[PATCH 0/4] drm/gem-dma: Support dedicated DMA device for allocation

[PATCH 0/4] drm/gem-dma: Support dedicated DMA device for allocation

[Chen-Yu Tsai]

Hi folks,

This series expands the "dedicated DMA device" support in DRM to the GEM
DMA helpers, and converts the MediaTek DRM driver to setting the DMA
device and dropping the custom GEM helpers that implemented this
function.

Various display drivers implement the "dedicated DMA device" support
with custom GEM helpers. These include Exynos, MediaTek, and Rockchip
to name a few. Allwinner does something entirely different, calling
of_dma_configure() on the virtual display device using the OF node of
the actual DMA device. Recently this causes a warning if IOMMUs are
involved.

This series intends to allow the core helpers to deal with it, and not
have every driver implement it in slightly different ways, duplicating
code.

Patch 1 adds dedicated DMA device support to drm_prime_pages_to_sg().
I believe this was missing from the original change that added dedicated
DMA devices for PRIME.

Patch 2 adds support for dedicated DMA device to the GEM DMA helpers
for GEM buffer allocation and mmap.

Patch 3 converts the MediaTek DRM driver to use the dedicated DMA device
support, and drop all the remaining custom GEM callbacks that deal with
it.

Patch 4 converts the Allwinner sun4i DRM driver to use the dedicated DMA
device support, instead of the of_dma_configure() hack it currently has.

The series should be merged through drm-misc-next so that other drivers
can take advantage of the change.

I also intend to try to convert the Exynos and Rockchip drivers, however
both also have options to set DMA_ATTR_NO_KERNEL_MAPPING when using
dma_alloc_attrs() to allocate memory for the buffers. I intend to
resurrect the DRM_MODE_DUMB_KERNEL_MAP work from Rob Herring [1]
to handle this. The Rockchip driver also has custom IOMMU attachment
that I'm still trying to understand.


Thanks
ChenYu


Chen-Yu Tsai (4):
  drm/prime: Limit scatter list size with dedicated DMA device
  drm/gem-dma: Support dedicated DMA device for allocation and mapping
  drm/mediatek: Set dedicated DMA device and drop custom GEM callbacks
  drm/sun4i: Use backend/mixer as dedicated DMA device

 drivers/gpu/drm/drm_gem_dma_helper.c   |  21 ++-
 drivers/gpu/drm/drm_prime.c            |   2 +-
 drivers/gpu/drm/mediatek/mtk_crtc.c    |   1 -
 drivers/gpu/drm/mediatek/mtk_drm_drv.c |  21 +--
 drivers/gpu/drm/mediatek/mtk_drm_drv.h |   1 -
 drivers/gpu/drm/mediatek/mtk_gem.c     | 231 -------------------------
 drivers/gpu/drm/mediatek/mtk_gem.h     |  17 --
 drivers/gpu/drm/sun4i/sun4i_backend.c  |  27 +--
 drivers/gpu/drm/sun4i/sun8i_mixer.c    |  27 +--
 9 files changed, 46 insertions(+), 302 deletions(-)
 delete mode 100644 drivers/gpu/drm/mediatek/mtk_gem.c
 delete mode 100644 drivers/gpu/drm/mediatek/mtk_gem.h

-- 
2.53.0.473.g4a7958ca14-goog

[PATCH 1/4] drm/prime: Limit scatter list size with dedicated DMA device

[Chen-Yu Tsai]

If a dedicated DMA device is specified for the DRM device, then the
scatter list size limit should pertain to the DMA device.

Use the dedicated DMA device, if given, to limit the scatter list size.
This only applies to drivers that have called drm_dev_set_dma_dev() and
are using drm_prime_pages_to_sg() either directly or through the SHMEM
helpers. At the time of this writing, the former case only includes the
Rockchip DRM driver, while the latter case includes the gud, udl, and
the tiny appletbdrm and gm12u320 drivers.

Signed-off-by: Chen-Yu Tsai <wenst@chromium.org>
---
 drivers/gpu/drm/drm_prime.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
index 51fdb06d3e9f..9b44c78cd77f 100644
--- a/drivers/gpu/drm/drm_prime.c
+++ b/drivers/gpu/drm/drm_prime.c
@@ -859,7 +859,7 @@ struct sg_table *drm_prime_pages_to_sg(struct drm_device *dev,
 		return ERR_PTR(-ENOMEM);
 
 	if (dev)
-		max_segment = dma_max_mapping_size(dev->dev);
+		max_segment = dma_max_mapping_size(drm_dev_dma_dev(dev));
 	if (max_segment == 0)
 		max_segment = UINT_MAX;
 	err = sg_alloc_table_from_pages_segment(sg, pages, nr_pages, 0,
-- 
2.53.0.473.g4a7958ca14-goog

Claude review: drm/prime: Limit scatter list size with dedicated DMA device

[Claude Code Review Bot]

Patch Review

A one-liner changing `dev->dev` to `drm_dev_dma_dev(dev)` in `drm_prime_pages_to_sg()`.

```c
-		max_segment = dma_max_mapping_size(dev->dev);
+		max_segment = dma_max_mapping_size(drm_dev_dma_dev(dev));
```

This is correct and consistent with how the rest of the PRIME code already uses the dedicated DMA device for imports (via `drm_gem_prime_import()`). The commit message properly documents which drivers are affected. No issues.

