[PATCH 2/2] Reimplement rx ring with EOQ markers to avoid reset storms
authorFelix Fietkau <nbd@openwrt.org>
Fri, 7 Dec 2007 00:03:10 +0000 (00:03 +0000)
committerFelix Fietkau <nbd@openwrt.org>
Fri, 7 Dec 2007 00:03:10 +0000 (00:03 +0000)
Signed-off-by: oliver@opencloud.com
SVN-Revision: 9665

target/linux/ar7/files/drivers/net/cpmac.c

index 45bd1eeb8e0ac32d04c1baca6094d5a9675735e6..4b6bfe1be6914dd52c83d4c429daa5f42da521ec 100644 (file)
@@ -192,6 +192,7 @@ struct cpmac_desc {
 #define CPMAC_EOQ                      0x1000
        struct sk_buff *skb;
        struct cpmac_desc *next;
+       struct cpmac_desc *prev;
        dma_addr_t mapping;
        dma_addr_t data_mapping;
 };
@@ -246,6 +247,16 @@ static void cpmac_dump_desc(struct net_device *dev, struct cpmac_desc *desc)
        printk("\n");
 }
 
+static void cpmac_dump_all_desc(struct net_device *dev)
+{
+       struct cpmac_priv *priv = netdev_priv(dev);
+       struct cpmac_desc *dump = priv->rx_head;
+       do {
+               cpmac_dump_desc(dev, dump);
+               dump = dump->next;
+       } while (dump != priv->rx_head);
+}
+
 static void cpmac_dump_skb(struct net_device *dev, struct sk_buff *skb)
 {
        int i;
@@ -416,8 +427,8 @@ static struct sk_buff *cpmac_rx_one(struct net_device *dev,
 static int cpmac_poll(struct net_device *dev, int *budget)
 {
        struct sk_buff *skb;
-       struct cpmac_desc *desc;
-       int received = 0, quota = min(dev->quota, *budget);
+       struct cpmac_desc *desc, *restart;
+       int received = 0, processed = 0, quota = min(dev->quota, *budget);
        struct cpmac_priv *priv = netdev_priv(dev);
 
        spin_lock(&priv->rx_lock);
@@ -425,12 +436,31 @@ static int cpmac_poll(struct net_device *dev, int *budget)
                if (netif_msg_rx_err(priv) && net_ratelimit())
                        printk(KERN_WARNING "%s: rx: polling, but no queue\n",
                               dev->name);
+               spin_unlock(&priv->rx_lock);
                netif_rx_complete(dev);
                return 0;
        }
 
        desc = priv->rx_head;
+       restart = NULL;
        while ((received < quota) && ((desc->dataflags & CPMAC_OWN) == 0)) {
+               processed++;
+
+               if ((desc->dataflags & CPMAC_EOQ) != 0) {
+                       /* The last update to eoq->hw_next didn't happen soon enough, and the
+                        * receiver stopped here. Remember this descriptor so we can restart
+                        * the receiver after freeing some space.
+                        */
+                       if (unlikely(restart)) {
+                               if (netif_msg_rx_err(priv))
+                                       printk(KERN_ERR "%s: poll found a duplicate EOQ: %p and %p\n",
+                                              dev->name, restart, desc);
+                               goto fatal_error;
+                       }
+
+                       restart = desc->next;
+               }
+                       
                skb = cpmac_rx_one(dev, priv, desc);
                if (likely(skb)) {
                        netif_receive_skb(skb);
@@ -438,6 +468,44 @@ static int cpmac_poll(struct net_device *dev, int *budget)
                }
                desc = desc->next;
        }
+       
+       if (desc != priv->rx_head) {
+               /* We freed some buffers, but not the whole ring, add what we did free to the rx list */
+               desc->prev->hw_next = (u32)0;
+               priv->rx_head->prev->hw_next = priv->rx_head->mapping;
+       }
+       
+       /* Optimization: If we did not actually process an EOQ (perhaps because of
+        * quota limits), check to see if the tail of the queue has EOQ set. We
+        * should immediately restart in that case so that the receiver can restart
+        * and run in parallel with more packet processing. This lets us handle slightly
+        * larger bursts before running out of ring space (assuming dev->weight < ring_size)
+        */
+       if (!restart && 
+           (priv->rx_head->prev->dataflags & (CPMAC_OWN|CPMAC_EOQ)) == CPMAC_EOQ &&
+           (priv->rx_head->dataflags & CPMAC_OWN) != 0) {
+               /* reset EOQ so the poll loop (above) doesn't try to restart this when it
+                * eventually gets to this descriptor.
+                */
+               priv->rx_head->prev->dataflags &= ~CPMAC_EOQ;
+               restart = priv->rx_head;
+       }
+
+       if (restart) {
+               dev->stats.rx_errors++;
+               dev->stats.rx_fifo_errors++;
+               if (netif_msg_rx_err(priv) && net_ratelimit())
+                       printk(KERN_WARNING "%s: rx dma ring overrun\n", dev->name);
+               
+               if (unlikely((restart->dataflags & CPMAC_OWN) == 0)) {
+                       if (netif_msg_drv(priv))
+                               printk(KERN_ERR "%s: cpmac_poll is trying to restart rx from a descriptor that's not free: %p\n",
+                                      dev->name, restart);
+                       goto fatal_error;
+               }
+
+               cpmac_write(priv->regs, CPMAC_RX_PTR(0), restart->mapping);
+       }
 
