tracing v2: Introduce TraceBufferWriter

components/tracing/core/trace_buffer_writer.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 "components/tracing/core/trace_buffer_writer.h"
+
+#include "base/compiler_specific.h"
+#include "base/logging.h"
+#include "components/tracing/core/proto_utils.h"
+
+namespace tracing {
+namespace v2 {
+
+namespace {
+
+// TODO(primiano) remove this in next CLs. This should just be taken from the
+// C++ class autogenerated from events_chunk.proto (crbug.com/608721).
+struct ChunkProto {
+  enum : uint32_t {
+    kWriterIdFieldNumber = 1,
+    kSeqIdInStreamFieldNumber = 2,
+    kEventsFieldNumber = 3,
+    kFirstEventContinuesFromPrevChunkFieldNumber = 4,
+    kLastEventContinuesOnNextChunkFieldNumber = 5
+  };
+};
+
+const size_t kEventPreambleSize = 1 + proto::kMessageLengthFieldSize;
+
+}  // namespace
+
+TraceBufferWriter::TraceBufferWriter(TraceRingBuffer* trace_ring_buffer,
+                                     uint32_t writer_id)
+    : trace_ring_buffer_(trace_ring_buffer),
+      writer_id_(writer_id),
+      chunk_seq_id_(0),
+      chunk_(nullptr),
+      event_data_start_in_current_chunk_(nullptr),
+      stream_writer_(this) {
+  event_.Reset(&stream_writer_);
+}
+
+TraceBufferWriter::~TraceBufferWriter() {}
+
+void TraceBufferWriter::FinalizeCurrentEvent() {
+  if (UNLIKELY(!chunk_))
+    return;
+
+  // Finalize the last event added. This ensures that it and all its nested
+  // fields are committed to the ring buffer and sealed. No further changes to
+  // the chunks's memory can be made from the |event_| after this point.
+  event_.Finalize();
+
+  // In the unlikely event that the last event did wrap over one or more chunks,
+  // is is now time to return those chunks (all but the active one) back.
+  TraceRingBuffer::Chunk* retained_chunk = chunk_->next_in_owner_list();
+  if (UNLIKELY(retained_chunk)) {
+    while (retained_chunk) {
+      TraceRingBuffer::Chunk* next = retained_chunk->next_in_owner_list();
+      retained_chunk->set_next_in_owner_list(nullptr);
+      trace_ring_buffer_->ReturnChunk(retained_chunk);
+      retained_chunk = next;
+    }
+    chunk_->set_next_in_owner_list(nullptr);
+  }
+}
+
+TraceEventHandle TraceBufferWriter::AddEvent() {
+  FinalizeCurrentEvent();
+
+  // In order to start a new event at least kMessageLengthFieldSize + 1 bytes
+  // are required in the chunk to write the preamble and size of the event
+  // itself. We take a bit more room here, it doesn't make a lot of sense
+  // starting a partial event that will fragment immediately after.
+  // TODO(primiano): replace 16 with a more reasonable size, that is, the size
+  // of a simple trace event with no args.
+  if (stream_writer_.bytes_available() < 16)

[oystein (OOO til 10th of July), 2016/08/05 23:04:35]

nit: make this a named constant at least? Or is the idea that it'll be replaced
by a sizeof()?

[Primiano Tucci (use gerrit), 2016/08/08 19:36:23]

Made a constant and moved the todo up there. It's going to be an estimation
based on the proto and the fields that we'll use for common (no arg) events.

