Re: [RFC PATCH V4] pci: virtio_pci: Add SR-IOV support for virtio_pci devices

[Michael S. Tsirkin]

On Mon, Feb 26, 2018 at 10:05:31AM -0800, Alexander Duyck wrote:
> On Mon, Feb 26, 2018 at 9:48 AM, Rustad, Mark D <mark.d.rustad@xxxxxxxxx> wrote:
> > Alex,
> >
> >> On Feb 26, 2018, at 7:26 AM, Alexander Duyck <alexander.duyck@xxxxxxxxx> wrote:
> >>
> >> Mark,
> >>
> >> In the future please don't put my "Reviewed-by" on a patch that I
> >> haven't reviewed. I believe I reviewed one of the earlier patches, but
> >> I hadn't reviewed this version.
> >
> > I'm very sorry. I completely spaced doing something about that. I think yours was the first Reviewed-by I ever had in this way. In the future I will remove such things from my changelog right after sending. Thanks for alerting me to what I had failed to do.
> >
> >> Also, after thinking about it over the weekend we may want to look at
> >> just coming up with a truly "generic" solution that is applied to
> >> SR-IOV capable devices that don't have a SR-IOV capable driver loaded
> >> on them. That would allow us to handle the uio, vfio, pci-stub, and
> >> virtio cases all in one fell swoop. I think us going though and
> >> modifying one patch at a time to do this kind of thing isn't going to
> >> scale.
> >
> > The notion of that kind of troubles me - at least pci-stub does. Having worked on ixgbe a bit, I have to wonder what kind of havoc would ensue if an ixgbe device were assigned to a guest, and an attempt was made to allocate VFs by the pci-stub. The guest could be running any version of the ixgbe driver, possibly even an old one that didn't support SR-IOV. Even if it did support SR-IOV, I don't know how it would respond to mailbox messages when it doesn't think it has VFs.
> 
> The assumption here is that the root user knows what they are doing.

There are tools that let non-root users load the stub.

People use that to e.g.  prevent a native driver from loading while also
assuming it won't break the kernel.


> We have already had some discussion on this in regards to VFIO. My
> thought is we look at adding a new PCI sysfs option called
> "sriov_unmanaged_autoprobe" that would be similar to
> "sriov_drivers_autoprobe" and is used to determine if we allow for
> auto probing of the VFs into the host kernel when SR-IOV is enabled.

I'm not sure how a global option can work for use-cases
such as containers though.

> I
> would want to default the value to false so that by default an
> unmanaged PF wouldn't have its VFs assigned to the host unless we
> specifically enable it by updating the sysfs value.
> 
> >> I'll try to do some digging and find the VFIO approach we had been
> >> working on. I think with a couple tweaks we can probably make that
> >> truly generic and ready for submission.
> >
> > I'd like to know more about you are thinking about.
> 
> Basic idea is to take your generic SR-IOV enable/disable bits from
> (http://patchwork.ozlabs.org/patch/877674/) and combine it with the
> some of the autoprobe bits and feedback comments from
> (http://patchwork.ozlabs.org/patch/846454/). The idea would be when
> either no driver is loaded, or a driver without the sriov_configure
> method we update the iov auto probe setting based on the value we set
> via sriov_unamanaged_autoprobe, and then call your generic
> configuration method to enable SR-IOV.