net: speed up skb_rbtree_purge()
authorEric Dumazet <edumazet@google.com>
Sat, 23 Sep 2017 19:39:12 +0000 (12:39 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 26 Sep 2017 03:35:11 +0000 (20:35 -0700)
As measured in my prior patch ("sch_netem: faster rb tree removal"),
rbtree_postorder_for_each_entry_safe() is nice looking but much slower
than using rb_next() directly, except when tree is small enough
to fit in CPU caches (then the cost is the same)

Also note that there is not even an increase of text size :
$ size net/core/skbuff.o.before net/core/skbuff.o
   text    data     bss     dec     hex filename
  40711    1298       0   42009    a419 net/core/skbuff.o.before
  40711    1298       0   42009    a419 net/core/skbuff.o

From: Eric Dumazet <edumazet@google.com>

Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/skbuff.c

index 16982de..000ce73 100644 (file)
@@ -2848,12 +2848,15 @@ EXPORT_SYMBOL(skb_queue_purge);
  */
 void skb_rbtree_purge(struct rb_root *root)
 {
-       struct sk_buff *skb, *next;
+       struct rb_node *p = rb_first(root);
 
-       rbtree_postorder_for_each_entry_safe(skb, next, root, rbnode)
-               kfree_skb(skb);
+       while (p) {
+               struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
 
-       *root = RB_ROOT;
+               p = rb_next(p);
+               rb_erase(&skb->rbnode, root);
+               kfree_skb(skb);
+       }
 }
 
 /**