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

[Thomas Hellström (Intel)]

On 3/23/21 3:00 PM, Jason Gunthorpe wrote:
> 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..

It's actually used all over the place. Both in drivers and also 
redefined with
CONFIG_MEM_SOFT_DIRTY which makes me think Daniels idea of 
vma_is_cow_mapping() is better since it won't clash and cause 
compilation failures...

>
> Can the user mprotect the write back on with the above logic?
No, it's blocked by mprotect.
> Do we
> need VM_DENYWRITE too?

Seems tied to MAP_DENYWRITE which is nowadays ignored according to man 
mmap().

Thanks,

Thomas

>
> Jason