On Thu, Sep 02, 2021 at 04:57:04PM +0200, Markus Armbruster wrote: > Peter Xu <pet...@redhat.com> writes: > > > Hi, Dan, > > > > On Thu, Sep 02, 2021 at 02:53:00PM +0100, Daniel P. Berrangé wrote: > >> The correct ordering of devices/backends is generally pretty obvious > >> for libvirt to determine. Most of the problems we've had related to > >> ordering are on the QEMU side, because the ARGV given to QEMU made > >> correct sense if parsed left-to-right, but QEMU didn't actually process > >> them in that order. We've patched QEMU to hack around its inability to > >> honour the CLI order repeatedly. > > > > Is there a pointer to the problem? > > Just an example: > > 9ea18ed25a "vl: Fix -drive / -blockdev persistent reservation management > cda4aa9a5a "vl: Create block backends before setting machine properties"
Thanks, same to Dan. > > >> Being completely self-ordering on the QEMU side using a topological > >> sort would be neat from a conceptual purity POV, but that is quite a > >> challenge to implement and I'm not convinced it is worth it, compared > >> to other problems we want to spend time on. > > > > I just noticed there can also be dependency between the buses; that cannot > > be > > fixed by ordering of classes indeed as either proposed in this series, or > > introduce a new priority. > > --verbose? Please ignore it - I just started to realize what kind of a rabbit hole I'm digging, and I am already prepared to run away. :) -- Peter Xu
