On 8/17/20, Jesper Dangaard Brouer <brouer@xxxxxxxxxx> wrote:
> On Sun, 16 Aug 2020 15:29:37 -0700 (PDT)
> David Miller <davem@xxxxxxxxxxxxx> wrote:
>
>> From: "Jason A. Donenfeld" <Jason@xxxxxxxxx>
>> Date: Sat, 15 Aug 2020 09:29:30 +0200
>>
>> > When an XDP program changes the ethernet header protocol field,
>> > eth_type_trans is used to recalculate skb->protocol. In order for
>> > eth_type_trans to work correctly, the ethernet header must actually be
>> > part of the skb data segment, so the code first pushes that onto the
>> > head of the skb. However, it subsequently forgets to pull it back off,
>> > making the behavior of the passed-on packet inconsistent between the
>> > protocol modifying case and the static protocol case. This patch fixes
>> > the issue by simply pulling the ethernet header back off of the skb
>> > head.
>> >
>> > Fixes: 297249569932 ("net: fix generic XDP to handle if eth header was
>> > mangled")
>> > Cc: Jesper Dangaard Brouer <brouer@xxxxxxxxxx>
>> > Cc: David S. Miller <davem@xxxxxxxxxxxxx>
>> > Signed-off-by: Jason A. Donenfeld <Jason@xxxxxxxxx>
>>
>> Applied and queued up for -stable, thanks.
>>
>> Jesper, I wonder how your original patch was tested because it pushes a
>> packet
>> with skb->data pointing at the ethernet header into the stack. That
>> should be
>> popped at this point as per this fix here.
>
> I think this patch is wrong, because eth_type_trans() also does a
> skb_pull_inline(skb, ETH_HLEN).
Huh, wow. That's one unusual and confusing function. But indeed it
seems like I'm the one who needs to reevaluate testing methodology
here. I'm very sorry for the noise and hassle.
Jason
