RE: [PATCH rdma-rc 3/3] IB/core: Fix ODP with IB_ACCESS_HUGETLB handling

[Artemy Kovalyov <artemyko@mellanox.com>]

-----Original Message-----
From: Jason Gunthorpe <jgg@ziepe.ca> 
Sent: Thursday, December 19, 2019 8:33 PM
To: Leon Romanovsky <leon@kernel.org>
Cc: Doug Ledford <dledford@redhat.com>; Leon Romanovsky <leonro@mellanox.com>; RDMA mailing list <linux-rdma@vger.kernel.org>; Artemy Kovalyov <artemyko@mellanox.com>; Aviad Yehezkel <aviadye@mellanox.com>; Yishai Hadas <yishaih@mellanox.com>
Subject: Re: [PATCH rdma-rc 3/3] IB/core: Fix ODP with IB_ACCESS_HUGETLB handling

On Thu, Dec 19, 2019 at 03:46:46PM +0200, Leon Romanovsky wrote:

> @@ -403,6 +390,7 @@ int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
>  	int j, k, ret = 0, start_idx, npages = 0;
>  	unsigned int flags = 0, page_shift;
>  	phys_addr_t p = 0;
> +	struct vm_area_struct **vmas;
>  
>  	if (access_mask == 0)
>  		return -EINVAL;
> @@ -415,6 +403,12 @@ int ib_umem_odp_map_dma_pages(struct ib_umem_odp *umem_odp, u64 user_virt,
>  	if (!local_page_list)
>  		return -ENOMEM;
>  
> +	vmas = (struct vm_area_struct **)__get_free_page(GFP_KERNEL);
> +	if (!vmas) {
> +		ret = -ENOMEM;
> +		goto out_free_page_list;
> +	}

I'd rather not do this on the fast path

> +			if ((1 << page_shift) > vma_kernel_pagesize(vmas[j])) {
> +				ret = -EFAULT;
> +				break;
> +			}

And vma's cannot be de-refenced outside the mmap_sem

There is already logic checking for linear contiguous pages:

                        if (user_virt & ~page_mask) {
                                p += PAGE_SIZE;
                                if (page_to_phys(local_page_list[j]) != p) {
                                        ret = -EFAULT;
                                        break;
                                }

Why do we need to add the vma check?

AK: checking for linear contiguity may be not enough, page may be transparent huge page, so in addition we also ensure its indeed persistent.

Jason