diff --git a/drivers/rknpu/include/rknpu_drv.h b/drivers/rknpu/include/rknpu_drv.h
index d13a990d23b2..9ec5444f8a6d 100644
--- a/drivers/rknpu/include/rknpu_drv.h
+++ b/drivers/rknpu/include/rknpu_drv.h
@@ -30,10 +30,10 @@
 
 #define DRIVER_NAME "rknpu"
 #define DRIVER_DESC "RKNPU driver"
-#define DRIVER_DATE "20230625"
+#define DRIVER_DATE "20230629"
 #define DRIVER_MAJOR 0
-#define DRIVER_MINOR 8
-#define DRIVER_PATCHLEVEL 9
+#define DRIVER_MINOR 9
+#define DRIVER_PATCHLEVEL 0
 
 #define LOG_TAG "RKNPU"
 
diff --git a/drivers/rknpu/include/rknpu_ioctl.h b/drivers/rknpu/include/rknpu_ioctl.h
index 6294ac59274d..3b0b857108ce 100644
--- a/drivers/rknpu/include/rknpu_ioctl.h
+++ b/drivers/rknpu/include/rknpu_ioctl.h
@@ -73,8 +73,8 @@ enum e_rknpu_mem_type {
 	RKNPU_MEM_ZEROING = 1 << 5,
 	/* allocate secure buffer */
 	RKNPU_MEM_SECURE = 1 << 6,
-	/* allocate from non-dma32 zone */
-	RKNPU_MEM_NON_DMA32 = 1 << 7,
+	/* allocate from dma32 zone */
+	RKNPU_MEM_DMA32 = 1 << 7,
 	/* request SRAM */
 	RKNPU_MEM_TRY_ALLOC_SRAM = 1 << 8,
 	/* request NBUF */
@@ -82,7 +82,7 @@ enum e_rknpu_mem_type {
 	RKNPU_MEM_MASK = RKNPU_MEM_NON_CONTIGUOUS | RKNPU_MEM_CACHEABLE |
 			 RKNPU_MEM_WRITE_COMBINE | RKNPU_MEM_KERNEL_MAPPING |
 			 RKNPU_MEM_IOMMU | RKNPU_MEM_ZEROING |
-			 RKNPU_MEM_SECURE | RKNPU_MEM_NON_DMA32 |
+			 RKNPU_MEM_SECURE | RKNPU_MEM_DMA32 |
 			 RKNPU_MEM_TRY_ALLOC_SRAM | RKNPU_MEM_TRY_ALLOC_NBUF
 };
 
diff --git a/drivers/rknpu/rknpu_drv.c b/drivers/rknpu/rknpu_drv.c
index 584bdc8dbec8..ad0e22eabaa8 100644
--- a/drivers/rknpu/rknpu_drv.c
+++ b/drivers/rknpu/rknpu_drv.c
@@ -224,7 +224,6 @@ int rknpu_power_get(struct rknpu_device *rknpu_dev)
 {
 	int ret = 0;
 
-	cancel_delayed_work(&rknpu_dev->power_off_work);
 	mutex_lock(&rknpu_dev->power_lock);
 	if (atomic_inc_return(&rknpu_dev->power_refcount) == 1)
 		ret = rknpu_power_on(rknpu_dev);
@@ -247,6 +246,9 @@ int rknpu_power_put(struct rknpu_device *rknpu_dev)
 
 static int rknpu_power_put_delay(struct rknpu_device *rknpu_dev)
 {
+	if (rknpu_dev->power_put_delay == 0)
+		return rknpu_power_put(rknpu_dev);
+
 	mutex_lock(&rknpu_dev->power_lock);
 	if (atomic_read(&rknpu_dev->power_refcount) == 1)
 		queue_delayed_work(
@@ -255,6 +257,7 @@ static int rknpu_power_put_delay(struct rknpu_device *rknpu_dev)
 	else
 		atomic_dec_if_positive(&rknpu_dev->power_refcount);
 	mutex_unlock(&rknpu_dev->power_lock);
+
 	return 0;
 }
 
diff --git a/drivers/rknpu/rknpu_gem.c b/drivers/rknpu/rknpu_gem.c
index 3bb84372c5c2..6c08734e0fbc 100644
--- a/drivers/rknpu/rknpu_gem.c
+++ b/drivers/rknpu/rknpu_gem.c
@@ -67,6 +67,7 @@ static int rknpu_gem_get_pages(struct rknpu_gem_object *rknpu_obj)
 			      rknpu_obj->size);
 		goto free_sgt;
 	}
+	iommu_flush_iotlb_all(iommu_get_domain_for_dev(drm->dev));
 
 	if (rknpu_obj->flags & RKNPU_MEM_KERNEL_MAPPING) {
 		rknpu_obj->cookie = vmap(rknpu_obj->pages, rknpu_obj->num_pages,
@@ -181,7 +182,9 @@ static int rknpu_gem_alloc_buf(struct rknpu_gem_object *rknpu_obj)
 	if (rknpu_obj->flags & RKNPU_MEM_ZEROING)
 		gfp_mask |= __GFP_ZERO;
 
-	if (!(rknpu_obj->flags & RKNPU_MEM_NON_DMA32)) {
+	if (!rknpu_dev->iommu_en ||
+	    rknpu_dev->config->dma_mask <= DMA_BIT_MASK(32) ||
+	    (rknpu_obj->flags & RKNPU_MEM_DMA32)) {
 		gfp_mask &= ~__GFP_HIGHMEM;
 		gfp_mask |= __GFP_DMA32;
 	}
@@ -360,6 +363,7 @@ static const struct drm_gem_object_funcs rknpu_gem_object_funcs = {
 static struct rknpu_gem_object *rknpu_gem_init(struct drm_device *drm,
 					       unsigned long size)
 {
+	struct rknpu_device *rknpu_dev = drm->dev_private;
 	struct rknpu_gem_object *rknpu_obj = NULL;
 	struct drm_gem_object *obj = NULL;
 	gfp_t gfp_mask;
@@ -388,7 +392,9 @@ static struct rknpu_gem_object *rknpu_gem_init(struct drm_device *drm,
 	if (rknpu_obj->flags & RKNPU_MEM_ZEROING)
 		gfp_mask |= __GFP_ZERO;
 
-	if (!(rknpu_obj->flags & RKNPU_MEM_NON_DMA32)) {
+	if (!rknpu_dev->iommu_en ||
+	    rknpu_dev->config->dma_mask <= DMA_BIT_MASK(32) ||
+	    (rknpu_obj->flags & RKNPU_MEM_DMA32)) {
 		gfp_mask &= ~__GFP_HIGHMEM;
 		gfp_mask |= __GFP_DMA32;
 	}