netdev: Allocate multiple queues for TX.
[deliverable/linux.git] / net / sched / sch_generic.c
index d355e5e47fe3a71aca1d151db1ce72429a069297..4e2b865cbba00b7971d27336c0f5e3ee94b5fd43 100644 (file)
 /* Main transmission queue. */
 
 /* Modifications to data participating in scheduling must be protected with
- * dev->queue_lock spinlock.
+ * queue->lock spinlock.
  *
  * The idea is the following:
  * - enqueue, dequeue are serialized via top level device
- *   spinlock dev->queue_lock.
+ *   spinlock queue->lock.
  * - ingress filtering is serialized via top level device
- *   spinlock dev->ingress_lock.
+ *   spinlock dev->rx_queue.lock.
  * - updates to tree and tree walking are only done under the rtnl mutex.
  */
 
 void qdisc_lock_tree(struct net_device *dev)
-       __acquires(dev->queue_lock)
-       __acquires(dev->ingress_lock)
+       __acquires(dev->rx_queue.lock)
 {
-       spin_lock_bh(&dev->queue_lock);
-       spin_lock(&dev->ingress_lock);
+       unsigned int i;
+
+       local_bh_disable();
+       for (i = 0; i < dev->num_tx_queues; i++) {
+               struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
+               spin_lock(&txq->lock);
+       }
+       spin_lock(&dev->rx_queue.lock);
 }
 EXPORT_SYMBOL(qdisc_lock_tree);
 
 void qdisc_unlock_tree(struct net_device *dev)
-       __releases(dev->ingress_lock)
-       __releases(dev->queue_lock)
+       __releases(dev->rx_queue.lock)
 {
-       spin_unlock(&dev->ingress_lock);
-       spin_unlock_bh(&dev->queue_lock);
+       unsigned int i;
+
+       spin_unlock(&dev->rx_queue.lock);
+       for (i = 0; i < dev->num_tx_queues; i++) {
+               struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
+               spin_unlock(&txq->lock);
+       }
+       local_bh_enable();
 }
 EXPORT_SYMBOL(qdisc_unlock_tree);
 
@@ -62,25 +72,26 @@ static inline int qdisc_qlen(struct Qdisc *q)
        return q->q.qlen;
 }
 
-static inline int dev_requeue_skb(struct sk_buff *skb, struct net_device *dev,
+static inline int dev_requeue_skb(struct sk_buff *skb,
+                                 struct netdev_queue *dev_queue,
                                  struct Qdisc *q)
 {
        if (unlikely(skb->next))
-               dev->gso_skb = skb;
+               dev_queue->gso_skb = skb;
        else
                q->ops->requeue(skb, q);
 
-       netif_schedule(dev);
+       netif_schedule_queue(dev_queue);
        return 0;
 }
 
-static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
-                                             struct Qdisc *q)
+static inline struct sk_buff *dequeue_skb(struct netdev_queue *dev_queue,
+                                         struct Qdisc *q)
 {
        struct sk_buff *skb;
 
-       if ((skb = dev->gso_skb))
-               dev->gso_skb = NULL;
+       if ((skb = dev_queue->gso_skb))
+               dev_queue->gso_skb = NULL;
        else
                skb = q->dequeue(q);
 
@@ -88,12 +99,12 @@ static inline struct sk_buff *dev_dequeue_skb(struct net_device *dev,
 }
 
 static inline int handle_dev_cpu_collision(struct sk_buff *skb,
-                                          struct net_device *dev,
+                                          struct netdev_queue *dev_queue,
                                           struct Qdisc *q)
 {
        int ret;
 
-       if (unlikely(dev->xmit_lock_owner == smp_processor_id())) {
+       if (unlikely(dev_queue->xmit_lock_owner == smp_processor_id())) {
                /*
                 * Same CPU holding the lock. It may be a transient
                 * configuration error, when hard_start_xmit() recurses. We
@@ -103,7 +114,7 @@ static inline int handle_dev_cpu_collision(struct sk_buff *skb,
                kfree_skb(skb);
                if (net_ratelimit())
                        printk(KERN_WARNING "Dead loop on netdevice %s, "
-                              "fix it urgently!\n", dev->name);
+                              "fix it urgently!\n", dev_queue->dev->name);
                ret = qdisc_qlen(q);
        } else {
                /*
@@ -111,22 +122,22 @@ static inline int handle_dev_cpu_collision(struct sk_buff *skb,
                 * some time.
                 */
                __get_cpu_var(netdev_rx_stat).cpu_collision++;
-               ret = dev_requeue_skb(skb, dev, q);
+               ret = dev_requeue_skb(skb, dev_queue, q);
        }
 
