修bug

src/utils/dma_alloc.cpp

@@ -164,14 +164,26 @@ DmaBufferPtr DmaAlloc(size_t alloc_size) {
         return DmaBufferPtr(reused, [](DmaBuffer* b) { Pool().Put(b); });
     }
 
-    // Try different dma_heap devices - prefer CMA for RGA compatibility (<4GB)
+    // Try different dma_heap devices.
+    // Notes:
+    // - Some RGA drivers cannot map buffers whose physical address is >4GB; prefer DMA32 heaps.
+    // - CMA heaps are also <4GB friendly but can be small/fragmented under multi-stream load.
     const char* heap_paths[] = {
+        // Prefer CMA first when available.
         "/dev/dma_heap/cma",
-        "/dev/dma_heap/cma-uncached", 
+        "/dev/dma_heap/cma-uncached",
         "/dev/dma_heap/linux,cma",
-        "/dev/dma_heap/reserved",
+
+        // Then prefer DMA32 heaps to keep allocations under 4GB.
+        "/dev/dma_heap/system-dma32",
+        "/dev/dma_heap/system-uncached-dma32",
+
+        // Generic system heaps (may allocate >4GB on high-mem systems).
         "/dev/dma_heap/system-uncached",
         "/dev/dma_heap/system",
+
+        // Last resort.
+        "/dev/dma_heap/reserved",
         nullptr
     };