PKT_SCHED: dsmark must take care of shared/cloned skbs

Correctly handle shared and cloned skbs by copying them before writing and dequeue unwriteable skbs unchanged. Assumes that IP/IPv6 header is always linear so no pulling required.

Dave,

Correctly handle shared and cloned skbs by copying them before writing and dequeue unwriteable skbs unchanged. Assumes that IP/IPv6 header is always linear so no pulling required.

Signed-off-by Thomas Graf <[email protected]>

net/sched/sch_dsmark.c +42 -4

@@ -195,7 +195,8 @@
 
 	D2PRINTK("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n",skb,sch,p);
 	if (p->set_tc_index) {
-		/* FIXME: Safe with non-linear skbs? --RR */
+		/* Safe with non-linear skbs? --RR
+		   IP/IPv6 header is always linear --TGR */
 		switch (skb->protocol) {
 			case __constant_htons(ETH_P_IP):
 				skb->tc_index = ipv4_get_dsfield(skb->nh.iph);

@@ -250,6 +251,34 @@
 	return ret;
 }
 
+static inline int
+dsmark_make_writeable(struct sk_buff **pskb, int offset)
+{
+	struct sk_buff *nskb;
+
+	if ((offset + (*pskb)->mac_len) > (*pskb)->len)
+		return 0;
+
+	if (skb_shared(*pskb) || skb_cloned(*pskb))
+		goto copy_skb;
+
+	/* IP/IPv6 header is always linear, no need to pull */
+	return 1;
+
+copy_skb:
+	nskb = skb_copy(*pskb, GFP_ATOMIC);
+	if (!nskb)
+		return 0;
+	BUG_ON(skb_is_nonlinear(nskb));
+
+	/* Rest of kernel will get very unhappy if we pass it a
+	   suddenly-orphaned skbuff */
+	if ((*pskb)->sk)
+		skb_set_owner_w(nskb, (*pskb)->sk);
+	kfree_skb(*pskb);
+	*pskb = nskb;
+	return 1;
+}
 
 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
 {

@@ -266,10 +295,14 @@
 	D2PRINTK("index %d->%d\n",skb->tc_index,index);
 	switch (skb->protocol) {
 		case __constant_htons(ETH_P_IP):
+			if (!dsmark_make_writeable(&skb, sizeof(struct iphdr)))
+				goto unwriteable;
 			ipv4_change_dsfield(skb->nh.iph,
 			    p->mask[index],p->value[index]);
 			break;
 		case __constant_htons(ETH_P_IPV6):
+			if (!dsmark_make_writeable(&skb, sizeof(struct ipv6hdr)))
+				goto unwriteable;
 			ipv6_change_dsfield(skb->nh.ipv6h,
 			    p->mask[index],p->value[index]);
 			break;

@@ -280,12 +313,17 @@
 			 * and don't need yet another qdisc as a bypass.
 			 */
 			if (p->mask[index] != 0xff || p->value[index])
-				printk(KERN_WARNING "dsmark_dequeue: "
-				       "unsupported protocol %d\n",
-				       htons(skb->protocol));
+				if (net_ratelimit())
+					printk(KERN_WARNING "dsmark_dequeue: "
+					       "unsupported protocol %d\n",
+					       htons(skb->protocol));
 			break;
 	};
 	return skb;
+unwriteable:
+	if (net_ratelimit())
+		printk(KERN_WARNING "dsmark_dequeue: skb not writable\n");
+	return skb;
 }
 
 

If the qdisc at that level muddies the packet thats fair game - thats what goes out on the wire. So we should leave the code as is.

* jamal <[email protected]> 2004-12-19 15:23

Quoted 2 lines

If the qdisc at that level muddies the packet thats fair game - thats what goes out on the wire. So we should leave the code as is.

Agreed for egress but I think it is needed for stuff like IMQ. It's debatable whether we should take care of IMQ and alike though.

Quoted 4 lines

* jamal <[email protected]> 2004-12-19 15:23

Quoted 2 lines

If the qdisc at that level muddies the packet thats fair game - thats what goes out on the wire. So we should leave the code as is.

Quoted 2 lines

Agreed for egress but I think it is needed for stuff like IMQ. It's debatable whether we should take care of IMQ and alike though.

You are right, it may cause an issue on an IMQ like device when we have one in kernel. Hang on to the patch for now is my opinion; "schedulers" should probably not be mucking with packets once they are queued in (dsmark aint really a scheduler). We should talk to Werner to see if we can move that functionality prequeueing ...

* jamal <[email protected]> 2004-12-19 15:23

Quoted 2 lines

If the qdisc at that level muddies the packet thats fair game - thats what goes out on the wire. So we should leave the code as is.

Agreed for egress but I think it is needed for stuff like IMQ. It's debatable whether we should take care of IMQ and alike though.

You shouldn't care about IMQ, but we still need to copy the packet before modifying it if the data is shared. Otherwise we have a race on SMP with AF_PACKET sockets, depending on when the packet is read it can be either modified or not.

Converting dsmark to an action sounds like the best long-term solution.

Certainly not a big deal; shouldnt care if once in a while tcpdump actually gets to see the real packet that went out the wire.

Quoted 2 lines · On 20 Dec 2004 09:13:46 -0500 jamal <[email protected]> wrote:

Certainly not a big deal; shouldnt care if once in a while tcpdump actually gets to see the real packet that went out the wire.

It's not just tcpdump. Any modification of a the packet data for a shared SKB is illegal, no matter where it occurs. This can corrupt TCP packets, which share the transmitted packet with the socket retransmit queue.

We have a similar problem with TSO and some gigabit cards whose drivers muck with the iphdr->tot_len field on transmit. I still am not sure how I want to address that case yet. Since transmitted TCP data packets are always shared/cloned, we'll have to do a data copy on every TSO send on these cards which frankly nullifies much of the performance gain TSO gives. If we end of fixing it via a copy we'll probably need to seriously consider not doing TSO unless we are doing sendfile.

Quoted 8 lines

We have a similar problem with TSO and some gigabit cards whose drivers muck with the iphdr->tot_len field on transmit. I still am not sure how I want to address that case yet. Since transmitted TCP data packets are always shared/cloned, we'll have to do a data copy on every TSO send on these cards which frankly nullifies much of the performance gain TSO gives. If we end of fixing it via a copy we'll probably need to seriously consider not doing TSO unless we are doing sendfile.

Ok, makes sense. I can see how TSO may not be useful if you have to copy. I suppose NICS with hardware based retransmits will behave differently.