ThreadWatcher - Added jankiness monitor for FILE thread.

chrome/browser/metrics/thread_watcher.h

Index: chrome/browser/metrics/thread_watcher.h
===================================================================
--- chrome/browser/metrics/thread_watcher.h	(revision 181723)
+++ chrome/browser/metrics/thread_watcher.h	(working copy)
@@ -308,30 +308,72 @@
   // A map from BrowserThread to the actual instances.
   typedef std::map<content::BrowserThread::ID, ThreadWatcher*> RegistrationList;
 
-  // A map from thread names (UI, IO, etc) to |live_threads_threshold|.
+  // A map from thread names (UI, IO, etc) to |CrashDataThresholds|.
   // |live_threads_threshold| specifies the maximum number of browser threads
   // that have to be responsive when we want to crash the browser because of
-  // hung watched thread.
+  // hung watched thread. This threshold allows us to either look for a system
+  // deadlock, or look for a solo hung thread. A small live_threads_threshold
+  // looks for a broad deadlock (few browser threads left running), and a large
+  // threshold looks for a single hung thread (this in only appropriate for a
+  // thread that *should* never have much jank, such as the IO).
   //
+  // |unresponsive_threshold| specifies the number of unanswered ping messages
+  // after which watched (UI, IO, etc) thread is considered as not responsive.
+  // We translate "time" (given in seconds) into a number of pings. As a result,
+  // we only declare a thread unresponsive when a lot of "time" has passed (many
+  // pings), and yet our pinging thread has continued to process messages (so we
+  // know the entire PC is not hung). Set this number higher to crash less
+  // often, and lower to crash more often.
+  //
   // The map lists all threads (by name) that can induce a crash by hanging. It
   // is populated from the command line, or given a default list.  See
   // InitializeAndStartWatching() for the separate list of all threads that are
   // watched, as they provide the system context of how hung *other* threads
   // are.
   //
-  // Example 1: If the value for "IO" was 3, then we would crash if at least one
-  // thread is responding and total responding threads is less than or equal to
-  // 3 (this thread, plus at least one other thread is unresponsive). We would
-  // not crash if none of the threads are not responding, as we'd assume such
-  // large hang counts mean that the system is generally unresponsive.
-  // Example 2: If the value for "UI" was INT_MAX, then we would always crash if
-  // the UI thread was hung, no matter what the other threads are doing.
-  // Example 3: If the value of "FILE" was 5, then we would only crash if the
-  // FILE thread was the ONLY hung thread (because we watch 6 threads). IF there
-  // was another unresponsive thread, we would not consider this a problem worth
-  // crashing for.
-  typedef std::map<std::string, uint32> CrashOnHangThreadMap;
+  // CrashOnHangThreadMap is populated by ParseCommandLineCrashOnHangThreads().
+  // It parses command line argument like "UI:3:18,IO:3:18,FILE:5:90". In this
+  // string, the first parameter specifies the thread_id: UI, IO or FILE. The
+  // second parameter specifies |live_threads_threshold|. For UI and IO threads,
+  // we would crash if the number of threads responding is less than or equal to
+  // 3. The third parameter specifies the unresponsive threshold seconds. This
+  // number is used to calculate |unresponsive_threshold|. In this example for
+  // UI and IO threads, we would crash if it doesn't respond for 18 seconds (or

[jar (doing other things), 2013/02/12 01:27:27]

nit: it doesn't --> those threads don't

[ramant (doing other things), 2013/02/12 01:54:30]

Done.

+  // 9 unanswered ping messages) and for FILE thread, crash_seconds is set to 90
+  // seconds (or 45 unanswered ping messages).
+  //
+  // ThreadWatcher watches six (UI, IO, DB, FILE, FILE_USER_BLOCKING and CACHE)

[jar (doing other things), 2013/02/12 01:27:27]

This paragraph should move up earlier, before you list the example on line 335
that starts "It parses command line..."

Change sentence to:
ThreadWatcher MONITORS six BROWSER THREADS (i.e., UI, IO, DB, FILE,
FILE_USER_BLOCKING and CACHE).  Out of the 6 threads, any subset may be watched,
to potentially cause a crash.  The following example's command line causes
exactly 3 threads to be watched.  The example command line argument consists of 
"UI...."  In that string, ....

[ramant (doing other things), 2013/02/12 01:54:30]

Done.

