[libvirt] How to avoid the disk still persistent after no persistent hotplug and with restarting libvirtd then.
Libvirt loads the domain conf from the status file (/var/run/libvirt/qemu/$dom.xml) if the domain is running, the problem if one restart libvirtd just after some changing on the domain conf, and then destroy/start the guest. libvirt will never known what the original domain conf is then. (vm->newdef is NULL at this moment, generally libvirt restore the domain conf to the original conf (vm->newDef) when destroying and shutdown). Following is a demo of the problem: 1) hotplug a disk to a running domain without "--persistent" specified. 2) restart libvirtd 3) Destroy/start the guest 4) you will see the hotpluged disk is still there. The solution in my mind is to add some attribute to the status XML, such as "live_updated=1", and do the checking when loading the domain conf from status XML: * if "live_updated=1", then loads the domain conf from the persistent domain XML. But the problem is: For a persistent domain, if one made changes on the domain conf both persistently and transiently, we have no way to known which part are changed. Loading from the persistent domain conf is not correct either. For a transient domain, we still have no way to get the original domain conf, that means one will still be surprised when seeing the the disk is still there. Any idea on this? Thanks Osier
On 08/19/2011 08:17 AM, Osier Yang wrote:
Libvirt loads the domain conf from the status file (/var/run/libvirt/qemu/$dom.xml) if the domain is running, the problem if one restart libvirtd just after some changing on the domain conf, and then destroy/start the guest. libvirt will never known what the original domain conf is then. (vm->newdef is NULL at this moment, generally libvirt restore the domain conf to the original conf (vm->newDef) when destroying and shutdown).
Sounds to me like this is a bug in libvirtd reloading. For a running persistent guest, libvirt should be looking in two files: both the running state (/var/run/libvirt/qemu/$dom.xml) and the next-boot state (/etc/lib/libvirt/qemu/$dom.xml); and if they differ, then assign vm->newDef to the persistent definition. In other words, the bug is that vm->newdef was NULL after libvirtd restart.
The solution in my mind is to add some attribute to the status XML, such as "live_updated=1", and do the checking when loading the domain conf from status XML:
Shouldn't need any XML changes; rather, libvirtd just needs to be taught to properly reconstruct its state from multiple files, instead of assuming that the first file that worked was all the state it needed.
For a transient domain, we still have no way to get the original domain conf, that means one will still be surprised when seeing the the disk is still there.
For a transient domain, it is already forbidden to request a persistent change. So you can't get into this situation in the first place unless you have a persistent domain. -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
于 2011年08月20日 01:16, Eric Blake 写道:
On 08/19/2011 08:17 AM, Osier Yang wrote:
Libvirt loads the domain conf from the status file (/var/run/libvirt/qemu/$dom.xml) if the domain is running, the problem if one restart libvirtd just after some changing on the domain conf, and then destroy/start the guest. libvirt will never known what the original domain conf is then. (vm->newdef is NULL at this moment, generally libvirt restore the domain conf to the original conf (vm->newDef) when destroying and shutdown).
Sounds to me like this is a bug in libvirtd reloading. For a running persistent guest, libvirt should be looking in two files: both the running state (/var/run/libvirt/qemu/$dom.xml) and the next-boot state (/etc/lib/libvirt/qemu/$dom.xml); and if they differ, then assign vm->newDef to the persistent definition.
Is this too expensive? Need to compare most of the properties of the domain conf (except the auto-generated info, and some other fixed info such as domain name, UUID, etc. All other info which could be changed lively should be compared). Instead remeber if there was live changes in XML cost low, though need to modify much of the functions I guess.
In other words, the bug is that vm->newdef was NULL after libvirtd restart.
Yes.
The solution in my mind is to add some attribute to the status XML, such as "live_updated=1", and do the checking when loading the domain conf from status XML:
Shouldn't need any XML changes; rather, libvirtd just needs to be taught to properly reconstruct its state from multiple files, instead of assuming that the first file that worked was all the state it needed.
For a transient domain, we still have no way to get the original domain conf, that means one will still be surprised when seeing the the disk is still there.
For a transient domain, it is already forbidden to request a persistent change. So you can't get into this situation in the first place unless you have a persistent domain.
On 08/22/2011 03:12 AM, Osier Yang wrote:
于 2011年08月20日 01:16, Eric Blake 写道:
On 08/19/2011 08:17 AM, Osier Yang wrote:
Libvirt loads the domain conf from the status file (/var/run/libvirt/qemu/$dom.xml) if the domain is running, the problem if one restart libvirtd just after some changing on the domain conf, and then destroy/start the guest. libvirt will never known what the original domain conf is then. (vm->newdef is NULL at this moment, generally libvirt restore the domain conf to the original conf (vm->newDef) when destroying and shutdown).
Sounds to me like this is a bug in libvirtd reloading. For a running persistent guest, libvirt should be looking in two files: both the running state (/var/run/libvirt/qemu/$dom.xml) and the next-boot state (/etc/lib/libvirt/qemu/$dom.xml); and if they differ, then assign vm->newDef to the persistent definition.
Is this too expensive? Need to compare most of the properties of the domain conf (except the auto-generated info, and some other fixed info such as domain name, UUID, etc. All other info which could be changed lively should be compared).
We don't necessarily have to do a comparison. Just blindly parsing in both files is a good enough approach, even in the case where the two files end up not describing any difference in inactive configuration. My main point was that since we track state in two files, we should be restoring two files when libvirtd restarts, but right now we are only restoring one. -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org
participants (2)
-
Eric Blake -
Osier Yang