On Tue, Mar 26, 2013 at 09:57:00AM -0700, Christoffer Dall wrote: > Hi guys, > > I have an x86 question that I really hope you can help with: > > Running the kvm-unit-tests on x86, I see a quite large difference > between I/O from the kernel, and I/O user (around 6,000 cycles) on two > separate platforms. > > This is probably expected, but what is the explanation? If I correlate > with system call overhead from lmbench, it's nowhere near that (in the > hundreds of cycles), so I'm sure I'm missing something obvious. > KVM does not restore full CPU host state on each vmexit since not all the state is used by the kernel while KVM is running on the CPU. Some state is restored only when vcpu thread is scheduled out from a host CPU and some state is restored only when CPU moves to userspace. To exit to userspace full host CPU state need to be restored. -- Gleb. -- 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