[BlobAsync] Patch 4: Browser Classes & Logic.

storage/browser/blob/blob_async_transport_strategy.cc

Index: storage/browser/blob/blob_async_transport_strategy.cc
diff --git a/storage/browser/blob/blob_async_transport_strategy.cc b/storage/browser/blob/blob_async_transport_strategy.cc
new file mode 100644
index 0000000000000000000000000000000000000000..bf1a535f0890f27ae9e4b83b22175bb1f2287dc4
--- /dev/null
+++ b/storage/browser/blob/blob_async_transport_strategy.cc
@@ -0,0 +1,308 @@
+// Copyright 2015 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 <algorithm>
+
+#include "storage/browser/blob/blob_async_transport_strategy.h"
+#include "storage/common/blob_storage/blob_storage_constants.h"
+
+namespace storage {
+namespace {
+bool IsBytes(DataElement::Type type) {
+  return type == DataElement::TYPE_BYTES ||
+         type == DataElement::TYPE_BYTES_DESCRIPTION;
+}
+}  // namespace
+
+// This class handles the logic of how transported memory is going to be
+// represented as storage in the browser.  The main idea is that all the memory
+// is now packed into file chunks, and the browser items will just reference
+// the file with offsets and sizes.
+class FileStorageStrategy
+    : public BlobAsyncTransportStrategy::BlobSegmentVisitor<uint64_t> {
+ public:
+  typedef uint64_t SizeType;
+
+  FileStorageStrategy(
+      std::vector<BlobAsyncTransportStrategy::RendererMemoryItemRequest>*
+          requests,
+      BlobDataBuilder* builder)
+      : requests(requests), builder(builder), current_storage_item(0) {}
+
+  ~FileStorageStrategy() override {}
+
+  void VisitBytesSegment(size_t element_index,
+                         SizeType element_offset,
+                         size_t segment_index,
+                         SizeType segment_offset,
+                         SizeType size) override {
+    BlobAsyncTransportStrategy::RendererMemoryItemRequest request;
+    request.browser_item_index = current_storage_item;
+    request.browser_item_offset = 0;
+    request.message.request_number = requests->size();
+    request.message.transport_strategy = IPCBlobItemRequestStrategy::FILE;
+    request.message.renderer_item_index = element_index;
+    request.message.renderer_item_offset = element_offset;
+    request.message.size = size;
+    request.message.handle_index = segment_index;
+    request.message.handle_offset = segment_offset;
+
+    requests->push_back(request);
+    builder->AppendFutureFile(segment_offset, size);

[michaeln, 2015/11/21 00:59:45]

AppendFuture and PopulateFuture, nice choice for names.

+    current_storage_item++;
+  }
+
+  void VisitNonBytesSegment(const DataElement& element,
+                            size_t element_index) override {
+    builder->AppendIPCDataElement(element);
+    current_storage_item++;
+  }
+
+  void Done() override {}
+
+  std::vector<BlobAsyncTransportStrategy::RendererMemoryItemRequest>* requests;
+  BlobDataBuilder* builder;
+
+  size_t current_storage_item;
+};
+
+// This class handles the logic of storing memory that is transported as
+// consolidated shared memory.  The main hurdle is to re-separate memory blocks
+// that cross file or blob boundaries.

[michaeln, 2015/11/21 00:59:45]

when are blob boundaries crossed?

[dmurph, 2015/11/23 20:07:02]

Blob reference boundaries.

+class SharedMemoryStorageStrategy
+    : public BlobAsyncTransportStrategy::BlobSegmentVisitor<size_t> {
+ public:
+  typedef size_t SizeType;
+
+  SharedMemoryStorageStrategy(
+      SizeType max_segment_size,
+      std::vector<BlobAsyncTransportStrategy::RendererMemoryItemRequest>*
+          requests,
+      BlobDataBuilder* builder)
+      : requests(requests),
+        max_segment_size(max_segment_size),
+        storage_element_offset(0),
+        builder(builder),
+        current_item_size(0),
+        current_storage_item(0) {}
+  ~SharedMemoryStorageStrategy() override {}
+
+  void VisitBytesSegment(size_t element_index,
+                         SizeType element_offset,
+                         size_t segment_index,
+                         SizeType segment_offset,
+                         SizeType size) override {
+    if (storage_element_offset + size > max_segment_size) {
+      builder->AppendFutureData(current_item_size);
+      current_storage_item++;
+      current_item_size = 0;
+      storage_element_offset = 0;
+    }
+    BlobAsyncTransportStrategy::RendererMemoryItemRequest request;
+    request.browser_item_index = current_storage_item;
+    request.browser_item_offset = storage_element_offset;
+    request.message.request_number = requests->size();
+    request.message.transport_strategy =
+        IPCBlobItemRequestStrategy::SHARED_MEMORY;
+    request.message.renderer_item_index = element_index;
+    request.message.renderer_item_offset = element_offset;
+    request.message.size = size;
+    request.message.handle_index = segment_index;
+    request.message.handle_offset = segment_offset;
+
+    requests->push_back(request);
+    storage_element_offset += size;
+    current_item_size += size;
+  };
+
+  void VisitNonBytesSegment(const DataElement& element,
+                            size_t element_index) override {
+    builder->AppendFutureData(current_item_size);

[michaeln, 2015/11/21 00:59:45]

what if current_item_size is 0?

