July 2025 - hotspot-jfr-dev

Re: RFR: 8356587: Missing object ID X in pool jdk.types.Method [v5]
by Markus Grönlund 28 Jul '25

28 Jul '25

> Greetings, > > The following change set addresses the data loss resulting in the assertion "Missing object ID X in pool jdk.types.Method". > > It involves three components: > > 1. Address a regression introduced by [JDK-835221](https://bugs.openjdk.org/browse/JDK-8352251). By locating JFR_ONLY(Jfr::check_and_process_sample_request(thread);) before the global_poll() in SafepointMechanism::process(), a stacktrace can be captured, and artifacts tagged, during a safepoint. This breaks an invariant as artifacts, i.e. methods, can be tagged in the wrong epoch (also a stacktrace can be stored in the wrong epoch). Must be moved to post global_poll(). > > 2. Retransform/Redefine classes include a non-safe copy of Method trace flags, leading to stale bits being set onto the new Methods. Method trace flags need to be copied, but must be done under a safepoint. > > 3. It seems that there has been an increase in the frequency of issuing calls to InstanceKlass::purge_previous_versions(), making scratch klasses and old methods disappear before JFR gets the chance to serialize also tagged old methods. Therefore, we need to ensure that we always tag the latest version of a Method. > > There is also a cleanup of gratuitous type conversions, from Klass* to InstanceKlass* in the newly introduced MethodTracing subsystem. > > Testing: jdk_jfr, stress testing > > Thank you > Markus Markus Grönlund has updated the pull request incrementally with one additional commit since the last revision: removed unused ------------- Changes: - all: https://git.openjdk.org/jdk/pull/26458/files - new: https://git.openjdk.org/jdk/pull/26458/files/e68e9f3c..6504e1b7 Webrevs: - full: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=04 - incr: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=03-04 Stats: 2 lines in 1 file changed: 0 ins; 2 del; 0 mod Patch: https://git.openjdk.org/jdk/pull/26458.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/26458/head:pull/26458 PR: https://git.openjdk.org/jdk/pull/26458

2 1

Re: RFR: 8356587: Missing object ID X in pool jdk.types.Method [v4]
by Markus Grönlund 28 Jul '25

28 Jul '25

> Greetings, > > The following change set addresses the data loss resulting in the assertion "Missing object ID X in pool jdk.types.Method". > > It involves three components: > > 1. Address a regression introduced by [JDK-835221](https://bugs.openjdk.org/browse/JDK-8352251). By locating JFR_ONLY(Jfr::check_and_process_sample_request(thread);) before the global_poll() in SafepointMechanism::process(), a stacktrace can be captured, and artifacts tagged, during a safepoint. This breaks an invariant as artifacts, i.e. methods, can be tagged in the wrong epoch (also a stacktrace can be stored in the wrong epoch). Must be moved to post global_poll(). > > 2. Retransform/Redefine classes include a non-safe copy of Method trace flags, leading to stale bits being set onto the new Methods. Method trace flags need to be copied, but must be done under a safepoint. > > 3. It seems that there has been an increase in the frequency of issuing calls to InstanceKlass::purge_previous_versions(), making scratch klasses and old methods disappear before JFR gets the chance to serialize also tagged old methods. Therefore, we need to ensure that we always tag the latest version of a Method. > > There is also a cleanup of gratuitous type conversions, from Klass* to InstanceKlass* in the newly introduced MethodTracing subsystem. > > Testing: jdk_jfr, stress testing > > Thank you > Markus Markus Grönlund has updated the pull request incrementally with one additional commit since the last revision: always use new_methods length ------------- Changes: - all: https://git.openjdk.org/jdk/pull/26458/files - new: https://git.openjdk.org/jdk/pull/26458/files/933a2261..e68e9f3c Webrevs: - full: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=03 - incr: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=02-03 Stats: 2 lines in 1 file changed: 0 ins; 1 del; 1 mod Patch: https://git.openjdk.org/jdk/pull/26458.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/26458/head:pull/26458 PR: https://git.openjdk.org/jdk/pull/26458