**Reviewed-by worthy**: Yes.

---
Generated by Claude Code Patch Reviewer

[PATCH 2/4] drm/gem-dma: Support dedicated DMA device for allocation and mapping

[Chen-Yu Tsai]

Support for a dedicated DMA device for prime imports was added in commit
143ec8d3f939 ("drm/prime: Support dedicated DMA device for dma-buf imports").
This allowed the DRM driver to provide a dedicated DMA device when its
own underlying device was not capable of DMA, for example when it is a
USB device (the original target) or a virtual device. The latter case is
common on embedded SoCs, on which the display pipeline is composed of
various fixed function blocks, and the DRM device is simply a made-up
device, an address space managing the routing between the blocks, or
whichever block the implementor thought made sense at the time. The
point is that the chosen device is often not the actual device doing
the DMA. Various drivers have used workarounds or reimplemented the
GEM DMA helpers to get the DMA addresses and IOMMUs to work correctly.

Add support for the dedicated DMA device to the GEM DMA helpers.

No existing driver currently uses the GEM DMA helpers and calls
drm_dev_set_dma_dev() to set a dedicated DMA device, so no existing
users should be affected.

Signed-off-by: Chen-Yu Tsai <wenst@chromium.org>
---
 drivers/gpu/drm/drm_gem_dma_helper.c | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/drm_gem_dma_helper.c b/drivers/gpu/drm/drm_gem_dma_helper.c
index ecb9746f4da8..70f83e464476 100644
--- a/drivers/gpu/drm/drm_gem_dma_helper.c
+++ b/drivers/gpu/drm/drm_gem_dma_helper.c
@@ -146,12 +146,13 @@ struct drm_gem_dma_object *drm_gem_dma_create(struct drm_device *drm,
 		return dma_obj;
 
 	if (dma_obj->map_noncoherent) {
-		dma_obj->vaddr = dma_alloc_noncoherent(drm->dev, size,
+		dma_obj->vaddr = dma_alloc_noncoherent(drm_dev_dma_dev(drm),
+						       size,
 						       &dma_obj->dma_addr,
 						       DMA_TO_DEVICE,
 						       GFP_KERNEL | __GFP_NOWARN);
 	} else {
-		dma_obj->vaddr = dma_alloc_wc(drm->dev, size,
+		dma_obj->vaddr = dma_alloc_wc(drm_dev_dma_dev(drm), size,
 					      &dma_obj->dma_addr,
 					      GFP_KERNEL | __GFP_NOWARN);
 	}
@@ -236,12 +237,14 @@ void drm_gem_dma_free(struct drm_gem_dma_object *dma_obj)
 		drm_prime_gem_destroy(gem_obj, dma_obj->sgt);
 	} else if (dma_obj->vaddr) {
 		if (dma_obj->map_noncoherent)
-			dma_free_noncoherent(gem_obj->dev->dev, dma_obj->base.size,
+			dma_free_noncoherent(drm_dev_dma_dev(gem_obj->dev),
+					     dma_obj->base.size,
 					     dma_obj->vaddr, dma_obj->dma_addr,
 					     DMA_TO_DEVICE);
 		else
-			dma_free_wc(gem_obj->dev->dev, dma_obj->base.size,
-				    dma_obj->vaddr, dma_obj->dma_addr);
+			dma_free_wc(drm_dev_dma_dev(gem_obj->dev),
+				    dma_obj->base.size, dma_obj->vaddr,
+				    dma_obj->dma_addr);
 	}
 
 	drm_gem_object_release(gem_obj);
@@ -432,7 +435,7 @@ struct sg_table *drm_gem_dma_get_sg_table(struct drm_gem_dma_object *dma_obj)
 	if (!sgt)
 		return ERR_PTR(-ENOMEM);
 
