You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@impala.apache.org by ta...@apache.org on 2018/07/13 06:03:37 UTC

[25/51] [abbrv] impala git commit: IMPALA-7006: Add KRPC folders from kudu@334ecafd

http://git-wip-us.apache.org/repos/asf/impala/blob/fcf190c4/be/src/kudu/util/debug/trace_event_impl.cc
----------------------------------------------------------------------
diff --git a/be/src/kudu/util/debug/trace_event_impl.cc b/be/src/kudu/util/debug/trace_event_impl.cc
new file mode 100644
index 0000000..155faf7
--- /dev/null
+++ b/be/src/kudu/util/debug/trace_event_impl.cc
@@ -0,0 +1,2436 @@
+// Copyright (c) 2012 The Chromium 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 "kudu/util/debug/trace_event_impl.h"
+
+#include <sched.h>
+#include <unistd.h>
+
+#include <algorithm>
+#include <cinttypes>
+#include <cstdlib>
+#include <cstring>
+#include <list>
+#include <sstream>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+#include <gflags/gflags.h>
+
+#include "kudu/gutil/bind.h"
+#include "kudu/gutil/dynamic_annotations.h"
+#include "kudu/gutil/map-util.h"
+#include "kudu/gutil/mathlimits.h"
+#include "kudu/gutil/port.h"
+#include "kudu/gutil/ref_counted_memory.h"
+#include "kudu/gutil/singleton.h"
+#include "kudu/gutil/stl_util.h"
+#include "kudu/gutil/stringprintf.h"
+#include "kudu/gutil/strings/join.h"
+#include "kudu/gutil/strings/split.h"
+#include "kudu/gutil/strings/stringpiece.h"
+#include "kudu/gutil/strings/substitute.h"
+#include "kudu/gutil/strings/util.h"
+#include "kudu/gutil/sysinfo.h"
+#include "kudu/gutil/walltime.h"
+
+#include "kudu/util/atomic.h"
+#include "kudu/util/debug/trace_event.h"
+#include "kudu/util/debug/trace_event_synthetic_delay.h"
+#include "kudu/util/flag_tags.h"
+#include "kudu/util/monotime.h"
+#include "kudu/util/status.h"
+#include "kudu/util/thread.h"
+#include "kudu/util/threadlocal.h"
+
+DEFINE_string(trace_to_console, "",
+              "Trace pattern specifying which trace events should be dumped "
+              "directly to the console");
+TAG_FLAG(trace_to_console, experimental);
+
+// The thread buckets for the sampling profiler.
+BASE_EXPORT TRACE_EVENT_API_ATOMIC_WORD g_trace_state[3];
+
+namespace kudu {
+namespace debug {
+
+using base::SpinLockHolder;
+
+using strings::SubstituteAndAppend;
+using std::string;
+using std::vector;
+
+__thread TraceLog::PerThreadInfo* TraceLog::thread_local_info_ = nullptr;
+
+namespace {
+
+// Controls the number of trace events we will buffer in-memory
+// before throwing them away.
+const size_t kTraceBufferChunkSize = TraceBufferChunk::kTraceBufferChunkSize;
+const size_t kTraceEventVectorBufferChunks = 256000 / kTraceBufferChunkSize;
+const size_t kTraceEventRingBufferChunks = kTraceEventVectorBufferChunks / 4;
+const size_t kTraceEventBatchChunks = 1000 / kTraceBufferChunkSize;
+// Can store results for 30 seconds with 1 ms sampling interval.
+const size_t kMonitorTraceEventBufferChunks = 30000 / kTraceBufferChunkSize;
+// ECHO_TO_CONSOLE needs a small buffer to hold the unfinished COMPLETE events.
+const size_t kEchoToConsoleTraceEventBufferChunks = 256;
+
+const char kSyntheticDelayCategoryFilterPrefix[] = "DELAY(";
+
+#define MAX_CATEGORY_GROUPS 100
+
+// Parallel arrays g_category_groups and g_category_group_enabled are separate
+// so that a pointer to a member of g_category_group_enabled can be easily
+// converted to an index into g_category_groups. This allows macros to deal
+// only with char enabled pointers from g_category_group_enabled, and we can
+// convert internally to determine the category name from the char enabled
+// pointer.
+const char* g_category_groups[MAX_CATEGORY_GROUPS] = {
+  "toplevel",
+  "tracing already shutdown",
+  "tracing categories exhausted; must increase MAX_CATEGORY_GROUPS",
+  "__metadata"};
+
+// The enabled flag is char instead of bool so that the API can be used from C.
+unsigned char g_category_group_enabled[MAX_CATEGORY_GROUPS] = { 0 };
+// Indexes here have to match the g_category_groups array indexes above.
+const int kCategoryAlreadyShutdown = 1;
+const int kCategoryCategoriesExhausted = 2;
+const int kCategoryMetadata = 3;
+const int kNumBuiltinCategories = 4;
+// Skip default categories.
+AtomicWord g_category_index = kNumBuiltinCategories;
+
+// The name of the current thread. This is used to decide if the current
+// thread name has changed. We combine all the seen thread names into the
+// output name for the thread.
+__thread const char* g_current_thread_name = "";
+
+static void NOTIMPLEMENTED() {
+  LOG(FATAL);
+}
+
+class TraceBufferRingBuffer : public TraceBuffer {
+ public:
+  explicit TraceBufferRingBuffer(size_t max_chunks)
+      : max_chunks_(max_chunks),
+        recyclable_chunks_queue_(new size_t[queue_capacity()]),
+        queue_head_(0),
+        queue_tail_(max_chunks),
+        current_iteration_index_(0),
+        current_chunk_seq_(1) {
+    chunks_.reserve(max_chunks);
+    for (size_t i = 0; i < max_chunks; ++i)
+      recyclable_chunks_queue_[i] = i;
+  }
+
+  ~TraceBufferRingBuffer() {
+    STLDeleteElements(&chunks_);
+  }
+
+  virtual gscoped_ptr<TraceBufferChunk> GetChunk(size_t* index) OVERRIDE {
+    // Because the number of threads is much less than the number of chunks,
+    // the queue should never be empty.
+    DCHECK(!QueueIsEmpty());
+
+    *index = recyclable_chunks_queue_[queue_head_];
+    queue_head_ = NextQueueIndex(queue_head_);
+    current_iteration_index_ = queue_head_;
+
+    if (*index >= chunks_.size())
+      chunks_.resize(*index + 1);
+
+    TraceBufferChunk* chunk = chunks_[*index];
+    chunks_[*index] = nullptr;  // Put NULL in the slot of a in-flight chunk.
+    if (chunk)
+      chunk->Reset(current_chunk_seq_++);
+    else
+      chunk = new TraceBufferChunk(current_chunk_seq_++);
+
+    return gscoped_ptr<TraceBufferChunk>(chunk);
+  }
+
+  virtual void ReturnChunk(size_t index,
+                           gscoped_ptr<TraceBufferChunk> chunk) OVERRIDE {
+    // When this method is called, the queue should not be full because it
+    // can contain all chunks including the one to be returned.
+    DCHECK(!QueueIsFull());
+    DCHECK(chunk);
+    DCHECK_LT(index, chunks_.size());
+    DCHECK(!chunks_[index]);
+    chunks_[index] = chunk.release();
+    recyclable_chunks_queue_[queue_tail_] = index;
+    queue_tail_ = NextQueueIndex(queue_tail_);
+  }
+
+  virtual bool IsFull() const OVERRIDE {
+    return false;
+  }
+
+  virtual size_t Size() const OVERRIDE {
+    // This is approximate because not all of the chunks are full.
+    return chunks_.size() * kTraceBufferChunkSize;
+  }
+
+  virtual size_t Capacity() const OVERRIDE {
+    return max_chunks_ * kTraceBufferChunkSize;
+  }
+
+  virtual TraceEvent* GetEventByHandle(TraceEventHandle handle) OVERRIDE {
+    if (handle.chunk_index >= chunks_.size())
+      return nullptr;
+    TraceBufferChunk* chunk = chunks_[handle.chunk_index];
+    if (!chunk || chunk->seq() != handle.chunk_seq)
+      return nullptr;
+    return chunk->GetEventAt(handle.event_index);
+  }
+
+  virtual const TraceBufferChunk* NextChunk() OVERRIDE {
+    if (chunks_.empty())
+      return nullptr;
+
+    while (current_iteration_index_ != queue_tail_) {
+      size_t chunk_index = recyclable_chunks_queue_[current_iteration_index_];
+      current_iteration_index_ = NextQueueIndex(current_iteration_index_);
+      if (chunk_index >= chunks_.size()) // Skip uninitialized chunks.
+        continue;
+      DCHECK(chunks_[chunk_index]);
+      return chunks_[chunk_index];
+    }
+    return nullptr;
+  }
+
+  virtual gscoped_ptr<TraceBuffer> CloneForIteration() const OVERRIDE {
+    gscoped_ptr<ClonedTraceBuffer> cloned_buffer(new ClonedTraceBuffer());
+    for (size_t queue_index = queue_head_; queue_index != queue_tail_;
+        queue_index = NextQueueIndex(queue_index)) {
+      size_t chunk_index = recyclable_chunks_queue_[queue_index];
+      if (chunk_index >= chunks_.size()) // Skip uninitialized chunks.
+        continue;
+      TraceBufferChunk* chunk = chunks_[chunk_index];
+      cloned_buffer->chunks_.push_back(chunk ? chunk->Clone().release() : nullptr);
+    }
+    return cloned_buffer.PassAs<TraceBuffer>();
+  }
+
+ private:
+  class ClonedTraceBuffer : public TraceBuffer {
+   public:
+    ClonedTraceBuffer() : current_iteration_index_(0) {}
+    ~ClonedTraceBuffer() {
+      STLDeleteElements(&chunks_);
+    }
+
+    // The only implemented method.
+    virtual const TraceBufferChunk* NextChunk() OVERRIDE {
+      return current_iteration_index_ < chunks_.size() ?
+          chunks_[current_iteration_index_++] : nullptr;
+    }
+
+    virtual gscoped_ptr<TraceBufferChunk> GetChunk(size_t* index) OVERRIDE {
+      NOTIMPLEMENTED();
+      return gscoped_ptr<TraceBufferChunk>();
+    }
+    virtual void ReturnChunk(size_t index,
+                             gscoped_ptr<TraceBufferChunk>) OVERRIDE {
+      NOTIMPLEMENTED();
+    }
+    virtual bool IsFull() const OVERRIDE { return false; }
+    virtual size_t Size() const OVERRIDE { return 0; }
+    virtual size_t Capacity() const OVERRIDE { return 0; }
+    virtual TraceEvent* GetEventByHandle(TraceEventHandle handle) OVERRIDE {
+      return nullptr;
+    }
+    virtual gscoped_ptr<TraceBuffer> CloneForIteration() const OVERRIDE {
+      NOTIMPLEMENTED();
+      return gscoped_ptr<TraceBuffer>();
+    }
+
+    size_t current_iteration_index_;
+    vector<TraceBufferChunk*> chunks_;
+  };
+
+  bool QueueIsEmpty() const {
+    return queue_head_ == queue_tail_;
+  }
+
+  size_t QueueSize() const {
+    return queue_tail_ > queue_head_ ? queue_tail_ - queue_head_ :
+        queue_tail_ + queue_capacity() - queue_head_;
+  }
+
+  bool QueueIsFull() const {
+    return QueueSize() == queue_capacity() - 1;
+  }
+
+  size_t queue_capacity() const {
+    // One extra space to help distinguish full state and empty state.
