RFR: 8231289: Disentangle JvmtiRawMonitor from ObjectMonitor and clean it up
david.holmes at oracle.com
Mon Oct 7 11:23:34 UTC 2019
On 7/10/2019 6:42 pm, serguei.spitsyn at oracle.com wrote:
> Hi David,
> Suppose, T1 thread is blocked on a raw monitor hold by thread T3 in JVMTI
> MonitorContendedEntered callback on an object monitor hold by thread T2.
> In theory, if only raw monitor is used in further deadlock detection
> then a possible dependency loop with T1 an T2 involved won't be detected.
> For full analysis the dependency chain through object monitor have to be
> also analyzed.
> This would add extra complexity into the deadlock detection code for
> very rare corner case.
> I do not care much about this, and so, like the approach.
To be clear the case you describe is not handled by either the current
code nor my proposed update.
> The fix itself looks good to me.
> On 10/3/19 20:58, David Holmes wrote:
>> Okay, to allow for me to make forward progress here I've decided to
>> follow the "principle of least brokenness" ;-)
>> Recap: Because of JVMTI event callbacks it is possible for a thread to
>> set its current pending monitor as a JvmtiRawMonitor when it was
>> already set for an ObjectMonitor. This is broken in at least two ways:
>> - when the raw monitor use completes the pending monitor is set to
>> NULL, not restored to the ObjectMonitor
>> - whilst the raw monitor is seen as the pending monitor the
>> ObjectMonitor is not considered by the deadlock detection logic
>> Ignoring what I'm currently doing with jvmtiRawMonitor, I do not know
>> how to fix the above brokenness and it is out of scope for what I am
>> trying to do.
>> So what I propose is to make things no more broken than they are now,
>> and actually improve things a little:
>> - the pending JvmtiRawMonitor is given preference over the
>> ObjectMonitor in the deadlock detection code (this emulates current
>> situation of the raw monitor overwriting the pending ObjectMonitor)
>> - we no longer lose the fact we were also pending on an ObjectMonitor
>> - the stack printing code in the deadlock detector prints information
>> about both the raw monitor and the ObjectMonitor
>> Updated webrev:
>> The only change is threadService.cpp
More information about the hotspot-runtime-dev