-	ret = dma_get_sgtable(obj->dev->dev, sgt, dma_obj->vaddr,
+	ret = dma_get_sgtable(drm_dev_dma_dev(obj->dev), sgt, dma_obj->vaddr,
 			      dma_obj->dma_addr, obj->size);
 	if (ret < 0)
 		goto out;
@@ -539,12 +542,12 @@ int drm_gem_dma_mmap(struct drm_gem_dma_object *dma_obj, struct vm_area_struct *
 	if (dma_obj->map_noncoherent) {
 		vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 
-		ret = dma_mmap_pages(dma_obj->base.dev->dev,
+		ret = dma_mmap_pages(drm_dev_dma_dev(dma_obj->base.dev),
 				     vma, vma->vm_end - vma->vm_start,
 				     virt_to_page(dma_obj->vaddr));
 	} else {
-		ret = dma_mmap_wc(dma_obj->base.dev->dev, vma, dma_obj->vaddr,
-				  dma_obj->dma_addr,
+		ret = dma_mmap_wc(drm_dev_dma_dev(dma_obj->base.dev), vma,
+				  dma_obj->vaddr, dma_obj->dma_addr,
 				  vma->vm_end - vma->vm_start);
 	}
 	if (ret)
-- 
2.53.0.473.g4a7958ca14-goog

Claude review: drm/gem-dma: Support dedicated DMA device for allocation and mapping

[Claude Code Review Bot]

Patch Review

Mechanically replaces all `drm->dev` / `gem_obj->dev->dev` with `drm_dev_dma_dev(drm)` / `drm_dev_dma_dev(gem_obj->dev)` in the GEM DMA helper for allocation, free, sgtable, and mmap paths.

The changes are consistent and complete -- all DMA API calls in `drm_gem_dma_helper.c` now go through `drm_dev_dma_dev()`:
- `dma_alloc_noncoherent()` / `dma_alloc_wc()` in `drm_gem_dma_create()`
- `dma_free_noncoherent()` / `dma_free_wc()` in `drm_gem_dma_free()`
- `dma_get_sgtable()` in `drm_gem_dma_get_sg_table()`
- `dma_mmap_pages()` / `dma_mmap_wc()` in `drm_gem_dma_mmap()`

The commit message correctly notes no existing driver currently combines GEM DMA helpers with `drm_dev_set_dma_dev()`, so this is safe to land without immediate behavioral changes.

**Reviewed-by worthy**: Yes.

---
Generated by Claude Code Patch Reviewer

[PATCH 3/4] drm/mediatek: Set dedicated DMA device and drop custom GEM callbacks

[Chen-Yu Tsai]

In commit 9b54a32c7c6a ("drm/mediatek: mtk_gem: Partial refactor and
use drm_gem_dma_object") the MediaTek DRM driver was refactored to use
drm_gem_dma_object, but custom callbacks were still needed to deal with
using the first device of the pipeline as the DMA device, instead of
the MMSYS device that the DRM driver binds to.

Turns out there is already partial support for dedicated DMA devices in
the DRM subsystem for PRIME imports. The preceding patches add support
for dedicated DMA devices to the GEM DMA helpers.

This allows us to just set the dedicated DMA device for the DRM device,
and drop all the custom GEM callbacks. Also drop the .dma_dev field
from the driver private data as it is no longer needed.

There are slight differences in the mmap helper: the VM_DONTDUMP and
VM_IO flags are no longer set. Both were lifted from drm_gem_mmap_obj().
VM_IO probably doesn't make sense since the buffer is allocated using
dma_alloc_attrs().

Signed-off-by: Chen-Yu Tsai <wenst@chromium.org>
---
 drivers/gpu/drm/mediatek/mtk_crtc.c    |   1 -
 drivers/gpu/drm/mediatek/mtk_drm_drv.c |  21 +--
 drivers/gpu/drm/mediatek/mtk_drm_drv.h |   1 -
 drivers/gpu/drm/mediatek/mtk_gem.c     | 231 -------------------------
 drivers/gpu/drm/mediatek/mtk_gem.h     |  17 --
 5 files changed, 3 insertions(+), 268 deletions(-)
 delete mode 100644 drivers/gpu/drm/mediatek/mtk_gem.c
 delete mode 100644 drivers/gpu/drm/mediatek/mtk_gem.h

diff --git a/drivers/gpu/drm/mediatek/mtk_crtc.c b/drivers/gpu/drm/mediatek/mtk_crtc.c
index 351d58c50b84..fcb16f3f7b23 100644
--- a/drivers/gpu/drm/mediatek/mtk_crtc.c
+++ b/drivers/gpu/drm/mediatek/mtk_crtc.c
@@ -23,7 +23,6 @@
 #include "mtk_crtc.h"
 #include "mtk_ddp_comp.h"
 #include "mtk_drm_drv.h"
-#include "mtk_gem.h"
 #include "mtk_plane.h"
 
