workqueue: relax lockdep annotation on flush_work()

Currently, the lockdep annotation in flush_work() requires exclusive access on the workqueue the target work is queued on and triggers warning if a work is trying to flush another work on the same workqueue; however, this is no longer true as workqueues can now execute multiple works concurrently.

This patch adds lock_map_acquire_read() and make process_one_work() hold read access to the workqueue while executing a work and start_flush_work() check for write access if concurrnecy level is one and read access if higher.

This better represents what's going on and removes spurious lockdep warnings which are triggered by fake dependency chain created through flush_work().

Signed-off-by Tejun Heo <[email protected]>

Reported-by "Rafael J. Wysocki" <[email protected]>

---
How should this one be routed? The lockdep part can be split, merged back into workqueue tree and so on but that seems a bit too much. If it's okay, I'll route this through the workqueue tree. Going through the lockdep tree is fine too.

Thanks.

 include/linux/lockdep.h |    3 +++
 kernel/workqueue.c      |    8 ++++++--
 2 files changed, 9 insertions(+), 2 deletions(-)

include/linux/lockdep.h +3 -0

index 71c09b2..9f19430 100644
@@ -522,12 +522,15 @@ static inline void print_irqtrace_events(struct task_struct *curr)
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
 # ifdef CONFIG_PROVE_LOCKING
 #  define lock_map_acquire(l)		lock_acquire(l, 0, 0, 0, 2, NULL, _THIS_IP_)
+#  define lock_map_acquire_read(l)	lock_acquire(l, 0, 0, 2, 2, NULL, _THIS_IP_)
 # else
 #  define lock_map_acquire(l)		lock_acquire(l, 0, 0, 0, 1, NULL, _THIS_IP_)
+#  define lock_map_acquire_read(l)	lock_acquire(l, 0, 0, 2, 1, NULL, _THIS_IP_)
 # endif
 # define lock_map_release(l)			lock_release(l, 1, _THIS_IP_)
 #else
 # define lock_map_acquire(l)			do { } while (0)
+# define lock_map_acquire_read(l)		do { } while (0)
 # define lock_map_release(l)			do { } while (0)
 #endif
 

kernel/workqueue.c +6 -2

index 8ee6ec8..85f8f7b 100644
@@ -1840,7 +1840,7 @@ __acquires(&gcwq->lock)
 	spin_unlock_irq(&gcwq->lock);
 
 	work_clear_pending(work);
-	lock_map_acquire(&cwq->wq->lockdep_map);
+	lock_map_acquire_read(&cwq->wq->lockdep_map);
 	lock_map_acquire(&lockdep_map);
 	trace_workqueue_execute_start(work);
 	f(work);

@@ -2384,8 +2384,12 @@ static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
 	insert_wq_barrier(cwq, barr, work, worker);
 	spin_unlock_irq(&gcwq->lock);
 
-	lock_map_acquire(&cwq->wq->lockdep_map);
+	if (cwq->wq->saved_max_active > 1)
+		lock_map_acquire_read(&cwq->wq->lockdep_map);
+	else
+		lock_map_acquire(&cwq->wq->lockdep_map);
 	lock_map_release(&cwq->wq->lockdep_map);
+
 	return true;
 already_gone:
 	spin_unlock_irq(&gcwq->lock);

Currently, the lockdep annotation in flush_work() requires exclusive access on the workqueue the target work is queued on and triggers warning if a work is trying to flush another work on the same workqueue; however, this is no longer true as workqueues can now execute multiple works concurrently.

This patch adds lock_map_acquire_read() and make process_one_work() hold read access to the workqueue while executing a work and start_flush_work() check for write access if concurrnecy level is one and read access if higher.

This better represents what's going on and removes spurious lockdep warnings which are triggered by fake dependency chain created through flush_work().

kernel/workqueue.c +5 -1

@@ -2384,8 +2384,18 @@ static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
 	insert_wq_barrier(cwq, barr, work, worker);
 	spin_unlock_irq(&gcwq->lock);
 
-	lock_map_acquire(&cwq->wq->lockdep_map);
+	/*
+	 * If @max_active is 1 or rescuer is in use, flushing another work
+	 * item on the same workqueue may lead to deadlock.  Make sure the
+	 * flusher is not running on the same workqueue by verifying write
+	 * access.

