SA for ZGC
Yasumasa Suenaga
yasuenag at gmail.com
Tue Jan 16 10:57:40 UTC 2018
Hi Stefan,
I've restored your change and use VM.getOS() and getCPU() in new webrev:
http://cr.openjdk.java.net/~ysuenaga/z/sa-universe/webrev.02/
I checked this patch with -XX:+ZUnmapBadViews on my Linux x64, it works fine.
Thanks,
Yasumasa
2018-01-16 18:26 GMT+09:00 Stefan Karlsson <stefan.karlsson at oracle.com>:
> Hi Yasumas,
>
> On 2018-01-16 08:14, Yasumasa Suenaga wrote:
>>
>> Hi Stefan,
>>
>>>> Stefan's patch seems to have resolved JDK-8194737.
>>>> Should we resolve this JDK-8194737 in upstream repo and merge it to zgc
>>>> at
>>>> first?
>>>
>>>
>>>
>>> I'll let Jini handle that bug.
>>
>>
>> Thanks!
>> I watch this bug.
>>
>>
>>> On SPARC we add a heap base when converting from "offsets" to "objects".
>>> See:
>>>
>>> http://hg.openjdk.java.net/zgc/zgc/file/89d447cff6c5/src/hotspot/os_cpu/solaris_sparc/zAddress_solaris_sparc.inline.hpp
>>
>>
>> I think we can use getOS() and getCPU() in VM class:
>>
>> http://hg.openjdk.java.net/zgc/zgc/file/89d447cff6c5/src/jdk.hotspot.agent/share/classes/sun/jvm/hotspot/runtime/VM.java#l511
>>
>>
>>> My first version of of ZCollectedHeap.oop_load_at passed an OopHandle to
>>> ZBarrier.weak_barrier, but I changed that, so my implementation of
>>> ZOop.to_address isn't needed there. However, it is still needed for
>>> ZCollectedHeap.oopAddressDescription:
>>>
>>> public String oopAddressDescription(OopHandle handle) {
>>> Address origOop = ZOop.to_address(handle);
>>> Address loadBarrieredOop = ZBarrier.weak_barrier(origOop);
>>>
>>> We either need to restore the previous implementation, or we need to fix
>>> oopAddressDescription.
>>
>>
>> OK, I will keep your original ZOop.java and related implementation.
>
>
> OK. We can revisit this later if we do decide to clean this up.
>
>>
>>
>>>> I added these values to constant values in VMStructs.
>>>> Thus I removed ZGlobalsForVMStructs.
>>>
>>>
>>>
>>>
>>> This is a bug, and causes my test to fail.
>>>
>>> The constants values in VMStructs are for constants in the C++ code. Some
>>> of
>>> the values you changed are initialized during the VM initializations, and
>>> some are changed throughout the run of the VM.
>>
>>
>> I guess we can use HotSpotVMAddress as JVMCI:
>>
>> http://hg.openjdk.java.net/zgc/zgc/file/89d447cff6c5/src/hotspot/share/jvmci/vmStructs_jvmci.cpp#l937
>>
>> However this symbol is not exported.
>> We need to add `gHotSpotVMAddress` to vmStructs.hpp .
>>
>> BTW, do ZGC team have a plan to refactor zGlobals.hpp?
>
>
>
> Currently, we don't have any plans to do that.
>
>> AFAICS HotSpot does not have global variables (not `const` value).
>> IMHO we can refactor zGlobals to the C++ class which extends AllStatic.
>> We can use VMStructs directly if we do so.
>
>
> HotSpot actually does have global variables. You'll find some of them in
> globalDefinitions.hpp. For example: heapOopSize, MinObjAlignment, and
> OopEncodingHeapMax.
>
> The SA seems to duplicate the calculations of these values instead of
> reading the actual values from the process/core.
>
> I'm a bit reluctant to change the ZGC code style just to adapt to the SA
> agent. Currently, I think it's fine to have an extra layer in vmStructs_z
> until the SA agent gets support for global variables.
>
> Thanks,
> StefanK
>
>
>>
>>
>>> Try to run with -XX:+ZUnmapBadViews. You will find any stray pointers
>>> faster
>>> with that flag.
>>
>>
>> Thanks!
>> After all concerns in above is resolved, I will test it.