 /*
diff --git a/drivers/gpu/drm/mediatek/mtk_drm_drv.c b/drivers/gpu/drm/mediatek/mtk_drm_drv.c
index a94c51a83261..6f6db2e1980e 100644
--- a/drivers/gpu/drm/mediatek/mtk_drm_drv.c
+++ b/drivers/gpu/drm/mediatek/mtk_drm_drv.c
@@ -19,6 +19,7 @@
 #include <drm/drm_fbdev_dma.h>
 #include <drm/drm_fourcc.h>
 #include <drm/drm_gem.h>
+#include <drm/drm_gem_dma_helper.h>
 #include <drm/drm_gem_framebuffer_helper.h>
 #include <drm/drm_ioctl.h>
 #include <drm/drm_of.h>
@@ -29,7 +30,6 @@
 #include "mtk_ddp_comp.h"
 #include "mtk_disp_drv.h"
 #include "mtk_drm_drv.h"
-#include "mtk_gem.h"
 
 #define DRIVER_NAME "mediatek"
 #define DRIVER_DESC "Mediatek SoC DRM"
@@ -565,8 +565,7 @@ static int mtk_drm_kms_init(struct drm_device *drm)
 		goto err_component_unbind;
 	}
 
-	for (i = 0; i < private->data->mmsys_dev_num; i++)
-		private->all_drm_private[i]->dma_dev = dma_dev;
+	drm_dev_set_dma_dev(drm, dma_dev);
 
 	/*
 	 * Configure the DMA segment size to make sure we get contiguous IOVA
@@ -600,26 +599,12 @@ static void mtk_drm_kms_deinit(struct drm_device *drm)
 
 DEFINE_DRM_GEM_FOPS(mtk_drm_fops);
 
-/*
- * We need to override this because the device used to import the memory is
- * not dev->dev, as drm_gem_prime_import() expects.
- */
-static struct drm_gem_object *mtk_gem_prime_import(struct drm_device *dev,
-						   struct dma_buf *dma_buf)
-{
-	struct mtk_drm_private *private = dev->dev_private;
-
-	return drm_gem_prime_import_dev(dev, dma_buf, private->dma_dev);
-}
-
 static const struct drm_driver mtk_drm_driver = {
 	.driver_features = DRIVER_MODESET | DRIVER_GEM | DRIVER_ATOMIC,
 
-	.dumb_create = mtk_gem_dumb_create,
+	DRM_GEM_DMA_DRIVER_OPS,
 	DRM_FBDEV_DMA_DRIVER_OPS,
 
-	.gem_prime_import = mtk_gem_prime_import,
-	.gem_prime_import_sg_table = mtk_gem_prime_import_sg_table,
 	.fops = &mtk_drm_fops,
 
 	.name = DRIVER_NAME,
diff --git a/drivers/gpu/drm/mediatek/mtk_drm_drv.h b/drivers/gpu/drm/mediatek/mtk_drm_drv.h
index 675cdc90a440..1fc3df4b5485 100644
--- a/drivers/gpu/drm/mediatek/mtk_drm_drv.h
+++ b/drivers/gpu/drm/mediatek/mtk_drm_drv.h
@@ -54,7 +54,6 @@ struct mtk_mmsys_driver_data {
 
