vfio-user/spec: report in VFIO_USER_DEVICE_GET_REGION_INFO that region contains I/O region FDs (?)

[tmakatos]

Currently the vfio-user client doesn't know whether a particular region contains I/O region FDs. The only way for it to tell is to send a VFIO_USER_DEVICE_GET_REGION_IO_FDS to the vfio-user server. This could be avoided if the vfio-user server could set a flag in the response or something. However setting a flag, e.g. VFIO_REGION_INFO_FLAG_CONTAINS_REGION_IO_FDS could conflict with new VFIO_REGION_INFO_FLAG_ flags added upstream (otherwise we'd have to have our own versions of VFIO_USER_REGION_INFO_FLAG_. Another way would be to introduce a new region capability type, we'd still have the same problem with it being re-defined upstream (unless we can introduce an upstream VFIO_REGION_TYPE_USER and then freely define VFIO_REGION_SUBTYPE_USER_XXX?).

At this stage I think it's find having the vfio-user client having to explicitly send a message just to tell whether or not the region contains I/O region FDs.

@jlevon @john-johnson-git thoughts?

[jlevon]

this is control path, so I don't think it's so bad to just have the client ask, I think.

[john-johnson-git]

On Jun 23, 2022, at 1:47 PM, Thanos Makatos ***@***.***> wrote: Currently the vfio-user client doesn't know whether a particular region contains I/O region FDs. The only way for it to tell is to send a VFIO_USER_DEVICE_GET_REGION_IO_FDS to the vfio-user server. This could be avoided if the vfio-user server could set a flag in the response or something. However setting a flag, e.g. VFIO_REGION_INFO_FLAG_CONTAINS_REGION_IO_FDS could conflict with new VFIO_REGION_INFO_FLAG_ flags added upstream (otherwise we'd have to have our own versions of VFIO_USER_REGION_INFO_FLAG_. Another way would be to introduce a new region capability type, we'd still have the same problem with it being re-defined upstream (unless we can introduce an upstream VFIO_REGION_TYPE_USER and then freely define VFIO_REGION_SUBTYPE_USER_XXX?).

The qemu upstream folks told me vfio.h changes have to go to linux first, then pulled to qemu, so I’ve avoided changes there to date. I suspect you’ll have better luck with a new capability, but I’d ping Alex.

At this stage I think it's find having the vfio-user client having to explicitly send a message just to tell whether or not the region contains I/O region FDs.

This seems like a feature the kernel implementation could potentially use, so I’d at least try to get it in vfio.h, but if you get a lot of pushback, a per-region message at startup seems resonable. JJ