Add UMA stats for hard-fault counts for Windows 7 and greater.

components/startup_metric_utils/startup_metric_utils.cc

 // Copyright 2013 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 "components/startup_metric_utils/startup_metric_utils.h"
 
 #include "base/containers/hash_tables.h"
 #include "base/environment.h"
 #include "base/logging.h"
 #include "base/metrics/histogram.h"
 #include "base/metrics/histogram_base.h"
 #include "base/metrics/statistics_recorder.h"
 #include "base/process/process_info.h"
 #include "base/strings/string_number_conversions.h"
 #include "base/synchronization/lock.h"
 #include "base/sys_info.h"
 #include "base/time/time.h"
 
+#if defined(OS_WIN)
+#include <winternl.h>
+#endif
+
 namespace {
 
 // Mark as volatile to defensively make sure usage is thread-safe.
 // Note that at the time of this writing, access is only on the UI thread.
 volatile bool g_non_browser_ui_displayed = false;
 
 base::Time* MainEntryPointTimeInternal() {
   static base::Time main_start_time = base::Time::Now();
   return &main_start_time;
 }
 
 typedef base::hash_map<std::string,base::TimeDelta> SubsystemStartupTimeHash;
 
 SubsystemStartupTimeHash* GetSubsystemStartupTimeHash() {
   static SubsystemStartupTimeHash* slow_startup_time_hash =
                                     new SubsystemStartupTimeHash;
   return slow_startup_time_hash;
 }
 
 base::Lock* GetSubsystemStartupTimeHashLock() {
   static base::Lock* slow_startup_time_hash_lock = new base::Lock;
   return slow_startup_time_hash_lock;
 }
 
+#if defined(OS_WIN)
+
+// The struct used to return system process information via the NT internal
+// QuerySystemInformation call. This is partially documented at
+// http://goo.gl/Ja9MrH and fully documented at http://goo.gl/QJ70rn
+// This structure is laid out in the same format on both 32-bit and 64-bit
+// systems, but has a different size due to the various pointer-sized fields.
+struct SYSTEM_PROCESS_INFORMATION_EX {
+  ULONG NextEntryOffset;
+  ULONG NumberOfThreads;
+  LARGE_INTEGER WorkingSetPrivateSize;
+  ULONG HardFaultCount;
+  BYTE Reserved1[36];
+  PVOID Reserved2[3];
+  // This is labeled a handle so that it expands to the correct size for 32-bit
+  // and 64-bit operating systems. However, under the hood it's a 32-bit DWORD
+  // containing the process ID.
+  HANDLE UniqueProcessId;
+  PVOID Reserved3;
+  ULONG HandleCount;
+  BYTE Reserved4[4];
+  PVOID Reserved5[11];
+  SIZE_T PeakPagefileUsage;
+  SIZE_T PrivatePageCount;
+  LARGE_INTEGER Reserved6[6];
+  // Array of SYSTEM_THREAD_INFORMATION structs follows.
+};
+
+// The signature of the NtQuerySystemInformation function.
+typedef NTSTATUS (WINAPI *NtQuerySystemInformationPtr)(
+    SYSTEM_INFORMATION_CLASS, PVOID, ULONG, PULONG);
+
+// Gets the hard fault count of the current process, returning it via
+// |hard_fault_count|. Returns true on success, false otherwise. Also returns
+// whether or not the system call was even possible for the current OS version
+// via |has_os_support|.
+bool GetHardFaultCountForCurrentProcess(uint32* hard_fault_count,

[Alexei Svitkine (slow), 2015/04/15 17:36:19]

Nit: uint32_t is best practice for new code.

[chrisha, 2015/04/15 17:56:01]

Done.

+                                        bool* has_os_support) {
+  DCHECK(hard_fault_count);
+  DCHECK(has_os_support);
+
+  *has_os_support = false;
+  OSVERSIONINFOEX version_info = {};
+  version_info.dwOSVersionInfoSize = sizeof(OSVERSIONINFOEX);
+  if (!::GetVersionEx(reinterpret_cast<LPOSVERSIONINFOW>(&version_info)))

[Alexei Svitkine (slow), 2015/04/15 17:36:19]

Can you use base/win/windows_version.h instead of doing it manually?

[chrisha, 2015/04/15 17:56:01]

I certainly can, and it's far simpler that way. Thanks!

Done.

+    return false;
+
+  // The hard fault counter is only available as of Windows 7 (version 6.1).
+  if (version_info.dwMajorVersion < 6)
+    return false;
+  if (version_info.dwMajorVersion == 6 && version_info.dwMinorVersion < 1)
+    return false;
+
+  // At this point the OS supports the required system call.
+  *has_os_support = true;
+
+  // Get the function pointer.
+  NtQuerySystemInformationPtr query_sys_info =
+      reinterpret_cast<NtQuerySystemInformationPtr>(
+          ::GetProcAddress(GetModuleHandle(L"ntdll.dll"),
+                           "NtQuerySystemInformation"));
+  if (query_sys_info == nullptr)
+    return false;
+
+  // The output of this system call depends on the number of threads and
+  // processes on the entire system, and this can change between calls. Retry
+  // a small handful of times growing the buffer along the way.
+  // NOTE: The actual required size depends entirely on the number of processes
+  //       and threads running on the system. The initial guess suffices for
+  //       ~100s of processes and ~1000s of threads.
+  std::vector<uint8> buffer(32 * 1024);
+  for (size_t tries = 0; tries < 3; ++tries) {
+    ULONG return_length = 0;
+    NTSTATUS status = query_sys_info(
+        SystemProcessInformation,
+        buffer.data(),
+        static_cast<ULONG>(buffer.size()),
+        &return_length);
+    // Insufficient space in the buffer.
+    if (return_length > buffer.size()) {
+      buffer.resize(return_length);
+      continue;
+    }
+    if (status == 0 && return_length <= buffer.size())

[Alexei Svitkine (slow), 2015/04/15 17:36:19]

Should use NT_SUCCESS() instead of status == 0 I think.

