[NET]: Convert init_timer into setup_timer
[linux-2.6.git] / net / sched / sch_generic.c
index 9a2f8e4..84c048a 100644 (file)
  *              - Ingress support
  */
 
-#include <asm/uaccess.h>
-#include <asm/system.h>
 #include <linux/bitops.h>
-#include <linux/config.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. */
 
-/* Main qdisc structure lock. 
-
-   However, modifications
-   to data, participating in scheduling must be additionally
-   protected with dev->queue_lock spinlock.
-
-   The idea is the following:
-   - enqueue, dequeue are serialized via top level device
-     spinlock dev->queue_lock.
-   - tree walking is protected by read_lock_bh(qdisc_tree_lock)
-     and this lock is used only in process context.
-   - updates to tree are made under rtnl semaphore or
-     from softirq context (__qdisc_destroy rcu-callback)
-     hence this lock needs local bh disabling.
-
-   qdisc_tree_lock must be grabbed BEFORE dev->queue_lock!
+/* Modifications to data participating in scheduling must be protected with
+ * dev->queue_lock spinlock.
+ *
+ * The idea is the following:
+ * - enqueue, dequeue are serialized via top level device
+ *   spinlock dev->queue_lock.
+ * - ingress filtering is serialized via top level device
+ *   spinlock dev->ingress_lock.
+ * - updates to tree and tree walking are only done under the rtnl mutex.
  */
-DEFINE_RWLOCK(qdisc_tree_lock);
 
 void qdisc_lock_tree(struct net_device *dev)
 {
-       write_lock_bh(&qdisc_tree_lock);
        spin_lock_bh(&dev->queue_lock);
+       spin_lock(&dev->ingress_lock);
 }
 
 void qdisc_unlock_tree(struct net_device *dev)
 {
+       spin_unlock(&dev->ingress_lock);
        spin_unlock_bh(&dev->queue_lock);
-       write_unlock_bh(&qdisc_tree_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;
+}
 
-   dev->xmit_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 dev->xmit_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.
-   Note, that this procedure can be called by a watchdog timer, so that
-   we do not check dev->tbusy flag here.
+       if ((skb = dev->gso_skb))
+               dev->gso_skb = NULL;
+       else
+               skb = q->dequeue(q);
 
-   Returns:  0  - queue is empty.
-            >0  - queue is not empty, but throttled.
-           <0  - queue is not empty. Device is throttled, if dev->tbusy != 0.
+       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;
 
-int qdisc_restart(struct net_device *dev)
+       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 = NETDEV_TX_BUSY;
 
        /* Dequeue packet */
-       if ((skb = q->dequeue(q)) != NULL) {
-               unsigned nolock = (dev->features & NETIF_F_LLTX);
-               /*
-                * 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 (!spin_trylock(&dev->xmit_lock)) {
-                       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);
-                                       return -1;
-                               }
-                               __get_cpu_var(netdev_rx_stat).cpu_collision++;
-                               goto requeue;
-                       }
-                       /* Remember that the driver is grabbed by us. */
-                       dev->xmit_lock_owner = smp_processor_id();
-               }
-               
-               {
-                       /* And release queue */
-                       spin_unlock(&dev->queue_lock);
-
-                       if (!netif_queue_stopped(dev)) {
-                               int ret;
-                               if (netdev_nit)
-                                       dev_queue_xmit_nit(skb, dev);
-
-                               ret = dev->hard_start_xmit(skb, dev);
-                               if (ret == NETDEV_TX_OK) { 
-                                       if (!nolock) {
-                                               dev->xmit_lock_owner = -1;
-                                               spin_unlock(&dev->xmit_lock);
-                                       }
-                                       spin_lock(&dev->queue_lock);
-                                       return -1;
-                               }
-                               if (ret == NETDEV_TX_LOCKED && nolock) {
-                                       spin_lock(&dev->queue_lock);
-                                       goto collision; 
-                               }
-                       }
+       if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
+               return 0;
 
-                       /* NETDEV_TX_BUSY - we need to requeue */
-                       /* Release the driver */
-                       if (!nolock) { 
-                               dev->xmit_lock_owner = -1;
-                               spin_unlock(&dev->xmit_lock);
-                       } 
-                       spin_lock(&dev->queue_lock);
-                       q = dev->qdisc;
-               }
 
-               /* Device kicked us out :(
-                  This is possible in three cases:
+       /* And release queue */
+       spin_unlock(&dev->queue_lock);
 
