[xdp-hints] Re: [PATCH RFCv2 bpf-next 00/18] XDP-hints: XDP gaining access to HW offload hints via BTF

[Maryam Tahhan <mtahhan@redhat.com>]

On 05/10/2022 19:47, sdf@google.com wrote:
> On 10/05, Toke H�iland-J�rgensen wrote:
>> Stanislav Fomichev <sdf@google.com> writes:
> 
>> > On Tue, Oct 4, 2022 at 5:59 PM Jakub Kicinski <kuba@kernel.org> wrote:
>> >>
>> >> On Tue, 4 Oct 2022 17:25:51 -0700 Martin KaFai Lau wrote:
>> >> > A intentionally wild question, what does it take for the driver 
>> to return the
>> >> > hints.  Is the rx_desc and rx_queue enough?  When the xdp prog is 
>> calling a
>> >> > kfunc/bpf-helper, like 'hwtstamp = bpf_xdp_get_hwtstamp()', can 
>> the driver
>> >> > replace it with some inline bpf code (like how the inline code is 
>> generated for
>> >> > the map_lookup helper).  The xdp prog can then store the hwstamp 
>> in the meta
>> >> > area in any layout it wants.
>> >>
>> >> Since you mentioned it... FWIW that was always my preference rather 
>> than
>> >> the BTF magic :)  The jited image would have to be per-driver like we
>> >> do for BPF offload but that's easy to do from the technical
>> >> perspective (I doubt many deployments bind the same prog to multiple
>> >> HW devices)..
>> >
>> > +1, sounds like a good alternative (got your reply while typing)
>> > I'm not too versed in the rx_desc/rx_queue area, but seems like worst
>> > case that bpf_xdp_get_hwtstamp can probably receive a xdp_md ctx and
>> > parse it out from the pre-populated metadata?
>> >
>> > Btw, do we also need to think about the redirect case? What happens
>> > when I redirect one frame from a device A with one metadata format to
>> > a device B with another?
> 
>> Yes, we absolutely do! In fact, to me this (redirects) is the main
>> reason why we need the ID in the packet in the first place: when running
>> on (say) a veth, an XDP program needs to be able to deal with packets
>> from multiple physical NICs.
> 
>> As far as API is concerned, my hope was that we could solve this with a
>> CO-RE like approach where the program author just writes something like:
> 
>> hw_tstamp = bpf_get_xdp_hint("hw_tstamp", u64);
> 
>> and bpf_get_xdp_hint() is really a macro (or a special kind of
>> relocation?) and libbpf would do the following on load:
> 
>> - query the kernel BTF for all possible xdp_hint structs
>> - figure out which of them have an 'u64 hw_tstamp' member
>> - generate the necessary conditionals / jump table to disambiguate on
>>    the BTF_ID in the packet
> 
> 
>> Now, if this is better done by a kfunc I'm not terribly opposed to that
>> either, but I'm not sure it's actually better/easier to do in the kernel
>> than in libbpf at load time?
> 
> Replied in the other thread, but to reiterate here: then btf_id in the
> metadata has to stay and we either pre-generate those bpf_get_xdp_hint()
> at libbpf or at kfunc load time level as you mention.
> 
> But the program essentially has to handle all possible hints' btf ids 
> thrown
> at it by the system. Not sure about the performance in this case :-/
> Maybe that's something that can be hidden behind "I might receive forwarded
> packets and I know how to handle all metadata format" flag? By default,
> we'll pre-generate parsing only for that specific device?

I did a simple POC of Jespers xdp-hints with AF-XDP and CNDP (Cloud 
Native Data Plane). In the cases where my app had access to the HW I 
didn't need to handle all possible hints... I knew what Drivers were on 
the system and they were the hints I needed to deal with.

So at program init time I registered the relevant BTF_IDs (and some 
callback functions to handle them) from the NICs that were available to 
me in a simple tailq (tbh there were so few I could've probably used a 
static array).

When processing the hints then I only needed to invoke the appropriate 
callback function based on the received BTF_ID. I didn't have a massive 
chains of if...else if... else statements.

In the case where we have redirection to a virtual NIC and we don't 
necessarily know the underlying hints that are exposed to the app, could 
we not still use the xdp_hints (as proposed by Jesper) themselves to 
indicate the relevant drivers to the application? or even indicate them 
via a map or something?