[PATCH bpf-next] bpf: fix xdp_generic for bpf_adjust_tail usecase

[PATCH bpf-next] bpf: fix xdp_generic for bpf_adjust_tail usecase

[Nikita V. Shirokov]

 when bpf_adjust_tail was introduced for generic xdp, it changed skb's tail
 pointer, so it was pointing to the new  "end of the packet". however skb's
 len field wasn't properly modified, so on the wire ethernet frame had
 original (or even bigger, if adjust_head was used) size. this diff is fixing
 this.

Fixes: 198d83bb3 (" bpf: make generic xdp compatible w/
bpf_xdp_adjust_tail")

Signed-off-by: Nikita V. Shirokov <tehnerd@tehnerd.com>
---

Notes:
    original tests missed this because it looks like tap interface
    ignores incorrect ethernet FCS (all tests were done in VM)
    and even w/ missaligned l3 and l2 lengths, kernel still were
    accepting this ICMP packet
    output was generated w/ bpf_adjust_tail prog from samples
    before this fix (see lengths field of the ethernet layer):
    
    tehnerd@maindev:~$ sudo tcpdump -ni tap0 icmp -vvv -eee
    tcpdump: listening on tap0, link-type EN10MB (Ethernet), capture size 262144 bytes
    06:38:15.546782 52:54:00:12:34:57 > 12:0e:a3:cc:78:b8, ethertype IPv4 (0x0800), length 1454: (tos 0x0, ttl 64, id 0, offset 0, flags [none], proto ICMP (1), length 112)
        172.16.0.2 > 172.16.0.1: ICMP 172.16.0.2 unreachable - need to frag (mtu 586), length 92
            (tos 0x0, ttl 64, id 48021, offset 0, flags [DF], proto TCP (6), length 1412)
        172.16.0.1.50916 > 172.16.0.2.22: Flags [P.], seq 427401155:427402515, ack 3567613893, win 229, options [nop,nop,TS val 1287434011 ecr 2176566223], length 1360
    
    after:
    tehnerd@maindev:~$ sudo tcpdump -ni tap0 icmp -vvv -eee
    tcpdump: listening on tap0, link-type EN10MB (Ethernet), capture size 262144 bytes
    06:47:37.226843 52:54:00:12:34:57 > 32:45:9f:69:35:ba, ethertype IPv4 (0x0800), length 126: (tos 0x0, ttl 64, id 0, offset 0, flags [none], proto ICMP (1), length 112)
    
    172.16.0.2 > 172.16.0.1: ICMP 172.16.0.2 unreachable - need to frag (mtu 586), length 92
        (tos 0x0, ttl 64, id 29964, offset 0, flags [DF], proto TCP (6), length 1412)
    172.16.0.1.50918 > 172.16.0.2.22: Flags [P.], seq 14171614:14172974, ack 1433043471, win 229, options [nop,nop,TS val 1287995744 ecr 3312743811], length 1360

 net/core/dev.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index c624a04dad1f..8f8931b93140 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4057,8 +4057,10 @@ static u32 netif_receive_generic_xdp(struct sk_buff *skb,
 	 * pckt.
 	 */
 	off = orig_data_end - xdp.data_end;
-	if (off != 0)
+	if (off != 0) {
 		skb_set_tail_pointer(skb, xdp.data_end - xdp.data);
+		skb->len -= off;
+	}
 
 	switch (act) {
 	case XDP_REDIRECT:
-- 
2.15.1

Re: [PATCH bpf-next] bpf: fix xdp_generic for bpf_adjust_tail usecase

[Daniel Borkmann]

On 04/25/2018 04:15 PM, Nikita V. Shirokov wrote:
>  when bpf_adjust_tail was introduced for generic xdp, it changed skb's tail
>  pointer, so it was pointing to the new  "end of the packet". however skb's
>  len field wasn't properly modified, so on the wire ethernet frame had
>  original (or even bigger, if adjust_head was used) size. this diff is fixing
>  this.
> 
> Fixes: 198d83bb3 (" bpf: make generic xdp compatible w/
> bpf_xdp_adjust_tail")
> 
> Signed-off-by: Nikita V. Shirokov <tehnerd@tehnerd.com>

Applied yesterday to bpf-next (and now in net-next), thanks Nikita!