// 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 #include #include #include #include #include #include #include #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" namespace base { using Frame = StackSamplingProfiler::Frame; using InternalFrame = StackSamplingProfiler::InternalFrame; using ProfileBuilder = StackSamplingProfiler::ProfileBuilder; // Stack recording functions -------------------------------------------------- namespace { // The thread environment block internal type. struct TEB { NT_TIB Tib; // 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; }; struct THREAD_BASIC_INFORMATION { NTSTATUS ExitStatus; TEB* Teb; CLIENT_ID ClientId; KAFFINITY AffinityMask; LONG Priority; LONG BasePriority; }; using NtQueryInformationThreadFunction = NTSTATUS(WINAPI*)(HANDLE, THREAD_INFORMATION_CLASS, PVOID, ULONG, PULONG); const auto nt_query_information_thread = reinterpret_cast(::GetProcAddress( ::GetModuleHandle(L"ntdll.dll"), "NtQueryInformationThread")); if (!nt_query_information_thread) return nullptr; THREAD_BASIC_INFORMATION basic_info = {0}; NTSTATUS status = nt_query_information_thread( thread_handle, ThreadBasicInformation, &basic_info, sizeof(THREAD_BASIC_INFORMATION), nullptr); 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(*pointer); if (value >= bottom && value < top) { *pointer = reinterpret_cast( static_cast(stack_copy) + (value - bottom)); } } #endif void CopyMemoryFromStack(void* to, const void* from, size_t length) NO_SANITIZE("address") { #if defined(ADDRESS_SANITIZER) // 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(to)[pos] = reinterpret_cast(from)[pos]; #else std::memcpy(to, from, length); #endif } // 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[] = { &CONTEXT::R12, &CONTEXT::R13, &CONTEXT::R14, &CONTEXT::R15, &CONTEXT::Rdi, &CONTEXT::Rsi, &CONTEXT::Rbx, &CONTEXT::Rbp, &CONTEXT::Rsp}; // 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(reg)); } // Rewrite pointers on the stack. const void** start = reinterpret_cast(stack_copy); const void** end = reinterpret_cast( reinterpret_cast(stack_copy) + (top - bottom)); for (const void** loc = start; loc < end; ++loc) RewritePointerIfInOriginalStack(top, bottom, stack_copy, loc); #endif } // 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; private: DISALLOW_COPY_AND_ASSIGN(RecordedFrame); }; // 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* stack) { #ifdef _WIN64 DCHECK(stack->empty()); // Reserve enough memory for most stacks, to avoid repeated // allocations. Approximately 99.9% of recorded stacks are 128 frames or // fewer. stack->reserve(128); Win32StackFrameUnwinder frame_unwinder; while (context->Rip) { const void* instruction_pointer = reinterpret_cast(context->Rip); ScopedModuleHandle module; if (!frame_unwinder.TryUnwind(context, &module)) return; RecordedFrame frame; frame.instruction_pointer = instruction_pointer; frame.module = std::move(module); stack->push_back(std::move(frame)); } #endif } // ScopedDisablePriorityBoost ------------------------------------------------- // Disables priority boost on a thread for the lifetime of the object. class ScopedDisablePriorityBoost { public: ScopedDisablePriorityBoost(HANDLE thread_handle); ~ScopedDisablePriorityBoost(); private: HANDLE thread_handle_; BOOL got_previous_boost_state_; BOOL boost_state_was_disabled_; DISALLOW_COPY_AND_ASSIGN(ScopedDisablePriorityBoost); }; ScopedDisablePriorityBoost::ScopedDisablePriorityBoost(HANDLE thread_handle) : thread_handle_(thread_handle), got_previous_boost_state_(false), 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 { public: ScopedSuspendThread(HANDLE thread_handle); ~ScopedSuspendThread(); bool was_successful() const { return was_successful_; } private: HANDLE thread_handle_; bool was_successful_; DISALLOW_COPY_AND_ASSIGN(ScopedSuspendThread); }; ScopedSuspendThread::ScopedSuspendThread(HANDLE thread_handle) : thread_handle_(thread_handle), was_successful_(::SuspendThread(thread_handle) != static_cast(-1)) {} ScopedSuspendThread::~ScopedSuspendThread() { if (!was_successful_) return; // 