+    return max_chunks_ + 1;
+  }
+
+  size_t NextQueueIndex(size_t index) const {
+    index++;
+    if (index >= queue_capacity())
+      index = 0;
+    return index;
+  }
+
+  size_t max_chunks_;
+  vector<TraceBufferChunk*> chunks_;
+
+  gscoped_ptr<size_t[]> recyclable_chunks_queue_;
+  size_t queue_head_;
+  size_t queue_tail_;
+
+  size_t current_iteration_index_;
+  uint32_t current_chunk_seq_;
+
+  DISALLOW_COPY_AND_ASSIGN(TraceBufferRingBuffer);
+};
+
+class TraceBufferVector : public TraceBuffer {
+ public:
+  TraceBufferVector()
+      : in_flight_chunk_count_(0),
+        current_iteration_index_(0) {
+    chunks_.reserve(kTraceEventVectorBufferChunks);
+  }
+  ~TraceBufferVector() {
+    STLDeleteElements(&chunks_);
+  }
+
+  virtual gscoped_ptr<TraceBufferChunk> GetChunk(size_t* index) OVERRIDE {
+    // This function may be called when adding normal events or indirectly from
+    // AddMetadataEventsWhileLocked(). We can not DECHECK(!IsFull()) because we
+    // have to add the metadata events and flush thread-local buffers even if
+    // the buffer is full.
+    *index = chunks_.size();
+    chunks_.push_back(nullptr);  // Put NULL in the slot of a in-flight chunk.
+    ++in_flight_chunk_count_;
+    // + 1 because zero chunk_seq is not allowed.
+    return gscoped_ptr<TraceBufferChunk>(
+        new TraceBufferChunk(static_cast<uint32_t>(*index) + 1));
+  }
+
+  virtual void ReturnChunk(size_t index,
+                           gscoped_ptr<TraceBufferChunk> chunk) OVERRIDE {
+    DCHECK_GT(in_flight_chunk_count_, 0u);
+    DCHECK_LT(index, chunks_.size());
+    DCHECK(!chunks_[index]);
+    --in_flight_chunk_count_;
+    chunks_[index] = chunk.release();
+  }
+
+  virtual bool IsFull() const OVERRIDE {
+    return chunks_.size() >= kTraceEventVectorBufferChunks;
+  }
+
+  virtual size_t Size() const OVERRIDE {
+    // This is approximate because not all of the chunks are full.
+    return chunks_.size() * kTraceBufferChunkSize;
+  }
+
+  virtual size_t Capacity() const OVERRIDE {
+    return kTraceEventVectorBufferChunks * kTraceBufferChunkSize;
+  }
+
+  virtual TraceEvent* GetEventByHandle(TraceEventHandle handle) OVERRIDE {
+    if (handle.chunk_index >= chunks_.size())
+      return nullptr;
+    TraceBufferChunk* chunk = chunks_[handle.chunk_index];
+    if (!chunk || chunk->seq() != handle.chunk_seq)
+      return nullptr;
+    return chunk->GetEventAt(handle.event_index);
+  }
+
+  virtual const TraceBufferChunk* NextChunk() OVERRIDE {
+    while (current_iteration_index_ < chunks_.size()) {
+      // Skip in-flight chunks.
+      const TraceBufferChunk* chunk = chunks_[current_iteration_index_++];
+      if (chunk)
+        return chunk;
+    }
+    return nullptr;
+  }
+
+  virtual gscoped_ptr<TraceBuffer> CloneForIteration() const OVERRIDE {
+    NOTIMPLEMENTED();
+    return gscoped_ptr<TraceBuffer>();
+  }
+
+ private:
+  size_t in_flight_chunk_count_;
+  size_t current_iteration_index_;
+  vector<TraceBufferChunk*> chunks_;
+
+  DISALLOW_COPY_AND_ASSIGN(TraceBufferVector);
+};
+
+template <typename T>
+void InitializeMetadataEvent(TraceEvent* trace_event,
+                             int thread_id,
+                             const char* metadata_name, const char* arg_name,
+                             const T& value) {
+  if (!trace_event)
+    return;
+
+  int num_args = 1;
+  unsigned char arg_type;
+  uint64_t arg_value;
+  ::trace_event_internal::SetTraceValue(value, &arg_type, &arg_value);
+  trace_event->Initialize(thread_id,
+                          MicrosecondsInt64(0), MicrosecondsInt64(0), TRACE_EVENT_PHASE_METADATA,
+                          &g_category_group_enabled[kCategoryMetadata],
+                          metadata_name, ::trace_event_internal::kNoEventId,
+                          num_args, &arg_name, &arg_type, &arg_value, nullptr,
+                          TRACE_EVENT_FLAG_NONE);
+}
+
+// RAII object which marks '*dst' with a non-zero value while in scope.
+// This assumes that no other threads write to '*dst'.
+class MarkFlagInScope {
+ public:
+  explicit MarkFlagInScope(Atomic32* dst)
+      : dst_(dst) {
+    // We currently use Acquire_AtomicExchange here because it appears
+    // to be the cheapest way of getting an "Acquire_Store" barrier. Actually
+    // using Acquire_Store generates more assembly instructions and benchmarks
+    // slightly slower.
+    //
+    // TODO: it would be even faster to avoid the memory barrier here entirely,
+    // and do an asymmetric barrier, for example by having the flusher thread
+    // send a signal to every registered thread, or wait until every other thread
+    // has experienced at least one context switch. A number of options for this
+    // are outlined in:
+    // http://home.comcast.net/~pjbishop/Dave/Asymmetric-Dekker-Synchronization.txt
+    Atomic32 old_val = base::subtle::Acquire_AtomicExchange(dst_, 1);
+    DCHECK_EQ(old_val, 0);
+  }
+  ~MarkFlagInScope() {
+    base::subtle::Release_Store(dst_, 0);
+  }
+
+ private:
+  Atomic32* dst_;
+  DISALLOW_COPY_AND_ASSIGN(MarkFlagInScope);
+};
+}  // anonymous namespace
+
+TraceLog::ThreadLocalEventBuffer* TraceLog::PerThreadInfo::AtomicTakeBuffer() {
+  return reinterpret_cast<TraceLog::ThreadLocalEventBuffer*>(
+    base::subtle::Acquire_AtomicExchange(
+      reinterpret_cast<AtomicWord*>(&event_buffer_),
+      0));
+}
+
+void TraceBufferChunk::Reset(uint32_t new_seq) {
+  for (size_t i = 0; i < next_free_; ++i)
+    chunk_[i].Reset();
+  next_free_ = 0;
+  seq_ = new_seq;
+}
+
+TraceEvent* TraceBufferChunk::AddTraceEvent(size_t* event_index) {
+  DCHECK(!IsFull());
+  *event_index = next_free_++;
+  return &chunk_[*event_index];
+}
+
+gscoped_ptr<TraceBufferChunk> TraceBufferChunk::Clone() const {
+  gscoped_ptr<TraceBufferChunk> cloned_chunk(new TraceBufferChunk(seq_));
+  cloned_chunk->next_free_ = next_free_;
+  for (size_t i = 0; i < next_free_; ++i)
+    cloned_chunk->chunk_[i].CopyFrom(chunk_[i]);
+  return std::move(cloned_chunk);
+}
+
+// A helper class that allows the lock to be acquired in the middle of the scope
+// and unlocks at the end of scope if locked.
+class TraceLog::OptionalAutoLock {
+ public:
+  explicit OptionalAutoLock(base::SpinLock& lock) // NOLINT(google-runtime-references)
+      : lock_(lock),
+        locked_(false) {
+  }
+
+  ~OptionalAutoLock() {
+    if (locked_)
+      lock_.Unlock();
+  }
+
+  void EnsureAcquired() {
+    if (!locked_) {
+      lock_.Lock();
+      locked_ = true;
+    }
+  }
+
+ private:
+  base::SpinLock& lock_;
+  bool locked_;
+  DISALLOW_COPY_AND_ASSIGN(OptionalAutoLock);
+};
+
+// Use this function instead of TraceEventHandle constructor to keep the
+// overhead of ScopedTracer (trace_event.h) constructor minimum.
+void MakeHandle(uint32_t chunk_seq, size_t chunk_index, size_t event_index,
+                TraceEventHandle* handle) {
+  DCHECK(chunk_seq);
+  DCHECK(chunk_index < (1u << 16));
+  DCHECK(event_index < (1u << 16));
+  handle->chunk_seq = chunk_seq;
+  handle->chunk_index = static_cast<uint16_t>(chunk_index);
+  handle->event_index = static_cast<uint16_t>(event_index);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// TraceEvent
+//
+////////////////////////////////////////////////////////////////////////////////
+
+namespace {
+
+size_t GetAllocLength(const char* str) { return str ? strlen(str) + 1 : 0; }
+
+// Copies |*member| into |*buffer|, sets |*member| to point to this new
+// location, and then advances |*buffer| by the amount written.
+void CopyTraceEventParameter(char** buffer,
+                             const char** member,
+                             const char* end) {
+  if (*member) {
+    size_t written = strings::strlcpy(*buffer, *member, end - *buffer) + 1;
+    DCHECK_LE(static_cast<int>(written), end - *buffer);
+    *member = *buffer;
+    *buffer += written;
+  }
+}
+
+}  // namespace
+
+TraceEvent::TraceEvent()
+    : duration_(-1),
+      thread_duration_(-1),
+      id_(0u),
+      category_group_enabled_(nullptr),
+      name_(nullptr),
+      thread_id_(0),
+      phase_(TRACE_EVENT_PHASE_BEGIN),
+      flags_(0) {
+  for (auto& arg_name : arg_names_) {
+    arg_name = nullptr;
+  }
+  memset(arg_values_, 0, sizeof(arg_values_));
+}
+
+TraceEvent::~TraceEvent() {
+}
+
+void TraceEvent::CopyFrom(const TraceEvent& other) {
+  timestamp_ = other.timestamp_;
+  thread_timestamp_ = other.thread_timestamp_;
+  duration_ = other.duration_;
+  id_ = other.id_;
+  category_group_enabled_ = other.category_group_enabled_;
+  name_ = other.name_;
+  thread_id_ = other.thread_id_;
+  phase_ = other.phase_;
+  flags_ = other.flags_;
+  parameter_copy_storage_ = other.parameter_copy_storage_;
+
+  for (int i = 0; i < kTraceMaxNumArgs; ++i) {
+    arg_names_[i] = other.arg_names_[i];
+    arg_types_[i] = other.arg_types_[i];
+    arg_values_[i] = other.arg_values_[i];
+    convertable_values_[i] = other.convertable_values_[i];
+  }
+}
+
+void TraceEvent::Initialize(
+    int thread_id,
+    MicrosecondsInt64 timestamp,
+    MicrosecondsInt64 thread_timestamp,
+    char phase,
+    const unsigned char* category_group_enabled,
+    const char* name,
+    uint64_t id,
+    int num_args,
+    const char** arg_names,
+    const unsigned char* arg_types,
+    const uint64_t* arg_values,
+    const scoped_refptr<ConvertableToTraceFormat>* convertable_values,
+    unsigned char flags) {
+  timestamp_ = timestamp;
+  thread_timestamp_ = thread_timestamp;
+  duration_ = -1;;
+  id_ = id;
+  category_group_enabled_ = category_group_enabled;
+  name_ = name;
+  thread_id_ = thread_id;
+  phase_ = phase;
+  flags_ = flags;
+
+  // Clamp num_args since it may have been set by a third_party library.
