Re: Add option to skip cpu feature and model verification in libvirt and rely on qemu 'enforce' option

On Wed, Nov 23, 2022 at 11:18:01 +0000, Daniel P. Berrangé wrote: > On Tue, Nov 22, 2022 at 01:48:39PM +0100, Jiri Denemark wrote: > > On Mon, Nov 21, 2022 at 12:47:58 +0530, manish.mishra wrote: > > > > > > On 18/11/22 5:00 pm, Jiri Denemark wrote: > > > > On Fri, Nov 18, 2022 at 10:18:59 +0000, John Levon wrote: > > > >> On Fri, Nov 18, 2022 at 10:52:32AM +0100, Jiri Denemark wrote: > > > >> > > > >>>> * Qemu already provide an option 'enforce' to validate if features > > > >>>> with which vm is started is exactly same as one provided and nothing > > > >>>> is silently dropped. > > > >>> Right, but it's not enough. In addition to removed features libvirt also > > > >>> checks for unexpectedly added features. And you really need to do both. > > > >>> Because if you ask for -cpu Model,feat1=on,feat2=on,enforce and QEMU > > > >>> says everything is fine, the guest might see more than what you asked. > > > >>> For example, if a feature is enabled only if a host supports it you may > > > >>> or may not get it without any complains from QEMU. But if you get it you > > > >>> really need to explicitly ask for it during migration, otherwise the > > > >>> feature can just silently disappear. Of course, this would be a really > > > >>> bad behavior from QEMU, but that does not mean it can't happen (I think > > > >>> SVM is a bit problematic in this way) and the whole point of libvirt's > > > >>> checks is to prevent this kind of issues. > > > >> Hi Jiri, I'm not following this very well. I think you're saying that qemu has > > > >> had bugs previously where features get silently enabled, > > > > Personally, I haven't actually witnessed any bug in this area (as far as > > > > I can remember, which is not that far :-)), but got some reports of at > > > > least one, even though without any proof. > > > > > > > > Specifically, SVM is quite strange as it is included in all AMD CPU > > > > models in QEMU and yet if you try to start it on a host without nesting > > > > enabled "enforce" does not complain. I saw the feature is enabled with > > > > older machine types, but I was told the magic behind this feature looks > > > > like not only machine type but even host configuration itself is > > > > involved. Anyway, although I saw reports of that the feature could be > > > > enabled without an explicit request even with new machine types I still > > > > haven't seen any proof of this happening. So I hope it just does not > > > > happen and users are only afraid of this possibility. > > > > > > > >> and it's libvirt's job/role to paper over those issues? Do you have > > > >> some specific cases of this? > > > > This is not about papering. It's actually the opposite, that is about > > > > detecting if something like this happens and reporting it as a failure > > > > rather than papering it and hoping everything goes well. So I think > > > > doing this is a good idea even though I don't think we actually saw any > > > > issue of this kind. > > > > > > > > > Hi Jiri, I see now with libvirt master, with check=='full' we verify > > > both silently dropped as well as added features. But as you already > > > stated Qemu silently adding feature is a Qemu bug and libvirt just > > > reports that bug, so it should be very unlikely, i agree that is not a > > > good reasoning :). Our requirement is that we want to use CPU Models > > > and features which are defined in Qemu but not in libvirt for e.g if > > > we want to use Icelake-Server-V4 directly or newly added vmx feature, > > > libvirt does not allow. Currently we take help of qemu to do > > > validations but for cpu feature verfication and model definations we > > > still use libvirt defined definations which prevent us to use anything > > > which is not defined in libvirt. I see there are already efforts going > > > on to get all model and feature defination from qemu itself, but not > > > sure how much time it will take. Till that happens we thought safest > > > option is to have an option to remove all validations from libvirt and > > > rely on qemu 'enforce' for migration safetly. I understand > > > qemu-enforce does not check for silently added features, but that case > > > we can assume is very unlikely and Qemu should fix otherwise VMs will > > > not poweron anyway with check=='full'. Basically we want it as an > > > modification of check='none' but also skipping things like > > > virCPUValidateFeatures and passing option 'enforce' to Qemu. Or if > > > silently adding features is that big concern we can have a check in > > > Qemu itself? I understand current qemu-enforce is not as migration > > > safe as check=='full' but probably suitable for our use case for time > > > being? > > > > I understand why you want this, but on the other hand adding just a > > plain pass-through for CPU model and features is quite dangerous as it > > can be used to set any -cpu option even if it's not a feature. And > > especially the parts that are configured in other areas of domain XML > > (such as hypervisor features). So I think instead of adding a new mode > > for <cpu> we should go another way. For things that are not yet > > supported by libvirt we support various elements in qemu namespace, > > e.g., setting QEMU command line options, environment variables, or even > > overriding options libvirt would normally set on the command line. So I > > guess we could have a special <qemu:cpu> element which would be used > > when a user wants full control over the -cpu option without any libvirt > > intervention. > > I really don't think we should allow this, especially not for something > which is clearly intended to be used for production deployment. Our > hypervisor CLI passthrough is there to allow for short term workarounds > for bugs, or to experiment with a feature before it is mapped into > libvirt in a supported manner. > > If there are aspects related to QEMU configuration thiat are not in > libvirt yet, we need to address those gaps, not provide yet another > way to bypass libvirt mapping. Indeed, this was definitely meant as a short term workaround for stuff we don't have support for yet, for testing or experimental purposes. The supported approach is for implement the missing parts in libvirt (and QEMU) as soon as possible. I don't see why would a properly documented support for experiments with -cpu would be an issue.