IB/ehca: Rework irq handler

Rework ehca interrupt handling to avoid/reduce missed irq events.

Signed-off-by: Hoang-Nam Nguyen <hnguyen@de.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/hw/ehca/ehca_classes.h

@@ -42,8 +42,6 @@
 #ifndef __EHCA_CLASSES_H__
 #define __EHCA_CLASSES_H__
 
-#include "ehca_classes.h"
-#include "ipz_pt_fn.h"
 
 struct ehca_module;
 struct ehca_qp;
@@ -54,15 +52,23 @@ struct ehca_mw;
 struct ehca_pd;
 struct ehca_av;
 
-#ifdef CONFIG_PPC64
-#include "ehca_classes_pSeries.h"
-#endif
-
 #include <rdma/ib_verbs.h>
 #include <rdma/ib_user_verbs.h>
 
+#ifdef CONFIG_PPC64
+#include "ehca_classes_pSeries.h"
+#endif
+#include "ipz_pt_fn.h"
+#include "ehca_qes.h"
 #include "ehca_irq.h"
 
+#define EHCA_EQE_CACHE_SIZE 20
+
+struct ehca_eqe_cache_entry {
+	struct ehca_eqe *eqe;
+	struct ehca_cq *cq;
+};
+
 struct ehca_eq {
 	u32 length;
 	struct ipz_queue ipz_queue;
@@ -74,6 +80,8 @@ struct ehca_eq {
 	spinlock_t spinlock;
 	struct tasklet_struct interrupt_task;
 	u32 ist;
+	spinlock_t irq_spinlock;
+	struct ehca_eqe_cache_entry eqe_cache[EHCA_EQE_CACHE_SIZE];
 };
 
 struct ehca_sport {

drivers/infiniband/hw/ehca/ehca_eq.c

@@ -61,6 +61,7 @@ int ehca_create_eq(struct ehca_shca *shca,
 	struct ib_device *ib_dev = &shca->ib_device;
 
 	spin_lock_init(&eq->spinlock);
+	spin_lock_init(&eq->irq_spinlock);
 	eq->is_initialized = 0;
 
 	if (type != EHCA_EQ && type != EHCA_NEQ) {

drivers/infiniband/hw/ehca/ehca_irq.c

@@ -401,87 +401,143 @@ irqreturn_t ehca_interrupt_eq(int irq, void *dev_id)
 	return IRQ_HANDLED;
 }
 
-void ehca_tasklet_eq(unsigned long data)
+
+static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
 {
-	struct ehca_shca *shca = (struct ehca_shca*)data;
+	u64 eqe_value;
-	struct ehca_eqe *eqe;
-	int int_state;
-	int query_cnt = 0;
-
-	do {
-		eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
-
-		if ((shca->hw_level >= 2) && eqe)
-			int_state = 1;
-		else
-			int_state = 0;
-
-		while ((int_state == 1) || eqe) {
-			while (eqe) {
-				u64 eqe_value = eqe->entry;
-
-				ehca_dbg(&shca->ib_device,
-					 "eqe_value=%lx", eqe_value);
-
-				/* TODO: better structure */
-				if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT,
-						   eqe_value)) {
-					unsigned long flags;
 	u32 token;
+	unsigned long flags;
 	struct ehca_cq *cq;
-
+	eqe_value = eqe->entry;
-					ehca_dbg(&shca->ib_device,
+	ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
-						 "... completion event");
+	if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
-					token =
+		ehca_dbg(&shca->ib_device, "... completion event");
-						EHCA_BMASK_GET(EQE_CQ_TOKEN,
+		token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
-							       eqe_value);
+		spin_lock_irqsave(&ehca_cq_idr_lock, flags);
-					spin_lock_irqsave(&ehca_cq_idr_lock,
-							  flags);
 		cq = idr_find(&ehca_cq_idr, token);
-
 		if (cq == NULL) {
-						spin_unlock_irqrestore(&ehca_cq_idr_lock,
+			spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
-								       flags);
+			ehca_err(&shca->ib_device,
-						break;
+				 "Invalid eqe for non-existing cq token=%x",
+				 token);
+			return;
 		}
-
 		reset_eq_pending(cq);
 #ifdef CONFIG_INFINIBAND_EHCA_SCALING
 		queue_comp_task(cq);
-					spin_unlock_irqrestore(&ehca_cq_idr_lock,
+		spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
-							       flags);
 #else
-					spin_unlock_irqrestore(&ehca_cq_idr_lock,
+		spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
-							       flags);
 		comp_event_callback(cq);
 #endif
 	} else {
 		ehca_dbg(&shca->ib_device,
-						 "... non completion event");
+			 "Got non completion event");
 		parse_identifier(shca, eqe_value);
 	}
-				eqe =
+}
-					(struct ehca_eqe *)ehca_poll_eq(shca,
-								    &shca->eq);
-			}
 
