On Mon, Mar 11, 2013 at 07:27:44PM +0100, Jan Kiszka wrote: > On 2013-03-11 19:13, Gleb Natapov wrote: > > On Mon, Mar 11, 2013 at 07:05:48PM +0100, Jan Kiszka wrote: > >> On 2013-03-11 18:41, Gleb Natapov wrote: > >>> On Mon, Mar 11, 2013 at 06:34:03PM +0100, Jan Kiszka wrote: > >>>> On 2013-03-11 18:23, Gleb Natapov wrote: > >>>>> On Mon, Mar 11, 2013 at 04:36:33PM +0100, Jan Kiszka wrote: > >>>>>> On 2013-03-11 15:23, Paolo Bonzini wrote: > >>>>>>> Il 11/03/2013 15:05, Gleb Natapov ha scritto: > >>>>>>>> On Mon, Mar 11, 2013 at 03:01:40PM +0100, Jan Kiszka wrote: > >>>>>>>>>> We are not moving away from mp_state, we are moving away from using > >>>>>>>>>> mp_state for signaling because with nested virt INIT does not always > >>>>>>>>>> change mp_state, not only that it can change mp_state long after signal > >>>>>>>>>> is received after vmx off is done. > >>>>>>>>> > >>>>>>>>> Right. > >>>>>>>>> > >>>>>>>>> BTW, for that to happen, we will also need to influence the INIT level. > >>>>>>>>> Unless I misread the spec, INIT is blocked while in root mode, and if > >>>>>>>>> you deassert INIT before leaving root (vmxoff, vmenter), nothing > >>>>>>>>> actually happens. So what matters is the INIT signal level at the exit > >>>>>>>>> of root mode. > >>>>>>>>> > >>>>>>>> You are talking about INIT# signal received via CPU pin, right? I think > >>>>>>>> INIT send by IPI cannot go away. > >>>>>>> > >>>>>>> Neither can go away. For INIT sent by IPI, 10.4.7 says: > >>>>>>> > >>>>>>> Only the Pentium and P6 family processors support the INIT-deassert IPI. > >>>>>>> An INIT-disassert IPI has no affect on the state of the APIC, other than > >>>>>>> to reload the arbitration ID register with the value in the APIC ID > >>>>>>> register. > >>>>>>> > >>>>>>> 18.27.1 also says that "In the local APIC, NMI and INIT (except for INIT > >>>>>>> deassert) are always treated as edge triggered interrupts". > >>>>>>> > >>>>>>> > >>>>>>> For INIT#, the ICH9 chipset says that "INIT# is driven low for 16 PCI > >>>>>>> clocks" when a soft reset is requested. So we can guess that INIT# is > >>>>>>> also edge-triggered. > >>>>>> > >>>>>> Ah, ok. So, virtually, INIT stays asserted until it can be delivered in > >>>>>> form of a reset or a vmexit. > >>>>>> > >>>>> vmexit clears it? > >>>> > >>>> It has to. Otherwise, it would hit the host on vmxoff. > >>>> > >>> Why do you thing this is not happening? > >>> > >>> Look at [1] page 10 "VMX and INIT blocking". Do you think they were > >>> lucky to hit CPU while it was in a root mode? > >>> > >>> [1] http://www.invisiblethingslab.com/resources/2011/Software%20Attacks%20on%20Intel%20VT-d.pdf > >> > >> Interesting. And confusing. If a VMM cannot "consume" INIT events by > >> reentering the guest nor postpone those events up to that point if they > >> arrived in root mode, the whole vmexit-on-INIT thing is practically > >> useless. I wonder what use case Intel had in mind while designing this. > >> > > I actually find it very useful. On INIT vmexit hypervisor may call > > vmxoff and do proper reset. I find it less useful on AMD where you need > > to send self INIT IPI, but then how you can send self SIPI? > > Where's the difference? On Intel, SIPI is also not deliverable until > after vmxoff. So that signal has to come from the INIT sender, just like > on AMD. > On Intel: CPU 1 CPU 2 in a guest mode send INIT send SIPI INIT vmexit vmxoff reset and start from SIPI vector On AMD; CPU 1 CPU 2 in a guest mode send INIT send SIPI INIT vmexit self INIT IPI ??? > However, AMD allows you to NOT do a reset after leaving virtualization > mode. On Intel, INIT is obviously irreversible, thus of limited use. > Why would hypervisor send INIT to one of its CPUs if it does not want to reset it? :) -- Gleb. -- 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