[PATCH] skge: transmit complete via IRQ not NAPI

The transmit side code has a number of ring problems that caused some
of the Bugzilla reports. Rather than trying to fix the details, it is safer
to rewrite the code that handles transmit completion and freeing.
Signed-off-by: Stephen Hemminger <shemminger@osdl.org>
Signed-off-by: Jeff Garzik <jeff@garzik.org>

drivers/net/skge.c

@@ -2303,21 +2303,20 @@ static int skge_xmit_frame(struct sk_buff *skb, struct net_device *dev)
 {
 	struct skge_port *skge = netdev_priv(dev);
 	struct skge_hw *hw = skge->hw;
-	struct skge_ring *ring = &skge->tx_ring;
 	struct skge_element *e;
 	struct skge_tx_desc *td;
 	int i;
 	u32 control, len;
 	u64 map;
+	unsigned long flags;
 
 	skb = skb_padto(skb, ETH_ZLEN);
 	if (!skb)
 		return NETDEV_TX_OK;
 
-	if (!spin_trylock(&skge->tx_lock)) {
+	if (!spin_trylock_irqsave(&skge->tx_lock, flags))
 		/* Collision - tell upper layer to requeue */
 		return NETDEV_TX_LOCKED;
-	}
 
 	if (unlikely(skge_avail(&skge->tx_ring) < skb_shinfo(skb)->nr_frags + 1)) {
 		if (!netif_queue_stopped(dev)) {
@@ -2326,12 +2325,13 @@ static int skge_xmit_frame(struct sk_buff *skb, struct net_device *dev)
 			printk(KERN_WARNING PFX "%s: ring full when queue awake!\n",
 			       dev->name);
 		}
-		spin_unlock(&skge->tx_lock);
+		spin_unlock_irqrestore(&skge->tx_lock, flags);
 		return NETDEV_TX_BUSY;
 	}
 
-	e = ring->to_use;
+	e = skge->tx_ring.to_use;
 	td = e->desc;
+	BUG_ON(td->control & BMU_OWN);
 	e->skb = skb;
 	len = skb_headlen(skb);
 	map = pci_map_single(hw->pdev, skb->data, len, PCI_DMA_TODEVICE);
@@ -2372,8 +2372,10 @@ static int skge_xmit_frame(struct sk_buff *skb, struct net_device *dev)
 					   frag->size, PCI_DMA_TODEVICE);
 
 			e = e->next;
-			e->skb = NULL;
+			e->skb = skb;
 			tf = e->desc;
+			BUG_ON(tf->control & BMU_OWN);
+
 			tf->dma_lo = map;
 			tf->dma_hi = (u64) map >> 32;
 			pci_unmap_addr_set(e, mapaddr, map);
@@ -2390,56 +2392,68 @@ static int skge_xmit_frame(struct sk_buff *skb, struct net_device *dev)
 
 	skge_write8(hw, Q_ADDR(txqaddr[skge->port], Q_CSR), CSR_START);
 
-	if (netif_msg_tx_queued(skge))
+	if (unlikely(netif_msg_tx_queued(skge)))
 		printk(KERN_DEBUG "%s: tx queued, slot %td, len %d\n",
-		       dev->name, e - ring->start, skb->len);
+		       dev->name, e - skge->tx_ring.start, skb->len);
 
-	ring->to_use = e->next;
+	skge->tx_ring.to_use = e->next;
 	if (skge_avail(&skge->tx_ring) <= TX_LOW_WATER) {
 		pr_debug("%s: transmit queue full\n", dev->name);
 		netif_stop_queue(dev);
 	}
 
-	mmiowb();
-	spin_unlock(&skge->tx_lock);
+	spin_unlock_irqrestore(&skge->tx_lock, flags);
 
 	dev->trans_start = jiffies;
 
 	return NETDEV_TX_OK;
 }
 
-static void skge_tx_complete(struct skge_port *skge, struct skge_element *last)
+
+/* Free resources associated with this reing element */
+static void skge_tx_free(struct skge_port *skge, struct skge_element *e,
+			 u32 control)
 {
 	struct pci_dev *pdev = skge->hw->pdev;
-	struct skge_element *e;
 
-	for (e = skge->tx_ring.to_clean; e != last; e = e->next) {
-		struct sk_buff *skb = e->skb;
-		int i;
+	BUG_ON(!e->skb);
 
-		e->skb = NULL;
+	/* skb header vs. fragment */
+	if (control & BMU_STF)
 		pci_unmap_single(pdev, pci_unmap_addr(e, mapaddr),
-				 skb_headlen(skb), PCI_DMA_TODEVICE);
-
-		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
-			e = e->next;
+				 pci_unmap_len(e, maplen),
+				 PCI_DMA_TODEVICE);
+	else
 		pci_unmap_page(pdev, pci_unmap_addr(e, mapaddr),
-				       skb_shinfo(skb)->frags[i].size,
+			       pci_unmap_len(e, maplen),
 			       PCI_DMA_TODEVICE);
-		}
 
