scsi: target: Move I/O path stats to per CPU

The atomic use in the main I/O path is causing perf issues when using
higher performance backend devices and multiple queues. This moves the
stats to per CPU. Combined with the next patch that moves the
non_ordered/delayed_cmd_count to per CPU, IOPS by up to 33% for 8K IOS
when using 4 or more queues.

Signed-off-by: Mike Christie <michael.christie@oracle.com>
Link: https://lore.kernel.org/r/20250424032741.16216-2-michael.christie@oracle.com
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

drivers/target/target_core_device.c

@@ -55,14 +55,14 @@ transport_lookup_cmd_lun(struct se_cmd *se_cmd)
 	rcu_read_lock();
 	deve = target_nacl_find_deve(nacl, se_cmd->orig_fe_lun);
 	if (deve) {
-		atomic_long_inc(&deve->total_cmds);
+		this_cpu_inc(deve->stats->total_cmds);
 
 		if (se_cmd->data_direction == DMA_TO_DEVICE)
-			atomic_long_add(se_cmd->data_length,
+			this_cpu_add(deve->stats->write_bytes,
-					&deve->write_bytes);
+				     se_cmd->data_length);
 		else if (se_cmd->data_direction == DMA_FROM_DEVICE)
-			atomic_long_add(se_cmd->data_length,
+			this_cpu_add(deve->stats->read_bytes,
-					&deve->read_bytes);
+				     se_cmd->data_length);
 
 		if ((se_cmd->data_direction == DMA_TO_DEVICE) &&
 		    deve->lun_access_ro) {
@@ -126,14 +126,14 @@ out_unlock:
 	 * target_core_fabric_configfs.c:target_fabric_port_release
 	 */
 	se_cmd->se_dev = rcu_dereference_raw(se_lun->lun_se_dev);
-	atomic_long_inc(&se_cmd->se_dev->num_cmds);
+	this_cpu_inc(se_cmd->se_dev->stats->total_cmds);
 
 	if (se_cmd->data_direction == DMA_TO_DEVICE)
-		atomic_long_add(se_cmd->data_length,
+		this_cpu_add(se_cmd->se_dev->stats->write_bytes,
-				&se_cmd->se_dev->write_bytes);
+			     se_cmd->data_length);
 	else if (se_cmd->data_direction == DMA_FROM_DEVICE)
-		atomic_long_add(se_cmd->data_length,
+		this_cpu_add(se_cmd->se_dev->stats->read_bytes,
-				&se_cmd->se_dev->read_bytes);
+			     se_cmd->data_length);
 
 	return ret;
 }
@@ -322,6 +322,7 @@ int core_enable_device_list_for_node(
 	struct se_portal_group *tpg)
 {
 	struct se_dev_entry *orig, *new;
+	int ret = 0;
 
 	new = kzalloc(sizeof(*new), GFP_KERNEL);
 	if (!new) {
@@ -329,6 +330,12 @@ int core_enable_device_list_for_node(
 		return -ENOMEM;
 	}
 
+	new->stats = alloc_percpu(struct se_dev_entry_io_stats);
+	if (!new->stats) {
+		ret = -ENOMEM;
+		goto free_deve;
+	}
+
 	spin_lock_init(&new->ua_lock);
 	INIT_LIST_HEAD(&new->ua_list);
 	INIT_LIST_HEAD(&new->lun_link);
@@ -351,8 +358,8 @@ int core_enable_device_list_for_node(
 			       " for dynamic -> explicit NodeACL conversion:"
 				" %s\n", nacl->initiatorname);
 			mutex_unlock(&nacl->lun_entry_mutex);
-			kfree(new);
+			ret = -EINVAL;
-			return -EINVAL;
+			goto free_stats;
 		}
 		if (orig->se_lun_acl != NULL) {
 			pr_warn_ratelimited("Detected existing explicit"
@@ -360,8 +367,8 @@ int core_enable_device_list_for_node(
 				" mapped_lun: %llu, failing\n",
 				 nacl->initiatorname, mapped_lun);
 			mutex_unlock(&nacl->lun_entry_mutex);
-			kfree(new);
+			ret = -EINVAL;
-			return -EINVAL;
+			goto free_stats;
 		}
 
 		new->se_lun = lun;
@@ -394,6 +401,20 @@ int core_enable_device_list_for_node(
 
 	target_luns_data_has_changed(nacl, new, true);
 	return 0;
+
+free_stats:
+	free_percpu(new->stats);
+free_deve:
+	kfree(new);
+	return ret;
+}
+
+static void target_free_dev_entry(struct rcu_head *head)
+{
+	struct se_dev_entry *deve = container_of(head, struct se_dev_entry,
+						 rcu_head);
+	free_percpu(deve->stats);
+	kfree(deve);
 }
 
 void core_disable_device_list_for_node(
@@ -443,7 +464,7 @@ void core_disable_device_list_for_node(
 	kref_put(&orig->pr_kref, target_pr_kref_release);
 	wait_for_completion(&orig->pr_comp);
 
