IdleWakeups tool

tools/win/IdleWakeups/idle_wakeups.cpp

+// Copyright 2016 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 "stdafx.h"
+
+#include <map>
+#include <vector>
+#include <algorithm>
+
+#include "power_sampler.h"
+
+// From ntdef.h
+typedef struct _UNICODE_STRING {
+  USHORT Length;
+  USHORT MaximumLength;
+  PWCH Buffer;
+} UNICODE_STRING;
+
+// From <wdm.h>
+typedef LONG KPRIORITY;
+typedef LONG KWAIT_REASON;  // Full definition is in wdm.h
+
+// From ntddk.h
+typedef struct _VM_COUNTERS {
+  SIZE_T PeakVirtualSize;
+  SIZE_T VirtualSize;
+  ULONG PageFaultCount;
+  // Padding here in 64-bit
+  SIZE_T PeakWorkingSetSize;
+  SIZE_T WorkingSetSize;
+  SIZE_T QuotaPeakPagedPoolUsage;
+  SIZE_T QuotaPagedPoolUsage;
+  SIZE_T QuotaPeakNonPagedPoolUsage;
+  SIZE_T QuotaNonPagedPoolUsage;
+  SIZE_T PagefileUsage;
+  SIZE_T PeakPagefileUsage;
+} VM_COUNTERS;
+
+// Two possibilities available from here:
+// http://stackoverflow.com/questions/28858849/where-is-system-information-class-defined
+
+typedef enum _SYSTEM_INFORMATION_CLASS {
+  SystemBasicInformation = 0,
+  SystemPerformanceInformation = 2,
+  SystemTimeOfDayInformation = 3,
+  SystemProcessInformation = 5,  // This is the number that we need
+  SystemProcessorPerformanceInformation = 8,
+  SystemInterruptInformation = 23,
+  SystemExceptionInformation = 33,
+  SystemRegistryQuotaInformation = 37,
+  SystemLookasideInformation = 45
+} SYSTEM_INFORMATION_CLASS;
+
+// https://msdn.microsoft.com/en-us/library/gg750647.aspx?f=255&MSPPError=-2147217396
+typedef struct {
+  HANDLE UniqueProcess;  // Actually process ID
+  HANDLE UniqueThread;   // Actually thread ID
+} CLIENT_ID;
+
+// From http://alax.info/blog/1182, with corrections and modifications
+// Originally from
+// http://undocumented.ntinternals.net/index.html?page=UserMode%2FUndocumented%20Functions%2FSystem%20Information%2FStructures%2FSYSTEM_THREAD.html
+struct SYSTEM_THREAD_INFORMATION {
+  ULONGLONG KernelTime;
+  ULONGLONG UserTime;
+  ULONGLONG CreateTime;
+  ULONG WaitTime;
+  // Padding here in 64-bit
+  PVOID StartAddress;
+  CLIENT_ID ClientId;
+  KPRIORITY Priority;
+  LONG BasePriority;
+  ULONG ContextSwitchCount;
+  ULONG State;
+  KWAIT_REASON WaitReason;
+};
+#if _M_X64
+static_assert(sizeof(SYSTEM_THREAD_INFORMATION) == 80,
+              "Structure size mismatch");
+#else
+static_assert(sizeof(SYSTEM_THREAD_INFORMATION) == 64,
+              "Structure size mismatch");
+#endif
+
+// From http://alax.info/blog/1182, with corrections and modifications
+// Originally from
+// http://undocumented.ntinternals.net/index.html?page=UserMode%2FUndocumented%20Functions%2FSystem%20Information%2FStructures%2FSYSTEM_THREAD.html
+struct SYSTEM_PROCESS_INFORMATION {
+  ULONG NextEntryOffset;
+  ULONG NumberOfThreads;
+  // http://processhacker.sourceforge.net/doc/struct___s_y_s_t_e_m___p_r_o_c_e_s_s___i_n_f_o_r_m_a_t_i_o_n.html
+  ULONGLONG WorkingSetPrivateSize;
+  ULONG HardFaultCount;
+  ULONG Reserved1;
+  ULONGLONG CycleTime;
+  ULONGLONG CreateTime;
+  ULONGLONG UserTime;
+  ULONGLONG KernelTime;
+  UNICODE_STRING ImageName;
+  KPRIORITY BasePriority;
+  HANDLE ProcessId;
+  HANDLE ParentProcessId;
+  ULONG HandleCount;
+  ULONG Reserved2[2];
+  // Padding here in 64-bit
+  VM_COUNTERS VirtualMemoryCounters;
+  size_t Reserved3;
+  IO_COUNTERS IoCounters;
+  SYSTEM_THREAD_INFORMATION Threads[1];
+};
+#if _M_X64
+static_assert(sizeof(SYSTEM_PROCESS_INFORMATION) == 336,
+              "Structure size mismatch");
+#else
+static_assert(sizeof(SYSTEM_PROCESS_INFORMATION) == 248,
+              "Structure size mismatch");
+#endif
+
+// ntstatus.h conflicts with windows.h so define this locally.
+#define STATUS_SUCCESS ((NTSTATUS)0x00000000L)
+#define STATUS_BUFFER_TOO_SMALL ((NTSTATUS)0xC0000023L)
+#define STATUS_INFO_LENGTH_MISMATCH ((NTSTATUS)0xC0000004L)
+
+typedef NTSTATUS(WINAPI* NTQUERYSYSTEMINFORMATION)(
+    SYSTEM_INFORMATION_CLASS SystemInformationClass,
+    PVOID SystemInformation,
+    ULONG SystemInformationLength,
+    PULONG ReturnLength);
+
+__declspec(noreturn) void oops(const char* pMessage) {
+  printf("%s\n", pMessage);
+  exit(0);
+}
+
+// Contains per thread data stored in each data snapshot.
+struct ThreadData {
+  HANDLE ThreadId;
+  ULONG ContextSwitches;
+};
+
+typedef std::vector<ThreadData> ThreadsVector;
+
+// Contains per process data stored in each data snapshot.
+struct ProcessData {
+  HANDLE ProcessId;
+  ULONGLONG CpuTime;
+  ULONGLONG WorkingSetPrivateSize;
+  ThreadsVector Threads;
+};
+
+// A vector of ProcessData represents one snapshot of perf data
+// collected by the tool each collection interval.
+typedef std::vector<ProcessData> ProcessesVector;
+
+// Result data structure contains a final set of values calculated based on
+// comparison of two snapshots. These are the values that the tool prints
+// in the output.
+struct Result {
+  ULONG IdleWakeupsPerSec;
+  double CpuUsage;
+  ULONGLONG WorkingSet;
+  double Power;
+};
+
+typedef std::vector<Result> ResultVector;
+
+// The following 4 functions are used for sorting of ResultVector.
+ULONG GetIdleWakeupsPerSec(const Result& r) {
+  return r.IdleWakeupsPerSec;
+}
+double GetCpuUsage(const Result& r) {
+  return r.CpuUsage;
+}
+ULONGLONG GetWorkingSet(const Result& r) {
+  return r.WorkingSet;
+}
+double GetPower(const Result& r) {
+  return r.Power;
+}
+
+template <typename T>
+T GetMedian(ResultVector* results, T (*getter)(const Result&)) {
+  std::sort(results->begin(), results->end(),
+            [&](const Result& lhs, const Result& rhs) {
+              return getter(lhs) < getter(rhs);
+            });
+
+  size_t median_index = results->size() / 2;
+  if (results->size() % 2 != 0) {
+    return getter((*results)[median_index]);
+  } else {
+    return (getter((*results)[median_index - 1]) +
+            getter((*results)[median_index])) /
+           2;
+  }
+}
+
+// This class holds the app state and constains a number of utilities for
+// collecting and diffing snapshots of data, handling processes, etc.