 struct mtk_drm_private {
 	struct drm_device *drm;
-	struct device *dma_dev;
 	bool mtk_drm_bound;
 	bool drm_master;
 	struct device *dev;
diff --git a/drivers/gpu/drm/mediatek/mtk_gem.c b/drivers/gpu/drm/mediatek/mtk_gem.c
deleted file mode 100644
index f059a1452220..000000000000
--- a/drivers/gpu/drm/mediatek/mtk_gem.c
+++ /dev/null
@@ -1,231 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0-only
-/*
- * Copyright (c) 2015 MediaTek Inc.
- * Copyright (c) 2025 Collabora Ltd.
- *                    AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
- */
-
-#include <linux/dma-buf.h>
-#include <linux/vmalloc.h>
-
-#include <drm/drm.h>
-#include <drm/drm_device.h>
-#include <drm/drm_gem.h>
-#include <drm/drm_gem_dma_helper.h>
-#include <drm/drm_prime.h>
-#include <drm/drm_print.h>
-
-#include "mtk_drm_drv.h"
-#include "mtk_gem.h"
-
-static int mtk_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma);
-
-static void mtk_gem_free_object(struct drm_gem_object *obj)
-{
-	struct drm_gem_dma_object *dma_obj = to_drm_gem_dma_obj(obj);
-	struct mtk_drm_private *priv = obj->dev->dev_private;
-
-	if (dma_obj->sgt)
-		drm_prime_gem_destroy(obj, dma_obj->sgt);
-	else
-		dma_free_wc(priv->dma_dev, dma_obj->base.size,
-			    dma_obj->vaddr, dma_obj->dma_addr);
-
-	/* release file pointer to gem object. */
-	drm_gem_object_release(obj);
-
-	kfree(dma_obj);
-}
-
-/*
- * Allocate a sg_table for this GEM object.
- * Note: Both the table's contents, and the sg_table itself must be freed by
- *       the caller.
- * Returns a pointer to the newly allocated sg_table, or an ERR_PTR() error.
- */
-static struct sg_table *mtk_gem_prime_get_sg_table(struct drm_gem_object *obj)
-{
-	struct drm_gem_dma_object *dma_obj = to_drm_gem_dma_obj(obj);
-	struct mtk_drm_private *priv = obj->dev->dev_private;
-	struct sg_table *sgt;
-	int ret;
-
-	sgt = kzalloc_obj(*sgt);
-	if (!sgt)
-		return ERR_PTR(-ENOMEM);
-
-	ret = dma_get_sgtable(priv->dma_dev, sgt, dma_obj->vaddr,
-			      dma_obj->dma_addr, obj->size);
-	if (ret) {
-		DRM_ERROR("failed to allocate sgt, %d\n", ret);
-		kfree(sgt);
-		return ERR_PTR(ret);
-	}
-
-	return sgt;
-}
-
-static const struct drm_gem_object_funcs mtk_gem_object_funcs = {
-	.free = mtk_gem_free_object,
-	.print_info = drm_gem_dma_object_print_info,
-	.get_sg_table = mtk_gem_prime_get_sg_table,
-	.vmap = drm_gem_dma_object_vmap,
-	.mmap = mtk_gem_object_mmap,
-	.vm_ops = &drm_gem_dma_vm_ops,
-};
-
-static struct drm_gem_dma_object *mtk_gem_init(struct drm_device *dev,
-					unsigned long size, bool private)
-{
-	struct drm_gem_dma_object *dma_obj;
-	int ret;
-
-	size = round_up(size, PAGE_SIZE);
-
-	if (size == 0)
-		return ERR_PTR(-EINVAL);
-
-	dma_obj = kzalloc_obj(*dma_obj);
-	if (!dma_obj)
-		return ERR_PTR(-ENOMEM);
-
-	dma_obj->base.funcs = &mtk_gem_object_funcs;
-
-	if (private) {
-		ret = 0;
-		drm_gem_private_object_init(dev, &dma_obj->base, size);
-	} else {
-		ret = drm_gem_object_init(dev, &dma_obj->base, size);
-	}
-	if (ret) {
-		DRM_ERROR("failed to initialize gem object\n");
-		kfree(dma_obj);
-		return ERR_PTR(ret);
-	}
-
-	return dma_obj;
-}
-
-static struct drm_gem_dma_object *mtk_gem_create(struct drm_device *dev, size_t size)
-{
-	struct mtk_drm_private *priv = dev->dev_private;
-	struct drm_gem_dma_object *dma_obj;
-	struct drm_gem_object *obj;
-	int ret;
-
-	dma_obj = mtk_gem_init(dev, size, false);
-	if (IS_ERR(dma_obj))
-		return ERR_CAST(dma_obj);
-
-	obj = &dma_obj->base;
-
-	dma_obj->vaddr = dma_alloc_wc(priv->dma_dev, obj->size,
-				      &dma_obj->dma_addr,
-				      GFP_KERNEL | __GFP_NOWARN);
-	if (!dma_obj->vaddr) {
-		DRM_ERROR("failed to allocate %zx byte dma buffer", obj->size);
-		ret = -ENOMEM;
-		goto err_gem_free;
-	}
-
-	DRM_DEBUG_DRIVER("vaddr = %p dma_addr = %pad size = %zu\n",
-			 dma_obj->vaddr, &dma_obj->dma_addr,
-			 size);
-
-	return dma_obj;
-
-err_gem_free:
-	drm_gem_object_release(obj);
-	kfree(dma_obj);
-	return ERR_PTR(ret);
-}
-
-int mtk_gem_dumb_create(struct drm_file *file_priv, struct drm_device *dev,
-			struct drm_mode_create_dumb *args)
-{
-	struct drm_gem_dma_object *dma_obj;
-	int ret;
-
-	args->pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
-
-	/*
-	 * Multiply 2 variables of different types,
-	 * for example: args->size = args->spacing * args->height;
-	 * may cause coverity issue with unintentional overflow.
-	 */
-	args->size = args->pitch;
-	args->size *= args->height;
-
-	dma_obj = mtk_gem_create(dev, args->size);
-	if (IS_ERR(dma_obj))
-		return PTR_ERR(dma_obj);
-
-	/*
-	 * allocate a id of idr table where the obj is registered
-	 * and handle has the id what user can see.
-	 */
-	ret = drm_gem_handle_create(file_priv, &dma_obj->base, &args->handle);
-	if (ret)
-		goto err_handle_create;
-
-	/* drop reference from allocate - handle holds it now. */
-	drm_gem_object_put(&dma_obj->base);
-
-	return 0;
-
-err_handle_create:
-	mtk_gem_free_object(&dma_obj->base);
-	return ret;
-}
-
-static int mtk_gem_object_mmap(struct drm_gem_object *obj,
-			       struct vm_area_struct *vma)
-
-{
-	struct drm_gem_dma_object *dma_obj = to_drm_gem_dma_obj(obj);
-	struct mtk_drm_private *priv = obj->dev->dev_private;
-	int ret;
-
-	/*
-	 * Set vm_pgoff (used as a fake buffer offset by DRM) to 0 and map the
-	 * whole buffer from the start.
-	 */
-	vma->vm_pgoff -= drm_vma_node_start(&obj->vma_node);
-
-	/*
-	 * dma_alloc_attrs() allocated a struct page table for mtk_gem, so clear
-	 * VM_PFNMAP flag that was set by drm_gem_mmap_obj()/drm_gem_mmap().
-	 */
-	vm_flags_mod(vma, VM_IO | VM_DONTEXPAND | VM_DONTDUMP, VM_PFNMAP);
-
-	vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
-	vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
-
-	ret = dma_mmap_wc(priv->dma_dev, vma, dma_obj->vaddr,
-			  dma_obj->dma_addr, obj->size);
-	if (ret)
-		drm_gem_vm_close(vma);
-
-	return ret;
-}
-
-struct drm_gem_object *mtk_gem_prime_import_sg_table(struct drm_device *dev,
-			struct dma_buf_attachment *attach, struct sg_table *sgt)
-{
-	struct drm_gem_dma_object *dma_obj;
-
-	/* check if the entries in the sg_table are contiguous */
-	if (drm_prime_get_contiguous_size(sgt) < attach->dmabuf->size) {
-		DRM_ERROR("sg_table is not contiguous");
-		return ERR_PTR(-EINVAL);
-	}
-
-	dma_obj = mtk_gem_init(dev, attach->dmabuf->size, true);
-	if (IS_ERR(dma_obj))
-		return ERR_CAST(dma_obj);
-
-	dma_obj->dma_addr = sg_dma_address(sgt->sgl);
-	dma_obj->sgt = sgt;
-
-	return &dma_obj->base;
-}
diff --git a/drivers/gpu/drm/mediatek/mtk_gem.h b/drivers/gpu/drm/mediatek/mtk_gem.h
deleted file mode 100644
index afebc3a970a8..000000000000
--- a/drivers/gpu/drm/mediatek/mtk_gem.h
+++ /dev/null
@@ -1,17 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0-only */
-/*
- * Copyright (c) 2015 MediaTek Inc.
- */
-
-#ifndef _MTK_GEM_H_
-#define _MTK_GEM_H_
-
-#include <drm/drm_gem.h>
-#include <drm/drm_gem_dma_helper.h>
-
-int mtk_gem_dumb_create(struct drm_file *file_priv, struct drm_device *dev,
-			struct drm_mode_create_dumb *args);
-struct drm_gem_object *mtk_gem_prime_import_sg_table(struct drm_device *dev,
-			struct dma_buf_attachment *attach, struct sg_table *sg);
-
-#endif
-- 
2.53.0.473.g4a7958ca14-goog

