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

["Jan Beulich" <JBeulich@suse.com>]

>>> On 12.05.16 at 15:29, <quan.xu@intel.com> wrote:
> 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,
> ....

Ah, good point. With the renaming going on I didn't realize these
are used here. So in fact they're not just wrappers. Please disregard
my respective comments then.

>> >> > @@ -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.

Such implications need to be commented on, so readers (like me)
don't assume brokenness.

> 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.

Indeed.

Jan

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