        return ret;
 }
 
 /*
- * NOTE: Called under dev->queue_lock with locally disabled BH.
+ * NOTE: Called under 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.
+ * __QUEUE_STATE_QDISC_RUNNING guarantees only one CPU can process
+ * this queue at a time. queue->lock serializes queue accesses for
+ * this queue AND txq->qdisc pointer itself.
  *
  *  netif_tx_lock serializes accesses to device driver.
  *
- *  dev->queue_lock and netif_tx_lock are mutually exclusive,
+ *  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
@@ -136,27 +147,30 @@ static inline int handle_dev_cpu_collision(struct sk_buff *skb,
  *                             >0 - queue is not empty.
  *
  */
-static inline int qdisc_restart(struct net_device *dev)
+static inline int qdisc_restart(struct netdev_queue *txq)
 {
-       struct Qdisc *q = dev->qdisc;
-       struct sk_buff *skb;
+       struct Qdisc *q = txq->qdisc;
        int ret = NETDEV_TX_BUSY;
+       struct net_device *dev;
+       struct sk_buff *skb;
 
        /* Dequeue packet */
-       if (unlikely((skb = dev_dequeue_skb(dev, q)) == NULL))
+       if (unlikely((skb = dequeue_skb(txq, q)) == NULL))
                return 0;
 
 
        /* And release queue */
-       spin_unlock(&dev->queue_lock);
+       spin_unlock(&txq->lock);
 
-       HARD_TX_LOCK(dev, smp_processor_id());
+       dev = txq->dev;
+
+       HARD_TX_LOCK(dev, txq, smp_processor_id());
        if (!netif_subqueue_stopped(dev, skb))
                ret = dev_hard_start_xmit(skb, dev);
-       HARD_TX_UNLOCK(dev);
+       HARD_TX_UNLOCK(dev, txq);
 
-       spin_lock(&dev->queue_lock);
-       q = dev->qdisc;
+       spin_lock(&txq->lock);
+       q = txq->qdisc;
 
        switch (ret) {
        case NETDEV_TX_OK:
@@ -166,7 +180,7 @@ static inline int qdisc_restart(struct net_device *dev)
 
        case NETDEV_TX_LOCKED:
                /* Driver try lock failed */
-               ret = handle_dev_cpu_collision(skb, dev, q);
+               ret = handle_dev_cpu_collision(skb, txq, q);
                break;
 
        default:
@@ -175,18 +189,19 @@ static inline int qdisc_restart(struct net_device *dev)
                        printk(KERN_WARNING "BUG %s code %d qlen %d\n",
                               dev->name, ret, q->q.qlen);
 
-               ret = dev_requeue_skb(skb, dev, q);
+               ret = dev_requeue_skb(skb, txq, q);
                break;
        }
 
        return ret;
 }
 
-void __qdisc_run(struct net_device *dev)
+void __qdisc_run(struct netdev_queue *txq)
 {
+       struct net_device *dev = txq->dev;
        unsigned long start_time = jiffies;
 
-       while (qdisc_restart(dev)) {
+       while (qdisc_restart(txq)) {
                if (netif_queue_stopped(dev))
                        break;
 
@@ -196,12 +211,12 @@ void __qdisc_run(struct net_device *dev)
                 * 2. we've been doing it for too long.
                 */
                if (need_resched() || jiffies != start_time) {
-                       netif_schedule(dev);
+                       netif_schedule_queue(txq);
                        break;
                }
        }
 
