Re: [PATCH 1/3] xen/vt-d: need barriers to workaround CLFLUSH

["Chen, Tiejun" <tiejun.chen@intel.com>]

On 2015/5/5 17:24, Jan Beulich wrote:
>>>> On 05.05.15 at 04:45, <tiejun.chen@intel.com> wrote:
>> Does this work for everyone?
>
> Please first of all explain why the interfaces in asm/flushtlb.h can't
> be used here (at least when flushing entire pages). Because - as

I also don't understand any reason we didn't use this previously on 
IOMMU side...

> said before - for a complete fix you'd need to deal with the CLFLUSH
> use(s) elsewhere in the system too.

Looks we need to do this in xen/arch/x86/flushtlb.c:flush_area_local().

>
> Btw, isn't the !iommus_incoherent check in that function inverted?
> I.e. why would we _not_ need to flush caches when IOMMUs
> are not coherent (and why would flushing be needed when they're
> coherent anyway)?
>

I guess you're misunderstanding this

     if ( !ecap_coherent(iommu->ecap) )
         iommus_incoherent = 1;

So here !iommus_incoherent means IOMMU is coherent and then we don't 
need to flush cache in this case.

Thanks
Tiejun