---

yaml
---
r: 154290
b: refs/heads/master
c: a620c16
h: refs/heads/master
v: v3

[refs]

@@ -1,2 +1,2 @@
 ---
-refs/heads/master: f696a10838ffab85e5bc07e7cff0d0e1870a30d7
+refs/heads/master: a620c1632629b42369e78448acc7b384fe1faf48

trunk/drivers/staging/octeon/ethernet-defines.h

@@ -117,6 +117,8 @@
 
 /* Maximum number of packets to process per interrupt. */
 #define MAX_RX_PACKETS 120
+/* Maximum number of SKBs to try to free per xmit packet. */
+#define MAX_SKB_TO_FREE 10
 #define MAX_OUT_QUEUE_DEPTH 1000
 
 #ifndef CONFIG_SMP

trunk/drivers/staging/octeon/ethernet-tx.c

@@ -47,6 +47,7 @@
 
 #include "ethernet-defines.h"
 #include "octeon-ethernet.h"
+#include "ethernet-tx.h"
 #include "ethernet-util.h"
 
 #include "cvmx-wqe.h"
@@ -82,8 +83,10 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	uint64_t old_scratch2;
 	int dropped;
 	int qos;
+	int queue_it_up;
 	struct octeon_ethernet *priv = netdev_priv(dev);
-	int32_t in_use;
+	int32_t skb_to_free;
+	int32_t undo;
 	int32_t buffers_to_free;
 #if REUSE_SKBUFFS_WITHOUT_FREE
 	unsigned char *fpa_head;
@@ -120,15 +123,15 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 		old_scratch2 = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 
 		/*
-		 * Assume we're going to be able t osend this
-		 * packet. Fetch and increment the number of pending
-		 * packets for output.
+		 * Fetch and increment the number of packets to be
+		 * freed.
 		 */
 		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH + 8,
 					       FAU_NUM_PACKET_BUFFERS_TO_FREE,
 					       0);
 		cvmx_fau_async_fetch_and_add32(CVMX_SCR_SCRATCH,
-					       priv->fau + qos * 4, 1);
+					       priv->fau + qos * 4,
+					       MAX_SKB_TO_FREE);
 	}
 
 	/*
@@ -286,15 +289,29 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 	if (USE_ASYNC_IOBDMA) {
 		/* Get the number of skbuffs in use by the hardware */
 		CVMX_SYNCIOBDMA;
-		in_use = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
+		skb_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH);
 		buffers_to_free = cvmx_scratch_read64(CVMX_SCR_SCRATCH + 8);
 	} else {
 		/* Get the number of skbuffs in use by the hardware */
-		in_use = cvmx_fau_fetch_and_add32(priv->fau + qos * 4, 1);
+		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
+						       MAX_SKB_TO_FREE);
 		buffers_to_free =
 		    cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
 	}
 
+	/*
+	 * We try to claim MAX_SKB_TO_FREE buffers.  If there were not
+	 * that many available, we have to un-claim (undo) any that
+	 * were in excess.  If skb_to_free is positive we will free
+	 * that many buffers.
+	 */
+	undo = skb_to_free > 0 ?
+		MAX_SKB_TO_FREE : skb_to_free + MAX_SKB_TO_FREE;
+	if (undo > 0)
+		cvmx_fau_atomic_add32(priv->fau+qos*4, -undo);
+	skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ?
+		MAX_SKB_TO_FREE : -skb_to_free;
+
 	/*
 	 * If we're sending faster than the receive can free them then
 	 * don't do the HW free.
@@ -330,38 +347,31 @@ int cvm_oct_xmit(struct sk_buff *skb, struct net_device *dev)
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH + 8, old_scratch2);
 	}
 
+	queue_it_up = 0;
 	if (unlikely(dropped)) {
 		dev_kfree_skb_any(skb);
-		cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
 		priv->stats.tx_dropped++;
 	} else {
 		if (USE_SKBUFFS_IN_HW) {
 			/* Put this packet on the queue to be freed later */
 			if (pko_command.s.dontfree)
-				skb_queue_tail(&priv->tx_free_list[qos], skb);
-			else {
+				queue_it_up = 1;
+			else
 				cvmx_fau_atomic_add32
 				    (FAU_NUM_PACKET_BUFFERS_TO_FREE, -1);
-				cvmx_fau_atomic_add32(priv->fau + qos * 4, -1);
-			}
 		} else {
 			/* Put this packet on the queue to be freed later */
-			skb_queue_tail(&priv->tx_free_list[qos], skb);
+			queue_it_up = 1;
 		}
 	}
 
-	/* Free skbuffs not in use by the hardware, possibly two at a time */
-	if (skb_queue_len(&priv->tx_free_list[qos]) > in_use) {
+	if (queue_it_up) {
 		spin_lock(&priv->tx_free_list[qos].lock);
-		/*
-		 * Check again now that we have the lock. It might
-		 * have changed.
-		 */
-		if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
-			dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
-		if (skb_queue_len(&priv->tx_free_list[qos]) > in_use)
-			dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
+		__skb_queue_tail(&priv->tx_free_list[qos], skb);
+		cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 0);
 		spin_unlock(&priv->tx_free_list[qos].lock);
+	} else {
+		cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 1);
 	}
 
 	return 0;

