[RFC PATCH 2/2] mm,drm/ttm: Use VM_PFNMAP for TTM vmas

[Jason Gunthorpe]

On Sun, Mar 21, 2021 at 07:45:29PM +0100, Thomas Hellström (Intel) wrote:
> To block fast gup we need to make sure TTM ptes are always special.
> With MIXEDMAP we, on architectures that don't support pte_special,
> insert normal ptes, but OTOH on those architectures, fast is not
> supported.
> At the same time, the function documentation to vm_normal_page() suggests
> that ptes pointing to system memory pages of MIXEDMAP vmas are always
> normal, but that doesn't seem consistent with what's implemented in
> vmf_insert_mixed(). I'm thus not entirely sure this patch is actually
> needed.
> 
> But to make sure and to avoid also normal (non-fast) gup, make all
> TTM vmas PFNMAP. With PFNMAP we can't allow COW mappings
> anymore so make is_cow_mapping() available and use it to reject
> COW mappigs at mmap time.
> 
> There was previously a comment in the code that WC mappings together
> with x86 PAT + PFNMAP was bad for performance. However from looking at
> vmf_insert_mixed() it looks like in the current code PFNMAP and MIXEDMAP
> are handled the same for architectures that support pte_special. This
> means there should not be a performance difference anymore, but this
> needs to be verified.
> 
> Cc: Christian Koenig <christian.koenig at amd.com>
> Cc: David Airlie <airlied at linux.ie>
> Cc: Daniel Vetter <daniel at ffwll.ch>
> Cc: Andrew Morton <akpm at linux-foundation.org>
> Cc: Jason Gunthorpe <jgg at nvidia.com>
> Cc: linux-mm at kvack.org
> Cc: dri-devel at lists.freedesktop.org
> Cc: linux-kernel at vger.kernel.org
> Signed-off-by: Thomas Hellström (Intel) <thomas_os at shipmail.org>
>  drivers/gpu/drm/ttm/ttm_bo_vm.c | 22 ++++++++--------------
>  include/linux/mm.h              |  5 +++++
>  mm/internal.h                   |  5 -----
>  3 files changed, 13 insertions(+), 19 deletions(-)
> 
> diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c
> index 1c34983480e5..708c6fb9be81 100644
> +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
> @@ -372,12 +372,7 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
>  		 * at arbitrary times while the data is mmap'ed.
>  		 * See vmf_insert_mixed_prot() for a discussion.
>  		 */
> -		if (vma->vm_flags & VM_MIXEDMAP)
> -			ret = vmf_insert_mixed_prot(vma, address,
> -						    __pfn_to_pfn_t(pfn, PFN_DEV),
> -						    prot);
> -		else
> -			ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
> +		ret = vmf_insert_pfn_prot(vma, address, pfn, prot);
>  
>  		/* Never error on prefaulted PTEs */
>  		if (unlikely((ret & VM_FAULT_ERROR))) {
> @@ -555,18 +550,14 @@ static void ttm_bo_mmap_vma_setup(struct ttm_buffer_object *bo, struct vm_area_s
>  	 * Note: We're transferring the bo reference to
>  	 * vma->vm_private_data here.
>  	 */
> -
>  	vma->vm_private_data = bo;
>  
>  	/*
> -	 * We'd like to use VM_PFNMAP on shared mappings, where
> -	 * (vma->vm_flags & VM_SHARED) != 0, for performance reasons,
> -	 * but for some reason VM_PFNMAP + x86 PAT + write-combine is very
> -	 * bad for performance. Until that has been sorted out, use
> -	 * VM_MIXEDMAP on all mappings. See freedesktop.org bug #75719
> +	 * PFNMAP forces us to block COW mappings in mmap(),
> +	 * and with MIXEDMAP we would incorrectly allow fast gup
> +	 * on TTM memory on architectures that don't have pte_special.
>  	 */
> -	vma->vm_flags |= VM_MIXEDMAP;
> -	vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
> +	vma->vm_flags |= VM_PFNMAP | VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
>  }
>  
>  int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
> @@ -579,6 +570,9 @@ int ttm_bo_mmap(struct file *filp, struct vm_area_struct *vma,
>  	if (unlikely(vma->vm_pgoff < DRM_FILE_PAGE_OFFSET_START))
>  		return -EINVAL;
>  
> +	if (unlikely(is_cow_mapping(vma->vm_flags)))
> +		return -EINVAL;
> +
>  	bo = ttm_bo_vm_lookup(bdev, vma->vm_pgoff, vma_pages(vma));
>  	if (unlikely(!bo))
>  		return -EINVAL;
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 77e64e3eac80..c6ebf7f9ddbb 100644
> +++ b/include/linux/mm.h
> @@ -686,6 +686,11 @@ static inline bool vma_is_accessible(struct vm_area_struct *vma)
>  	return vma->vm_flags & VM_ACCESS_FLAGS;
>  }
>  
> +static inline bool is_cow_mapping(vm_flags_t flags)
> +{
> +	return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
> +}

Most driver places are just banning VM_SHARED.

I see you copied this from remap_pfn_range(), but that logic is so
special I'm not sure..

Can the user mprotect the write back on with the above logic? Do we
need VM_DENYWRITE too?

Jason