Support higher CPU profiler sampling rate on posix systems

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 @@
 #include "../include/v8-profiler.h"
 
 namespace v8 {
 namespace internal {
 
 static const int kTickSamplesBufferChunkSize = 64 * KB;
 static const int kTickSamplesBufferChunksCount = 16;
 static const int kProfilerStackSize = 64 * KB;
 
 
-ProfilerEventsProcessor::ProfilerEventsProcessor(ProfileGenerator* generator)
+ProfilerEventsProcessor::ProfilerEventsProcessor(
+    ProfileGenerator* generator,
+    Sampler* sampler,
+    int period_in_useconds)
     : Thread(Thread::Options("v8:ProfEvntProc", kProfilerStackSize)),
       generator_(generator),
+      sampler_(sampler),
       running_(true),
+      period_in_useconds_(period_in_useconds),
       ticks_buffer_(sizeof(TickSampleEventRecord),
                     kTickSamplesBufferChunkSize,
                     kTickSamplesBufferChunksCount),
       last_code_event_id_(0), last_processed_code_event_id_(0) {
 }
 
 
 void ProfilerEventsProcessor::Enqueue(const CodeEventsContainer& event) {
   event.generic.order = ++last_code_event_id_;
   events_buffer_.Enqueue(event);
@@ skipping 67 matching lines @@
     // 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();
   }
 }
 
 
+void ProfilerEventsProcessor::ProcessEventsAndDoSample() {
+  int64_t stop_time = OS::Ticks() + period_in_useconds_;
+  // Keep processing existing events until we need to do next sample.
+  while (OS::Ticks() < stop_time) {
+    if (ProcessTicks()) {
+      // All ticks of the current dequeue_order are processed,
+      // proceed to the next code event.
+      ProcessCodeEvent();
+    }
+  }
+  // Schedule next sample. sampler_ is NULL in tests.
+  if (sampler_) sampler_->DoSample();
+}
+
+
+void ProfilerEventsProcessor::ProcessEventsAndYield() {
+  // Process ticks until we have any.
+  if (ProcessTicks()) {
+    // All ticks of the current dequeue_order are processed,
+    // proceed to the next code event.
+    ProcessCodeEvent();
+  }
+  YieldCPU();
+}
+
+
 void ProfilerEventsProcessor::Run() {
   while (running_) {
-    // Process ticks until we have any.
-    if (ProcessTicks()) {
-      // All ticks of the current last_processed_code_event_id_ are processed,
-      // proceed to the next code event.
-      ProcessCodeEvent();
+    if (Sampler::CanSampleOnProfilerEventsProcessorThread()) {
+      ProcessEventsAndDoSample();
+    } else {
+      ProcessEventsAndYield();

[Benedikt Meurer, 2013/07/30 06:41:04]

I think there are two severe issues with this loop, first of all running_ is not
volatile and there's no memory barrier for running_. And YieldCPU() should not
be used. Maybe unrelated to the change at hand, but how about using proper
synchronization primitives (i.e. monitors/condition variables or semaphores)
instead?

[yurys, 2013/07/30 08:52:26]

The profiler code tries to avoid usage of synchronization primitives to minimize
effect it imposes on the thread that is being profiled. In particular both
queues used to pass code events and tick samples from the VM thread to the
processing one are lock-free (I didn't see any measurements that would justify
this approach though). Adding a mutex/semaphore here may have a noticeable
performance impact so I'd prefer to stay with a more light-weight solution. I
believe changing running_ into Atomic32 and using NoBarrier_Load to read it here
should be enough, wdyt?

Also can you elaborate on why you think that YieldCPU should not be used? Among
other things I believe that a side-effect of using it here may be that it
implies a memory barrier and makes this code work correctly:)

Anyways I'm going to address this in a separate CL as all the raised issues are
applicable to the current implementation.

     }
-    YieldCPU();
   }
 
   // Process remaining tick events.
   ticks_buffer_.FlushResidualRecords();
   do {
     ProcessTicks();
   } while (ProcessCodeEvent());
 }
 
 
@@ skipping 263 matching lines @@
 }
 
 
 void CpuProfiler::StartProcessorIfNotStarted() {
   if (processor_ == NULL) {
     Logger* logger = isolate_->logger();
     // Disable logging when using the new implementation.
     saved_logging_nesting_ = logger->logging_nesting_;
     logger->logging_nesting_ = 0;
     generator_ = new ProfileGenerator(profiles_);
-    processor_ = new ProfilerEventsProcessor(generator_);
+    Sampler* sampler = logger->sampler();
+    processor_ = new ProfilerEventsProcessor(
+        generator_, sampler, FLAG_cpu_profiler_sampling_interval);
     is_profiling_ = true;
     processor_->StartSynchronously();
     // Enumerate stuff we already have in the heap.
     ASSERT(isolate_->heap()->HasBeenSetUp());
     if (!FLAG_prof_browser_mode) {
       logger->LogCodeObjects();
     }
     logger->LogCompiledFunctions();
     logger->LogAccessorCallbacks();
     LogBuiltins();
     // Enable stack sampling.
-    Sampler* sampler = logger->sampler();
+    if (Sampler::CanSampleOnProfilerEventsProcessorThread()) {
+      sampler->SetHasProcessingThread(true);
+    }
     sampler->IncreaseProfilingDepth();
     if (!sampler->IsActive()) {
       sampler->Start();
       need_to_stop_sampler_ = true;
     }
   }
 }
 
 
 CpuProfile* CpuProfiler::StopProfiling(const char* title) {
@@ skipping 19 matching lines @@
 
 void CpuProfiler::StopProcessorIfLastProfile(const char* title) {
   if (profiles_->IsLastProfile(title)) StopProcessor();
 }
 
 
 void CpuProfiler::StopProcessor() {
   Logger* logger = isolate_->logger();
   Sampler* sampler = reinterpret_cast<Sampler*>(logger->ticker_);
   sampler->DecreaseProfilingDepth();
-  if (need_to_stop_sampler_) {
-    sampler->Stop();
-    need_to_stop_sampler_ = false;
-  }
   is_profiling_ = false;
   processor_->StopSynchronously();
   delete processor_;
   delete generator_;
   processor_ = NULL;
   generator_ = NULL;
+  if (Sampler::CanSampleOnProfilerEventsProcessorThread()) {
+    sampler->SetHasProcessingThread(false);
+  }
+  if (need_to_stop_sampler_) {
+    sampler->Stop();
+    need_to_stop_sampler_ = false;
+  }
   logger->logging_nesting_ = saved_logging_nesting_;
 }
 
 
 void CpuProfiler::LogBuiltins() {
   Builtins* builtins = isolate_->builtins();
   ASSERT(builtins->is_initialized());
   for (int i = 0; i < Builtins::builtin_count; i++) {
     CodeEventsContainer evt_rec(CodeEventRecord::REPORT_BUILTIN);
     ReportBuiltinEventRecord* rec = &evt_rec.ReportBuiltinEventRecord_;
     Builtins::Name id = static_cast<Builtins::Name>(i);
     rec->start = builtins->builtin(id)->address();
     rec->builtin_id = id;
     processor_->Enqueue(evt_rec);
   }
 }
 
 
 } }  // namespace v8::internal