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. Signed-off-by: Matt Coster <matt.coster@xxxxxxxxxx> --- Changes in v3: - Change from a workaround to a regular codepath - Add missing <linux/property.h> include - Link to v2: https://lore.kernel.org/r/20241118-sets-bxs-4-64-patch-v1-v2-19-3fd45d9fb0cf@xxxxxxxxxx Changes in v2: - None - Link to v1: https://lore.kernel.org/r/20241105-sets-bxs-4-64-patch-v1-v1-19-4ed30e865892@xxxxxxxxxx --- drivers/gpu/drm/imagination/pvr_gem.c | 10 +++++++--- drivers/gpu/drm/imagination/pvr_gem.h | 6 ++++-- drivers/gpu/drm/imagination/pvr_mmu.c | 8 +++++++- 3 files changed, 18 insertions(+), 6 deletions(-) diff --git a/drivers/gpu/drm/imagination/pvr_gem.c b/drivers/gpu/drm/imagination/pvr_gem.c index 6a8c81fe8c1e85c2130a4fe90fce35b6a2be35aa..9467cd563339cd5bf62ff92edba115ed5721ee76 100644 --- a/drivers/gpu/drm/imagination/pvr_gem.c +++ b/drivers/gpu/drm/imagination/pvr_gem.c @@ -19,6 +19,7 @@ #include <linux/log2.h> #include <linux/mutex.h> #include <linux/pagemap.h> +#include <linux/property.h> #include <linux/refcount.h> #include <linux/scatterlist.h> @@ -336,6 +337,7 @@ struct drm_gem_object *pvr_gem_create_object(struct drm_device *drm_dev, size_t 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; @@ -345,7 +347,10 @@ pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags) 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); @@ -360,8 +365,7 @@ pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags) 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 diff --git a/drivers/gpu/drm/imagination/pvr_gem.h b/drivers/gpu/drm/imagination/pvr_gem.h index e0e5ea509a2e88a437b8d241ea13c7bab2220f56..c99f30cc62088c030bd8a806df79b738b62a968f 100644 --- a/drivers/gpu/drm/imagination/pvr_gem.h +++ b/drivers/gpu/drm/imagination/pvr_gem.h @@ -44,8 +44,10 @@ struct pvr_file; * 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 diff --git a/drivers/gpu/drm/imagination/pvr_mmu.c b/drivers/gpu/drm/imagination/pvr_mmu.c index 4fe70610ed94cf707e631f8148af081a94f97327..450d476d183f0173d0ef03f0d8897fbeb04831a2 100644 --- a/drivers/gpu/drm/imagination/pvr_mmu.c +++ b/drivers/gpu/drm/imagination/pvr_mmu.c @@ -17,6 +17,7 @@ #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_)) @@ -259,6 +260,7 @@ pvr_mmu_backing_page_init(struct pvr_mmu_backing_page *page, struct device *dev = from_pvr_device(pvr_dev)->dev; struct page *raw_page; + pgprot_t prot; int err; dma_addr_t dma_addr; @@ -268,7 +270,11 @@ pvr_mmu_backing_page_init(struct pvr_mmu_backing_page *page, 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; -- 2.48.1
