On 14/10/13 15:16, Alexander Graf wrote: > > On 14.10.2013, at 16:13, Marc Zyngier <marc.zyngier@xxxxxxx> wrote: > >> On 14/10/13 15:05, Michael S. Tsirkin wrote: >>> On Mon, Oct 14, 2013 at 02:49:10PM +0100, Marc Zyngier wrote: >>>> On 14/10/13 14:39, Alexander Graf wrote: >>>>> >>>>> On 14.10.2013, at 15:24, Marc Zyngier <marc.zyngier@xxxxxxx> wrote: >>>>> >>>>>> On 14/10/13 14:10, Alexander Graf wrote: >>>>>>> >>>>>>> On 14.10.2013, at 15:03, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: >>>>>>> >>>>>>>> Il 11/10/2013 16:36, Marc Zyngier ha scritto: >>>>>>>>> This small patch series adds just enough kernel infrastructure and >>>>>>>>> fixes to allow a BE guest to use virtio-mmio on a LE host, provided >>>>>>>>> that the host actually supports such madness. >>>>>>>> >>>>>>>> More precisely, it allows the guest drivers to pick the endianness they >>>>>>>> prefer. Mixed-endian virtio works fine on QEMU with e.g. a mips guest >>>>>>>> in emulation mode, because then any given QEMU binary will always use >>>>>>>> the same endianness (e.g. big for qemu-system-mips). >>>>>>> >>>>>>> We have the same problem (runtime switchable endianness) on PowerPC. IBM POWER is gaining Little Endian support in Linux now, so we could easily end up with an LE guest on a BE host. >>>>>>> >>>>>>> IIRC the way we're going to solve this is to hack up virtio_is_big_endian() to evaluate the first CPU's endianness mode (which will always be the same as all other CPU's endianness mode due to hypercall restrictions). >>>>>> >>>>>> I have implemented something similar for MMIO emulation in KVM/arm >>>>>> (except that I only care about the faulting CPU). >>>>>> >>>>>> See my initial patch for that: >>>>>> https://lists.cs.columbia.edu/pipermail/kvmarm/2013-October/007359.html >>>>>> >>>>>> That doesn't really change the non-trapping virtio accesses, though. >>>>>> Where is this virtio_is_big_endian() thing? >>>>> >>>>> It's in QEMU's exec.c. It only gets used for config space access that goes through PCI though. Is there any other place where virtio specifies native endianness today? >>>> >>>> That's the main problem. Today's virtio flavour doesn't specify anything >>>> about endianness, and that is what I'm adding. Or rather (as Paolo put >>>> it), the prefered endianness of the virtio driver. >>>> >>>> So once (and if) this flags are in place, you always know what you're >>>> dealing with. And because it is virtio-centric, you can implement it in >>>> an architecture independent way. >>>> >>>> Also, most of my life revolves around kvmtool. QEMU is hardly on my >>>> radar, these days (for reasons that are neither technical, nor relevant >>>> to this forum). So it is important to me that the solution is platform >>>> emulation agnostic. >>>> >>>> M. >>> >>> f you like, you should be able to implement virtio_is_big_endian >>> in kvmtool too. >> >> Sure. And I imagine this traps back into the kernel to read some >> register and find out what the endianness of the accessing CPU is? > > Not yet. To be exact, it does the below today. But all virtio device > emulation is 100% guest endianness unaware. This helper is the only > piece of code where it gets any idea what endianness the guest has. So > by checking for references to it in the code you know where endianness > is an issue. And that's only in the config space. Only config space? How do you deal with virtio ring descriptors, for example? M. -- Jazz is not dead. It just smells funny... -- 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
