Initial GetSample implementation.

src/sampler.cc

Index: src/sampler.cc
diff --git a/src/sampler.cc b/src/sampler.cc
index 0ed4b379626a9746038b6e05b1661cee522f4f1f..c0b25f102bbb5856f6d87856ce5e6130e192a74c 100644
--- a/src/sampler.cc
+++ b/src/sampler.cc
@@ -1,7 +1,6 @@
 // Copyright 2013 the V8 project 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 "src/sampler.h"
 
 #if V8_OS_POSIX && !V8_OS_CYGWIN
@@ -44,6 +43,7 @@
 
 #include "src/v8.h"
 
+#include "src/base/logging.h"
 #include "src/base/platform/platform.h"
 #include "src/cpu-profiler-inl.h"
 #include "src/flags.h"
@@ -269,8 +269,21 @@ class SimulatorHelper {
 
 class SignalHandler : public AllStatic {
  public:
-  static void SetUp() { if (!mutex_) mutex_ = new base::Mutex(); }
-  static void TearDown() { delete mutex_; }
+  static void SetUp() {
+    ASSERT_EQ(NULL, mutex_);
+    ASSERT_EQ(NULL, sampling_semaphore_);
+
+    mutex_ = new base::Mutex();
+    sampling_semaphore_ = new base::Semaphore(0);
+  }
+
+  static void TearDown() {
+    delete mutex_;
+    delete sampling_semaphore_;
+
+    mutex_ = NULL;
+    sampling_semaphore_ = NULL;
+  }
 
   static void IncreaseSamplerCount() {
     base::LockGuard<base::Mutex> lock_guard(mutex_);
@@ -308,14 +321,45 @@ class SignalHandler : public AllStatic {
   }
 
   static void HandleProfilerSignal(int signal, siginfo_t* info, void* context);
+
+  // The SIGPROF could have come from DoSample or from GetSample.
+  // GetSample is the internal endpoint for the public
+  // GetSample API, which aims to bypass all the internal buffers
+  // and return just one sample instead, rightaway.
+  // TODO(gholap): Eventually, get rid of DoSample.
+  //               Only GetSample should remain.
+
+  // This is the sample which will be filled by the call from GetSample.
+  static TickSample sample_;
+
+  // For the SIGPROF handler to know whether the call was from GetSample.
+  static volatile bool called_from_get_sample_;
+
+  // GetSample waits synchronously till the SIGPROF handler returns.
+  // this semaphore is used to signal GetSample.
+  static base::Semaphore* sampling_semaphore_;
+
+  // It is not that every time HandleProfilerSignal is invoked,
+  // it succeeds in obtaining a sample.
+  // This provides GetSample with the information whether
+  // the handler finished with or without getting the sample.
+  static bool sample_available_;
+
   // Protects the process wide state below.
   static base::Mutex* mutex_;
   static int client_count_;
   static bool signal_handler_installed_;
   static struct sigaction old_signal_handler_;
+
+  friend class Sampler;
 };
 
 
+TickSample SignalHandler::sample_;
+volatile bool SignalHandler::called_from_get_sample_ = false;
+base::Semaphore* SignalHandler::sampling_semaphore_ = NULL;
+bool SignalHandler::sample_available_ = false;
+
 base::Mutex* SignalHandler::mutex_ = NULL;
 int SignalHandler::client_count_ = 0;
 struct sigaction SignalHandler::old_signal_handler_;
@@ -329,6 +373,11 @@ void SignalHandler::HandleProfilerSignal(int signal, siginfo_t* info,
   // is disabled.
   return;
 #else
+  // Even if we return prematurely,
+  // we need to signal the sampling_semaphore_
+  sample_available_ = false;
+  ScopedSemaphore scoped_semaphore(sampling_semaphore_);
+
   USE(info);
   if (signal != SIGPROF) return;
   Isolate* isolate = Isolate::UncheckedReentrantCurrent();
@@ -466,7 +515,12 @@ void SignalHandler::HandleProfilerSignal(int signal, siginfo_t* info,
 #endif  // V8_HOST_ARCH_*
 #endif  // V8_OS_QNX
 #endif  // USE_SIMULATOR
-  sampler->SampleStack(state);
+  if (called_from_get_sample_) {
+    sample_.Init(sampler->isolate(), state);
+  } else {
+    sampler->SampleStack(state);
+  }
+  sample_available_ = true;
 #endif  // V8_OS_NACL
 }
 
