blob: d9dd3dd98761ea37de02acc9ee2c1efe9cfdb990 [file] [log] [blame]
// Copyright 2015 The Chromium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "base/profiler/native_stack_sampler.h"
#include <windows.h>
#include <winternl.h>
#include <cstdlib>
#include <map>
#include <memory>
#include <utility>
#include <vector>
#include "base/lazy_instance.h"
#include "base/logging.h"
#include "base/macros.h"
#include "base/memory/ptr_util.h"
#include "base/profiler/win32_stack_frame_unwinder.h"
#include "base/sampling_heap_profiler/module_cache.h"
#include "base/stl_util.h"
#include "base/time/time.h"
#include "base/trace_event/trace_event.h"
#include "starboard/types.h"
namespace base {
using Frame = StackSamplingProfiler::Frame;
using ProfileBuilder = StackSamplingProfiler::ProfileBuilder;
// Stack recording functions --------------------------------------------------
namespace {
// The thread environment block internal type.
struct TEB {
// Rest of struct is ignored.
// Returns the thread environment block pointer for |thread_handle|.
const TEB* GetThreadEnvironmentBlock(HANDLE thread_handle) {
// Define the internal types we need to invoke NtQueryInformationThread.
enum THREAD_INFORMATION_CLASS { ThreadBasicInformation };
struct CLIENT_ID {
HANDLE UniqueProcess;
HANDLE UniqueThread;
NTSTATUS ExitStatus;
TEB* Teb;
KAFFINITY AffinityMask;
LONG Priority;
LONG BasePriority;
using NtQueryInformationThreadFunction =
const auto nt_query_information_thread =
::GetModuleHandle(L"ntdll.dll"), "NtQueryInformationThread"));
if (!nt_query_information_thread)
return nullptr;
NTSTATUS status = nt_query_information_thread(
thread_handle, ThreadBasicInformation, &basic_info,
if (status != 0)
return nullptr;
return basic_info.Teb;
#if defined(_WIN64)
// If the value at |pointer| points to the original stack, rewrite it to point
// to the corresponding location in the copied stack.
void RewritePointerIfInOriginalStack(uintptr_t top,
uintptr_t bottom,
void* stack_copy,
const void** pointer) {
const auto value = reinterpret_cast<uintptr_t>(*pointer);
if (value >= bottom && value < top) {
*pointer = reinterpret_cast<const void*>(
static_cast<unsigned char*>(stack_copy) + (value - bottom));
void CopyMemoryFromStack(void* to, const void* from, size_t length)
NO_SANITIZE("address") {
// The following loop is an inlined version of memcpy. The code must be
// inlined to avoid instrumentation when using ASAN (memory sanitizer). The
// stack profiler is generating false positive when walking the stack.
for (size_t pos = 0; pos < length; ++pos)
reinterpret_cast<char*>(to)[pos] = reinterpret_cast<const char*>(from)[pos];
std::memcpy(to, from, length);
// Rewrites possible pointers to locations within the stack to point to the
// corresponding locations in the copy, and rewrites the non-volatile registers
// in |context| likewise. This is necessary to handle stack frames with dynamic
// stack allocation, where a pointer to the beginning of the dynamic allocation
// area is stored on the stack and/or in a non-volatile register.
// Eager rewriting of anything that looks like a pointer to the stack, as done
// in this function, does not adversely affect the stack unwinding. The only
// other values on the stack the unwinding depends on are return addresses,
// which should not point within the stack memory. The rewriting is guaranteed
// to catch all pointers because the stacks are guaranteed by the ABI to be
// sizeof(void*) aligned.
// Note: this function must not access memory in the original stack as it may
// have been changed or deallocated by this point. This is why |top| and
// |bottom| are passed as uintptr_t.
void RewritePointersToStackMemory(uintptr_t top,
uintptr_t bottom,
CONTEXT* context,
void* stack_copy) {
#if defined(_WIN64)
DWORD64 CONTEXT::*const nonvolatile_registers[] = {
// Rewrite pointers in the context.
for (size_t i = 0; i < size(nonvolatile_registers); ++i) {
DWORD64* const reg = &(context->*nonvolatile_registers[i]);
RewritePointerIfInOriginalStack(top, bottom, stack_copy,
reinterpret_cast<const void**>(reg));
// Rewrite pointers on the stack.
const void** start = reinterpret_cast<const void**>(stack_copy);
const void** end = reinterpret_cast<const void**>(
reinterpret_cast<char*>(stack_copy) + (top - bottom));
for (const void** loc = start; loc < end; ++loc)
RewritePointerIfInOriginalStack(top, bottom, stack_copy, loc);
// Movable type representing a recorded stack frame.
