[libvirt] mdevctl: A shoestring mediated device management and persistence utility

[Halil Pasic]

On Wed, 12 Jun 2019 09:14:39 +0200
Cornelia Huck <cohuck at redhat.com> wrote:

> On Tue, 11 Jun 2019 14:28:22 -0600
> Alex Williamson <alex.williamson at redhat.com> wrote:
> 
> > On Tue, 11 Jun 2019 21:45:08 +0200
> > Cornelia Huck <cohuck at redhat.com> wrote:
> > 
> > > On Fri, 7 Jun 2019 18:06:30 +0200
> > > Halil Pasic <pasic at linux.ibm.com> wrote:
> 
> > > > I guess for vfio-ccw one needs to make sure that the ccw device is bound
> > > > to the vfio-ccw driver first, and only after that can one use  
> > > > create-mdev to create the mdev on top of the subchannel.
> > > > 
> > > > So to make this work persistently (survive a reboot) one would need to
> > > > take care of the subchannel getting bound to the right vfio_ccw driver
> > > > before mdevctl is called. Right?
> > > > 
> > > > BTW how does this concurrence situation between the drivers io_subchannel
> > > > and vfio_ccw work? Especially if both are build in?    
> > > 
> > > If you have two drivers that match to the same device type, you'll
> > > always have the issue that the driver that is first matched with the
> > > device will bind to it and you have to do the unbind/rebind dance to
> > > get it bound to the correct device driver. (I guess that this was the
> > > basic motivation behind the ap bus default driver infrastructure,
> > > right?) I think that in our case the io_subchannel driver will be
> > > called first (alphabetical order and the fact that vfio-ccw will often
> > > be a module). I'm not sure if it is within the scope of mdevctl to
> > > ensure that the device is bound to the correct driver, or if it rather
> > > should work with devices already bound to the correct driver only.
> > > Maybe a separate udev-rules generator?  
> > 
> > Getting a device bound to a specific driver is exactly the domain of
> > driverctl.  Implement the sysfs interfaces driverctl uses and see if it
> > works.  Driverctl defaults to PCI and knows some extra things about
> > PCI, but appears to be written to be generally bus agnostic.  Thanks,
> > 
> > Alex
> 

@Alex: Thanks! I was not aware of driverctl.

> Ok, looked at driverctl. Extending this one for non-PCI seems like a
> reasonable path. However, we would also need to extend any non-PCI
> device type we want to support with a driver_override attribute like
> you did for PCI in 782a985d7af26db39e86070d28f987cad2 -- so this is
> only for newer kernels. Adding that attribute for subchannels looks
> feasible at a glance, but I have not tried to actually do it :)
> 
> Halil, do you think that would make sense?

Looks doable. Did not quite figure out the details yet, but it seems
that for PCI driver_override has more benefits than for cio (compared
to simple unbind/bind), as matching and probing seems to be more
elaborate for PCI. The benefit I see are
1) the ability to exclude the device form driver binding, and
2) having the same mechanism and thus consistent experience for pci and
cio.

What we IMHO should not do is make driver_override the override the
sch->st == id->type check.

Regards,
Halil

> 
> [This might also help with the lcs vs. ctc confusion on a certain 3088
> cu model if this is added for ccw devices as well; but I'm not sure if
> these are still out in the wild at all. Probably not worth the effort
> for that.]