Temporary commit to evaluate perf impact of prototype CPU profiler

base/profiler/cpu_profiler_win.cc

Index: base/profiler/cpu_profiler_win.cc
diff --git a/base/profiler/cpu_profiler_win.cc b/base/profiler/cpu_profiler_win.cc
new file mode 100644
index 0000000000000000000000000000000000000000..1ef7aa3964f811b2196c36756b7b7ea4e21ba401
--- /dev/null
+++ b/base/profiler/cpu_profiler_win.cc
@@ -0,0 +1,221 @@
+// 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/cpu_profiler.h"
+
+#include <Tlhelp32.h>
+#include <dbghelp.h>
+#include <stddef.h>
+#include <windows.h>
+
+#include "base/debug/stack_trace.h"
+#include "base/threading/platform_thread.h"
+#include "base/threading/thread_id_name_manager.h"
+#include "base/time/time.h"
+
+namespace base {
+
+const DWORD kSuspendFailed = static_cast<DWORD>(-1);
+
+CpuProfiler::CpuProfiler()
+    : ui_thread_(0),
+      io_thread_(0),
+      ui_thread_stack_depth_(0),
+      io_thread_stack_depth_(0) {
+#ifdef _WIN64
+  // Initialiazation
+  if (RtlVirtualUnwind == NULL && RtlLookupFunctionEntry == NULL) {
+    const HMODULE hNtDll = ::GetModuleHandle(L"ntdll.dll");

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

hNTDLL does not follow the naming conventions

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+    reinterpret_cast<void*&>(RtlVirtualUnwind) =
+        ::GetProcAddress(hNtDll, "RtlVirtualUnwind");
+    reinterpret_cast<void*&>(RtlLookupFunctionEntry) =
+        ::GetProcAddress(hNtDll, "RtlLookupFunctionEntry");
+  }
+#endif
+
+  //SymSetOptions(SYMOPT_UNDNAME | SYMOPT_LOAD_LINES | SYMOPT_DEFERRED_LOADS);

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

remove commented code.

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+  //if (!SymInitialize(GetCurrentProcess(), NULL, TRUE)) {
+  //  LOG(ERROR) << "SymInitialize failed with error: " << GetLastError();
+  //}
+}
+
+CpuProfiler::~CpuProfiler() {}
+
+bool CpuProfiler::IsPlatformSupported() {
+#ifdef _WIN64
+  bool supported = true;
+#else
+  bool supported = false;
+#endif
+  return supported;
+}
+
+void CpuProfiler::OnTimer() {
+  if (!ui_thread_ || !io_thread_) {
+    GetThreadIds();
+  }

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

racy, you should open the thread handles and keep them, rather than the thread
ids. Also it will be faster since you avoid OpenThread and CloseHandle being
called all the time.

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+
+  ui_thread_stack_depth_ = SampleThread(ui_thread_, arraysize(ui_thread_stack_),
+      ui_thread_stack_);
+  io_thread_stack_depth_ = SampleThread(io_thread_, arraysize(io_thread_stack_),
+      io_thread_stack_);
+  ResolveModules();
+
+  modules_.clear();
+  ui_thread_stack_depth_ = 0;
+  io_thread_stack_depth_ = 0;
+}
+
+int CpuProfiler::SampleThread(DWORD thread, int max_stack_size,
+                              StackTraceEntry* stack) {

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

this method does not need to be a member of the class

[Mike Wittman, 2015/02/04 01:37:09]

Done.

+  HANDLE thread_handle = ::OpenThread(THREAD_GET_CONTEXT |
+      THREAD_QUERY_INFORMATION | THREAD_SUSPEND_RESUME, FALSE, thread);
+  if (!thread_handle)
+    return 0;
+
+  if (::SuspendThread(thread_handle) == kSuspendFailed) {
+    LOG(ERROR) << "SuspendThread failed: " << GetLastError();
+    return 0;
+  }
+
+  CONTEXT thread_context;
+  memset(&thread_context, 0, sizeof(thread_context));
+  thread_context.ContextFlags = CONTEXT_ALL;
+  if (!::GetThreadContext(thread_handle, &thread_context)) {
+    LOG(ERROR) << "GetThreadContext failed: " << GetLastError();
+  }
+
+  //debug::StackTrace trace(&thread_context);

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

don't keep commented code. It is in or out.

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+
+  int stack_depth = StackTrace64(&thread_context, max_stack_size, stack);
+
+  //GetNames(stack_trace, stack_depth);
+
+  if (::ResumeThread(thread_handle) == kSuspendFailed) {
+    LOG(ERROR) << "ResumeThread failed: " << GetLastError();
+  }
+  ::CloseHandle(thread_handle);
+  return stack_depth;
+}
+
+void CpuProfiler::GetThreadIds() {
+  HANDLE snapshot = ::CreateToolhelp32Snapshot(TH32CS_SNAPTHREAD, 0);
+  if (INVALID_HANDLE_VALUE == snapshot) {
+    LOG(ERROR) << "CreateToolhelp32Snapshot failed";
+    return;
+  }
+
+  base::ThreadIdNameManager* manager = base::ThreadIdNameManager::GetInstance();
+  DWORD current_process = GetCurrentProcessId();
+  THREADENTRY32 thread_entry = {0};
+  thread_entry.dwSize = sizeof(THREADENTRY32);
+  BOOL result = ::Thread32First(snapshot, &thread_entry);
+  while (result) {
+    if (thread_entry.th32OwnerProcessID == current_process) {
+      const char* name = manager->GetName(thread_entry.th32ThreadID);
+      if (strcmp(name, "CrBrowserMain") == 0) {

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

I am afraid that this "CrBrowserMain" knowledge is not allowed at this layer,
the user of the profiler needs to pass the names or better the thread ids.

