diff --git a/block/blk-flush.c b/block/blk-flush.c
index f78bb39e589e831d71952d6ddc037771392b6d37..e4df894189cedbaa98c121dc050bb85ff2dd20df 100644
--- a/block/blk-flush.c
+++ b/block/blk-flush.c
@@ -229,7 +229,7 @@ static void flush_end_io(struct request *flush_rq, blk_status_t error)
 	/* release the tag's ownership to the req cloned from */
 	spin_lock_irqsave(&fq->mq_flush_lock, flags);
 
-	if (!refcount_dec_and_test(&flush_rq->ref)) {
+	if (!req_ref_put_and_test(flush_rq)) {
 		fq->rq_status = error;
 		spin_unlock_irqrestore(&fq->mq_flush_lock, flags);
 		return;
@@ -349,7 +349,7 @@ static void blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq,
 	 * and READ flush_rq->end_io
 	 */
 	smp_wmb();
-	refcount_set(&flush_rq->ref, 1);
+	req_ref_set(flush_rq, 1);
 
 	blk_flush_queue_rq(flush_rq, false);
 }
diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 995336abee3341c36d3c25584462fb617b7d79ec..380e2dd31bfcd6884420bfdcec87283301146c1f 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -228,7 +228,7 @@ static struct request *blk_mq_find_and_get_req(struct blk_mq_tags *tags,
 
 	spin_lock_irqsave(&tags->lock, flags);
 	rq = tags->rqs[bitnr];
-	if (!rq || rq->tag != bitnr || !refcount_inc_not_zero(&rq->ref))
+	if (!rq || rq->tag != bitnr || !req_ref_inc_not_zero(rq))
 		rq = NULL;
 	spin_unlock_irqrestore(&tags->lock, flags);
 	return rq;
diff --git a/block/blk-mq.c b/block/blk-mq.c
index fc4520e992b1f9a2be0c9b93b900e2b2dbe18b9e..8c7cab75229ebaa5dbe6ad2195d5b73fe1bd296a 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -394,7 +394,7 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 	INIT_LIST_HEAD(&rq->queuelist);
 	/* tag was already set */
 	WRITE_ONCE(rq->deadline, 0);
-	refcount_set(&rq->ref, 1);
+	req_ref_set(rq, 1);
 
 	if (rq->rq_flags & RQF_ELV) {
 		struct elevator_queue *e = data->q->elevator;
@@ -642,7 +642,7 @@ void blk_mq_free_request(struct request *rq)
 	rq_qos_done(q, rq);
 
 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
-	if (refcount_dec_and_test(&rq->ref))
+	if (req_ref_put_and_test(rq))
 		__blk_mq_free_request(rq);
 }
 EXPORT_SYMBOL_GPL(blk_mq_free_request);
@@ -938,7 +938,7 @@ void blk_mq_end_request_batch(struct io_comp_batch *iob)
 		rq_qos_done(rq->q, rq);
 
 		WRITE_ONCE(rq->state, MQ_RQ_IDLE);
-		if (!refcount_dec_and_test(&rq->ref))
+		if (!req_ref_put_and_test(rq))
 			continue;
 
 		blk_crypto_free_request(rq);
@@ -1401,7 +1401,7 @@ void blk_mq_put_rq_ref(struct request *rq)
 {
 	if (is_flush_rq(rq))
 		rq->end_io(rq, 0);
-	else if (refcount_dec_and_test(&rq->ref))
+	else if (req_ref_put_and_test(rq))
 		__blk_mq_free_request(rq);
 }
 
@@ -3049,7 +3049,7 @@ static void blk_mq_clear_rq_mapping(struct blk_mq_tags *drv_tags,
 			unsigned long rq_addr = (unsigned long)rq;
 
 			if (rq_addr >= start && rq_addr < end) {
-				WARN_ON_ONCE(refcount_read(&rq->ref) != 0);
+				WARN_ON_ONCE(req_ref_read(rq) != 0);
 				cmpxchg(&drv_tags->rqs[i], rq, NULL);
 			}
 		}
@@ -3383,7 +3383,7 @@ static void blk_mq_clear_flush_rq_mapping(struct blk_mq_tags *tags,
 	if (!tags)
 		return;
 
-	WARN_ON_ONCE(refcount_read(&flush_rq->ref) != 0);
+	WARN_ON_ONCE(req_ref_read(flush_rq) != 0);
 
 	for (i = 0; i < queue_depth; i++)
 		cmpxchg(&tags->rqs[i], flush_rq, NULL);
diff --git a/block/blk.h b/block/blk.h
index a55d82c3d1c217a339ce8c32e764b9e46a9d6ac6..24d8b333bb03743d917ed45243c65ad923f0f532 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -461,4 +461,35 @@ static inline bool should_fail_request(struct block_device *part,
 }
 #endif /* CONFIG_FAIL_MAKE_REQUEST */
 
+/*
+ * Optimized request reference counting. Ideally we'd make timeouts be more
+ * clever, as that's the only reason we need references at all... But until
+ * this happens, this is faster than using refcount_t. Also see:
+ *
+ * abc54d634334 ("io_uring: switch to atomic_t for io_kiocb reference count")
+ */
+#define req_ref_zero_or_close_to_overflow(req)	\
+	((unsigned int) atomic_read(&(req->ref)) + 127u <= 127u)
+
+static inline bool req_ref_inc_not_zero(struct request *req)
+{
+	return atomic_inc_not_zero(&req->ref);
+}
+
+static inline bool req_ref_put_and_test(struct request *req)
+{
+	WARN_ON_ONCE(req_ref_zero_or_close_to_overflow(req));
+	return atomic_dec_and_test(&req->ref);
+}
+
+static inline void req_ref_set(struct request *req, int value)
+{
+	atomic_set(&req->ref, value);
+}
+
+static inline int req_ref_read(struct request *req)
+{
+	return atomic_read(&req->ref);
+}
+
 #endif /* BLK_INTERNAL_H */
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 1b87b7c8bbffa9fc09d8d0b3ce864182f807969b..561beb5be7ec07708c0b02dd5b6719131bd52556 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -139,7 +139,7 @@ struct request {
 	unsigned short ioprio;
 
 	enum mq_rq_state state;
-	refcount_t ref;
+	atomic_t ref;
 
 	unsigned long deadline;