On Tue, 16 Mar 2021 14:24:38 +0000, Andrew Scull <ascull@xxxxxxxxxx> wrote: > > On Tue, Mar 16, 2021 at 10:13:10AM +0000, Marc Zyngier wrote: > > ZCR_EL2 controls the upper bound for ZCR_EL1, and is set to > > a potentially lower limit when the guest uses SVE. > > > > In order to restore the SVE state on the EL1 host, we must first > > reset ZCR_EL2 to its original value. > > > > Provide a hypervall that perform this reset. > > Is there a good reason to have an explicit hypercall vs trapping the > host access to SVE and restoring in that event? Trapping ZCR_EL2 isn't possible, as it would UNDEF at EL1. Trapping ZCR_EL1 accesses is possible though, but it'd mean leaving the SVE traps enabled on guest exit, something we currently don't do. > It's quite easy to do trap handling at EL2 now and it could let things > be even lazier, if that's any benefit in this case. We don't really have a good infrastructure for dealing with individual sysregs (pKVM will eventually change that, but we're not there yet, and it isn't clear how that'd apply to non-protected), so we'd have to deal with the whole SVE EC. What we could do is: - set CPTR_EL2.TZ set on guest exit - on SVE trap, reset ZCR_EL2, clear CPTR_EL2.TZ, reexecute the faulting instruction. I can have a look at how badly it looks. > Trapping seems to have had a bad rep in other conversations but I'm not > sure the same reasoning applies to this as well, or not. HVC and traps have the same basic cost. I seriously doubt you can measure the difference on any real CPU. Thanks, M. -- Without deviation from the norm, progress is not possible.
