On Tue, Aug 02, 2022, Vitaly Kuznetsov wrote: > While it seems reasonable to not expose LOAD_IA32_PERF_GLOBAL_CTRL controls > to L1 hypervisor on buggy CPUs, such change would inevitably break live > migration from older KVMs where the controls are exposed. Keep the status quo > for now, L1 hypervisor itself is supposed to take care of the errata. As noted before, this statement is wrong as it requires guest FMS == host FMS, but it's irrelevant because KVM can emulate the control unconditionally. I'll test and fold in my suggested patch[*] (assuming it works) and reword this part of the changelog. Ah, and I'll also need to fold in a patch to actually emulate the controls without hardware support. [*] https://lore.kernel.org/all/YtnZmCutdd5tpUmz@xxxxxxxxxx > Reviewed-by: Maxim Levitsky <mlevitsk@xxxxxxxxxx> > Signed-off-by: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> > --- > arch/x86/kvm/vmx/vmx.c | 59 +++++++++++++++++++++++++----------------- > 1 file changed, 35 insertions(+), 24 deletions(-) > ... > @@ -8192,6 +8199,10 @@ static __init int hardware_setup(void) > if (setup_vmcs_config(&vmcs_config, &vmx_capability) < 0) > return -EIO; > > + if (cpu_has_perf_global_ctrl_bug()) > + pr_warn_once("kvm: VM_EXIT_LOAD_IA32_PERF_GLOBAL_CTRL " > + "does not work properly. Using workaround\n"); Any objections to opportunistically tweaking this? pr_warn_once("kvm: CPU has VM_EXIT_LOAD_IA32_PERF_GLOBAL_CTRL erratum," "using MSR load/store lists for PERF_GLOBAL_CTRL\n"); > + > if (boot_cpu_has(X86_FEATURE_NX)) > kvm_enable_efer_bits(EFER_NX); > > -- > 2.35.3 >
