scsi: megaraid_sas: IRQ poll to avoid CPU hard lockups
authorShivasharan S <shivasharan.srikanteshwara@broadcom.com>
Tue, 7 May 2019 17:05:35 +0000 (10:05 -0700)
committerMartin K. Petersen <martin.petersen@oracle.com>
Tue, 18 Jun 2019 23:46:19 +0000 (19:46 -0400)
Issue Description:

We have seen cpu lock up issues from field if system has a large (more than
96) logical cpu count.  SAS3.0 controller (Invader series) supports max 96
MSI-X vector and SAS3.5 product (Ventura) supports max 128 MSI-X vectors.

This may be a generic issue (if PCI device support completion on multiple
reply queues).

Let me explain it w.r.t megaraid_sas supported h/w just to simplify the
problem and possible changes to handle such issues.  MegaRAID controller
supports multiple reply queues in completion path.  Driver creates MSI-X
vectors for controller as "minimum of (FW supported Reply queues, Logical
CPUs)".  If submitter is not interrupted via completion on same CPU, there
is a loop in the IO path. This behavior can cause hard/soft CPU lockups, IO
timeout, system sluggish etc.

Example - one CPU (e.g. CPU A) is busy submitting the IOs and another CPU
(e.g. CPU B) is busy with processing the corresponding IO's reply
descriptors from reply descriptor queue upon receiving the interrupts from
HBA.  If CPU A is continuously pumping the IOs then always CPU B (which is
executing the ISR) will see the valid reply descriptors in the reply
descriptor queue and it will be continuously processing those reply
descriptor in a loop without quitting the ISR handler.

megaraid_sas driver will exit ISR handler if it finds unused reply
descriptor in the reply descriptor queue.  Since CPU A will be continuously
sending the IOs, CPU B may always see a valid reply descriptor (posted by
HBA Firmware after processing the IO) in the reply descriptor queue. In
worst case, driver will not quit from this loop in the ISR handler.
Eventually, CPU lockup will be detected by watchdog.

Above mentioned behavior is not common if "rq_affinity" set to 2 or
affinity_hint is honored by irqbalancer as "exact".  If rq_affinity is set
to 2, submitter will be always interrupted via completion on same CPU.  If
irqbalancer is using "exact" policy, interrupt will be delivered to
submitter CPU.

Problem statement:

If CPU count to MSI-X vectors (reply descriptor Queues) count ratio is not
1:1, we still have exposure of issue explained above and for that we don't
have any solution.

Exposure of soft/hard lockup is seen if CPU count is more than MSI-X
supported by device.

If CPUs count to MSI-X vectors count ratio is not 1:1, (Other way, if
CPU counts to MSI-X vector count ratio is something like X:1, where X > 1)
then 'exact' irqbalance policy OR rq_affinity = 2 won't help to avoid CPU
hard/soft lockups. There won't be any one to one mapping between
CPU to MSI-X vector instead one MSI-X interrupt (or reply descriptor queue)
is shared with group/set of CPUs and there is a possibility of having a
loop in the IO path within that CPU group and may observe lockups.

For example: Consider a system having two NUMA nodes and each node having
four logical CPUs and also consider that number of MSI-X vectors enabled on
the HBA is two, then CPUs count to MSI-X vector count ratio as 4:1.
e.g.
MSI-X vector 0 is affinity to CPU 0, CPU 1, CPU 2 & CPU 3 of NUMA node 0 and
MSI-X vector 1 is affinity to CPU 4, CPU 5, CPU 6 & CPU 7 of NUMA node 1.

numactl --hardware
available: 2 nodes (0-1)
node 0 cpus: 0 1 2 3                 --> MSI-X 0
node 0 size: 65536 MB
node 0 free: 63176 MB
node 1 cpus: 4 5 6 7                 --> MSI-X 1
node 1 size: 65536 MB
node 1 free: 63176 MB

Assume that user started an application which uses all the CPUs of NUMA
node 0 for issuing the IOs.  Only one CPU from affinity list (it can be any
cpu since this behavior depends upon irqbalance) CPU0 will receive the
interrupts from MSI-X 0 for all the IOs. Eventually, CPU 0 IO submission
percentage will be decreasing and ISR processing percentage will be
increasing as it is more busy with processing the interrupts.  Gradually IO
submission percentage on CPU 0 will be zero and it's ISR processing
percentage will be 100% as IO loop has already formed within the
NUMA node 0, i.e. CPU 1, CPU 2 & CPU 3 will be continuously busy with
submitting the heavy IOs and only CPU 0 is busy in the ISR path as it
always find the valid reply descriptor in the reply descriptor queue.
Eventually, we will observe the hard lockup here.

