Thread-local trace-event buffers

base/debug/trace_event_impl.cc

 // Copyright (c) 2012 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/debug/trace_event_impl.h"
 
 #include <algorithm>
 
 #include "base/base_switches.h"
 #include "base/bind.h"
 #include "base/command_line.h"
 #include "base/debug/leak_annotations.h"
 #include "base/debug/trace_event.h"
 #include "base/format_macros.h"
 #include "base/lazy_instance.h"
 #include "base/memory/singleton.h"
+#include "base/message_loop/message_loop.h"
 #include "base/process/process_metrics.h"
 #include "base/stl_util.h"
 #include "base/strings/string_split.h"
 #include "base/strings/string_tokenizer.h"
 #include "base/strings/string_util.h"
 #include "base/strings/stringprintf.h"
 #include "base/strings/utf_string_conversions.h"
 #include "base/synchronization/cancellation_flag.h"
 #include "base/synchronization/waitable_event.h"
 #include "base/sys_info.h"
 #include "base/third_party/dynamic_annotations/dynamic_annotations.h"
 #include "base/threading/platform_thread.h"
 #include "base/threading/thread_id_name_manager.h"
-#include "base/threading/thread_local.h"
 #include "base/time/time.h"
 
 #if defined(OS_WIN)
 #include "base/debug/trace_event_win.h"
 #endif
 
 class DeleteTraceLogForTesting {
  public:
   static void Delete() {
     Singleton<base::debug::TraceLog,
               LeakySingletonTraits<base::debug::TraceLog> >::OnExit(0);
   }
 };
 
 // The thread buckets for the sampling profiler.
 BASE_EXPORT TRACE_EVENT_API_ATOMIC_WORD g_trace_state[3];
 
 namespace base {
 namespace debug {
 
+namespace {
+
+// The overhead of TraceEvent above this threshold will be reported in the
+// trace.
+const int kOverheadReportThresholdInMicroseconds = 50;
+
 // Controls the number of trace events we will buffer in-memory
 // before throwing them away.
 const size_t kTraceEventVectorBufferSize = 250000;
 const size_t kTraceEventRingBufferSize = kTraceEventVectorBufferSize / 4;
+const size_t kTraceEventThreadLocalBufferSize = 1024;
 const size_t kTraceEventBatchSize = 1000;
 const size_t kTraceEventInitialBufferSize = 1024;
 
 #define MAX_CATEGORY_GROUPS 100
 
-namespace {
-
 // Parallel arrays g_category_groups and g_category_group_enabled are separate
 // so that a pointer to a member of g_category_group_enabled can be easily
 // converted to an index into g_category_groups. This allows macros to deal
 // only with char enabled pointers from g_category_group_enabled, and we can
 // convert internally to determine the category name from the char enabled
 // pointer.
 const char* g_category_groups[MAX_CATEGORY_GROUPS] = {
   "tracing already shutdown",
   "tracing categories exhausted; must increase MAX_CATEGORY_GROUPS",
   "__metadata",
-};
+  // For reporting trace_event overhead. For thread local event buffers only.

[dsinclair, 2013/08/29 14:24:32]

After this patch, will we have non-thread-local event buffers? If not, let's
just drop this comment.

[Xianzhu, 2013/08/29 17:42:12]

Yes, trace events from a thread without a message loop (e.g. some short-lived
worker thread) will still use the main buffer directly.

+  "trace_event_overhead"};
 
 // The enabled flag is char instead of bool so that the API can be used from C.
 unsigned char g_category_group_enabled[MAX_CATEGORY_GROUPS] = { 0 };
 const int g_category_already_shutdown = 0;
 const int g_category_categories_exhausted = 1;
 const int g_category_metadata = 2;
-const int g_num_builtin_categories = 3;
+const int g_category_trace_event_overhead = 3;
+const int g_num_builtin_categories = 4;
 int g_category_index = g_num_builtin_categories; // Skip default categories.
 
 // The name of the current thread. This is used to decide if the current
 // thread name has changed. We combine all the seen thread names into the
 // output name for the thread.
 LazyInstance<ThreadLocalPointer<const char> >::Leaky
     g_current_thread_name = LAZY_INSTANCE_INITIALIZER;
 
 const char kRecordUntilFull[] = "record-until-full";
 const char kRecordContinuously[] = "record-continuously";
 const char kEnableSampling[] = "enable-sampling";
 
-}  // namespace
+TimeTicks ThreadNow() {
+  return TimeTicks::IsThreadNowSupported() ?
+         TimeTicks::ThreadNow() : TimeTicks();

[dsinclair, 2013/08/29 14:24:32]

nit: indented too far.

