Re: [Qemu-devel] [PATCH v5 0/4] Add QEMU support for Intel local MCE

[Laszlo Ersek]

On 07/06/16 10:35, Haozhong Zhang wrote:
> On 07/06/16 10:26, Laszlo Ersek wrote:
>> On 06/22/16 08:56, Haozhong Zhang wrote:
>>> Changes in v5:
>>>  * v4 Patch 1&2 are now merged into v5 Patch 1. (Paolo Bonzini)
>>>  * Add a separate Patch 3 to automatically enable LMCE for "-cpu host"
>>>    if host does support. (Eduardo Habkost)
>>>  * Move the sanity check of misconfiguration of LMCE to a separate
>>>    Patch 4. I'm fine to drop it if it's considered to be
>>>    unnecessary. (Eduardo Habkost)
>>>  * Move check of MCG_LMCE_P capability to the existing code in
>>>    kvm_arch_init_vcpu(). The difference is that if LMCE is not
>>>    supported by host but MCG_LMCE_P is included in env->mcg_cap,
>>>    kvm_arch_init_vcpu() return with error instead of just reporting a
>>>    warning. (Eduardo Habkost)
>>>  * Fix setting has_msr_feature_control in kvm_arch_init_vcpu(). (Paolo
>>>    Bonzini)
>>>  * Fix the type error in fw_cfg file "etc/msr_feature_control".
>>>
>>> Changes in v4:
>>>  * Abort starting QEMU if lmce option is present but host does support
>>>    LMCE. (Eduardo Habkost)
>>>  * Remove setting MSR_IA32_FEATURE_CONTROL, which should be left to
>>>    guest. (Radim Krčmá, Paolo Bonzini
>>>  * Adjust error messages in mce_init(). (Boris Petkov)
>>>  * Move adding option 'lmce' to patch 1. (Eduardo Habkost, Paolo Bonzini)
>>>  * Adjust LMCE error message in cpu_post_load(). (Eduardo Habkost)
>>>  * (Patch 3) Add a fw_cfg file 'etc/msr_feature_control' to advise
>>>    bits should be set in MSR_IA32_FEATURE_CONTROL. (Paolo Bonzini)
>>>  * Fix SOB chain in patch 1.
>>>
>>> Changes in v3:
>>>  * LMCE can be enabled only for non-intel guests.
>>>  * LMCE is disabled by default and a cpu option 'lmce=on/off' is added
>>>    to explicitly enable/disable LMCE.
>>>  * LMCE is disabled if KVM does not support (even though 'lmce=on').
>>>  * VM on LMCE-enabled QEMU can be only migrated to LMCE-enabled QEMU.
>>>  * MCG_LMCE_P is not included in MCE_CAP_DEF and instead added to
>>>    env->mcg_cap if LMCE is enabled.
>>>  * Code style fix.
>>>
>>> This QEMU patch series along with the corresponding KVM patch series
>>> (sent via another email with title "[PATCH v3 0/3] Add KVM support for
>>> Intel local MCE") enables Intel local MCE feature for guest.
>>>
>>> Intel Local MCE (LMCE) is a feature on Intel Skylake Server CPU that
>>> can deliver MCE to a single processor thread instead of broadcasting
>>> to all threads, which can reduce software's load when processing MCE
>>> on machines with a large number of processor threads.
>>>
>>> The technical details of LMCE can be found in Intel SDM Vol 3, Chapter
>>> "Machine-Check Architecture" (search for 'LMCE'). Basically,
>>>  * The capability of LMCE is indicated by bit 27 (MCG_LMCE_P) of
>>>    MSR_IA32_MCG_CAP.
>>>  * LMCE is enabled by setting bit 20 (MSR_IA32_FEATURE_CONTROL_LMCE)
>>>    of MSR_IA32_FEATURE_CONTROL and bit 0 (MCG_EXT_CTL_LMCE_EN) of
>>>    MSR_IA32_MCG_EXT_CTL.
>>>  * Software can determine if a MCE is local to the current processor
>>>    thread by checking bit 2 (MCG_STATUS_LMCE) of MSR_IA32_MCG_STATUS.
>>>
>>> Ashok Raj (1):
>>>   target-i386: KVM: add basic Intel LMCE support
>>>
>>> Haozhong Zhang (3):
>>>   i386: publish advised value of MSR_IA32_FEATURE_CONTROL via fw_cfg
>>>   target-i386: enable LMCE for '-cpu host' if supported by host
>>>   target-i386: abort migration if LMCE config mismatch
>>>
>>>  hw/i386/pc.c          | 29 +++++++++++++++++++++++++++++
>>>  target-i386/cpu.c     | 19 ++++++++++++++++++-
>>>  target-i386/cpu.h     | 16 ++++++++++++++++
>>>  target-i386/kvm.c     | 36 +++++++++++++++++++++++++++++++++---
>>>  target-i386/machine.c | 25 +++++++++++++++++++++++++
>>>  5 files changed, 121 insertions(+), 4 deletions(-)
>>>
>>
>> How can I test this series?
>>
>> I'm trying to satisfy the conditions in pc_build_feature_control_file(), so 
>> that "etc/msr_feature_control" is created. However, I cannot even trigger
>>
>>     if (ecx & CPUID_EXT_VMX) {
>>         feature_control_bits |= FEATURE_CONTROL_VMXON_ENABLED_OUTSIDE_SMX;
>>     }
>>
>> I tried various -cpu switches (with and without -enable-kvm); nothing seems 
>> to work. When I add "+vmx", QEMU always complains that vmx is not available. 
>> Any hints?
>>
>> (My laptop does support VMX, and it has 
>> FEATURE_CONTROL_VMXON_ENABLED_OUTSIDE_SMX set in the corresponding host MSR.)
>>
> 
> You may check if kvm-intel module is loaded with nested support by
> 'cat /sys/module/kvm_intel/parameters/nested'. If it's 0, you need to
> reload kvm-intel with nested parameter: modprobe kvm-intel
> nested=1.

Ah, that's what is missing. I'll look into it, thanks!

> Moreover, you may also try qemu -cpu host to get the most
> your host hardware supports.

Yep, I tried that, but it wasn't enough.

Thanks!
Laszlo