-			if (shca->hw_level >= 2) {
+void ehca_process_eq(struct ehca_shca *shca, int is_irq)
-				int_state =
+{
-				    hipz_h_query_int_state(shca->ipz_hca_handle,
+	struct ehca_eq *eq = &shca->eq;
-							   shca->eq.ist);
+	struct ehca_eqe_cache_entry *eqe_cache = eq->eqe_cache;
+	u64 eqe_value;
+	unsigned long flags;
+	int eqe_cnt, i;
+	int eq_empty = 0;
+
+	spin_lock_irqsave(&eq->irq_spinlock, flags);
+	if (is_irq) {
+		const int max_query_cnt = 100;
+		int query_cnt = 0;
+		int int_state = 1;
+		do {
+			int_state = hipz_h_query_int_state(
+				shca->ipz_hca_handle, eq->ist);
 			query_cnt++;
 			iosync();
-				if (query_cnt >= 100) {
+		} while (int_state && query_cnt < max_query_cnt);
-					query_cnt = 0;
+		if (unlikely((query_cnt == max_query_cnt)))
-					int_state = 0;
+			ehca_dbg(&shca->ib_device, "int_state=%x query_cnt=%x",
+				 int_state, query_cnt);
 	}
+
+	/* read out all eqes */
+	eqe_cnt = 0;
+	do {
+		u32 token;
+		eqe_cache[eqe_cnt].eqe =
+			(struct ehca_eqe *)ehca_poll_eq(shca, eq);
+		if (!eqe_cache[eqe_cnt].eqe)
+			break;
+		eqe_value = eqe_cache[eqe_cnt].eqe->entry;
+		if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
+			token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
+			spin_lock(&ehca_cq_idr_lock);
+			eqe_cache[eqe_cnt].cq = idr_find(&ehca_cq_idr, token);
+			if (!eqe_cache[eqe_cnt].cq) {
+				spin_unlock(&ehca_cq_idr_lock);
+				ehca_err(&shca->ib_device,
+					 "Invalid eqe for non-existing cq "
+					 "token=%x", token);
+				continue;
 			}
+			spin_unlock(&ehca_cq_idr_lock);
+		} else
+			eqe_cache[eqe_cnt].cq = NULL;
+		eqe_cnt++;
+	} while (eqe_cnt < EHCA_EQE_CACHE_SIZE);
+	if (!eqe_cnt) {
+		if (is_irq)
+			ehca_dbg(&shca->ib_device,
+				 "No eqe found for irq event");
+		goto unlock_irq_spinlock;
+	} else if (!is_irq)
+		ehca_dbg(&shca->ib_device, "deadman found %x eqe", eqe_cnt);
+	if (unlikely(eqe_cnt == EHCA_EQE_CACHE_SIZE))
+		ehca_dbg(&shca->ib_device, "too many eqes for one irq event");
+	/* enable irq for new packets */
+	for (i = 0; i < eqe_cnt; i++) {
+		if (eq->eqe_cache[i].cq)
+			reset_eq_pending(eq->eqe_cache[i].cq);
+	}
+	/* check eq */
+	spin_lock(&eq->spinlock);
+	eq_empty = (!ipz_eqit_eq_peek_valid(&shca->eq.ipz_queue));
+	spin_unlock(&eq->spinlock);
+	/* call completion handler for cached eqes */
+	for (i = 0; i < eqe_cnt; i++)
+		if (eq->eqe_cache[i].cq) {
+#ifdef CONFIG_INFINIBAND_EHCA_SCALING
+			spin_lock(&ehca_cq_idr_lock);
+			queue_comp_task(eq->eqe_cache[i].cq);
+			spin_unlock(&ehca_cq_idr_lock);
+#else
+			comp_event_callback(eq->eqe_cache[i].cq);
+#endif
+		} else {
+			ehca_dbg(&shca->ib_device, "Got non completion event");
+			parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
+		}
+	/* poll eq if not empty */
+	if (eq_empty)
+		goto unlock_irq_spinlock;
+	do {
+		struct ehca_eqe *eqe;
 		eqe = (struct ehca_eqe *)ehca_poll_eq(shca, &shca->eq);
+		if (!eqe)
+			break;
+		process_eqe(shca, eqe);
+		eqe_cnt++;
+	} while (1);
 
