On Tue, 16 May 2023 14:33:05 GMT, Patricio Chilano Mateo <pchilanom...@openjdk.org> wrote:
>> The following patch fixes a bug introduced while refactoring the >> VirtualThreadStart/End events. Specifically, the code to delete the >> JvmtiThreadState of a terminating vthread was moved before we start the VTMS >> transition. That allowed said code to run concurrently with >> recompute_enabled() leading to different crashing modes. I wrote the >> detailed sequence of events leading to the crash in the bug comments. >> >> To fix it I moved the cleanup code back after the call to >> VTMS_unmount_begin(). Now, since the rebinding of the JvmtiThreadState to >> that of the carrier has to be done after this cleanup code is executed, >> otherwise we would delete the wrong JvmtiThreadState state, I had to add a >> boolean argument to VTMS_unmount_begin() to differentiate the last unmount >> call from the other ones. This is unfortunate since ideally >> VTMS_unmount_begin() would be oblivious to these two cases as with >> VTMS_mount_end() where we don't need to check if this is the first mount. >> I looked for other ways to solve it instead of the extra boolean argument >> but wasn't convinced. One way would be to have another >> JvmtiExport::cleanup_thread() that would handle this case. Another way which >> is very simple is to move the rebind_to_jvmti_thread_state_of() call to >> VTMS_unmount_end() instead. But that means during the transition the >> _jvmti_thread_state field of the carrier would be either null or that of the >> vthread, unlike today which is always that of the carrier during the >> transitions. I didn't want to change that behavior in this fix but I can >> also explore that route. >> >> I tested the patch with the reproducer I attached to the bug, plus I also >> run tiers1-3 in mach5. >> >> Thanks, >> Patricio > > Patricio Chilano Mateo has updated the pull request incrementally with one > additional commit since the last revision: > > remove extra assert Summary: Thumbs up. This is most definitely NOT a trivial fix. Normally the way I like to review these kinds of fixes is to map the failure modes back to the fix just to make sure I understand how each of the failure modes is covered by the changes. For this particular bug that has NOT been an easy thing to do. I believe the failure modes are complicated to follow because we "lost" the synchronization of being in the `VTMS_unmount_begin` transition which allowed a thread calling `recompute_enabled()` to race with our ending/exiting vthread/cthread combo. Ouch. The fix does restore the synchronization of being in the `VTMS_unmount_begin` transition so I can see how we would no longer be racing with the `recompute_enabled()` calling thread. ------------- Marked as reviewed by dcubed (Reviewer). PR Review: https://git.openjdk.org/jdk/pull/13949#pullrequestreview-1429333688
