Am 21.01.2019 um 17:05 hat Dr. David Alan Gilbert geschrieben: > * Kevin Wolf (kw...@redhat.com) wrote: > > Am 18.01.2019 um 16:57 hat Dr. David Alan Gilbert geschrieben: > > > * Kevin Wolf (kw...@redhat.com) wrote: > > > > Am 14.01.2019 um 11:51 hat Dr. David Alan Gilbert geschrieben: > > > > > * Michael Tokarev (m...@tls.msk.ru) wrote: > > > > > > $ qemu-system-x86_64 -monitor stdio -hda foo.img > > > > > > QEMU 3.1.0 monitor - type 'help' for more information > > > > > > (qemu) stop > > > > > > (qemu) migrate "exec:cat >/dev/null" > > > > > > (qemu) migrate "exec:cat >/dev/null" > > > > > > qemu-system-x86_64: /build/qemu/qemu-3.1/block.c:4647: > > > > > > bdrv_inactivate_recurse: Assertion `!(bs->open_flags & > > > > > > BDRV_O_INACTIVE)' failed. > > > > > > Aborted > > > > > > > > > > And on head as well; it only happens if the 1st migrate is succesful; > > > > > if it got cancelled the 2nd one works, so it's not too bad. > > > > > > > > > > I suspect the problem here is all around locking/ownership - the block > > > > > devices get shutdown at the end of migration since the assumption is > > > > > that the other end has them open now and we had better release them. > > > > > > > > Yes, only "cont" gets control back to the source VM. > > > > > > > > I think we really should limit the possible monitor commands in the > > > > postmigrate status, and possibly provide a way to get back to the > > > > regular paused state (which means getting back control of the resources) > > > > without resuming the VM first. > > > > > > This error is a little interesting if you'd done something like: > > > > > > > > > src: > > > stop > > > migrate > > > > > > dst: > > > <kill qemu for some reason> > > > start a new qemu > > > > > > src: > > > migrate > > > > > > Now that used to work (safely) - note we've not started > > > a VM succesfully anywhere else. > > > > > > Now the source refuses to let that happen - with a rather > > > nasty abort. > > > > Essentially it's another effect of the problem that migration has always > > lacked a proper model of ownership transfer. And it's still treating > > this as a block layer problem rather than making it a core concept of > > migration as it should. > > > > We can stack another one-off fix on top, and get back control of the > > block devices automatically on a second 'migrate'. But it feels like a > > hack and not like VMs had a properly designed and respected state > > machine. > > Hmm; I don't like to get back to this argument because I think > we've got a perfectly servicable model that's implemented at higher > levels outside qemu, and the real problem is the block layer added > new assumptions about the semantics without checking they were really > true. > qemu only has the view from a single host; it takes the higher level > view from something like libvirt to have the view across multiple hosts > to understand who has the ownership when.
Obviously the upper layer is not handling this without the help of QEMU or we wouldn't have had bugs that images were accessed by two QEMU processes at the same time. We didn't change the assumptions, but we only started to actually check the preconditions that have always been necessary to perform live migration correctly. But if you like to think the upper layer should handle all of this, then it's on libvirt to handle the ownership transfer manually. If we really want, we can add explicit QMP commands to activate and inactivate block nodes. This can be done and requiring that the management layer does all of this would be a consistent interface, too. I just don't like this design much for two reasons: The first is that you can't migrate a VM that has disks with a simple 'migrate' command any more. The second is that if you implement it consistently, this has an impact on compatibility. I think it's a design that could be considered if we were adding live migration as a new feature, but it's probably hard to switch to it now. In any case, I do think we should finally make a decision how ownership of resources should work in the context of migration, and then implement that. Kevin