+class IdleWakeups {
+ public:
+  IdleWakeups(const wchar_t* process_name);
+  ~IdleWakeups();
+
+  double TakeSnapshot(ProcessesVector* processes);
+  Result DiffSnapshots(double time_delta,
+                       const ProcessesVector& prev_processes,
+                       const ProcessesVector& processes);
+
+  void OpenProcesses(const ProcessesVector& processes);
+  void CloseProcesses();
+
+  const wchar_t* target_process_name_filter() const {
+    return target_process_name;
+  }
+
+ private:
+  HANDLE GetProcessHandle(const ProcessData& process);
+  void OpenProcess(const ProcessData& process);
+  void CloseProcess(const ProcessData& process);
+  bool GetFinishedProcessCpuTime(const ProcessData& process,
+                                 ULONGLONG* cpu_usage);
+
+  static void SortThreads(ThreadsVector* processes);
+  static void SortProcesses(ProcessesVector* processes);
+
+  static ULONG CountContextSwitches(const ProcessData& process_data);
+  static ULONG DiffContextSwitches(const ProcessData& prev_process_data,
+                                   const ProcessData& process_data);
+
+  DWORD NumberOfprocessors() const { return number_of_processors; }
+
+  NTQUERYSYSTEMINFORMATION nt_query_system_information_ptr;
+  DWORD number_of_processors;
+  wchar_t target_process_name[256];
+  LARGE_INTEGER perf_frequency;
+  LARGE_INTEGER previous_perf_counter_value;
+  ULONG previous_buffer_size = 0;
+
+  // The first argument of HANDLE type is actually used for process IDs. This is
+  // consistent with data structures above.
+  std::map<HANDLE, HANDLE> process_id_to_hanle_map;
+};
+
+IdleWakeups::IdleWakeups(const wchar_t* process_name) {
+  lstrcpyn(target_process_name, process_name,
+           sizeof(target_process_name) / sizeof(wchar_t));
+
+  HMODULE ntdll = GetModuleHandle(_T("ntdll.dll"));
+  if (!ntdll)
+    oops("Couldn't load ntdll.dll");
+  nt_query_system_information_ptr = (NTQUERYSYSTEMINFORMATION)GetProcAddress(
+      ntdll, "NtQuerySystemInformation");
+  if (!nt_query_system_information_ptr)
+    oops("Couldn't find NtQuerySystemInformation");
+
+  SYSTEM_INFO system_info;
+  GetNativeSystemInfo(&system_info);
+  number_of_processors = system_info.dwNumberOfProcessors;
+  printf("Number of processors: %d\n", number_of_processors);
+
+  QueryPerformanceFrequency(&perf_frequency);
+}
+
+IdleWakeups::~IdleWakeups() {
+  CloseProcesses();
+}
+
+void IdleWakeups::OpenProcesses(const ProcessesVector& processes) {
+  for (auto& process : processes) {
+    OpenProcess(process);
+  }
+}
+
+void IdleWakeups::CloseProcesses() {
+  for (auto& pair : process_id_to_hanle_map) {
+    CloseHandle(pair.second);
+  }
+  process_id_to_hanle_map.clear();
+}
+
+HANDLE IdleWakeups::GetProcessHandle(const ProcessData& process_data) {
+  return process_id_to_hanle_map[process_data.ProcessId];
+}
+
+void IdleWakeups::OpenProcess(const ProcessData& process_data) {
+  process_id_to_hanle_map[process_data.ProcessId] =
+      ::OpenProcess(PROCESS_QUERY_LIMITED_INFORMATION, FALSE,
+                    (DWORD)(ULONGLONG)process_data.ProcessId);
+}
+
+void IdleWakeups::CloseProcess(const ProcessData& process) {
+  HANDLE handle = GetProcessHandle(process);
+  CloseHandle(handle);
+  process_id_to_hanle_map.erase(process.ProcessId);
+}
+
+double IdleWakeups::TakeSnapshot(ProcessesVector* processes) {
+  ULONG data_size;
+
+  LARGE_INTEGER perf_counter_value;
+  std::vector<BYTE> process_data(previous_buffer_size);
+
+  for (;;) {
+    data_size = 0;
+    NTSTATUS result = nt_query_system_information_ptr(
+        SystemProcessInformation,
+        previous_buffer_size > 0 ? &process_data[0] : NULL,
+        previous_buffer_size, &data_size);
+    if (result == STATUS_INFO_LENGTH_MISMATCH ||
+        result == STATUS_BUFFER_TOO_SMALL) {
+      // Reallocate the buffer.
