[libvirt] [PATCH] mdev: point user to mdevctl for missing devices

[Jonathon Jongsma]

On Tue, 2019-08-20 at 08:44 +0200, Erik Skultety wrote:
> On Mon, Aug 19, 2019 at 02:30:05PM -0500, Jonathon Jongsma wrote:
> > When a host is rebooted, any mediated devices that were previously
> > configured will disappear. There have been requests for libvirt to
> > handle persisting these mediated devices across reboots, but the
> > decision was made that this should be handled at a lower level.
> > mdevctl
> > is a new tool that handles registration and persistence of mediated
> > devices. If desired, mdevctl can automatically start these mediated
> > devices when the parent device becomes available.
> > 
> > Since mdevctl is the recommended solution for handling persistent
> > mediated devices, point users there when they encounter an error
> > for a
> > missing mediated device.
> 
> Is mdevctl already being packaged for distros?

Alex said that he's currently working on getting it into fedora. But as
far as I know it's not officially in any distros yet, so the patch may
be a bit premature. 

> 
> > Signed-off-by: Jonathon Jongsma <jjongsma at redhat.com>
> > ---
> > 
> > NOTE:
> > - previous patch which attempted to start missing devices using
> > mdevctl has
> >   been withdrawn.
> 
> 
> >  src/util/virmdev.c | 4 +++-
> >  1 file changed, 3 insertions(+), 1 deletion(-)
> > 
> > diff --git a/src/util/virmdev.c b/src/util/virmdev.c
> > index 3d5488cdae..df02fc8d28 100644
> > --- a/src/util/virmdev.c
> > +++ b/src/util/virmdev.c
> > @@ -149,7 +149,9 @@ virMediatedDeviceNew(const char *uuidstr,
> > virMediatedDeviceModelType model)
> > 
> >      if (!virFileExists(sysfspath)) {
> >          virReportError(VIR_ERR_DEVICE_MISSING,
> > -                       _("mediated device '%s' not found"),
> > uuidstr);
> > +                       _("mediated device '%s' not found. "
> > +                         "Persistent devices can be managed with
> > 'mdevctl'."),
> > +                       uuidstr);
> 
> mdevctl is very cool, but unless it's being packaged for distros
> which to me is
> kind of a guarantee (maintenance commitment may be a better wording)
> that it
> will not stay just a cool experiment, I'm not sure we want to
> reference it
> directly in the error messages. Of course, it's been a few months
> that I
> followed up on mdevctl, so I don't know where it stands as of now,
> can you
> refresh my memory?
> 
> Now, my response comes quite late (sorry for that) as I've been
> thinking about
> the original patch you sent and I was also trying to remember all the
> details
> (I already know I surely forgot some bits). Last time we discussed
> this, I
> think the issue with libvirt in this case is migration of domains
> with mdevs in
> the future.  Right now, libvirt domains only hold the UUID to
> reference an
> mdev, so as not to duplicate the information we already have about
> the mdev in
> the nodedev driver.
> 
> I'd very happily delegate the responsibility onto mdevctl, but do we
> have a
> guarantee that mdevctl is available on the destination machine?
> Because if not,
> then libvirt wouldn't be able to resume the machine because the
> device would
> likely not exist if someone didn't pre-create it manually. Then
> there's the
> problem with the UUID->mdev_type mapping because the state of mdevctl
> would not
> be transferred.  On the source, the UUID maps to some type, but the
> destination
> doesn't know about this mapping, does it? This is why we discussed
> that even
> though mdevctl provides us with clear benefits, libvirt would still
> need to
> keep track of the mdevs. The initial idea was to re-work the nodedev-
> driver (I
> have a GSoC project listed on libvirt.org), so that it:
> a) supports the define;start;destroy;etc. commands which all the
> drivers do
> b) specifies what an mdev XML looks like and allows to define it
> c) extends the current Create (from XML) function so that it supports
> creation
>    of more than just NPIV/vHBAs - currently it's tied to NPIV
> 
> Once in place, the user/management layer would then have to make sure
> that they
> migrate the persistent configs across all the compute nodes before
> the
> migration actually takes place. Libvirt would then "start" (aka
> create) the
> mdevs on demand when the domain is started.
> I haven't seen the migration discussion for a while, so I don't know
> where we
> stand ATM, I'm saying this because I don't know if and how the
> mdev_types on
> newer revisions of vGPU HW are compatible which plays a role in the
> XML
> structure we'd select for an mdev in libvirt.
> 
> Like I said, this is roughly where we were standing at the last time
> we had a
> discussion about this, situation may have changed already, these were
> just my
> 2 cents and I myself need to find time to catch up with mdevs a bit
> :).

Erik and I had a brief discussion on IRC. One of his questions was
whether mdevctl exposed any configuration:

<eskultet> I can imagine libvirt depending on mdevctl if it exposes
some configs or profiles, because once we have those, LP can migrate
those and libvirt should IMO be fine

And the answer is that mdevctl does provide a way to export
configuration in JSON format. For example, here's mdevctl showing the
mediated device that I configured with my laptop GPU:

$ mdevctl list -d --dumpjson
[
  {
    "0000:00:02.0": [
      {
        "20e75edd-1c34-43a2-a951-f1ee1e77e789": {
          "mdev_type": "i915-GVTg_V5_8",
          "start": "auto"
        }
      }
    ]
  }
]

Something above libvirt would need to ensure that a suitable device
exists on the migration destination and then translate the PCI address
somehow, but it looks like mdevctl probably gives us enough
information. I need to think a bit more deeply about the exact
requirements and implementation details. I'd welcome any guidance from
those with more experience in this area.

Jonathon