+  num_args = (num_args > kTraceMaxNumArgs) ? kTraceMaxNumArgs : num_args;
+  int i = 0;
+  for (; i < num_args; ++i) {
+    arg_names_[i] = arg_names[i];
+    arg_types_[i] = arg_types[i];
+
+    if (arg_types[i] == TRACE_VALUE_TYPE_CONVERTABLE)
+      convertable_values_[i] = convertable_values[i];
+    else
+      arg_values_[i].as_uint = arg_values[i];
+  }
+  for (; i < kTraceMaxNumArgs; ++i) {
+    arg_names_[i] = nullptr;
+    arg_values_[i].as_uint = 0u;
+    convertable_values_[i] = nullptr;
+    arg_types_[i] = TRACE_VALUE_TYPE_UINT;
+  }
+
+  bool copy = !!(flags & TRACE_EVENT_FLAG_COPY);
+  size_t alloc_size = 0;
+  if (copy) {
+    alloc_size += GetAllocLength(name);
+    for (i = 0; i < num_args; ++i) {
+      alloc_size += GetAllocLength(arg_names_[i]);
+      if (arg_types_[i] == TRACE_VALUE_TYPE_STRING)
+        arg_types_[i] = TRACE_VALUE_TYPE_COPY_STRING;
+    }
+  }
+
+  bool arg_is_copy[kTraceMaxNumArgs];
+  for (i = 0; i < num_args; ++i) {
+    // No copying of convertable types, we retain ownership.
+    if (arg_types_[i] == TRACE_VALUE_TYPE_CONVERTABLE)
+      continue;
+
+    // We only take a copy of arg_vals if they are of type COPY_STRING.
+    arg_is_copy[i] = (arg_types_[i] == TRACE_VALUE_TYPE_COPY_STRING);
+    if (arg_is_copy[i])
+      alloc_size += GetAllocLength(arg_values_[i].as_string);
+  }
+
+  if (alloc_size) {
+    parameter_copy_storage_ = new RefCountedString;
+    parameter_copy_storage_->data().resize(alloc_size);
+    char* ptr = string_as_array(&parameter_copy_storage_->data());
+    const char* end = ptr + alloc_size;
+    if (copy) {
+      CopyTraceEventParameter(&ptr, &name_, end);
+      for (i = 0; i < num_args; ++i) {
+        CopyTraceEventParameter(&ptr, &arg_names_[i], end);
+      }
+    }
+    for (i = 0; i < num_args; ++i) {
+      if (arg_types_[i] == TRACE_VALUE_TYPE_CONVERTABLE)
+        continue;
+      if (arg_is_copy[i])
+        CopyTraceEventParameter(&ptr, &arg_values_[i].as_string, end);
+    }
+    DCHECK_EQ(end, ptr) << "Overrun by " << ptr - end;
+  }
+}
+
+void TraceEvent::Reset() {
+  // Only reset fields that won't be initialized in Initialize(), or that may
+  // hold references to other objects.
+  duration_ = -1;;
+  parameter_copy_storage_ = nullptr;
+  for (int i = 0; i < kTraceMaxNumArgs && arg_names_[i]; ++i)
+    convertable_values_[i] = nullptr;
+}
+
+void TraceEvent::UpdateDuration(const MicrosecondsInt64& now,
+                                const MicrosecondsInt64& thread_now) {
+  DCHECK(duration_ == -1);
+  duration_ = now - timestamp_;
+  thread_duration_ = thread_now - thread_timestamp_;
+}
+
+namespace {
+// Escape the given string using JSON rules.
+void JsonEscape(StringPiece s, string* out) {
+  out->reserve(out->size() + s.size() * 2);
+  const char* p_end = s.data() + s.size();
+  for (const char* p = s.data(); p != p_end; p++) {
+    // Only the following characters need to be escaped, according to json.org.
+    // In particular, it's illegal to escape the single-quote character, and
+    // JSON does not support the "\x" escape sequence like C/Java.
+    switch (*p) {
+      case '"':
+      case '\\':
+        out->push_back('\\');
+        out->push_back(*p);
+        break;
+      case '\b':
+        out->append("\\b");
+        break;
+      case '\f':
+        out->append("\\f");
+        break;
+      case '\n':
+        out->append("\\n");
+      case '\r':
+        out->append("\\r");
+        break;
+      case '\t':
+        out->append("\\t");
+        break;
+      default:
+        out->push_back(*p);
+    }
+  }
+}
+} // anonymous namespace
+
+// static
+void TraceEvent::AppendValueAsJSON(unsigned char type,
+                                   TraceEvent::TraceValue value,
+                                   std::string* out) {
+  switch (type) {
+    case TRACE_VALUE_TYPE_BOOL:
+      *out += value.as_bool ? "true" : "false";
+      break;
+    case TRACE_VALUE_TYPE_UINT:
+      SubstituteAndAppend(out, "$0", static_cast<uint64_t>(value.as_uint));
+      break;
+    case TRACE_VALUE_TYPE_INT:
+      SubstituteAndAppend(out, "$0", static_cast<int64_t>(value.as_int));
+      break;
+    case TRACE_VALUE_TYPE_DOUBLE: {
+      // FIXME: base/json/json_writer.cc is using the same code,
+      //        should be made into a common method.
+      std::string real;
+      double val = value.as_double;
+      if (MathLimits<double>::IsFinite(val)) {
+        real = strings::Substitute("$0", val);
+        // Ensure that the number has a .0 if there's no decimal or 'e'.  This
+        // makes sure that when we read the JSON back, it's interpreted as a
+        // real rather than an int.
+        if (real.find('.') == std::string::npos &&
+            real.find('e') == std::string::npos &&
+            real.find('E') == std::string::npos) {
+          real.append(".0");
+        }
+        // The JSON spec requires that non-integer values in the range (-1,1)
+        // have a zero before the decimal point - ".52" is not valid, "0.52" is.
+        if (real[0] == '.') {
+          real.insert(0, "0");
+        } else if (real.length() > 1 && real[0] == '-' && real[1] == '.') {
+          // "-.1" bad "-0.1" good
+          real.insert(1, "0");
+        }
+      } else if (MathLimits<double>::IsNaN(val)){
+        // The JSON spec doesn't allow NaN and Infinity (since these are
+        // objects in EcmaScript).  Use strings instead.
+        real = "\"NaN\"";
+      } else if (val < 0) {
+        real = "\"-Infinity\"";
+      } else {
+        real = "\"Infinity\"";
+      }
+      SubstituteAndAppend(out, "$0", real);
+      break;
+    }
+    case TRACE_VALUE_TYPE_POINTER:
+      // JSON only supports double and int numbers.
+      // So as not to lose bits from a 64-bit pointer, output as a hex string.
+      StringAppendF(out, "\"0x%" PRIx64 "\"", static_cast<uint64_t>(
+                                     reinterpret_cast<intptr_t>(
+                                     value.as_pointer)));
+      break;
+    case TRACE_VALUE_TYPE_STRING:
+    case TRACE_VALUE_TYPE_COPY_STRING:
+      *out += "\"";
+      JsonEscape(value.as_string ? value.as_string : "NULL", out);
+      *out += "\"";
+      break;
+    default:
+      LOG(FATAL) << "Don't know how to print this value";
+      break;
+  }
+}
+
+void TraceEvent::AppendAsJSON(std::string* out) const {
+  int64_t time_int64 = timestamp_;
+  int process_id = TraceLog::GetInstance()->process_id();
+  // Category group checked at category creation time.
+  DCHECK(!strchr(name_, '"'));
+  StringAppendF(out,
+      "{\"cat\":\"%s\",\"pid\":%i,\"tid\":%i,\"ts\":%" PRId64 ","
+      "\"ph\":\"%c\",\"name\":\"%s\",\"args\":{",
+      TraceLog::GetCategoryGroupName(category_group_enabled_),
+      process_id,
+      thread_id_,
+      time_int64,
+      phase_,
+      name_);
+
+  // Output argument names and values, stop at first NULL argument name.
+  for (int i = 0; i < kTraceMaxNumArgs && arg_names_[i]; ++i) {
+    if (i > 0)
+      *out += ",";
+    *out += "\"";
+    *out += arg_names_[i];
+    *out += "\":";
+
+    if (arg_types_[i] == TRACE_VALUE_TYPE_CONVERTABLE)
+      convertable_values_[i]->AppendAsTraceFormat(out);
+    else
+      AppendValueAsJSON(arg_types_[i], arg_values_[i], out);
+  }
+  *out += "}";
+
+  if (phase_ == TRACE_EVENT_PHASE_COMPLETE) {
+    int64_t duration = duration_;
+    if (duration != -1)
+      StringAppendF(out, ",\"dur\":%" PRId64, duration);
+    if (thread_timestamp_ >= 0) {
+      int64_t thread_duration = thread_duration_;
+      if (thread_duration != -1)
+        StringAppendF(out, ",\"tdur\":%" PRId64, thread_duration);
+    }
+  }
+
+  // Output tts if thread_timestamp is valid.
+  if (thread_timestamp_ >= 0) {
+    int64_t thread_time_int64 = thread_timestamp_;
+    StringAppendF(out, ",\"tts\":%" PRId64, thread_time_int64);
+  }
+
+  // If id_ is set, print it out as a hex string so we don't loose any
+  // bits (it might be a 64-bit pointer).
+  if (flags_ & TRACE_EVENT_FLAG_HAS_ID)
+    StringAppendF(out, ",\"id\":\"0x%" PRIx64 "\"", static_cast<uint64_t>(id_));
+
+  // Instant events also output their scope.
+  if (phase_ == TRACE_EVENT_PHASE_INSTANT) {
+    char scope = '?';
+    switch (flags_ & TRACE_EVENT_FLAG_SCOPE_MASK) {
+      case TRACE_EVENT_SCOPE_GLOBAL:
+        scope = TRACE_EVENT_SCOPE_NAME_GLOBAL;
+        break;
+
+      case TRACE_EVENT_SCOPE_PROCESS:
+        scope = TRACE_EVENT_SCOPE_NAME_PROCESS;
+        break;
+
+      case TRACE_EVENT_SCOPE_THREAD:
+        scope = TRACE_EVENT_SCOPE_NAME_THREAD;
+        break;
+    }
+    StringAppendF(out, ",\"s\":\"%c\"", scope);
+  }
+
+  *out += "}";
+}
+
+void TraceEvent::AppendPrettyPrinted(std::ostringstream* out) const {
+  *out << name_ << "[";
+  *out << TraceLog::GetCategoryGroupName(category_group_enabled_);
+  *out << "]";
+  if (arg_names_[0]) {
+    *out << ", {";
+    for (int i = 0; i < kTraceMaxNumArgs && arg_names_[i]; ++i) {
+      if (i > 0)
+        *out << ", ";
+      *out << arg_names_[i] << ":";
+      std::string value_as_text;
+
+      if (arg_types_[i] == TRACE_VALUE_TYPE_CONVERTABLE)
+        convertable_values_[i]->AppendAsTraceFormat(&value_as_text);
+      else
+        AppendValueAsJSON(arg_types_[i], arg_values_[i], &value_as_text);
+
+      *out << value_as_text;
+    }
+    *out << "}";
+  }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// TraceResultBuffer
+//
+////////////////////////////////////////////////////////////////////////////////
+
+string TraceResultBuffer::FlushTraceLogToString() {
+  return DoFlush(false);
+}
+
+string TraceResultBuffer::FlushTraceLogToStringButLeaveBufferIntact() {
+  return DoFlush(true);
+}
+
+string TraceResultBuffer::DoFlush(bool leave_intact) {
+  TraceResultBuffer buf;
+  TraceLog* tl = TraceLog::GetInstance();
+  if (leave_intact) {
+    tl->FlushButLeaveBufferIntact(Bind(&TraceResultBuffer::Collect, Unretained(&buf)));
+  } else {
+    tl->Flush(Bind(&TraceResultBuffer::Collect, Unretained(&buf)));
+  }
+  buf.json_.append("]}\n");
+  return buf.json_;
+}
+
+TraceResultBuffer::TraceResultBuffer()
+  : first_(true) {
+}
+TraceResultBuffer::~TraceResultBuffer() {
+}
+
+void TraceResultBuffer::Collect(
+  const scoped_refptr<RefCountedString>& s,
+  bool has_more_events) {
+  if (first_) {
+    json_.append("{\"traceEvents\": [\n");
+    first_ = false;
+  } else if (!s->data().empty()) {
+    // Sometimes we get sent an empty chunk at the end,
+    // and we don't want to end up with an extra trailing ','
+    json_.append(",\n");
+  }
+  json_.append(s->data());
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// TraceSamplingThread
+//
+////////////////////////////////////////////////////////////////////////////////
+class TraceBucketData;
+typedef Callback<void(TraceBucketData*)> TraceSampleCallback;
+
+class TraceBucketData {
+ public:
+  TraceBucketData(AtomicWord* bucket,
+                  const char* name,
+                  TraceSampleCallback callback);
+  ~TraceBucketData();
+
+  TRACE_EVENT_API_ATOMIC_WORD* bucket;
+  const char* bucket_name;
+  TraceSampleCallback callback;
+};
+
+// This object must be created on the IO thread.