+      previous_buffer_size = data_size;
+      process_data.resize(data_size);
+      continue;
+    }
+
+    if (result != STATUS_SUCCESS)
+      oops("NtQuerySystemInformation failed");
+
+    QueryPerformanceCounter(&perf_counter_value);
+    break;
+  }
+
+  for (size_t offset = 0; offset < data_size;) {
+    auto pi = reinterpret_cast<const SYSTEM_PROCESS_INFORMATION*>(
+        process_data.data() + offset);
+
+    // Validate that the offset is valid and all needed data is within
+    // the buffer boundary.
+    if (offset + sizeof(SYSTEM_PROCESS_INFORMATION) > data_size)
+      break;
+    if (offset + sizeof(SYSTEM_PROCESS_INFORMATION) +
+            (pi->NumberOfThreads - 1) * sizeof(SYSTEM_THREAD_INFORMATION) >
+        data_size)
+      break;
+
+    if (pi->ImageName.Buffer &&
+        wcsncmp(target_process_name_filter(), pi->ImageName.Buffer,
+                lstrlen(target_process_name_filter())) == 0) {
+      // There is no point in recording per-process idle wakeups with any more
+      // precision than they are recorded per-thread because the per-thread
+      // numbers may have wrapped. Therefore idleWakeups can only really be used
+      // to calculate diffs, or for processes that have not run long enough to
+      // accumulate four billion context switches.
+      ProcessData process_data;
+      process_data.ProcessId = pi->ProcessId;
+      process_data.CpuTime = pi->KernelTime + pi->UserTime;
+      process_data.WorkingSetPrivateSize = pi->WorkingSetPrivateSize;
+      for (ULONG thread_index = 0; thread_index < pi->NumberOfThreads;
+           ++thread_index) {
+        const SYSTEM_THREAD_INFORMATION* ti = &pi->Threads[thread_index];
+        if (ti->ClientId.UniqueProcess != pi->ProcessId)
+          continue;
+
+        ThreadData thread_data;
+        thread_data.ThreadId = ti->ClientId.UniqueThread;
+        thread_data.ContextSwitches = ti->ContextSwitchCount;
+        process_data.Threads.push_back(thread_data);
+      }
+
+      SortThreads(&process_data.Threads);
+
+      processes->push_back(process_data);
+    }
+
+    // Check for end of the list.
+    if (!pi->NextEntryOffset)
+      break;
+
+    // Jump to the next entry.
+    offset += pi->NextEntryOffset;
+  }
+
+  SortProcesses(processes);
+
+  double time_delta = 0.0;
+
+  if (previous_perf_counter_value.QuadPart != 0) {
+    time_delta = static_cast<double>(perf_counter_value.QuadPart -
+                                     previous_perf_counter_value.QuadPart) /
+                 perf_frequency.QuadPart;
+  }
+  previous_perf_counter_value = perf_counter_value;
+
+  return time_delta;
+}
+
+void IdleWakeups::SortThreads(ThreadsVector* threads) {
+  std::sort(threads->begin(), threads->end(),
+            [](const ThreadData& lhs, const ThreadData& rhs) {
+              return lhs.ThreadId < rhs.ThreadId;
+            });
+}
+
+void IdleWakeups::SortProcesses(ProcessesVector* processes) {
+  std::sort(processes->begin(), processes->end(),
+            [](const ProcessData& lhs, const ProcessData& rhs) {
+              return lhs.ProcessId < rhs.ProcessId;
+            });
+}
+
+ULONG IdleWakeups::CountContextSwitches(const ProcessData& process_data) {
+  ULONG context_switches = 0;