-	kfree_rcu(orig, rcu_head);
+	call_rcu(&orig->rcu_head, target_free_dev_entry);
 
 	core_scsi3_free_pr_reg_from_nacl(dev, nacl);
 	target_luns_data_has_changed(nacl, NULL, false);
@@ -689,11 +710,13 @@ struct se_device *target_alloc_device(struct se_hba *hba, const char *name)
 	if (!dev)
 		return NULL;
 
+	dev->stats = alloc_percpu(struct se_dev_io_stats);
+	if (!dev->stats)
+		goto free_device;
+
 	dev->queues = kcalloc(nr_cpu_ids, sizeof(*dev->queues), GFP_KERNEL);
-	if (!dev->queues) {
+	if (!dev->queues)
-		hba->backend->ops->free_device(dev);
+		goto free_stats;
-		return NULL;
-	}
 
 	dev->queue_cnt = nr_cpu_ids;
 	for (i = 0; i < dev->queue_cnt; i++) {
@@ -707,6 +730,7 @@ struct se_device *target_alloc_device(struct se_hba *hba, const char *name)
 		INIT_WORK(&q->sq.work, target_queued_submit_work);
 	}
 
+
 	dev->se_hba = hba;
 	dev->transport = hba->backend->ops;
 	dev->transport_flags = dev->transport->transport_flags_default;
@@ -791,6 +815,12 @@ struct se_device *target_alloc_device(struct se_hba *hba, const char *name)
 		sizeof(dev->t10_wwn.revision));
 
 	return dev;
+
+free_stats:
+	free_percpu(dev->stats);
+free_device:
+	hba->backend->ops->free_device(dev);
+	return NULL;
 }
 
 /*
@@ -1001,6 +1031,7 @@ void target_free_device(struct se_device *dev)
 		dev->transport->free_prot(dev);
 
 	kfree(dev->queues);
+	free_percpu(dev->stats);
 	dev->transport->free_device(dev);
 }
 

drivers/target/target_core_stat.c

@@ -280,30 +280,51 @@ static ssize_t target_stat_lu_num_cmds_show(struct config_item *item,
 		char *page)
 {
 	struct se_device *dev = to_stat_lu_dev(item);
+	struct se_dev_io_stats *stats;
+	unsigned int cpu;
+	u32 cmds = 0;
+
+	for_each_possible_cpu(cpu) {
+		stats = per_cpu_ptr(dev->stats, cpu);
+		cmds += stats->total_cmds;
+	}
 
 	/* scsiLuNumCommands */
-	return snprintf(page, PAGE_SIZE, "%lu\n",
+	return snprintf(page, PAGE_SIZE, "%u\n", cmds);
-			atomic_long_read(&dev->num_cmds));
 }
 
 static ssize_t target_stat_lu_read_mbytes_show(struct config_item *item,
 		char *page)
 {
 	struct se_device *dev = to_stat_lu_dev(item);
+	struct se_dev_io_stats *stats;
+	unsigned int cpu;
+	u32 bytes = 0;
+
+	for_each_possible_cpu(cpu) {
+		stats = per_cpu_ptr(dev->stats, cpu);
+		bytes += stats->read_bytes;
+	}
 
 	/* scsiLuReadMegaBytes */
-	return snprintf(page, PAGE_SIZE, "%lu\n",
+	return snprintf(page, PAGE_SIZE, "%u\n", bytes >> 20);
-			atomic_long_read(&dev->read_bytes) >> 20);
 }
 
 static ssize_t target_stat_lu_write_mbytes_show(struct config_item *item,
 		char *page)
 {
 	struct se_device *dev = to_stat_lu_dev(item);
+	struct se_dev_io_stats *stats;
+	unsigned int cpu;
+	u32 bytes = 0;
+
+	for_each_possible_cpu(cpu) {
+		stats = per_cpu_ptr(dev->stats, cpu);
+		bytes += stats->write_bytes;
+	}
 
 	/* scsiLuWrittenMegaBytes */
-	return snprintf(page, PAGE_SIZE, "%lu\n",
+	return snprintf(page, PAGE_SIZE, "%u\n", bytes >> 20);
-			atomic_long_read(&dev->write_bytes) >> 20);
 }
 
 static ssize_t target_stat_lu_resets_show(struct config_item *item, char *page)
@@ -1019,8 +1040,11 @@ static ssize_t target_stat_auth_num_cmds_show(struct config_item *item,
 {
 	struct se_lun_acl *lacl = auth_to_lacl(item);
 	struct se_node_acl *nacl = lacl->se_lun_nacl;
+	struct se_dev_entry_io_stats *stats;
 	struct se_dev_entry *deve;
+	unsigned int cpu;
 	ssize_t ret;
+	u32 cmds = 0;
 