Claude review: drm/mediatek: Set dedicated DMA device and drop custom GEM callbacks

[Claude Code Review Bot]

Patch Review

This is the big payoff: replaces the loop setting `private->all_drm_private[i]->dma_dev` and drops `mtk_gem.c` / `mtk_gem.h` entirely.

```c
-	for (i = 0; i < private->data->mmsys_dev_num; i++)
-		private->all_drm_private[i]->dma_dev = dma_dev;
+	drm_dev_set_dma_dev(drm, dma_dev);
```

One observation: the old code set `.dma_dev` on **all** `mmsys_dev_num` drm_private instances, while the new code only calls `drm_dev_set_dma_dev()` on the single `drm` device. For multi-MMSYS configurations (e.g., MT8195 which has 2 MMSYS blocks), are there multiple `drm_device` instances that need the DMA device set? Looking at the code, all the `all_drm_private[i]` share the same `drm` device (they're all private data for sub-components bound to the same master), and `dma_dev` is stored in the private struct, not in `drm_device`. The new code sets it on `drm->dma_dev` which is the single DRM device, so this should be fine -- all paths now go through `drm_dev_dma_dev(drm)` which reads the single `drm->dma_dev` field. But it's worth confirming that none of the old `private->dma_dev` references remain elsewhere in the mediatek driver code beyond what's being removed.

The replacement of custom driver ops with `DRM_GEM_DMA_DRIVER_OPS` is clean:

```c
-	.dumb_create = mtk_gem_dumb_create,
+	DRM_GEM_DMA_DRIVER_OPS,
...
-	.gem_prime_import = mtk_gem_prime_import,
-	.gem_prime_import_sg_table = mtk_gem_prime_import_sg_table,
```

The commit message correctly notes the VM_DONTDUMP and VM_IO flag difference in the mmap path. The GEM DMA helpers don't set these because `dma_mmap_wc()` handles the mapping directly and the backing pages come from DMA-coherent allocations, making VM_IO inappropriate as the author notes.

