fix(profiler): recover frozen virtual-thread callers

AGENTS.md

@@ -71,6 +71,8 @@ and `:ddprof-lib:benchmarks`), `ddprof-test` (Java integration tests), `ddprof-t
 
 **Supported JVMs:** HotSpot, J9, Zing; HotSpot continuation recovery uses sampled instruction phases (compiled/interpreted entries and arm64 exception exits), validated metadata for return-barrier entries and interiors, and retires consumed Java anchors — see [doc/architecture/JvmSupportLayer.md](doc/architecture/JvmSupportLayer.md).
 **Platforms:** Linux x64/arm64 (primary), macOS arm64/x64; musl libc handled automatically.
+Frozen virtual-thread caller recovery uses bounded, read-only heap chunk traversal
+on supported HotSpot layouts and collectors; see [doc/architecture/FrozenStackWalker.md](doc/architecture/FrozenStackWalker.md).
 
 ## Operating Rules for Agent Sessions
 
@@ -154,6 +156,7 @@ before working on a subsystem.
 
 | Subsystem | Details |
 |---|---|
+| Frozen virtual-thread callers (bounded HotSpot heap chunk traversal, inline scopes and deoptimization) | [doc/architecture/FrozenStackWalker.md](doc/architecture/FrozenStackWalker.md) |
 | Call trace storage (triple-buffered ACTIVE/STANDBY/SCRATCH, 64-bit instance+slot trace IDs, lock-free rotation) | [doc/architecture/CallTraceStorage.md](doc/architecture/CallTraceStorage.md) |
 | String dictionary (arena chunks, rotation/clear protocols) | [doc/architecture/StringDictionary.md](doc/architecture/StringDictionary.md) |
 | RefCountGuard (lock-free RAII reader-drain before reclamation) | [doc/architecture/RefCountGuard.md](doc/architecture/RefCountGuard.md) |

ddprof-lib/src/main/cpp/counters.h

@@ -145,6 +145,11 @@
   X(WALKVM_FP_CHAIN_REACHED_CODEHEAP, "walkvm_fp_chain_reached_codeheap")    \
   X(WALKVM_ANCHOR_NOT_IN_JAVA,  "walkvm_anchor_not_in_java")                   \
   X(WALKVM_CONT_BARRIER_HIT,    "walkvm_cont_barrier_hit")                     \
+  X(WALKVM_FROZEN_RECOVERED,    "walkvm_frozen_recovered")                     \
+  X(WALKVM_FROZEN_FAILED,       "walkvm_frozen_failed")                        \
+  X(WALKVM_FROZEN_INTERPRETED,  "walkvm_frozen_interpreted")                   \
+  X(WALKVM_FROZEN_COMPILED,     "walkvm_frozen_compiled")                      \
+  X(WALKVM_FROZEN_PARENT,       "walkvm_frozen_parent")                        \
   X(WALKVM_ENTER_SPECIAL_HIT,   "walkvm_enter_special_hit")                    \
   X(WALKVM_CONT_SPECULATIVE_HIT,"walkvm_cont_speculative_hit")                 \
   X(WALKVM_CONT_ENTRY_NULL,     "walkvm_cont_entry_null")                      \

