Separate shared service ownership from thread-local logging caches

This commit is contained in:
jdb committed 2026-09-22 07:18:22 -05:00
1 parent c1c8a3c072
commit 761bd66793
6 files changed
+386 -157

No files matched your search

+19 -6
View File
@@ -9,8 +9,8 @@ Run from the repository root:
| --- | --- |
| `nimble test` | All unit, regression, and ThreadSanitizer tests under ORC with threads enabled |
| `nimble testUnit` | Existing embedded unit tests |
| `nimble testRegression` | Configuration freshness, refresh failure recovery, appender-state cleanup, and writer shutdown |
| `nimble testThreadSafety` | Separate ThreadSanitizer processes for service ownership, custom state, and formatter closures, plus an independent-service control |
| `nimble testRegression` | Handle lifetime, configuration freshness, refresh failure recovery, appender-state cleanup, and writer shutdown |
| `nimble testThreadSafety` | Separate ThreadSanitizer processes for service ownership, last-owner teardown, reconfiguration, custom state, and formatter closures, plus an independent-service control |
`nimble test` attempts every suite even if an earlier suite fails to compile or
run, summarizes the failing suites, and exits unsuccessfully if any suite fails.
@@ -23,8 +23,8 @@ The other tests do not need background I/O to exercise ownership.
The full run and the sanitizer subset require Linux, Bash, GNU `timeout`, and a
C compiler and runtime supporting `-fsanitize=thread`. The sanitizer instruments
the generated Nim C code, including the ORC runtime. A sanitizer report, timeout,
crash, or runtime startup
failure causes a nonzero result; consult the diagnostic to distinguish these.
crash, or runtime startup failure causes a nonzero result; consult the diagnostic
to distinguish these.
Each scenario runs separately so one failure does not hide another. To inspect
one scenario after building:
@@ -33,14 +33,22 @@ TSAN_OPTIONS=halt_on_error=1:exitcode=66 timeout 30s tests/bin/thread_safety ser
```
The worker start barrier ensures argument setup finishes before the concurrent
operations begin. Parent owners survive until all workers join. Custom-state
operations begin. In `service-lifetime`, workers first acquire local services;
the parent then releases the original handle before allowing them to proceed.
Thread arguments contain borrowed pointers so no hidden owning argument keeps
the service alive. Atomic construction/destruction counters verify that every
tracked appender is released exactly once after the workers join.
`service-reconfigure` exercises concurrent updates and local cache refresh.
The other scenarios retain parent owners until all workers join. Custom-state
and closure tests only read application data; their concurrent writes are the
implicit ownership operations caused by cloning and destruction. The independent
case creates a separate service inside each worker and should pass. Custom-state
and formatter cloning is serialized with a lock, matching configuration refresh;
the copied appenders are used and destroyed outside that lock.
The three in-process regression tests are deterministic. Shutdown tests launch
The in-process regression tests are deterministic. The lifetime suite checks
copying, self-assignment, replacement, moving, container copies, and exporting a
handle from a local service. Shutdown tests launch
a fresh process for each destination and check all 10,000 messages after normal
exit, using isolated temporary paths and a 15-second deadline. They exercise the
real exit hook without sleeps to hide queued-message loss. Like the race tests,
@@ -51,3 +59,8 @@ This suite does not yet force an active producer/channel-close race, separately
exercise the shared `testutil.LoggedMessages` graph, or exhaust ORC cycle-collector
interleavings. It provides targeted regressions for the reproduced ownership
defects and the configuration, cleanup, and shutdown findings.
After the shared-handle implementation, the service sanitizer scenarios, lifetime
suite, and appender-state cleanup regression should pass. The custom-state and
formatter sanitizer cases, `isEnabled` freshness, failed-refresh recovery, and
writer shutdown remain failing regressions for subsequent work.