On Fri, 18 Nov 2022 04:55:10 GMT, Serguei Spitsyn <sspit...@openjdk.org> wrote:
>> I think you need a flag that tells you if
>> `init_static_notify_jvmti_events()` has been called.
>
> A part of the initialization sequence we need to know is:
>
> create_vm() {
> . . .
> // Launch -agentlib/-agentpath and converted -Xrun agents
> if (Arguments::init_agents_at_startup()) {
> create_vm_init_agents(); => {
> <loads all agents and calls AgentOnLoad entry points> =>
> get_jvmti_interface() => set_notify_jvmti_events(true)
> }
> . . .
> init_globals() => javaClasses_init() =>
> java_lang_VirtualThread::init_static_notify_jvmti_events()
>
> The `create_vm_init_agents()` is called in the context of unattaching thread.
> In this context a call to
> `java_lang_VirtualThread::init_static_notify_jvmti_events()` is guaranteed to
> happen after all the agents were successfully loaded at startup and executed
> their `AgentOnLoad` entree points which make calls to `vm->GetEnv()` that
> transitively call to `get_jvmti_interface()` and
> `java_lang_VirtualThread::set_notify_jvmti_events(true)`.
>
> The conclusion is that the
> `java_lang_VirtualThread::init_static_notify_jvmti_events()` is always called
> at startup (single-threaded execution mode) after load of all the agents.
> In opposite, all calls to `get_jvmti_interface()` from the `AgentOnAttach`
> entry points have to in context of attached threads. I'm thinking if we could
> add an assert to ensure it is always the case.
> We can add a comment on this but I'm puzzled on how to make it clear and
> simple.
If there are no command line agents, then on startup
`vthread_notify_jvmti_events` is not set true. Because it is not true, when
`javaClasses_init()` calls `init_static_notify_jvmti_events()`, it does
nothing. The whole point of the code we are reviewing here is to make sure
`init_static_notify_jvmti_events()` is called while
`vthread_notify_jvmti_events == true` so it actually does something. However,
the code here does not bother calling `init_static_notify_jvmti_events()` if
the current thread is detached, but it does still set
`vthread_notify_jvmti_events = true`. This means that if this code gets called
a second time, this time with the current thread attached, it will not call
`init_static_notify_jvmti_events()` due to `vthread_notify_jvmti_events ==
true`, but it seems it should be calling it.
What I believe to be the flaw here is that you call
`set_notify_jvmti_events(true)` even if you don't call
`init_static_notify_jvmti_events()`.
-------------
PR: https://git.openjdk.org/jdk/pull/11204