[Xianzhu, 2013/08/29 17:42:12]

Done.

+}
 
 class TraceBufferRingBuffer : public TraceBuffer {
  public:
   TraceBufferRingBuffer()
       : unused_event_index_(0),
         oldest_event_index_(0) {
     logged_events_.reserve(kTraceEventInitialBufferSize);
   }
 
   virtual ~TraceBufferRingBuffer() {}
@@ skipping 73 matching lines @@
 class TraceBufferVector : public TraceBuffer {
  public:
   TraceBufferVector() : current_iteration_index_(0) {
     logged_events_.reserve(kTraceEventInitialBufferSize);
   }
 
   virtual ~TraceBufferVector() {
   }
 
   virtual void AddEvent(const TraceEvent& event) OVERRIDE {
-    // Note, we have two callers which need to be handled. The first is
-    // AddTraceEventWithThreadIdAndTimestamp() which checks Size() and does an
-    // early exit if full. The second is AddThreadNameMetadataEvents().
+    // Note, we have two callers which need to be handled:
+    // - AddEventToMainBufferWhileLocked() which has two cases:
+    //   - called directly from AddTraceEventWithThreadIdAndTimeStamp()
+    //     which checks if buffer is full and does an early exit if full;
+    //   - called from ThreadLocalEventBuffer::FlushWhileLocked();
+    // - AddThreadNameMetadataEvents().
     // We can not DECHECK(!IsFull()) because we have to add the metadata
-    // events even if the buffer is full.
+    // events and flush thread-local buffers even if the buffer is full.
     logged_events_.push_back(event);
   }
 
   virtual bool HasMoreEvents() const OVERRIDE {
     return current_iteration_index_ < Size();
   }
 
   virtual const TraceEvent& NextEvent() OVERRIDE {
     DCHECK(HasMoreEvents());
     return GetEventAt(current_iteration_index_++);
@@ skipping 61 matching lines @@
 
   // As this buffer is never full, we can return any positive number.
   virtual size_t Capacity() const OVERRIDE { return 1; }
 
   virtual const TraceEvent& GetEventAt(size_t index) const OVERRIDE {
     NOTREACHED();
     return *static_cast<TraceEvent*>(NULL);
   }
 };
 
+}  // namespace
+
 ////////////////////////////////////////////////////////////////////////////////
 //
 // TraceEvent
 //
 ////////////////////////////////////////////////////////////////////////////////
 
 namespace {
 
 size_t GetAllocLength(const char* str) { return str ? strlen(str) + 1 : 0; }
 
@@ skipping 475 matching lines @@
 
 TraceBucketData::~TraceBucketData() {
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 //
 // TraceLog
 //
 ////////////////////////////////////////////////////////////////////////////////
 
+class TraceLog::ThreadLocalEventBuffer
+    : public MessageLoop::DestructionObserver {
+ public:
+  ThreadLocalEventBuffer(TraceLog* trace_log);
+  virtual ~ThreadLocalEventBuffer();
+
+  void AddEvent(const TraceEvent& event, NotificationHelper* notifier);
+  void ReportOverhead(const TimeTicks& event_timestamp,
+                      const TimeTicks& event_thread_timestamp);
+
+ private:
+  // MessageLoop::DestructionObserver
+  virtual void WillDestroyCurrentMessageLoop() OVERRIDE;
+
+  void FlushWhileLocked(NotificationHelper* notifier);
+
+  void CheckThisIsCurrentBuffer() {
+    DCHECK(trace_log_->thread_local_event_buffer_.Get() == this);
+  }
+
+  // Since TraceLog is a leaky singleton, trace_log_ will always be valid
+  // as long as the thread exists.
+  TraceLog* trace_log_;
+  std::vector<TraceEvent> logged_events_;
+  int event_count_;
+  TimeDelta overhead_;
+
+  DISALLOW_COPY_AND_ASSIGN(ThreadLocalEventBuffer);
+};
+
+TraceLog::ThreadLocalEventBuffer::ThreadLocalEventBuffer(TraceLog* trace_log)
+    : trace_log_(trace_log),
+      event_count_(0) {
+  logged_events_.reserve(kTraceEventThreadLocalBufferSize);
+
+  if (g_category_group_enabled[g_category_trace_event_overhead]) {
+    int thread_id = static_cast<int>(PlatformThread::CurrentId());
+    logged_events_.push_back(TraceEvent(
+        thread_id,
+        TimeTicks::NowFromSystemTraceTime() - trace_log->time_offset_,
+        ThreadNow(),
+        TRACE_EVENT_PHASE_ASYNC_BEGIN,
+        &g_category_group_enabled[g_category_trace_event_overhead],
+        "thread_trace_event",
+        thread_id,
+        0, NULL, NULL, NULL, NULL,
+        TRACE_EVENT_FLAG_HAS_ID));
+  }
+
+  // ThreadLocalEventBuffer is created only if the thread has message loop, so

