Re: [PATCH v4 01/10] vt-d: fix the IOMMU flush issue

["Xu, Quan" <quan.xu@intel.com>]

On May 12, 2016 4:53 PM, Jan Beulich <JBeulich@suse.com> wrote:
> >>> On 12.05.16 at 09:50, <quan.xu@intel.com> wrote:
> > On May 10, 2016 12:10 AM, Jan Beulich <JBeulich@suse.com> wrote:
> >> >>> On 06.05.16 at 10:54, <quan.xu@intel.com> wrote:
> >> > -static void intel_iommu_iotlb_flush(struct domain *d, unsigned
> >> > long gfn, unsigned int page_count)
> >> > +static void iommu_flush_iotlb_page(struct domain *d, unsigned long
> gfn,
> >> > +                                   unsigned int page_count)
> >>
> >> The new name suggests just one page. Please use e.g.
> >> iommu_flush_iotlb_pages() instead.
> >>
> >
> > Make sense.
> >
> >> >  {
> >> > -    __intel_iommu_iotlb_flush(d, gfn, 1, page_count);
> >> > +    iommu_flush_iotlb(d, gfn, 1, page_count);
> >> >  }
> >>
> >> But of course the question is whether having this wrapper is useful
> >> in the first place,
> >
> >
> > This wrapper assumes the 'dma_old_pte_present' is '1', but in another
> > caller intel_iommu_map_page(), i.e.
> >
> >
> >      intel_iommu_map_page()
> >     {
> >        ...
> >              if ( !this_cpu(iommu_dont_flush_iotlb) )
> >                   iommu_flush_iotlb(d, gfn, dma_pte_present(old), 1);
> >        ...
> >     }
> >
> >
> > the 'dma_old_pte_present' is not sure.
> 
> I'm sorry, but you're looking at this backwards: I suggested to remove the
> wrapper, not to move any check into iommu_flush_iotlb().
> Removing the wrapper simply means to move the passing of the hard coded 1
> into the current callers of that wrapper.
> 

A little bit confused.
Check one thing, do the wrappers refer to iommu_flush_iotlb_page() and iommu_flush_iotlb_all() ?

If yes, we can't ignore another thing:

These two wrappers are also initialized for 2 .callbacks at the bottom of this file:

....
    .iotlb_flush = iommu_flush_iotlb_pages,
    .iotlb_flush_all = iommu_flush_iotlb_all,
....




> >> > @@ -1391,13 +1399,19 @@ int domain_context_mapping_one(
> >> >      spin_unlock(&iommu->lock);
> >> >
> >> >      /* Context entry was previously non-present (with domid 0). */
> >> > -    if ( iommu_flush_context_device(iommu, 0, (((u16)bus) << 8) | devfn,
> >> > -                                    DMA_CCMD_MASK_NOBIT, 1) )
> >> > -        iommu_flush_write_buffer(iommu);
> >> > -    else
> >> > +    rc = iommu_flush_context_device(iommu, 0, (((u16)bus) << 8) |
> devfn,
> >> > +                                    DMA_CCMD_MASK_NOBIT, 1);
> >> > +
> >> > +    if ( !rc )
> >> >      {
> >> >          int flush_dev_iotlb = find_ats_dev_drhd(iommu) ? 1 : 0;
> >> > -        iommu_flush_iotlb_dsi(iommu, 0, 1, flush_dev_iotlb);
> >> > +        rc = iommu_flush_iotlb_dsi(iommu, 0, 1, flush_dev_iotlb);
> >>
> >> Please take the opportunity and add the missing blank line (between
> >> declaration(s) and statement(s) in cases like this.
> >>
> >> > +    }
> >> > +
> >> > +    if ( rc > 0 )
> >>
> >> Can iommu_flush_context_device() return a positive value? If so, the
> >> logic is now likely wrong. If not (which is what I assume) I'd like
> >> to suggest adding a respective ASSERT() (even if only to document the
> >> fact). Or alternatively this
> >> if() could move into the immediately preceding one.
> >
> > Check it again. iommu_flush_context_device() can return a positive value.
> > [...]
> > Could you tell me why the logic is now likely wrong? I will fix it first.
> 
> With
> 
>     rc = iommu_flush_context_device(iommu, 0, (((u16)bus) << 8) | devfn,
>                                     DMA_CCMD_MASK_NOBIT, 1);
> 
>     if ( !rc )
>     {
>         int flush_dev_iotlb = find_ats_dev_drhd(iommu) ? 1 : 0;
>         rc = iommu_flush_iotlb_dsi(iommu, 0, 1, flush_dev_iotlb);
>     }
> 
>     if ( rc > 0 )
>     {
>         iommu_flush_write_buffer(iommu);
>         rc = 0;
>     }
> 
> it seems pretty clear that you won't call iommu_flush_iotlb_dsi() if
> iommu_flush_context_device() returned 1, which doesn't look like what is
> wanted at the first glance. But I may be wrong, hence the "likely" in my earlier
> reply.
> 

Oh, this was on purpose.

If iommu_flush_context_device() returned 1,  the iommu_flush_iotlb_dsi() returned 1 too.
As both flush_context_qi() and  flush_iotlb_qi () are the same at the beginning of the  functions.

One concern is if iommu_flush_context_device() is failed, then we won't call iommu_flush_iotlb_dsi(),  which is not best effort to flush.


Quan










_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel