Re: [PATCH] vhost: configure all host notifiers in a single MR transaction

[Longpeng (Mike, Cloud Infrastructure Service Product Dept.)]

在 2022/11/21 12:01, Jason Wang 写道:
> On Fri, Nov 18, 2022 at 10:49 PM Longpeng(Mike) <longpeng2@huawei.com> wrote:
> > From: Longpeng <longpeng2@huawei.com>
> >
> > This allows the vhost device to batch the setup of all its host notifiers.
> > This significantly reduces the device starting time, e.g. the vhost-vDPA
> > generic device [1] start time reduce from 376ms to 9.1ms for a VM with
> > 64 vCPUs and 3 vDPA device(64vq per device).
>
> Great, I think we need to do this for host_notifiers_mr as well. This
> helps for the case when the notification area could be mapped directly
> to guests.
Batch and commit once for host_notifiers_mrs can reduce the cost from 
423ms to 32ms, testing on vdpasim_blk (3 devices and 64 vqs per device) 
with doorbell passthrough support.
I'll append a patch in the next version.

> > [1] https://www.mail-archive.com/qemu-devel@nongnu.org/msg921541.html
> >
> > Signed-off-by: Longpeng <longpeng2@huawei.com>
> > ---
> >   hw/virtio/vhost.c | 39 ++++++++++++++++++++++++++++++++++++++-
> >   1 file changed, 38 insertions(+), 1 deletion(-)
> >
> > diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
> > index d1c4c20b8c..bf82d9b176 100644
> > --- a/hw/virtio/vhost.c
> > +++ b/hw/virtio/vhost.c
> > @@ -1507,6 +1507,7 @@ void vhost_dev_cleanup(struct vhost_dev *hdev)
> >   int vhost_dev_enable_notifiers(struct vhost_dev *hdev, VirtIODevice *vdev)
> >   {
> >       BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
> > +    int vq_init_count = 0;
> >       int i, r, e;
> >
> >       /* We will pass the notifiers to the kernel, make sure that QEMU
> > @@ -1518,6 +1519,12 @@ int vhost_dev_enable_notifiers(struct vhost_dev *hdev, 
> > VirtIODevice *vdev)
> >           goto fail;
> >       }
> >
> > +    /*
> > +     * Batch all the host notifiers in a single transaction to avoid
> > +     * quadratic time complexity in address_space_update_ioeventfds().
> > +     */
> > +    memory_region_transaction_begin();
> > +
> >       for (i = 0; i < hdev->nvqs; ++i) {
> >           r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
> >                                            true);
> > @@ -1525,19 +1532,33 @@ int vhost_dev_enable_notifiers(struct vhost_dev *hdev, 
> > VirtIODevice *vdev)
> >               error_report("vhost VQ %d notifier binding failed: %d", i, -r);
> >               goto fail_vq;
> >           }
> > +
> > +        vq_init_count++;
> >       }
>
> Nit, the name needs some tweak, it's actually the number of the host
> notifiers that is initialized. And we can count it out of the loop.
Ok, I will refer to virtio_device_start_ioeventfd_impl().

> > +    memory_region_transaction_commit();
> > +
> >       return 0;
> >   fail_vq:
> > -    while (--i >= 0) {
> > +    for (i = 0; i < vq_init_count; i++) {
>
> It looks to me there's no need for this change.
>
> Others look good.
>
> Thanks
>
> >           e = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
> >                                            false);
> >           if (e < 0) {
> >               error_report("vhost VQ %d notifier cleanup error: %d", i, -r);
> >           }
> >           assert (e >= 0);
> > +    }
> > +
> > +    /*
> > +     * The transaction expects the ioeventfds to be open when it
> > +     * commits. Do it now, before the cleanup loop.
> > +     */
> > +    memory_region_transaction_commit();
> > +
> > +    for (i = 0; i < vq_init_count; i++) {
> >           virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + 
> > i);
> >       }
> > +
> >       virtio_device_release_ioeventfd(vdev);
> >   fail:
> >       return r;
> > @@ -1553,6 +1574,12 @@ void vhost_dev_disable_notifiers(struct vhost_dev *hdev, 
> > VirtIODevice *vdev)
> >       BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
> >       int i, r;
> >
> > +    /*
> > +     * Batch all the host notifiers in a single transaction to avoid
> > +     * quadratic time complexity in address_space_update_ioeventfds().
> > +     */
> > +    memory_region_transaction_begin();
> > +
> >       for (i = 0; i < hdev->nvqs; ++i) {
> >           r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + i,
> >                                            false);
> > @@ -1560,8 +1587,18 @@ void vhost_dev_disable_notifiers(struct vhost_dev *hdev, 
> > VirtIODevice *vdev)
> >               error_report("vhost VQ %d notifier cleanup failed: %d", i, -r);
> >           }
> >           assert (r >= 0);
> > +    }
> > +
> > +    /*
> > +     * The transaction expects the ioeventfds to be open when it
> > +     * commits. Do it now, before the cleanup loop.
> > +     */
> > +    memory_region_transaction_commit();
> > +
> > +    for (i = 0; i < hdev->nvqs; ++i) {
> >           virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), hdev->vq_index + 
> > i);
> >       }
> > +
> >       virtio_device_release_ioeventfd(vdev);
> >   }
> >
> > --
> > 2.23.0
>
> .