Re: [PATCH] uio/uio_pci_generic: Add SR-IOV support

[Don Dutile]

On 10/02/2017 03:10 PM, David Woodhouse wrote:
> On Mon, 2017-10-02 at 14:52 -0400, Don Dutile wrote:
> > On 10/02/2017 08:35 AM, David Woodhouse wrote:
> > > This would allow you to enable SR-IOV on a PF before its driver is
> > > loaded, right? Even when that driver *is* going to need to perform
> > > resource management for those VFs?
> > >
> > > Would existing drivers cope with SR-IOV being enabled, and VFs being
> > > assigned to guests, before they're loaded? If so then sure, let's do it
> > > generically. But I'm not sure that's the case.
> > No better than a uio driver/mgmt api that may have to configure a PF
> > before a VF is enabled.
>
> Conceptually, the current model is that you don't have SR-IOV until you
> have a driver loaded for the physical function which can do any
> necessary resource management.
>
> That's *why* the generic "sriov_numvfs" interface in sysfs isn't
> present until such a driver is loaded.
>
> In the UIO case, *userspace* is responsible for the PF. So it's not an
> "attack vector"; we let userspace do what it likes with the PF and that
> includes enabling SR-IOV too.
>
> Do we actually *disable* SR-IOV when a (UIO or in-kernel) driver for
> the PF is unloaded? If not, that's the only "attack vector" I see â to
> load a driver which permits SR-IOV to be enabled, and do so, and then
> unload it and load a different driver which doesn't cope.
>
> And each driver in that scenario can be either an in-kernel driver or
> UIO+userspace; it doesn't matter either way. The patch I sent is just
> following the *existing* model.
>
> But sure, my question was intended to ask whether we want to *stick*
> with that model. Given the answers I got, my own conclusion was that we
> probably do...
ok. got the whole picture now.
+1 to your reply.