>>
>>
>> Yasumasa
>>
>>
>> 2018-01-15 17:48 GMT+09:00 Stefan Karlsson <stefan.karlsson at oracle.com>:
>>>
>>> Hi Yasumasa,
>>>
>>> On 2018-01-13 04:39, Yasumasa Suenaga wrote:
>>>>
>>>>
>>>> Hi Stefan, Per,
>>>>
>>>> Thank you for your comments and patches!
>>>> I uploaded new webrev:
>>>>
>>>> all:
>>>> http://cr.openjdk.java.net/~ysuenaga/z/sa-universe/webrev.01/all/
>>>>
>>>> diff from Stefan's patch:
>>>> http://cr.openjdk.java.net/~ysuenaga/z/sa-universe/webrev.01/diff/
>>>>
>>>>
>>>> Stefan's patch seems to have resolved JDK-8194737.
>>>> Should we resolve this JDK-8194737 in upstream repo and merge it to zgc
>>>> at
>>>> first?
>>>
>>>
>>>
>>> I'll let Jini handle that bug.
>>>
>>>>
>>>>
>>>>> Just one comment, in ZPage.java:
>>>>>
>>>>> 91 // There's no relocate operation in the SA. Simply use
>>>>> the
>>>>> from address and hope for the best.
>>>>> 92 return from;
>>>>>
>>>>> we should return ZAddress.good(from) here. And then we can remove the
>>>>> last part of the comment. Maybe say something about in-place forwarding
>>>>> instead.
>>>>
>>>>
>>>>
>>>> I've fixed.
>>>>
>>>>
>>>>>> 1) Some of the constants are Linux-specific and there's no platform
>>>>>> abstraction in place.
>>>>
>>>>
>>>>
>>>> I added some constant values to vmStructs_z.hpp . So we can remove
>>>> Linux-specific values from SA Java sources.
>>>
>>>
>>>
>>> OK. Sounds like the right thing to do for true constants. More about this
>>> later.
>>>
>>>> But I've not remove address() from ZAddress.java because I'm not sure
>>>> why
>>>> this method is needed.
>>>
>>>
>>>
>>> On SPARC we add a heap base when converting from "offsets" to "objects".
>>> See:
>>>
>>> http://hg.openjdk.java.net/zgc/zgc/file/89d447cff6c5/src/hotspot/os_cpu/solaris_sparc/zAddress_solaris_sparc.inline.hpp
>>>
>>>
>>>>
>>>>
>>>>>> 2) ZOop.to_address subverts the type system so that we can get from
>>>>>> OopHandles and longs to Addresses. This seems to be frowned upon if
>>>>>> you read
>>>>>> the SA code. Maybe there are more correct ways to implement this?
>>>>
>>>>
>>>>
>>>> I changed ZOop#to_address() which have one argument which type is
>>>> Address.
>>>> I guess this change is type-safe.
>>>
>>>
>>>
>>> My version of ZOop.to_address converted a OopHandle into an non-OopHandle
>>> Address. Without it you get an UnsupportedOperationException when you try
>>> to
>>> do bit-manipulation on the OopHandle. From LinuxOopHandle:
>>> public Address addOffsetTo (long offset) throws
>>> UnsupportedOperationException {
>>> throw new UnsupportedOperationException("addOffsetTo not applicable
>>> to
>>> OopHandles (interior object pointers not allowed)");
>>> }
>>>
>>>
>>> My first version of of ZCollectedHeap.oop_load_at passed an OopHandle to
>>> ZBarrier.weak_barrier, but I changed that, so my implementation of
>>> ZOop.to_address isn't needed there. However, it is still needed for
>>> ZCollectedHeap.oopAddressDescription:
>>>
>>> public String oopAddressDescription(OopHandle handle) {
>>> Address origOop = ZOop.to_address(handle);
>>> Address loadBarrieredOop = ZBarrier.weak_barrier(origOop);
>>>
>>> We either need to restore the previous implementation, or we need to fix
>>> oopAddressDescription.
>>>
>>> My version of ZOop.to_address performed bit manipulations just to ensure
>>> that we never got a all-zero value. You have left the
>>> andWithMask(REMOVE_MSB_MASK), which isn't necessary when you didn't add
>>> the
>>> bit in the first place.
>>>
>>>
>>>>
>>>>
>>>>>> 4) I see that you used VMObjectFactory.newObject, I skipped that part,
>>>>>> but we can think about unifying our patches.
>>>>
>>>>
>>>>
>>>> I think the classes which extends VMObject should be instantiated via
>>>> VMObjectFactory#newObject() because other SA codes seems to be so.