[dmurph, 2015/11/23 20:07:02]

Fixed, and added test.

+    current_storage_item++;
+    builder->AppendIPCDataElement(element);
+    current_storage_item++;
+    storage_element_offset = 0;
+    current_item_size = 0;
+  }
+
+  void Done() override {
+    if (current_item_size != 0) {
+      builder->AppendFutureData(current_item_size);
+    }
+  }
+
+  std::vector<BlobAsyncTransportStrategy::RendererMemoryItemRequest>* requests;
+
+  SizeType max_segment_size;
+  SizeType storage_element_offset;
+  BlobDataBuilder* builder;
+  SizeType current_item_size;
+  size_t current_storage_item;

[michaeln, 2015/11/21 00:59:45]

naminga this as current_item_index might help

+};
+
+BlobAsyncTransportStrategy::RendererMemoryItemRequest::
+    RendererMemoryItemRequest()
+    : browser_item_index(0), browser_item_offset(0), received(false) {}
+
+BlobAsyncTransportStrategy::BlobAsyncTransportStrategy()
+    : error_(BlobAsyncTransportStrategy::ERROR_NONE), total_bytes_size_(0) {}
+
+BlobAsyncTransportStrategy::~BlobAsyncTransportStrategy() {}
+
+// if total_blob_size > kMaxBlobSize (say 400MB)
+//   Request all data in files
+//   (Segment all of the existing data into
+//   file blocks, of <= kMaxFileSize)
+// else if total_blob_size > kMaxIPCSize (this is 150KB)
+//   Request all data in shared memory
+//   (Segment all of the existing data into
+//   shared memory blocks, of <= kMaxSharedMemorySize)
+// else
+//   Request all data to be sent over IPC
+void BlobAsyncTransportStrategy::Initialize(

[michaeln, 2015/11/21 00:59:45]

this method reads a lot nicer now, thnx!

+    size_t max_ipc_memory_size,
+    size_t max_shared_memory_size,
+    uint64_t max_file_size,
+    uint64_t disk_space_left,
+    size_t memory_available,
+    const std::string& uuid,
+    const std::vector<DataElement>& blob_item_infos) {
+  file_handle_sizes_.clear();
+  shared_memory_handle_sizes_.clear();
+  requests_.clear();
+  builder_.reset(new BlobDataBuilder(uuid));
+  error_ = BlobAsyncTransportStrategy::ERROR_NONE;

[kinuko, 2015/11/20 15:19:42]

I think I've once asked this but is an instance of this class going to be used
multiple times?  The current usage doesn't look so, and if that's not the case
resetting everything in initialize feels a bit wasting.

[dmurph, 2015/11/20 22:10:05]

Sounds good, I'll add DCHECKs then.

+
+  size_t memory_items = 0;
+  total_bytes_size_ = 0;
+  for (const auto& info : blob_item_infos) {
+    if (!IsBytes(info.type())) {
+      continue;
+    }
+    total_bytes_size_ += info.length();
+    ++memory_items;
+  }
+
+  // See if we have enough memory
+  if (total_bytes_size_ >
+      disk_space_left + static_cast<uint64_t>(memory_available)) {
+    error_ = BlobAsyncTransportStrategy::ERROR_TOO_LARGE;
+    return;
+  }
+
+  // If we're more than the available memory, then we're going straight to disk.
+  if (total_bytes_size_ > memory_available) {
+    // First, handle the case where we go to disk.

[michaeln, 2015/11/21 00:59:45]

redundant comment?

[dmurph, 2015/11/23 20:07:02]

Done.

+    if (total_bytes_size_ < disk_space_left) {
+      ComputeHandleSizes(total_bytes_size_, max_file_size, &file_handle_sizes_);
+      FileStorageStrategy strategy(&requests_, builder_.get());
+      ForEachWithSegment(blob_item_infos, max_file_size, &strategy);
+      return;
+    }
+    error_ = BlobAsyncTransportStrategy::ERROR_TOO_LARGE;
+    return;

[kinuko, 2015/11/20 15:19:42]

nit: prefer early return (handle error case first) in general

[dmurph, 2015/11/20 22:10:05]

Done.

