Currently pskb_trim_rcsum() is always done on the RX path. However, IPv6
jumbo packets hide the real packet len in the Hop-by-hop option header,
which should be parsed before doing the trim.
In ip6_rcv_core() it calls ipv6_parse_hopopts() to handle the Hop-by-hop
option header then do pskb_trim_rcsum(). The similar process should also
be done properly before pskb_trim_rcsum() on the RX path of bridge and
openvswitch and tc.
This patchset improves the function handling the Hop-by-hop option header
in bridge, and moves this function into netfilter utils, and then uses it
in nf_conntrack_ovs for openvswitch and and tc.
Note that this patch is especially needed after the IPv6 BIG TCP was
supported in kernel, which is using IPv6 Jumbo packets, and the last
patch adds a big tcp selftest, which also covers it.
v1->v2:
- use the proper cast type in Patch 1, as Simon suggested.
- simplify the return path in Patch 2, as Simon suggested.
- move pkt_len definition into a smaller scope in Patch 3,
as Simon suggested.
- move err definition into a smaller scope in Patch 5, as
Simon suggested.
Xin Long (6):
netfilter: bridge: call pskb_may_pull in br_nf_check_hbh_len
netfilter: bridge: check len before accessing more nh data
netfilter: bridge: move pskb_trim_rcsum out of br_nf_check_hbh_len
netfilter: move br_nf_check_hbh_len to utils
netfilter: use nf_ip6_check_hbh_len in nf_ct_skb_network_trim
selftests: add a selftest for big tcp
include/linux/netfilter_ipv6.h | 2 +
net/bridge/br_netfilter_ipv6.c | 79 ++---------
net/netfilter/nf_conntrack_ovs.c | 11 +-
net/netfilter/utils.c | 52 +++++++
tools/testing/selftests/net/Makefile | 1 +
tools/testing/selftests/net/big_tcp.sh | 180 +++++++++++++++++++++++++
6 files changed, 254 insertions(+), 71 deletions(-)
create mode 100755 tools/testing/selftests/net/big_tcp.sh
--
2.39.1