>>>> I changed to use this factory method as possible.
>>>
>>>
>>>
>>> OK.
>>>
>>>>
>>>>
>>>>>> 5) I had to invent a way to read global variables from HotSpot. See
>>>>>> the
>>>>>> usage of ZGlobalsForVMStructs. Maybe there's a more correct way to do
>>>>>> this?
>>>>
>>>>
>>>>
>>>> I added these values to constant values in VMStructs.
>>>> Thus I removed ZGlobalsForVMStructs.
>>>
>>>
>>>
>>>
>>> This is a bug, and causes my test to fail.
>>>
>>> The constants values in VMStructs are for constants in the C++ code. Some
>>> of
>>> the values you changed are initialized during the VM initializations, and
>>> some are changed throughout the run of the VM.
>>>
>>>>
>>>>
>>>>>> I stress tested the thread dumping part that threw exceptions for me
>>>>>> earlier by running SPECjbb2005:
>>>>>> $ java -XX:+UseZGC -Xlog:gc -cp jbb.jar:check.jar spec.jbb.JBBmain
>>>>>>
>>>>>> and then continuously dumping the threads:
>>>>>> $ while true; do echo -e "verbose true\nthreads" | jhsdb -J-ea clhsdb
>>>>>> --pid=`jps | grep JBBmain | awk '{print $1}'`; done
>>>>
>>>>
>>>>
>>>> Sorry, I do not have SPECjbb.
>>>> Can you share the result?
>>>
>>>
>>>
>>>
>>> You can use any program that starts a number of threads an continuously
>>> run
>>> GCs.
>>>
>>>>
>>>> I tested `universe` and `jstack` and `jstack -v` on CLHSDB with simple
>>>> Java app.
>>>> This change works fine on my Linux x64 box.
>>>
>>>
>>>
>>> Try to run with -XX:+ZUnmapBadViews. You will find any stray pointers
>>> faster
>>> with that flag.
>>>
>>> This reproduces the problem:
>>>
>>> public class HelloSleep {
>>> public static void main(String... args) throws Exception {
>>> Thread.sleep(100000000000000L);
>>> }
>>> }
>>>
>>> DIR=~/hg/z/build/release/jdk
>>> $DIR/bin/java -XX:+ZUnmapBadViews -XX:+UseZGC HelloSleep &
>>> $DIR/bin/jcmd $! GC.run
>>> echo -e "verbose true\nthreads" | $DIR/bin/jhsdb -J-ea hsdb --pid=$
>>>
>>> Thanks,
>>> StefanK
>>>
>>>
>>>>
>>>>
>>>> Thanks,
>>>>
>>>> Yasumasa
>>>>
>>>>
>>>> On 2018/01/11 21:40, Per Liden wrote:
>>>>>
>>>>>
>>>>> Hi,
>>>>>
>>>>> On 01/10/2018 05:55 PM, Stefan Karlsson wrote:
>>>>>>
>>>>>>
>>>>>> Hi Yasumasa,
>>>>>>
>>>>>> Here's my first stab at adding the ZGC load barries to the SA:
>>>>>>
>>>>>> http://cr.openjdk.java.net/~stefank/zgc/zSALoadBarriers/webrev.01/
>>>>>
>>>>>
>>>>>
>>>>> Just one comment, in ZPage.java:
>>>>>
>>>>> 91 // There's no relocate operation in the SA. Simply use
>>>>> the
>>>>> from address and hope for the best.
>>>>> 92 return from;
>>>>>
>>>>> we should return ZAddress.good(from) here. And then we can remove the
>>>>> last part of the comment. Maybe say something about in-place forwarding
>>>>> instead.
>>>>>
>>>>> cheers,
>>>>> Per
>>>>>
>>>>>>
>>>>>> I find it unfortunate that we have to duplicate the C++ code into
>>>>>> Java,
>>>>>> to get this to work. But it is what it is.
>>>>>>
>>>>>> Some comments about the code:
>>>>>>
>>>>>> 1) Some of the constants are Linux-specific and there's no platform
>>>>>> abstraction in place.
>>>>>>
>>>>>> 2) ZOop.to_address subverts the type system so that we can get from
>>>>>> OopHandles and longs to Addresses. This seems to be frowned upon if
>>>>>> you read
>>>>>> the SA code. Maybe there are more correct ways to implement this?
