RFR: 8215491: ICStubInterface::finalize finds zombie nmethod with ZGC concurrent class unloading

[dean.long at oracle.com]

+1

dl

On 12/19/18 11:48 AM, Coleen Phillimore wrote:
> I think this looks good.  Then we don’t have to know who has already cleaned them or if they did it in the right order.
>   
> Thanks
> Coleen
>
>> On Dec 18, 2018, at 4:30 AM, Erik Österlund <erik.osterlund at oracle.com> wrote:
>>
>> Hi Coleen,
>>
>>> On 2018-12-17 22:31, coleen.phillimore at oracle.com wrote:
>>>> On 12/17/18 2:57 PM, Erik Österlund wrote:
>>>> Hi Dean,
>>>>
>>>>> On 2018-12-17 20:27, dean.long at oracle.com wrote:
>>>>>> On 12/17/18 8:56 AM, Erik Österlund wrote:
>>>>>> Hi,
>>>>>>
>>>>>> Today, when converting a not_entrant nmethod to zombie, the sweeper first clears the IC stubs of that nmethod, and then make_zombie(). With concurrent class unloading, this ordering is a bit unfortunate. Between clearing the IC stubs and making it zombie, the concurrent GC threads may hit a timing window where they will create IC stubs for concurrently cleaning the ICs of this still is_alive() nmethod.
>>>>> Do we normally have to creates stubs to clean an IC?
>>>> Yes if you clean an IC in an nmethod that is_alive(), we currently have to do that using IC stubs. I have some ideas going forward for how we can stop doing that, but that is outside of scope for now.
>>>>
>>>>>> The result is that during safepoint cleanup, when we finalize the IC stubs, we find one associated with a zombie.
>>>>>>
>>>>>> The unregistering of the nmethod from the GC will block during concurrent nmethod unloading, so clearing the IC stubs after the nmethod has become zombie is a lot more sane as there can not be any such races then.
>>>>>>
>>>>> Where is the unregistering happening?
>>>> Inside of make_zombie() we unregister_nmethod() on CollectedHeap. In there, ZGC makes sure we wait until concurrent nmethod unloading is over before continuing.
>>> Can you move clean_ic_stubs() inside of make_not_entrant_or_zombie() to maintain the order dependence?
>> Sure. The implication of that is that the sweeper will redundantly check for IC stubs to clear on unloaded nmethods that can't have any. But I agree that the model is more clear if it is inside of make_not_entrant_or_zombie(), and think that is worth the cost.
>>
>> New webrev:
>> http://cr.openjdk.java.net/~eosterlund/8215491/webrev.01/
>>
>> Thanks for reviewing.
>>
>> /Erik
>>
>>> thanks,
>>> Coleen
>>>>> Doing the clearing after it's a zombie does sound safer.
>>>> Agreed!
>>>>
>>>> Thanks for reviewing.
>>>>
>>>> /Erik
>>>>
>>>>> dl
>>>>>
>>>>>> Bug:
>>>>>> https://bugs.openjdk.java.net/browse/JDK-8215491
>>>>>>
>>>>>> Webrev:
>>>>>> http://cr.openjdk.java.net/~eosterlund/8215491/webrev.00/index.html
>>>>>>
>>>>>> Thanks,
>>>>>> /Erik