userptr support in drm drivers

[Jean Delvare]

Hi Christian,

On Thu, 18 Feb 2016 10:48:18 +0100, Christian König wrote:
> Am 18.02.2016 um 09:59 schrieb Jean Delvare:
> > Maybe I was not clear enough in my original post, but I am really
> > advocating for FEWER kconfig options, not more. Plus I just explained
> > why I think the radeon and amdgpu drivers do it wrong, so I'm certainly
> > not going to align i915 on that.
> >
> >> Distro kernels pretty much all select MMU_NOTIFIER already for unrelated
> >> reasons, but it's good to be less suprising for everyone who builds their
> >> own custom kernel.
> > The current situation is actually very surprising and this is what I
> > would like to change. The options of one driver depending on choices
> > made somewhere else is utterly confusing. CONFIG_RADEON_USERPTR and
> > CONFIG_AMDGPU_USERPTR give the user a little more control where i915
> > gives none, but that's only in one direction (you can force full
> > userptr support in, but you can't force it out.)
> 
> As I said, that is perfectly fine and exactly what we want.

I don't know what "we" represent in your sentence, but I doubt a
majority of users want that.

> > I'm still waiting for someone to explain why we can't just
> > unconditionally select MMU_NOTIFIER in these 3 drivers and be done with
> > it. That's less work at kernel configuration time AND fewer
> > preprocessing conditionals within the code, so easier/better testing
> > coverage and more readable code. We would need a very good reason for
> > NOT doing it.
> 
> Adding the MMU Notifier causes overhead in the MM which people want to 
> avoid when they don't need userptr support.
>
> We used to select MMU_NOTIFIER unconditionally before and explicitly 
> added this option on user request. So yes it is clearly necessary.

Thanks for the explanation.

Can you tell me more about the overhead? Is it actually measurable?
Which people need this degree of optimization?

Are the i915, radeon or amdgpu drm drivers used on embedded systems?

Just because people asked for something doesn't mean that implementing
it is a good idea. Keeping things simple is a valid goal on its own.
Anyone could ask for features they don't need to become options they
can disable, but in the long run this approach can't be sustained.

> The additional userptr support code in the drivers are negligible so 
> always enabling the code when the MMU Notifier is available is also 
> perfectly fine.

Indeed this is a small amount of code. OTOH you have ifdefs in place
anyway. Using the right ones would turn the option into a real option,
i.e. the user decides to enable a feature or not. Not a "you may or may
not get the feature" option as we have at the moment. More control to
the user and less confusion.

> So if you want to clean it up make a common DRM option which selects MMU 
> Notifier if it isn't already selected.

I will, as it will still be an improvement compared to the current
situation. But I still believe this is confusing in its current form.

If a minority of users need to opt out from MMU_NOTIFIER, then
DRM_RADEON_USERPTR and friends should default to y and be hidden behind
EXPERT. That way you avoid asking yet more questions to everyone else.
Would that work for you?

-- 
Jean Delvare
SUSE L3 Support