Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 27 Jul 2023 17:25:11 +0200
From:   Richard Gobert <richardbgobert@gmail.com>
To:     davem@davemloft.net, edumazet@google.com, kuba@kernel.org,
        pabeni@redhat.com, willemdebruijn.kernel@gmail.com,
        dsahern@kernel.org, tom@herbertland.com, netdev@vger.kernel.org,
        linux-kernel@vger.kernel.org, gal@nvidia.com
Subject: [PATCH v3 0/1] net: gro: fix misuse of CB in udp socket lookup
Message-ID: <20230727152503.GA32010@debian>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
User-Agent: Mutt/1.10.1 (2018-07-13)
Precedence: bulk

GRO stack uses `udp_lib_lookup_skb` which relies on IP/IPv6 CB's info, and
at the GRO stage, CB holds `napi_gro_cb` info. Specifically,
`udp_lib_lookup_skb` tries to fetch `iff` and `flags` information from the
CB, to find the relevant udp tunnel socket (GENEVE/VXLAN/..). Up until a
patch I submitted recently [0], it worked merely by luck, due
to the layouts of `napi_gro_cb` and IP6CB.

AFAIU it worked because:
`IP6CB(skb)->flags` is at offset 16 inside IP6CB:
 - Before the patch: `flags` was mapped to `flush`.
 - After the patch: `flags` was mapped to `data_offset`.

`IP6CB(skb)->iff` is at offset 0 inside IP6CB:
 - Before the patch: `iif` was mapped to `frag0`.
 - After the patch: `iif` was mapped to a union of `frag0` and `last`.

After my patch, on the receive phase, while `data_offset` is 40 (since IPv6
header is 40 bytes), `inet_iif` calls `ipv6_l3mdev_skb`, which checks
whether `IP6CB(skb)->flags`'s `IP6SKB_L3SLAVE` bit is on or off (in our
case its off). If it is off, `inet_iif` returns `IP6CB(skb)->iif`, which is
mapped to `napi_gro_cb->frag0`, making `inet_iif` return 0 most of the
times. `inet_sdif` returns zero due to a similar reason caused by
`data_offset` being equal to 40 (and less than 64).

On the other hand, the complete phase behaves differently.
`data_offset` is usually greater than 64 and less than 128 so the
`IP6SKB_L3SLAVE` flag is on.  Thus, `inet_sdif` returns `IP6CB(skb)->iif`,
which is mapped to `last` which contains a pointer. This causes
`udp_sk_bound_dev_eq` to fail, which leads to `udp6_lib_lookup2` failing
and not returning a socket. This leads the receive phase of GRO
to find the right socket, and on the complete phase, it fails to find it 
and makes the throughput go down to nearly zero.

Before [0] `flags` was mapped to `flush`. `flush`'s possible
values were 1 and 0, making `inet6_iff` always returning `skb->skb_iif` and
`inet6_sdif` returning 0, and leading to `udp_sk_bound_dev_eq` returning
true.

A fix is to not rely on CB, and get `iff` and `sdif` using skb->dev. l3mdev
case requires special attention since it has a master and a slave device.

[0] https://lore.kernel.org/netdev/20230601160924.GA9194@debian/

Changelog:

v2 -> v3:
  * change functions to static inline in the header file

v1 -> v2:
  * make functions inline
  * fix logical bug
  * add a comment when we can use the new functions
  * checkpatch fixes

Richard Gobert (1):
  net: gro: fix misuse of CB in udp socket lookup

 include/net/gro.h      | 43 ++++++++++++++++++++++++++++++++++++++++++
 net/ipv4/udp.c         |  8 ++++++--
 net/ipv4/udp_offload.c |  7 +++++--
 net/ipv6/udp.c         |  8 ++++++--
 net/ipv6/udp_offload.c |  7 +++++--
 5 files changed, 65 insertions(+), 8 deletions(-)

-- 
2.36.1