[heifner]

Consumes Wire-Network/wire-vcpkg-registry#23 (merged as 81b3b452 on top of #22): registry baseline/reference → 81b3b452, which carries both softfloat 3.0.0#3 (#701) and wire-sys-vm 1.2.0#2; wire-sys-vm override → 1.2.0#2 (Wire-Network/eos-vm@f760622, eos-vm#14 with #15).

Why

read-only-trx-irreversible-test timeoutTest aborted nodeop with SIGSEGV in CI run 39807473127. Read-only threads share one compiled sys-vm JIT module; a timed-out run revoked its code pages while a peer's timed_run exit cleared the process-wide timeout flag, so the faulting thread was handed to the previous SIGSEGV handler. Any API node with more than one read-only thread can hit it. The pinned eos-vm classifies faults in generated code by location, serializes timed_run registration and exit transitions under a mutex (timeout publication stays lock-free in signal context), and backports the upstream signal-mask and <cstdint> fixes.

Verification

A fresh worktree build configured against registry 81b3b452 with both overrides: vcpkg installs [email protected]#3 (exporting THREAD_LOCAL=__thread) and [email protected]#2 (port tree 4aca5b06), and the installed sysio/vm headers carry the fix. Read-only test matrix on the final headers (CI gcc-14 chroot, unit/plugin read-only and checktime suites under all three runtimes, the read-only/interrupt cluster tests repeated until-fail:3, both perf tests): all green on the final headers — unit and plugin suites rc=0 under sys-vm, sys-vm-jit and sys-vm-oc, 11/11 ctest entries in pass 1, the 8 read-only/interrupt cluster tests 3/3 each, both perf tests passed, no crash signatures in any log.

Regression test (second commit)

read-only-trx-irreversible-test could pass after the API node died: pushTransaction turns every clio failure, a refused connection included, into a (False, message) result; the "should fail" cases accepted any false result; worker threads reported through one shared errorInThread flag that every worker reset and that two call sites never checked, so a failing thread only printed a traceback; nothing checked that the node was still running. tests/read_only_trx_test.py now runs workers on a WorkerThread that keeps its exception (joinAll fails the test on any of them), validates expected failures by exception code (tx_cpu_usage_exceeded / deadline_exception for the forever and infinite transactions and for fast or slow ones that lose to the read window, wasm_execution_error for the segv action), and asserts apiNode.verifyAlive() after every concurrent phase and after the segv cases.

Checked against the pinned build: the irreversible variant passes (-p 2 -n 3 --read-only-threads 2 --api-node-read-mode irreversible); the 16-thread parallel variant passes (--read-only-threads 16 --num-test-runs 2); a run in which the API node is sent SIGSEGV during timeoutTest fails within two seconds of the kill: worker thread failures: … clio failed without the expected exception, is nodeop still up? Failed http request to nodeop ….

Overlap

#701 (softfloat 3.0.0#3) edits the same two files; whichever lands second re-pins to a registry commit that carries both bumps.

[huangminghuang]

Medium — Make the regression test fail if nodeop exits

This PR pins the dependency to fix a SIGSEGV reproduced by read-only-trx-irreversible-test, but that test can still report success after the process dies:

• transactions.py:253-265 converts every CLI failure, including connection refusal after node death, into a normal (False, message) result.
• read_only_trx_test.py:428-442 accepts any false result as the expected timeout/error.
• Worker threads are only joined at lines 471-484, so their assertion failures do not propagate, and the main path later marks the test successful at line 507.

This occurred in the GCC CI job: the test artifact records repeated connection refusals followed by a passing result, while the accompanying core artifact contains a nodeop core generated during the test. An ineffective dependency pin could therefore still merge green.

Please propagate worker failures, validate the specific expected timeout/VM error instead of any false result, and assert apiNode.verifyAlive() after the concurrent workload.

[heifner]

Addressed in f7abf65 (tests/read_only_trx_test.py, second commit): worker threads run on a WorkerThread that keeps its exception and joinAll fails the test on any of them; expected failures are validated by exception code (tx_cpu_usage_exceeded / deadline_exception for the forever and infinite transactions and for fast or slow ones that lose to the read window, wasm_execution_error for the segv action), so a refused connection no longer passes as an expected failure; apiNode.verifyAlive() is asserted after every concurrent phase and after the segv cases.

Checked against the pinned build: the irreversible and the 16-thread parallel variants pass; a run in which the API node is sent SIGSEGV during timeoutTest fails within two seconds with worker thread failures: … clio failed without the expected exception, is nodeop still up? Failed http request to nodeop …. The read-only matrix on the final headers is in the PR body.

[huangminghuang]

Reviewed the dependency pin and follow-up regression-test hardening. The worker failures now propagate, expected transaction errors are distinguished from transport failures, and API-node liveness is checked. No actionable findings remain.

[huangminghuang]

Reviewed current head cca582d. No actionable findings.