scsi: lpfc: Add auto EQ delay logic

Administrator intervention is currently required to get good numbers
when switching from running latency tests to IOPS tests.

The configured interrupt coalescing values will greatly effect the
results of these tests.  Currently, the driver has a single coalescing
value set by values of the module attribute.  This patch changes the
driver to support auto-configuration of the coalescing value based on
the total number of outstanding IOs and average number of CQEs processed
per interrupt for an EQ.  Values are checked every 5 seconds.

The driver defaults to the automatic selection. Automatic selection can
be disabled by the new lpfc_auto_imax module_parameter.

Older hardware can only change interrupt coalescing by mailbox
command. Newer hardware supports change via a register. The patch
support both.

Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

8 files changed, 177 insertions, 13 deletions

diff --git a/drivers/scsi/lpfc/lpfc.h b/drivers/scsi/lpfc/lpfc.h
index a9d7372..562dc01 100644
--- a/drivers/scsi/lpfc/lpfc.h
+++ b/drivers/scsi/lpfc/lpfc.h
@@ -756,6 +756,7 @@ struct lpfc_hba {
 	uint8_t  nvmet_support;	/* driver supports NVMET */
 #define LPFC_NVMET_MAX_PORTS	32
 	uint8_t  mds_diags_support;
+	uint32_t initial_imax;
 
 	/* HBA Config Parameters */
 	uint32_t cfg_ack0;
@@ -777,6 +778,7 @@ struct lpfc_hba {
 	uint32_t cfg_poll_tmo;
 	uint32_t cfg_task_mgmt_tmo;
 	uint32_t cfg_use_msi;
+	uint32_t cfg_auto_imax;
 	uint32_t cfg_fcp_imax;
 	uint32_t cfg_fcp_cpu_map;
 	uint32_t cfg_fcp_io_channel;
@@ -1050,6 +1052,7 @@ struct lpfc_hba {
 
 	uint8_t temp_sensor_support;
 	/* Fields used for heart beat. */
+	unsigned long last_eqdelay_time;
 	unsigned long last_completion_time;
 	unsigned long skipped_hb;
 	struct timer_list hb_tmofunc;
diff --git a/drivers/scsi/lpfc/lpfc_attr.c b/drivers/scsi/lpfc/lpfc_attr.c
index 8eee39d..66269e3 100644
--- a/drivers/scsi/lpfc/lpfc_attr.c
+++ b/drivers/scsi/lpfc/lpfc_attr.c
@@ -4481,9 +4481,11 @@ lpfc_fcp_imax_store(struct device *dev, struct device_attribute *attr,
 		return -EINVAL;
 
 	phba->cfg_fcp_imax = (uint32_t)val;
+	phba->initial_imax = phba->cfg_fcp_imax;
 
 	for (i = 0; i < phba->io_channel_irqs; i += LPFC_MAX_EQ_DELAY_EQID_CNT)
-		lpfc_modify_hba_eq_delay(phba, i);
+		lpfc_modify_hba_eq_delay(phba, i, LPFC_MAX_EQ_DELAY_EQID_CNT,
+					 val);
 
 	return strlen(buf);
 }
@@ -4538,6 +4540,16 @@ lpfc_fcp_imax_init(struct lpfc_hba *phba, int val)
 static DEVICE_ATTR(lpfc_fcp_imax, S_IRUGO | S_IWUSR,
 		   lpfc_fcp_imax_show, lpfc_fcp_imax_store);
 
