On Thu, Jul 4, 2013 at 8:43 PM, Gleb Natapov <gleb@xxxxxxxxxx> wrote: > On Thu, Jul 04, 2013 at 02:34:11PM +0200, Paolo Bonzini wrote: >> Il 04/07/2013 13:31, Gleb Natapov ha scritto: >> > On Thu, Jul 04, 2013 at 01:21:51PM +0200, Paolo Bonzini wrote: >> >> Il 04/07/2013 13:12, Gleb Natapov ha scritto: >> >>>> I don't like that it requires a firmware change in order to use nested >> >>>> VMX (at least for hypervisors that read the MSR). "Worse emulation" and >> >>>> "better emulation + new firmware" are indistiguishable from the point of >> >>>> view of anyone except the firmware. >> >>>> >> >>>> IMO there is no reason for a better emulation that no one would care >> >>>> about _and_ could look like a regression when updating to a newer kernel. >> >>> >> >>> That is why now is the good time to do that since nested vmx is not >> >>> widely used. When it will be widely used the change will be impossible >> >>> to do for reason you age giving. So it is now or never. >> >> >> >> I think it is a can of worms. For example, should this be >> >> conditionalized on running under QEMU? Under UEFI, TianoCore should be >> >> doing it, not SeaBIOS. And for CoreBoot, should it be done by CoreBoot >> >> or SeaBIOS? (How do people use KVM together with CoreBoot?) >> >> >> > This is not the first thing that firmware need to initialize. I let >> > firmware guys fight over who is doing it, we just model HW. FWIW for >> > Seabios patch would be trivial. >> >> Trivial but still depending on the question "who is doing it". If >> CoreBoot should (also) be doing it, the SeaBIOS patch should be >> conditional on CONFIG_QEMU. Also, should it be unconditional or depend >> on some external configuration knob (as on a bare-metal firmware)? >> > Let firmware developers solve firmware problems. They have all the same > problems when running on real HW and they will have to figure out a > solution regardless. Making things different on virt will only cause > people to treat virt differently (remember irqbalance?). I prefer to Gleb's idea. As nested virt is trying to provide a framework the same as the HW, we just need to model HW. If initialization works are done in KVM, the responsibility will be confused and some features are hard to expand in the future. e.g. if we want to add SMX support and let BIOS to configure MSR_IA32_FEATURE_CONTROL, it may hard to handle this at that time. Arthur > >> Actually KVM probes MSR_IA32_FEATURE_CONTROL itself and sets the bits, >> so we can sidestep the whole firmware thing, and go with a fixed version >> of Nadav's patch. >> > Indeed, so no regression will be seen even temporary. > >> >> So I still prefer never... :) >> > >> > This is a "can of worms" IMO. What we decide to init in KVM next to >> > relieve firmware from its duty? This is "other hypervisor" way, in KVM >> > we just model HW. >> >> FWIW, I now checked Xen nested VMX and it just returns 5, but this has >> nothing to do with paravirtualization). >> >> Paolo > > -- > 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
