Re: [PATCH v2] virtio_balloon: stay awake while adjusting balloon

[David Stevens]

On Mon, Dec 18, 2023 at 12:33 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
>
> On 18.12.23 16:18, David Stevens wrote:
> > From: David Stevens <stevensd@xxxxxxxxxxxx>
> >
> > A virtio_balloon's parent device may be configured so that a
> > configuration change interrupt is a wakeup event. Extend the processing
> > of such a wakeup event until the balloon finishes inflating or deflating
> > by calling pm_stay_awake/pm_relax in the virtio_balloon driver. Note
> > that these calls are no-ops if the parent device doesn't support wakeup
> > events or if the wakeup events are not enabled.
> >
> > This change allows the guest to use system power states such as s2idle
> > without running the risk the virtio_balloon's cooperative memory
> > management becoming unresponsive to the host's requests.
> >
> > Signed-off-by: David Stevens <stevensd@xxxxxxxxxxxx>
> > ---
> > v1 -> v2:
> >   - Use adjustment_signal_pending flag instead of a sequence number
> >   - Call pm_stay_awake/pm_relax on parent device instead of adding a wake
> >     event to the virtio balloon device
> >
> >   drivers/virtio/virtio_balloon.c | 57 +++++++++++++++++++++++++++------
> >   1 file changed, 47 insertions(+), 10 deletions(-)
> >
> > diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c
> > index 1fe93e93f5bc..a3c11159cbe0 100644
> > --- a/drivers/virtio/virtio_balloon.c
> > +++ b/drivers/virtio/virtio_balloon.c
> > @@ -119,6 +119,11 @@ struct virtio_balloon {
> >       /* Free page reporting device */
> >       struct virtqueue *reporting_vq;
> >       struct page_reporting_dev_info pr_dev_info;
> > +
> > +     /* State for keeping the wakeup_source active while adjusting the balloon */
> > +     spinlock_t adjustment_lock;
> > +     bool adjustment_signal_pending;
> > +     bool adjustment_in_progress;
> >   };
> >
> >   static const struct virtio_device_id id_table[] = {
> > @@ -437,6 +442,31 @@ static void virtio_balloon_queue_free_page_work(struct virtio_balloon *vb)
> >       queue_work(vb->balloon_wq, &vb->report_free_page_work);
> >   }
> >
> > +static void start_update_balloon_size(struct virtio_balloon *vb)
> > +{
> > +     unsigned long flags;
> > +
> > +     spin_lock_irqsave(&vb->adjustment_lock, flags);
> > +     vb->adjustment_signal_pending = true;
> > +     if (!vb->adjustment_in_progress) {
> > +             vb->adjustment_in_progress = true;
> > +             pm_stay_awake(vb->vdev->dev.parent);
> > +     }
> > +     spin_unlock_irqrestore(&vb->adjustment_lock, flags);
> > +
> > +     queue_work(system_freezable_wq, &vb->update_balloon_size_work);
> > +}
> > +
> > +static void end_update_balloon_size(struct virtio_balloon *vb)
> > +{
> > +     spin_lock(&vb->adjustment_lock);
> > +     if (!vb->adjustment_signal_pending && vb->adjustment_in_progress) {
>
> How could vb->adjustment_in_progress ever not be set at this point?
>
> > +             vb->adjustment_in_progress = false;
> > +             pm_relax(vb->vdev->dev.parent);
> > +     }
> > +     spin_unlock(&vb->adjustment_lock);
> > +}
> > +
>
> LGTM, although I wonder what happens when calling pm_stay_awake() etc.
> on a parent device that is not wakeup-even-capable?

If the parent device is not wakeup capable or if wakeup isn't enabled,
then the vb->vdev->dev.parent->power.wakeup pointer will be NULL, so
the NULL checks in __pm_relax/__pm_stay_awake will immediately return.

-David