scsi: lpfc: Add registration for CPU Offline/Online events

The recent affinitization didn't address cpu offlining/onlining.  If an
interrupt vector is shared and the low order cpu owning the vector is
offlined, as interrupts are managed, the vector is taken offline. This
causes the other CPUs sharing the vector will hang as they can't get io
completions.

Correct by registering callbacks with the system for Offline/Online
events. When a cpu is taken offline, its eq, which is tied to an interrupt
vector is found. If the cpu is the "owner" of the vector and if the
eq/vector is shared by other CPUs, the eq is placed into a polled mode.
Additionally, code paths that perform io submission on the "sharing CPUs"
will check the eq state and poll for completion after submission of new io
to a wq that uses the eq.

Similarly, when a cpu comes back online and owns an offlined vector, the eq
is taken out of polled mode and rearmed to start driving interrupts for eq.

Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Dick Kennedy <[email protected]>
Signed-off-by: James Smart <[email protected]>
Signed-off-by: Martin K. Petersen <[email protected]>

drivers/scsi/lpfc/lpfc.h

@@ -1215,6 +1215,13 @@ struct lpfc_hba {
 	uint64_t ktime_seg10_min;
 	uint64_t ktime_seg10_max;
 #endif
+
+	struct hlist_node cpuhp;	/* used for cpuhp per hba callback */
+	struct timer_list cpuhp_poll_timer;
+	struct list_head poll_list;	/* slowpath eq polling list */
+#define LPFC_POLL_HB	1		/* slowpath heartbeat */
+#define LPFC_POLL_FASTPATH	0	/* called from fastpath */
+#define LPFC_POLL_SLOWPATH	1	/* called from slowpath */
 };
 
 static inline struct Scsi_Host *

drivers/scsi/lpfc/lpfc_crtn.h

@@ -215,6 +215,12 @@ irqreturn_t lpfc_sli_fp_intr_handler(int, void *);
 irqreturn_t lpfc_sli4_intr_handler(int, void *);
 irqreturn_t lpfc_sli4_hba_intr_handler(int, void *);
 
+inline void lpfc_sli4_cleanup_poll_list(struct lpfc_hba *phba);
+int lpfc_sli4_poll_eq(struct lpfc_queue *q, uint8_t path);
+void lpfc_sli4_poll_hbtimer(struct timer_list *t);
+void lpfc_sli4_start_polling(struct lpfc_queue *q);
+void lpfc_sli4_stop_polling(struct lpfc_queue *q);
+
 void lpfc_read_rev(struct lpfc_hba *, LPFC_MBOXQ_t *);
 void lpfc_sli4_swap_str(struct lpfc_hba *, LPFC_MBOXQ_t *);
 void lpfc_config_ring(struct lpfc_hba *, int, LPFC_MBOXQ_t *);

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/cpuhotplug.h>
 
 #include <scsi/scsi.h>
 #include <scsi/scsi_device.h>
@@ -66,9 +67,13 @@
 #include "lpfc_version.h"
 #include "lpfc_ids.h"
 
+static enum cpuhp_state lpfc_cpuhp_state;
 /* Used when mapping IRQ vectors in a driver centric manner */
 static uint32_t lpfc_present_cpu;
 
+static void __lpfc_cpuhp_remove(struct lpfc_hba *phba);
+static void lpfc_cpuhp_remove(struct lpfc_hba *phba);
+static void lpfc_cpuhp_add(struct lpfc_hba *phba);
 static void lpfc_get_hba_model_desc(struct lpfc_hba *, uint8_t *, uint8_t *);
 static int lpfc_post_rcv_buf(struct lpfc_hba *);
 static int lpfc_sli4_queue_verify(struct lpfc_hba *);
@@ -3379,6 +3384,8 @@ lpfc_online(struct lpfc_hba *phba)
 	if (phba->cfg_xri_rebalancing)
 		lpfc_create_multixri_pools(phba);
 