**Minor**: No remaining references to `mtk_gem.h` in other mediatek files besides `mtk_crtc.c` (which is cleaned up)? The patch looks complete.

**Reviewed-by worthy**: Yes, with confirmation that multi-MMSYS setups work correctly with only one `drm_dev_set_dma_dev()` call.

---
Generated by Claude Code Patch Reviewer

[PATCH 4/4] drm/sun4i: Use backend/mixer as dedicated DMA device

[Chen-Yu Tsai]

The sun4i DRM driver deals with DMA constraints in a peculiar way.
Instead of using the actual DMA device in various helpers, it justs
reconfigures the DMA constraints of the virtual display device using
the DMA device's device tree node by calling of_dma_configure().

Turns out of_dma_configure() should only be called from bus code.
Lately this also triggers a big warning through of_iommu_configure()
and ultimately __iommu_probe_device():

    late IOMMU probe at driver bind, something fishy here!

Now that the GEM DMA helpers have proper support for allocating
and mapping buffers with a dedicated DMA device, switch over to
it as the proper solution.

The mixer change was tested on a Pine H64 model B. The backend change
was only compile tested. Though I don't expect any issues, help testing
on an older device would be appreciated.

Signed-off-by: Chen-Yu Tsai <wenst@chromium.org>
---
 drivers/gpu/drm/sun4i/sun4i_backend.c | 27 +++++++++++++++------------
 drivers/gpu/drm/sun4i/sun8i_mixer.c   | 27 +++++++++++++++------------
 2 files changed, 30 insertions(+), 24 deletions(-)

diff --git a/drivers/gpu/drm/sun4i/sun4i_backend.c b/drivers/gpu/drm/sun4i/sun4i_backend.c
index 6391bdc94a5c..a57fb5151def 100644
--- a/drivers/gpu/drm/sun4i/sun4i_backend.c
+++ b/drivers/gpu/drm/sun4i/sun4i_backend.c
@@ -798,18 +798,21 @@ static int sun4i_backend_bind(struct device *dev, struct device *master,
 	dev_set_drvdata(dev, backend);
 	spin_lock_init(&backend->frontend_lock);
 
-	if (of_property_present(dev->of_node, "interconnects")) {
-		/*
-		 * This assume we have the same DMA constraints for all our the
-		 * devices in our pipeline (all the backends, but also the
-		 * frontends). This sounds bad, but it has always been the case
-		 * for us, and DRM doesn't do per-device allocation either, so
-		 * we would need to fix DRM first...
-		 */
-		ret = of_dma_configure(drm->dev, dev->of_node, true);
-		if (ret)
-			return ret;
-	}
+	/*
+	 * This assume we have the same DMA constraints for all our the
+	 * devices in our pipeline (all the backends, but also the
+	 * frontends). This sounds bad, but it has always been the case
+	 * for us, and DRM doesn't do per-device allocation either, so
+	 * we would need to fix DRM first...
+	 *
+	 * Always use the first bound backend as the DMA device. While
+	 * our device trees always have all backends enabled, some in
+	 * the wild may actually have the first one disabled. If both
+	 * are enabled, the order in which they are bound is guaranteed
+	 * since the driver adds components in order.
+	 */
+	if (drm_dev_dma_dev(drm) == drm->dev)
+		drm_dev_set_dma_dev(drm, dev);
 
 	backend->engine.node = dev->of_node;
 	backend->engine.ops = &sun4i_backend_engine_ops;
diff --git a/drivers/gpu/drm/sun4i/sun8i_mixer.c b/drivers/gpu/drm/sun4i/sun8i_mixer.c
index 02acc7cbdb97..4071ab38b4ae 100644
--- a/drivers/gpu/drm/sun4i/sun8i_mixer.c
+++ b/drivers/gpu/drm/sun4i/sun8i_mixer.c
@@ -536,18 +536,21 @@ static int sun8i_mixer_bind(struct device *dev, struct device *master,
 	mixer->engine.ops = &sun8i_engine_ops;
 	mixer->engine.node = dev->of_node;
 
