RFR (S) 8197844: JVMTI GetLoadedClasses should use the Access API

Mon Mar 19 23:38:22 UTC 2018

On 3/19/18 7:32 PM, Roman Kennke wrote:
> The change looks good to me.
>
> I have also verified that Shenandoah does not call into any of the
> iterators that might resurrect CLDs by mistake.

That's good!  Thank you for verifying this.  These iterators should now 
be only runtime iterators.
thanks,
Coleen

>
> Thanks, Roman
>
>
>> After reading my own reply and Erik's, I want to reconsider my change
>> for this bug.
>>
>> There are many places that iterate over ClassLoaderDataGraph, and I've
>> found another that is buggy in the same way that JVMTI GetLoadedClasses
>> should have a G1 barrier.  I've also spent more time trying to find if
>> other uses are buggy, with dissatisfying inconclusion.
>>
>> The safest thing to do is mark loaders that we are iterating over as
>> alive.  This alive barrier doesn't affect oops_do, cld_do or
>> do_unloading, which are the cases that would be walked by the GC.   It
>> also can't include verify, which I've removed from the original change.
>>
>> I can't assert that the thread isn't a GC thread because
>> VM_GC_HeapInspection builds a KlassInfoTable as a GC thread.
>>
>> I think a klass walk of the ClassLoaderDataGraph when done for profiling
>> and assorted runtime code, must necessarily keep the classes alive
>> during the interval between GCs.   The class can be unloaded in a
>> further GC.
>>
>> open webrev at http://cr.openjdk.java.net/~coleenp/8197844.02/webrev
>>
>> This is the same patch with the ensure_loader_alive removed from
>> ClassLoaderDataGraph::verify.
>>
>> thanks,
>> Coleen
>>
>> On 3/16/18 11:29 AM, Erik Österlund wrote:
>>> Hi Coleen,
>>>
>>> I for one would like to see your version of the fix, plus some
>>> suitable assert to catch bad use of these APIs.
>>> The motivation is that it seems unfortunate to expose APIs for getting
>>> Metadata that when retrieved is unstable and unsafe to use unless you
>>> have deep knowledge about GC races and class unloading, and hope all
>>> callsites know that they can't just use the things they retrieved
>>> because they are dangerously explosive. The only problem with the more
>>> conservative API is if the GC itself calls into the code and
>>> accidentally keeps everything alive preventing class unloading. But I
>>> think a suitable assert could catch that.
>>>
>>> I'm not 100% sure what the assert condition should be. But I'm
>>> thinking something along the lines of:
>>> Thread::current()->is_Java_thread() ||
>>> Thread::current()->is_VM_thread() &&
>>> (!SafepointSynchronize::is_at_safepoint() || !HeapLock->is_locked())
>>> Other suggestions for a good condition are welcome - it is just an
>>> example that may or may not work.
>>>
>>> Thanks,
>>> /Erik
>>>
>>> On 2018-03-16 16:01, coleen.phillimore at oracle.com wrote:
>>>> Hi Roman, thank you for your comments.
>>>>
>>>> On 3/15/18 9:46 AM, Roman Kennke wrote:
>>>>> Am 14.03.2018 um 18:37 schrieb coleen.phillimore at oracle.com:
>>>>>> Summary: Make sure the holder of a class loader is accessed during
>>>>>> iteration of CLDG
>>>>>>
>>>>>> This is where we should have put the GC barrier.  This can be cleaned
>>>>>> somewhat when we have a WeakHandle holder in the ClassLoaderData, then
>>>>>> the code in ensure_loader_alive() becomes _holder.resolve();
>>>>>>
>>>>>> Tested with tier1-5.
>>>>>>
>>>>>> open webrev at http://cr.openjdk.java.net/~coleenp/8197844.01/webrev
>>>>>> bug link https://bugs.openjdk.java.net/browse/JDK-8197844
>>>>>>
>>>>>> Thanks,
>>>>>> Coleen
>>>>> I am not sure at all that this wouldn't have adverse effects. Suppose a