-       clear_bit(__LINK_STATE_QDISC_RUNNING, &dev->state);
+       clear_bit(__QUEUE_STATE_QDISC_RUNNING, &txq->state);
 }
 
 static void dev_watchdog(unsigned long arg)
@@ -209,19 +224,35 @@ static void dev_watchdog(unsigned long arg)
        struct net_device *dev = (struct net_device *)arg;
 
        netif_tx_lock(dev);
-       if (dev->qdisc != &noop_qdisc) {
+       if (!qdisc_tx_is_noop(dev)) {
                if (netif_device_present(dev) &&
                    netif_running(dev) &&
                    netif_carrier_ok(dev)) {
-                       if (netif_queue_stopped(dev) &&
-                           time_after(jiffies, dev->trans_start + dev->watchdog_timeo)) {
+                       int some_queue_stopped = 0;
+                       unsigned int i;
+
+                       for (i = 0; i < dev->num_tx_queues; i++) {
+                               struct netdev_queue *txq;
 
-                               printk(KERN_INFO "NETDEV WATCHDOG: %s: transmit timed out\n",
+                               txq = netdev_get_tx_queue(dev, i);
+                               if (netif_tx_queue_stopped(txq)) {
+                                       some_queue_stopped = 1;
+                                       break;
+                               }
+                       }
+
+                       if (some_queue_stopped &&
+                           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);
                                WARN_ON_ONCE(1);
                        }
-                       if (!mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + dev->watchdog_timeo)))
+                       if (!mod_timer(&dev->watchdog_timer,
+                                      round_jiffies(jiffies +
+                                                    dev->watchdog_timeo)))
                                dev_hold(dev);
                }
        }
@@ -364,7 +395,7 @@ static int pfifo_fast_enqueue(struct sk_buff *skb, struct Qdisc* qdisc)
 {
        struct sk_buff_head *list = prio2list(skb, qdisc);
 
-       if (skb_queue_len(list) < qdisc->dev->tx_queue_len) {
+       if (skb_queue_len(list) < qdisc_dev(qdisc)->tx_queue_len) {
                qdisc->q.qlen++;
                return __qdisc_enqueue_tail(skb, qdisc, list);
        }
@@ -440,7 +471,8 @@ static struct Qdisc_ops pfifo_fast_ops __read_mostly = {
        .owner          =       THIS_MODULE,
 };
 
-struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
+struct Qdisc *qdisc_alloc(struct netdev_queue *dev_queue,
+                         struct Qdisc_ops *ops)
 {
        void *p;
        struct Qdisc *sch;
@@ -462,24 +494,25 @@ struct Qdisc *qdisc_alloc(struct net_device *dev, struct Qdisc_ops *ops)
        sch->ops = ops;
        sch->enqueue = ops->enqueue;
        sch->dequeue = ops->dequeue;
-       sch->dev = dev;
-       dev_hold(dev);
+       sch->dev_queue = dev_queue;
+       dev_hold(qdisc_dev(sch));
        atomic_set(&sch->refcnt, 1);
 
        return sch;
 errout:
-       return ERR_PTR(-err);
+       return ERR_PTR(err);
 }
 
-struct Qdisc * qdisc_create_dflt(struct net_device *dev, struct Qdisc_ops *ops,
+struct Qdisc * qdisc_create_dflt(struct net_device *dev,
+                                struct netdev_queue *dev_queue,
+                                struct Qdisc_ops *ops,
                                 unsigned int parentid)
 {
        struct Qdisc *sch;
 
-       sch = qdisc_alloc(dev, ops);
+       sch = qdisc_alloc(dev_queue, ops);
        if (IS_ERR(sch))
                goto errout;
-       sch->stats_lock = &dev->queue_lock;
        sch->parent = parentid;
 
        if (!ops->init || ops->init(sch, NULL) == 0)
@@ -491,7 +524,7 @@ errout:
 }
 EXPORT_SYMBOL(qdisc_create_dflt);
 
-/* Under dev->queue_lock and BH! */
+/* Under queue->lock and BH! */
 
 void qdisc_reset(struct Qdisc *qdisc)
 {
@@ -511,7 +544,7 @@ static void __qdisc_destroy(struct rcu_head *head)
        kfree((char *) qdisc - qdisc->padded);
 }
 
-/* Under dev->queue_lock and BH! */
+/* Under queue->lock and BH! */
 
 void qdisc_destroy(struct Qdisc *qdisc)
 {
@@ -529,65 +562,135 @@ void qdisc_destroy(struct Qdisc *qdisc)
                ops->destroy(qdisc);
 
        module_put(ops->owner);
-       dev_put(qdisc->dev);
+       dev_put(qdisc_dev(qdisc));
        call_rcu(&qdisc->q_rcu, __qdisc_destroy);
 }
 EXPORT_SYMBOL(qdisc_destroy);
 
