[RFC PATCH 3/5] iommu: Add support for the unmap_pages IOMMU callback

[Robin Murphy]

On 2021-03-31 04:00, Isaac J. Manjarres wrote:
> The IOMMU framework currently unmaps memory one page block at a time,
> per the page block sizes that are supported by the IOMMU hardware.
> Now that IOMMU drivers can supply a callback for unmapping multiple
> in one call, add support in the IOMMU framework to calculate how many
> page mappings of the same size can be unmapped in one shot, and invoke the
> IOMMU driver's unmap_pages callback if it has one. Otherwise, the
> existing behavior will be used.
> 
> Signed-off-by: Isaac J. Manjarres <isaacm at codeaurora.org>
> Suggested-by: Will Deacon <will at kernel.org>
> ---
>   drivers/iommu/iommu.c | 44 +++++++++++++++++++++++++++++++++++++------
>   1 file changed, 38 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index d0b0a15dba84..dc4295f6bc7f 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -2356,8 +2356,8 @@ phys_addr_t iommu_iova_to_phys(struct iommu_domain *domain, dma_addr_t iova)
>   }
>   EXPORT_SYMBOL_GPL(iommu_iova_to_phys);
>   
> -static size_t iommu_pgsize(struct iommu_domain *domain,
> -			   unsigned long addr_merge, size_t size)
> +static size_t __iommu_pgsize(struct iommu_domain *domain,
> +			     unsigned long addr_merge, size_t size)
>   {
>   	unsigned int pgsize_idx;
>   	size_t pgsize;
> @@ -2388,6 +2388,24 @@ static size_t iommu_pgsize(struct iommu_domain *domain,
>   	return pgsize;
>   }
>   
> +static size_t iommu_pgsize(struct iommu_domain *domain,
> +			   unsigned long addr_merge, size_t size,
> +			   size_t *pgcount)
> +{
> +	size_t pgsize = __iommu_pgsize(domain, addr_merge, size);
> +	size_t pgs = 0;
> +
> +	do {
> +		pgs++;
> +		size -= pgsize;
> +		addr_merge += pgsize;
> +	} while (size && __iommu_pgsize(domain, addr_merge, size) == pgsize);

This looks horrifically inefficient. As part of calculating the best 
current page size it should then be pretty trivial to calculate "(size & 
next_pgsize_up - 1) >> pgsize_idx" for the number of current-size pages 
up to the next-better-size boundary (with next_pgsize_up being 0 if 
pgsize is already the largest possible for the relative alignment of 
physical and virtual address). A loop is just... yuck :(

> +
> +	*pgcount = pgs;
> +
> +	return pgsize;
> +}
> +
>   static int __iommu_map(struct iommu_domain *domain, unsigned long iova,
>   		       phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
>   {
> @@ -2422,7 +2440,7 @@ static int __iommu_map(struct iommu_domain *domain, unsigned long iova,
>   	pr_debug("map: iova 0x%lx pa %pa size 0x%zx\n", iova, &paddr, size);
>   
>   	while (size) {
> -		size_t pgsize = iommu_pgsize(domain, iova | paddr, size);
> +		size_t pgsize = __iommu_pgsize(domain, iova | paddr, size);
>   
>   		pr_debug("mapping: iova 0x%lx pa %pa pgsize 0x%zx\n",
>   			 iova, &paddr, pgsize);
> @@ -2473,6 +2491,21 @@ int iommu_map_atomic(struct iommu_domain *domain, unsigned long iova,
>   }
>   EXPORT_SYMBOL_GPL(iommu_map_atomic);
>   
> +static size_t __iommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
> +				  size_t size, struct iommu_iotlb_gather *iotlb_gather)
> +{
> +	const struct iommu_ops *ops = domain->ops;
> +	size_t pgsize, pgcount;
> +
> +	if (ops->unmap_pages) {
> +		pgsize = iommu_pgsize(domain, iova, size, &pgcount);
> +		return ops->unmap_pages(domain, iova, pgsize, pgcount, iotlb_gather);
> +	}
> +
> +	pgsize = __iommu_pgsize(domain, iova, size);
> +	return ops->unmap(domain, iova, pgsize, iotlb_gather);
> +}
> +
>   static size_t __iommu_unmap(struct iommu_domain *domain,
>   			    unsigned long iova, size_t size,
>   			    struct iommu_iotlb_gather *iotlb_gather)
> @@ -2510,9 +2543,8 @@ static size_t __iommu_unmap(struct iommu_domain *domain,
>   	 * or we hit an area that isn't mapped.
>   	 */
>   	while (unmapped < size) {
> -		size_t pgsize = iommu_pgsize(domain, iova, size - unmapped);
> -
> -		unmapped_page = ops->unmap(domain, iova, pgsize, iotlb_gather);
> +		unmapped_page = __iommu_unmap_pages(domain, iova, size - unmapped,
> +						    iotlb_gather);

I think it would make more sense to restructure the basic function 
around handling a page range, then just have a little inner loop to 
iterate over the individual pages if the driver doesn't provide the new 
callback.

Robin.

>   		if (!unmapped_page)
>   			break;
>   
>