Wrap large IndexedDB values into Blobs before writing to LevelDB.

third_party/WebKit/Source/modules/indexeddb/IDBValueWrapping.cpp

Index: third_party/WebKit/Source/modules/indexeddb/IDBValueWrapping.cpp
diff --git a/third_party/WebKit/Source/modules/indexeddb/IDBValueWrapping.cpp b/third_party/WebKit/Source/modules/indexeddb/IDBValueWrapping.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..a4c744e0a6cf05c706247a3668bd2a37b371a628
--- /dev/null
+++ b/third_party/WebKit/Source/modules/indexeddb/IDBValueWrapping.cpp
@@ -0,0 +1,251 @@
+// Copyright 2017 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 "modules/indexeddb/IDBValueWrapping.h"
+
+#include <utility>
+
+#include "bindings/core/v8/ScriptValue.h"
+#include "bindings/core/v8/serialization/SerializationTag.h"
+#include "bindings/modules/v8/V8BindingForModules.h"
+#include "core/fileapi/Blob.h"
+#include "modules/indexeddb/IDBRequest.h"
+#include "modules/indexeddb/IDBValue.h"
+#include "platform/blob/BlobData.h"
+#include "platform/wtf/text/WTFString.h"
+
+namespace blink {
+
+namespace {
+
+// V8 values are stored on disk by IndexedDB using the format implemented in

[jsbell, 2017/05/22 21:54:20]

Docs should probably mention that the blob is pushed into the value's blob_info
vector, and thus if the unwrapped value has N blob references, the wrapped value
has N+1

[pwnall, 2017/05/25 13:27:12]

I tried to add the details to the documentation in IDBValueWrapper.h. Is this
clearer?

+// SerializedScriptValue (SSV). The wrapping detection logic in
+// IDBValueUnwrapper::IsWrapped() must be able to distinguish between SSV byte
+// sequences produced and byte sequences expressing the fact that an IDBValue
+// has been wrapped and requires post-processing.
+//
+// The detection logic takes advantage of the highly regular structure around
+// SerializedScriptValue. A version 17 byte sequence always starts with the
+// following four bytes:
+//
+// 1) 0xFF - version tag
+// 2) 0x11 - Blink wrapper version, 17
+// 3) 0xFF - version tag
+// 4) 0x0D - V8 serialization version, currently 13, doesn't matter
+//
+// It follows that SSV will never produce byte sequences starting with 0xFF,
+// 0x11, and any value except for 0xFF. If the SSV format changes, the version
+// will have to be bumped.

[jbroman, 2017/05/23 20:13:11]

Yes, I believe this is true. It saddens me somewhat to have so many clever
tricks used here, basically arising from the fact that we didn't originally have
a magic number.

It might be simpler to just grab and reserve a tag that was never used (0xFE or
something), and just use that to mark the value. Anything current should begin
with 0xFF, and it seems to me that this requires slightly less of a history
lesson. Am I forgetting something?

[pwnall, 2017/05/25 13:27:12]

I'm not sure I understand your proposal.

Did you mean I could use a 0xFE prefix instead of the current 0xFF 0x11
[anything but 0xFF]? If I'd go that route, I'd need a list of forbidden
follow-up bytes, similar to SSV's list of burned versions, in order to make sure
I don't conflict with byte-swapped v0 SSVs.

Using 0xFE later on (0xFF 0x11 0xFE or 0xFF 0x11 0xFF 0x0D 0xFE) seems to result
in longer prefixes. I don't see any benefit in return for the longer prefix
length. Specifically, SSV should fail on the prefix I'm proposing, and this
would be used outside the SSV framework, so there's no benefit in getting a tag
that would integrate well with SSV.

Can you please explain in more detail? Thanks for putting up with me!

[jbroman, 2017/05/25 14:50:58]

Yeah, okay. You're right that it's slightly more complicated than I'd thought
(byte-swapping, yaaay). This sequence sgtm. (I just wish we didn't have to have
to many of these tricks around, but it's hard to avoid.)

+
+// The SSV format version whose encoding hole is (ab)used for wrapping.
+const static uint8_t kRequiresProcessingSSVPseudoVersion = 17;
+
+// Identifies IndexedDB values that were wrapped in Blobs. The wrapper has the

[jsbell, 2017/05/22 21:54:20]

Can you explicitly write out the wrapped blob format header? i.e.

1) 0xFF - kVersionTag
2) 0x11 - kRequiresProcessingSSVPseudoVersion
3) 0x01 - kBlobWrappedValue

Followed by...