+  // browser threads. The following examples explain how the data in
+  // |CrashDataThresholds| controls the crashes.
+  //
+  // Example 1: If the |live_threads_threshold| value for "IO" was 3 and
+  // unresponsive threshold seconds is 18 (or |unresponsive_threshold| is 9),
+  // then we would crash if the IO thread was hung (9 unanswered ping messages)
+  // and if at least one thread is responding and total responding threads is
+  // less than or equal to 3 (this thread, plus at least one other thread is
+  // unresponsive). We would not crash if none of the threads are not
+  // responding, as we'd assume such large hang counts mean that the system is

[jar (doing other things), 2013/02/12 01:27:27]

nit: should have read:

We would not crash if none of the threads are responding.

There is an extra "not" before "responding" in your text.

[ramant (doing other things), 2013/02/12 01:54:30]

Done.

+  // generally unresponsive.
+  // Example 2: If the |live_threads_threshold| value for "UI" was INT_MAX and

[jar (doing other things), 2013/02/12 01:27:27]

Instead of INT_MAX, "any number higher than 6"

[ramant (doing other things), 2013/02/12 01:54:30]

Done.

+  // unresponsive threshold seconds is 18 (or |unresponsive_threshold| is 9),
+  // then we would always crash if the UI thread was hung (9 unanswered ping
+  // messages), no matter what the other threads are doing.
+  // Example 3: If the |live_threads_threshold| value of "FILE" was 5 and
+  // unresponsive threshold seconds is 90 (or |unresponsive_threshold| is 45),
+  // then we would only crash if the FILE thread was the ONLY hung thread
+  // (because we watch 6 threads). If there was another unresponsive thread, we
+  // would not consider this a problem worth crashing for. FILE thread would be
+  // considered as hung if it didn't respond for 45 ping messages.
+  struct CrashDataThresholds {
+    CrashDataThresholds(uint32 live_threads_threshold,
+                        uint32 unresponsive_threshold);
+    CrashDataThresholds();
 
+    uint32 live_threads_threshold;
+    uint32 unresponsive_threshold;
+  };
+  typedef std::map<std::string, CrashDataThresholds> CrashOnHangThreadMap;
+
   // This method posts a task on WatchDogThread to start watching all browser
   // threads.
   // This method is accessible on UI thread.
@@ -361,7 +403,9 @@
   // Allow tests to access our innards for testing purposes.
   friend class CustomThreadWatcher;
   friend class ThreadWatcherTest;
-  FRIEND_TEST_ALL_PREFIXES(ThreadWatcherTest, CommandLineArgs);
+  FRIEND_TEST_ALL_PREFIXES(ThreadWatcherTest, ThreadNamesOnlyArgs);
+  FRIEND_TEST_ALL_PREFIXES(ThreadWatcherTest, ThreadNamesAndLiveThresholdArgs);
+  FRIEND_TEST_ALL_PREFIXES(ThreadWatcherTest, CrashOnHangThreadsAllArgs);
 
   // This singleton holds the global list of registered ThreadWatchers.
   ThreadWatcherList();
@@ -369,16 +413,26 @@
   // Destructor deletes all registered ThreadWatcher instances.
   virtual ~ThreadWatcherList();
 
-  // Parses the command line to get |unresponsive_threshold| from
-  // switches::kCrashOnHangSeconds, |crash_on_hang| thread names from
-  // switches::kCrashOnHangThreads and |live_threads_threshold| from
-  // switches::kCrashOnLive. |crash_on_hang_threads| is a map of
-  // |crash_on_hang| thread's names to |live_threads_threshold|.
+  // Parses the command line to get |crash_on_hang_threads| map from
+  // switches::kCrashOnHangThreads. |crash_on_hang_threads| is a map of
+  // |crash_on_hang| thread's names to |CrashDataThresholds|.
   static void ParseCommandLine(
       const CommandLine& command_line,
       uint32* unresponsive_threshold,
       CrashOnHangThreadMap* crash_on_hang_threads);
 
+  // Parses the argument |crash_on_hang_thread_names| and creates
+  // |crash_on_hang_threads| map of |crash_on_hang| thread's names to
+  // |CrashDataThresholds|. If |crash_on_hang_thread_names| doesn't specify
+  // |live_threads_threshold|, then it uses |default_live_threads_threshold| as
+  // the value. If |crash_on_hang_thread_names| doesn't specify |crash_seconds|,
+  // then it uses |default_crash_seconds| as the value.
+  static void ParseCommandLineCrashOnHangThreads(
+      const std::string& crash_on_hang_thread_names,
+      uint32 default_live_threads_threshold,
+      uint32 default_crash_seconds,
+      CrashOnHangThreadMap* crash_on_hang_threads);
+
   // This constructs the |ThreadWatcherList| singleton and starts watching
   // browser threads by calling StartWatching() on each browser thread that is
   // watched. It disarms StartupTimeBomb.