scsi: smartpqi: Add stream detection

Enhance performance by adding sequential stream detection for RAID5/RAID6
sequential write requests. Reduce stripe lock contention with full-stripe
write operations.

There is one common stripe lock for each RAID volume that can be set by
either the RAID engine or the AIO engine. The AIO path has I/O request
sizes well below the stripe size resulting in many Read-Modify-Write
operations.

Sending the request to the RAID engine allows for coalescing requests into
full stripe operations resulting in reduced Read-Modify-Write operations.

Link: https://lore.kernel.org/r/161549375693.25025.2962141451773219796.stgit@brunhildaReviewed-by: Scott Benesh <scott.benesh@microchip.com>
Reviewed-by: Mike McGowen <mike.mcgowen@microchip.com>
Reviewed-by: Scott Teel <scott.teel@microchip.com>
Reviewed-by: Kevin Barnett <kevin.barnett@microchip.com>
Signed-off-by: Don Brace <don.brace@microchip.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

drivers/scsi/smartpqi/smartpqi.h

@@ -1043,6 +1043,13 @@ struct pqi_scsi_dev_raid_map_data {
 
 #define RAID_CTLR_LUNID		"\0\0\0\0\0\0\0\0"
 
+#define NUM_STREAMS_PER_LUN	8
+
+struct pqi_stream_data {
+	u64	next_lba;
+	u32	last_accessed;
+};
+
 struct pqi_scsi_dev {
 	int	devtype;		/* as reported by INQUIRY commmand */
 	u8	device_type;		/* as reported by */
@@ -1097,6 +1104,7 @@ struct pqi_scsi_dev {
 	struct list_head add_list_entry;
 	struct list_head delete_list_entry;
 
+	struct pqi_stream_data stream_data[NUM_STREAMS_PER_LUN];
 	atomic_t scsi_cmds_outstanding;
 	atomic_t raid_bypass_cnt;
 };
@@ -1296,6 +1304,7 @@ struct pqi_ctrl_info {
 	u8		enable_r5_writes : 1;
 	u8		enable_r6_writes : 1;
 	u8		lv_drive_type_mix_valid : 1;
+	u8		enable_stream_detection : 1;
 
 	u8		ciss_report_log_flags;
 	u32		max_transfer_encrypted_sas_sata;

drivers/scsi/smartpqi/smartpqi_init.c

@@ -5688,8 +5688,82 @@ void pqi_prep_for_scsi_done(struct scsi_cmnd *scmd)
 	atomic_dec(&device->scsi_cmds_outstanding);
 }
 
-static int pqi_scsi_queue_command(struct Scsi_Host *shost,
+static bool pqi_is_parity_write_stream(struct pqi_ctrl_info *ctrl_info,
 	struct scsi_cmnd *scmd)
+{
+	u32 oldest_jiffies;
+	u8 lru_index;
+	int i;
+	int rc;
+	struct pqi_scsi_dev *device;
+	struct pqi_stream_data *pqi_stream_data;
+	struct pqi_scsi_dev_raid_map_data rmd;
+
+	if (!ctrl_info->enable_stream_detection)
+		return false;
+
+	rc = pqi_get_aio_lba_and_block_count(scmd, &rmd);
+	if (rc)
+		return false;
+
+	/* Check writes only. */
+	if (!rmd.is_write)
+		return false;
+
+	device = scmd->device->hostdata;
+
+	/* Check for RAID 5/6 streams. */
+	if (device->raid_level != SA_RAID_5 && device->raid_level != SA_RAID_6)
+		return false;
+
+	/*
+	 * If controller does not support AIO RAID{5,6} writes, need to send
+	 * requests down non-AIO path.
+	 */
+	if ((device->raid_level == SA_RAID_5 && !ctrl_info->enable_r5_writes) ||
+		(device->raid_level == SA_RAID_6 && !ctrl_info->enable_r6_writes))
+		return true;
+
+	lru_index = 0;
+	oldest_jiffies = INT_MAX;
+	for (i = 0; i < NUM_STREAMS_PER_LUN; i++) {
+		pqi_stream_data = &device->stream_data[i];
+		/*
+		 * Check for adjacent request or request is within
+		 * the previous request.
+		 */
+		if ((pqi_stream_data->next_lba &&
+			rmd.first_block >= pqi_stream_data->next_lba) &&
+			rmd.first_block <= pqi_stream_data->next_lba +
+				rmd.block_cnt) {
+			pqi_stream_data->next_lba = rmd.first_block +
+				rmd.block_cnt;
+			pqi_stream_data->last_accessed = jiffies;
+			return true;
+		}
+
+		/* unused entry */
+		if (pqi_stream_data->last_accessed == 0) {
+			lru_index = i;
+			break;
+		}
+
+		/* Find entry with oldest last accessed time. */
+		if (pqi_stream_data->last_accessed <= oldest_jiffies) {
+			oldest_jiffies = pqi_stream_data->last_accessed;
+			lru_index = i;
+		}
+	}
+
+	/* Set LRU entry. */
+	pqi_stream_data = &device->stream_data[lru_index];
+	pqi_stream_data->last_accessed = jiffies;
+	pqi_stream_data->next_lba = rmd.first_block + rmd.block_cnt;
+
+	return false;
+}
+
+static int pqi_scsi_queue_command(struct Scsi_Host *shost, struct scsi_cmnd *scmd)
 {
 	int rc;
 	struct pqi_ctrl_info *ctrl_info;
@@ -5736,11 +5810,12 @@ static int pqi_scsi_queue_command(struct Scsi_Host *shost,
 		raid_bypassed = false;
 		if (device->raid_bypass_enabled &&
 			!blk_rq_is_passthrough(scmd->request)) {
-			rc = pqi_raid_bypass_submit_scsi_cmd(ctrl_info, device,
-				scmd, queue_group);
-			if (rc == 0 || rc == SCSI_MLQUEUE_HOST_BUSY) {
-				raid_bypassed = true;
-				atomic_inc(&device->raid_bypass_cnt);
+			if (!pqi_is_parity_write_stream(ctrl_info, scmd)) {
+				rc = pqi_raid_bypass_submit_scsi_cmd(ctrl_info, device, scmd, queue_group);
+				if (rc == 0 || rc == SCSI_MLQUEUE_HOST_BUSY) {
+					raid_bypassed = true;
+					atomic_inc(&device->raid_bypass_cnt);
+				}
 			}
 		}
 		if (!raid_bypassed)