+class TraceSamplingThread {
+ public:
+  TraceSamplingThread();
+  virtual ~TraceSamplingThread();
+
+  void ThreadMain();
+
+  static void DefaultSamplingCallback(TraceBucketData* bucekt_data);
+
+  void Stop();
+
+ private:
+  friend class TraceLog;
+
+  void GetSamples();
+  // Not thread-safe. Once the ThreadMain has been called, this can no longer
+  // be called.
+  void RegisterSampleBucket(TRACE_EVENT_API_ATOMIC_WORD* bucket,
+                            const char* const name,
+                            TraceSampleCallback callback);
+  // Splits a combined "category\0name" into the two component parts.
+  static void ExtractCategoryAndName(const char* combined,
+                                     const char** category,
+                                     const char** name);
+  std::vector<TraceBucketData> sample_buckets_;
+  bool thread_running_;
+  AtomicBool cancellation_flag_;
+};
+
+
+TraceSamplingThread::TraceSamplingThread()
+  : thread_running_(false),
+    cancellation_flag_(false) {
+}
+
+TraceSamplingThread::~TraceSamplingThread() {
+}
+
+void TraceSamplingThread::ThreadMain() {
+  thread_running_ = true;
+  const MonoDelta sleepDelta = MonoDelta::FromMicroseconds(1000);
+  while (!cancellation_flag_.Load()) {
+    SleepFor(sleepDelta);
+    GetSamples();
+  }
+}
+
+// static
+void TraceSamplingThread::DefaultSamplingCallback(
+    TraceBucketData* bucket_data) {
+  TRACE_EVENT_API_ATOMIC_WORD category_and_name =
+      TRACE_EVENT_API_ATOMIC_LOAD(*bucket_data->bucket);
+  if (!category_and_name)
+    return;
+  const char* const combined =
+      reinterpret_cast<const char* const>(category_and_name);
+  const char* category_group;
+  const char* name;
+  ExtractCategoryAndName(combined, &category_group, &name);
+  TRACE_EVENT_API_ADD_TRACE_EVENT(TRACE_EVENT_PHASE_SAMPLE,
+      TraceLog::GetCategoryGroupEnabled(category_group),
+      name, 0, 0, nullptr, nullptr, nullptr, nullptr, 0);
+}
+
+void TraceSamplingThread::GetSamples() {
+  for (auto& sample_bucket : sample_buckets_) {
+    TraceBucketData* bucket_data = &sample_bucket;
+    bucket_data->callback.Run(bucket_data);
+  }
+}
+
+void TraceSamplingThread::RegisterSampleBucket(
+    TRACE_EVENT_API_ATOMIC_WORD* bucket,
+    const char* const name,
+    TraceSampleCallback callback) {
+  // Access to sample_buckets_ doesn't cause races with the sampling thread
+  // that uses the sample_buckets_, because it is guaranteed that
+  // RegisterSampleBucket is called before the sampling thread is created.
+  DCHECK(!thread_running_);
+  sample_buckets_.emplace_back(bucket, name, callback);
+}
+
+// static
+void TraceSamplingThread::ExtractCategoryAndName(const char* combined,
+                                                 const char** category,
+                                                 const char** name) {
+  *category = combined;
+  *name = &combined[strlen(combined) + 1];
+}
+
+void TraceSamplingThread::Stop() {
+  cancellation_flag_.Store(true);
+}
+
+TraceBucketData::TraceBucketData(AtomicWord* bucket, const char* name,
+                                 TraceSampleCallback callback)
+    : bucket(bucket), bucket_name(name), callback(std::move(callback)) {}
+
+TraceBucketData::~TraceBucketData() {
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// TraceLog
+//
+////////////////////////////////////////////////////////////////////////////////
+
+class TraceLog::ThreadLocalEventBuffer {
+ public:
+  explicit ThreadLocalEventBuffer(TraceLog* trace_log);
+  virtual ~ThreadLocalEventBuffer();
+
+  TraceEvent* AddTraceEvent(TraceEventHandle* handle);
+
+  TraceEvent* GetEventByHandle(TraceEventHandle handle) {
+    if (!chunk_ || handle.chunk_seq != chunk_->seq() ||
+        handle.chunk_index != chunk_index_)
+      return nullptr;
+
+    return chunk_->GetEventAt(handle.event_index);
+  }
+
+  int generation() const { return generation_; }
+
+  void Flush(int64_t tid);
+
+ private:
+  // Check that the current thread is the one that constructed this trace buffer.
+  void CheckIsOwnerThread() const {
+    DCHECK_EQ(kudu::Thread::UniqueThreadId(), owner_tid_);
+  }
+
+  // Since TraceLog is a leaky singleton, trace_log_ will always be valid
+  // as long as the thread exists.
+  TraceLog* trace_log_;
+  gscoped_ptr<TraceBufferChunk> chunk_;
+  size_t chunk_index_;
+  int generation_;
+
+  // The TID of the thread that constructed this event buffer. Only this thread
+  // may add trace events.
+  int64_t owner_tid_;
+
+  DISALLOW_COPY_AND_ASSIGN(ThreadLocalEventBuffer);
+};
+
+TraceLog::ThreadLocalEventBuffer::ThreadLocalEventBuffer(TraceLog* trace_log)
+    : trace_log_(trace_log),
+      chunk_index_(0),
+      generation_(trace_log->generation()),
+      owner_tid_(kudu::Thread::UniqueThreadId()) {
+}
+
+TraceLog::ThreadLocalEventBuffer::~ThreadLocalEventBuffer() {
+}
+
+TraceEvent* TraceLog::ThreadLocalEventBuffer::AddTraceEvent(
+    TraceEventHandle* handle) {
+  CheckIsOwnerThread();
+
+  if (chunk_ && chunk_->IsFull()) {
+    SpinLockHolder lock(&trace_log_->lock_);
+    Flush(Thread::UniqueThreadId());
+    chunk_.reset();
+  }
+  if (!chunk_) {
+    SpinLockHolder lock(&trace_log_->lock_);
+    chunk_ = trace_log_->logged_events_->GetChunk(&chunk_index_);
+    trace_log_->CheckIfBufferIsFullWhileLocked();
+  }
+  if (!chunk_)
+    return nullptr;
+
+  size_t event_index;
+  TraceEvent* trace_event = chunk_->AddTraceEvent(&event_index);
+  if (trace_event && handle)
+    MakeHandle(chunk_->seq(), chunk_index_, event_index, handle);
+
+  return trace_event;
+}
+
+void TraceLog::ThreadLocalEventBuffer::Flush(int64_t tid) {
+  DCHECK(trace_log_->lock_.IsHeld());
+
+  if (!chunk_)
+    return;
+
+  if (trace_log_->CheckGeneration(generation_)) {
+    // Return the chunk to the buffer only if the generation matches.
+    trace_log_->logged_events_->ReturnChunk(chunk_index_, std::move(chunk_));
+  }
+}
+
+// static
+TraceLog* TraceLog::GetInstance() {
+  return Singleton<TraceLog>::get();
+}
+
+TraceLog::TraceLog()
+    : mode_(DISABLED),
+      num_traces_recorded_(0),
+      event_callback_(0),
+      dispatching_to_observer_list_(false),
+      process_sort_index_(0),
+      process_id_hash_(0),
+      process_id_(0),
+      time_offset_(0),
+      watch_category_(0),
+      trace_options_(RECORD_UNTIL_FULL),
+      sampling_thread_handle_(nullptr),
+      category_filter_(CategoryFilter::kDefaultCategoryFilterString),
+      event_callback_category_filter_(
+          CategoryFilter::kDefaultCategoryFilterString),
+      thread_shared_chunk_index_(0),
+      generation_(0) {
+  // Trace is enabled or disabled on one thread while other threads are
+  // accessing the enabled flag. We don't care whether edge-case events are
+  // traced or not, so we allow races on the enabled flag to keep the trace
+  // macros fast.
+  ANNOTATE_BENIGN_RACE_SIZED(g_category_group_enabled,
+                             sizeof(g_category_group_enabled),
+                             "trace_event category enabled");
+  for (int i = 0; i < MAX_CATEGORY_GROUPS; ++i) {
+    ANNOTATE_BENIGN_RACE(&g_category_group_enabled[i],
+                         "trace_event category enabled");
+  }
+  SetProcessID(static_cast<int>(getpid()));
+
+  string filter = FLAGS_trace_to_console;
+  if (!filter.empty()) {
+    SetEnabled(CategoryFilter(filter), RECORDING_MODE, ECHO_TO_CONSOLE);
+    LOG(ERROR) << "Tracing to console with CategoryFilter '" << filter << "'.";
+  }
+
+  logged_events_.reset(CreateTraceBuffer());
+}
+
+TraceLog::~TraceLog() {
+}
+
+const unsigned char* TraceLog::GetCategoryGroupEnabled(
+    const char* category_group) {
+  TraceLog* tracelog = GetInstance();
+  if (!tracelog) {
+    DCHECK(!g_category_group_enabled[kCategoryAlreadyShutdown]);
+    return &g_category_group_enabled[kCategoryAlreadyShutdown];
+  }
+  return tracelog->GetCategoryGroupEnabledInternal(category_group);
+}
+
+const char* TraceLog::GetCategoryGroupName(
+    const unsigned char* category_group_enabled) {
+  // Calculate the index of the category group by finding
+  // category_group_enabled in g_category_group_enabled array.
+  uintptr_t category_begin =
+      reinterpret_cast<uintptr_t>(g_category_group_enabled);
+  uintptr_t category_ptr = reinterpret_cast<uintptr_t>(category_group_enabled);
+  DCHECK(category_ptr >= category_begin &&
+         category_ptr < reinterpret_cast<uintptr_t>(
+             g_category_group_enabled + MAX_CATEGORY_GROUPS)) <<
+      "out of bounds category pointer";
+  uintptr_t category_index =
+      (category_ptr - category_begin) / sizeof(g_category_group_enabled[0]);
+  return g_category_groups[category_index];
+}
+
+void TraceLog::UpdateCategoryGroupEnabledFlag(int category_index) {
+  unsigned char enabled_flag = 0;
+  const char* category_group = g_category_groups[category_index];
+  if (mode_ == RECORDING_MODE &&
+      category_filter_.IsCategoryGroupEnabled(category_group))
+    enabled_flag |= ENABLED_FOR_RECORDING;
+  else if (mode_ == MONITORING_MODE &&
+      category_filter_.IsCategoryGroupEnabled(category_group))
+    enabled_flag |= ENABLED_FOR_MONITORING;
+  if (event_callback_ &&
+      event_callback_category_filter_.IsCategoryGroupEnabled(category_group))
+    enabled_flag |= ENABLED_FOR_EVENT_CALLBACK;
+  g_category_group_enabled[category_index] = enabled_flag;
+}
+
+void TraceLog::UpdateCategoryGroupEnabledFlags() {
+  int category_index = base::subtle::NoBarrier_Load(&g_category_index);
+  for (int i = 0; i < category_index; i++)
+    UpdateCategoryGroupEnabledFlag(i);
+}
+
+void TraceLog::UpdateSyntheticDelaysFromCategoryFilter() {
+  ResetTraceEventSyntheticDelays();
+  const CategoryFilter::StringList& delays =
+      category_filter_.GetSyntheticDelayValues();
+  CategoryFilter::StringList::const_iterator ci;
+  for (ci = delays.begin(); ci != delays.end(); ++ci) {
+    std::list<string> tokens = strings::Split(*ci, ";");
+    if (tokens.empty()) continue;
+
+    TraceEventSyntheticDelay* delay =
+        TraceEventSyntheticDelay::Lookup(tokens.front());
+    tokens.pop_front();
+    while (!tokens.empty()) {
+      std::string token = tokens.front();
+      tokens.pop_front();
+      char* duration_end;
+      double target_duration = strtod(token.c_str(), &duration_end);
+      if (duration_end != token.c_str()) {
+        delay->SetTargetDuration(MonoDelta::FromSeconds(target_duration));
+      } else if (token == "static") {
+        delay->SetMode(TraceEventSyntheticDelay::STATIC);
+      } else if (token == "oneshot") {
+        delay->SetMode(TraceEventSyntheticDelay::ONE_SHOT);
+      } else if (token == "alternating") {
+        delay->SetMode(TraceEventSyntheticDelay::ALTERNATING);
+      }
+    }
+  }
+}
+
+const unsigned char* TraceLog::GetCategoryGroupEnabledInternal(
+    const char* category_group) {
+  DCHECK(!strchr(category_group, '"')) <<
+      "Category groups may not contain double quote";
+  // The g_category_groups is append only, avoid using a lock for the fast path.