+/*
+ * lpfc_auto_imax: Controls Auto-interrupt coalescing values support.
+ *       0       No auto_imax support
+ *       1       auto imax on
+ * Auto imax will change the value of fcp_imax on a per EQ basis, using
+ * the EQ Delay Multiplier, depending on the activity for that EQ.
+ * Value range [0,1]. Default value is 1.
+ */
+LPFC_ATTR_RW(auto_imax, 1, 0, 1, "Enable Auto imax");
+
 /**
  * lpfc_state_show - Display current driver CPU affinity
  * @dev: class converted to a Scsi_host structure.
@@ -5164,6 +5176,7 @@ struct device_attribute *lpfc_hba_attrs[] = {
 	&dev_attr_lpfc_task_mgmt_tmo,
 	&dev_attr_lpfc_use_msi,
 	&dev_attr_lpfc_nvme_oas,
+	&dev_attr_lpfc_auto_imax,
 	&dev_attr_lpfc_fcp_imax,
 	&dev_attr_lpfc_fcp_cpu_map,
 	&dev_attr_lpfc_fcp_io_channel,
@@ -6182,6 +6195,7 @@ lpfc_get_cfgparam(struct lpfc_hba *phba)
 	lpfc_enable_SmartSAN_init(phba, lpfc_enable_SmartSAN);
 	lpfc_use_msi_init(phba, lpfc_use_msi);
 	lpfc_nvme_oas_init(phba, lpfc_nvme_oas);
+	lpfc_auto_imax_init(phba, lpfc_auto_imax);
 	lpfc_fcp_imax_init(phba, lpfc_fcp_imax);
 	lpfc_fcp_cpu_map_init(phba, lpfc_fcp_cpu_map);
 	lpfc_enable_hba_reset_init(phba, lpfc_enable_hba_reset);
@@ -6226,6 +6240,10 @@ lpfc_get_cfgparam(struct lpfc_hba *phba)
 			phba->cfg_enable_fc4_type |= LPFC_ENABLE_FCP;
 	}
 
+	if (phba->cfg_auto_imax && !phba->cfg_fcp_imax)
+		phba->cfg_auto_imax = 0;
+	phba->initial_imax = phba->cfg_fcp_imax;
+
 	/* A value of 0 means use the number of CPUs found in the system */
 	if (phba->cfg_fcp_io_channel == 0)
 		phba->cfg_fcp_io_channel = phba->sli4_hba.num_present_cpu;
diff --git a/drivers/scsi/lpfc/lpfc_debugfs.c b/drivers/scsi/lpfc/lpfc_debugfs.c
index bd45c50..cc49850 100644
--- a/drivers/scsi/lpfc/lpfc_debugfs.c
+++ b/drivers/scsi/lpfc/lpfc_debugfs.c
@@ -3265,9 +3265,9 @@ __lpfc_idiag_print_eq(struct lpfc_queue *qp, char *eqtype,
 
 	len += snprintf(pbuffer + len, LPFC_QUE_INFO_GET_BUF_SIZE - len,
 			"\n%s EQ info: EQ-STAT[max:x%x noE:x%x "
-			"bs:x%x proc:x%llx]\n",
+			"bs:x%x proc:x%llx eqd %d]\n",
 			eqtype, qp->q_cnt_1, qp->q_cnt_2, qp->q_cnt_3,
-			(unsigned long long)qp->q_cnt_4);
+			(unsigned long long)qp->q_cnt_4, qp->q_mode);
 	len += snprintf(pbuffer + len, LPFC_QUE_INFO_GET_BUF_SIZE - len,
 			"EQID[%02d], QE-CNT[%04d], QE-SZ[%04d], "
 			"HST-IDX[%04d], PRT-IDX[%04d], PST[%03d]",
diff --git a/drivers/scsi/lpfc/lpfc_hw4.h b/drivers/scsi/lpfc/lpfc_hw4.h
index e0a5fce..bb47157 100644
--- a/drivers/scsi/lpfc/lpfc_hw4.h
+++ b/drivers/scsi/lpfc/lpfc_hw4.h
@@ -197,6 +197,7 @@ struct lpfc_sli_intf {
 
 /* Delay Multiplier constant */
 #define LPFC_DMULT_CONST       651042
+#define LPFC_DMULT_MAX         1023
 
 /* Configuration of Interrupts / sec for entire HBA port */
 #define LPFC_MIN_IMAX          5000
@@ -657,6 +658,15 @@ struct lpfc_register {
 #define LPFC_CTL_PORT_ER1_OFFSET	0x40C
 #define LPFC_CTL_PORT_ER2_OFFSET	0x410
 
