RFR: 8255384: Remove special_runtime_exit_condition() check from SS::block() [v2]

[Patricio Chilano]

Hi Richard,

On 11/3/20 12:22 PM, Richard Reingruber wrote:
> On Thu, 29 Oct 2020 22:10:58 GMT, Patricio Chilano Mateo <pchilanomate at openjdk.org> wrote:
>
>>> Hi all,
>>>
>>> Please review the following patch that removes the call to handle_special_runtime_exit_condition() from SS::block(). This avoids recursive calls when transitioning and stopping for safepoints and also makes the code simpler to read since it is not trivial to deduce why we need to execute the check for certain states but not others, i.e. what exact scenarios we are trying to guard against.
>>>
>>> Method handle_special_runtime_exit_condition() checks for external suspends, object deoptimization, async exceptions and JFR suspends. All these need to be checked when transitioning to java and when transitioning out of native (except for async exceptions when transitioning to thread_in_vm). In SS::block() this check is executed for the _thread_new_trans, _thread_in_native_trans and thread_in_Java cases. For _thread_new_trans, we know the JT will have to go through JavaCallWrapper() the first time it transitions to Java and that already has a check for handle_special_runtime_exit_condition(). For _thread_in_native_trans, transitioning out of native already has checks for external suspends, object deoptimization and JFR suspends in check_safepoint_and_suspend_for_native_trans() which is called from ThreadStateTransition::transition_from_native()(called either directly or through the ThreadStateTransition wrappers) and check_special_condition_for_native_trans (for native w!
>   rappers c
>   ase). So that leaves the thread_in_Java case.
>>> Careful analysis shows the handle_special_runtime_exit_condition() call in SS::block() prevents JTs transitioning back to Java from escaping after being externally suspended. This can happen when calling SafepointMechanism::process_if_requested() while transitiong back to java without a later check for external suspend. Looking at the callers of SafepointMechanism::process_if_requested() we see that this can happen from handle_polling_page_exception(), java_suspend_self_with_safepoint_check() and check_safepoint_and_suspend_for_native_trans(). An example of this can be shown for the handle_polling_page_exception() case:
>>>      - JT hits a poll exception while executing nmethod.
>>>      - JT calls handle_polling_page_exception() ( which doesn't use ThreadStateTransition wrappers) and calls SafepointMechanism::process_if_requested()
>>>      - Stops for a safepoint due to a VM_ThreadSuspend request
>>>      - Upon unblocking from the safepoint, unless we have the check in SS::block() the JT will transition back to java without actually suspending
>>>
>>> The "escape from suspend" scenarios for the other callers of SafepointMechanism::process_if_requested() are described in the comments of the bug as well as the proper fixes.
>>>
>>> I have tested the patch several times in mach5 tiers1-7 and saw no issues. Let me know if you think I should run any other special tests.
>>>
>>> Thanks,
>>> Patricio
>> Patricio Chilano Mateo has updated the pull request incrementally with one additional commit since the last revision:
>>
>>    Make direct calls instead of using transition wrappers
> Hi Patricio,
>
> the change looks good to me.
Thanks for looking at this.

> In `JavaThread::check_and_handle_async_exceptions()` the block depending on is_at_poll_safepoint() looks like dead code now. I wonder if `ThreadSafepointState::_at_poll_safepoint` could even be DEBUG_ONLY?
Yes, I actually thought about doing that in the first version but then I 
realized that code was already dead even before this change. We only 
call set_at_poll_safepoint() in handle_polling_page_exception() and the 
handle_special_runtime_exit_condition() call in SS::block() already 
excludes checking async exceptions for that case. The call I removed 
from ~TIVMFH was exactly the same. So I don't see a path where it could 
be called where is_at_poll_safepoint() returned true.
I agree that _at_poll_safepoint should probably be DEBUG_ONLY. Then we 
should add an assert in check_and_handle_async_exceptions(). Do you 
think I should do that here or in another bug?

Thanks,
Patricio
> Thanks, Richard.
>
> -------------
>
> PR: https://git.openjdk.java.net/jdk/pull/913