[chrisha, 2015/04/15 17:56:01]

Done.

+      break;
+    return false;
+  }
+
+  // Look for the struct housing information for the current process.
+  DWORD proc_id = ::GetCurrentProcessId();
+  size_t index = 0;
+  while (index < buffer.size()) {
+    SYSTEM_PROCESS_INFORMATION_EX* proc_info =
+        reinterpret_cast<SYSTEM_PROCESS_INFORMATION_EX*>(buffer.data() + index);
+    if (reinterpret_cast<DWORD>(proc_info->UniqueProcessId) == proc_id) {
+      *hard_fault_count = proc_info->HardFaultCount;
+      return true;
+    }
+    // The list ends with NextEntryOffset is zero. This also prevents busy
+    // looping if the data is in fact invalid.
+    if (proc_info->NextEntryOffset == 0)
+      return false;
+    index += proc_info->NextEntryOffset;
+  }
+
+  return false;
+}
+
+#endif  // defined(OS_WIN)
+
+// On Windows, records the number of hard-faults that have occurred in the
+// current chrome.exe process since it was started. This is a nop on other
+// platforms.
+// crbug.com/476923
+// TODO(chrisha): If this proves useful, use it to split startup stats in two.
+void RecordHardFaultHistogram(bool is_first_run) {
+#if defined(OS_WIN)
+  uint32 hard_fault_count = 0;
+  bool has_os_support = false;
+  bool success = GetHardFaultCountForCurrentProcess(
+      &hard_fault_count, &has_os_support);
+
+  // Log whether or not the system call was successful, assuming the OS was
+  // detected to support it.
+  if (has_os_support) {
+    UMA_HISTOGRAM_BOOLEAN(
+        "Startup.BrowserMessageLoopStartHardFaultCount.Success",
+        success);

[Alexei Svitkine (slow), 2015/04/15 17:36:19]

Since there are different cases available, I suggest making a custom enum
histogram outlining the possible ones, which would provide all the different
things we might be interested in.

i.e.

NO_OS_SUPPORT,
SUCCEEDED,
FAILED,

[chrisha, 2015/04/15 17:56:01]

The "no OS support" case is completely deterministic (depends only on the OS
version), and doesn't really provide any useful information. I'm not sure what
use it would have?

[Alexei Svitkine (slow), 2015/04/15 18:08:10]

I was thinking it would be simpler to reason about what percentage of users fall
into each case - but that has more value when this actually affects behavior -
since we're only logging the value, current set up seems fine to me now that I
think about it again. Thanks.

+  }
+
+  // Don't log a histogram value if unable to get the hard fault count.
+  if (!success)
+    return;
+
+  // Hard fault counts are expected to be in the thousands range,
+  // corresponding to faulting in ~10s of MBs of code ~10s of KBs at a time.
+  // (Observed to vary from 1000 to 10000 on various test machines and
+  // platforms.)
+  if (is_first_run) {
+    UMA_HISTOGRAM_CUSTOM_COUNTS(
+        "Startup.BrowserMessageLoopStartHardFaultCount.FirstRun",
+        hard_fault_count,
+        0, 40000, 50);
+  } else {
+    UMA_HISTOGRAM_CUSTOM_COUNTS(
+        "Startup.BrowserMessageLoopStartHardFaultCount",
+        hard_fault_count,
+        0, 40000, 50);
+  }
+#endif  // defined(OS_WIN)
+}
+
 // Record time of main entry so it can be read from Telemetry performance
 // tests.
 // TODO(jeremy): Remove once crbug.com/317481 is fixed.
 void RecordMainEntryTimeHistogram() {
   const int kLowWordMask = 0xFFFFFFFF;
   const int kLower31BitsMask = 0x7FFFFFFF;
   base::TimeDelta browser_main_entry_time_absolute =
       *MainEntryPointTimeInternal() - base::Time::UnixEpoch();
 
   uint64 browser_main_entry_time_raw_ms =
@@ skipping 55 matching lines @@
 }
 #endif // OS_ANDROID
 
 // Return the time recorded by RecordMainEntryPointTime().
 const base::Time MainEntryStartTime() {
   DCHECK(g_main_entry_time_was_recorded);
   return *MainEntryPointTimeInternal();
 }
 
 void OnBrowserStartupComplete(bool is_first_run) {
+  RecordHardFaultHistogram(is_first_run);
   RecordMainEntryTimeHistogram();
 
   // Bail if uptime < 7 minutes, to filter out cases where Chrome may have been
   // autostarted and the machine is under io pressure.
   const int64 kSevenMinutesInMilliseconds =
       base::TimeDelta::FromMinutes(7).InMilliseconds();
   if (base::SysInfo::Uptime() < kSevenMinutesInMilliseconds) {
     g_startup_stats_collection_finished = true;
     return;
   }
@@ skipping 114 matching lines @@
   SubsystemStartupTimeHash* hash = GetSubsystemStartupTimeHash();
   // Only record the initial sample for a given histogram.
   if (hash->find(histogram_name_) !=  hash->end())
     return;
 
   (*hash)[histogram_name_] =
       base::TimeTicks::Now() - start_time_;
 }
 
 }  // namespace startup_metric_utils