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(-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) { MEMORY_BASIC_INFORMATION memory_info; SIZE_T result = ::VirtualQuery(reinterpret_cast(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* stack, ProfileBuilder* profile_builder, NativeStackSamplerTestDelegate* test_delegate) { DCHECK(stack->empty()); 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(base_address); uintptr_t bottom = 0u; { ScopedSuspendThread suspend_thread(thread_handle); if (!suspend_thread.was_successful()) return; if (!::GetThreadContext(thread_handle, &thread_context)) return; #if defined(_WIN64) bottom = thread_context.Rsp; #else bottom = thread_context.Esp; #endif if ((top - bottom) > stack_copy_buffer_size) return; // 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)) return; profile_builder->RecordAnnotations(); CopyMemoryFromStack(stack_copy_buffer, reinterpret_cast(bottom), top - bottom); } if (test_delegate) test_delegate->OnPreStackWalk(); RewritePointersToStackMemory(top, bottom, &thread_context, stack_copy_buffer); RecordStack(&thread_context, stack); } } // namespace // NativeStackSamplerWin ------------------------------------------------------ class NativeStackSamplerWin : public NativeStackSampler { public: NativeStackSamplerWin(win::ScopedHandle thread_handle, NativeStackSamplerTestDelegate* test_delegate); ~NativeStackSamplerWin() override; // StackSamplingProfiler::NativeStackSampler: void ProfileRecordingStarting() override; std::vector RecordStackFrames( StackBuffer* stack_buffer, ProfileBuilder* profile_builder) override; private: // Creates a set of internal frames with the information represented by // |stack|. std::vector CreateInternalFrames( const std::vector& 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 internal module objects, indexed by the module handle. std::map module_cache_; DISALLOW_COPY_AND_ASSIGN(NativeStackSamplerWin); }; NativeStackSamplerWin::NativeStackSamplerWin( win::ScopedHandle thread_handle, NativeStackSamplerTestDelegate* test_delegate) : thread_handle_(thread_handle.Take()), test_delegate_(test_delegate), thread_stack_base_address_( GetThreadEnvironmentBlock(thread_handle_.Get())->Tib.StackBase) {} NativeStackSamplerWin::~NativeStackSamplerWin() {} void NativeStackSamplerWin::ProfileRecordingStarting() { module_cache_.clear(); } std::vector NativeStackSamplerWin::RecordStackFrames( StackBuffer* stack_buffer, ProfileBuilder* profile_builder) { DCHECK(stack_buffer); std::vector stack; SuspendThreadAndRecordStack(thread_handle_.Get(), thread_stack_base_address_, stack_buffer->buffer(), stack_buffer->size(), &stack, profile_builder, test_delegate_); return CreateInternalFrames(stack); } std::vector NativeStackSamplerWin::CreateInternalFrames( const std::vector& stack) { std::vector internal_frames; internal_frames.reserve(stack.size()); for (const auto& frame : stack) { auto frame_ip = reinterpret_cast(frame.instruction_pointer); HMODULE module_handle = frame.module.Get(); if (!module_handle) { internal_frames.emplace_back(frame_ip, ModuleCache::Module()); continue; } auto loc = module_cache_.find(module_handle); if (loc != module_cache_.end()) { internal_frames.emplace_back(frame_ip, loc->second); continue; } ModuleCache::Module internal_module = ModuleCache::CreateModuleForHandle(module_handle); if (internal_module.is_valid) module_cache_.insert(std::make_pair(module_handle, internal_module)); internal_frames.emplace_back(frame_ip, std::move(internal_module)); } return internal_frames; } // NativeStackSampler --------------------------------------------------------- // static std::unique_ptr NativeStackSampler::Create( PlatformThreadId thread_id, NativeStackSamplerTestDelegate* test_delegate) { #if _WIN64 // Get the thread's handle. HANDLE thread_handle = ::OpenThread( THREAD_GET_CONTEXT | THREAD_SUSPEND_RESUME | THREAD_QUERY_INFORMATION, FALSE, thread_id); if (thread_handle) { return std::unique_ptr(new NativeStackSamplerWin( win::ScopedHandle(thread_handle), test_delegate)); } #endif return std::unique_ptr(); } // 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