On Wed, Nov 14, 2018 at 02:22:12PM -0500, John Ferlan wrote:
On 11/14/18 4:25 AM, Daniel P. Berrangé wrote:
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.
OK - so that more or less obviates patch2 and patch3...
Of course the fact that we cache stuff hasn't been completely invisible and telling someone to fix the problem by "simply" removing the cache files and pointing them to the cache location seems a bit "awkward" once you figure out that is the problem of course. Many times it's just not intuitively obvious!
OTOH, baking in the idea that a "reload" could remove the chance that caching was the problem could be useful. I guess it just felt like it was a perhaps less used option. Assuming of course most would use stop/start or restart instead.
Looking at this from a more general POV, cache invalidation bugs are just one of many different bugs that can & have impacted libvirt over the years. Adding a force reload API is essentially saying we want to have virConnectWhackPossibleBug() because we think we might have a future bug. I don't think that is a good design precedent or rationale - essentially its admitting failure. If caching really were so terribly implemented that this is considered needed, then I'd argue caching should be deleted. I don't think we are in such a bad case though - the kind of problems have been fairly niche in impact. 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 :|