Re: [RFC PATCH 0/9] vfio: Introduce mmap maple tree

[Alex Williamson <alex.williamson@redhat.com>]

On Tue, 5 Aug 2025 17:48:05 +0200
Mahmoud Nagy Adam <mngyadam@amazon.de> wrote:

> Jason Gunthorpe <jgg@ziepe.ca> writes:
> 
> > map2 should not exist, once you introduced a vfio_mmap_ops for free
> > then the mmap op should be placed into that structure.  
> 
> Does this mean dropping mmap op completely from vfio ops? I think we
> could update mmap op in vfio to have the vmmap structure, no?
> 
> > ioctl2 is nasty, that should be the "new function op for
> > VFIO_DEVICE_GET_REGION_INFO" instead. We have been slowly moving
> > towards the core code doing more decode and dispatch of ioctls instead
> > of duplicating in drivers.  
> 
> ack.
> 
> > I still stand by the patch plan I gave in the above email. Clean up
> > how VFIO_DEVICE_GET_REGION_INFO is handled by drivers and a maple tree
> > change will trivially drop on top.
> >  
> 
> but I think also prior of migrating to use packed offsets, we would need
> to fix the previous offset calculations, which means read & write ops
> also need to access the vmmap object to convert the offset.
> 
> 
> 
> Another question is: since VFIO_DEVICE_GET_REGION_INFO will use mt and
> technically will create and return different cookie offset everytime
> it's called for the same region, do we expect that this will not break
> any userspace usage?  I'm not sure but could be that some user usage
> relying on calling the get_region_info to produce the same offset as the
> initial call, instead of caching the region offset for example?

If we're returning a different cookie every time region_info is called
we're already on the wrong track.  Userspace can call region_info an
arbitrary number of times and should get the same offset for the same
region every time.  IMO we need an interface that triggers a new
mapping cookie to be created with specific mmap attributes.  If we're
using region_info then we're dynamically introducing device specific
regions.

I think we're too concerned about this vmmap object, which ought to be
obtained via an offset into the maple tree in place of how we're using
fixed region indexes today.  Some offsets will be pre-populated to
provide the existing regions we have today, optionally those
pre-populated entries may be at compatible offsets for broken
userspaces.  New cookies will be dynamically created alongside.  Thanks,

Alex