scsi: lpfc: Change default IRQ model on AMD architectures

The current driver attempts to allocate an interrupt vector per cpu using
the systems managed IRQ allocator (flag PCI_IRQ_AFFINITY). The system IRQ
allocator will either provide the per-cpu vector, or return fewer
vectors. When fewer vectors, they are evenly spread between the numa nodes
on the system.  When run on an AMD architecture, if interrupts occur to a
cpu that is not in the same numa node as the adapter generating the
interrupt, there are extreme costs and overheads in performance.  Thus, if
1:1 vector allocation is used, or the "balanced" vectors in the other numa
nodes, performance can be hit significantly.

A much more performant model is to allocate interrupts only on the cpus
that are in the numa node where the adapter resides.  I/O completion is
still performed by the cpu where the I/O was generated. Unfortunately,
there is no flag to request the managed IRQ subsystem allocate vectors only
for the CPUs in the numa node as the adapter.

On AMD architecture, revert the irq allocation to the normal style
(non-managed) and then use irq_set_affinity_hint() to set the cpu
affinity and disable user-space rebalancing.

Tie the support into CPU offline/online. If the cpu being offlined owns a
vector, the vector is re-affinitized to one of the other CPUs on the same
numa node. If there are no more CPUs on the numa node, the vector has all
affinity removed and lets the system determine where it's serviced.
Similarly, when the cpu that owned a vector comes online, the vector is
reaffinitized to the cpu.

Link: https://lore.kernel.org/r/20191105005708.7399-10-jsmart2021@gmail.com
Signed-off-by: Dick Kennedy <dick.kennedy@broadcom.com>
Signed-off-by: James Smart <jsmart2021@gmail.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

1 files changed, 332 insertions, 118 deletions

diff --git a/drivers/scsi/lpfc/lpfc_init.c b/drivers/scsi/lpfc/lpfc_init.c
index 888ad32d5267..28e6a763f106 100644
--- a/drivers/scsi/lpfc/lpfc_init.c
+++ b/drivers/scsi/lpfc/lpfc_init.c
@@ -40,6 +40,7 @@
 #include <linux/irq.h>
 #include <linux/bitops.h>
 #include <linux/crash_dump.h>
+#include <linux/cpu.h>
 #include <linux/cpuhotplug.h>
 
 #include <scsi/scsi.h>
@@ -5996,6 +5997,35 @@ static void lpfc_log_intr_mode(struct lpfc_hba *phba, uint32_t intr_mode)
 }
 
 /**
+ * lpfc_cpumask_of_node_init - initalizes cpumask of phba's NUMA node
+ * @phba: Pointer to HBA context object.
+ *
+ **/
+static void
+lpfc_cpumask_of_node_init(struct lpfc_hba *phba)
+{
+	unsigned int cpu, numa_node;
+	struct cpumask *numa_mask = NULL;
+
+#ifdef CONFIG_NUMA
+	numa_node = phba->pcidev->dev.numa_node;
+#else
+	numa_node = NUMA_NO_NODE;
+#endif
+	numa_mask = &phba->sli4_hba.numa_mask;
+
+	cpumask_clear(numa_mask);
+
+	/* Check if we're a NUMA architecture */
+	if (!cpumask_of_node(numa_node))
+		return;
+
+	for_each_possible_cpu(cpu)
+		if (cpu_to_node(cpu) == numa_node)
+			cpumask_set_cpu(cpu, numa_mask);
+}
+
+/**
  * lpfc_enable_pci_dev - Enable a generic PCI device.
  * @phba: pointer to lpfc hba data structure.
  *
@@ -6438,6 +6468,7 @@ lpfc_sli4_driver_resource_setup(struct lpfc_hba *phba)
 	phba->sli4_hba.num_present_cpu = lpfc_present_cpu;
 	phba->sli4_hba.num_possible_cpu = num_possible_cpus();
 	phba->sli4_hba.curr_disp_cpu = 0;
+	lpfc_cpumask_of_node_init(phba);
 
 	/* Get all the module params for configuring this host */
 	lpfc_get_cfgparam(phba);
@@ -6973,6 +7004,7 @@ lpfc_sli4_driver_resource_unset(struct lpfc_hba *phba)
 	phba->sli4_hba.num_possible_cpu = 0;
 	phba->sli4_hba.num_present_cpu = 0;
 	phba->sli4_hba.curr_disp_cpu = 0;
+	cpumask_clear(&phba->sli4_hba.numa_mask);
 
