RFR: 8361912: ThreadsListHandle::cv_internal_thread_to_JavaThread does not deal with a virtual thread's carrier thread [v2]

[David Holmes]

On Thu, 24 Jul 2025 12:29:29 GMT, David Holmes <dholmes at openjdk.org> wrote:

>> src/hotspot/share/runtime/threadSMR.cpp line 839:
>> 
>>> 837:        }
>>> 838:        if (java_thread == nullptr) {
>>> 839:          // Virtual thread was unmounted, or else carrier has now terminated.
>> 
>> Nit: If `java_thread` for a virtual thread is null then this thread is unmounted. The words about the terminated carrier thread are confusing.
>> 
>> I have one concern here. Nothing protects this code from facing some unexpected conditions as this function is racy with mounting and unmounting of target virtual thread
>>  - this function can observe a `JavaThread` in `mount` or `unmount` transition where the thread identity of a `java.lang.Thread` associated with the `JavaThread` is not consistent (in a VTMS transition the thread identity might not match the stack trace)
>>  - nothing guaranties that the result returned by this function remains the same upon return as `mount`/`unmount` of the target virtual thread can be executed concurrently: a mounted virtual thread can be quickly unmounted or an unmounted virtual thread can be quickly mounted
>>  
>> So, it seems that this function is going to work correctly if used for target platform threads only or if the `JavaThread*` pointer is not really needed. Otherwise, the association with the `JavaThread` needs to be rechecked and its stability somehow guaranteed with any form of scheduling suspension or a VTMS transition disabler.
>
> The java_thread can be null because the carrier of a mounted vthread was not contained in the ThreadsListHandle.
> 
> The only thing this function guarantees is that if the vthread appeared to have a carrier and that carrier is protected by the TLH, then the caller can safely interact with the carrier knowing it can't terminate - same as for regular threads. The caller has to account for the potential async mounting/unmounting of  vthread - e.g. by handshaking the reported carrier and then confirming it is still the carrier.

Note that `java_thread` may already be null so we don't get to execute line 836.

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

PR Review Comment: https://git.openjdk.org/jdk/pull/26287#discussion_r2228385947