Re: [PATCH v1 1/6] drm/panfrost: Set IOMMU_CACHE flag

From: Steven Price
Date: Fri Mar 14 2025 - 12:17:44 EST


On 10/03/2025 19:59, Ariel D'Alessandro wrote:
> Panfrost does not support uncached mappings, so flag them properly. Also
> flag the pages that are mapped as response to a page fault as cached.

As I understand it the hardware only sort-of supports uncached mappings.
The legacy page table format has two options: cached, or "implementation
defined" (ARM_MALI_LPAE_MEMATTR_IMP_DEF). When selecting IMP_DEF that
means that the internal units signal to the cache whether a particular
access should be cached or not. I believe the theory is that they can
decide whether it makes sense to store in the cache or not.

Having said all that, I have never observed any actual performance
difference and I suspect it won't make any actual difference. And of
course the AArch64 page tables need this to be set. I just thought I
should explain the background and that this wasn't wrong for the legacy
page tables.

Reviewed-by: Steven Price <steven.price@xxxxxxx>

Steve

> Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx>
> Signed-off-by: Ariel D'Alessandro <ariel.dalessandro@xxxxxxxxxxxxx>
> ---
> drivers/gpu/drm/panfrost/panfrost_mmu.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/panfrost/panfrost_mmu.c b/drivers/gpu/drm/panfrost/panfrost_mmu.c
> index b91019cd5acb1..9e6f198ef5c1b 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_mmu.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_mmu.c
> @@ -327,7 +327,7 @@ int panfrost_mmu_map(struct panfrost_gem_mapping *mapping)
> struct drm_gem_object *obj = &shmem->base;
> struct panfrost_device *pfdev = to_panfrost_device(obj->dev);
> struct sg_table *sgt;
> - int prot = IOMMU_READ | IOMMU_WRITE;
> + int prot = IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE;
>
> if (WARN_ON(mapping->active))
> return 0;
> @@ -528,7 +528,7 @@ static int panfrost_mmu_map_fault_addr(struct panfrost_device *pfdev, int as,
> goto err_map;
>
> mmu_map_sg(pfdev, bomapping->mmu, addr,
> - IOMMU_WRITE | IOMMU_READ | IOMMU_NOEXEC, sgt);
> + IOMMU_WRITE | IOMMU_READ | IOMMU_CACHE | IOMMU_NOEXEC, sgt);
>
> bomapping->active = true;
> bo->heap_rss_size += SZ_2M;