On Wed, Aug 19, 2020 at 6:17 PM Hohensee, Paul <[email protected]> wrote: > > Please review this backport to jdk8u. I especially need a CSR review, since > the CSR approval process can be a bottleneck. The patch significantly reduces > fleet profiling overhead, and a version of it has been in production at > Amazon for over 3 years. > > > > Original JBS issue: https://bugs.openjdk.java.net/browse/JDK-8185003 > > Original CSR: https://bugs.openjdk.java.net/browse/JDK-8185705 > > Original patch: http://hg.openjdk.java.net/jdk10/master/rev/68d46cb9be45 > > > > Backport JBS issue: https://bugs.openjdk.java.net/browse/JDK-8251494 > > Backport CSR: https://bugs.openjdk.java.net/browse/JDK-8251498 > > Backport JDK webrev: http://cr.openjdk.java.net/~phh/8185003/webrev.8u.jdk.05/ >
JDK part looks good to me. > Backport Hotspot webrev: > http://cr.openjdk.java.net/~phh/8185003/webrev.8u.hotspot.05/ > HotSpot part looks good to me but see discussion below. > > > Details of the interface changes needed for the backport are in the > Description of the Backport CSR 8251498. The actual functional changes are > minimal and low risk. > I've also reviewed the CSR yesterday which I think is fine. But now, when looking at the implementation, I'm a little concerned about changing JMM_VERSION from " 0x20010203" to "0x20020000" in "jmm.h". This might be especially problematic in combination with the changes in "Management::get_jmm_interface()" which is called by JVM_GetManagement(): void* Management::get_jmm_interface(int version) { #if INCLUDE_MANAGEMENT - if (version == JMM_VERSION_1_0) { + if (version == JMM_VERSION) { return (void*) &jmm_interface; } #endif // INCLUDE_MANAGEMENT return NULL; } You've correctly fixed the single caller of "JVM_GetManagement()" in the JDK (in "JNI_OnLoad()" in "management.c"): - jmm_interface = (JmmInterface*) JVM_GetManagement(JMM_VERSION_1_0); + jmm_interface = (JmmInterface*) JVM_GetManagement(JMM_VERSION); but I wonder if there are other monitoring/serviceability tools out there which use this interface and which will break after this change. A quick search revealed at least two StackOverflow entries which recommend using "JVM_GetManagement(JMM_VERSION_1_0)" [1,2] and there's a talk and a blog entry doing the same [3, 4]. I'm not sure how relevant this is but I think a much safer and backwards-compatible way of doing this downport would be the following: - don't change "Management::get_jmm_interface()" (i.e. still check for "JMM_VERSION_1_0") but return the "new" JMM_VERSION in "jmm_GetVersion()". This won't break anything but will make it possible for clients to detect the new version if they want. - don't change the signature of "DumpThreads()". Instead add a new version (e.g. "DumpThreadsMaxDepth()/jmm_DumpThreadsMaxDepth()") to the "JMMInterface" struct and to "jmm_interface" in "management.cpp". You can do this in one of the two first, reserved fields of "JMMInterface" so you won't break binary compatibility. "jmm_DumpThreads()" will then be a simple wrapper which calls "jmm_DumpThreadsMaxDepth()" with Integer.MAX_VALUE as depth. - in the jdk you then simply call "DumpThreadsMaxDepth()" in "Java_sun_management_ThreadImpl_dumpThreads0()" I think this way we can maintain full binary compatibility while still using the new feature. What do you think? Best regards, Volker [1] https://stackoverflow.com/questions/23632653/generate-java-heap-dump-on-uncaught-exception [2] https://stackoverflow.com/questions/60887816/jvm-options-printnmtstatistics-save-info-to-file [3] https://sudonull.com/post/25841-JVM-TI-how-to-make-a-plugin-for-a-virtual-machine-Odnoklassniki-company-blog [4] https://2019.jpoint.ru/talks/2o8scc5jbaurnqqlsydzxv/ > > Passes the included (suitably modified) test, as well as the tests in > > > > jdk/test/java/lang/management/ThreadMXBean > > jdk/test/com/sun/management/ThreadMXBean > > > > Thanks, > > Paul
