neigh: speedup neigh_hh_init()

When a new dst is used to send a frame, neigh_resolve_output() tries to
associate an struct hh_cache to this dst, calling neigh_hh_init() with
the neigh rwlock write locked.

Most of the time, hh_cache is already known and linked into neighbour,
so we find it and increment its refcount.

This patch changes the logic so that we call neigh_hh_init() with
neighbour lock read locked only, so that fast path can be run in
parallel by concurrent cpus.

This brings part of the speedup we got with commit c7d4426a98a5f
(introduce DST_NOCACHE flag) for non cached dsts, even for cached ones,
removing one of the contention point that routers hit on multiqueue
enabled machines.

Further improvements would need to use a seqlock instead of an rwlock to
protect neigh->ha[], to not dirty neigh too often and remove two atomic
ops.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

3 files changed, 69 insertions, 40 deletions

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 6abcef67b178..4160db3721ba 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -281,6 +281,12 @@ struct hh_cache {
 	unsigned long	hh_data[HH_DATA_ALIGN(LL_MAX_HEADER) / sizeof(long)];
 };
 
+static inline void hh_cache_put(struct hh_cache *hh)
+{
+	if (atomic_dec_and_test(&hh->hh_refcnt))
+		kfree(hh);
+}
+
 /* Reserve HH_DATA_MOD byte aligned hard_header_len, but at least that much.
  * Alternative is:
  *   dev->hard_header_len ? (dev->hard_header_len +
diff --git a/net/core/dst.c b/net/core/dst.c
index 6c41b1fac3db..978a1ee1f7d0 100644
--- a/net/core/dst.c
+++ b/net/core/dst.c
@@ -228,8 +228,8 @@ again:
 	child = dst->child;
 
 	dst->hh = NULL;
-	if (hh && atomic_dec_and_test(&hh->hh_refcnt))
-		kfree(hh);
+	if (hh)
+		hh_cache_put(hh);
 
 	if (neigh) {
 		dst->neighbour = NULL;
diff --git a/net/core/neighbour.c b/net/core/neighbour.c
index 3ffafaa0414c..2044906ecd1a 100644
--- a/net/core/neighbour.c
+++ b/net/core/neighbour.c
@@ -709,8 +709,7 @@ void neigh_destroy(struct neighbour *neigh)
 		write_seqlock_bh(&hh->hh_lock);
 		hh->hh_output = neigh_blackhole;
 		write_sequnlock_bh(&hh->hh_lock);
-		if (atomic_dec_and_test(&hh->hh_refcnt))
-			kfree(hh);
+		hh_cache_put(hh);
 	}
 
 	skb_queue_purge(&neigh->arp_queue);
@@ -1210,39 +1209,67 @@ struct neighbour *neigh_event_ns(struct neigh_table *tbl,
 }
 EXPORT_SYMBOL(neigh_event_ns);
 
+static inline bool neigh_hh_lookup(struct neighbour *n, struct dst_entry *dst,
+				   __be16 protocol)
+{
+	struct hh_cache *hh;
+
+	for (hh = n->hh; hh; hh = hh->hh_next) {
+		if (hh->hh_type == protocol) {
+			atomic_inc(&hh->hh_refcnt);
+			if (unlikely(cmpxchg(&dst->hh, NULL, hh) != NULL))
+				hh_cache_put(hh);
+			return true;
+		}
+	}
+	return false;
+}
+
+/* called with read_lock_bh(&n->lock); */
 static void neigh_hh_init(struct neighbour *n, struct dst_entry *dst,
 			  __be16 protocol)
 {
 	struct hh_cache	*hh;
 	struct net_device *dev = dst->dev;
 
-	for (hh = n->hh; hh; hh = hh->hh_next)
-		if (hh->hh_type == protocol)
-			break;
+	if (likely(neigh_hh_lookup(n, dst, protocol)))
+		return;
 
-	if (!hh && (hh = kzalloc(sizeof(*hh), GFP_ATOMIC)) != NULL) {
-		seqlock_init(&hh->hh_lock);
-		hh->hh_type = protocol;
-		atomic_set(&hh->hh_refcnt, 0);
-		hh->hh_next = NULL;
+	/* slow path */
+	hh = kzalloc(sizeof(*hh), GFP_ATOMIC);
+	if (!hh)
+		return;
 
-		if (dev->header_ops->cache(n, hh)) {
-			kfree(hh);
-			hh = NULL;
-		} else {
-			atomic_inc(&hh->hh_refcnt);
-			hh->hh_next = n->hh;
-			n->hh	    = hh;
-			if (n->nud_state & NUD_CONNECTED)
-				hh->hh_output = n->ops->hh_output;
-			else
-				hh->hh_output = n->ops->output;
-		}
+	seqlock_init(&hh->hh_lock);
+	hh->hh_type = protocol;
+	atomic_set(&hh->hh_refcnt, 2);
+
+	if (dev->header_ops->cache(n, hh)) {
+		kfree(hh);
+		return;
 	}
-	if (hh)	{
-		atomic_inc(&hh->hh_refcnt);
-		dst->hh = hh;
+	read_unlock(&n->lock);
+	write_lock(&n->lock);
+
+	/* must check if another thread already did the insert */
+	if (neigh_hh_lookup(n, dst, protocol)) {
+		kfree(hh);
+		goto end;
 	}
+
+	if (n->nud_state & NUD_CONNECTED)
+		hh->hh_output = n->ops->hh_output;
+	else
+		hh->hh_output = n->ops->output;
+
+	hh->hh_next = n->hh;
+	n->hh	    = hh;
+
+	if (unlikely(cmpxchg(&dst->hh, NULL, hh) != NULL))
+		hh_cache_put(hh);
+end:
+	write_unlock(&n->lock);
+	read_lock(&n->lock);
 }
 
 /* This function can be used in contexts, where only old dev_queue_xmit
@@ -1281,21 +1308,17 @@ int neigh_resolve_output(struct sk_buff *skb)
 	if (!neigh_event_send(neigh, skb)) {
 		int err;
 		struct net_device *dev = neigh->dev;
+
+		read_lock_bh(&neigh->lock);
 		if (dev->header_ops->cache &&
 		    !dst->hh &&
-		    !(dst->flags & DST_NOCACHE)) {
-			write_lock_bh(&neigh->lock);
-			if (!dst->hh)
-				neigh_hh_init(neigh, dst, dst->ops->protocol);
-			err = dev_hard_header(skb, dev, ntohs(skb->protocol),
-					      neigh->ha, NULL, skb->len);
-			write_unlock_bh(&neigh->lock);
-		} else {
-			read_lock_bh(&neigh->lock);
-			err = dev_hard_header(skb, dev, ntohs(skb->protocol),
-					      neigh->ha, NULL, skb->len);
-			read_unlock_bh(&neigh->lock);
-		}
+		    !(dst->flags & DST_NOCACHE))
+			neigh_hh_init(neigh, dst, dst->ops->protocol);
+
+		err = dev_hard_header(skb, dev, ntohs(skb->protocol),
+				      neigh->ha, NULL, skb->len);
+		read_unlock_bh(&neigh->lock);
+
 		if (err >= 0)
 			rc = neigh->ops->queue_xmit(skb);
 		else