On 17.07.2013, at 17:47, Bhushan Bharat-R65777 wrote: > >>>>>> On 17.07.2013, at 13:00, Gleb Natapov wrote: >>>>>> >>>>>>> On Tue, Jul 16, 2013 at 06:04:34PM -0500, Scott Wood wrote: >>>>>>>> On 07/16/2013 01:35:55 AM, Gleb Natapov wrote: >>>>>>>>> On Mon, Jul 15, 2013 at 01:17:33PM -0500, Scott Wood wrote: >>>>>>>>>> On 07/15/2013 06:30:20 AM, Gleb Natapov wrote: >>>>>>>>>>> There is no much sense to share hypercalls between architectures. >>>>>>>>>>> There >>>>>>>>>>> is zero probability x86 will implement those for instance >>>>>>>>>> >>>>>>>>>> This is similar to the question of whether to keep device API >>>>>>>>>> enumerations per-architecture... It costs very little to keep >>>>>>>>>> it in a common place, and it's hard to go back in the other >>>>>>>>>> direction if we later realize there are things that should be shared. >>>>>>>>>> >>>>>>>>> This is different from device API since with device API all >>>>>>>>> arches have to create/destroy devices, so it make sense to put >>>>>>>>> device lifecycle management into the common code, and device API >>>>>>>>> has single entry point to the code - device fd ioctl - where it >>>>>>>>> makes sense to handle common tasks, if any, and despatch others >>>>>>>>> to specific device implementation. >>>>>>>>> >>>>>>>>> This is totally unlike hypercalls which are, by definition, very >>>>>>>>> architecture specific (the way they are triggered, the way >>>>>>>>> parameter are passed from guest to host, what hypercalls arch needs...). >>>>>>>> >>>>>>>> The ABI is architecture specific. The API doesn't need to be, >>>>>>>> any more than it does with syscalls (I consider the >>>>>>>> architecture-specific definition of syscall numbers and similar >>>>>>>> constants in Linux to be unfortunate, especially for tools such >>>>>>>> as strace or QEMU's linux-user emulation). >>>>>>>> >>>>>>> Unlike syscalls different arches have very different ideas what >>>>>>> hypercalls they need to implement, so while with unified syscall >>>>>>> space I can see how it may benefit (very) small number of tools, I >>>>>>> do not see what advantage it will give us. The disadvantage is one >>>>>>> more global name space to manage. >>>>>>> >>>>>>>>>> Keeping it in a common place also makes it more visible to >>>>>>>>>> people looking to add new hcalls, which could cut down on >>>>>>>>>> reinventing the wheel. >>>>>>>>> I do not want other arches to start using hypercalls in the way >>>>>>>>> powerpc started to use them: separate device io space, so it is >>>>>>>>> better to hide this as far away from common code as possible :) >>>>>>>>> But on a more serious note hypercalls should be a last resort >>>>>>>>> and added only when no other possibility exists, so people >>>>>>>>> should not look what hcalls others implemented, so they can add >>>>>>>>> them to their favorite arch, but they should have a problem at >>>>>>>>> hand that they cannot solve without hcall, but at this point >>>>>>>>> they will have pretty good idea what this hcall should do. >>>>>>>> >>>>>>>> Why are hcalls such a bad thing? >>>>>>>> >>>>>>> Because they often used to do non architectural things making OSes >>>>>>> behave different from how they runs on real HW and real HW is what >>>>>>> OSes are designed and tested for. Example: there once was a KVM >>>>>>> (XEN have/had similar one) hypercall to accelerate MMU operation. >>>>>>> One thing it allowed is to to flush tlb without doing IPI if vcpu >>>>>>> is not running. Later optimization was added to Linux MMU code >>>>>>> that _relies_ on those IPIs for synchronisation. Good that at that >>>>>>> point those hypercalls were already deprecated on KVM (IIRC XEN >>>>>>> was broke for some time in that regard). Which brings me to >>>>>>> another point: they often get obsoleted by code improvement and HW >>>>>>> advancement (happened to aforementioned MMU hypercalls), but they >>>>>>> hard to deprecate if hypervisor supports live migration, without >>>>>>> live migration it is less of a problem. Next point is that people >>>>>>> often try to use them instead of emulate PV or real device just >>>>>>> because they think it is easier, but it is often not so. Example: >>>>>>> pvpanic device was initially proposed as hypercall, so lets say we >>>>>>> would implement it as such. It would have been KVM specific, >>>>>>> implementation would touch core guest KVM code and would have been >>>>>>> Linux guest specific. Instead it was implemented as platform >>>>>>> device with very small platform driver confined in drivers/ >>>>>>> directory, immediately usable by XEN and QEMU tcg in addition >>>>>> >>>>>> This is actually a very good point. How do we support reboot and >>>>>> shutdown for TCG guests? We surely don't want to expose TCG as KVM >> hypervisor. >>>>> >>>>> Hmm...so are you proposing that we abandon the current approach, and >>>>> switch to a device-based mechanism for reboot/shutdown? >>>> >>>> Reading Gleb's email it sounds like the more future proof approach, >>>> yes. I'm not quite sure yet where we should plug this though. >>> >>> What do you mean...where the paravirt device would go in the physical >>> address map?? >> >> Right. Either we >> >> - let the guest decide (PCI) >> - let QEMU decide, but potentially break the SoC layout (SysBus) >> - let QEMU decide, but only for the virt machine so that we don't break anyone >> (PlatBus) > > Can you please elaborate above two points ? If we emulate an MPC8544DS, we need to emulate an MPC8544DS. Any time we diverge from the layout of the original chip, things can break. However, for our PV machine (-M ppce500 / e500plat) we don't care about real hardware layouts. We simply emulate a machine that is 100% described through the device tree. So guests that can't deal with the machine looking different from real hardware don't really matter anyways, since they'd already be broken. Alex -- 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
