On 11/08/2016 04:13, Xulei (Stone) wrote:
Following your suggestion, I found this problem may be caused by the flag of HF_SMM_MASK. I'm now sure QEMU is sending the KVM_SMI ioctl, and kmod already handles this ioctl.
I add printk in inject_pending_event(), like this:
/* try to inject new event if pending */
- if(get_smi)
- printk(KERN_INFO "is_smm:%d\n", is_smm(vcpu));
- if (vcpu->arch.smi_pending && !is_smm(vcpu)) { vcpu->arch.smi_pending = false; process_smi(vcpu); }
Then I found that the normal output is (is_smm is 0): 2016-08-11T09:44:45.090078+08:00|info|kernel[-]|[1269634.151054get smi ioctl from qemu 2016-08-11T09:44:45.090097+08:00|info|kernel[-]|[1269634.151056in process_smi_request 2016-08-11T09:44:45.090114+08:00|info|kernel[-]|[1269634.151057is_smm: 0 2016-08-11T09:44:45.090131+08:00|info|kernel[-]|[1269634.151058process smi 2016-08-11T09:44:45.090914+08:00|info|kernel[-]|[1269634.151575get smi ioctl from qemu 2016-08-11T09:44:45.090947+08:00|info|kernel[-]|[1269634.151578in process_smi_request 2016-08-11T09:44:45.090972+08:00|info|kernel[-]|[1269634.151579is_smm: 0
When problem occurs the output is (is_smm is 1): 2016-08-11T10:07:11.755982+08:00|info|kernel[-]|[1270981.916056get smi ioctl from qemu 2016-08-11T10:07:11.756035+08:00|info|kernel[-]|[1270981.916062in process_smi_request 2016-08-11T10:07:11.756078+08:00|info|kernel[-]|[1270981.916064is_smm: 1 2016-08-11T10:07:13.896977+08:00|info|kernel[-]|[1270984.058117get smi ioctl from qemu 2016-08-11T10:07:13.897063+08:00|info|kernel[-]|[1270984.058121in process_smi_request 2016-08-11T10:07:13.897091+08:00|info|kernel[-]|[1270984.058123is_smm: 1
So, it seems kvm cannot set HF_SMM_MASK flag correctly when we
continuously reset VM.
Any futher suggestion? Paolo :)
I guess a KVM_SET_VCPU_EVENTS is missing at reset time.
This maybe a kvm bug of SMM supporting feature: assuming kvm SMI injection process is interrupted at the point between process_smi() and bios executing RSM. Then bios runs from the beginning again because of resetting and kvm keeps is_smm() =1. When bios writes 0xb2 to raise a SMI, kvm thinks is_smm() = 1, and will not process this request, and also will not call kvm_vcpu_reset() to clear HF_SMM_MASK flag any more. Bios will not executing RSM because of it does not get this SMI injecting from kvm. This condition cannot break until we destroy this VM.
After testing 2 days, following patch can solve this problem, but obviously it is not a suitable solution: static int kvm_vcpu_ioctl_smi(struct kvm_vcpu *vcpu) { + if (is_smm(vcpu)) { + vcpu->arch.hflags &= ~HF_SMM_MASK; + vcpu->arch.smi_pending = 0; + } kvm_make_request(KVM_REQ_SMI, vcpu);
return 0; }
Paolo, could you post a suitable patch to solve this problem based on the information I mentioned above?
Paolo
-
Xulei (Stone)