        priv->rx_head = desc;
        spin_unlock(&priv->rx_lock);
@@ -446,14 +514,37 @@ static int cpmac_poll(struct net_device *dev, int *budget)
        if (unlikely(netif_msg_rx_status(priv)))
                printk(KERN_DEBUG "%s: poll processed %d packets\n", dev->name,
                       received);
-       if (desc->dataflags & CPMAC_OWN) {
+       
+       if (processed == 0) {
+               /* we ran out of packets to read, revert to interrupt-driven mode */
                netif_rx_complete(dev);
-               cpmac_write(priv->regs, CPMAC_RX_PTR(0), (u32)desc->mapping);
                cpmac_write(priv->regs, CPMAC_RX_INT_ENABLE, 1);
                return 0;
        }
 
        return 1;
+
+fatal_error:
+       /* Something went horribly wrong. Reset hardware to try to recover rather than wedging. */
+
+       if (netif_msg_drv(priv)) {
+               printk(KERN_ERR "%s: cpmac_poll is confused. Resetting hardware\n", dev->name);
+               cpmac_dump_all_desc(dev);
+               printk(KERN_DEBUG "%s: RX_PTR(0)=0x%08x RX_ACK(0)=0x%08x\n",
+                      dev->name,
+                      cpmac_read(priv->regs, CPMAC_RX_PTR(0)),
+                      cpmac_read(priv->regs, CPMAC_RX_ACK(0)));
+       }
+
+       spin_unlock(&priv->rx_lock);
+       netif_rx_complete(dev);
+       netif_stop_queue(dev);
+
+       atomic_inc(&priv->reset_pending);
+       cpmac_hw_stop(dev);
+       if (!schedule_work(&priv->reset_work))
+               atomic_dec(&priv->reset_pending);
+       return 0;
 }
 
 static int cpmac_start_xmit(struct sk_buff *skb, struct net_device *dev)
@@ -630,8 +721,11 @@ static void cpmac_clear_rx(struct net_device *dev)
                        desc->dataflags = CPMAC_OWN;
                        dev->stats.rx_dropped++;
                }
+               desc->hw_next = desc->next->mapping;
                desc = desc->next;
        }
+
+       priv->rx_head->prev->hw_next = 0;
 }
 
 static void cpmac_clear_tx(struct net_device *dev)
@@ -941,9 +1035,12 @@ static int cpmac_open(struct net_device *dev)
                desc->buflen = CPMAC_SKB_SIZE;
                desc->dataflags = CPMAC_OWN;
                desc->next = &priv->rx_head[(i + 1) % priv->ring_size];
+               desc->next->prev = desc;
                desc->hw_next = (u32)desc->next->mapping;
        }
 
+       priv->rx_head->prev->hw_next = (u32)0;
+
        if ((res = request_irq(dev->irq, cpmac_irq, IRQF_SHARED,
                               dev->name, dev))) {
                if (netif_msg_drv(priv))
@@ -1083,13 +1180,13 @@ static int __devinit cpmac_probe(struct platform_device *pdev)
        dev->tx_timeout         = cpmac_tx_timeout;
        dev->ethtool_ops        = &cpmac_ethtool_ops;
        dev->poll = cpmac_poll;
-       dev->weight = 64;
        dev->features |= NETIF_F_MULTI_QUEUE;
 
        spin_lock_init(&priv->lock);
        spin_lock_init(&priv->rx_lock);
        priv->dev = dev;
        priv->ring_size = 64;
+       dev->weight = max(4, priv->ring_size/4);
        priv->msg_enable = netif_msg_init(debug_level, 0xff);
        memcpy(dev->dev_addr, pdata->dev_addr, sizeof(dev->dev_addr));
        if (phy_id == 31) {