On 04/14/2014 02:49 PM, Alexander Graf wrote: > > On 14.04.14 14:46, Michael S. Tsirkin wrote: >> On Mon, Apr 14, 2014 at 02:40:04PM +0200, Alexander Graf wrote: >>> On 14.04.14 14:37, Michael S. Tsirkin wrote: >>>> On Mon, Apr 14, 2014 at 02:29:20PM +0200, Alexander Graf wrote: >>>>> On 14.04.14 14:24, Michael S. Tsirkin wrote: >>>>> >>>> This will have to be measured and proved by whoever's proposing the >>>> patch, not by reviewers. Platforms such as AMD which don't do >>>> prediction well would be especially interesting to test on. >>> Sure, Greg, can you do that? I'm sure Michael has test cases >>> available he can give you to measure performance on this. >>> >>> Speaking of which, how does all of this work with vhost? >>> >>> >>> Alex >> I think that's missing. >> As a first step, we need to disable vhost when >> host/guest endian-ness do not match. >> >> We could also add cross-endian support to vhost. >> >> Or just wait a couple more months for virtio 1.0 which is fixed >> endian-ness. > > That won't help for current ppc64le guests, so I suppose we will > need cross-endian vhost.
Yes. For the moment, vhost=off is needed and even with that, qemu needs to byteswap a few attributes (see virtio-net header in patch 3). I am looking for a simple way to propagate the vring endianness to the host without breaking the compatibility with the current virtio drivers. Hopefully this is feasible. C.
