workqueue: clear POOL_DISASSOCIATED in rebind_workers()
authorLai Jiangshan <laijs@cn.fujitsu.com>
Tue, 3 Jun 2014 07:33:27 +0000 (15:33 +0800)
committerTejun Heo <tj@kernel.org>
Thu, 19 Jun 2014 19:43:33 +0000 (15:43 -0400)
The commit a9ab775bcadf ("workqueue: directly restore CPU affinity of
workers from CPU_ONLINE") moved the pool->lock into rebind_workers()
without also moving "pool->flags &= ~POOL_DISASSOCIATED".

There is nothing wrong with "pool->flags &= ~POOL_DISASSOCIATED" not
being moved together, but there isn't any benefit either. We move it
into rebind_workers() and achieve these benefits:

1) Better readability.  POOL_DISASSOCIATED is cleared in
   rebind_workers() as expected.

2) When POOL_DISASSOCIATED is cleared, we can ensure that all the
   running workers of the pool are on the local CPU (pool->cpu).

tj: Cosmetic updates to the code and description.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kernel/workqueue.c

index 6203d29008772e9ee647eae939e9b4d070a841e6..129e4c3452ee362ef21a19363251eba981c90ba1 100644 (file)
@@ -4542,6 +4542,7 @@ static void rebind_workers(struct worker_pool *pool)
                                                  pool->attrs->cpumask) < 0);
 
        spin_lock_irq(&pool->lock);
+       pool->flags &= ~POOL_DISASSOCIATED;
 
        for_each_pool_worker(worker, pool) {
                unsigned int worker_flags = worker->flags;
@@ -4643,15 +4644,10 @@ static int workqueue_cpu_up_callback(struct notifier_block *nfb,
                for_each_pool(pool, pi) {
                        mutex_lock(&pool->attach_mutex);
 
-                       if (pool->cpu == cpu) {
-                               spin_lock_irq(&pool->lock);
-                               pool->flags &= ~POOL_DISASSOCIATED;
-                               spin_unlock_irq(&pool->lock);
-
+                       if (pool->cpu == cpu)
                                rebind_workers(pool);
-                       } else if (pool->cpu < 0) {
+                       else if (pool->cpu < 0)
                                restore_unbound_workers_cpumask(pool, cpu);
-                       }
 
                        mutex_unlock(&pool->attach_mutex);
                }