+  int current_category_index = base::subtle::Acquire_Load(&g_category_index);
+
+  // Search for pre-existing category group.
+  for (int i = 0; i < current_category_index; ++i) {
+    if (strcmp(g_category_groups[i], category_group) == 0) {
+      return &g_category_group_enabled[i];
+    }
+  }
+
+  unsigned char* category_group_enabled = nullptr;
+  // This is the slow path: the lock is not held in the case above, so more
+  // than one thread could have reached here trying to add the same category.
+  // Only hold to lock when actually appending a new category, and
+  // check the categories groups again.
+  SpinLockHolder lock(&lock_);
+  int category_index = base::subtle::Acquire_Load(&g_category_index);
+  for (int i = 0; i < category_index; ++i) {
+    if (strcmp(g_category_groups[i], category_group) == 0) {
+      return &g_category_group_enabled[i];
+    }
+  }
+
+  // Create a new category group.
+  DCHECK(category_index < MAX_CATEGORY_GROUPS) <<
+      "must increase MAX_CATEGORY_GROUPS";
+  if (category_index < MAX_CATEGORY_GROUPS) {
+    // Don't hold on to the category_group pointer, so that we can create
+    // category groups with strings not known at compile time (this is
+    // required by SetWatchEvent).
+    const char* new_group = strdup(category_group);
+    // NOTE: new_group is leaked, but this is a small finite amount of data
+    g_category_groups[category_index] = new_group;
+    DCHECK(!g_category_group_enabled[category_index]);
+    // Note that if both included and excluded patterns in the
+    // CategoryFilter are empty, we exclude nothing,
+    // thereby enabling this category group.
+    UpdateCategoryGroupEnabledFlag(category_index);
+    category_group_enabled = &g_category_group_enabled[category_index];
+    // Update the max index now.
+    base::subtle::Release_Store(&g_category_index, category_index + 1);
+  } else {
+    category_group_enabled =
+        &g_category_group_enabled[kCategoryCategoriesExhausted];
+  }
+  return category_group_enabled;
+}
+
+void TraceLog::GetKnownCategoryGroups(
+    std::vector<std::string>* category_groups) {
+  SpinLockHolder lock(&lock_);
+  int category_index = base::subtle::NoBarrier_Load(&g_category_index);
+  for (int i = kNumBuiltinCategories; i < category_index; i++)
+    category_groups->push_back(g_category_groups[i]);
+}
+
+void TraceLog::SetEnabled(const CategoryFilter& category_filter,
+                          Mode mode,
+                          Options options) {
+  std::vector<EnabledStateObserver*> observer_list;
+  {
+    SpinLockHolder lock(&lock_);
+
+    // Can't enable tracing when Flush() is in progress.
+    Options old_options = trace_options();
+
+    if (IsEnabled()) {
+      if (options != old_options) {
+        DLOG(ERROR) << "Attempting to re-enable tracing with a different "
+                    << "set of options.";
+      }
+
+      if (mode != mode_) {
+        DLOG(ERROR) << "Attempting to re-enable tracing with a different mode.";
+      }
+
+      category_filter_.Merge(category_filter);
+      UpdateCategoryGroupEnabledFlags();
+      return;
+    }
+
+    if (dispatching_to_observer_list_) {
+      DLOG(ERROR) <<
+          "Cannot manipulate TraceLog::Enabled state from an observer.";
+      return;
+    }
+
+    mode_ = mode;
+
+    if (options != old_options) {
+      base::subtle::NoBarrier_Store(&trace_options_, options);
+      UseNextTraceBuffer();
+    }
+
+    num_traces_recorded_++;
+
+    category_filter_ = CategoryFilter(category_filter);
+    UpdateCategoryGroupEnabledFlags();
+    UpdateSyntheticDelaysFromCategoryFilter();
+
+    if (options & ENABLE_SAMPLING) {
+      sampling_thread_.reset(new TraceSamplingThread);
+      sampling_thread_->RegisterSampleBucket(
+          &g_trace_state[0],
+          "bucket0",
+          Bind(&TraceSamplingThread::DefaultSamplingCallback));
+      sampling_thread_->RegisterSampleBucket(
+          &g_trace_state[1],
+          "bucket1",
+          Bind(&TraceSamplingThread::DefaultSamplingCallback));
+      sampling_thread_->RegisterSampleBucket(
+          &g_trace_state[2],
+          "bucket2",
+          Bind(&TraceSamplingThread::DefaultSamplingCallback));
+
+      Status s = Thread::Create("tracing", "sampler",
+                                &TraceSamplingThread::ThreadMain,
+                                sampling_thread_.get(),
+                                &sampling_thread_handle_);
+      if (!s.ok()) {
+        LOG(DFATAL) << "failed to create trace sampling thread: " << s.ToString();
+      }
+    }
+
+    dispatching_to_observer_list_ = true;
+    observer_list = enabled_state_observer_list_;
+  }
+  // Notify observers outside the lock in case they trigger trace events.
+  for (const auto& observer : observer_list)
+    observer->OnTraceLogEnabled();
+
+  {
+    SpinLockHolder lock(&lock_);
+    dispatching_to_observer_list_ = false;
+  }
+}
+
+CategoryFilter TraceLog::GetCurrentCategoryFilter() {
+  SpinLockHolder lock(&lock_);
+  return category_filter_;
+}
+
+void TraceLog::SetDisabled() {
+  SpinLockHolder lock(&lock_);
+  SetDisabledWhileLocked();
+}
+
+void TraceLog::SetDisabledWhileLocked() {
+  DCHECK(lock_.IsHeld());
+
+  if (!IsEnabled())
+    return;
+
+  if (dispatching_to_observer_list_) {
+    DLOG(ERROR)
+        << "Cannot manipulate TraceLog::Enabled state from an observer.";
+    return;
+  }
+
+  mode_ = DISABLED;
+
+  if (sampling_thread_.get()) {
+    // Stop the sampling thread.
+    sampling_thread_->Stop();
+    lock_.Unlock();
+    sampling_thread_handle_->Join();
+    lock_.Lock();
+    sampling_thread_handle_.reset();
+    sampling_thread_.reset();
+  }
+
+  category_filter_.Clear();
+  base::subtle::NoBarrier_Store(&watch_category_, 0);
+  watch_event_name_ = "";
+  UpdateCategoryGroupEnabledFlags();
+  AddMetadataEventsWhileLocked();
+
+  dispatching_to_observer_list_ = true;
+  std::vector<EnabledStateObserver*> observer_list =
+      enabled_state_observer_list_;
+
+  {
+    // Dispatch to observers outside the lock in case the observer triggers a
+    // trace event.
+    lock_.Unlock();
+    for (const auto& observer : observer_list)
+      observer->OnTraceLogDisabled();
+    lock_.Lock();
+  }
+  dispatching_to_observer_list_ = false;
+}
+
+int TraceLog::GetNumTracesRecorded() {
+  SpinLockHolder lock(&lock_);
+  if (!IsEnabled())
+    return -1;
+  return num_traces_recorded_;
+}
+
+void TraceLog::AddEnabledStateObserver(EnabledStateObserver* listener) {
+  enabled_state_observer_list_.push_back(listener);
+}
+
+void TraceLog::RemoveEnabledStateObserver(EnabledStateObserver* listener) {
+  auto it = std::find(enabled_state_observer_list_.begin(),
+                      enabled_state_observer_list_.end(), listener);
+  if (it != enabled_state_observer_list_.end())
+    enabled_state_observer_list_.erase(it);
+}
+
+bool TraceLog::HasEnabledStateObserver(EnabledStateObserver* listener) const {
+  auto it = std::find(enabled_state_observer_list_.begin(),
+                      enabled_state_observer_list_.end(), listener);
+  return it != enabled_state_observer_list_.end();
+}
+
+float TraceLog::GetBufferPercentFull() const {
+  SpinLockHolder lock(&lock_);
+  return static_cast<float>(static_cast<double>(logged_events_->Size()) /
+                            logged_events_->Capacity());
+}
+
+bool TraceLog::BufferIsFull() const {
+  SpinLockHolder lock(&lock_);
+  return logged_events_->IsFull();
+}
+
+TraceBuffer* TraceLog::CreateTraceBuffer() {
+  Options options = trace_options();
+  if (options & RECORD_CONTINUOUSLY)
+    return new TraceBufferRingBuffer(kTraceEventRingBufferChunks);
+  else if ((options & ENABLE_SAMPLING) && mode_ == MONITORING_MODE)
+    return new TraceBufferRingBuffer(kMonitorTraceEventBufferChunks);
+  else if (options & ECHO_TO_CONSOLE)
+    return new TraceBufferRingBuffer(kEchoToConsoleTraceEventBufferChunks);
+  return new TraceBufferVector();
+}
+
+TraceEvent* TraceLog::AddEventToThreadSharedChunkWhileLocked(
+    TraceEventHandle* handle, bool check_buffer_is_full) {
+  DCHECK(lock_.IsHeld());
+
+  if (thread_shared_chunk_ && thread_shared_chunk_->IsFull()) {
+    logged_events_->ReturnChunk(thread_shared_chunk_index_,
+                                std::move(thread_shared_chunk_));
+  }
+
+  if (!thread_shared_chunk_) {
+    thread_shared_chunk_ = logged_events_->GetChunk(
+        &thread_shared_chunk_index_);
+    if (check_buffer_is_full)
+      CheckIfBufferIsFullWhileLocked();
+  }
+  if (!thread_shared_chunk_)
+    return nullptr;
+
+  size_t event_index;
+  TraceEvent* trace_event = thread_shared_chunk_->AddTraceEvent(&event_index);
+  if (trace_event && handle) {
+    MakeHandle(thread_shared_chunk_->seq(), thread_shared_chunk_index_,
+               event_index, handle);
+  }
+  return trace_event;
+}
+
+void TraceLog::CheckIfBufferIsFullWhileLocked() {
+  DCHECK(lock_.IsHeld());
+  if (logged_events_->IsFull())
+    SetDisabledWhileLocked();
+}
+
+void TraceLog::SetEventCallbackEnabled(const CategoryFilter& category_filter,
+                                       EventCallback cb) {
+  SpinLockHolder lock(&lock_);
+  base::subtle::NoBarrier_Store(&event_callback_,
+                          reinterpret_cast<AtomicWord>(cb));
+  event_callback_category_filter_ = category_filter;
+  UpdateCategoryGroupEnabledFlags();
+};
+
+void TraceLog::SetEventCallbackDisabled() {
+  SpinLockHolder lock(&lock_);
+  base::subtle::NoBarrier_Store(&event_callback_, 0);
+  UpdateCategoryGroupEnabledFlags();
+}
+
+// Flush() works as the following:
+//
+// We ensure by taking the global lock that we have exactly one Flusher thread
+// (the caller of this function) and some number of "target" threads. We do
+// not want to block the target threads, since they are running application code,
+// so this implementation takes an approach based on asymmetric synchronization.
