JVM SIGSEGV in vframe::java_sender via allocation-profiler JVMTI GetStackTrace on a virtual thread (JDK 25, Alpine/musl) — regression in 1.62.0 (1.60.1 OK)

添付された hs_err_pid1.log から始め、仮想スレッド上で Profiler::recordJVMTISample、ObjectSampler::recordAllocation、JVMTI GetStackTrace を通るアロケーションプロファイラの経路を追跡します。Alpine/musl を使用する JDK 25 で、tracer のバージョン 1.60.1 と 1.62.0 の動作を比較します。完了条件は、アロケーションプロファイリングで報告された SIGSEGV が発生しなくなることです。

hs_err_pid1.redacted.log

Tracer Version(s)

1.62.0 (crashes). Reverting to 1.60.1 resolves it — so this is a regression introduced between 1.60.1 and 1.62.0.

Java Version(s)

OpenJDK Runtime Environment Temurin-25.0.1+8 (build 25.0.1+8-LTS), linux-amd64

JVM Vendor

Eclipse Adoptium / Temurin

OS / Environment

Alpine Linux v3.22 (musl libc), QEMU 4 cores / 11G, container -Xmx6g, G1 GC. Spring Boot 4.0.6 app, virtual threads in use. Continuous profiler enabled (native ddprof, allocation profiling on).

Bug Report

After upgrading the agent to 1.62.0, the JVM hard-crashes (SIGSEGV, is_crash) intermittently — ~10 min after startup under normal traffic. The crash is in the Datadog allocation profiler: on a sampled allocation it calls JVMTI GetStackTrace, and walking a virtual thread's frames segfaults inside vframe::java_sender().

This appears to be a sibling of #9830 but on a different stackwalk path: #9830 was the async/ASGCT path (vframeStreamForte::forte_next), mitigated by -Ddd.profiling.ddprof.cstack=vm (default since 1.55.0). This one is the JVMTI GetStackTrace path used by the allocation sampler on virtual threads, which cstack=vm does not appear to cover (we are on 1.62.0, well past that default, and still crashing).

si_code: 128 (SI_KERNEL), si_addr: 0x0; the faulting thread is a virtual-thread carrier (ForkJoinPool-1-worker), _thread_in_vm. The ArrayList.grow / QueryExecutorImpl.processResults Java frames are just the innocent allocation being sampled (no large allocation — verified, the underlying collection is small).

Workaround: downgrade to 1.60.1 (clean). We expect DD_PROFILING_ALLOCATION_ENABLED=false (or DD_PROFILING_DDPROF_ENABLED=false, given musl) would also avoid it.

Crash stack (top frames from hs_err_pid1.log):

#  SIGSEGV (0xb) at pc=..., pid=1, tid=197
# Java VM: OpenJDK 64-Bit Server VM Temurin-25.0.1+8 (25.0.1+8-LTS, mixed mode, sharing, tiered, compressed oops, compressed class ptrs, g1 gc, linux-amd64)
# Problematic frame:
# V  [libjvm.so+0x1190128]  vframe::java_sender() const+0x38

Current thread: JavaThread "ForkJoinPool-1-worker-18" daemon [_thread_in_vm, id=197]

Native frames: (J=compiled Java code, j=interpreted, Vv=VM code, C=native code)
V  [libjvm.so]  vframe::java_sender() const+0x38
V  [libjvm.so]  JvmtiEnvBase::get_stack_trace(javaVFrame*, int, int, jvmtiFrameInfo*, int*)
V  [libjvm.so]  GetStackTraceClosure::do_vthread(Handle)
V  [libjvm.so]  JvmtiHandshake::execute(...)
V  [libjvm.so]  JvmtiEnv::GetStackTrace(...)
V  [libjvm.so]  jvmti_GetStackTrace
C  [libjavaProfiler-dd-*.so]  Profiler::recordJVMTISample(...)
C  [libjavaProfiler-dd-*.so]  ObjectSampler::recordAllocation(...)
C  [libjavaProfiler-dd-*.so]  ObjectSampler::SampledObjectAlloc(...)
V  [libjvm.so]  JvmtiExport::post_sampled_object_alloc(JavaThread*, oopDesc*)
V  [libjvm.so]  MemAllocator::allocate() / InstanceKlass::allocate_objArray(...)
V  [libjvm.so]  OptoRuntime::new_array_C(...)

Java frames: (J=compiled Java code, j=interpreted, Vv=VM code)
J  java.util.ArrayList.grow()  java.base@25.0.1
J  org.postgresql.core.v3.QueryExecutorImpl.processResults(...)
J  jdk.internal.vm.Continuation.run()  java.base@25.0.1
J  java.lang.VirtualThread.runContinuation()  java.base@25.0.1
J  java.util.concurrent.ForkJoinPool.runWorker(...)  java.base@25.0.1
j  java.util.concurrent.ForkJoinWorkerThread.run()  java.base@25.0.1

siginfo: si_signo: 11 (SIGSEGV), si_code: 128 (SI_KERNEL), si_addr: 0x0

I can attach the full hs_err_pid1.log (and the .jfr) if useful.