On 10/04/15 10:12, Christoffer Dall wrote: > On Thu, Apr 09, 2015 at 05:53:59PM +0100, Marc Zyngier wrote: >> The world switch spends quite some time dealing with the FP/SIMD >> registers, as the state is quite sizeable (32 128bit registers, >> plus some crumbs on the side). We save/restore them on each >> entry/exit, so that both the host and the guest always see >> the state they expect. >> >> But let's face it: the host kernel doesn't care. It is the host >> userspace that actually cares about FP. An obvious improvement is >> to remove the save/restore from the world switch, and only perform >> it when we're about to enter/exit the guest (by plugging it into >> vcpu_load/vcpu_put). The effect is pretty spectacular when running >> hackbench (which is the only benchmark worth looking at): >> > so the kernel never uses fp/simd registers for stuff like memcopies > etc.? It can, but wraps it with kernel_neon_begin/kernel_neon_end, which does its own save-restore. And when we perform our own save/restore, we do it with interrupts disabled so that the kernel cannot change the state during that time. > Can we also make a similar change for ARM on the 32-bit side? We could, but it is worth evaluating how much better this is compared to what we have already. Thanks, 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
