Just want to send out a quick reply after looking at this.

On Wed, Jun 03, 2026 at 01:25:15AM +0800, Zhenzhong Wu wrote:

Hi Shung-Hsi,

...

I ran the suggested checks with the same reproducer, where BAD means the
program ran and observed the unexpected error, and GOOD means no error was
observed:

- latest 6.6.y, v6.6.142 (924b4a879cbb): BAD
- bpf-next at b93c55b4932d: GOOD
- bpf-next with the d028f87517d6 JNE refinement reverted: still GOOD

So the issue still reproduces on the latest 6.6.y, but d028f87517d6 alone
does not explain why bpf-next passes. I'll do more narrowing and update the
candidate backport set accordingly.

...

I think it possibly comes down to commit 4bf79f9be434e ("bpf: Track
equal scalars history on per-instruction level") added in v6.12. Without
that, the precise mark wasn't propogated (for scalars with the same ID),
and that likely made the state comparison (invalidly) go through.

Shung-Hsi