Thanks a lot Marcelo! On Thu, Jan 9, 2014 at 1:46 PM, Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote: > On Wed, Jan 08, 2014 at 06:35:00PM -0500, Hu Yaohui wrote: >> Thanks a lot Marcelo! >> >> On Wed, Jan 8, 2014 at 6:25 PM, Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote: >> > On Wed, Jan 08, 2014 at 06:14:15PM -0500, Hu Yaohui wrote: >> >> Hi guys, >> >> I think you should be pretty familiar with lapic. I would really >> >> appreciate it if someone could shed some lights on my problem >> >> regarding Guest TLB flush IPI. >> >> Supposed we get two vcpus 0 and 1. >> >> When vcpu#0 wants to invalidate the tlb entry on vcpu#1. An IPI will >> >> be generated by lapic on vcpu#0 by writing to ICR which will cause a >> >> vmexit. >> >> apic_send_ipi->kvm_irq_delivery_to_apic->kvm_apic_set_irq->__apic_accept_irq >> >> In __apic_accept_irq, it will call kvm_make_request, kvm_vcpu_kick. >> >> If vcpu#1 in guest mode, how can it receives this IPI immediately, or >> >> the stale tlb entry could be accessed. Thanks for your time! >> > >> I am using kvm-kmod-3.2 >> > Two possibilities: >> > >> > 2) Hardware does not support APIC virtualization: kvm_vcpu_kick sends an >> > host-IPI to the remote vcpu, and if that vcpu is in guest mode, a VM-exit >> > (exit reason: external interrupt) will be triggered due to the host-IPI. >> > Then on VM-entry (inject_pending_event) the guest-IPI is injected. >> > >> if vcpu#1 is not on the same pcpu as the vcpu#0, a host-IPI will be sent. > > Yes. > >> But if they are on the same pcpu, if vcpu#1 is in guest mode. > > If vcpu#0 and vcpu#1 are on the same pcpu, then either one of them > is guest mode at one given moment, but not both. > >> Then the >> guest tlb flush IPI >> will wait until the next vcpu#1 vmexit. If that's the case. they are >> some time that the tlb entry has been >> invalidated in vcpu#0, but the corresponding entry in vcpu#1 could >> still been accessed, which seems cause some problem. > > The TLB flush is performed synchronously, see the effect of the wait > parameter to the smp_call_function_many function, and how that function > is called at arch/x86/mm/tlb.c. > if kvm_vcpu_kick sends an host-IPI to the remote vcpu, and if that vcpu is in guest mode. Is it possible that the remote vcpu has accessed the stale tlb entry before the host-IPI arrives? If that's the case, how this problem to be solved in KVM? Thanks for your time! >> > 2) Host CPU supports APIC virtualization (see commit 83d4c286931c and >> > Intel's documentation): >> > A bit is set in the posted interrupt section, and a special host-IPI is >> > delivered to the target cpu where the guest vcpu is scheduled >> > (vmx_deliver_posted_interrupt) which causes the hardware to >> > inject the vector (without VM-exit). >> > >> > >> I did not find this function (vmx_deliver_posted_interrupt) in my kvm >> kernel module. >> Does that mean my hardware doesn't support APIC virtualization? > > No, it means the kvm codebase you are looking at does not support it. > >> Thanks for your time! >> >> Best Wishes, >> Yaohui Hu >> -- >> To unsubscribe from this list: send the line "unsubscribe kvm" in >> the body of a message to majordomo@xxxxxxxxxxxxxxx >> More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html