add a fix for memleaks caused by failed memory allocations
[openwrt.git] / target / linux / ar7 / files / drivers / net / cpmac.c
index 2c8d529..4b6bfe1 100644 (file)
@@ -37,6 +37,7 @@
 #include <linux/platform_device.h>
 #include <linux/dma-mapping.h>
 #include <asm/gpio.h>
+#include <asm/atomic.h>
 
 MODULE_AUTHOR("Eugene Konev <ejka@imfi.kspu.ru>");
 MODULE_DESCRIPTION("TI AR7 ethernet driver (CPMAC)");
@@ -191,6 +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;
 };
@@ -211,6 +213,7 @@ struct cpmac_priv {
        struct net_device *dev;
        struct work_struct reset_work;
        struct platform_device *pdev;
+       atomic_t reset_pending;
 };
 
 static irqreturn_t cpmac_irq(int, void *);
@@ -244,6 +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;
@@ -414,8 +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);
@@ -423,12 +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);
@@ -436,6 +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);
@@ -444,14 +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)
@@ -460,6 +553,9 @@ static int cpmac_start_xmit(struct sk_buff *skb, struct net_device *dev)
        struct cpmac_desc *desc;
        struct cpmac_priv *priv = netdev_priv(dev);
 
+       if (unlikely(atomic_read(&priv->reset_pending)))
+               return NETDEV_TX_BUSY;
+
        if (unlikely(skb_padto(skb, ETH_ZLEN)))
                return NETDEV_TX_OK;
 
@@ -625,8 +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)
@@ -639,14 +738,14 @@ static void cpmac_clear_tx(struct net_device *dev)
                priv->desc_ring[i].dataflags = 0;
                if (priv->desc_ring[i].skb) {
                        dev_kfree_skb_any(priv->desc_ring[i].skb);
-                       if (netif_subqueue_stopped(dev, i))
-                           netif_wake_subqueue(dev, i);
+                       priv->desc_ring[i].skb = NULL;
                }
        }
 }
 
 static void cpmac_hw_error(struct work_struct *work)
 {
+       int i;
        struct cpmac_priv *priv =
                container_of(work, struct cpmac_priv, reset_work);
 
@@ -655,7 +754,45 @@ static void cpmac_hw_error(struct work_struct *work)
        spin_unlock(&priv->rx_lock);
        cpmac_clear_tx(priv->dev);
        cpmac_hw_start(priv->dev);
-       netif_start_queue(priv->dev);
+       barrier();
+       atomic_dec(&priv->reset_pending);
+       
+       for (i = 0; i < CPMAC_QUEUES; i++) {
+               netif_wake_subqueue(priv->dev, i);
+       }
+       netif_wake_queue(priv->dev);
+}
+
+static void cpmac_check_status(struct net_device *dev)
+{
+       struct cpmac_priv *priv = netdev_priv(dev);
+
+       u32 macstatus = cpmac_read(priv->regs, CPMAC_MAC_STATUS);
+       int rx_channel = (macstatus >> 8) & 7;
+       int rx_code = (macstatus >> 12) & 15;
+       int tx_channel = (macstatus >> 16) & 7;
+       int tx_code = (macstatus >> 20) & 15;
+
+       if (rx_code || tx_code) {
+               if (netif_msg_drv(priv) && net_ratelimit()) {
+                       /* Can't find any documentation on what these error codes actually are.
+                        * So just log them and hope..
+                        */
+                       if (rx_code)
+                               printk(KERN_WARNING "%s: host error %d on rx channel %d (macstatus %08x), resetting\n",
+                                      dev->name, rx_code, rx_channel, macstatus);
+                       if (tx_code)
+                               printk(KERN_WARNING "%s: host error %d on tx channel %d (macstatus %08x), resetting\n",
+                                      dev->name, tx_code, tx_channel, macstatus);
+               }
+               
+               netif_stop_queue(dev);
+               cpmac_hw_stop(dev);
+               if (schedule_work(&priv->reset_work))
+                       atomic_inc(&priv->reset_pending);                       
+               if (unlikely(netif_msg_hw(priv)))
+                       cpmac_dump_regs(dev);
+       }
 }
 
 static irqreturn_t cpmac_irq(int irq, void *dev_id)
@@ -687,48 +824,33 @@ static irqreturn_t cpmac_irq(int irq, void *dev_id)
 
        cpmac_write(priv->regs, CPMAC_MAC_EOI_VECTOR, 0);
 
-       if (unlikely(status & (MAC_INT_HOST | MAC_INT_STATUS))) {
-               if (netif_msg_drv(priv) && net_ratelimit())
-                       printk(KERN_ERR "%s: hw error, resetting...\n",
-                              dev->name);
-               netif_stop_queue(dev);
-               cpmac_hw_stop(dev);
-               schedule_work(&priv->reset_work);
-               if (unlikely(netif_msg_hw(priv)))
-                       cpmac_dump_regs(dev);
-       }
+       if (unlikely(status & (MAC_INT_HOST | MAC_INT_STATUS)))
+               cpmac_check_status(dev);
 
        return IRQ_HANDLED;
 }
 
 static void cpmac_tx_timeout(struct net_device *dev)
 {
-       struct cpmac_priv *priv = netdev_priv(dev);
        int i;
+       struct cpmac_priv *priv = netdev_priv(dev);
 
        spin_lock(&priv->lock);
        dev->stats.tx_errors++;
        spin_unlock(&priv->lock);
        if (netif_msg_tx_err(priv) && net_ratelimit())
                printk(KERN_WARNING "%s: transmit timeout\n", dev->name);
-       /* 
-        * FIXME: waking up random queue is not the best thing to
-        * do... on the other hand why we got here at all?
-        */
-#ifdef CONFIG_NETDEVICES_MULTIQUEUE
-       for (i = 0; i < CPMAC_QUEUES; i++)
-               if (priv->desc_ring[i].skb) {
-                       priv->desc_ring[i].dataflags = 0;
-                       dev_kfree_skb_any(priv->desc_ring[i].skb);
-                       netif_wake_subqueue(dev, i);
-                       break;
-               }
-#else
-       priv->desc_ring[0].dataflags = 0;
-       if (priv->desc_ring[0].skb)
-               dev_kfree_skb_any(priv->desc_ring[0].skb);
-       netif_wake_queue(dev);
-#endif
+
+       atomic_inc(&priv->reset_pending);
+       barrier();
+       cpmac_clear_tx(dev);
+       barrier();
+       atomic_dec(&priv->reset_pending);
+
+       netif_wake_queue(priv->dev);
+       for (i = 0; i < CPMAC_QUEUES; i++) {
+               netif_wake_subqueue(dev, i);
+       }
 }
 
 static int cpmac_ioctl(struct net_device *dev, struct ifreq *ifr, int cmd)
@@ -913,9 +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))
@@ -924,6 +1049,7 @@ static int cpmac_open(struct net_device *dev)
                goto fail_irq;
        }
 
+       atomic_set(&priv->reset_pending, 0);
        INIT_WORK(&priv->reset_work, cpmac_hw_error);
        cpmac_hw_start(dev);
 
@@ -1054,13 +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) {
This page took 0.060879 seconds and 4 git commands to generate.