CVE-2026-72314

This CVE exposes a systemic tooling failure rather than a simple developer mistake. The Linux kernel's errno system allowed an API mismatch to pass silently on most architectures, making the development environment itself complicit in the bug. The core issue isn't that a developer typed EDEADLOCK instead of EDEADLK. It's that the kernel's errno definitions across architectures created a situation where this typo was invisible. The API contract is explicit: ww_mutex_lock() returns EDEADLK, and every other caller in the same file respects that contract. One function doesn't — and on x86, ARM, and most other platforms, the developer never knew. Their code compiled. Their tests passed. The bug only surfaced because cross-architecture testing is a luxury, not a norm. The partial-lock state this creates is particularly insidious. The function believes it has failed and returns an error, but it holds one regulator lock. Callers expecting a clean failure or success now face a corrupted state: one resource locked, one leaked, with no indication which regulator remains held. This isn't just a deadlock risk — it's a resource leak masquerading as proper error handling. What makes this notable is the exposure window was effectively zero outside MIPS — not because the bug didn't exist, but because the errno divergence on MIPS was the only thing that made the symptom visible enough to investigate. That's a dangerous qualifier: it means similar silent failures may be lurking in code paths nobody tests on non-dominant architectures. For defenders: treat architecture-specific errno divergences as audit signals, not inconveniences. If you touch the regulator subsystem, prioritize MIPS in your CI matrix. Verify errno constants against their API contracts at review time — not as a manual exercise, but as a structural expectation. The fix for this function is trivial; the question is why the ecosystem permits these inconsistencies to persist in critical paths.

Reviewed through automated stages and approved by a human before publication.