Re: [PATCH vfio 0/6] Add support for PRE_COPY initial bytes re-initialization

[Alex Williamson <alex@shazbot.org>]

On Tue, 10 Mar 2026 14:44:04 +0200
Yishai Hadas <yishaih@nvidia.com> wrote:

> On 10/03/2026 12:09, Cédric Le Goater wrote:
> > On 3/10/26 10:46, Yishai Hadas wrote:  
> >> On 01/03/2026 14:43, Yishai Hadas wrote:  
> >>> On 28/02/2026 8:27, Cédric Le Goater wrote:  
> >>>> Hello,
> >>>>
> >>>> On 2/27/26 21:23, Alex Williamson wrote:  
> >>>>>
> >>>>> +Cédric, +Peter, please see what you think of this approach 
> >>>>> relative to
> >>>>> QEMU.  The broken uAPI for flags on the PRECOPY_INFO ioctl is
> >>>>> unfortunate, but we need an opt-in for the driver to enable REINIT
> >>>>> reporting anyway.  Thanks,
> >>>>>
> >>>>> Alex  
> >>>>
> >>>>
> >>>> I took a quick look. The series would be a little cleaner if
> >>>> vfio_check_precopy_ioctl() came first  
> >>>
> >>> The motivation to introduce that core helper and adapt all the 
> >>> drivers to use it, was to centralize the common code and ensures that 
> >>> output flags are cleared on entry.
> >>>
> >>> This can be done only after that we have the previous opt-in patch as 
> >>> we would like to keep the V1 behavior for compatibility reasons.
> >>>  
> >>>> and some parts are little ugly
> >>>> (precopy_info_flags_fix). Will take a closer look when back from PTO.
> >>>>  
> >>>
> >>> I'm open for any better name, any specific suggestion ?  
> >>
> >> How about renaming it to precopy_info_v2, which is closer to the 
> >> feature name ?  
> > 
> > I found the prefix '_fix'  confusing and saw no connection with
> > the names of the new flag (VFIO_PRECOPY_INFO_REINIT) or ioctl
> > VFIO_DEVICE_FEATURE_MIG_PRECOPY_INFOv2.
> > 
> > 
> > Based on how 'precopy_info_flags_fix' is used in mlx5vf_precopy_ioctl() :
> > 
> >    +    /*
> >    +     * opt-in for VFIO_DEVICE_FEATURE_MIG_PRECOPY_INFOv2 serves
> >    +     * as opt-in for VFIO_PRECOPY_INFO_REINIT as well
> >    +     */
> >    +    reinit_state = mvdev->core_device.vdev.precopy_info_flags_fix &&
> >    +            migration_state == 
> > MLX5_QUERY_VHCA_MIG_STATE_OPER_MIGRATION_INIT;
> > 
> > 
> > a name like 'precopy_reinit_allow' seems appropriate to me.  
> 
> The opt-in allows using the 'flags' field not only for the reinit case, 
> but for any further usage.

Yes.  It started out as an opt-in for reinit, but then the failure to
sanitize flags was discovered and it seemed to make sense that if we're
defining a v2 interface where flags is valid, reinit could be a
fundamental part of the new base v2 protocol.

If a driver doesn't want/need to support reinit, it never needs to
raise that flag.  Userspace can also always choose to ignore the flag
and transition to stop-copy at any point.  It didn't seem worth a
separate opt-in beyond v2, but do comment if you have other ideas.

This could have also gone the route of making a new v2 ioctl, but this
seemed slightly cleaner.  Thanks,

Alex