reimplement flush_workqueue()
authorOleg Nesterov <oleg@tv-sign.ru>
Wed, 9 May 2007 09:33:51 +0000 (02:33 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 9 May 2007 19:30:50 +0000 (12:30 -0700)
Remove ->remove_sequence, ->insert_sequence, and ->work_done from struct
cpu_workqueue_struct.  To implement flush_workqueue() we can queue a
barrier work on each CPU and wait for its completition.

The barrier is queued under workqueue_mutex to ensure that per cpu
wq->cpu_wq is alive, we drop this mutex before going to sleep.  If CPU goes
down while we are waiting for completition, take_over_work() will move the
barrier on another CPU, and the handler will wake up us eventually.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kernel/workqueue.c

index 1ea4bcb..b7bb37a 100644 (file)
 /*
  * The per-CPU workqueue (if single thread, we always use the first
  * possible cpu).
- *
- * The sequence counters are for flush_scheduled_work().  It wants to wait
- * until all currently-scheduled works are completed, but it doesn't
- * want to be livelocked by new, incoming ones.  So it waits until
- * remove_sequence is >= the insert_sequence which pertained when
- * flush_scheduled_work() was called.
  */
 struct cpu_workqueue_struct {
 
        spinlock_t lock;
 
-       long remove_sequence;   /* Least-recently added (next to run) */
-       long insert_sequence;   /* Next to add */
-
        struct list_head worklist;
        wait_queue_head_t more_work;
-       wait_queue_head_t work_done;
 
        struct workqueue_struct *wq;
        struct task_struct *thread;
@@ -138,8 +128,6 @@ static int __run_work(struct cpu_workqueue_struct *cwq, struct work_struct *work
                f(work);
 
                spin_lock_irqsave(&cwq->lock, flags);
-               cwq->remove_sequence++;
-               wake_up(&cwq->work_done);
                ret = 1;
        }
        spin_unlock_irqrestore(&cwq->lock, flags);
@@ -187,7 +175,6 @@ static void __queue_work(struct cpu_workqueue_struct *cwq,
        spin_lock_irqsave(&cwq->lock, flags);
        set_wq_data(work, cwq);
        list_add_tail(&work->entry, &cwq->worklist);
-       cwq->insert_sequence++;
        wake_up(&cwq->more_work);
        spin_unlock_irqrestore(&cwq->lock, flags);
 }
@@ -338,8 +325,6 @@ static void run_workqueue(struct cpu_workqueue_struct *cwq)
                }
 
                spin_lock_irqsave(&cwq->lock, flags);
-               cwq->remove_sequence++;
-               wake_up(&cwq->work_done);
        }
        cwq->run_depth--;
        spin_unlock_irqrestore(&cwq->lock, flags);
@@ -394,6 +379,25 @@ static int worker_thread(void *__cwq)
        return 0;
 }
 
+struct wq_barrier {
+       struct work_struct      work;
+       struct completion       done;
+};
+
+static void wq_barrier_func(struct work_struct *work)
+{
+       struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
+       complete(&barr->done);
+}
+
+static inline void init_wq_barrier(struct wq_barrier *barr)
+{
+       INIT_WORK(&barr->work, wq_barrier_func);
+       __set_bit(WORK_STRUCT_PENDING, work_data_bits(&barr->work));
+
+       init_completion(&barr->done);
+}
+
 static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
 {
        if (cwq->thread == current) {
@@ -401,23 +405,18 @@ static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
                 * Probably keventd trying to flush its own queue. So simply run
                 * it by hand rather than deadlocking.
                 */
+               mutex_unlock(&workqueue_mutex);
                run_workqueue(cwq);
+               mutex_lock(&workqueue_mutex);
        } else {
-               DEFINE_WAIT(wait);
-               long sequence_needed;
+               struct wq_barrier barr;
 
-               spin_lock_irq(&cwq->lock);
-               sequence_needed = cwq->insert_sequence;
+               init_wq_barrier(&barr);
+               __queue_work(cwq, &barr.work);
 
-               while (sequence_needed - cwq->remove_sequence > 0) {
-                       prepare_to_wait(&cwq->work_done, &wait,
-                                       TASK_UNINTERRUPTIBLE);
-                       spin_unlock_irq(&cwq->lock);
-                       schedule();
-                       spin_lock_irq(&cwq->lock);
-               }
-               finish_wait(&cwq->work_done, &wait);
-               spin_unlock_irq(&cwq->lock);
+               mutex_unlock(&workqueue_mutex);
+               wait_for_completion(&barr.done);
+               mutex_lock(&workqueue_mutex);
        }
 }
 
@@ -428,29 +427,25 @@ static void flush_cpu_workqueue(struct cpu_workqueue_struct *cwq)
  * Forces execution of the workqueue and blocks until its completion.
  * This is typically used in driver shutdown handlers.
  *
- * This function will sample each workqueue's current insert_sequence number and
- * will sleep until the head sequence is greater than or equal to that.  This
- * means that we sleep until all works which were queued on entry have been
- * handled, but we are not livelocked by new incoming ones.
+ * We sleep until all works which were queued on entry have been handled,
+ * but we are not livelocked by new incoming ones.
  *
  * This function used to run the workqueues itself.  Now we just wait for the
  * helper threads to do it.
  */
 void fastcall flush_workqueue(struct workqueue_struct *wq)
 {
-       might_sleep();
-
+       mutex_lock(&workqueue_mutex);
        if (is_single_threaded(wq)) {
                /* Always use first cpu's area. */
                flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
        } else {
                int cpu;
 
-               mutex_lock(&workqueue_mutex);
                for_each_online_cpu(cpu)
                        flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));
-               mutex_unlock(&workqueue_mutex);
        }
+       mutex_unlock(&workqueue_mutex);
 }
 EXPORT_SYMBOL_GPL(flush_workqueue);
 
@@ -463,12 +458,9 @@ static struct task_struct *create_workqueue_thread(struct workqueue_struct *wq,
        spin_lock_init(&cwq->lock);
        cwq->wq = wq;
        cwq->thread = NULL;
-       cwq->insert_sequence = 0;
-       cwq->remove_sequence = 0;
        cwq->freezeable = freezeable;
        INIT_LIST_HEAD(&cwq->worklist);
        init_waitqueue_head(&cwq->more_work);
-       init_waitqueue_head(&cwq->work_done);
 
        if (is_single_threaded(wq))
                p = kthread_create(worker_thread, cwq, "%s", wq->name);