Re: [PATCH v3 23/32] vfio: use iova_to_phys_length for efficient unmap

[Jason Gunthorpe]

On Wed, Jun 03, 2026 at 11:17:55PM +0800, Guanghui Feng wrote:
> Use iommu_iova_to_phys_length() to get PTE page size, allowing
> traversal by actual mapping granularity instead of PAGE_SIZE steps.
> 
> Signed-off-by: Guanghui Feng <guanghuifeng@xxxxxxxxxxxxxxxxx>
> Acked-by: Shiqiang Zhang <shiyu.zsq@xxxxxxxxxxxxxxxxx>
> Acked-by: Simon Guo <wei.guo.simon@xxxxxxxxxxxxxxxxx>
> ---
>  drivers/vfio/vfio_iommu_type1.c | 27 ++++++++++++++++++++++-----
>  1 file changed, 22 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> index c8151ba54de3..115d88d7003e 100644
> --- a/drivers/vfio/vfio_iommu_type1.c
> +++ b/drivers/vfio/vfio_iommu_type1.c
> @@ -1177,25 +1177,42 @@ static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
>  
>  	iommu_iotlb_gather_init(&iotlb_gather);
>  	while (pos < dma->size) {
> -		size_t unmapped, len;
> +		size_t unmapped, len, pgsize;
>  		phys_addr_t phys, next;
>  		dma_addr_t iova = dma->iova + pos;
>  
> -		phys = iommu_iova_to_phys(domain->domain, iova);
> -		if (WARN_ON(!phys)) {
> +		/* Single page table walk returns both phys and PTE size */
> +		phys = iommu_iova_to_phys_length(domain->domain, iova,
> +						  &pgsize);
> +		if (WARN_ON(phys == PHYS_ADDR_MAX)) {
>  			pos += PAGE_SIZE;
>  			continue;
>  		}
> +		if (WARN_ON(!pgsize || pgsize < PAGE_SIZE))
> +			pgsize = PAGE_SIZE;
>  
>  		/*
>  		 * To optimize for fewer iommu_unmap() calls, each of which
>  		 * may require hardware cache flushing, try to find the
>  		 * largest contiguous physical memory chunk to unmap.
> +		 *
> +		 * mapped_length already accounts for contiguous entries
> +		 * from iova, then try to join following physically
> +		 * contiguous PTEs.
>  		 */
> -		for (len = PAGE_SIZE; pos + len < dma->size; len += PAGE_SIZE) {
> -			next = iommu_iova_to_phys(domain->domain, iova + len);
> +		len = min_t(size_t, pgsize, dma->size - pos);
> +		for (; pos + len < dma->size; ) {
> +			size_t next_pgsize;
> +
> +			next = iommu_iova_to_phys_length(domain->domain,
> +							  iova + len,
> +							  &next_pgsize);

vfio should not be calling it twice, the core code needs to give the
best length as efficiently as it can. not open coding this in callers.

I think I've said this three times now

Jason