Re: [PATCH] pci: Save and restore VFs as a part of a reset

[Don Dutile]

On 05/28/2014 04:14 PM, Bjorn Helgaas wrote:
> On Wed, May 28, 2014 at 10:39 AM, Alexander Duyck
> <alexander.h.duyck@xxxxxxxxx> wrote:
> > On 05/27/2014 09:12 PM, Alex Williamson wrote:
> > > On Tue, 2014-05-27 at 19:19 -0600, Bjorn Helgaas wrote:
>
> > > > Maybe resetting the PF should just fail if there's an active VF.  If
> > > > you need to reset the PF, you'd have to unbind the VFs first.
> > >
> > > The use case is certainly questionable, personally I'm not going to
> > > expect VFs to continue working after the PF is reset.  Driver binding
> > > gets complicated, especially when KVM doesn't actually bind devices to
> > > use them.  Hopefully we'll get that out of the tree some day though.  I
> > > suppose we could -EBUSY the PF reset as long as VFs are enabled.
> >
> > What I could do is go through and notify the VFs that they are about to
> > get hit by a reset.  What they do with that information would be up to them.
> >
> > So if the VFs are loaded on the host I could then at least allow them to
> > recover by saving and restoring the config space within the driver
> > themselves.
>
> I really like the idea of punting by failing the PF reset if there are
> any active VFs.  That's a really easy way of making sure we aren't
> going to blow up any guests.  What problems would it cause if we went
> this route?
I think this is the safest route.  PF<->VF interaction isn't architected,
and resetting the PF with active VFs will probably hang a number of SRIOV
implementations, requiring a system-level reset to correct the compounded problem.

> > > > This reminds me about an open problem: VFs can be on "virtual" buses,
> > > > which aren't really connected in the hierarchy, and I don't think we
> > > > have a nice way to iterate over them.  So probably pci_get_device() is
> > > > the best we can do now.
> > >
> > > Yeah, those virtual buses don't have a bus->self, we just have to skip
> > > to bus->parent->self.  pci_walk_bus() goes in the opposite direction,
> > > but without an actual device hosting the bus, I don't see how it finds
> > > it.  Thanks,
> >
> > It seems like we should be able to come up with something like
> > pci_walk_vbus() though or something similar.  All we would need to do is
> > search the VFs on the bus of the PF and all child busses to that bus if
> > I am not mistaken.
>
> I don't think that's going to work because the virtual buses don't
> appear as the child bus of anything.
+1.

> Bjorn

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/