On 2020-05-06 10:26, Guillaume Nault wrote:
> On Wed, May 06, 2020 at 11:52:48AM +0200, David Balažic wrote:
>> (the last line is from my fix, the connection is now not dropped when
>> the PADT is not for us; works fine, my connection stays up and
>> working)
>>
> Looks like a more fundamental issue. This frame shouldn't have been
> accepted in the first place. Can you also print the packet class
> ("... pkt_type %u", ..., skb->pkt_type)?
>
> Testing the destination MAC here is likely to just paper over the
> problem.
Having either sketchy (e.g., CRC hash filter based) or just absent
hardware filtering of unicast addresses isn't an unknown property for
some Ethernet devices. Relying on exact match from the underlying
hardware seems like a bad idea. (Some, for instance, can either receive
"all multicast" or can filter using the low N bits of the CRC to index
into a 2^N-bit hash table, but can't really exclude a unicast packet
that happens to match the low N bits of a selectively enabled multicast
address. See, for example, the PowerQUICC implementation. But there
are quite a few that work like this. It wouldn't surprise me in the
least that 'ar71xx' [Atheros/Qualcomm?] is like this.)
I agree that having to test the MAC address right here is a bit lame and
that the driver itself ought to assist the hardware in doing the Right
Thing.
--
James Carlson 42.703N 71.076W <carlsonj@xxxxxxxxxxxxxxx>
