[Qemu-devel] [PATCH 0/2] target-i386: Fix default Hypervisor level for kvm

[Qemu-devel] [PATCH 0/2] target-i386: Fix default Hypervisor level for kvm

[Don Slutz]

Looking at http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
The new value for EAX is 0x40000001.

This depends on http://lists.gnu.org/archive/html/qemu-devel/2012-09/msg02497.html

As far as I known it is #5.  It depends on (1), (2), (3) and (4).

Based on cpu-queue[1] branch.
(From http://lists.gnu.org/archive/html/qemu-devel/2012-09/msg02639.html)

[1] https://github.com/ehabkost/qemu/commits/cpu-queue
    My branch is now based on Andreas's qom-cpu branch from
    https://github.com/afaerber/qemu-cpu/commits/qom-cpu


Don Slutz (2):
  target-i386: Fix default Hypervisor level for accel=kvm.
  target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

 target-i386/cpu.c |   12 +++++++++++-
 target-i386/kvm.c |    2 +-
 2 files changed, 12 insertions(+), 2 deletions(-)

[Qemu-devel] [PATCH 1/2] target-i386: Fix default Hypervisor level for accel=kvm.

[Don Slutz]

>From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
---
 target-i386/kvm.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/target-i386/kvm.c b/target-i386/kvm.c
index 761a9b1..0c9f5dd 100644
--- a/target-i386/kvm.c
+++ b/target-i386/kvm.c
@@ -392,7 +392,7 @@ int kvm_arch_init_vcpu(CPUX86State *env)
     c->function = KVM_CPUID_SIGNATURE;
     if (env->cpuid_hv_level == 0) {
         memcpy(signature, "KVMKVMKVM\0\0\0", 12);
-        c->eax = 0;
+        c->eax = KVM_CPUID_FEATURES;
         c->ebx = signature[0];
         c->ecx = signature[1];
         c->edx = signature[2];
-- 
1.7.1

Re: [Qemu-devel] [PATCH 1/2] target-i386: Fix default Hypervisor level for accel=kvm.

[Eduardo Habkost]

On Tue, Sep 18, 2012 at 10:49:52AM -0400, Don Slutz wrote:
> From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
> EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
> ---
>  target-i386/kvm.c |    2 +-
>  1 files changed, 1 insertions(+), 1 deletions(-)
> 
> diff --git a/target-i386/kvm.c b/target-i386/kvm.c
> index 761a9b1..0c9f5dd 100644
> --- a/target-i386/kvm.c
> +++ b/target-i386/kvm.c
> @@ -392,7 +392,7 @@ int kvm_arch_init_vcpu(CPUX86State *env)
>      c->function = KVM_CPUID_SIGNATURE;
>      if (env->cpuid_hv_level == 0) {
>          memcpy(signature, "KVMKVMKVM\0\0\0", 12);
> -        c->eax = 0;
> +        c->eax = KVM_CPUID_FEATURES;

This makes the CPUID bits to suddenly change, when live-migrating to a
newer QEMU version.

Strictly speaking, this is never supposed to happen, but... on both
cases the meaning of the bits are the same (0 is documented as
equivalent to KVM_CPUID_FEATURES) and probably the guest will look at
them only once on boot. Do we really want to add migration-compatibility
code for this?


>          c->ebx = signature[0];
>          c->ecx = signature[1];
>          c->edx = signature[2];
> -- 
> 1.7.1
> 

-- 
Eduardo

Re: [Qemu-devel] [PATCH 1/2] target-i386: Fix default Hypervisor level for accel=kvm.

[Don Slutz]

On 09/18/12 11:05, Eduardo Habkost wrote:
> On Tue, Sep 18, 2012 at 10:49:52AM -0400, Don Slutz wrote:
>>  From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
>> EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
>> ---
>>   target-i386/kvm.c |    2 +-
>>   1 files changed, 1 insertions(+), 1 deletions(-)
>>
>> diff --git a/target-i386/kvm.c b/target-i386/kvm.c
>> index 761a9b1..0c9f5dd 100644
>> --- a/target-i386/kvm.c
>> +++ b/target-i386/kvm.c
>> @@ -392,7 +392,7 @@ int kvm_arch_init_vcpu(CPUX86State *env)
>>       c->function = KVM_CPUID_SIGNATURE;
>>       if (env->cpuid_hv_level == 0) {
>>           memcpy(signature, "KVMKVMKVM\0\0\0", 12);
>> -        c->eax = 0;
>> +        c->eax = KVM_CPUID_FEATURES;
> This makes the CPUID bits to suddenly change, when live-migrating to a
> newer QEMU version.
>
> Strictly speaking, this is never supposed to happen, but... on both
> cases the meaning of the bits are the same (0 is documented as
> equivalent to KVM_CPUID_FEATURES) and probably the guest will look at
> them only once on boot. Do we really want to add migration-compatibility
> code for this?
>
My vote would be no; because this should be ok and the tests that I know 
of are all at boot time.  I will look into adding 
migration-compatibility.  I also do not have any direct need for this 
change, it just looked like the right thing to do.
>>           c->ebx = signature[0];
>>           c->ecx = signature[1];
>>           c->edx = signature[2];
>> -- 
>> 1.7.1
>>

[Qemu-devel] [PATCH 2/2] target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

[Don Slutz]

>From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.

If kvm is not configured, the additional option of hypervisor-level=1
(or hypervisor-level=0x40000001) needs to be specified to get this.
---
 target-i386/cpu.c |   12 +++++++++++-
 1 files changed, 11 insertions(+), 1 deletions(-)

diff --git a/target-i386/cpu.c b/target-i386/cpu.c
index 6e43eff..d73b0a8 100644
--- a/target-i386/cpu.c
+++ b/target-i386/cpu.c
@@ -1248,7 +1248,12 @@ static char *x86_cpuid_get_hv_vendor(Object *obj, Error **errp)
                env->cpuid_hv_level == CPUID_HV_LEVEL_XEN) {
         pstrcpy(value, sizeof(value), "xen");
     } else if (!strcmp(value, CPUID_HV_VENDOR_KVM) &&
-               env->cpuid_hv_level == 0) {
+#if defined(CONFIG_KVM)
+               env->cpuid_hv_level == KVM_CPUID_FEATURES
+#else
+               env->cpuid_hv_level == 0
+#endif
+	    ) {
         pstrcpy(value, sizeof(value), "kvm");
     }
     return value;
@@ -1281,6 +1286,11 @@ static void x86_cpuid_set_hv_vendor(Object *obj, const char *value,
         }
         pstrcpy(adj_value, sizeof(adj_value), CPUID_HV_VENDOR_XEN);
     } else if (!strcmp(value, "kvm")) {
+#if defined(CONFIG_KVM)
+        if (env->cpuid_hv_level == 0) {
+            env->cpuid_hv_level = KVM_CPUID_FEATURES;
+        }
+#endif
         pstrcpy(adj_value, sizeof(adj_value), CPUID_HV_VENDOR_KVM);
     } else {
         pstrcpy(adj_value, sizeof(adj_value), value);
-- 
1.7.1

Re: [Qemu-devel] [PATCH 2/2] target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

[Eduardo Habkost]

On Tue, Sep 18, 2012 at 10:49:53AM -0400, Don Slutz wrote:
> From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
> EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
> 
> If kvm is not configured, the additional option of hypervisor-level=1
> (or hypervisor-level=0x40000001) needs to be specified to get this.
> ---
>  target-i386/cpu.c |   12 +++++++++++-
>  1 files changed, 11 insertions(+), 1 deletions(-)
> 
> diff --git a/target-i386/cpu.c b/target-i386/cpu.c
> index 6e43eff..d73b0a8 100644
> --- a/target-i386/cpu.c
> +++ b/target-i386/cpu.c
> @@ -1248,7 +1248,12 @@ static char *x86_cpuid_get_hv_vendor(Object *obj, Error **errp)
>                 env->cpuid_hv_level == CPUID_HV_LEVEL_XEN) {
>          pstrcpy(value, sizeof(value), "xen");
>      } else if (!strcmp(value, CPUID_HV_VENDOR_KVM) &&
> -               env->cpuid_hv_level == 0) {
> +#if defined(CONFIG_KVM)
> +               env->cpuid_hv_level == KVM_CPUID_FEATURES
> +#else
> +               env->cpuid_hv_level == 0
> +#endif
> +	    ) {
>          pstrcpy(value, sizeof(value), "kvm");
>      }
>      return value;
> @@ -1281,6 +1286,11 @@ static void x86_cpuid_set_hv_vendor(Object *obj, const char *value,
>          }
>          pstrcpy(adj_value, sizeof(adj_value), CPUID_HV_VENDOR_XEN);
>      } else if (!strcmp(value, "kvm")) {
> +#if defined(CONFIG_KVM)
> +        if (env->cpuid_hv_level == 0) {
> +            env->cpuid_hv_level = KVM_CPUID_FEATURES;
> +        }
> +#endif

If CPUID[0x40000000].EAX set to 0 is documented as equivalent to having
it set to 0x40000001 (KVM_CPUID_FEATURES), why the confusing checks for
CONFIG_KVM? Why not always set it to KVM_CPUID_FEATURES?

-- 
Eduardo

Re: [Qemu-devel] [PATCH 2/2] target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

[Don Slutz]

On 09/18/12 13:00, Eduardo Habkost wrote:
> On Tue, Sep 18, 2012 at 10:49:53AM -0400, Don Slutz wrote:
>>  From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
>> EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
>>
>> If kvm is not configured, the additional option of hypervisor-level=1
>> (or hypervisor-level=0x40000001) needs to be specified to get this.
>> ---
>>   target-i386/cpu.c |   12 +++++++++++-
>>   1 files changed, 11 insertions(+), 1 deletions(-)
>>
>> diff --git a/target-i386/cpu.c b/target-i386/cpu.c
>> index 6e43eff..d73b0a8 100644
>> --- a/target-i386/cpu.c
>> +++ b/target-i386/cpu.c
>> @@ -1248,7 +1248,12 @@ static char *x86_cpuid_get_hv_vendor(Object *obj, Error **errp)
>>                  env->cpuid_hv_level == CPUID_HV_LEVEL_XEN) {
>>           pstrcpy(value, sizeof(value), "xen");
>>       } else if (!strcmp(value, CPUID_HV_VENDOR_KVM) &&
>> -               env->cpuid_hv_level == 0) {
>> +#if defined(CONFIG_KVM)
>> +               env->cpuid_hv_level == KVM_CPUID_FEATURES
>> +#else
>> +               env->cpuid_hv_level == 0
>> +#endif
>> +	    ) {
>>           pstrcpy(value, sizeof(value), "kvm");
>>       }
>>       return value;
>> @@ -1281,6 +1286,11 @@ static void x86_cpuid_set_hv_vendor(Object *obj, const char *value,
>>           }
>>           pstrcpy(adj_value, sizeof(adj_value), CPUID_HV_VENDOR_XEN);
>>       } else if (!strcmp(value, "kvm")) {
>> +#if defined(CONFIG_KVM)
>> +        if (env->cpuid_hv_level == 0) {
>> +            env->cpuid_hv_level = KVM_CPUID_FEATURES;
>> +        }
>> +#endif
> If CPUID[0x40000000].EAX set to 0 is documented as equivalent to having
> it set to 0x40000001 (KVM_CPUID_FEATURES), why the confusing checks for
> CONFIG_KVM? Why not always set it to KVM_CPUID_FEATURES?
>
At line 36 of the file:

#if defined(CONFIG_KVM)
#include <linux/kvm_para.h>
#endif

So without the check you get compile failures (i386-linux-user).

I have no issue with removing the check(s) for 0.  Currently 
hypervisor-level=0 will not force the old values; I can make a change so 
that works.
      -Don Slutz

Re: [Qemu-devel] [PATCH 2/2] target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

[Eduardo Habkost]

On Tue, Sep 18, 2012 at 03:32:04PM -0400, Don Slutz wrote:
> On 09/18/12 13:00, Eduardo Habkost wrote:
> >On Tue, Sep 18, 2012 at 10:49:53AM -0400, Don Slutz wrote:
> >> From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
> >>EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
> >>
> >>If kvm is not configured, the additional option of hypervisor-level=1
> >>(or hypervisor-level=0x40000001) needs to be specified to get this.
> >>---
> >>  target-i386/cpu.c |   12 +++++++++++-
> >>  1 files changed, 11 insertions(+), 1 deletions(-)
> >>
> >>diff --git a/target-i386/cpu.c b/target-i386/cpu.c
> >>index 6e43eff..d73b0a8 100644
> >>--- a/target-i386/cpu.c
> >>+++ b/target-i386/cpu.c
> >>@@ -1248,7 +1248,12 @@ static char *x86_cpuid_get_hv_vendor(Object *obj, Error **errp)
> >>                 env->cpuid_hv_level == CPUID_HV_LEVEL_XEN) {
> >>          pstrcpy(value, sizeof(value), "xen");
> >>      } else if (!strcmp(value, CPUID_HV_VENDOR_KVM) &&
> >>-               env->cpuid_hv_level == 0) {
> >>+#if defined(CONFIG_KVM)
> >>+               env->cpuid_hv_level == KVM_CPUID_FEATURES
> >>+#else
> >>+               env->cpuid_hv_level == 0
> >>+#endif
> >>+	    ) {
> >>          pstrcpy(value, sizeof(value), "kvm");
> >>      }
> >>      return value;
> >>@@ -1281,6 +1286,11 @@ static void x86_cpuid_set_hv_vendor(Object *obj, const char *value,
> >>          }
> >>          pstrcpy(adj_value, sizeof(adj_value), CPUID_HV_VENDOR_XEN);
> >>      } else if (!strcmp(value, "kvm")) {
> >>+#if defined(CONFIG_KVM)
> >>+        if (env->cpuid_hv_level == 0) {
> >>+            env->cpuid_hv_level = KVM_CPUID_FEATURES;
> >>+        }
> >>+#endif
> >If CPUID[0x40000000].EAX set to 0 is documented as equivalent to having
> >it set to 0x40000001 (KVM_CPUID_FEATURES), why the confusing checks for
> >CONFIG_KVM? Why not always set it to KVM_CPUID_FEATURES?
> >
> At line 36 of the file:
> 
> #if defined(CONFIG_KVM)
> #include <linux/kvm_para.h>
> #endif
> 
> So without the check you get compile failures (i386-linux-user).
> 

Right. We can't include linux/kvm_para.h on a non-Linux build host, but
on the other hand I don't see any reason to not have
cpuid_hv_level=0x40000001 on builds without CONFIG_KVM too.

