Re: [PATCH] x86: Conditionalise init_dom0_cpu_policy()

["Daniel P. Smith" <dpsmith@xxxxxxxxxxxxxxxxxxxx>]

On 7/25/25 06:56, Roger Pau Monné wrote:
> On Fri, Jul 25, 2025 at 12:02:18PM +0200, Alejandro Vallejo wrote:
> > On Wed Jul 23, 2025 at 9:18 AM CEST, Roger Pau Monné wrote:
> > > On Thu, Jul 17, 2025 at 07:58:24PM +0200, Alejandro Vallejo wrote:
> > > > Later patches will keep refactoring create_dom0()
> > > > until it can create arbitrary domains. This is one
> > > > small step in that direction.
> > > >
> > > > Signed-off-by: Alejandro Vallejo <alejandro.garciavallejo@xxxxxxx>
> > > > ---
> > > >   xen/arch/x86/setup.c | 3 ++-
> > > >   1 file changed, 2 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/xen/arch/x86/setup.c b/xen/arch/x86/setup.c
> > > > index c6890669b9..6943ffba79 100644
> > > > --- a/xen/arch/x86/setup.c
> > > > +++ b/xen/arch/x86/setup.c
> > > > @@ -1054,7 +1054,8 @@ static struct domain *__init create_dom0(struct boot_info 
> > > > *bi)
> > > >       if ( IS_ERR(d) )
> > > >           panic("Error creating d%u: %ld\n", bd->domid, PTR_ERR(d));
> > > >   
> > > > -    init_dom0_cpuid_policy(d);
> > > > +    if ( pv_shim || d->cdf & (CDF_privileged | CDF_hardware) )
> > > You possibly want this to be:
> > >
> > > (d->cdf & (CDF_privileged | CDF_hardware)) == (CDF_privileged | CDF_hardware)
> > >
> > > To ensure the contents of dom0_cpuid_cmdline is only applied to dom0,
> > > and not to the hardware or control domains.  I assume it should be
> > > possible to pass a different set of cpuid options for the hardware vs
> > > the control domains.
> > >
> > > Thanks, Roger.
> > Why only a hwdom+ctldom, surely a single hwdom should get it too.
> hm, not really I think: a late hardware domain would get any custom
> cpuid options from the toolstack that created it, or in the
> hyperlaunch case from the provided configuration, but not from the
> dom0-cpuid command line option I would expect.  Otherwise you have two
> different sources of cpuid options, the inheritance from dom0-cpuid,
> plus whatever is provided from the hardware domain configuration.
Yes, this has been a sticking point for me and never got any good 
answers thus far. Should the dom0 related xen command line options only 
apply when not booting via hyperlaunch. If the answer is no, then you're 
in this area with some dom0 options that really are applicable to hwdom 
vs ctldom and vice-a-versa. Some could even be suggested to apply to 
both. And then, I don't believe there really is a consensus one which 
options apply to which domains. Over the years working on this, I have 
been an advocate that commandline adjustments allow for quicker 
troubleshooting by the user/administrator. In the last version of the 
multidomain construction RFC, I am growing more and more to advocate for 
my initial proposition, that dom0 options only apply when not using 
hyperlaunch.
> > I can see
> > the argument for a ctldom not getting it. For our use case having dom0
> > disaggregated is of the essence, so what happens with a hwdom that is not a
> > ctldom is fairly important.
> >
> > That said, I'm thinking moving in a different direction and have a generic
> > init_cpuid_policy() that internally checks for hw or control, or leave the
> > default policy or something else.
> Right, so instead of introducing init_{hwdom,ctrldom}_cpuid_policy()
> equivalents you would rather generalize init_dom0_cpuid_policy() so
> it's used by all domains.
>
> > This would remove the conditional and allow much finer selection. e.g: a domain
> > brought up through a "nomigrate" DTB node (TBD: nonexisting binding atm) would
> > get itsc reported, just as dom0 does today.
> We might want to enforce such no migration attribute at the hypervisor
> level (by adding a new domain field to signal it possibly?), as this
> is all toolstack knowledge ATM.
>
> On a related tangent: domains brought up using hyperlaunch will have a
> config file, capable of expressing options like cpuid features I
> expect, at which point ITSC could be set in the config file and Xen
> won't need to do any guessing?
The goal has never been to support all the configuration options that a 
full toolstack can adjust for the VM. I have always felt it should be a 
light-weight toolstack to allow essential configuration. I still 
advocate for the boot domain that would have capabilities like parsing 
the AML, walking PCI device tree, etc. It would then handle any more 
complicated initialization/configuration of the VMs before exiting and 
allowing all the domains to start
v/r,
dps