netem: use rb tree to implement the time queue
authorEric Dumazet <edumazet@google.com>
Fri, 28 Jun 2013 14:40:57 +0000 (07:40 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 2 Jul 2013 01:07:15 +0000 (18:07 -0700)
Following typical setup to implement a ~100 ms RTT and big
amount of reorders has very poor performance because netem
implements the time queue using a linked list.
-----------------------------------------------------------
ETH=eth0
IFB=ifb0
modprobe ifb
ip link set dev $IFB up
tc qdisc add dev $ETH ingress 2>/dev/null
tc filter add dev $ETH parent ffff: \
   protocol ip u32 match u32 0 0 flowid 1:1 action mirred egress \
   redirect dev $IFB
ethtool -K $ETH gro off tso off gso off
tc qdisc add dev $IFB root netem delay 50ms 10ms limit 100000
tc qd add dev $ETH root netem delay 50ms limit 100000
---------------------------------------------------------

Switch netem time queue to a rb tree, so this kind of setup can work at
high speed.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/sched/sch_netem.c

index 3d2acc7a9c8099e677412276eeafe5a8b3f90fcb..ed0082cf8effd3702ed2f3b4c481948339268565 100644 (file)
@@ -23,6 +23,7 @@
 #include <linux/vmalloc.h>
 #include <linux/rtnetlink.h>
 #include <linux/reciprocal_div.h>
+#include <linux/rbtree.h>
 
 #include <net/netlink.h>
 #include <net/pkt_sched.h>
@@ -68,7 +69,8 @@
 */
 
 struct netem_sched_data {
-       /* internal t(ime)fifo qdisc uses sch->q and sch->limit */
+       /* internal t(ime)fifo qdisc uses t_root and sch->limit */
+       struct rb_root t_root;
 
        /* optional qdisc for classful handling (NULL at netem init) */
        struct Qdisc    *qdisc;
@@ -128,10 +130,35 @@ struct netem_sched_data {
  */
 struct netem_skb_cb {
        psched_time_t   time_to_send;
+       ktime_t         tstamp_save;
 };
 
+/* Because space in skb->cb[] is tight, netem overloads skb->next/prev/tstamp
+ * to hold a rb_node structure.
+ *
+ * If struct sk_buff layout is changed, the following checks will complain.
+ */
+static struct rb_node *netem_rb_node(struct sk_buff *skb)
+{
+       BUILD_BUG_ON(offsetof(struct sk_buff, next) != 0);
+       BUILD_BUG_ON(offsetof(struct sk_buff, prev) !=
+                    offsetof(struct sk_buff, next) + sizeof(skb->next));
+       BUILD_BUG_ON(offsetof(struct sk_buff, tstamp) !=
+                    offsetof(struct sk_buff, prev) + sizeof(skb->prev));
+       BUILD_BUG_ON(sizeof(struct rb_node) > sizeof(skb->next) +
+                                             sizeof(skb->prev) +
+                                             sizeof(skb->tstamp));
+       return (struct rb_node *)&skb->next;
+}
+
+static struct sk_buff *netem_rb_to_skb(struct rb_node *rb)
+{
+       return (struct sk_buff *)rb;
+}
+
 static inline struct netem_skb_cb *netem_skb_cb(struct sk_buff *skb)
 {
+       /* we assume we can use skb next/prev/tstamp as storage for rb_node */
        qdisc_cb_private_validate(skb, sizeof(struct netem_skb_cb));
        return (struct netem_skb_cb *)qdisc_skb_cb(skb)->data;
 }
@@ -333,20 +360,23 @@ static psched_time_t packet_len_2_sched_time(unsigned int len, struct netem_sche
 
 static void tfifo_enqueue(struct sk_buff *nskb, struct Qdisc *sch)
 {
-       struct sk_buff_head *list = &sch->q;
+       struct netem_sched_data *q = qdisc_priv(sch);
        psched_time_t tnext = netem_skb_cb(nskb)->time_to_send;
-       struct sk_buff *skb = skb_peek_tail(list);
+       struct rb_node **p = &q->t_root.rb_node, *parent = NULL;
 
-       /* Optimize for add at tail */
-       if (likely(!skb || tnext >= netem_skb_cb(skb)->time_to_send))
-               return __skb_queue_tail(list, nskb);
+       while (*p) {
+               struct sk_buff *skb;
 
