[PATCH 7/8] iommu/intel: Support the gfp argument to the map_pages op

[Tian, Kevin]

> From: Jason Gunthorpe <jgg at nvidia.com>
> Sent: Saturday, January 7, 2023 12:43 AM
> 
> @@ -2368,7 +2372,7 @@ static int iommu_domain_identity_map(struct
> dmar_domain *domain,
> 
>  	return __domain_mapping(domain, first_vpfn,
>  				first_vpfn, last_vpfn - first_vpfn + 1,
> -				DMA_PTE_READ|DMA_PTE_WRITE);
> +				DMA_PTE_READ|DMA_PTE_WRITE,
> GFP_KERNEL);
>  }

Baolu, can you help confirm whether switching from GFP_ATOMIC to
GFP_KERNEL is OK in this path? it looks fine to me in a quick glance
but want to be conservative here.

> @@ -4333,7 +4337,8 @@ static size_t intel_iommu_unmap(struct
> iommu_domain *domain,
> 
>  	/* Cope with horrid API which requires us to unmap more than the
>  	   size argument if it happens to be a large-page mapping. */
> -	BUG_ON(!pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> &level));
> +	BUG_ON(!pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> &level,
> +			       GFP_ATOMIC));

with level==0 it implies it's only lookup w/o pgtable allocation. From this
angle it reads better to use a more relaxed gfp e.g. GFP_KERNEL here.

> @@ -4392,7 +4397,8 @@ static phys_addr_t
> intel_iommu_iova_to_phys(struct iommu_domain *domain,
>  	int level = 0;
>  	u64 phys = 0;
> 
> -	pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> &level);
> +	pte = pfn_to_dma_pte(dmar_domain, iova >> VTD_PAGE_SHIFT,
> &level,
> +			     GFP_ATOMIC);

ditto