On 03/12/2014 14:00, Christian Borntraeger wrote: > Am 03.12.2014 um 13:54 schrieb Paolo Bonzini: >> >> >> On 03/12/2014 13:12, David Hildenbrand wrote: >>>> This series improves yielding on architectures that cannot disable preemption >>>> while entering the guest and makes the creating thread of a VCPU the owning >>>> thread and therefore the yield target when yielding to that VCPU. >>>> >>>> We should focus on the case creating thread == executing thread and therefore >>>> remove the complicated handling of PIDs involving synchronize_rcus. >>>> >>>> This way we can speed up the creation of VCPUs and directly yield to the >>>> executing vcpu threads. >>>> >>>> Please note that - in theory - all VCPU ioctls should be triggered from the same >>>> VCPU thread, so changing threads is not a scenario we should optimize. >>>> >>>> >>>> David Hildenbrand (2): >>>> KVM: don't check for PF_VCPU when yielding >>>> KVM: thread creating a vcpu is the owner of that vcpu >>>> >>>> include/linux/kvm_host.h | 1 + >>>> virt/kvm/kvm_main.c | 22 ++-------------------- >>>> 2 files changed, 3 insertions(+), 20 deletions(-) >>>> >>> >>> Hi Paolo, >>> >>> would be good if you could have a look at these patches. >> >> Sure. >> >> I think patch 1 is fine and I am applying it. For patch 2, what about >> moving the ->pid assignment in the KVM_RUN case of kvm_vcpu_ioctl? > > That was my initial patch for the rcu specific latencies (do you > remember?) But IMHO patch 2 is actually the proper thing to do, no? Was it? :) Found it: http://permalink.gmane.org/gmane.comp.emulators.kvm.devel/125694 I don't know... I think it's feasible to have a userspace that creates all VCPUs in the main thread, and then runs them from multiple threads. Sure, those people would not have read the docs carefully, but it has worked until now. Wrongly-directed yields are a much better reason to apply your patch than QEMU slowness, and I've done that now. Paolo -- 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