+#define LPFC_CTL_PORT_EQ_DELAY_OFFSET	0x418
+#define lpfc_sliport_eqdelay_delay_SHIFT 16
+#define lpfc_sliport_eqdelay_delay_MASK	0xffff
+#define lpfc_sliport_eqdelay_delay_WORD	word0
+#define lpfc_sliport_eqdelay_id_SHIFT	0
+#define lpfc_sliport_eqdelay_id_MASK	0xfff
+#define lpfc_sliport_eqdelay_id_WORD	word0
+#define LPFC_SEC_TO_USEC		1000000
+
 /* The following Registers apply to SLI4 if_type 0 UCNAs. They typically
  * reside in BAR 2.
  */
@@ -3258,6 +3268,10 @@ struct lpfc_sli4_parameters {
 #define cfg_xib_SHIFT				4
 #define cfg_xib_MASK				0x00000001
 #define cfg_xib_WORD				word19
+#define cfg_eqdr_SHIFT				8
+#define cfg_eqdr_MASK				0x00000001
+#define cfg_eqdr_WORD				word19
+#define LPFC_NODELAY_MAX_IO		32
 };
 
 #define LPFC_SET_UE_RECOVERY		0x10
diff --git a/drivers/scsi/lpfc/lpfc_init.c b/drivers/scsi/lpfc/lpfc_init.c
index a8258060..9d3a126 100644
--- a/drivers/scsi/lpfc/lpfc_init.c
+++ b/drivers/scsi/lpfc/lpfc_init.c
@@ -1249,6 +1249,12 @@ lpfc_hb_timeout_handler(struct lpfc_hba *phba)
 	int retval, i;
 	struct lpfc_sli *psli = &phba->sli;
 	LIST_HEAD(completions);
+	struct lpfc_queue *qp;
+	unsigned long time_elapsed;
+	uint32_t tick_cqe, max_cqe, val;
+	uint64_t tot, data1, data2, data3;
+	struct lpfc_register reg_data;
+	void __iomem *eqdreg = phba->sli4_hba.u.if_type2.EQDregaddr;
 
 	vports = lpfc_create_vport_work_array(phba);
 	if (vports != NULL)
@@ -1263,6 +1269,95 @@ lpfc_hb_timeout_handler(struct lpfc_hba *phba)
 		(phba->pport->fc_flag & FC_OFFLINE_MODE))
 		return;
 
+	if (phba->cfg_auto_imax) {
+		if (!phba->last_eqdelay_time) {
+			phba->last_eqdelay_time = jiffies;
+			goto skip_eqdelay;
+		}
+		time_elapsed = jiffies - phba->last_eqdelay_time;
+		phba->last_eqdelay_time = jiffies;
+
+		tot = 0xffff;
+		/* Check outstanding IO count */
+		if (phba->cfg_enable_fc4_type & LPFC_ENABLE_NVME) {
+			if (phba->nvmet_support) {
+				spin_lock(&phba->sli4_hba.nvmet_io_lock);
+				tot = phba->sli4_hba.nvmet_xri_cnt -
+					phba->sli4_hba.nvmet_ctx_cnt;
+				spin_unlock(&phba->sli4_hba.nvmet_io_lock);
+			} else {
+				tot = atomic_read(&phba->fc4NvmeIoCmpls);
+				data1 = atomic_read(
+					&phba->fc4NvmeInputRequests);
+				data2 = atomic_read(
+					&phba->fc4NvmeOutputRequests);
+				data3 = atomic_read(
+					&phba->fc4NvmeControlRequests);
+				tot =  (data1 + data2 + data3) - tot;
+			}
+		}
+
+		/* Interrupts per sec per EQ */
+		val = phba->cfg_fcp_imax / phba->io_channel_irqs;
+		tick_cqe = val / CONFIG_HZ; /* Per tick per EQ */
+
+		/* Assume 1 CQE/ISR, calc max CQEs allowed for time duration */
+		max_cqe = time_elapsed * tick_cqe;
+
+		for (i = 0; i < phba->io_channel_irqs; i++) {
+			/* Fast-path EQ */
+			qp = phba->sli4_hba.hba_eq[i];
+			if (!qp)
+				continue;
+
+			/* Use no EQ delay if we don't have many outstanding
+			 * IOs, or if we are only processing 1 CQE/ISR or less.
+			 * Otherwise, assume we can process up to lpfc_fcp_imax
+			 * interrupts per HBA.
+			 */
+			if (tot < LPFC_NODELAY_MAX_IO ||
+			    qp->EQ_cqe_cnt <= max_cqe)
+				val = 0;
+			else
+				val = phba->cfg_fcp_imax;
+
+			if (phba->sli.sli_flag & LPFC_SLI_USE_EQDR) {
+				/* Use EQ Delay Register method */
+
+				/* Convert for EQ Delay register */
+				if (val) {
+					/* First, interrupts per sec per EQ */
+					val = phba->cfg_fcp_imax /
+						phba->io_channel_irqs;
+
+					/* us delay between each interrupt */
+					val = LPFC_SEC_TO_USEC / val;
+				}
+				if (val != qp->q_mode) {
+					reg_data.word0 = 0;
+					bf_set(lpfc_sliport_eqdelay_id,
+					       &reg_data, qp->queue_id);
+					bf_set(lpfc_sliport_eqdelay_delay,
+					       &reg_data, val);
+					writel(reg_data.word0, eqdreg);
+				}
+			} else {
+				/* Use mbox command method */
+				if (val != qp->q_mode)
+					lpfc_modify_hba_eq_delay(phba, i,
+								 1, val);
+			}
+
+			/*
+			 * val is cfg_fcp_imax or 0 for mbox delay or us delay
+			 * between interrupts for EQDR.
+			 */
+			qp->q_mode = val;
+			qp->EQ_cqe_cnt = 0;
+		}
+	}
+
+skip_eqdelay:
 	spin_lock_irq(&phba->pport->work_port_lock);
 
