Re: XDP-hints: Howto support multiple BTF types per packet basis?

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Thu, 24 Jun 2021 18:04:48 +0200
Toke Høiland-Jørgensen <toke@redhat.com> wrote:

> Zvi Effron via xdp-hints <xdp-hints@xdp-project.net> writes:
> 
> > On Thu, Jun 24, 2021 at 5:23 AM Toke Høiland-Jørgensen <toke@redhat.com> wrote:  
> >>
> >> Michal Swiatkowski <michal.swiatkowski@linux.intel.com> writes:
> >>  
> >> > On Tue, Jun 22, 2021 at 01:53:33PM +0200, Toke Høiland-Jørgensen wrote:  
> >> >> Michal Swiatkowski <michal.swiatkowski@linux.intel.com> writes:
> >> >>  
> >> >> > On Wed, Jun 02, 2021 at 09:18:37AM -0700, Jakub Kicinski wrote:  
> >> >> >> On Tue, 01 Jun 2021 17:22:51 -0700 John Fastabend wrote:  
> >> >> >> > > If we do this, the BPF program obviously needs to know which fields are
> >> >> >> > > valid and which are not. AFAICT you're proposing that this should be
> >> >> >> > > done out-of-band (i.e., by the system administrator manually ensuring
> >> >> >> > > BPF program config fits system config)? I think there are a couple of
> >> >> >> > > problems with this:
> >> >> >> > >
> >> >> >> > > - It requires the system admin to coordinate device config with all of
> >> >> >> > >   their installed XDP applications. This is error-prone, especially as
> >> >> >> > >   the number of applications grows (say if different containers have
> >> >> >> > >   different XDP programs installed on their virtual devices).  
> >> >> >> >
> >> >> >> > A complete "system" will need to be choerent. If I forward into a veth
> >> >> >> > device the orchestration component needs to ensure program sending
> >> >> >> > bits there is using the same format the program installed there expects.
> >> >> >> >
> >> >> >> > If I tailcall/fentry into another program that program the callee and
> >> >> >> > caller need to agree on the metadata protocol.
> >> >> >> >
> >> >> >> > I don't see any way around this. Someone has to manage the network.  
> >> >> >>
> >> >> >> FWIW I'd like to +1 Toke's concerns.
> >> >> >>
> >> >> >> In large deployments there won't be a single arbiter. Saying there
> >> >> >> is seems to contradict BPF maintainers' previous stand which lead
> >> >> >> to addition of bpf_links for XDP.
> >> >> >>
> >> >> >> In practical terms person rolling out an NTP config change may not
> >> >> >> be aware that in some part of the network some BPF program expects
> >> >> >> descriptor not to contain time stamps. Besides features may depend
> >> >> >> or conflict so the effects of feature changes may not be obvious
> >> >> >> across multiple drivers in a heterogeneous environment.
> >> >> >>
> >> >> >> IMO guarding from obvious mis-configuration provides obvious value.  
> >> >> >
> >> >> > Hi,
> >> >> >
> >> >> > Thanks for a lot of usefull information about CO-RE. I have read
> >> >> > recommended articles, but still don't understand everything, so sorry if
> >> >> > my questions are silly.
> >> >> >
> >> >> > As introduction, I wrote small XDP example using CO-RE (autogenerated
> >> >> > vmlinux.h and getting rid of skeleton etc.) based on runqslower
> >> >> > implementation. Offset reallocation of hints works great, I built CO-RE
> >> >> > application, added new field to hints struct, changed struct layout and
> >> >> > without rebuilding application everything still works fine. Is it worth
> >> >> > to add XDP sample using CO-RE in kernel or this isn't good place for
> >> >> > this kind of sample?
> >> >> >
> >> >> > First question not stricte related to hints. How to get rid of #define
> >> >> > and macro when I am using generated vmlinux.h? For example I wanted to
> >> >> > use htons macro and ethtype definition. They are located in headers that
> >> >> > also contains few struct definition. Because of that I have redefinition
> >> >> > error when I am trying to include them (redefinition in vmlinux.h and
> >> >> > this included file). What can I do with this besides coping definitions
> >> >> > to bpf code?  
> >> >>
> >> >> One way is to only include the structs you actually need from vmlinux.h.
> >> >> You can even prune struct members, since CO-RE works just fine with
> >> >> partial struct definitions as long as the member names match.
> >> >>
> >> >> Jesper has an example on how to handle this here:
> >> >> https://github.com/netoptimizer/bpf-examples/blob/ktrace01-CO-RE.public/headers/vmlinux_local.h

Above links to my experimental "learning-by-doing" branch.  I've
created a PR to merge this officially here:
 https://github.com/xdp-project/bpf-examples/pull/24/