Chances of occurring of hard/soft lockups are directly proportional to
value of X. If value of X is high, then chances of observing CPU lockups is
high.

Solution:

Use IRQ poll interface defined in "irq_poll.c".

megaraid_sas driver will execute ISR routine in softirq context and it will
always quit the loop based on budget provided in IRQ poll interface.
Driver will switch to IRQ poll only when more than a threshold number of
reply descriptors are handled in one ISR. Currently threshold is set as
1/4th of HBA queue depth.

In these scenarios (i.e. where CPUs count to MSI-X vectors count ratio is
X:1 (where X >  1)), IRQ poll interface will avoid CPU hard lockups due to
voluntary exit from the reply queue processing based on budget.
Note - Only one MSI-X vector is busy doing processing.

Select CONFIG_IRQ_POLL from driver Kconfig for driver compilation.

Signed-off-by: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: Shivasharan S <shivasharan.srikanteshwara@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
drivers/scsi/megaraid/Kconfig.megaraid
drivers/scsi/megaraid/megaraid_sas.h
drivers/scsi/megaraid/megaraid_sas_base.c
drivers/scsi/megaraid/megaraid_sas_fp.c
drivers/scsi/megaraid/megaraid_sas_fusion.c
drivers/scsi/megaraid/megaraid_sas_fusion.h

index 17419e3..1fa9d09 100644 (file)
@@ -78,6 +78,7 @@ config MEGARAID_LEGACY
 config MEGARAID_SAS
        tristate "LSI Logic MegaRAID SAS RAID Module"
        depends on PCI && SCSI
+       select IRQ_POLL
        help
        Module for LSI Logic's SAS based RAID controllers.
        To compile this driver as a module, choose 'm' here.
index 990ee23..8d6a9c5 100644 (file)
@@ -2173,6 +2173,9 @@ struct megasas_aen_event {
 struct megasas_irq_context {
        struct megasas_instance *instance;
        u32 MSIxIndex;
+       u32 os_irq;
+       struct irq_poll irqpoll;
+       bool irq_poll_scheduled;
 };
 
 struct MR_DRV_SYSTEM_INFO {
@@ -2303,6 +2306,7 @@ struct megasas_instance {
        struct pci_dev *pdev;
        u32 unique_id;
        u32 fw_support_ieee;
+       u32 threshold_reply_count;
 
        atomic_t fw_outstanding;
        atomic_t ldio_outstanding;
@@ -2639,4 +2643,5 @@ void megasas_set_dma_settings(struct megasas_instance *instance,
 int megasas_adp_reset_wait_for_ready(struct megasas_instance *instance,
                                     bool do_adp_reset,
                                     int ocr_context);
+int megasas_irqpoll(struct irq_poll *irqpoll, int budget);
 #endif                         /*LSI_MEGARAID_SAS_H */
index 34abd51..ec1259f 100644 (file)
@@ -48,6 +48,7 @@
 #include <linux/mutex.h>
 #include <linux/poll.h>
 #include <linux/vmalloc.h>
+#include <linux/irq_poll.h>
 
 #include <scsi/scsi.h>
 #include <scsi/scsi_cmnd.h>
@@ -5271,6 +5272,25 @@ fail_alloc_cmds:
        return 1;
 }
 
+static
+void megasas_setup_irq_poll(struct megasas_instance *instance)
+{
+       struct megasas_irq_context *irq_ctx;
+       u32 count, i;
+
+       count = instance->msix_vectors > 0 ? instance->msix_vectors : 1;
+
+       /* Initialize IRQ poll */
+       for (i = 0; i < count; i++) {
+               irq_ctx = &instance->irq_context[i];
+               irq_ctx->os_irq = pci_irq_vector(instance->pdev, i);
+               irq_ctx->irq_poll_scheduled = false;
+               irq_poll_init(&irq_ctx->irqpoll,
+                             instance->threshold_reply_count,
+                             megasas_irqpoll);
+       }
+}
+
 /*
  * megasas_setup_irqs_ioapic -         register legacy interrupts.
  * @instance:                          Adapter soft state
@@ -5349,6 +5369,16 @@ static void
 megasas_destroy_irqs(struct megasas_instance *instance) {
 
