On Thu, 6 Apr 2023 03:35:17 GMT, Chris Plummer <cjplum...@openjdk.org> wrote:
>> The real purpose of this PR is to add virtual thread support to >> ThreadMemoryLeakTest.java, but this exposed bugs in both the debug agent and >> in TestScaffold, so those are being fixed also (and the debug agent bug is >> the CR being used). >> >> The debug agent bug is due to a race condition during VM exit. The VM is in >> the process of shutting down. The debug agent has already disabled JVMTI >> callbacks and has sent the VMDeathEvent. At this point in time there are >> also threads exiting that the debug agent knows about, but it will not get a >> ThreadEndEvent for because of the callbacks being disabled. Thus these >> threads remain in the debug agent's list of known threads, even though they >> have exited. The debuggee receives the VMDeathEvent and does a VM.resume(). >> During the debug agent's handing of the VM.Resume command, it iterates over >> all known threads and needs to map each to its ThreadNode so it can be >> resumed, and this mapping requires accessing the JVMTI TLS for the thread. >> The problem is some of the threads may have exited already, and therefore no >> longer have TLS. This results in the assert in the debug agent. This debug >> agent issue was already addressed for platform threads, but not for virtual >> threads, which is why w e started seeing this issue when this test was modified. The fix is to just replicate what is done for platform threads for virtual threads also. >> >> The TestScaffold bug is that if the debuggee crashes/asserts, this is likely >> to go unnoticed, especially if it happens during VM exit (and the test >> essentially has already completed). Because of this TestScaffold bug, the >> debug agent bug above did not result in a test failure. After fixing >> TestScaffold to check the exitCode of the debuggee process, the test started >> to appropriately fail until the debug agent was fixed. >> >> One other thing to point out is the OOME issue I started getting frequently >> when testing with virtual threads. Since virtual threads are created at a >> much higher rate than platform threads, their creation started to overwhelm >> the debugger (actually the JDI implementation). There is already a mechanism >> in place to do a VM.HoldEvents if JDI has queue up 10,000 events. The >> problem is that events are coming in so fast that even after doing the >> VM.HoldEvents, the number of queued events continues to go up for a while, >> and sometimes reaches 30,000 or more. This raises the peak memory usage of >> the test quite a bit. Since the test purposely uses a small heap so a memory >> leak is quickly and reliably detected, the large queue often results in an >> OOME. Because of this I make virtual threads sleep for 100ms instead of 50ms >> to slow down their creation, and this resolved the issue. >> >> I tested by running all of test/jdk/com/sun/jdi 25 times on each platform >> with and without virtual thread testing enabled. > > Chris Plummer has updated the pull request incrementally with one additional > commit since the last revision: > > TestScaffold now waits indefinitely for process exit. Simpler coding of > sleep time. Marked as reviewed by lmesnik (Reviewer). ------------- PR Review: https://git.openjdk.org/jdk/pull/13246#pullrequestreview-1374011838
