[Intel-gfx] [PATCH v3 00/10] Update vfio_pin/unpin_pages API

[Alex Williamson]

On Fri, 22 Jul 2022 17:38:25 -0700
Nicolin Chen <nicolinc at nvidia.com> wrote:

> On Fri, Jul 22, 2022 at 06:18:00PM -0600, Alex Williamson wrote:
> > External email: Use caution opening links or attachments
> > 
> > 
> > On Fri, 22 Jul 2022 16:12:19 -0700
> > Nicolin Chen <nicolinc at nvidia.com> wrote:
> >   
> > > On Fri, Jul 22, 2022 at 04:11:29PM -0600, Alex Williamson wrote:
> > >  
> > > > GVT-g explodes for me with this series on my Broadwell test system,
> > > > continuously spewing the following:  
> > >
> > > Thank you for running additional tests.
> > >  
> > > > [   47.348778] WARNING: CPU: 3 PID: 501 at drivers/vfio/vfio_iommu_type1.c:978 vfio_iommu_type1_unpin_pages+0x7b/0x100 [vfio_iommu_type1]  
> > >  
> > > > Line 978 is the WARN_ON(i != npage) line.  For the cases where we don't
> > > > find a matching vfio_dma, I'm seeing addresses that look maybe like
> > > > we're shifting  a value that's already an iova by PAGE_SHIFT somewhere.  
> > >
> > > Hmm..I don't understand the PAGE_SHIFT part. Do you mind clarifying?  
> > 
> > The iova was a very large address for a 4GB VM with a lot of zeros on
> > the low order bits, ex. 0x162459000000.  Thanks,  
> 
> Ah! Thanks for the hint. The following commit did a double shifting:
>    "vfio: Pass in starting IOVA to vfio_pin/unpin_pages AP"
> 
> And the following change should fix:
> -------------------
> diff --git a/drivers/gpu/drm/i915/gvt/kvmgt.c b/drivers/gpu/drm/i915/gvt/kvmgt.c
> index 481dd2aeb40e..4790c7f35b88 100644
> --- a/drivers/gpu/drm/i915/gvt/kvmgt.c
> +++ b/drivers/gpu/drm/i915/gvt/kvmgt.c
> @@ -293,7 +293,7 @@ static int gvt_dma_map_page(struct intel_vgpu *vgpu, unsigned long gfn,
>         if (dma_mapping_error(dev, *dma_addr)) {
>                 gvt_vgpu_err("DMA mapping failed for pfn 0x%lx, ret %d\n",
>                              page_to_pfn(page), ret);
> -               gvt_unpin_guest_page(vgpu, gfn << PAGE_SHIFT, size);
> +               gvt_unpin_guest_page(vgpu, gfn, size);
>                 return -ENOMEM;
>         }
> 
> @@ -306,7 +306,7 @@ static void gvt_dma_unmap_page(struct intel_vgpu *vgpu, unsigned long gfn,
>         struct device *dev = vgpu->gvt->gt->i915->drm.dev;
> 
>         dma_unmap_page(dev, dma_addr, size, DMA_BIDIRECTIONAL);
> -       gvt_unpin_guest_page(vgpu, gfn << PAGE_SHIFT, size);
> +       gvt_unpin_guest_page(vgpu, gfn, size);
>  }
> 
>  static struct gvt_dma *__gvt_cache_find_dma_addr(struct intel_vgpu *vgpu,
> -------------------

Looks likely.  Not sure how Terrance was able to test this successfully
though.

> So, I think that I should send a v4, given that the patches aren't
> officially applied?

Yep, please rebase on current vfio next branch.  Thanks,

Alex