so bing pragmatic i think the simplest path forward given my other patches have not laned in almost 2 years is to quickly add a workaround config option to disable mulitple port bindign which we can backport and then we can try and work on the actual fix after. acording to https://bugs.launchpad.net/neutron/+bug/1815989 that shoudl serve as a workaround for thos that hav this issue but its a regression in functionality.
i can create a patch that will do that in an hour or so and submit a followup DNM patch to enabel the workaound in one of the gate jobs that tests live migration. i have a meeting in 10 mins and need to finish the pacht im currently updating but ill submit a poc once that is done.
im not sure if i will be able to spend time on the actul fix which i proposed last year but ill see what i can do.
On Wed, 2020-04-29 at 16:37 +0200, Ignazio Cassano wrote:
PS I have testing environment on queens,rocky and stein and I can make test as you need. Ignazio
Il giorno mer 29 apr 2020 alle ore 16:19 Ignazio Cassano < ignaziocassano@gmail.com> ha scritto:
Hello Sean, the following is the configuration on my compute nodes: [root@podiscsivc-kvm01 network-scripts]# rpm -qa|grep libvirt libvirt-daemon-driver-storage-iscsi-4.5.0-33.el7.x86_64 libvirt-daemon-kvm-4.5.0-33.el7.x86_64 libvirt-libs-4.5.0-33.el7.x86_64 libvirt-daemon-driver-network-4.5.0-33.el7.x86_64 libvirt-daemon-driver-nodedev-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-gluster-4.5.0-33.el7.x86_64 libvirt-client-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-core-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-logical-4.5.0-33.el7.x86_64 libvirt-daemon-driver-secret-4.5.0-33.el7.x86_64 libvirt-daemon-4.5.0-33.el7.x86_64 libvirt-daemon-driver-nwfilter-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-scsi-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-rbd-4.5.0-33.el7.x86_64 libvirt-daemon-config-nwfilter-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-disk-4.5.0-33.el7.x86_64 libvirt-bash-completion-4.5.0-33.el7.x86_64 libvirt-daemon-driver-qemu-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-4.5.0-33.el7.x86_64 libvirt-python-4.5.0-1.el7.x86_64 libvirt-daemon-driver-interface-4.5.0-33.el7.x86_64 libvirt-daemon-driver-storage-mpath-4.5.0-33.el7.x86_64 [root@podiscsivc-kvm01 network-scripts]# rpm -qa|grep qemu qemu-kvm-common-ev-2.12.0-44.1.el7_8.1.x86_64 qemu-kvm-ev-2.12.0-44.1.el7_8.1.x86_64 libvirt-daemon-driver-qemu-4.5.0-33.el7.x86_64 centos-release-qemu-ev-1.0-4.el7.centos.noarch ipxe-roms-qemu-20180825-2.git133f4c.el7.noarch qemu-img-ev-2.12.0-44.1.el7_8.1.x86_64
As far as firewall driver /etc/neutron/plugins/ml2/openvswitch_agent.ini:
firewall_driver = iptables_hybrid
I have same libvirt/qemu version on queens, on rocky and on stein testing environment and the same firewall driver. Live migration on provider network on queens works fine. It does not work fine on rocky and stein (vm lost connection after it is migrated and start to respond only when the vm send a network packet , for example when chrony pools the time server).
Ignazio
Il giorno mer 29 apr 2020 alle ore 14:36 Sean Mooney smooney@redhat.com ha scritto:
On Wed, 2020-04-29 at 10:39 +0200, Ignazio Cassano wrote:
Hello, some updated about this issue. I read someone has got same issue as reported here:
https://bugs.launchpad.net/neutron/+bug/1866139
If you read the discussion, someone tells that the garp must be sent by qemu during live miration. If this is true, this means on rocky/stein the qemu/libvirt are bugged.
it is not correct. qemu/libvir thas alsway used RARP which predates GARP to serve as its mac learning frames instead https://en.wikipedia.org/wiki/Reverse_Address_Resolution_Protocol https://lists.gnu.org/archive/html/qemu-devel/2009-10/msg01457.html however it looks like this was broken in 2016 in qemu 2.6.0 https://lists.gnu.org/archive/html/qemu-devel/2016-07/msg04645.html but was fixed by https://github.com/qemu/qemu/commit/ca1ee3d6b546e841a1b9db413eb8fa09f13a061b can you confirm you are not using the broken 2.6.0 release and are using 2.7 or newer or 2.4 and older.