        int i;
+       int count;
+       struct megasas_irq_context *irq_ctx;
+
+       count = instance->msix_vectors > 0 ? instance->msix_vectors : 1;
+       if (instance->adapter_type != MFI_SERIES) {
+               for (i = 0; i < count; i++) {
+                       irq_ctx = &instance->irq_context[i];
+                       irq_poll_disable(&irq_ctx->irqpoll);
+               }
+       }
 
        if (instance->msix_vectors)
                for (i = 0; i < instance->msix_vectors; i++) {
@@ -5721,6 +5751,9 @@ static int megasas_init_fw(struct megasas_instance *instance)
                megasas_setup_irqs_ioapic(instance))
                goto fail_init_adapter;
 
+       if (instance->adapter_type != MFI_SERIES)
+               megasas_setup_irq_poll(instance);
+
        instance->instancet->enable_intr(instance);
 
        dev_info(&instance->pdev->dev, "INIT adapter done\n");
@@ -7185,6 +7218,9 @@ megasas_resume(struct pci_dev *pdev)
                        megasas_setup_irqs_ioapic(instance))
                goto fail_init_mfi;
 
+       if (instance->adapter_type != MFI_SERIES)
+               megasas_setup_irq_poll(instance);
+
        /* Re-launch SR-IOV heartbeat timer */
        if (instance->requestorId) {
                if (!megasas_sriov_start_heartbeat(instance, 0))
index 87c2c04..9ac3576 100644 (file)
@@ -45,6 +45,7 @@
 #include <linux/compat.h>
 #include <linux/blkdev.h>
 #include <linux/poll.h>
+#include <linux/irq_poll.h>
 
 #include <scsi/scsi.h>
 #include <scsi/scsi_cmnd.h>
index 5dc324c..5fdba15 100644 (file)
@@ -47,6 +47,7 @@
 #include <linux/poll.h>
 #include <linux/vmalloc.h>
 #include <linux/workqueue.h>
+#include <linux/irq_poll.h>
 
 #include <scsi/scsi.h>
 #include <scsi/scsi_cmnd.h>
@@ -1788,6 +1789,7 @@ megasas_init_adapter_fusion(struct megasas_instance *instance)
 
        instance->flag_ieee = 1;
        instance->r1_ldio_hint_default =  MR_R1_LDIO_PIGGYBACK_DEFAULT;
+       instance->threshold_reply_count = instance->max_fw_cmds / 4;
        fusion->fast_path_io = 0;
 
        if (megasas_allocate_raid_maps(instance))
@@ -3421,7 +3423,8 @@ megasas_complete_r1_command(struct megasas_instance *instance,
  * Completes all commands that is in reply descriptor queue
  */
 int
-complete_cmd_fusion(struct megasas_instance *instance, u32 MSIxIndex)
+complete_cmd_fusion(struct megasas_instance *instance, u32 MSIxIndex,
+                   struct megasas_irq_context *irq_context)
 {
        union MPI2_REPLY_DESCRIPTORS_UNION *desc;
        struct MPI2_SCSI_IO_SUCCESS_REPLY_DESCRIPTOR *reply_desc;
@@ -3554,7 +3557,7 @@ complete_cmd_fusion(struct megasas_instance *instance, u32 MSIxIndex)
                 * number of reply counts and still there are more replies in reply queue
                 * pending to be completed
                 */
-               if (threshold_reply_count >= THRESHOLD_REPLY_COUNT) {
+               if (threshold_reply_count >= instance->threshold_reply_count) {
                        if (instance->msix_combined)
                                writel(((MSIxIndex & 0x7) << 24) |
                                        fusion->last_reply_idx[MSIxIndex],
@@ -3564,23 +3567,46 @@ complete_cmd_fusion(struct megasas_instance *instance, u32 MSIxIndex)
                                        fusion->last_reply_idx[MSIxIndex],
                                        instance->reply_post_host_index_addr[0]);
                        threshold_reply_count = 0;
+                       if (irq_context) {
+                               if (!irq_context->irq_poll_scheduled) {
+                                       irq_context->irq_poll_scheduled = true;
+                                       disable_irq_nosync(irq_context->os_irq);
+                                       irq_poll_sched(&irq_context->irqpoll);
+                               }
+                               return num_completed;
+                       }
                }
        }
 
-       if (!num_completed)
-               return IRQ_NONE;
+       if (num_completed) {
+               wmb();
+               if (instance->msix_combined)
+                       writel(((MSIxIndex & 0x7) << 24) |
+                               fusion->last_reply_idx[MSIxIndex],
+                               instance->reply_post_host_index_addr[MSIxIndex/8]);
+               else
+                       writel((MSIxIndex << 24) |
+                               fusion->last_reply_idx[MSIxIndex],
+                               instance->reply_post_host_index_addr[0]);
+               megasas_check_and_restore_queue_depth(instance);
+       }
+       return num_completed;
+}
 
-       wmb();
-       if (instance->msix_combined)
-               writel(((MSIxIndex & 0x7) << 24) |
-                       fusion->last_reply_idx[MSIxIndex],
-                       instance->reply_post_host_index_addr[MSIxIndex/8]);
-       else
-               writel((MSIxIndex << 24) |
-                       fusion->last_reply_idx[MSIxIndex],
-                       instance->reply_post_host_index_addr[0]);
-       megasas_check_and_restore_queue_depth(instance);
-       return IRQ_HANDLED;
+/**
+ * megasas_enable_irq_poll() - enable irqpoll
+ */
+static void megasas_enable_irq_poll(struct megasas_instance *instance)
+{
+       u32 count, i;
+       struct megasas_irq_context *irq_ctx;
+
+       count = instance->msix_vectors > 0 ? instance->msix_vectors : 1;
+
+       for (i = 0; i < count; i++) {
+               irq_ctx = &instance->irq_context[i];
+               irq_poll_enable(&irq_ctx->irqpoll);
+       }
 }
 
 /**
@@ -3592,11 +3618,46 @@ void megasas_sync_irqs(unsigned long instance_addr)
        u32 count, i;
        struct megasas_instance *instance =
                (struct megasas_instance *)instance_addr;
+       struct megasas_irq_context *irq_ctx;
 
        count = instance->msix_vectors > 0 ? instance->msix_vectors : 1;
 
-       for (i = 0; i < count; i++)
+       for (i = 0; i < count; i++) {
                synchronize_irq(pci_irq_vector(instance->pdev, i));
+               irq_ctx = &instance->irq_context[i];
+               irq_poll_disable(&irq_ctx->irqpoll);
+               if (irq_ctx->irq_poll_scheduled) {
+                       irq_ctx->irq_poll_scheduled = false;
+                       enable_irq(irq_ctx->os_irq);
+               }
+       }
+}
+
+/**
+ * megasas_irqpoll() - process a queue for completed reply descriptors
+ * @irqpoll:   IRQ poll structure associated with queue to poll.
+ * @budget:    Threshold of reply descriptors to process per poll.
+ *
+ * Return: The number of entries processed.
+ */
+
+int megasas_irqpoll(struct irq_poll *irqpoll, int budget)
+{
+       struct megasas_irq_context *irq_ctx;
+       struct megasas_instance *instance;
+       int num_entries;
+
+       irq_ctx = container_of(irqpoll, struct megasas_irq_context, irqpoll);
+       instance = irq_ctx->instance;
+
+       num_entries = complete_cmd_fusion(instance, irq_ctx->MSIxIndex, irq_ctx);
+       if (num_entries < budget) {
+               irq_poll_complete(irqpoll);
+               irq_ctx->irq_poll_scheduled = false;
+               enable_irq(irq_ctx->os_irq);
+       }
+
+       return num_entries;
 }
 
 /**
@@ -3619,7 +3680,7 @@ megasas_complete_cmd_dpc_fusion(unsigned long instance_addr)
                return;
 
        for (MSIxIndex = 0 ; MSIxIndex < count; MSIxIndex++)
-               complete_cmd_fusion(instance, MSIxIndex);
+               complete_cmd_fusion(instance, MSIxIndex, NULL);
 }
 
 /**
@@ -3646,7 +3707,8 @@ irqreturn_t megasas_isr_fusion(int irq, void *devp)
                return IRQ_HANDLED;
        }
 
