Skip to contentSkip to conversation

net: increase skb->users instead of skb_clone()

PATCH Applied Changli Gao 4 participants 15 messages Dec 16, 2010 – Dec 20, 2010

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net: increase skb->users instead of skb_clone()

In dev_queue_xmit_nit(), we have to clone skbs as we need to mangle skbs, however, we don't need to clone skbs for all the packet_types.

Except for the first packet_type, we increase skb->users instead of skb_clone().

---
 net/core/dev.c |   30 ++++++++++++++++++++----------
 1 file changed, 20 insertions(+), 10 deletions(-)
net/core/dev.c +20 -10
index bf5ced5..888cb74 100644
@@ -1496,6 +1496,14 @@ int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
 }
 EXPORT_SYMBOL_GPL(dev_forward_skb);
 
+static inline int deliver_skb(struct sk_buff *skb,
+			      struct packet_type *pt_prev,
+			      struct net_device *orig_dev)
+{
+	atomic_inc(&skb->users);
+	return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
+}
+
 /*
  *	Support routine. Sends outgoing frames to any network
  *	taps currently in use.
@@ -1504,6 +1512,8 @@ EXPORT_SYMBOL_GPL(dev_forward_skb);
 static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct packet_type *ptype;
+	struct sk_buff *skb2 = NULL;
+	struct packet_type *pt_prev = NULL;
 
 #ifdef CONFIG_NET_CLS_ACT
 	if (!(skb->tstamp.tv64 && (G_TC_FROM(skb->tc_verd) & AT_INGRESS)))
@@ -1520,7 +1530,13 @@ static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
 		if ((ptype->dev == dev || !ptype->dev) &&
 		    (ptype->af_packet_priv == NULL ||
 		     (struct sock *)ptype->af_packet_priv != skb->sk)) {
-			struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
+			if (pt_prev) {
+				deliver_skb(skb2, pt_prev, skb->dev);
+				pt_prev = ptype;
+				continue;
+			}
+
+			skb2 = skb_clone(skb, GFP_ATOMIC);
 			if (!skb2)
 				break;
 
@@ -1542,9 +1558,11 @@ static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
 
 			skb2->transport_header = skb2->network_header;
 			skb2->pkt_type = PACKET_OUTGOING;
-			ptype->func(skb2, skb->dev, ptype, skb->dev);
+			pt_prev = ptype;
 		}
 	}
+	if (pt_prev)
+		pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
 	rcu_read_unlock();
 }
 
@@ -2788,14 +2806,6 @@ static void net_tx_action(struct softirq_action *h)
 	}
 }
 
-static inline int deliver_skb(struct sk_buff *skb,
-			      struct packet_type *pt_prev,
-			      struct net_device *orig_dev)
-{
-	atomic_inc(&skb->users);
-	return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
-}
-
 #if (defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)) && \
     (defined(CONFIG_ATM_LANE) || defined(CONFIG_ATM_LANE_MODULE))
 /* This hook is defined here for ATM LANE */
#1 lore raw
Quoted 71 lines · Changli Gao

In dev_queue_xmit_nit(), we have to clone skbs as we need to mangle skbs, however, we don't need to clone skbs for all the packet_types.

Except for the first packet_type, we increase skb->users instead of skb_clone().

---
 net/core/dev.c |   30 ++++++++++++++++++++----------
 1 file changed, 20 insertions(+), 10 deletions(-)
net/core/dev.c +20 -3
index bf5ced5..888cb74 100644
@@ -1496,6 +1496,14 @@ int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
 }
 EXPORT_SYMBOL_GPL(dev_forward_skb);
 
+static inline int deliver_skb(struct sk_buff *skb,
+			      struct packet_type *pt_prev,
+			      struct net_device *orig_dev)
+{
+	atomic_inc(&skb->users);
+	return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
+}
+
 /*
  *	Support routine. Sends outgoing frames to any network
  *	taps currently in use.
@@ -1504,6 +1512,8 @@ EXPORT_SYMBOL_GPL(dev_forward_skb);
 static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct packet_type *ptype;
+	struct sk_buff *skb2 = NULL;
+	struct packet_type *pt_prev = NULL;
 
 #ifdef CONFIG_NET_CLS_ACT
 	if (!(skb->tstamp.tv64 && (G_TC_FROM(skb->tc_verd) & AT_INGRESS)))
@@ -1520,7 +1530,13 @@ static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
 		if ((ptype->dev == dev || !ptype->dev) &&
 		    (ptype->af_packet_priv == NULL ||
 		     (struct sock *)ptype->af_packet_priv != skb->sk)) {
-			struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC);
+			if (pt_prev) {
+				deliver_skb(skb2, pt_prev, skb->dev);
+				pt_prev = ptype;
+				continue;
+			}
+
+			skb2 = skb_clone(skb, GFP_ATOMIC);
 			if (!skb2)
 				break;
 