 	/* Free memory allocated for fast-path work queue handles */
 	kfree(phba->sli4_hba.hba_eq_hdl);
@@ -10686,7 +10718,6 @@ lpfc_find_cpu_handle(struct lpfc_hba *phba, uint16_t id, int match)
 		 */
 		if ((match == LPFC_FIND_BY_EQ) &&
 		    (cpup->flag & LPFC_CPU_FIRST_IRQ) &&
-		    (cpup->irq != LPFC_VECTOR_MAP_EMPTY) &&
 		    (cpup->eq == id))
 			return cpu;
 
@@ -10724,6 +10755,75 @@ lpfc_find_hyper(struct lpfc_hba *phba, int cpu,
 }
 #endif
 
+/*
+ * lpfc_assign_eq_map_info - Assigns eq for vector_map structure
+ * @phba: pointer to lpfc hba data structure.
+ * @eqidx: index for eq and irq vector
+ * @flag: flags to set for vector_map structure
+ * @cpu: cpu used to index vector_map structure
+ *
+ * The routine assigns eq info into vector_map structure
+ */
+static inline void
+lpfc_assign_eq_map_info(struct lpfc_hba *phba, uint16_t eqidx, uint16_t flag,
+			unsigned int cpu)
+{
+	struct lpfc_vector_map_info *cpup = &phba->sli4_hba.cpu_map[cpu];
+	struct lpfc_hba_eq_hdl *eqhdl = lpfc_get_eq_hdl(eqidx);
+
+	cpup->eq = eqidx;
+	cpup->flag |= flag;
+
+	lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
+			"3336 Set Affinity: CPU %d irq %d eq %d flag x%x\n",
+			cpu, eqhdl->irq, cpup->eq, cpup->flag);
+}
+
+/**
+ * lpfc_cpu_map_array_init - Initialize cpu_map structure
+ * @phba: pointer to lpfc hba data structure.
+ *
+ * The routine initializes the cpu_map array structure
+ */
+static void
+lpfc_cpu_map_array_init(struct lpfc_hba *phba)
+{
+	struct lpfc_vector_map_info *cpup;
+	struct lpfc_eq_intr_info *eqi;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		cpup = &phba->sli4_hba.cpu_map[cpu];
+		cpup->phys_id = LPFC_VECTOR_MAP_EMPTY;
+		cpup->core_id = LPFC_VECTOR_MAP_EMPTY;
+		cpup->hdwq = LPFC_VECTOR_MAP_EMPTY;
+		cpup->eq = LPFC_VECTOR_MAP_EMPTY;
+		cpup->flag = 0;
+		eqi = per_cpu_ptr(phba->sli4_hba.eq_info, cpu);
+		INIT_LIST_HEAD(&eqi->list);
+		eqi->icnt = 0;
+	}
+}
+
+/**
+ * lpfc_hba_eq_hdl_array_init - Initialize hba_eq_hdl structure
+ * @phba: pointer to lpfc hba data structure.
+ *
+ * The routine initializes the hba_eq_hdl array structure
+ */
+static void
+lpfc_hba_eq_hdl_array_init(struct lpfc_hba *phba)
+{
+	struct lpfc_hba_eq_hdl *eqhdl;
+	int i;
+
+	for (i = 0; i < phba->cfg_irq_chann; i++) {
+		eqhdl = lpfc_get_eq_hdl(i);
+		eqhdl->irq = LPFC_VECTOR_MAP_EMPTY;
+		eqhdl->phba = phba;
+	}
+}
+
 /**
  * lpfc_cpu_affinity_check - Check vector CPU affinity mappings
  * @phba: pointer to lpfc hba data structure.
@@ -10742,22 +10842,10 @@ lpfc_cpu_affinity_check(struct lpfc_hba *phba, int vectors)
 	int max_core_id, min_core_id;
 	struct lpfc_vector_map_info *cpup;
 	struct lpfc_vector_map_info *new_cpup;
-	const struct cpumask *maskp;
 #ifdef CONFIG_X86
 	struct cpuinfo_x86 *cpuinfo;
 #endif
 
