[PATCH 2/2 v3] drm/exynos: added userptr feature.

[KOSAKI Motohiro]

On Fri, May 11, 2012 at 7:29 PM, Jerome Glisse <j.glisse at gmail.com> wrote:
> On Fri, May 11, 2012 at 6:59 PM, KOSAKI Motohiro
> <kosaki.motohiro at gmail.com> wrote:
>>> My point is this ioctl will be restricted to one user (Xserver if i
>>> understand) and only this user, there is no fork in it so no need to
>>> worry about fork, just setting the vma as locked will be enough.
>>>
>>> But i don't want people reading this driver suddenly think that what
>>> it's doing is ok, it's not, it's hack and can never make to work
>>> properly on a general case, that's why it needs a big comment stating,
>>> stressing that. I just wanted to make sure Inki and Kyungmin
>>> understood that this kind of ioctl should be restricted to carefully
>>> selected user and that there is no way to make it general or reliable
>>> outside that.
>>
>>
>> first off, I'm not drm guy and then I don't intend to insist you. but if
>> application don't use fork, get_user_pages() has no downside. I guess we
>> don't need VM_LOCKED hack.
>>
>> but again, up to drm folks.
>
> You need the VM_LOCKED hack to mare sure that the xorg vma still point
> to the same page, afaict with get_user_pages pages can be migrated out
> of the anonymous vma so the vma might point to new page, while old
> page are still in use by the gpu and not recycle until their refcount
> drop to 0.

afaik, get_user_pages() prevent page migration. (see
migrate_page_move_mapping). but mlock doesn't.