struct RecordedFrame {
RecordedFrame() {}
RecordedFrame(RecordedFrame&& other)
: instruction_pointer(other.instruction_pointer),
module(std::move(other.module)) {}
RecordedFrame& operator=(RecordedFrame&& other) {
instruction_pointer = other.instruction_pointer;
module = std::move(other.module);
return *this;
const void* instruction_pointer;
ScopedModuleHandle module;
// Walks the stack represented by |context| from the current frame downwards,
// recording the instruction pointer and associated module for each frame in
// |stack|.
void RecordStack(CONTEXT* context, std::vector<RecordedFrame>* stack) {
#ifdef _WIN64
// Reserve enough memory for most stacks, to avoid repeated
// allocations. Approximately 99.9% of recorded stacks are 128 frames or
// fewer.
Win32StackFrameUnwinder frame_unwinder;
while (context->Rip) {
const void* instruction_pointer =
reinterpret_cast<const void*>(context->Rip);
ScopedModuleHandle module;
if (!frame_unwinder.TryUnwind(context, &module))
RecordedFrame frame;
frame.instruction_pointer = instruction_pointer;
frame.module = std::move(module);
// ScopedDisablePriorityBoost -------------------------------------------------
// Disables priority boost on a thread for the lifetime of the object.
class ScopedDisablePriorityBoost {
ScopedDisablePriorityBoost(HANDLE thread_handle);
HANDLE thread_handle_;
BOOL got_previous_boost_state_;
BOOL boost_state_was_disabled_;
ScopedDisablePriorityBoost::ScopedDisablePriorityBoost(HANDLE thread_handle)
: thread_handle_(thread_handle),
boost_state_was_disabled_(false) {
got_previous_boost_state_ =
::GetThreadPriorityBoost(thread_handle_, &boost_state_was_disabled_);
if (got_previous_boost_state_) {
// Confusingly, TRUE disables priority boost.
::SetThreadPriorityBoost(thread_handle_, TRUE);
ScopedDisablePriorityBoost::~ScopedDisablePriorityBoost() {
if (got_previous_boost_state_)
::SetThreadPriorityBoost(thread_handle_, boost_state_was_disabled_);
// ScopedSuspendThread --------------------------------------------------------
// Suspends a thread for the lifetime of the object.
class ScopedSuspendThread {
ScopedSuspendThread(HANDLE thread_handle);
bool was_successful() const { return was_successful_; }
HANDLE thread_handle_;
bool was_successful_;
ScopedSuspendThread::ScopedSuspendThread(HANDLE thread_handle)
: thread_handle_(thread_handle),
was_successful_(::SuspendThread(thread_handle) !=
static_cast<DWORD>(-1)) {}
ScopedSuspendThread::~ScopedSuspendThread() {
if (!was_successful_)
// Disable the priority boost that the thread would otherwise receive on
// resume. We do this to avoid artificially altering the dynamics of the
// executing application any more than we already are by suspending and
// resuming the thread.
// Note that this can racily disable a priority boost that otherwise would
// have been given to the thread, if the thread is waiting on other wait
// conditions at the time of SuspendThread and those conditions are satisfied
// before priority boost is reenabled. The measured length of this window is
// ~100us, so this should occur fairly rarely.
ScopedDisablePriorityBoost disable_priority_boost(thread_handle_);
bool resume_thread_succeeded =
::ResumeThread(thread_handle_) != static_cast<DWORD>(-1);
CHECK(resume_thread_succeeded) << "ResumeThread failed: " << GetLastError();
// Tests whether |stack_pointer| points to a location in the guard page.
// IMPORTANT NOTE: This function is invoked while the target thread is
// suspended so it must not do any allocation from the default heap, including
// indirectly via use of DCHECK/CHECK or other logging statements. Otherwise
// this code can deadlock on heap locks in the default heap acquired by the
// target thread before it was suspended.
bool PointsToGuardPage(uintptr_t stack_pointer) {
SIZE_T result = ::VirtualQuery(reinterpret_cast<LPCVOID>(stack_pointer),
&memory_info, sizeof(memory_info));
return result != 0 && (memory_info.Protect & PAGE_GUARD);
// Suspends the thread with |thread_handle|, copies its stack and resumes the
// thread, then records the stack frames and associated modules into |stack|.
// IMPORTANT NOTE: No allocations from the default heap may occur in the
// ScopedSuspendThread scope, including indirectly via use of DCHECK/CHECK or
// other logging statements. Otherwise this code can deadlock on heap locks in
// the default heap acquired by the target thread before it was suspended.
void SuspendThreadAndRecordStack(
HANDLE thread_handle,
const void* base_address,
void* stack_copy_buffer,
size_t stack_copy_buffer_size,
std::vector<RecordedFrame>* stack,
ProfileBuilder* profile_builder,
NativeStackSamplerTestDelegate* test_delegate) {
CONTEXT thread_context = {0};
thread_context.ContextFlags = CONTEXT_FULL;
// The stack bounds are saved to uintptr_ts for use outside
// ScopedSuspendThread, as the thread's memory is not safe to dereference
// beyond that point.
const auto top = reinterpret_cast<uintptr_t>(base_address);
uintptr_t bottom = 0u;
TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("cpu_profiler"), "SuspendThread");
ScopedSuspendThread suspend_thread(thread_handle);
if (!suspend_thread.was_successful())
if (!::GetThreadContext(thread_handle, &thread_context))
#if defined(_WIN64)
bottom = thread_context.Rsp;
bottom = thread_context.Esp;
if ((top - bottom) > stack_copy_buffer_size)
// Dereferencing a pointer in the guard page in a thread that doesn't own
// the stack results in a STATUS_GUARD_PAGE_VIOLATION exception and a
// crash. This occurs very rarely, but reliably over the population.
if (PointsToGuardPage(bottom))
reinterpret_cast<const void*>(bottom), top - bottom);
if (test_delegate)
TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("cpu_profiler"), "RecordStack");
RewritePointersToStackMemory(top, bottom, &thread_context,
RecordStack(&thread_context, stack);
} // namespace
// NativeStackSamplerWin ------------------------------------------------------
class NativeStackSamplerWin : public NativeStackSampler {
NativeStackSamplerWin(win::ScopedHandle thread_handle,
NativeStackSamplerTestDelegate* test_delegate);
~NativeStackSamplerWin() override;
// StackSamplingProfiler::NativeStackSampler:
void ProfileRecordingStarting() override;
std::vector<Frame> RecordStackFrames(
StackBuffer* stack_buffer,
ProfileBuilder* profile_builder) override;
// Creates a set of frames with the information represented by |stack|.
std::vector<Frame> CreateFrames(const std::vector<RecordedFrame>& stack);
win::ScopedHandle thread_handle_;
NativeStackSamplerTestDelegate* const test_delegate_;
// The stack base address corresponding to |thread_handle_|.
const void* const thread_stack_base_address_;
// The module objects, indexed by the module handle.
std::map<HMODULE, ModuleCache::Module> module_cache_;
win::ScopedHandle thread_handle,
NativeStackSamplerTestDelegate* test_delegate)
: thread_handle_(thread_handle.Take()),
GetThreadEnvironmentBlock(thread_handle_.Get())->Tib.StackBase) {}
NativeStackSamplerWin::~NativeStackSamplerWin() {}
void NativeStackSamplerWin::ProfileRecordingStarting() {
std::vector<Frame> NativeStackSamplerWin::RecordStackFrames(
StackBuffer* stack_buffer,
ProfileBuilder* profile_builder) {
std::vector<RecordedFrame> stack;
SuspendThreadAndRecordStack(thread_handle_.Get(), thread_stack_base_address_,
stack_buffer->buffer(), stack_buffer->size(),
&stack, profile_builder, test_delegate_);
return CreateFrames(stack);
std::vector<Frame> NativeStackSamplerWin::CreateFrames(
const std::vector<RecordedFrame>& stack) {
std::vector<Frame> frames;
for (const auto& frame : stack) {
auto frame_ip = reinterpret_cast<uintptr_t>(frame.instruction_pointer);
HMODULE module_handle = frame.module.Get();
if (!module_handle) {
frames.emplace_back(frame_ip, ModuleCache::Module());
auto loc = module_cache_.find(module_handle);
if (loc != module_cache_.end()) {
frames.emplace_back(frame_ip, loc->second);
ModuleCache::Module module =
if (module.is_valid)
module_cache_.insert(std::make_pair(module_handle, module));
frames.emplace_back(frame_ip, std::move(module));
return frames;
// NativeStackSampler ---------------------------------------------------------
// static
std::unique_ptr<NativeStackSampler> NativeStackSampler::Create(
PlatformThreadId thread_id,
NativeStackSamplerTestDelegate* test_delegate) {
#if _WIN64
// Get the thread's handle.
HANDLE thread_handle = ::OpenThread(
FALSE, thread_id);
if (thread_handle) {
return std::unique_ptr<NativeStackSampler>(new NativeStackSamplerWin(
win::ScopedHandle(thread_handle), test_delegate));
return std::unique_ptr<NativeStackSampler>();
// static
size_t NativeStackSampler::GetStackBufferSize() {
// The default Win32 reserved stack size is 1 MB and Chrome Windows threads
// currently always use the default, but this allows for expansion if it
// occurs. The size beyond the actual stack size consists of unallocated
// virtual memory pages so carries little cost (just a bit of wasted address
// space).
return 2 << 20; // 2 MiB
} // namespace base