>>>>> GC is iterating all CLDs and marks all of them alive, we'd probably
>>>>> never unload it? I probably wouldn't do the keep-alive stuff wholesale
>>>>> in the iterator methods.
>>>>>
>>>>> I've made (and later withdraw) a (IMO) more straightforward patch to
>>>>> address the same:
>>>>>
>>>>> http://cr.openjdk.java.net/~rkennke/8199612/webrev.00/
>>>>>
>>>>> It has the advantage that it doesn't to a bogus load, just to keep
>>>>> something alive. It loads the mirror and at the same time communicates
>>>>> the GC to keep it alive. Maybe better?
>>>> So when I made this change, I was being conservative.  Anytime we
>>>> iterate over the CLDG, and take out oops to put somewhere else, we
>>>> need to mark that the CLD needs to stay alive for that GC, until the
>>>> somewhere else is walked.  We also have to worry about these classes
>>>> being unloaded if we are still holding metadata after walking the CLDG.
>>>>
>>>> I've gone through all the CLDG::classes_do, loaded_classes_do,
>>>> dictionary_all_entries_do, methods_do, modules_do and packages_do at
>>>> least 100 times.  (also oops_do variants and cld_do).   All these
>>>> calls do something with metadata, and it appears that these places
>>>> are unaffected by GC or class unloading, except the instance in
>>>> jvmtiGetLoadedClasses that Poonam found.
>>>>
>>>> Also if the closure or function passed to the CLDG iterator functions
>>>> stop for a safepoint and unload, the walk will be broken.  I didn't
>>>> find any instances of this in the code on my 100 traversals.  My
>>>> change is trying to make it so I don't have to do this analysis anymore.
>>>>
>>>> Poonams, Erik's change and your change puts the responsibility on the
>>>> caller of ClassLoaderDataGraph to keep the class alive that it is
>>>> processing through the walk.   I think I can accept that as a fix,
>>>> and withdraw mine, since we seem to have no instances of bugs in
>>>> other walks.  If it turns out to be something that becomes buggy, I
>>>> will suggest my fix again as an alternative.
>>>>
>>>> This fix doesn't affect the walks done during GC.  As of this
>>>> morning, there are no KlassClosures in GC anymore.   Your concern is
>>>> well placed though because it could help us introduce bugs which
>>>> could negatively affect class unloading.
>>>>
>>>> So on the balance, there's risk in both versions of the fix, so I'll
>>>> withdraw mine.  I think the best fix is yours:
>>>>
>>>> http://cr.openjdk.java.net/~rkennke/8199612/webrev.00/src/hotspot/share/prims/jvmtiGetLoadedClasses.cpp.udiff.html
>>>>
>>>> with Erik's function java_mirror_phantom() that does the RootAccess load.
>>>>
>>>> http://cr.openjdk.java.net/~eosterlund/8197844/webrev.00/src/hotspot/share/oops/klass.cpp.udiff.html
>>>>
>>>> but not have it mark the class_loader, only do below.
>>>>
>>>> +oop Klass::java_mirror_phantom() {
>>>> + // Loading the klass_holder as a phantom oop ref keeps the class alive.
>>>> + // After the holder has been kept alive, it is safe to return the
>>>> mirror.
>>>> + oop mirror =
>>>> RootAccess<ON_PHANTOM_OOP_REF>::oop_load(k->java_mirror_handle().ptr_raw());
>>>> + return mirror;
>>>> +}
>>>>
>>>> I didn't like that Klass was a friend of ClassLoaderData and could
>>>> get to the class_loader field (in case I move it).
>>>>
>>>> Your change here is covered by Kim's work to accessorize jni.
>>>>
>>>> http://cr.openjdk.java.net/~rkennke/8199612/webrev.00/src/hotspot/share/runtime/jniHandles.cpp.udiff.html
>>>>
>>>> So I withdraw and I'll unassign myself and you and/or Erik can take
>>>> this bug back again.
>>>>
>>>> Thanks!
>>>> Coleen
>>>>
>>>>
>>>>> Roman
>>>>>
>