trunk/drivers/staging/octeon/ethernet-tx.h

@@ -30,3 +30,28 @@ int cvm_oct_xmit_pow(struct sk_buff *skb, struct net_device *dev);
 int cvm_oct_transmit_qos(struct net_device *dev, void *work_queue_entry,
 			 int do_free, int qos);
 void cvm_oct_tx_shutdown(struct net_device *dev);
+
+/**
+ * Free dead transmit skbs.
+ *
+ * @priv:		The driver data
+ * @skb_to_free:	The number of SKBs to free (free none if negative).
+ * @qos:		The queue to free from.
+ * @take_lock:		If true, acquire the skb list lock.
+ */
+static inline void cvm_oct_free_tx_skbs(struct octeon_ethernet *priv,
+					int skb_to_free,
+					int qos, int take_lock)
+{
+	/* Free skbuffs not in use by the hardware.  */
+	if (skb_to_free > 0) {
+		if (take_lock)
+			spin_lock(&priv->tx_free_list[qos].lock);
+		while (skb_to_free > 0) {
+			dev_kfree_skb(__skb_dequeue(&priv->tx_free_list[qos]));
+			skb_to_free--;
+		}
+		if (take_lock)
+			spin_unlock(&priv->tx_free_list[qos].lock);
+	}
+}

trunk/drivers/staging/octeon/ethernet.c

@@ -37,13 +37,14 @@
 #include <asm/octeon/octeon.h>
 
 #include "ethernet-defines.h"
+#include "octeon-ethernet.h"
 #include "ethernet-mem.h"
 #include "ethernet-rx.h"
 #include "ethernet-tx.h"
 #include "ethernet-mdio.h"
 #include "ethernet-util.h"
 #include "ethernet-proc.h"
-#include "octeon-ethernet.h"
+
 
 #include "cvmx-pip.h"
 #include "cvmx-pko.h"
@@ -130,53 +131,55 @@ extern struct semaphore mdio_sem;
  */
 static void cvm_do_timer(unsigned long arg)
 {
+	int32_t skb_to_free, undo;
+	int queues_per_port;
+	int qos;
+	struct octeon_ethernet *priv;
 	static int port;
-	if (port < CVMX_PIP_NUM_INPUT_PORTS) {
-		if (cvm_oct_device[port]) {
-			int queues_per_port;
-			int qos;
-			struct octeon_ethernet *priv =
-				netdev_priv(cvm_oct_device[port]);
-			if (priv->poll) {
-				/* skip polling if we don't get the lock */
-				if (!down_trylock(&mdio_sem)) {
-					priv->poll(cvm_oct_device[port]);
-					up(&mdio_sem);
-				}
-			}
 
-			queues_per_port = cvmx_pko_get_num_queues(port);
-			/* Drain any pending packets in the free list */
-			for (qos = 0; qos < queues_per_port; qos++) {
-				if (skb_queue_len(&priv->tx_free_list[qos])) {
-					spin_lock(&priv->tx_free_list[qos].
-						  lock);
-					while (skb_queue_len
-					       (&priv->tx_free_list[qos]) >
-					       cvmx_fau_fetch_and_add32(priv->
-									fau +
-									qos * 4,
-									0))
-						dev_kfree_skb(__skb_dequeue
-							      (&priv->
-							       tx_free_list
-							       [qos]));
-					spin_unlock(&priv->tx_free_list[qos].
-						    lock);
-				}
-			}
-			cvm_oct_device[port]->netdev_ops->ndo_get_stats(cvm_oct_device[port]);
-		}
-		port++;
-		/* Poll the next port in a 50th of a second.
-		   This spreads the polling of ports out a little bit */
-		mod_timer(&cvm_oct_poll_timer, jiffies + HZ / 50);
-	} else {
+	if (port >= CVMX_PIP_NUM_INPUT_PORTS) {
+		/*
+		 * All ports have been polled. Start the next
+		 * iteration through the ports in one second.
+		 */
 		port = 0;
-		/* All ports have been polled. Start the next iteration through
-		   the ports in one second */
 		mod_timer(&cvm_oct_poll_timer, jiffies + HZ);
+		return;
+	}
+	if (!cvm_oct_device[port])
+		goto out;
+
+	priv = netdev_priv(cvm_oct_device[port]);
+	if (priv->poll) {
+		/* skip polling if we don't get the lock */
+		if (!down_trylock(&mdio_sem)) {
+			priv->poll(cvm_oct_device[port]);
+			up(&mdio_sem);
+		}
+	}
+
+	queues_per_port = cvmx_pko_get_num_queues(port);
+	/* Drain any pending packets in the free list */
+	for (qos = 0; qos < queues_per_port; qos++) {
+		if (skb_queue_len(&priv->tx_free_list[qos]) == 0)
+			continue;
+		skb_to_free = cvmx_fau_fetch_and_add32(priv->fau + qos * 4,
+						       MAX_SKB_TO_FREE);
+		undo = skb_to_free > 0 ?
+			MAX_SKB_TO_FREE : skb_to_free + MAX_SKB_TO_FREE;
+		if (undo > 0)
+			cvmx_fau_atomic_add32(priv->fau+qos*4, -undo);
+		skb_to_free = -skb_to_free > MAX_SKB_TO_FREE ?
+			MAX_SKB_TO_FREE : -skb_to_free;
+		cvm_oct_free_tx_skbs(priv, skb_to_free, qos, 1);
 	}
+	cvm_oct_device[port]->netdev_ops->ndo_get_stats(cvm_oct_device[port]);
+
+out:
+	port++;
+	/* Poll the next port in a 50th of a second.
+	   This spreads the polling of ports out a little bit */
+	mod_timer(&cvm_oct_poll_timer, jiffies + HZ / 50);
 }
 
 /**