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

Promises: bound host rejection queue draining #874

Closed

Description

@chrisbbreuer

Parent: #493
Related: #474, #480, #482, #495, #884, #885, #888

Problem

The realm-local unhandled- and handled-rejection queues are growable arrays, and each host checkpoint removes index zero with orderedRemove(0). The private ABI drains those functions in a loop, so N JavaScript-created rejection notifications copy Θ(N²) Promise pointers while repeatedly holding realm_lock. Rejection fanout is script-controlled, making this an avoidable latency and contention surface.

The same growable-array append is the fallible publication boundary tracked by #884. A cursor would fix front shifting but retain that boundary; the final design should solve both issues without adding a second queue representation.

Required implementation

  • Replace array front-removal with one allocation-free realm-owned FIFO: Context head/tail roots plus a single intrusive successor per queued Promise.
  • Keep append and dequeue O(1), clear every removed successor immediately, and reset head/tail/length exactly at exhaustion.
  • Enforce single membership: a Promise may enter the handled queue only after leaving the unhandled queue, and may never be linked twice.
  • Trace every pending intrusive node under realm_lock; relocate both Context endpoints and every Promise successor while mutators are stopped.
  • Preserve early-handled skipping, one-time notification state, and reentrant ordering: callbacks may enqueue more notifications, which follow the pending suffix and drain in the same checkpoint.
  • Keep Promise in its existing 256-byte GC allocation class.
  • Do not reverse events, use unordered removal, cap valid fanout, defer required notifications, weaken locking, add an emergency allocation pool, or recognize particular workloads.

Acceptance

  • A deterministic 4,096-entry witness proves allocation-free O(1) FIFO append/pop, length-bounded iteration, cleared links, reentrant tail insertion, exact exhaustion reset, and reuse.
  • Unhandled and later-handled notifications retain FIFO order, early-handled entries are skipped, and reentrant enqueues follow pending entries.
  • Precise tracing and moving relocation cover all pending nodes and rewrite both endpoints/successors without retaining consumed nodes.
  • Promise allocation density, empty/repeated behavior, teardown, and publication invariants remain exact.
  • Focused Promise/private-ABI tests pass in Debug, ReleaseSafe, and suppression-free TSan; relevant threadfuzz and the full unit suite have zero failures/leaks.
  • Exact affected Test262 Promise coverage has zero regressions; documentation states the allocation-free FIFO contract without unsupported throughput claims.

Activity

  1. chrisbbreuer commented on Sep 3, 2026

    @chrisbbreuer
    MemberAuthor

    Implementation evidence so far:

    • The 4,096-entry queue witness passes and proves O(1) cursor advancement, unchanged backing storage during partial drain, FIFO reentrant append, exact exhaustion reset, capacity reuse, empty behavior, and append OOM invariants.
    • The private rejection-dispatch witness passes a 64-entry unhandled FIFO plus a reentrant 65th rejection, followed by 64 ordered handled notifications.
    • Promise/rejection focused tests pass in Debug and ReleaseSafe; the rejection filter passes under suppression-free TSan.
    • Moving-GC relocation proves consumed prefixes are untouched while pending suffixes are rewritten; tracing uses the same pending suffix.
    • The pinned Home private ABI passes 392/392 in Debug, ReleaseSafe, and TSan after an audit of 538 symbols (447 implemented, zero pending or unclassified).
    • Relevant default threadfuzz passes 400/400. Documentation builds 50 pages and validates 328 links; the manifest change is limited to the updated Home integration page and tree hash.

    The full lifecycle threadfuzz profile is currently running because it exposed the independent test-harness race tracked in #875. The full unit suite and exact Promise Test262 coverage remain intentionally unchecked until completed.

  2. chrisbbreuer commented on Sep 22, 2026

    @chrisbbreuer
    MemberAuthor

    Closing as implemented and accepted. The final allocation-free intrusive FIFO is part of 4a1e5d82; its rendered Home integration manifest was reconciled by 09279120. No user-owned #874 stash was read, applied, or modified for this closure.

    Final evidence:

    • The 4,096-entry FIFO witness passes in Debug, ReleaseSafe, and TSan. It covers O(1) append/pop, length-bounded traversal, cleared links, reentrant tail insertion after the pending suffix, exact empty reset, and reuse.
    • The tracker witness proves unhandled publication, repeated settlement, early handling, and later-handled publication perform zero allocator calls while flags and queue membership commit exactly.
    • Single membership is enforced by the overlaid rejection link and atomic linked bit; dequeue clears the successor before the Promise may enter the handled queue.
    • Precise GC traces both endpoints and every length-bounded node under realm_lock; relocation rewrites endpoints and linked successors. Teardown unlinks both queues before heap destruction, and consumed nodes are not retained.
    • Promise remains in the 256-byte production GC allocation class. The private rejection dispatch preserves exact Promise/reason identities and reentrant event order.
    • Fresh standalone ReleaseSafe rejection coverage: 13/13. Focused Debug/TSan queue and tracker witnesses pass, and Home private ABI links/runs 392/392 shims.
    • Linux integration run 34287554681 passed full Debug/TSan units, all Debug/ReleaseSafe/TSan no-GIL corpus shards, and every configured threadfuzz/TSan-fuzz profile. Its sole unrelated failure was the Errors: Proxy and detached-buffer messages are V8-derived, not JavaScriptCore's #903 detached-buffer reference expectation.
    • Fresh exact-parent Promise Test262 from the implementation boundary: 677/677 -> 677/677, 0 flips and 0 regressions.
    • docs/HOME_INTEGRATION.md states the allocation-free intrusive FIFO contract without publishing a throughput claim; the pinned BunPress build and all 328 links passed.

    The quadratic front removal and its fallible queue-publication boundary are both gone; no cursor compatibility path, event cap, emergency pool, or second queue representation remains.

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