[dsinclair, 2013/08/29 14:24:32]

nit: if the thread has a message loop

[Xianzhu, 2013/08/29 17:42:12]

Done.

+  // the following message_loop won't be NULL.
+  MessageLoop* message_loop = MessageLoop::current();
+  message_loop->AddDestructionObserver(this);
+
+  AutoLock lock(trace_log->lock_);
+  trace_log->thread_message_loops_.insert(message_loop);
+}
+
+TraceLog::ThreadLocalEventBuffer::~ThreadLocalEventBuffer() {
+  CheckThisIsCurrentBuffer();
+  MessageLoop::current()->RemoveDestructionObserver(this);
+
+  if (event_count_) {
+    const char* arg_names[2] = { "event_count", "average_overhead" };
+    unsigned char arg_types[2];
+    unsigned long long arg_values[2];
+    trace_event_internal::SetTraceValue(
+        event_count_, &arg_types[0], &arg_values[0]);
+    trace_event_internal::SetTraceValue(
+        overhead_.InMillisecondsF() / event_count_,

[dsinclair, 2013/08/29 14:24:32]

Does overhead_ get an initial value of 0 by default? What happens if
ReportOverhead is never called for some reason?

[Xianzhu, 2013/08/29 17:42:12]

Yes, set by the default ctor of TimeDelta.
If event_count_ is zero, this block won't be executed. The trace-viewer will
discard the async begin event sent in the constructor because there is no
matching async end.

This happens in either of the following cases:
- no event generated for the thread;
- the thread has no message loop;
- trace_event_overhead is disabled;

Added comments above the if (event_count_) condition.

+        &arg_types[1], &arg_values[1]);
+    int thread_id = static_cast<int>(PlatformThread::CurrentId());
+    logged_events_.push_back(TraceEvent(
+        thread_id,
+        TimeTicks::NowFromSystemTraceTime() - trace_log_->time_offset_,
+        ThreadNow(),
+        TRACE_EVENT_PHASE_ASYNC_END,
+        &g_category_group_enabled[g_category_trace_event_overhead],
+        "thread_trace_event",
+        thread_id,
+        2, arg_names, arg_types, arg_values, NULL,
+        TRACE_EVENT_FLAG_HAS_ID));
+  }
+
+  NotificationHelper notifier(trace_log_);
+  {
+    AutoLock lock(trace_log_->lock_);
+    FlushWhileLocked(&notifier);
+    trace_log_->thread_message_loops_.erase(MessageLoop::current());
+  }
+  notifier.SendNotificationIfAny();
+}
+
+void TraceLog::ThreadLocalEventBuffer::AddEvent(const TraceEvent& event,
+                                                NotificationHelper* notifier) {
+  CheckThisIsCurrentBuffer();
+  logged_events_.push_back(event);
+  if (logged_events_.size() >= kTraceEventThreadLocalBufferSize) {
+    AutoLock lock(trace_log_->lock_);
+    FlushWhileLocked(notifier);
+  }
+}
+
+void TraceLog::ThreadLocalEventBuffer::ReportOverhead(
+    const TimeTicks& event_timestamp, const TimeTicks& event_thread_timestamp) {
+  if (g_category_group_enabled[g_category_trace_event_overhead]) {

[dsinclair, 2013/08/29 14:24:32]

if (!g_category_group_enabled[g_category_trace_event_overhead])
  return;

[Xianzhu, 2013/08/29 17:42:12]

Done.