-       skb_queue_reverse_walk(list, skb) {
+               parent = *p;
+               skb = netem_rb_to_skb(parent);
                if (tnext >= netem_skb_cb(skb)->time_to_send)
-                       break;
+                       p = &parent->rb_right;
+               else
+                       p = &parent->rb_left;
        }
-
-       __skb_queue_after(list, skb, nskb);
+       rb_link_node(netem_rb_node(nskb), parent, p);
+       rb_insert_color(netem_rb_node(nskb), &q->t_root);
+       sch->q.qlen++;
 }
 
 /*
@@ -436,23 +466,28 @@ static int netem_enqueue(struct sk_buff *skb, struct Qdisc *sch)
                now = psched_get_time();
 
                if (q->rate) {
-                       struct sk_buff_head *list = &sch->q;
+                       struct sk_buff *last;
 
-                       if (!skb_queue_empty(list)) {
+                       if (!skb_queue_empty(&sch->q))
+                               last = skb_peek_tail(&sch->q);
+                       else
+                               last = netem_rb_to_skb(rb_last(&q->t_root));
+                       if (last) {
                                /*
                                 * Last packet in queue is reference point (now),
                                 * calculate this time bonus and subtract
                                 * from delay.
                                 */
-                               delay -= netem_skb_cb(skb_peek_tail(list))->time_to_send - now;
+                               delay -= netem_skb_cb(last)->time_to_send - now;
                                delay = max_t(psched_tdiff_t, 0, delay);
-                               now = netem_skb_cb(skb_peek_tail(list))->time_to_send;
+                               now = netem_skb_cb(last)->time_to_send;
                        }
 
                        delay += packet_len_2_sched_time(skb->len, q);
                }
 
                cb->time_to_send = now + delay;
+               cb->tstamp_save = skb->tstamp;
                ++q->counter;
                tfifo_enqueue(skb, sch);
        } else {
@@ -476,6 +511,21 @@ static unsigned int netem_drop(struct Qdisc *sch)
        unsigned int len;
 
        len = qdisc_queue_drop(sch);
+
+       if (!len) {
+               struct rb_node *p = rb_first(&q->t_root);
+
+               if (p) {
+                       struct sk_buff *skb = netem_rb_to_skb(p);
+
+                       rb_erase(p, &q->t_root);
+                       sch->q.qlen--;
+                       skb->next = NULL;
+                       skb->prev = NULL;
+                       len = qdisc_pkt_len(skb);
+                       kfree_skb(skb);
+               }
+       }
        if (!len && q->qdisc && q->qdisc->ops->drop)
            len = q->qdisc->ops->drop(q->qdisc);
        if (len)
@@ -488,19 +538,32 @@ static struct sk_buff *netem_dequeue(struct Qdisc *sch)
 {
        struct netem_sched_data *q = qdisc_priv(sch);
        struct sk_buff *skb;
+       struct rb_node *p;
 
        if (qdisc_is_throttled(sch))
                return NULL;
 
 tfifo_dequeue:
-       skb = qdisc_peek_head(sch);
+       skb = __skb_dequeue(&sch->q);
        if (skb) {
-               const struct netem_skb_cb *cb = netem_skb_cb(skb);
+deliver:
+               sch->qstats.backlog -= qdisc_pkt_len(skb);
+               qdisc_unthrottled(sch);
+               qdisc_bstats_update(sch, skb);
+               return skb;
+       }
+       p = rb_first(&q->t_root);
+       if (p) {
+               skb = netem_rb_to_skb(p);
 
                /* if more time remaining? */
-               if (cb->time_to_send <= psched_get_time()) {
-                       __skb_unlink(skb, &sch->q);
-                       sch->qstats.backlog -= qdisc_pkt_len(skb);
+               if (netem_skb_cb(skb)->time_to_send <= psched_get_time()) {
+                       rb_erase(p, &q->t_root);
+
+                       sch->q.qlen--;
+                       skb->next = NULL;
+                       skb->prev = NULL;
+                       skb->tstamp = netem_skb_cb(skb)->tstamp_save;
 
 #ifdef CONFIG_NET_CLS_ACT
                        /*
@@ -522,10 +585,7 @@ tfifo_dequeue:
                                }
                                goto tfifo_dequeue;
                        }
-deliver:
-                       qdisc_unthrottled(sch);
-                       qdisc_bstats_update(sch, skb);
-                       return skb;
+                       goto deliver;
                }
 
                if (q->qdisc) {
@@ -533,7 +593,8 @@ deliver:
                        if (skb)
                                goto deliver;
                }
-               qdisc_watchdog_schedule(&q->watchdog, cb->time_to_send);
+               qdisc_watchdog_schedule(&q->watchdog,
+                                       netem_skb_cb(skb)->time_to_send);
        }
 
        if (q->qdisc) {