpus_allowed() and move_task_off_dead_cpu().
__migrate_task() doesn't report any err code, so task can be left on its runqueue if its cpus_allowed mask changed so that dest_cpu is not longer a possible target. Also, chaning cpus_allowed mask requires rq->lock being held.

Kirill
P.S. against 2.6.17-mm1

Hi!

Since you've got

-static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) +static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)

 {
        runqueue_t *rq_dest, *rq_src;

+ int res = 0;

if (unlikely(cpu_is_offline(dest_cpu)))
- return;
+ return 0;

why not return res here?

oh and ret is a more commonly used name than res (your choice of course).

and an addition to the comment such as "returns non-zero only when it fails to migrate the task" would be nice.

thanks!

--
-ck

Signature · 1 line
-ck
#3 lore raw
In reply to Kirill Korotaev
Quoted 7 lines · On Fri, 23 Jun 2006 12:16:09 +0400 Kirill Korotaev <[email protected]> wrote:

Looks like there is a race between set_cpus_allowed() and move_task_off_dead_cpu().
__migrate_task() doesn't report any err code, so task can be left on its runqueue if its cpus_allowed mask changed so that dest_cpu is not longer a possible target. Also, chaning cpus_allowed mask requires rq->lock being held.

Quoted 5 lines

Kirill
P.S. against 2.6.17-mm1

That's not against 2.6.17-mm1.

kernel/sched.c int sigstop_on_cpu_lost; · line 4973

Quoted 9 lines

static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *tsk) {

kernel/sched.c +3 -1
-	int dest_cpu;
+	runqueue_t *rq;
+	unsigned long flags;
+	int dest_cpu, res;
 	cpumask_t mask;
 	int force = 0;

Your kernel has extra goodies: the `force' stuff.

Please check that the patch which I merged is correct.