On Fri, Nov 30, 2018 at 11:37:02AM +0000, Jean-Philippe Brucker wrote:
> Commit 78139c94dc8c ("net: vhost: lock the vqs one by one") moved the vq
> lock to improve scalability, but introduced a possible deadlock in
> vhost-iotlb. vhost_iotlb_notify_vq() now takes vq->mutex while holding
> the device's IOTLB spinlock.
Indeed spin_lock is just outside this snippet. Yack.
> And on the vhost_iotlb_miss() path, the
> spinlock is taken while holding vq->mutex.
>
> As long as we hold dev->mutex to prevent an ioctl from modifying
> vq->poll concurrently, we can safely call vhost_poll_queue() without
> holding vq->mutex. Since vhost_process_iotlb_msg() holds dev->mutex when
> calling vhost_iotlb_notify_vq(), avoid the deadlock by not taking
> vq->mutex.
>
> Fixes: 78139c94dc8c ("net: vhost: lock the vqs one by one")
> Signed-off-by: Jean-Philippe Brucker <jean-philippe.brucker@xxxxxxx>
Acked-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
but see below for a minor comment.
I guess we now need this on stable?
> ---
> drivers/vhost/vhost.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> index 3a5f81a66d34..1cbb17f898f7 100644
> --- a/drivers/vhost/vhost.c
> +++ b/drivers/vhost/vhost.c
> @@ -944,10 +944,10 @@ static void vhost_iotlb_notify_vq(struct vhost_dev *d,
> if (msg->iova <= vq_msg->iova &&
> msg->iova + msg->size - 1 >= vq_msg->iova &&
> vq_msg->type == VHOST_IOTLB_MISS) {
> - mutex_lock(&node->vq->mutex);
> + /* Safe to call outside vq->mutex as long as dev->mutex
> + * is held.
> + */
> vhost_poll_queue(&node->vq->poll);
> - mutex_unlock(&node->vq->mutex);
> -
In fact vhost_poll_queue is generally lockless so it's
safe to call without any locks.
> list_del(&node->node);
> kfree(node);
> }
> --
> 2.19.1
