Re: [PATCH bpf-next 1/5] bpf: Handle 8-byte values in DEVMAP and DEVMAP_HASH

[David Ahern <dsahern@gmail.com>]

On 5/27/20 8:57 AM, Toke Høiland-Jørgensen wrote:
> 
> Either way you're creating a contract where the kernel says "first four
> bytes is the ifindex, second four bytes is the fd/id". BTF just makes
> that explicit, and allows userspace to declare that it agrees this is
> what the fields should mean. And gives us more flexibility when
> extending the API later than just adding stuff at the end and looking at
> the size...
> 
>> You need to know precisely which 4 bytes is the program fd that needs
>> to be looked up, and that AFAIK is beyond the scope of BTF.
> 
> Exactly - BTF is a way for userspace to explicitly tell the kernel "I am
> going to put the fd into these four bytes of the value field", instead
> of the kernel implicitly assuming it's always bytes 5-8.
> 

Really, I should define that struct in uapi/linux/bpf.h. The ifindex
field has special meaning: the kernel needs to convert it to a
net_device. The prog_fd field has special meaning: it should map to a
bpf program.

This use case is not in BTF's scope. But, prove me wrong. Ideas are
cheap; code is hard. Show me the code that implements your idea.