[PATCH v2.6.29 v3 1/5] gianfar: Fix packet drop when out of memory

[PATCH v2.6.29 v3 1/5] gianfar: Fix packet drop when out of memory

[Andy Fleming]

The patch which fixed gianfar so it drops packets when it runs out
of memory left in the code which frees the skb when it drops packets.
Change the code so that we only free the skb if the new skb was successfully
created.

Signed-off-by: Andy Fleming <afleming@freescale.com>
---
 drivers/net/gianfar.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)

diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
index 3e611a6..a6efabc 100644
--- a/drivers/net/gianfar.c
+++ b/drivers/net/gianfar.c
@@ -1732,8 +1732,7 @@ int gfar_clean_rx_ring(struct net_device *dev, int rx_work_limit)
 
 			if (unlikely(!newskb))
 				newskb = skb;
-
-			if (skb)
+			else if (skb)
 				dev_kfree_skb_any(skb);
 		} else {
 			/* Increment the number of packets */
-- 
1.5.4.GIT

[PATCH v2.6.29 v3 2/5] gianfar: Add Scatter Gather support

[Andy Fleming]

From: Haruki Dai <Dai.Haruki@freescale.com>

Scatter Gather support in gianfar driver to handle fragmented frames on
the transmit side.

Signed-off-by: Poonam Aggrwal <poonam.aggrwal@freescale.com>
Signed-off-by: Dai Haruki <dai.haruki@freescale.com>
Signed-off-by: Andy Fleming <afleming@freescale.com>
---
 drivers/net/gianfar.c         |  217 +++++++++++++++++++++++++++-------------
 drivers/net/gianfar.h         |    1 +
 drivers/net/gianfar_ethtool.c |    2 +
 3 files changed, 149 insertions(+), 71 deletions(-)

diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
index a6efabc..094387f 100644
--- a/drivers/net/gianfar.c
+++ b/drivers/net/gianfar.c
@@ -368,7 +368,7 @@ static int gfar_probe(struct of_device *ofdev,
 
 	if (priv->device_flags & FSL_GIANFAR_DEV_HAS_CSUM) {
 		priv->rx_csum_enable = 1;
-		dev->features |= NETIF_F_IP_CSUM;
+		dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_HIGHDMA;
 	} else
 		priv->rx_csum_enable = 0;
 
@@ -426,6 +426,7 @@ static int gfar_probe(struct of_device *ofdev,
 	priv->rx_buffer_size = DEFAULT_RX_BUFFER_SIZE;
 	priv->tx_ring_size = DEFAULT_TX_RING_SIZE;
 	priv->rx_ring_size = DEFAULT_RX_RING_SIZE;
+	priv->num_txbdfree = DEFAULT_TX_RING_SIZE;
 
 	priv->txcoalescing = DEFAULT_TX_COALESCE;
 	priv->txic = DEFAULT_TXIC;
@@ -819,22 +820,26 @@ static void free_skb_resources(struct gfar_private *priv)
 {
 	struct rxbd8 *rxbdp;
 	struct txbd8 *txbdp;
-	int i;
+	int i, j;
 
 	/* Go through all the buffer descriptors and free their data buffers */
 	txbdp = priv->tx_bd_base;
 
 	for (i = 0; i < priv->tx_ring_size; i++) {
+		if (!priv->tx_skbuff[i])
+			continue;
 
-		if (priv->tx_skbuff[i]) {
-			dma_unmap_single(&priv->dev->dev, txbdp->bufPtr,
-					txbdp->length,
-					DMA_TO_DEVICE);
-			dev_kfree_skb_any(priv->tx_skbuff[i]);
-			priv->tx_skbuff[i] = NULL;
+		dma_unmap_single(&priv->dev->dev, txbdp->bufPtr,
+				txbdp->length, DMA_TO_DEVICE);
+		txbdp->lstatus = 0;
+		for (j = 0; j < skb_shinfo(priv->tx_skbuff[i])->nr_frags; j++) {
+			txbdp++;
+			dma_unmap_page(&priv->dev->dev, txbdp->bufPtr,
+					txbdp->length, DMA_TO_DEVICE);
 		}
-
 		txbdp++;
+		dev_kfree_skb_any(priv->tx_skbuff[i]);
+		priv->tx_skbuff[i] = NULL;
 	}
 
 	kfree(priv->tx_skbuff);
@@ -967,6 +972,7 @@ int startup_gfar(struct net_device *dev)
 		priv->rx_skbuff[i] = NULL;
 
 	/* Initialize some variables in our dev structure */
+	priv->num_txbdfree = priv->tx_ring_size;
 	priv->dirty_tx = priv->cur_tx = priv->tx_bd_base;
 	priv->cur_rx = priv->rx_bd_base;
 	priv->skb_curtx = priv->skb_dirtytx = 0;
@@ -1207,28 +1213,84 @@ void inline gfar_tx_vlan(struct sk_buff *skb, struct txfcb *fcb)
 	fcb->vlctl = vlan_tx_tag_get(skb);
 }
 
+static inline struct txbd8 *skip_txbd(struct txbd8 *bdp, int stride,
+			       struct txbd8 *base, int ring_size)
+{
+	struct txbd8 *new_bd = bdp + stride;
+
+	return (new_bd >= (base + ring_size)) ? (new_bd - ring_size) : new_bd;
+}
+
+static inline struct txbd8 *next_txbd(struct txbd8 *bdp, struct txbd8 *base,
+		int ring_size)
+{
+	return skip_txbd(bdp, 1, base, ring_size);
+}
+
 /* This is called by the kernel when a frame is ready for transmission. */
 /* It is pointed to by the dev->hard_start_xmit function pointer */
 static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct gfar_private *priv = netdev_priv(dev);
 	struct txfcb *fcb = NULL;
-	struct txbd8 *txbdp, *base;
+	struct txbd8 *txbdp, *txbdp_start, *base;
 	u32 lstatus;
+	int i;
+	u32 bufaddr;
 	unsigned long flags;
+	unsigned int nr_frags, length;
+
+	base = priv->tx_bd_base;
+
+	/* total number of fragments in the SKB */
+	nr_frags = skb_shinfo(skb)->nr_frags;
+
+	spin_lock_irqsave(&priv->txlock, flags);
+
+	/* check if there is space to queue this packet */
+	if (nr_frags > priv->num_txbdfree) {
+		/* no space, stop the queue */
+		netif_stop_queue(dev);
+		dev->stats.tx_fifo_errors++;
+		spin_unlock_irqrestore(&priv->txlock, flags);
+		return NETDEV_TX_BUSY;
+	}
 
 	/* Update transmit stats */
 	dev->stats.tx_bytes += skb->len;
 
-	/* Lock priv now */
-	spin_lock_irqsave(&priv->txlock, flags);
+	txbdp = txbdp_start = priv->cur_tx;
 
-	/* Point at the first free tx descriptor */
-	txbdp = priv->cur_tx;
-	base = priv->tx_bd_base;
+	if (nr_frags == 0) {
+		lstatus = txbdp->lstatus | BD_LFLAG(TXBD_LAST | TXBD_INTERRUPT);
+	} else {
+		/* Place the fragment addresses and lengths into the TxBDs */
+		for (i = 0; i < nr_frags; i++) {
+			/* Point at the next BD, wrapping as needed */
+			txbdp = next_txbd(txbdp, base, priv->tx_ring_size);
+
+			length = skb_shinfo(skb)->frags[i].size;
+
+			lstatus = txbdp->lstatus | length |
+				BD_LFLAG(TXBD_READY);
+
+			/* Handle the last BD specially */
+			if (i == nr_frags - 1)
+				lstatus |= BD_LFLAG(TXBD_LAST | TXBD_INTERRUPT);
+
+			bufaddr = dma_map_page(&dev->dev,
+					skb_shinfo(skb)->frags[i].page,
+					skb_shinfo(skb)->frags[i].page_offset,
+					length,
+					DMA_TO_DEVICE);
+
+			/* set the TxBD length and buffer pointer */
+			txbdp->bufPtr = bufaddr;
+			txbdp->lstatus = lstatus;
+		}
 
-	/* Clear all but the WRAP status flags */
-	lstatus = txbdp->lstatus & BD_LFLAG(TXBD_WRAP);
+		lstatus = txbdp_start->lstatus;
+	}
 
 	/* Set up checksumming */
 	if (CHECKSUM_PARTIAL == skb->ip_summed) {
@@ -1246,48 +1308,45 @@ static int gfar_start_xmit(struct sk_buff *skb, struct net_device *dev)
 		gfar_tx_vlan(skb, fcb);
 	}
 
