[4/5] block: always verify unfreeze lock on the owner task

Commit Message

Ming Lei Oct. 30, 2024, 12:42 p.m. UTC

commit f1be1788a32e ("block: model freeze & enter queue as lock for
supporting lockdep") tries to apply lockdep for verifying freeze &
unfreeze. However, the verification is only done the outmost freeze and
unfreeze. This way is actually not correct because q->mq_freeze_depth
still may drop to zero on other task instead of the freeze owner task.

Fix this issue by always verifying the last unfreeze lock on the owner
task context, and freeze lock is still verified on the outmost one.

Fixes: f1be1788a32e ("block: model freeze & enter queue as lock for supporting lockdep")
Signed-off-by: Ming Lei <[e-mail]>
---
 block/blk-core.c       |  2 +-
 block/blk-mq.c         | 64 ++++++++++++++++++++++++++++++++++++------
 block/blk.h            |  3 +-
 include/linux/blkdev.h |  4 +++
 4 files changed, 62 insertions(+), 11 deletions(-)

Comments

Christoph Hellwig Oct. 30, 2024, 2:46 p.m. UTC | #1

On Wed, Oct 30, 2024 at 08:42:36PM +0800, Ming Lei wrote:
> commit f1be1788a32e ("block: model freeze & enter queue as lock for
> supporting lockdep") tries to apply lockdep for verifying freeze &
> unfreeze. However, the verification is only done the outmost freeze and
> unfreeze. This way is actually not correct because q->mq_freeze_depth
> still may drop to zero on other task instead of the freeze owner task.

Well, that's how non-owner functions work in general.

> Fix this issue by always verifying the last unfreeze lock on the owner
> task context, and freeze lock is still verified on the outmost one.
> 
> Fixes: f1be1788a32e ("block: model freeze & enter queue as lock for supporting lockdep")

What does this actually fix vs just improving coverage?  Because the
hacks in here look pretty horrible and I'd be much happier if we didn't
have them.

Ming Lei Oct. 30, 2024, 2:53 p.m. UTC | #2

On Wed, Oct 30, 2024 at 03:46:14PM +0100, Christoph Hellwig wrote:
> On Wed, Oct 30, 2024 at 08:42:36PM +0800, Ming Lei wrote:
> > commit f1be1788a32e ("block: model freeze & enter queue as lock for
> > supporting lockdep") tries to apply lockdep for verifying freeze &
> > unfreeze. However, the verification is only done the outmost freeze and
> > unfreeze. This way is actually not correct because q->mq_freeze_depth
> > still may drop to zero on other task instead of the freeze owner task.
> 
> Well, that's how non-owner functions work in general.
> 
> > Fix this issue by always verifying the last unfreeze lock on the owner
> > task context, and freeze lock is still verified on the outmost one.
> > 
> > Fixes: f1be1788a32e ("block: model freeze & enter queue as lock for supporting lockdep")
> 
> What does this actually fix vs just improving coverage?  Because the
> hacks in here look pretty horrible and I'd be much happier if we didn't
> have them.
 
task A								task B

blk_mq_freeze_queue()
									blk_mq_freeze_queue()

blk_mq_unfreeze_queue()
									blk_mq_unfreeze_queue()


freeze_queue is verified on task A, but unfreeze_queue is verified on
task B, this way is definitely wrong.

This patch moves unfreeze_queue verification on task A.


Thanks, 
Ming

Patch

diff --git a/block/blk-core.c b/block/blk-core.c
index 09d10bb95fda..4f791a3114a1 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -287,7 +287,7 @@  bool blk_queue_start_drain(struct request_queue *q)
 	 * entering queue, so we call blk_freeze_queue_start() to
 	 * prevent I/O from crossing blk_queue_enter().
 	 */
-	bool freeze = __blk_freeze_queue_start(q);
+	bool freeze = __blk_freeze_queue_start(q, current);
 	if (queue_is_mq(q))
 		blk_mq_wake_waiters(q);
 	/* Make blk_queue_enter() reexamine the DYING flag. */
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 8e18284ede8f..0ec3b2db1d00 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -120,20 +120,66 @@  void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
 	inflight[1] = mi.inflight[1];
 }
 