-                  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)
-                */
+       HARD_TX_LOCK(dev, smp_processor_id());
+       if (!netif_subqueue_stopped(dev, skb))
+               ret = dev_hard_start_xmit(skb, dev);
+       HARD_TX_UNLOCK(dev);
 
-requeue:
-               q->ops->requeue(skb, q);
-               netif_schedule(dev);
-               return 1;
+       spin_lock(&dev->queue_lock);
+       q = dev->qdisc;
+
+       switch (ret) {
+       case NETDEV_TX_OK:
+               /* Driver sent out skb successfully */
+               ret = qdisc_qlen(q);
+               break;
+
+       case NETDEV_TX_LOCKED:
+               /* Driver try lock failed */
+               ret = handle_dev_cpu_collision(skb, dev, q);
+               break;
+
+       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 q->q.qlen;
+
+       return ret;
+}
+
+void __qdisc_run(struct net_device *dev)
+{
+       do {
+               if (!qdisc_restart(dev))
+                       break;
+       } while (!netif_queue_stopped(dev));
+
+       clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
 }
 
 static void dev_watchdog(unsigned long arg)
 {
        struct net_device *dev = (struct net_device *)arg;
 
-       spin_lock(&dev->xmit_lock);
+       netif_tx_lock(dev);
        if (dev->qdisc != &noop_qdisc) {
                if (netif_device_present(dev) &&
                    netif_running(dev) &&
                    netif_carrier_ok(dev)) {
                        if (netif_queue_stopped(dev) &&
-                           (jiffies - dev->trans_start) > dev->watchdog_timeo) {
-                               printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n", dev->name);
+                           time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
+
+                               printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
+                                      dev->name);
                                dev->tx_timeout(dev);
                        }
-                       if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
+                       if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
                                dev_hold(dev);
                }
        }
-       spin_unlock(&dev->xmit_lock);
+       netif_tx_unlock(dev);
 
        dev_put(dev);
 }
 
-static void dev_watchdog_init(struct net_device *dev)
-{
-       init_timer(&dev->watchdog_timer);
-       dev->watchdog_timer.data = (unsigned long)dev;
-       dev->watchdog_timer.function = dev_watchdog;
-}
-
 void __netdev_watchdog_up(struct net_device *dev)
 {
        if (dev->tx_timeout) {
                if (dev->watchdog_timeo <= 0)
                        dev->watchdog_timeo = 5*HZ;
-               if (!mod_timer(&dev->watchdog_timer, jiffies + dev->watchdog_timeo))
+               if (!mod_timer(&dev->watchdog_timer,
+                              round_jiffies(jiffies + dev->watchdog_timeo)))
                        dev_hold(dev);
        }
 }
 
 static void dev_watchdog_up(struct net_device *dev)
 {
-       spin_lock_bh(&dev->xmit_lock);
        __netdev_watchdog_up(dev);
-       spin_unlock_bh(&dev->xmit_lock);
 }
 
 static void dev_watchdog_down(struct net_device *dev)
 {
-       spin_lock_bh(&dev->xmit_lock);
+       netif_tx_lock_bh(dev);
        if (del_timer(&dev->watchdog_timer))
-               __dev_put(dev);
-       spin_unlock_bh(&dev->xmit_lock);
+               dev_put(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)) {
+               linkwatch_fire_event(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))
+               linkwatch_fire_event(dev);
 }
 
 /* "NOOP" scheduler: the best scheduler, recommended for all interfaces
@@ -242,31 +267,27 @@ static void dev_watchdog_down(struct net_device *dev)
    cheaper.
  */
 
-static int
-noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
+static int noop_enqueue(struct sk_buff *skb, struct Qdisc * qdisc)
 {
        kfree_skb(skb);
        return NET_XMIT_CN;
 }
 
-static struct sk_buff *
-noop_dequeue(struct Qdisc * qdisc)
+static struct sk_buff *noop_dequeue(struct Qdisc * qdisc)
 {
        return NULL;
 }
 
