On Fri, 15 Jun 2018 10:06:07 -0700
Siwei Liu [off-list ref] wrote:

On Fri, Jun 15, 2018 at 4:48 AM, Cornelia Huck [off-list ref] wrote:

quoted

On Thu, 14 Jun 2018 18:57:11 -0700
Siwei Liu [off-list ref] wrote:
 

quoted

Thank you for sharing your thoughts, Cornelia. With questions below, I
think you raised really good points, some of which I don't have answer
yet and would also like to explore here.

First off, I don't want to push the discussion to the extreme at this
point, or sell anything about having QEMU manage everything
automatically. Don't get me wrong, it's not there yet. Let's don't
assume we are tied to a specific or concerte solution. I think the key
for our discussion might be to define or refine the boundary between
VM and guest,  e.g. what each layer is expected to control and manage
exactly.

In my view, there might be possibly 3 different options to represent
the failover device conceipt to QEMU and libvirt (or any upper layer
software):

a. Seperate device: in this model, virtio and passthough remains
separate devices just as today. QEMU exposes the standby feature bit
for virtio, and publish status/event around the negotiation process of
this feature bit for libvirt to react upon. Since Libvirt has the
pairing relationship itself, maybe through MAC address or something
else, it can control the presence of primary by hot plugging or
unplugging the passthrough device, although it has to work tightly
with virtio's feature negotation process. Not just for migration but
also various corner scenarios (driver/feature ok, device reset,
reboot, legacy guest etc) along virtio's feature negotiation.  

Yes, that one has obvious tie-ins to virtio's modus operandi.
 

quoted

b. Coupled device: in this model, virtio and passthough devices are
weakly coupled using some group ID, i.e. QEMU match the passthough
device for a standby virtio instance by comparing the group ID value
present behind each device's bridge. Libvirt provides QEMU the group
ID for both type of devices, and only deals with hot plug for
migration, by checking some migration status exposed (e.g. the feature
negotiation status on the virtio device) by QEMU. QEMU manages the
visibility of the primary in guest along virtio's feature negotiation
process.  

I'm a bit confused here. What, exactly, ties the two devices together?  

The group UUID. Since QEMU VFIO dvice does not have insight of MAC
address (which it doesn't have to), the association between VFIO
passthrough and standby must be specificed for QEMU to understand the
relationship with this model. Note, standby feature is no longer
required to be exposed under this model.

Isn't that a bit limiting, though?

With this model, you can probably tie a vfio-pci device and a
virtio-net-pci device together. But this will fail if you have
different transports: Consider tying together a vfio-pci device and a
virtio-net-ccw device on s390, for example. The standby feature bit is
on the virtio-net level and should not have any dependency on the
transport used.

quoted

If libvirt already has the knowledge that it should manage the two as a
couple, why do we need the group id (or something else for other
architectures)? (Maybe I'm simply missing something because I'm not
that familiar with pci.)  

The idea is to have QEMU control the visibility and enumeration order
of the passthrough VFIO for the failover scenario. Hotplug can be one
way to achieve it, and perhaps there's other way around also. The
group ID is not just for QEMU to couple devices, it's also helpful to
guest too as grouping using MAC address is just not safe.

Sorry about dragging mainframes into this, but this will only work for
homogenous device coupling, not for heterogenous. Consider my vfio-pci
+ virtio-net-ccw example again: The guest cannot find out that the two
belong together by checking some group ID, it has to either use the MAC
or some needs-to-be-architectured property.

Alternatively, we could propose that mechanism as pci-only, which means
we can rely on mechanisms that won't necessarily work on non-pci
transports. (FWIW, I don't see a use case for using vfio-ccw to pass
through a network card anytime in the near future, due to the nature of
network cards currently in use on s390.)

quoted

quoted

c. Fully combined device: in this model, virtio and passthough devices
are viewed as a single VM interface altogther. QEMU not just controls
the visibility of the primary in guest, but can also manage the
exposure of the passthrough for migratability. It can be like that
libvirt supplies the group ID to QEMU. Or libvirt does not even have
to provide group ID for grouping the two devices, if just one single
combined device is exposed by QEMU. In either case, QEMU manages all
aspect of such internal construct, including virtio feature
negotiation, presence of the primary, and live migration.  

Same question as above.
 

quoted

It looks like to me that, in your opinion, you seem to prefer go with
(a). While I'm actually okay with either (b) or (c). Do I understand
your point correctly?  

I'm not yet preferring anything, as I'm still trying to understand how
this works :) I hope we can arrive at a model that covers the use case
and that is also flexible enough to be extended to other platforms.
 

quoted

The reason that I feel that (a) might not be ideal, just as Michael
alluded to (quoting below), is that as management stack, it really
doesn't need to care about the detailed process of feature negotiation
(if we view the guest presence of the primary as part of feature
negotiation at an extended level not just virtio). All it needs to be
done is to hand in the required devices to QEMU and that's all. Why do
we need to addd various hooks, events for whichever happens internally
within the guest?

''
Primary device is added with a special "primary-failover" flag.
A virtual machine is then initialized with just a standby virtio
device. Primary is not yet added.

Later QEMU detects that guest driver device set DRIVER_OK.
It then exposes the primary device to the guest, and triggers
a device addition event (hot-plug event) for it.

If QEMU detects guest driver removal, it initiates a hot-unplug sequence
to remove the primary driver.  In particular, if QEMU detects guest
re-initialization (e.g. by detecting guest reset) it immediately removes
the primary device.
''

and,

''
management just wants to give the primary to guest and later take it back,
it really does not care about the details of the process,
so I don't see what does pushing it up the stack buy you.

So I don't think it *needs* to be done in libvirt. It probably can if you
add a bunch of hooks so it knows whenever vm reboots, driver binds and
unbinds from device, and can check that backup flag was set.
If you are pushing for a setup like that please get a buy-in
from libvirt maintainers or better write a patch.
''  

This actually seems to mean the opposite to me: We need to know what
the guest is doing and when, as it directly drives what we need to do
with the devices. If we switch to a visibility vs a hotplug model (see
the other mail), we might be able to handle that part within qemu.  

In the model of (b), I think it essentially turns hotplug to one of
mechanisms for QEMU to control the visibility. The libvirt can still
manage the hotplug of individual devices during live migration or in
normal situation to hot add/remove devices. Though the visibility of
the VFIO is under the controll of QEMU, and it's possible that the hot
add/remove request does not involve actual hot plug activity in guest
at all.

That depends on how you model visibility, I guess. You'll probably want
to stop traffic flowing through one or the other of the cards; would
link down or similar be enough for the virtio device?

In the model of (c), the hotplug semantics of the combined device
would mean differently - it would end up with devices plugged in or
out altogther. To make this work, we either have to build a brand new
bond-like QEMU device consist of virtio and VFIO internally, or need
to have some abstraction in place for libvirt to manipulate the
combined device (and prohibit libvirt from operating on individual
internal device directly). Note with this model the group ID doesn't
even need to get exposed to libvirt, just imagine libvirt to supply
all options required to configure two regular virtio-net and VFIO
devices for a single device object, and QEMU will deal with the
device's visibility and enumeration, such when to hot plug VFIO device
in to or out from the guest.

It might be complicated to implement (c) though.

I think (c) would be even more complicated for heterogenous setups.