deadlock probe: re-check pendingHostCalls before the verdict - #77
Merged
Conversation
The probe's precondition (pendingHostCalls empty, no resuming thread) can expire during the probe's own macrotask turn WITHOUT the awaiting set changing: the same activation resumes off an engine continuation chunk (jspi pin (j) — a sync-completing Suspending import defers its continuation), runs, and re-parks through the A1 arm, which registers a fresh pendingHostCalls entry. The membership re-check sees the same single activation, readyCandidates is empty (the new park is genuinely waiting on an external event), and the verdict fires against a live workload — with hostCalls=1 at trap time. Exposed by wasi-shims' A5 poll: its sync fast path lets the guest run stretches of ready-polls with no registered park, so the probe can sample the store between a settled A1 promise (which self-deletes from pendingHostCalls) and the next genuine park. The always-Promise poll shape A5 replaced never left that window open, which is why the same workload (polymorph-iroh's upstream-iroh spike: tokio parked on timer + synthetic-socket pollables, woken by detached host bridges) was green before A5 and trapped after: [drive #3] driveAsync branch=deadlock-probe ready=0 waiting=0{} awaiting=1 hostCalls=0 claim=false [drive #4] driveAsync branch=deadlock-probe:progressed=false ready=0 waiting=1{SuspensionPoint[not-ready]} awaiting=1 hostCalls=1 claim=false -> trap (old code): 'every suspended activation is waiting on a suspension only this scheduler could resume' Re-checking the full precondition turns the sampled race into a re-probe, exactly like a membership change. A genuine deadlock still traps: its re-check finds hostCalls=0 again (nothing external exists to register one), pinned by tests/jspi/deadlock_test.ts. Gates: test-runtime 363/0 (deadlock pins green); conformance 1254 passed / 0 failed / 95 xfail; the exposing workload (iroh spike: relay echo + WebRTC migration) runs end-to-end against this runtime with the A5 kernel.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A5's sync-fast-path
pollexposes a sampling race indriveAsync's deadlock probe: the probe's precondition (pendingHostCalls.size === 0 && !hasResumingThread()) can expire during the probe's own macrotask turn without theawaitingmembership changing — the same activation resumes off an engine continuation chunk (pin (j)), runs sync ready-polls, and re-parks through the A1 arm, registering a freshpendingHostCallsentry. The re-check only comparesawaitingmembership andreadyCandidates(), so the verdict fires against a live workload withhostCalls=1at trap time.Observed on polymorph-iroh's upstream-iroh spike (tokio parked on timer + synthetic-socket pollables, woken by detached host bridges) the moment its host adopted the A5 kernel —
DELTIC_DRIVE_TRACE=1:The same workload under the pre-A5 always-Promise poll never left the window open (every park held a registered A1 promise at every sample instant), which is why this only surfaced now.
Fix: re-check the full precondition at verdict time; a grown
pendingHostCalls(or a resuming thread) ⇒ re-probe, exactly like a membership change. A genuine deadlock still traps — nothing external exists to register a host call, so its re-check findshostCalls=0again (deadlock_test.ts pins stay green).Gates:
just test-runtime363/0 (deadlock pins exercised, not ignored — fixtures built);just conformance1254 passed / 0 failed / 95 xfail; the exposing workload (relay echo + WebRTC live migration) runs end-to-end against this branch with the A5 kernel.On a pinned regression test: the race needs "probe macrotask turn interleaved between an A1 settle and the next A1 park of the same activation" — I did not find a deterministic in-suite shape for it (the exposing repro is a real scheduler race under a live tokio guest). polymorph-iroh's spike gates it externally as of its next pin bump; if you want an in-repo pin, a
traceDrive-level injection point seems like the least-contrived option — happy to iterate.