+  }
+
+  if (total_bytes_size_ > max_ipc_memory_size) {
+    CHECK_LE(total_bytes_size_, std::numeric_limits<size_t>::max());
+    ComputeHandleSizes(static_cast<size_t>(total_bytes_size_),
+                       max_shared_memory_size, &shared_memory_handle_sizes_);
+    SharedMemoryStorageStrategy strategy(max_shared_memory_size, &requests_,
+                                         builder_.get());
+    ForEachWithSegment(blob_item_infos, max_shared_memory_size, &strategy);
+    return;
+  }

[kinuko, 2015/11/20 15:19:42]

nit: let's add one empty line here

[dmurph, 2015/11/20 22:10:05]

Done.

+  // Since they can all fit in IPC memory, we don't need to segment anything,
+  // and just request them straight in IPC.
+  size_t items_length = blob_item_infos.size();
+  for (size_t i = 0; i < items_length; i++) {
+    const auto& info = blob_item_infos.at(i);
+    if (!IsBytes(info.type())) {
+      builder_->AppendIPCDataElement(info);
+      continue;
+    }
+    BlobAsyncTransportStrategy::RendererMemoryItemRequest request;
+    request.browser_item_index = i;
+    request.browser_item_offset = 0;
+    request.message.request_number = requests_.size();
+    request.message.transport_strategy = IPCBlobItemRequestStrategy::IPC;
+    request.message.renderer_item_index = i;
+    request.message.renderer_item_offset = 0;
+    request.message.size = info.length();
+    requests_.push_back(request);
+    builder_->AppendFutureData(info.length());
+  }
+}
+
+// Splits each |element| into one or more |segments| of a max_size, invokes the
+// strategy to determine the request to make for each |segment| produced. A
+// |segment| can also span multiple |elements|.
+/* static */
+template <typename SizeType>
+void BlobAsyncTransportStrategy::ForEachWithSegment(
+    const std::vector<DataElement>& elements,
+    SizeType max_segment_size,
+    BlobSegmentVisitor<SizeType>* visitor) {
+  DCHECK_GT(max_segment_size, 0ull);
+  size_t segment_index = 0;
+  SizeType segment_offset = 0;
+  size_t elements_length = elements.size();
+  for (size_t element_index = 0; element_index < elements_length;
+       ++element_index) {
+    const auto& element = elements.at(element_index);
+    DataElement::Type type = element.type();
+    if (!IsBytes(type)) {
+      visitor->VisitNonBytesSegment(element, element_index);
+      continue;
+    }
+    SizeType element_memory_left = element.length();
+    SizeType element_offset = 0;
+    while (element_memory_left > 0) {
+      if (segment_offset == max_segment_size) {
+        ++segment_index;
+        segment_offset = 0;
+      }
+      SizeType memory_writing =
+          std::min(max_segment_size - segment_offset, element_memory_left);
+      visitor->VisitBytesSegment(element_index, element_offset, segment_index,
+                                 segment_offset, memory_writing);
+      element_memory_left -= memory_writing;
+      segment_offset += memory_writing;
+      element_offset += memory_writing;
+    }
+  }
+  visitor->Done();
+}
+
+/* static */
+bool BlobAsyncTransportStrategy::ShouldBeShortcut(
+    const std::vector<DataElement>& elements,
+    size_t memory_available) {
+  size_t shortcut_bytes = 0;
+  for (const auto& element : elements) {
+    DataElement::Type type = element.type();
+    if (type == DataElement::TYPE_BYTES_DESCRIPTION) {
+      return false;
+    }
+    if (type == DataElement::TYPE_BYTES) {
+      shortcut_bytes += element.length();
+    }
+  }
+  return shortcut_bytes <= memory_available;
+}
+
+/* static */
+template <typename SizeType>
+void BlobAsyncTransportStrategy::ComputeHandleSizes(
+    SizeType total_memory_size,
+    SizeType max_segment_size,
+    std::vector<SizeType>* segment_sizes) {

[kinuko, 2015/11/20 15:19:42]

nit: probably resize the segment_sizes first before inserting for
micro-optimization

[dmurph, 2015/11/20 22:10:05]

Done.

[dmurph, 2015/11/20 22:10:05]

Done.

+  segment_sizes->insert(segment_sizes->begin(),
+                        total_memory_size / max_segment_size, max_segment_size);
+  if (total_memory_size % max_segment_size > 0) {
+    segment_sizes->push_back(total_memory_size % max_segment_size);
+  }
+}
+
+}  // namespace storage