+static bool dev_all_qdisc_sleeping_noop(struct net_device *dev)
+{
+       unsigned int i;
+
+       for (i = 0; i < dev->num_tx_queues; i++) {
+               struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
+
+               if (txq->qdisc_sleeping != &noop_qdisc)
+                       return false;
+       }
+       return true;
+}
+
+static void attach_one_default_qdisc(struct net_device *dev,
+                                    struct netdev_queue *dev_queue,
+                                    void *_unused)
+{
+       struct Qdisc *qdisc;
+
+       if (dev->tx_queue_len) {
+               qdisc = qdisc_create_dflt(dev, dev_queue,
+                                         &pfifo_fast_ops, TC_H_ROOT);
+               if (!qdisc) {
+                       printk(KERN_INFO "%s: activation failed\n", dev->name);
+                       return;
+               }
+               list_add_tail(&qdisc->list, &dev_queue->qdisc_list);
+       } else {
+               qdisc =  &noqueue_qdisc;
+       }
+       dev_queue->qdisc_sleeping = qdisc;
+}
+
+static void transition_one_qdisc(struct net_device *dev,
+                                struct netdev_queue *dev_queue,
+                                void *_need_watchdog)
+{
+       int *need_watchdog_p = _need_watchdog;
+
+       spin_lock_bh(&dev_queue->lock);
+       rcu_assign_pointer(dev_queue->qdisc, dev_queue->qdisc_sleeping);
+       if (dev_queue->qdisc != &noqueue_qdisc)
+               *need_watchdog_p = 1;
+       spin_unlock_bh(&dev_queue->lock);
+}
+
 void dev_activate(struct net_device *dev)
 {
+       int need_watchdog;
+
        /* No queueing discipline is attached to device;
           create default one i.e. pfifo_fast for devices,
           which need queueing and noqueue_qdisc for
           virtual interfaces
         */
 
-       if (dev->qdisc_sleeping == &noop_qdisc) {
-               struct Qdisc *qdisc;
-               if (dev->tx_queue_len) {
-                       qdisc = qdisc_create_dflt(dev, &pfifo_fast_ops,
-                                                 TC_H_ROOT);
-                       if (qdisc == NULL) {
-                               printk(KERN_INFO "%s: activation failed\n", dev->name);
-                               return;
-                       }
-                       list_add_tail(&qdisc->list, &dev->qdisc_list);
-               } else {
-                       qdisc =  &noqueue_qdisc;
-               }
-               dev->qdisc_sleeping = qdisc;
-       }
+       if (dev_all_qdisc_sleeping_noop(dev))
+               netdev_for_each_tx_queue(dev, attach_one_default_qdisc, NULL);
 
        if (!netif_carrier_ok(dev))
                /* Delay activation until next carrier-on event */
                return;
 
-       spin_lock_bh(&dev->queue_lock);
-       rcu_assign_pointer(dev->qdisc, dev->qdisc_sleeping);
-       if (dev->qdisc != &noqueue_qdisc) {
+       need_watchdog = 0;
+       netdev_for_each_tx_queue(dev, transition_one_qdisc, &need_watchdog);
+
+       if (need_watchdog) {
                dev->trans_start = jiffies;
                dev_watchdog_up(dev);
        }
-       spin_unlock_bh(&dev->queue_lock);
 }
 
-void dev_deactivate(struct net_device *dev)
+static void dev_deactivate_queue(struct net_device *dev,
+                                struct netdev_queue *dev_queue,
+                                void *_qdisc_default)
 {
+       struct Qdisc *qdisc_default = _qdisc_default;
        struct Qdisc *qdisc;
        struct sk_buff *skb;
-       int running;
 
-       spin_lock_bh(&dev->queue_lock);
-       qdisc = dev->qdisc;
-       dev->qdisc = &noop_qdisc;
+       spin_lock_bh(&dev_queue->lock);
 
