improve cache behavior of gralloc allocations on qsd8k

make sure to use cached buffers only when frequent cpu access is needed.
also, always flush buffers after they are allocated and cleared.
diff --git a/libgralloc-qsd8k/gralloc.cpp b/libgralloc-qsd8k/gralloc.cpp
index 38f3885..f1d8f55 100644
--- a/libgralloc-qsd8k/gralloc.cpp
+++ b/libgralloc-qsd8k/gralloc.cpp
@@ -279,9 +279,16 @@
                 err = -ENOMEM;
             } else {
                 struct pmem_region sub = { offset, size };
-                
+                int openFlags = O_RDWR | O_SYNC;
+                uint32_t uread = usage & GRALLOC_USAGE_SW_READ_MASK;
+                uint32_t uwrite = usage & GRALLOC_USAGE_SW_WRITE_MASK;
+                if (uread == GRALLOC_USAGE_SW_READ_OFTEN ||
+                    uwrite == GRALLOC_USAGE_SW_WRITE_OFTEN) {
+                    openFlags &= ~O_SYNC;
+                }
+
                 // now create the "sub-heap"
-                fd = open("/dev/pmem", O_RDWR, 0);
+                fd = open("/dev/pmem", openFlags, 0);
                 err = fd < 0 ? fd : 0;
                 
                 // and connect to it
@@ -297,8 +304,11 @@
                     close(fd);
                     sAllocator.deallocate(offset);
                     fd = -1;
+                } else {
+                    memset((char*)base + offset, 0, size);
+                    // clean and invalidate the new allocation
+                    cacheflush(intptr_t(base) + offset, size, 0);
                 }
-                memset((char*)base + offset, 0, size);
                 //LOGD_IF(!err, "allocating pmem size=%d, offset=%d", size, offset);
             }
         } else {