On Mon, 23 Dec 2024 15:26:13 GMT, Robert Toyonaga <d...@openjdk.org> wrote:
>> This is a redo of [JDK-8304824](https://bugs.openjdk.org/browse/JDK-8304824) >> which was backed out by >> [JDK-8343726](https://bugs.openjdk.org/browse/JDK-8343726) due to problems >> documented in [JDK-8343244](https://bugs.openjdk.org/browse/JDK-8343244). >> >> The problem was that `NmtVirtualMemoryLocker` was not locking when the >> current thread is not attached by checking `Thread::current_or_null_safe() >> != nullptr`. This is necessary during VM init, but should not be allowed >> afterward. NMT may be used in `attach_current_thread` before the current >> thread is set. The lock was not being acquired in that case, which >> intermittently caused NMT accounting to become corrupted, triggering various >> assertions when future NMT operations are done. To fix this, I've adopted >> [Thomas' >> suggestion](https://github.com/openjdk/jdk/pull/21928#issuecomment-2460238057) >> to reverse the order of >> >> >> thread->register_thread_stack_with_NMT(); >> thread->initialize_thread_current(); >> >> >> in `attach_current_thread`. This allows `NmtVirtualMemoryLocker` to be >> locked after current thread is set. >> >> To allow for `NmtVirtualMemoryLocker` to be used during VM init, I've >> replaced the `ConditionalMutexLocker` check `Thread::current_or_null_safe() >> != nullptr` with a new flag: `_done_bootstrap`. This flag prevents the lock >> from being used during VM init, at which point we are single threaded >> anyway. This avoids errors due to Hotspot mutexes and current thread not yet >> being ready. >> >> I also added new asserts in `virtualMemoryTracker.cpp` to catch future bugs >> like this where the lock is not held when it should be. I updated the >> appropriate VMT tests to also lock (there were a few cases where locking was >> being bypassed) so they can pass the new asserts. >> >> I also removed the unused `_query_lock` variable in `MemTracker`. >> >> Testing: >> >> - On Linux amd64, I was able to consistently reproduce the errors described >> in [JDK-8343244](https://bugs.openjdk.org/browse/JDK-8343244) by increasing >> the number of test threads in >> `java/lang/Thread/jni/AttachCurrentThread/AttachTest.java`. The test >> consistently passes with the new changes in this PR. >> - hotspot_nmt , gtest:VirtualSpace, tier1 > > Robert Toyonaga has updated the pull request incrementally with one > additional commit since the last revision: > > comments, remove unneeded ifdef, remove typo In case my comment within the existing conversations gets missed I will re-state that I don't think you need any new "is bootstrapping" logic because you can just use `Threads::number_of_threads() > 0` to tell you if you need to acquire the NMT lock. Though that assumes that the `WatcherThread` does not use NMT ... but in that case you can use `WatcherThread::watcher_thread() != nullptr` as the bootstrap condition instead. ------------- Changes requested by dholmes (Reviewer). PR Review: https://git.openjdk.org/jdk/pull/22745#pullrequestreview-2531327078
