liquidio: Per queue oom work queue

Removed oom task unconditional rescheduling every 250ms and created per
queue oom work queue for refilling buffers.

The oom task refills only if the available descriptors is fallen to 64.
There will be no packets coming in after hitting this level. So NAPI will
not run until oom task refills the buffers.
Signed-off-by: Intiyaz Basha <intiyaz.basha@cavium.com>
Acked-by: Derek Chickles <derek.chickles@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/cavium/liquidio/lio_core.c

@@ -425,56 +425,73 @@ void octeon_pf_changed_vf_macaddr(struct octeon_device *oct, u8 *mac)
 	 */
 }
 
+void octeon_schedule_rxq_oom_work(struct octeon_device *oct,
+				  struct octeon_droq *droq)
+{
+	struct net_device *netdev = oct->props[0].netdev;
+	struct lio *lio = GET_LIO(netdev);
+	struct cavium_wq *wq = &lio->rxq_status_wq[droq->q_no];
+
+	queue_delayed_work(wq->wq, &wq->wk.work,
+			   msecs_to_jiffies(LIO_OOM_POLL_INTERVAL_MS));
+}
+
 static void octnet_poll_check_rxq_oom_status(struct work_struct *work)
 {
 	struct cavium_wk *wk = (struct cavium_wk *)work;
 	struct lio *lio = (struct lio *)wk->ctxptr;
 	struct octeon_device *oct = lio->oct_dev;
-	struct octeon_droq *droq;
-	int q, q_no = 0;
+	int q_no = wk->ctxul;
+	struct octeon_droq *droq = oct->droq[q_no];
 
-	if (ifstate_check(lio, LIO_IFSTATE_RUNNING)) {
-		for (q = 0; q < lio->linfo.num_rxpciq; q++) {
-			q_no = lio->linfo.rxpciq[q].s.q_no;
-			droq = oct->droq[q_no];
-			if (!droq)
-				continue;
-			octeon_droq_check_oom(droq);
-		}
-	}
-	queue_delayed_work(lio->rxq_status_wq.wq,
-			   &lio->rxq_status_wq.wk.work,
-			   msecs_to_jiffies(LIO_OOM_POLL_INTERVAL_MS));
+	if (!ifstate_check(lio, LIO_IFSTATE_RUNNING) || !droq)
+		return;
+
+	if (octeon_retry_droq_refill(droq))
+		octeon_schedule_rxq_oom_work(oct, droq);
 }
 
 int setup_rx_oom_poll_fn(struct net_device *netdev)
 {
 	struct lio *lio = GET_LIO(netdev);
 	struct octeon_device *oct = lio->oct_dev;
+	struct cavium_wq *wq;
+	int q, q_no;
 
-	lio->rxq_status_wq.wq = alloc_workqueue("rxq-oom-status",
-						WQ_MEM_RECLAIM, 0);
-	if (!lio->rxq_status_wq.wq) {
-		dev_err(&oct->pci_dev->dev, "unable to create cavium rxq oom status wq\n");
-		return -ENOMEM;
+	for (q = 0; q < oct->num_oqs; q++) {
+		q_no = lio->linfo.rxpciq[q].s.q_no;
+		wq = &lio->rxq_status_wq[q_no];
+		wq->wq = alloc_workqueue("rxq-oom-status",
+					 WQ_MEM_RECLAIM, 0);
+		if (!wq->wq) {
+			dev_err(&oct->pci_dev->dev, "unable to create cavium rxq oom status wq\n");
+			return -ENOMEM;
+		}
+
+		INIT_DELAYED_WORK(&wq->wk.work,
+				  octnet_poll_check_rxq_oom_status);
+		wq->wk.ctxptr = lio;
+		wq->wk.ctxul = q_no;
 	}
-	INIT_DELAYED_WORK(&lio->rxq_status_wq.wk.work,
-			  octnet_poll_check_rxq_oom_status);
-	lio->rxq_status_wq.wk.ctxptr = lio;
-	queue_delayed_work(lio->rxq_status_wq.wq,
-			   &lio->rxq_status_wq.wk.work,
-			   msecs_to_jiffies(LIO_OOM_POLL_INTERVAL_MS));
+
 	return 0;
 }
 
 void cleanup_rx_oom_poll_fn(struct net_device *netdev)
 {
 	struct lio *lio = GET_LIO(netdev);
-
-	if (lio->rxq_status_wq.wq) {
-		cancel_delayed_work_sync(&lio->rxq_status_wq.wk.work);
-		flush_workqueue(lio->rxq_status_wq.wq);
-		destroy_workqueue(lio->rxq_status_wq.wq);
+	struct octeon_device *oct = lio->oct_dev;
+	struct cavium_wq *wq;
+	int q_no;
+
+	for (q_no = 0; q_no < oct->num_oqs; q_no++) {
+		wq = &lio->rxq_status_wq[q_no];
+		if (wq->wq) {
+			cancel_delayed_work_sync(&wq->wk.work);
+			flush_workqueue(wq->wq);
+			destroy_workqueue(wq->wq);
+			wq->wq = NULL;
+		}
 	}
 }
 

drivers/net/ethernet/cavium/liquidio/lio_ethtool.c

@@ -1115,6 +1115,8 @@ static int lio_reset_queues(struct net_device *netdev, uint32_t num_qs)
 	 * steps like updating sriov_info for the octeon device need to be done.
 	 */
 	if (queue_count_update) {
+		cleanup_rx_oom_poll_fn(netdev);
+
 		lio_delete_glists(lio);
 
 		/* Delete mbox for PF which is SRIOV disabled because sriov_info
@@ -1214,6 +1216,11 @@ static int lio_reset_queues(struct net_device *netdev, uint32_t num_qs)
 			return -1;
 		}
 
+		if (setup_rx_oom_poll_fn(netdev)) {
+			dev_err(&oct->pci_dev->dev, "lio_setup_rx_oom_poll_fn failed\n");
+			return 1;
+		}
+
 		/* Send firmware the information about new number of queues
 		 * if the interface is a VF or a PF that is SRIOV enabled.
 		 */

drivers/net/ethernet/cavium/liquidio/octeon_droq.c

@@ -333,8 +333,6 @@ int octeon_init_droq(struct octeon_device *oct,
  * Returns:
  *  Success: Pointer to recv_info_t
  *  Failure: NULL.
- * Locks:
- *  The droq->lock is held when this routine is called.
  */
 static inline struct octeon_recv_info *octeon_create_recv_info(
 		struct octeon_device *octeon_dev,
@@ -433,8 +431,6 @@ octeon_droq_refill_pullup_descs(struct octeon_droq *droq,
  *  up buffers (that were not dispatched) to form a contiguous ring.
  * Returns:
  *  No of descriptors refilled.
- * Locks:
- *  This routine is called with droq->lock held.
  */
 static u32
 octeon_droq_refill(struct octeon_device *octeon_dev, struct octeon_droq *droq)
@@ -449,8 +445,7 @@ octeon_droq_refill(struct octeon_device *octeon_dev, struct octeon_droq *droq)
 
 	while (droq->refill_count && (desc_refilled < droq->max_count)) {
 		/* If a valid buffer exists (happens if there is no dispatch),
-		 * reuse
-		 * the buffer, else allocate.
+		 * reuse the buffer, else allocate.
 		 */
 		if (!droq->recv_buf_list[droq->refill_idx].buffer) {
 			pg_info =
@@ -503,28 +498,33 @@ octeon_droq_refill(struct octeon_device *octeon_dev, struct octeon_droq *droq)
 
 /** check if we can allocate packets to get out of oom.
  *  @param  droq - Droq being checked.
- *  @return does not return anything
+ *  @return 1 if fails to refill minimum
  */
-void octeon_droq_check_oom(struct octeon_droq *droq)
+int octeon_retry_droq_refill(struct octeon_droq *droq)
 {
-	int desc_refilled;
 	struct octeon_device *oct = droq->oct_dev;
+	int desc_refilled, reschedule = 1;
+	u32 pkts_credit;
+
+	spin_lock_bh(&droq->lock);
+	pkts_credit = readl(droq->pkts_credit_reg);
+	desc_refilled = octeon_droq_refill(oct, droq);
+	if (desc_refilled) {
+		/* Flush the droq descriptor data to memory to be sure
+		 * that when we update the credits the data in memory
+		 * is accurate.
+		 */
+		wmb();
+		writel(desc_refilled, droq->pkts_credit_reg);
+		/* make sure mmio write completes */
+		mmiowb();
 
-	if (readl(droq->pkts_credit_reg) <= CN23XX_SLI_DEF_BP) {
-		spin_lock_bh(&droq->lock);
-		desc_refilled = octeon_droq_refill(oct, droq);
-		if (desc_refilled) {
-			/* Flush the droq descriptor data to memory to be sure
-			 * that when we update the credits the data in memory
-			 * is accurate.
-			 */
-			wmb();
-			writel(desc_refilled, droq->pkts_credit_reg);
-			/* make sure mmio write completes */
-			mmiowb();
-		}
-		spin_unlock_bh(&droq->lock);
+		if (pkts_credit + desc_refilled >= CN23XX_SLI_DEF_BP)
+			reschedule = 0;
 	}
+	spin_unlock_bh(&droq->lock);
+
+	return reschedule;
 }
 
 static inline u32
@@ -603,9 +603,9 @@ octeon_droq_fast_process_packets(struct octeon_device *oct,
 				 struct octeon_droq *droq,
 				 u32 pkts_to_process)
 {
+	u32 pkt, total_len = 0, pkt_count, retval;
 	struct octeon_droq_info *info;
 	union octeon_rh *rh;
-	u32 pkt, total_len = 0, pkt_count;
 
 	pkt_count = pkts_to_process;
 
