On Fri, 30 Jun 2023 11:27:58 GMT, Serguei Spitsyn <[email protected]> wrote:
> The JVMTI function `SetEventNotificationMode` can set notification mode
> globally (`event_thread == nullptr`) for all threads or for a specific thread
> (`event_thread != nullptr`). To get a stable mount/unmount vision of virtual
> threads a JvmtiVTMSTransitionDisabler helper object is created :
> `JvmtiVTMSTransitionDisabler disabler(event_thread);`
>
> In a case if `event_thread == nullptr` the VTMS transitions are disabled for
> all virtual thread,
> otherwise they are disabled for a specific thread if it is virtual.
> The call to `JvmtiEventController::set_user_enabled()` makes a call to
> `recompute_enabled()` at the end of its work to do a required bookkeeping. As
> part of this work, the `recompute_thread_enabled(state)` is called for each
> thread from the `ThreadsListHandle`, not only for the given `event_thread`:
>
> ThreadsListHandle tlh;
> for (; state != nullptr; state = state->next()) {
> any_env_thread_enabled |= recompute_thread_enabled(state);
> }
>
> This can cause crashes as VTMS transitions for other virtual threads are
> allowed.
> Crashes are observed in this small function:
>
> bool is_interp_only_mode() {
> return _thread == nullptr ? _saved_interp_only_mode != 0 :
> _thread->is_interp_only_mode();
> }
>
> In a case `_thread != nullptr` then the call needs to be executed:
> `_thread->is_interp_only_mode()`.
> But the filed `_thread` can be already changed to `nullptr` by a VTMS
> transition.
>
> The fix is to always disable all transitions.
> Thanks to Dan and Patricio for great analysis of this crash!
>
> Testing:
> - In progress: mach5 tiers 1-6
Marked as reviewed by cjplummer (Reviewer).
-------------
PR Review: https://git.openjdk.org/jdk/pull/14728#pullrequestreview-1507636034
