On Wed, Oct 13, 2021 at 08:30:40PM +0800, Yongji Xie wrote:
> On Wed, Oct 13, 2021 at 7:10 PM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >
> > On Wed, Sep 29, 2021 at 04:30:50PM +0800, Xie Yongji wrote:
> > > The interrupt might be triggered after a reset since there is
> > > no synchronization between resetting and irq injecting.
> >
> > In fact, irq_lock is already used to synchronize with
> > irqs. Why isn't taking and releasing it enough?
> >
>
> For example:
>
> CPU 0
> CPU1
> ---------
> --------
> vduse_dev_ioctl()
> check DRIVER_OK
>
> vduse_dev_reset()
>
> flush_work(&vq->inject);
> queue_work(vduse_irq_wq, &vq->inject);
>
> virtio_vdpa_probe()
>
> virtio_vdpa_find_vqs()
> vduse_vq_irq_inject()
> vq->cb.callback(vq->cb.private);
>
> set DRIVER_OK
>
> In the above case, the irq callback is still triggered before DRIVER_OK is set.
>
> But now I found it seems to be better to just check DRIVER_OK again in
> vduse_vq_irq_inject().
And then pesumably make sure each time we set status
it's done under the irq lock?
> > > And it
> > > might break something if the interrupt is delayed until a new
> > > round of device initialization.
> > >
> > > Fixes: c8a6153b6c59 ("vduse: Introduce VDUSE - vDPA Device in Userspace")
> > > Signed-off-by: Xie Yongji <xieyongji@xxxxxxxxxxxxx>
> > > ---
> > > drivers/vdpa/vdpa_user/vduse_dev.c | 37 +++++++++++++++++++++++++------------
> > > 1 file changed, 25 insertions(+), 12 deletions(-)
> > >
> > > diff --git a/drivers/vdpa/vdpa_user/vduse_dev.c b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > index cefb301b2ee4..841667a896dd 100644
> > > --- a/drivers/vdpa/vdpa_user/vduse_dev.c
> > > +++ b/drivers/vdpa/vdpa_user/vduse_dev.c
> > > @@ -80,6 +80,7 @@ struct vduse_dev {
> > > struct vdpa_callback config_cb;
> > > struct work_struct inject;
> > > spinlock_t irq_lock;
> > > + struct rw_semaphore rwsem;
> > > int minor;
> > > bool broken;
> > > bool connected;
> >
> > What does this lock protect? Use a more descriptive name pls,
> > and maybe add a comment.
> >
>
> This lock is used to ensure there is no more inflight irq kwork after reset.
>
> >
> > > @@ -410,6 +411,8 @@ static void vduse_dev_reset(struct vduse_dev *dev)
> > > if (domain->bounce_map)
> > > vduse_domain_reset_bounce_map(domain);
> > >
> > > + down_write(&dev->rwsem);
> > > +
> > > dev->status = 0;
> > > dev->driver_features = 0;
> > > dev->generation++;
> > > @@ -443,6 +446,8 @@ static void vduse_dev_reset(struct vduse_dev *dev)
> > > flush_work(&vq->inject);
> > > flush_work(&vq->kick);
> > > }
> > > +
> > > + up_write(&dev->rwsem);
> > > }
> > >
> > > static int vduse_vdpa_set_vq_address(struct vdpa_device *vdpa, u16 idx,
> > > @@ -885,6 +890,23 @@ static void vduse_vq_irq_inject(struct work_struct *work)
> > > spin_unlock_irq(&vq->irq_lock);
> > > }
> > >
> > > +static int vduse_dev_queue_irq_work(struct vduse_dev *dev,
> > > + struct work_struct *irq_work)
> > > +{
> > > + int ret = -EINVAL;
> > > +
> > > + down_read(&dev->rwsem);
> > > + if (!(dev->status & VIRTIO_CONFIG_S_DRIVER_OK))
> > > + goto unlock;
> > > +
> > > + ret = 0;
> > > + queue_work(vduse_irq_wq, irq_work);
> > > +unlock:
> > > + up_read(&dev->rwsem);
> > > +
> > > + return ret;
> > > +}
> > > +
> > > static long vduse_dev_ioctl(struct file *file, unsigned int cmd,
> > > unsigned long arg)
> > > {
> >
> >
> > so that's a lot of overhead for an irq.
> > Normally the way to address races like this is to add
> > flushing to the reset path, not locking to irq path.
> >
>
> Yes, we already call flush_work() in the reset path.
>
> Thanks,
> Yongji
_______________________________________________
Virtualization mailing list
Virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
