On 18/01/2019 13:29, Pierre Morel wrote:
On 17/01/2019 14:02, Robin Murphy wrote:
On 15/01/2019 17:37, Pierre Morel wrote:
The s390 iommu can only allow DMA transactions between the zPCI device
entries start_dma and end_dma.
Let's declare the regions before start_dma and after end_dma as
reserved regions using the appropriate callback in iommu_ops.
The reserved region may later be retrieved from sysfs or from
the vfio iommu internal interface.
For this particular case, I think the best solution is to give VFIO
the ability to directly interrogate the domain geometry (which s390
appears to set correctly already). The idea of reserved regions was
really for 'unexpected' holes inside the usable address space - using
them to also describe places that are entirely outside that address
space rather confuses things IMO.
Furthermore, even if we *did* end up going down the route of actively
reserving addresses beyond the usable aperture, it doesn't seem
sensible for individual drivers to do it themselves when the core API
already describes the relevant information generically.
Robin.
Robin,
I already posted a patch retrieving the geometry through
VFIO_IOMMU_GET_INFO using a specific capability for the geometry [1],
and AFAIU, Alex did not agree with this.
What is different in what you propose?
I didn't mean to imply that aperture and reserved regions are mutually
exclusive, just that they are conceptually distinct things, i.e. there
is a fundamental difference between "address which could in theory be
mapped but wouldn't work as expected" and "address which is physically
impossible to map at all".
Admittedly I hadn't closely followed all of the previous discussions,
and Alex has a fair point - for VFIO users who will mostly care about
checking whether two address maps are compatible, it probably is more
useful to just describe a single list of usable regions, rather than the
absolute bounds plus a list of unusable holes within them. That still
doesn't give us any need to conflate things throughout the kernel
internals, though - the typical usage there is to size an IOVA allocator
or page table based on the aperture, then carve out any necessary
reservations. In that context, having to be aware of and handle
'impossible' reservations outside the aperture just invites bugs and
adds complexity that would be better avoided.
Robin.
@Alex: I was hoping that this patch goes in your direction. What do you
think?
Thanks,
Pierre
[1]: https://lore.kernel.org/patchwork/patch/1030369/
This seems to me related with the work Shameer has started on
vfio_iommu_type1 so I add Alex and Shameer to the CC list.
Pierre Morel (1):
iommu/s390: Declare s390 iommu reserved regions
drivers/iommu/s390-iommu.c | 29 +++++++++++++++++++++++++++++
1 file changed, 29 insertions(+)