+    stream_writer_.Reset(AcquireNewChunk(false /* is_fragmenting_event */));
+
+  event_.Reset(&stream_writer_);
+  WriteEventPreambleForNewChunk(
+      stream_writer_.ReserveBytesUnsafe(kEventPreambleSize));
+  DCHECK_EQ(stream_writer_.write_ptr(), event_data_start_in_current_chunk_);
+  return TraceEventHandle(static_cast<::tracing::proto::Event*>(&event_));
+}
+
+// This is invoked by the ProtoZeroMessage write methods when reaching the
+// end of the current chunk during a write.
+ContiguousMemoryRange TraceBufferWriter::GetNewBuffer() {
+  return AcquireNewChunk(true /* is_fragmenting_event */);
+}
+
+void TraceBufferWriter::FinalizeCurrentChunk(bool is_fragmenting_event) {
+  DCHECK(!is_fragmenting_event || chunk_);
+  if (!chunk_)
+    return;
+  DCHECK_GE(stream_writer_.write_ptr(), chunk_->payload());
+  DCHECK_LE(stream_writer_.write_ptr(), chunk_->end() - 2);
+  size_t used_size = stream_writer_.write_ptr() - chunk_->payload();

[oystein (OOO til 10th of July), 2016/08/05 23:04:32]

nit: size_t and then casting to uint32 after using it, feels weird to me. Maybe
DCHECK on write_ptr() - chunk_->payload() being within uint32_t range, then use
uint32_t instead of size_t?

[Primiano Tucci (use gerrit), 2016/08/08 19:36:23]

Done.

+
+  if (is_fragmenting_event) {
+    uint8_t* write_ptr = stream_writer_.write_ptr();
+    proto::StaticAssertSingleBytePreamble<
+      ChunkProto::kLastEventContinuesOnNextChunkFieldNumber>();
+    *write_ptr++ = static_cast<uint8_t>(proto::MakeTagVarInt(
+      ChunkProto::kLastEventContinuesOnNextChunkFieldNumber));
+    *write_ptr = 1;

[oystein (OOO til 10th of July), 2016/08/05 23:04:35]

What's the 1?

[Primiano Tucci (use gerrit), 2016/08/08 19:36:23]

1 == true -> the event does continue on the next chunk. 
The statement above is just the preamble that says "this field will tell whether
the event continues or not".
(I know, shame that a byte has to be wasted to encode one bit)

+    used_size += 2;

[oystein (OOO til 10th of July), 2016/08/05 23:04:32]

Can you just store an original_write_ptr instead of used_size, then calculate
used_size below this conditional?

[Primiano Tucci (use gerrit), 2016/08/08 19:36:23]

Done.

+  }
+
+  chunk_->set_used_size(static_cast<uint32_t>(used_size));
+}
+
+// There are paths that lead to AcquireNewChunk():
+// When |is_fragmenting_event| = false:
+//   AddEvent() is called and there isn't enough room in the current chunk to
+//   start a new event (or we don't have a chunk yet).
+// When |is_fragmenting_event| = true:
+//   The client is writing an event, a ProtoZeroMessage::Append* method hits
+//   the boundary of the chunk and requests a new one via GetNewBuffer().
+ContiguousMemoryRange TraceBufferWriter::AcquireNewChunk(
+    bool is_fragmenting_event) {
+  FinalizeCurrentChunk(is_fragmenting_event);
+  TraceRingBuffer::Chunk* new_chunk = trace_ring_buffer_->TakeChunk(writer_id_);
+  if (is_fragmenting_event) {
+    // Backfill the size field of the event with the partial size accumulated
+    // so far in the old chunk. WriteEventPreambleForNewChunk() will take care
+    // of resetting the |size_field| of the event to the new chunk.
+    DCHECK_GE(event_data_start_in_current_chunk_, chunk_->payload());
+    DCHECK_LE(event_data_start_in_current_chunk_,
+              chunk_->end() - proto::kMessageLengthFieldSize);
+    const uint32_t event_partial_size = static_cast<uint32_t>(
+        stream_writer_.write_ptr() - event_data_start_in_current_chunk_);
+    proto::WriteRedundantVarIntU32<proto::kMessageLengthFieldSize>(
+        event_partial_size, event_.size_field().begin);
+    event_.inc_size_already_written(event_partial_size);
+
+    // If this is a continuation of an event, this writer needs to retain the
+    // old chunk. The client might still be able to write to it. This is to deal
+    // with the case of a nested message which is started in one chunk and
+    // ends in another one. The finalization needs to write-back the size field
+    // in the old chunk.
+    new_chunk->set_next_in_owner_list(chunk_);
+  } else if (chunk_) {
+    // Otherwise, if this is a new event, the previous chunk can be returned.
+    trace_ring_buffer_->ReturnChunk(chunk_);
+  }
+  chunk_ = new_chunk;
+
+  // Write the protobuf for the chunk header. The generated C++ stub for
+  // events_chunk.proto cannot be used here because that would re-enter this
+  // class and make this code extremely hard to reason about.
+  uint8_t* chunk_proto = new_chunk->payload();
+
+  proto::StaticAssertSingleBytePreamble<ChunkProto::kWriterIdFieldNumber>();
+  *chunk_proto++ = static_cast<uint8_t>(
+      proto::MakeTagVarInt(ChunkProto::kWriterIdFieldNumber));
+  chunk_proto = proto::WriteVarIntU32(writer_id_, chunk_proto);
+
+  proto::StaticAssertSingleBytePreamble<
+      ChunkProto::kSeqIdInStreamFieldNumber>();
+  *chunk_proto++ = static_cast<uint8_t>(
+      proto::MakeTagVarInt(ChunkProto::kSeqIdInStreamFieldNumber));
+  chunk_proto = proto::WriteVarIntU32(chunk_seq_id_, chunk_proto);
+
+  if (is_fragmenting_event) {
+    proto::StaticAssertSingleBytePreamble<
+        ChunkProto::kFirstEventContinuesFromPrevChunkFieldNumber>();
+    *chunk_proto++ = static_cast<uint8_t>(proto::MakeTagVarInt(
+        ChunkProto::kFirstEventContinuesFromPrevChunkFieldNumber));
+    *chunk_proto++ = 1;

