On Thu, Jun 28, 2018 at 07:59:38PM +0100, Dr. David Alan Gilbert wrote: > * Daniel P. Berrangé (berrange@xxxxxxxxxx) wrote: > > This post is to raise question about helping use of named CPU models with > > KVM ie any case not using -cpu host. > > > > In the old days (ie before 2018), the world was innocent and we had a nice > > set of named CPU models that corresponded to different Intel/AMD physical > > CPU families/generations (lets temporarily ignore the -noTSX fiasco). > > > > An application could query libvirt to determine what the host CPU model > > was/is and use that model name in the guest XML and be fairly happy. If > > they wanted to, they could explicitly include the extra features listed > > by capabilities XML, or just rely on the host-model. > > > > Then Spectre happened, and QEMU took the decision to almost double the > > number of x86 models, adding in -IBRS / -IBPB variants for most CPU model, > > so that applications could get the spec_ctrl / ibpb flags set without > > having to manually list them. > > > > In retrospect this was somewhat pointless, at least at the QEMU level, > > because there is little difference in complexity between the two approaches: > > > > -cpu Westmere,+spec-ctrl > > -cpu Westmere-IBRS > > > > At a higher level the extra named CPU models were slightly useful in so much > > as many application developers had taken a lazy approach and not provided > > users any way to explicitly turn on extra flags. This affected oVirt, > > OpenStack and virt-manager, and probably more. Though OpenStack since added > > ability to turn on arbitrary flags in response to the Spectre flaw, others > > have not. > > > > Then a recently along came the Speculative Store Bypass hardware vulnerability > > requiring addition of yet another CPU flag to guest configs. This required use > > of 'ssbd' on Intel and 'virt-ssbd' on AMD. While QEMU could have now added yet > > more CPU models, eg Westmere-SSBD, this does not feel like a winning strategy > > long term. Looking at the models how would a user have any clue whether the > > -IBRS or -SSBD or -NEXT-FLAW or -YET-ANOTHER-FLAW suffix is "better" ? So QEMU > > and libvirt took the joint decision to stop adding new named CPU models when > > CPU vulnerabilities are discovered from this point forwards. Applications / > > users would be expected to turn on CPU features explicitly as needed and are > > considered broken if they don't provide this functionality. > > > > As briefly mentioned above though, even before Spectre we had the pain of > > dealing with the -noTSX CPU models working around brokenness in the Intel TSX > > impl where they had to delete a CPU feature during microcode updates. This was > > rather painful to roll out at the time. > > > > An alternative to adding CPU models is to change meaning of existing CPU > > models. QEMU has a way todo this by tieing the change to machine types, and > > it has in fact been used to correct mistakes in the specification of CPU > > models in the past, when those mistakes have not had dependancies on microcode > > changes. This is not a particularly attractice way to deal with the errata. > > Short life distros tend to stick with upstream QEMU machine types and won't > > want to diverge by adding their own machine types. This gates them on having > > upstream define the extra machine types which is tricky under embargo. Long > > life distros do typically take on the burden of defining custom machine types, > > but usually only add them when doing major updates. > > > > The pain point with machine types is that the testing matrix grows at O(n^2) > > Using machine types for CPU security errata would significant increase the > > number of machine types and thus the testing matrix. eg if a security fix > > is needed in rhel-7.3, 7.4, 7.5 we can't just add a pc-rhel-7.5.1 machine > > with the fix, as it would not be possible to implement that in 7.3. So we > > would need would need pc-rhel-7.3.1, pc-rhel-7.4.1, pc-rhel-7.5.1, machine > > types, with 7.5 gaining all three. Finally CPU model changes have host > > hardware dependancies and machine types need to be independant of the host, > > since they are decided statically are build time. The only nice thing about > > machine type is that it is reasonably obvious what the "best" machine type > > is as they include a version number in the name, and users automatically get > > the best if they use an unversioned name. > > > > > > What if we can borrow the concept of versioning from machine types and apply > > it to CPU models directly. For example, considering the history of "Haswell" > > in QEMU, if we had versioned things, we would by now have: > > > > Haswell-1.3.0 - first version (37507094f350b75c62dc059f998e7185de3ab60a) > > Haswell-2.2.0 - added 'rdrand' (78a611f1936b3eac8ed78a2be2146a742a85212c_ > > Haswell-2.3.0 - removed 'hle' & 'rtm' (a356850b80b3d13b2ef737dad2acb05e6da03753) > > Haswell-2.5.0 - added 'abm' (becb66673ec30cb604926d247ab9449a60ad8b11 > > Haswell-2.12.0 - added 'spec-ctrl' (ac96c41354b7e4c70b756342d9b686e31ab87458) > > Haswell-3.0.0 - added 'ssbd' (never done) > > OK. > Note that this isn't that different to what happens on some real > hardware where you have different 'steppings' > > > If we followed the machine type approach, then a bare "Haswell" would > > statically resolve at build time to the most recent Haswell-X.X.X version > > associated with the QEMU release. This is unhelpful as we have a direct > > dependancy on the host hardware features. Better would be for a bare > > "Haswell" to be dynamically resolved at runtime, picking the most recent > > version that is capable of launching given the current hardware, KVM/TCG impl > > and QEMU version. > > > > ie -cpu Haswell > > > > should use Haswell-2.5.0 if on silicon with the TSX errata applied, > > but use Haswell-2.12.0 if the Spectre errata is applied in microcode, > > and use Haswell-3.0.0 once Intel finally releases SSBD microcode errata. > > > > Versioning of CPU models as opposed to using arbitrary string suffixes > > (-noTSX, -IBRS) has a number of usability improvements that we would > > gain with versioned machine types, while avoiding exploding the machine > > type matrix. With versioned CPU models we can > > > > - Automatically tailor the best model based on hardware support > > > > - Users always get the best model if they use the bare CPU name > > > > - It is obvious to users which is the "best" / "newest" CPU model > > > > - Avoid combinatorial expansion of machines since same CPU model > > version can be added to all releases without adding machine types. > > > > - Users can still force a specific downgraded model by using the > > fully versioned name. > > > > Such versioning of CPU models would largely "just work" with existing > > libvirt versions, but to libvirt would really want to expand the bare > > CPU name to a versioned CPU name when recording new guest XML, so the > > ABI is preserved long term. > > > > An application like virt-manager which wants a simple UI can forever be > > happy simply giving users a list of bare CPU model names, and allowing > > libvirt / QEMU to automatically expand to the best versioned model for > > their host. > > > > An application like oVirt/OpenStack which wants direct control can allow > > the admin to choice if a bare name, or explicitly picking a versioned name > > if they need to cope with possibility of outdated hosts. > > I fear people are going to find this out the hard way, when they add > a new system into their cluster, a little bit later it gets a VM started > on it, and then they try and migrate it to one of the older machines. > > Now if there was something that could take the CPU defintions from all > the machines in the cluster and tell it which to use/which problems > they had then that might make sense. It would be best for each > higher level not to reinvent that. Libvirt / QEMU have ability to let apps do that. Nova could choose to do that, but it does not wish to take advantage of that right now. The nova admin already has make an explicit config to set the maximum Nova database object schema version, and maximum machine type version across the cluster. The max CPU model version fits in with what they are already doing quite well. Essentially the admin knows what versions of QEMU they have across their cluster - generally it is jus N, and N+1 during course of a version upgrade. So it is quite easy for the admin to know what the max version should be. In fact the admin doens't really need to be involved in it at all - the cluster provisioning tool already knows the old + new host versions, so can set it automatically at the right times during upgrade process. > Would you restrict the combinations to cut down the test matrix - e.g. > not allow Haswell-3.0.0 on anything prior to a 2.12 machine type? The key point of having CPU versions separate from machine type versions, is to allow arbitrary mixing of versions. For example when we issued Spectre updates for RHEL, we had to add new CPU model to something like 12 different QEMU versions, and we needed the same CPU model changes to work with all the different machine types we had across those versions. So restricting CPU versions to only certain machine types would defeat the puporse of having CPU versions. 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 :| -- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list
