Re: [PATCH 1/6] vduse: ensure vq->ready access is smp safe

[Jason Wang <jasowang@redhat.com>]

On Fri, Jan 30, 2026 at 3:56 PM Eugenio Perez Martin
<eperezma@redhat.com> wrote:
>
> On Fri, Jan 30, 2026 at 3:18 AM Jason Wang <jasowang@redhat.com> wrote:
> >
> > On Thu, Jan 29, 2026 at 2:21 PM Eugenio Perez Martin
> > <eperezma@redhat.com> wrote:
> > >
> > > On Thu, Jan 29, 2026 at 2:17 AM Jason Wang <jasowang@redhat.com> wrote:
> > > >
> > > > On Wed, Jan 28, 2026 at 8:45 PM Eugenio Pérez <eperezma@redhat.com> wrote:
> > > > >
> > > > > The vduse_vdpa_set_vq_ready can be called in the lifetime of the device
> > > > > well after initial setup, and the device can read it afterwards.
> > > > >
> > > > > Ensure that reads and writes to vq->ready are SMP safe so that the
> > > > > caller can trust that virtqueue kicks and calls behave as expected
> > > > > immediately after the operation returns.
> > > > >
> > > > > Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
> > > > > ---
> > > > >  drivers/vdpa/vdpa_user/vduse_dev.c | 34 +++++++++++++++++++++++-------
> > > > >  1 file changed, 26 insertions(+), 8 deletions(-)
> > > > >
> > > > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > > index 73d1d517dc6c..a4963aaf9332 100644
> > > > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > > > @@ -460,6 +460,24 @@ static __poll_t vduse_dev_poll(struct file *file, poll_table *wait)
> > > > >         return mask;
> > > > >  }
> > > > >
> > > > > +static bool vduse_vq_get_ready(const struct vduse_virtqueue *vq)
> > > > > +{
> > > > > +       /*
> > > > > +        * Paired with vduse_vq_set_ready smp_store, as the driver may modify
> > > > > +        * it while the VDUSE instance is reading it.
> > > > > +        */
> > > > > +       return smp_load_acquire(&vq->ready);
> > > > > +}
> > > > > +
> > > > > +static void vduse_vq_set_ready(struct vduse_virtqueue *vq, bool ready)
> > > > > +{
> > > > > +       /*
> > > > > +        * Paired with vduse_vq_get_ready smp_load, as the driver may modify
> > > > > +        * it while the VDUSE instance is reading it.
> > > > > +        */
> > > > > +       smp_store_release(&vq->ready, ready);
> > > >
> > > > Assuming this is not used in the datapath, I wonder if we can simply
> > > > use vq_lock mutex.
> > > >
> > >
> > > The function vduse_vq_set/get_ready are not in the datapath, but
> > > vduse_vq_kick and vduse_vq_signal_irqfd are.  I'm ok if you want to
> > > switch to vq_mutex if you want though, maybe it's even comparable with
> > > the cost of the ioctls or eventfd signaling.
> >
> > I'd like to use mutex for simplicity.
> >
>
> I cannot move it to a mutex, as we need to take it in the critical
> sections of the kick_lock and irq_lock spinlocks.
>
> I can move it to a spinlock, but it seems more complicated to me. We
> need to make sure we always take these kick_lock and irq_lock in the
> same order as the ready_lock, to not create deadlocks, and they always
> protect just the ready boolean. But sure, I'll send the spinlock
> version for V2.

Thinking about this, I'm not sure I understand the issue.

Maybe you can give me an example of the race.

THanks

>