RFR: 8361103: java_lang_Thread::async_get_stack_trace does not properly protect JavaThread [v4]

[David Holmes]

On Wed, 16 Jul 2025 07:33:08 GMT, Serguei Spitsyn <sspitsyn at openjdk.org> wrote:

>> Sorry for being late to the party. I still need to walk through all the comments but it seems you have already caught the main problem with virtual threads. Nothing protects them from being concurrently mounted/unmounted/remounted. It is why the `JvmtiVTMSTransitionDisabler` is used in the JVMTI code. But, please, note that the `JvmtiVTMSTransitionDisabler` mechanism is enabled only when there is a JVMTI agent. Otherwise, it has been disabled for scalability purposes to exclude potentially high performance overhead at the VTMS transition points.
>
> With `JvmtiVTMSTransitionDisabler` there is normally no need in the carrier thread protection. A virtual thread is either mounted or unmounted. There is no carrier thread when virtual thread is unmounted. If virtual thread is mounted and VTMS transitions are disabled then the carrier thread is blocked until VTMS transitions are enabled again and the execution control is returned to the carrier. But JVMTI is playing by the established rules and always using TLH to protect the JavaThread associated with the virtual thread and its carrier.

If transition disabling is only available with JVMTI then I think we have a significant problem trying to deal with virtual threads in a number of APIs. Without disabling transitions you would basically need a loop that safely extracts the current carrier, handshakes with it, checks the VT is still mounted, and if not repeat the process. At the moment it appears that if the carrier is seen to have changed then we can just skip the VT - which would be very wrong IMO for most APIs e.g. thread dumps.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/26119#discussion_r2211933753