scsi: mpi3mr: Add support for threaded ISR

Register driver for threaded interrupts.

By default the driver will attempt I/O completion from interrupt context
(primary handler). Since the driver tracks per reply queue outstanding
I/Os, it will schedule threaded ISR if there are any outstanding I/Os
expected on that particular reply queue.

Threaded ISR (secondary handler) will loop for I/O completion as long as
there are outstanding I/Os (speculative method using same per reply queue
outstanding counter) or it has completed some X amount of commands
(something like budget).

Link: https://lore.kernel.org/r/20210520152545.2710479-18-kashyap.desai@broadcom.com
Cc: sathya.prakash@broadcom.com
Reviewed-by: Hannes Reinecke <hare@suse.de>
Reviewed-by: Tomas Henzl <thenzl@redhat.com>
Reviewed-by: Himanshu Madhani <himanshu.madhani@oracle.com>
Signed-off-by: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

drivers/scsi/mpi3mr/mpi3mr.h

@@ -149,6 +149,10 @@ extern struct list_head mrioc_list;
 /* Default target device queue depth */
 #define MPI3MR_DEFAULT_SDEV_QD	32
 
+/* Definitions for Threaded IRQ poll*/
+#define MPI3MR_IRQ_POLL_SLEEP			2
+#define MPI3MR_IRQ_POLL_TRIGGER_IOCOUNT		8
+
 /* SGE Flag definition */
 #define MPI3MR_SGEFLAGS_SYSTEM_SIMPLE_END_OF_LIST \
 	(MPI3_SGE_FLAGS_ELEMENT_TYPE_SIMPLE | MPI3_SGE_FLAGS_DLAS_SYSTEM | \
@@ -300,6 +304,9 @@ struct op_req_qinfo {
  * @q_segment_list: Segment list base virtual address
  * @q_segment_list_dma: Segment list base DMA address
  * @ephase: Expected phased identifier for the reply queue
+ * @pend_ios: Number of IOs pending in HW for this queue
+ * @enable_irq_poll: Flag to indicate polling is enabled
+ * @in_use: Queue is handled by poll/ISR
  */
 struct op_reply_qinfo {
 	u16 ci;
@@ -311,6 +318,9 @@ struct op_reply_qinfo {
 	void *q_segment_list;
 	dma_addr_t q_segment_list_dma;
 	u8 ephase;
+	atomic_t pend_ios;
+	bool enable_irq_poll;
+	atomic_t in_use;
 };
 
 /**
@@ -562,6 +572,7 @@ struct scmd_priv {
  * @shost: Scsi_Host pointer
  * @id: Controller ID
  * @cpu_count: Number of online CPUs
+ * @irqpoll_sleep: usleep unit used in threaded isr irqpoll
  * @name: Controller ASCII name
  * @driver_name: Driver ASCII name
  * @sysif_regs: System interface registers virtual address
@@ -663,6 +674,7 @@ struct mpi3mr_ioc {
 	u8 id;
 	int cpu_count;
 	bool enable_segqueue;
+	u32 irqpoll_sleep;
 
 	char name[MPI3MR_NAME_LENGTH];
 	char driver_name[MPI3MR_NAME_LENGTH];

drivers/scsi/mpi3mr/mpi3mr_fw.c

@@ -345,12 +345,16 @@ static int mpi3mr_process_op_reply_q(struct mpi3mr_ioc *mrioc,
 
 	reply_qidx = op_reply_q->qid - 1;
 
+	if (!atomic_add_unless(&op_reply_q->in_use, 1, 1))
+		return 0;
+
 	exp_phase = op_reply_q->ephase;
 	reply_ci = op_reply_q->ci;
 
 	reply_desc = mpi3mr_get_reply_desc(op_reply_q, reply_ci);
 	if ((le16_to_cpu(reply_desc->reply_flags) &
 	    MPI3_REPLY_DESCRIPT_FLAGS_PHASE_MASK) != exp_phase) {
+		atomic_dec(&op_reply_q->in_use);
 		return 0;
 	}
 
@@ -361,6 +365,7 @@ static int mpi3mr_process_op_reply_q(struct mpi3mr_ioc *mrioc,
 		WRITE_ONCE(op_req_q->ci, le16_to_cpu(reply_desc->request_queue_ci));
 		mpi3mr_process_op_reply_desc(mrioc, reply_desc, &reply_dma,
 		    reply_qidx);
+		atomic_dec(&op_reply_q->pend_ios);
 		if (reply_dma)
 			mpi3mr_repost_reply_buf(mrioc, reply_dma);
 		num_op_reply++;
@@ -375,6 +380,14 @@ static int mpi3mr_process_op_reply_q(struct mpi3mr_ioc *mrioc,
 		if ((le16_to_cpu(reply_desc->reply_flags) &
 		    MPI3_REPLY_DESCRIPT_FLAGS_PHASE_MASK) != exp_phase)
 			break;
+		/*
+		 * Exit completion loop to avoid CPU lockup
+		 * Ensure remaining completion happens from threaded ISR.
+		 */
+		if (num_op_reply > mrioc->max_host_ios) {
+			intr_info->op_reply_q->enable_irq_poll = true;
+			break;
+		}
 
 	} while (1);
 
@@ -383,6 +396,7 @@ static int mpi3mr_process_op_reply_q(struct mpi3mr_ioc *mrioc,
 	op_reply_q->ci = reply_ci;
 	op_reply_q->ephase = exp_phase;
 
+	atomic_dec(&op_reply_q->in_use);
 	return num_op_reply;
 }
 
@@ -391,7 +405,7 @@ static irqreturn_t mpi3mr_isr_primary(int irq, void *privdata)
 	struct mpi3mr_intr_info *intr_info = privdata;
 	struct mpi3mr_ioc *mrioc;
 	u16 midx;
-	u32 num_admin_replies = 0;
+	u32 num_admin_replies = 0, num_op_reply = 0;
 
 	if (!intr_info)
 		return IRQ_NONE;
@@ -405,8 +419,10 @@ static irqreturn_t mpi3mr_isr_primary(int irq, void *privdata)
 
 	if (!midx)
 		num_admin_replies = mpi3mr_process_admin_reply_q(mrioc);
+	if (intr_info->op_reply_q)
+		num_op_reply = mpi3mr_process_op_reply_q(mrioc, intr_info);
 
-	if (num_admin_replies)
+	if (num_admin_replies || num_op_reply)
 		return IRQ_HANDLED;
 	else
 		return IRQ_NONE;
@@ -415,15 +431,32 @@ static irqreturn_t mpi3mr_isr_primary(int irq, void *privdata)
 static irqreturn_t mpi3mr_isr(int irq, void *privdata)
 {
 	struct mpi3mr_intr_info *intr_info = privdata;
+	struct mpi3mr_ioc *mrioc;
+	u16 midx;
 	int ret;
 
 	if (!intr_info)
 		return IRQ_NONE;
 
+	mrioc = intr_info->mrioc;
+	midx = intr_info->msix_index;
 	/* Call primary ISR routine */
 	ret = mpi3mr_isr_primary(irq, privdata);
 
-	return ret;
+	/*
+	 * If more IOs are expected, schedule IRQ polling thread.
+	 * Otherwise exit from ISR.
+	 */
+	if (!intr_info->op_reply_q)
+		return ret;
+
+	if (!intr_info->op_reply_q->enable_irq_poll ||
+	    !atomic_read(&intr_info->op_reply_q->pend_ios))
+		return ret;
+
+	disable_irq_nosync(pci_irq_vector(mrioc->pdev, midx));
+
+	return IRQ_WAKE_THREAD;
 }
 
 /**
@@ -438,6 +471,36 @@ static irqreturn_t mpi3mr_isr(int irq, void *privdata)
  */
 static irqreturn_t mpi3mr_isr_poll(int irq, void *privdata)
 {
+	struct mpi3mr_intr_info *intr_info = privdata;
+	struct mpi3mr_ioc *mrioc;
+	u16 midx;
+	u32 num_op_reply = 0;
+
+	if (!intr_info || !intr_info->op_reply_q)
+		return IRQ_NONE;
+
+	mrioc = intr_info->mrioc;
+	midx = intr_info->msix_index;
+
+	/* Poll for pending IOs completions */
+	do {
+		if (!mrioc->intr_enabled)
+			break;
+
+		if (!midx)
+			mpi3mr_process_admin_reply_q(mrioc);
+		if (intr_info->op_reply_q)
+			num_op_reply +=
+			    mpi3mr_process_op_reply_q(mrioc, intr_info);
+
+		usleep_range(mrioc->irqpoll_sleep, 10 * mrioc->irqpoll_sleep);
+
+	} while (atomic_read(&intr_info->op_reply_q->pend_ios) &&
+	    (num_op_reply < mrioc->max_host_ios));
+
+	intr_info->op_reply_q->enable_irq_poll = false;
+	enable_irq(pci_irq_vector(mrioc->pdev, midx));
+
 	return IRQ_HANDLED;
 }
 
@@ -1147,6 +1210,9 @@ static int mpi3mr_create_op_reply_q(struct mpi3mr_ioc *mrioc, u16 qidx)
 	op_reply_q->num_replies = MPI3MR_OP_REP_Q_QD;
 	op_reply_q->ci = 0;
 	op_reply_q->ephase = 1;
+	atomic_set(&op_reply_q->pend_ios, 0);
+	atomic_set(&op_reply_q->in_use, 0);
+	op_reply_q->enable_irq_poll = false;
 
 	if (!op_reply_q->q_segments) {
 		retval = mpi3mr_alloc_op_reply_q_segments(mrioc, qidx);
@@ -1465,6 +1531,10 @@ int mpi3mr_op_request_post(struct mpi3mr_ioc *mrioc,
 		pi = 0;
 	op_req_q->pi = pi;
 
+	if (atomic_inc_return(&mrioc->op_reply_qinfo[reply_qidx].pend_ios)
+	    > MPI3MR_IRQ_POLL_TRIGGER_IOCOUNT)
+		mrioc->op_reply_qinfo[reply_qidx].enable_irq_poll = true;
+
 	writel(op_req_q->pi,
 	    &mrioc->sysif_regs->oper_queue_indexes[reply_qidx].producer_index);
 
@@ -2795,6 +2865,7 @@ int mpi3mr_init_ioc(struct mpi3mr_ioc *mrioc, u8 re_init)
 	u32 ioc_status, ioc_config, i;
 	struct mpi3_ioc_facts_data facts_data;
 
+	mrioc->irqpoll_sleep = MPI3MR_IRQ_POLL_SLEEP;
 	mrioc->change_count = 0;
 	if (!re_init) {
 		mrioc->cpu_count = num_online_cpus();
@@ -3081,6 +3152,8 @@ static void mpi3mr_memset_buffers(struct mpi3mr_ioc *mrioc)
 		mrioc->op_reply_qinfo[i].ci = 0;
 		mrioc->op_reply_qinfo[i].num_replies = 0;
 		mrioc->op_reply_qinfo[i].ephase = 0;
+		atomic_set(&mrioc->op_reply_qinfo[i].pend_ios, 0);
+		atomic_set(&mrioc->op_reply_qinfo[i].in_use, 0);
 		mpi3mr_memset_op_reply_q_buffers(mrioc, i);
 
 		mrioc->req_qinfo[i].ci = 0;