[PATCH v2 1/1] block/blk-mq: fix RT kernel regression with dedicated quiesce_sync_lock

From: Ionut Nechita <[email protected]>

In RT kernel (PREEMPT_RT), commit 679b1874eba7 ("block: fix ordering
between checking QUEUE_FLAG_QUIESCED request adding") causes severe
performance regression on systems with multiple MSI-X interrupt vectors.

The above change added spinlock_t queue_lock to blk_mq_run_hw_queue()
to synchronize QUEUE_FLAG_QUIESCED checks with blk_mq_unquiesce_queue().
While this works correctly in standard kernel, it causes catastrophic
serialization in RT kernel where spinlock_t converts to sleeping
rt_mutex.

Problem in RT kernel:
- blk_mq_run_hw_queue() is called from IRQ thread context (I/O completion)
- With 8 MSI-X vectors, all 8 IRQ threads contend on the same queue_lock
- queue_lock becomes rt_mutex (sleeping) in RT kernel
- IRQ threads serialize and enter D-state waiting for lock
- Throughput drops from 640 MB/s to 153 MB/s

The original commit message noted that memory barriers were considered
but rejected because "memory barrier is not easy to be maintained" -
barriers would need to be added at multiple call sites throughout the
block layer where work is added before calling blk_mq_run_hw_queue().

Solution:
Instead of using the general-purpose queue_lock or attempting complex
memory barrier pairing across many call sites, introduce a dedicated
raw_spinlock_t quiesce_sync_lock specifically for synchronizing the
quiesce state between:
- blk_mq_quiesce_queue_nowait()
- blk_mq_unquiesce_queue()
- blk_mq_run_hw_queue()

Why raw_spinlock is safe:
- Critical section is provably short (only flag and counter checks)
- No sleeping operations under lock
- raw_spinlock does not convert to rt_mutex in RT kernel
- Provides same ordering guarantees as original queue_lock approach

This approach:
- Maintains correctness of original synchronization
- Avoids sleeping in RT kernel's IRQ thread context
- Limits scope to only quiesce-related synchronization
- Simpler than auditing all call sites for memory barrier pairing

Additionally, change blk_freeze_queue_start to use async=true for better
performance in RT kernel by avoiding synchronous queue runs during freeze.

Test results on RT kernel (megaraid_sas with 8 MSI-X vectors):
- Before: 153 MB/s, 6-8 IRQ threads in D-state
- After:  640 MB/s, 0 IRQ threads blocked

Fixes: 679b1874eba7 ("block: fix ordering between checking QUEUE_FLAG_QUIESCED request adding")
Cc: [email protected]
Signed-off-by: Ionut Nechita <[email protected]>
---
 block/blk-core.c       |  1 +
 block/blk-mq.c         | 27 ++++++++++++++++-----------
 include/linux/blkdev.h |  6 ++++++
 3 files changed, 23 insertions(+), 11 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 474700ffaa1c8..fd615aeb5c463 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -434,6 +434,7 @@ struct request_queue *blk_alloc_queue(struct queue_limits *lim, int node_id)
 	mutex_init(&q->limits_lock);
 	mutex_init(&q->rq_qos_mutex);
 	spin_lock_init(&q->queue_lock);
+	raw_spin_lock_init(&q->quiesce_sync_lock);
 
 	init_waitqueue_head(&q->mq_freeze_wq);
 	mutex_init(&q->mq_freeze_lock);
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 0ad3dd3329db7..888718a782f88 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -171,7 +171,7 @@ bool __blk_freeze_queue_start(struct request_queue *q,
 		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);
+			blk_mq_run_hw_queues(q, true);
 	} else {
 		mutex_unlock(&q->mq_freeze_lock);
 	}
@@ -262,10 +262,10 @@ void blk_mq_quiesce_queue_nowait(struct request_queue *q)
 {
 	unsigned long flags;
 
-	spin_lock_irqsave(&q->queue_lock, flags);
+	raw_spin_lock_irqsave(&q->quiesce_sync_lock, flags);
 	if (!q->quiesce_depth++)
 		blk_queue_flag_set(QUEUE_FLAG_QUIESCED, q);
-	spin_unlock_irqrestore(&q->queue_lock, flags);
+	raw_spin_unlock_irqrestore(&q->quiesce_sync_lock, flags);
 }
 EXPORT_SYMBOL_GPL(blk_mq_quiesce_queue_nowait);
 
@@ -317,14 +317,14 @@ void blk_mq_unquiesce_queue(struct request_queue *q)
 	unsigned long flags;
 	bool run_queue = false;
 
-	spin_lock_irqsave(&q->queue_lock, flags);
+	raw_spin_lock_irqsave(&q->quiesce_sync_lock, flags);
 	if (WARN_ON_ONCE(q->quiesce_depth <= 0)) {
 		;
 	} else if (!--q->quiesce_depth) {
 		blk_queue_flag_clear(QUEUE_FLAG_QUIESCED, q);
 		run_queue = true;
 	}
-	spin_unlock_irqrestore(&q->queue_lock, flags);
+	raw_spin_unlock_irqrestore(&q->quiesce_sync_lock, flags);
 
 	/* dispatch requests which are inserted during quiescing */
 	if (run_queue)
@@ -2368,14 +2368,19 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 		unsigned long flags;
 
 		/*
-		 * Synchronize with blk_mq_unquiesce_queue(), because we check
-		 * if hw queue is quiesced locklessly above, we need the use
-		 * ->queue_lock to make sure we see the up-to-date status to
-		 * not miss rerunning the hw queue.
+		 * Synchronize with blk_mq_unquiesce_queue(). We check if hw
+		 * queue is quiesced locklessly above, so we need to use
+		 * quiesce_sync_lock to ensure we see the up-to-date status
+		 * and don't miss rerunning the hw queue.
+		 *
+		 * Uses raw_spinlock to avoid sleeping in RT kernel's IRQ
+		 * thread context during I/O completion. Critical section is
+		 * short (only flag and counter checks), making raw_spinlock
+		 * safe.
 		 */
-		spin_lock_irqsave(&hctx->queue->queue_lock, flags);
+		raw_spin_lock_irqsave(&hctx->queue->quiesce_sync_lock, flags);
 		need_run = blk_mq_hw_queue_need_run(hctx);
-		spin_unlock_irqrestore(&hctx->queue->queue_lock, flags);
+		raw_spin_unlock_irqrestore(&hctx->queue->quiesce_sync_lock, flags);
 
 		if (!need_run)
 			return;
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 99ef8cd7673c2..ab9e62aa3ae42 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -514,6 +514,12 @@ struct request_queue {
 	struct request		*last_merge;
 
 	spinlock_t		queue_lock;
+	/*
+	 * Synchronizes quiesce state checks between blk_mq_run_hw_queue()
+	 * and blk_mq_unquiesce_queue(). Uses raw_spinlock to avoid sleeping
+	 * in RT kernel's IRQ thread context during I/O completion.
+	 */
+	raw_spinlock_t		quiesce_sync_lock;
 
 	int			quiesce_depth;
 
-- 
2.52.0