-	/* Set buffer length and pointer */
-	txbdp->bufPtr = dma_map_single(&dev->dev, skb->data,
-			skb->len, DMA_TO_DEVICE);
-
-	/* Save the skb pointer so we can free it later */
+	/* setup the TxBD length and buffer pointer for the first BD */
 	priv->tx_skbuff[priv->skb_curtx] = skb;
+	txbdp_start->bufPtr = dma_map_single(&dev->dev, skb->data,
+			skb_headlen(skb), DMA_TO_DEVICE);
 
-	/* Update the current skb pointer (wrapping if this was the last) */
-	priv->skb_curtx =
-	    (priv->skb_curtx + 1) & TX_RING_MOD_MASK(priv->tx_ring_size);
-
-	/* Flag the BD as ready, interrupt-causing, last, and in need of CRC */
-	lstatus |=
-		BD_LFLAG(TXBD_READY | TXBD_LAST | TXBD_CRC | TXBD_INTERRUPT) |
-		skb->len;
-
-	dev->trans_start = jiffies;
+	lstatus |= BD_LFLAG(TXBD_CRC | TXBD_READY) | skb_headlen(skb);
 
-	/* The powerpc-specific eieio() is used, as wmb() has too strong
+	/*
+	 * The powerpc-specific eieio() is used, as wmb() has too strong
 	 * semantics (it requires synchronization between cacheable and
 	 * uncacheable mappings, which eieio doesn't provide and which we
 	 * don't need), thus requiring a more expensive sync instruction.  At
 	 * some point, the set of architecture-independent barrier functions
 	 * should be expanded to include weaker barriers.
 	 */
-
 	eieio();
-	txbdp->lstatus = lstatus;
 
-	txbdp = next_bd(txbdp, base, priv->tx_ring_size);
+	txbdp_start->lstatus = lstatus;
+
+	/* Update the current skb pointer to the next entry we will use 
+	 * (wrapping if necessary) */
+	priv->skb_curtx = (priv->skb_curtx + 1) &
+		TX_RING_MOD_MASK(priv->tx_ring_size);
+
+	priv->cur_tx = next_txbd(txbdp, base, priv->tx_ring_size);
+
+	/* reduce TxBD free count */
+	priv->num_txbdfree -= (nr_frags + 1);
+
+	dev->trans_start = jiffies;
 
 	/* If the next BD still needs to be cleaned up, then the bds
 	   are full.  We need to tell the kernel to stop sending us stuff. */
-	if (txbdp == priv->dirty_tx) {
+	if (!priv->num_txbdfree) {
 		netif_stop_queue(dev);
 
 		dev->stats.tx_fifo_errors++;
 	}
 
-	/* Update the current txbd to the next one */
-	priv->cur_tx = txbdp;
-
 	/* Tell the DMA to go go go */
 	gfar_write(&priv->regs->tstat, TSTAT_CLEAR_THALT);
 
@@ -1461,50 +1520,66 @@ static void gfar_timeout(struct net_device *dev)
 /* Interrupt Handler for Transmit complete */
 static int gfar_clean_tx_ring(struct net_device *dev)
 {
-	struct txbd8 *bdp, *base;
 	struct gfar_private *priv = netdev_priv(dev);
+	struct txbd8 *bdp;
+	struct txbd8 *lbdp = NULL;
+	struct txbd8 *base = priv->tx_bd_base;
+	struct sk_buff *skb;
+	int skb_dirtytx;
+	int tx_ring_size = priv->tx_ring_size;
+	int frags = 0;
+	int i;
 	int howmany = 0;
+	u32 lstatus;
 
 	bdp = priv->dirty_tx;
-	base = priv->tx_bd_base;
-	while ((bdp->status & TXBD_READY) == 0) {
-		/* If dirty_tx and cur_tx are the same, then either the */
-		/* ring is empty or full now (it could only be full in the beginning, */
-		/* obviously).  If it is empty, we are done. */
-		if ((bdp == priv->cur_tx) && (netif_queue_stopped(dev) == 0))
-			break;
+	skb_dirtytx = priv->skb_dirtytx;
 
-		howmany++;
+	while ((skb = priv->tx_skbuff[skb_dirtytx])) {
+		frags = skb_shinfo(skb)->nr_frags;
+		lbdp = skip_txbd(bdp, frags, base, tx_ring_size);
 
-		/* Deferred means some collisions occurred during transmit, */
-		/* but we eventually sent the packet. */
-		if (bdp->status & TXBD_DEF)
-			dev->stats.collisions++;
+		lstatus = lbdp->lstatus;
 