-bool __blk_freeze_queue_start(struct request_queue *q)
+#ifdef CONFIG_LOCKDEP
+static bool blk_freeze_set_owner(struct request_queue *q,
+				 struct task_struct *owner)
 {
-	int freeze;
+	if (!owner)
+		return false;
+
+	if (!q->mq_freeze_depth) {
+		q->mq_freeze_owner = owner;
+		q->mq_freeze_owner_depth = 1;
+		return true;
+	}
+
+	if (owner == q->mq_freeze_owner)
+		q->mq_freeze_owner_depth += 1;
+	return false;
+}
+
+/* verify the last unfreeze in owner context */
+static bool blk_unfreeze_check_owner(struct request_queue *q)
+{
+	if (!q->mq_freeze_owner)
+		return false;
+	if (q->mq_freeze_owner != current)
+		return false;
+	if (--q->mq_freeze_owner_depth == 0) {
+		q->mq_freeze_owner = NULL;
+		return true;
+	}
+	return false;
+}
+
+#else
+
+static bool blk_freeze_set_owner(struct request_queue *q,
+				 struct task_struct *owner)
+{
+	return false;
+}
+
+static bool blk_unfreeze_check_owner(struct request_queue *q)
+{
+	return false;
+}
+#endif
+
+bool __blk_freeze_queue_start(struct request_queue *q,
+			      struct task_struct *owner)
+{
+	bool freeze;
 
 	mutex_lock(&q->mq_freeze_lock);
+	freeze = blk_freeze_set_owner(q, owner);
 	if (++q->mq_freeze_depth == 1) {
 		percpu_ref_kill(&q->q_usage_counter);
 		mutex_unlock(&q->mq_freeze_lock);
 		if (queue_is_mq(q))
 			blk_mq_run_hw_queues(q, false);
-		freeze = true;
 	} else {
 		mutex_unlock(&q->mq_freeze_lock);
-		freeze = false;
 	}
 
 	return freeze;
@@ -141,7 +187,7 @@  bool __blk_freeze_queue_start(struct request_queue *q)
 
 void blk_freeze_queue_start(struct request_queue *q)
 {
-	if (__blk_freeze_queue_start(q))
+	if (__blk_freeze_queue_start(q, current))
 		blk_freeze_acquire_lock(q, false, false);
 }
 EXPORT_SYMBOL_GPL(blk_freeze_queue_start);
@@ -170,7 +216,7 @@  EXPORT_SYMBOL_GPL(blk_mq_freeze_queue);
 
 bool __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
 {
-	int unfreeze = false;
+	bool unfreeze;
 
 	mutex_lock(&q->mq_freeze_lock);
 	if (force_atomic)
@@ -180,8 +226,8 @@  bool __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic)
 	if (!q->mq_freeze_depth) {
 		percpu_ref_resurrect(&q->q_usage_counter);
 		wake_up_all(&q->mq_freeze_wq);
-		unfreeze = true;
 	}
+	unfreeze = blk_unfreeze_check_owner(q);
 	mutex_unlock(&q->mq_freeze_lock);
 
 	return unfreeze;
@@ -203,7 +249,7 @@  EXPORT_SYMBOL_GPL(blk_mq_unfreeze_queue);
  */
 void blk_mq_freeze_queue_non_owner(struct request_queue *q)
 {
-	__blk_freeze_queue_start(q);
+	__blk_freeze_queue_start(q, NULL);
 	blk_mq_freeze_queue_wait(q);
 }
 EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_non_owner);
@@ -211,7 +257,7 @@  EXPORT_SYMBOL_GPL(blk_mq_freeze_queue_non_owner);
 /* non_owner variant of blk_freeze_queue_start */
 void blk_freeze_queue_start_non_owner(struct request_queue *q)
 {
-	__blk_freeze_queue_start(q);
+	__blk_freeze_queue_start(q, NULL);
 }
 EXPORT_SYMBOL_GPL(blk_freeze_queue_start_non_owner);
 
diff --git a/block/blk.h b/block/blk.h
index ac48b79cbf80..57fc035620d6 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -37,7 +37,8 @@  void blk_free_flush_queue(struct blk_flush_queue *q);
 
 bool __blk_mq_unfreeze_queue(struct request_queue *q, bool force_atomic);
 bool blk_queue_start_drain(struct request_queue *q);
-bool __blk_freeze_queue_start(struct request_queue *q);
+bool __blk_freeze_queue_start(struct request_queue *q,
+			      struct task_struct *owner);
 int __bio_queue_enter(struct request_queue *q, struct bio *bio);
 void submit_bio_noacct_nocheck(struct bio *bio);
 void bio_await_chain(struct bio *bio);
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 7bfc877e159e..379cd8eebdd9 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -575,6 +575,10 @@  struct request_queue {
 	struct throtl_data *td;
 #endif
 	struct rcu_head		rcu_head;
+#ifdef CONFIG_LOCKDEP
+	struct task_struct	*mq_freeze_owner;
+	int			mq_freeze_owner_depth;
+#endif
 	wait_queue_head_t	mq_freeze_wq;
 	/*
 	 * Protect concurrent access to q_usage_counter by