RFR: 8078644: CDS needs to support JVMTI CFLH
Daniel D. Daugherty
daniel.daugherty at oracle.com
Wed Sep 14 02:00:32 UTC 2016
On 9/13/16 12:02 PM, Jiangli Zhou wrote:
> Hi Karen,
>
> Thank you!
>
> Just in case if you want to double check, here is the updated webrev with the add_package() fix for boot loader.
>
> http://cr.openjdk.java.net/~jiangli/8078644/webrev.02/
src/share/vm/classfile/klassFactory.cpp
No comments.
src/share/vm/classfile/klassFactory.hpp
No comments.
src/share/vm/classfile/systemDictionary.cpp
No comments.
src/share/vm/memory/filemap.cpp
No comments.
src/share/vm/memory/filemap.hpp
No comments.
src/share/vm/memory/metaspace.cpp
No comments.
src/share/vm/memory/metaspaceShared.cpp
No comments.
src/share/vm/memory/metaspaceShared.hpp
No comments.
src/share/vm/oops/instanceKlass.cpp
L2119: // deallocate the cached class file
L2120: if (_cached_class_file != NULL) {
L2121: os::free(_cached_class_file);
L2122: _cached_class_file = NULL;
L2123: }
Sorry, this free() call still bothers me and now I think I
see why. This getter:
L3652: JvmtiCachedClassFileData*
InstanceKlass::get_cached_class_file() {
L3653: if
(MetaspaceShared::is_in_shared_space(_cached_class_file)) {
L3654: // Ignore the archived class stream data
L3655: return NULL;
L3656: } else {
shows that the _cached_class_file field can refer to a shared
class and when it does the getter returns NULL. Yes, this is
the deallocation code path and we should never get here for a
shared instanceKlass. Perhaps we need an assert at the top of
function:
assert(!this->is_shared(), "should not be called for a shared
class");
Feel free to say that I'm being way too paranoid here. :-)
L3671: assert(this->is_shared(), "class is not shared");
Perhaps: "class should be shared" instead.
src/share/vm/oops/instanceKlass.hpp
No comments.
src/share/vm/utilities/debug.cpp
No comments.
src/share/vm/utilities/debug.hpp
No comments.
Thumbs up. Don't need to see a new webrev if you take any of my
minor suggestions above. If you do add the new assert() to the
deallocation code path, you'll need a new round of testing.
Dan
>
> The KlassFactory::check_shared_class_file_load_hook() is updated to call add_package() for the boot loader. It is done right before running the ‘new_ik’ to make sure there is no other loading failure once the package is added. No other change was made to the webrev.
>
> if (class_loader.is_null()) {
> ResourceMark rm;
> ClassLoader::add_package(class_name->as_C_string(), path_index, THREAD);
> }
>
> On the test side for this particular issue, I modified the new TransformSuperAndSubClasses test and verified the correct system packages were obtained by Package.getPackages(). I’ll work with Misha to incorporate the test update.
>
> Thanks,
> Jiangli
>
>> On Sep 13, 2016, at 7:01 AM, Karen Kinnear <karen.kinnear at oracle.com> wrote:
>>
>> Jiangli,
>>
>> I finished my review and found the answers to my other questions - so all set!
>>
>> thanks,
>> Karen
>>
>>> On Sep 12, 2016, at 7:26 PM, Jiangli Zhou <jiangli.zhou at oracle.com <mailto:jiangli.zhou at oracle.com>> wrote:
>>>
>>> Hi Karen,
>>>
>>> Thank you so much for the review!
>>>
>>>> On Sep 12, 2016, at 1:58 PM, Karen Kinnear <karen.kinnear at oracle.com <mailto:karen.kinnear at oracle.com>> wrote:
>>>>
>>>> Jiangli,
>>>>
>>>> This looks great! Many thanks. Very creative solution!
>>>>
>>>> Question:
>>>>
>>>> If you load_shared_class for the boot loader and you use the CFLH such that you
>>>> return the new_ik, where do you call ClassLoader::add_package?