>>>>>>
>>>>>> 3) I use the ZGC weak barrier, since the SA agent only needs to be
>>>>>> able
>>>>>> to access the Objects, not mark them. Which is what the weak barriers
>>>>>> provide.
>>>>>>
>>>>>> 4) I see that you used VMObjectFactory.newObject, I skipped that part,
>>>>>> but we can think about unifying our patches.
>>>>>>
>>>>>> 5) I had to invent a way to read global variables from HotSpot. See
>>>>>> the
>>>>>> usage of ZGlobalsForVMStructs. Maybe there's a more correct way to do
>>>>>> this?
>>>>>>
>>>>>> I stress tested the thread dumping part that threw exceptions for me
>>>>>> earlier by running SPECjbb2005:
>>>>>> $ java -XX:+UseZGC -Xlog:gc -cp jbb.jar:check.jar spec.jbb.JBBmain
>>>>>>
>>>>>> and then continuously dumping the threads:
>>>>>> $ while true; do echo -e "verbose true\nthreads" | jhsdb -J-ea clhsdb
>>>>>> --pid=`jps | grep JBBmain | awk '{print $1}'`; done
>>>>>>
>>>>>>
>>>>>> Cheers,
>>>>>> StefanK
>>>>>>
>>>>>> On 2018-01-09 13:59, Yasumasa Suenaga wrote:
>>>>>>>
>>>>>>>
>>>>>>> Hi Stefan,
>>>>>>>
>>>>>>>
>>>>>>>> Let me try to create an initial prototype for load barriers in the
>>>>>>>> SA
>>>>>>>> case and see where this takes us.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Sure!
>>>>>>>
>>>>>>>
>>>>>>>> FYI i also found that compressed oops are not fully supported in the
>>>>>>>> SA:
>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8194737
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Thanks, I didn't know that.
>>>>>>> I hope this issue will be fixed soon :-)
>>>>>>>
>>>>>>>
>>>>>>> Yasumasa
>>>>>>>
>>>>>>>
>>>>>>> On 2018/01/09 18:36, Stefan Karlsson wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>> Hi Yasumasa,
>>>>>>>>
>>>>>>>> Thanks for your contribution!
>>>>>>>>
>>>>>>>> I tested your patch together with jhsdb hsdb and immediately hit
>>>>>>>> problems because of the lack of load barriers in the SA code. Let me
>>>>>>>> try to
>>>>>>>> create an initial prototype for load barriers in the SA case and see
>>>>>>>> where
>>>>>>>> this takes us.
>>>>>>>>
>>>>>>>> FYI i also found that compressed oops are not fully supported in the
>>>>>>>> SA:
>>>>>>>> https://bugs.openjdk.java.net/browse/JDK-8194737
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>> StefanK
>>>>>>>>
>>>>>>>> On 2017-12-21 23:38, Yasumasa Suenaga wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Thanks Per!
>>>>>>>>>
>>>>>>>>> I'm waiting for comments and sponsorship.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Yasumasa
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On 2017/12/22 4:33, Per Liden wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Hi Yasumasa,
>>>>>>>>>>
>>>>>>>>>> On 2017-12-21 09:40, Yasumasa Suenaga wrote:
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Hi all,
>>>>>>>>>>>
>>>>>>>>>>> I checked SA implementation for ZGC, but it is not yet available.
>>>>>>>>>>> For example, we can see WrongTypeException when CLHSDB `universe`
>>>>>>>>>>> command is executed.
>>>>>>>>>>>
>>>>>>>>>>> As the first step, I propose to implement ZCollectedHeap and
>>>>>>>>>>> related classes for
>>>>>>>>>>> SA as this webrev:
>>>>>>>>>>>
>>>>>>>>>>> http://cr.openjdk.java.net/~ysuenaga/z/sa-universe/
>>>>>>>>>>>
>>>>>>>>>>> After applying this patch, we can use `universe` command on
>>>>>>>>>>> CLHSDB.
>>>>>>>>>>> (I followed `VM.info` jcmd output format)
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Thanks! Most of us working on ZGC are currently on vacation, but
>>>>>>>>>> we'll have a closer look at the patch after the holidays.
>>>>>>>>>>
>>>>>>>>>> cheers,
>>>>>>>>>> Per
>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Of course, it is not all. We need to work more for it.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> Thanks,
>>>>>>>>>>>
>>>>>>>>>>> Yasumasa
>>>>>>>>>>>
>>>
>
More information about the zgc-dev
mailing list