On Mon, Mar 23, 2020 at 02:20:01AM -0700, Kang, Luwei wrote: > > On Wed, Mar 18, 2020 at 11:48:18AM +0800, Luwei Kang wrote: > > > If the logical processor is operating with Intel PT enabled ( > > > IA32_RTIT_CTL.TraceEn = 1) at the time of VM entry, the “load > > > IA32_RTIT_CTL” VM-entry control must be 0(SDM 26.2.1.1). > > > > > > The first disabled the host Intel PT(Clear TraceEn) will make all the > > > buffered packets are flushed out of the processor and it may cause an > > > Intel PT PMI. The host Intel PT will be re-enabled in the host Intel > > > PT PMI handler. > > > > > > handle_pmi_common() > > > -> intel_pt_interrupt() > > > -> pt_config_start() > > > > IIUC, this is only possible when PT "plays nice" with VMX, correct? > > Otherwise pt->vmx_on will be true and pt_config_start() would skip the > > WRMSR. > > > > And IPT PMI must be delivered via NMI (though maybe they're always > > delivered via NMI?). > > > > In any case, redoing WRMSR doesn't seem safe, and it certainly isn't > > performant, e.g. what prevents the second WRMSR from triggering a second > > IPT PMI? > > > > pt_guest_enter() is called after the switch to the vCPU has already been > > recorded, can't this be handled in the IPT code, e.g. something like this? > > > > diff --git a/arch/x86/events/intel/pt.c b/arch/x86/events/intel/pt.c index > > 1db7a51d9792..e38ddae9f0d1 100644 > > --- a/arch/x86/events/intel/pt.c > > +++ b/arch/x86/events/intel/pt.c > > @@ -405,7 +405,7 @@ static void pt_config_start(struct perf_event *event) > > ctl |= RTIT_CTL_TRACEEN; > > if (READ_ONCE(pt->vmx_on)) > > perf_aux_output_flag(&pt->handle, PERF_AUX_FLAG_PARTIAL); > > - else > > + else (!(current->flags & PF_VCPU)) > > wrmsrl(MSR_IA32_RTIT_CTL, ctl); > > Intel PT can work in SYSTEM and HOST_GUEST mode by setting the kvm-intel.ko > parameter "pt_mode". In SYSTEM mode, the host and guest PT trace will be > saved in the host buffer. The KVM do nothing during VM-entry/exit in SYSTEM > mode and Intel PT PMI may happened on any place. The PT trace may be disabled > when running in KVM(PT only needs to be disabled before VM-entry in > HOST_GUEST mode). Ah, right. What about enhancing intel_pt_handle_vmx() and 'struct pt' to replace vmx_on with a field that incorporates the KVM mode? From an outsider's perspective, that'd be an improvment irrespective of this bug fix as 'vmx_on' is misleading, e.g. it can be %false when the CPU is post- VMXON, and really means "post-VMXON and Intel PT can't trace it".
