diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
index e7662ba9abfb41233ecf82786e0ad042bb643d54..3060ce14071e2c35e9da4bbf72a80f6f187a0e46 100644
--- a/drivers/gpu/drm/radeon/radeon.h
+++ b/drivers/gpu/drm/radeon/radeon.h
@@ -624,6 +624,7 @@ struct radeon_device {
 	bool				gpu_lockup;
 	bool				shutdown;
 	bool				suspend;
+	bool				need_dma32;
 };
 
 int radeon_device_init(struct radeon_device *rdev,
diff --git a/drivers/gpu/drm/radeon/radeon_device.c b/drivers/gpu/drm/radeon/radeon_device.c
index 332911267ebe1e3dd2e0b865099ce1d82b8a53b9..27a5ac969953333714ee345387c7aa666c8ece04 100644
--- a/drivers/gpu/drm/radeon/radeon_device.c
+++ b/drivers/gpu/drm/radeon/radeon_device.c
@@ -450,6 +450,7 @@ int radeon_device_init(struct radeon_device *rdev,
 		       uint32_t flags)
 {
 	int r, ret;
+	int dma_bits;
 
 	DRM_INFO("radeon: Initializing kernel modesetting.\n");
 	rdev->shutdown = false;
@@ -492,8 +493,20 @@ int radeon_device_init(struct radeon_device *rdev,
 		return r;
 	}
 
-	/* Report DMA addressing limitation */
-	r = pci_set_dma_mask(rdev->pdev, DMA_BIT_MASK(32));
+	/* set DMA mask + need_dma32 flags.
+	 * PCIE - can handle 40-bits.
+	 * IGP - can handle 40-bits (in theory)
+	 * AGP - generally dma32 is safest
+	 * PCI - only dma32
+	 */
+	rdev->need_dma32 = false;
+	if (rdev->flags & RADEON_IS_AGP)
+		rdev->need_dma32 = true;
+	if (rdev->flags & RADEON_IS_PCI)
+		rdev->need_dma32 = true;
+
+	dma_bits = rdev->need_dma32 ? 32 : 40;
+	r = pci_set_dma_mask(rdev->pdev, DMA_BIT_MASK(dma_bits));
 	if (r) {
 		printk(KERN_WARNING "radeon: No suitable DMA available.\n");
 	}
diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index 1227a97f5169dc3f6c0f34da6bea4d89076e067f..4ca9aa9203d0fb74b533851e0228d4c5f7c67114 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -442,7 +442,8 @@ int radeon_ttm_init(struct radeon_device *rdev)
 	/* No others user of address space so set it to 0 */
 	r = ttm_bo_device_init(&rdev->mman.bdev,
 			       rdev->mman.mem_global_ref.object,
-			       &radeon_bo_driver, DRM_FILE_PAGE_OFFSET);
+			       &radeon_bo_driver, DRM_FILE_PAGE_OFFSET,
+			       rdev->need_dma32);
 	if (r) {
 		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
 		return r;
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index a753598a5e35004c3759160a244fd75fc1e49ecf..e55e7972c8979753d4b6062d813ea5924cba6588 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -224,6 +224,9 @@ static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
 	TTM_ASSERT_LOCKED(&bo->mutex);
 	bo->ttm = NULL;
 
+	if (bdev->need_dma32)
+		page_flags |= TTM_PAGE_FLAG_DMA32;
+
 	switch (bo->type) {
 	case ttm_bo_type_device:
 		if (zero_alloc)
@@ -1332,7 +1335,8 @@ EXPORT_SYMBOL(ttm_bo_device_release);
 
 int ttm_bo_device_init(struct ttm_bo_device *bdev,
 		       struct ttm_mem_global *mem_glob,
-		       struct ttm_bo_driver *driver, uint64_t file_page_offset)
+		       struct ttm_bo_driver *driver, uint64_t file_page_offset,
+		       bool need_dma32)
 {
 	int ret = -EINVAL;
 
@@ -1369,6 +1373,7 @@ int ttm_bo_device_init(struct ttm_bo_device *bdev,
 	INIT_LIST_HEAD(&bdev->ddestroy);
 	INIT_LIST_HEAD(&bdev->swap_lru);
 	bdev->dev_mapping = NULL;
+	bdev->need_dma32 = need_dma32;
 	ttm_mem_init_shrink(&bdev->shrink, ttm_bo_swapout);
 	ret = ttm_mem_register_shrink(mem_glob, &bdev->shrink);
 	if (unlikely(ret != 0)) {
diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
index 75dc8bd245920695d77e5f6fd92d5f9d3b150e5c..81ab81f030a3df28029b003f0d6be4346611a875 100644
--- a/drivers/gpu/drm/ttm/ttm_tt.c
+++ b/drivers/gpu/drm/ttm/ttm_tt.c
@@ -131,10 +131,15 @@ static void ttm_tt_free_page_directory(struct ttm_tt *ttm)
 
 static struct page *ttm_tt_alloc_page(unsigned page_flags)
 {
+	gfp_t gfp_flags = GFP_HIGHUSER;
+
 	if (page_flags & TTM_PAGE_FLAG_ZERO_ALLOC)
-		return alloc_page(GFP_HIGHUSER | __GFP_ZERO);
+		gfp_flags |= __GFP_ZERO;
+
+	if (page_flags & TTM_PAGE_FLAG_DMA32)
+		gfp_flags |= __GFP_DMA32;
 
-	return alloc_page(GFP_HIGHUSER);
+	return alloc_page(gfp_flags);
 }
 
 static void ttm_tt_free_user_pages(struct ttm_tt *ttm)
diff --git a/include/drm/ttm/ttm_bo_driver.h b/include/drm/ttm/ttm_bo_driver.h
index 62ed733c52a23898a893c51bf05eb80e17995bf4..ea83dd23a4d71b4ec4ca12d8eca0a675057fce2d 100644
--- a/include/drm/ttm/ttm_bo_driver.h
+++ b/include/drm/ttm/ttm_bo_driver.h
@@ -121,6 +121,7 @@ struct ttm_backend {
 #define TTM_PAGE_FLAG_SWAPPED         (1 << 4)
 #define TTM_PAGE_FLAG_PERSISTANT_SWAP (1 << 5)
 #define TTM_PAGE_FLAG_ZERO_ALLOC      (1 << 6)
+#define TTM_PAGE_FLAG_DMA32           (1 << 7)
 
 enum ttm_caching_state {
 	tt_uncached,
@@ -429,6 +430,8 @@ struct ttm_bo_device {
 	 */
 
 	struct delayed_work wq;
+
+	bool need_dma32;
 };
 
 /**
@@ -648,7 +651,7 @@ extern int ttm_bo_device_release(struct ttm_bo_device *bdev);
 extern int ttm_bo_device_init(struct ttm_bo_device *bdev,
 			      struct ttm_mem_global *mem_glob,
 			      struct ttm_bo_driver *driver,
-			      uint64_t file_page_offset);
+			      uint64_t file_page_offset, bool need_dma32);
 
 /**
  * ttm_bo_reserve: