Re: [RFC] hw/nvme: Use irqfd to send interrupts

[Jinhao Fan]

at 12:18 PM, Klaus Jensen <its@irrelevant.dk> wrote:

> On Jul 12 14:26, Klaus Jensen wrote:
>> On Jul  9 12:35, Jinhao Fan wrote:
>>> Use irqfd to directly notify KVM to inject interrupts. This is done by
>>> registering a virtual IRQ(virq) in KVM and associate the virq with an
>>> irqfd, so that KVM can directly inject the interrupt when it receives
>>> notification from the irqfd. This approach is supposed to improve 
>>> performance because it bypasses QEMU's MSI interrupt emulation logic.
>>> 
>>> However, I did not see an obvious improvement of the emulation KIOPS:
>>> 
>>> QD      1   4  16  64 
>>> QEMU   38 123 210 329
>>> irqfd  40 129 219 328
>>> 
>>> I found this problem quite hard to diagnose since irqfd's workflow
>>> involves both QEMU and the in-kernel KVM. 
>>> 
>>> Could you help me figure out the following questions:
>>> 
>>> 1. How much performance improvement can I expect from using irqfd?
>> 
>> This is a level of QEMU/KVM that I am by no means an expert on and I
>> would have to let the broader QEMU community comment on this.
> 
> In any case, I'm wary about adding this level of kvm-dependence in the
> device. This wont work on non-kvm platforms any more.

Yes, irqfd seems only useful on KVM-based systems. Maybe it is more suitable
for vhost or VFIO based solutions which need irqfd to deliver interrupts.

> I think you should put irqfd on hold and focus on iothreads :)

I’m working on iothread currently. But I also observed a performance
regression with iothread enabled. I found ftrace, which is supported by both
QEMU and KVM, seems good for analyzing performance issues. I’m currently
exploring with ftrace.