René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"Non-blocking `_PyMutex_LockTimed` spins and may fail unnecessarily in no-GIL builds","articleBody":"# Bug report\n\n### Bug description:\n\n### Non-blocking `_PyMutex_LockTimed` spins and may fail unnecessarily (no-GIL build)\n\n**Branch:** `main` (commit `2793b68f758c10fb63b264787f10d46a71fc8086`)  \n**Build:** configured with `--disable-gil` (`MAX_SPIN_COUNT \u003e 0`)  \n**OS:** all\n\n---\n\n## Summary\n\n`_PyMutex_LockTimed()` is supposed to return immediately when called with\n`timeout == 0` (non-blocking).  In a no-GIL build it:\n\n1. **Spins for up to `MAX_SPIN_COUNT` yields** before returning;  \n2. Can still return **`PY_LOCK_FAILURE` even though the lock was released** during that spin.\n\nTimed/blocking calls (`timeout \u003e 0`) also waste CPU because the spin loop\nnever reloads the lock word, so they can’t notice an unlock until after the\nmaximum spin count.\n\nThe bug is invisible in GIL builds because `MAX_SPIN_COUNT` is 0.\n\n---\n\n## Root cause\n\n```c\n/* Python/lock.c — excerpts */\n\nif ((v \u0026 _Py_LOCKED) == 0) {\n    if (_Py_atomic_compare_exchange_uint8(\u0026m-\u003e_bits, \u0026v, v | _Py_LOCKED))\n        return PY_LOCK_ACQUIRED;\n}\nelse if (timeout == 0) {            // executes only if FIRST load saw LOCKED\n    return PY_LOCK_FAILURE;         // non-blocking, OK\n}\n\n/* … later … */\nif (!(v \u0026 _Py_HAS_PARKED) \u0026\u0026 spin_count \u003c MAX_SPIN_COUNT) {\n    _Py_yield();\n    spin_count++;\n    continue;                       // BUG: never refreshes v\n}\n```\n\n* If our fast CAS loses a race, the `else if` guard is skipped, so a\n  *non-blocking* call drops into the spin loop.  \n* Inside that loop `v` is never reloaded, so the thread can’t see that the\n  lock became free.  After `MAX_SPIN_COUNT` iterations it falls through to\n  the same `timeout == 0` guard and fails spuriously.\n\n---\n\n## Proposed fix\n\n```c\n/* 1 — reload v each spin */\nif (!(v \u0026 _Py_HAS_PARKED) \u0026\u0026 spin_count \u003c MAX_SPIN_COUNT) {\n    _Py_yield();\n    spin_count++;\n    v = _Py_atomic_load_uint8_relaxed(\u0026m-\u003e_bits);   // ← added\n    continue;\n}\n\n/* 2 — early-out for timeout == 0 */\nif ((v \u0026 _Py_LOCKED) == 0) {\n    if (_Py_atomic_compare_exchange_uint8(\u0026m-\u003e_bits, \u0026v, v | _Py_LOCKED))\n        return PY_LOCK_ACQUIRED;\n}\nif (timeout == 0) {                               // ← moved outside else\n    return PY_LOCK_FAILURE;\n}\n```\n\n### Result\n\n* **Non-blocking calls** now return immediately: success if the CAS wins,\n  failure if it loses – no spinning, no parking.\n* **Timed/blocking calls** still spin for fairness, but they now reload the\n  lock word each iteration, so they acquire promptly once the lock is free.\n\n\n\n### CPython versions tested on:\n\nCPython main branch\n\n### Operating systems tested on:\n\nmacOS\n\n\u003c!-- gh-linked-prs --\u003e\n### Linked PRs\n* gh-135872\n* gh-135946\n* gh-135947\n* gh-146064\n\u003c!-- /gh-linked-prs --\u003e\n","author":{"url":"https://github.com/jtibbertsma","@type":"Person","name":"jtibbertsma"},"datePublished":"2025-06-24T00:39:45.000Z","interactionStatistic":{"@type":"InteractionCounter","interactionType":"https://schema.org/CommentAction","userInteractionCount":9},"url":"https://github.com/135871/cpython/issues/135871"}