On 1/24/2019 3:15 AM, Eduardo Habkost wrote:
On Mon, Jan 21, 2019 at 05:29:32PM +0800, Tao Xu wrote:
On 1/15/2019 2:35 AM, Eduardo Habkost wrote:
Sorry, we do have a problem here:
On Thu, Dec 27, 2018 at 10:43:04AM +0800, Tao Xu wrote:
[...]
#define PC_COMPAT_3_0 \
diff --git a/target/i386/cpu.c b/target/i386/cpu.c
index 09706ad51a..5296c73cd5 100644
--- a/target/i386/cpu.c
+++ b/target/i386/cpu.c
@@ -2499,7 +2499,8 @@ static X86CPUDefinition builtin_x86_defs[] = {
CPUID_7_0_ECX_PKU | CPUID_7_0_ECX_OSPKE |
CPUID_7_0_ECX_AVX512VNNI,
.features[FEAT_7_0_EDX] =
- CPUID_7_0_EDX_SPEC_CTRL | CPUID_7_0_EDX_SPEC_CTRL_SSBD,
+ CPUID_7_0_EDX_SPEC_CTRL | CPUID_7_0_EDX_SPEC_CTRL_SSBD |
+ CPUID_7_0_EDX_ARCH_CAPABILITIES,
CPUID_7_0_EDX_ARCH_CAPABILITIES is still set on
unmigratable_flags. We need to make it migratable before adding
it by default to a named CPU model.
Hi Eduardo,
Do you mean I need to remove CPUID_7_0_EDX_ARCH_CAPABILITIES
from .migratable_flags? Or CPUID_7_0_EDX_ARCH_CAPABILITIES can not
support migration now?
We need to remove it from .unmigratable_flags, but only after
confirm it is really migration-safe. To make it migration-safe,
the MSR value seen by the guest must have a predictable value
that is 100% independent from host hardware or host software
version.
This is easy to ensure if MSR_IA32_ARCH_CAPABILITIES is in
KVM_GET_MSR_INDEX_LIST (meaning QEMU can actually configure the
MSR value seen by the guest). If MSR_IA32_ARCH_CAPABILITIES is
not on KVM_GET_MSR_INDEX_LIST, arch-capabilities must not be
returned by x86_cpu_get_supported_feature_word() when
migratable_only is true.
Thank you. I have seen your patch to migratable it:
[PATCH 0/2] i386: arch_capabilities fixes + migratability
http://lists.nongnu.org/archive/html/qemu-devel/2019-01/msg06701.html
So now can arch-capabilities and features exposed by it be added in
Cascadelake CPU model? Because Cascadelake CPU model can support it in
hardware. And for Icelake CPU model, Robert will add in the future.
Also, why are you setting this only on Cascadelake-Server and not
on all the other Intel CPUs?
Thank you for your notice. I reviewed the git log of KVM.
"MSR_IA32_ARCH_CAPABILITIES is emulated in kvm, there is no dependency on
hardware support for this feature". So do you mean we should also add
features based on ARCH_CAPABILITIES in former CPU(such as skylake)?
It doesn't depend on host hardware features, but it does depend
on host KVM code added in Linux v4.17, which makes this more
tricky.
We don't want a QEMU upgrade to make an existing VM configuration
to stop running under Linux v4.16. This will probably require
versioned CPU models[1] to work well.
[1] https://lists.gnu.org/archive/html/qemu-devel/2018-06/msg08422.html
Hi Eduardo, I think the versioned CPU models solution is great. But I
haven't seen further patches in the mail list. What is the recent
progress of this work? Need me to do something?
Looking forward to your reply.
Tao
