Skip to content

memory: lock heartbeat refreshes a lock file the process no longer owns #5418

Open

Description

@soarkit
Field Value
Labels (suggested) bug, memory, concurrency
Reproduced on commit a5587f2835dd4a2ba7f8fbcb6d8bdca6d5f4b8ba (default branch at the time of analysis)
Introduced by commit 4ddd6ce6d89ad3a9b5a4294fadc8bd2d993fc5e3 — "fix(memory): serialize stale-lock breaking behind a guard"

Summary

withFileLock() in src/memory/index.ts keeps its lock alive with a heartbeat interval that calls
fs.utimes(lockPath, ...) unconditionally. The release path a few lines below already knows that the
lock file may have changed hands — it re-reads the file and only unlinks it when the contents are still
its own token — but the heartbeat has no such check. If the lock was broken as stale and re-taken by
another waiter (or released and re-created), the original holder keeps refreshing the replacement
owner's lock file for as long as its own operation runs.

Affected code

src/memory/index.ts, lines 304–322 at a5587f2835dd:

  // Keep the lock fresh while the operation runs, so a mutation that outlasts
  // staleMs (a large graph, a slow disk) is not mistaken for an abandoned one.
  const heartbeat = setInterval(() => {
    const now = new Date();
    fs.utimes(lockPath, now, now).catch(ignoreError);   // <- no ownership check
  }, staleMs / 2);
  heartbeat.unref();

  try {
    return await operation();
  } finally {
    clearInterval(heartbeat);
    // Release only our own lock: if it was broken as stale while we held it,
    // the file may now belong to another process.
    const current = await fs.readFile(lockPath, "utf-8").catch(() => null);
    if (current === token) {                             // <- release DOES check
      await fs.unlink(lockPath).catch(ignoreError);
    }
  }

The comment on the release path states the exact invariant the heartbeat violates: after a stale break
"the file may now belong to another process".

Expected behavior

After the lock has been broken and re-taken, A's heartbeat must leave B's lock file alone. A must not
touch a lock file whose contents are not its token (the release path already enforces this).

Actual behavior

Holder A is inside withFileLock with staleMs = 4000 (heartbeat every 2000 ms) when the lock file's mtime
is backdated by 60 s. Holder B (staleMs = 59000; its 29.5 s heartbeat never fires inside the ~2.6 s
observation) judges the lock stale, breaks it and re-takes it with a new token — and A's own heartbeat still
fires inside that window:

{"mode":"bug","beatsOnReplacement":1,"beatsOnOther":0,"replacementTokenDiffers":true,
 "lockSurvivedARelease":true,"lockAfterBRelease":"gone"}
HEARTBEAT_REFRESHED_REPLACEMENT_LOCK
  • beatsOnReplacement: 1 — A's fs.utimes hit the lock file whose contents were B's token.
  • beatsOnOther: 0 — the observation is clean; the single hit is attributable to A.
  • replacementTokenDiffers: true — B's token is demonstrably different from A's.
  • lockSurvivedARelease: true — A's release correctly declined to delete B's lock.

Control: adding the ownership re-read to the heartbeat (the guard the release path already uses) makes the same
run report {"mode":"control","beatsOnReplacement":0,...} — CONTROL_NO_REPLACEMENT_TOUCH.

Why it matters

The heartbeat exists so a legitimately slow mutation is not mistaken for an abandoned one. Once the
lock has changed hands, the same heartbeat instead keeps a lock file fresh on behalf of a process that
may already be dead:

  • A's remaining work can extend B's staleMs deadline indefinitely. If B crashes while A is still
    running, a third process waiting on the lock will see a continuously refreshed mtime and will not
    break the abandoned lock; it waits until its own timeoutMs and then fails with
    Timed out after … ms waiting for the memory file lock.
  • The effect is the opposite of the intent: instead of making progress possible, the heartbeat delays
    recovery from a crashed peer by the length of the previous holder's operation.

This is a robustness/availability bug rather than a data-corruption bug: the token-guarded release
still prevents deleting someone else's lock.

Suggested fix

Revalidate ownership before refreshing, exactly as the release path does — and keep the check cheap:

const heartbeat = setInterval(() => {
  const now = new Date();
  void fs.readFile(lockPath, "utf-8")
    .then((current) => {
      if (current !== token) {           // lock changed hands: stop refreshing it
        clearInterval(heartbeat);
        return;
      }
      return fs.utimes(lockPath, now, now);
    })
    .catch(ignoreError);
}, staleMs / 2);

(Terminating the interval on token mismatch also stops the pointless reads. Holding the created
FileHandle open and using handle.utimes() is an alternative, but it changes unlink semantics on
Windows, where an open handle keeps the file from being removed — the token re-read keeps the current
behaviour.)

Activity

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