Merge branch 'liquidio-Removed-droq-lock-from-Rx-path'

Intiyaz Basha says:

====================
liquidio: Removed droq lock from Rx path

Series of patches for removing droq lock from Rx Path.
====================
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/lio_main.c

@@ -1239,8 +1239,10 @@ static void send_rx_ctrl_cmd(struct lio *lio, int start_stop)
 static void liquidio_destroy_nic_device(struct octeon_device *oct, int ifidx)
 {
 	struct net_device *netdev = oct->props[ifidx].netdev;
-	struct lio *lio;
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct napi_struct *napi, *n;
+	struct lio *lio;
 
 	if (!netdev) {
 		dev_err(&oct->pci_dev->dev, "%s No netdevice ptr for index %d\n",
@@ -1269,6 +1271,8 @@ static void liquidio_destroy_nic_device(struct octeon_device *oct, int ifidx)
 	list_for_each_entry_safe(napi, n, &netdev->napi_list, dev_list)
 		netif_napi_del(napi);
 
+	tasklet_enable(&oct_priv->droq_tasklet);
+
 	if (atomic_read(&lio->ifstate) & LIO_IFSTATE_REGISTERED)
 		unregister_netdev(netdev);
 
@@ -1805,9 +1809,13 @@ static int liquidio_open(struct net_device *netdev)
 {
 	struct lio *lio = GET_LIO(netdev);
 	struct octeon_device *oct = lio->oct_dev;
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct napi_struct *napi, *n;
 
 	if (oct->props[lio->ifidx].napi_enabled == 0) {
+		tasklet_disable(&oct_priv->droq_tasklet);
+
 		list_for_each_entry_safe(napi, n, &netdev->napi_list, dev_list)
 			napi_enable(napi);
 
@@ -1861,6 +1869,8 @@ static int liquidio_stop(struct net_device *netdev)
 {
 	struct lio *lio = GET_LIO(netdev);
 	struct octeon_device *oct = lio->oct_dev;
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct napi_struct *napi, *n;
 
 	ifstate_reset(lio, LIO_IFSTATE_RUNNING);
@@ -1907,6 +1917,8 @@ static int liquidio_stop(struct net_device *netdev)
 
 		if (OCTEON_CN23XX_PF(oct))
 			oct->droq[0]->ops.poll_mode = 0;
+
+		tasklet_enable(&oct_priv->droq_tasklet);
 	}
 
 	dev_info(&oct->pci_dev->dev, "%s interface is stopped\n", netdev->name);

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

@@ -444,6 +444,8 @@ static void octeon_pci_flr(struct octeon_device *oct)
  */
 static void octeon_destroy_resources(struct octeon_device *oct)
 {
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct msix_entry *msix_entries;
 	int i;
 
@@ -587,6 +589,8 @@ static void octeon_destroy_resources(struct octeon_device *oct)
 		/* Nothing to be done here either */
 		break;
 	}
+
+	tasklet_kill(&oct_priv->droq_tasklet);
 }
 
 /**
@@ -652,6 +656,8 @@ static void send_rx_ctrl_cmd(struct lio *lio, int start_stop)
 static void liquidio_destroy_nic_device(struct octeon_device *oct, int ifidx)
 {
 	struct net_device *netdev = oct->props[ifidx].netdev;
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct napi_struct *napi, *n;
 	struct lio *lio;
 
@@ -681,6 +687,8 @@ static void liquidio_destroy_nic_device(struct octeon_device *oct, int ifidx)
 	list_for_each_entry_safe(napi, n, &netdev->napi_list, dev_list)
 		netif_napi_del(napi);
 
+	tasklet_enable(&oct_priv->droq_tasklet);
+
 	if (atomic_read(&lio->ifstate) & LIO_IFSTATE_REGISTERED)
 		unregister_netdev(netdev);
 
@@ -898,9 +906,13 @@ static int liquidio_open(struct net_device *netdev)
 {
 	struct lio *lio = GET_LIO(netdev);
 	struct octeon_device *oct = lio->oct_dev;
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct napi_struct *napi, *n;
 
 	if (!oct->props[lio->ifidx].napi_enabled) {
+		tasklet_disable(&oct_priv->droq_tasklet);
+
 		list_for_each_entry_safe(napi, n, &netdev->napi_list, dev_list)
 			napi_enable(napi);
 
@@ -938,6 +950,8 @@ static int liquidio_stop(struct net_device *netdev)
 {
 	struct lio *lio = GET_LIO(netdev);
 	struct octeon_device *oct = lio->oct_dev;
+	struct octeon_device_priv *oct_priv =
+		(struct octeon_device_priv *)oct->priv;
 	struct napi_struct *napi, *n;
 
 	/* tell Octeon to stop forwarding packets to host */
@@ -967,6 +981,8 @@ static int liquidio_stop(struct net_device *netdev)
 		oct->props[lio->ifidx].napi_enabled = 0;
 
 		oct->droq[0]->ops.poll_mode = 0;
+
+		tasklet_enable(&oct_priv->droq_tasklet);
 	}
 
 	cancel_delayed_work_sync(&lio->stats_wk.work);

