On Mon, 2009-04-13 at 21:02 +0800, Avi Kivity wrote: > Huang Ying wrote: > > On Sat, 2009-04-11 at 20:04 +0800, Avi Kivity wrote: > > > >>>> Why not always allocate it on vcpu setup? > >>>> > >>>> > >>> Because the MCE bank number is not fixed, it is based on mcg_cap from > >>> user space. > >>> > >>> > >> Right, but we can allocate the maximum number, no? it's a fairly small > >> amount of memory. > >> > > > > OK. If you are OK with 8k extra memory usage. > > > > Let's make it 64 banks like Andi suggests. Return the number in the > KVM_CHECK_EXTENSION call (we do the same for memory slots) so userspace > can adjust. OK. I will change this. > >> > >> kvm does feature negotiation (really, feature advertising) using > >> KVM_CAP_... and KVM_CHECK_EXTENSION. So there's no need for this. > >> > > > > It is not only feature advertising, but also something like CPUID. Based > > on CPU type specified in command line (qemu -cpu ?), different mcg_cap > > can be used to determine number of MCE banks and various features of MCE > > (MCG_CTL_P, MCG_TES_P, MCG_CMCI, etc) for specified CPU type. With KVM > > feature advertising, we can only check whether MCE simulation is > > supported, but can not set parameters of MCE simulation. > > > > Maybe this interface can be changed to KVM_SET_MSRS, with > > MSR_IA32_MCG_CAP be set by kvm-qemu. But MSR_IA32_MCG_CAP is read-only > > by spec, and current interface can not distinguishes between MSR set > > from guest system and MSR setup from kvm-qemu. I think there should be > > other read-only MSRs need to be setup during guest system > > creation/setup. How about add a interface such as KVM_SETUP_MSRS to > > setup the value of read-only MSRs? > > > > Then we would need to tell which read-only MSRs are setup writeable and > which aren't... Maybe something like only if KVM_CAP_MCE is advertised, the MCE related read-only RO-MSRs can be setup. Other RO-MSRs setup code can be added following the similar rules. And interface is something like follow? long kvm_arch_vcpu_ioctl(...) { ... case KVM_SETUP_RO_MSRS: r = msr_io(vcpu, argp, kvm_setup_ro_msr, 0); ... } int kvm_setup_ro_msr(...) { /* return 1 for setup un-writable MSR, otherwise 0 */ } > I'm okay with an ioctl to setup MCE, but just make sure userspace has > all the information to know what the kernel can do rather than the > try-and-see-if-it-works approach. We can publish this information via > KVM_CAP things, or via another ioctl (see KVM_GET_SUPPORTED_CPUID2 for > an example). Yes. MCE support should be published by KVM_CAP_MCE and other features can be published via reading the default value of MSR_IA32_MCG_CAP. Best Regards, Huang Ying
Attachment:
signature.asc
Description: This is a digitally signed message part