Re: [PATCH v2] accel/kvm: Specify default IPA size for arm64

[Philippe Mathieu-Daudé]

Hi Akihiko,

On 21/7/23 08:24, Akihiko Odaki wrote:
> libvirt uses "none" machine type to test KVM availability. Before this
> change, QEMU used to pass 0 as machine type when calling KVM_CREATE_VM.
>
> The kernel documentation says:
> > On arm64, the physical address size for a VM (IPA Size limit) is
> > limited to 40bits by default. The limit can be configured if the host
> > supports the extension KVM_CAP_ARM_VM_IPA_SIZE. When supported, use
> > KVM_VM_TYPE_ARM_IPA_SIZE(IPA_Bits) to set the size in the machine type
> > identifier, where IPA_Bits is the maximum width of any physical
> > address used by the VM. The IPA_Bits is encoded in bits[7-0] of the
> > machine type identifier.
> >
> > e.g, to configure a guest to use 48bit physical address size::
> >
> >      vm_fd = ioctl(dev_fd, KVM_CREATE_VM, KVM_VM_TYPE_ARM_IPA_SIZE(48));
> >
> > The requested size (IPA_Bits) must be:
> >
> >   ==   =========================================================
> >    0   Implies default size, 40bits (for backward compatibility)
> >    N   Implies N bits, where N is a positive integer such that,
> >        32 <= N <= Host_IPA_Limit
> >   ==   =========================================================
>
> > Host_IPA_Limit is the maximum possible value for IPA_Bits on the host
> > and is dependent on the CPU capability and the kernel configuration.
> > The limit can be retrieved using KVM_CAP_ARM_VM_IPA_SIZE of the
> > KVM_CHECK_EXTENSION ioctl() at run-time.
> >
> > Creation of the VM will fail if the requested IPA size (whether it is
> > implicit or explicit) is unsupported on the host.
> https://docs.kernel.org/virt/kvm/api.html#kvm-create-vm
>
> So if Host_IPA_Limit < 40, such KVM_CREATE_VM will fail, and libvirt
> incorrectly thinks KVM is not available. This actually happened on M2
> MacBook Air.
>
> Fix this by specifying 32 for IPA_Bits as any arm64 system should
> support the value according to the documentation.
>
> Signed-off-by: Akihiko Odaki <akihiko.odaki@daynix.com>
> ---
> V1 -> V2: Introduced an arch hook
>
>   include/sysemu/kvm.h   | 1 +
>   accel/kvm/kvm-all.c    | 2 +-
>   target/arm/kvm.c       | 2 ++
>   target/i386/kvm/kvm.c  | 2 ++
>   target/mips/kvm.c      | 2 ++
>   target/ppc/kvm.c       | 2 ++
>   target/riscv/kvm.c     | 2 ++
>   target/s390x/kvm/kvm.c | 2 ++
>   8 files changed, 14 insertions(+), 1 deletion(-)

My understanding of Peter's suggestion would be smth like:

-- >8 --
diff --git a/include/sysemu/kvm.h b/include/sysemu/kvm.h
index 115f0cca79..c0af15eb6c 100644
--- a/include/sysemu/kvm.h
+++ b/include/sysemu/kvm.h
@@ -201,10 +201,15 @@ typedef struct KVMCapabilityInfo {

 struct KVMState;

+struct KVMClass {
+    AccelClass parent_class;
+
+    int default_vm_type;
+};
+
 #define TYPE_KVM_ACCEL ACCEL_CLASS_NAME("kvm")
 typedef struct KVMState KVMState;
-DECLARE_INSTANCE_CHECKER(KVMState, KVM_STATE,
-                         TYPE_KVM_ACCEL)
+OBJECT_DECLARE_TYPE(KVMState, KVMClass, KVM_ACCEL)

 extern KVMState *kvm_state;
 typedef struct Notifier Notifier;
diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
index 373d876c05..fdd424e1a5 100644
--- a/accel/kvm/kvm-all.c
+++ b/accel/kvm/kvm-all.c
@@ -2458,12 +2458,13 @@ static int kvm_init(MachineState *ms)
     KVMState *s;
     const KVMCapabilityInfo *missing_cap;
     int ret;
-    int type = 0;
+    int type;
     uint64_t dirty_log_manual_caps;

     qemu_mutex_init(&kml_slots_lock);

     s = KVM_STATE(ms->accelerator);
+    type = KVM_GET_CLASS(s)->default_vm_type;

     /*
      * On systems where the kernel can support different base page
diff --git a/target/arm/kvm.c b/target/arm/kvm.c
index b4c7654f49..5c13594fdf 100644
--- a/target/arm/kvm.c
+++ b/target/arm/kvm.c
@@ -1064,4 +1064,8 @@ bool kvm_arch_cpu_check_are_resettable(void)

 void kvm_arch_accel_class_init(ObjectClass *oc)
 {
+    KVMClass *kc = KVM_CLASS(oc);
+
+    /* Host_IPA_Limit ... */
+    kc->default_vm_type = 32;
 }
---