RFR: 8215773: applications/kitchensink/Kitchensink.java crash with "assert(ZAddress::is_marked(addr)) failed: Should be marked"

[Erik Österlund]

Hi Coleen,

On 2019-01-04 15:16, coleen.phillimore at oracle.com wrote:
>
>
> On 1/4/19 8:28 AM, Erik Österlund wrote:
>> Hi Coleen,
>>
>> On 2019-01-04 13:37, coleen.phillimore at oracle.com wrote:
>>>
>>> Hi Erik,  This is kind of creepy.  They use the dead mirror and call 
>>> identity hash on it, which writes to the dead mirror.
>>
>> Yup, that's right. This is equivalent to how during code cache 
>> unloading, possibly dead java.lang.invoke.CallSite objects are chased 
>> down, and their possibly dead CallSiteContext are queried so that we 
>> can write to its dependency context. It is creepy indeed, but is 
>> completely safe to do in the context it is done in, as the same 
>> thread performing the accesses, is the thread that will eventually 
>> free the memory of the dead objects. And in that incident, I did the 
>> same thing: use AS_NO_KEEPALIVE when peeking at the dead objects.
>>
>>> Are these functions called at a safepoint?  Maybe JFR could do 
>>> something different here.
>>
>> Yeah, they are called from safepoints with STW-unloading, and called 
>> by the ZDriver (the concurrent GC thread controlling the GC phases), 
>> during concurrent class unloading.
>
> Ah, ok.  I was getting this mixed up with the cld_unloading_do() calls 
> in jfr.  This is called by do_unloading.
> Looks strange but good. 

Thanks for the review!

> This is for 12 right?

Yup.

Thanks,
/Erik

> Thanks,
> Coleen
>>
>> Thanks,
>> /Erik
>>
>>> Coleen
>>>
>>>
>>> On 1/4/19 5:10 AM, Erik Österlund wrote:
>>>> Hi,
>>>>
>>>> During SystemDictionary::do_unloading(), JFR pokes around at the 
>>>> dead mirror oops that died due to class unloading.
>>>> The mirror is loaded without AS_NO_KEEPALIVE, which makes ZGC load 
>>>> barriers a bit sad, because they dislike loading dead oops.
>>>>
>>>> The solution is to load these mirrors with AS_NO_KEEPALIVE 
>>>> decorators, indicating that it is okay for them to be dead as we 
>>>> are only peeking at them. This is fine to do during concurrent 
>>>> execution by the GC thread that is performing class unloading, and 
>>>> hence controls the GC cycle, making sure that the memory of the 
>>>> dead oops is not reclaimed until the accesses have safely finished.
>>>>
>>>> Bug:
>>>> https://bugs.openjdk.java.net/browse/JDK-8215773
>>>>
>>>> Webrev:
>>>> http://cr.openjdk.java.net/~eosterlund/8215773/webrev.00/
>>>>
>>>> Manually ran kitchensink 24 hours to make sure this doesn't happen 
>>>> again. And hs-tier1-6 on linux.
>>>>
>>>> Thanks,
>>>> /Erik
>>>
>>
>