diff --git a/drivers/gpu/nvgpu/gk20a/fifo_gk20a.c b/drivers/gpu/nvgpu/gk20a/fifo_gk20a.c
index 298f59332..289c1c6c5 100644
--- a/drivers/gpu/nvgpu/gk20a/fifo_gk20a.c
+++ b/drivers/gpu/nvgpu/gk20a/fifo_gk20a.c
@@ -623,8 +623,6 @@ static int init_runlist(struct gk20a *g, struct fifo_gk20a *f)
 	u32 i;
 	size_t runlist_size;
 	u32 active_engine_id, pbdma_id, engine_id;
-	u32 flags = nvgpu_is_enabled(g, NVGPU_MM_USE_PHYSICAL_SG) ?
-		NVGPU_DMA_FORCE_CONTIGUOUS : 0U;
 	int err = 0;
 
 	nvgpu_log_fn(g, " ");
@@ -664,9 +662,10 @@ static int init_runlist(struct gk20a *g, struct fifo_gk20a *f)
 				f->num_runlist_entries, runlist_size);
 
 		for (i = 0; i < MAX_RUNLIST_BUFFERS; i++) {
-			err = nvgpu_dma_alloc_flags_sys(g, flags,
-							    runlist_size,
-							    &runlist->mem[i]);
+			err = nvgpu_dma_alloc_flags_sys(g,
+					NVGPU_DMA_PHYSICALLY_ADDRESSED,
+					runlist_size,
+					&runlist->mem[i]);
 			if (err != 0) {
 				nvgpu_err(g, "memory allocation failed");
 				goto clean_up_runlist;
@@ -967,10 +966,8 @@ int gk20a_init_fifo_setup_sw(struct gk20a *g)
 				   (size_t)f->num_channels,
 				   &f->userd);
 	} else {
-		u32 flags = nvgpu_is_enabled(g, NVGPU_MM_USE_PHYSICAL_SG) ?
-	                NVGPU_DMA_FORCE_CONTIGUOUS : 0U;
-
-		err = nvgpu_dma_alloc_flags_sys(g, flags,
+		err = nvgpu_dma_alloc_flags_sys(g,
+				NVGPU_DMA_PHYSICALLY_ADDRESSED,
 				(size_t)f->userd_entry_size *
 				(size_t)f->num_channels, &f->userd);
 	}
diff --git a/drivers/gpu/nvgpu/include/nvgpu/dma.h b/drivers/gpu/nvgpu/include/nvgpu/dma.h
index d7ece3949..28f2159d5 100644
--- a/drivers/gpu/nvgpu/include/nvgpu/dma.h
+++ b/drivers/gpu/nvgpu/include/nvgpu/dma.h
@@ -54,6 +54,15 @@ struct nvgpu_mem;
  */
 #define NVGPU_DMA_READ_ONLY		BIT32(2)
 
+/*
+ * Buffer is physically addressed from the GPU.
+ * If device is not iommuable, or nvlink is enabled, don't allow building
+ * the buffer from individual pages, but require a physically contiguous
+ * block.
+ */
+#define NVGPU_DMA_PHYSICALLY_ADDRESSED	BIT32(3)
+
+
 /**
  * nvgpu_iommuable - Check if GPU is behind IOMMU
  *
diff --git a/drivers/gpu/nvgpu/os/linux/linux-dma.c b/drivers/gpu/nvgpu/os/linux/linux-dma.c
index 348f2cdc7..6dcc40662 100644
--- a/drivers/gpu/nvgpu/os/linux/linux-dma.c
+++ b/drivers/gpu/nvgpu/os/linux/linux-dma.c
@@ -74,8 +74,9 @@ static char *nvgpu_dma_flags_to_str(struct gk20a *g, unsigned long flags)
 		}							\
 	} while (false)
 
-	APPEND_FLAG(NVGPU_DMA_NO_KERNEL_MAPPING, "NO_KERNEL_MAPPING ");
-	APPEND_FLAG(NVGPU_DMA_FORCE_CONTIGUOUS,  "FORCE_CONTIGUOUS ");
+	APPEND_FLAG(NVGPU_DMA_NO_KERNEL_MAPPING,    "NO_KERNEL_MAPPING ");
+	APPEND_FLAG(NVGPU_DMA_FORCE_CONTIGUOUS,     "FORCE_CONTIGUOUS ");
+	APPEND_FLAG(NVGPU_DMA_PHYSICALLY_ADDRESSED, "PHYSICALLY_ADDRESSED");
 #undef APPEND_FLAG
 
 	return buf;
@@ -187,6 +188,12 @@ int nvgpu_dma_alloc_flags_sys(struct gk20a *g, unsigned long flags,
 		WARN_ON(1);
 	}
 
+	if ((flags & NVGPU_DMA_PHYSICALLY_ADDRESSED) &&
+	    (!nvgpu_iommuable(g) ||
+	     nvgpu_is_enabled(g, NVGPU_MM_USE_PHYSICAL_SG))) {
+		flags |= NVGPU_DMA_FORCE_CONTIGUOUS;
+	}
+
 	/*
 	 * WAR for IO coherent chips: the DMA API does not seem to generate
 	 * mappings that work correctly. Unclear why - Bug ID: 2040115.
diff --git a/drivers/gpu/nvgpu/os/linux/ltc.c b/drivers/gpu/nvgpu/os/linux/ltc.c
index baeb20b2e..9b51632df 100644
--- a/drivers/gpu/nvgpu/os/linux/ltc.c
+++ b/drivers/gpu/nvgpu/os/linux/ltc.c
@@ -31,7 +31,6 @@ int nvgpu_ltc_alloc_cbc(struct gk20a *g, size_t compbit_backing_size,
 			bool vidmem_alloc)
 {
 	struct gr_gk20a *gr = &g->gr;
-	unsigned long flags = 0;
 
 	if (nvgpu_mem_is_valid(&gr->compbit_store.mem))
 		return 0;
@@ -49,11 +48,8 @@ int nvgpu_ltc_alloc_cbc(struct gk20a *g, size_t compbit_backing_size,
 					 compbit_backing_size,
 					 &gr->compbit_store.mem);
 	} else {
-		if (!nvgpu_iommuable(g))
-			flags = NVGPU_DMA_FORCE_CONTIGUOUS;
-
 		return nvgpu_dma_alloc_flags_sys(g,
-					 flags,
+					 NVGPU_DMA_PHYSICALLY_ADDRESSED,
 					 compbit_backing_size,
 					 &gr->compbit_store.mem);
 	}