On Thu, 2025-07-17 at 17:10 -0700, Andrii Nakryiko wrote:
> Initial __arean global variable support implementation in libbpf
> contains a bug: it remembers struct bpf_map pointer for arena, which is
> used later on to process relocations. Recording this pointer is
> problematic because map pointers are not stable during ELF relocation
> collection phase, as an array of struct bpf_map's can be reallocated,
> invalidating all the pointers. Libbpf is dealing with similar issues by
> using a stable internal map index, though for BPF arena map specifically
> this approach wasn't used due to an oversight.
>
> The resulting behavior is non-deterministic issue which depends on exact
> layout of ELF object file, number of actual maps, etc. We didn't hit
> this until very recently, when this bug started triggering crash in BPF
> CI when validating one of sched-ext BPF programs.
>
> The fix is rather straightforward: we just follow an established pattern
> of remembering map index (just like obj->kconfig_map_idx, for example)
> instead of `struct bpf_map *`, and resolving index to a pointer at the
> point where map information is necessary.
>
> While at it also add debug-level message for arena-related relocation
> resolution information, which we already have for all other kinds of
> maps.
>
> Fixes: 2e7ba4f8fd1f ("libbpf: Recognize __arena global variables.")
> Signed-off-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
> ---
Tested using scx jobs failing on CI:
https://github.com/kernel-patches/bpf/actions/runs/16354530864/artifacts/3558110958
The change lgtm.
Tested-by: Eduard Zingerman <eddyz87@xxxxxxxxx>
[...]
