Skip to contentSkip to conversation

liquidio: Removed droq lock from Rx path

PATCH In review Felix Manlunas 2 participants 5 messages Sep 10, 2018 – Sep 12, 2018

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover liquidio: Removed droq lock from Rx path

From: Intiyaz Basha <[email protected]>

Series of patches for removing droq lock from Rx Path.

Intiyaz Basha (3):
liquidio: Disabling tasklet when NAPI is active
liquidio: Per queue oom work queue
liquidio: Removed droq lock

 drivers/net/ethernet/cavium/liquidio/lio_core.c    |  77 +++++++------
 drivers/net/ethernet/cavium/liquidio/lio_ethtool.c |   7 ++
 drivers/net/ethernet/cavium/liquidio/lio_main.c    |  14 ++-
 drivers/net/ethernet/cavium/liquidio/lio_vf_main.c |  16 +++
 .../net/ethernet/cavium/liquidio/octeon_device.c   |   4 -
 drivers/net/ethernet/cavium/liquidio/octeon_droq.c | 120 +++++++++------------
 drivers/net/ethernet/cavium/liquidio/octeon_droq.h |   5 +-
 drivers/net/ethernet/cavium/liquidio/octeon_main.h |   4 +
 .../net/ethernet/cavium/liquidio/octeon_network.h  |   2 +-
 9 files changed, 140 insertions(+), 109 deletions(-)
Signature · 1 line
1.8.3.1
#1 lore raw

PATCH 1/3 liquidio: Disabling tasklet when NAPI is active

From: Intiyaz Basha <[email protected]>

Control packets are processed in tasklet when interface is down and in NAPI when interface is up. So tasklet can be disabled when interface up and re-enabled when interface is down.

---
 drivers/net/ethernet/cavium/liquidio/lio_main.c    | 14 +++++++++++++-
 drivers/net/ethernet/cavium/liquidio/lio_vf_main.c | 16 ++++++++++++++++
 2 files changed, 29 insertions(+), 1 deletion(-)
drivers/net/ethernet/cavium/liquidio/lio_main.c +13 -1
index 40f941f..0aba1f7 100644
@@ -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 +16 -0
index 8fa7ac3..0ec4bfe 100644
@@ -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);
Signature · 1 line
1.8.3.1
#2 lore raw

PATCH 2/3 liquidio: Per queue oom work queue

In reply to Felix Manlunas

From: Intiyaz Basha <[email protected]>

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.

---
 drivers/net/ethernet/cavium/liquidio/lio_core.c    | 77 +++++++++++-------
 drivers/net/ethernet/cavium/liquidio/lio_ethtool.c |  7 ++
 drivers/net/ethernet/cavium/liquidio/octeon_droq.c | 91 ++++++++++++----------
 drivers/net/ethernet/cavium/liquidio/octeon_droq.h |  2 +-
 drivers/net/ethernet/cavium/liquidio/octeon_main.h |  4 +
 .../net/ethernet/cavium/liquidio/octeon_network.h  |  2 +-
 6 files changed, 110 insertions(+), 73 deletions(-)
drivers/net/ethernet/cavium/liquidio/lio_core.c +47 -30
index 0284204..55ed20b 100644
@@ -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 +7 -0
index e1e5808..9e53cdb 100644
@@ -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 +50 -41

165 lines not on this page · raw message

drivers/net/ethernet/cavium/liquidio/octeon_droq.h +1 -1
index f28f262..b201936 100644
@@ -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 +4 -0
index 38c055f..073d064 100644
@@ -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 +1 -1
index 9117b5a..beb3eec 100644
@@ -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;
Signature · 1 line
1.8.3.1
#3 lore raw

PATCH 3/3 liquidio: Removed droq lock

In reply to Felix Manlunas

From: Intiyaz Basha <[email protected]>

With the changes in patch 1 and 2, droq lock is not required. So removing droq lock.

---
 .../net/ethernet/cavium/liquidio/octeon_device.c   |  4 ---
 drivers/net/ethernet/cavium/liquidio/octeon_droq.c | 33 ++--------------------
 drivers/net/ethernet/cavium/liquidio/octeon_droq.h |  3 --
 3 files changed, 3 insertions(+), 37 deletions(-)
drivers/net/ethernet/cavium/liquidio/octeon_device.c +0 -4
index d0ed6c4..0f0275c 100644
@@ -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 +3 -30
index 53c25ee..ad621aa 100644
@@ -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. */
@@ -506,7 +504,6 @@ int octeon_retry_droq_refill(struct octeon_droq *droq)
 	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) {
@@ -522,7 +519,6 @@ int octeon_retry_droq_refill(struct octeon_droq *droq)
 		if (pkts_credit + desc_refilled >= CN23XX_SLI_DEF_BP)
 			reschedule = 0;
 	}
-	spin_unlock_bh(&droq->lock);
 
 	return reschedule;
 }
@@ -756,25 +752,17 @@ static inline void octeon_droq_drop_packets(struct octeon_device *oct,
 	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;
 
 	octeon_droq_fast_process_packets(oct, droq, pkt_count);
 
-	/* Release the spin lock */
-	spin_unlock(&droq->lock);
-
 	list_for_each_safe(tmp, tmp2, &droq->dispatch_list) {
 		struct __dispatch *rdisp = (struct __dispatch *)tmp;
 
@@ -809,8 +797,6 @@ static inline void octeon_droq_drop_packets(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);
 
@@ -827,8 +813,6 @@ static inline void octeon_droq_drop_packets(struct octeon_device *oct,
 		total_pkts_processed += pkts_processed;
 	}
 
-	spin_unlock(&droq->lock);
-
 	list_for_each_safe(tmp, tmp2, &droq->dispatch_list) {
 		struct __dispatch *rdisp = (struct __dispatch *)tmp;
 
@@ -888,9 +872,8 @@ static inline void octeon_droq_drop_packets(struct octeon_device *oct,
 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);
 
@@ -910,21 +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);
 
@@ -945,14 +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 +0 -3
index b201936..c9b19e6 100644
@@ -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;
Signature · 1 line
1.8.3.1