-static int
-noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
+static int noop_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 {
        if (net_ratelimit())
-               printk(KERN_DEBUG "%s deferred output. It is buggy.\n", skb->dev->name);
+               printk(KERN_DEBUG "%s deferred output. It is buggy.\n",
+                      skb->dev->name);
        kfree_skb(skb);
        return NET_XMIT_CN;
 }
 
 struct Qdisc_ops noop_qdisc_ops = {
-       .next           =       NULL,
-       .cl_ops         =       NULL,
        .id             =       "noop",
        .priv_size      =       0,
        .enqueue        =       noop_enqueue,
@@ -279,13 +300,11 @@ struct Qdisc noop_qdisc = {
        .enqueue        =       noop_enqueue,
        .dequeue        =       noop_dequeue,
        .flags          =       TCQ_F_BUILTIN,
-       .ops            =       &noop_qdisc_ops,        
+       .ops            =       &noop_qdisc_ops,
        .list           =       LIST_HEAD_INIT(noop_qdisc.list),
 };
 
 static struct Qdisc_ops noqueue_qdisc_ops = {
-       .next           =       NULL,
-       .cl_ops         =       NULL,
        .id             =       "noqueue",
        .priv_size      =       0,
        .enqueue        =       noop_enqueue,
@@ -310,97 +329,86 @@ static const u8 prio2band[TC_PRIO_MAX+1] =
    generic prio+fifo combination.
  */
 
-static int
-pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
+#define PFIFO_FAST_BANDS 3
+
+static inline struct sk_buff_head *prio2list(struct sk_buff *skb,
+                                            struct Qdisc *qdisc)
 {
        struct sk_buff_head *list = qdisc_priv(qdisc);
+       return list + prio2band[skb->priority & TC_PRIO_MAX];
+}
 
-       list += prio2band[skb->priority&TC_PRIO_MAX];
+static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
+{
+       struct sk_buff_head *list = prio2list(skb, qdisc);
 
-       if (list->qlen < qdisc->dev->tx_queue_len) {
-               __skb_queue_tail(list, skb);
+       if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
                qdisc->q.qlen++;
-               qdisc->bstats.bytes += skb->len;
-               qdisc->bstats.packets++;
-               return 0;
+               return __qdisc_enqueue_tail(skb, qdisc, list);
        }
-       qdisc->qstats.drops++;
-       kfree_skb(skb);
-       return NET_XMIT_DROP;
+
+       return qdisc_drop(skb, qdisc);
 }
 
-static struct sk_buff *
-pfifo_fast_dequeue(struct Qdisc* qdisc)
+static struct sk_buff *pfifo_fast_dequeue(struct Qdisc* qdisc)
 {
        int prio;
        struct sk_buff_head *list = qdisc_priv(qdisc);
-       struct sk_buff *skb;
 
-       for (prio = 0; prio < 3; prio++, list++) {
-               skb = __skb_dequeue(list);
-               if (skb) {
+       for (prio = 0; prio < PFIFO_FAST_BANDS; prio++) {
+               if (!skb_queue_empty(list + prio)) {
                        qdisc->q.qlen--;
-                       return skb;
+                       return __qdisc_dequeue_head(qdisc, list + prio);
                }
        }
+
        return NULL;
 }
 
-static int
-pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
+static int pfifo_fast_requeue(struct sk_buff *skb, struct Qdisc* qdisc)
 {
-       struct sk_buff_head *list = qdisc_priv(qdisc);
-
-       list += prio2band[skb->priority&TC_PRIO_MAX];
-
-       __skb_queue_head(list, skb);
        qdisc->q.qlen++;
-       qdisc->qstats.requeues++;
-       return 0;
+       return __qdisc_requeue(skb, qdisc, prio2list(skb, qdisc));
 }
 
-static void
-pfifo_fast_reset(struct Qdisc* qdisc)
+static void pfifo_fast_reset(struct Qdisc* qdisc)
 {
        int prio;
        struct sk_buff_head *list = qdisc_priv(qdisc);
 
-       for (prio=0; prio < 3; prio++)
-               skb_queue_purge(list+prio);
+       for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
+               __qdisc_reset_queue(qdisc, list + prio);
+
+       qdisc->qstats.backlog = 0;
        qdisc->q.qlen = 0;
 }
 
 static int pfifo_fast_dump(struct Qdisc *qdisc, struct sk_buff *skb)
 {
-       unsigned char    *b = skb->tail;
-       struct tc_prio_qopt opt;
+       struct tc_prio_qopt opt = { .bands = PFIFO_FAST_BANDS };
 
-       opt.bands = 3; 
        memcpy(&opt.priomap, prio2band, TC_PRIO_MAX+1);
        RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
        return skb->len;
 
 rtattr_failure:
-       skb_trim(skb, b - skb->data);
        return -1;
 }
 
 static int pfifo_fast_init(struct Qdisc *qdisc, struct rtattr *opt)
 {
-       int i;
+       int prio;
        struct sk_buff_head *list = qdisc_priv(qdisc);
 
