Re: [PATCH v2 bpf 1/5] net: ethtool: add xdp properties flag set

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

John Fastabend <john.fastabend@gmail.com> writes:

> Toke Høiland-Jørgensen wrote:
>> Jesper Dangaard Brouer <jbrouer@redhat.com> writes:
>> 
>> > On Mon, 7 Dec 2020 18:01:00 -0700
>> > David Ahern <dsahern@gmail.com> wrote:
>> >
>> >> On 12/7/20 1:52 PM, John Fastabend wrote:
>> >> >>
>> >> >> I think we need to keep XDP_TX action separate, because I think that
>> >> >> there are use-cases where the we want to disable XDP_TX due to end-user
>> >> >> policy or hardware limitations.  
>> >> > 
>> >> > How about we discover this at load time though. 
>> >
>> > Nitpick at XDP "attach" time. The general disconnect between BPF and
>> > XDP is that BPF can verify at "load" time (as kernel knows what it
>> > support) while XDP can have different support/features per driver, and
>> > cannot do this until attachment time. (See later issue with tail calls).
>> > (All other BPF-hooks don't have this issue)
>> >
>> >> > Meaning if the program
>> >> > doesn't use XDP_TX then the hardware can skip resource allocations for
>> >> > it. I think we could have verifier or extra pass discover the use of
>> >> > XDP_TX and then pass a bit down to driver to enable/disable TX caps.
>> >> >   
>> >> 
>> >> This was discussed in the context of virtio_net some months back - it is
>> >> hard to impossible to know a program will not return XDP_TX (e.g., value
>> >> comes from a map).
>> >
>> > It is hard, and sometimes not possible.  For maps the workaround is
>> > that BPF-programmer adds a bound check on values from the map. If not
>> > doing that the verifier have to assume all possible return codes are
>> > used by BPF-prog.
>> >
>> > The real nemesis is program tail calls, that can be added dynamically
>> > after the XDP program is attached.  It is at attachment time that
>> > changing the NIC resources is possible.  So, for program tail calls the
>> > verifier have to assume all possible return codes are used by BPF-prog.
>> 
>> We actually had someone working on a scheme for how to express this for
>> programs some months ago, but unfortunately that stalled out (Jesper
>> already knows this, but FYI to the rest of you). In any case, I view
>> this as a "next step". Just exposing the feature bits to userspace will
>> help users today, and as a side effect, this also makes drivers declare
>> what they support, which we can then incorporate into the core code to,
>> e.g., reject attachment of programs that won't work anyway. But let's
>> do this in increments and not make the perfect the enemy of the good
>> here.
>> 
>> > BPF now have function calls and function replace right(?)  How does
>> > this affect this detection of possible return codes?
>> 
>> It does have the same issue as tail calls, in that the return code of
>> the function being replaced can obviously change. However, the verifier
>> knows the target of a replace, so it can propagate any constraints put
>> upon the caller if we implement it that way.
>
> OK I'm convinced its not possible to tell at attach time if a program
> will use XDP_TX or not in general. And in fact for most real programs it
> likely will not be knowable. At least most programs I look at these days
> use either tail calls or function calls so seems like a dead end.
>
> Also above somewhere it was pointed out that XDP_REDIRECT would want
> the queues and it seems even more challenging to sort that out.

Yeah. Doesn't mean that all hope is lost for "reject stuff that doesn't
work". We could either do pessimistic return code detection (if we don't
know for sure assume all codes are used), or we could add metadata where
the program declares what it wants to do...

-Toke