Re: [PATCH 6/9] iommu/amd: Implement map_atomic

[Robin Murphy]

On 11/04/2019 19:47, Tom Murphy wrote:
> Instead of using a spin lock I removed the mutex lock from both the
> amd_iommu_map and amd_iommu_unmap path as well. iommu_map doesnât lock
> while mapping and so if iommu_map is called by two different threads on
> the same iova region it results in a race condition even with the locks.
> So the locking in amd_iommu_map and amd_iommu_unmap doesn't add any real
> protection. The solution to this is for whatever manages the allocated
> iovaâs externally to make sure iommu_map isnât called twice on the same
> region at the same time.

Note that that assumption is not necessarily sufficient - even with 
correct address space management you can have cases like two threads 
mapping adjacent pages, where even thought they are targeting different 
PTEs they can race to install/modify intermediate levels of the 
pagetable. I believe AMD is actually OK in that regard, but some drivers 
*are* relying on locking for correctness so it can't just be 
unequivocally removed everywhere.

Robin.

> Signed-off-by: Tom Murphy <tmurphy@xxxxxxxxxx>
> ---
>   drivers/iommu/amd_iommu.c | 25 ++++++++++++++++++-------
>   1 file changed, 18 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/iommu/amd_iommu.c b/drivers/iommu/amd_iommu.c
> index 2d4ee10626b4..b45e0e033adc 100644
> --- a/drivers/iommu/amd_iommu.c
> +++ b/drivers/iommu/amd_iommu.c
> @@ -3089,12 +3089,12 @@ static int amd_iommu_attach_device(struct iommu_domain *dom,
>   	return ret;
>   }
>   
> -static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
> -			 phys_addr_t paddr, size_t page_size, int iommu_prot)
> +static int __amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
> +			 phys_addr_t paddr, size_t page_size, int iommu_prot,
> +			 gfp_t gfp)
>   {
>   	struct protection_domain *domain = to_pdomain(dom);
>   	int prot = 0;
> -	int ret;
>   
>   	if (domain->mode == PAGE_MODE_NONE)
>   		return -EINVAL;
> @@ -3104,11 +3104,21 @@ static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
>   	if (iommu_prot & IOMMU_WRITE)
>   		prot |= IOMMU_PROT_IW;
>   
> -	mutex_lock(&domain->api_lock);
> -	ret = iommu_map_page(domain, iova, paddr, page_size, prot, GFP_KERNEL);
> -	mutex_unlock(&domain->api_lock);
> +	return iommu_map_page(domain, iova, paddr, page_size, prot, gfp);
> +}
>   
> -	return ret;
> +static int amd_iommu_map(struct iommu_domain *dom, unsigned long iova,
> +			 phys_addr_t paddr, size_t page_size, int iommu_prot)
> +{
> +	return __amd_iommu_map(dom, iova, paddr, page_size, iommu_prot,
> +			GFP_KERNEL);
> +}
> +
> +static int amd_iommu_map_atomic(struct iommu_domain *dom, unsigned long iova,
> +			 phys_addr_t paddr, size_t page_size, int iommu_prot)
> +{
> +	return __amd_iommu_map(dom, iova, paddr, page_size, iommu_prot,
> +			GFP_ATOMIC);
>   }
>   
>   static size_t amd_iommu_unmap(struct iommu_domain *dom, unsigned long iova,
> @@ -3262,6 +3272,7 @@ const struct iommu_ops amd_iommu_ops = {
>   	.attach_dev = amd_iommu_attach_device,
>   	.detach_dev = amd_iommu_detach_device,
>   	.map = amd_iommu_map,
> +	.map_atomic = amd_iommu_map_atomic,
>   	.unmap = amd_iommu_unmap,
>   	.iova_to_phys = amd_iommu_iova_to_phys,
>   	.add_device = amd_iommu_add_device,