Re: [PATCH 28/31] vdpa: Expose VHOST_F_LOG_ALL on SVQ

[Eugenio Perez Martin]

On Thu, Feb 17, 2022 at 7:02 AM Jason Wang <jasowang@redhat.com> wrote:
>
> On Wed, Feb 16, 2022 at 11:54 PM Eugenio Perez Martin
> <eperezma@redhat.com> wrote:
> >
> > On Tue, Feb 8, 2022 at 9:25 AM Jason Wang <jasowang@redhat.com> wrote:
> > >
> > >
> > > 在 2022/2/1 下午7:45, Eugenio Perez Martin 写道:
> > > > On Sun, Jan 30, 2022 at 7:50 AM Jason Wang <jasowang@redhat.com> wrote:
> > > >>
> > > >> 在 2022/1/22 上午4:27, Eugenio Pérez 写道:
> > > >>> SVQ is able to log the dirty bits by itself, so let's use it to not
> > > >>> block migration.
> > > >>>
> > > >>> Also, ignore set and clear of VHOST_F_LOG_ALL on set_features if SVQ 
> > > >>> is
> > > >>> enabled. Even if the device supports it, the reports would be nonsense
> > > >>> because SVQ memory is in the qemu region.
> > > >>>
> > > >>> The log region is still allocated. Future changes might skip that, but
> > > >>> this series is already long enough.
> > > >>>
> > > >>> Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
> > > >>> ---
> > > >>>    hw/virtio/vhost-vdpa.c | 20 ++++++++++++++++++++
> > > >>>    1 file changed, 20 insertions(+)
> > > >>>
> > > >>> diff --git a/hw/virtio/vhost-vdpa.c b/hw/virtio/vhost-vdpa.c
> > > >>> index fb0a338baa..75090d65e8 100644
> > > >>> --- a/hw/virtio/vhost-vdpa.c
> > > >>> +++ b/hw/virtio/vhost-vdpa.c
> > > >>> @@ -1022,6 +1022,9 @@ static int vhost_vdpa_get_features(struct 
> > > >>> vhost_dev *dev, uint64_t *features)
> > > >>>        if (ret == 0 && v->shadow_vqs_enabled) {
> > > >>>            /* Filter only features that SVQ can offer to guest */
> > > >>>            vhost_svq_valid_guest_features(features);
> > > >>> +
> > > >>> +        /* Add SVQ logging capabilities */
> > > >>> +        *features |= BIT_ULL(VHOST_F_LOG_ALL);
> > > >>>        }
> > > >>>
> > > >>>        return ret;
> > > >>> @@ -1039,8 +1042,25 @@ static int vhost_vdpa_set_features(struct 
> > > >>> vhost_dev *dev,
> > > >>>
> > > >>>        if (v->shadow_vqs_enabled) {
> > > >>>            uint64_t dev_features, svq_features, acked_features;
> > > >>> +        uint8_t status = 0;
> > > >>>            bool ok;
> > > >>>
> > > >>> +        ret = vhost_vdpa_call(dev, VHOST_VDPA_GET_STATUS, &status);
> > > >>> +        if (unlikely(ret)) {
> > > >>> +            return ret;
> > > >>> +        }
> > > >>> +
> > > >>> +        if (status & VIRTIO_CONFIG_S_DRIVER_OK) {
> > > >>> +            /*
> > > >>> +             * vhost is trying to enable or disable _F_LOG, and the 
> > > >>> device
> > > >>> +             * would report wrong dirty pages. SVQ handles it.
> > > >>> +             */
> > > >>
> > > >> I fail to understand this comment, I'd think there's no way to disable
> > > >> dirty page tracking for SVQ.
> > > >>
> > > > vhost_log_global_{start,stop} are called at the beginning and end of
> > > > migration. To inform the device that it should start logging, they set
> > > > or clean VHOST_F_LOG_ALL at vhost_dev_set_log.
> > >
> > >
> > > Yes, but for SVQ, we can't disable dirty page tracking, isn't it? The
> > > only thing is to ignore or filter out the F_LOG_ALL and pretend to be
> > > enabled and disabled.
> > >
> >
> > Yes, that's what this patch does.
> >
> > >
> > > >
> > > > While SVQ does not use VHOST_F_LOG_ALL, it exports the feature bit so
> > > > vhost does not block migration. Maybe we need to look for another way
> > > > to do this?
> > >
> > >
> > > I'm fine with filtering since it's much more simpler, but I fail to
> > > understand why we need to check DRIVER_OK.
> > >
> >
> > Ok maybe I can make that part more clear,
> >
> > Since both operations use vhost_vdpa_set_features we must just filter
> > the one that actually sets or removes VHOST_F_LOG_ALL, without
> > affecting other features.
> >
> > In practice, that means to not forward the set features after
> > DRIVER_OK. The device is not expecting them anymore.
>
> I wonder what happens if we don't do this.
>

If we simply delete the check vhost_dev_set_features will return an
error, failing the start of the migration. More on this below.

> So kernel had this check:
>
>         /*
>          * It's not allowed to change the features after they have
>          * been negotiated.
>          */
> if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
>         return -EBUSY;
>
> So is it FEATURES_OK actually?
>

Yes, FEATURES_OK seems more appropriate actually so I will switch to
it for the next version.

But it should be functionally equivalent, since
vhost.c:vhost_dev_start sets both and the setting of _F_LOG_ALL cannot
be concurrent with it.

> For this patch, I wonder if the thing we need to do is to see whether
> it is a enable/disable F_LOG_ALL and simply return.
>

Yes, that's the intention of the patch.

We have 4 cases here:
a) We're being called from vhost_dev_start, with enable_log = false
b) We're being called from vhost_dev_start, with enable_log = true
c) We're being called from vhost_dev_set_log, with enable_log = false
d) We're being called from vhost_dev_set_log, with enable_log = true

The way to tell the difference between a/b and c/d is to check if
{FEATURES,DRIVER}_OK is set. And, as you point out in previous mails,
F_LOG_ALL must be filtered unconditionally since SVQ tracks dirty
memory through the memory unmapping, so we clear the bit
unconditionally if we detect that VHOST_SET_FEATURES will be called
(cases a and b).

Another possibility is to track if features have been set with a bool
in vhost_vdpa or something like that. But it seems cleaner to me to
only store that in the actual device.

> Thanks
>
> >
> > Does that make more sense?
> >
> > Thanks!
> >
> > > Thanks
> > >
> > >
> > > >
> > > > Thanks!
> > > >
> > > >> Thanks
> > > >>
> > > >>
> > > >>> +            return 0;
> > > >>> +        }
> > > >>> +
> > > >>> +        /* We must not ack _F_LOG if SVQ is enabled */
> > > >>> +        features &= ~BIT_ULL(VHOST_F_LOG_ALL);
> > > >>> +
> > > >>>            ret = vhost_vdpa_get_dev_features(dev, &dev_features);
> > > >>>            if (ret != 0) {
> > > >>>                error_report("Can't get vdpa device features, got 
> > > >>> (%d)", ret);
> > >
> >
>