]> pilppa.org Git - linux-2.6-omap-h63xx.git/blobdiff - net/sched/sch_generic.c
Merge git://git.linux-nfs.org/pub/linux/nfs-2.6
[linux-2.6-omap-h63xx.git] / net / sched / sch_generic.c
index f4d34480a093d126061920e6f53e7b4fc8dfe83b..fa1a6f45dc416173f8b6d853f9ce0c5fa3565bc7 100644 (file)
  *              - Ingress support
  */
 
-#include <asm/uaccess.h>
-#include <asm/system.h>
 #include <linux/bitops.h>
 #include <linux/module.h>
 #include <linux/types.h>
 #include <linux/kernel.h>
 #include <linux/sched.h>
 #include <linux/string.h>
-#include <linux/mm.h>
-#include <linux/socket.h>
-#include <linux/sockios.h>
-#include <linux/in.h>
 #include <linux/errno.h>
-#include <linux/interrupt.h>
 #include <linux/netdevice.h>
 #include <linux/skbuff.h>
 #include <linux/rtnetlink.h>
 #include <linux/init.h>
 #include <linux/rcupdate.h>
 #include <linux/list.h>
-#include <net/sock.h>
 #include <net/pkt_sched.h>
 
 /* Main transmission queue. */
@@ -59,122 +51,128 @@ void qdisc_unlock_tree(struct net_device *dev)
        spin_unlock_bh(&dev->queue_lock);
 }
 
-/*
-   dev->queue_lock serializes queue accesses for this device
-   AND dev->qdisc pointer itself.
+static inline int qdisc_qlen(struct Qdisc *q)
+{
+       return q->q.qlen;
+}
 
-   netif_tx_lock serializes accesses to device driver.
+static inline int dev_requeue_skb(struct sk_buff *skb, struct net_device *dev,
+                                 struct Qdisc *q)
+{
+       if (unlikely(skb->next))
+               dev->gso_skb = skb;
+       else
+               q->ops->requeue(skb, q);
 
-   dev->queue_lock and netif_tx_lock are mutually exclusive,
-   if one is grabbed, another must be free.
- */
+       netif_schedule(dev);
+       return 0;
+}
 
+static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
+                                             struct Qdisc *q)
+{
+       struct sk_buff *skb;
 
-/* Kick device.
+       if ((skb = dev->gso_skb))
+               dev->gso_skb = NULL;
+       else
+               skb = q->dequeue(q);
 
-   Returns:  0  - queue is empty or throttled.
-           >0  - queue is not empty.
+       return skb;
+}
 
-   NOTE: Called under dev->queue_lock with locally disabled BH.
-*/
+static inline int handle_dev_cpu_collision(struct sk_buff *skb,
+                                          struct net_device *dev,
+                                          struct Qdisc *q)
+{
+       int ret;
 
+       if (unlikely(dev->xmit_lock_owner == smp_processor_id())) {
+               /*
+                * Same CPU holding the lock. It may be a transient
+                * configuration error, when hard_start_xmit() recurses. We
+                * detect it by checking xmit owner and drop the packet when
+                * deadloop is detected. Return OK to try the next skb.
+                */
+               kfree_skb(skb);
+               if (net_ratelimit())
+                       printk(KERN_WARNING "Dead loop on netdevice %s, "
+                              "fix it urgently!\n", dev->name);
+               ret = qdisc_qlen(q);
+       } else {
+               /*
+                * Another cpu is holding lock, requeue & delay xmits for
+                * some time.
+                */
+               __get_cpu_var(netdev_rx_stat).cpu_collision++;
+               ret = dev_requeue_skb(skb, dev, q);
+       }
+
+       return ret;
+}
+
+/*
+ * NOTE: Called under dev->queue_lock with locally disabled BH.
+ *
+ * __LINK_STATE_QDISC_RUNNING guarantees only one CPU can process this
+ * device at a time. dev->queue_lock serializes queue accesses for
+ * this device AND dev->qdisc pointer itself.
+ *
+ *  netif_tx_lock serializes accesses to device driver.
+ *
+ *  dev->queue_lock and netif_tx_lock are mutually exclusive,
+ *  if one is grabbed, another must be free.
+ *
+ * Note, that this procedure can be called by a watchdog timer
+ *
+ * Returns to the caller:
+ *                             0  - queue is empty or throttled.
+ *                             >0 - queue is not empty.
+ *
+ */
 static inline int qdisc_restart(struct net_device *dev)
 {
        struct Qdisc *q = dev->qdisc;
        struct sk_buff *skb;
+       int ret;
 
        /* Dequeue packet */
-       if (((skb = dev->gso_skb)) || ((skb = q->dequeue(q)))) {
-               unsigned nolock = (dev->features & NETIF_F_LLTX);
+       if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
+               return 0;
 
-               dev->gso_skb = NULL;
 
-               /*
-                * When the driver has LLTX set it does its own locking
-                * in start_xmit. No need to add additional overhead by
-                * locking again. These checks are worth it because
-                * even uncongested locks can be quite expensive.
-                * The driver can do trylock like here too, in case
-                * of lock congestion it should return -1 and the packet
-                * will be requeued.
-                */
-               if (!nolock) {
-                       if (!netif_tx_trylock(dev)) {
-                       collision:
-                               /* So, someone grabbed the driver. */
-
-                               /* It may be transient configuration error,
-                                  when hard_start_xmit() recurses. We detect
-                                  it by checking xmit owner and drop the
-                                  packet when deadloop is detected.
-                               */
-                               if (dev->xmit_lock_owner == smp_processor_id()) {
-                                       kfree_skb(skb);
-                                       if (net_ratelimit())
-                                               printk(KERN_DEBUG "Dead loop on netdevice %s, fix it urgently!\n", dev->name);
-                                       goto out;
-                               }
-                               __get_cpu_var(netdev_rx_stat).cpu_collision++;
-                               goto requeue;
-                       }
-               }
+       /* And release queue */
+       spin_unlock(&dev->queue_lock);
 