+    event_count_++;
+    TimeTicks now =
+        TimeTicks::NowFromSystemTraceTime() - trace_log_->time_offset_;
+    TimeDelta overhead = now - event_timestamp;
+    if (overhead.InMicroseconds() >= kOverheadReportThresholdInMicroseconds) {
+      int thread_id = static_cast<int>(PlatformThread::CurrentId());
+      // TODO(wangxianzhu): Use X event when it's ready.
+      logged_events_.push_back(TraceEvent(
+          thread_id, event_timestamp, event_thread_timestamp,
+          TRACE_EVENT_PHASE_BEGIN,
+          &g_category_group_enabled[g_category_trace_event_overhead],
+          "overhead",
+          0, 0, NULL, NULL, NULL, NULL, 0));
+      logged_events_.push_back(TraceEvent(
+          thread_id, now, ThreadNow(),
+          TRACE_EVENT_PHASE_END,
+          &g_category_group_enabled[g_category_trace_event_overhead],
+          "overhead",
+          0, 0, NULL, NULL, NULL, NULL, 0));
+    }
+    overhead_ += overhead;
+  }
+}
+
+void TraceLog::ThreadLocalEventBuffer::WillDestroyCurrentMessageLoop() {
+  delete this;
+}
+
+void TraceLog::ThreadLocalEventBuffer::FlushWhileLocked(
+    NotificationHelper* notifier) {
+  trace_log_->lock_.AssertAcquired();
+  for (size_t i = 0; i < logged_events_.size(); ++i) {
+    trace_log_->AddEventToMainBufferWhileLocked(logged_events_[i]);
+  }
+  logged_events_.resize(0);
+  trace_log_->CheckIfBufferIsFullWhileLocked(notifier);
+}
+
 TraceLog::NotificationHelper::NotificationHelper(TraceLog* trace_log)
     : trace_log_(trace_log),
       notification_(0) {
 }
 
 TraceLog::NotificationHelper::~NotificationHelper() {
 }
 
 void TraceLog::NotificationHelper::AddNotificationWhileLocked(
     int notification) {
+  trace_log_->lock_.AssertAcquired();
   if (trace_log_->notification_callback_.is_null())
     return;
   if (notification_ == 0)
     callback_copy_ = trace_log_->notification_callback_;
   notification_ |= notification;
 }
 
 void TraceLog::NotificationHelper::SendNotificationIfAny() {
   if (notification_)
     callback_copy_.Run(notification_);
@@ skipping 26 matching lines @@
   }
   if (!(ret & RECORD_UNTIL_FULL) && !(ret & RECORD_CONTINUOUSLY))
     ret |= RECORD_UNTIL_FULL;  // Default when no options are specified.
 
   return static_cast<Options>(ret);
 }
 
 TraceLog::TraceLog()
     : enable_count_(0),
       num_traces_recorded_(0),
-      event_callback_(NULL),
+      buffer_is_full_(0),
+      event_callback_(0),
       dispatching_to_observer_list_(false),
       process_sort_index_(0),
-      watch_category_(NULL),
+      process_id_hash_(0),
+      process_id_(0),
+      watch_category_(0),
       trace_options_(RECORD_UNTIL_FULL),
       sampling_thread_handle_(0),
       category_filter_(CategoryFilter::kDefaultCategoryFilterString) {
   // Trace is enabled or disabled on one thread while other threads are
   // accessing the enabled flag. We don't care whether edge-case events are
   // traced or not, so we allow races on the enabled flag to keep the trace
   // macros fast.
   // TODO(jbates): ANNOTATE_BENIGN_RACE_SIZED crashes windows TSAN bots:
   // ANNOTATE_BENIGN_RACE_SIZED(g_category_group_enabled,
   //                            sizeof(g_category_group_enabled),
@@ skipping 117 matching lines @@
       category_group_enabled =
           &g_category_group_enabled[g_category_categories_exhausted];
     }
   }
   return category_group_enabled;
 }
 
 void TraceLog::GetKnownCategoryGroups(
     std::vector<std::string>* category_groups) {
   AutoLock lock(lock_);
+  category_groups->push_back(
+      g_category_groups[g_category_trace_event_overhead]);
   for (int i = g_num_builtin_categories; i < g_category_index; i++)
     category_groups->push_back(g_category_groups[i]);
 }
 
 void TraceLog::SetEnabled(const CategoryFilter& category_filter,
                           Options options) {
   std::vector<EnabledStateObserver*> observer_list;
   {
     AutoLock lock(lock_);
 
