Moving cpu profiling into its own thread.

src/cpu-profiler.cc

Index: src/cpu-profiler.cc
diff --git a/src/cpu-profiler.cc b/src/cpu-profiler.cc
index 3cbac77858d5560feb3cab420f08c9e461f2422e..02d41a1fb616e2b06b20ad8d0a965f8ab21f5159 100644
--- a/src/cpu-profiler.cc
+++ b/src/cpu-profiler.cc
@@ -45,10 +45,11 @@ 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),
@@ -184,7 +185,7 @@ void ProfilerEventsProcessor::AddCurrentStack() {
 }
 
 
-bool ProfilerEventsProcessor::ProcessCodeEvent(unsigned* dequeue_order) {
+bool ProfilerEventsProcessor::ProcessCodeEvent(unsigned& dequeue_order) {
   if (!events_buffer_.IsEmpty()) {
     CodeEventsContainer record;
     events_buffer_.Dequeue(&record);
@@ -199,15 +200,15 @@ bool ProfilerEventsProcessor::ProcessCodeEvent(unsigned* dequeue_order) {
 #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;
@@ -236,26 +237,36 @@ bool ProfilerEventsProcessor::ProcessTicks(unsigned dequeue_order) {
       return true;
     }
   }
+  return false;
 }
 
 
-void ProfilerEventsProcessor::Run() {
-  unsigned dequeue_order = 0;
-
-  while (running_) {
+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)) {
+    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);
+      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.

+}
+
+
+void ProfilerEventsProcessor::Run() {
+  unsigned dequeue_order = 0;
+
+  while (running_) {
+    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)) { }
 }
 
 
@@ -486,11 +497,18 @@ 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.
@@ -504,13 +522,6 @@ void CpuProfiler::StartProcessorIfNotStarted() {
       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();
   }
 }