Re: [PATCH bpf-next 5/5] libbpf: add selftests for TC-BPF API

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Sun, Mar 28, 2021 at 07:38:42PM -0700, Andrii Nakryiko wrote:
> 
> See above. I don't know which hassle is libbpf for users today. You
> were implying code size used for functionality users might not use
> (e.g., linker). Libbpf is a very small library, <300KB. There are
> users building tools for constrained embedded systems that use libbpf.
> There are/were various problems mentioned, but the size of libbpf
> wasn't yet one of them. We should certainly watch the code bloat, but
> we are not yet at the point where library is too big for users to be
> turned off. 

It's true that today sizeof(libbpf + libelf + libz) ~ 500k is not a concern.
I'm worried what it will get to if we don't start splitting things up.

Why split libxdp into its own lib?
If tc attach is going to part of libbpf all things xdp should be
part of libbpf as well.

But af_xdp folks are probably annoyed that they need to add -lelf an -lz
though they're not using them. Just a tech debt that eventually need to be paid.

> > I would take this opportunity to split libbpf into maintainable pieces:
> > - libsysbpf - sys_bpf wrappers (pretty much tools/lib/bpf/bpf.c)
> > - libbpfutil - hash, strset
> 
> strset and hash are internal data structures, I never intended to
> expose them through public APIs. I haven't investigated, but if we
> have a separate shared library (libbpfutil), I imagine we won't be
> able to hide those APIs, right?

In the other thread you've proposed to copy paste hash implemenation
into pahole. That's not ideal. If we had libbpfutil other projects
could have used that without copy-paste.

> But again, let's just reflect for a second that we are talking about
> the library that takes less than 300KB total. 

that's today. Plus mandatory libelf and libz.
I would like to have libsysbpf that doesn't depend on libelf/libz
for folks that don't need it.
Also I'd like to see symbolizer to be included in "libbpf package".
Currently it's the main component that libbcc offers, but libbpf doesn't.
Say we don't split libbpf. Then symbolizer will bring some dwarf library
(say libdwarves ~ 1Mbyte) and libiberty ~ 500k (for c++ demangle).
Now we're looking at multi megabyte libbpf package.
I think the users would benefit from smaller building blocks.
Splitting into 10 mini libs is overkill, of course,
but some split is necessary.
I agree that moving static linking into separate lib won't really
affect .text size. The point is not to reduce text, but to establish
a framework where such things are possible. Then symbolizer and
anything fancier that would depend on other libs can be part
of "libbpf package". I mean single rpm that contains all libbpf libs.
Basic libsysbpf wouldn't need libelf/z.
libbpfsymbolizer would need libdwarf, etc.
Maybe some libbpfnet would depend on libnl or what not.