[PATCHv4] sbitmap: fix batched wait_cnt accounting

[PATCHv4] sbitmap: fix batched wait_cnt accounting

[Keith Busch]

From: Keith Busch <kbusch@kernel.org>

Batched completions can clear multiple bits, but we're only decrementing
the wait_cnt by one each time. This can cause waiters to never be woken,
stalling IO. Use the batched count instead.

Link: https://bugzilla.kernel.org/show_bug.cgi?id=215679
Signed-off-by: Keith Busch <kbusch@kernel.org>
---
v3->v4:

  Merged to the latest block for-6.1/block that contains the dependency
  fix that I attempted to implement in v3.

 block/blk-mq-tag.c      |  2 +-
 include/linux/sbitmap.h |  3 ++-
 lib/sbitmap.c           | 34 ++++++++++++++++++----------------
 3 files changed, 21 insertions(+), 18 deletions(-)

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 8e3b36d1cb57..9eb968e14d31 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -196,7 +196,7 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 		 * other allocations on previous queue won't be starved.
 		 */
 		if (bt != bt_prev)
-			sbitmap_queue_wake_up(bt_prev);
+			sbitmap_queue_wake_up(bt_prev, 1);
 
 		ws = bt_wait_ptr(bt, data->hctx);
 	} while (1);
diff --git a/include/linux/sbitmap.h b/include/linux/sbitmap.h
index 8f5a86e210b9..4d2d5205ab58 100644
--- a/include/linux/sbitmap.h
+++ b/include/linux/sbitmap.h
@@ -575,8 +575,9 @@ void sbitmap_queue_wake_all(struct sbitmap_queue *sbq);
  * sbitmap_queue_wake_up() - Wake up some of waiters in one waitqueue
  * on a &struct sbitmap_queue.
  * @sbq: Bitmap queue to wake up.
+ * @nr: Number of bits cleared.
  */
-void sbitmap_queue_wake_up(struct sbitmap_queue *sbq);
+void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr);
 
 /**
  * sbitmap_queue_show() - Dump &struct sbitmap_queue information to a &struct
diff --git a/lib/sbitmap.c b/lib/sbitmap.c
index cbfd2e677d87..cd631d2e4737 100644
--- a/lib/sbitmap.c
+++ b/lib/sbitmap.c
@@ -599,24 +599,27 @@ static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
 	return NULL;
 }
 
-static bool __sbq_wake_up(struct sbitmap_queue *sbq)
+static bool __sbq_wake_up(struct sbitmap_queue *sbq, int nr)
 {
 	struct sbq_wait_state *ws;
 	unsigned int wake_batch;
-	int wait_cnt;
+	int wait_cnt, cur;
 	bool ret;
 
 	ws = sbq_wake_ptr(sbq);
-	if (!ws)
+	if (!ws || nr <= 0)
 		return false;
 
-	wait_cnt = atomic_dec_return(&ws->wait_cnt);
-	/*
-	 * For concurrent callers of this, callers should call this function
-	 * again to wakeup a new batch on a different 'ws'.
-	 */
-	if (wait_cnt < 0)
-		return true;
+	cur = atomic_read(&ws->wait_cnt);
+	do {
+		/*
+		 * For concurrent callers of this, callers should call this
+		 * function again to wakeup a new batch on a different 'ws'.
+		 */
+		if (cur <= 0)
+			return true;
+		wait_cnt = cur - nr;
+	} while (!atomic_try_cmpxchg(&ws->wait_cnt, &cur, wait_cnt));
 
 	/*
 	 * If we decremented queue without waiters, retry to avoid lost
@@ -638,7 +641,7 @@ static bool __sbq_wake_up(struct sbitmap_queue *sbq)
 	 * Wake up first in case that concurrent callers decrease wait_cnt
 	 * while waitqueue is empty.
 	 */
-	wake_up_nr(&ws->wait, wake_batch);
+	wake_up_nr(&ws->wait, wake_batch - wait_cnt);
 
 	/*
 	 * Pairs with the memory barrier in sbitmap_queue_resize() to
@@ -663,12 +666,11 @@ static bool __sbq_wake_up(struct sbitmap_queue *sbq)
 	return ret;
 }
 
-void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
+void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
 {
-	while (__sbq_wake_up(sbq))
+	while (__sbq_wake_up(sbq, nr))
 		;
 }
-EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
 
 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
 {
@@ -705,7 +707,7 @@ void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
 		atomic_long_andnot(mask, (atomic_long_t *) addr);
 
 	smp_mb__after_atomic();
-	sbitmap_queue_wake_up(sbq);
+	sbitmap_queue_wake_up(sbq, nr_tags);
 	sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
 					tags[nr_tags - 1] - offset);
 }
@@ -733,7 +735,7 @@ void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
 	 * waiter. See the comment on waitqueue_active().
 	 */
 	smp_mb__after_atomic();
-	sbitmap_queue_wake_up(sbq);
+	sbitmap_queue_wake_up(sbq, 1);
 	sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
 }
 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
-- 
2.30.2

Re: [PATCHv4] sbitmap: fix batched wait_cnt accounting

[Keith Busch]

On Thu, Sep 08, 2022 at 02:51:32PM -0700, Keith Busch wrote:
> From: Keith Busch <kbusch@kernel.org>
> 
> Batched completions can clear multiple bits, but we're only decrementing
> the wait_cnt by one each time. This can cause waiters to never be woken,
> stalling IO. Use the batched count instead.
> 
> Link: https://bugzilla.kernel.org/show_bug.cgi?id=215679

A longer running test finally failed in a similar manner as before. It use to
fail much quicker, but something is still off, so please don't apply or test
this patch yet.

Re: [PATCHv4] sbitmap: fix batched wait_cnt accounting

[Keith Busch]

On Thu, Sep 08, 2022 at 02:51:32PM -0700, Keith Busch wrote:
> @@ -638,7 +641,7 @@ static bool __sbq_wake_up(struct sbitmap_queue *sbq)
>  	 * Wake up first in case that concurrent callers decrease wait_cnt
>  	 * while waitqueue is empty.
>  	 */
> -	wake_up_nr(&ws->wait, wake_batch);
> +	wake_up_nr(&ws->wait, wake_batch - wait_cnt);

The reason it was failing is this was not spreading the wake up among enough
wait states. I'll have to go back to the same method that patch 1 was doing
with clamping the decrement count..