In the current implementation there are subtle bugs that have to do with
the termination of subleaves, resulting in features reported on the
guest that it should not have according to the model.

If a guest model has fewer valid subleaves stopping too early leaves
unspecified gaps that are filled by the host, resulting in potentially
invalid feature combinations, for example:

The host reports features:

CPUID[EAX=01H].ECX.AVX [bit 28] and
CPUID[EAX=07H,ECX=01H].EAX.AVX_VNNI [bit 4]

In the current implementation we would skip over 7.1 completely, not
registering CPUID responses for the subleaf, resulting in a guest with
qemu64 model (which does not feature AVX), reporting AVX_VNNI as
available (passthrough from host).

We need to walk through the max subleaves of leaf 7 and register them
explicitly (also if they're 0).

A second bug was about the propagation of "terminator" subleaves,
all-zero responses that signal the end. We also want to register this
terminator response , so the guest stops enumerating. In the current
implementation we weren't doing that.

Fixes: 4fa04dd1621
Signed-off-by: Magnus Kulke <[email protected]>
---
 target/i386/mshv/mshv-cpu.c | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/target/i386/mshv/mshv-cpu.c b/target/i386/mshv/mshv-cpu.c
index 126ca40b48..a3b5f4474f 100644
--- a/target/i386/mshv/mshv-cpu.c
+++ b/target/i386/mshv/mshv-cpu.c
@@ -997,7 +997,7 @@ static void collect_cpuid_entries(const CPUState *cpu, 
GList **cpuid_entries)
     CPUX86State *env = &x86_cpu->env;
     uint32_t eax, ebx, ecx, edx;
     uint32_t leaf, subleaf;
-    uint32_t max_basic_leaf, max_extended_leaf;
+    uint32_t max_basic_leaf, max_extended_leaf, max_subleaf_7;
     uint32_t max_subleaf = 0x20;
     uint32_t leaves_with_subleaves[] = {0x04, 0x07, 0x0d, 0x0f, 0x10};
     int n_subleaf_leaves = ARRAY_SIZE(leaves_with_subleaves);
@@ -1035,14 +1035,31 @@ static void collect_cpuid_entries(const CPUState *cpu, 
GList **cpuid_entries)
             continue;
         }
 
+        /*
+         * Valid subleaves for are reported in 7.0:EAX. We need to register all
+         * subleaves to the maximum subleaf, even if they are 0. Otherwise the
+         * host will supply its own values for a unregistered subleaf, which
+         * can result in an inconsistent feature set.
+         */
+        if (leaf == 0x07) {
+            cpu_x86_cpuid(env, leaf, 0, &max_subleaf_7, &ebx, &ecx, &edx);
+            for (subleaf = 0; subleaf <= max_subleaf_7; subleaf++) {
+                cpu_x86_cpuid(env, leaf, subleaf, &eax, &ebx, &ecx, &edx);
+                add_cpuid_entry(cpuid_entries, leaf, subleaf,
+                                eax, ebx, ecx, edx);
+            }
+            continue;
+        }
+
         subleaf = 0;
         while (subleaf < max_subleaf) {
             cpu_x86_cpuid(env, leaf, subleaf, &eax, &ebx, &ecx, &edx);
 
+            /* register the "terminator" to the guest, before breaking */
+            add_cpuid_entry(cpuid_entries, leaf, subleaf, eax, ebx, ecx, edx);
             if (eax == 0 && ebx == 0 && ecx == 0 && edx == 0) {
                 break;
             }
-            add_cpuid_entry(cpuid_entries, leaf, subleaf, eax, ebx, ecx, edx);
             subleaf++;
         }
     }
-- 
2.34.1


Reply via email to