If we want to allow fake KVM CPUID leaves on non-CONFIG_KVM builds like
the code above does, QEMU needs its own constant for
KVM_CPUID_FEATURES/0x40000001, so it doesn't depend on the Linux header.


> I have no issue with removing the check(s) for 0.  Currently
> hypervisor-level=0 will not force the old values; I can make a change
> so that works.
>      -Don Slutz

-- 
Eduardo

Re: [Qemu-devel] [PATCH 2/2] target-i386: Fix default Hypervisor level for hypervisor-vendor=kvm.

[Don Slutz]

On 09/19/12 09:20, Eduardo Habkost wrote:
> On Tue, Sep 18, 2012 at 03:32:04PM -0400, Don Slutz wrote:
>> On 09/18/12 13:00, Eduardo Habkost wrote:
>>> On Tue, Sep 18, 2012 at 10:49:53AM -0400, Don Slutz wrote:
>>>>  From http://lkml.indiana.edu/hypermail/linux/kernel/1205.0/00100.html
>>>> EAX should be KVM_CPUID_FEATURES (0x40000001) not 0.
>>>>
>>>> If kvm is not configured, the additional option of hypervisor-level=1
>>>> (or hypervisor-level=0x40000001) needs to be specified to get this.
>>>> ---
>>>>   target-i386/cpu.c |   12 +++++++++++-
>>>>   1 files changed, 11 insertions(+), 1 deletions(-)
>>>>
>>>> diff --git a/target-i386/cpu.c b/target-i386/cpu.c
>>>> index 6e43eff..d73b0a8 100644
>>>> --- a/target-i386/cpu.c
>>>> +++ b/target-i386/cpu.c
>>>> @@ -1248,7 +1248,12 @@ static char *x86_cpuid_get_hv_vendor(Object *obj, Error **errp)
>>>>                  env->cpuid_hv_level == CPUID_HV_LEVEL_XEN) {
>>>>           pstrcpy(value, sizeof(value), "xen");
>>>>       } else if (!strcmp(value, CPUID_HV_VENDOR_KVM) &&
>>>> -               env->cpuid_hv_level == 0) {
>>>> +#if defined(CONFIG_KVM)
>>>> +               env->cpuid_hv_level == KVM_CPUID_FEATURES
>>>> +#else
>>>> +               env->cpuid_hv_level == 0
>>>> +#endif
>>>> +	    ) {
>>>>           pstrcpy(value, sizeof(value), "kvm");
>>>>       }
>>>>       return value;
>>>> @@ -1281,6 +1286,11 @@ static void x86_cpuid_set_hv_vendor(Object *obj, const char *value,
>>>>           }
>>>>           pstrcpy(adj_value, sizeof(adj_value), CPUID_HV_VENDOR_XEN);
>>>>       } else if (!strcmp(value, "kvm")) {
>>>> +#if defined(CONFIG_KVM)
>>>> +        if (env->cpuid_hv_level == 0) {
>>>> +            env->cpuid_hv_level = KVM_CPUID_FEATURES;
>>>> +        }
>>>> +#endif
>>> If CPUID[0x40000000].EAX set to 0 is documented as equivalent to having
>>> it set to 0x40000001 (KVM_CPUID_FEATURES), why the confusing checks for
>>> CONFIG_KVM? Why not always set it to KVM_CPUID_FEATURES?
>>>
>> At line 36 of the file:
>>
>> #if defined(CONFIG_KVM)
>> #include <linux/kvm_para.h>
>> #endif
>>
>> So without the check you get compile failures (i386-linux-user).
>>
> Right. We can't include linux/kvm_para.h on a non-Linux build host, but
> on the other hand I don't see any reason to not have
> cpuid_hv_level=0x40000001 on builds without CONFIG_KVM too.
>
> If we want to allow fake KVM CPUID leaves on non-CONFIG_KVM builds like
> the code above does, QEMU needs its own constant for
> KVM_CPUID_FEATURES/0x40000001, so it doesn't depend on the Linux header.
>
Will do a V2 with a local define.
>> I have no issue with removing the check(s) for 0.  Currently
>> hypervisor-level=0 will not force the old values; I can make a change
>> so that works.
>>       -Don Slutz
   -Don Slutz