-		/* Unmap the DMA memory */
-		dma_unmap_single(&priv->dev->dev, bdp->bufPtr,
-				bdp->length, DMA_TO_DEVICE);
+		/* Only clean completed frames */
+		if ((lstatus & BD_LFLAG(TXBD_READY)) &&
+				(lstatus & BD_LENGTH_MASK))
+			break;
+
+		dma_unmap_single(&dev->dev,
+				bdp->bufPtr,
+				bdp->length,
+				DMA_TO_DEVICE);
 
-		/* Free the sk buffer associated with this TxBD */
-		dev_kfree_skb_irq(priv->tx_skbuff[priv->skb_dirtytx]);
+		bdp->lstatus &= BD_LFLAG(TXBD_WRAP);
+		bdp = next_txbd(bdp, base, tx_ring_size);
 
-		priv->tx_skbuff[priv->skb_dirtytx] = NULL;
-		priv->skb_dirtytx =
-		    (priv->skb_dirtytx +
-		     1) & TX_RING_MOD_MASK(priv->tx_ring_size);
+		for (i = 0; i < frags; i++) {
+			dma_unmap_page(&dev->dev,
+					bdp->bufPtr,
+					bdp->length,
+					DMA_TO_DEVICE);
+			bdp->lstatus &= BD_LFLAG(TXBD_WRAP);
+			bdp = next_txbd(bdp, base, tx_ring_size);
+		}
 
-		/* Clean BD length for empty detection */
-		bdp->length = 0;
+		dev_kfree_skb_any(skb);
+		priv->tx_skbuff[skb_dirtytx] = NULL;
 
-		bdp = next_bd(bdp, base, priv->tx_ring_size);
+		skb_dirtytx = (skb_dirtytx + 1) &
+			TX_RING_MOD_MASK(tx_ring_size);
+
+		howmany++;
+		priv->num_txbdfree += frags + 1;
+	}
 
-		/* Move dirty_tx to be the next bd */
-		priv->dirty_tx = bdp;
+	/* If we freed a buffer, we can restart transmission, if necessary */
+	if (netif_queue_stopped(dev) && priv->num_txbdfree)
+		netif_wake_queue(dev);
 
-		/* We freed a buffer, so now we can restart transmission */
-		if (netif_queue_stopped(dev))
-			netif_wake_queue(dev);
-	} /* while ((bdp->status & TXBD_READY) == 0) */
+	/* Update dirty indicators */
+	priv->skb_dirtytx = skb_dirtytx;
+	priv->dirty_tx = bdp;
 
 	dev->stats.tx_packets += howmany;
 
diff --git a/drivers/net/gianfar.h b/drivers/net/gianfar.h
index 9c8974d..7ef3cc5 100644
--- a/drivers/net/gianfar.h
+++ b/drivers/net/gianfar.h
@@ -728,6 +728,7 @@ struct gfar_private {
 	struct txbd8 *dirty_tx;		/* First buffer in line
 					   to be transmitted */
 	unsigned int tx_ring_size;
+	unsigned int num_txbdfree;	/* number of TxBDs free */
 
 	/* RX Locked fields */
 	spinlock_t rxlock;
diff --git a/drivers/net/gianfar_ethtool.c b/drivers/net/gianfar_ethtool.c
index 3021057..59b3b5d 100644
--- a/drivers/net/gianfar_ethtool.c
+++ b/drivers/net/gianfar_ethtool.c
@@ -475,6 +475,7 @@ static int gfar_sringparam(struct net_device *dev, struct ethtool_ringparam *rva
 	/* Change the size */
 	priv->rx_ring_size = rvals->rx_pending;
 	priv->tx_ring_size = rvals->tx_pending;
+	priv->num_txbdfree = priv->tx_ring_size;
 
 	/* Rebuild the rings with the new size */
 	if (dev->flags & IFF_UP) {
@@ -623,6 +624,7 @@ const struct ethtool_ops gfar_ethtool_ops = {
 	.get_tx_csum = gfar_get_tx_csum,
 	.set_rx_csum = gfar_set_rx_csum,
 	.set_tx_csum = gfar_set_tx_csum,
+	.set_sg = ethtool_op_set_sg,
 	.get_msglevel = gfar_get_msglevel,
 	.set_msglevel = gfar_set_msglevel,
 #ifdef CONFIG_PM
-- 
1.5.4.GIT

[PATCH v2.6.29 v3 3/5] gianfar: Use interface name in interrupt name to distinguish the source.

[Andy Fleming]

From: Haruki Dai <Dai.Haruki@freescale.com>

Interface name (ex. eth0) is used as the prefix for the interrupt name,
with _rx, _tx, and _er appended to distinguish multiple interrupts on
the same interface.

Signed-off-by: Dai Haruki <dai.haruki@freescale.com>
Signed-off-by: Andy Fleming <afleming@freescale.com>
---
 drivers/net/gianfar.c |   30 ++++++++++++++++++++++++------
 drivers/net/gianfar.h |    7 +++++++
 2 files changed, 31 insertions(+), 6 deletions(-)

diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
index 094387f..4d7c225 100644
--- a/drivers/net/gianfar.c
+++ b/drivers/net/gianfar.c
@@ -304,8 +304,9 @@ static int gfar_probe(struct of_device *ofdev,
 	u32 tempval;
 	struct net_device *dev = NULL;
 	struct gfar_private *priv = NULL;
-	int err = 0;
 	DECLARE_MAC_BUF(mac);
+	int err = 0;
+	int len_devname;
 
 	/* Create an ethernet device instance */
 	dev = alloc_etherdev(sizeof (*priv));
@@ -447,6 +448,23 @@ static int gfar_probe(struct of_device *ofdev,
 		goto register_fail;
 	}
 
+	/* fill out IRQ number and name fields */
+	len_devname = strlen(dev->name);
+	strncpy(&priv->int_name_tx[0], dev->name, len_devname);
+	if (priv->device_flags & FSL_GIANFAR_DEV_HAS_MULTI_INTR) {
+		strncpy(&priv->int_name_tx[len_devname],
+			"_tx", sizeof("_tx") + 1);
+
+		strncpy(&priv->int_name_rx[0], dev->name, len_devname);
+		strncpy(&priv->int_name_rx[len_devname],
+			"_rx", sizeof("_rx") + 1);
+
+		strncpy(&priv->int_name_er[0], dev->name, len_devname);
+		strncpy(&priv->int_name_er[len_devname],
+			"_er", sizeof("_er") + 1);
+	} else
+		priv->int_name_tx[len_devname] = '\0';
+
 	/* Create all the sysfs files */
 	gfar_init_sysfs(dev);
 
@@ -1020,7 +1038,7 @@ int startup_gfar(struct net_device *dev)
 		/* Install our interrupt handlers for Error,
 		 * Transmit, and Receive */
 		if (request_irq(priv->interruptError, gfar_error,
-				0, "enet_error", dev) < 0) {
+				0, priv->int_name_er, dev) < 0) {
 			if (netif_msg_intr(priv))
 				printk(KERN_ERR "%s: Can't get IRQ %d\n",
 					dev->name, priv->interruptError);
@@ -1030,7 +1048,7 @@ int startup_gfar(struct net_device *dev)
 		}
 
 		if (request_irq(priv->interruptTransmit, gfar_transmit,
-				0, "enet_tx", dev) < 0) {
+				0, priv->int_name_tx, dev) < 0) {
 			if (netif_msg_intr(priv))
 				printk(KERN_ERR "%s: Can't get IRQ %d\n",
 					dev->name, priv->interruptTransmit);
@@ -1041,7 +1059,7 @@ int startup_gfar(struct net_device *dev)
 		}
 
 		if (request_irq(priv->interruptReceive, gfar_receive,
-				0, "enet_rx", dev) < 0) {
+				0, priv->int_name_rx, dev) < 0) {
 			if (netif_msg_intr(priv))
 				printk(KERN_ERR "%s: Can't get IRQ %d (receive0)\n",
 						dev->name, priv->interruptReceive);
@@ -1051,10 +1069,10 @@ int startup_gfar(struct net_device *dev)
 		}
 	} else {
 		if (request_irq(priv->interruptTransmit, gfar_interrupt,
-				0, "gfar_interrupt", dev) < 0) {
+				0, priv->int_name_tx, dev) < 0) {
 			if (netif_msg_intr(priv))
 				printk(KERN_ERR "%s: Can't get IRQ %d\n",
-					dev->name, priv->interruptError);
+					dev->name, priv->interruptTransmit);
 
 			err = -1;
 			goto err_irq_fail;
diff --git a/drivers/net/gianfar.h b/drivers/net/gianfar.h
index 7ef3cc5..06bac34 100644
--- a/drivers/net/gianfar.h
+++ b/drivers/net/gianfar.h
@@ -374,6 +374,8 @@ extern const char gfar_driver_version[];
 #define RXFCB_PERR_MASK		0x000c
 #define RXFCB_PERR_BADL3	0x0008
 
+#define GFAR_INT_NAME_MAX	IFNAMSIZ + 4
+
 struct txbd8
 {
 	union {
@@ -796,6 +798,11 @@ struct gfar_private {
 	uint32_t msg_enable;
 
 	struct work_struct reset_task;
+
+	char int_name_tx[GFAR_INT_NAME_MAX];
+	char int_name_rx[GFAR_INT_NAME_MAX];
+	char int_name_er[GFAR_INT_NAME_MAX];
+
 	/* Network Statistics */
 	struct gfar_extra_stats extra_stats;
 };
-- 
1.5.4.GIT

[PATCH v2.6.29 v3 4/5] gianfar: Merge Tx and Rx interrupt for scheduling clean up ring

[Andy Fleming]

From: Haruki Dai <Dai.Haruki@freescale.com>

No clean up function is executed in the interrupt context by this patch.

Signed-off-by: Dai Haruki <dai.haruki@freescale.com>
Acked-by: Andy Fleming <afleming@freescale.com>
---
 drivers/net/gianfar.c |   59 +++++++++++++++----------------------------------
 drivers/net/gianfar.h |    2 +-
 2 files changed, 19 insertions(+), 42 deletions(-)

diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
index 4d7c225..939e38c 100644
--- a/drivers/net/gianfar.c
+++ b/drivers/net/gianfar.c
@@ -1604,29 +1604,19 @@ static int gfar_clean_tx_ring(struct net_device *dev)
 	return howmany;
 }
 
-/* Interrupt Handler for Transmit complete */
-static irqreturn_t gfar_transmit(int irq, void *dev_id)
+static void gfar_schedule_cleanup(struct net_device *dev)
 {
-	struct net_device *dev = (struct net_device *) dev_id;
 	struct gfar_private *priv = netdev_priv(dev);
-
-	/* Clear IEVENT */
-	gfar_write(&priv->regs->ievent, IEVENT_TX_MASK);
-
-	/* Lock priv */
-	spin_lock(&priv->txlock);
-
-	gfar_clean_tx_ring(dev);
-
-	/* If we are coalescing the interrupts, reset the timer */
-	/* Otherwise, clear it */
-	if (likely(priv->txcoalescing)) {
-		gfar_write(&priv->regs->txic, 0);
-		gfar_write(&priv->regs->txic, priv->txic);
+	if (netif_rx_schedule_prep(dev, &priv->napi)) {
+		gfar_write(&priv->regs->imask, IMASK_RTX_DISABLED);
+		__netif_rx_schedule(dev, &priv->napi);
 	}
+}
 
-	spin_unlock(&priv->txlock);
-
+/* Interrupt Handler for Transmit complete */
+static irqreturn_t gfar_transmit(int irq, void *dev_id)
+{
+	gfar_schedule_cleanup((struct net_device *)dev_id);
 	return IRQ_HANDLED;
 }
 
@@ -1713,28 +1703,7 @@ static inline void count_errors(unsigned short status, struct net_device *dev)
 
 irqreturn_t gfar_receive(int irq, void *dev_id)
 {
-	struct net_device *dev = (struct net_device *) dev_id;
-	struct gfar_private *priv = netdev_priv(dev);
-	u32 tempval;
-
-	/* support NAPI */
-	/* Clear IEVENT, so interrupts aren't called again
-	 * because of the packets that have already arrived */
-	gfar_write(&priv->regs->ievent, IEVENT_RTX_MASK);
-
-	if (netif_rx_schedule_prep(dev, &priv->napi)) {
-		tempval = gfar_read(&priv->regs->imask);
-		tempval &= IMASK_RTX_DISABLED;
-		gfar_write(&priv->regs->imask, tempval);
-
-		__netif_rx_schedule(dev, &priv->napi);
-	} else {
-		if (netif_msg_rx_err(priv))
-			printk(KERN_DEBUG "%s: receive called twice (%x)[%x]\n",
-				dev->name, gfar_read(&priv->regs->ievent),
-				gfar_read(&priv->regs->imask));
-	}
-
+	gfar_schedule_cleanup((struct net_device *)dev_id);
 	return IRQ_HANDLED;
 }
 