>>>> (I did find a call to notify_class_loaded from fill_instance_klass which is called by create_instance_klass)
>>> Good catch!! In ClassLoader::load_class(), the add_packege() is done through ClassLoaderExt::Context::record_result(). For shared classes, SystemDictionary::load_shared_class() calls add_package() explicitly for boot loader:
>>>
>>> // For boot loader, ensure that GetSystemPackage knows that a class in this
>>> // package was loaded.
>>> if (class_loader.is_null()) {
>>> int path_index = ik->shared_classpath_index();
>>> ResourceMark rm;
>>> ClassLoader::add_package(ik->name()->as_C_string(), path_index, THREAD);
>>> }
>>>
>>> We need to do that before returning the transformed class. Will fix.
>>>
>>> Thanks!
>>>
>>> Jiangli
>>>
>>>
>>>
>>>> thanks,
>>>> Karen
>>>>
>>>> p.s. I may have a couple more questions - I will try to get them to you tomorrow. Thanks.
>>>>
>>>>> On Sep 9, 2016, at 11:42 PM, Jiangli Zhou <jiangli.zhou at oracle.com <mailto:jiangli.zhou at oracle.com>> wrote:
>>>>>
>>>>> Here is the updated webrev.
>>>>>
>>>>> http://cr.openjdk.java.net/~jiangli/8078644/webrev.01/ <http://cr.openjdk.java.net/~jiangli/8078644/webrev.01/>
>>>>>
>>>>> Thanks everyone again for the timely feedbacks!
>>>>>
>>>>> Jiangli
>>>>>
>>>>>> On Sep 9, 2016, at 5:53 PM, Jiangli Zhou <jiangli.zhou at Oracle.COM <mailto:jiangli.zhou at oracle.com>> wrote:
>>>>>>
>>>>>> Hi Dan,
>>>>>>
>>>>>> Thanks for the review!
>>>>>>
>>>>>>> On Sep 9, 2016, at 3:40 PM, Daniel D. Daugherty <daniel.daugherty at oracle.com <mailto:daniel.daugherty at oracle.com>> wrote:
>>>>>>>
>>>>>>> On 9/6/16 6:14 PM, Jiangli Zhou wrote:
>>>>>>>> Hi,
>>>>>>>>
>>>>>>>> Please review the following change thats support JVMTI class_file_load_hook(CFLH) during initial loading of shared classes. The webrev also removes the temporary workaround that disables CDS when JVMTI CFLH is required (JDK-8141341 <https://bugs.openjdk.java.net/browse/JDK-8141341 <https://bugs.openjdk.java.net/browse/JDK-8141341>>).
>>>>>>>>
>>>>>>>> webrev:http://cr.openjdk.java.net/~jiangli/8078644/webrev.00/ <http://cr.openjdk.java.net/~jiangli/8078644/webrev.00/>
>>>>>>> General comment
>>>>>>>
>>>>>>> - please make sure all copyright years are updated.
>>>>>>> - there are some jcheck space at end-of-line violations
>>>>>>> (There are more than what I flagged in the comments below).
>>>>>> Will make sure to fix those before commit.
>>>>>>
>>>>>>> - I think there might be a data leak with the _cached_class_file
>>>>>>> field, but that pre-dates your changes and I have to think about it.
>>>>>>>
>>>>>>>
>>>>>>> src/share/vm/oops/instanceKlass.hpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/oops/instanceKlass.cpp
>>>>>>> L2119: // deallocate the cached class file
>>>>>>> L2120: if (_cached_class_file != NULL) {
>>>>>>> L2121: os::free(_cached_class_file);
>>>>>>> L2122: _cached_class_file = NULL;
>>>>>>> L2123: }
>>>>>>> Doesn't this use of _cached_class_file need to be checked
>>>>>>> against the shared spaces? I don't think os::free() will
>>>>>>> be happy if _cached_class_file points into shared space.
>>>>>> That’s a good question. Only shared InstanceKlass might have _cached_class_file pointing to shared space. The deallocation call path never happen for shared classes at runtime. So if we get here, we must not be dealing with a shared class and _cached_class_file is not shared. Otherwise we would see issues for other data being freed in InstanceKlass::release_C_heap_structures().
