[xdp-hints] Re: [RFC bpf-next v2 06/14] xdp: Carry over xdp metadata into skb context

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

Snipping a bit of context to reply to this bit:

>>>> Can the xdp prog still change the metadata through xdp->data_meta? tbh, I am not
>>>> sure it is solid enough by asking the xdp prog not to use the same random number
>>>> in its own metadata + not to change the metadata through xdp->data_meta after
>>>> calling bpf_xdp_metadata_export_to_skb().
>>>
>>> What do you think the usecase here might be? Or are you suggesting we
>>> reject further access to data_meta after
>>> bpf_xdp_metadata_export_to_skb somehow?
>>>
>>> If we want to let the programs override some of this
>>> bpf_xdp_metadata_export_to_skb() metadata, it feels like we can add
>>> more kfuncs instead of exposing the layout?
>>>
>>> bpf_xdp_metadata_export_to_skb(ctx);
>>> bpf_xdp_metadata_export_skb_hash(ctx, 1234);

There are several use cases for needing to access the metadata after
calling bpf_xdp_metdata_export_to_skb():

- Accessing the metadata after redirect (in a cpumap or devmap program,
  or on a veth device)
- Transferring the packet+metadata to AF_XDP
- Returning XDP_PASS, but accessing some of the metadata first (whether
  to read or change it)

The last one could be solved by calling additional kfuncs, but that
would be less efficient than just directly editing the struct which
will be cache-hot after the helper returns.

And yeah, this will allow the XDP program to inject arbitrary metadata
into the netstack; but it can already inject arbitrary *packet* data
into the stack, so not sure if this is much of an additional risk? If it
does lead to trivial crashes, we should probably harden the stack
against that?

As for the random number, Jesper and I discussed replacing this with the
same BTF-ID scheme that he was using in his patch series. I.e., instead
of just putting in a random number, we insert the BTF ID of the metadata
struct at the end of it. This will allow us to support multiple
different formats in the future (not just changing the layout, but
having multiple simultaneous formats in the same kernel image), in case
we run out of space.

We should probably also have a flag set on the xdp_frame so the stack
knows that the metadata area contains relevant-to-skb data, to guard
against an XDP program accidentally hitting the "magic number" (BTF_ID)
in unrelated stuff it puts into the metadata area.

> After re-reading patch 6, have another question. The 'void
> bpf_xdp_metadata_export_to_skb();' function signature. Should it at
> least return ok/err? or even return a 'struct xdp_to_skb_metadata *'
> pointer and the xdp prog can directly read (or even write) it?

Hmm, I'm not sure returning a failure makes sense? Failure to read one
or more fields just means that those fields will not be populated? We
should probably have a flags field inside the metadata struct itself to
indicate which fields are set or not, but I'm not sure returning an
error value adds anything? Returning a pointer to the metadata field
might be convenient for users (it would just be an alias to the
data_meta pointer, but the verifier could know its size, so the program
doesn't have to bounds check it).

> A related question, why 'struct xdp_to_skb_metadata' needs
> __randomize_layout?

The __randomize_layout thing is there to force BPF programs to use CO-RE
to access the field. This is to avoid the struct layout accidentally
ossifying because people in practice rely on a particular layout, even
though we tell them to use CO-RE. There are lots of examples of this
happening in other domains (IP header options, TCP options, etc), and
__randomize_layout seemed like a neat trick to enforce CO-RE usage :)

>>>> Does xdp_to_skb_metadata have a use case for XDP_PASS (like patch 7) or the
>>>> xdp_to_skb_metadata can be limited to XDP_REDIRECT only?
>>>
>>> XDP_PASS cases where we convert xdp_buff into skb in the drivers right
>>> now usually have C code to manually pull out the metadata (out of hw
>>> desc) and put it into skb.
>>>
>>> So, currently, if we're calling bpf_xdp_metadata_export_to_skb() for
>>> XDP_PASS, we're doing a double amount of work:
>>> skb_metadata_import_from_xdp first, then custom driver code second.
>>>
>>> In theory, maybe we should completely skip drivers custom parsing when
>>> there is a prog with BPF_F_XDP_HAS_METADATA?
>>> Then both xdp->skb paths (XDP_PASS+XDP_REDIRECT) will be bpf-driven
>>> and won't require any mental work (plus, the drivers won't have to
>>> care either in the future).
>>>  > WDYT?
>> 
>> 
>> Yeah, not sure if it can solely depend on BPF_F_XDP_HAS_METADATA but it makes 
>> sense to only use the hints (if ever written) from xdp prog especially if it 
>> will eventually support xdp prog changing some of the hints in the future.  For 
>> now, I think either way is fine since they are the same and the xdp prog is sort 
>> of doing extra unnecessary work anyway by calling 
>> bpf_xdp_metadata_export_to_skb() with XDP_PASS and knowing nothing can be 
>> changed now.

I agree it would be best if the drivers also use the XDP metadata (if
present) on XDP_PASS. Longer term my hope is we can make the XDP
metadata support the only thing drivers need to implement (i.e., have
the stack call into that code even when no XDP program is loaded), but
for now just for consistency (and allowing the XDP program to update the
metadata), we should probably at least consume it on XDP_PASS.

-Toke