KVM support for AMX includes a new system attribute, KVM_X86_XCOMP_GUEST_SUPP.
Commit 19db68ca68 ("x86: Grant AMX permission for guest", 2022-03-15) however
did not fully consider the behavior on older kernels.  First, it warns
too aggressively.  Second, it invokes the KVM_GET_DEVICE_ATTR ioctl
unconditionally and then uses the "bitmask" variable, which remains
uninitialized if the ioctl fails.  Third, kvm_ioctl returns -errno rather
than -1 on errors.

While at it, explain why the ioctl is needed and KVM_GET_SUPPORTED_CPUID
is not enough.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
        v1->v2: for error condition for kvm_ioctl [Volker]
 target/i386/kvm/kvm.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/target/i386/kvm/kvm.c b/target/i386/kvm/kvm.c
index ef2c68a6f4..06901c2a43 100644
--- a/target/i386/kvm/kvm.c
+++ b/target/i386/kvm/kvm.c
@@ -411,6 +411,12 @@ uint32_t kvm_arch_get_supported_cpuid(KVMState *s, uint32_t function,
         }
     } else if (function == 0xd && index == 0 &&
                (reg == R_EAX || reg == R_EDX)) {
+        /*
+         * The value returned by KVM_GET_SUPPORTED_CPUID does not include
+         * features that still have to be enabled with the arch_prctl
+         * system call.  QEMU needs the full value, which is retrieved
+         * with KVM_GET_DEVICE_ATTR.
+         */
         struct kvm_device_attr attr = {
             .group = 0,
             .attr = KVM_X86_XCOMP_GUEST_SUPP,
@@ -419,13 +425,16 @@ uint32_t kvm_arch_get_supported_cpuid(KVMState *s, uint32_t function,
 
         bool sys_attr = kvm_check_extension(s, KVM_CAP_SYS_ATTRIBUTES);
         if (!sys_attr) {
-            warn_report("cannot get sys attribute capabilities %d", sys_attr);
+            return ret;
         }
 
         int rc = kvm_ioctl(s, KVM_GET_DEVICE_ATTR, &attr);
-        if (rc == -1 && (errno == ENXIO || errno == EINVAL)) {
-            warn_report("KVM_GET_DEVICE_ATTR(0, KVM_X86_XCOMP_GUEST_SUPP) "
-                        "error: %d", rc);
+        if (rc < 0) {
+            if (rc != -ENXIO) {
+                warn_report("KVM_GET_DEVICE_ATTR(0, KVM_X86_XCOMP_GUEST_SUPP) "
+                            "error: %d", rc);
+            }
+            return ret;
         }
         ret = (reg == R_EAX) ? bitmask : bitmask >> 32;
     } else if (function == 0x80000001 && reg == R_ECX) {
-- 
2.35.1

On Sun, Mar 20, 2022 at 20:39:14 +0100, Paolo Bonzini wrote:
> KVM support for AMX includes a new system attribute, KVM_X86_XCOMP_GUEST_SUPP.
> Commit 19db68ca68 ("x86: Grant AMX permission for guest", 2022-03-15) however
> did not fully consider the behavior on older kernels.  First, it warns
> too aggressively.  Second, it invokes the KVM_GET_DEVICE_ATTR ioctl
> unconditionally and then uses the "bitmask" variable, which remains
> uninitialized if the ioctl fails.  Third, kvm_ioctl returns -errno rather
> than -1 on errors.
> 
> While at it, explain why the ioctl is needed and KVM_GET_SUPPORTED_CPUID
> is not enough.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>         v1->v2: for error condition for kvm_ioctl [Volker]
>  target/i386/kvm/kvm.c | 17 +++++++++++++----
>  1 file changed, 13 insertions(+), 4 deletions(-)

Based on the commit message it seems that this is only for the problem
that Michal reported, right?

Because it doesn't still fix the assertion failure on my AMD box:

$  ./build/qemu-system-x86_64 -accel kvm
qemu-system-x86_64: ../target/i386/kvm/kvm-cpu.c:105: kvm_cpu_xsave_init: Assertion `esa->size == eax' failed.
Aborted (core dumped)

On 3/20/22 12:39, Paolo Bonzini wrote:
> +        if (rc < 0) {
> +            if (rc != -ENXIO) {
> +                warn_report("KVM_GET_DEVICE_ATTR(0, KVM_X86_XCOMP_GUEST_SUPP) "
> +                            "error: %d", rc);

Better as "error: %s", strerror(-rc) ?


r~