: Update MTU to all related cache entries in ip_rt_frag_needed()

PATCH Applied Timo Teräs 2 participants 4 messages Apr 24, 2008 – Apr 29, 2008

Subsystems (3)

Networking Network drivers IPv4/IPv6

lore

Share & tools

lore

Timo Teräs

author

2008-04-24 10:05 UTC #0 lore raw

PATCH 1/1 : Update MTU to all related cache entries in ip_rt_frag_needed()

Add struct net_device parameter to ip_rt_frag_needed() and update MTU to cache entries where ifindex is specified. This is similar to what is already done in ip_rt_redirect().

Signed-off-by Timo Teras <[email protected]>

---
I noticed this bug when using GRE tunnel bound to specific device. This causes ip_gre to use ifindex in ip_route_output_key() lookups and the results are not proper unless the MTU is updated for entries where ifindex is specified (causing my tunneled packets to be dropped).

I have a patch for older kernels too if this is good for -stable.

 include/net/route.h |    2 +-
 net/ipv4/icmp.c     |    3 ++-
 net/ipv4/route.c    |   38 ++++++++++++++++++++++----------------
 3 files changed, 25 insertions(+), 18 deletions(-)

include/net/route.h +1 -1

index c633880..fc836ff 100644
@@ -116,7 +116,7 @@ extern int		__ip_route_output_key(struct net *, struct rtable **, const struct f
 extern int		ip_route_output_key(struct net *, struct rtable **, struct flowi *flp);
 extern int		ip_route_output_flow(struct net *, struct rtable **rp, struct flowi *flp, struct sock *sk, int flags);
 extern int		ip_route_input(struct sk_buff*, __be32 dst, __be32 src, u8 tos, struct net_device *devin);
-extern unsigned short	ip_rt_frag_needed(struct net *net, struct iphdr *iph, unsigned short new_mtu);
+extern unsigned short	ip_rt_frag_needed(struct net *net, struct iphdr *iph, unsigned short new_mtu, struct net_device *dev);
 extern void		ip_rt_send_redirect(struct sk_buff *skb);
 
 extern unsigned		inet_addr_type(struct net *net, __be32 addr);

net/ipv4/icmp.c +2 -1

index c67d00e..8739735 100644
@@ -691,7 +691,8 @@ static void icmp_unreach(struct sk_buff *skb)
 					       NIPQUAD(iph->daddr));
 			} else {
 				info = ip_rt_frag_needed(net, iph,
-						     ntohs(icmph->un.frag.mtu));
+							 ntohs(icmph->un.frag.mtu),
+							 skb->dev);
 				if (!info)
 					goto out;
 			}

net/ipv4/route.c +22 ~~-16~~

index 780e948..0d2f4da 100644
@@ -1429,11 +1429,13 @@ static inline unsigned short guess_mtu(unsigned short old_mtu)
 }
 
 unsigned short ip_rt_frag_needed(struct net *net, struct iphdr *iph,
-				 unsigned short new_mtu)
+				 unsigned short new_mtu,
+				 struct net_device *dev)
 {
-	int i;
+	int i, k;
 	unsigned short old_mtu = ntohs(iph->tot_len);
 	struct rtable *rth;
+	int  ikeys[2] = { dev->ifindex, 0 };
 	__be32  skeys[2] = { iph->saddr, 0, };
 	__be32  daddr = iph->daddr;
 	unsigned short est_mtu = 0;

@@ -1441,22 +1443,26 @@ unsigned short ip_rt_frag_needed(struct net *net, struct iphdr *iph,
 	if (ipv4_config.no_pmtu_disc)
 		return 0;
 
-	for (i = 0; i < 2; i++) {
-		unsigned hash = rt_hash(daddr, skeys[i], 0);
+	for (k = 0; k < 2; k++) {
+		for (i = 0; i < 2; i++) {
+			unsigned hash = rt_hash(daddr, skeys[i], ikeys[k]);
 
-		rcu_read_lock();
-		for (rth = rcu_dereference(rt_hash_table[hash].chain); rth;
-		     rth = rcu_dereference(rth->u.dst.rt_next)) {
-			if (rth->fl.fl4_dst == daddr &&
-			    rth->fl.fl4_src == skeys[i] &&
-			    rth->rt_dst  == daddr &&
-			    rth->rt_src  == iph->saddr &&
-			    rth->fl.iif == 0 &&
-			    !(dst_metric_locked(&rth->u.dst, RTAX_MTU)) &&
-			    net_eq(dev_net(rth->u.dst.dev), net) &&
-			    rth->rt_genid == atomic_read(&rt_genid)) {
+			rcu_read_lock();
+			for (rth = rcu_dereference(rt_hash_table[hash].chain); rth;
+			     rth = rcu_dereference(rth->u.dst.rt_next)) {
 				unsigned short mtu = new_mtu;
 