+    Options old_options = trace_options();
+
     if (enable_count_++ > 0) {
-      if (options != trace_options_) {
+      if (options != old_options) {
         DLOG(ERROR) << "Attemting to re-enable tracing with a different "
                     << "set of options.";
       }
 
       category_filter_.Merge(category_filter);
       UpdateCategoryGroupEnabledFlags();
       return;
     }
 
-    if (options != trace_options_) {
-      trace_options_ = options;
+    if (options != old_options) {
+      subtle::NoBarrier_Store(&trace_options_, options);
       logged_events_.reset(GetTraceBuffer());
+      subtle::NoBarrier_Store(&buffer_is_full_, 0);
     }
 
     if (dispatching_to_observer_list_) {
       DLOG(ERROR) <<
           "Cannot manipulate TraceLog::Enabled state from an observer.";
       return;
     }
 
     num_traces_recorded_++;
 
@@ skipping 37 matching lines @@
   AutoLock lock(lock_);
   DCHECK(enable_count_ > 0);
   return category_filter_;
 }
 
 void TraceLog::SetDisabled() {
   std::vector<EnabledStateObserver*> observer_list;
   {
     AutoLock lock(lock_);
     DCHECK(enable_count_ > 0);
+
     if (--enable_count_ != 0)
       return;
 
     if (dispatching_to_observer_list_) {
       DLOG(ERROR)
           << "Cannot manipulate TraceLog::Enabled state from an observer.";
       return;
     }
 
     if (sampling_thread_.get()) {
       // Stop the sampling thread.
       sampling_thread_->Stop();
       lock_.Release();
       PlatformThread::Join(sampling_thread_handle_);
       lock_.Acquire();
       sampling_thread_handle_ = PlatformThreadHandle();
       sampling_thread_.reset();
     }
 
     category_filter_.Clear();
-    watch_category_ = NULL;
+    subtle::NoBarrier_Store(&watch_category_, 0);
     watch_event_name_ = "";
     UpdateCategoryGroupEnabledFlags();
     AddMetadataEvents();
 
     dispatching_to_observer_list_ = true;
     observer_list = enabled_state_observer_list_;
   }
 
   // Dispatch to observers outside the lock in case the observer triggers a
   // trace event.
@@ skipping 39 matching lines @@
                             logged_events_->Capacity());
 }
 
 void TraceLog::SetNotificationCallback(
     const TraceLog::NotificationCallback& cb) {
   AutoLock lock(lock_);
   notification_callback_ = cb;
 }
 
 TraceBuffer* TraceLog::GetTraceBuffer() {
-  if (trace_options_ & RECORD_CONTINUOUSLY)
+  Options options = trace_options();
+  if (options & RECORD_CONTINUOUSLY)
     return new TraceBufferRingBuffer();
-  else if (trace_options_ & ECHO_TO_CONSOLE)
+  else if (options & ECHO_TO_CONSOLE)
     return new TraceBufferDiscardsEvents();
   return new TraceBufferVector();
 }
 
+void TraceLog::AddEventToMainBufferWhileLocked(const TraceEvent& trace_event) {
+  // Don't check buffer_is_full_ because we want the remaining thread-local
+  // events to be flushed into the main buffer with this method, otherwise
+  // we may lose some early events of a thread that generates events sparsely.
+  lock_.AssertAcquired();
+  logged_events_->AddEvent(trace_event);
+}
+
+void TraceLog::CheckIfBufferIsFullWhileLocked(NotificationHelper* notifier) {

[dsinclair, 2013/08/29 14:24:32]

lock_.AssertAcquired();

[Xianzhu, 2013/08/29 17:42:12]

Done.

+  if (!subtle::NoBarrier_Load(&buffer_is_full_) && logged_events_->IsFull()) {
+    subtle::NoBarrier_Store(&buffer_is_full_,
+                            static_cast<subtle::AtomicWord>(1));
+    notifier->AddNotificationWhileLocked(TRACE_BUFFER_FULL);
+  }
+}
+
 void TraceLog::SetEventCallback(EventCallback cb) {
-  AutoLock lock(lock_);
-  event_callback_ = cb;
+  subtle::NoBarrier_Store(&event_callback_,
+                          reinterpret_cast<subtle::AtomicWord>(cb));
 };
 
