Re: [PATCH v9 5/6] drm/panthor: Support sparse mappings

[Boris Brezillon]

On Fri, 24 Apr 2026 11:09:27 +0100
Steven Price <steven.price@xxxxxxx> wrote:

> Hi Adrián,
> 
> On 22/04/2026 13:25, Adrián Larumbe wrote:
> > Allow UM to bind sparsely populated memory regions by cyclically mapping
> > virtual ranges over a kernel-allocated dummy BO. This alternative is
> > preferable to the old method of handling sparseness in the UMD, because it
> > relied on the creation of a buffer object to the same end, despite the fact
> > Vulkan sparse resources don't need to be backed by a driver BO.
> > 
> > The choice of backing sparsely-bound regions with a Panhtor BO was made so
> > as to profit from the existing shrinker reclaim code. That way no special
> > treatment must be given to the dummy sparse BOs when reclaiming memory, as
> > would be the case if we had chosen a raw kernel page implementation.
> > 
> > A new dummy BO is allocated per open file context, because even though the
> > Vulkan spec mandates that writes into sparsely bound regions must be
> > discarded, our implementation is still a workaround over the fact Mali CSF
> > GPUs cannot support this behaviour on the hardware level, so writes still
> > make it into the backing BO. If we had a global one, then it could be a
> > venue for information leaks between file contexts, which should never
> > happen in DRM.
> > 
> > Reviewed-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx>
> > Signed-off-by: Adrián Larumbe <adrian.larumbe@xxxxxxxxxxxxx>  
> 
> Looks good, a few issues below.
> 
> I'm worried about remap_evicted_vma() and how that interacts with sparse
> mappings. Does that need to be fixed up to handle sparse mappings? Or is
> there something to prevent the dummy BO being reclaimed? I might be
> missing something here.

Given the sparse mappings still have a vm_bo+gem object attached to them,
I think reclaim is fine, but I'll double check.

> > +static int
> > +panthor_vm_map_sparse(struct panthor_vm *vm, u64 iova, int prot,
> > +		      struct sg_table *sgt, u64 size)
> > +{
> > +	u64 start_iova = iova;
> > +	int ret;
> > +
> > +	if (iova & (SZ_2M - 1)) {
> > +		u64 unaligned_size = min(ALIGN(iova, SZ_2M) - iova, size);
> > +
> > +		ret = panthor_vm_map_pages(vm, iova, prot, sgt,
> > +					   0, unaligned_size);
> > +		if (ret)
> > +			return ret;
> > +
> > +		size -= unaligned_size;
> > +		iova += unaligned_size;
> > +	}
> > +
> > +	/* TODO: we should probably optimize this at the io_pgtable level. */
> > +	while (size > 0) {
> > +		u64 next_size = min(size, sg_dma_len(sgt->sgl));  
> 
> Here we're only using the first entry of the scatter list. So I think in
> the fragmented case we don't end up using the full 2MB.

It should just be

		u32 chunk_size = min(size, SZ_2M);

really. The fact the BO is backed by physically contiguous memory
doesn't matter because panthor_vm_map_pages() can cope with that
already.

> 
> > +
> > +		ret = panthor_vm_map_pages(vm, iova, prot,
> > +					   sgt, 0, next_size);
> > +		if (ret)
> > +			goto err_unmap;
> > +
> > +		size -= next_size;
> > +		iova += next_size;
> > +	}

To sum up, the whole thing can be simplified to something like:

static int
panthor_vm_map_sparse(struct panthor_vm *vm, u64 iova, int prot,
		      struct sg_table *sgt, u64 size)
	u64 offset = 0;

	while (offset < size) {
		u32 chunk_size = min(size - offset, SZ_2M - (iova & (SZ_2M - 1)));

		ret = panthor_vm_map_pages(vm, iova + offset, prot,
					   sgt, 0, chunk_size);
		if (ret) {
			panthor_vm_unmap_pages(vm, iova, offset);
			return ret;
		}

		offset += chunk_size;
	}

	return 0;
}