Tracing V2 prototype [NOT FOR REVIEW]

base/trace_event/v2/ring_buffer.cc

+// Copyright 2016 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 "base/trace_event/v2/ring_buffer.h"
+
+#include "base/threading/platform_thread.h"
+#include "base/trace_event/common/proto/events_chunk.tracing-pb.h"
+#include "base/trace_event/v2/proto_utils.h"
+#include "base/trace_event/v2/append_only_proto_message.h"
+
+namespace base {
+namespace trace_event {
+namespace v2 {
+
+using ChunkProto = tracing::EventsChunk;
+
+namespace {
+
+const size_t kChunkSize = 32*1024;
+const size_t kEventPreambleSize = AppendOnlyProtoMessage::kMaxMessageLengthFieldSize + 1;
+
+uint8_t* WriteProtoVarInt(uint8_t* dst, uint32_t field_id, uint32_t value) {
+  dst += ProtoUtils::EncodeVarIntUnsigned(
+      ProtoUtils::GetVarIntFieldHeader(field_id), dst);
+  dst += ProtoUtils::EncodeVarIntUnsigned(value, dst);
+  return dst;
+}
+
+uint8_t* WriteEventPreamble(TraceEvent* event, uint8_t* dst) {
+  static_assert(
+      ChunkProto::kEventsFieldNumber <= 31,
+      "Chunk.event must have a field id <= 31 to stay within 1 byte preamble");
+  uint8_t* buf = dst;
+  *buf = static_cast<uint8_t>(ProtoUtils::GetLengthLimitedFieldHeader(ChunkProto::kEventsFieldNumber));
+  ++buf;
+  ScatteredBuffer::ContiguousMemoryRange reserved_range;
+  reserved_range.begin = buf;
+  buf += AppendOnlyProtoMessage::kMaxMessageLengthFieldSize;
+  reserved_range.end = buf;
+  event->set_write_size_field_on_finalization(reserved_range);
+  DCHECK_EQ(kEventPreambleSize, static_cast<size_t>(buf - dst));
+  return buf;
+}
+
+}  // namespace
+
+RingBuffer::RingBuffer(uint8_t* begin, size_t size)
+    : begin_(begin),
+      end_(begin + size),
+      num_chunks_(size / kChunkSize),
+      current_chunk_idx_(0) {
+  DCHECK_GE(size, kChunkSize);
+  DCHECK_EQ(0ul, size % kChunkSize);
+  chunks_.reset(new Chunk[num_chunks_]);
+  uint8_t* chunk_begin = begin;
+  for (size_t i = 0; i < num_chunks_; ++i) {
+    Chunk& chunk = chunks_[i];
+    chunk.begin_ = chunk_begin;
+    chunk.end_ = chunk_begin + kChunkSize;
+    chunk_begin = chunk.end_;
+  }
+}
+
+RingBuffer::~RingBuffer() {}
+
+RingBuffer::Chunk* RingBuffer::TakeChunk() {
+  AutoLock lock(lock_);
+  DCHECK_GT(num_chunks_, 0ul);
+  DCHECK_LT(current_chunk_idx_, num_chunks_);
+  Chunk* chunk = nullptr;
+  for (size_t i = 0; i < num_chunks_; ++i) {
+    chunk = &chunks_[current_chunk_idx_];
+    ++current_chunk_idx_;
+    if (current_chunk_idx_ >= num_chunks_)
+      current_chunk_idx_ = 0;
+    if (chunk->is_returned()) {
+      chunk->set_owner(PlatformThread::CurrentId());
+      chunk->set_proto_used_size(0);
+      return chunk;
+    }
+  }
+
+  // Bankrupcy: there are more threads than chunks. All chunks were on flight.
+  return nullptr;  // TODO return a fallback chunk where everybody locks?
+}
+
+void RingBuffer::ReturnChunk(RingBuffer::Chunk* chunk, uint32_t used_size) {
+  AutoLock lock(lock_);
+  chunk->set_proto_used_size(used_size);
+  chunk->set_returned();
+}
+
+RingBuffer::Chunk::Chunk()
+    : begin_(nullptr), end_(nullptr), owner_(kInvalidThreadId) {}
+RingBuffer::Chunk::~Chunk() {}
+
+ZeroCopyTraceBufferWriter::ZeroCopyTraceBufferWriter(RingBuffer* ring_buffer,
+                                                     uint32_t stream_id)
+    : ring_buffer_(ring_buffer),
+      chunk_(nullptr),
+      stream_id_(stream_id),
+      chunk_seq_id_(0),
+      continue_on_next_chunk_ptr_(nullptr),
+      event_start_addr_(nullptr),
+      event_(nullptr),
+      scattered_buffer_(this) {
+  memset(event_storage_, 0, sizeof(event_storage_));
+}
+
+ZeroCopyTraceBufferWriter::~ZeroCopyTraceBufferWriter() {}
+
+TraceEventHandle ZeroCopyTraceBufferWriter::AddEvent() {
+  if (event_)
+    event_->~TraceEvent();
+  event_ = new (event_storage_) TraceEvent();
+  // event_.reset(new TraceEvent());
+
+  // In order to start a new event at least kMaxMessageLengthFieldSize + 1 bytes
+  // are required in the chunk, to write at least the size of the partial event.
