On 14.09.2017 12:48, Greg Kurz wrote:
> If the host has both KVM PR and KVM HV loaded and we pass:
>
> -machine pseries,accel=kvm,kvm-type=PR
>
> the kvmppc_is_pr() returns false instead of true. Since the helper
> is mostly used as fallback, it doesn't have any real impact with
> recent kernels. A notable exception is the workaround to allow
> migration between compatible hosts with different PVRs (eg, POWER8
> and POWER8E), since KVM still doesn't provide a way to check if a
> specific PVR is supported (see commit c363a37a450f for details).
>
> According to the official KVM API documentation [1], KVM_PPC_GET_PVINFO
> is "vm ioctl", but we check it as a global ioctl. The following function
> in KVM is hence called with kvm == NULL and considers we're in HV mode.
>
> int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
> {
> int r;
> /* Assume we're using HV mode when the HV module is loaded */
> int hv_enabled = kvmppc_hv_ops ? 1 : 0;
>
> if (kvm) {
> /*
> * Hooray - we know which VM type we're running on. Depend on
> * that rather than the guess above.
> */
> hv_enabled = is_kvmppc_hv_enabled(kvm);
> }
>
> Let's use kvm_vm_check_extension() to fix the issue.
By the way, what about the other CAPs that rely on hv_enabled? grepping
through the QEMU sources, I can see:
cap_ppc_smt = kvm_check_extension(s, KVM_CAP_PPC_SMT);
cap_htab_fd = kvm_check_extension(s, KVM_CAP_PPC_HTAB_FD);
int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
!kvm_check_extension(cs->kvm_state, KVM_CAP_SW_TLB)) {
... do we need to fix them, too?
Thomas
