On Mon, 24 Mar 2025 12:12:10 GMT, Michael McMahon <micha...@openjdk.org> wrote:
>> Guards against multi-deletes in `UnixSocketChannelReadWrite`. >> >> The fix can be verified as follows: >> >> >> make build-microbenchmark >> build/linux-x64/jdk/bin/java \ >> -jar build/linux-x64/images/test/micro/benchmarks.jar \ >> -f 1 -wi 1 -i 1 -t 2 UnixSocketChannelReadWrite >> >> >> Note that `-t 2` denotes the number of worker threads, i.e., parallelism. > >> > > Seems like the only explanation for why that is happening is someone >> > > cleaning out /tmp while the test is running >> > >> > >> > @Michael-Mc-Mahon No, the reason for the failure is multiple threads >> > creating a Unix socket in the very same `tempDir`, and then trying to >> > delete the `tempDir`. See the following execution flow involving multiple >> > threads: >> > ``` >> > 1. `tempDir` is statically assigned during class initialization (say >> > `/tmp/readWriteTest2414375588689416060`) >> > >> > 2. `Thread1` runs `beforeRun()`, which creates the Unix socket >> > `/tmp/readWriteTest2414375588689416060/1` >> > >> > 3. `Thread2` runs `beforeRun()`, which creates the Unix socket >> > `/tmp/readWriteTest2414375588689416060/2` >> > >> > 4. Both `Thread1` and `Thread2` runs `afterRun()` in parallel >> > >> > 5. `Thread1` runs `Files.delete(/tmp/readWriteTest2414375588689416060/1)` >> > and succeeds >> > >> > 6. `Thread2` runs `Files.delete(/tmp/readWriteTest2414375588689416060/2)` >> > and succeeds >> > >> > 7. `Thread1` runs `Files.delete(/tmp/readWriteTest2414375588689416060)` >> > and succeeds (since both `/1` and `/2` socket files are removed above, >> > and, hence, the directory is empty) >> > >> > 8. `Thread2` runs `Files.delete(/tmp/readWriteTest2414375588689416060)` >> > and fails, since the directory has already been deleted above >> > ``` >> > >> > >> > >> > >> > >> > >> > >> > >> > >> > >> > >> > Ideally, `tempDir` deletion should be performed by a single thread. But >> > replacing `Files::delete` with `Files::deleteIfExists` also does the job, >> > at a lower cost. >> >> What would happen then if thread1 completes before thread2 starts? Maybe, we >> _should_ delete the directory in a separate block, executed in a single >> thread at the every end. Is there a variant of the @teardown() annotation >> which executes at the end of the test? > > maybe the directory could be created and deleted in separate > Setup/Teardown(Level.Trial) invocations ? @Michael-Mc-Mahon, I use dedicated temporary directory per socket in cabb4145759a982ba32e9ade82066cfc75b5f05d. This reduces the amount of boilerplate for socket file administration (-13 LoC), and avoids any synchronization due to its shared-nothing state. Is this okay with you? ------------- PR Comment: https://git.openjdk.org/jdk/pull/24126#issuecomment-2749159972
