Re: [PATCH net-next] bpf: fix cb access in socket filter programs

[Alexei Starovoitov <ast@plumgrid.com>]

On 10/7/15 6:09 AM, Daniel Borkmann wrote:
>> bpf_prog_run_clear_cb() wouldn't work on dev_forward_skb() as
>> skb->pkt_type is then being scrubbed to PACKET_HOST, so on the
>> receive path, AF_PACKET might not always see clean skbs->cb[]
>> as assumed ... I think that the skb->pkt_type part needs to be
>> dropped, no?

right. will respin.

> Thinking a bit more about this part, which only accounts for
> fanout_demux_bpf() and run_filter(), so AF_PACKET only, this
> logic still needs to be slightly different:
>
> You currently can have eBPF on packet fanout as a demux and behind
> that eBPF on the actual packet socket. So, for some reason, fanout
> could transfer some state to the socket along the way, which could
> break when cleared as-is via bpf_prog_run_clear_cb().
>
> So we need to make sure to only clear this once, either in front
> of fanout, or when not present, in front of the socket filter.

no. that would be anti-feature. user space shouldn't rely on such
things. If somebody wants to pass data between two disjoint
programs (not called via tail_call), they should be using maps or
some future per-cpu scratch area that will guarantee such semantics.
cb as part of skb is not suitable for that, since it will limit
what kernel can do in-between.