GCC BPF support

[theihor]

Successor of #144

[theihor]

@chantra this one makes GCC BPF specific changes. It should be reviewed/merged after #166

[chantra]

LGTM, I am just not sure if the BPF-GCC is complementary to the current bpf build, e.g after this build, we have 2 distinctive bpf objects, 1 for the llvm build, and one for the gcc one?

[chantra]

is this related to this diff?

[theihor]

No, I'll clean it up, thanks.

[chantra]

isn't this going to build bpf-gcc for all combination of x86 builds?

shouldn't we instead define a "bpf-toolchain" that defaults to llvm (current behaviour), and can potentially be changd to gcc and use this build? Or is this handled transparently by the "test_progs-bpf_gcc" target?

[theihor]

Right. This is going to build GCC for all x86 builds, but since it is cached, it's not an issue. When cached, it gets pulled in about 6 seconds: https://github.com/libbpf/ci/actions/runs/12777030966/job/35616953725

I thought about making a separate job (a dependency of kernel-test, when BUILD_BPF_GCC is set), since GCC doesn't really depend on anything within our workflows. But I concluded it won't make a difference, because the tests still depend on kernel to be built.

[theihor]

@chantra please see an explanation re these two questions below

I am just not sure if the BPF-GCC is complementary to the current bpf build, e.g after this build, we have 2 distinctive bpf objects, 1 for the llvm build, and one for the gcc one?

shouldn't we instead define a "bpf-toolchain" that defaults to llvm (current behaviour), and can potentially be changd to gcc and use this build? Or is this handled transparently by the "test_progs-bpf_gcc" target?

Yes, it's handled transparently in selftests/bpf/Makefile.

The "bpf-toolchain" doesn't really change when we set BPF_GCC. This is why I decided to not include the changes from #144 that add "bpf-toolchain" action arguments.

For example, let's say the CI toolchain-full is set to "llvm-17". So, the kernel is built with LLVM=-17 (see here), and the test_progs binaries are built with llvm-17 (not the default).

Now, on x86 BUILD_BPF_GCC is true, so BPF_GCC will be set to where freshly built gcc binaries are located. The test_progs-bpf_gcc binary is still built with clang 17 in this case (make target just uses $(CC)). What changes is:

1. test_progs-bpf_gcc is built (by default it's not)
2. The BPF objects for test_progs-bpf_gcc are built with provided GCC compiler. Later they are automagically included in the test runner binary as static arrays, IIRC.

So from the perspective of the CI, BPF_GCC is an additional test runner, built conditionally.

Now to the question wether we want to run this on all variants of the toolchain, I think yes. This way we are testing BPF emitted by GCC against kernels built by different toolchains.