So I tried to use stein and rocky with the same version of libvirt/qemu packages I installed on queens (I updated compute and controllers node
on
queens for obtaining same libvirt/qemu version deployed on rocky and
stein).
On queens live migration on provider network continues to work fine. On rocky and stein not, so I think the issue is related to openstack components .
on queens we have only a singel prot binding and nova blindly assumes that the port binding details wont change when it does a live migration and does not update the xml for the netwrok interfaces.
the port binding is updated after the migration is complete in post_livemigration in rocky+ neutron optionally uses the multiple port bindings flow to prebind the port to the destiatnion so it can update the xml if needed and if post copy live migration is enable it will asyconsly activate teh dest port binding before post_livemigration shortenting the downtime.
if you are using the iptables firewall os-vif will have precreated the ovs port and intermediate linux bridge before the migration started which will allow neutron to wire it up (put it on the correct vlan and install security groups) before the vm completes the migraton.
if you are using the ovs firewall os-vif still precreates teh ovs port but libvirt deletes it and recreats it too. as a result there is a race when using openvswitch firewall that can result in the RARP packets being lost.
Best Regards Ignazio Cassano
Il giorno lun 27 apr 2020 alle ore 19:50 Sean Mooney <
smooney@redhat.com>
ha scritto:
On Mon, 2020-04-27 at 18:19 +0200, Ignazio Cassano wrote:
Hello, I have this problem with rocky or newer with iptables_hybrid firewall. So, can I solve using post copy live migration ???
so this behavior has always been how nova worked but rocky the
https://specs.openstack.org/openstack/nova-specs/specs/rocky/implemented/neu...
spec intoduced teh ablity to shorten the outage by pre biding the
port and
activating it when the vm is resumed on the destiation host before we get to pos live
migrate.
this reduces the outage time although i cant be fully elimiated as
some
level of packet loss is always expected when you live migrate.
so yes enabliy post copy live migration should help but be aware that
if a
network partion happens during a post copy live migration the vm will crash and need to be restarted. it is generally safe to use and will imporve the migration performace
but
unlike pre copy migration if the guess resumes on the dest and the mempry page has not been copied
yet
then it must wait for it to be copied and retrive it form the souce host. if the connection too the souce
host
is intrupted then the vm cant do that and the migration will fail and the instance will crash. if
you
are using precopy migration if there is a network partaion during the migration the migration will fail but the instance will continue to run on the source host.
so while i would still recommend using it, i it just good to be aware
of
that behavior change.
Thanks Ignazio
Il Lun 27 Apr 2020, 17:57 Sean Mooney smooney@redhat.com ha
scritto:
> On Mon, 2020-04-27 at 17:06 +0200, Ignazio Cassano wrote: > > Hello, I have a problem on stein neutron. When a vm migrate
from one
node
> > to another I cannot ping it for several minutes. If in the vm I
put a
> > script that ping the gateway continously, the live migration
works
fine
> > and > > I can ping it. Why this happens ? I read something about
gratuitous
arp.
> > qemu does not use gratuitous arp but instead uses an older
protocal
called
> RARP > to do mac address learning. > > what release of openstack are you using. and are you using
iptables
> firewall of openvswitch firewall. > > if you are using openvswtich there is is nothing we can do until
we
> finally delegate vif pluging to os-vif. > currently libvirt handels interface plugging for kernel ovs when
using
the
> openvswitch firewall driver > https://review.opendev.org/#/c/602432/ would adress that but it
and
the
> neutron patch are > https://review.opendev.org/#/c/640258 rather out dated. while
libvirt
is
> pluging the vif there will always be > a race condition where the RARP packets sent by qemu and then mac
learning
> packets will be lost. > > if you are using the iptables firewall and you have opnestack
rock or
> later then if you enable post copy live migration > it should reduce the downtime. in this conficution we do not have
the
race
> betwen neutron and libvirt so the rarp > packets should not be lost. > > > > Please, help me ? > > Any workaround , please ? > > > > Best Regards > > Ignazio > >