-       for (i=0; i<3; i++)
-               skb_queue_head_init(list+i);
+       for (prio = 0; prio < PFIFO_FAST_BANDS; prio++)
+               skb_queue_head_init(list + prio);
 
        return 0;
 }
 
 static struct Qdisc_ops pfifo_fast_ops = {
-       .next           =       NULL,
-       .cl_ops         =       NULL,
        .id             =       "pfifo_fast",
-       .priv_size      =       3 * sizeof(struct sk_buff_head),
+       .priv_size      =       PFIFO_FAST_BANDS * sizeof(struct sk_buff_head),
        .enqueue        =       pfifo_fast_enqueue,
        .dequeue        =       pfifo_fast_dequeue,
        .requeue        =       pfifo_fast_requeue,
@@ -410,24 +418,22 @@ static struct Qdisc_ops pfifo_fast_ops = {
        .owner          =       THIS_MODULE,
 };
 
-struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
+struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
 {
        void *p;
        struct Qdisc *sch;
-       int size;
+       unsigned int size;
+       int err = -ENOBUFS;
 
        /* ensure that the Qdisc and the private data are 32-byte aligned */
-       size = ((sizeof(*sch) + QDISC_ALIGN_CONST) & ~QDISC_ALIGN_CONST);
-       size += ops->priv_size + QDISC_ALIGN_CONST;
+       size = QDISC_ALIGN(sizeof(*sch));
+       size += ops->priv_size + (QDISC_ALIGNTO - 1);
 
-       p = kmalloc(size, GFP_KERNEL);
+       p = kzalloc(size, GFP_KERNEL);
        if (!p)
-               return NULL;
-       memset(p, 0, size);
-
-       sch = (struct Qdisc *)(((unsigned long)p + QDISC_ALIGN_CONST) 
-                              & ~QDISC_ALIGN_CONST);
-       sch->padded = (char *)sch - (char *)p;
+               goto errout;
+       sch = (struct Qdisc *) QDISC_ALIGN((unsigned long) p);
+       sch->padded = (char *) sch - (char *) p;
 
        INIT_LIST_HEAD(&sch->list);
        skb_queue_head_init(&sch->q);
@@ -436,13 +442,29 @@ struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops)
        sch->dequeue = ops->dequeue;
        sch->dev = dev;
        dev_hold(dev);
-       sch->stats_lock = &dev->queue_lock;
        atomic_set(&sch->refcnt, 1);
+
+       return sch;
+errout:
+       return ERR_PTR(-err);
+}
+
+struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
+                                unsigned int parentid)
+{
+       struct Qdisc *sch;
+
+       sch = qdisc_alloc(dev, ops);
+       if (IS_ERR(sch))
+               goto errout;
+       sch->stats_lock = &dev->queue_lock;
+       sch->parent = parentid;
+
        if (!ops->init || ops->init(sch, NULL) == 0)
                return sch;
 
-       dev_put(dev);
-       kfree(p);
+       qdisc_destroy(sch);
+errout:
        return NULL;
 }
 
@@ -456,26 +478,12 @@ void qdisc_reset(struct Qdisc *qdisc)
                ops->reset(qdisc);
 }
 
-/* this is the rcu callback function to clean up a qdisc when there 
+/* this is the rcu callback function to clean up a qdisc when there
  * are no further references to it */
 
 static void __qdisc_destroy(struct rcu_head *head)
 {
        struct Qdisc *qdisc = container_of(head, struct Qdisc, q_rcu);
-       struct Qdisc_ops  *ops = qdisc->ops;
-
-#ifdef CONFIG_NET_ESTIMATOR
-       gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
-#endif
-       write_lock(&qdisc_tree_lock);
-       if (ops->reset)
-               ops->reset(qdisc);
-       if (ops->destroy)
-               ops->destroy(qdisc);
-       write_unlock(&qdisc_tree_lock);
-       module_put(ops->owner);
-
-       dev_put(qdisc->dev);
        kfree((char *) qdisc - qdisc->padded);
 }
 
@@ -483,32 +491,21 @@ static void __qdisc_destroy(struct rcu_head *head)
 
 void qdisc_destroy(struct Qdisc *qdisc)
 {
-       struct list_head cql = LIST_HEAD_INIT(cql);
-       struct Qdisc *cq, *q, *n;
+       struct Qdisc_ops  *ops = qdisc->ops;
 
        if (qdisc->flags & TCQ_F_BUILTIN ||
-               !atomic_dec_and_test(&qdisc->refcnt))
+           !atomic_dec_and_test(&qdisc->refcnt))
                return;
 
