Possible IPv6 fragmentation/pmtud bug
Subsystems (2)
Hi,
I have the following test network configuration (all linux nodes running the 3.10.17 kernel):
h1 <-----------> r1 <-----------> r2 <-----------> h2
8000 MTU link 4000 MTU link 8000 MTU link
When host h1 ping6's host h2, routers r1 and r2 forward the ICMPv6 echo request/reply packets as expected. But, when I set a ping6 packet size that is larger than the router-to-router link MTU:
h1# ping6 -s 5000 h2
router r1 does NOT send an ICMPv6 Packet Too Big message back to host h1 but rather fragments the ICMPv6 echo request packets itself before forwarding on to router r2 then ultimately host h1. The ping6's still succeed because of IPv6 fragmentation, but it is *router r1* that is doing the fragmentation instead of host h1.
This seems to be related to the setting of the "skb->local_df" flag, which should only be set for locally-generated packets; not for packets that are being forwarded by a router. Does anyone have any idea as to why my router is performing fragmentation when it forwards an ICMPv6 packet?
Thanks - Fred
[email protected]
Another clue that I should have mentioned in my initial message; the "hosts", "routers" and "links" are not physical nodes but rather network namespace virtual nodes running within a single linux 3.10.17 emulation server. Could there be some sort of weird "cross-talk" between VMs that is making the router believe it is a host?
Thanks - Fred
[email protected]
Quoted 39 lines
-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Templin, Fred L
Sent: Thursday, December 05, 2013 11:40 AM
To: [email protected]
Subject: Possible IPv6 fragmentation/pmtud bugHi,
I have the following test network configuration (all linux nodes running the 3.10.17 kernel):
h1 <-----------> r1 <-----------> r2 <-----------> h2
8000 MTU link 4000 MTU link 8000 MTU linkWhen host h1 ping6's host h2, routers r1 and r2 forward the ICMPv6 echo request/reply packets as expected. But, when I set a ping6 packet size that is larger than the router-to-router link MTU:
h1# ping6 -s 5000 h2
router r1 does NOT send an ICMPv6 Packet Too Big message back to host h1 but rather fragments the ICMPv6 echo request packets itself before forwarding on to router r2 then ultimately host h1. The ping6's still succeed because of IPv6 fragmentation, but it is *router r1* that is doing the fragmentation instead of host h1.
This seems to be related to the setting of the "
skb->local_df" flag, which should only be set for locally-generated packets; not for packets that are being forwarded by a router. Does anyone have any idea as to why my router is performing fragmentation when it forwards an ICMPv6 packet?Thanks - Fred
[email protected]
--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 5 lines · Fred L
Another clue that I should have mentioned in my initial message; the "hosts", "routers" and "links" are not physical nodes but rather network namespace virtual nodes running within a single linux 3.10.17 emulation server. Could there be some sort of weird "cross-talk" between VMs that is making the router believe it is a host?
That is probable. With a quick test I could not reproduce it. I guess we don't reset skb->local_df after creation. I'll check it out.
Thanks for the report,
Hannes
PATCH 1/1 net: clear local_df when passing skb between namespaces
We must clear local_df when passing the skb between namespaces as the packet is not local to the new namespace any more and thus may not get fragmented by local rules. Fred Templin noticed that other namespaces do fragment IPv6 packets while forwarding. Instead they should have send back a PTB.
The same problem should be present when forwarding DF-IPv4 packets between namespaces.
--- net/core/skbuff.c | 1 + 1 file changed, 1 insertion(+)
net/core/skbuff.c +1 -0
@@ -3584,6 +3584,7 @@ void skb_scrub_packet(struct sk_buff *skb, bool xnet) skb->tstamp.tv64 =0 ; skb->pkt_type =PACKET_HOST ; skb->skb_iif =0 ; + skb->local_df =0 ;skb_dst_drop (skb); skb->mark =0 ;secpath_reset (skb);
Signature · 1 line
1.8.3.1
RE: [PATCH] net: clear local_df when passing skb between namespaces
Concur - testing on an earlier kernel version shows that clearing local_df solves the problem.
Fred
Quoted 35 lines
-----Original Message-----
From: Hannes Frederic Sowa [mailto:[email protected]]
Sent: Thursday, December 05, 2013 2:29 PM
To: Templin, Fred L; [email protected]
Subject: [PATCH] net: clearlocal_dfwhen passing skb between namespacesWe must clear
local_dfwhen passing the skb between namespaces as the packet is not local to the new namespace any more and thus may not get fragmented by local rules. Fred Templin noticed that other namespaces do fragment IPv6 packets while forwarding. Instead they should have send back a PTB.The same problem should be present when forwarding DF-IPv4 packets between namespaces.
Reported-by Templin, Fred L <[email protected]>Signed-off-by Hannes Frederic Sowa <[email protected]>--- net/core/skbuff.c | 1 + 1 file changed, 1 insertion(+)net/core/skbuff.c +1
-1@@ -3584,6 +3584,7 @@ void skb_scrub_packet(struct sk_buff *skb, bool xnet) skb->tstamp.tv64 =0 ; skb->pkt_type =PACKET_HOST ; skb->skb_iif =0 ; + skb->local_df =0 ;skb_dst_drop (skb); skb->mark =0 ;secpath_reset (skb); --1.8.3.1
David Miller
maintainerRe: [PATCH] net: clear local_df when passing skb between namespaces
In reply to Hannes Frederic SowaFrom: Hannes Frederic Sowa <[email protected]>
Date: Thu, 5 Dec 2013 23:29:19 +0100
Quoted 11 lines
We must clear
local_dfwhen passing the skb between namespaces as the packet is not local to the new namespace any more and thus may not get fragmented by local rules. Fred Templin noticed that other namespaces do fragment IPv6 packets while forwarding. Instead they should have send back a PTB.The same problem should be present when forwarding DF-IPv4 packets between namespaces.
Reported-by Templin, Fred L <[email protected]>Signed-off-by Hannes Frederic Sowa <[email protected]>
Applied and queued up for -stable, thanks Hannes.