RFR: 8257831: Suspend with handshakes [v4]

[Richard Reingruber]

On Fri, 9 Apr 2021 16:12:16 GMT, Daniel D. Daugherty <dcubed at openjdk.org> wrote:

>> src/hotspot/share/runtime/handshake.cpp line 630:
>> 
>>> 628:     // exiting.
>>> 629:   }
>>> 630: }
>> 
>> I need a little help learning the steps of this dance :)
>> 
>> We reach here in _thread_in_vm. We cannot be suspended in this state. There
>> might be another thread waiting to handshake us to suspend us but why can't we
>> just ignore that and do the `_handshakee->set_exiting();` even without locking
>> HandshakeState::_lock?
>> 
>> Adding a handshake operation is lock free, so even if the check
>> `SafepointMechanism::should_process(_handshakee)` in L619 returns false a new
>> operation to process could have been added concurrently such that
>> `SafepointMechanism::should_process(_handshakee)` would return true when we
>> execute `_handshakee->set_exiting();` in L620. What am I missing?
>
> `HandshakeState::suspend()` is a synchronous handshake and adding the
> handshake to the queue is lock free, but the execution of the synchronous
> handshake itself requires a `HandshakeState::claim_handshake()` call which
> does acquire the lock in question. We (the suspend requester) hold the lock
> while the handshake is being processed so we either detect that
> _handshakee->set_exiting() won the race (in the target thread) or we (the
> suspend requester) win the race of setting the suspend flag so the target
> thread can't exit yet.
> 
> Hopefully that helps explain this dance.

Hi Dan,

thanks for picking up my question!

> `HandshakeState::suspend()` is a synchronous handshake and adding the
> handshake to the queue is lock free, but the execution of the synchronous
> handshake itself requires a `HandshakeState::claim_handshake()` call which
> does acquire the lock in question.

My point would be that the attempt to execute the synchronous handshake for
suspending a thread that is just about to call HandshakeState::thread_exit()
cannot make progress (blocks) while the target thread is not safe
(_thread_in_vm).

A synchronous handshake requires the target thread to be in a safe state for the
requester to execute the handshake operation.  When executing
HandshakeState::thread_exit() the suspendee is _thread_in_vm. And the requester
will find it to be `_not_safe` when calling `possibly_can_process_handshake()`
before calling `HandshakeState::claim_handshake()` or when calling
`can_process_handshake()` afterwards. In both cases try_process() returns with
failure _not_safe and the lock is not held.

++
 546	  if (!possibly_can_process_handshake()) {
 547	    // JT is observed in an unsafe state, it must notice the handshake itself
 548	    return HandshakeState::_not_safe;
 549	  }
 550	
 551	  // Claim the mutex if there still an operation to be executed.
 552	  if (!claim_handshake()) {
 553	    return HandshakeState::_claim_failed;
 554	  }
 555	
 556	  // If we own the mutex at this point and while owning the mutex we
 557	  // can observe a safe state the thread cannot possibly continue without
 558	  // getting caught by the mutex.
 559	  if (!can_process_handshake()) {
 560	    _lock.unlock();
 561	    return HandshakeState::_not_safe;
 562	  }

So isn't being unsafe sufficient to sync with suspend requests?

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

PR: https://git.openjdk.java.net/jdk/pull/3191