Re: [PATCH bpf-next 00/16] Add libbpf full support for BPF-to-BPF calls

On Fri, Aug 21, 2020 at 4:00 PM Alexei Starovoitov
[off-list ref] wrote:

On Thu, Aug 20, 2020 at 04:12:34PM -0700, Andrii Nakryiko wrote:

quoted

Currently, libbpf supports a limited form of BPF-to-BPF subprogram calls. The
restriction is that entry-point BPF program should use *all* of defined
sub-programs in BPF .o file. If any of the subprograms is not used, such
entry-point BPF program will be rejected by verifier as containing unreachable
dead code. This is not a big limitation for cases with single entry-point BPF
programs, but is quite a havy restriction for multi-programs that use only
partially overlapping set of subprograms.

This patch sets removes all such restrictions and adds complete support for
using BPF sub-program calls on BPF side. This is achieved through libbpf
tracking subprograms individually and detecting which subprograms are used by
any given entry-point BPF program, and subsequently only appending and
relocating code for just those used subprograms.

In addition, libbpf now also supports multiple entry-point BPF programs within
the same ELF section. This allows to structure code so that there are few
variants of BPF programs of the same type and attaching to the same target
(e.g., for tracepoints and kprobes) without the need to worry about ELF
section name clashes.

This patch set opens way for more wider adoption of BPF subprogram calls,
especially for real-world production use-cases with complicated net of
subprograms. This will allow to further scale BPF verification process through
good use of global functions, which can be verified independently. This is
also important prerequisite for static linking which allows static BPF
libraries to not worry about naming clashes for section names, as well as use
static non-inlined functions (subprograms) without worries of verifier
rejecting program due to dead code.

Patch set is structured as follows:
- patches 1-5 contain various smaller improvements to logging and selftests;
- patched 6-11 contain all the libbpf changes necessary to support multi-prog
  sections and bpf2bpf subcalls;
- patch 12 adds dedicated selftests validating all combinations of possible
  sub-calls (within and across sections, static vs global functions);
- patch 13 deprecated bpf_program__title() in favor of
  bpf_program__section_name(). The intent was to also deprecate
  bpf_object__find_program_by_title() as it's now non-sensical with multiple
  programs per section. But there were too many selftests uses of this and
  I didn't want to delay this patches further and make it even bigger, so left
  it for a follow up cleanup;
- patches 14-15 remove uses for title-related APIs from bpftool and
  bpf_program__title() use from selftests;
- patch 16 is converting fexit_bpf2bpf to have explicit subtest (it does
  contain 4 subtests, which are not handled as sub-tests).

I've applied the first 5 patches. Cleanup of 'elf:' logs is nice.
Thanks for doing it.
The rest of the patches look fine as well, but minimalistic selftest is
a bit concerning for such major update to libbpf.
Please consider expanding the tests.

That test is not that minimalistic, actually. It tests all
combinations of bpf2bpf calls (global/static * same/other section),
plus with only subsets of functions used by entry-point BPF programs.
Similarly, fentry_bpf2bpf tests have also pretty complicated patterns,
as well as test_pkt_access.c.

May be cloudflare's test_cls_redirect.c can be adopted for this purpose?
test_xdp_noinline.c can also be extended by doing two copies of
balancer_ingress(). One to process ipv4 another ipv6.

I'll take a look at those, if they are using sub-program calls (not
__always_inline), they are already testing this. I'll see if I can
un-inline some more functions, though.

Then it will make libbpf to do plenty of intersting call adjustments
and function munipulations for three programs in "xdp-test" section
that use different sets of sub-programs.
test_l4lb_noinline.c can be another candidate.
The selftest that is part of this set is nice for targeted debugging, but would
be great to see production bpf prog adopting this exciting libbpf feature.

Sure, I'll also see if strobemeta examples can be modified minimally
to allow non-inlined functions.