[Mike Wittman, 2015/02/04 01:37:09]

Changed this to just collect stacks from the main thread, whose handle we can
capture in the constructor. This should give us enough information to evaluate
performance.

+        ui_thread_ = thread_entry.th32ThreadID;
+      }
+      if (strcmp(name, "Chrome_IOThread") == 0) {
+        io_thread_ = thread_entry.th32ThreadID;
+      }
+    }
+    result = Thread32Next(snapshot, &thread_entry);
+  }
+  CloseHandle(snapshot);
+}
+
+int CpuProfiler::StackTrace64(CONTEXT* context, int stack_depth,
+                              StackTraceEntry* stack_trace) {
+#ifdef _WIN64
+  IMAGEHLP_SYMBOL64 sym;
+  sym.SizeOfStruct  = sizeof(sym);
+  sym.MaxNameLength = 0;
+
+  for (int i = 0; i < stack_depth; i++, stack_trace++) {

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

for ( ... , ++stack_trace) is the idiomatic use of our for loops.

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+    // Try to look up unwind metadata for the current function.
+    ULONG64 image_base;
+    PRUNTIME_FUNCTION runtime_function =
+        RtlLookupFunctionEntry(context->Rip, &image_base, NULL);
+
+    stack_trace->rsp = context->Rsp;
+    stack_trace->rip = context->Rip;
+    stack_trace->module = NULL;
+
+    if (runtime_function) {
+      HMODULE module = NULL;
+      if (GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS,

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

careful with side effects for example here
GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT 

[Mike Wittman, 2015/02/04 01:37:10]

Added a call to FreeLibrary, which should be sufficient for the purposes of
testing performance. GET_MODULE_HANDLE_EX_FLAG_UNCHANGED_REFCOUNT looks up the
module based on name. We need to look up based on IP.

+              reinterpret_cast<LPCTSTR >(context->Rip), &module)) {
+        stack_trace->module = module;
+      }
+
+      KNONVOLATILE_CONTEXT_POINTERS nvcontext;
+      RtlZeroMemory(&nvcontext, sizeof(KNONVOLATILE_CONTEXT_POINTERS));

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

154 not needed, KNONVOLATILE_CONTEXT_POINTERS nvcontext = {0};

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+      PVOID handler_data;
+      ULONG64 establisher_frame;
+      RtlVirtualUnwind(0, image_base, context->Rip, runtime_function, context,
+          &handler_data, &establisher_frame, &nvcontext);
+    } else {
+      // If we don't have a RUNTIME_FUNCTION, then we've encountered
+      // a leaf function.  Adjust the stack appropriately.
+      context->Rip  = (ULONG64)(*(PULONG64)context->Rsp);

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

don't use C style casts. In general this code has the copied from msdn smell,
please do a pass removing vestigial C code.

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+      context->Rsp += 8;
+    }
+
+    if (!context->Rip)
+      return i;
+  }
+  return stack_depth;
+#else
+  return 0;
+#endif
+}
+
+void CpuProfiler::GetNames(StackTraceEntry* stack_trace, int stack_depth) {

[cpu_(ooo_6.6-7.5), 2015/02/03 02:40:53]

seems also this one does not need to be part of the class, it can be standalone.

[Mike Wittman, 2015/02/04 01:37:10]

Done.

+  HANDLE process = GetCurrentProcess();
+  DWORD64 displacement = 0;
+  DWORD line_displacement = 0;
+  DWORD64 address = 0;
+  char buffer[sizeof(SYMBOL_INFO) + MAX_SYM_NAME * sizeof(wchar_t)];
+  PSYMBOL_INFO symbol_info = (PSYMBOL_INFO)buffer;
+  symbol_info->SizeOfStruct = sizeof(SYMBOL_INFO);
+  symbol_info->MaxNameLen = MAX_SYM_NAME;
+  std::string file_name;
+
+  IMAGEHLP_LINE64 line;
+  line.SizeOfStruct = sizeof(IMAGEHLP_LINE64);
+
+  for (int i = 0; i < stack_depth; i++, stack_trace++) {
+    address = stack_trace->rip;
+    if (!SymFromAddr(process, address, &displacement, symbol_info)) {
+      wcscpy(reinterpret_cast<wchar_t*>(symbol_info->Name), L"failed");
+    }
+    if (!SymGetLineFromAddr64(process, address, &line_displacement, &line)) {
+      line.LineNumber = 0;
+      file_name.clear();
+    } else {
+      file_name.assign(line.FileName);
+    }
+  }
+}
+
+void CpuProfiler::ResolveModules() {
+  ProcessStack(ui_thread_stack_, ui_thread_stack_depth_);
+  ProcessStack(io_thread_stack_, io_thread_stack_depth_);
+}
+
+void CpuProfiler::ProcessStack(StackTraceEntry* stack, int stack_depth) {
+  for (int i = 0; i < stack_depth; i++, stack++) {
+    HMODULE module = stack->module;
+    if (!module || (modules_.find(module) != modules_.end()))
+      continue;
+
+    wchar_t module_name[MAX_PATH];
+    if (GetModuleFileName(module, module_name, arraysize(module_name))) {
+      modules_.insert(std::pair<HMODULE, base::string16>(module, module_name));
+    }
+  }
+}
+
+}