-	/* Init cpu_map array */
-	for_each_possible_cpu(cpu) {
-		cpup = &phba->sli4_hba.cpu_map[cpu];
-		cpup->phys_id = LPFC_VECTOR_MAP_EMPTY;
-		cpup->core_id = LPFC_VECTOR_MAP_EMPTY;
-		cpup->hdwq = LPFC_VECTOR_MAP_EMPTY;
-		cpup->eq = LPFC_VECTOR_MAP_EMPTY;
-		cpup->irq = LPFC_VECTOR_MAP_EMPTY;
-		cpup->flag = 0;
-	}
-
 	max_phys_id = 0;
 	min_phys_id = LPFC_VECTOR_MAP_EMPTY;
 	max_core_id = 0;
@@ -10793,65 +10881,6 @@ lpfc_cpu_affinity_check(struct lpfc_hba *phba, int vectors)
 			min_core_id = cpup->core_id;
 	}
 
-	for_each_possible_cpu(i) {
-		struct lpfc_eq_intr_info *eqi =
-			per_cpu_ptr(phba->sli4_hba.eq_info, i);
-
-		INIT_LIST_HEAD(&eqi->list);
-		eqi->icnt = 0;
-	}
-
-	/* This loop sets up all CPUs that are affinitized with a
-	 * irq vector assigned to the driver. All affinitized CPUs
-	 * will get a link to that vectors IRQ and EQ.
-	 *
-	 * NULL affinity mask handling:
-	 * If irq count is greater than one, log an error message.
-	 * If the null mask is received for the first irq, find the
-	 * first present cpu, and assign the eq index to ensure at
-	 * least one EQ is assigned.
-	 */
-	for (idx = 0; idx <  phba->cfg_irq_chann; idx++) {
-		/* Get a CPU mask for all CPUs affinitized to this vector */
-		maskp = pci_irq_get_affinity(phba->pcidev, idx);
-		if (!maskp) {
-			if (phba->cfg_irq_chann > 1)
-				lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
-						"3329 No affinity mask found "
-						"for vector %d (%d)\n",
-						idx, phba->cfg_irq_chann);
-			if (!idx) {
-				cpu = cpumask_first(cpu_present_mask);
-				cpup = &phba->sli4_hba.cpu_map[cpu];
-				cpup->eq = idx;
-				cpup->irq = pci_irq_vector(phba->pcidev, idx);
-				cpup->flag |= LPFC_CPU_FIRST_IRQ;
-			}
-			break;
-		}
-
-		i = 0;
-		/* Loop through all CPUs associated with vector idx */
-		for_each_cpu_and(cpu, maskp, cpu_present_mask) {
-			/* Set the EQ index and IRQ for that vector */
-			cpup = &phba->sli4_hba.cpu_map[cpu];
-			cpup->eq = idx;
-			cpup->irq = pci_irq_vector(phba->pcidev, idx);
-
-			/* If this is the first CPU thats assigned to this
-			 * vector, set LPFC_CPU_FIRST_IRQ.
-			 */
-			if (!i)
-				cpup->flag |= LPFC_CPU_FIRST_IRQ;
-			i++;
-
-			lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
-					"3336 Set Affinity: CPU %d "
-					"irq %d eq %d flag x%x\n",
-					cpu, cpup->irq, cpup->eq, cpup->flag);
-		}
-	}
-
 	/* After looking at each irq vector assigned to this pcidev, its
 	 * possible to see that not ALL CPUs have been accounted for.
 	 * Next we will set any unassigned (unaffinitized) cpu map
@@ -10877,7 +10906,7 @@ lpfc_cpu_affinity_check(struct lpfc_hba *phba, int vectors)
 			for (i = 0; i < phba->sli4_hba.num_present_cpu; i++) {
 				new_cpup = &phba->sli4_hba.cpu_map[new_cpu];
 				if (!(new_cpup->flag & LPFC_CPU_MAP_UNASSIGN) &&
-				    (new_cpup->irq != LPFC_VECTOR_MAP_EMPTY) &&
+				    (new_cpup->eq != LPFC_VECTOR_MAP_EMPTY) &&
 				    (new_cpup->phys_id == cpup->phys_id))
 					goto found_same;
 				new_cpu = cpumask_next(
@@ -10890,7 +10919,6 @@ lpfc_cpu_affinity_check(struct lpfc_hba *phba, int vectors)
 found_same:
 			/* We found a matching phys_id, so copy the IRQ info */
 			cpup->eq = new_cpup->eq;
-			cpup->irq = new_cpup->irq;
 
