Re: [PATCH v10 2/9] KVM: Introduce per-page memory attributes

[Sean Christopherson]

On Mon, Feb 13, 2023, Isaku Yamahata wrote:
> On Fri, Feb 10, 2023 at 12:35:30AM +0000,
> Sean Christopherson <seanjc@google.com> wrote:
> 
> > On Wed, Feb 08, 2023, Isaku Yamahata wrote:
> > > On Fri, Dec 02, 2022 at 02:13:40PM +0800,
> > > Chao Peng <chao.p.peng@linux.intel.com> wrote:
> > > 
> > > > +static int kvm_vm_ioctl_set_mem_attributes(struct kvm *kvm,
> > > > +                                          struct kvm_memory_attributes 
> > > > *attrs)
> > > > +{
> > > > +       gfn_t start, end;
> > > > +       unsigned long i;
> > > > +       void *entry;
> > > > +       u64 supported_attrs = kvm_supported_mem_attributes(kvm);
> > > > +
> > > > +       /* flags is currently not used. */
> > > > +       if (attrs->flags)
> > > > +               return -EINVAL;
> > > > +       if (attrs->attributes & ~supported_attrs)
> > > > +               return -EINVAL;
> > > > +       if (attrs->size == 0 || attrs->address + attrs->size < 
> > > > attrs->address)
> > > > +               return -EINVAL;
> > > > +       if (!PAGE_ALIGNED(attrs->address) || !PAGE_ALIGNED(attrs->size))
> > > > +               return -EINVAL;
> > > > +
> > > > +       start = attrs->address >> PAGE_SHIFT;
> > > > +       end = (attrs->address + attrs->size - 1 + PAGE_SIZE) >> 
> > > > PAGE_SHIFT;
> > > > +
> > > > +       entry = attrs->attributes ? xa_mk_value(attrs->attributes) : 
> > > > NULL;
> > > > +
> > > > +       mutex_lock(&kvm->lock);
> > > > +       for (i = start; i < end; i++)
> > > > +               if (xa_err(xa_store(&kvm->mem_attr_array, i, entry,
> > > > +                                   GFP_KERNEL_ACCOUNT)))
> > > > +                       break;
> > > > +       mutex_unlock(&kvm->lock);
> > > > +
> > > > +       attrs->address = i << PAGE_SHIFT;
> > > > +       attrs->size = (end - i) << PAGE_SHIFT;
> > > > +
> > > > +       return 0;
> > > > +}
> > > > +#endif /* CONFIG_HAVE_KVM_MEMORY_ATTRIBUTES */
> > > > +
> > > 
> > > If memslot isn't private, it should return error if private attribute is 
> > > set.
> > 
> > Why?  I'd rather keep the two things separate.  If we enforce this sort of 
> > thing
> > at KVM_SET_MEMORY_ATTRIBUTES, then we also have to enforce it at
> > KVM_SET_USER_MEMORY_REGION.
> 
> For device assignment via shared GPA, non-private memory slot needs to be
> allowed.

That doesn't say anything about why setting attributes needs to poke into the
memslot.  The page fault path already kicks out to userspace if there's a
discrepancy between the attributes and the memslot, why is that insufficient?