+// Flush() works as the following:
+// 1. Flush() is called in threadA whose message loop is saved in
+//    flush_message_loop_proxy_;
+// 2. In the thread, FlushNextThreadOrFinish gets the first message loop from
+//    thread_message_loops_. If any, post a task to run
+//    FlushCurrentThreadAndContinue() in the message loop; otherwise finish the
+//    flush (step 4);
+// 3. FlushCurrentThreadAndContinue() deletes the thread local event buffer:
+//    - The last batch of events of the thread are flushed into the main
+//      buffer;
+//    - The message loop will be removed from thread_message_loops_;
+//    and posts FlushNextThreadOrFinish() in flush_message_loop_proxy_ to
+//    continue the flush procedure (step 2);
+// 4. When all thread local buffers have been flushed (and deleted), finish
+//    the flush by calling the OutputCallback with all events converted into
+//    JSON format.
 void TraceLog::Flush(const TraceLog::OutputCallback& cb) {
-  // Ignore memory allocations from here down.
-  INTERNAL_TRACE_MEMORY(TRACE_DISABLED_BY_DEFAULT("memory"),
-                        TRACE_MEMORY_IGNORE);
+  if (IsEnabled()) {

[dsinclair, 2013/08/29 14:24:32]

Is this, basically, to stop us from doing any post tasks while we're tracing?

[Xianzhu, 2013/08/29 17:42:12]

Yes.

In addition, as posting tasks will generate more trace events, it will badly
interfere with the flush process. For example, just after a thread-local buffer
is flushed and destroyed, new trace event from the thread will cause a new
thread-local buffer to be created, then FlushNextThreadOrFinish() will never
finish.

Flushing while enabled happens on chrome-android because the gpu process and the
browser process are the same process. As the gpu process is still registered as
a subprocess, the TraceLog in the process will be enabled twice, and be disabled
and flushed twice on the end of tracing.

+    scoped_refptr<RefCountedString> empty_result = new RefCountedString;
+    cb.Run(empty_result, false);

[dsinclair, 2013/08/29 14:24:32]

Should we check if cb.isNull() here before we try to run it?

[Xianzhu, 2013/08/29 17:42:12]

Done.

+    return;
+  }
+
+  {
+    AutoLock lock(lock_);
+    DCHECK(!flush_message_loop_proxy_.get());
+    flush_message_loop_proxy_ = MessageLoopProxy::current();
+    DCHECK(!thread_message_loops_.size() || flush_message_loop_proxy_.get());
+    flush_output_callback_ = cb;
+  }
+  FlushNextThreadOrFinish();
+}
+
+void TraceLog::FlushNextThreadOrFinish() {
   scoped_ptr<TraceBuffer> previous_logged_events;
   {
     AutoLock lock(lock_);
+    hash_set<MessageLoop*>::const_iterator next_message_loop =
+        thread_message_loops_.begin();
+    if (next_message_loop != thread_message_loops_.end()) {
+      // Destroy the next thread local buffer. The buffer will be flushed into
+      // the main event buffer and the message loop will be removed from
+      // thread_message_loops.
+      (*next_message_loop)->PostTask(
+          FROM_HERE,
+          Bind(&TraceLog::FlushCurrentThreadAndContinue,
+               Unretained(this)));
+      return;
+    }
+
+    // All thread local buffers have been flushed (and destroyed).
+    // From here to the end of the function finishes the whole flush procedure.
     previous_logged_events.swap(logged_events_);
     logged_events_.reset(GetTraceBuffer());
+    subtle::NoBarrier_Store(&buffer_is_full_, 0);
   }  // release lock
 
-  while (previous_logged_events->HasMoreEvents()) {
+  flush_message_loop_proxy_ = NULL;
+
+  bool has_more_events = previous_logged_events->HasMoreEvents();
+  // The callback need to be called at least once even if there is no events
+  // to let the caller know the completion of flush.
+  do {
     scoped_refptr<RefCountedString> json_events_str_ptr =
         new RefCountedString();
 
-    for (size_t i = 0; i < kTraceEventBatchSize; ++i) {
+    for (size_t i = 0; has_more_events && i < kTraceEventBatchSize; ++i) {
       if (i > 0)
         *(&(json_events_str_ptr->data())) += ",";
 
       previous_logged_events->NextEvent().AppendAsJSON(
           &(json_events_str_ptr->data()));
 
-      if (!previous_logged_events->HasMoreEvents())
-        break;
+      has_more_events = previous_logged_events->HasMoreEvents();
     }
 
-    cb.Run(json_events_str_ptr);
-  }
+    flush_output_callback_.Run(json_events_str_ptr, has_more_events);

[dsinclair, 2013/08/29 14:24:32]

Should we check if flush_output_callback_ exists?