@@ -569,6 +623,7 @@ DISABLE_ASAN void TickSample::Init(Isolate* isolate,
   timestamp = base::TimeTicks::HighResolutionNow();
   pc = regs.pc;
   state = isolate->current_vm_state();
+  frames_count = 0;
 
   // Avoid collecting traces while doing GC.
   if (state == GC) return;
@@ -693,19 +748,33 @@ void Sampler::DoSample() {
   pthread_kill(platform_data()->vm_tid(), SIGPROF);
 }
 
+
+void Sampler::GetSample(v8::Sample* sample) {
+  SignalHandler::called_from_get_sample_ = true;
+  pthread_kill(platform_data()->vm_tid(), SIGPROF);
+  SignalHandler::sampling_semaphore_->Wait();
+  if (SignalHandler::sample_available_) {
+    *sample = v8::Sample(SignalHandler::sample_.stack,

[Benedikt Meurer, 2014/09/05 03:50:00]

Nit: Use

 v8::Sample(&SignalHandler::sample_.stack[0],
&SignalHandler::sample_.stack[SignalHandler::sample_.frames_count])

syntax instead to make it easier to spot the array arithmetic here.

[alph, 2014/09/08 12:47:40]

Note it will always copy all the 255 frames.
The profiler code is quite performance critical. Yury was squeezing it to make
it possible to do 10kHz sampling.

I wonder if it's possible to avoid such a huge memory copying, e.g. by using a
placement new?

[gholap, 2014/09/08 23:57:38]

Done.

[gholap, 2014/09/08 23:57:38]

Done.

+                         SignalHandler::sample_.stack
+                         + SignalHandler::sample_.frames_count);
+  }
+  SignalHandler::called_from_get_sample_ = false;
+}
+
 #elif V8_OS_WIN || V8_OS_CYGWIN
 
-void Sampler::DoSample() {
+void Sampler::GetSampleHelper(TickSample * sample,

[alph, 2014/09/08 12:47:40]

nit: it doesn't match the declaration in .h

[gholap, 2014/09/08 23:57:38]

Done.

+                              bool called_from_get_sample) {
   HANDLE profiled_thread = platform_data()->profiled_thread();
-  if (profiled_thread == NULL) return;
+  if (profiled_thread == NULL) return NULL;
 
 #if defined(USE_SIMULATOR)
   SimulatorHelper helper;
-  if (!helper.Init(this, isolate())) return;
+  if (!helper.Init(this, isolate())) return NULL;
 #endif
 
   const DWORD kSuspendFailed = static_cast<DWORD>(-1);
-  if (SuspendThread(profiled_thread) == kSuspendFailed) return;
+  if (SuspendThread(profiled_thread) == kSuspendFailed) return NULL;
 
   // Context used for sampling the register state of the profiled thread.
   CONTEXT context;
@@ -726,11 +795,32 @@ void Sampler::DoSample() {
     state.fp = reinterpret_cast<Address>(context.Ebp);
 #endif
 #endif  // USE_SIMULATOR
-    SampleStack(state);
+    if (called_from_get_sample) {
+      sample->Init(sampler->isolate(), state);

[alph, 2014/09/08 12:47:40]

hmm... what is sampler?
Does it compile on Win?

[gholap, 2014/09/08 23:57:38]

Corrected.

+    } else {
+      SampleStack(state);
+    }
   }
   ResumeThread(profiled_thread);
 }
 
+
+void Sampler::DoSample() {
+  GetSampleHelper(NULL, false);
+}
+
+
+void Sampler::GetSample(v8::Sample* sample) {
+  // We need this tick_sample because,
+  // GetSampleHelper calls Init, which must be called on
+  // a TickSample and never on Sample.
+  TickSample tick_sample;
+  sample->frames_count = 0;
+  GetSampleHelper(&tick_sample, true);
+  *sample = v8::Sample(tick_sample.stack,

[Benedikt Meurer, 2014/09/05 03:50:00]

Nit: Use

 v8::Sample(&tick_sample.stack[0], &tick_sample.stack[tick_sample.frames_count])

syntax instead to make it easier to spot the array arithmetic here.

[gholap, 2014/09/08 23:57:38]

Done.

+                       tick_sample.stack + tick_sample.frames_count);
+}
+
 #endif  // USE_SIGNALS