workqueue: disable irq while manipulating PENDING
authorTejun Heo <tj@kernel.org>
Fri, 3 Aug 2012 17:30:45 +0000 (10:30 -0700)
committerTejun Heo <tj@kernel.org>
Fri, 3 Aug 2012 17:30:45 +0000 (10:30 -0700)
Queueing operations use WORK_STRUCT_PENDING_BIT to synchronize access
to the target work item.  They first try to claim the bit and proceed
with queueing only after that succeeds and there's a window between
PENDING being set and the actual queueing where the task can be
interrupted or preempted.

There's also a similar window in process_one_work() when clearing
PENDING.  A work item is dequeued, gcwq->lock is released and then
PENDING is cleared and the worker might get interrupted or preempted
between releasing gcwq->lock and clearing PENDING.

cancel[_delayed]_work_sync() tries to claim or steal PENDING.  The
function assumes that a work item with PENDING is either queued or in
the process of being [de]queued.  In the latter case, it busy-loops
until either the work item loses PENDING or is queued.  If canceling
coincides with the above described interrupts or preemptions, the
canceling task will busy-loop while the queueing or executing task is
preempted.

This patch keeps irq disabled across claiming PENDING and actual
queueing and moves PENDING clearing in process_one_work() inside
gcwq->lock so that busy looping from PENDING && !queued doesn't wait
for interrupted/preempted tasks.  Note that, in process_one_work(),
setting last CPU and clearing PENDING got merged into single
operation.

This removes possible long busy-loops and will allow using
try_to_grab_pending() from bh and irq contexts.

v2: __queue_work() was testing preempt_count() to ensure that the
    caller has disabled preemption.  This triggers spuriously if
    !CONFIG_PREEMPT_COUNT.  Use preemptible() instead.  Reported by
    Fengguang Wu.

v3: Disable irq instead of preemption.  IRQ will be disabled while
    grabbing gcwq->lock later anyway and this allows using
    try_to_grab_pending() from bh and irq contexts.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Fengguang Wu <fengguang.wu@intel.com>
kernel/workqueue.c

index 5c26d36..30474c4 100644 (file)
@@ -537,9 +537,10 @@ static int work_next_color(int color)
  * work is on queue.  Once execution starts, WORK_STRUCT_CWQ is
  * cleared and the work data contains the cpu number it was last on.
  *
- * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
- * cwq, cpu or clear work->data.  These functions should only be
- * called while the work is owned - ie. while the PENDING bit is set.
+ * set_work_cwq(), set_work_cpu_and_clear_pending() and clear_work_data()
+ * can be used to set the cwq, cpu or clear work->data.  These functions
+ * should only be called while the work is owned - ie. while the PENDING
+ * bit is set.
  *
  * get_work_[g]cwq() can be used to obtain the gcwq or cwq
  * corresponding to a work.  gcwq is available once the work has been
@@ -561,9 +562,10 @@ static void set_work_cwq(struct work_struct *work,
                      WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
 }
 
-static void set_work_cpu(struct work_struct *work, unsigned int cpu)
+static void set_work_cpu_and_clear_pending(struct work_struct *work,
+                                          unsigned int cpu)
 {
-       set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
+       set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, 0);
 }
 
 static void clear_work_data(struct work_struct *work)
@@ -981,7 +983,14 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
        struct cpu_workqueue_struct *cwq;
        struct list_head *worklist;
        unsigned int work_flags;
-       unsigned long flags;
+
+       /*
+        * While a work item is PENDING && off queue, a task trying to
+        * steal the PENDING will busy-loop waiting for it to either get
+        * queued or lose PENDING.  Grabbing PENDING and queueing should
+        * happen with IRQ disabled.
+        */
+       WARN_ON_ONCE(!irqs_disabled());
 
        debug_work_activate(work);
 
@@ -1008,7 +1017,7 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
                    (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
                        struct worker *worker;
 
-                       spin_lock_irqsave(&last_gcwq->lock, flags);
+                       spin_lock(&last_gcwq->lock);
 
                        worker = find_worker_executing_work(last_gcwq, work);
 
@@ -1016,14 +1025,15 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
                                gcwq = last_gcwq;
                        else {
                                /* meh... not running there, queue here */
-                               spin_unlock_irqrestore(&last_gcwq->lock, flags);
-                               spin_lock_irqsave(&gcwq->lock, flags);
+                               spin_unlock(&last_gcwq->lock);
+                               spin_lock(&gcwq->lock);
                        }
-               } else
-                       spin_lock_irqsave(&gcwq->lock, flags);
+               } else {
+                       spin_lock(&gcwq->lock);
+               }
        } else {
                gcwq = get_gcwq(WORK_CPU_UNBOUND);
-               spin_lock_irqsave(&gcwq->lock, flags);
+               spin_lock(&gcwq->lock);
        }
 