1 0

Re: RFR: 8356587: Missing object ID X in pool jdk.types.Method [v3]
by Markus Grönlund 28 Jul '25

28 Jul '25

> Greetings, > > The following change set addresses the data loss resulting in the assertion "Missing object ID X in pool jdk.types.Method". > > It involves three components: > > 1. Address a regression introduced by [JDK-835221](https://bugs.openjdk.org/browse/JDK-8352251). By locating JFR_ONLY(Jfr::check_and_process_sample_request(thread);) before the global_poll() in SafepointMechanism::process(), a stacktrace can be captured, and artifacts tagged, during a safepoint. This breaks an invariant as artifacts, i.e. methods, can be tagged in the wrong epoch (also a stacktrace can be stored in the wrong epoch). Must be moved to post global_poll(). > > 2. Retransform/Redefine classes include a non-safe copy of Method trace flags, leading to stale bits being set onto the new Methods. Method trace flags need to be copied, but must be done under a safepoint. > > 3. It seems that there has been an increase in the frequency of issuing calls to InstanceKlass::purge_previous_versions(), making scratch klasses and old methods disappear before JFR gets the chance to serialize also tagged old methods. Therefore, we need to ensure that we always tag the latest version of a Method. > > There is also a cleanup of gratuitous type conversions, from Klass* to InstanceKlass* in the newly introduced MethodTracing subsystem. > > Testing: jdk_jfr, stress testing > > Thank you > Markus Markus Grönlund has updated the pull request incrementally with one additional commit since the last revision: improved predicate ------------- Changes: - all: https://git.openjdk.org/jdk/pull/26458/files - new: https://git.openjdk.org/jdk/pull/26458/files/a4d6ecc3..933a2261 Webrevs: - full: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=02 - incr: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=01-02 Stats: 2 lines in 1 file changed: 0 ins; 1 del; 1 mod Patch: https://git.openjdk.org/jdk/pull/26458.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/26458/head:pull/26458 PR: https://git.openjdk.org/jdk/pull/26458

1 0

Re: RFR: 8364090: Dump JFR recording on CrashOnOutOfMemoryError [v2]
by Yasumasa Suenaga 27 Jul '25

27 Jul '25

> JFR emergency dump would happen when OOM was thrown. However it would not contain most recent `OldObjectSample` events emitted by LeakProfiler. > > I [reported this issue in past](https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2019-January/00038…, and it seems to be difficult to fix soon, and also JDK codebase has been updated in several years. It brings us to fix this issue easier than past. > > So I propose again to emit the events from LeakProfiler when OOM happened. > This change passed `jdk_jfr` tests on Linux x64 (excepts TestHeapSummaryEventPSParOld.java reported in [JDK-8364082](https://bugs.openjdk.org/browse/JDK-8364082)) > > Related email thread: https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2025-July/008007.html Yasumasa Suenaga has updated the pull request incrementally with one additional commit since the last revision: Rename is_oom to emit_old_object_samples ------------- Changes: - all: https://git.openjdk.org/jdk/pull/26468/files - new: https://git.openjdk.org/jdk/pull/26468/files/c3198690..ee27149e Webrevs: - full: https://webrevs.openjdk.org/?repo=jdk&pr=26468&range=01 - incr: https://webrevs.openjdk.org/?repo=jdk&pr=26468&range=00-01 Stats: 9 lines in 4 files changed: 0 ins; 0 del; 9 mod Patch: https://git.openjdk.org/jdk/pull/26468.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/26468/head:pull/26468 PR: https://git.openjdk.org/jdk/pull/26468

2 3

Re: RFR: 8364090: Dump JFR recording on CrashOnOutOfMemoryError
by Yasumasa Suenaga 26 Jul '25

