sched: Remove rq->clock coupling from set_task_cpu()
Peter Zijlstra [Fri, 27 Nov 2009 13:12:25 +0000 (14:12 +0100)]
set_task_cpu() should be rq invariant and only touch task state, it
currently fails to do so, which opens up a few races, since not all
callers hold both rq->locks.

Remove the relyance on rq->clock, as any site calling set_task_cpu()
should also do a remote clock update, which should ensure the observed
time between these two cpus is monotonic, as per
kernel/sched_clock.c:sched_clock_remote().

Therefore we can simply remove the clock_offset bits and be happy.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/sched.c

index 01fd131..1f9c6d9 100644 (file)
@@ -2060,23 +2060,12 @@ task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 {
        int old_cpu = task_cpu(p);
-       struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
+       struct rq *old_rq = cpu_rq(old_cpu);
        struct cfs_rq *old_cfsrq = task_cfs_rq(p),
                      *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
-       u64 clock_offset;
-
-       clock_offset = old_rq->clock - new_rq->clock;
 
        trace_sched_migrate_task(p, new_cpu);
 
-#ifdef CONFIG_SCHEDSTATS
-       if (p->se.wait_start)
-               p->se.wait_start -= clock_offset;
-       if (p->se.sleep_start)
-               p->se.sleep_start -= clock_offset;
-       if (p->se.block_start)
-               p->se.block_start -= clock_offset;
-#endif
        if (old_cpu != new_cpu) {
                p->se.nr_migrations++;
 #ifdef CONFIG_SCHEDSTATS