Re: [RFC PATCH 6/7] drm/panfrost: Use shmem sparse allocation for heap BOs

[Boris Brezillon <boris.brezillon@collabora.com>]

On Tue, 18 Feb 2025 23:25:36 +0000
Adrián Larumbe <adrian.larumbe@collabora.com> wrote:

> Panfrost heap BOs grow on demand when the GPU triggers a page fault after
> accessing an address within the BO's virtual range.
> 
> We still store the sgts we get back from the shmem sparse allocation function,
> since it was decided management of sparse memory SGTs should be done by client
> drivers rather than the shmem subsystem.
> 
> Signed-off-by: Adrián Larumbe <adrian.larumbe@collabora.com>
> ---
>  drivers/gpu/drm/panfrost/panfrost_gem.c | 12 ++--
>  drivers/gpu/drm/panfrost/panfrost_gem.h |  2 +-
>  drivers/gpu/drm/panfrost/panfrost_mmu.c | 85 +++++--------------------
>  3 files changed, 25 insertions(+), 74 deletions(-)
> 
> diff --git a/drivers/gpu/drm/panfrost/panfrost_gem.c b/drivers/gpu/drm/panfrost/panfrost_gem.c
> index 8e0ff3efede7..0cda2c4e524f 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_gem.c
> +++ b/drivers/gpu/drm/panfrost/panfrost_gem.c
> @@ -40,10 +40,10 @@ static void panfrost_gem_free_object(struct drm_gem_object *obj)
>  		int n_sgt = bo->base.base.size / SZ_2M;
>  
>  		for (i = 0; i < n_sgt; i++) {
> -			if (bo->sgts[i].sgl) {
> -				dma_unmap_sgtable(pfdev->dev, &bo->sgts[i],
> +			if (bo->sgts[i]) {
> +				dma_unmap_sgtable(pfdev->dev, bo->sgts[i],
>  						  DMA_BIDIRECTIONAL, 0);
> -				sg_free_table(&bo->sgts[i]);
> +				sg_free_table(bo->sgts[i]);
>  			}
>  		}
>  		kvfree(bo->sgts);
> @@ -274,7 +274,11 @@ panfrost_gem_create(struct drm_device *dev, size_t size, u32 flags)
>  	if (flags & PANFROST_BO_HEAP)
>  		size = roundup(size, SZ_2M);
>  
> -	shmem = drm_gem_shmem_create(dev, size);
> +	if (flags & PANFROST_BO_HEAP)
> +		shmem = drm_gem_shmem_create_sparse(dev, size);
> +	else
> +		shmem = drm_gem_shmem_create(dev, size);
> +
>  	if (IS_ERR(shmem))
>  		return ERR_CAST(shmem);
>  
> diff --git a/drivers/gpu/drm/panfrost/panfrost_gem.h b/drivers/gpu/drm/panfrost/panfrost_gem.h
> index 7516b7ecf7fe..2a8d0752011e 100644
> --- a/drivers/gpu/drm/panfrost/panfrost_gem.h
> +++ b/drivers/gpu/drm/panfrost/panfrost_gem.h
> @@ -11,7 +11,7 @@ struct panfrost_mmu;
>  
>  struct panfrost_gem_object {
>  	struct drm_gem_shmem_object base;
> -	struct sg_table *sgts;
> +	struct sg_table **sgts;

I guess using an xarray here would make sense. Or maybe even an
sg_append_table, since we don't expect holes in the populated pages.
This makes me wonder if we really want the gem_shmem layer to automate
sgt creation for sparse GEM objects. Looks like something the driver
can easily optimize for its use-case.