 	if (time_after(phba->last_completion_time +
@@ -7257,6 +7352,9 @@ lpfc_sli4_bar0_register_memmap(struct lpfc_hba *phba, uint32_t if_type)
 			phba->sli4_hba.conf_regs_memmap_p + LPFC_SLI_INTF;
 		break;
 	case LPFC_SLI_INTF_IF_TYPE_2:
+		phba->sli4_hba.u.if_type2.EQDregaddr =
+			phba->sli4_hba.conf_regs_memmap_p +
+						LPFC_CTL_PORT_EQ_DELAY_OFFSET;
 		phba->sli4_hba.u.if_type2.ERR1regaddr =
 			phba->sli4_hba.conf_regs_memmap_p +
 						LPFC_CTL_PORT_ER1_OFFSET;
@@ -8783,7 +8881,8 @@ lpfc_sli4_queue_setup(struct lpfc_hba *phba)
 	}
 
 	for (qidx = 0; qidx < io_channel; qidx += LPFC_MAX_EQ_DELAY_EQID_CNT)
-		lpfc_modify_hba_eq_delay(phba, qidx);
+		lpfc_modify_hba_eq_delay(phba, qidx, LPFC_MAX_EQ_DELAY_EQID_CNT,
+					 phba->cfg_fcp_imax);
 
 	return 0;
 
@@ -10252,6 +10351,9 @@ lpfc_get_sli4_parameters(struct lpfc_hba *phba, LPFC_MBOXQ_t *mboxq)
 	if (bf_get(cfg_xib, mbx_sli4_parameters) && phba->cfg_suppress_rsp)
 		phba->sli.sli_flag |= LPFC_SLI_SUPPRESS_RSP;
 
+	if (bf_get(cfg_eqdr, mbx_sli4_parameters))
+		phba->sli.sli_flag |= LPFC_SLI_USE_EQDR;
+
 	/* Make sure that sge_supp_len can be handled by the driver */
 	if (sli4_params->sge_supp_len > LPFC_MAX_SGE_SIZE)
 		sli4_params->sge_supp_len = LPFC_MAX_SGE_SIZE;
diff --git a/drivers/scsi/lpfc/lpfc_sli.c b/drivers/scsi/lpfc/lpfc_sli.c
index f60c9e3..040575a 100644
--- a/drivers/scsi/lpfc/lpfc_sli.c
+++ b/drivers/scsi/lpfc/lpfc_sli.c
@@ -13478,6 +13478,7 @@ process_cq:
 	/* Track the max number of CQEs processed in 1 EQ */
 	if (ecount > cq->CQ_max_cqe)
 		cq->CQ_max_cqe = ecount;
+	cq->assoc_qp->EQ_cqe_cnt += ecount;
 