> >> >
> >> > I see, thanks, I will take a look at other examples.
> >> >  
> >> >> > I defined hints struct in driver code, is it right place for that? All
> >> >> > vendors will define their own hints struct or the idea is to have one
> >> >> > big hints struct with flags informing about availability of each fields?
> >> >> >
> >> >> > For me defining it in driver code was easier because I can have used
> >> >> > module btf to generate vmlinux.h with hints struct inside. However this
> >> >> > break portability if other vendors will have different struct name etc,
> >> >> > am I right?  
> >> >>
> >> >> I would expect the easiest is for drivers to just define their own
> >> >> structs and maybe have some infrastructure in the core to let userspace
> >> >> discover the right BTF IDs to use for a particular netdev. However, as
> >> >> you say it's not going to work if every driver just invents their own
> >> >> field names, so we'll need to coordinate somehow. We could do this by
> >> >> convention, though, it'll need manual intervention to make sure the
> >> >> semantics of identically-named fields match anyway.
> >> >>
> >> >> Cf the earlier discussion with how many BTF IDs each driver might
> >> >> define, I think we *also* need a way to have flags that specify which
> >> >> fields of a given BTF ID are currently used; and having some common
> >> >> infrastructure for that would be good...
> >> >>  
> >> >
> >> > Sounds good.
> >> >
> >> > Sorry, but I feel that I don't fully understand the idea. Correct me if
> >> > I am wrong:
> >> >
> >> > In building CO-RE application step we can defined big struct with
> >> > all possible fields or even empty struct (?) and use
> >> > bpf_core_field_exists.
> >> >
> >> > bpf_core_field_exists will be resolve before loading program by libbpf
> >> > code. In normal case libbpf will look for btf with hints name in vmlinux
> >> > of running kernel and do offset rewrite and exsistence check. But as the
> >> > same hints struct will be define in multiple modules we want to add more
> >> > logic to libbpf to discover correct BTF ID based on netdev on which program
> >> > will be loaded?  
> >>
> >> I would expect that the program would decide ahead-of-time which BTF IDs
> >> it supports, by something like including the relevant structs from
> >> vmlinux.h. And then we need the BTF ID encoded into the packet metadata
> >> as well, so that it is possible to check at run-time which driver the
> >> packet came from (since a packet can be redirected, so you may end up
> >> having to deal with multiple formats in the same XDP program).
> >>
> >> Which would allow you to write code like:
> >>
> >> if (ctx->has_driver_meta) {
> >>   /* this should be at a well-known position, like first (or last) in meta area */
> >>   __u32 *meta_btf_id = ctx->data_meta;
> >>
> >>   if (*meta_btf_id == BTF_ID_MLX5) {
> >>     struct meta_mlx5 *meta = ctx->data_meta;
> >>     /* do something with meta */
> >>   } else if (meta_btf_id == BTF_ID_I40E) {
> >>     struct meta_i40e *meta = ctx->data_meta;
> >>     /* do something with meta */
> >>   } /* etc */
> >> }
> >>
> >> and libbpf could do relocations based on the different meta structs,
> >> even removing the code for the ones that don't exist on the running
> >> kernel.
> >>
> >> -Toke
> >>  
> >
> > How does putting the BTF ID and the driver metadata into the XDP metadata
> > section interact with programs that are already using the metadata section
> > for other purposes. For example, programs that use the XDP metadata to pass
> > information through BPF tail calls?
> >
> > Would this break existing programs that aren't aware of the new driver
> > metadata? Do we need to make driver metadata opt-in at XDP program
> > load?  
> 
> Well, XDP applications would be free to just ignore the driver-provided
> metadata and overwrite it with its own data? And I guess any application
> that doesn't know about it will just implicitly do that? :)

Remember to wrap your head around: That metadata area "grows" via minus
offset as ctx->data_meta points to area before ctx->data. See[1] where
bpf_xdp_adjust_meta() helper does a minus adjustment.

[1] https://github.com/torvalds/linux/blob/v5.13-rc7/samples/bpf/xdp2skb_meta_kern.c#L41

Thus, AFAIK if the driver already added some metadata before your
XDP-prog, then this call[1] will just move ctx->data_meta some-more to
make room for *your* metadata (and driver metadata will be "after").
When using this metadata area, e.g.[2] then it will point to the
metadata you added.

[2] https://github.com/torvalds/linux/blob/v5.13-rc7/samples/bpf/xdp2skb_meta_kern.c#L78

Notice, this is also the reason, we (Bjørn, Magnus + I) suggested that
the btf_id (for AF_XDP use-case) should be placed as the "last" element
in the metadata struct, as it will be located at (ctx->data - 4 bytes).

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer