[SRU][Xenial][PATCH 0/1] skb_warn_bad_offload kernel splat due to CHECKSUM target not compatible with GSO skbs

[Matthew Ruffell]

BugLink: https://bugs.launchpad.net/bugs/1840619

[Impact]

In environments which have CHECKSUM iptables rules set, the following kernel 
call trace will be created when a GSO skb is processed by the CHECKSUM target:

WARNING: CPU: 34 PID: 806048 at /build/linux-zdslHp/linux-4.4.0/net/core/dev.c
:2456 skb_warn_bad_offload+0xcf/0x110()
qr-f78bfdf7-fe: caps=(0x000000000fdb58e9, 0x000000000fdb58e9) len=1955 
data_len=479 gso_size=1448 gso_type=1 ip_summed=3
CPU: 34 PID: 806048 Comm: haproxy Tainted: G W OE 4.4.0-138-generic #164-Ubuntu
Call Trace:
 dump_stack+0x63/0x90
 warn_slowpath_common+0x82/0xc0
 warn_slowpath_fmt+0x5c/0x80
 ? ___ratelimit+0xa2/0xe0
 skb_warn_bad_offload+0xcf/0x110
 skb_checksum_help+0x185/0x1a0
 checksum_tg+0x22/0x29 [xt_CHECKSUM]
 ipt_do_table+0x301/0x730 [ip_tables]
 ? ipt_do_table+0x349/0x730 [ip_tables]
 iptable_mangle_hook+0x39/0x107 [iptable_mangle]
 nf_iterate+0x68/0x80
 nf_hook_slow+0x73/0xd0
 ip_output+0xcf/0xe0
 ? __ip_flush_pending_frames.isra.43+0x90/0x90
 ip_local_out+0x3b/0x50
 ip_queue_xmit+0x154/0x390
 __tcp_transmit_skb+0x52b/0x9b0
 tcp_write_xmit+0x1dd/0xf50
 __tcp_push_pending_frames+0x31/0xd0
 tcp_push+0xec/0x110
 tcp_sendmsg+0x749/0xba0
 inet_sendmsg+0x6b/0xa0
 sock_sendmsg+0x3e/0x50
 SYSC_sendto+0x101/0x190
 ? __sys_sendmsg+0x51/0x90
 SyS_sendto+0xe/0x10
 entry_SYSCALL_64_fastpath+0x22/0xc1

The CHECKSUM target does not support GSO skbs, and when a GSO skb is passed to 
skb_checksum_help(), it errors out and skb_warn_bad_offload() is called.

The above call trace was found in a customer environment which has an Openstack 
deployment, with the following sorts of iptables rules set:

-A neutron-l3-agent-POSTROUTING -o qr-+ -p tcp -m tcp --sport 9697 -j CHECKSUM 
--checksum-fill
-A neutron-dhcp-age-POSTROUTING -p udp -m udp --dport 68 -j CHECKSUM 
--checksum-fill

This was causing haproxy running on the node to crash and restart every time a 
GSO skb was processed by the CHECKSUM target.

I recommend reading the netdev mailing list thread for more details:
https://www.spinics.net/lists/netdev/msg517366.html

[Fix]

This was fixed in 4.19 upstream with the below commit:

commit 10568f6c5761db24249c610c94d6e44d5505a0ba
Author: Florian Westphal <fw at strlen.de>
Date: Wed Aug 22 11:33:27 2018 +0200
Subject: netfilter: xt_checksum: ignore gso skbs

This commit adds a check to see if the current skb is a gso skb, and if it is, 
skips skb_checksum_help(). It then continues on to check if the packet uses udp,
and if it does, exits early. Otherwise it prints a single warning that CHECKSUM
should be avoided, and if really needed, only for use with outbound udp.

Note, 10568f6c5761db24249c610c94d6e44d5505a0ba was included in upstream stable 
version 4.18.13, and was backported to bionic in 4.15.0-58.64 by LP #1836426.

This patch required minor backporting for 4.4, by slightly adjusting the context
in the final patch hunk.

[Testcase]

You can reproduce this by adding the following iptables rule to the mangle 
table:

-t mangle -A POSTROUTING -p tcp -m tcp --sport 80 -j CHECKSUM --checksum-fill

and running traffic over port 80 with incorrect checksums in the ip header.

I built a test kernel, which is available here:

https://launchpad.net/~mruffell/+archive/ubuntu/sf216537-test

For unpatched kernels, this causes the process which was handling the socket to
crash, as seen by haproxy crashing on a node in production which hits this 
issue.

On patched kernels you see the below warning printed to dmesg and no crashes 
occur.

xt_CHECKSUM: CHECKSUM should be avoided.  If really needed, restrict with 
"-p udp" and only use in OUTPUT

[Regression Potential]

The changes are limited only to users which have CHECKSUM rules enabled in their
iptables configs. Openstack commonly configures such rules on deployment, even 
though they are not necessary, as almost all packets have their checksum 
calculated by NICs these days, and CHECKSUM is only around to service old dhcp 
clients which would discard UDP packets with empty checksums.

This commit was selected for upstream -stable 4.18.13, and has made its way into
bionic 4.15.0-58.64 by LP #1836426. There have been no reported problems and 
those kernels would have had sufficient testing with Openstack and its 
configured iptables rules.

If any users are affected by regression, then they can simply delete any 
CHECKSUM entries in their iptables configs.

Florian Westphal (1):
  netfilter: xt_checksum: ignore gso skbs

 net/netfilter/Kconfig       | 12 ++++++------
 net/netfilter/xt_CHECKSUM.c | 23 ++++++++++++++++++++++-
 2 files changed, 28 insertions(+), 7 deletions(-)

-- 
2.20.1