On Fri, Oct 14, 2016 at 08:21:26PM +0200, Paolo Bonzini wrote: > On some benchmarks (e.g. netperf with ioeventfd disabled), APICv > posted interrupts turn out to be slower than interrupt injection via > KVM_REQ_EVENT. > > These patches save 40 cycles on kvm-unit-tests "inl" tests (1-2%), but I > suspect the effect is bigger on workloads that inject interrupts heavily. > > Patches 1-2 are micro-optimization of the APICv vmentry code. > > The interesting one is patch 3. The aim is for APICv to not use > KVM_REQ_EVENT at all for interrupts, therefore turning APICv's weakness > (having to look at PIR on every vmentry) into a strength (because > checking PIR.ON is cheaper than processing KVM_REQ_EVENT). I find that > the new code overall makes more sense than the old one, and I like that > more of the logic is contained within vmx_deliver_posted_interrupt and > vmx_sync_pir_to_irr. KVM_REQ_EVENT did protect from mistakes, so the new > code is more delicate, but I still prefer that to duct tape. :) > > Patches 4 and 5 are other cleanups.
FWIW, series Acked-by: Michael S. Tsirkin <[email protected]> I still think we should make pir volatile but we can argue about that later. > Paolo > > Paolo Bonzini (5): > KVM: x86: avoid atomic operations on APICv vmentry > KVM: x86: do not scan IRR twice on APICv vmentry > KVM: x86: do not use KVM_REQ_EVENT for APICv interrupt injection > KVM: x86: remove unnecessary sync_pir_to_irr > KVM: vmx: clear pending interrupts on KVM_SET_LAPIC > > arch/x86/kvm/lapic.c | 45 +++++++++++++++++++++++++++++---------------- > arch/x86/kvm/lapic.h | 4 ++-- > arch/x86/kvm/vmx.c | 44 ++++++++++++++++++++++++++++++++------------ > arch/x86/kvm/x86.c | 45 +++++++++++++++++++++++++-------------------- > 4 files changed, 88 insertions(+), 50 deletions(-) > > -- > 1.8.3.1