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

@@ -1440,12 +1440,8 @@ void lio_enable_irq(struct octeon_droq *droq, struct octeon_instr_queue *iq)
 	/* the whole thing needs to be atomic, ideally */
 	if (droq) {
 		pkts_pend = (u32)atomic_read(&droq->pkts_pending);
-		spin_lock_bh(&droq->lock);
 		writel(droq->pkt_count - pkts_pend, droq->pkts_sent_reg);
 		droq->pkt_count = pkts_pend;
-		/* this write needs to be flushed before we release the lock */
-		mmiowb();
-		spin_unlock_bh(&droq->lock);
 		oct = droq->oct_dev;
 	}
 	if (iq) {

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

@@ -301,8 +301,6 @@ int octeon_init_droq(struct octeon_device *oct,
 	dev_dbg(&oct->pci_dev->dev, "DROQ INIT: max_empty_descs: %d\n",
 		droq->max_empty_descs);
 
-	spin_lock_init(&droq->lock);
-
 	INIT_LIST_HEAD(&droq->dispatch_list);
 
 	/* For 56xx Pass1, this function won't be called, so no checks. */
@@ -333,8 +331,6 @@ 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 +429,6 @@ 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 +443,7 @@ 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 +496,31 @@ 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;
+
+	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;
 	}
+
+	return reschedule;
 }
 
 static inline u32
@@ -603,9 +599,9 @@ 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 +705,43 @@ 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;
+	}
+
+	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 pkt;
+	return retval;
 }
 
 int
@@ -740,29 +749,19 @@ 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 */
-	spin_lock(&droq->lock);
-
 	octeon_droq_check_hw_for_pkts(droq);
 	pkt_count = atomic_read(&droq->pkts_pending);
 
-	if (!pkt_count) {
-		spin_unlock(&droq->lock);
+	if (!pkt_count)
 		return 0;
-	}
 
 	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);
-
-	/* Release the spin lock */
-	spin_unlock(&droq->lock);
+	octeon_droq_fast_process_packets(oct, droq, pkt_count);
 
 	list_for_each_safe(tmp, tmp2, &droq->dispatch_list) {
 		struct __dispatch *rdisp = (struct __dispatch *)tmp;
@@ -798,8 +797,6 @@ octeon_droq_process_poll_pkts(struct octeon_device *oct,
 	if (budget > droq->max_count)
 		budget = droq->max_count;
 
-	spin_lock(&droq->lock);
-
 	while (total_pkts_processed < budget) {
 		octeon_droq_check_hw_for_pkts(droq);
 
@@ -813,13 +810,9 @@ 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;
 	}
 
-	spin_unlock(&droq->lock);
-
 	list_for_each_safe(tmp, tmp2, &droq->dispatch_list) {
 		struct __dispatch *rdisp = (struct __dispatch *)tmp;
 
@@ -879,9 +872,8 @@ octeon_enable_irq(struct octeon_device *oct, u32 q_no)
 int octeon_register_droq_ops(struct octeon_device *oct, u32 q_no,
 			     struct octeon_droq_ops *ops)
 {
-	struct octeon_droq *droq;
-	unsigned long flags;
 	struct octeon_config *oct_cfg = NULL;
+	struct octeon_droq *droq;
 
 	oct_cfg = octeon_get_conf(oct);
 
@@ -901,21 +893,15 @@ int octeon_register_droq_ops(struct octeon_device *oct, u32 q_no,
 	}
 
 	droq = oct->droq[q_no];
-
-	spin_lock_irqsave(&droq->lock, flags);
-
 	memcpy(&droq->ops, ops, sizeof(struct octeon_droq_ops));
 
-	spin_unlock_irqrestore(&droq->lock, flags);
-
 	return 0;
 }
 
 int octeon_unregister_droq_ops(struct octeon_device *oct, u32 q_no)
 {
-	unsigned long flags;
-	struct octeon_droq *droq;
 	struct octeon_config *oct_cfg = NULL;
+	struct octeon_droq *droq;
 
 	oct_cfg = octeon_get_conf(oct);
 
@@ -936,14 +922,10 @@ int octeon_unregister_droq_ops(struct octeon_device *oct, u32 q_no)
 		return 0;
 	}
 
-	spin_lock_irqsave(&droq->lock, flags);
-
 	droq->ops.fptr = NULL;
 	droq->ops.farg = NULL;
 	droq->ops.drop_on_max = 0;
 
-	spin_unlock_irqrestore(&droq->lock, flags);
-
 	return 0;
 }
 

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

@@ -245,9 +245,6 @@ struct octeon_droq_ops {
  *  Octeon DROQ.
  */
 struct octeon_droq {
-	/** A spinlock to protect access to this ring. */
-	spinlock_t lock;
-
 	u32 q_no;
 
 	u32 pkt_count;
@@ -414,6 +411,6 @@ 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;