Re: [PATCH bpf-next v2 0/3] Add XDP support for bpf_load_hdr_opt

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Martin KaFai Lau <kafai@fb.com> writes:

> On Sat, Oct 09, 2021 at 12:20:27AM +0200, Toke Høiland-Jørgensen wrote:
>> So if we can't fix the verifier, maybe we could come up with a more
>> general helper for packet parsing? Something like:
>> 
>> bpf_for_each_pkt_chunk(ctx, offset, callback_fn, callback_arg)
>> {
>>   ptr = ctx->data + offset;
>>   while (ptr < ctx->data_end) {
>>     offset = callback_fn(ptr, ctx->data_end, callback_arg);
>>     if (offset == 0)
>>       return 0;
>>     ptr += offset;
>>   }
>>   
>>   // out of bounds before callback was done
>>   return -EINVAL;
>> }
>>    
>> This would work for parsing any kind of packet header or TLV-style data
>> without having to teach the kernel about each header type. It'll have
>> quite a bit of overhead if all the callbacks happen via indirect calls,
>> but maybe the verifier can inline the calls (or at least turn them into
>> direct CALL instructions)?
> Direct call different callback_fn?  bpf_for_each_pkt_chunk() is a kernel
> function.  It would be nice if the verifier could do that.

Ohh, right, think-o on my part. It could be done if the helper was
inlined in its entirety, though? Not sure if that's feasible?

> This for_each helper had been considered also. Other than the need to
> callback in a loop, the thought was to extend the existing
> bpf_load_hdr_opt() because our initial feedback is the same header
> handling logic cannot be used in xdp which is confusing.

TBH, I had not noticed this helper before. Now that I have, it does
seems like the kind of thing that belongs as a BPF library function
rather than a helper in the first place :)

> I don't mind to go with the for_each helper.  However, with another
> thought, if it needs to call a function in the loop anyway, I think
> it could also be done in bpf by putting a global function in a loop.
> Need to try and double check.

Hmm, that would be interesting if possible!

-Toke