Moving cpu profiling into its own thread.

src/cpu-profiler.cc

 // Copyright 2012 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 27 matching lines @@
 
 namespace v8 {
 namespace internal {
 
 static const int kEventsBufferSize = 256 * KB;
 static const int kTickSamplesBufferChunkSize = 64 * KB;
 static const int kTickSamplesBufferChunksCount = 16;
 static const int kProfilerStackSize = 64 * KB;
 
 
-ProfilerEventsProcessor::ProfilerEventsProcessor(ProfileGenerator* generator)
-    : Thread(Thread::Options("v8:ProfEvntProc", kProfilerStackSize)),
+ProfilerEventsProcessor::ProfilerEventsProcessor(ProfileGenerator* generator, Sampler* sampler, int interval)
+    : CpuProfilerThread(sampler),
       generator_(generator),
       running_(true),
+      interval_(interval),
       ticks_buffer_(sizeof(TickSampleEventRecord),
                     kTickSamplesBufferChunkSize,
                     kTickSamplesBufferChunksCount),
       enqueue_order_(0) {
 }
 
 
 void ProfilerEventsProcessor::CallbackCreateEvent(Logger::LogEventsAndTags tag,
                                                   const char* prefix,
                                                   String* name,
@@ skipping 115 matching lines @@
   sample->pc = reinterpret_cast<Address>(sample);  // Not NULL.
   for (StackTraceFrameIterator it(isolate);
        !it.done() && sample->frames_count < TickSample::kMaxFramesCount;
        it.Advance()) {
     sample->stack[sample->frames_count++] = it.frame()->pc();
   }
   ticks_from_vm_buffer_.Enqueue(record);
 }
 
 
-bool ProfilerEventsProcessor::ProcessCodeEvent(unsigned* dequeue_order) {
+bool ProfilerEventsProcessor::ProcessCodeEvent(unsigned& dequeue_order) {
   if (!events_buffer_.IsEmpty()) {
     CodeEventsContainer record;
     events_buffer_.Dequeue(&record);
     switch (record.generic.type) {
 #define PROFILER_TYPE_CASE(type, clss)                          \
       case CodeEventRecord::type:                               \
         record.clss##_.UpdateCodeMap(generator_->code_map());   \
         break;
 
       CODE_EVENTS_TYPE_LIST(PROFILER_TYPE_CASE)
 
 #undef PROFILER_TYPE_CASE
       default: return true;  // Skip record.
     }
-    *dequeue_order = record.generic.order;
+    dequeue_order = record.generic.order;
     return true;
   }
   return false;
 }
 
 
-bool ProfilerEventsProcessor::ProcessTicks(unsigned dequeue_order) {
-  while (true) {
+bool ProfilerEventsProcessor::ProcessTicks(unsigned dequeue_order, int64_t start, int64_t time_limit) {
+  while (time_limit == -1 || OS::Ticks() - start < time_limit) {
     if (!ticks_from_vm_buffer_.IsEmpty()
         && ticks_from_vm_buffer_.Peek()->order == dequeue_order) {
       TickSampleEventRecord record;
       ticks_from_vm_buffer_.Dequeue(&record);
       generator_->RecordTickSample(record.sample);
     }
 
     const TickSampleEventRecord* rec =
         TickSampleEventRecord::cast(ticks_buffer_.StartDequeue());
     if (rec == NULL) return !ticks_from_vm_buffer_.IsEmpty();
     // Make a local copy of tick sample record to ensure that it won't
     // be modified as we are processing it. This is possible as the
     // sampler writes w/o any sync to the queue, so if the processor
     // will get far behind, a record may be modified right under its
     // feet.
     TickSampleEventRecord record = *rec;
     if (record.order == dequeue_order) {
       // A paranoid check to make sure that we don't get a memory overrun
       // in case of frames_count having a wild value.
       if (record.sample.frames_count < 0
           || record.sample.frames_count > TickSample::kMaxFramesCount)
         record.sample.frames_count = 0;
       generator_->RecordTickSample(record.sample);
       ticks_buffer_.FinishDequeue();
     } else {
       return true;
     }
   }
+  return false;
+}
+
+
+void ProfilerEventsProcessor::ProcessEventsQueue(unsigned& dequeue_order, int64_t time_limit) {
+  int64_t start = OS::Ticks();
+  while (OS::Ticks() - start < time_limit)

[caseq, 2012/08/16 14:39:22]

I would just use absolute time limit, so that we compute it based on when we
started previous sample, not when we started processing.

+    // Process ticks until we have any.
+    if (ProcessTicks(dequeue_order, start, time_limit)) {
+      // All ticks of the current dequeue_order are processed,
+      // proceed to the next code event.
+      ProcessCodeEvent(dequeue_order);
+    }
+    YieldCPU();
 }
 
 
 void ProfilerEventsProcessor::Run() {
   unsigned dequeue_order = 0;
 
   while (running_) {
-    // Process ticks until we have any.
-    if (ProcessTicks(dequeue_order)) {
-      // All ticks of the current dequeue_order are processed,
-      // proceed to the next code event.
-      ProcessCodeEvent(&dequeue_order);
-    }
-    YieldCPU();

[caseq, 2012/08/16 14:39:22]

So we're running the loop with 100% usage until it's time to do next sample,
then perform YieldCPU(). This does not seem to make sense. I guess if we're
looking to increase sampling frequency, we can't yield CPU as the scheduler
quantum is probably noticeable larger than sampling period.

+    int64_t start = OS::Ticks();
+    DoCpuProfile();
+    ProcessEventsQueue(dequeue_order, static_cast<int64_t>(interval_) - (OS::Ticks() - start));
   }
 
   // Process remaining tick events.
   ticks_buffer_.FlushResidualRecords();
   // Perform processing until we have tick events, skip remaining code events.
-  while (ProcessTicks(dequeue_order) && ProcessCodeEvent(&dequeue_order)) { }
+  while (ProcessTicks(dequeue_order, -1, -1) && ProcessCodeEvent(dequeue_order)) { }
 }
 
 
 void CpuProfiler::StartProfiling(const char* title) {
   ASSERT(Isolate::Current()->cpu_profiler() != NULL);
   Isolate::Current()->cpu_profiler()->StartCollectingProfile(title);
 }
 
 
 void CpuProfiler::StartProfiling(String* title) {
@@ skipping 210 matching lines @@
 
 void CpuProfiler::StartCollectingProfile(String* title) {
   StartCollectingProfile(profiles_->GetName(title));
 }
 
 
 void CpuProfiler::StartProcessorIfNotStarted() {
   if (processor_ == NULL) {
     Isolate* isolate = Isolate::Current();
 
+    // Enable stack sampling.
+    Sampler* sampler = reinterpret_cast<Sampler*>(isolate->logger()->ticker_);
+    if (!sampler->IsActive()) {
+      sampler->Start();
+      need_to_stop_sampler_ = true;
+    }
+    sampler->IncreaseProfilingDepth();
     // Disable logging when using the new implementation.
     saved_logging_nesting_ = isolate->logger()->logging_nesting_;
     isolate->logger()->logging_nesting_ = 0;
     generator_ = new ProfileGenerator(profiles_);
-    processor_ = new ProfilerEventsProcessor(generator_);
+    processor_ = new ProfilerEventsProcessor(generator_, sampler, sampler->interval() * 1000);
     NoBarrier_Store(&is_profiling_, true);
     processor_->Start();
     // Enumerate stuff we already have in the heap.
     if (isolate->heap()->HasBeenSetUp()) {
       if (!FLAG_prof_browser_mode) {
         bool saved_log_code_flag = FLAG_log_code;
         FLAG_log_code = true;
         isolate->logger()->LogCodeObjects();
         FLAG_log_code = saved_log_code_flag;
       }
       isolate->logger()->LogCompiledFunctions();
       isolate->logger()->LogAccessorCallbacks();
     }
-    // Enable stack sampling.
-    Sampler* sampler = reinterpret_cast<Sampler*>(isolate->logger()->ticker_);
-    if (!sampler->IsActive()) {
-      sampler->Start();
-      need_to_stop_sampler_ = true;
-    }
-    sampler->IncreaseProfilingDepth();
   }
 }
 
 
 CpuProfile* CpuProfiler::StopCollectingProfile(const char* title) {
   const double actual_sampling_rate = generator_->actual_sampling_rate();
   StopProcessorIfLastProfile(title);
   CpuProfile* result =
       profiles_->StopProfiling(TokenEnumerator::kNoSecurityToken,
                                title,
@@ skipping 49 matching lines @@
 
 void CpuProfiler::TearDown() {
   Isolate* isolate = Isolate::Current();
   if (isolate->cpu_profiler() != NULL) {
     delete isolate->cpu_profiler();
   }
   isolate->set_cpu_profiler(NULL);
 }
 
 } }  // namespace v8::internal