5:20 p.m.
(Several of us started an offline discussion on this topic, and it
quickly became complicated, so we decided it should continue upstream.
Here is a synopsis of the discussion so far (as *I've* interpreted it,
so corrections are welcome and apologies in advance for anything I got
wrong!) Some of the things are stated here as givens, but feel free to
rip them apart.)
Summary of the problem:
1) We want to persuade libvirt+QEMU users to move away from the i440fx
machinetype in favor of Q35. (NB: Someday this *might* lead to the
ability to deprecate and even remove the 440fx machinetype, but even if
that were to happen, it would be a *very long* time from now, so this
discussion is *not* about that!)
2) When Q35 machinetype is used, libvirt assigns virtio devices to a
slot on a PCI Express controller (because why have modern PCIe
controllers/slots available but force everything onto clunky old legacy
controllers?).
3) When a virtio device is plugged into an Express controller, QEMU
disables the device's IO port space, and it is put into "modern-only"
mode (this is done to avoid a rapid exhaustion of limited IO port space).
4) modern-only virtio devices won't work with a legacy (virtio-0.9-only)
guest driver, because virtio-0.9 requires IO port space.
5) Some guest OSes that we still want to support (and which would
otherwise work okay on a Q35 virtual machine) have virtio drivers too
old to support virtio-1.0 (CentOS6 and RHEL6 are examples of such OSes),
but due to the chain of reasons listed above, the "standard" config for
a Q35 guest generated by libvirt doesn't support virtio-0.9, hence
doesn't support these guest OSes.
And here's a list of possible solutions to this problem (note that
"consumers" means management applications such as OpenStack, oVirt,
virt-manager, virt-install, gnome-boxes, etc. In all cases, it's assumed
that the consumer's decision on the action to take will be based on
information from libosinfo). For completeness, I've included even the
possibilities that have been rejected, along with a brief synopsis of
(at least part of) the reason for rejection:
(1) Add some way libvirt consumers can ask libvirt to place
virtio devices on a legacy pci slot instead of pcie when
the machinetype is q35 (qemu sets virtio devices in legacy
PCI slots to transitional mode, so io port space is enabled
and virtio-0.0 drivers will work).
This has been proposed on libvir-list, but rejected. Here is
the most elquently stated reasoning for the rejection I could
find (with thanks to Dan Berrange):
The domain XML is a way to express the configuration
of the guest virtual machine. What we're talking about
here is a policy tunable for an internal libvirt QEMU
driver algorithm, as so does not belong anywhere in the
domain XML.
(2) Add full-blown pci enumeration support to all libvirt consumers
(i.e. they will need to build a model of the PCI bus topology
of each guest, and keep track of which addresses are in use).
They can then manually place virtio devices on legacy pci slots
(again, triggering transitional mode) when the intended guest
OS doesn't support virtio-0.9.
(This is seen as requiring too much duplicated effort for
development and support/maintenance, since up until now libvirt
has been the single point of action for PCI address assignment
(well, QEMU can do it too, but for > 10 years libvirt has
*always* provided full PCI addresses for all devices)
(3) Add virtio-1.0 support to all guest OSes. If this is done,
existing libvirt configs will work.
(Aside from the difficulty of backporting, and the fact that
there are going to be some OSes that don't get it *at all*,
there will always be older releases that haven't gotten the
backport. So this isn't a complete solution).
(4) Consumers can continue using the 440fx machinetype for guest
OSes that don't support virtio-0.9
(This would work, but perpetuates use of the 440fx
machinetype, and all for just this one reason (at least in
the case of CentOS6/RHEL6, which otherwise work just fine with
Q35)).
(5) Introduce virtio-0.9, virtio-1.0 models in libvirt
which are explicitly legacy-only and modern-only.
QEMU doesn't need to change, as libvirt can simply set
the right params on existing QEMU models to force the
behavior.
(NB: it's unclear to me whether virtio-0.9 simply won't
work without forcing the device to be on a legacy PCI
slot, or if that's just "a very bad idea" because it
will mean that the device uses up extra io port space)
The offline discussion had basically come to the point of saying that
options (4) and (5) were the only reasonable ones, with option (5) being
preferred (I think).
As a starter for continuing the discussion, it seems to me that for
option (5):
a) we don't really need the virtio-1.0 model, since that's what you
currently get anyway when you ask for "virtio" on Q35 (and on 440fx,
"virtio" gives you transitional, which works for everybody).
b) Rather than a "legacy-only" model for virtio-0.9, it would be more
useful to have "transitional". This way the config would work for older
OSes that don't support virtio-1.0, and when/if the OS was upgraded such
that it supported virtio-1.0, that would be automatically used without
needing to change the config.
c) Even if it's possible to force a device on an Express slot into
transitional mode, this is extremely wasteful of io port space, so
libvirt should consider virtio-0.9 devices to be legacy PCI, and thus
plug them into legacy PCI slots. And once we're doing this, it's
unnecessary to add any extra option to the qemu commandline to force
legacy support (i.e. transitional mode), as that is what QEMU already
does when the device is connected to a legacy PCI slot.
So making the naive assumption that we agree on implementing option (5)
and there are no objections to my points a-c (Hah! As if!), how does
this sound as a plan:
A) libosinfo starts telling consumers that the preferred virtio device
model for the relevant OSes is "virtio-0.9", and leaves the
recommendation for other OSes as "virtio".
B) libvirt adds a "virtio-0.9" model for all virtio devices that
actually have virtio-0.9 support (a couple of devices never existed
prior to virtio-1.0 (rng and ???) so virtio-0.9 would be nonsensical for
them).
C) inside libvirt, the implementation of the "virtio-0.9" model is
identical to "virtio", except that the VIR_PCI_CONNECT_TYPE flags for
these devices contain VIR_PCI_CONNECT_TYPE_PCI rather than
VIR_PCI_CONNECT_TYPE_PCIE, resulting in those devices being assigned to
a legacy PCI slot, and thus they would be transitional mode by default.
(If there is disagreement about putting these devices on a legacy PCI
slot, then (C) could be changed to add "disable-legacy=off" to the qemu
commandline. But again, even if that works, it would use up 4k of IO
port space for each device, causing it to rapidly run out, and I don't
think that should be the default mode of operation).