net: avoid RCU for NOCACHE dst
There is no point using RCU for dst we allocate for a very short time
(used once).
Change dst_release() to take DST_NOCACHE into account, but also change
skb_dst_set_noref() to force a refcount increment for such dst.
This is a _huge_ gain, because we dont waste memory to store xx thousand
of dsts. Instead of queueing them to RCU, we can free them instantly.
CPU caches can stay hot, re-using same memory blocks to hold temporary
dsts.
Note : remove unneeded smp_mb__before_atomic_dec(); in dst_release(),
since atomic_dec_return() implies a full memory barrier.
Stress test, 160.000.000 udp frames sent, IP route cache disabled
(DDOS).
Before:
real 0m38.091s
user 0m13.189s
sys 7m53.018s
After:
real 0m29.946s
user 0m12.157s
sys 7m40.605s
For reference, if IP route cache was enabled :
real 0m32.030s
user 0m10.521s
sys 8m15.243s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 05a358f..e6ba898 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -460,19 +460,7 @@
skb->_skb_refdst = (unsigned long)dst;
}
-/**
- * skb_dst_set_noref - sets skb dst, without a reference
- * @skb: buffer
- * @dst: dst entry
- *
- * Sets skb dst, assuming a reference was not taken on dst
- * skb_dst_drop() should not dst_release() this dst
- */
-static inline void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
-{
- WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
- skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
-}
+extern void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst);
/**
* skb_dst_is_noref - Test if skb dst isnt refcounted
diff --git a/net/core/dst.c b/net/core/dst.c
index 32e542d..8abe628 100644
--- a/net/core/dst.c
+++ b/net/core/dst.c
@@ -271,13 +271,40 @@
if (dst) {
int newrefcnt;
- smp_mb__before_atomic_dec();
newrefcnt = atomic_dec_return(&dst->__refcnt);
WARN_ON(newrefcnt < 0);
+ if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
+ dst = dst_destroy(dst);
+ if (dst)
+ __dst_free(dst);
+ }
}
}
EXPORT_SYMBOL(dst_release);
+/**
+ * skb_dst_set_noref - sets skb dst, without a reference
+ * @skb: buffer
+ * @dst: dst entry
+ *
+ * Sets skb dst, assuming a reference was not taken on dst
+ * skb_dst_drop() should not dst_release() this dst
+ */
+void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
+{
+ WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
+ /* If dst not in cache, we must take a reference, because
+ * dst_release() will destroy dst as soon as its refcount becomes zero
+ */
+ if (unlikely(dst->flags & DST_NOCACHE)) {
+ dst_hold(dst);
+ skb_dst_set(skb, dst);
+ } else {
+ skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
+ }
+}
+EXPORT_SYMBOL(skb_dst_set_noref);
+
/* Dirty hack. We did it in 2.2 (in __dst_free),
* we have _very_ good reasons not to repeat
* this mistake in 2.3, but we have no choice
diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index ff98983d..d6cb2bf 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -1105,9 +1105,9 @@
* Note that we do rt_free on this new route entry, so that
* once its refcount hits zero, we are still able to reap it
* (Thanks Alexey)
- * Note also the rt_free uses call_rcu. We don't actually
- * need rcu protection here, this is just our path to get
- * on the route gc list.
+ * Note: To avoid expensive rcu stuff for this uncached dst,
+ * we set DST_NOCACHE so that dst_release() can free dst without
+ * waiting a grace period.
*/
rt->dst.flags |= DST_NOCACHE;
@@ -1117,12 +1117,11 @@
if (net_ratelimit())
printk(KERN_WARNING
"Neighbour table failure & not caching routes.\n");
- rt_drop(rt);
+ ip_rt_put(rt);
return err;
}
}
- rt_free(rt);
goto skip_hashing;
}