Luiz Capitulino <lcapitul...@redhat.com> wrote: > Hi there, > > Libvirt has a new snapshot API which uses savevm, loadvm and delvm, so we > need to convert them to QMP. > > I thought this wouldn't be difficult, but while doing and testing this work > I hit a number of problems and had to made not so easy decisions. > > Most of the time, the problem is that the handler implementation is not > as consistent as it should be. Usually, this doesn't affect users very much > but the headache begins when you have to make this available under QMP. > > Here goes a list of macro issues/decisions. Details (and other small > problems) > in the patches. > > 1. Multiple failures: do_delvm() and do_savevm() report errors in a > QTAILQ_FOREACH() loop and they don't return when an error happens. > > Although QMP will end up reporting only the first error, this is > considered a bug because handlers are not expected to continue > executing when an error happens.
I haven't looked at all the error cases, but I think that part of the problem are devices like cdrom, that are by definition _non_ snapshotable (what an ugly word). > There are two solutions: > > 1) Return right away in the first error > 2) Identify fatal errors and only report those > > I don't know the implications of doing 1) and don't know how to > to do 2) (although do_loadvm() works this way). > > Can someone from the block layer help here? Use case: We want to save snapshot foo. It can be: - snapshot foo already exist in some device. What to do: - just fail and ask the user for other name - overwrite the name on the devices that already have it - any other good idea? if we just fail the save, then we tell the user: devices bar1 and bar2 already have that snapshot name. What to do? - force the user to use a different name (not such a bad idea) - ask the user to stop the vm and remove the snaphost from the ofended devices with qemu-img. High avalavility people don't like to be forced to stop machines. - delvm just ignore the devices that don't have the snapshot name (basically what we have now) - we add a delvm_force or del_device_vm <device> <name> operations.\ - any other great idea? Except where commented, I like the rest of the patches. The loadvm is a change that I don't agree with. The StateVM name is something that I can live with. Thanks, Juan.
