i386: Mask SVM features if nested SVM is disabled

QEMU incorrectly validates FEAT_SVM feature flags against
GET_SUPPORTED_CPUID even if SVM features are being masked out by
cpu_x86_cpuid().  This can make QEMU print warnings on most AMD
CPU models, even when SVM nesting is disabled (which is the
default).

This bug was never detected before because of a Linux KVM bug:
until Linux v5.6, KVM was not filtering out SVM features in
GET_SUPPORTED_CPUID when nested was disabled.  This KVM bug was
fixed in Linux v5.7-rc1, on Linux commit a50718cc3f43 ("KVM:
nSVM: Expose SVM features to L1 iff nested is enabled").

Fix the problem by adding a CPUID_EXT3_SVM dependency to all
FEAT_SVM feature flags in the feature_dependencies table.

Reported-by: Yanan Fu <yfu@redhat.com>
Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
Message-Id: <20200623230116.277409-1-ehabkost@redhat.com>
[Fix testcase. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

target/i386/cpu.c

@@ -1404,6 +1404,10 @@ static FeatureDep feature_dependencies[] = {
         .from = { FEAT_VMX_SECONDARY_CTLS,  VMX_SECONDARY_EXEC_ENABLE_VMFUNC },
         .to = { FEAT_VMX_VMFUNC,            ~0ull },
     },
+    {
+        .from = { FEAT_8000_0001_ECX,       CPUID_EXT3_SVM },
+        .to = { FEAT_SVM,                   ~0ull },
+    },
 };
 
 typedef struct X86RegisterInfo32 {