3:14 a.m.
On 02/01/2017 09:35 AM, Maxime Coquelin wrote:
Hi,
Few months ago, Michael reported a problem about migrating VMs relying
on vhost-user between hosts supporting different backend versions:
- Message-Id: <20161011173526-mutt-send-email-mst(a)kernel.org>
- https://lists.gnu.org/archive/html/qemu-devel/2016-10/msg03026.html
The goal of this thread is to draft a proposal based on the outcomes
of discussions with contributors of the different parties (DPDK/OVS
/libvirt/...).
Problem statement:
==================
When migrating a VM from one host to another, the interfaces exposed by
QEMU must stay unchanged in order to guarantee a successful migration.
In the case of vhost-user interface, parameters like supported Virtio
feature set, max number of queues, max vring sizes,... must remain
compatible. Indeed, the frontend not being re-initialized, no
renegotiation happens at migration time.
For example, we have a VM that runs on host A, which has its vhost-user
backend advertising VIRTIO_F_RING_INDIRECT_DESC feature. Since the Guest
also support this feature, it is successfully negotiated, and guest
transmit packets using indirect descriptor tables, that the backend
knows to handle.
At some point, the VM is being migrated to host B, which runs an older
version of the backend not supporting this VIRTIO_F_RING_INDIRECT_DESC
feature. The migration would break, because the Guest still have the
VIRTIO_F_RING_INDIRECT_DESC bit sets, and the virtqueue contains some
decriptors pointing to indirect tables, that backend B doesn't know to
handle.
This is just an example about Virtio features compatibility, but other
backend implementation details could cause other failures.
Exactly. Libvirt can't possibly know which virtio features has guest
negotiated to use. Therefore I don't think this falls into libvirt scope.
What we need is to be able to query the destination host's backend to
ensure migration is possible. Also, we would need to query this
statically, even before the VM is started, to be sure it could be
migrated elsewhere for any reason.
Again, if you have more than two hosts, say A-Z, I don't see how libvirt
could know what hosts to asks (where you will migrate your guest), and
what combination of virtio features is okay and which is a deal breaker.
Solution 1: Libvirt queries DPDK vhost lib: *KO*
================================================
Initial idea was to have the management tool (libvirt) to query DPDK
vhost lib and get key/value pairs and check whether migration is
possible. This solution doesn't work for several reasons:
1. Vhost lib API provide a way for the application to disable features
at runtime (understand, not at build time). So coming back to previous
example, DPDK v16.11 supports indirect descriptor features, but it could
be disabled by OVS. We had a look at whether this API was mandatory, and
it turns out to be, as TSO feature is supported on DPDK but not in OVS.
So we cannot rely on DPDK only.
2. Some parameter may be not only DPDK specific, such as the maximum
number of queues for example.
Solution 2: Libvirt queries OVS for vhost backend key/value pairs: *KO*
=======================================================================
Second idea was for OVS to expose its vhost backend implementation
parameters as key/value pairs, for example in the DB or by a dedicated
tool. For example, you could have this kind of information:
- virtio-features: 0x12045694
- max-rx-queues: 1024
- max-rx-ring-size: 512
Doing this, Libvirt has the information to take decision whether
migration is possible or not.
The problem is that Libvirt doesn't know (and want) to interpret these
values (should it be equal/lower/greater/...?), and each time a new
key is introduced in OVS, Libvirt will have to be updated to handle it,
adding an unwanted synchronization constraint between the projects.
Solution 3: Libvirt queries OVS for vhost backend version string: *OK*
======================================================================
The idea is to have a table of supported versions, associated to
key/value pairs. Libvirt could query the list of supported versions
strings for each hosts, and select the first common one among all hosts.
How does libvirt know what hosts to ask? Libvirt aims on managing a
single host. It has no knowledge of other hosts on the network. That's
task for upper layers like RHEV, OpenStack, etc.
Then, libvirt would ask OVS to probe the vhost-user interfaces in the
selected version (compatibility mode). For example host A runs OVS-2.7,
and host B OVS-2.6. Host A's OVS-2.7 has an OVS-2.6 compatibility mode
(e.g. with indirect descriptors disabled), which should be selected at
vhost-user interface probe time.
Advantage of doing so is that libvirt does not need any update if new
keys are introduced (i.e. it does not need to know how the new keys have
to be handled), all these checks remain in OVS's vhost-user implementation.
And that's where they should stay. Duplicating code between projects
will inevitably lead to a divergence.
Ideally, we would support per vhost-user interface compatibility mode,
which may have an impact also on DPDK API, as the Virtio feature update
API is global, and not per port.
In general, I don't think we want any kind of this logic in libvirt. Either:
a) fallback logic should be implemented in qemu (e.g. upon migration it
should detect that the migrated guest uses certain version and thus set
backend to use that version or error out and cancel migration), or
b) libvirt would grew another element/attribute to specify version of
vhost-user backend in use and do nothing more than pass it to qemu. At
the same time, we can provide an API (or extend and existing one, e.g.
virsh domcapabilities) to list all available versions on given host.
Upper layer, which knows what are the possible hosts suitable for
virtualization, can then use this API to ask all the hosts, construct
the matrix, select preferred version and put it into libvirt's domain XML.
But frankly, I don't like b) that much. Lets put the fact this is OVS
aside for a moment. Just pretend this is a generic device in qemu. Would
we do the same magic with it? No! Or lets talk about machine types. You
spawn -M type$((X+1)) guest and then decide to migrate it to a host with
older qemu wich supports just typeX. Well, you get an error. Do we care?
Not at all! It's your responsibility (as user/admin) to upgrade the qemu
so that it supports new machine type. I think the same applies to OVS.
Sorry.
Michal