[PATCH 3/5] iommu/arm-smmu-v3: add IOMMU_CAP_BYPASS to the ARM SMMUv3 driver

[alex.williamson@redhat.com (Alex Williamson)]

On Mon, 24 Jul 2017 18:23:20 +0100
Robin Murphy <robin.murphy@arm.com> wrote:

> On 24/07/17 18:16, Alex Williamson wrote:
> > On Thu, 20 Jul 2017 12:17:12 +0100
> > Robin Murphy <robin.murphy@arm.com> wrote:
> >   
> >> On 20/07/17 10:10, Will Deacon wrote:  
> >>> On Thu, Jul 20, 2017 at 09:32:00AM +0530, Anup Patel wrote:    
> >>>> On Wed, Jul 19, 2017 at 5:23 PM, Will Deacon <will.deacon@arm.com> wrote:    
> >>>>> There are two things here:
> >>>>>
> >>>>>   1. iommu_present() is pretty useless, because it applies to a "bus" which
> >>>>>      doesn't actually tell you what you need to know for things like the
> >>>>>      platform_bus, where some masters might be upstream of an SMMU and
> >>>>>      others might not be.    
> >>>>
> >>>> I agree with you. The iommu_present() check in vfio_iommu_group_get()
> >>>> is not much useful. We only reach line which checks iommu_present()
> >>>> when iommu_group_get() returns NULL for given "struct device *". If there
> >>>> is no IOMMU group for a "struct device *" then it means there is no IOMMU
> >>>> HW doing translations for such device.
> >>>>
> >>>> If we drop the iommu_present() check (due to above reasons) in
> >>>> vfio_iommu_group_get() then we don't require the IOMMU_CAP_BYPASS
> >>>> and we can happily drop PATCH1, PATCH2, and PATCH3.
> >>>>
> >>>> I will remove the iommu_present() check in vfio_iommu_group_get()
> >>>> because it is only comes into actions when VFIO_NOIOMMU is
> >>>> enabled. This will also help us drop PATCH1-to-PATCH3.    
> >>>
> >>> I don't think that's the right answer. Whilst iommu_present has obvious
> >>> shortcomings, its intention is clear: it should tell you whether a given
> >>> *device* is upstream of an IOMMU. So the right fix is to make this
> >>> per-device, instead of per-bus. Removing it altogether is worse than leaving
> >>> it like it is.    
> >>
> >> Not really - if there is an IOMMU up and running to the point of setting
> >> bus ops, every device it cares about can be expected to have a group
> >> already (there are only a couple of drivers left that don't use groups,
> >> and they're hardly relevant to VFIO). Thus iommu_group_get() already is
> >> the de-facto per-device IOMMU check.
> >>
> >> And having looked into it, I'm now spinning a couple of patches to
> >> finish off making groups truly mandatory so that that can be less
> >> de-facto ;)  
> > 
> > No, look at vfio-noiommu and even vfio-mdev devices for devices which
> > have an iommu group but there is no physical iommu supporting them.
> > iommu_present() is how we can distinguish these groups and therefore
> > not generate a segfault in trying to use the full IOMMU API on them.  
> 
> OK, so that means that the combination of vfio-noiommu and vfio-platform
> is simply unusable, because iommu_present(&platform_bus_type) can give
> such dangerous false positives too.

Yep, this kinda falls apart since platform_bus_type doesn't really map
to a physical bus, nor does the presence of a group canonically
demonstrate that an iommu is present since anyone can create a group
for a device.  We really do need to migrate to per-device iommu_ops.
Thanks,

Alex