 	/* Catch the no cq entry condition */
 	if (unlikely(ecount == 0))
@@ -13569,6 +13570,7 @@ lpfc_sli4_fof_handle_eqe(struct lpfc_hba *phba, struct lpfc_eqe *eqe)
 	/* Track the max number of CQEs processed in 1 EQ */
 	if (ecount > cq->CQ_max_cqe)
 		cq->CQ_max_cqe = ecount;
+	cq->assoc_qp->EQ_cqe_cnt += ecount;
 
 	/* Catch the no cq entry condition */
 	if (unlikely(ecount == 0))
@@ -13629,7 +13631,6 @@ lpfc_sli4_fof_intr_handler(int irq, void *dev_id)
 
 	/* Check device state for handling interrupt */
 	if (unlikely(lpfc_intr_state_check(phba))) {
-		eq->EQ_badstate++;
 		/* Check again for link_state with lock held */
 		spin_lock_irqsave(&phba->hbalock, iflag);
 		if (phba->link_state < LPFC_LINK_DOWN)
@@ -13741,7 +13742,6 @@ lpfc_sli4_hba_intr_handler(int irq, void *dev_id)
 
 	/* Check device state for handling interrupt */
 	if (unlikely(lpfc_intr_state_check(phba))) {
-		fpeq->EQ_badstate++;
 		/* Check again for link_state with lock held */
 		spin_lock_irqsave(&phba->hbalock, iflag);
 		if (phba->link_state < LPFC_LINK_DOWN)
@@ -14000,14 +14000,15 @@ lpfc_dual_chute_pci_bar_map(struct lpfc_hba *phba, uint16_t pci_barset)
  * fails this function will return -ENXIO.
  **/
 int
-lpfc_modify_hba_eq_delay(struct lpfc_hba *phba, uint32_t startq)
+lpfc_modify_hba_eq_delay(struct lpfc_hba *phba, uint32_t startq,
+			 uint32_t numq, uint32_t imax)
 {
 	struct lpfc_mbx_modify_eq_delay *eq_delay;
 	LPFC_MBOXQ_t *mbox;
 	struct lpfc_queue *eq;
 	int cnt, rc, length, status = 0;
 	uint32_t shdr_status, shdr_add_status;
-	uint32_t result;
+	uint32_t result, val;
 	int qidx;
 	union lpfc_sli4_cfg_shdr *shdr;
 	uint16_t dmult;
@@ -14026,22 +14027,45 @@ lpfc_modify_hba_eq_delay(struct lpfc_hba *phba, uint32_t startq)
 	eq_delay = &mbox->u.mqe.un.eq_delay;
 
 	/* Calculate delay multiper from maximum interrupt per second */
-	result = phba->cfg_fcp_imax / phba->io_channel_irqs;
+	result = imax / phba->io_channel_irqs;
 	if (result > LPFC_DMULT_CONST || result == 0)
 		dmult = 0;
 	else
 		dmult = LPFC_DMULT_CONST/result - 1;
+	if (dmult > LPFC_DMULT_MAX)
+		dmult = LPFC_DMULT_MAX;
 
