[Intel-gfx] Building GVT-g as a sub-module of i915

[Wang, Zhi A]

Thanks Daniel and Joonas. : p See my replies below.

> -----Original Message-----
> From: daniel.vetter at ffwll.ch [mailto:daniel.vetter at ffwll.ch] On Behalf Of
> Daniel Vetter
> Sent: Monday, May 23, 2016 5:18 PM
> To: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> Cc: Wang, Zhi A <zhi.a.wang at intel.com>; Vetter, Daniel
> <daniel.vetter at intel.com>; intel-gfx at lists.freedesktop.org
> Subject: Re: [Intel-gfx] Building GVT-g as a sub-module of i915
> 
> On Mon, May 23, 2016 at 4:16 PM, Joonas Lahtinen
> <joonas.lahtinen at linux.intel.com> wrote:
> > On ma, 2016-05-23 at 07:03 +0000, Wang, Zhi A wrote:
> >> Hi Guys:
> >>   I'm trying to make GVT-g as a sub-module of i915 in the next
> >> version patchset. The basic idea is to introduce a "gvt-g pre-enabled
> >> state" in i915. I think it should be a kernel option.
> >>
> >
> > Could not the GGTT partitioning be done ad hoc by moving objects out
> > of the memory areas to be ballooned? This way gvt module could just be
> > loaded and it would work, instead of having to reboot and change
> > kernel parameters.
> 
> Yeah, if we want to make gvt loadable, then it should indeed not reserve
> anything if it's not loaded. Otherwise there's no point in that option, and it's no
> better than just a static Kconfig+ maybe i915 module option.
> 
> If dynamic loading is too hard for v1, then I'd say we should postpone it to
> post-merging. GVT-g is already tricky to integrate as-is.
> 

Yes. We can try to reserve some portion of GGTT by allocating 2 gem object and pin them into mappable / high GGTT memory.

I think better we can postpone it to post-merging. For now statically partition only requires little changes. :)

> >> When this kernel option is enabled by user, i915 will do GGTT
> >> partition and save HW initial MMIO snapshot for gvt-g module during
> >> loading.
> >
> > Like discussed in the F2F, I really think taking a MMIO snapshot in
> > Dom0 at boot sounds a little suspicious to me as changing Dom0 BIOS
> > settings could very obscurely break VM booting, especially if
> > migration is at some point wanted. It will also leak the Dom0 boot
> > state to a VM, which I do not like either.
> >
> > I would be more comfortable if the VMs are booting to a driver-fixed
> > MMIO state.
> >
> > Any thoughts by others on these?
> 
> Golden MMIO state sounds like a good idea.

Yes. It's a good idea. I agree with that. But consider that each GT stepping/generation might need a dedicated MMIO snapshot. They could become huge, probably at the very beginning 2MB is OK, in future we should get them out of kernel.

BTW: It's not only MMIO bar needs a snapshot, PCI configuration space also needs a golden snapshot. :) So consider the GT/SKUs we have, maybe we should figure out a way to store them at first. :)

> -Daniel
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch