poc: use frame pointer unwinding for go arm64

nativeunwind/elfunwindinfo/elfgopclntab.go

@@ -141,6 +141,14 @@
 	return magic == magicGo1_18 || magic == magicGo1_20
 }
 
+// IsGo121OrNewer returns true if magic matches with the Go 1.21 or newer.
+// TODO(fg): This is actually checking for go1.21 right now, I need to figure
+// out how to detect go1.21+ here. But it "works" b/c the existing coredump test
+// cases are for go1.18.
+func IsGo21orNewer(magic uint32) bool {
+	return magic == magicGo1_20
+}
+
 // pclntabHeaderSignature returns a byte slice that can be
 // used to verify if some bytes represent a valid pclntab header.
 func pclntabHeaderSignature(arch elf.Machine) []byte {
@@ -550,7 +558,19 @@
 				return err
 			}
 		case elf.EM_AARCH64:
-			if err := parseArm64pclntabFunc(ee.deltas, fun, dataLen, pctab, i,
+			if !IsGo21orNewer(hdr.magic) {
+				// Before go1.21 frame pointers were not properly kept on arm64
+				// when the Go runtime copies the stack during
+				// `runtime.morestack` calls: all old frame pointers are set to
+				// 0.
+				//
+				// https://github.com/golang/go/blob/c318f191/src/runtime/stack.go#L676
+				// https://blog.felixge.de/waiting-for-go1-21-execution-tracing-with-less-than-one-percent-overhead/
+				//
+				// We thus need to unwind with stack delta offsets.
+				strategy = strategyDeltasWithoutRBP
+			}
+			if err := parseArm64pclntabFunc(ee.deltas, fun, dataLen, pctab, strategy, i,
 				hdr.quantum); err != nil {
 				return err
 			}
@@ -628,7 +648,8 @@
 
 // parseArm64pclntabFunc extracts interval information from ARM64 based pclntabFunc.
 func parseArm64pclntabFunc(deltas *sdtypes.StackDeltaArray, fun *pclntabFunc,
-	dataLen uintptr, pctab []byte, i uint64, quantum uint8) error {
+	dataLen uintptr, pctab []byte, strategy int, i uint64, quantum uint8) error {
+
 	if fun.pcspOff == 0 {
 		// Some CGO functions don't have PCSP info: skip them.
 		return nil
@@ -637,20 +658,25 @@
 		return fmt.Errorf(".gopclntab func %v pcspOff = %d is invalid", i, fun.pcspOff)
 	}
 
-	// On ARM64, frame pointers are not properly kept when the Go runtime copies the stack during
-	// `runtime.morestack` calls: all old frame pointers are set to 0.
-	//
-	// https://github.com/golang/go/blob/c318f191/src/runtime/stack.go#L676
-	//
-	// We thus need to unwind with stack delta offsets.
-
 	hint := sdtypes.UnwindHintKeep
 	p := newPcval(pctab[fun.pcspOff:], uint(fun.startPc), quantum)
 	for ok := true; ok; ok = p.step() {
 		var info sdtypes.UnwindInfo
 		if p.val == 0 {
 			// Return instruction, function prologue or leaf function body: unwind via LR.
 			info = sdtypes.UnwindInfoLR
+		} else if strategy == strategyFramePointer {
+			param, ok := sdtypes.PackDerefParam(0, 8)
+			if !ok {
+				panic("bad")
+			}
+			// Use stack frame-pointer delta
+			info = sdtypes.UnwindInfo{
+				Opcode:   sdtypes.UnwindOpcodeBaseFP | sdtypes.UnwindOpcodeFlagDeref,
+				Param:    param,
+				FPOpcode: sdtypes.UnwindOpcodeBaseSP,
+				FPParam:  0,
+			}
 		} else {
 			// Regular basic block in the function body: unwind via SP.
 			info = sdtypes.UnwindInfo{

support/ebpf/native_stack_trace.ebpf.c

@@ -527,6 +527,11 @@ static ErrorCode unwind_one_frame(u64 pid, u32 frame_idx, struct UnwindState *st
     }
   }
 
+  if (state->fp == 0) {
+    *stop = true;
+    return ERR_OK;
+  }
+
   UnwindInfo *info = bpf_map_lookup_elem(&unwind_info_array, &unwindInfo);
   if (!info) {
     increment_metric(metricID_UnwindNativeErrBadUnwindInfoIndex);
@@ -589,7 +594,7 @@ static ErrorCode unwind_one_frame(u64 pid, u32 frame_idx, struct UnwindState *st
   // this implies that if no other changes are applied to the stack such
   // as alloca(), following the prolog SP/FP points to the frame record
   // itself, in such a case FP offset will be equal to 8
-  if (info->fpParam == 8) {
+  if (info->fpParam == 8 || (info->opcode == (UNWIND_OPCODEF_DEREF | UNWIND_OPCODE_BASE_FP))) {
     // we can assume the presence of frame pointers
     if (info->fpOpcode != UNWIND_OPCODE_BASE_LR) {
       // FP precedes the RA on the stack (Aarch64 ABI requirement)

tools/coredump/testdata/arm64/hello.1900091.hello5.json

@@ -0,0 +1,75 @@
+{
+  "coredump-ref": "f941d6492b38f927820f32ab65196c88a357154e4d04a1502a02e268421448c0",
+  "threads": [
+    {
+      "lwp": 1900091,
+      "frames": [
+        "hello+0x92d4c",
+        "hello+0x92ceb",
+        "hello+0x92c33",
+        "hello+0x92bef",
+        "hello+0x92afb",
+        "hello+0x92ac7",
+        "hello+0x92dcb",
+        "hello+0x456fb",
+        "hello+0x73b03"
+      ]
+    },
+    {
+      "lwp": 1900101,
+      "frames": [
+        "hello+0x748e4",
+        "hello+0x530fb",
+        "hello+0x48a47",
+        "hello+0x48997",
+        "hello+0x716af"
+      ]
+    },
+    {
+      "lwp": 1900102,
+      "frames": [
+        "hello+0x74f4c",
+        "hello+0x3feeb",
+        "hello+0x1aefb",
+        "hello+0x4a723",
+        "hello+0x4c69f",
+        "hello+0x4dae7",
+        "hello+0x4e17f",
+        "hello+0x71723"
+      ]
+    },
+    {
+      "lwp": 1900103,
+      "frames": [
+        "hello+0x74f4c",
+        "hello+0x3feeb",
+        "hello+0x1aefb",
+        "hello+0x4a723",
+        "hello+0x4b37b",
+        "hello+0x4daab",
+        "hello+0x4e17f",
+        "hello+0x71723"
+      ]
+    },
+    {
+      "lwp": 1900104,
+      "frames": [
+        "hello+0x74f4c",
+        "hello+0x3feeb",
+        "hello+0x1aefb",
+        "hello+0x4a723",
+        "hello+0x4c69f",
+        "hello+0x4dae7",
+        "hello+0x48a77",
+        "hello+0x48997",
+        "hello+0x716af"
+      ]
+    }
+  ],
+  "modules": [
+    {
+      "ref": "a2dbc7c1f120e00e2da9881afad3eccb2a4b6d0469d28851acb59c4e570f4012",
+      "local-path": "/home/bits/go/src/github.com/open-telemetry/opentelemetry-ebpf-profiler/tools/coredump/hello"
+    }
+  ]
+}