Re: [PATCH 2/2] domain: expose newly introduced hypercalls as XENFEAT

["Roger Pau Monné" <roger.pau@citrix.com>]

On Fri, Oct 06, 2023 at 11:47:48AM +0100, Andrew Cooper wrote:
> On 06/10/2023 10:13 am, Roger Pau Monne wrote:
> > XENFEAT_runstate_phys_area is exposed to all architectures, while
> > XENFEAT_vcpu_time_phys_area is currnelty only implemented for x86, and hence
> > the feature flag is also only exposed on x86.
> >
> > Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
> 
> Thanks for doing this.
> 
> > ---
> >  CHANGELOG.md                  | 2 ++
> >  xen/common/kernel.c           | 6 +++++-
> >  xen/include/public/features.h | 9 +++++++++
> >  3 files changed, 16 insertions(+), 1 deletion(-)
> >
> > diff --git a/CHANGELOG.md b/CHANGELOG.md
> > index e33cf4e1b113..41da710426f6 100644
> > --- a/CHANGELOG.md
> > +++ b/CHANGELOG.md
> > @@ -31,6 +31,8 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
> >   - Add Intel Hardware P-States (HWP) cpufreq driver.
> >   - On Arm, experimental support for dynamic addition/removal of Xen device tree
> >     nodes using a device tree overlay binary (.dtbo).
> > + - Introduce two new hypercalls to map the vCPU runstate and time areas by
> > +   physical rather than linear addresses.
> 
> I'd suggest linear/virtual here.  Linear is the correct term in x86, but
> virtual is the correct term in pretty much every other architecture.
> 
> > diff --git a/xen/include/public/features.h b/xen/include/public/features.h
> > index d2a9175aae67..cffb2f14a562 100644
> > --- a/xen/include/public/features.h
> > +++ b/xen/include/public/features.h
> > @@ -111,6 +111,15 @@
> >  #define XENFEAT_not_direct_mapped         16
> >  #define XENFEAT_direct_mapped             17
> >  
> > +/*
> > + * Signal whether the hypervisor implements the following hypercalls:
> 
> This is not what the hypervisor implements.  It's what the domain is
> permitted to use.
> 
> There also needs to be a matching patch to public/vcpu.h to require
> implementations to check for these feature bits before using the new
> hypercalls.
> 
> Also,
> 
> diff --git a/xen/common/domain.c b/xen/common/domain.c
> index b8281d7cff9d..df994bd30fd2 100644
> --- a/xen/common/domain.c
> +++ b/xen/common/domain.c
> @@ -1998,6 +1998,10 @@ long common_vcpu_op(int cmd, struct vcpu *v
>      {
>          struct vcpu_register_runstate_memory_area area;
>  
> +        rc = -ENOSYS;
> +        if ( 0 /* TODO: Dom's XENFEAT_runstate_phys_area setting */ )
> +            break;
> +
>          rc = -EFAULT;
>          if ( copy_from_guest(&area.addr.p, arg, 1) )
>              break;
> 
> and a matching one for XENFEAT_vcpu_time_phys_area because I'm even more
> serious about this becoming a domain controllable setting following what
> OSSTest had to say overnight.

While this is all fine, please note that the newly added code
{,un}map_guest_area() is also used by the existing
VCPUOP_register_vcpu_info hypercall, and that one can't be disabled.

IOW: even if we add knobs to make the new hypercalls selectable, most
of the newly added code could still be reached from
VCPUOP_register_vcpu_info.

Thanks, Roger.