[oystein (OOO til 10th of July), 2016/08/05 23:04:32]

What does the 1 represent?

[Primiano Tucci (use gerrit), 2016/08/08 19:36:23]

as above, the value of the bool.

+  }
+
+  ++chunk_seq_id_;
+
+  // If the new chunk was requested while writing an event (the event spans
+  // across chunks) write a new preamble for the partial event in the new chunk.
+  if (is_fragmenting_event)
+    chunk_proto = WriteEventPreambleForNewChunk(chunk_proto);
+
+  // We reserve 2 bytes from the end, so that FinalizeCurrentChunk() can use
+  // them to write the |last_event_continues_on_next_chunk| field.
+  return {chunk_proto, new_chunk->end() - 2};
+}
+
+// Writes the one-byte preamble for the start of either a new or a partial
+// event and reserves kMessageLengthFieldSize bytes for its length. Also
+// keeps size-field the bookeeping up to date. Returns the pointer in the chunk

[oystein (OOO til 10th of July), 2016/08/05 23:04:33]

nit: bookkeeping

[Primiano Tucci (use gerrit), 2016/08/08 19:36:23]

Done.

+// past the event preamble, where the event proto should be written.
+uint8_t* TraceBufferWriter::WriteEventPreambleForNewChunk(uint8_t* begin) {
+  // The caller must have ensured to have enough room in the chunk. The event
+  // preamble itself cannot be fragmented.
+  uint8_t* const end = begin + kEventPreambleSize;
+  proto::StaticAssertSingleBytePreamble<ChunkProto::kEventsFieldNumber>();
+  *begin++ = static_cast<uint8_t>(
+      proto::MakeTagLengthDelimited(ChunkProto::kEventsFieldNumber));
+  ContiguousMemoryRange range = {begin, end};
+  event_.set_size_field(range);
+  event_data_start_in_current_chunk_ = end;
+  return end;
+}
+
+void TraceBufferWriter::Flush() {
+  FinalizeCurrentEvent();
+  FinalizeCurrentChunk(false /* is_fragmenting_event */);
+  trace_ring_buffer_->ReturnChunk(chunk_);
+  chunk_ = nullptr;
+}
+
+}  // namespace v2
+}  // namespace tracing