-       if (!list_empty(&qdisc->list)) {
-               if (qdisc->ops->cl_ops == NULL)
-                       list_del(&qdisc->list);
-               else
-                       list_move(&qdisc->list, &cql);
-       }
-
-       /* unlink inner qdiscs from dev->qdisc_list immediately */
-       list_for_each_entry(cq, &cql, list)
-               list_for_each_entry_safe(q, n, &qdisc->dev->qdisc_list, list)
-                       if (TC_H_MAJ(q->parent) == TC_H_MAJ(cq->handle)) {
-                               if (q->ops->cl_ops == NULL)
-                                       list_del_init(&q->list);
-                               else
-                                       list_move_tail(&q->list, &cql);
-                       }
-       list_for_each_entry_safe(cq, n, &cql, list)
-               list_del_init(&cq->list);
+       list_del(&qdisc->list);
+       gen_kill_estimator(&qdisc->bstats, &qdisc->rate_est);
+       if (ops->reset)
+               ops->reset(qdisc);
+       if (ops->destroy)
+               ops->destroy(qdisc);
 
+       module_put(ops->owner);
+       dev_put(qdisc->dev);
        call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 }
 
@@ -523,20 +520,17 @@ void dev_activate(struct net_device *dev)
        if (dev->qdisc_sleeping == &noop_qdisc) {
                struct Qdisc *qdisc;
                if (dev->tx_queue_len) {
-                       qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops);
+                       qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
+                                                 TC_H_ROOT);
                        if (qdisc == NULL) {
                                printk(KERN_INFO "%s: activation failed\n", dev->name);
                                return;
                        }
-                       write_lock_bh(&qdisc_tree_lock);
                        list_add_tail(&qdisc->list, &dev->qdisc_list);
-                       write_unlock_bh(&qdisc_tree_lock);
                } else {
                        qdisc =  &noqueue_qdisc;
                }
-               write_lock_bh(&qdisc_tree_lock);
                dev->qdisc_sleeping = qdisc;
-               write_unlock_bh(&qdisc_tree_lock);
        }
 
        if (!netif_carrier_ok(dev))
@@ -555,6 +549,8 @@ void dev_activate(struct net_device *dev)
 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;
@@ -562,14 +558,39 @@ void dev_deactivate(struct net_device *dev)
 
        qdisc_reset(qdisc);
 
+       skb = dev->gso_skb;
+       dev->gso_skb = NULL;
        spin_unlock_bh(&dev->queue_lock);
 
+       kfree_skb(skb);
+
        dev_watchdog_down(dev);
 
-       while (test_bit(__LINK_STATE_SCHED, &dev->state))
-               yield();
+       /* Wait for outstanding qdisc-less dev_queue_xmit calls. */
+       synchronize_rcu();
 
-       spin_unlock_wait(&dev->xmit_lock);
+       /* Wait for outstanding qdisc_run calls. */
+       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)
@@ -580,7 +601,7 @@ void dev_init_scheduler(struct net_device *dev)
        INIT_LIST_HEAD(&dev->qdisc_list);
        qdisc_unlock_tree(dev);
 
-       dev_watchdog_init(dev);
+       setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 }
 
 void dev_shutdown(struct net_device *dev)
@@ -593,21 +614,20 @@ void dev_shutdown(struct net_device *dev)
        dev->qdisc_sleeping = &noop_qdisc;
        qdisc_destroy(qdisc);
 #if defined(CONFIG_NET_SCH_INGRESS) || defined(CONFIG_NET_SCH_INGRESS_MODULE)
-        if ((qdisc = dev->qdisc_ingress) != NULL) {
+       if ((qdisc = dev->qdisc_ingress) != NULL) {
                dev->qdisc_ingress = NULL;
                qdisc_destroy(qdisc);
-        }
+       }
 #endif
        BUG_TRAP(!timer_pending(&dev->watchdog_timer));
        qdisc_unlock_tree(dev);
 }
 
-EXPORT_SYMBOL(__netdev_watchdog_up);
+EXPORT_SYMBOL(netif_carrier_on);
+EXPORT_SYMBOL(netif_carrier_off);
 EXPORT_SYMBOL(noop_qdisc);
-EXPORT_SYMBOL(noop_qdisc_ops);
 EXPORT_SYMBOL(qdisc_create_dflt);
 EXPORT_SYMBOL(qdisc_destroy);
 EXPORT_SYMBOL(qdisc_reset);
-EXPORT_SYMBOL(qdisc_restart);
 EXPORT_SYMBOL(qdisc_lock_tree);
 EXPORT_SYMBOL(qdisc_unlock_tree);