[Xianzhu, 2013/08/29 17:42:12]

Done.

+  } while (has_more_events);
+}
+
+// Run in each thread holding a local event buffer.
+void TraceLog::FlushCurrentThreadAndContinue() {
+  delete thread_local_event_buffer_.Get();
+  thread_local_event_buffer_.Set(NULL);
+
+  // Continue the flush procedure.
+  flush_message_loop_proxy_->PostTask(
+      FROM_HERE,
+      Bind(&TraceLog::FlushNextThreadOrFinish, Unretained(this)));
 }
 
 void TraceLog::AddTraceEvent(
     char phase,
     const unsigned char* category_group_enabled,
     const char* name,
     unsigned long long id,
     int num_args,
     const char** arg_names,
     const unsigned char* arg_types,
@@ skipping 29 matching lines @@
 #if defined(OS_ANDROID)
   SendToATrace(phase, GetCategoryGroupName(category_group_enabled), name, id,
                num_args, arg_names, arg_types, arg_values, convertable_values,
                flags);
 #endif
 
   if (!IsCategoryGroupEnabled(category_group_enabled))
     return;
 
   TimeTicks now = timestamp - time_offset_;
-  base::TimeTicks thread_now;
-  if (base::TimeTicks::IsThreadNowSupported())
-    thread_now = base::TimeTicks::ThreadNow();
-  EventCallback event_callback_copy;
+  TimeTicks thread_now = ThreadNow();
 
   NotificationHelper notifier(this);
 
+  ThreadLocalEventBuffer* thread_local_event_buffer = NULL;
+  // A ThreadLocalEventBuffer needs the message loop
+  // - to know when the thread exits;
+  // - to handle the final flush.
+  // For a thread without a message loop, the trace events will be added into
+  // the main buffer directly.
+  if (MessageLoop::current()) {
+    thread_local_event_buffer = thread_local_event_buffer_.Get();
+    if (!thread_local_event_buffer) {
+      thread_local_event_buffer = new ThreadLocalEventBuffer(this);
+      thread_local_event_buffer_.Set(thread_local_event_buffer);
+    }
+  }
+
   // Check and update the current thread name only if the event is for the
   // current thread to avoid locks in most cases.
   if (thread_id == static_cast<int>(PlatformThread::CurrentId())) {
     const char* new_name = ThreadIdNameManager::GetInstance()->
         GetName(thread_id);
     // Check if the thread name has been set or changed since the previous
     // call (if any), but don't bother if the new name is empty. Note this will
     // not detect a thread name change within the same char* buffer address: we
     // favor common case performance over corner case correctness.
     if (new_name != g_current_thread_name.Get().Get() &&
@@ skipping 15 matching lines @@
                                existing_names.end(),
                                new_name) != existing_names.end();
         if (!found) {
           existing_name->second.push_back(',');
           existing_name->second.append(new_name);
         }
       }
     }
   }
 
-  TraceEvent trace_event(thread_id,
-      now, thread_now, phase, category_group_enabled, name, id,
-      num_args, arg_names, arg_types, arg_values,
-      convertable_values, flags);
+  if (!subtle::NoBarrier_Load(&buffer_is_full_)) {
+    TraceEvent trace_event(thread_id, now, thread_now, phase,
+                           category_group_enabled, name, id,
+                           num_args, arg_names, arg_types, arg_values,
+                           convertable_values, flags);
 
-  do {
-    AutoLock lock(lock_);
+    if (thread_local_event_buffer) {
+      thread_local_event_buffer->AddEvent(trace_event, &notifier);
+    } else {
+      AutoLock lock(lock_);
+      AddEventToMainBufferWhileLocked(trace_event);
+      CheckIfBufferIsFullWhileLocked(&notifier);
+    }
 
-    event_callback_copy = event_callback_;
-    if (logged_events_->IsFull())
-      break;
+    if (trace_options() & ECHO_TO_CONSOLE) {
+      AutoLock lock(lock_);
 
