workqueue: protect wq->pwqs and iteration with wq->mutex
authorLai Jiangshan <laijs@cn.fujitsu.com>
Mon, 25 Mar 2013 23:57:18 +0000 (16:57 -0700)
committerTejun Heo <tj@kernel.org>
Mon, 25 Mar 2013 23:57:18 +0000 (16:57 -0700)
We're expanding wq->mutex to cover all fields specific to each
workqueue with the end goal of replacing pwq_lock which will make
locking simpler and easier to understand.

init_and_link_pwq() and pwq_unbound_release_workfn() already grab
wq->mutex when adding or removing a pwq from wq->pwqs list.  This
patch makes it official that the list is wq->mutex protected for
writes and updates readers accoridingly.  Explicit IRQ toggles for
sched-RCU read-locking in flush_workqueue_prep_pwqs() and
drain_workqueues() are removed as the surrounding wq->mutex can
provide sufficient synchronization.

Also, assert_rcu_or_pwq_lock() is renamed to assert_rcu_or_wq_mutex()
and checks for wq->mutex too.

pwq_lock locking and assertion are not removed by this patch and a
couple of for_each_pwq() iterations are still protected by it.
They'll be removed by future patches.

tj: Rebased on top of the current dev branch.  Updated description.
    Folded in assert_rcu_or_wq_mutex() renaming from a later patch
    along with associated comment updates.

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

index 3ac2c4d856074e8f811533db34b383d530210946..9c32fd171d5c7e3a31309be5dc363c3fc503d4a2 100644 (file)
@@ -204,7 +204,7 @@ struct pool_workqueue {
         * Release of unbound pwq is punted to system_wq.  See put_pwq()
         * and pwq_unbound_release_workfn() for details.  pool_workqueue
         * itself is also sched-RCU protected so that the first pwq can be
-        * determined without grabbing pwq_lock.
+        * determined without grabbing wq->mutex.
         */
        struct work_struct      unbound_release_work;
        struct rcu_head         rcu;
@@ -298,10 +298,11 @@ static void copy_workqueue_attrs(struct workqueue_attrs *to,
                           lockdep_is_held(&wq_pool_mutex),             \
                           "sched RCU or wq_pool_mutex should be held")
 
-#define assert_rcu_or_pwq_lock()                                       \
+#define assert_rcu_or_wq_mutex(wq)                                     \
        rcu_lockdep_assert(rcu_read_lock_sched_held() ||                \
+                          lockdep_is_held(&wq->mutex) ||               \
                           lockdep_is_held(&pwq_lock),                  \
-                          "sched RCU or pwq_lock should be held")
+                          "sched RCU or wq->mutex should be held")
 
 #ifdef CONFIG_LOCKDEP
 #define assert_manager_or_pool_lock(pool)                              \
@@ -356,7 +357,7 @@ static void copy_workqueue_attrs(struct workqueue_attrs *to,
  * @pwq: iteration cursor
  * @wq: the target workqueue
  *
- * This must be called either with pwq_lock held or sched RCU read locked.
+ * This must be called either with wq->mutex held or sched RCU read locked.
  * If the pwq needs to be used beyond the locking in effect, the caller is
  * responsible for guaranteeing that the pwq stays online.
  *
@@ -365,7 +366,7 @@ static void copy_workqueue_attrs(struct workqueue_attrs *to,
  */
 #define for_each_pwq(pwq, wq)                                          \
        list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)          \
-               if (({ assert_rcu_or_pwq_lock(); false; })) { }         \
+               if (({ assert_rcu_or_wq_mutex(wq); false; })) { }       \
                else
 
 #ifdef CONFIG_DEBUG_OBJECTS_WORK
@@ -504,13 +505,13 @@ static int worker_pool_assign_id(struct worker_pool *pool)
  * first_pwq - return the first pool_workqueue of the specified workqueue
  * @wq: the target workqueue
  *
- * This must be called either with pwq_lock held or sched RCU read locked.
+ * This must be called either with wq->mutex held or sched RCU read locked.
  * If the pwq needs to be used beyond the locking in effect, the caller is
  * responsible for guaranteeing that the pwq stays online.
  */
 static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
 {
-       assert_rcu_or_pwq_lock();
+       assert_rcu_or_wq_mutex(wq);
        return list_first_or_null_rcu(&wq->pwqs, struct pool_workqueue,
                                      pwqs_node);
 }
@@ -2477,12 +2478,10 @@ static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
                atomic_set(&wq->nr_pwqs_to_flush, 1);
        }
 
-       local_irq_disable();
-
        for_each_pwq(pwq, wq) {
                struct worker_pool *pool = pwq->pool;
 
-               spin_lock(&pool->lock);
+               spin_lock_irq(&pool->lock);
 
                if (flush_color >= 0) {
                        WARN_ON_ONCE(pwq->flush_color != -1);
@@ -2499,11 +2498,9 @@ static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
                        pwq->work_color = work_color;
                }
 
-               spin_unlock(&pool->lock);
+               spin_unlock_irq(&pool->lock);
        }
 
-       local_irq_enable();
-
        if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
                complete(&wq->first_flusher->done);
 
@@ -2691,14 +2688,14 @@ void drain_workqueue(struct workqueue_struct *wq)
 reflush:
        flush_workqueue(wq);
 
-       local_irq_disable();
+       mutex_lock(&wq->mutex);
 
        for_each_pwq(pwq, wq) {
                bool drained;
 
-               spin_lock(&pwq->pool->lock);
+               spin_lock_irq(&pwq->pool->lock);
                drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
-               spin_unlock(&pwq->pool->lock);
+               spin_unlock_irq(&pwq->pool->lock);
 
                if (drained)
                        continue;
@@ -2708,13 +2705,10 @@ reflush:
                        pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
                                wq->name, flush_cnt);
 
-               local_irq_enable();
+               mutex_unlock(&wq->mutex);
                goto reflush;
        }
 
-       local_irq_enable();
-
-       mutex_lock(&wq->mutex);
        if (!--wq->nr_drainers)
                wq->flags &= ~__WQ_DRAINING;
        mutex_unlock(&wq->mutex);
@@ -3843,13 +3837,13 @@ void destroy_workqueue(struct workqueue_struct *wq)
        drain_workqueue(wq);
 
        /* sanity checks */
-       spin_lock_irq(&pwq_lock);
+       mutex_lock(&wq->mutex);
        for_each_pwq(pwq, wq) {
                int i;
 
                for (i = 0; i < WORK_NR_COLORS; i++) {
                        if (WARN_ON(pwq->nr_in_flight[i])) {
-                               spin_unlock_irq(&pwq_lock);
+                               mutex_unlock(&wq->mutex);
                                return;
                        }
                }
@@ -3857,11 +3851,11 @@ void destroy_workqueue(struct workqueue_struct *wq)
                if (WARN_ON(pwq->refcnt > 1) ||
                    WARN_ON(pwq->nr_active) ||
                    WARN_ON(!list_empty(&pwq->delayed_works))) {
-                       spin_unlock_irq(&pwq_lock);
+                       mutex_unlock(&wq->mutex);
                        return;
                }
        }
-       spin_unlock_irq(&pwq_lock);
+       mutex_unlock(&wq->mutex);
 
        /*
         * wq list is used to freeze wq, remove from list after