 			/* Bump start_cpu to the next slot to minmize the
 			 * chance of having multiple unassigned CPU entries
@@ -10902,9 +10930,10 @@ found_same:
 
 			lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
 					"3337 Set Affinity: CPU %d "
-					"irq %d from id %d same "
+					"eq %d from peer cpu %d same "
 					"phys_id (%d)\n",
-					cpu, cpup->irq, new_cpu, cpup->phys_id);
+					cpu, cpup->eq, new_cpu,
+					cpup->phys_id);
 		}
 	}
 
@@ -10928,7 +10957,7 @@ found_same:
 			for (i = 0; i < phba->sli4_hba.num_present_cpu; i++) {
 				new_cpup = &phba->sli4_hba.cpu_map[new_cpu];
 				if (!(new_cpup->flag & LPFC_CPU_MAP_UNASSIGN) &&
-				    (new_cpup->irq != LPFC_VECTOR_MAP_EMPTY))
+				    (new_cpup->eq != LPFC_VECTOR_MAP_EMPTY))
 					goto found_any;
 				new_cpu = cpumask_next(
 					new_cpu, cpu_present_mask);
@@ -10938,13 +10967,12 @@ found_same:
 			/* We should never leave an entry unassigned */
 			lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
 					"3339 Set Affinity: CPU %d "
-					"irq %d UNASSIGNED\n",
-					cpup->hdwq, cpup->irq);
+					"eq %d UNASSIGNED\n",
+					cpup->hdwq, cpup->eq);
 			continue;
 found_any:
 			/* We found an available entry, copy the IRQ info */
 			cpup->eq = new_cpup->eq;
-			cpup->irq = new_cpup->irq;
 
 			/* Bump start_cpu to the next slot to minmize the
 			 * chance of having multiple unassigned CPU entries
@@ -10956,8 +10984,8 @@ found_any:
 
 			lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
 					"3338 Set Affinity: CPU %d "
-					"irq %d from id %d (%d/%d)\n",
-					cpu, cpup->irq, new_cpu,
+					"eq %d from peer cpu %d (%d/%d)\n",
+					cpu, cpup->eq, new_cpu,
 					new_cpup->phys_id, new_cpup->core_id);
 		}
 	}
@@ -10978,9 +11006,9 @@ found_any:
 		idx++;
 		lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
 				"3333 Set Affinity: CPU %d (phys %d core %d): "
-				"hdwq %d eq %d irq %d flg x%x\n",
+				"hdwq %d eq %d flg x%x\n",
 				cpu, cpup->phys_id, cpup->core_id,
-				cpup->hdwq, cpup->eq, cpup->irq, cpup->flag);
+				cpup->hdwq, cpup->eq, cpup->flag);
 	}
 	/* Finally we need to associate a hdwq with each cpu_map entry
 	 * This will be 1 to 1 - hdwq to cpu, unless there are less
@@ -11056,9 +11084,9 @@ found_any:
  logit:
 		lpfc_printf_log(phba, KERN_ERR, LOG_INIT,
 				"3335 Set Affinity: CPU %d (phys %d core %d): "
-				"hdwq %d eq %d irq %d flg x%x\n",
+				"hdwq %d eq %d flg x%x\n",
 				cpu, cpup->phys_id, cpup->core_id,
-				cpup->hdwq, cpup->eq, cpup->irq, cpup->flag);
+				cpup->hdwq, cpup->eq, cpup->flag);
 	}
 
 	/* The cpu_map array will be used later during initialization
@@ -11078,10 +11106,8 @@ static void
 lpfc_cpuhp_get_eq(struct lpfc_hba *phba, unsigned int cpu,
 		  struct list_head *eqlist)
 {
-	struct lpfc_vector_map_info *map;
 	const struct cpumask *maskp;
 	struct lpfc_queue *eq;
-	unsigned int i;
 	cpumask_t tmp;
 	u16 idx;
 