+				if (rth->fl.fl4_dst != daddr ||
+				    rth->fl.fl4_src != skeys[i] ||
+				    rth->rt_dst != daddr ||
+				    rth->rt_src != iph->saddr ||
+				    rth->fl.oif != ikeys[k] ||
+				    rth->fl.iif != 0 ||
+				    dst_metric_locked(&rth->u.dst, RTAX_MTU) ||
+				    !net_eq(dev_net(rth->u.dst.dev), net) ||
+				    rth->rt_genid != atomic_read(&rt_genid))
+					continue;
+
 				if (new_mtu < 68 || new_mtu >= old_mtu) {
 
 					/* BSD 4.2 compatibility hack :-( */

@@ -1482,8 +1488,8 @@ unsigned short ip_rt_frag_needed(struct net *net, struct iphdr *iph,
 					est_mtu = mtu;
 				}
 			}
+			rcu_read_unlock();
 		}
-		rcu_read_unlock();
 	}
 	return est_mtu ? : new_mtu;
 }

Signature · 1 line

1.5.2.5

David Miller

maintainer

2008-04-27 06:27 UTC #1 lore raw

From: Timo_Teräs <[email protected]>
Date: Thu, 24 Apr 2008 13:05:27 +0300

Quoted 4 lines

Add struct net_device parameter to ip_rt_frag_needed() and update MTU to cache entries where ifindex is specified. This is similar to what is already done in ip_rt_redirect().

Quoted 8 lines

Signed-off-by Timo Teras <[email protected]>
---
I noticed this bug when using GRE tunnel bound to specific device. This causes ip_gre to use ifindex in ip_route_output_key() lookups and the results are not proper unless the MTU is updated for entries where ifindex is specified (causing my tunneled packets to be dropped).
I have a patch for older kernels too if this is good for -stable.

I haven't forgotten about this patch, I'm just still thinking about possible negative side effects. I don't think there are any, but I want to be sure.

This function has been a constant source of consternation. It also doesn't do the right thing when the TOS is modified mid-network, and that's arguably wrong too but more difficult to solve efficiently than this ifindex keying case.

Timo Teräs

author

2008-04-27 08:35 UTC #2 lore raw

Quoted 9 lines · David Miller

From: Timo_Teräs <[email protected]>
Date: Thu, 24 Apr 2008 13:05:27 +0300
Quoted 5 lines
Add struct net_device parameter to ip_rt_frag_needed() and update MTU to cache entries where ifindex is specified. This is similar to what is already done in ip_rt_redirect().
Signed-off-by Timo Teras <[email protected]>

Quoted 8 lines

I haven't forgotten about this patch, I'm just still thinking about possible negative side effects. I don't think there are any, but I want to be sure.
This function has been a constant source of consternation. It also doesn't do the right thing when the TOS is modified mid-network, and that's arguably wrong too but more difficult to solve efficiently than this ifindex keying case.

I understand your concern. It is pretty easy break routing cache with some seemingly unrelated changes. I tried to think about possible side effects, but concluded that this should be safe. Then again, I don't know all the tiny details and unusual scenarios. But I have been running a 2.6.23 kernel with (backported) patch on a production machine (router connected to two ISPs, internal subnet and also having multiple gre tunnels), so far I have not encountered any problems.

Cheers,
Timo

David Miller

maintainer Applied

2008-04-29 10:32 UTC #3 lore raw

From: Timo_Teräs <[email protected]>
Date: Sun, 27 Apr 2008 11:35:42 +0300

Quoted 7 lines

I understand your concern. It is pretty easy break routing cache with some seemingly unrelated changes. I tried to think about possible side effects, but concluded that this should be safe. Then again, I don't know all the tiny details and unusual scenarios. But I have been running a 2.6.23 kernel with (backported) patch on a production machine (router connected to two ISPs, internal subnet and also having multiple gre tunnels), so far I have not encountered any problems.

I thought more about your patch and it should be OK.

I've applied it, thanks a lot.