[pwnall, 2017/05/25 13:27:12]

Done.
Ah, good call. That was really unclear. Thank you so much for catching the
confusion!

+// following format:
+//
+// 1) varint - Blob size
+// 2) varint - the offset of the SSV-wrapping Blob in the IDBValue list of Blobs
+//             (should always be the last Blob)
+const static uint8_t kBlobWrappedValue = 1;
+
+}  // namespace
+
+IDBValueWrapper::IDBValueWrapper(
+    v8::Isolate* isolate,
+    v8::Local<v8::Value> value,
+    SerializedScriptValue::SerializeOptions::WasmSerializationPolicy
+        wasm_policy,
+    ExceptionState& exception_state) {
+  SerializedScriptValue::SerializeOptions options;
+  options.blob_info = &blob_info_;
+  options.for_storage = SerializedScriptValue::kForStorage;
+  options.wasm_policy = wasm_policy;
+
+  serialized_value_ = SerializedScriptValue::Serialize(isolate, value, options,
+                                                       exception_state);
+
+#if DCHECK_IS_ON()
+  if (exception_state.HadException())
+    had_exception_ = true;
+#endif  // DCHECK_IS_ON()
+}
+
+void IDBValueWrapper::Clone(ScriptState* script_state, ScriptValue* clone) {
+#if DCHECK_IS_ON()
+  DCHECK(!had_exception_) << __FUNCTION__
+                          << " called on wrapper with serialization exception";
+  DCHECK(!wrap_called_) << "Clone() called after WrapIfBiggerThan()";
+#endif  // DCHECK_IS_ON()
+  *clone = DeserializeScriptValue(script_state, serialized_value_.Get(),
+                                  &blob_info_);
+}
+
+void IDBValueWrapper::WriteVarint(unsigned value, Vector<char>& output) {
+  // Writes an unsigned integer as a base-128 varint.
+  // The number is written, 7 bits at a time, from the least significant to
+  // the most significant 7 bits. Each byte, except the last, has the MSB set.
+  // See also https://developers.google.com/protocol-buffers/docs/encoding
+  do {
+    output.push_back((value & 0x7F) | 0x80);
+    value >>= 7;
+  } while (value);
+  output.back() &= 0x7F;
+}
+
+bool IDBValueWrapper::WrapIfBiggerThan(unsigned max_bytes) {
+#if DCHECK_IS_ON()
+  DCHECK(!had_exception_) << __FUNCTION__
+                          << " called on wrapper with serialization exception";
+  DCHECK(!wrap_called_) << __FUNCTION__ << " called twice on the same wrapper";
+  wrap_called_ = true;
+#endif  // DCHECK_IS_ON()
+
+  serialized_value_->ToWireBytes(wire_bytes_);
+  if (wire_bytes_.size() <= max_bytes)
+    return false;
+
+  // TODO(pwnall): The MIME type should probably be an atomic string.
+  String mime_type(kWrapMimeType);
+  // TODO(crbug.com/721516): Use WebBlobRegistry::CreateBuilder instead of
+  //                         Blob::Create to avoid a buffer copy.
+  Blob* wrapper =
+      Blob::Create(reinterpret_cast<unsigned char*>(wire_bytes_.data()),
+                   wire_bytes_.size(), mime_type);
+
+  wrapper_handle_ = wrapper->GetBlobDataHandle();
+  blob_info_.emplace_back(wrapper_handle_->Uuid(), wrapper_handle_->GetType(),
+                          wrapper->size());
+
+  wire_bytes_.clear();
+
+  wire_bytes_.push_back(kVersionTag);
+  wire_bytes_.push_back(kRequiresProcessingSSVPseudoVersion);
+  wire_bytes_.push_back(kBlobWrappedValue);
+  IDBValueWrapper::WriteVarint(wrapper->size(), wire_bytes_);
+  IDBValueWrapper::WriteVarint(serialized_value_->BlobDataHandles().size(),
+                               wire_bytes_);
+  return true;
+}
+
+void IDBValueWrapper::ExtractBlobDataHandles(
+    Vector<RefPtr<BlobDataHandle>>* blob_data_handles) {
+  for (const auto& kvp : serialized_value_->BlobDataHandles())
+    blob_data_handles->push_back(kvp.value);
+  if (wrapper_handle_)
+    blob_data_handles->push_back(std::move(wrapper_handle_));
+}
+
+RefPtr<SharedBuffer> IDBValueWrapper::ExtractWireBytes() {
+#if DCHECK_IS_ON()
+  DCHECK(!had_exception_) << __FUNCTION__
+                          << " called on wrapper with serialization exception";
+#endif  // DCHECK_IS_ON()
+
+  return SharedBuffer::AdoptVector(wire_bytes_);
+}
+
+IDBValueUnwrapper::IDBValueUnwrapper() {
+  Reset();
+}
+
+bool IDBValueUnwrapper::IsWrapped(IDBValue* value) {

[jsbell, 2017/05/22 21:54:20]

Verify that blob_info size is >= 1 ?

