Re: [PATCH] intel-iommu: Default to non-coherent for domains unattached to iommus

[Chris Wright <chrisw@sous-sol.org>]

* David Woodhouse (dwmw2@infradead.org) wrote:
> On Fri, 2011-11-11 at 15:49 -0700, Alex Williamson wrote:
> > To fix this, switch domain_update_iommu_coherency() to use the
> > safer, non-coherent default for domains not attached to iommus. 
> 
> That isn't a fix for the problem you described.
> 
> The problem is that changing a domain from coherent to non-coherent is
> *broken*. It probably needs to flush the cache for the *entire* set of
> page tables — not just the new context entry it adds.

For a guest domain, the page tables aren't actually changing.
And for the snoop mode change, we remap the pages.

> You might have removed the *common* case where we trigger that bug, but
> it certainly isn't a fix.
> 
> However, I'd be receptive to an argument that the situation you describe
> is in fact the *only* time we'd have to switch from coherent to
> non-coherent at run time, because the coherency is an all-or-nothing
> characteristic of the chipset. Either all the IOMMUs are coherent, or
> none of them, right? This brain-damage only affects the first chipsets
> from before we worked out that cache incoherency was a *really* f*cking
> stupid idea, doesn't it?

Dunno if it exists going forward (I've stopped being surprised by the
brain damage in this area ;), but those machines are still out there.

> So if you were to ditch the whole idea of a per-domain runtime update,
> and instead calculate a global value for 'iommu_coherency' at boot time,
> by iterating over for_each_active_iommu()¹, I think that would be a
> better way to deal with the issue. And you *could* really call that a
> 'fix'.
> 
> Make sense?

Ideally, yes.  Not sure we can practically do it though.  Would have to
be sure we force incoherent access mode for the busted hw.

thanks,
-chris