ddprof-lib/src/main/cpp/hotspot/frozenStackWalker.cpp

@@ -0,0 +1,250 @@
+/*
+ * Copyright 2026, Datadog, Inc.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "hotspot/frozenStackWalker.h"
+#include "hotspot/hotspotSupport.h"
+#include "hotspot/vmStructs.inline.h"
+#include "stackWalker.inline.h"
+
+namespace {
+
+struct ChunkLayout {
+    int carrier_cont;
+    int cont_parent;
+    int cont_tail;
+    int parent;
+    int size;
+    int sp;
+    int bottom;
+    int argsize;
+    int stack;
+    uintptr_t oop_base;
+    int oop_shift;
+    bool supported;
+};
+
+ChunkLayout layout = {};
+
+int readOffset(CodeCache* libjvm, const char* name) {
+    const int* address = (const int*)libjvm->findSymbol(name);
+    if (address == nullptr || *address <= 0 || *address > 4096) return -1;
+    return *address;
+}
+
+uintptr_t readOop(uintptr_t object, int offset) {
+    if (layout.oop_shift >= 0) {
+        unsigned int narrow = __atomic_load_n((unsigned int*)(object + offset), __ATOMIC_ACQUIRE);
+        return narrow == 0 ? 0 : layout.oop_base + ((uintptr_t)narrow << layout.oop_shift);
+    }
+    return __atomic_load_n((uintptr_t*)(object + offset), __ATOMIC_ACQUIRE);
+}
+
+// Interpreted frames in a chunk store FP, last_sp, sender_sp and locals as
+// word offsets, rather than the native-stack pointers used by walkVM.
+// See OpenJDK's StackChunkFrameStream::{fp,next_for_interpreter_frame}.
+bool relativePointer(uintptr_t base, intptr_t words, uintptr_t low, uintptr_t high, uintptr_t& result) {
+    if (words < -(intptr_t)((base - low) / sizeof(uintptr_t)) ||
+            words > (intptr_t)((high - base) / sizeof(uintptr_t))) return false;
+    result = base + words * sizeof(uintptr_t);
+    return true;
+}
+
+bool appendCompiled(VMNMethod* nm, const void* pc, uintptr_t unextended_sp, uintptr_t low,
+                    uintptr_t frame_end, ASGCT_CallFrame* frames,
+                    volatile int& depth, int max_depth, bool details) {
+    const void* original_pc = nm->originalPC(pc, unextended_sp, low, frame_end);
+    if (original_pc == nullptr || (original_pc != pc && CodeHeap::findNMethod(original_pc) != nm)) return false;
+    int level = nm->level();
+    FrameTypeId type = details && level >= 1 && level <= 3 ? FRAME_C1_COMPILED : FRAME_JIT_COMPILED;
+    int scope_offset = nm->findScopeOffset(original_pc);
+    // A deoptimized Java frame needs a decoded scope to account for its inline callers.
+    if (original_pc != pc && scope_offset <= 0) return false;
+    if (scope_offset > 0) {
+        ScopeDesc scope(nm);
+        do {
+            scope_offset = scope.decode(scope_offset);
+            VMMethod* method = scope.method();
+            if (method == nullptr) return false;
+            jmethodID id = method->validatedId();
+            if (id == JMETHODID_NOT_WALKABLE) return false;
+            FrameTypeId frame_type = details && scope_offset > 0 ? FRAME_INLINED : type;
+            HotspotSupport::fillJavaFrame(frames[depth++], frame_type, scope.bci(), id, method);
+        } while (scope_offset > 0 && depth < max_depth);
+        return scope_offset <= 0;
+    }
+    VMMethod* method = nm->method();
+    if (method == nullptr) return false;
+    jmethodID id = method->validatedId();
+    if (id == JMETHODID_NOT_WALKABLE) return false;
+    HotspotSupport::fillJavaFrame(frames[depth++], type, 0, id, method);
+    return true;
+}
+
+} // namespace
+
+bool FrozenStackWalker::available() {
+    return __atomic_load_n(&layout.supported, __ATOMIC_ACQUIRE);
+}
+
+void FrozenStackWalker::init() {
+    // VM::ready serializes initialization. Once readers can use this layout,
+    // later agent/API initialization must not rewrite its oop width or offsets.
+    if (available()) return;
+    if (_libjvm == nullptr || VM::java_version() < 21 || VM::java_version() > 27) return;
+#if !defined(__x86_64__) && !defined(__aarch64__)
+    return;
+#endif
+    // G1/Serial/Parallel relocate objects at safepoints. A sampled Java thread
+    // cannot reach a safepoint while its signal handler is reading these oops.
+    // Concurrently relocating collectors require load barriers not provided here.
+    bool collector_supported = false;
+    for (const char* name : {"UseG1GC", "UseSerialGC", "UseParallelGC"}) {
+        VMFlag* flag = VMFlag::find(name, {VMFlag::Type::Bool});
+        if (flag != nullptr && flag->get()) collector_supported = true;
+    }
+    TEST_LOG("Frozen stack prerequisites: collector=%d thread_oop=%d handle=%d", collector_supported,
+             _thread_oop_offset, _oop_handle_offset);
+    if (!collector_supported || _thread_oop_offset < 0 || _oop_handle_offset < 0) return;
+
+    layout.carrier_cont = readOffset(_libjvm, "_ZN16java_lang_Thread20_continuation_offsetE");
