On Mon, Feb 10, 2020 at 07:52:13PM -0800, Andy Lutomirski wrote: > On Tue, Feb 4, 2020 at 1:37 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > > > On Tue, Feb 04, 2020 at 05:19:26PM +0800, Xiaoyao Li wrote: > > > > > > > + case MSR_IA32_CORE_CAPS: > > > > > + if (!msr_info->host_initiated) > > > > > > > > Shouldn't @data be checked against kvm_get_core_capabilities()? > > > > > > Maybe it's for the case that userspace might have the ability to emulate SLD > > > feature? And we usually let userspace set whatever it wants, e.g., > > > ARCH_CAPABILITIES. > > > > If the 'sq_misc.split_lock' event is sufficiently accurate, I suppose > > the host could use that to emulate the feature at the cost of one > > counter used. > > I would be impressed if the event were to fire before executing the > offending split lock. Wouldn't the best possible result be for it to > fire with RIP pointing to the *next* instruction? This seems like it > could be quite confusing to a guest. True; and I see no indication the event is PEBS capable, so even that is pushing it. However, it's virt; isn't that confused per definition? ;-))