 	rcu_read_lock();
 	deve = target_nacl_find_deve(nacl, lacl->mapped_lun);
@@ -1028,9 +1052,14 @@ static ssize_t target_stat_auth_num_cmds_show(struct config_item *item,
 		rcu_read_unlock();
 		return -ENODEV;
 	}
+
+	for_each_possible_cpu(cpu) {
+		stats = per_cpu_ptr(deve->stats, cpu);
+		cmds += stats->total_cmds;
+	}
+
 	/* scsiAuthIntrOutCommands */
-	ret = snprintf(page, PAGE_SIZE, "%lu\n",
+	ret = snprintf(page, PAGE_SIZE, "%u\n", cmds);
-		       atomic_long_read(&deve->total_cmds));
 	rcu_read_unlock();
 	return ret;
 }
@@ -1040,8 +1069,11 @@ static ssize_t target_stat_auth_read_mbytes_show(struct config_item *item,
 {
 	struct se_lun_acl *lacl = auth_to_lacl(item);
 	struct se_node_acl *nacl = lacl->se_lun_nacl;
+	struct se_dev_entry_io_stats *stats;
 	struct se_dev_entry *deve;
+	unsigned int cpu;
 	ssize_t ret;
+	u32 bytes = 0;
 
 	rcu_read_lock();
 	deve = target_nacl_find_deve(nacl, lacl->mapped_lun);
@@ -1049,9 +1081,14 @@ static ssize_t target_stat_auth_read_mbytes_show(struct config_item *item,
 		rcu_read_unlock();
 		return -ENODEV;
 	}
+
+	for_each_possible_cpu(cpu) {
+		stats = per_cpu_ptr(deve->stats, cpu);
+		bytes += stats->read_bytes;
+	}
+
 	/* scsiAuthIntrReadMegaBytes */
-	ret = snprintf(page, PAGE_SIZE, "%u\n",
+	ret = snprintf(page, PAGE_SIZE, "%u\n", bytes >> 20);
-		      (u32)(atomic_long_read(&deve->read_bytes) >> 20));
 	rcu_read_unlock();
 	return ret;
 }
@@ -1061,8 +1098,11 @@ static ssize_t target_stat_auth_write_mbytes_show(struct config_item *item,
 {
 	struct se_lun_acl *lacl = auth_to_lacl(item);
 	struct se_node_acl *nacl = lacl->se_lun_nacl;
+	struct se_dev_entry_io_stats *stats;
 	struct se_dev_entry *deve;
+	unsigned int cpu;
 	ssize_t ret;
+	u32 bytes = 0;
 
 	rcu_read_lock();
 	deve = target_nacl_find_deve(nacl, lacl->mapped_lun);
@@ -1070,9 +1110,14 @@ static ssize_t target_stat_auth_write_mbytes_show(struct config_item *item,
 		rcu_read_unlock();
 		return -ENODEV;
 	}
+
+	for_each_possible_cpu(cpu) {
+		stats = per_cpu_ptr(deve->stats, cpu);
+		bytes += stats->write_bytes;
+	}
+
 	/* scsiAuthIntrWrittenMegaBytes */
-	ret = snprintf(page, PAGE_SIZE, "%u\n",
+	ret = snprintf(page, PAGE_SIZE, "%u\n", bytes >> 20);
-		      (u32)(atomic_long_read(&deve->write_bytes) >> 20));
 	rcu_read_unlock();
 	return ret;
 }

include/target/target_core_base.h

@@ -669,15 +669,19 @@ struct se_lun_acl {
 	struct se_ml_stat_grps	ml_stat_grps;
 };
 
+struct se_dev_entry_io_stats {
+	u32			total_cmds;
+	u32			read_bytes;
+	u32			write_bytes;
+};
+
 struct se_dev_entry {
 	u64			mapped_lun;
 	u64			pr_res_key;
 	u64			creation_time;
 	bool			lun_access_ro;
 	u32			attach_count;
-	atomic_long_t		total_cmds;
+	struct se_dev_entry_io_stats __percpu	*stats;
-	atomic_long_t		read_bytes;
-	atomic_long_t		write_bytes;
 	/* Used for PR SPEC_I_PT=1 and REGISTER_AND_MOVE */
 	struct kref		pr_kref;
 	struct completion	pr_comp;
@@ -800,6 +804,12 @@ struct se_device_queue {
 	struct se_cmd_queue	sq;
 };
 
+struct se_dev_io_stats {
+	u32			total_cmds;
+	u32			read_bytes;
+	u32			write_bytes;
+};
+
 struct se_device {
 	/* Used for SAM Task Attribute ordering */
 	u32			dev_cur_ordered_id;
@@ -821,9 +831,7 @@ struct se_device {
 	atomic_long_t		num_resets;
 	atomic_long_t		aborts_complete;
 	atomic_long_t		aborts_no_task;
-	atomic_long_t		num_cmds;
+	struct se_dev_io_stats __percpu	*stats;
-	atomic_long_t		read_bytes;
-	atomic_long_t		write_bytes;
 	/* Active commands on this virtual SE device */
 	atomic_t		non_ordered;
 	bool			ordered_sync_in_progress;