-		dev_kfree_skb(skb);
+	if (control & BMU_EOF) {
+		if (unlikely(netif_msg_tx_done(skge)))
+			printk(KERN_DEBUG PFX "%s: tx done slot %td\n",
+			       skge->netdev->name, e - skge->tx_ring.start);
+
+		dev_kfree_skb_any(e->skb);
 	}
-	skge->tx_ring.to_clean = e;
+	e->skb = NULL;
 }
 
+/* Free all buffers in transmit ring */
 static void skge_tx_clean(struct skge_port *skge)
 {
+	struct skge_element *e;
+	unsigned long flags;
+
+	spin_lock_irqsave(&skge->tx_lock, flags);
+	for (e = skge->tx_ring.to_clean; e != skge->tx_ring.to_use; e = e->next) {
+		struct skge_tx_desc *td = e->desc;
+		skge_tx_free(skge, e, td->control);
+		td->control = 0;
+	}
 
-	spin_lock_bh(&skge->tx_lock);
-	skge_tx_complete(skge, skge->tx_ring.to_use);
+	skge->tx_ring.to_clean = e;
 	netif_wake_queue(skge->netdev);
-	spin_unlock_bh(&skge->tx_lock);
+	spin_unlock_irqrestore(&skge->tx_lock, flags);
 }
 
 static void skge_tx_timeout(struct net_device *dev)
@@ -2665,32 +2679,28 @@ static inline struct sk_buff *skge_rx_get(struct skge_port *skge,
 	return NULL;
 }
 
-static void skge_tx_done(struct skge_port *skge)
+/* Free all buffers in Tx ring which are no longer owned by device */
+static void skge_txirq(struct net_device *dev)
 {
+	struct skge_port *skge = netdev_priv(dev);
 	struct skge_ring *ring = &skge->tx_ring;
-	struct skge_element *e, *last;
+	struct skge_element *e;
+
+	rmb();
 
 	spin_lock(&skge->tx_lock);
-	last = ring->to_clean;
 	for (e = ring->to_clean; e != ring->to_use; e = e->next) {
 		struct skge_tx_desc *td = e->desc;
 
 		if (td->control & BMU_OWN)
 			break;
 
-		if (td->control & BMU_EOF) {
-			last = e->next;
-			if (unlikely(netif_msg_tx_done(skge)))
-				printk(KERN_DEBUG PFX "%s: tx done slot %td\n",
-				       skge->netdev->name, e - ring->start);
+		skge_tx_free(skge, e, td->control);
 	}
-	}
-
-	skge_tx_complete(skge, last);
-
-	skge_write8(skge->hw, Q_ADDR(txqaddr[skge->port], Q_CSR), CSR_IRQ_CL_F);
+	skge->tx_ring.to_clean = e;
 
-	if (skge_avail(&skge->tx_ring) > TX_LOW_WATER)
+	if (netif_queue_stopped(skge->netdev)
+	    && skge_avail(&skge->tx_ring) > TX_LOW_WATER)
 		netif_wake_queue(skge->netdev);
 
 	spin_unlock(&skge->tx_lock);
@@ -2705,8 +2715,6 @@ static int skge_poll(struct net_device *dev, int *budget)
 	int to_do = min(dev->quota, *budget);
 	int work_done = 0;
 
-	skge_tx_done(skge);
-
 	for (e = ring->to_clean; prefetch(e->next), work_done < to_do; e = e->next) {
 		struct skge_rx_desc *rd = e->desc;
 		struct sk_buff *skb;
@@ -2738,10 +2746,12 @@ static int skge_poll(struct net_device *dev, int *budget)
 		return 1; /* not done */
 
 	netif_rx_complete(dev);
-	mmiowb();
 
-  	hw->intr_mask |= skge->port == 0 ? (IS_R1_F|IS_XA1_F) : (IS_R2_F|IS_XA2_F);
+	spin_lock_irq(&hw->hw_lock);
+	hw->intr_mask |= rxirqmask[skge->port];
   	skge_write32(hw, B0_IMSK, hw->intr_mask);
+	mmiowb();
+	spin_unlock_irq(&hw->hw_lock);
 
 	return 0;
 }
@@ -2871,8 +2881,10 @@ static void skge_extirq(void *arg)
 	}
 	mutex_unlock(&hw->phy_mutex);
 
+	spin_lock_irq(&hw->hw_lock);
 	hw->intr_mask |= IS_EXT_REG;
 	skge_write32(hw, B0_IMSK, hw->intr_mask);
+	spin_unlock_irq(&hw->hw_lock);
 }
 
 static irqreturn_t skge_intr(int irq, void *dev_id, struct pt_regs *regs)
@@ -2885,54 +2897,68 @@ static irqreturn_t skge_intr(int irq, void *dev_id, struct pt_regs *regs)
 	if (status == 0)
 		return IRQ_NONE;
 
