bpf: always inline .bss in absence of a .data section

[ti-mo]

Due to an early vars == nil check in inlineGlobalData, ELFs without a .data section would not get their .bss references rewritten. This would cause verifier errors on pre-5.x kernels like 'unrecognized bpf_ld_imm64 insn', since those don't support pseudo-register load insns yet.

Delay the check and add a regression test.

bpf: always inline .bss in absence of a .data section

[ti-mo]

/test

[qmonnet]

Thanks!

[gentoo-root]

OpCode -> Dst

[qmonnet]

The change breaks some of the unit tests, it introduces patterns such as:

             	150: (18) r2 = 0x0
            	152: (69) r2 = *(u16 *)(r2 +0)
            	R2 invalid mem access 'inv'

Observable here for this test PR. Unit tests were skipped on your PR because it doesn't touch bpf/ files.

[dylandreimerink]

It looks like this change breaks global data which we use in the tests. For example in bpf/tests/xdp_nodeport_lb4_nat_lb.c we have a variable static __be16 nat_source_port;. It is zero initialized so it ends up in .bss, this change inlines it which should not happen. Giving the variables an initial value does not work, relocation entry moves to .data and then it also gets inlined.

I guess the tests always worked because we were hitting the edge case this PR is removing of not having any global data in .data so our global .bss variables would behave as expected.

I don't really know how to fix this one since we use global data for mocking and to confirm that mocked functions were invoked. Ideally we would mark our oddball Cilium variables in such a way that we can selectively inline them and not dynamic global variables.

[ti-mo]

This is a bad approach. We should instead try to avoid anything ending up in .bss by annotating all static variable declarations with .data or another snowflake section like .inline now that iproute2 is no longer used for loading.

We can then reject references to any other sections on pseudo-load instructions because variables likely ended up there by mistake. This would include rejecting references to .bss, which is a lot clearer than the verifier returning unrecognized bpf_ld_imm64 insn.