[PATCH 0/6] Clean up virtio device object handling [was Re: [PATCH] virtio: make PCI devices take a virtio_pci module ref]

[Mark McLoughlin <markmc@redhat.com>]

(Moved from kvm@vger to virtualization@linux-foundation, changed
subject, cleaned up cc list)

On Wed, 2008-12-10 at 13:02 +0100, Kay Sievers wrote:
> On Wed, Dec 10, 2008 at 10:49, Mark McLoughlin <markmc@redhat.com> wrote:
> > On Tue, 2008-12-09 at 19:16 +0100, Kay Sievers wrote:
> >> On Tue, Dec 9, 2008 at 17:41, Mark McLoughlin <markmc@redhat.com> wrote:
> >> > On Mon, 2008-12-08 at 08:46 -0600, Anthony Liguori wrote:
> >> >> Mark McLoughlin wrote:
> >> >> > On Sun, 2008-12-07 at 18:52 +1030, Rusty Russell wrote:
> >> >> >> On Saturday 06 December 2008 01:37:06 Mark McLoughlin wrote:
> >> >> >>
> >> >> >>> Another example of a lack of an explicit dependency causing problems is
> >> >> >>> Fedora's mkinitrd having this hack:
> >> >> >>>
> >> >> >>>     if echo $PWD | grep -q /virtio-pci/ ; then
> >> >> >>>         findmodule virtio_pci
> >> >> >>>     fi
> >> >> >>>
> >> >> >>> which basically says "if this is a virtio device, don't forget to
> >> >> >>> include virtio_pci in the initrd too!". Now, mkinitrd is full of hacks,
> >> >> >>> but this is a particularly unusual one.
> >> >> >>>
> >> >> >> Um, I don't know what this does, sorry.
> >> >> >>
> >> >> >> I have no idea how Fedora chooses what to put in an initrd; I can't think
> >> >> >> of a sensible way of deciding what goes in and what doesn't other than
> >> >> >> lists and heuristics.
> >> >> >>
> >> >> >
> >> >> > Fedora's mkinitrd creates an initrd suitable to boot the machine you run
> >> >> > mkinitrd on, rather than creating an initrd suitable to boot any
> >> >> > machine.
> >> >> >
> >> >> > So, it goes "ah, / is mounted from /dev/vda, we need to include
> >> >> > virtio_blk and it's dependencies". It does that in a generic way that
> >> >> > works well for most setups:
> >> >> >
> >> >> >   1) Find the device name (e.g. vda) below /sys/block
> >> >> >
> >> >> >   2) Follow the 'device' link to e.g. /sys/devices/virtio-pci/virtio1
> >> >> >
> >> >> >   3) Find the module need for this through either 'modalias' or the
> >> >> >      'driver/module' symlink
> >> >> >
> >> >> >   4) Use modprobe to list any dependencies of that module
> >> >> >
> >> >> > Clearly, virtio-pci won't be pulled in by any of this so we've added a
> >> >> > hack to say "oh, it's a virtio device, let's include virtio_pci just in
> >> >> > case".
> >> >> >
> >> >> > It's not even the case that mkinitrd needs to know how to include the
> >> >> > the module for the bus, because in our case that's virtio.ko ... we've
> >> >> > pretty effectively hidden the the bus *implementation* from userspace.
> >> >> >
> >> >> > I don't think this is worth wasting too much time fixing, that's why I'm
> >> >> > thinking we should just make virtio_pci built-in by default with
> >> >> > CONFIG_KVM_GUEST.
> >> >> >
> >> >>
> >> >> What if we have multiple virtio transports?
> >> >
> >> > I don't think that's so much an an issue (just build in any transport
> >> > supported by KVM), but rather that you might build a non-pv_ops kernel
> >> > to run on QEMU which would benefit from using virtio drivers ...
> >> >
> >> >>   Is there a way that we can
> >> >> expose the relationship with virtio-blk and virtio-pci in sysfs?  We
> >> >> have a struct device for the PCI device, it's just a matter of making
> >> >> the link visible.
> >> >
> >> > It feels a bit like busy work to generalise this since only virtio_pci
> >> > can be built as a module, but here's a patch.
> >> >
> >> > The mkinitrd hack turns into:
> >> >
> >> >        # Handle finding virtio bus implementations
> >> >        if [ -L ./virtio_module ] ; then
> >> >            findmodule $(basename $(readlink ./virtio_module))
> >> >        else if echo $PWD | grep -q /virtio-pci/ ; then
> >> >            findmodule virtio_pci
> >> >        fi; fi
> >> >
> >> > [PATCH] virtio: add a 'virtio_module' sysfs symlink
> >>
> >> Doesn't the device have a "driver" link already? If yes, the driver it
> >> points to should have a "module" link.
> >
> > The virtio bus is an abstraction that has several different backend
> > implementations - currently virtio-pci, lguest and kvm-s390.
> >
> > So yes, the driver/module link gives us the device driver, but the
> > virtio_module link is to the virtio bus driver (aka implementation,
> > transport, backend, ...):
> >
> >  $> basename $(readlink virtio_module)
> >  virtio_pci
> >  $> basename $(readlink driver/module)
> >  virtio_net
> 
> I see. But why not just call it "module", like we do in all other
> places, when it points to /sys/module/.
> 
> To find dependent modules, you would walk up the chain of parents, and
> include everything that is found by looking for "driver/module" and
> "module" links?
> 
> Wouldn't that make it completely generic, without any virtio specific hacks?

Yeah, that sounds much better - a minor detail is that it'd be better to
hang the symlink off each virtio implementation's root object rather
than off each device.

To that end, I've hacked up register_virtio_root_device() which fixes
the fact that we statically allocate root objects and gives us a sane
place to add this generic symlink.

It might make sense to add this to the core, though - e.g.
device_register_root() - and that would also allow us use the same
approach as module_add_driver() to add the module symlink for built-in
modules.

Cheers,
Mark.