-	if (of_property_present(dev->of_node, "iommus")) {
-		/*
-		 * This assume we have the same DMA constraints for
-		 * all our the mixers in our pipeline. This sounds
-		 * bad, but it has always been the case for us, and
-		 * DRM doesn't do per-device allocation either, so we
-		 * would need to fix DRM first...
-		 */
-		ret = of_dma_configure(drm->dev, dev->of_node, true);
-		if (ret)
-			return ret;
-	}
+	/*
+	 * This assume we have the same DMA constraints for all our the
+	 * devices in our pipeline (all the backends, but also the
+	 * frontends). This sounds bad, but it has always been the case
+	 * for us, and DRM doesn't do per-device allocation either, so
+	 * we would need to fix DRM first...
+	 *
+	 * Always use the first bound backend as the DMA device. While
+	 * our device trees always have all backends enabled, some in
+	 * the wild may actually have the first one disabled. If both
+	 * are enabled, the order in which they are bound is guaranteed
+	 * since the driver adds components in order.
+	 */
+	if (drm_dev_dma_dev(drm) == drm->dev)
+		drm_dev_set_dma_dev(drm, dev);
 
 	/*
 	 * While this function can fail, we shouldn't do anything
-- 
2.53.0.473.g4a7958ca14-goog

Claude review: drm/sun4i: Use backend/mixer as dedicated DMA device

[Claude Code Review Bot]

Patch Review

This replaces the `of_dma_configure()` hack with proper `drm_dev_set_dma_dev()` calls. The approach is:

```c
-	if (of_property_present(dev->of_node, "interconnects")) {
-		ret = of_dma_configure(drm->dev, dev->of_node, true);
-		if (ret)
-			return ret;
-	}
+	if (drm_dev_dma_dev(drm) == drm->dev)
+		drm_dev_set_dma_dev(drm, dev);
```

The guard `drm_dev_dma_dev(drm) == drm->dev` ensures only the first backend/mixer to bind sets itself as the DMA device, which matches the comment explaining the rationale.

**Behavioral change to note**: The old backend code was gated on `of_property_present(dev->of_node, "interconnects")`, and the old mixer code was gated on `of_property_present(dev->of_node, "iommus")`. The new code unconditionally sets the DMA device. This changes behavior for devices *without* those DT properties -- previously the DRM device's own (parent) device would be used for DMA, now it will use the backend/mixer device. This should generally be fine because the backend/mixer *is* the actual DMA device regardless of whether an IOMMU or interconnect is involved, but it's a semantic change worth calling out. If any board has a backend/mixer that truly can't do DMA and relies on the parent device, this could be a problem, though that scenario seems unlikely.

**Comment nit**: The comment in `sun8i_mixer.c` says "backends" and "frontends" (copy-pasted from `sun4i_backend.c`) but should probably say "mixers" to be accurate for the mixer context. This existed partially before but was made worse by the patch expanding the comment. Line 1175-1176:

```c
+	 * This assume we have the same DMA constraints for all our the
+	 * devices in our pipeline (all the backends, but also the
+	 * frontends).
```

This text makes sense for `sun4i_backend.c` but is inaccurate in `sun8i_mixer.c` where the original comment correctly said "mixers" instead of "backends"/"frontends". The old mixer comment was already more appropriate.

**Unbind path**: The old code modified `drm->dev`'s DMA configuration directly, so there was no cleanup needed on unbind. With the new approach, `drm_dev_set_dma_dev()` takes a reference on the device via `get_device()`. Is this reference properly released when the DRM device is torn down? Looking at `drm_dev_set_dma_dev()`, it does `get_device(dma_dev)` and `put_device(dev->dma_dev)`. The DRM device cleanup in `drm_dev_release()` should handle this, but it's worth verifying that `dma_dev` is properly cleared. This applies to all drivers calling `drm_dev_set_dma_dev()` from a component bind function -- if the component unbinds before the DRM device is destroyed, the reference on the component's device will keep it alive, which should be fine since the DRM device will outlive individual components.

**Reviewed-by worthy**: Yes, with the comment nit about "backends"/"frontends" in the mixer code.

---
Generated by Claude Code Patch Reviewer

Claude review: drm/gem-dma: Support dedicated DMA device for allocation

[Claude Code Review Bot]

Overall Series Review

Subject: drm/gem-dma: Support dedicated DMA device for allocation
Author: Chen-Yu Tsai <wenst@chromium.org>
Patches: 5
Reviewed: 2026-03-11T13:48:35.189516

---

This is a well-structured 4-patch series that extends the existing `drm_dev_dma_dev()` / `drm_dev_set_dma_dev()` infrastructure to the GEM DMA helpers for allocation and mmap operations. The motivation is clear: several embedded SoC display drivers (MediaTek, Allwinner, Exynos, Rockchip) have reinvented this wheel with custom GEM callbacks, and this series aims to consolidate that into the core helpers. The approach is sound and the code is straightforward. The MediaTek conversion is a satisfying deletion of ~250 lines of custom code. The sun4i conversion properly fixes a misuse of `of_dma_configure()` that triggers warnings with IOMMUs.

Overall the series looks good. I have a few observations below, mostly minor.

---
Generated by Claude Code Patch Reviewer