Re: [Qemu-devel] [RFC 1/5] hw/vfio: Add function for getting reserved_region of device iommu group

[Alex Williamson]

On Wed, 15 Nov 2017 09:49:41 +0000
Shameerali Kolothum Thodi <address@hidden> wrote:

> Hi Alex,
> 
> > -----Original Message-----
> > From: Alex Williamson [mailto:address@hidden
> > Sent: Tuesday, November 14, 2017 3:48 PM
> > To: Zhuyijun <address@hidden>
> > Cc: address@hidden; address@hidden;
> > address@hidden; address@hidden; Shameerali Kolothum
> > Thodi <address@hidden>; Zhaoshenglong
> > <address@hidden>
> > Subject: Re: [Qemu-devel] [RFC 1/5] hw/vfio: Add function for getting
> > reserved_region of device iommu group
> > 
> > On Tue, 14 Nov 2017 09:15:50 +0800
> > <address@hidden> wrote:
> >   
> > > From: Zhu Yijun <address@hidden>
> > >
> > > With kernel 4.11, iommu/smmu will populate the MSI IOVA reserved
> > > window and PCI reserved window which has to be excluded from Guest iova  
> > allocations.  
> > >
> > > However, If it falls within the Qemu default virtual memory address
> > > space, then reserved regions may get allocated for a Guest VF DMA iova
> > > and it will fail.
> > >
> > > So get those reserved regions in this patch and create some holes in
> > > the Qemu ram address in next patchset.
> > >
> > > Signed-off-by: Zhu Yijun <address@hidden>
> > > ---
> > >  hw/vfio/common.c              | 67  
> > +++++++++++++++++++++++++++++++++++++++++++  
> > >  hw/vfio/pci.c                 |  2 ++
> > >  hw/vfio/platform.c            |  2 ++
> > >  include/exec/memory.h         |  7 +++++
> > >  include/hw/vfio/vfio-common.h |  3 ++
> > >  5 files changed, 81 insertions(+)  
> > 
> > I generally prefer the vfio interface to be more self sufficient, if there 
> > are
> > regions the IOMMU cannot map, we should be describing those via capabilities
> > on the container through the vfio interface.  If we're just scraping 
> > together
> > things from sysfs, the user can just as easily do that and provide an 
> > explicit
> > memory map for the VM taking the devices into account.   
> 
> Ok. I was under the impression that the purpose of introducing the 
> /sys/kernel/iommu_groups/reserved_regions was to get the IOVA regions 
> that are reserved(MSI or non-mappable) for Qemu or other apps to
> make use of.  I think this was introduced as part of the "KVM/MSI passthrough
> support on ARM" patch series. And if I remember correctly, Eric had 
> an approach where the user space can retrieve all the reserved regions through
> the VFIO_IOMMU_GET_INFO ioctl and later this idea was replaced with the 
> sysfs interface.
> 
> May be I am missing something here.

And sysfs is a good interface if the user wants to use it to configure
the VM in a way that's compatible with a device.  For instance, in your
case, a user could evaluate these reserved regions across all devices in
a system, or even across an entire cluster, and instantiate the VM with
a memory map compatible with hotplugging any of those evaluated
devices (QEMU implementation of allowing the user to do this TBD).
Having the vfio device evaluate these reserved regions only helps in
the cold-plug case.  So the proposed solution is limited in scope and
doesn't address similar needs on other platforms.  There is value to
verifying that a device's IOVA space is compatible with a VM memory map
and modifying the memory map on cold-plug or rejecting the device on
hot-plug, but isn't that why we have an ioctl within vfio to expose
information about the IOMMU?  Why take the path of allowing QEMU to
rummage through sysfs files outside of vfio, implying additional
security and access concerns, rather than filling the gap within the
vifo API?  Thanks,

Alex