On 19/08/14 14:14, David Hildenbrand wrote: >> Il 19/08/2014 13:28, David Hildenbrand ha scritto: >>> Looking at the code, kvm_cpu_synchronize_state() seems to do these ioctls in >>> the vcpu thread (e.g. comming from cpu_synchronize_all_states()), any reasons >>> why kvm_cpu_synchronize_post_reset() doesn't do the same (e.g. called from >>> cpu_synchronize_all_post_reset())? >> >> No reason, feel free to post a patch for QEMU kvm-all.c. >> Documentation/virtual/kvm/api.txt clearly says: >> >> Only run vcpu ioctls from the same thread that was used to create the >> vcpu. >> >> Paolo >> > > Thanks! A little more tweaking in the other parts of s390x resets > and we should be able to reduce the number of "wrong" ioctls (I think I found > most cases that are responsible for the performance degradation). Hmm. We want to not only reduce, we want them be zero. In addition to a reworked MP_STATE patch set, we might be able to change the code to call "KVM_S390_INITIAL_RESET" only from the cpu thread itself. If that simplifies things, we could avoid doing KVM_S390_INITIAL_RESET on CPU creation, because we know that all kernel version will do an implicit cpu reset on cpu creation anyway. Can you have a try on this as well when reworking that code? We could then fix this rcu performance penalty independent from getting rid of that ioctl. Christian -- 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
