On Thu, 20 Jul 2023 06:53:26 GMT, Serguei Spitsyn <sspit...@openjdk.org> wrote:
> This problem is encountered when a JVMTI agent is loaded into running VM. The > JvmtiExport::get_jvmti_interface() is called from the agent's Agent_OnAttach > entrypoint. To support virtual threads it enables JVMTI notifications from > the VirtualThread class with a call to: > `JvmtiEnvBase::enable_virtual_threads_notify_jvmti()`. > The problem is that there is no JVMTI environments at this point yet. This > assert is hit when a virtual thread is created concurrently after the JVMTI > notifications have been enabled but the requested JVMTI environment has not > been created yet. > The fix is to create a JVMTI env first and only then to enable the JVMTI > notifications. > > This issue is very hard to reproduce. I had to use some tricks with adding > `os::naked_short_nanosleep()` and also by refactoring the test > `VThreadTLSTest.java`. At least, I was able to verify this test does not fail > with my fix anymore. > > Testing: > - submitted hundreds of `VThreadTLSTest.java` mach5 runs on several > platforms in the `fastdebug` mode > - in progress: mach5 tiers 1-6 This seems a reasonable solution. Thanks. ------------- Marked as reviewed by dholmes (Reviewer). PR Review: https://git.openjdk.org/jdk/pull/14945#pullrequestreview-1538610965
