Re: [PATCH 1/2] x86/io: add interface to reserve io memtype for a resource range. (v1.1)

From: Daniel Vetter
Date: Wed Oct 26 2016 - 01:49:27 EST


On Tue, Oct 25, 2016 at 07:31:29PM +0200, Luis R. Rodriguez wrote:
> On Mon, Oct 24, 2016 at 04:31:45PM +1000, Dave Airlie wrote:
> > A recent change to the mm code in:
> > 87744ab3832b83ba71b931f86f9cfdb000d07da5
> > mm: fix cache mode tracking in vm_insert_mixed()
> >
> > started enforcing checking the memory type against the registered list for
> > amixed pfn insertion mappings. It happens that the drm drivers for a number
> > of gpus relied on this being broken. Currently the driver only inserted
> > VRAM mappings into the tracking table when they came from the kernel,
> > and userspace mappings never landed in the table. This led to a regression
> > where all the mapping end up as UC instead of WC now.
>
> Eek.
>
> > I've considered a number of solutions but since this needs to be fixed
> > in fixes and not next, and some of the solutions were going to introduce
> > overhead that hadn't been there before I didn't consider them viable at
> > this stage. These mainly concerned hooking into the TTM io reserve APIs,
> > but these API have a bunch of fast paths I didn't want to unwind to add
> > this to.
> >
> > The solution I've decided on is to add a new API like the arch_phys_wc
> > APIs (these would have worked but wc_del didn't take a range), and
> > use them from the drivers to add a WC compatible mapping to the table
> > for all VRAM on those GPUs. This means we can then create userspace
> > mapping that won't get degraded to UC.
>
> Is anything on a driver to be able to tell when this is actually needed ?
> How will driver developers know? Can you add a bit of documentation to
> the API? If its transitive towards a secondary solution indicating so
> would help driver developers.

I'll plug the io-mapping stuff again here, and more specifically the
userspace pte wrangling stuff we've added in 4.9 to i915_mm.c. Should
probably move that one to the core. That way io_mapping takes care of the
full reservartion, and allows you to on-demand kmap (for kernel) and write
ptes. All nicely fast and all, and for bonus, also nicely encapsulated.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch