Re: Caching qemu capabilities and KubeVirt

On Thu, Sep 08, 2022 at 02:23:31PM +0100, Daniel P. Berrangé wrote: > On Thu, Sep 08, 2022 at 03:10:09PM +0200, Roman Mohr wrote: > > On Thu, Sep 8, 2022 at 2:56 PM Daniel P. Berrangé <berrange(a)redhat.com > > > wrote: > > > > > On Thu, Sep 08, 2022 at 02:24:00PM +0200, Roman Mohr wrote: > > > > Hi, > > > > > > > > > > > > I have a question regarding capability caching in the context of > > > KubeVirt. > > > > Since we start in KubeVirt one libvirt instance per VM, libvirt has to > > > > re-discover on every VM start the qemu capabilities which leads to a > > > 1-2s+ > > > > delay in startup. > > > > > > > > We already discover the features in a dedicated KubeVirt pod on each > > > node. > > > > Therefore I tried to copy the capabilities over to see if that would > > > work. > > > > > > > > It looks like in general it could work, but libvirt seems to detect a > > > > mismatch in the exposed KVM CPU ID in every pod. Therefore it invalidates > > > > the cache. The recreated capability cache looks esctly like the original > > > > one though ... > > > > > > > > The check responsible for the invalidation is this: > > > > > > > > ``` > > > > Outdated capabilities for '%s': host cpuid changed > > > > ``` > > > > > > > > So the KVM_GET_SUPPORTED_CPUID call seems to return > > > > slightly different values in different containers. > > > > > > > > After trying out the attached golang scripts in different containers, I > > > > could indeed see differences. > > > > > > > > I can however not really judge what the differences in these KVM function > > > > registers mean and I am curious if someone else knows. The files are > > > > attached too (as json for easy diffing). > > > > > > Can you confirm whether the two attached data files were captured > > > by containers running on the same physical host, or could each > > > container have run on a different host. > > > > > > > They are coming from the same host, that is the most surprising bit for me. > > I am also very sure that this is the case, because I only had one k8s node > > from where I took these. > > The containers however differ (obviously) on namespaces and on the > > privilege level (less obvious). The handler dump is from a fully privileged > > container. > > The privilege level sounds like something that might be impactful, > so I'll investigate that. I'd be pretty surprised for namespaces > to have any impact thnough. The privilege level is a red herring. Peter reminded me that we have to filter out some parts of CPUID because the APIC IDs vary depending on what host CPU the task executes on. https://gitlab.com/libvirt/libvirt/-/blob/master/src/util/virhostcpu.c#L1346 In the 2 jSON files you provide, the differences i see should already be matched by /* filter out local apic id */ if (entry->function == 0x01 && entry->index == 0x00) entry->ebx &= 0x00ffffff; if (entry->function == 0x0b) entry->edx &= 0xffffff00; so those differences ought not to be causing the cache to be invalidated.

With regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|