@@ -709,30 +709,43 @@ octeon_droq_fast_process_packets(struct octeon_device *oct,
 		if (droq->refill_count >= droq->refill_threshold) {
 			int desc_refilled = octeon_droq_refill(oct, droq);
 
-			/* Flush the droq descriptor data to memory to be sure
-			 * that when we update the credits the data in memory
-			 * is accurate.
-			 */
-			wmb();
-			writel((desc_refilled), droq->pkts_credit_reg);
-			/* make sure mmio write completes */
-			mmiowb();
+			if (desc_refilled) {
+				/* Flush the droq descriptor data to memory to
+				 * be sure that when we update the credits the
+				 * data in memory is accurate.
+				 */
+				wmb();
+				writel(desc_refilled, droq->pkts_credit_reg);
+				/* make sure mmio write completes */
+				mmiowb();
+			}
 		}
-
 	}                       /* for (each packet)... */
 
 	/* Increment refill_count by the number of buffers processed. */
 	droq->stats.pkts_received += pkt;
 	droq->stats.bytes_received += total_len;
 
+	retval = pkt;
 	if ((droq->ops.drop_on_max) && (pkts_to_process - pkt)) {
 		octeon_droq_drop_packets(oct, droq, (pkts_to_process - pkt));
 
 		droq->stats.dropped_toomany += (pkts_to_process - pkt);
-		return pkts_to_process;
+		retval = pkts_to_process;
 	}
 
-	return pkt;
+	atomic_sub(retval, &droq->pkts_pending);
+
+	if (droq->refill_count >= droq->refill_threshold &&
+	    readl(droq->pkts_credit_reg) < CN23XX_SLI_DEF_BP) {
+		octeon_droq_check_hw_for_pkts(droq);
+
+		/* Make sure there are no pkts_pending */
+		if (!atomic_read(&droq->pkts_pending))
+			octeon_schedule_rxq_oom_work(oct, droq);
+	}
+
+	return retval;
 }
 
 int
@@ -740,7 +753,7 @@ octeon_droq_process_packets(struct octeon_device *oct,
 			    struct octeon_droq *droq,
 			    u32 budget)
 {
-	u32 pkt_count = 0, pkts_processed = 0;
+	u32 pkt_count = 0;
 	struct list_head *tmp, *tmp2;
 
 	/* Grab the droq lock */
@@ -757,9 +770,7 @@ octeon_droq_process_packets(struct octeon_device *oct,
 	if (pkt_count > budget)
 		pkt_count = budget;
 
-	pkts_processed = octeon_droq_fast_process_packets(oct, droq, pkt_count);
-
-	atomic_sub(pkts_processed, &droq->pkts_pending);
+	octeon_droq_fast_process_packets(oct, droq, pkt_count);
 
 	/* Release the spin lock */
 	spin_unlock(&droq->lock);
@@ -813,8 +824,6 @@ octeon_droq_process_poll_pkts(struct octeon_device *oct,
 			octeon_droq_fast_process_packets(oct, droq,
 							 pkts_available);
 
-		atomic_sub(pkts_processed, &droq->pkts_pending);
-
 		total_pkts_processed += pkts_processed;
 	}
 

drivers/net/ethernet/cavium/liquidio/octeon_droq.h

@@ -414,6 +414,6 @@ int octeon_droq_process_poll_pkts(struct octeon_device *oct,
 
 int octeon_enable_irq(struct octeon_device *oct, u32 q_no);
 
-void octeon_droq_check_oom(struct octeon_droq *droq);
+int octeon_retry_droq_refill(struct octeon_droq *droq);
 
 #endif	/*__OCTEON_DROQ_H__ */

drivers/net/ethernet/cavium/liquidio/octeon_main.h

@@ -70,6 +70,10 @@ void octeon_update_tx_completion_counters(void *buf, int reqtype,
 void octeon_report_tx_completion_to_bql(void *txq, unsigned int pkts_compl,
 					unsigned int bytes_compl);
 void octeon_pf_changed_vf_macaddr(struct octeon_device *oct, u8 *mac);
+
+void octeon_schedule_rxq_oom_work(struct octeon_device *oct,
+				  struct octeon_droq *droq);
+
 /** Swap 8B blocks */
 static inline void octeon_swap_8B_data(u64 *data, u32 blocks)
 {

drivers/net/ethernet/cavium/liquidio/octeon_network.h

@@ -173,7 +173,7 @@ struct lio {
 	struct cavium_wq	txq_status_wq;
 
 	/* work queue for  rxq oom status */
-	struct cavium_wq	rxq_status_wq;
+	struct cavium_wq rxq_status_wq[MAX_POSSIBLE_OCTEON_OUTPUT_QUEUES];
 
 	/* work queue for  link status */
 	struct cavium_wq	link_status_wq;