        /* gcwq determined, get cwq and queue */
@@ -1031,7 +1041,7 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
        trace_workqueue_queue_work(cpu, cwq, work);
 
        if (WARN_ON(!list_empty(&work->entry))) {
-               spin_unlock_irqrestore(&gcwq->lock, flags);
+               spin_unlock(&gcwq->lock);
                return;
        }
 
@@ -1049,7 +1059,7 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
 
        insert_work(cwq, work, worklist, work_flags);
 
-       spin_unlock_irqrestore(&gcwq->lock, flags);
+       spin_unlock(&gcwq->lock);
 }
 
 /**
@@ -1067,11 +1077,16 @@ bool queue_work_on(int cpu, struct workqueue_struct *wq,
                   struct work_struct *work)
 {
        bool ret = false;
+       unsigned long flags;
+
+       local_irq_save(flags);
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
                __queue_work(cpu, wq, work);
                ret = true;
        }
+
+       local_irq_restore(flags);
        return ret;
 }
 EXPORT_SYMBOL_GPL(queue_work_on);
@@ -1102,7 +1117,9 @@ static void delayed_work_timer_fn(unsigned long __data)
        struct delayed_work *dwork = (struct delayed_work *)__data;
        struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
 
+       local_irq_disable();
        __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
+       local_irq_enable();
 }
 
 /**
@@ -1120,6 +1137,10 @@ bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
        struct timer_list *timer = &dwork->timer;
        struct work_struct *work = &dwork->work;
        bool ret = false;
+       unsigned long flags;
+
+       /* read the comment in __queue_work() */
+       local_irq_save(flags);
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
                unsigned int lcpu;
@@ -1156,6 +1177,8 @@ bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
                        add_timer(timer);
                ret = true;
        }
+
+       local_irq_restore(flags);
        return ret;
 }
 EXPORT_SYMBOL_GPL(queue_delayed_work_on);
@@ -1970,15 +1993,13 @@ __acquires(&gcwq->lock)
                return;
        }
 
-       /* claim and process */
+       /* claim and dequeue */
        debug_work_deactivate(work);
        hlist_add_head(&worker->hentry, bwh);
        worker->current_work = work;
        worker->current_cwq = cwq;
        work_color = get_work_color(work);
 
-       /* record the current cpu number in the work data and dequeue */
-       set_work_cpu(work, gcwq->cpu);
        list_del_init(&work->entry);
 
        /*
@@ -1995,10 +2016,18 @@ __acquires(&gcwq->lock)
        if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
                wake_up_worker(pool);
 
-       spin_unlock_irq(&gcwq->lock);
+       /*
+        * Record the last CPU and clear PENDING.  The following wmb is
+        * paired with the implied mb in test_and_set_bit(PENDING) and
+        * ensures all updates to @work made here are visible to and
+        * precede any updates by the next PENDING owner.  Also, clear
+        * PENDING inside @gcwq->lock so that PENDING and queued state
+        * changes happen together while IRQ is disabled.
+        */
+       smp_wmb();
+       set_work_cpu_and_clear_pending(work, gcwq->cpu);
 
-       smp_wmb();      /* paired with test_and_set_bit(PENDING) */
-       work_clear_pending(work);
+       spin_unlock_irq(&gcwq->lock);
 
        lock_map_acquire_read(&cwq->wq->lockdep_map);
        lock_map_acquire(&lockdep_map);
@@ -2836,9 +2865,11 @@ EXPORT_SYMBOL_GPL(cancel_work_sync);
  */
 bool flush_delayed_work(struct delayed_work *dwork)
 {
+       local_irq_disable();
        if (del_timer_sync(&dwork->timer))
                __queue_work(raw_smp_processor_id(),
                             get_work_cwq(&dwork->work)->wq, &dwork->work);
+       local_irq_enable();
        return flush_work(&dwork->work);
 }
 EXPORT_SYMBOL(flush_delayed_work);
@@ -2857,9 +2888,11 @@ EXPORT_SYMBOL(flush_delayed_work);
  */
 bool flush_delayed_work_sync(struct delayed_work *dwork)
 {
+       local_irq_disable();
        if (del_timer_sync(&dwork->timer))
                __queue_work(raw_smp_processor_id(),
                             get_work_cwq(&dwork->work)->wq, &dwork->work);
+       local_irq_enable();
        return flush_work_sync(&dwork->work);
 }
 EXPORT_SYMBOL(flush_delayed_work_sync);