-       qdisc_reset(qdisc);
+       qdisc = dev_queue->qdisc;
+       if (qdisc) {
+               dev_queue->qdisc = qdisc_default;
+               qdisc_reset(qdisc);
+       }
+       skb = dev_queue->gso_skb;
+       dev_queue->gso_skb = NULL;
 
-       skb = dev->gso_skb;
-       dev->gso_skb = NULL;
-       spin_unlock_bh(&dev->queue_lock);
+       spin_unlock_bh(&dev_queue->lock);
 
        kfree_skb(skb);
+}
+
+static bool some_qdisc_is_running(struct net_device *dev, int lock)
+{
+       unsigned int i;
+
+       for (i = 0; i < dev->num_tx_queues; i++) {
+               struct netdev_queue *dev_queue;
+               int val;
+
+               dev_queue = netdev_get_tx_queue(dev, i);
+
+               if (lock)
+                       spin_lock_bh(&dev_queue->lock);
+
+               val = test_bit(__QUEUE_STATE_QDISC_RUNNING, &dev_queue->state);
+
+               if (lock)
+                       spin_unlock_bh(&dev_queue->lock);
+
+               if (val)
+                       return true;
+       }
+       return false;
+}
+
+void dev_deactivate(struct net_device *dev)
+{
+       bool running;
+
+       netdev_for_each_tx_queue(dev, dev_deactivate_queue, &noop_qdisc);
 
        dev_watchdog_down(dev);
 
@@ -596,16 +699,14 @@ void dev_deactivate(struct net_device *dev)
 
        /* Wait for outstanding qdisc_run calls. */
        do {
-               while (test_bit(__LINK_STATE_QDISC_RUNNING, &dev->state))
+               while (some_qdisc_is_running(dev, 0))
                        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);
+               running = some_qdisc_is_running(dev, 1);
 
                /*
                 * The running flag should never be set at this point because
@@ -618,32 +719,47 @@ void dev_deactivate(struct net_device *dev)
        } while (WARN_ON_ONCE(running));
 }
 
+static void dev_init_scheduler_queue(struct net_device *dev,
+                                    struct netdev_queue *dev_queue,
+                                    void *_qdisc)
+{
+       struct Qdisc *qdisc = _qdisc;
+
+       dev_queue->qdisc = qdisc;
+       dev_queue->qdisc_sleeping = qdisc;
+       INIT_LIST_HEAD(&dev_queue->qdisc_list);
+}
+
 void dev_init_scheduler(struct net_device *dev)
 {
        qdisc_lock_tree(dev);
-       dev->qdisc = &noop_qdisc;
-       dev->qdisc_sleeping = &noop_qdisc;
-       INIT_LIST_HEAD(&dev->qdisc_list);
+       netdev_for_each_tx_queue(dev, dev_init_scheduler_queue, &noop_qdisc);
+       dev_init_scheduler_queue(dev, &dev->rx_queue, NULL);
        qdisc_unlock_tree(dev);
 
        setup_timer(&dev->watchdog_timer, dev_watchdog, (unsigned long)dev);
 }
 
-void dev_shutdown(struct net_device *dev)
+static void shutdown_scheduler_queue(struct net_device *dev,
+                                    struct netdev_queue *dev_queue,
+                                    void *_qdisc_default)
 {
-       struct Qdisc *qdisc;
+       struct Qdisc *qdisc = dev_queue->qdisc_sleeping;
+       struct Qdisc *qdisc_default = _qdisc_default;
+
+       if (qdisc) {
+               dev_queue->qdisc = qdisc_default;
+               dev_queue->qdisc_sleeping = qdisc_default;
 
-       qdisc_lock_tree(dev);
-       qdisc = dev->qdisc_sleeping;
-       dev->qdisc = &noop_qdisc;
-       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) {
-               dev->qdisc_ingress = NULL;
                qdisc_destroy(qdisc);
        }
-#endif
+}
+
+void dev_shutdown(struct net_device *dev)
+{
+       qdisc_lock_tree(dev);
+       netdev_for_each_tx_queue(dev, shutdown_scheduler_queue, &noop_qdisc);
+       shutdown_scheduler_queue(dev, &dev->rx_queue, NULL);
        BUG_TRAP(!timer_pending(&dev->watchdog_timer));
        qdisc_unlock_tree(dev);
 }
This page took 0.030822 seconds and 5 git commands to generate.