Fix lock-free queue bugs, add GitHub Actions CI, and add ABA stress test

[Copilot]

Several correctness bugs in the hazard pointer implementation and test harness, plus missing CI since Travis CI is no longer free for OSS.

Bug Fixes

• Wrong calloc element sizes in lfq_init: HP (array of pointers) and tid_map (array of int) were both allocated with sizeof(struct lfq_node) — 3× overallocation that masked out-of-bounds writes:

  // Before (wrong)
  ctx->HP      = calloc(max_consume_thread, sizeof(struct lfq_node));
  ctx->tid_map = calloc(max_consume_thread, sizeof(struct lfq_node));
  
  // After (correct)
  ctx->HP      = calloc(max_consume_thread, sizeof(struct lfq_node *));
  ctx->tid_map = calloc(max_consume_thread, sizeof(int));

• Thread ID off-by-one in test consumer: ATOMIC_ADD(&cn_t, 1) returns 1-based values (1…MAX_CONSUMER), causing the last consumer to write HP[MAX_CONSUMER] — one past the end. Fixed: ATOMIC_ADD(&cn_t, 1) - 1.

• smb()/lmb() undefined on non-x86 platforms: Added #else fallback to mb() so the code compiles on ARM/RISC-V.

• Memory leak in lfq_init failure path: If the second calloc failed, tmpnode was leaked. Added free(tmpnode) before the early return.

CI

• Replaced .travis.yml with a GitHub Actions workflow (.github/workflows/ci.yml) using the official C/C++ CI starter template, running make and make test on ubuntu-latest with contents: read permissions.

ABA Stress Test

Added test_aba.c: a dedicated test designed to maximise the conditions under which the ABA problem would manifest in a CAS-based lock-free queue.

The ABA problem: a CAS on a pointer succeeds spuriously when a node is dequeued, freed, and a new node is allocated at the same address before a stale CAS completes — causing the queue to operate on a logically different node without detecting the change.

How this test maximises ABA probability compared to the existing tests:

Property	Existing test	ABA stress test
Thread roles	Separate producers + consumers	Every thread is both producer AND consumer
Queue depth	Grows large (500k items per producer)	Near-zero (1 item per thread in flight)
Yielding	THREAD_YIELD() when queue empty	No yielding — tight spin after enqueue
Memory pattern	Many different addresses	Rapid alloc/free → allocator reuses same addresses
Corruption detection	Count mismatch only	LIVE_MAGIC/DEAD_MAGIC sentinel per node

Each node carries a LIVE_MAGIC value on enqueue, poisoned to DEAD_MAGIC before free(). Any ABA-induced use-after-free or stale-pointer read surfaces immediately as a magic mismatch. The queue passes the test (8,000,000 ops across 16 threads, zero corruption), confirming the hazard pointer scheme correctly prevents ABA.

The Makefile is updated to build bin/test_aba and include it in make test.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[darkautism]

CI改成最新的github提供的workflow @copilot

[Copilot]

CI改成最新的github提供的workflow @copilot

Updated in ad326ff — the workflow now follows GitHub's official C/C++ CI starter template (name: C/C++ CI, branch filters on HP, job named build, steps using make / make test).

[darkautism]

設計一款更在multithread更容易發生ABBA的測試方法，來測試該LFQUEUE有無此弱點 @copilot

[Copilot]

設計一款更在multithread更容易發生ABBA的測試方法，來測試該LFQUEUE有無此弱點 @copilot

Added test_aba.c in 32a815f. It's wired into make test via the Makefile.

The test maximises ABA exposure compared to the existing tests by:

• All threads simultaneously produce AND consume (no separation) — tight enqueue → spin-dequeue loops with no yield, keeping the queue depth near 0–1 items
• Rapid malloc/free of uniform small objects — causes the allocator to hand back the same addresses frequently, creating the classic A→B→A pointer value cycle on head
• 16 competing threads (configurable via ABA_THREADS) all racing on the head CAS

Each node carries a LIVE_MAGIC value on enqueue, which is poisoned to DEAD_MAGIC before free(). Any ABA-induced use-after-free or stale-pointer read surfaces immediately as a magic mismatch and is reported as a failure.

The queue passes the test (8 000 000 ops, no corruption detected), confirming the hazard pointer scheme correctly prevents ABA.