aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/sched.c
diff options
context:
space:
mode:
authorPeter Zijlstra <a.p.zijlstra@chello.nl>2009-12-16 12:04:37 -0500
committerIngo Molnar <mingo@elte.hu>2009-12-16 13:01:56 -0500
commit3802290628348674985d14914f9bfee7b9084548 (patch)
treeb6b513fa5651f570013f3eff86e843a9d52d1dcb /kernel/sched.c
parente2912009fb7b715728311b0d8fe327a1432b3f79 (diff)
sched: Fix sched_exec() balancing
Since we access ->cpus_allowed without holding rq->lock we need a retry loop to validate the result, this comes for near free when we merge sched_migrate_task() into sched_exec() since that already does the needed check. Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> LKML-Reference: <20091216170517.884743662@chello.nl> Signed-off-by: Ingo Molnar <mingo@elte.hu>
Diffstat (limited to 'kernel/sched.c')
-rw-r--r--kernel/sched.c45
1 files changed, 23 insertions, 22 deletions
diff --git a/kernel/sched.c b/kernel/sched.c
index 33d7965f63f0..63e55ac242d1 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -2322,7 +2322,7 @@ void task_oncpu_function_call(struct task_struct *p,
2322 * 2322 *
2323 * - fork, @p is stable because it isn't on the tasklist yet 2323 * - fork, @p is stable because it isn't on the tasklist yet
2324 * 2324 *
2325 * - exec, @p is unstable XXX 2325 * - exec, @p is unstable, retry loop
2326 * 2326 *
2327 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so 2327 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2328 * we should be good. 2328 * we should be good.
@@ -3132,21 +3132,36 @@ static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3132} 3132}
3133 3133
3134/* 3134/*
3135 * If dest_cpu is allowed for this process, migrate the task to it. 3135 * sched_exec - execve() is a valuable balancing opportunity, because at
3136 * This is accomplished by forcing the cpu_allowed mask to only 3136 * this point the task has the smallest effective memory and cache footprint.
3137 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3138 * the cpu_allowed mask is restored.
3139 */ 3137 */
3140static void sched_migrate_task(struct task_struct *p, int dest_cpu) 3138void sched_exec(void)
3141{ 3139{
3140 struct task_struct *p = current;
3142 struct migration_req req; 3141 struct migration_req req;
3142 int dest_cpu, this_cpu;
3143 unsigned long flags; 3143 unsigned long flags;
3144 struct rq *rq; 3144 struct rq *rq;
3145 3145
3146again:
3147 this_cpu = get_cpu();
3148 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3149 if (dest_cpu == this_cpu) {
3150 put_cpu();
3151 return;
3152 }
3153
3146 rq = task_rq_lock(p, &flags); 3154 rq = task_rq_lock(p, &flags);
3155 put_cpu();
3156
3157 /*
3158 * select_task_rq() can race against ->cpus_allowed
3159 */
3147 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed) 3160 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3148 || unlikely(!cpu_active(dest_cpu))) 3161 || unlikely(!cpu_active(dest_cpu))) {
3149 goto out; 3162 task_rq_unlock(rq, &flags);
3163 goto again;
3164 }
3150 3165
3151 /* force the process onto the specified CPU */ 3166 /* force the process onto the specified CPU */
3152 if (migrate_task(p, dest_cpu, &req)) { 3167 if (migrate_task(p, dest_cpu, &req)) {
@@ -3161,24 +3176,10 @@ static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3161 3176
3162 return; 3177 return;
3163 } 3178 }
3164out:
3165 task_rq_unlock(rq, &flags); 3179 task_rq_unlock(rq, &flags);
3166} 3180}
3167 3181
3168/* 3182/*
3169 * sched_exec - execve() is a valuable balancing opportunity, because at
3170 * this point the task has the smallest effective memory and cache footprint.
3171 */
3172void sched_exec(void)
3173{
3174 int new_cpu, this_cpu = get_cpu();
3175 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3176 put_cpu();
3177 if (new_cpu != this_cpu)
3178 sched_migrate_task(current, new_cpu);
3179}
3180
3181/*
3182 * pull_task - move a task from a remote runqueue to the local runqueue. 3183 * pull_task - move a task from a remote runqueue to the local runqueue.
3183 * Both runqueues must be locked. 3184 * Both runqueues must be locked.
3184 */ 3185 */