From: Ashish Kalra <[email protected]>
Some BIOSes allow the end user to set the minimum SEV ASID value
(CPUID 0x8000001F_EDX) to be greater than the maximum number of
encrypted guests, or maximum SEV ASID value (CPUID 0x8000001F_ECX)
in order to dedicate all the SEV ASIDs to SEV-ES or SEV-SNP.
The SEV support, as coded, does not handle the case where the minimum
SEV ASID value can be greater than the maximum SEV ASID value.
As a result, the following confusing message is issued:
[ 30.715724] kvm_amd: SEV enabled (ASIDs 1007 - 1006)
Fix the support to properly handle this case.
Fixes: 916391a2d1dc ("KVM: SVM: Add support for SEV-ES capability in KVM")
Signed-off-by: Ashish Kalra <[email protected]>
Cc: [email protected]
---
arch/x86/kvm/svm/sev.c | 26 ++++++++++++++++++--------
1 file changed, 18 insertions(+), 8 deletions(-)
diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
index 4900c078045a..ad41008ca0d9 100644
--- a/arch/x86/kvm/svm/sev.c
+++ b/arch/x86/kvm/svm/sev.c
@@ -59,10 +59,14 @@ module_param_named(sev_es, sev_es_enabled, bool, 0444);
/* enable/disable SEV-ES DebugSwap support */
static bool sev_es_debug_swap_enabled = true;
module_param_named(debug_swap, sev_es_debug_swap_enabled, bool, 0444);
+
+/* When true, at least one type of SEV guest is enabled to run */
+static bool sev_guests_enabled;
#else
#define sev_enabled false
#define sev_es_enabled false
#define sev_es_debug_swap_enabled false
+#define sev_guests_enabled false
#endif /* CONFIG_KVM_AMD_SEV */
static u8 sev_enc_bit;
@@ -1854,7 +1858,7 @@ int sev_mem_enc_ioctl(struct kvm *kvm, void __user *argp)
struct kvm_sev_cmd sev_cmd;
int r;
- if (!sev_enabled)
+ if (!sev_guests_enabled)
return -ENOTTY;
if (!argp)
@@ -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) {
kvm_cpu_cap_clear(X86_FEATURE_SEV);
+ return;
+ }
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);
if (boot_cpu_has(X86_FEATURE_SEV_ES))
pr_info("SEV-ES %s (ASIDs %u - %u)\n",
sev_es_supported ? "enabled" : "disabled",
@@ -2270,6 +2279,7 @@ void __init sev_hardware_setup(void)
sev_enabled = sev_supported;
sev_es_enabled = sev_es_supported;
+ sev_guests_enabled = sev_enabled || sev_es_enabled;
if (!sev_es_enabled || !cpu_feature_enabled(X86_FEATURE_DEBUG_SWAP) ||
!cpu_feature_enabled(X86_FEATURE_NO_NESTED_DATA_BP))
sev_es_debug_swap_enabled = false;
@@ -2278,7 +2288,7 @@ void __init sev_hardware_setup(void)
void sev_hardware_unsetup(void)
{
- if (!sev_enabled)
+ if (!sev_guests_enabled)
return;
/* No need to take sev_bitmap_lock, all VMs have been destroyed. */
@@ -2293,7 +2303,7 @@ void sev_hardware_unsetup(void)
int sev_cpu_init(struct svm_cpu_data *sd)
{
- if (!sev_enabled)
+ if (!sev_guests_enabled)
return 0;
sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
--
2.34.1
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.
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.
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.
> + }
> 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;
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);
if (boot_cpu_has(X86_FEATURE_SEV_ES))
pr_info("SEV-ES %s (ASIDs %u - %u)\n",
sev_es_supported ? "enabled" : "disabled",
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.
>
> 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.
>
>> + }
>> 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. 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.
> 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 ?
Thanks, Ashish
> if (boot_cpu_has(X86_FEATURE_SEV_ES))
> pr_info("SEV-ES %s (ASIDs %u - %u)\n",
> sev_es_supported ? "enabled" : "disabled",
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.
On 1/3/2024 3:10 PM, Sean Christopherson wrote:
> 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.
Actually, sev_asid_new() returning failure to sev_guest_init() will
cause it to return -EBUSY to user.
Thanks, Ashish
>> 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.
On Wed, Jan 03, 2024, Ashish Kalra wrote:
> On 1/3/2024 3:10 PM, Sean Christopherson wrote:
> > > > 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.
>
> Actually, sev_asid_new() returning failure to sev_guest_init() will cause it
> to return -EBUSY to user.
Argh, I see it now. That too should be fixed, e.g.
diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
index d0c580607f00..79eb11083ad5 100644
--- a/arch/x86/kvm/svm/sev.c
+++ b/arch/x86/kvm/svm/sev.c
@@ -246,21 +246,20 @@ static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
{
struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
- int asid, ret;
+ int ret;
if (kvm->created_vcpus)
return -EINVAL;
- ret = -EBUSY;
if (unlikely(sev->active))
- return ret;
+ return -EINVAL;
sev->active = true;
sev->es_active = argp->id == KVM_SEV_ES_INIT;
- asid = sev_asid_new(sev);
- if (asid < 0)
+ ret = sev_asid_new(sev);
+ if (ret < 0)
goto e_no_asid;
- sev->asid = asid;
+ sev->asid = ret;
ret = sev_platform_init(&argp->error);
if (ret)