26 Jul '25

On Fri, 25 Jul 2025 00:52:29 GMT, Yasumasa Suenaga <ysuenaga(a)openjdk.org> wrote: > JFR emergency dump would happen when OOM was thrown. However it would not contain most recent `OldObjectSample` events emitted by LeakProfiler. > > I [reported this issue in past](https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2019-January/00038…, and it seems to be difficult to fix soon, and also JDK codebase has been updated in several years. It brings us to fix this issue easier than past. > > So I propose again to emit the events from LeakProfiler when OOM happened. > This change passed `jdk_jfr` tests on Linux x64 (excepts TestHeapSummaryEventPSParOld.java reported in [JDK-8364082](https://bugs.openjdk.org/browse/JDK-8364082)) > > Related email thread: https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2025-July/008007.html I renamed a parameter `is_oom` to `emit_old_object_samples`. @egahlin Could you review again? ------------- PR Comment: https://git.openjdk.org/jdk/pull/26468#issuecomment-3120997144

1 0

Re: RFR: 8356587: Missing object ID X in pool jdk.types.Method [v2]
by Markus Grönlund 25 Jul '25

25 Jul '25

> Greetings, > > The following change set addresses the data loss resulting in the assertion "Missing object ID X in pool jdk.types.Method". > > It involves three components: > > 1. Address a regression introduced by [JDK-835221](https://bugs.openjdk.org/browse/JDK-8352251). By locating JFR_ONLY(Jfr::check_and_process_sample_request(thread);) before the global_poll() in SafepointMechanism::process(), a stacktrace can be captured, and artifacts tagged, during a safepoint. This breaks an invariant as artifacts, i.e. methods, can be tagged in the wrong epoch (also a stacktrace can be stored in the wrong epoch). Must be moved to post global_poll(). > > 2. Retransform/Redefine classes include a non-safe copy of Method trace flags, leading to stale bits being set onto the new Methods. Method trace flags need to be copied, but must be done under a safepoint. > > 3. It seems that there has been an increase in the frequency of issuing calls to InstanceKlass::purge_previous_versions(), making scratch klasses and old methods disappear before JFR gets the chance to serialize also tagged old methods. Therefore, we need to ensure that we always tag the latest version of a Method. > > There is also a cleanup of gratuitous type conversions, from Klass* to InstanceKlass* in the newly introduced MethodTracing subsystem. > > Testing: jdk_jfr, stress testing > > Thank you > Markus Markus Grönlund has updated the pull request incrementally with one additional commit since the last revision: Must support AllowRedefinitionToAddDeleteMethods ------------- Changes: - all: https://git.openjdk.org/jdk/pull/26458/files - new: https://git.openjdk.org/jdk/pull/26458/files/4b77421b..a4d6ecc3 Webrevs: - full: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=01 - incr: https://webrevs.openjdk.org/?repo=jdk&pr=26458&range=00-01 Stats: 25 lines in 4 files changed: 16 ins; 1 del; 8 mod Patch: https://git.openjdk.org/jdk/pull/26458.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/26458/head:pull/26458 PR: https://git.openjdk.org/jdk/pull/26458

1 0

Re: RFR: 8364090: Dump JFR recording on CrashOnOutOfMemoryError
by Erik Gahlin 25 Jul '25

25 Jul '25

On Fri, 25 Jul 2025 00:52:29 GMT, Yasumasa Suenaga <ysuenaga(a)openjdk.org> wrote: > JFR emergency dump would happen when OOM was thrown. However it would not contain most recent `OldObjectSample` events emitted by LeakProfiler. > > I [reported this issue in past](https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2019-January/00038…, and it seems to be difficult to fix soon, and also JDK codebase has been updated in several years. It brings us to fix this issue easier than past. > > So I propose again to emit the events from LeakProfiler when OOM happened. > This change passed `jdk_jfr` tests on Linux x64 (excepts TestHeapSummaryEventPSParOld.java reported in [JDK-8364082](https://bugs.openjdk.org/browse/JDK-8364082)) > > Related email thread: https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2025-July/008007.html src/hotspot/share/runtime/java.cpp line 466: > 464: } > 465: > 466: JFR_ONLY(Jfr::on_vm_shutdown(true, halt);) You set oom to true here, why? ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/26468#discussion_r2230419356