@@ -11111,15 +11137,8 @@ lpfc_cpuhp_get_eq(struct lpfc_hba *phba, unsigned int cpu,
 		 * the software can share an eq, but eventually
 		 * only eq will be mapped to this vector
 		 */
-		for_each_possible_cpu(i) {
-			map = &phba->sli4_hba.cpu_map[i];
-			if (!(map->irq == pci_irq_vector(phba->pcidev, idx)))
-				continue;
-			eq = phba->sli4_hba.hdwq[map->hdwq].hba_eq;
-			list_add(&eq->_poll_list, eqlist);
-			/* 1 is good enough. others will be a copy of this */
-			break;
-		}
+		eq = phba->sli4_hba.hba_eq_hdl[idx].eq;
+		list_add(&eq->_poll_list, eqlist);
 	}
 }
 
@@ -11181,6 +11200,99 @@ static int __lpfc_cpuhp_checks(struct lpfc_hba *phba, int *retval)
 	return false;
 }
 
+/**
+ * lpfc_irq_set_aff - set IRQ affinity
+ * @eqhdl: EQ handle
+ * @cpu: cpu to set affinity
+ *
+ **/
+static inline void
+lpfc_irq_set_aff(struct lpfc_hba_eq_hdl *eqhdl, unsigned int cpu)
+{
+	cpumask_clear(&eqhdl->aff_mask);
+	cpumask_set_cpu(cpu, &eqhdl->aff_mask);
+	irq_set_status_flags(eqhdl->irq, IRQ_NO_BALANCING);
+	irq_set_affinity_hint(eqhdl->irq, &eqhdl->aff_mask);
+}
+
+/**
+ * lpfc_irq_clear_aff - clear IRQ affinity
+ * @eqhdl: EQ handle
+ *
+ **/
+static inline void
+lpfc_irq_clear_aff(struct lpfc_hba_eq_hdl *eqhdl)
+{
+	cpumask_clear(&eqhdl->aff_mask);
+	irq_clear_status_flags(eqhdl->irq, IRQ_NO_BALANCING);
+	irq_set_affinity_hint(eqhdl->irq, &eqhdl->aff_mask);
+}
+
+/**
+ * lpfc_irq_rebalance - rebalances IRQ affinity according to cpuhp event
+ * @phba: pointer to HBA context object.
+ * @cpu: cpu going offline/online
+ * @offline: true, cpu is going offline. false, cpu is coming online.
+ *
+ * If cpu is going offline, we'll try our best effort to find the next
+ * online cpu on the phba's NUMA node and migrate all offlining IRQ affinities.
+ *
+ * If cpu is coming online, reaffinitize the IRQ back to the onlineng cpu.
+ *
+ * Note: Call only if cfg_irq_numa is enabled, otherwise rely on
+ *	 PCI_IRQ_AFFINITY to auto-manage IRQ affinity.
+ *
+ **/
+static void
+lpfc_irq_rebalance(struct lpfc_hba *phba, unsigned int cpu, bool offline)
+{
+	struct lpfc_vector_map_info *cpup;
+	struct cpumask *aff_mask;
+	unsigned int cpu_select, cpu_next, idx;
+	const struct cpumask *numa_mask;
+
+	if (!phba->cfg_irq_numa)
+		return;
+
+	numa_mask = &phba->sli4_hba.numa_mask;
+
+	if (!cpumask_test_cpu(cpu, numa_mask))
+		return;
+
+	cpup = &phba->sli4_hba.cpu_map[cpu];
+
+	if (!(cpup->flag & LPFC_CPU_FIRST_IRQ))
+		return;
+
+	if (offline) {
+		/* Find next online CPU on NUMA node */
+		cpu_next = cpumask_next_wrap(cpu, numa_mask, cpu, true);
+		cpu_select = lpfc_next_online_numa_cpu(numa_mask, cpu_next);
+
+		/* Found a valid CPU */
+		if ((cpu_select < nr_cpu_ids) && (cpu_select != cpu)) {
+			/* Go through each eqhdl and ensure offlining
+			 * cpu aff_mask is migrated
+			 */
+			for (idx = 0; idx < phba->cfg_irq_chann; idx++) {
+				aff_mask = lpfc_get_aff_mask(idx);
+
+				/* Migrate affinity */
+				if (cpumask_test_cpu(cpu, aff_mask))
+					lpfc_irq_set_aff(lpfc_get_eq_hdl(idx),
+							 cpu_select);
+			}
+		} else {
+			/* Rely on irqbalance if no online CPUs left on NUMA */
+			for (idx = 0; idx < phba->cfg_irq_chann; idx++)
+				lpfc_irq_clear_aff(lpfc_get_eq_hdl(idx));
+		}
+	} else {
+		/* Migrate affinity back to this CPU */
+		lpfc_irq_set_aff(lpfc_get_eq_hdl(cpup->eq), cpu);
+	}
+}
+
 static int lpfc_cpu_offline(unsigned int cpu, struct hlist_node *node)
 {
 	struct lpfc_hba *phba = hlist_entry_safe(node, struct lpfc_hba, cpuhp);
@@ -11196,6 +11308,8 @@ static int lpfc_cpu_offline(unsigned int cpu, struct hlist_node *node)
 	if (__lpfc_cpuhp_checks(phba, &retval))
 		return retval;
 