+//
+// For each active thread, we grab its PerThreadInfo object, which may contain
+// a pointer to its active trace chunk. We use an AtomicExchange to swap this
+// out for a null pointer. This ensures that, on the *next* TRACE call made by
+// that thread, it will see a NULL buffer and create a _new_ trace buffer. That
+// new buffer would be assigned the generation of the next collection and we don't
+// have to worry about it in the current Flush().
+//
+// However, the swap doesn't ensure that the thread doesn't already have a local copy of
+// the 'event_buffer_' that we are trying to flush. So, if the thread is in the
+// middle of a Trace call, we have to wait until it exits. We do that by spinning
+// on the 'is_in_trace_event_' member of that thread's thread-local structure.
+//
+// After we've swapped the buffer pointer and waited on the thread to exit any
+// concurrent Trace() call, we know that no other thread can hold a pointer to
+// the trace buffer, and we can safely flush it and delete it.
+void TraceLog::Flush(const TraceLog::OutputCallback& cb) {
+  if (IsEnabled()) {
+    // Can't flush when tracing is enabled because otherwise PostTask would
+    // - generate more trace events;
+    // - deschedule the calling thread on some platforms causing inaccurate
+    //   timing of the trace events.
+    scoped_refptr<RefCountedString> empty_result = new RefCountedString;
+    if (!cb.is_null())
+      cb.Run(empty_result, false);
+    LOG(WARNING) << "Ignored TraceLog::Flush called when tracing is enabled";
+    return;
+  }
+
+  int generation = this->generation();
+  {
+    // Holding the active threads lock ensures that no thread will exit and
+    // delete its own PerThreadInfo object.
+    MutexLock l(active_threads_lock_);
+    for (const ActiveThreadMap::value_type& entry : active_threads_) {
+      int64_t tid = entry.first;
+      PerThreadInfo* thr_info = entry.second;
+
+      // Swap out their buffer from their thread-local data.
+      // After this, any _future_ trace calls on that thread will create a new buffer
+      // and not use the one we obtain here.
+      ThreadLocalEventBuffer* buf = thr_info->AtomicTakeBuffer();
+
+      // If this thread hasn't traced anything since our last
+      // flush, we can skip it.
+      if (!buf) {
+        continue;
+      }
+
+      // The buffer may still be in use by that thread if they're in a call. Sleep until
+      // they aren't, so we can flush/delete their old buffer.
+      //
+      // It's important that we do not hold 'lock_' here, because otherwise we can get a
+      // deadlock: a thread may be in the middle of a trace event (is_in_trace_event_ ==
+      // true) and waiting to take lock_, while we are holding the lock and waiting for it
+      // to not be in the trace event.
+      while (base::subtle::Acquire_Load(&thr_info->is_in_trace_event_)) {
+        sched_yield();
+      }
+
+      {
+        SpinLockHolder lock(&lock_);
+        buf->Flush(tid);
+      }
+      delete buf;
+    }
+  }
+
+  {
+    SpinLockHolder lock(&lock_);
+
+    if (thread_shared_chunk_) {
+      logged_events_->ReturnChunk(thread_shared_chunk_index_,
+                                  std::move(thread_shared_chunk_));
+    }
+  }
+
+  FinishFlush(generation, cb);
+}
+
+void TraceLog::ConvertTraceEventsToTraceFormat(
+    gscoped_ptr<TraceBuffer> logged_events,
+    const TraceLog::OutputCallback& flush_output_callback) {
+
+  if (flush_output_callback.is_null())
+    return;
+
+  // The callback need to be called at least once even if there is no events
+  // to let the caller know the completion of flush.
+  bool has_more_events = true;
+  do {
+    scoped_refptr<RefCountedString> json_events_str_ptr =
+        new RefCountedString();
+
+    for (size_t i = 0; i < kTraceEventBatchChunks; ++i) {
+      const TraceBufferChunk* chunk = logged_events->NextChunk();
+      if (!chunk) {
+        has_more_events = false;
+        break;
+      }
+      for (size_t j = 0; j < chunk->size(); ++j) {
+        if (i > 0 || j > 0)
+          json_events_str_ptr->data().append(",");
+        chunk->GetEventAt(j)->AppendAsJSON(&(json_events_str_ptr->data()));
+      }
+    }
+
+    flush_output_callback.Run(json_events_str_ptr, has_more_events);
+  } while (has_more_events);
+  logged_events.reset();
+}
+
+void TraceLog::FinishFlush(int generation,
+                           const TraceLog::OutputCallback& flush_output_callback) {
+  gscoped_ptr<TraceBuffer> previous_logged_events;
+
+  if (!CheckGeneration(generation))
+    return;
+
+  {
+    SpinLockHolder lock(&lock_);
+
+    previous_logged_events.swap(logged_events_);
+    UseNextTraceBuffer();
+  }
+
+  ConvertTraceEventsToTraceFormat(std::move(previous_logged_events),
+                                  flush_output_callback);
+}
+
+void TraceLog::FlushButLeaveBufferIntact(
+    const TraceLog::OutputCallback& flush_output_callback) {
+  gscoped_ptr<TraceBuffer> previous_logged_events;
+  {
+    SpinLockHolder lock(&lock_);
+    if (mode_ == DISABLED || (trace_options_ & RECORD_CONTINUOUSLY) == 0) {
+      scoped_refptr<RefCountedString> empty_result = new RefCountedString;
+      flush_output_callback.Run(empty_result, false);
+      LOG(WARNING) << "Ignored TraceLog::FlushButLeaveBufferIntact when monitoring is not enabled";
+      return;
+    }
+
+    AddMetadataEventsWhileLocked();
+    if (thread_shared_chunk_) {
+      // Return the chunk to the main buffer to flush the sampling data.
+      logged_events_->ReturnChunk(thread_shared_chunk_index_,
+                                  std::move(thread_shared_chunk_));
+    }
+    previous_logged_events = logged_events_->CloneForIteration();
+  }
+
+  ConvertTraceEventsToTraceFormat(std::move(previous_logged_events),
+                                  flush_output_callback);
+}
+
+void TraceLog::UseNextTraceBuffer() {
+  logged_events_.reset(CreateTraceBuffer());
+  base::subtle::NoBarrier_AtomicIncrement(&generation_, 1);
+  thread_shared_chunk_.reset();
+  thread_shared_chunk_index_ = 0;
+}
+
+TraceEventHandle TraceLog::AddTraceEvent(
+    char phase,
+    const unsigned char* category_group_enabled,
+    const char* name,
+    uint64_t id,
+    int num_args,
+    const char** arg_names,
+    const unsigned char* arg_types,
+    const uint64_t* arg_values,
+    const scoped_refptr<ConvertableToTraceFormat>* convertable_values,
+    unsigned char flags) {
+  int thread_id = static_cast<int>(kudu::Thread::UniqueThreadId());
+  MicrosecondsInt64 now = GetMonoTimeMicros();
+  return AddTraceEventWithThreadIdAndTimestamp(phase, category_group_enabled,
+                                               name, id, thread_id, now,
+                                               num_args, arg_names,
+                                               arg_types, arg_values,
+                                               convertable_values, flags);
+}
+
+TraceLog::PerThreadInfo* TraceLog::SetupThreadLocalBuffer() {
+  int64_t cur_tid = Thread::UniqueThreadId();
+
+  auto thr_info = new PerThreadInfo();
+  thr_info->event_buffer_ = nullptr;
+  thr_info->is_in_trace_event_ = 0;
+  thread_local_info_ = thr_info;
+
+  threadlocal::internal::AddDestructor(&TraceLog::ThreadExitingCB, this);
+
+  {
+    MutexLock lock(active_threads_lock_);
+    InsertOrDie(&active_threads_, cur_tid, thr_info);
+  }
+  return thr_info;
+}
+
+void TraceLog::ThreadExitingCB(void* arg) {
+  static_cast<TraceLog*>(arg)->ThreadExiting();
+}
+
+void TraceLog::ThreadExiting() {
+  PerThreadInfo* thr_info = thread_local_info_;
+  if (!thr_info) {
+    return;
+  }
+
+  int64_t cur_tid = Thread::UniqueThreadId();
+
+  // Flush our own buffer back to the central event buffer.
+  // We do the atomic exchange because a flusher thread may
+  // also be trying to flush us at the same time, and we need to avoid
+  // conflict.
+  ThreadLocalEventBuffer* buf = thr_info->AtomicTakeBuffer();
+  if (buf) {
+    SpinLockHolder lock(&lock_);
+    buf->Flush(Thread::UniqueThreadId());
+  }
+  delete buf;
+
+  {
+    MutexLock lock(active_threads_lock_);
+    active_threads_.erase(cur_tid);
+  }
+  delete thr_info;
+}
+
+TraceEventHandle TraceLog::AddTraceEventWithThreadIdAndTimestamp(
+    char phase,
+    const unsigned char* category_group_enabled,
+    const char* name,
+    uint64_t id,
+    int thread_id,
+    const MicrosecondsInt64& timestamp,
+    int num_args,
+    const char** arg_names,
+    const unsigned char* arg_types,
+    const uint64_t* arg_values,
+    const scoped_refptr<ConvertableToTraceFormat>* convertable_values,
+    unsigned char flags) {
+  TraceEventHandle handle = { 0, 0, 0 };
+  if (!*category_group_enabled)
+    return handle;
+
+  DCHECK(name);
+
+  if (flags & TRACE_EVENT_FLAG_MANGLE_ID)
+    id ^= process_id_hash_;
+
+  MicrosecondsInt64 now = OffsetTimestamp(timestamp);
+  MicrosecondsInt64 thread_now = GetThreadCpuTimeMicros();
+
+  PerThreadInfo* thr_info = thread_local_info_;
+  if (PREDICT_FALSE(!thr_info)) {
+    thr_info = SetupThreadLocalBuffer();
+  }
+
+  // Avoid re-entrance of AddTraceEvent. This may happen in GPU process when
+  // ECHO_TO_CONSOLE is enabled: AddTraceEvent -> LOG(ERROR) ->
+  // GpuProcessLogMessageHandler -> PostPendingTask -> TRACE_EVENT ...
+  if (base::subtle::NoBarrier_Load(&thr_info->is_in_trace_event_))
+    return handle;
+
+  MarkFlagInScope thread_is_in_trace_event(&thr_info->is_in_trace_event_);
+
+  ThreadLocalEventBuffer* thread_local_event_buffer =
+    reinterpret_cast<ThreadLocalEventBuffer*>(
+      base::subtle::NoBarrier_Load(
+        reinterpret_cast<AtomicWord*>(&thr_info->event_buffer_)));
+
+  // If we have an event buffer, but it's a left-over from a previous trace,
+  // delete it.
+  if (PREDICT_FALSE(thread_local_event_buffer &&
+                    !CheckGeneration(thread_local_event_buffer->generation()))) {
+    // We might also race against a flusher thread, so we have to atomically
+    // take the buffer.
+    thread_local_event_buffer = thr_info->AtomicTakeBuffer();
+    delete thread_local_event_buffer;
+    thread_local_event_buffer = nullptr;
+  }
+
+  // If there is no current buffer, create one for this event.
+  if (PREDICT_FALSE(!thread_local_event_buffer)) {
+    thread_local_event_buffer = new ThreadLocalEventBuffer(this);
+
+    base::subtle::NoBarrier_Store(
+      reinterpret_cast<AtomicWord*>(&thr_info->event_buffer_),
+      reinterpret_cast<AtomicWord>(thread_local_event_buffer));
+  }
+
+  // Check and update the current thread name only if the event is for the
+  // current thread to avoid locks in most cases.
