* [PATCH net-next 1/1] net: neighbour: skip the stale dst lookup for a redirected skb
2026-11-20 15:42 [PATCH net-next 0/1] net: neighbour: skip the stale dst lookup for a redirected skb Jamal Hadi Salim
@ 2026-11-20 15:42 ` Jamal Hadi Salim
2026-11-20 15:47 ` [PATCH net-next 0/1] " netdev-bot+sinfo
1 sibling, 0 replies; 4+ messages in thread
From: Jamal Hadi Salim @ 2026-11-20 15:42 UTC (permalink / raw)
To: netdev
Cc: Jamal Hadi Salim, David S. Miller, Eric Dumazet, Jakub Kicinski,
Paolo Abeni, Simon Horman, bpf, Alexei Starovoitov, Jiri Pirko,
Victor Nogueira
Alexei Starovoitov suggested guarding the re-lookup in
neigh_update_process_arp_queue() with the skb's device instead of changing
every producer that steers a packet to another device. This is that guard.
The issue:
neigh_update_process_arp_queue() re-looks the neighbour up from the skb's
dst, because shapers, eql and sch_teql can output a packet through a
different neigh object and it has to be reinjected at the top of the path.
A packet steered to another device by bpf_redirect_neigh(), or by an nft
"fwd ... device" rule, is parked on that device's neighbour while its dst
is still the route that named the device it arrived on. The lookup returns
the neighbour of that route, so the drained frame leaves the parking
device with the arrived-on device's source address and that route's
neighbour as the destination.
The Fix:
Skip the lookup when skb->dev is already neigh->dev; the parking neighbour
is the top-level one in that case and can reinject the packet itself. A
dst that is a route via another device, which is the sch_teql shape, still
takes the lookup.
Some context:
A route for one device is a normal thing to find on a frame that leaves
another one; bpf_redirect() and mirred have always produced it, and the
packet has to keep it for the xfrm, tunnel and lwtstate consumers further
down. lwtstate, NBMA tunnels, the route realm and ICMP errors are
unaffected, and the fast path does not change.
Conditions to recreate the bug:
Two network namespaces joined by two veth pairs, va/vb in the first and
va_peer/vb_peer in the second. vb_peer carries no address, so the
neighbour for its side stays unresolved.
ip netns add ns1
ip netns add ns2
ip link add va type veth peer name va_peer
ip link add vb type veth peer name vb_peer
ip link set va netns ns1
ip link set va_peer netns ns2
ip link set vb netns ns1
ip link set vb_peer netns ns2
ip -n ns1 link set va address 02:aa:00:00:00:01
ip -n ns1 link set vb address 02:bb:00:00:00:01
ip -n ns1 addr add 10.10.0.1/24 dev va
ip -n ns1 addr add 10.1.0.1/24 dev vb
ip -n ns2 addr add 10.10.0.2/24 dev va_peer
ip -n ns1 link set va up
ip -n ns1 link set vb up
ip -n ns2 link set va_peer up
ip -n ns2 link set vb_peer up
ip -n ns1 neigh replace 10.10.0.2 lladdr 02:aa:00:00:00:ab dev va nud permanent
ip netns exec ns1 nft add table netdev t
ip netns exec ns1 nft "add chain netdev t c { type filter hook egress device va priority 0; policy accept; }"
ip netns exec ns1 nft add rule netdev t c fwd ip to 10.1.0.2 device vb
One packet to 10.10.0.2 egresses va and the rule steers it through vb to
10.1.0.2, whose neighbour is incomplete, so it is parked there with va's
route. A cls_bpf program returning
struct bpf_redir_neigh nh = { .nh_family = AF_INET,
.ipv4_nh = htonl(0x0a010002) };
return bpf_redirect_neigh(vb, &nh, sizeof(nh), 0);
in the same hook produces the same park. Making that neighbour valid
drains the queue:
ip -n ns1 neigh replace 10.1.0.2 lladdr 02:bb:00:00:00:cd dev vb nud permanent
The single frame that leaves vb_peer carries 02:aa:00:00:00:01 as its
source and 02:aa:00:00:00:ab as its destination instead of vb's own
addresses.
Link: https://lore.kernel.org/netdev/DMX7RM93TMGR.AXFN4YJ55REF@gmail.com/
Suggested-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
---
net/core/neighbour.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/net/core/neighbour.c b/net/core/neighbour.c
index 887d2e35cd04..219dc89716f8 100644
--- a/net/core/neighbour.c
+++ b/net/core/neighbour.c
@@ -1334,9 +1334,11 @@ static void neigh_update_process_arp_queue(struct neighbour *neigh)
* the packet in the output path. So what we need to do
* here is re-lookup the top-level neigh in the path so
* we can reinject the packet there.
+ * When skb->dev is neigh->dev 'neigh' is the top-level one.
+ * dst can be a route via another device after a redirect.
*/
n2 = NULL;
- if (dst &&
+ if (dst && skb->dev != neigh->dev &&
READ_ONCE(dst->obsolete) != DST_OBSOLETE_DEAD) {
n2 = dst_neigh_lookup_skb(dst, skb);
if (n2)
--
2.43.0
^ permalink raw reply related [flat|nested] 4+ messages in thread* Re: [PATCH net-next 0/1] net: neighbour: skip the stale dst lookup for a redirected skb
2026-11-20 15:42 [PATCH net-next 0/1] net: neighbour: skip the stale dst lookup for a redirected skb Jamal Hadi Salim
2026-11-20 15:42 ` [PATCH net-next 1/1] " Jamal Hadi Salim
@ 2026-11-20 15:47 ` netdev-bot+sinfo
2026-11-20 16:05 ` Jamal Hadi Salim
1 sibling, 1 reply; 4+ messages in thread
From: netdev-bot+sinfo @ 2026-11-20 15:47 UTC (permalink / raw)
To: Jamal Hadi Salim
Cc: netdev, David S. Miller, Eric Dumazet, Jakub Kicinski,
Paolo Abeni, Simon Horman, bpf, Alexei Starovoitov, Jiri Pirko,
Victor Nogueira
Hi!
This is an automated message. This series looks like a fix, but its
commit messages seem to be missing some information:
- How the issue was discovered, e.g. hit in production, hit during
development, syzbot report, manual code inspection, LLM or static
analysis tool scan.
Please do not repost the series just to address the above. Instead,
reply to this email with the missing information, so that reviewers
can take it into account.
*IMPORTANT*: If you do repost (for other reasons) - please include
the relevant info in the cover letter or commit messages.
--
The evaluation is done by an LLM so it may be wrong, if you think
that is the case please reply and explain.
^ permalink raw reply [flat|nested] 4+ messages in thread