+    layout.cont_parent = readOffset(_libjvm, "_ZN28jdk_internal_vm_Continuation14_parent_offsetE");
+    layout.cont_tail = readOffset(_libjvm, "_ZN28jdk_internal_vm_Continuation12_tail_offsetE");
+    layout.parent = readOffset(_libjvm, "_ZN26jdk_internal_vm_StackChunk14_parent_offsetE");
+    layout.size = readOffset(_libjvm, "_ZN26jdk_internal_vm_StackChunk12_size_offsetE");
+    layout.sp = readOffset(_libjvm, "_ZN26jdk_internal_vm_StackChunk10_sp_offsetE");
+    layout.bottom = readOffset(_libjvm, "_ZN26jdk_internal_vm_StackChunk14_bottom_offsetE");
+    layout.argsize = readOffset(_libjvm, "_ZN26jdk_internal_vm_StackChunk15_argsize_offsetE");
+    layout.stack = readOffset(_libjvm, "_ZN23InstanceStackChunkKlass16_offset_of_stackE");
+    layout.oop_shift = -1;
+    VMFlag* compressed = VMFlag::find("UseCompressedOops", {VMFlag::Type::Bool});
+    if (compressed == nullptr) return;
+    if (compressed->get()) {
+        if (_narrow_oop_base_addr == nullptr || _narrow_oop_shift_addr == nullptr) return;
+        layout.oop_base = *(uintptr_t*)_narrow_oop_base_addr;
+        layout.oop_shift = *(int*)_narrow_oop_shift_addr;
+        if (layout.oop_shift < 0 || layout.oop_shift > 16) return;
+    }
+    bool supported = layout.carrier_cont > 0 && layout.cont_parent > 0 && layout.cont_tail > 0
+            && layout.parent > 0 && layout.size > 0 && layout.sp > 0 && layout.stack > 0
+            && layout.stack % sizeof(uintptr_t) == 0 && (layout.bottom > 0 || layout.argsize > 0);
+    __atomic_store_n(&layout.supported, supported, __ATOMIC_RELEASE);
+    TEST_LOG("Frozen stack layout: supported=%d carrier=%d tail=%d parent=%d size=%d sp=%d bottom=%d argsize=%d stack=%d oop_shift=%d",
+             layout.supported, layout.carrier_cont, layout.cont_tail, layout.parent, layout.size, layout.sp,
+             layout.bottom, layout.argsize, layout.stack, layout.oop_shift);
+}
+
+bool FrozenStackWalker::append(VMThread* thread, const void* sampled_pc, unsigned int continuation_depth,
+                              ASGCT_CallFrame* frames, volatile int& depth, int max_depth,
+                              bool details) {
+    if (!available() || thread == nullptr || thread->state() != _thread_in_Java
+            || continuation_depth != 0 || depth >= max_depth) return false;
+
+    // Freeze/thaw runs without a Java-to-VM transition. During doYield or its
+    // runtime calls, copied frames can still be on the carrier as well as in a
+    // chunk. Only an interrupted Java body has finished those stack updates.
+    VMNMethod* origin = CodeHeap::findNMethod(sampled_pc);
+    if (origin == nullptr || (!origin->isInterpreter() &&
+            (!origin->isNMethod() || origin->level() <= 0))) return false;
+
+    uintptr_t carrier = thread->carrierThreadOop();
+    if (carrier == 0) return false;
+    // Continuation.run publishes a child in Thread.cont before enterSpecial,
+    // and restores the parent after it returns. In those Java intervals the
+    // native boundary belongs to the parent. Without matching native-entry
+    // identity, recover only a single, outermost continuation.
+    uintptr_t continuation = readOop(carrier, layout.carrier_cont);
+    if (continuation == 0 || readOop(continuation, layout.cont_parent) != 0) return false;
+    uintptr_t chunk = readOop(continuation, layout.cont_tail);
+    if (chunk == 0) return false;
+    int interpreted_frames = 0;
+    int compiled_frames = 0;
+    int parent_chunks = 0;
+
+    // Empty chunks do not consume output depth. Bound their traversal as well
+    // so corrupt parent links cannot loop indefinitely in a signal handler.
+    for (int chunks = 0; chunk != 0 && chunks < 256; chunks++) {
+        int size = *(int*)(chunk + layout.size);
+        int sp_offset = *(int*)(chunk + layout.sp);
+        int bottom_offset = layout.bottom > 0 ? *(int*)(chunk + layout.bottom)
+                : size - *(int*)(chunk + layout.argsize);
+        if (size <= 0 || size > (1 << 24) || sp_offset < 0 || sp_offset > size
+                || bottom_offset < 0 || bottom_offset > size) return false;
+
+        uintptr_t low = chunk + layout.stack;
+        uintptr_t high = low + (uintptr_t)size * sizeof(uintptr_t);
+        uintptr_t end = low + (uintptr_t)bottom_offset * sizeof(uintptr_t);
+        uintptr_t sp = low + (uintptr_t)sp_offset * sizeof(uintptr_t);
+        uintptr_t unextended_sp = sp;
+
+        while (sp < end) {
+            if (depth >= max_depth || sp < low + 2 * sizeof(uintptr_t) || sp > high) return false;
+            const void* pc = stripPointer(((const void**)sp)[-1]);
+            VMNMethod* nm = CodeHeap::findNMethod(pc);
+            if (nm == nullptr) return false;
+            if (nm->isInterpreter()) {
+                uintptr_t fp;
+                uintptr_t fp_slot = sp - 2 * sizeof(uintptr_t);
+                if (!relativePointer(fp_slot, *(intptr_t*)fp_slot, low, high, fp)
+                        || fp < low + 9 * sizeof(uintptr_t) || fp > high - 2 * sizeof(uintptr_t)) return false;
+                VMMethod* method = ((VMMethod**)fp)[InterpreterFrame::method_offset];
+                if (method == nullptr) return false;
+                jmethodID id = method->validatedId();
+                if (id == JMETHODID_NOT_WALKABLE) return false;
+                // Method::bcp_from accepts either a bytecode pointer or an index.
+                uintptr_t bcp = ((uintptr_t*)fp)[InterpreterFrame::bcp_offset()];
+                uintptr_t bci = bcp <= 65535 ? bcp : bcp - (uintptr_t)method->bytecode();
+                if (bci > 65535) return false;
+                HotspotSupport::fillJavaFrame(frames[depth++], FRAME_INTERPRETED, bci, id, method);
+                interpreted_frames++;
+
+                uintptr_t locals;
+                // In both supported architectures locals immediately precedes bcp.
+                int locals_offset = InterpreterFrame::bcp_offset() + 1;
+                if (!relativePointer(fp, ((intptr_t*)fp)[locals_offset], low, high, locals)) return false;
+                if (locals + sizeof(uintptr_t) >= end) {
+                    sp = end;
+                } else {
+                    uintptr_t next_sp = fp + 2 * sizeof(uintptr_t);
+                    if (next_sp <= sp || !relativePointer(fp, ((intptr_t*)fp)[InterpreterFrame::sender_sp_offset],
+                            low, high, unextended_sp)) return false;
+                    sp = next_sp;
+                }
+            } else if (nm->isNMethod()) {
+                int frame_size = nm->frameSize();
+                if (frame_size <= 0 || frame_size > MAX_FRAME_SIZE_WORDS) return false;
+                // If the previous frame was interpreted, its sender_sp is the
+                // compiled caller's unextended SP (StackChunkFrameStream::next).
+                uintptr_t next_sp = unextended_sp + (uintptr_t)frame_size * sizeof(uintptr_t);
+                if (next_sp <= sp || next_sp > high) return false;
+                if (!appendCompiled(nm, pc, unextended_sp, low, next_sp,
+                                    frames, depth, max_depth, details)) return false;
+                compiled_frames++;
+                sp = next_sp >= end - 2 * sizeof(uintptr_t) ? end : next_sp;
+                unextended_sp = sp;
+            } else {
+                return false;
+            }
+            if (sp < end) {
+                const void* next_pc = stripPointer(((const void**)sp)[-1]);
+                VMNMethod* next = CodeHeap::findNMethod(next_pc);
+                if (next != nullptr && next->isInterpreter()) {
+                    uintptr_t fp;
+                    uintptr_t fp_slot = sp - 2 * sizeof(uintptr_t);
+                    if (!relativePointer(fp_slot, *(intptr_t*)fp_slot, low, high, fp)
+                            || fp < low + 2 * sizeof(uintptr_t) || fp > high) return false;
+                    if (!relativePointer(fp, ((intptr_t*)fp)[-2], low, high, unextended_sp)) return false;
+                }
+            }
+        }
+        uintptr_t parent = readOop(chunk, layout.parent);
+        if (parent == chunk) return false;
+        if (parent != 0) parent_chunks++;
+        chunk = parent;
+    }
+    Counters::increment(WALKVM_FROZEN_INTERPRETED, interpreted_frames);
+    Counters::increment(WALKVM_FROZEN_COMPILED, compiled_frames);
+    Counters::increment(WALKVM_FROZEN_PARENT, parent_chunks);
+    return chunk == 0;
+}

ddprof-lib/src/main/cpp/hotspot/frozenStackWalker.h

@@ -0,0 +1,24 @@
+/*
+ * Copyright 2026, Datadog, Inc.
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#ifndef _HOTSPOT_FROZENSTACKWALKER_H
+#define _HOTSPOT_FROZENSTACKWALKER_H
+
+#include "hotspot/vmStructs.h"
+
+// Reads frozen callers while walkVM's fault recovery is installed. No JNI,
+// allocation, heap mutation, or JVM calls are permitted during the walk.
+class FrozenStackWalker : private VMStructs {
+public:
+    // Called under VM::ready's lock. A successful layout is immutable thereafter.
+    static void init();
+    // Acquires the published layout without allocation or synchronization waits.
+    static bool available();
+    static bool append(VMThread* thread, const void* sampled_pc, unsigned int continuation_depth,
+                       ASGCT_CallFrame* frames, volatile int& depth, int max_depth,
+                       bool details);
+};
+
+#endif // _HOTSPOT_FROZENSTACKWALKER_H