Ganesh: images upload to GPU performance fix (skip copying encoded data)

third_party/WebKit/Source/platform/graphics/ImageFrameGenerator.cpp

Index: third_party/WebKit/Source/platform/graphics/ImageFrameGenerator.cpp
diff --git a/third_party/WebKit/Source/platform/graphics/ImageFrameGenerator.cpp b/third_party/WebKit/Source/platform/graphics/ImageFrameGenerator.cpp
index 0d29bce0f18e6463f7fcbfd94cabe702c6455231..165100c9eccca3c44aebdf7121997800cc870ac9 100644
--- a/third_party/WebKit/Source/platform/graphics/ImageFrameGenerator.cpp
+++ b/third_party/WebKit/Source/platform/graphics/ImageFrameGenerator.cpp
@@ -27,6 +27,7 @@
 
 #include "platform/graphics/ImageFrameGenerator.h"
 
+#include "SkData.h"
 #include "platform/SharedBuffer.h"
 #include "platform/TraceEvent.h"
 #include "platform/graphics/ImageDecodingStore.h"
@@ -101,6 +102,7 @@ static bool updateYUVComponentSizes(ImageDecoder* decoder, SkISize componentSize
 
 ImageFrameGenerator::ImageFrameGenerator(const SkISize& fullSize, PassRefPtr<SharedBuffer> data, bool allDataReceived, bool isMultiFrame)
     : m_fullSize(fullSize)
+    , m_data(adoptRef(new ThreadSafeDataTransport()))

[aleksandar.stojiljkovic, 2015/12/04 15:36:59]

haraken@
The problem wasn't in fast alloc - it was in adoptRef

previous code had m_data(new ThreadSafeDataTransport()) (without adoptRef)

Monitored the test you asked:

1) with    adoptRef, with or without fast alloc - no leak
2) without adoptRef, with or without fast alloc - leak

Noel, thanks.

     , m_isMultiFrame(isMultiFrame)
     , m_decodeFailedAndEmpty(false)
     , m_decodeCount(0)
@@ -116,15 +118,56 @@ ImageFrameGenerator::~ImageFrameGenerator()
 
 void ImageFrameGenerator::setData(PassRefPtr<SharedBuffer> data, bool allDataReceived)
 {
-    m_data.setData(data.get(), allDataReceived);
+    m_data->setData(data.get(), allDataReceived);
 }
 
-void ImageFrameGenerator::copyData(RefPtr<SharedBuffer>* data, bool* allDataReceived)
+static void sharedSkDataReleaseCallback(const void* address, void* context)
 {
+    ThreadSafeDataTransport* dataTransport = static_cast<ThreadSafeDataTransport*>(context);
+#if ENABLE(ASSERT)
+    ASSERT(dataTransport);
     SharedBuffer* buffer = 0;
-    m_data.data(&buffer, allDataReceived);
-    if (buffer)
-        *data = buffer->copy();
+    bool allDataReceived = false;
+    dataTransport->data(&buffer, &allDataReceived);
+    ASSERT(allDataReceived && buffer && buffer->data() == address);
+#endif
+    // Deref m_data now.
+    dataTransport->deref();
+}
+
+SkData* ImageFrameGenerator::refEncodedData() const
+{
+    // Every SkData instance and SharedBuffer need to hold reference to buffer->data().
+    // Both SkData and m_data->m_readBuffer are having separate reference counting
+    // implementations.
+    //
+    // SkData's SkData::NewWithProc is designed with similar use case in mind,
+    // unmapping wrapped shmem data once all SkData instances are disposed.
+    //
+    // In this case, sharedSkDataReleaseCallback would get called when SkData is disposed.
+    // In sharedSkDataReleaseCallback, we deref ThreadSafeDataTransport (and ImageFrameGenerator
+    // might still hold a reference to ThreadSafeDataTransport).
+    //
+    // Multithreading concerns:
+    // ThreadSafeDataTransport disposal needs to be thread safe - SkData could get disposed
+    // in any thread, and SkData would still need to hold the reference to SharedBuffer when
+    // ImageFrameGenerator is deleted. This is why m_data (ThreadSafeRefCounted) is passed to
+    // client - to hold the reference to SharedBuffer (accompanying returned SkData).
+    //
+    // A note about allDataReceived:
+    // SkData is returned only when full image (encoded) data is downloaded. This is important
+    // since DeferredImageDecoder::setData is called only once with allDataReceived set to true,
+    // and after that m_data->m_readBuffer.data() is not changed. See also RELEASE_ASSERT used in
+    // ThreadSafeDataTransport::data().
+    SharedBuffer* buffer = 0;
+    bool allDataReceived = false;
+    m_data->data(&buffer, &allDataReceived);
+    if (!allDataReceived)
+        return nullptr;
+
+    // While SkData is holding reference to underlying data, prevent disposing m_data and its content.
+    m_data->ref();
+    return SkData::NewWithProc(buffer->data(), buffer->size(), sharedSkDataReleaseCallback, m_data.get());
 }
 
 bool ImageFrameGenerator::decodeAndScale(const SkImageInfo& info, size_t index, void* pixels, size_t rowBytes)
@@ -184,7 +227,7 @@ bool ImageFrameGenerator::decodeToYUV(SkISize componentSizes[3], void* planes[3]
 
     SharedBuffer* data = 0;
     bool allDataReceived = false;
-    m_data.data(&data, &allDataReceived);
+    m_data->data(&data, &allDataReceived);
 
     // FIXME: YUV decoding does not currently support progressive decoding.
     ASSERT(allDataReceived);
@@ -290,7 +333,7 @@ bool ImageFrameGenerator::decode(size_t index, ImageDecoder** decoder, SkBitmap*
     SharedBuffer* data = 0;
     bool allDataReceived = false;
     bool newDecoder = false;
-    m_data.data(&data, &allDataReceived);
+    m_data->data(&data, &allDataReceived);
 
     // Try to create an ImageDecoder if we are not given one.
     if (!*decoder) {
@@ -358,7 +401,7 @@ bool ImageFrameGenerator::getYUVComponentSizes(SkISize componentSizes[3])
 
     SharedBuffer* data = 0;
     bool allDataReceived = false;
-    m_data.data(&data, &allDataReceived);
+    m_data->data(&data, &allDataReceived);
 
     // FIXME: YUV decoding does not currently support progressive decoding.
     if (!allDataReceived)