drm/imagination: Use cached memory with dma_coherent

The TI k3-j721s2 platform does not allow us to use uncached memory
(which is what the driver currently does) without disabling cache snooping
on the AXI ACE-Lite interface, which would be too much of a performance
hit.

Given the platform is dma-coherent, we can simply force all
device-accessible memory allocations through the CPU cache. In fact, this
can be done whenever the dma_coherent attribute is present.

Reviewed-by: Frank Binns <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Matt Coster <[e-mail]>

drivers/gpu/drm/imagination/pvr_gem.c

···
 #include <linux/log2.h>
 #include <linux/mutex.h>
 #include <linux/pagemap.h>
+#include <linux/property.h>
 #include <linux/refcount.h>
 #include <linux/scatterlist.h>
 
···
 struct pvr_gem_object *
 pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags)
 {
+	struct drm_device *drm_dev = from_pvr_device(pvr_dev);
 	struct drm_gem_shmem_object *shmem_obj;
 	struct pvr_gem_object *pvr_obj;
 	struct sg_table *sgt;
···
 	if (size == 0 || !pvr_gem_object_flags_validate(flags))
 		return ERR_PTR(-EINVAL);
 
-	shmem_obj = drm_gem_shmem_create(from_pvr_device(pvr_dev), size);
+	if (device_get_dma_attr(drm_dev->dev) == DEV_DMA_COHERENT)
+		flags |= PVR_BO_CPU_CACHED;
+
+	shmem_obj = drm_gem_shmem_create(drm_dev, size);
 	if (IS_ERR(shmem_obj))
 		return ERR_CAST(shmem_obj);
 
···
 		goto err_shmem_object_free;
 	}
 
-	dma_sync_sgtable_for_device(shmem_obj->base.dev->dev, sgt,
-				    DMA_BIDIRECTIONAL);
+	dma_sync_sgtable_for_device(drm_dev->dev, sgt, DMA_BIDIRECTIONAL);
 
 	/*
 	 * Do this last because pvr_gem_object_zero() requires a fully

drivers/gpu/drm/imagination/pvr_gem.h

···
  * Bits not defined anywhere are "undefined".
  *
  * CPU mapping options
- *    :PVR_BO_CPU_CACHED: By default, all GEM objects are mapped write-combined on the CPU. Set this
- *       flag to override this behaviour and map the object cached.
+ *    :PVR_BO_CPU_CACHED: By default, all GEM objects are mapped write-combined on the CPU. Set
+ *       this flag to override this behaviour and map the object cached. If the dma_coherent
+ *       property is present in devicetree, all allocations will be mapped as if this flag was set.
+ *       This does not require any additional consideration at allocation time.
  *
  * Firmware options
  *    :PVR_BO_FW_NO_CLEAR_ON_RESET: By default, all FW objects are cleared and reinitialised on hard

drivers/gpu/drm/imagination/pvr_mmu.c

···
 #include <linux/dma-mapping.h>
 #include <linux/kmemleak.h>
 #include <linux/minmax.h>
+#include <linux/property.h>
 #include <linux/sizes.h>
 
 #define PVR_SHIFT_FROM_SIZE(size_) (__builtin_ctzll(size_))
···
 	struct device *dev = from_pvr_device(pvr_dev)->dev;
 
 	struct page *raw_page;
+	pgprot_t prot;
 	int err;
 
 	dma_addr_t dma_addr;
···
 	if (!raw_page)
 		return -ENOMEM;
 
-	host_ptr = vmap(&raw_page, 1, VM_MAP, pgprot_writecombine(PAGE_KERNEL));
+	prot = PAGE_KERNEL;
+	if (device_get_dma_attr(dev) != DEV_DMA_COHERENT)
+		prot = pgprot_writecombine(prot);
+
+	host_ptr = vmap(&raw_page, 1, VM_MAP, prot);
 	if (!host_ptr) {
 		err = -ENOMEM;
 		goto err_free_page;