@@ -1542,9 +1558,11 @@ static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
 
 			skb2->transport_header = skb2->network_header;
 			skb2->pkt_type = PACKET_OUTGOING;
-			ptype->func(skb2, skb->dev, ptype, skb->dev);
+			pt_prev = ptype;
 		}
 	}
+	if (pt_prev)
+		pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
 	rcu_read_unlock();
 }
 
@@ -2788,14 +2806,6 @@ static void net_tx_action(struct softirq_action *h)
 	}
 }
 
-static inline int deliver_skb(struct sk_buff *skb,

net/core/dev.c static void net_tx_action(struct softirq_action *h) · line 2811

Quoted 10 lines
net/core/dev.c +0 -7
-			      struct packet_type *pt_prev,
-			      struct net_device *orig_dev)
-{
-	atomic_inc(&skb->users);
-	return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
-}
-
 #if (defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)) && \
     (defined(CONFIG_ATM_LANE) || defined(CONFIG_ATM_LANE_MODULE))
 /* This hook is defined here for ATM LANE */

You beat me, but I was thinking of a different way, adding a new pt_prev->xmit_func(), handling all the details (no need for atomic ops on skb users if packet is not delivered at all).

By the way, your patch is not 100% safe/OK, because af_packet rcv() handler writes on skb (skb_pull() and all)

#2 lore raw
Quoted 8 lines · Eric Dumazet

You beat me, but I was thinking of a different way, adding a new pt_prev->xmit_func(), handling all the details (no need for atomic ops on skb users if packet is not delivered at all).

By the way, your patch is not 100% safe/OK, because af_packet rcv() handler writes on skb (skb_pull() and all)

But af_packet_rcv() restores skbs at last.

        if (skb_head != skb->data && skb_shared(skb)) {
                skb->data = skb_head;
                skb->len = skb_len;
        }
Signature · 2 lines
Regards,
Changli Gao([email protected])
Acked-by
#3 lore raw
Quoted 10 lines · Changli Gao
Quoted 8 lines · Eric Dumazet

You beat me, but I was thinking of a different way, adding a new pt_prev->xmit_func(), handling all the details (no need for atomic ops on skb users if packet is not delivered at all).

By the way, your patch is not 100% safe/OK, because af_packet rcv() handler writes on skb (skb_pull() and all)

Quoted 7 lines

But af_packet_rcv() restores skbs at last.

        if (skb_head != skb->data && skb_shared(skb)) {
                skb->data = skb_head;
                skb->len = skb_len;
        }

Thats right, your patch seems fine, thanks !

#4 lore raw
In reply to Changli Gao
Quoted 8 lines · Changli Gao
Quoted 7 lines

You beat me, but I was thinking of a different way, adding a new pt_prev->xmit_func(), handling all the details (no need for atomic ops on skb users if packet is not delivered at all).

By the way, your patch is not 100% safe/OK, because af_packet rcv() handler writes on skb (skb_pull() and all)

Quoted 7 lines

But af_packet_rcv() restores skbs at last.

       if (skb_head != skb->data && skb_shared(skb)) { skb->data = skb_head;
               skb->len = skb_len;
       }

If af packet_rcv invokes skb_clone, this skb is differ from the original one. Eric's warning is right.

Signature · 1 line
--Junchang
#5 lore raw
Quoted 12 lines · Junchang Wang
Quoted 10 lines · Changli Gao
Quoted 7 lines

You beat me, but I was thinking of a different way, adding a new pt_prev->xmit_func(), handling all the details (no need for atomic ops on skb users if packet is not delivered at all).

By the way, your patch is not 100% safe/OK, because af_packet rcv() handler writes on skb (skb_pull() and all)

But af_packet_rcv() restores skbs at last.

Quoted 7 lines
Quoted 5 lines
       if (skb_head != skb->data && skb_shared(skb)) {
               skb->data = skb_head;
               skb->len = skb_len;
       }

If af packet_rcv invokes skb_clone, this skb is differ from the original one. Eric's warning is right.

It was a false alarm.

If packet_rcv() invokes skb_clone(), skb still points to original skb. No worry.

#6 lore raw
In reply to Changli Gao
Quoted 5 lines · Changli Gao

In dev_queue_xmit_nit(), we have to clone skbs as we need to mangle skbs, however, we don't need to clone skbs for all the packet_types.

Except for the first packet_type, we increase skb->users instead of skb_clone().

Hi Changli,
Take af_packet for example, I can't see benefit from this patch.

net/core/dev.c int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) · line 1505

Quoted 7 lines
net/core/dev.c +7 -0
+static inline int deliver_skb(struct sk_buff *skb,
+                             struct packet_type *pt_prev,
+                             struct net_device *orig_dev)
+{
+       atomic_inc(&skb->users);
+       return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
+}

The increment call will incur skb_shared() failure in packet_rcv. In reality, packet_rcv has to clone this packet by itself.

Thanks

Signature · 1 line
--Junchang
#7 lore raw
Quoted 4 lines · Junchang Wang

Hi Changli,
Take af_packet for example, I can't see benefit from this patch.

