bneradt opened a new pull request, #13535:
URL: https://github.com/apache/trafficserver/pull/13535

   The cache unit test harness starts the event and net processors but
   never stops them, so every test binary reaches exit() with ET_NET
   threads still running. Static destruction then frees globals out from
   under those threads: the records table in RecCore.cc is destroyed while
   a still-initializing event thread reads it through RecGetRecordInt(),
   and the ts::Metrics storage blob is released while NetHandler's activity
   loop increments a counter into it. Both are heap-use-after-frees, and
   under ASan the reporting thread races the exiting main thread. Usually
   the process dies first and the report is truncated to two lines with a
   zero exit status, so ctest reports a pass; occasionally the report
   deadlocks instead and the test hangs until ctest times it out after
   1500 seconds. The short tests that never touch the cache lose this race
   most often, which is why CacheAggregateWriteBuffer and CacheStripe are
   the ones that fail.
   
   This addresses the deadlock at its source by giving the harness's Catch2
   listener a testRunEnded hook that shuts the event system down and joins
   the event threads before the test binary returns from main. Once the
   threads are gone, static destruction has no concurrent reader to race,
   so neither use-after-free can be reported and the ASan reporting
   deadlock cannot arise.
   
   ### Verification
   
   Built with `--preset ci-rocky` (ASan, Debug) and looped the affected
   binaries directly, since a full `ctest` pass rarely trips the race and the
   failing runs still exit zero:
   
   | Test | Before | After |
   | --- | --- | --- |
   | `CacheAggregateWriteBuffer` | 15 ASan reports / 4000 runs | 0 / 2000 runs |
   | `CacheStripe` | 3 ASan reports / 600 runs | 0 / 600 runs |
   
   The 15 reports split into the two use-after-frees described above: six
   `READ of size 8` in `g_records_ht.find()`, reached from
   `initialize_thread_for_net()` on a thread that has not finished starting,
   and nine `WRITE of size 8` in `ts::Metrics::Counter::increment()` from
   `NetHandler::waitForActivity()`. Both are freed by the main thread from
   static destruction.
   
   All 20 `test_cache*` plus `test_ConfigVolumes` targets pass with the fix, so
   joining the event threads does not hang the tests that drive the cache and
   already call `TSSystemState::shut_down_event_system()` themselves.
   
   Ten full `ctest -j4` passes were clean, which is expected: this bug reports
   a zero exit status in the overwhelming majority of cases and is only visible
   in CI when the ASan report deadlocks instead, as in the Rocky Linux build
   that prompted this patch.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to