Re: [libvirt] [RFC PATCH] Use disable-modern=on for disk device='lun'

On 09/20/2016 08:57 AM, Daniel P. Berrange wrote: Another open question is whether or not this is a problem if you start an existing domain that uses a pre-virtio-1.0 machinetype but you're using a qemu binary that does have virtio-1.0. I'm guessing that it should work with no extra "disable-modern=on" required (that's the whole point of versioned machinetypes after all). If so, then existing configs will continue to work properly with no change, and for new configs that use virtio-1.0-capable machinetypes maybe it will be more merciful to immediately complain when somebody tries to use device='lun', thus forcing people who want the "lun" functionality to switch to virtio-scsi right away, rather than propagating its use with virtio-blk-pci even further into the future (which of course will lead to it living *forever*). Trying to restate it more clearly: the problem is that the only way for libvirt to detect whether or not qemu is going to fail is by querying for presence of virtio-1.0 (which we do indirectly by looking for "disable-legacy" in the options), and that only tells us whether or not the qemu binary supports it, *not* whether a particular machinetype supported by the qemu binary supports it. The same binary might support several machinetypes that support virtio-1.0 (and thus don't support device='lun' without adding in disable-modern=on) and several machinetypes that *don't* support virtio-1.0 (and thus should have no problem with device='lun'), and libvirt has no way of knowing which is which, so libvirt *can't* be the one to complain about this, or even to automatically add the "disable-modern=on" (what happens if you add "disable-modern=on" to the virtio-blk-pci device of a machinetype that is too old for virtio-1.0? Is it ignored? Or does qemu give an error and exit? I haven't tried it, so don't know). In the end, until qemu provides us with a way to query capabilities *per machinetype*, there's realistically nothing libvirt can do about this problem that isn't going to be wrong for somebody. (Hi Eduardo!! :-) TL;DR - I've actually come to the opinion that the only right thing for libvirt to do here is "nothing" (unless you want to write code to try and interpret the qemu error message after the fact.... Ewwwww, never mind, scrap that!) To be fair, that's assuming it was an application that silently added "device='lun'" rather than a user explicitly requesting it. Still, even though it was me who originally forced poor Jan to add the warning, the more I think about the details of qemu binary capabilities vs. capabilities of specific versions of machinetypes supported by that qemu binary, the more I realize that it's pointless for libvirt to report this particular warning or try to correct for the problem, because the chances are very high that it will be wrong about it :-/