Jason Wang
2021-Mar-08 07:01 UTC
[RFC v4 10/11] vduse: Introduce a workqueue for irq injection
On 2021/3/8 12:50 ??, Yongji Xie wrote:> On Mon, Mar 8, 2021 at 11:04 AM Jason Wang <jasowang at redhat.com> wrote: >> >> On 2021/3/5 4:12 ??, Yongji Xie wrote: >>> On Fri, Mar 5, 2021 at 3:37 PM Jason Wang <jasowang at redhat.com> wrote: >>>> On 2021/3/5 3:27 ??, Yongji Xie wrote: >>>>> On Fri, Mar 5, 2021 at 3:01 PM Jason Wang <jasowang at redhat.com> wrote: >>>>>> On 2021/3/5 2:36 ??, Yongji Xie wrote: >>>>>>> On Fri, Mar 5, 2021 at 11:42 AM Jason Wang <jasowang at redhat.com> wrote: >>>>>>>> On 2021/3/5 11:30 ??, Yongji Xie wrote: >>>>>>>>> On Fri, Mar 5, 2021 at 11:05 AM Jason Wang <jasowang at redhat.com> wrote: >>>>>>>>>> On 2021/3/4 4:58 ??, Yongji Xie wrote: >>>>>>>>>>> On Thu, Mar 4, 2021 at 2:59 PM Jason Wang <jasowang at redhat.com> wrote: >>>>>>>>>>>> On 2021/2/23 7:50 ??, Xie Yongji wrote: >>>>>>>>>>>>> This patch introduces a workqueue to support injecting >>>>>>>>>>>>> virtqueue's interrupt asynchronously. This is mainly >>>>>>>>>>>>> for performance considerations which makes sure the push() >>>>>>>>>>>>> and pop() for used vring can be asynchronous. >>>>>>>>>>>> Do you have pref numbers for this patch? >>>>>>>>>>>> >>>>>>>>>>> No, I can do some tests for it if needed. >>>>>>>>>>> >>>>>>>>>>> Another problem is the VIRTIO_RING_F_EVENT_IDX feature will be useless >>>>>>>>>>> if we call irq callback in ioctl context. Something like: >>>>>>>>>>> >>>>>>>>>>> virtqueue_push(); >>>>>>>>>>> virtio_notify(); >>>>>>>>>>> ioctl() >>>>>>>>>>> ------------------------------------------------- >>>>>>>>>>> irq_cb() >>>>>>>>>>> virtqueue_get_buf() >>>>>>>>>>> >>>>>>>>>>> The used vring is always empty each time we call virtqueue_push() in >>>>>>>>>>> userspace. Not sure if it is what we expected. >>>>>>>>>> I'm not sure I get the issue. >>>>>>>>>> >>>>>>>>>> THe used ring should be filled by virtqueue_push() which is done by >>>>>>>>>> userspace before? >>>>>>>>>> >>>>>>>>> After userspace call virtqueue_push(), it always call virtio_notify() >>>>>>>>> immediately. In traditional VM (vhost-vdpa) cases, virtio_notify() >>>>>>>>> will inject an irq to VM and return, then vcpu thread will call >>>>>>>>> interrupt handler. But in container (virtio-vdpa) cases, >>>>>>>>> virtio_notify() will call interrupt handler directly. So it looks like >>>>>>>>> we have to optimize the virtio-vdpa cases. But one problem is we don't >>>>>>>>> know whether we are in the VM user case or container user case. >>>>>>>> Yes, but I still don't get why used ring is empty after the ioctl()? >>>>>>>> Used ring does not use bounce page so it should be visible to the kernel >>>>>>>> driver. What did I miss :) ? >>>>>>>> >>>>>>> Sorry, I'm not saying the kernel can't see the correct used vring. I >>>>>>> mean the kernel will consume the used vring in the ioctl context >>>>>>> directly in the virtio-vdpa case. In userspace's view, that means >>>>>>> virtqueue_push() is used vring's producer and virtio_notify() is used >>>>>>> vring's consumer. They will be called one by one in one thread rather >>>>>>> than different threads, which looks odd and has a bad effect on >>>>>>> performance. >>>>>> Yes, that's why we need a workqueue (WQ_UNBOUND you used). Or do you >>>>>> want to squash this patch into patch 8? >>>>>> >>>>>> So I think we can see obvious difference when virtio-vdpa is used. >>>>>> >>>>> But it looks like we don't need this workqueue in vhost-vdpa cases. >>>>> Any suggestions? >>>> I haven't had a deep thought. But I feel we can solve this by using the >>>> irq bypass manager (or something similar). Then we don't need it to be >>>> relayed via workqueue and vdpa. But I'm not sure how hard it will be. >>>> >>> Or let vdpa bus drivers give us some information? >> >> This kind of 'type' is proposed in the early RFC of vDPA series. One >> issue is that at device level, we should not differ virtio from vhost, >> so if we introduce that, it might encourge people to design a device >> that is dedicated to vhost or virtio which might not be good. >> >> But we can re-visit this when necessary. >> > OK, I see. How about adding some information in ops.set_vq_cb()?I'm not sure I get this, maybe you can explain a little bit more? Thanks> > Thanks, > Yongji >