Re: [PATCH] iommu/sun50i: Convert to map_pages/unmap_pages

[Robin Murphy]

On 2023-09-12 19:19, Jernej Skrabec wrote:
> Convert driver to use map_pages and unmap_pages. Since functions operate
> on page table, extend them to be able to operate on whole table, not
> just one entry.
>
> Signed-off-by: Jernej Skrabec <jernej.skrabec@xxxxxxxxx>
> ---
> Hi!
>
> I'm not sure if it makes sense to check validity of page table entry when
> unmaping pages makes sense. What do you think?
>
> Best regards,
> Jernej
>
>   drivers/iommu/sun50i-iommu.c | 55 +++++++++++++++++++++---------------
>   1 file changed, 33 insertions(+), 22 deletions(-)
>
> diff --git a/drivers/iommu/sun50i-iommu.c b/drivers/iommu/sun50i-iommu.c
> index 74c5cb93e900..be6102730a56 100644
> --- a/drivers/iommu/sun50i-iommu.c
> +++ b/drivers/iommu/sun50i-iommu.c
> @@ -589,11 +589,12 @@ static u32 *sun50i_dte_get_page_table(struct sun50i_iommu_domain *sun50i_domain,
>   }
>   
>   static int sun50i_iommu_map(struct iommu_domain *domain, unsigned long iova,
> -			    phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
> +			    phys_addr_t paddr, size_t pgsize, size_t pgcount,
> +			    int prot, gfp_t gfp, size_t *mapped)
>   {
>   	struct sun50i_iommu_domain *sun50i_domain = to_sun50i_domain(domain);
>   	struct sun50i_iommu *iommu = sun50i_domain->iommu;
> -	u32 pte_index;
> +	u32 pte_index, count, i;
>   	u32 *page_table, *pte_addr;
>   	int ret = 0;
>   
> @@ -604,45 +605,55 @@ static int sun50i_iommu_map(struct iommu_domain *domain, unsigned long iova,
>   	}
>   
>   	pte_index = sun50i_iova_get_pte_index(iova);
> -	pte_addr = &page_table[pte_index];
> -	if (unlikely(sun50i_pte_is_page_valid(*pte_addr))) {
> -		phys_addr_t page_phys = sun50i_pte_get_page_address(*pte_addr);
> -		dev_err(iommu->dev,
> -			"iova %pad already mapped to %pa cannot remap to %pa prot: %#x\n",
> -			&iova, &page_phys, &paddr, prot);
> -		ret = -EBUSY;
> -		goto out;
> +	count = min(pgcount, (size_t)NUM_PT_ENTRIES - pte_index);
> +	for (i = 0; i < count; i++) {
> +		pte_addr = &page_table[pte_index + i];
> +		if (unlikely(sun50i_pte_is_page_valid(*pte_addr))) {
> +			phys_addr_t page_phys = sun50i_pte_get_page_address(*pte_addr);
> +
> +			dev_err(iommu->dev,
> +				"iova %pad already mapped to %pa cannot remap to %pa prot: %#x\n",
> +				&iova, &page_phys, &paddr, prot);
> +			ret = -EBUSY;
> +			goto out;
> +		}
> +		*pte_addr = sun50i_mk_pte(paddr, prot);
> +		paddr += SPAGE_SIZE;
>   	}
>   
> -	*pte_addr = sun50i_mk_pte(paddr, prot);
> -	sun50i_table_flush(sun50i_domain, pte_addr, 1);
> +	sun50i_table_flush(sun50i_domain, &page_table[pte_index], i);

I think you want the "out" label here, so that "mapped" is still updated 
and the core knows how much to undo on failure. I guess one could argue 
that the pagetables are already messed up if that happens, but it still 
doesn't seem like a good reason to deliberately compound the problem by 
leaving behind even more "unexpected" entries.

Thanks,
Robin.

> +	*mapped = i * SPAGE_SIZE;
>   
>   out:
>   	return ret;
>   }
>   
>   static size_t sun50i_iommu_unmap(struct iommu_domain *domain, unsigned long iova,
> -				 size_t size, struct iommu_iotlb_gather *gather)
> +				 size_t pgsize, size_t pgcount,
> +				 struct iommu_iotlb_gather *gather)
>   {
>   	struct sun50i_iommu_domain *sun50i_domain = to_sun50i_domain(domain);
> +	u32 dte, count, i, pte_index;
>   	phys_addr_t pt_phys;
>   	u32 *pte_addr;
> -	u32 dte;
>   
>   	dte = sun50i_domain->dt[sun50i_iova_get_dte_index(iova)];
>   	if (!sun50i_dte_is_pt_valid(dte))
>   		return 0;
>   
>   	pt_phys = sun50i_dte_get_pt_address(dte);
> -	pte_addr = (u32 *)phys_to_virt(pt_phys) + sun50i_iova_get_pte_index(iova);
> +	pte_index = sun50i_iova_get_pte_index(iova);
> +	pte_addr = (u32 *)phys_to_virt(pt_phys) + pte_index;
>   
> -	if (!sun50i_pte_is_page_valid(*pte_addr))
> -		return 0;
> +	count = min(pgcount, (size_t)NUM_PT_ENTRIES - pte_index);
> +	for (i = 0; i < count; i++)
> +		if (!sun50i_pte_is_page_valid(pte_addr[i]))
> +			break;
>   
> -	memset(pte_addr, 0, sizeof(*pte_addr));
> -	sun50i_table_flush(sun50i_domain, pte_addr, 1);
> +	memset(pte_addr, 0, sizeof(*pte_addr) * i);
> +	sun50i_table_flush(sun50i_domain, pte_addr, i);
>   
> -	return SZ_4K;
> +	return i * SPAGE_SIZE;
>   }
>   
>   static phys_addr_t sun50i_iommu_iova_to_phys(struct iommu_domain *domain,
> @@ -838,8 +849,8 @@ static const struct iommu_ops sun50i_iommu_ops = {
>   		.iotlb_sync_map = sun50i_iommu_iotlb_sync_map,
>   		.iotlb_sync	= sun50i_iommu_iotlb_sync,
>   		.iova_to_phys	= sun50i_iommu_iova_to_phys,
> -		.map		= sun50i_iommu_map,
> -		.unmap		= sun50i_iommu_unmap,
> +		.map_pages	= sun50i_iommu_map,
> +		.unmap_pages	= sun50i_iommu_unmap,
>   		.free		= sun50i_iommu_domain_free,
>   	}
>   };