-               {
-                       /* And release queue */
-                       spin_unlock(&dev->queue_lock);
-
-                       if (!netif_queue_stopped(dev)) {
-                               int ret;
-
-                               ret = dev_hard_start_xmit(skb, dev);
-                               if (ret == NETDEV_TX_OK) {
-                                       if (!nolock) {
-                                               netif_tx_unlock(dev);
-                                       }
-                                       spin_lock(&dev->queue_lock);
-                                       q = dev->qdisc;
-                                       goto out;
-                               }
-                               if (ret == NETDEV_TX_LOCKED && nolock) {
-                                       spin_lock(&dev->queue_lock);
-                                       q = dev->qdisc;
-                                       goto collision;
-                               }
-                       }
+       HARD_TX_LOCK(dev, smp_processor_id());
+       ret = dev_hard_start_xmit(skb, dev);
+       HARD_TX_UNLOCK(dev);
 
-                       /* NETDEV_TX_BUSY - we need to requeue */
-                       /* Release the driver */
-                       if (!nolock) {
-                               netif_tx_unlock(dev);
-                       }
-                       spin_lock(&dev->queue_lock);
-                       q = dev->qdisc;
-               }
+       spin_lock(&dev->queue_lock);
+       q = dev->qdisc;
 
-               /* Device kicked us out :(
-                  This is possible in three cases:
+       switch (ret) {
+       case NETDEV_TX_OK:
+               /* Driver sent out skb successfully */
+               ret = qdisc_qlen(q);
+               break;
 
-                  0. driver is locked
-                  1. fastroute is enabled
-                  2. device cannot determine busy state
-                     before start of transmission (f.e. dialout)
-                  3. device is buggy (ppp)
-                */
+       case NETDEV_TX_LOCKED:
+               /* Driver try lock failed */
+               ret = handle_dev_cpu_collision(skb, dev, q);
+               break;
 
-requeue:
-               if (unlikely(q == &noop_qdisc))
-                       kfree_skb(skb);
-               else if (skb->next)
-                       dev->gso_skb = skb;
-               else
-                       q->ops->requeue(skb, q);
-               netif_schedule(dev);
+       default:
+               /* Driver returned NETDEV_TX_BUSY - requeue skb */
+               if (unlikely (ret != NETDEV_TX_BUSY && net_ratelimit()))
+                       printk(KERN_WARNING "BUG %s code %d qlen %d\n",
+                              dev->name, ret, q->q.qlen);
+
+               ret = dev_requeue_skb(skb, dev, q);
+               break;
        }
-       return 0;
 
-out:
-       BUG_ON((int) q->q.qlen < 0);
-       return q->q.qlen;
+       return ret;
 }
 
 void __qdisc_run(struct net_device *dev)
@@ -243,14 +241,27 @@ static void dev_watchdog_down(struct net_device *dev)
        netif_tx_unlock_bh(dev);
 }
 
+/**
+ *     netif_carrier_on - set carrier
+ *     @dev: network device
+ *
+ * Device has detected that carrier.
+ */
 void netif_carrier_on(struct net_device *dev)
 {
-       if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state))
+       if (test_and_clear_bit(__LINK_STATE_NOCARRIER, &dev->state)) {
                linkwatch_fire_event(dev);
-       if (netif_running(dev))
-               __netdev_watchdog_up(dev);
+               if (netif_running(dev))
+                       __netdev_watchdog_up(dev);
+       }
 }
 
+/**
+ *     netif_carrier_off - clear carrier
+ *     @dev: network device
+ *
+ * Device has detected loss of carrier.
+ */
 void netif_carrier_off(struct net_device *dev)
 {
        if (!test_and_set_bit(__LINK_STATE_NOCARRIER, &dev->state))
@@ -493,9 +504,7 @@ void qdisc_destroy(struct Qdisc *qdisc)
                return;
 
        list_del(&qdisc->list);
-#ifdef CONFIG_NET_ESTIMATOR
        gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
-#endif
        if (ops->reset)
                ops->reset(qdisc);
        if (ops->destroy)
@@ -547,6 +556,7 @@ void dev_deactivate(struct net_device *dev)
 {
        struct Qdisc *qdisc;
        struct sk_buff *skb;
+       int running;
 
        spin_lock_bh(&dev->queue_lock);
        qdisc = dev->qdisc;
@@ -562,12 +572,31 @@ void dev_deactivate(struct net_device *dev)
 
        dev_watchdog_down(dev);
 
-       /* Wait for outstanding dev_queue_xmit calls. */
+       /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
        synchronize_rcu();
 
        /* Wait for outstanding qdisc_run calls. */
-       while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
-               yield();
+       do {
+               while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
+                       yield();
+
+               /*
+                * Double-check inside queue lock to ensure that all effects
+                * of the queue run are visible when we return.
+                */
+               spin_lock_bh(&dev->queue_lock);
+               running = test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
+               spin_unlock_bh(&dev->queue_lock);
+
+               /*
+                * The running flag should never be set at this point because
+                * we've already set dev->qdisc to noop_qdisc *inside* the same
+                * pair of spin locks.  That is, if any qdisc_run starts after
+                * our initial test it should see the noop_qdisc and then
+                * clear the RUNNING bit before dropping the queue lock.  So
+                * if it is set here then we've found a bug.
+                */
+       } while (WARN_ON_ONCE(running));
 }
 
 void dev_init_scheduler(struct net_device *dev)