镜像站点 · 本页由第三方 GitHub 只读镜像提供,非 GitHub 官方站点,不接受任何登录或凭据输入。前往 github.com
Skip to content

Threads: serialize shared Map and Set iterator cursor advancement #643

Closed

Description

@chrisbbreuer

Problem

Fresh required TSan CI reports three races in the shared Map iterator cursor witness: concurrent next() calls copy the iterator cursor's three-word CollectionIteratorState in Object.collectionIteratorState while peer threads update index, current, and traversal state in collectionCursorIterNext.

The semantic test happens to finish successfully, but these are engine-state races. A multiword cursor snapshot can be torn even when individual word reads/writes appear valid, violating the repository's no-GIL synchronization contract.

Required implementation

  • Give each published collection iterator cursor an explicit synchronization ruling.
  • Make state inspection and advancement one atomic logical transaction without serializing unrelated Map/Set instances or the underlying collection's independent readers.
  • Preserve iterator ordering, deletion/insertion semantics, exhausted-state idempotence, and GC tracing/relocation of the cursor's managed edges.
  • Do not suppress TSan, make the fields independently atomic, restore the context GIL, or weaken the shared-cursor witness.

Acceptance

  • parallel_js: Map iterator cursor advances atomically without GIL passes with zero TSan reports.
  • Focused tests cover concurrent next(), exhaustion, mutation during iteration, Map keys/values/entries, Set values/entries, and GC while a cursor is shared.
  • Normal, TSan, no-GIL, and Test262 Map/Set iterator coverage remain green.

Parent: #478
Related CI: tsan-parallel-js

Activity

  1. chrisbbreuer commented on Aug 20, 2026

    @chrisbbreuer
    MemberAuthor

    Fresh Linux suppression-free TSan still reproduces the cursor race on current main: tsan-parallel-js job 96534732523. The new report is at direct writes in collectionCursorIterNext (interpreter.zig lines 34513, 34523, and 34524), with 3 TSan warnings; the functional test itself returns OK and the remaining 61-case focused gate completes 54 pass / 8 skip / 0 fail / 0 leak. I am leaving implementation ownership with the user as requested.

  2. chrisbbreuer commented on Aug 20, 2026

    @chrisbbreuer
    MemberAuthor

    The #643 witness recurred unchanged in workflow 32423835718 for unrelated JSON commit 0052b2aa: tsan-parallel-js reported 2 races and TSan unit shard 1 reported 3, all in the Map iterator cursor path (value.Object.collectionIteratorState / interpreter.collectionCursorIterNext). The other 49 jobs succeeded and 1 skipped. No iterator fix was attempted here; this is linked CI evidence for the owner.

  3. chrisbbreuer commented on Aug 21, 2026

    @chrisbbreuer
    MemberAuthor

    Fresh descendant CI evidence from cancelled/superseded run 32488007100 reproduces this exact cursor race in both sanitizer gates:

    • tsan-unit (shard 3/4): context.test.parallel_js: Map iterator cursor advances atomically without GIL, two TSan warnings; both summaries point through value.Object.collectionIteratorState at src/value.zig:3090 and __tsan_memcpy.
    • tsan-parallel-js: the same witness, one TSan warning and the same stack boundary.

    Functional assertions still completed (415 passed / 32 skipped / 0 failed in the shard; 54 / 8 / 0 in focused parallel_js), but TSan correctly exited 66. I made no Map/Set iterator, stack, queue, Promise, microtask, Worker, or job changes; this is evidence only for the user-owned #643 fix.

  4. chrisbbreuer commented on Aug 22, 2026

    @chrisbbreuer
    MemberAuthor

    Landed in 04d46c59 (`fix(collections): make shared cursor dispatch one locked transaction (#643)`).

    Synchronization ruling (recorded on the advance helper): a published collection-iterator cursor is guarded by its owning iterator object's `property_lock`; inspection and advancement are one atomic logical transaction inside that critical section; concurrent GC trace reads `source` through the backing lock and relocation rewrites it only at the world-stopped commit.

    The fix removes the unlocked rare-state probe that Linux TSan flagged: dispatch first runs a content-free atomic tag check (the rare tag is immutable once published, so non-collection iterators — the hot array path — never acquire a lock), then probes and advances under the same property lock that serializes every cursor-field store. No unrelated Map/Set instance or the collection's own readers are serialized; per-entry exactly-once semantics, exhaustion idempotence, insertion-order/

  5. chrisbbreuer commented on Aug 22, 2026

    @chrisbbreuer
    MemberAuthor

    Linux CI confirmation for the landing commit 04d46c59: run 32592838252 passes `tsan-unit` (all four shards) and `tsan-parallel-js` with zero cursor-race diagnostics — the same gates that previously reported the `collectionIteratorState`/`__tsan_memcpy` race on every run since 32356955396. The run's only failure is the pre-existing `semantics/oom-one-thread.js`, now tracked in #722.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions