Re: [libvirt] [PATCH RFC 0/3] Add mechanisms to force QEMU capabilities refetches

On Tue, Nov 13, 2018 at 03:21:03PM -0500, John Ferlan wrote: > Sending as an RFC primarily because I'm looking for whether either > or both mechanisms in the series is more or less desired. Likewise, > if it's felt that the current process of telling customers to just > delete the cache is acceptible, then so be it. If there's other ideas > I'm willing to give them a go too. I did consider adding a virsh > option to "virsh capabilities" (still possible) and/or a virt-admin > option to force the refresh. These just were "easier" and didn't > require an API adjustment to implement. > > Patch1 is essentially a means to determine if the kernel config > was changed to allow nested virtualization and to force a refresh > of the capabilities in that case. Without doing so the CPU settings > for a guest may not add the vmx=on depending on configuration and > for the user that doesn't make sense. There is a private bz on this > so I won't bother posting it. > > Patch2 and Patch3 make use of the 'service libvirtd reload' function > in order to invalidate all the entries in the internal QEMU capabilities > hash table and then to force a reread. This perhaps has downsides related > to guest usage and previous means to use reload and not refresh if a guest > was running. On the other hand, we tell people to just clear the QEMU > capabilities cache (e.g. rm /var/cache/libvirt/qemu/capabilities/*.xml) > and restart libvirtd, so in essence, the same result. It's not clear > how frequently this is used (it's essentially a SIGHUP to libvirtd). IMHO the fact that we cache stuff should be completely invisible outside of libvirt. Sure we've had some bugs in this area, but they are not very frequent so I'm not enthusiastic to expose any knob to force rebuild beyond just deleting files.