 	cnt = 0;
 	for (qidx = startq; qidx < phba->io_channel_irqs; qidx++) {
 		eq = phba->sli4_hba.hba_eq[qidx];
 		if (!eq)
 			continue;
+		eq->q_mode = imax;
 		eq_delay->u.request.eq[cnt].eq_id = eq->queue_id;
 		eq_delay->u.request.eq[cnt].phase = 0;
 		eq_delay->u.request.eq[cnt].delay_multi = dmult;
 		cnt++;
-		if (cnt >= LPFC_MAX_EQ_DELAY_EQID_CNT)
+
+		/* q_mode is only used for auto_imax */
+		if (phba->sli.sli_flag & LPFC_SLI_USE_EQDR) {
+			/* Use EQ Delay Register method for q_mode */
+
+			/* Convert for EQ Delay register */
+			val =  phba->cfg_fcp_imax;
+			if (val) {
+				/* First, interrupts per sec per EQ */
+				val = phba->cfg_fcp_imax /
+					phba->io_channel_irqs;
+
+				/* us delay between each interrupt */
+				val = LPFC_SEC_TO_USEC / val;
+			}
+			eq->q_mode = val;
+		} else {
+			eq->q_mode = imax;
+		}
+
+		if (cnt >= numq)
 			break;
 	}
 	eq_delay->u.request.num_eq = cnt;
diff --git a/drivers/scsi/lpfc/lpfc_sli.h b/drivers/scsi/lpfc/lpfc_sli.h
index 9085306d..a3b1b51 100644
--- a/drivers/scsi/lpfc/lpfc_sli.h
+++ b/drivers/scsi/lpfc/lpfc_sli.h
@@ -321,6 +321,7 @@ struct lpfc_sli {
 #define LPFC_MENLO_MAINT          0x1000 /* need for menl fw download */
 #define LPFC_SLI_ASYNC_MBX_BLK    0x2000 /* Async mailbox is blocked */
 #define LPFC_SLI_SUPPRESS_RSP     0x4000 /* Suppress RSP feature is supported */
+#define LPFC_SLI_USE_EQDR         0x8000 /* EQ Delay Register is supported */
 
 	struct lpfc_sli_ring *sli3_ring;
 
diff --git a/drivers/scsi/lpfc/lpfc_sli4.h b/drivers/scsi/lpfc/lpfc_sli4.h
index 28b75e0..830dc83 100644
--- a/drivers/scsi/lpfc/lpfc_sli4.h
+++ b/drivers/scsi/lpfc/lpfc_sli4.h
@@ -168,7 +168,7 @@ struct lpfc_queue {
 	struct lpfc_sli_ring *pring; /* ptr to io ring associated with q */
 	struct lpfc_rqb *rqbp;	/* ptr to RQ buffers */
 
-	uint16_t sgl_list_cnt;
+	uint32_t q_mode;
 	uint16_t db_format;
 #define LPFC_DB_RING_FORMAT	0x01
 #define LPFC_DB_LIST_FORMAT	0x02
@@ -181,7 +181,7 @@ struct lpfc_queue {
 /* defines for EQ stats */
 #define	EQ_max_eqe		q_cnt_1
 #define	EQ_no_entry		q_cnt_2
-#define	EQ_badstate		q_cnt_3
+#define	EQ_cqe_cnt		q_cnt_3
 #define	EQ_processed		q_cnt_4
 
 /* defines for CQ stats */
@@ -523,6 +523,7 @@ struct lpfc_sli4_hba {
 #define SLIPORT_ERR2_REG_FAILURE_CQ		0x4
 #define SLIPORT_ERR2_REG_FAILURE_BUS		0x5
 #define SLIPORT_ERR2_REG_FAILURE_RQ		0x6
+			void __iomem *EQDregaddr;
 		} if_type2;
 	} u;
 
@@ -755,7 +756,8 @@ struct lpfc_queue *lpfc_sli4_queue_alloc(struct lpfc_hba *, uint32_t,
 			uint32_t);
 void lpfc_sli4_queue_free(struct lpfc_queue *);
 int lpfc_eq_create(struct lpfc_hba *, struct lpfc_queue *, uint32_t);
-int lpfc_modify_hba_eq_delay(struct lpfc_hba *phba, uint32_t startq);
+int lpfc_modify_hba_eq_delay(struct lpfc_hba *phba, uint32_t startq,
+			     uint32_t numq, uint32_t imax);
 int lpfc_cq_create(struct lpfc_hba *, struct lpfc_queue *,
 			struct lpfc_queue *, uint32_t, uint32_t);
 int lpfc_cq_create_set(struct lpfc_hba *phba, struct lpfc_queue **cqp,