-    logged_events_->AddEvent(trace_event);
-
-    if (trace_options_ & ECHO_TO_CONSOLE) {
       TimeDelta duration;
       if (phase == TRACE_EVENT_PHASE_END) {
         duration = timestamp - thread_event_start_times_[thread_id].top();
         thread_event_start_times_[thread_id].pop();
       }
 
       std::string thread_name = thread_names_[thread_id];
       if (thread_colors_.find(thread_name) == thread_colors_.end())
         thread_colors_[thread_name] = (thread_colors_.size() % 6) + 1;
 
@@ skipping 12 matching lines @@
 
       trace_event.AppendPrettyPrinted(&log);
       if (phase == TRACE_EVENT_PHASE_END)
         log << base::StringPrintf(" (%.3f ms)", duration.InMillisecondsF());
 
       LOG(ERROR) << log.str() << "\x1b[0;m";
 
       if (phase == TRACE_EVENT_PHASE_BEGIN)
         thread_event_start_times_[thread_id].push(timestamp);
     }
+  }
 
-    if (logged_events_->IsFull())
-      notifier.AddNotificationWhileLocked(TRACE_BUFFER_FULL);
-
-    if (watch_category_ == category_group_enabled && watch_event_name_ == name)
+  if (reinterpret_cast<const unsigned char*>(subtle::NoBarrier_Load(
+      &watch_category_)) == category_group_enabled) {
+    AutoLock lock(lock_);
+    if (watch_event_name_ == name)
       notifier.AddNotificationWhileLocked(EVENT_WATCH_NOTIFICATION);
-  } while (0); // release lock
+  }
 
   notifier.SendNotificationIfAny();
-  if (event_callback_copy != NULL) {
-    event_callback_copy(phase, category_group_enabled, name, id,
-        num_args, arg_names, arg_types, arg_values,
-        flags);
+  EventCallback event_callback = reinterpret_cast<EventCallback>(
+      subtle::NoBarrier_Load(&event_callback_));
+  if (event_callback) {
+    event_callback(phase, category_group_enabled, name, id,
+                   num_args, arg_names, arg_types, arg_values,
+                   flags);
   }
+
+  if (thread_local_event_buffer)
+    thread_local_event_buffer->ReportOverhead(now, thread_now);
 }
 
 void TraceLog::AddTraceEventEtw(char phase,
                                 const char* name,
                                 const void* id,
                                 const char* extra) {
 #if defined(OS_WIN)
   TraceEventETWProvider::Trace(name, phase, id, extra);
 #endif
   INTERNAL_TRACE_EVENT_ADD(phase, "ETW Trace Event", name,
@@ skipping 12 matching lines @@
                            TRACE_EVENT_FLAG_COPY, "id", id, "extra", extra);
 }
 
 void TraceLog::SetWatchEvent(const std::string& category_name,
                              const std::string& event_name) {
   const unsigned char* category = GetCategoryGroupEnabled(
       category_name.c_str());
   size_t notify_count = 0;
   {
     AutoLock lock(lock_);
-    watch_category_ = category;
+    subtle::NoBarrier_Store(&watch_category_,
+                            reinterpret_cast<subtle::AtomicWord>(category));
     watch_event_name_ = event_name;
 
     // First, search existing events for watch event because we want to catch
     // it even if it has already occurred.
     notify_count = logged_events_->CountEnabledByName(category, event_name);
   }  // release lock
 
   // Send notification for each event found.
   for (size_t i = 0; i < notify_count; ++i) {
     NotificationHelper notifier(this);
     lock_.Acquire();
     notifier.AddNotificationWhileLocked(EVENT_WATCH_NOTIFICATION);
     lock_.Release();
     notifier.SendNotificationIfAny();
   }
 }
 
 void TraceLog::CancelWatchEvent() {
   AutoLock lock(lock_);
-  watch_category_ = NULL;
+  subtle::NoBarrier_Store(&watch_category_, 0);
   watch_event_name_ = "";
 }
 
 namespace {
 
 template <typename T>
 void AddMetadataEventToBuffer(
     TraceBuffer* logged_events,
     int thread_id,
     const char* metadata_name, const char* arg_name,
@@ skipping 330 matching lines @@
         0,                       // num_args
         NULL,                    // arg_names
         NULL,                    // arg_types
         NULL,                    // arg_values
         NULL,                    // convertable values
         TRACE_EVENT_FLAG_NONE);  // flags
   }
 }
 
 }  // namespace trace_event_internal