netfilter: br_netfilter: Drop dst references before setting.
authorHarsh Modi <harshmodi@google.com>
Wed, 31 Aug 2022 05:36:03 +0000 (22:36 -0700)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 15 Sep 2022 10:39:46 +0000 (12:39 +0200)
commit92c4fe831a72c432b66193699a7664fa6a229e50
treee152bbeae31d0692fefde38c019a0920d75e0e0a
parentd574fee3d4ff51daaa8591e8f61a2bab666ef880
netfilter: br_netfilter: Drop dst references before setting.

[ Upstream commit d047283a7034140ea5da759a494fd2274affdd46 ]

The IPv6 path already drops dst in the daddr changed case, but the IPv4
path does not. This change makes the two code paths consistent.

Further, it is possible that there is already a metadata_dst allocated from
ingress that might already be attached to skbuff->dst while following
the bridge path. If it is not released before setting a new
metadata_dst, it will be leaked. This is similar to what is done in
bpf_set_tunnel_key() or ip6_route_input().

It is important to note that the memory being leaked is not the dst
being set in the bridge code, but rather memory allocated from some
other code path that is not being freed correctly before the skb dst is
overwritten.

An example of the leakage fixed by this commit found using kmemleak:

unreferenced object 0xffff888010112b00 (size 256):
  comm "softirq", pid 0, jiffies 4294762496 (age 32.012s)
  hex dump (first 32 bytes):
    00 00 00 00 00 00 00 00 80 16 f1 83 ff ff ff ff  ................
    e1 4e f6 82 ff ff ff ff 00 00 00 00 00 00 00 00  .N..............
  backtrace:
    [<00000000d79567ea>] metadata_dst_alloc+0x1b/0xe0
    [<00000000be113e13>] udp_tun_rx_dst+0x174/0x1f0
    [<00000000a36848f4>] geneve_udp_encap_recv+0x350/0x7b0
    [<00000000d4afb476>] udp_queue_rcv_one_skb+0x380/0x560
    [<00000000ac064aea>] udp_unicast_rcv_skb+0x75/0x90
    [<000000009a8ee8c5>] ip_protocol_deliver_rcu+0xd8/0x230
    [<00000000ef4980bb>] ip_local_deliver_finish+0x7a/0xa0
    [<00000000d7533c8c>] __netif_receive_skb_one_core+0x89/0xa0
    [<00000000a879497d>] process_backlog+0x93/0x190
    [<00000000e41ade9f>] __napi_poll+0x28/0x170
    [<00000000b4c0906b>] net_rx_action+0x14f/0x2a0
    [<00000000b20dd5d4>] __do_softirq+0xf4/0x305
    [<000000003a7d7e15>] __irq_exit_rcu+0xc3/0x140
    [<00000000968d39a2>] sysvec_apic_timer_interrupt+0x9e/0xc0
    [<000000009e920794>] asm_sysvec_apic_timer_interrupt+0x16/0x20
    [<000000008942add0>] native_safe_halt+0x13/0x20

Florian Westphal says: "Original code was likely fine because nothing
ever did set a skb->dst entry earlier than bridge in those days."

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Signed-off-by: Harsh Modi <harshmodi@google.com>
Acked-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Sasha Levin <sashal@kernel.org>
net/bridge/br_netfilter_hooks.c
net/bridge/br_netfilter_ipv6.c