On Tue, Aug 07, 2018 at 10:08:28PM +0200, Christoffer Dall wrote: > On Tue, Aug 07, 2018 at 12:23:45PM +0100, Dave Martin wrote: > > On Mon, Aug 06, 2018 at 03:41:33PM +0200, Christoffer Dall wrote: [...] > > > I'm personally fine with both feature flags and vcpu device ioctls. If > > > using vcpu device ioctls gives you an obvious way to set attributes > > > relating to SVE, e.g. the vector length, then I think that's a strong > > > argument for that approach. > > > > There is another option I'm tending towards, which is simply to have > > a "set vector lengths" ioctl (whether presented as a vcpu device > > ioctl or a random arch ioctl). > > Someone complained once about adding too many arch ioctls because there > is a limited number space for doing so, but I'm not sure if that was and > still a valid concern. I have no strong opinion on this. I may stick with the separate ioctl approach for the next spin just to reduce the amount of churn, but I'm not overly committed to it. > > > > If that ioctl() fails then SVE support is not available. > > > > If it succeeds, it will update its arguments to indicate which > > vector lengths are enabled (if different). > > > > Old userspace, or userspace that doesn't want to use SVE, would > > not use this ioctl at all. > > > > It would also do no harm additionally to advertise this as a > > capability, though I wonder whether it's necessary to do so (?) > > > > It is customary to expose features via capabilities. I have a vague > recollection that tools like libvirt negotiate capabilities across > systems and would need more plumbing to discover features by probing an > ioctl instead. OK, fair enough. Cheers ---Dave _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