>>>>>>
>>>>>>> src/share/vm/memory/metaspaceShared.hpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/memory/metaspaceShared.cpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/memory/metaspace.cpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/classfile/klassFactory.hpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/classfile/klassFactory.cpp
>>>>>>> L219: if ((result->get_cached_class_file()) != NULL) {
>>>>>>> L220: // JFR might set cached_class_file
>>>>>>> L221: len = result->get_cached_class_file_len();
>>>>>>> L222: bytes = result->get_cached_class_file_bytes();
>>>>>>> L223: } else {
>>>>>>> Perhaps:
>>>>>>> if ((bytes = result->get_cached_class_file()) != NULL) {
>>>>>>> // event based tracing might set cached_class_file
>>>>>>> len = result->get_cached_class_file_len();
>>>>>>> } else {
>>>>>>>
>>>>>>> so two changes:
>>>>>>> - mv init of bytes into the if-statement
>>>>>>> - change 'JFR' -> 'event based tracing’
>>>>>> Updated.
>>>>>>
>>>>>>> src/share/vm/classfile/systemDictionary.cpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/memory/filemap.hpp
>>>>>>> L260: }
>>>>>>> L261: // The estimated optional space size. Only the portion containning data is
>>>>>>>
>>>>>>> Please add a blank line between L260 and L261.
>>>>>>>
>>>>>>> Typo: 'containning' -> ‘containing'
>>>>>> Fixed.
>>>>>>
>>>>>>> There's a trailing space on L261; jcheck won't like it.
>>>>>> Fixed.
>>>>>>
>>>>>>> L266: }
>>>>>>> L267: // Total shared_spaces size includes the ro, rw, md, mc and od spaces
>>>>>>>
>>>>>>> Please add a blank line between L266 and L267.
>>>>>> Done.
>>>>>>
>>>>>>> L264: static size_t optional_space_size() {
>>>>>>> L265: return core_spaces_size();
>>>>>>>
>>>>>>> It is not clear why core_spaces_size() is being returned as
>>>>>>> an estimate of the optional space size.
>>>>>> Coleen also pointed it out. I added more details.
>>>>>>
>>>>>>> There's a trailing space on L265; jcheck won't like it.
>>>>>> Fixed.
>>>>>>
>>>>>>> src/share/vm/memory/filemap.cpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/utilities/debug.hpp
>>>>>>> No comments.
>>>>>>>
>>>>>>> src/share/vm/utilities/debug.cpp
>>>>>>> L289: "shared miscellaneous code space",
>>>>>>> L290: };
>>>>>>> The name string for the new space is missing.
>>>>>>> Perhaps "shared optional space"…
>>>>>> Also fixed when incorporating Coleen’s feedbacks. The estimated shared space size should be large enough and never run out space. Added a check.
>>>>>>
>>>>>> Thanks!
>>>>>>
>>>>>> Jiangli
>>>>>>
>>>>>>> Dan
>>>>>>>
>>>>>>>
>>>>>>>> bug: JDK-8078644 <https://bugs.openjdk.java.net/browse/JDK-8078644 <https://bugs.openjdk.java.net/browse/JDK-8078644>>
>>>>>>>>
>>>>>>>> Class file data is needed when posting CFLH. For shared classes, the class file data are now archived at dump time. That avoids re-reading the class file data from the JAR file or reconstituting the data at runtime, which would add class loading overhead for shared classes.
>>>>>>>>
>>>>>>>> To minimize the runtime memory overhead, the archived class file data are stored in a separate shared space, called ‘optional data space’ (od). The ‘od’ space a new region in the archive. It does not increase runtime memory usage when JvmtiExport::should_post_class_file_load_hook() is false. The memory contains the archived class file data is only paged in when the VM needs to post CFLH. The ‘od’ space can be shared by different processes.
>>>>>>>>
>>>>>>>> When loading shared class at runtime, we now call JvmtiExport::post_class_file_load_hook() with the archive class file data if needed. If JVMTI agent modifies the class, new InstanceKlass is created using the modified class data and the archived class is ignored.
>>>>>>>>
>>>>>>>> Tested with JPRT, CDS tests and QA tests.
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>> Jiangli
More information about the hotspot-runtime-dev
mailing list