[pwnall, 2017/05/25 13:27:12]

Done.
I added a check IDBValueUnwrapper::Parse(), because it already does a
size-related check. This is only relevant in case of disk corruption, so I hope
not failing fast is acceptable.

+  DCHECK(value);
+
+  uint8_t header[3];
+  if (!value->data_ || value->data_->size() < sizeof(header))
+    return false;
+
+  value->data_->GetPartAsBytes(header, static_cast<size_t>(0), sizeof(header));
+  return header[0] == kVersionTag &&
+         header[1] == kRequiresProcessingSSVPseudoVersion &&
+         header[2] == kBlobWrappedValue;
+}
+
+bool IDBValueUnwrapper::IsWrapped(const Vector<RefPtr<IDBValue>>& values) {
+  for (const auto& value : values) {
+    if (IsWrapped(value.Get()))
+      return true;
+  }
+  return false;
+}
+
+RefPtr<IDBValue> IDBValueUnwrapper::CreateUnwrapped(
+    IDBValue* wrapped_value,
+    RefPtr<SharedBuffer>&& wrapper_blob_content) {
+  DCHECK(wrapped_value);
+  DCHECK(wrapped_value->data_);
+  DCHECK_GT(wrapped_value->blob_info_->size(), 0U);
+  DCHECK_EQ(wrapped_value->blob_info_->size(),
+            wrapped_value->blob_data_->size());
+
+  // Create an IDBValue with the same blob information, minus the last blob.
+  unsigned blob_count = wrapped_value->BlobInfo()->size() - 1;
+  std::unique_ptr<Vector<RefPtr<BlobDataHandle>>> blob_data =
+      WTF::MakeUnique<Vector<RefPtr<BlobDataHandle>>>();
+  blob_data->ReserveCapacity(blob_count);
+  std::unique_ptr<Vector<WebBlobInfo>> blob_info =
+      WTF::MakeUnique<Vector<WebBlobInfo>>();
+  blob_info->ReserveCapacity(blob_count);
+
+  for (unsigned i = 0; i < blob_count; ++i) {
+    blob_data->push_back((*wrapped_value->blob_data_)[i]);
+    blob_info->push_back((*wrapped_value->blob_info_)[i]);
+  }
+
+  return AdoptRef(new IDBValue(std::move(wrapper_blob_content),
+                               std::move(blob_data), std::move(blob_info)));
+}
+
+bool IDBValueUnwrapper::Parse(IDBValue* value) {
+  // Fast path that avoids unnecessary dynamic allocations.
+  if (!IDBValueUnwrapper::IsWrapped(value))
+    return false;
+
+  const uint8_t* data = reinterpret_cast<const uint8_t*>(value->data_->Data());
+  end_ = data + value->data_->size();
+  current_ = data + 3;
+
+  if (!ReadVarint(blob_size_))
+    return Reset();
+
+  unsigned blob_offset;
+  if (!ReadVarint(blob_offset))
+    return Reset();
+  if (blob_offset != value->blob_data_->size() - 1)
+    return Reset();
+
+  blob_handle_ = value->blob_data_->back();
+  if (blob_handle_->size() != blob_size_)
+    return Reset();
+
+  return true;
+}
+
+RefPtr<BlobDataHandle> IDBValueUnwrapper::WrapperBlobHandle() {
+  DCHECK(blob_handle_);
+
+  return std::move(blob_handle_);
+}
+
+bool IDBValueUnwrapper::ReadVarint(unsigned& value) {
+  value = 0;
+  unsigned shift = 0;
+  bool has_another_byte;
+  do {
+    if (current_ >= end_)
+      return false;
+
+    if (shift >= sizeof(unsigned) * 8)
+      return false;
+    uint8_t byte = *current_;
+    ++current_;
+    value |= static_cast<unsigned>(byte & 0x7F) << shift;
+    shift += 7;
+
+    has_another_byte = byte & 0x80;
+  } while (has_another_byte);
+  return true;
+}
+
+}  // namespace blink