+	lpfc_cpuhp_add(phba);
+
 	lpfc_unblock_mgmt_io(phba);
 	return 0;
 }
@@ -3542,6 +3549,7 @@ lpfc_offline(struct lpfc_hba *phba)
 			spin_unlock_irq(shost->host_lock);
 		}
 	lpfc_destroy_vport_work_array(phba, vports);
+	__lpfc_cpuhp_remove(phba);
 
 	if (phba->cfg_xri_rebalancing)
 		lpfc_destroy_multixri_pools(phba);
@@ -9255,6 +9263,8 @@ lpfc_sli4_queue_destroy(struct lpfc_hba *phba)
 	}
 	spin_unlock_irq(&phba->hbalock);
 
+	lpfc_sli4_cleanup_poll_list(phba);
+
 	/* Release HBA eqs */
 	if (phba->sli4_hba.hdwq)
 		lpfc_sli4_release_hdwq(phba);
@@ -11057,6 +11067,170 @@ lpfc_cpu_affinity_check(struct lpfc_hba *phba, int vectors)
 	return;
 }
 
+/**
+ * lpfc_cpuhp_get_eq
+ *
+ * @phba:   pointer to lpfc hba data structure.
+ * @cpu:    cpu going offline
+ * @eqlist:
+ */
+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;
+
+	for (idx = 0; idx < phba->cfg_irq_chann; idx++) {
+		maskp = pci_irq_get_affinity(phba->pcidev, idx);
+		if (!maskp)
+			continue;
+		/*
+		 * if irq is not affinitized to the cpu going
+		 * then we don't need to poll the eq attached
+		 * to it.
+		 */
+		if (!cpumask_and(&tmp, maskp, cpumask_of(cpu)))
+			continue;
+		/* get the cpus that are online and are affini-
+		 * tized to this irq vector.  If the count is
+		 * more than 1 then cpuhp is not going to shut-
+		 * down this vector.  Since this cpu has not
+		 * gone offline yet, we need >1.
+		 */
+		cpumask_and(&tmp, maskp, cpu_online_mask);
+		if (cpumask_weight(&tmp) > 1)
+			continue;
+
+		/* Now that we have an irq to shutdown, get the eq
+		 * mapped to this irq.  Note: multiple hdwq's in
+		 * 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;
+		}
+	}
+}
+
+static void __lpfc_cpuhp_remove(struct lpfc_hba *phba)
+{
+	if (phba->sli_rev != LPFC_SLI_REV4)
+		return;
+
+	cpuhp_state_remove_instance_nocalls(lpfc_cpuhp_state,
+					    &phba->cpuhp);
+	/*
+	 * unregistering the instance doesn't stop the polling
+	 * timer. Wait for the poll timer to retire.
+	 */
+	synchronize_rcu();
+	del_timer_sync(&phba->cpuhp_poll_timer);
+}
+
+static void lpfc_cpuhp_remove(struct lpfc_hba *phba)
+{
+	if (phba->pport->fc_flag & FC_OFFLINE_MODE)
+		return;
+
+	__lpfc_cpuhp_remove(phba);
+}
+
+static void lpfc_cpuhp_add(struct lpfc_hba *phba)
+{
+	if (phba->sli_rev != LPFC_SLI_REV4)
+		return;
+
+	rcu_read_lock();
+
+	if (!list_empty(&phba->poll_list)) {
+		timer_setup(&phba->cpuhp_poll_timer, lpfc_sli4_poll_hbtimer, 0);
+		mod_timer(&phba->cpuhp_poll_timer,
+			  jiffies + msecs_to_jiffies(LPFC_POLL_HB));
+	}
+
+	rcu_read_unlock();
+
+	cpuhp_state_add_instance_nocalls(lpfc_cpuhp_state,
+					 &phba->cpuhp);
+}
+
+static int __lpfc_cpuhp_checks(struct lpfc_hba *phba, int *retval)
+{
+	if (phba->pport->load_flag & FC_UNLOADING) {
+		*retval = -EAGAIN;
+		return true;
+	}
+
+	if (phba->sli_rev != LPFC_SLI_REV4) {
+		*retval = 0;
+		return true;
+	}
+
+	/* proceed with the hotplug */
+	return false;
+}
+
+static int lpfc_cpu_offline(unsigned int cpu, struct hlist_node *node)
+{
+	struct lpfc_hba *phba = hlist_entry_safe(node, struct lpfc_hba, cpuhp);
+	struct lpfc_queue *eq, *next;
+	LIST_HEAD(eqlist);
+	int retval;
+
+	if (!phba) {
+		WARN_ONCE(!phba, "cpu: %u. phba:NULL", raw_smp_processor_id());
+		return 0;
+	}
+
+	if (__lpfc_cpuhp_checks(phba, &retval))
+		return retval;
+
+	lpfc_cpuhp_get_eq(phba, cpu, &eqlist);
+
+	/* start polling on these eq's */
+	list_for_each_entry_safe(eq, next, &eqlist, _poll_list) {
+		list_del_init(&eq->_poll_list);
+		lpfc_sli4_start_polling(eq);
+	}
+
+	return 0;
+}
+
+static int lpfc_cpu_online(unsigned int cpu, struct hlist_node *node)
+{
+	struct lpfc_hba *phba = hlist_entry_safe(node, struct lpfc_hba, cpuhp);
+	struct lpfc_queue *eq, *next;
+	unsigned int n;
+	int retval;
+
+	if (!phba) {
+		WARN_ONCE(!phba, "cpu: %u. phba:NULL", raw_smp_processor_id());
+		return 0;
+	}
+
+	if (__lpfc_cpuhp_checks(phba, &retval))
+		return retval;
+
+	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)
+			lpfc_sli4_stop_polling(eq);
+	}
+
+	return 0;
+}
+
 /**
  * lpfc_sli4_enable_msix - Enable MSI-X interrupt mode to SLI-4 device
  * @phba: pointer to lpfc hba data structure.
@@ -11460,6 +11634,9 @@ lpfc_sli4_hba_unset(struct lpfc_hba *phba)
 	/* Wait for completion of device XRI exchange busy */
 	lpfc_sli4_xri_exchange_busy_wait(phba);
 
