RFR: 8203817: Monitor::try_lock() should not call check_prelock_state()

[David Holmes]

On 25/05/2018 7:33 PM, Per Liden wrote:
> Hi David,
> 
> On 05/25/2018 09:51 AM, David Holmes wrote:
>> Hi Erik,
>>
>> On 25/05/2018 5:17 PM, Erik Osterlund wrote:
>>> Hi David,
>>>
>>> The change Per is proposing would make try_lock perform the same 
>>> checks that locking without safepoint checks does.
>>
>> Yet locking without a safepoint check also ensures 
>> _safepoint_check_required != Monitor::_safepoint_check_always. But 
>> try_lock does not and can be applied to both kinds of 
>> Monitors/Mutexes. Further it also has a stated, but not checked, 
>> precondition that the thread is not _thread_in_vm - though that may 
>> only be an issue if we were to block.
>>
>> I'm looking at the checks in check_prelock_state wondering which are 
>> only truly needed if we risk blocking? I think most of them, with the 
>> exception of is_crash_protected. Are we guaranteed never to run this 
>> JFR code on the WatcherThread?
>>
>> Though I'm also concerned about lock-ranking and deadlocks if we 
>> try_lock multiple locks. Notwithstanding the rank/deadlock code is far 
>> from ideal already.
>>
>> I don't mind removing checks/guards that really don't apply in the 
>> try_lock case, but I'm wary of removing all guards without being more 
>> sure of that.
> 
> Ok, so how about we just avoid the bad parts, by doing this:
> 
> http://cr.openjdk.java.net/~pliden/8203817/webrev.1

That seems a reasonable compromise.

There may still be possible deadlock issues if we use try-lock to 
acquire multiple locks, but we probably don't do that ... and as Erik 
noted there are issues with the ranking/deadlock detection anyway.

Thanks,
David

> /Per
> 
>>
>> Thanks,
>> David
>> -----
>>
>>> Perhaps locking without safepont checks could perform more sanity 
>>> checks, but that is a separate issue. I think try_lock should perform 
>>> the same checks that locking without safepoint checks does. The 
>>> alternatives are then to
>>>
>>> 1) Remove checking the prelock state like Per suggests for try_lock 
>>> (then they do the same checks), or
>>> 2) Overhaul the safepoint checking refactoring out the bits that 
>>> check the safepointing sanity chrcka from other deadlock checks (and 
>>> correct those to check for rank <= special, and not == special), 
>>> remove the safepoint checking part from try_lock and adding the 
>>> deadlock checking parts to lock without safepoints.
>>>
>>> Doing #2 seems like a different RFE. In fact I believe that would be 
>>> https://bugs.openjdk.java.net/browse/JDK-8184732 
>>> <https://bugs.openjdk.java.net/browse/JDK-8184732?jql=text%20~%20%22deadlock%20check%22> that 
>>> I filed a while back.
>>>
>>> In summary, there is a whole bunch of problems in the deadlock 
>>> detection system, and #2 makes it hard to not get dragged down in the 
>>> rabbit hole. #1 is sufficient to  make try_lock check as much (or 
>>> little) as locking without safepoint checking. And I think that is 
>>> enough for the scope of this change.
>>>
>>> Looks good to me.
>>>
>>> Thanks,
>>> /Erik
>>>
>>> On 25 May 2018, at 08:53, David Holmes <david.holmes at oracle.com 
>>> <mailto:david.holmes at oracle.com>> wrote:
>>>
>>>> Hi Per,
>>>>
>>>> Exactly what condition(s) does JFR violate? This is throwing away 
>>>> all the checks that guard against incorrect monitor use. It's not 
>>>> just about whether you'd block trying to acquire the Monitor, it's 
>>>> also about whether it is safe to acquire it from that code/thread in 
>>>> the first place. (Though I think some of the checks in there should 
>>>> also be considering the value of _safepoint_check_required.)
>>>>
>>>> Thanks,
>>>> David
>>>>
>>>>
>>>> On 25/05/2018 4:39 PM, Per Liden wrote:
>>>>> In debug builds, Monitor::try_lock() calls check_prelock_state() to 
>>>>> check the thread state, etc. The intention is to verify that the 
>>>>> call is made from the correct context, a context where we're 
>>>>> allowed to block and potentially safepoint. Unlike Monitor::lock(), 
>>>>> Monitor::try_lock() will never block, hence the call to 
>>>>> check_prelock_state() is overly strict and we should remove it. 
>>>>> Removing it would match the behavior of all other non-blocking 
>>>>> functions, like Monitor::lock_without_safepoint_check(), which 
>>>>> doesn't call check_prelock_state() either (for a good reason).
>>>>> The specific problem I've run into with this is related to JFR. 
>>>>> Monitor::try_lock() is by JFR to allow non-blocking event 
>>>>> generation, so that you can generate JFR events from "any" context 
>>>>> without risk blocking/safepointing (the logic is doing something 
>>>>> like, if try_lock() fails then put the event on a different queue 
>>>>> and let the next owner of the lock handle it). The overly strict 
>>>>> checks done by check_prelock_state() in try_lock() breaks this 
>>>>> logic, which in turn means that you can't generate JFR event from 
>>>>> "any" context as was intended.
>>>>> The patch to fix this is a one-liner, just remove the call to 
>>>>> check_prelock_state().
>>>>> Bug: https://bugs.openjdk.java.net/browse/JDK-8203817
>>>>> Webrev: http://cr.openjdk.java.net/~pliden/8203817/webrev.0
>>>>> /Per