RE: [PATCH v5 1/2] powerpc/pseries/iommu: Share the per-cpu TCE page with the hypervisor.

[Ram Pai]

On Wed, Dec 11, 2019 at 07:15:44PM +1100, Alexey Kardashevskiy wrote:
> 
> 
> On 11/12/2019 02:35, Ram Pai wrote:
> > On Tue, Dec 10, 2019 at 04:32:10PM +1100, Alexey Kardashevskiy wrote:
> >>
..snip..
> >> As discussed in slack, by default we do not need to clear the entire TCE
> >> table and we only have to map swiotlb buffer using the small window. It
> >> is a guest kernel change only. Thanks,
> > 
> > Can you tell me what code you are talking about here.  Where is the TCE
> > table getting cleared? What code needs to be changed to not clear it?
> 
> 
> pci_dma_bus_setup_pSeriesLP()
> 	iommu_init_table()
> 		iommu_table_clear()
> 			for () tbl->it_ops->get()
> 
> We do not really need to clear it there, we only need it for VFIO with
> IOMMU SPAPR TCE v1 which reuses these tables but there are
> iommu_take_ownership/iommu_release_ownership to clear these tables. I'll
> send a patch for this.

Did some experiments. It spent the first 9s in tce_free_pSeriesLP()
clearing the tce entries.  And the second 13s in 
tce_setrange_multi_pSeriesLP_walk().  BTW: the code in
tce_setrange_multi_pSeriesLP_walk() is modified to use DIRECT_TCE.

So it looks like the amount of time spent in
tce_setrange_multi_pSeriesLP_walk() is a function of the size of the
memory that is mapped in the ddw.


> 
..snip..
> 
> > But before I close, you have not told me clearly, what is the problem
> > with;  'share the page, make the H_PUT_INDIRECT_TCE hcall, unshare the page'.
> 
> Between share and unshare you have a (tiny) window of opportunity to
> attack the guest. No, I do not know how exactly.
> 
> For example, the hypervisor does a lot of PHB+PCI hotplug-unplug with
> 64bit devices - each time this will create a huge window which will
> share/unshare the same page.  No, I do not know how exactly how this can
> be exploited either, we cannot rely of what you or myself know today. My
> point is that we should not be sharing pages at all unless we really
> really have to, and this does not seem to be the case.
> 
> But since this seems to an acceptable compromise anyway,
> 
> Reviewed-by: Alexey Kardashevskiy <aik@xxxxxxxxx>
> 

Thanks!
RP