+	spin_lock(&hw->hw_lock);
+	status &= hw->intr_mask;
 	if (status & IS_EXT_REG) {
 		hw->intr_mask &= ~IS_EXT_REG;
 		schedule_work(&hw->phy_work);
 	}
 
-	if (status & (IS_R1_F|IS_XA1_F)) {
-		skge_write8(hw, Q_ADDR(Q_R1, Q_CSR), CSR_IRQ_CL_F);
-		hw->intr_mask &= ~(IS_R1_F|IS_XA1_F);
-		netif_rx_schedule(hw->dev[0]);
+	if (status & IS_XA1_F) {
+		skge_write8(hw, Q_ADDR(Q_XA1, Q_CSR), CSR_IRQ_CL_F);
+		skge_txirq(hw->dev[0]);
 	}
 
-	if (status & (IS_R2_F|IS_XA2_F)) {
-		skge_write8(hw, Q_ADDR(Q_R2, Q_CSR), CSR_IRQ_CL_F);
-		hw->intr_mask &= ~(IS_R2_F|IS_XA2_F);
-		netif_rx_schedule(hw->dev[1]);
+	if (status & IS_R1_F) {
+		skge_write8(hw, Q_ADDR(Q_R1, Q_CSR), CSR_IRQ_CL_F);
+		hw->intr_mask &= ~IS_R1_F;
+		netif_rx_schedule(hw->dev[0]);
 	}
 
-	if (likely((status & hw->intr_mask) == 0))
-		return IRQ_HANDLED;
+	if (status & IS_PA_TO_TX1)
+		skge_write16(hw, B3_PA_CTRL, PA_CLR_TO_TX1);
 
 	if (status & IS_PA_TO_RX1) {
 		struct skge_port *skge = netdev_priv(hw->dev[0]);
+
 		++skge->net_stats.rx_over_errors;
 		skge_write16(hw, B3_PA_CTRL, PA_CLR_TO_RX1);
 	}
 
+
+	if (status & IS_MAC1)
+		skge_mac_intr(hw, 0);
+
+	if (hw->dev[1]) {
+		if (status & IS_XA2_F) {
+			skge_write8(hw, Q_ADDR(Q_XA2, Q_CSR), CSR_IRQ_CL_F);
+			skge_txirq(hw->dev[1]);
+		}
+
+		if (status & IS_R2_F) {
+			skge_write8(hw, Q_ADDR(Q_R2, Q_CSR), CSR_IRQ_CL_F);
+			hw->intr_mask &= ~IS_R2_F;
+			netif_rx_schedule(hw->dev[1]);
+		}
+
 		if (status & IS_PA_TO_RX2) {
 			struct skge_port *skge = netdev_priv(hw->dev[1]);
 			++skge->net_stats.rx_over_errors;
 			skge_write16(hw, B3_PA_CTRL, PA_CLR_TO_RX2);
 		}
 
-	if (status & IS_PA_TO_TX1)
-		skge_write16(hw, B3_PA_CTRL, PA_CLR_TO_TX1);
-
 		if (status & IS_PA_TO_TX2)
 			skge_write16(hw, B3_PA_CTRL, PA_CLR_TO_TX2);
 
-	if (status & IS_MAC1)
-		skge_mac_intr(hw, 0);
-
 		if (status & IS_MAC2)
 			skge_mac_intr(hw, 1);
+	}
 
 	if (status & IS_HW_ERR)
 		skge_error_irq(hw);
 
 	skge_write32(hw, B0_IMSK, hw->intr_mask);
+	spin_unlock(&hw->hw_lock);
 
 	return IRQ_HANDLED;
 }
@@ -3083,6 +3109,7 @@ static int skge_reset(struct skge_hw *hw)
 	else
 		hw->ram_size = t8 * 4096;
 
+	spin_lock_init(&hw->hw_lock);
 	hw->intr_mask = IS_HW_ERR | IS_EXT_REG | IS_PORT_1;
 	if (hw->ports > 1)
 		hw->intr_mask |= IS_PORT_2;
@@ -3389,7 +3416,11 @@ static void __devexit skge_remove(struct pci_dev *pdev)
 	dev0 = hw->dev[0];
 	unregister_netdev(dev0);
 
+	spin_lock_irq(&hw->hw_lock);
+	hw->intr_mask = 0;
 	skge_write32(hw, B0_IMSK, 0);
+	spin_unlock_irq(&hw->hw_lock);
+
 	skge_write16(hw, B0_LED, LED_STAT_OFF);
 	skge_write8(hw, B0_CTST, CS_RST_SET);
 

drivers/net/skge.h

@@ -2388,6 +2388,7 @@ struct skge_ring {
 struct skge_hw {
 	void __iomem  	     *regs;
 	struct pci_dev	     *pdev;
+	spinlock_t	     hw_lock;
 	u32		     intr_mask;
 	struct net_device    *dev[2];