+
+  for (const auto& thread_data : process_data.Threads) {
+    context_switches += thread_data.ContextSwitches;
+  }
+
+  return context_switches;
+}
+
+ULONG IdleWakeups::DiffContextSwitches(const ProcessData& prev_process_data,
+                                       const ProcessData& process_data) {
+  ULONG context_switches = 0;
+  size_t prev_index = 0;
+
+  for (const auto& thread_data : process_data.Threads) {
+    ULONG prev_context_switches = 0;
+
+    for (; prev_index < prev_process_data.Threads.size(); ++prev_index) {
+      const auto& prev_thread_data = prev_process_data.Threads[prev_index];
+      if (prev_thread_data.ThreadId == thread_data.ThreadId) {
+        prev_context_switches = prev_thread_data.ContextSwitches;
+        ++prev_index;
+        break;
+      }
+
+      if (prev_thread_data.ThreadId > thread_data.ThreadId)
+        break;
+    }
+
+    context_switches += thread_data.ContextSwitches - prev_context_switches;
+  }
+
+  return context_switches;
+}
+
+bool IdleWakeups::GetFinishedProcessCpuTime(const ProcessData& process,
+                                            ULONGLONG* cpu_time) {
+  HANDLE process_handle = GetProcessHandle(process);
+
+  FILETIME creation_time, exit_time, kernel_time, user_time;
+  if (GetProcessTimes(process_handle, &creation_time, &exit_time, &kernel_time,
+                      &user_time)) {
+    ULARGE_INTEGER ul_kernel_time, ul_user_time;
+    ul_kernel_time.LowPart = kernel_time.dwLowDateTime;
+    ul_kernel_time.HighPart = kernel_time.dwHighDateTime;
+    ul_user_time.LowPart = user_time.dwLowDateTime;
+    ul_user_time.HighPart = user_time.dwHighDateTime;
+    *cpu_time = ul_kernel_time.QuadPart + ul_user_time.QuadPart;
+    return true;
+  }
+
+  *cpu_time = 0;
+  return false;
+}
+
+Result IdleWakeups::DiffSnapshots(double time_delta,
+                                  const ProcessesVector& prev_processes,
+                                  const ProcessesVector& processes) {
+  ULONG idle_wakeups_delta = 0;
+  ULONGLONG cpu_usage_delta = 0;
+  ULONGLONG total_working_set = 0;
+
+  size_t prev_index = 0;
+
+  for (const auto& process_data : processes) {
+    const ProcessData* prev_process_data_to_diff = nullptr;
+    ULONGLONG prev_process_cpu_time = 0;
+
+    for (; prev_index < prev_processes.size(); ++prev_index) {
+      const auto& prev_process_data = prev_processes[prev_index];
+
+      if (prev_process_data.ProcessId == process_data.ProcessId) {
+        prev_process_data_to_diff = &prev_process_data;
+        prev_process_cpu_time = prev_process_data.CpuTime;
+        ++prev_index;
+        break;
+      }
+
+      if (prev_process_data.ProcessId > process_data.ProcessId)
+        break;
+
+      // Prev process disappeared.
+      ULONGLONG last_known_cpu_time;
+      if (GetFinishedProcessCpuTime(prev_process_data, &last_known_cpu_time)) {
+        cpu_usage_delta += last_known_cpu_time - prev_process_data.CpuTime;
+      }
+      CloseProcess(prev_process_data);
+    }
+
+    if (prev_process_data_to_diff) {
+      idle_wakeups_delta +=
+          DiffContextSwitches(*prev_process_data_to_diff, process_data);
+    } else {
+      // New process that we haven't seen before.