Quoted 16 lines · Tejun Heo

kernel/workqueue.c +10 -1

@@ -2384,8 +2384,18 @@ static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
 	insert_wq_barrier(cwq, barr, work, worker);
 	spin_unlock_irq(&gcwq->lock);
 
-	lock_map_acquire(&cwq->wq->lockdep_map);
+	/*
+	 * If @max_active is 1 or rescuer is in use, flushing another work
+	 * item on the same workqueue may lead to deadlock.  Make sure the
+	 * flusher is not running on the same workqueue by verifying write
+	 * access.
+	 */
+	if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
+		lock_map_acquire(&cwq->wq->lockdep_map);
+	else
+		lock_map_acquire_read(&cwq->wq->lockdep_map);
 	lock_map_release(&cwq->wq->lockdep_map);

kernel/workqueue.c static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, · line 2401

Quoted 4 lines

kernel/workqueue.c +1 -0

+
 	return true;
 already_gone:
 	spin_unlock_irq(&gcwq->lock);

Ah, but this violates the rule that you must always use the most strict constraints. Code doesn't know if it will run in a rescue thread or not, hence it must assume it does.

Hello,

Quoted 18 lines · Peter Zijlstra

Quoted 16 lines · Tejun Heo

kernel/workqueue.c +10 -1

@@ -2384,8 +2384,18 @@ static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
 	insert_wq_barrier(cwq, barr, work, worker);
 	spin_unlock_irq(&gcwq->lock);
 
-	lock_map_acquire(&cwq->wq->lockdep_map);
+	/*
+	 * If @max_active is 1 or rescuer is in use, flushing another work
+	 * item on the same workqueue may lead to deadlock.  Make sure the
+	 * flusher is not running on the same workqueue by verifying write
+	 * access.
+	 */
+	if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
+		lock_map_acquire(&cwq->wq->lockdep_map);
+	else
+		lock_map_acquire_read(&cwq->wq->lockdep_map);
 	lock_map_release(&cwq->wq->lockdep_map);

Quoted 8 lines

kernel/workqueue.c static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, · line 2401

Quoted 4 lines

kernel/workqueue.c +1 -0

+
 	return true;
 already_gone:
 	spin_unlock_irq(&gcwq->lock);

Ah, but this violates the rule that you must always use the most strict constraints. Code doesn't know if it will run in a rescue thread or not, hence it must assume it does.

Hmmm? The code applies the most strict contraints. If the workqueue has a rescuer, flushing another work from the workqueue will always trigger lockdep warning. The rule is relaxed only for workqueues which aren't used for memory reclaiming && support parallel execution.

Quoted 4 lines

Hmmm? The code applies the most strict contraints. If the workqueue has a rescuer, flushing another work from the workqueue will always trigger lockdep warning. The rule is relaxed only for workqueues which aren't used for memory reclaiming && support parallel execution.

Ah, ok. I read it like: if the current thread is a rescue thread.

Quoted 6 lines · Peter Zijlstra

Quoted 4 lines

Hmmm? The code applies the most strict contraints. If the workqueue has a rescuer, flushing another work from the workqueue will always trigger lockdep warning. The rule is relaxed only for workqueues which aren't used for memory reclaiming && support parallel execution.

Ah, ok. I read it like: if the current thread is a rescue thread.

Ah, I see. Rafael, can you please verify the updated version is okay too (it should be but just in case)? Peter, shall I route this through workqueue tree with your ack?

Quoted 6 lines · Peter Zijlstra

Quoted 4 lines

Hmmm? The code applies the most strict contraints. If the workqueue has a rescuer, flushing another work from the workqueue will always trigger lockdep warning. The rule is relaxed only for workqueues which aren't used for memory reclaiming && support parallel execution.

Ah, ok. I read it like: if the current thread is a rescue thread.

Ah, I see. Rafael, can you please verify the updated version is okay too (it should be but just in case)?

Quoted 6 lines · Tejun Heo

Quoted 5 lines · Peter Zijlstra

Quoted 4 lines

Hmmm? The code applies the most strict contraints. If the workqueue has a rescuer, flushing another work from the workqueue will always trigger lockdep warning. The rule is relaxed only for workqueues which aren't used for memory reclaiming && support parallel execution.

Quoted 4 lines

Quoted 1 line

Ah, ok. I read it like: if the current thread is a rescue thread.

Ah, I see. Rafael, can you please verify the updated version is okay too (it should be but just in case)?

It still appears to be fine.