+  if (thread_id == static_cast<int>(Thread::UniqueThreadId())) {
+    Thread* kudu_thr = Thread::current_thread();
+    if (kudu_thr) {
+      const char* new_name = kudu_thr->name().c_str();
+      // Check if the thread name has been set or changed since the previous
+      // call (if any), but don't bother if the new name is empty. Note this will
+      // not detect a thread name change within the same char* buffer address: we
+      // favor common case performance over corner case correctness.
+      if (PREDICT_FALSE(new_name != g_current_thread_name &&
+                        new_name && *new_name)) {
+        g_current_thread_name = new_name;
+
+        SpinLockHolder thread_info_lock(&thread_info_lock_);
+
+        auto existing_name = thread_names_.find(thread_id);
+        if (existing_name == thread_names_.end()) {
+          // This is a new thread id, and a new name.
+          thread_names_[thread_id] = new_name;
+        } else {
+          // This is a thread id that we've seen before, but potentially with a
+          // new name.
+          std::vector<StringPiece> existing_names = strings::Split(existing_name->second, ",");
+          bool found = std::find(existing_names.begin(),
+                                 existing_names.end(),
+                                 new_name) != existing_names.end();
+          if (!found) {
+            if (existing_names.size())
+              existing_name->second.push_back(',');
+            existing_name->second.append(new_name);
+          }
+        }
+      }
+    }
+  }
+
+  std::string console_message;
+  if (*category_group_enabled &
+      (ENABLED_FOR_RECORDING | ENABLED_FOR_MONITORING)) {
+    TraceEvent* trace_event = thread_local_event_buffer->AddTraceEvent(&handle);
+
+    if (trace_event) {
+      trace_event->Initialize(thread_id, now, thread_now, phase,
+                              category_group_enabled, name, id,
+                              num_args, arg_names, arg_types, arg_values,
+                              convertable_values, flags);
+
+#if defined(OS_ANDROID)
+      trace_event->SendToATrace();
+#endif
+    }
+
+    if (trace_options() & ECHO_TO_CONSOLE) {
+      console_message = EventToConsoleMessage(
+          phase == TRACE_EVENT_PHASE_COMPLETE ? TRACE_EVENT_PHASE_BEGIN : phase,
+          timestamp, trace_event);
+    }
+  }
+
+  if (PREDICT_FALSE(console_message.size()))
+    LOG(ERROR) << console_message;
+
+  if (PREDICT_FALSE(reinterpret_cast<const unsigned char*>(
+                      base::subtle::NoBarrier_Load(&watch_category_)) == category_group_enabled)) {
+    bool event_name_matches;
+    WatchEventCallback watch_event_callback_copy;
+    {
+      SpinLockHolder lock(&lock_);
+      event_name_matches = watch_event_name_ == name;
+      watch_event_callback_copy = watch_event_callback_;
+    }
+    if (event_name_matches) {
+      if (!watch_event_callback_copy.is_null())
+        watch_event_callback_copy.Run();
+    }
+  }
+
+  if (PREDICT_FALSE(*category_group_enabled & ENABLED_FOR_EVENT_CALLBACK)) {
+    EventCallback event_callback = reinterpret_cast<EventCallback>(
+      base::subtle::NoBarrier_Load(&event_callback_));
+    if (event_callback) {
+      event_callback(now,
+                     phase == TRACE_EVENT_PHASE_COMPLETE ?
+                         TRACE_EVENT_PHASE_BEGIN : phase,
+                     category_group_enabled, name, id,
+                     num_args, arg_names, arg_types, arg_values,
+                     flags);
+    }
+  }
+
+  return handle;
+}
+
+// May be called when a COMPELETE event ends and the unfinished event has been
+// recycled (phase == TRACE_EVENT_PHASE_END and trace_event == NULL).
+std::string TraceLog::EventToConsoleMessage(unsigned char phase,
+                                            const MicrosecondsInt64& timestamp,
+                                            TraceEvent* trace_event) {
+  SpinLockHolder thread_info_lock(&thread_info_lock_);
+
+  // The caller should translate TRACE_EVENT_PHASE_COMPLETE to
+  // TRACE_EVENT_PHASE_BEGIN or TRACE_EVENT_END.
+  DCHECK(phase != TRACE_EVENT_PHASE_COMPLETE);
+
+  MicrosecondsInt64 duration;
+  int thread_id = trace_event ?
+      trace_event->thread_id() : Thread::UniqueThreadId();
+  if (phase == TRACE_EVENT_PHASE_END) {
+    duration = timestamp - thread_event_start_times_[thread_id].top();
+    thread_event_start_times_[thread_id].pop();
+  }
+
+  std::string thread_name = thread_names_[thread_id];
+  if (thread_colors_.find(thread_name) == thread_colors_.end())
+    thread_colors_[thread_name] = (thread_colors_.size() % 6) + 1;
+
+  std::ostringstream log;
+  log << StringPrintf("%s: \x1b[0;3%dm",
+                            thread_name.c_str(),
+                            thread_colors_[thread_name]);
+
+  size_t depth = 0;
+  if (thread_event_start_times_.find(thread_id) !=
+      thread_event_start_times_.end())
+    depth = thread_event_start_times_[thread_id].size();
+
+  for (size_t i = 0; i < depth; ++i)
+    log << "| ";
+
+  if (trace_event)
+    trace_event->AppendPrettyPrinted(&log);
+  if (phase == TRACE_EVENT_PHASE_END)
+    log << StringPrintf(" (%.3f ms)", duration / 1000.0f);
+
+  log << "\x1b[0;m";
+
+  if (phase == TRACE_EVENT_PHASE_BEGIN)
+    thread_event_start_times_[thread_id].push(timestamp);
+
+  return log.str();
+}
+
+void TraceLog::AddTraceEventEtw(char phase,
+                                const char* name,
+                                const void* id,
+                                const char* extra) {
+#if defined(OS_WIN)
+  TraceEventETWProvider::Trace(name, phase, id, extra);
+#endif
+  INTERNAL_TRACE_EVENT_ADD(phase, "ETW Trace Event", name,
+                           TRACE_EVENT_FLAG_COPY, "id", id, "extra", extra);
+}
+
+void TraceLog::AddTraceEventEtw(char phase,
+                                const char* name,
+                                const void* id,
+                                const std::string& extra) {
+#if defined(OS_WIN)
+  TraceEventETWProvider::Trace(name, phase, id, extra);
+#endif
+  INTERNAL_TRACE_EVENT_ADD(phase, "ETW Trace Event", name,
+                           TRACE_EVENT_FLAG_COPY, "id", id, "extra", extra);
+}
+
+void TraceLog::UpdateTraceEventDuration(
+    const unsigned char* category_group_enabled,
+    const char* name,
+    TraceEventHandle handle) {
+
+  PerThreadInfo* thr_info = thread_local_info_;
+  if (!thr_info) {
+    thr_info = SetupThreadLocalBuffer();
+  }
+
+  // Avoid re-entrance of AddTraceEvent. This may happen in GPU process when
+  // ECHO_TO_CONSOLE is enabled: AddTraceEvent -> LOG(ERROR) ->
+  // GpuProcessLogMessageHandler -> PostPendingTask -> TRACE_EVENT ...
+  if (base::subtle::NoBarrier_Load(&thr_info->is_in_trace_event_))
+    return;
+  MarkFlagInScope thread_is_in_trace_event(&thr_info->is_in_trace_event_);
+
+  MicrosecondsInt64 thread_now = GetThreadCpuTimeMicros();
+  MicrosecondsInt64 now = OffsetNow();
+
+  std::string console_message;
+  if (*category_group_enabled & ENABLED_FOR_RECORDING) {
+    OptionalAutoLock lock(lock_);
+
+    TraceEvent* trace_event = GetEventByHandleInternal(handle, &lock);
+    if (trace_event) {
+      DCHECK(trace_event->phase() == TRACE_EVENT_PHASE_COMPLETE);
+      trace_event->UpdateDuration(now, thread_now);
+#if defined(OS_ANDROID)
+      trace_event->SendToATrace();
+#endif
+    }
+
+    if (trace_options() & ECHO_TO_CONSOLE) {
+      console_message = EventToConsoleMessage(TRACE_EVENT_PHASE_END,
+                                              now, trace_event);
+    }
+  }
+
+  if (console_message.size())
+    LOG(ERROR) << console_message;
+
+  if (*category_group_enabled & ENABLED_FOR_EVENT_CALLBACK) {
+    EventCallback event_callback = reinterpret_cast<EventCallback>(
+      base::subtle::NoBarrier_Load(&event_callback_));
+    if (event_callback) {
+      event_callback(now, TRACE_EVENT_PHASE_END, category_group_enabled, name,
+                     trace_event_internal::kNoEventId, 0, nullptr, nullptr, nullptr,
+                     TRACE_EVENT_FLAG_NONE);
+    }
+  }
+}
+
+void TraceLog::SetWatchEvent(const std::string& category_name,
+                             const std::string& event_name,
+                             const WatchEventCallback& callback) {
+  const unsigned char* category = GetCategoryGroupEnabled(
+      category_name.c_str());
+  SpinLockHolder lock(&lock_);
+  base::subtle::NoBarrier_Store(&watch_category_,
+                          reinterpret_cast<AtomicWord>(category));
+  watch_event_name_ = event_name;
+  watch_event_callback_ = callback;
+}
+
+void TraceLog::CancelWatchEvent() {
+  SpinLockHolder lock(&lock_);
+  base::subtle::NoBarrier_Store(&watch_category_, 0);
+  watch_event_name_ = "";
+  watch_event_callback_.Reset();
+}
+
+void TraceLog::AddMetadataEventsWhileLocked() {
+  DCHECK(lock_.IsHeld());
+
+#if !defined(OS_NACL)  // NaCl shouldn't expose the process id.
+  InitializeMetadataEvent(AddEventToThreadSharedChunkWhileLocked(nullptr, false),
+                          0,
+                          "num_cpus", "number",
+                          base::NumCPUs());
+#endif
+
+
+  int current_thread_id = static_cast<int>(kudu::Thread::UniqueThreadId());
+  if (process_sort_index_ != 0) {
+    InitializeMetadataEvent(AddEventToThreadSharedChunkWhileLocked(nullptr, false),
+                            current_thread_id,
+                            "process_sort_index", "sort_index",
+                            process_sort_index_);
+  }
+
+  if (process_name_.size()) {
+    InitializeMetadataEvent(AddEventToThreadSharedChunkWhileLocked(nullptr, false),
+                            current_thread_id,
+                            "process_name", "name",
+                            process_name_);
+  }
+
+  if (process_labels_.size() > 0) {
+    std::vector<std::string> labels;
+    for(auto& label : process_labels_) {
+      labels.push_back(label.second);
+    }
+    InitializeMetadataEvent(AddEventToThreadSharedChunkWhileLocked(nullptr, false),
+                            current_thread_id,
+                            "process_labels", "labels",
+                            JoinStrings(labels, ","));
+  }
+
+  // Thread sort indices.
+  for(auto& sort_index : thread_sort_indices_) {
+    if (sort_index.second == 0)
+      continue;
+    InitializeMetadataEvent(AddEventToThreadSharedChunkWhileLocked(nullptr, false),
+                            sort_index.first,
+                            "thread_sort_index", "sort_index",
+                            sort_index.second);
+  }
+
+  // Thread names.