@@ -1877,6 +1846,10 @@ static int gfar_poll(struct napi_struct *napi, int budget)
 	int howmany;
 	unsigned long flags;
 
+	/* Clear IEVENT, so interrupts aren't called again
+	 * because of the packets that have already arrived */
+	gfar_write(&priv->regs->ievent, IEVENT_RTX_MASK);
+
 	/* If we fail to get the lock, don't bother with the TX BDs */
 	if (spin_trylock_irqsave(&priv->txlock, flags)) {
 		gfar_clean_tx_ring(dev);
@@ -1899,6 +1872,10 @@ static int gfar_poll(struct napi_struct *napi, int budget)
 			gfar_write(&priv->regs->rxic, 0);
 			gfar_write(&priv->regs->rxic, priv->rxic);
 		}
+		if (likely(priv->txcoalescing)) {
+			gfar_write(&priv->regs->txic, 0);
+			gfar_write(&priv->regs->txic, priv->txic);
+		}
 	}
 
 	return howmany;
diff --git a/drivers/net/gianfar.h b/drivers/net/gianfar.h
index 06bac34..b1a8334 100644
--- a/drivers/net/gianfar.h
+++ b/drivers/net/gianfar.h
@@ -251,7 +251,7 @@ extern const char gfar_driver_version[];
 #define IEVENT_FIQ		0x00000004
 #define IEVENT_DPE		0x00000002
 #define IEVENT_PERR		0x00000001
-#define IEVENT_RX_MASK          (IEVENT_RXB0 | IEVENT_RXF0)
+#define IEVENT_RX_MASK          (IEVENT_RXB0 | IEVENT_RXF0 | IEVENT_BSY)
 #define IEVENT_TX_MASK          (IEVENT_TXB | IEVENT_TXF)
 #define IEVENT_RTX_MASK         (IEVENT_RX_MASK | IEVENT_TX_MASK)
 #define IEVENT_ERR_MASK         \
-- 
1.5.4.GIT

[PATCH v2.6.29 v3 5/5] gianfar: Continue polling until both tx and rx are empty

[Andy Fleming]

gfar_poll would declare polling done once the rx queue was empty,
but the tx queue could still have packets left.

Stolen mostly from the e1000 driver.

Signed-off-by: Andy Fleming <afleming@freescale.com>
---
 drivers/net/gianfar.c |   14 +++++++++-----
 1 files changed, 9 insertions(+), 5 deletions(-)

diff --git a/drivers/net/gianfar.c b/drivers/net/gianfar.c
index 939e38c..144f941 100644
--- a/drivers/net/gianfar.c
+++ b/drivers/net/gianfar.c
@@ -1843,7 +1843,8 @@ static int gfar_poll(struct napi_struct *napi, int budget)
 {
 	struct gfar_private *priv = container_of(napi, struct gfar_private, napi);
 	struct net_device *dev = priv->dev;
-	int howmany;
+	int tx_cleaned = 0;
+	int rx_cleaned = 0;
 	unsigned long flags;
 
 	/* Clear IEVENT, so interrupts aren't called again
@@ -1852,13 +1853,16 @@ static int gfar_poll(struct napi_struct *napi, int budget)
 
 	/* If we fail to get the lock, don't bother with the TX BDs */
 	if (spin_trylock_irqsave(&priv->txlock, flags)) {
-		gfar_clean_tx_ring(dev);
+		tx_cleaned = gfar_clean_tx_ring(dev);
 		spin_unlock_irqrestore(&priv->txlock, flags);
 	}
 
-	howmany = gfar_clean_rx_ring(dev, budget);
+	rx_cleaned = gfar_clean_rx_ring(dev, budget);
 
-	if (howmany < budget) {
+	if (tx_cleaned)
+		return budget;
+
+	if (rx_cleaned < budget) {
 		netif_rx_complete(dev, napi);
 
 		/* Clear the halt bit in RSTAT */
@@ -1878,7 +1882,7 @@ static int gfar_poll(struct napi_struct *napi, int budget)
 		}
 	}
 
-	return howmany;
+	return rx_cleaned;
 }
 
 #ifdef CONFIG_NET_POLL_CONTROLLER
-- 
1.5.4.GIT

Re: [PATCH v2.6.29 v3 1/5] gianfar: Fix packet drop when out of memory

[David Miller]

From: Andy Fleming <afleming@freescale.com>
Date: Wed, 17 Dec 2008 15:42:51 -0600

> The patch which fixed gianfar so it drops packets when it runs out
> of memory left in the code which frees the skb when it drops packets.
> Change the code so that we only free the skb if the new skb was successfully
> created.
> 
> Signed-off-by: Andy Fleming <afleming@freescale.com>

Applied.

Re: [PATCH v2.6.29 v3 2/5] gianfar: Add Scatter Gather support

[David Miller]

From: Andy Fleming <afleming@freescale.com>
Date: Wed, 17 Dec 2008 15:42:52 -0600

> Scatter Gather support in gianfar driver to handle fragmented frames on
> the transmit side.
> 
> Signed-off-by: Poonam Aggrwal <poonam.aggrwal@freescale.com>
> Signed-off-by: Dai Haruki <dai.haruki@freescale.com>
> Signed-off-by: Andy Fleming <afleming@freescale.com>

Applied.

Re: [PATCH v2.6.29 v3 3/5] gianfar: Use interface name in interrupt name to distinguish the source.

[David Miller]

From: Andy Fleming <afleming@freescale.com>
Date: Wed, 17 Dec 2008 15:42:53 -0600

> Interface name (ex. eth0) is used as the prefix for the interrupt name,
> with _rx, _tx, and _er appended to distinguish multiple interrupts on
> the same interface.
> 
> Signed-off-by: Dai Haruki <dai.haruki@freescale.com>
> Signed-off-by: Andy Fleming <afleming@freescale.com>

Applied.

Re: [PATCH v2.6.29 v3 4/5] gianfar: Merge Tx and Rx interrupt for scheduling clean up ring

[David Miller]

From: Andy Fleming <afleming@freescale.com>
Date: Wed, 17 Dec 2008 15:42:54 -0600

> No clean up function is executed in the interrupt context by this patch.
> 
> Signed-off-by: Dai Haruki <dai.haruki@freescale.com>
> Acked-by: Andy Fleming <afleming@freescale.com>

Applied.

Re: [PATCH v2.6.29 v3 5/5] gianfar: Continue polling until both tx and rx are empty

[David Miller]

From: Andy Fleming <afleming@freescale.com>
Date: Wed, 17 Dec 2008 15:42:55 -0600

> gfar_poll would declare polling done once the rx queue was empty,
> but the tx queue could still have packets left.
> 
> Stolen mostly from the e1000 driver.
> 
> Signed-off-by: Andy Fleming <afleming@freescale.com>

Applied.

But I absolutely do not recommend this polling technique at all.

The best scheme is to first purge your TX ring completely, and do not
apply any quota to this work.  It's relatively cheap and batches well.

Then, you process the RX ring and apply the quota only to the RX ring
work.

This is what every NAPI driver I've written does and it gives the best
results.

Re: [PATCH v2.6.29 v3 5/5] gianfar: Continue polling until both tx and rx are empty

[Fleming Andy-AFLEMING]

On Dec 17, 2008, at 18:55, "David Miller" <davem@davemloft.net> wrote:

> From: Andy Fleming <afleming@freescale.com>
> Date: Wed, 17 Dec 2008 15:42:55 -0600
>
>> gfar_poll would declare polling done once the rx queue was empty,
>> but the tx queue could still have packets left.
>>
>> Stolen mostly from the e1000 driver.
>>
>> Signed-off-by: Andy Fleming <afleming@freescale.com>
>
> Applied.
>
> But I absolutely do not recommend this polling technique at all.
>
> The best scheme is to first purge your TX ring completely, and do not
> apply any quota to this work.  It's relatively cheap and batches well.
>
> Then, you process the RX ring and apply the quota only to the RX ring
> work.
>
> This is what every NAPI driver I've written does and it gives the best
> results.

I agree it seems a bit screwy, but I got a 4-5% performance increase.   
It may just require better tuning of my tx coalescing parameters-I was  
getting 7x the number of tx irqs as rx.

Re: [PATCH v2.6.29 v3 5/5] gianfar: Continue polling until both tx and rx are empty

[David Miller]

From: "Fleming Andy-AFLEMING" <afleming@freescale.com>
Date: Wed, 17 Dec 2008 20:40:20 -0600

> It may just require better tuning of my tx coalescing parameters-I
> was getting 7x the number of tx irqs as rx.

It's hard to decrease TX irqs, but yes some tuning in that area
might help.