Re: [PATCH bpf-next v2] selftests/bpf: Fix linked_externs with bpf-gcc

[Ihor Solodrai]

On 2026-09-28 4:14 p.m., Vineet Gupta wrote:
> test_progs-bpf_gcc fails to build:
>
>    prog_tests/linked_externs.c:17:23: error: 'struct linked_arena' has no
>        member named 'arena'
>
> Several problems, all from bpf-gcc not supporting address_space(1).
>
> First, the arena variables are declared with __arena, but that macro is
> for pointers -- it marks the pointee address space. The one that carries
> placement is __arena_global:
>
>    #if defined(__BPF_FEATURE_ADDR_SPACE_CAST) && !defined(BPF_ARENA_FORCE_ASM)
>    #define __arena        __attribute__((address_space(1))) __attribute__((btf_type_tag("arena")))
>    #define __arena_global __attribute__((address_space(1)))
>    #else
>    #define __arena        __attribute__((btf_type_tag("arena")))
>    #define __arena_global SEC(".addr_space.1")
>    #endif
>
> __BPF_FEATURE_ADDR_SPACE_CAST is a clang predefine. Under clang the two
> are interchangeable here, since address_space(1) both places the
> variable and is what btf_type_tag would have described. Under bpf-gcc
> __arena is only a BTF type tag, so the definitions land in .data, the
> arena map gets no initial value, and is_skel_data() in bpftool does not
> emit the typed arena member -- hence the build error. Note this is
> distinct from skel->maps.arena, which comes from SEC(".maps") and is
> always present.
>
> The extern declarations need __arena_global too. Without a section, gcc
> does not place an extern in any DATASEC, so it stays a free-floating
> extern VAR rather than an entry of .addr_space.1. test_relink() only
> walks that DATASEC:
>
>      id = btf__find_by_name_kind(btf, sec_name, BTF_KIND_DATASEC);
>      ...
>          ASSERT_NEQ(btf_var(t)->linkage, BTF_VAR_GLOBAL_EXTERN, "var_resolved");
>
> so relink_arena would report OK under bpf-gcc whether or not the linker
> resolved anything -- it would not be testing the allocated-section path
> this test exists for.
>
> Second, placement alone is not enough to run the test. Dereferencing an
> arena global needs an addr_space_cast, which only clang emits:
>
>    clang:                                  bpf-gcc:
>      r1 = 0x0 ll                             r1 = 0x0 ll
>      r1 = addr_space_cast(r1, 0x0, 0x1)      r2 = *(u64 *)(r1 + 0x0)
>      r1 = *(u64 *)(r1 + 0x0)
>
> so the verifier sees a scalar and rejects the program:
>
>    4: (79) r2 = *(u64 *)(r1 +0)
>    R1 invalid mem access 'scalar'
>
> Guard the program bodies and skip the skeleton subtest, as
> arena_atomics.c already does. The variables stay unconditional so the
> arena is still populated and the skeleton still has its member. The
> fallback bodies reference the externs without dereferencing them:
> 'used' is ignored on an extern, so an unreferenced one is dropped and
> .addr_space.1 loses it again.
>
> With this, bpf-gcc reports:
>
>    #217/1   linked_externs/skel_arena:SKIP
>    #217/2   linked_externs/relink_arena:OK
>    #217/3   linked_externs/relink_maps:OK
>
> with b_val present in .addr_space.1 as linkage=extern, same as clang, so
> relink_arena covers the same path in both. clang still passes all three.
>
> Fixes: 34354db1b148 ("selftests/bpf: Add linked_externs test for externs in allocated sections")
> Signed-off-by: Vineet Gupta <vineet.gupta@xxxxxxxxx>

Tested-by: Ihor Solodrai <ihor.solodrai@xxxxxxxxx>

This fixes the current GCC BPF selftests build failure on CI:
https://github.com/kernel-patches/vmtest/actions/runs/36498905340/job/109187331070

nit: The commit message is kinda too long. If you'd like it to be read
by a human, try making it shorter :)

Thanks!

> [...]