+  SpinLockHolder thread_info_lock(&thread_info_lock_);
+  for(auto& name : thread_names_) {
+    if (name.second.empty())
+      continue;
+    InitializeMetadataEvent(AddEventToThreadSharedChunkWhileLocked(nullptr, false),
+                            name.first,
+                            "thread_name", "name",
+                            name.second);
+  }
+}
+
+
+TraceEvent* TraceLog::GetEventByHandle(TraceEventHandle handle) {
+  return GetEventByHandleInternal(handle, nullptr);
+}
+
+TraceEvent* TraceLog::GetEventByHandleInternal(TraceEventHandle handle,
+                                               OptionalAutoLock* lock) {
+  TraceLog::PerThreadInfo* thr_info = TraceLog::thread_local_info_;
+
+  if (!handle.chunk_seq)
+    return nullptr;
+
+  if (thr_info) {
+    ThreadLocalEventBuffer* buf =
+      reinterpret_cast<ThreadLocalEventBuffer*>(
+        base::subtle::NoBarrier_Load(
+          reinterpret_cast<AtomicWord*>(&thr_info->event_buffer_)));
+
+    if (buf) {
+      DCHECK_EQ(1, ANNOTATE_UNPROTECTED_READ(thr_info->is_in_trace_event_));
+
+      TraceEvent* trace_event = buf->GetEventByHandle(handle);
+      if (trace_event)
+        return trace_event;
+    }
+  }
+
+  // The event has been out-of-control of the thread local buffer.
+  // Try to get the event from the main buffer with a lock.
+  if (lock)
+    lock->EnsureAcquired();
+
+  if (thread_shared_chunk_ &&
+      handle.chunk_index == thread_shared_chunk_index_) {
+    return handle.chunk_seq == thread_shared_chunk_->seq() ?
+        thread_shared_chunk_->GetEventAt(handle.event_index) : nullptr;
+  }
+
+  return logged_events_->GetEventByHandle(handle);
+}
+
+ATTRIBUTE_NO_SANITIZE_INTEGER
+void TraceLog::SetProcessID(int process_id) {
+  process_id_ = process_id;
+  // Create a FNV hash from the process ID for XORing.
+  // See http://isthe.com/chongo/tech/comp/fnv/ for algorithm details.
+  uint64_t offset_basis = 14695981039346656037ull;
+  uint64_t fnv_prime = 1099511628211ull;
+  uint64_t pid = static_cast<uint64_t>(process_id_);
+  process_id_hash_ = (offset_basis ^ pid) * fnv_prime;
+}
+
+void TraceLog::SetProcessSortIndex(int sort_index) {
+  SpinLockHolder lock(&lock_);
+  process_sort_index_ = sort_index;
+}
+
+void TraceLog::SetProcessName(const std::string& process_name) {
+  SpinLockHolder lock(&lock_);
+  process_name_ = process_name;
+}
+
+void TraceLog::UpdateProcessLabel(
+    int label_id, const std::string& current_label) {
+  if(!current_label.length())
+    return RemoveProcessLabel(label_id);
+
+  SpinLockHolder lock(&lock_);
+  process_labels_[label_id] = current_label;
+}
+
+void TraceLog::RemoveProcessLabel(int label_id) {
+  SpinLockHolder lock(&lock_);
+  auto it = process_labels_.find(label_id);
+  if (it == process_labels_.end())
+    return;
+
+  process_labels_.erase(it);
+}
+
+void TraceLog::SetThreadSortIndex(int64_t thread_id, int sort_index) {
+  SpinLockHolder lock(&lock_);
+  thread_sort_indices_[static_cast<int>(thread_id)] = sort_index;
+}
+
+void TraceLog::SetTimeOffset(MicrosecondsInt64 offset) {
+  time_offset_ = offset;
+}
+
+size_t TraceLog::GetObserverCountForTest() const {
+  return enabled_state_observer_list_.size();
+}
+
+bool CategoryFilter::IsEmptyOrContainsLeadingOrTrailingWhitespace(
+    const std::string& str) {
+  return  str.empty() ||
+          str.at(0) == ' ' ||
+          str.at(str.length() - 1) == ' ';
+}
+
+bool CategoryFilter::DoesCategoryGroupContainCategory(
+    const char* category_group,
+    const char* category) const {
+  DCHECK(category);
+  vector<string> pieces = strings::Split(category_group, ",");
+  for (const string& category_group_token : pieces) {
+    // Don't allow empty tokens, nor tokens with leading or trailing space.
+    DCHECK(!CategoryFilter::IsEmptyOrContainsLeadingOrTrailingWhitespace(
+        category_group_token))
+        << "Disallowed category string";
+
+    if (MatchPattern(category_group_token.c_str(), category))
+      return true;
+  }
+  return false;
+}
+
+CategoryFilter::CategoryFilter(const std::string& filter_string) {
+  if (!filter_string.empty())
+    Initialize(filter_string);
+  else
+    Initialize(CategoryFilter::kDefaultCategoryFilterString);
+}
+
+CategoryFilter::CategoryFilter(const CategoryFilter& cf)
+    : included_(cf.included_),
+      disabled_(cf.disabled_),
+      excluded_(cf.excluded_),
+      delays_(cf.delays_) {
+}
+
+CategoryFilter::~CategoryFilter() {
+}
+
+CategoryFilter& CategoryFilter::operator=(const CategoryFilter& rhs) {
+  if (this == &rhs)
+    return *this;
+
+  included_ = rhs.included_;
+  disabled_ = rhs.disabled_;
+  excluded_ = rhs.excluded_;
+  delays_ = rhs.delays_;
+  return *this;
+}
+
+void CategoryFilter::Initialize(const std::string& filter_string) {
+  // Tokenize list of categories, delimited by ','.
+  vector<string> tokens = strings::Split(filter_string, ",");
+  // Add each token to the appropriate list (included_,excluded_).
+  for (string category : tokens) {
+    // Ignore empty categories.
+    if (category.empty())
+      continue;
+    // Synthetic delays are of the form 'DELAY(delay;option;option;...)'.
+    if (category.find(kSyntheticDelayCategoryFilterPrefix) == 0 &&
+        category.at(category.size() - 1) == ')') {
+      category = category.substr(
+          strlen(kSyntheticDelayCategoryFilterPrefix),
+          category.size() - strlen(kSyntheticDelayCategoryFilterPrefix) - 1);
+      size_t name_length = category.find(';');
+      if (name_length != std::string::npos && name_length > 0 &&
+          name_length != category.size() - 1) {
+        delays_.push_back(category);
+      }
+    } else if (category.at(0) == '-') {
+      // Excluded categories start with '-'.
+      // Remove '-' from category string.
+      category = category.substr(1);
+      excluded_.push_back(category);
+    } else if (category.compare(0, strlen(TRACE_DISABLED_BY_DEFAULT("")),
+                                TRACE_DISABLED_BY_DEFAULT("")) == 0) {
+      disabled_.push_back(category);
+    } else {
+      included_.push_back(category);
+    }
+  }
+}
+
+void CategoryFilter::WriteString(const StringList& values,
+                                 std::string* out,
+                                 bool included) const {
+  bool prepend_comma = !out->empty();
+  int token_cnt = 0;
+  for (const auto& value : values) {
+    if (token_cnt > 0 || prepend_comma)
+      StringAppendF(out, ",");
+    StringAppendF(out, "%s%s", (included ? "" : "-"), value.c_str());
+    ++token_cnt;
+  }
+}
+
+void CategoryFilter::WriteString(const StringList& delays,
+                                 std::string* out) const {
+  bool prepend_comma = !out->empty();
+  int token_cnt = 0;
+  for (const auto& delay : delays) {
+    if (token_cnt > 0 || prepend_comma)
+      StringAppendF(out, ",");
+    StringAppendF(out, "%s%s)", kSyntheticDelayCategoryFilterPrefix,
+                  delay.c_str());
+    ++token_cnt;
+  }
+}
+
+std::string CategoryFilter::ToString() const {
+  std::string filter_string;
+  WriteString(included_, &filter_string, true);
+  WriteString(disabled_, &filter_string, true);
+  WriteString(excluded_, &filter_string, false);
+  WriteString(delays_, &filter_string);
+  return filter_string;
+}
+
+bool CategoryFilter::IsCategoryGroupEnabled(
+    const char* category_group_name) const {
+  // TraceLog should call this method only as  part of enabling/disabling
+  // categories.
+  StringList::const_iterator ci;
+
+  // Check the disabled- filters and the disabled-* wildcard first so that a
+  // "*" filter does not include the disabled.
+  for (ci = disabled_.begin(); ci != disabled_.end(); ++ci) {
+    if (DoesCategoryGroupContainCategory(category_group_name, ci->c_str()))
+      return true;
+  }
+  if (DoesCategoryGroupContainCategory(category_group_name,
+                                       TRACE_DISABLED_BY_DEFAULT("*")))
+    return false;
+
+  for (ci = included_.begin(); ci != included_.end(); ++ci) {
+    if (DoesCategoryGroupContainCategory(category_group_name, ci->c_str()))
+      return true;
+  }
+
+  for (ci = excluded_.begin(); ci != excluded_.end(); ++ci) {
+    if (DoesCategoryGroupContainCategory(category_group_name, ci->c_str()))
+      return false;
+  }
+  // If the category group is not excluded, and there are no included patterns
+  // we consider this pattern enabled.
+  return included_.empty();
+}
+
+bool CategoryFilter::HasIncludedPatterns() const {
+  return !included_.empty();
+}
+
+void CategoryFilter::Merge(const CategoryFilter& nested_filter) {
+  // Keep included patterns only if both filters have an included entry.
+  // Otherwise, one of the filter was specifying "*" and we want to honour the
+  // broadest filter.
+  if (HasIncludedPatterns() && nested_filter.HasIncludedPatterns()) {
+    included_.insert(included_.end(),
+                     nested_filter.included_.begin(),
+                     nested_filter.included_.end());
+  } else {
+    included_.clear();
+  }
+
+  disabled_.insert(disabled_.end(),
+                   nested_filter.disabled_.begin(),
+                   nested_filter.disabled_.end());
+  excluded_.insert(excluded_.end(),
+                   nested_filter.excluded_.begin(),
+                   nested_filter.excluded_.end());
+  delays_.insert(delays_.end(),
+                 nested_filter.delays_.begin(),
+                 nested_filter.delays_.end());
+}
+
+void CategoryFilter::Clear() {
+  included_.clear();
+  disabled_.clear();
+  excluded_.clear();
+}
+
+const CategoryFilter::StringList&
+    CategoryFilter::GetSyntheticDelayValues() const {
+  return delays_;
+}
+
+}  // namespace debug
+}  // namespace kudu
+
+namespace trace_event_internal {
+
+ScopedTraceBinaryEfficient::ScopedTraceBinaryEfficient(
+    const char* category_group, const char* name) {
+  // The single atom works because for now the category_group can only be "gpu".
+  DCHECK(strcmp(category_group, "gpu") == 0);
+  static TRACE_EVENT_API_ATOMIC_WORD atomic = 0;
+  INTERNAL_TRACE_EVENT_GET_CATEGORY_INFO_CUSTOM_VARIABLES(
+      category_group, atomic, category_group_enabled_);
+  name_ = name;
+  if (*category_group_enabled_) {
+    event_handle_ =
+        TRACE_EVENT_API_ADD_TRACE_EVENT_WITH_THREAD_ID_AND_TIMESTAMP(
+            TRACE_EVENT_PHASE_COMPLETE, category_group_enabled_, name,
+            trace_event_internal::kNoEventId,
+            static_cast<int>(kudu::Thread::UniqueThreadId()),
+            GetMonoTimeMicros(),
+            0, nullptr, nullptr, nullptr, nullptr, TRACE_EVENT_FLAG_NONE);
+  }
+}
+
+ScopedTraceBinaryEfficient::~ScopedTraceBinaryEfficient() {
+  if (*category_group_enabled_) {
+    TRACE_EVENT_API_UPDATE_TRACE_EVENT_DURATION(category_group_enabled_,
+                                                name_, event_handle_);
+  }
+}
+
+}  // namespace trace_event_internal