tracing: Send bigger proto trace slices via mojo pipe in ConsumerHost

Previously, we were writing individual packet headers + payloads one by
one onto the mojo pipe. Internally, each write causes a memcpy into a
shmem buffer and a mojo notification to be sent to the Consumer to read
the data from the buffer.

This was probably not a big problem when a trace packet contained many
trace events, but now that each packet is only a single event (and
especially now that each COMPLETE event is split into separate
begin/end events, i.e. separate packets), this causes a lot of
overhead.

Instead, this patch modifies ConsumerHost to write a single slice per
ReadBuffer callback to the mojo pipe, which makes proto trace
collection significantly faster.

Bug: 1039726
Change-Id: I057c68ffafb99cbd2b41d6c6308391cc11cf1883
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/1991448
Auto-Submit: Eric Seckler <eseckler@chromium.org>
Commit-Queue: Sami Kyöstilä <skyostil@chromium.org>
Reviewed-by: Sami Kyöstilä <skyostil@chromium.org>
Cr-Commit-Position: refs/heads/master@{#730178}

services/tracing/perfetto/consumer_host.cc

@@ -84,7 +84,7 @@ class JsonStringOutputWriter
 
 class ConsumerHost::StreamWriter {
  public:
-  using Slices = std::vector<std::string>;
+  using Slice = std::string;
 
   static scoped_refptr<base::SequencedTaskRunner> CreateTaskRunner() {
     return base::CreateSequencedTaskRunner({base::ThreadPool(),
@@ -101,16 +101,15 @@ class ConsumerHost::StreamWriter {
         disconnect_callback_(std::move(disconnect_callback)),
         callback_task_runner_(callback_task_runner) {}
 
-  void WriteToStream(std::unique_ptr<Slices> slices, bool has_more) {
+  void WriteToStream(std::unique_ptr<Slice> slice, bool has_more) {
     DCHECK(stream_.is_valid());
-    for (const auto& slice : *slices) {
-      uint32_t write_position = 0;
 
-      while (write_position < slice.size()) {
-        uint32_t write_bytes = slice.size() - write_position;
+    uint32_t write_position = 0;
+    while (write_position < slice->size()) {
+      uint32_t write_bytes = slice->size() - write_position;
 
       MojoResult result =
-            stream_->WriteData(slice.data() + write_position, &write_bytes,
+          stream_->WriteData(slice->data() + write_position, &write_bytes,
                              MOJO_WRITE_DATA_FLAG_NONE);
 
       if (result == MOJO_RESULT_OK) {
@@ -130,7 +129,6 @@ class ConsumerHost::StreamWriter {
         return;
       }
     }
-    }
 
     if (!has_more && !read_buffers_callback_.is_null()) {
       callback_task_runner_->PostTask(FROM_HERE,
@@ -465,11 +463,10 @@ void ConsumerHost::TracingSession::ExportJson() {
 
 void ConsumerHost::TracingSession::OnJSONTraceData(std::string json,
                                                    bool has_more) {
-  auto slices = std::make_unique<StreamWriter::Slices>();
-  slices->push_back(std::string());
-  slices->back().swap(json);
+  auto slice = std::make_unique<StreamWriter::Slice>();
+  slice->swap(json);
   read_buffers_stream_writer_.Post(FROM_HERE, &StreamWriter::WriteToStream,
-                                   std::move(slices), has_more);
+                                   std::move(slice), has_more);
 
   if (!has_more) {
     read_buffers_stream_writer_.Reset();
@@ -480,7 +477,7 @@ void ConsumerHost::TracingSession::OnTraceData(
     std::vector<perfetto::TracePacket> packets,
     bool has_more) {
   DCHECK_CALLED_ON_VALID_SEQUENCE(sequence_checker_);
-  if (trace_processor_) {
+
   // Calculate space needed for trace chunk. Each packet has a preamble and
   // payload size.
   size_t max_size = packets.size() * perfetto::TracePacket::kMaxPreambleBytes;
@@ -488,6 +485,7 @@ void ConsumerHost::TracingSession::OnTraceData(
     max_size += packet.size();
   }
 
+  if (trace_processor_) {
     // Copy packets into a trace file chunk.
     size_t position = 0;
     std::unique_ptr<uint8_t[]> data(new uint8_t[max_size]);
@@ -518,19 +516,21 @@ void ConsumerHost::TracingSession::OnTraceData(
     return;
   }
 
-  auto copy = std::make_unique<StreamWriter::Slices>();
+  // Copy packets into a trace slice.
+  auto chunk = std::make_unique<StreamWriter::Slice>();
+  chunk->reserve(max_size);
   for (auto& packet : packets) {
     char* data;
     size_t size;
     std::tie(data, size) = packet.GetProtoPreamble();
-    copy->emplace_back(data, size);
+    chunk->append(data, size);
     auto& slices = packet.slices();
     for (auto& slice : slices) {
-      copy->emplace_back(static_cast<const char*>(slice.start), slice.size);
+      chunk->append(static_cast<const char*>(slice.start), slice.size);
     }
   }
   read_buffers_stream_writer_.Post(FROM_HERE, &StreamWriter::WriteToStream,
-                                   std::move(copy), has_more);
+                                   std::move(chunk), has_more);
   if (!has_more) {
     read_buffers_stream_writer_.Reset();
   }