+	lpfc_irq_rebalance(phba, cpu, true);
+
 	lpfc_cpuhp_get_eq(phba, cpu, &eqlist);
 
 	/* start polling on these eq's */
@@ -11222,6 +11336,8 @@ static int lpfc_cpu_online(unsigned int cpu, struct hlist_node *node)
 	if (__lpfc_cpuhp_checks(phba, &retval))
 		return retval;
 
+	lpfc_irq_rebalance(phba, cpu, false);
+
 	list_for_each_entry_safe(eq, next, &phba->poll_list, _poll_list) {
 		n = lpfc_find_cpu_handle(phba, eq->hdwq, LPFC_FIND_BY_HDWQ);
 		if (n == cpu)
@@ -11236,7 +11352,24 @@ static int lpfc_cpu_online(unsigned int cpu, struct hlist_node *node)
  * @phba: pointer to lpfc hba data structure.
  *
  * This routine is invoked to enable the MSI-X interrupt vectors to device
- * with SLI-4 interface spec.
+ * with SLI-4 interface spec.  It also allocates MSI-X vectors and maps them
+ * to cpus on the system.
+ *
+ * When cfg_irq_numa is enabled, the adapter will only allocate vectors for
+ * the number of cpus on the same numa node as this adapter.  The vectors are
+ * allocated without requesting OS affinity mapping.  A vector will be
+ * allocated and assigned to each online and offline cpu.  If the cpu is
+ * online, then affinity will be set to that cpu.  If the cpu is offline, then
+ * affinity will be set to the nearest peer cpu within the numa node that is
+ * online.  If there are no online cpus within the numa node, affinity is not
+ * assigned and the OS may do as it pleases. Note: cpu vector affinity mapping
+ * is consistent with the way cpu online/offline is handled when cfg_irq_numa is
+ * configured.
+ *
+ * If numa mode is not enabled and there is more than 1 vector allocated, then
+ * the driver relies on the managed irq interface where the OS assigns vector to
+ * cpu affinity.  The driver will then use that affinity mapping to setup its
+ * cpu mapping table.
  *
  * Return codes
  * 0 - successful
@@ -11247,13 +11380,31 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba)
 {
 	int vectors, rc, index;
 	char *name;
+	const struct cpumask *numa_mask = NULL;
+	unsigned int cpu = 0, cpu_cnt = 0, cpu_select = nr_cpu_ids;
+	struct lpfc_hba_eq_hdl *eqhdl;
+	const struct cpumask *maskp;
+	bool first;
+	unsigned int flags = PCI_IRQ_MSIX;
 
 	/* Set up MSI-X multi-message vectors */
 	vectors = phba->cfg_irq_chann;
 
-	rc = pci_alloc_irq_vectors(phba->pcidev,
-				1,
-				vectors, PCI_IRQ_MSIX | PCI_IRQ_AFFINITY);
+	if (phba->cfg_irq_numa) {
+		numa_mask = &phba->sli4_hba.numa_mask;
+		cpu_cnt = cpumask_weight(numa_mask);
+		vectors = min(phba->cfg_irq_chann, cpu_cnt);
+
+		/* cpu: iterates over numa_mask including offline or online
+		 * cpu_select: iterates over online numa_mask to set affinity
+		 */
+		cpu = cpumask_first(numa_mask);
+		cpu_select = lpfc_next_online_numa_cpu(numa_mask, cpu);
+	} else {
+		flags |= PCI_IRQ_AFFINITY;
+	}
+
+	rc = pci_alloc_irq_vectors(phba->pcidev, 1, vectors, flags);
 	if (rc < 0) {
 		lpfc_printf_log(phba, KERN_INFO, LOG_INIT,
 				"0484 PCI enable MSI-X failed (%d)\n", rc);
@@ -11263,23 +11414,61 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba)
 
