OSDN Git Service

partially fix [3306150] HTML5 video with H/W acceleration blackout
authorMathias Agopian <mathias@google.com>
Sat, 15 Jan 2011 01:37:42 +0000 (17:37 -0800)
committerMathias Agopian <mathias@google.com>
Sat, 15 Jan 2011 02:19:02 +0000 (18:19 -0800)
We used to guarantee that a layer in SurfaceFlinger would never be
destroyed before all references (to its ISurface) on the client
side would be released. At some point, this guarantee got
relaxed to allow to free gralloc resources sooner. This last
change was incorrect, because:
- in implementations with reference-counting the gralloc resources
wouldn't be released anyways, until all the mapping were gone
- in implementations without ref counting, the client side
would most likely crash or do something bad
- it also caused the SharedBufferStack slot to be reallocated
to another surface, which could be problematic if the client
continued to use the surface after the window manager destroyed it.

So, we essentially reinstate the guarantee that layers won't be
destroyed until after all references to their ISurface are
released.

NOTE: This doesn't entirely fix 3306150 because there is another
problem there where the Browser continues to use a surface after it
has been destroyed.

Change-Id: I305c830dd722b30a6d53cbf3a9c714fd3cf7eb06

services/surfaceflinger/SurfaceFlinger.cpp
services/surfaceflinger/SurfaceFlinger.h

index c982ea5..b411000 100644 (file)
@@ -1099,8 +1099,12 @@ status_t SurfaceFlinger::removeLayer_l(const sp<LayerBase>& layerBase)
 
 status_t SurfaceFlinger::purgatorizeLayer_l(const sp<LayerBase>& layerBase)
 {
-    // remove the layer from the main list (through a transaction).
+    // First add the layer to the purgatory list, which makes sure it won't
+    // go away, then remove it from the main list (through a transaction).
     ssize_t err = removeLayer_l(layerBase);
+    if (err >= 0) {
+        mLayerPurgatory.add(layerBase);
+    }
 
     layerBase->onRemoved();
 
@@ -1349,6 +1353,19 @@ status_t SurfaceFlinger::destroySurface(const sp<LayerBaseClient>& layer)
              * to use the purgatory.
              */
             status_t err = flinger->removeLayer_l(l);
+            if (err == NAME_NOT_FOUND) {
+                // The surface wasn't in the current list, which means it was
+                // removed already, which means it is in the purgatory,
+                // and need to be removed from there.
+                // This needs to happen from the main thread since its dtor
+                // must run from there (b/c of OpenGL ES). Additionally, we
+                // can't really acquire our internal lock from
+                // destroySurface() -- see postMessage() below.
+                ssize_t idx = flinger->mLayerPurgatory.remove(l);
+                LOGE_IF(idx < 0,
+                        "layer=%p is not in the purgatory list", l.get());
+            }
+
             LOGE_IF(err<0 && err != NAME_NOT_FOUND,
                     "error removing layer=%p (%s)", l.get(), strerror(-err));
             return true;
index 48642d4..549e282 100644 (file)
@@ -359,6 +359,7 @@ private:
     volatile    int32_t                 mTransactionFlags;
     volatile    int32_t                 mTransactionCount;
                 Condition               mTransactionCV;
+                SortedVector< sp<LayerBase> > mLayerPurgatory;
                 bool                    mResizeTransationPending;
 
                 // protected by mStateLock (but we could use another lock)