+	/* per-phba callback de-registration for hotplug event */
+	lpfc_cpuhp_remove(phba);
+
 	/* Disable PCI subsystem interrupt */
 	lpfc_sli4_disable_intr(phba);
 
@@ -12752,6 +12929,9 @@ lpfc_pci_probe_one_s4(struct pci_dev *pdev, const struct pci_device_id *pid)
 	/* Enable RAS FW log support */
 	lpfc_sli4_ras_setup(phba);
 
+	INIT_LIST_HEAD(&phba->poll_list);
+	cpuhp_state_add_instance_nocalls(lpfc_cpuhp_state, &phba->cpuhp);
+
 	return 0;
 
 out_free_sysfs_attr:
@@ -13569,11 +13749,24 @@ lpfc_init(void)
 	/* Initialize in case vector mapping is needed */
 	lpfc_present_cpu = num_present_cpus();
 
+	error = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
+					"lpfc/sli4:online",
+					lpfc_cpu_online, lpfc_cpu_offline);
+	if (error < 0)
+		goto cpuhp_failure;
+	lpfc_cpuhp_state = error;
+
 	error = pci_register_driver(&lpfc_driver);
-	if (error) {
-		fc_release_transport(lpfc_transport_template);
-		fc_release_transport(lpfc_vport_transport_template);
-	}
+	if (error)
+		goto unwind;
+
+	return error;
+
+unwind:
+	cpuhp_remove_multi_state(lpfc_cpuhp_state);
+cpuhp_failure:
+	fc_release_transport(lpfc_transport_template);
+	fc_release_transport(lpfc_vport_transport_template);
 
 	return error;
 }
@@ -13590,6 +13783,7 @@ lpfc_exit(void)
 {
 	misc_deregister(&lpfc_mgmt_dev);
 	pci_unregister_driver(&lpfc_driver);
+	cpuhp_remove_multi_state(lpfc_cpuhp_state);
 	fc_release_transport(lpfc_transport_template);
 	fc_release_transport(lpfc_vport_transport_template);
 	idr_destroy(&lpfc_hba_index);