 	/* Assign MSI-X vectors to interrupt handlers */
 	for (index = 0; index < vectors; index++) {
-		name = phba->sli4_hba.hba_eq_hdl[index].handler_name;
+		eqhdl = lpfc_get_eq_hdl(index);
+		name = eqhdl->handler_name;
 		memset(name, 0, LPFC_SLI4_HANDLER_NAME_SZ);
 		snprintf(name, LPFC_SLI4_HANDLER_NAME_SZ,
 			 LPFC_DRIVER_HANDLER_NAME"%d", index);
 
-		phba->sli4_hba.hba_eq_hdl[index].idx = index;
-		phba->sli4_hba.hba_eq_hdl[index].phba = phba;
+		eqhdl->idx = index;
 		rc = request_irq(pci_irq_vector(phba->pcidev, index),
 			 &lpfc_sli4_hba_intr_handler, 0,
-			 name,
-			 &phba->sli4_hba.hba_eq_hdl[index]);
+			 name, eqhdl);
 		if (rc) {
 			lpfc_printf_log(phba, KERN_WARNING, LOG_INIT,
 					"0486 MSI-X fast-path (%d) "
 					"request_irq failed (%d)\n", index, rc);
 			goto cfg_fail_out;
 		}
+
+		eqhdl->irq = pci_irq_vector(phba->pcidev, index);
+
+		if (phba->cfg_irq_numa) {
+			/* If found a neighboring online cpu, set affinity */
+			if (cpu_select < nr_cpu_ids)
+				lpfc_irq_set_aff(eqhdl, cpu_select);
+
+			/* Assign EQ to cpu_map */
+			lpfc_assign_eq_map_info(phba, index,
+						LPFC_CPU_FIRST_IRQ,
+						cpu);
+
+			/* Iterate to next offline or online cpu in numa_mask */
+			cpu = cpumask_next(cpu, numa_mask);
+
+			/* Find next online cpu in numa_mask to set affinity */
+			cpu_select = lpfc_next_online_numa_cpu(numa_mask, cpu);
+		} else if (vectors == 1) {
+			cpu = cpumask_first(cpu_present_mask);
+			lpfc_assign_eq_map_info(phba, index, LPFC_CPU_FIRST_IRQ,
+						cpu);
+		} else {
+			maskp = pci_irq_get_affinity(phba->pcidev, index);
+
+			first = true;
+			/* Loop through all CPUs associated with vector index */
+			for_each_cpu_and(cpu, maskp, cpu_present_mask) {
+				/* If this is the first CPU thats assigned to
+				 * this vector, set LPFC_CPU_FIRST_IRQ.
+				 */
+				lpfc_assign_eq_map_info(phba, index,
+							first ?
+							LPFC_CPU_FIRST_IRQ : 0,
+							cpu);
+				if (first)
+					first = false;
+			}
+		}
 	}
 
 	if (vectors != phba->cfg_irq_chann) {
@@ -11295,9 +11484,12 @@ lpfc_sli4_enable_msix(struct lpfc_hba *phba)
 
 cfg_fail_out:
 	/* free the irq already requested */
-	for (--index; index >= 0; index--)
-		free_irq(pci_irq_vector(phba->pcidev, index),
-				&phba->sli4_hba.hba_eq_hdl[index]);
+	for (--index; index >= 0; index--) {
+		eqhdl = lpfc_get_eq_hdl(index);
+		lpfc_irq_clear_aff(eqhdl);
+		irq_set_affinity_hint(eqhdl->irq, NULL);
+		free_irq(eqhdl->irq, eqhdl);
+	}
 
 	/* Unconfigure MSI-X capability structure */
 	pci_free_irq_vectors(phba->pcidev);
@@ -11324,6 +11516,8 @@ static int
 lpfc_sli4_enable_msi(struct lpfc_hba *phba)
 {
 	int rc, index;
+	unsigned int cpu;
+	struct lpfc_hba_eq_hdl *eqhdl;
 
 	rc = pci_alloc_irq_vectors(phba->pcidev, 1, 1,
 				   PCI_IRQ_MSI | PCI_IRQ_AFFINITY);
@@ -11345,9 +11539,15 @@ lpfc_sli4_enable_msi(struct lpfc_hba *phba)
 		return rc;
 	}
 
