[libvirt] Migration API broken for xenmigr://
2675 * Returns the new domain object if the migration was successful, 2676 * or NULL in case of error. Note that the new domain object 2677 * exists in the scope of the destination connection (dconn). This is obviously impossible for xenmigr:///. As a result, virsh migrate always returns an error code: # virsh migrate --live domu-220 xen:/// interpol 18:44:40.974: error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName thewhip:xend # echo $? 1 Since virDomainLookupByName() returns NULL. I will request again, can I remove the 'error' code from xenUnifiedDomainLookupByName? More importantly, how can this be fixed? Perhaps we can return 'domain' instead of 'ddomain' if a migrateuri was specified? Also, where does 'domain' get freed in the case where 'ddomain' is returned? Wish I'd spotted this when it went in :( regards john
On Thu, Jan 29, 2009 at 10:50:18PM -0500, John Levon wrote:
2675 * Returns the new domain object if the migration was successful, 2676 * or NULL in case of error. Note that the new domain object 2677 * exists in the scope of the destination connection (dconn).
This is obviously impossible for xenmigr:///. As a result, virsh migrate always returns an error code:
I'm not sure I understand why this is impossible ? You always have two virConnectPtr objects - one of the source, and one for the destination host. The reason for the optional URI is to cope with the scenario where the clients view of what the destination hostname is, has to be different from the source host's view. eg, if a node is multi-homed, 10.0.0.1 and 192.168.1.1 Then the libvirt client might connect to source and estination hosts with xen://10.0.0.1/ (source) xen://10.0.0.2/ (dest) But, the source & dest may only be able to talk to each other using 192.168.1.1/.2, so yu'd have to give a extra URI for xenmigr://192.168.1.2/ so the source host used the correct network address for the actual migration. Either way, the libvirt client can get the resulting domain object, from the destination hosts virConnectPtr instance
# virsh migrate --live domu-220 xen:/// interpol 18:44:40.974: error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName
thewhip:xend # echo $? 1
Since virDomainLookupByName() returns NULL. I will request again, can I remove the 'error' code from xenUnifiedDomainLookupByName?
I'm not sure why it returns NULL there ?!?!
More importantly, how can this be fixed? Perhaps we can return 'domain' instead of 'ddomain' if a migrateuri was specified? Also, where does 'domain' get freed in the case where 'ddomain' is returned?
The passed in 'domain' object is owned by the caller, so the caller is responsible for free'ing it once its done using it. Likewise the returned ddomain object should also be free'd by the caller when it no longer requires it Danoel -- |: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :| |: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|
On Fri, Jan 30, 2009 at 03:41:04PM +0000, Daniel P. Berrange wrote:
This is obviously impossible for xenmigr:///. As a result, virsh migrate always returns an error code:
I'm not sure I understand why this is impossible ?
I thought the purpose of xenmigr:/// was to allow live migration the traditional way without needing a remotely accessible libvirtd. At least, that's how we're using it. If it's the intention of libvirt that you must always have a listening remote API, then we'll have to work out a private fix for our tree.
# virsh migrate --live domu-220 xen:/// interpol 18:44:40.974: error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName
thewhip:xend # echo $? 1
Since virDomainLookupByName() returns NULL. I will request again, can I remove the 'error' code from xenUnifiedDomainLookupByName?
I'm not sure why it returns NULL there ?!?!
xen:/// (the local source host) no longer has the domain, since it moved to interpol.
More importantly, how can this be fixed? Perhaps we can return 'domain' instead of 'ddomain' if a migrateuri was specified? Also, where does 'domain' get freed in the case where 'ddomain' is returned?
The passed in 'domain' object is owned by the caller, so the caller is responsible for free'ing it once its done using it. Likewise the returned ddomain object should also be free'd by the caller when it no longer requires it
Yuck, so that won't do. Perhaps we could duplicate domain and pass that back. regards john
On Fri, Jan 30, 2009 at 11:04:12AM -0500, John Levon wrote:
On Fri, Jan 30, 2009 at 03:41:04PM +0000, Daniel P. Berrange wrote:
This is obviously impossible for xenmigr:///. As a result, virsh migrate always returns an error code:
I'm not sure I understand why this is impossible ?
I thought the purpose of xenmigr:/// was to allow live migration the traditional way without needing a remotely accessible libvirtd. At least, that's how we're using it. If it's the intention of libvirt that you must always have a listening remote API, then we'll have to work out a private fix for our tree.
# virsh migrate --live domu-220 xen:/// interpol 18:44:40.974: error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName libvir: Xen error : Domain not found: xenUnifiedDomainLookupByName
thewhip:xend # echo $? 1
Since virDomainLookupByName() returns NULL. I will request again, can I remove the 'error' code from xenUnifiedDomainLookupByName?
I'm not sure why it returns NULL there ?!?!
xen:/// (the local source host) no longer has the domain, since it moved to interpol.
Yeah, so you've passed invalid data to the migrate call by giving it two connections which both point to the source host. Would should really add a sanity check to try & detect that mistake. Daniel -- |: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :| |: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|
On Fri, Jan 30, 2009 at 04:19:06PM +0000, Daniel P. Berrange wrote:
I'm not sure I understand why this is impossible ?
I thought the purpose of xenmigr:/// was to allow live migration the traditional way without needing a remotely accessible libvirtd. At least, that's how we're using it. If it's the intention of libvirt that you must always have a listening remote API, then we'll have to work out a private fix for our tree.
Could you answer this bit? It sounds like from your other comment that the above is true. We're not going to *require* libvirtd to listen across the network when there's an existing mechanism, so it sounds like we'll have to fork this part. regards, john
On Fri, Jan 30, 2009 at 11:36:16AM -0500, John Levon wrote:
On Fri, Jan 30, 2009 at 04:19:06PM +0000, Daniel P. Berrange wrote:
I'm not sure I understand why this is impossible ?
I thought the purpose of xenmigr:/// was to allow live migration the traditional way without needing a remotely accessible libvirtd. At least, that's how we're using it. If it's the intention of libvirt that you must always have a listening remote API, then we'll have to work out a private fix for our tree.
Could you answer this bit? It sounds like from your other comment that the above is true.
Yes, the migrate API requires a connection to both libvirtd daemons, source and destination.
We're not going to *require* libvirtd to listen across the network when there's an existing mechanism, so it sounds like we'll have to fork this part.
You can use the xen+ssh://hostname/ URI for the destination if you don't want libvirtd to listen publically Daniel -- |: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :| |: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :| |: http://autobuild.org -o- http://search.cpan.org/~danberr/ :| |: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|
On Fri, Jan 30, 2009 at 04:42:04PM +0000, Daniel P. Berrange wrote:
We're not going to *require* libvirtd to listen across the network when there's an existing mechanism, so it sounds like we'll have to fork this part.
You can use the xen+ssh://hostname/ URI for the destination if you don't want libvirtd to listen publically
Which requires setting up ssh keys and is significantly slower. Again, we're not going forbid use of an existing mechanism, but we definitely don't want to make people use xm, so we'll just hack it in. regards, john
participants (2)
-
Daniel P. Berrange -
John Levon