Re: [PATCH] x86/sev: Add support for allowing zero SEV ASIDs.

[Sean Christopherson]

On Wed, Jan 03, 2024, Ashish Kalra wrote:
> Hello Sean,
> 
> On 1/2/2024 6:30 PM, Sean Christopherson wrote:
> > On Tue, Jan 02, 2024, Ashish Kalra wrote:
> > > @@ -2172,8 +2176,10 @@ void sev_vm_destroy(struct kvm *kvm)
> > >   void __init sev_set_cpu_caps(void)
> > >   {
> > > -	if (!sev_enabled)
> > > +	if (!sev_guests_enabled) {
> > Ugh, what a mess.  The module param will show sev_enabled=false, but the caps
> > and CPUID will show SEV=true.
> > 
> > And this is doubly silly because "sev_enabled" is never actually checked, e.g.
> > if misc cgroup support is disabled, KVM_SEV_INIT will try to reclaim ASIDs and
> > eventually fail with -EBUSY, which is super confusing to users.
> 
> But this is what we expect that KVM_SEV_INIT will fail. In this case,
> sev_asid_new() will not actually try to reclaim any ASIDs as sev_misc_cg_try_charge()
> will fail before any ASID bitmap walking/reclamation and return an error which
> will eventually return -EBUSY to the user.

Please read what I wrote.  "if misc cgroup support is disabled", i.e. if
CONFIG_CGROUP_MISC=n, then sev_misc_cg_try_charge() is a nop.

> > The other weirdness is that KVM can cause sev_enabled=false && sev_es_enabled=true,
> > but if *userspace* sets sev_enabled=false then sev_es_enabled is also forced off.
> But that is already the behavior without this patch applied.
> > 
> > In other words, the least awful option seems to be to keep sev_enabled true :-(
> > 
> > >   		kvm_cpu_cap_clear(X86_FEATURE_SEV);
> > > +		return;
> > This is blatantly wrong, as it can result in KVM advertising SEV-ES if SEV is
> > disabled by the user.
> No, this ensures that we don't advertise any SEV capability if neither
> SEV/SEV-ES or in future SNP is enabled.

No, it does not.  There is an early return statement here that prevents KVM from
invoking kvm_cpu_cap_clear() for X86_FEATURE_SEV_ES.  Do I think userspace will
actually be tripped up by seeing SEV_ES without SEV?  No.  Is it unnecessarily
confusing?  Yes.

> > > +	}
> > >   	if (!sev_es_enabled)
> > >   		kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
> > >   }
> > > @@ -2229,9 +2235,11 @@ void __init sev_hardware_setup(void)
> > >   		goto out;
> > >   	}
> > > -	sev_asid_count = max_sev_asid - min_sev_asid + 1;
> > > -	WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count));
> > > -	sev_supported = true;
> > > +	if (min_sev_asid <= max_sev_asid) {
> > > +		sev_asid_count = max_sev_asid - min_sev_asid + 1;
> > > +		WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count));
> > > +		sev_supported = true;
> > > +	}
> > >   	/* SEV-ES support requested? */
> > >   	if (!sev_es_enabled)
> > > @@ -2262,7 +2270,8 @@ void __init sev_hardware_setup(void)
> > >   	if (boot_cpu_has(X86_FEATURE_SEV))
> > >   		pr_info("SEV %s (ASIDs %u - %u)\n",
> > >   			sev_supported ? "enabled" : "disabled",
> > > -			min_sev_asid, max_sev_asid);
> > > +			sev_supported ? min_sev_asid : 0,
> > > +			sev_supported ? max_sev_asid : 0);
> > I honestly think we should print the "garbage" values.  The whole point of
> > printing the min/max SEV ASIDs was to help users understand why SEV is disabled,
> > i.e. printing zeroes is counterproductive.
> > 
> > >   	if (boot_cpu_has(X86_FEATURE_SEV_ES))
> > >   		pr_info("SEV-ES %s (ASIDs %u - %u)\n",
> > >   			sev_es_supported ? "enabled" : "disabled",
> > It's all a bit gross, but I think we want something like this (I'm definitely
> > open to suggestions though):
> > 
> > diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
> > index d0c580607f00..bfac6d17462a 100644
> > --- a/arch/x86/kvm/svm/sev.c
> > +++ b/arch/x86/kvm/svm/sev.c
> > @@ -143,8 +143,20 @@ static void sev_misc_cg_uncharge(struct kvm_sev_info *sev)
> >   static int sev_asid_new(struct kvm_sev_info *sev)
> >   {
> > -       int asid, min_asid, max_asid, ret;
> > +       /*
> > +        * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
> > +        * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.  Note, the
> > +        * min ASID can end up larger than the max if basic SEV support is
> > +        * effectively disabled by disallowing use of ASIDs for SEV guests.
> > +        */
> > +       unsigned int min_asid = sev->es_active ? 1 : min_sev_asid;
> > +       unsigned int max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
> > +       unsigned int asid;
> >          bool retry = true;
> > +       int ret;
> > +
> > +       if (min_asid > max_asid)
> > +               return -ENOTTY;
> 
> This will still return -EBUSY to user.

Huh?  The above is obviously -ENOTTY, and I don't see anything in the call stack
that will convert it to -EBUSY.

> This check here or the failure return from sev_misc_cg_try_charge() are quite
> similar in that sense.
> 
> My point is that the same is achieved quite cleanly with
> sev_misc_cg_try_charge() too.

"Without additional effort" is not synonymous with "cleanly".  Relying on an
accounting restriction that is completely orthogonal to basic functionality is
not "clean".

> >          WARN_ON(sev->misc_cg);
> >          sev->misc_cg = get_current_misc_cg();
> > @@ -157,12 +169,6 @@ static int sev_asid_new(struct kvm_sev_info *sev)
> >          mutex_lock(&sev_bitmap_lock);
> > -       /*
> > -        * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
> > -        * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
> > -        */
> > -       min_asid = sev->es_active ? 1 : min_sev_asid;
> > -       max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
> >   again:
> >          asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
> >          if (asid > max_asid) {
> > @@ -2232,8 +2238,10 @@ void __init sev_hardware_setup(void)
> >                  goto out;
> >          }
> > -       sev_asid_count = max_sev_asid - min_sev_asid + 1;
> > -       WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count));
> > +       if (min_sev_asid <= max_sev_asid) {
> > +               sev_asid_count = max_sev_asid - min_sev_asid + 1;
> > +               WARN_ON_ONCE(misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count));
> > +       }
> >          sev_supported = true;
> >          /* SEV-ES support requested? */
> > @@ -2264,8 +2272,9 @@ void __init sev_hardware_setup(void)
> >   out:
> >          if (boot_cpu_has(X86_FEATURE_SEV))
> >                  pr_info("SEV %s (ASIDs %u - %u)\n",
> > -                       sev_supported ? "enabled" : "disabled",
> > -                       min_sev_asid, max_sev_asid);
> > +                       sev_supported ? (min_sev_asid <= max_sev_asid ? "enabled" : "unusable") : "disabled",
> > +                       sev_supported ? min_sev_asid : 0,
> > +                       sev_supported ? max_sev_asid : 0);
> 
> We are not showing min and max ASIDs for SEV as {0,0} with this patch as
> sev_supported is true ?

Yes, and that is deliberate.  See this from above:

 : I honestly think we should print the "garbage" values.  The whole point of  
 : printing the min/max SEV ASIDs was to help users understand why SEV is disabled,
 : i.e. printing zeroes is counterproductive.