Add histogram for native allocations

Shows up in traces.txt, sample output:
Histogram of native allocation 0:4315,131072:33,1179648:3 bucket size 131072
Histogram of native free 0:995,131072:19,1179648:2 bucket size 131072

Bug: 28680116

Change-Id: I584eb85709d3ae28f5142a01af11075a2a6436d7

runtime/base/histogram.h

@@ -85,6 +85,10 @@ template <class Value> class Histogram {
     return max_value_added_;
   }
 
+  Value BucketWidth() const {
+    return bucket_width_;
+  }
+
   const std::string& Name() const {
     return name_;
   }

runtime/gc/heap.cc

@@ -119,6 +119,8 @@ static constexpr uint32_t kAllocSpaceBeginForDeterministicAoT = 0x40000000;
 // Dump the rosalloc stats on SIGQUIT.
 static constexpr bool kDumpRosAllocStatsOnSigQuit = false;
 
+static constexpr size_t kNativeAllocationHistogramBuckets = 16;
+
 static inline bool CareAboutPauseTimes() {
   return Runtime::Current()->InJankPerceptibleProcessState();
 }
@@ -186,6 +188,11 @@ Heap::Heap(size_t initial_size,
       total_objects_freed_ever_(0),
       num_bytes_allocated_(0),
       native_bytes_allocated_(0),
+      native_histogram_lock_("Native allocation lock"),
+      native_allocation_histogram_("Native allocation sizes",
+                                   1U,
+                                   kNativeAllocationHistogramBuckets),
+      native_free_histogram_("Native free sizes", 1U, kNativeAllocationHistogramBuckets),
       num_bytes_freed_revoke_(0),
       verify_missing_card_marks_(false),
       verify_system_weaks_(false),
@@ -1185,6 +1192,20 @@ void Heap::DumpGcPerformanceInfo(std::ostream& os) {
     rosalloc_space_->DumpStats(os);
   }
 
+  {
+    MutexLock mu(Thread::Current(), native_histogram_lock_);
+    if (native_allocation_histogram_.SampleSize() > 0u) {
+      os << "Histogram of native allocation ";
+      native_allocation_histogram_.DumpBins(os);
+      os << " bucket size " << native_allocation_histogram_.BucketWidth() << "\n";
+    }
+    if (native_free_histogram_.SampleSize() > 0u) {
+      os << "Histogram of native free ";
+      native_free_histogram_.DumpBins(os);
+      os << " bucket size " << native_free_histogram_.BucketWidth() << "\n";
+    }
+  }
+
   BaseMutex::DumpAll(os);
 }
 
@@ -3848,6 +3869,10 @@ void Heap::RunFinalization(JNIEnv* env, uint64_t timeout) {
 
 void Heap::RegisterNativeAllocation(JNIEnv* env, size_t bytes) {
   Thread* self = ThreadForEnv(env);
+  {
+    MutexLock mu(self, native_histogram_lock_);
+    native_allocation_histogram_.AddValue(bytes);
+  }
   if (native_need_to_run_finalization_) {
     RunFinalization(env, kNativeAllocationFinalizeTimeout);
     UpdateMaxNativeFootprint();
@@ -3892,6 +3917,10 @@ void Heap::RegisterNativeAllocation(JNIEnv* env, size_t bytes) {
 
 void Heap::RegisterNativeFree(JNIEnv* env, size_t bytes) {
   size_t expected_size;
+  {
+    MutexLock mu(Thread::Current(), native_histogram_lock_);
+    native_free_histogram_.AddValue(bytes);
+  }
   do {
     expected_size = native_bytes_allocated_.LoadRelaxed();
     if (UNLIKELY(bytes > expected_size)) {

runtime/gc/heap.h

@@ -241,9 +241,9 @@ class Heap {
       SHARED_REQUIRES(Locks::mutator_lock_);
 
   void RegisterNativeAllocation(JNIEnv* env, size_t bytes)
-      REQUIRES(!*gc_complete_lock_, !*pending_task_lock_);
+      REQUIRES(!*gc_complete_lock_, !*pending_task_lock_, !native_histogram_lock_);
   void RegisterNativeFree(JNIEnv* env, size_t bytes)
-      REQUIRES(!*gc_complete_lock_, !*pending_task_lock_);
+      REQUIRES(!*gc_complete_lock_, !*pending_task_lock_, !native_histogram_lock_);
 
   // Change the allocator, updates entrypoints.
   void ChangeAllocator(AllocatorType allocator)
@@ -532,7 +532,7 @@ class Heap {
   space::Space* FindSpaceFromObject(const mirror::Object*, bool fail_ok) const
       SHARED_REQUIRES(Locks::mutator_lock_);
 
-  void DumpForSigQuit(std::ostream& os) REQUIRES(!*gc_complete_lock_);
+  void DumpForSigQuit(std::ostream& os) REQUIRES(!*gc_complete_lock_, !native_histogram_lock_);
 
   // Do a pending collector transition.
   void DoPendingCollectorTransition() REQUIRES(!*gc_complete_lock_);
@@ -654,7 +654,8 @@ class Heap {
   std::string SafePrettyTypeOf(mirror::Object* obj) NO_THREAD_SAFETY_ANALYSIS;
 
   // GC performance measuring
-  void DumpGcPerformanceInfo(std::ostream& os) REQUIRES(!*gc_complete_lock_);
+  void DumpGcPerformanceInfo(std::ostream& os)
+      REQUIRES(!*gc_complete_lock_, !native_histogram_lock_);
   void ResetGcPerformanceInfo() REQUIRES(!*gc_complete_lock_);
 
   // Thread pool.
@@ -1156,6 +1157,11 @@ class Heap {
   // Bytes which are allocated and managed by native code but still need to be accounted for.
   Atomic<size_t> native_bytes_allocated_;
 
+  // Native allocation stats.
+  Mutex native_histogram_lock_;
+  Histogram<uint64_t> native_allocation_histogram_;
+  Histogram<uint64_t> native_free_histogram_;
+
   // Number of bytes freed by thread local buffer revokes. This will
   // cancel out the ahead-of-time bulk counting of bytes allocated in
   // rosalloc thread-local buffers.  It is temporarily accumulated