Re: [Qemu-devel] Cascadelake-Server missing MSR based features ?

[Eduardo Habkost]

On Wed, Mar 20, 2019 at 01:30:26PM +0000, Daniel P. Berrangé wrote:
> On Tue, Mar 19, 2019 at 02:20:01PM -0300, Eduardo Habkost wrote:
> > On Tue, Mar 19, 2019 at 04:16:14PM +0000, Daniel P. Berrangé wrote:
> > > On Tue, Mar 19, 2019 at 10:35:33PM +0800, Tao Xu wrote:
> > > > On 3/19/2019 8:16 PM, Daniel P. Berrangé wrote:
> > > > > The Cascadelake-Server CPU was added last year in
> > > > > 
> > > > >    commit c7a88b52f62b30c04158eeb07f73e3f72221b6a8
> > > > >    Author: Tao Xu <address@hidden>
> > > > >    Date:   Wed Sep 19 11:11:22 2018 +0800
> > > > > 
> > > > >      i386: Add new model of Cascadelake-Server
> > > > > 
> > > > > The commit message says that MSR based features are initially
> > > > > omitted from its definition:
> > > > > 
> > > > >    [quote]
> > > > >      On Cascadelake, some capabilities (RDCL_NO, IBRS_ALL, RSBA,
> > > > >      SKIP_L1DFL_VMENTRY and SSB_NO) are enumerated by MSR.
> > > > >      These features rely on MSR based feature support patch.
> > > > >      Will be added later after that patch's in.
> > > > >      
> > > > > http://lists.nongnu.org/archive/html/qemu-devel/2018-09/msg00074.html
> > > > >    [/quote]
> > > 
> > > [snip]
> > > 
> > > > > Can anyone shed light on what the status is wrt exposing the MSR based
> > > > > features in Cascadelake-Server ?
> > > > 
> > > > Hi, I sent the patch to add the MSR based features with the Update 
> > > > stepping
> > > > patch together last year, and we discussed a lot about it:
> > > > 
> > > > [1] https://patchwork.kernel.org/patch/10743427/
> > > 
> > > Ahh, I see, so none of the MSR based feature stuff is migratable.
> > > 
> > > If migration isn't needed, then they can use "-cpu host" which gets all
> > > features automatically.
> > > 
> > > If migration is needed, then the app is using a named CPU model, and no
> > > MSR based features should ever be used.
> > > 
> > > Also means that libvirt must be careful to *not* enable MSR based
> > > features with "host-model" as this is supposde to be a migratable
> > > version of "host-passthrough"
> > > 
> > > Is there any ETA on fixing things so that the MSR features can be
> > > used with named models ?
> > > 
> > 
> > Making it migratable and usable in named CPU models is the easy
> > part.  This series is probably enough:
> > 
> > http://lists.nongnu.org/archive/html/qemu-devel/2019-01/msg06701.html
> > ("[PATCH 0/2] i386: arch_capabilities fixes + migratability")
> > 
> > I have just realized that I have left my own series fall through
> > the cracks.  At least patch 1/2 is a bug fix, so I will merge it
> > for -rc1.

I've decided to queue both patches for -rc1.

> > 
> > The hard part is enabling the flag by default on existing CPU
> > models.  We probably need to implement CPU model versioning in
> > QEMU 4.1 to do that properly.  See discussion at:
> > 
> > https://www.mail-archive.com/address@hidden/msg603291.html
> > ("Re: [Qemu-devel] [PATCH 2/2] i386: Add some MSR based features
> > on Cascadelake-Server CPU model")
> 
> Ok, so once the migration problem is fixed, IIUC the main blocker is
> that we want to be able to support use of Cascadelake-Server CPU model
> on older kernels that lacked the ARCH_CAPABILITIES support in KVM.
> We don't have a nice way to achieve that right now, but the versioned
> CPU models concept would allow for a solution.
> 
> IOW, once migration is fixed, it is fine for mgmt apps to enable the
> use of the MSR features by turning them on manually for Cascadelake-Server
> if they know they don't care about older kernels.

After we let management software enable arch_capabilities and set
individual bits, the main challenge would be dealing with RSBA
safely.  In theory we shouldn't let a VM with RSBA=off run on a
host with RSBA=on.  But we have been allowing this to happen for
a few QEMU releases, and it would be difficult to change this
behavior without breaking existing configurations.

-- 
Eduardo