Re: [PATCH] intel-iommu: Manage iommu_coherency globally

[Alex Williamson <alex.williamson@redhat.com>]

On Fri, 2011-11-18 at 17:00 +0000, David Woodhouse wrote:
> On Fri, 2011-11-18 at 09:03 -0700, Alex Williamson wrote:
> > I can't help but thinking we're just taking the easy, lazy path for VM
> > domains when it sounds like we really would prefer to keep the coherency
> > set to the least common denominator of the domain rather than the
> > platform.  Couldn't we instead force a flush of the domain when we
> > transition from coherent to non-coherent?  Not sure I'm qualified to
> > write that, but seems like it would keep the efficiency of VM domains
> > with no effect to native DMA domains since they'd never trigger such a
> > transition.  The below appears as if it would work and probably be OK
> > since the unnecessary cache flushes are rare, but they're still
> > unnecessary... and the comments/commit log are now wrong.  Thanks,
> 
> Yeah, that would make some sense. I was about to knock up some code
> which would walk the page tables and use clflush to flush every one...
> but wouldn't it be saner just to use wbinvd?

A bit heavy handed, but obviously easier.  It feels like we could safely
be more strategic, but maybe we'd end up trashing the cache anyway in a
drawn out attempt to flush the context and all page tables.  However, do
we actually need a wbinvd_on_all_cpus()?  Probably better to trash one
cache than all of them.  Thanks,

Alex