-		}
+unlock_irq_spinlock:
-	} while (int_state != 0);
+	spin_unlock_irqrestore(&eq->irq_spinlock, flags);
+}
 
-	return;
+void ehca_tasklet_eq(unsigned long data)
+{
+	ehca_process_eq((struct ehca_shca*)data, 1);
 }
 
 #ifdef CONFIG_INFINIBAND_EHCA_SCALING

drivers/infiniband/hw/ehca/ehca_irq.h

@@ -56,6 +56,7 @@ void ehca_tasklet_neq(unsigned long data);
 
 irqreturn_t ehca_interrupt_eq(int irq, void *dev_id);
 void ehca_tasklet_eq(unsigned long data);
+void ehca_process_eq(struct ehca_shca *shca, int is_irq);
 
 struct ehca_cpu_comp_task {
 	wait_queue_head_t wait_queue;

drivers/infiniband/hw/ehca/ehca_main.c

@@ -52,7 +52,7 @@
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_AUTHOR("Christoph Raisch <raisch@de.ibm.com>");
 MODULE_DESCRIPTION("IBM eServer HCA InfiniBand Device Driver");
-MODULE_VERSION("SVNEHCA_0020");
+MODULE_VERSION("SVNEHCA_0021");
 
 int ehca_open_aqp1     = 0;
 int ehca_debug_level   = 0;
@@ -778,8 +778,24 @@ void ehca_poll_eqs(unsigned long data)
 
 	spin_lock(&shca_list_lock);
 	list_for_each_entry(shca, &shca_list, shca_list) {
-		if (shca->eq.is_initialized)
+		if (shca->eq.is_initialized) {
-			ehca_tasklet_eq((unsigned long)(void*)shca);
+			/* call deadman proc only if eq ptr does not change */
+			struct ehca_eq *eq = &shca->eq;
+			int max = 3;
+			volatile u64 q_ofs, q_ofs2;
+			u64 flags;
+			spin_lock_irqsave(&eq->spinlock, flags);
+			q_ofs = eq->ipz_queue.current_q_offset;
+			spin_unlock_irqrestore(&eq->spinlock, flags);
+			do {
+				spin_lock_irqsave(&eq->spinlock, flags);
+				q_ofs2 = eq->ipz_queue.current_q_offset;
+				spin_unlock_irqrestore(&eq->spinlock, flags);
+				max--;
+			} while (q_ofs == q_ofs2 && max > 0);
+			if (q_ofs == q_ofs2)
+				ehca_process_eq(shca, 0);
+		}
 	}
 	mod_timer(&poll_eqs_timer, jiffies + HZ);
 	spin_unlock(&shca_list_lock);
@@ -790,7 +806,7 @@ int __init ehca_module_init(void)
 	int ret;
 
 	printk(KERN_INFO "eHCA Infiniband Device Driver "
-	                 "(Rel.: SVNEHCA_0020)\n");
+	       "(Rel.: SVNEHCA_0021)\n");
 	idr_init(&ehca_qp_idr);
 	idr_init(&ehca_cq_idr);
 	spin_lock_init(&ehca_qp_idr_lock);

drivers/infiniband/hw/ehca/ipz_pt_fn.h

@@ -247,6 +247,15 @@ static inline void *ipz_eqit_eq_get_inc_valid(struct ipz_queue *queue)
 	return ret;
 }
 
+static inline void *ipz_eqit_eq_peek_valid(struct ipz_queue *queue)
+{
+	void *ret = ipz_qeit_get(queue);
+	u32 qe = *(u8 *) ret;
+	if ((qe >> 7) != (queue->toggle_state & 1))
+		return NULL;
+	return ret;
+}
+
 /* returns address (GX) of first queue entry */
 static inline u64 ipz_qpt_get_firstpage(struct ipz_qpt *qpt)
 {