+  // We stay a bit more conservative here purely for performance reasons. It
+  // doesn't make a lot of sense starting a partial event that will fragment
+  // immediately.
+  // TODO bump up 8 to a reasonable value.
+  if (scattered_buffer_.contiguous_bytes_available() < 8) {
+    scattered_buffer_.Reset(
+        AcquireNewChunk(false /* event_continues_from_prev_chunk */));
+  }
+
+  auto range = scattered_buffer_.ReserveBytes(kEventPreambleSize);
+  event_start_addr_= WriteEventPreamble(event_, range.begin);
+  DCHECK_EQ(scattered_buffer_.write_ptr(), event_start_addr_);
+  event_->set_buffer_writer(&scattered_buffer_);
+
+  // TODO keep alive chunks for reservation until event is finalized. Hmmmmm.
+  return TraceEventHandle(event_);
+}
+
+ScatteredBuffer::ContiguousMemoryRange
+ZeroCopyTraceBufferWriter::GetNewContiguousMemoryBuffer() {
+  return AcquireNewChunk(true /* event_continues_from_prev_chunk */);
+}
+
+ScatteredBuffer::ContiguousMemoryRange
+ZeroCopyTraceBufferWriter::AcquireNewChunk(bool event_continues_from_prev_chunk) {
+  if (event_continues_from_prev_chunk) {
+    // The |event_| spawn across multiple chunks.
+    *continue_on_next_chunk_ptr_ = 1;
+    const size_t event_partial_size =
+        static_cast<size_t>(scattered_buffer_.write_ptr() - event_start_addr_);
+
+    ProtoUtils::EncodeRedundantVarIntUnsigned(
+        event_partial_size,
+        AppendOnlyProtoMessage::kMaxMessageLengthFieldSize,
+        event_->size_field().begin);
+    event_->inc_size_already_written(event_partial_size);
+  }
+
+  // Return the current chunk and acquire a new one.
+  if (chunk_) {
+    DCHECK_GE(scattered_buffer_.write_ptr(), chunk_->proto_begin());
+    DCHECK_LE(scattered_buffer_.write_ptr(), chunk_->end());
+    const uint32_t used_size = static_cast<uint32_t>(scattered_buffer_.write_ptr() - chunk_->proto_begin());
+    ring_buffer_->ReturnChunk(chunk_, used_size);
+  }
+  chunk_ = ring_buffer_->TakeChunk();
+
+  // Write the protobuf for the chunk header. The proto for events_chunk is
+  // filled manually instead of using the generated C++ stub. Rationale: the
+  // stub require this class to perform the buffer write operations. Using this
+  // to write our own proto would make this code extremely hard to reason about.
+  uint8_t* header = reinterpret_cast<uint8_t*>(chunk_->proto_begin());
+
+  header =
+      WriteProtoVarInt(header, ChunkProto::kStreamIdFieldNumber, stream_id_);
+  header = WriteProtoVarInt(header, ChunkProto::kSeqIdInStreamFieldNumber,
+                            chunk_seq_id_);
+
+  header = WriteProtoVarInt(
+      header, ChunkProto::kFirstEventContinuesFromPrevChunkFieldNumber,
+      event_continues_from_prev_chunk ? 1 : 0);
+
+  // At this point we don't know yet whether the last event in the chunk is
+  // going to continue on the next chunk. For the moment we put a zero as a
+  // placeholder and remember its position in the chunk. The actual value will
+  // be written the next time we take a new chunk (above in this function).
+  header = WriteProtoVarInt(
+      header, ChunkProto::kLastEventContinuesOnNextChunkFieldNumber, 0);
+  continue_on_next_chunk_ptr_ = header - 1;
+  ++chunk_seq_id_;
+
+  // If the new chunk happened while writing an event (the event is spread over
+  // multiple chunks) write a new proto preamble for the new partial byte array.
+  if (event_continues_from_prev_chunk) {
+    header = WriteEventPreamble(event_, header);
+  }
+
+  return {header, chunk_->end()};
+}
+
+// TODO: Things to test:
+// events beginning and ending precisely on a chunk boundary.
+// events longer than 2 chunks.
+// presence of last event in a non-full chunk.
+// test a long message which nested submessages where the size of the nested
+// subm is in a previous chunk. (chunk retaining logic).
+// reserved size field num bytes != for main event in chunk vs nested messages.
+
+// TODO possible optimization: in the codegen for add_args_simple instead of
+// ::tracing::EventArgsSimple* inst = new ::tracing::EventArgsSimple();
+// use a placement new.
+
+}  // namespace v2
+}  // namespace trace_event
+}  // namespace base