Re: [PATCH] iommu: Improve the performance for direct_mapping

[Will Deacon]

On Fri, Nov 20, 2020 at 05:06:28PM +0800, Yong Wu wrote:
> Currently direct_mapping always use the smallest pgsize which is SZ_4K
> normally to mapping. This is unnecessary. we could gather the size, and
> call iommu_map then, iommu_map could decide how to map better with the
> just right pgsize.
> 
> From the original comment, we should take care overlap, otherwise,
> iommu_map may return -EEXIST. In this overlap case, we should map the
> previous region before overlap firstly. then map the left part.
> 
> Each a iommu device will call this direct_mapping when its iommu
> initialize, This patch is effective to improve the boot/initialization
> time especially while it only needs level 1 mapping.
> 
> Signed-off-by: Anan Sun <anan.sun@xxxxxxxxxxxx>
> Signed-off-by: Yong Wu <yong.wu@xxxxxxxxxxxx>
> ---
>  drivers/iommu/iommu.c | 20 ++++++++++++++++++--
>  1 file changed, 18 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> index df87c8e825f7..854a8fcb928d 100644
> --- a/drivers/iommu/iommu.c
> +++ b/drivers/iommu/iommu.c
> @@ -737,6 +737,7 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
>  	/* We need to consider overlapping regions for different devices */
>  	list_for_each_entry(entry, &mappings, list) {
>  		dma_addr_t start, end, addr;
> +		size_t unmapped_sz = 0;

I think "unmapped" is the wrong word here, as this variable actually
represents the amount we want to map! I suggest "map_size" instead.

>  		if (domain->ops->apply_resv_region)
>  			domain->ops->apply_resv_region(dev, domain, entry);
> @@ -752,10 +753,25 @@ static int iommu_create_device_direct_mappings(struct iommu_group *group,
>  			phys_addr_t phys_addr;
>  
>  			phys_addr = iommu_iova_to_phys(domain, addr);
> -			if (phys_addr)
> +			if (phys_addr == 0) {
> +				unmapped_sz += pg_size; /* Gather the size. */
>  				continue;
> +			}
>  
> -			ret = iommu_map(domain, addr, addr, pg_size, entry->prot);
> +			if (unmapped_sz) {
> +				/* Map the region before the overlap. */
> +				ret = iommu_map(domain, start, start,
> +						unmapped_sz, entry->prot);
> +				if (ret)
> +					goto out;
> +				start += unmapped_sz;

I think it's a bit confusing to update start like this. Can we call
iommu_map(domain, addr - map_size, addr - map_size, map_size, entry->prot)
instead?

> +				unmapped_sz = 0;
> +			}
> +			start += pg_size;
> +		}
> +		if (unmapped_sz) {
> +			ret = iommu_map(domain, start, start, unmapped_sz,
> +					entry->prot);

Can you avoid this hunk by changing your loop check to something like:

	if (!phys_addr) {
		map_size += pg_size;
		if (addr + pg_size < end)
			continue;
	}

Will