On Tue, 19 Mar 2019 21:25:47 +0800 "Tangnianyao (ICT)" <tangnianyao@xxxxxxxxxx> wrote: > Hi, all > > Using gicv4, when guest is waiting for irq, it sends wfi and traps to kvm. > When vlpi is forwarded to PE after its_vpe_deschedule, before halt_poll in > kvm_vcpu_block, halt_poll may increase latency for this vlpi getting to guest. > In halt_poll process, it checks if there's pending irq for vcpu using pending_last. > However, doorbell is not enable at this moment and vlpi or doorbell can not set > pending_last true, to stop halt_poll. It will run until halt_poll time ends, if > there's no other physical irq coming in the meantime. And then vcpu is scheduled out. > This pending vlpi has to wait for vcpu getting schedule in next time. > > Should we enable doorbell before halt_poll process ? Enabling doorbells can be quite expensive. Depending on the HW, this is either: - a write to memory (+DSB, potential cache maintenance), a write to the INVLPI register, and a poll of the SYNC register - a write to memory (+DSB, potential cache maintenance), potentially a string of DISCARD+SYNC+MAPI+SYNC commands, and an INV+SYNC command Frankly, you want to be careful with that. I'd rather enable them late and have a chance of not blocking because of another (virtual) interrupt, which saves us the doorbell business. I wonder if you wouldn't be in a better position by drastically reducing halt_poll_ns for vcpu that can have directly injected interrupts. In any case, this is something that we should measure, not guess. Thanks, M. -- Without deviation from the norm, progress is not possible.
