On Mon, Nov 17, 2014 at 10:11:14AM +0000, Marc Zyngier wrote: > On 17/11/14 09:27, wanghaibin wrote: > > After recalculated the irq state of per vcpu, maybe, there is some vcpus > > became the state that have none pending irq form the state have pending > > irqs. > > So, it should be better that clear the irq_pending_on_cpu flag when > > recalculated result is false. > > Can you demonstrate a case where this is actually useful? > My guess is that it is extremely hard to measure an effect of this, but if you have a vcpu that clears the enable register or removes itself from the target register (masks the irq) then this change would cause one kick less from vgic_kick_vcpus(), otherwise it only saves a few other checks of the bitmaps. While I don't particularly mind merging this patch, I think a more interesting study would be exactly why we need this bit of state in the first place as opposed to just calculating in when needed. A measurement of the effects of that would be interesting. -Christoffer _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