+      OpenProcess(process_data);
+      idle_wakeups_delta += CountContextSwitches(process_data);
+    }
+
+    cpu_usage_delta += process_data.CpuTime - prev_process_cpu_time;
+    total_working_set += process_data.WorkingSetPrivateSize / 1024;
+  }
+
+  Result result;
+  result.IdleWakeupsPerSec =
+      static_cast<ULONG>(idle_wakeups_delta / time_delta);
+  // brucedawson: Don't divide by number of processors so that all numbers are
+  // percentage of a core
+  // result.CpuUsage = (double)cpu_usage_delta * 100 / (time_delta * 10000000 *
+  // NumberOfprocessors());
+  result.CpuUsage = (double)cpu_usage_delta * 100 / (time_delta * 10000000);
+  result.WorkingSet = total_working_set;
+
+  return result;
+}
+
+HANDLE ctrl_c_pressed = NULL;
+
+BOOL WINAPI HandlerFunction(DWORD ctrl_type) {
+  if (ctrl_type == CTRL_C_EVENT) {
+    printf("Ctrl+C pressed...\n");
+    SetEvent(ctrl_c_pressed);
+    return TRUE;
+  }
+
+  return FALSE;
+}
+
+const DWORD sleep_time_sec = 2;
+
+void PrintHeader() {
+  printf(
+      "------------------------------------------------------------------------"
+      "----------\n");
+  printf(
+      "                       Context switches/sec    CPU usage    Working set "
+      "     Power\n");
+  printf(
+      "------------------------------------------------------------------------"
+      "----------\n");
+}
+
+#define RESULT_FORMAT_STRING "    %20lu    %8.2f%c    %6.2f MiB    %4.2f W\n"
+
+int wmain(int argc, wchar_t* argv[]) {
+  ctrl_c_pressed = CreateEvent(NULL, FALSE, FALSE, NULL);
+  SetConsoleCtrlHandler(HandlerFunction, TRUE);
+
+  PowerSampler power_sampler;
+
+  IdleWakeups the_app(argc > 1 ? argv[1] : L"chrome.exe");
+
+  ProcessesVector prev_chrome_processes;
+  ProcessesVector chrome_processes;
+
+  // Take the initial snapshot.
+  the_app.TakeSnapshot(&chrome_processes);
+  the_app.OpenProcesses(chrome_processes);
+
+  ULONG cumulative_idle_wakeups_per_sec = 0;
+  double cumulative_cpu_usage = 0.0;
+  ULONGLONG cumulative_working_set = 0;
+  double cumulative_energy = 0.0;
+
+  ResultVector results;
+
+  printf("Capturing perf data for all processes matching %ls\n",
+         the_app.target_process_name_filter());
+
+  PrintHeader();
+
+  for (;;) {
+    prev_chrome_processes.swap(chrome_processes);
+    chrome_processes.clear();
+
+    if (WaitForSingleObject(ctrl_c_pressed, sleep_time_sec * 1000) ==
+        WAIT_OBJECT_0)
+      break;
+
+    double time_delta = the_app.TakeSnapshot(&chrome_processes);
+
+    Result result = the_app.DiffSnapshots(time_delta, prev_chrome_processes,
+                                          chrome_processes);
+    power_sampler.SampleCPUPowerState();
+    result.Power = power_sampler.get_power(L"Processor");
+
+    printf("%9u processes" RESULT_FORMAT_STRING, (DWORD)chrome_processes.size(),
+           result.IdleWakeupsPerSec, result.CpuUsage, '%',
+           result.WorkingSet / 1024.0, result.Power);
+
+    cumulative_idle_wakeups_per_sec += result.IdleWakeupsPerSec;
+    cumulative_cpu_usage += result.CpuUsage;
+    cumulative_working_set += result.WorkingSet;
+    cumulative_energy += result.Power;
+
+    results.push_back(result);
+  }
+
+  CloseHandle(ctrl_c_pressed);
+
+  ULONG sample_count = (ULONG)results.size();
+  if (sample_count == 0)
+    return 0;
+
+  PrintHeader();
+
+  printf("            Average" RESULT_FORMAT_STRING,
+         cumulative_idle_wakeups_per_sec / sample_count,
+         cumulative_cpu_usage / sample_count, '%',
+         (cumulative_working_set / 1024.0) / sample_count,
+         cumulative_energy / sample_count);
+
+  Result median_result;
+
+  median_result.IdleWakeupsPerSec =
+      GetMedian<ULONG>(&results, GetIdleWakeupsPerSec);
+  median_result.CpuUsage = GetMedian<double>(&results, GetCpuUsage);
+  median_result.WorkingSet = GetMedian<ULONGLONG>(&results, GetWorkingSet);
+  median_result.Power = GetMedian<double>(&results, GetPower);
+
+  printf("             Median" RESULT_FORMAT_STRING,
+         median_result.IdleWakeupsPerSec, median_result.CpuUsage, '%',
+         median_result.WorkingSet / 1024.0, median_result.Power);
+
+  return 0;
+}