2 4

JFR emergency dump didn't happen at OOM
by Yasumasa Suenaga 25 Jul '25

25 Jul '25

Hi all, I reported and proposed the fix about 6 years ago (!), I faced the issue again that would be helpful if this is fixed. 6 years ago, we discussed about this issue in [1], then I hear JDK-8196050 might fix this issue, but it hasn't yet happened unfortunately... JDK codebase has been made a lot of changes in 6 years, so I think we can fix simply than I proposed before. All of changes is [2]. It is not so big change and it is good because no flags are added in .jfc files. What do you think? I will send PR if it is ok, but I want to hear your opinions before PR (because we discussed before). For example, we can pass a flag for OOM rather than exception handler as following: ``` --- a/src/hotspot/share/jfr/recorder/repository/jfrEmergencyDump.cpp +++ b/src/hotspot/share/jfr/recorder/repository/jfrEmergencyDump.cpp @@ -556,22 +556,21 @@ class JavaThreadInVMAndNative : public StackObj { } }; -static void post_events(bool exception_handler, Thread* thread) { - if (exception_handler) { +static void post_events(bool is_oom, Thread* thread) { + if (is_oom) { + LeakProfiler::emit_events(max_jlong, false, false); + } else { EventShutdown e; e.set_reason("VM Error"); e.commit(); - } else { - // OOM - LeakProfiler::emit_events(max_jlong, false, false); } EventDumpReason event; - event.set_reason(exception_handler ? "Crash" : "Out of Memory"); + event.set_reason(is_oom ? "Out of Memory" : "Crash"); event.set_recordingId(-1); event.commit(); } ``` And also we can identify whether OOM or not (crash) via `id` in argument of `report_and_die`: ``` --- a/src/hotspot/share/utilities/vmError.cpp +++ b/src/hotspot/share/utilities/vmError.cpp @@ -1856,7 +1856,7 @@ void VMError::report_and_die(int id, const char* message, const char* detail_fmt log.set_fd(-1); } - JFR_ONLY(Jfr::on_vm_shutdown(true);) + JFR_ONLY(Jfr::on_vm_shutdown(static_cast<VMErrorType>(_id) == OOM_JAVA_HEAP_FATAL);) ``` Thanks, Yasumasa [1] https://mail.openjdk.org/pipermail/hotspot-jfr-dev/2019-January/000381.html [2] https://github.com/YaSuenag/jdk/commit/3f18fd74679387c385a8f79c8bea5b6b1f28…

2 2

Question about JEP 509 and thread state
by Milan Mimica 23 Jul '25

23 Jul '25

Hello I have always assumed that JFR thread sampler will ignore threads that are blocked on socket read, but the description on https://openjdk.org/jeps/509 states that this is exactly the problem that the JEP 509 aims to resolve. So my question is specifically about JfrSamplerThread::sample_java_thread/sample_native_thread. These explicitly skip sampling unless the thread state is _thread_in_Java/ _thread_in_native. Which state is a thread in while blocked on socket read? I assumed it was _thread_blocked. -- Milan Mimica

2 1

RFR: 8362836: JFR: Broken pipe in jdk/jfr/event/io/TestIOTopFrame.java
by Erik Gahlin 21 Jul '25

21 Jul '25

Could I have a review of a test fix? The problem is that the socket may be closed before all content has been written. Testing: 1000 * jdk/jfr/event/io/TestIOTopFrame.java Thanks Erik ------------- Commit messages: - Initial Changes: https://git.openjdk.org/jdk/pull/26405/files Webrev: https://webrevs.openjdk.org/?repo=jdk&pr=26405&range=00 Issue: https://bugs.openjdk.org/browse/JDK-8362836 Stats: 6 lines in 1 file changed: 4 ins; 0 del; 2 mod Patch: https://git.openjdk.org/jdk/pull/26405.diff Fetch: git fetch https://git.openjdk.org/jdk.git pull/26405/head:pull/26405 PR: https://git.openjdk.org/jdk/pull/26405

2 2