Quoted 11 lines

net/core/dev.c int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) · line 1505

Quoted 7 lines
net/core/dev.c +7 -0
+static inline int deliver_skb(struct sk_buff *skb,
+                             struct packet_type *pt_prev,
+                             struct net_device *orig_dev)
+{
+       atomic_inc(&skb->users);
+       return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
+}

The increment call will incur skb_shared() failure in packet_rcv. In reality, packet_rcv has to clone this packet by itself.

This happens when run_filter returns non zero. For your case, only small parts of packets match bpf filter.

Signature · 2 lines
Regards,
Changli Gao([email protected])
#8 lore raw
Quoted 3 lines · Changli Gao

This happens when run_filter returns non zero. For your case, only small parts of packets match bpf filter.

Hi Changli,
In most cases, I want user-space applications see everything. :)

Thanks.

Signature · 1 line
--Junchang
#9 lore raw
In reply to Junchang Wang
Quoted 10 lines · Junchang Wang
Quoted 5 lines · Changli Gao

In dev_queue_xmit_nit(), we have to clone skbs as we need to mangle skbs, however, we don't need to clone skbs for all the packet_types.

Except for the first packet_type, we increase skb->users instead of skb_clone().

Hi Changli,
Take af_packet for example, I can't see benefit from this patch.

Quoted 10 lines

net/core/dev.c int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) · line 1505

Quoted 7 lines
net/core/dev.c +7 -0
+static inline int deliver_skb(struct sk_buff *skb,
+                             struct packet_type *pt_prev,
+                             struct net_device *orig_dev)
+{
+       atomic_inc(&skb->users);
+       return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
+}

The increment call will incur skb_shared() failure in packet_rcv. In reality, packet_rcv has to clone this packet by itself.

Yes, and no.

Consider the case you have one receiver.

Packet given after Changli patch wont be shared, so packet_rcv wont clone it : Thats a win. Only one skb_clone() done instead of two.

Consider case with 2 receivers :

First time we call packet_rcv, packet is shared (because we call deliver_skb(), so packet_rcv clones it. Normal situation, we really need to clone it.

Second time, we give a non shared packet : Thats a win over previous situation.

#10 lore raw
Quoted 5 lines · Eric Dumazet

Yes, and no.

Consider the case you have one receiver.

Quoted 8 lines

Packet given after Changli patch wont be shared, so packet_rcv wont clone it : Thats a win. Only one skb_clone() done instead of two.

Consider case with 2 receivers :

First time we call packet_rcv, packet is shared (because we call deliver_skb(), so packet_rcv clones it. Normal situation, we really need to clone it.

Got it. Thanks.

Quoted 4 lines

Second time, we give a non shared packet : Thats a win over previous situation.

But, if we have N receivers, we get only the last one win - the first N-1 will call deliver_skb().

Thanks.

Signature · 1 line
--Junchang
#11 lore raw
Quoted 15 lines · Junchang Wang
Quoted 5 lines · Eric Dumazet

Yes, and no.

Consider the case you have one receiver.

Quoted 8 lines

Packet given after Changli patch wont be shared, so packet_rcv wont clone it : Thats a win. Only one skb_clone() done instead of two.

Consider case with 2 receivers :

First time we call packet_rcv, packet is shared (because we call deliver_skb(), so packet_rcv clones it. Normal situation, we really need to clone it.

Quoted 9 lines

Got it. Thanks.

Quoted 4 lines

Second time, we give a non shared packet : Thats a win over previous situation.

But, if we have N receivers, we get only the last one win - the first N-1 will call deliver_skb().

Yes, but you want to, because each receiver has to make a private copy of the skb.

The big win is that if packet if filtered out (not accepted by the socket filter), you end with no extra skb_clone() at all.

Say you have 8 receivers, with a filter matching some hash/cpu, and only one af_packet socket will take the message.

Before patch : 8 skb_clones()

After patch : one skb_clone()

If I undertood patch intent ;)

#12 lore raw
Quoted 7 lines · Eric Dumazet

Yes, but you want to, because each receiver has to make a private copy of the skb.

The big win is that if packet if filtered out (not accepted by the socket filter), you end with no extra skb_clone() at all.

Quoted 9 lines

Say you have 8 receivers, with a filter matching some hash/cpu, and only one af_packet socket will take the message.

Before patch : 8 skb_clones()

After patch : one skb_clone()

If I undertood patch intent ;)

Yes, you got it. :)

Signature · 2 lines
Regards,
Changli Gao([email protected])
#13 lore raw
In reply to Eric Dumazet
Quoted 10 lines · Eric Dumazet
Quoted 3 lines

But, if we have N receivers, we get only the last one win - the first N-1 will call deliver_skb().

Yes, but you want to, because each receiver has to make a private copy of the skb.

The big win is that if packet if filtered out (not accepted by the socket filter), you end with no extra skb_clone() at all.

Quoted 7 lines

Say you have 8 receivers, with a filter matching some hash/cpu, and only one af_packet socket will take the message.

Before patch : 8 skb_clones()

After patch : one skb_clone()

Now I understand. :)
The patch is fine. Thanks Changli and Eric.

Signature · 1 line
--Junchang