+	eqhdl = lpfc_get_eq_hdl(0);
+	eqhdl->irq = pci_irq_vector(phba->pcidev, 0);
+
+	cpu = cpumask_first(cpu_present_mask);
+	lpfc_assign_eq_map_info(phba, 0, LPFC_CPU_FIRST_IRQ, cpu);
+
 	for (index = 0; index < phba->cfg_irq_chann; index++) {
-		phba->sli4_hba.hba_eq_hdl[index].idx = index;
-		phba->sli4_hba.hba_eq_hdl[index].phba = phba;
+		eqhdl = lpfc_get_eq_hdl(index);
+		eqhdl->idx = index;
 	}
 
 	return 0;
@@ -11380,7 +11580,9 @@ lpfc_sli4_enable_intr(struct lpfc_hba *phba, uint32_t cfg_mode)
 		retval = 0;
 		if (!retval) {
 			/* Now, try to enable MSI-X interrupt mode */
+			get_online_cpus();
 			retval = lpfc_sli4_enable_msix(phba);
+			put_online_cpus();
 			if (!retval) {
 				/* Indicate initialization to MSI-X mode */
 				phba->intr_type = MSIX;
@@ -11405,15 +11607,21 @@ lpfc_sli4_enable_intr(struct lpfc_hba *phba, uint32_t cfg_mode)
 				     IRQF_SHARED, LPFC_DRIVER_NAME, phba);
 		if (!retval) {
 			struct lpfc_hba_eq_hdl *eqhdl;
+			unsigned int cpu;
 
 			/* Indicate initialization to INTx mode */
 			phba->intr_type = INTx;
 			intr_mode = 0;
 
+			eqhdl = lpfc_get_eq_hdl(0);
+			eqhdl->irq = pci_irq_vector(phba->pcidev, 0);
+
+			cpu = cpumask_first(cpu_present_mask);
+			lpfc_assign_eq_map_info(phba, 0, LPFC_CPU_FIRST_IRQ,
+						cpu);
 			for (idx = 0; idx < phba->cfg_irq_chann; idx++) {
-				eqhdl = &phba->sli4_hba.hba_eq_hdl[idx];
+				eqhdl = lpfc_get_eq_hdl(idx);
 				eqhdl->idx = idx;
-				eqhdl->phba = phba;
 			}
 		}
 	}
@@ -11435,14 +11643,14 @@ lpfc_sli4_disable_intr(struct lpfc_hba *phba)
 	/* Disable the currently initialized interrupt mode */
 	if (phba->intr_type == MSIX) {
 		int index;
+		struct lpfc_hba_eq_hdl *eqhdl;
 
 		/* Free up MSI-X multi-message vectors */
 		for (index = 0; index < phba->cfg_irq_chann; index++) {
-			irq_set_affinity_hint(
-				pci_irq_vector(phba->pcidev, index),
-				NULL);
-			free_irq(pci_irq_vector(phba->pcidev, index),
-					&phba->sli4_hba.hba_eq_hdl[index]);
+			eqhdl = lpfc_get_eq_hdl(index);
+			lpfc_irq_clear_aff(eqhdl);
+			irq_set_affinity_hint(eqhdl->irq, NULL);
+			free_irq(eqhdl->irq, eqhdl);
 		}
 	} else {
 		free_irq(phba->pcidev->irq, phba);
@@ -12848,6 +13056,12 @@ lpfc_pci_probe_one_s4(struct pci_dev *pdev, const struct pci_device_id *pid)
 	phba->pport = NULL;
 	lpfc_stop_port(phba);
 
+	/* Init cpu_map array */
+	lpfc_cpu_map_array_init(phba);
+
+	/* Init hba_eq_hdl array */
+	lpfc_hba_eq_hdl_array_init(phba);
+
 	/* Configure and enable interrupt */
 	intr_mode = lpfc_sli4_enable_intr(phba, cfg_mode);
 	if (intr_mode == LPFC_INTR_ERROR) {