-       return complete_cmd_fusion(instance, irq_context->MSIxIndex);
+       return complete_cmd_fusion(instance, irq_context->MSIxIndex, irq_context)
+                       ? IRQ_HANDLED : IRQ_NONE;
 }
 
 /**
@@ -4333,6 +4395,7 @@ megasas_issue_tm(struct megasas_instance *instance, u16 device_handle,
                        instance->instancet->disable_intr(instance);
                        megasas_sync_irqs((unsigned long)instance);
                        instance->instancet->enable_intr(instance);
+                       megasas_enable_irq_poll(instance);
                        if (scsi_lookup->scmd == NULL)
                                break;
                }
@@ -4346,6 +4409,7 @@ megasas_issue_tm(struct megasas_instance *instance, u16 device_handle,
                megasas_sync_irqs((unsigned long)instance);
                rc = megasas_track_scsiio(instance, id, channel);
                instance->instancet->enable_intr(instance);
+               megasas_enable_irq_poll(instance);
 
                break;
        case MPI2_SCSITASKMGMT_TASKTYPE_ABRT_TASK_SET:
@@ -4734,10 +4798,7 @@ int megasas_reset_fusion(struct Scsi_Host *shost, int reason)
                        dev_warn(&instance->pdev->dev, "Reset not supported"
                               ", killing adapter scsi%d.\n",
                                instance->host->host_no);
-                       megaraid_sas_kill_hba(instance);
-                       instance->skip_heartbeat_timer_del = 1;
-                       retval = FAILED;
-                       goto out;
+                       goto kill_hba;
                }
 
                /* Let SR-IOV VF & PF sync up if there was a HB failure */
@@ -4775,9 +4836,7 @@ int megasas_reset_fusion(struct Scsi_Host *shost, int reason)
                                dev_info(&instance->pdev->dev,
                                        "Failed from %s %d\n",
                                        __func__, __LINE__);
-                               megaraid_sas_kill_hba(instance);
-                               retval = FAILED;
-                               goto out;
+                               goto kill_hba;
                        }
 
                        megasas_refire_mgmt_cmd(instance);
@@ -4806,7 +4865,7 @@ int megasas_reset_fusion(struct Scsi_Host *shost, int reason)
                        clear_bit(MEGASAS_FUSION_IN_RESET,
                                  &instance->reset_flags);
                        instance->instancet->enable_intr(instance);
-
+                       megasas_enable_irq_poll(instance);
                        shost_for_each_device(sdev, shost) {
                                if ((instance->tgt_prop) &&
                                    (instance->nvme_page_size))
@@ -4857,9 +4916,7 @@ int megasas_reset_fusion(struct Scsi_Host *shost, int reason)
                /* Reset failed, kill the adapter */
                dev_warn(&instance->pdev->dev, "Reset failed, killing "
                       "adapter scsi%d.\n", instance->host->host_no);
-               megaraid_sas_kill_hba(instance);
-               instance->skip_heartbeat_timer_del = 1;
-               retval = FAILED;
+               goto kill_hba;
        } else {
                /* For VF: Restart HB timer if we didn't OCR */
                if (instance->requestorId) {
@@ -4867,8 +4924,15 @@ int megasas_reset_fusion(struct Scsi_Host *shost, int reason)
                }
                clear_bit(MEGASAS_FUSION_IN_RESET, &instance->reset_flags);
                instance->instancet->enable_intr(instance);
+               megasas_enable_irq_poll(instance);
                atomic_set(&instance->adprecovery, MEGASAS_HBA_OPERATIONAL);
+               goto out;
        }
+kill_hba:
+       megaraid_sas_kill_hba(instance);
+       megasas_enable_irq_poll(instance);
+       instance->skip_heartbeat_timer_del = 1;
+       retval = FAILED;
 out:
        clear_bit(MEGASAS_FUSION_IN_RESET, &instance->reset_flags);
        mutex_unlock(&instance->reset_mutex);
index 1481bf0..160ac16 100644 (file)
@@ -100,7 +100,6 @@ enum MR_RAID_FLAGS_IO_SUB_TYPE {
 
 #define MEGASAS_FP_CMD_LEN     16
 #define MEGASAS_FUSION_IN_RESET 0
-#define THRESHOLD_REPLY_COUNT 50
 #define RAID_1_PEER_CMDS 2
 #define JBOD_MAPS_COUNT        2
 #define MEGASAS_REDUCE_QD_COUNT 64