scsi: qla2xxx: Select qpair depending on which CPU post_cmd() gets called

In current I/O path, Tx and Rx may not be processed on same CPU. This may
lead to thrashing and optimum performance may not be achieved.

Pick qpair such that Tx and Rx are processed on same CPU.
Signed-off-by: Shreyas Deodhar <sdeodhar@marvell.com>
Signed-off-by: Nilesh Javali <njavali@marvell.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

drivers/scsi/qla2xxx/qla_def.h

@@ -3469,6 +3469,7 @@ struct qla_msix_entry {
 	int have_irq;
 	int in_use;
 	uint32_t vector;
+	uint32_t vector_base0;
 	uint16_t entry;
 	char name[30];
 	void *handle;
@@ -4125,6 +4126,7 @@ struct qla_hw_data {
 	struct req_que **req_q_map;
 	struct rsp_que **rsp_q_map;
 	struct qla_qpair **queue_pair_map;
+	struct qla_qpair **qp_cpu_map;
 	unsigned long req_qid_map[(QLA_MAX_QUEUES / 8) / sizeof(unsigned long)];
 	unsigned long rsp_qid_map[(QLA_MAX_QUEUES / 8) / sizeof(unsigned long)];
 	unsigned long qpair_qid_map[(QLA_MAX_QUEUES / 8)

drivers/scsi/qla2xxx/qla_init.c

@@ -9426,8 +9426,6 @@ struct qla_qpair *qla2xxx_create_qpair(struct scsi_qla_host *vha, int qos,
 		qpair->req = ha->req_q_map[req_id];
 		qpair->rsp->req = qpair->req;
 		qpair->rsp->qpair = qpair;
-		/* init qpair to this cpu. Will adjust at run time. */
-		qla_cpu_update(qpair, raw_smp_processor_id());
 
 		if (IS_T10_PI_CAPABLE(ha) && ql2xenabledif) {
 			if (ha->fw_attributes & BIT_4)

drivers/scsi/qla2xxx/qla_inline.h

@@ -515,3 +515,58 @@ fcport_is_bigger(fc_port_t *fcport)
 {
 	return !fcport_is_smaller(fcport);
 }
+
+static inline struct qla_qpair *
+qla_mapq_nvme_select_qpair(struct qla_hw_data *ha, struct qla_qpair *qpair)
+{
+	int cpuid = smp_processor_id();
+
+	if (qpair->cpuid != cpuid &&
+	    ha->qp_cpu_map[cpuid]) {
+		qpair = ha->qp_cpu_map[cpuid];
+	}
+	return qpair;
+}
+
+static inline void
+qla_mapq_init_qp_cpu_map(struct qla_hw_data *ha,
+			 struct qla_msix_entry *msix,
+			 struct qla_qpair *qpair)
+{
+	const struct cpumask *mask;
+	unsigned int cpu;
+
+	if (!ha->qp_cpu_map)
+		return;
+	mask = pci_irq_get_affinity(ha->pdev, msix->vector_base0);
+	qpair->cpuid = cpumask_first(mask);
+	for_each_cpu(cpu, mask) {
+		ha->qp_cpu_map[cpu] = qpair;
+	}
+	msix->cpuid = qpair->cpuid;
+}
+
+static inline void
+qla_mapq_free_qp_cpu_map(struct qla_hw_data *ha)
+{
+	if (ha->qp_cpu_map) {
+		kfree(ha->qp_cpu_map);
+		ha->qp_cpu_map = NULL;
+	}
+}
+
+static inline int qla_mapq_alloc_qp_cpu_map(struct qla_hw_data *ha)
+{
+	scsi_qla_host_t *vha = pci_get_drvdata(ha->pdev);
+
+	if (!ha->qp_cpu_map) {
+		ha->qp_cpu_map = kcalloc(NR_CPUS, sizeof(struct qla_qpair *),
+					 GFP_KERNEL);
+		if (!ha->qp_cpu_map) {
+			ql_log(ql_log_fatal, vha, 0x0180,
+			       "Unable to allocate memory for qp_cpu_map ptrs.\n");
+			return -1;
+		}
+	}
+	return 0;
+}

drivers/scsi/qla2xxx/qla_isr.c

@@ -3769,7 +3769,6 @@ void qla24xx_process_response_queue(struct scsi_qla_host *vha,
 
 	if (rsp->qpair->cpuid != smp_processor_id() || !rsp->qpair->rcv_intr) {
 		rsp->qpair->rcv_intr = 1;
-		qla_cpu_update(rsp->qpair, smp_processor_id());
 	}
 
 #define __update_rsp_in(_is_shadow_hba, _rsp, _rsp_in)			\
@@ -4377,6 +4376,7 @@ qla24xx_enable_msix(struct qla_hw_data *ha, struct rsp_que *rsp)
 	for (i = 0; i < ha->msix_count; i++) {
 		qentry = &ha->msix_entries[i];
 		qentry->vector = pci_irq_vector(ha->pdev, i);
+		qentry->vector_base0 = i;
 		qentry->entry = i;
 		qentry->have_irq = 0;
 		qentry->in_use = 0;
@@ -4604,5 +4604,6 @@ int qla25xx_request_irq(struct qla_hw_data *ha, struct qla_qpair *qpair,
 	}
 	msix->have_irq = 1;
 	msix->handle = qpair;
+	qla_mapq_init_qp_cpu_map(ha, msix, qpair);
 	return ret;
 }

drivers/scsi/qla2xxx/qla_nvme.c

@@ -609,6 +609,7 @@ static int qla_nvme_post_cmd(struct nvme_fc_local_port *lport,
 	fc_port_t *fcport;
 	struct srb_iocb *nvme;
 	struct scsi_qla_host *vha;
+	struct qla_hw_data *ha;
 	int rval;
 	srb_t *sp;
 	struct qla_qpair *qpair = hw_queue_handle;
@@ -629,6 +630,7 @@ static int qla_nvme_post_cmd(struct nvme_fc_local_port *lport,
 		return -ENODEV;
 
 	vha = fcport->vha;
+	ha = vha->hw;
 
 	if (test_bit(ABORT_ISP_ACTIVE, &vha->dpc_flags))
 		return -EBUSY;
@@ -643,6 +645,8 @@ static int qla_nvme_post_cmd(struct nvme_fc_local_port *lport,
 	if (fcport->nvme_flag & NVME_FLAG_RESETTING)
 		return -EBUSY;
 
+	qpair = qla_mapq_nvme_select_qpair(ha, qpair);
+
 	/* Alloc SRB structure */
 	sp = qla2xxx_get_qpair_sp(vha, qpair, fcport, GFP_ATOMIC);
 	if (!sp)

drivers/scsi/qla2xxx/qla_os.c

@@ -472,6 +472,11 @@ static int qla2x00_alloc_queues(struct qla_hw_data *ha, struct req_que *req,
 			    "Unable to allocate memory for queue pair ptrs.\n");
 			goto fail_qpair_map;
 		}
+		if (qla_mapq_alloc_qp_cpu_map(ha) != 0) {
+			kfree(ha->queue_pair_map);
+			ha->queue_pair_map = NULL;
+			goto fail_qpair_map;
+		}
 	}
 
 	/*
@@ -546,6 +551,7 @@ static void qla2x00_free_queues(struct qla_hw_data *ha)
 		ha->base_qpair = NULL;
 	}
 
+	qla_mapq_free_qp_cpu_map(ha);
 	spin_lock_irqsave(&ha->hardware_lock, flags);
 	for (cnt = 0; cnt < ha->max_req_queues; cnt++) {
 		if (!test_bit(cnt, ha->req_qid_map))