On 30 January 2015 at 01:36, Edgar E. Iglesias <edgar.igles...@gmail.com> wrote: > IIRC, last time the dedicated S-EL0 and S-EL1 MMU idx came up the > discussion went around flushing the qemu tlbs when switching between > S/NS. Having the dedicated MMU-idx is faster but for Aarch64 I think > we would need logic in at least the TTBRx access handlers to make use > of the dedicated secure MMU idx as Aarch64 secure monitors need to > reprogram the MMU when world switching. > > Another thing around the ARMMMUIdx_S2NS index. > From what I've seen, what would really help is having a fast > way to go from VM mode to non-vm mode. In particular for KVM. > For example when a guest writes to a virtio console there is alot > of ping-ponging between NS-S12(Guest) and NS-S1(Linux/KVM). > > Similary for XEN, it would really help to have that ASID/VMID indexed TLB I > think you suggested at some point. In XEN's case the ping-ponging > goes between two guests, domUs and dom0.
Yes, the lack of ASID/VMID restrictions on the TLB hurts us. It's tricky to implement in s/w without it being slower than what we have now, though. > I'm not try to indicate that you should add any of that now, > I'm just not sure sure it's worth adding the ARMMMUIdx_S2NS without > trying if it will actually give any real life improvements in > QEMU. Mostly my aim here was to make sure we were actually treating separately the various virt-to-phys mappings which the architecture says are separate, so we don't have nasty "hit something stale in the TLB" bugs to track down. -- PMM
