[Intel-gfx] [PATCH 3/7] drm/i915: Always enable execlists on BDW for vgpu

[Joonas Lahtinen]

Hi,

On to, 2015-08-20 at 17:40 +0800, Zhiyuan Lv wrote:
> On Thu, Aug 20, 2015 at 10:22:37AM +0100, Chris Wilson wrote:
> > On Thu, Aug 20, 2015 at 04:55:08PM +0800, Zhiyuan Lv wrote:
> > > Hi Chris,
> > > 
> > > On Thu, Aug 20, 2015 at 09:34:05AM +0100, Chris Wilson wrote:
> > > > On Thu, Aug 20, 2015 at 03:45:20PM +0800, Zhiyuan Lv wrote:
> > > > > Broadwell hardware supports both ring buffer mode and 
> > > > > execlist mode.
> > > > > When i915 runs inside a VM with Intel GVT-g, we allow 
> > > > > execlist mode
> > > > > only. The reason is that GVT-g does not support the dynamic 
> > > > > mode
> > > > > switch between ring buffer mode and execlist mode when 
> > > > > running
> > > > > multiple virtual machines. Consider that ring buffer mode is 
> > > > > legacy
> > > > > mode, it makes sense to drop it inside virtual machines.
> > > > 
> > > > If that is the case, you should query the host as to what mode 
> > > > it is
> > > > running.
> > > 
> > > Thanks for the reply! You mean we query the host mode, then tell 
> > > the
> > > guest driver inside VM, so that it could use the same mode as 
> > > host
> > > right? That might be a little complicated, and the only benefit 
> > > is to
> > > support legacy ring buffer mode ...
> > 
> > The only benefit being that the guest works no matter what the host
> > does?
> 
> Supporting ring buffer mode may need more work in GVT-g. When we 
> started to
> enable BDW support, ring buffer mode used to work but was not well 
> tested. And
> the inter-VM switch (all in ringbuffer mode) may be tricker comparing 
> with
> driver context switch with "MI_SET_CONTEXT", because we need to 
> switch ring
> buffer whereas driver does not. Regarding this, the EXECLIST mode 
> looks
> cleaner. In order to support that, we may have to: 1, change more LRI 
> commands
> to MMIO in current driver; 2, more testing/debugging of inter-VM 
> context
> switch flow.
> 
> Based on that, I think we should really make statement that "ring 
> buffer mode"
> is not supported by GVT-g on BDW :-)
> 

I think just move the vpgu test even before test for GEN9 just making
it return true on intel_vgpu_active(dev) and add a comment that
currently vGPU only supports execlist command submission, and then add
an error early in the init in some appropriate spot if
intel_vgpu_active is true but logical ring context are not available
(which would practically mean GEN < 8).

Does that sound OK?

Regards, Joonas

> > -Chris
> >