scsi: blk-mq: Return budget token from .get_budget callback

SCSI uses a global atomic variable to track queue depth for each
LUN/request queue.

This doesn't scale well when there are lots of CPU cores and the disk is
very fast. It has been observed that IOPS is affected a lot by tracking
queue depth via sdev->device_busy in the I/O path.

Return budget token from .get_budget callback. The budget token can be
passed to driver so that we can replace the atomic variable with
sbitmap_queue and alleviate the scaling problems that way.

Link: https://lore.kernel.org/r/20210122023317.687987-9-ming.lei@redhat.com
Cc: Omar Sandoval <osandov@fb.com>
Cc: Kashyap Desai <kashyap.desai@broadcom.com>
Cc: Sumanesh Samanta <sumanesh.samanta@broadcom.com>
Cc: Ewan D. Milne <emilne@redhat.com>
Tested-by: Sumanesh Samanta <sumanesh.samanta@broadcom.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

block/blk-mq-sched.c

@@ -131,6 +131,7 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 
 	do {
 		struct request *rq;
+		int budget_token;
 
 		if (e->type->ops.has_work && !e->type->ops.has_work(hctx))
 			break;
@@ -140,12 +141,13 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 			break;
 		}
 
-		if (!blk_mq_get_dispatch_budget(q))
+		budget_token = blk_mq_get_dispatch_budget(q);
+		if (budget_token < 0)
 			break;
 
 		rq = e->type->ops.dispatch_request(hctx);
 		if (!rq) {
-			blk_mq_put_dispatch_budget(q);
+			blk_mq_put_dispatch_budget(q, budget_token);
 			/*
 			 * We're releasing without dispatching. Holding the
 			 * budget could have blocked any "hctx"s with the
@@ -157,6 +159,8 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 			break;
 		}
 
+		blk_mq_set_rq_budget_token(rq, budget_token);
+
 		/*
 		 * Now this rq owns the budget which has to be released
 		 * if this rq won't be queued to driver via .queue_rq()
@@ -230,6 +234,8 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
 	struct request *rq;
 
 	do {
+		int budget_token;
+
 		if (!list_empty_careful(&hctx->dispatch)) {
 			ret = -EAGAIN;
 			break;
@@ -238,12 +244,13 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
 		if (!sbitmap_any_bit_set(&hctx->ctx_map))
 			break;
 
-		if (!blk_mq_get_dispatch_budget(q))
+		budget_token = blk_mq_get_dispatch_budget(q);
+		if (budget_token < 0)
 			break;
 
 		rq = blk_mq_dequeue_from_ctx(hctx, ctx);
 		if (!rq) {
-			blk_mq_put_dispatch_budget(q);
+			blk_mq_put_dispatch_budget(q, budget_token);
 			/*
 			 * We're releasing without dispatching. Holding the
 			 * budget could have blocked any "hctx"s with the
@@ -255,6 +262,8 @@ static int blk_mq_do_dispatch_ctx(struct blk_mq_hw_ctx *hctx)
 			break;
 		}
 
+		blk_mq_set_rq_budget_token(rq, budget_token);
+
 		/*
 		 * Now this rq owns the budget which has to be released
 		 * if this rq won't be queued to driver via .queue_rq()

block/blk-mq.c

@@ -1277,11 +1277,16 @@ static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
 						  bool need_budget)
 {
 	struct blk_mq_hw_ctx *hctx = rq->mq_hctx;
+	int budget_token = -1;
 
-	if (need_budget && !blk_mq_get_dispatch_budget(rq->q)) {
+	if (need_budget) {
+		budget_token = blk_mq_get_dispatch_budget(rq->q);
+		if (budget_token < 0) {
 			blk_mq_put_driver_tag(rq);
 			return PREP_DISPATCH_NO_BUDGET;
 		}
+		blk_mq_set_rq_budget_token(rq, budget_token);
+	}
 
 	if (!blk_mq_get_driver_tag(rq)) {
 		/*
@@ -1297,7 +1302,7 @@ static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
 			 * together during handling partial dispatch
 			 */
 			if (need_budget)
-				blk_mq_put_dispatch_budget(rq->q);
+				blk_mq_put_dispatch_budget(rq->q, budget_token);
 			return PREP_DISPATCH_NO_TAG;
 		}
 	}
@@ -1307,12 +1312,16 @@ static enum prep_dispatch blk_mq_prep_dispatch_rq(struct request *rq,
 
 /* release all allocated budgets before calling to blk_mq_dispatch_rq_list */
 static void blk_mq_release_budgets(struct request_queue *q,
-		unsigned int nr_budgets)
+		struct list_head *list)
 {
-	int i;
+	struct request *rq;
 
-	for (i = 0; i < nr_budgets; i++)
-		blk_mq_put_dispatch_budget(q);
+	list_for_each_entry(rq, list, queuelist) {
+		int budget_token = blk_mq_get_rq_budget_token(rq);
+
+		if (budget_token >= 0)
+			blk_mq_put_dispatch_budget(q, budget_token);
+	}
 }
 
 /*
@@ -1410,7 +1419,8 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
 			(hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED);
 		bool no_budget_avail = prep == PREP_DISPATCH_NO_BUDGET;
 
-		blk_mq_release_budgets(q, nr_budgets);
+		if (nr_budgets)
+			blk_mq_release_budgets(q, list);
 
 		spin_lock(&hctx->lock);
 		list_splice_tail_init(list, &hctx->dispatch);
@@ -2009,6 +2019,7 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 {
 	struct request_queue *q = rq->q;
 	bool run_queue = true;
+	int budget_token;
 
 	/*
 	 * RCU or SRCU read lock is needed before checking quiesced flag.
@@ -2026,11 +2037,14 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 	if (q->elevator && !bypass_insert)
 		goto insert;
 
-	if (!blk_mq_get_dispatch_budget(q))
+	budget_token = blk_mq_get_dispatch_budget(q);
+	if (budget_token < 0)
 		goto insert;
 
+	blk_mq_set_rq_budget_token(rq, budget_token);
+
 	if (!blk_mq_get_driver_tag(rq)) {
-		blk_mq_put_dispatch_budget(q);
+		blk_mq_put_dispatch_budget(q, budget_token);
 		goto insert;
 	}
 

block/blk-mq.h

@@ -187,17 +187,34 @@ unsigned int blk_mq_in_flight(struct request_queue *q,
 void blk_mq_in_flight_rw(struct request_queue *q, struct block_device *part,
 		unsigned int inflight[2]);
 
-static inline void blk_mq_put_dispatch_budget(struct request_queue *q)
+static inline void blk_mq_put_dispatch_budget(struct request_queue *q,
+					      int budget_token)
 {
 	if (q->mq_ops->put_budget)
-		q->mq_ops->put_budget(q);
+		q->mq_ops->put_budget(q, budget_token);
 }
 
-static inline bool blk_mq_get_dispatch_budget(struct request_queue *q)
+static inline int blk_mq_get_dispatch_budget(struct request_queue *q)
 {
 	if (q->mq_ops->get_budget)
 		return q->mq_ops->get_budget(q);
-	return true;
+	return 0;
+}
+
+static inline void blk_mq_set_rq_budget_token(struct request *rq, int token)
+{
+	if (token < 0)
+		return;
+
+	if (rq->q->mq_ops->set_rq_budget_token)
+		rq->q->mq_ops->set_rq_budget_token(rq, token);
+}
+
+static inline int blk_mq_get_rq_budget_token(struct request *rq)
+{
+	if (rq->q->mq_ops->get_rq_budget_token)
+		return rq->q->mq_ops->get_rq_budget_token(rq);
+	return -1;
 }
 
 static inline void __blk_mq_inc_active_requests(struct blk_mq_hw_ctx *hctx)

drivers/scsi/scsi_lib.c

@@ -329,6 +329,7 @@ void scsi_device_unbusy(struct scsi_device *sdev, struct scsi_cmnd *cmd)
 		atomic_dec(&starget->target_busy);
 
 	atomic_dec(&sdev->device_busy);
+	cmd->budget_token = -1;
 }
 
 static void scsi_kick_queue(struct request_queue *q)
@@ -1143,6 +1144,7 @@ void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
 	unsigned long jiffies_at_alloc;
 	int retries, to_clear;
 	bool in_flight;
+	int budget_token = cmd->budget_token;
 
 	if (!blk_rq_is_scsi(rq) && !(flags & SCMD_INITIALIZED)) {
 		flags |= SCMD_INITIALIZED;
@@ -1171,6 +1173,7 @@ void scsi_init_command(struct scsi_device *dev, struct scsi_cmnd *cmd)
 	cmd->retries = retries;
 	if (in_flight)
 		__set_bit(SCMD_STATE_INFLIGHT, &cmd->state);
+	cmd->budget_token = budget_token;
 
 }
 
@@ -1605,19 +1608,19 @@ static void scsi_mq_done(struct scsi_cmnd *cmd)
 	blk_mq_complete_request(cmd->request);
 }
 
-static void scsi_mq_put_budget(struct request_queue *q)
+static void scsi_mq_put_budget(struct request_queue *q, int budget_token)
 {
 	struct scsi_device *sdev = q->queuedata;
 
 	atomic_dec(&sdev->device_busy);
 }
 
-static bool scsi_mq_get_budget(struct request_queue *q)
+static int scsi_mq_get_budget(struct request_queue *q)
 {
 	struct scsi_device *sdev = q->queuedata;
 
 	if (scsi_dev_queue_ready(q, sdev))
-		return true;
+		return 0;
 
 	atomic_inc(&sdev->restarts);
 
@@ -1639,7 +1642,7 @@ static bool scsi_mq_get_budget(struct request_queue *q)
 	if (unlikely(atomic_read(&sdev->device_busy) == 0 &&
 				!scsi_device_blocked(sdev)))
 		blk_mq_delay_run_hw_queues(sdev->request_queue, SCSI_QUEUE_DELAY);
-	return false;
+	return -1;
 }
 
 static void scsi_mq_set_rq_budget_token(struct request *req, int token)
@@ -1667,6 +1670,8 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
 	blk_status_t ret;
 	int reason;
 
+	WARN_ON_ONCE(cmd->budget_token < 0);
+
 	/*
 	 * If the device is not in running state we will reject some or all
 	 * commands.
@@ -1718,7 +1723,8 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
 	if (scsi_target(sdev)->can_queue > 0)
 		atomic_dec(&scsi_target(sdev)->target_busy);
 out_put_budget:
-	scsi_mq_put_budget(q);
+	scsi_mq_put_budget(q, cmd->budget_token);
+	cmd->budget_token = -1;
 	switch (ret) {
 	case BLK_STS_OK:
 		break;

include/linux/blk-mq.h

@@ -306,12 +306,12 @@ struct blk_mq_ops {
 	 * reserved budget. Also we have to handle failure case
 	 * of .get_budget for avoiding I/O deadlock.
 	 */
-	bool (*get_budget)(struct request_queue *);
+	int (*get_budget)(struct request_queue *);
 
 	/**
 	 * @put_budget: Release the reserved budget.
 	 */
-	void (*put_budget)(struct request_queue *);
+	void (*put_budget)(struct request_queue *, int);
 
 	/*
 	 * @set_rq_budget_toekn: store rq's budget token