Re: [PATCH net 2/4] vhost_net: rework on the lock ordering for busy polling

[Michael S. Tsirkin]

On Tue, Dec 11, 2018 at 11:06:43AM +0800, Jason Wang wrote:
> 
> On 2018/12/11 äå9:34, Michael S. Tsirkin wrote:
> > On Mon, Dec 10, 2018 at 05:44:52PM +0800, Jason Wang wrote:
> > > When we try to do rx busy polling in tx path in commit 441abde4cd84
> > > ("net: vhost: add rx busy polling in tx path"), we lock rx vq mutex
> > > after tx vq mutex is held. This may lead deadlock so we try to lock vq
> > > one by one in commit 78139c94dc8c ("net: vhost: lock the vqs one by
> > > one"). With this commit, we avoid the deadlock with the assumption
> > > that handle_rx() and handle_tx() run in a same process. But this
> > > commit remove the protection for IOTLB updating which requires the
> > > mutex of each vq to be held.
> > > 
> > > To solve this issue, the first step is to have a exact same lock
> > > ordering for vhost_net. This is done through:
> > > 
> > > - For handle_rx(), if busy polling is enabled, lock tx vq immediately.
> > > - For handle_tx(), always lock rx vq before tx vq, and unlock it if
> > >    busy polling is not enabled.
> > > - Remove the tricky locking codes in busy polling.
> > > 
> > > With this, we can have a exact same lock ordering for vhost_net, this
> > > allows us to safely revert commit 78139c94dc8c ("net: vhost: lock the
> > > vqs one by one") in next patch.
> > > 
> > > The patch will add two more atomic operations on the tx path during
> > > each round of handle_tx(). 1 byte TCP_RR does not notice such
> > > overhead.
> > > 
> > > Fixes: commit 78139c94dc8c ("net: vhost: lock the vqs one by one")
> > > Cc: Tonghao Zhang<xiangxia.m.yue@xxxxxxxxx>
> > > Signed-off-by: Jason Wang<jasowang@xxxxxxxxxx>
> > > ---
> > >   drivers/vhost/net.c | 18 +++++++++++++++---
> > >   1 file changed, 15 insertions(+), 3 deletions(-)
> > > 
> > > diff --git a/drivers/vhost/net.c b/drivers/vhost/net.c
> > > index ab11b2bee273..5f272ab4d5b4 100644
> > > --- a/drivers/vhost/net.c
> > > +++ b/drivers/vhost/net.c
> > > @@ -513,7 +513,6 @@ static void vhost_net_busy_poll(struct vhost_net *net,
> > >   	struct socket *sock;
> > >   	struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
> > > -	mutex_lock_nested(&vq->mutex, poll_rx ? VHOST_NET_VQ_TX: VHOST_NET_VQ_RX);
> > >   	vhost_disable_notify(&net->dev, vq);
> > >   	sock = rvq->private_data;
> > > @@ -543,8 +542,6 @@ static void vhost_net_busy_poll(struct vhost_net *net,
> > >   		vhost_net_busy_poll_try_queue(net, vq);
> > >   	else if (!poll_rx) /* On tx here, sock has no rx data. */
> > >   		vhost_enable_notify(&net->dev, rvq);
> > > -
> > > -	mutex_unlock(&vq->mutex);
> > >   }
> > >   static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
> > > @@ -913,10 +910,16 @@ static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
> > >   static void handle_tx(struct vhost_net *net)
> > >   {
> > >   	struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
> > > +	struct vhost_net_virtqueue *nvq_rx = &net->vqs[VHOST_NET_VQ_RX];
> > >   	struct vhost_virtqueue *vq = &nvq->vq;
> > > +	struct vhost_virtqueue *vq_rx = &nvq_rx->vq;
> > >   	struct socket *sock;
> > > +	mutex_lock_nested(&vq_rx->mutex, VHOST_NET_VQ_RX);
> > >   	mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
> > > +	if (!vq->busyloop_timeout)
> > > +		mutex_unlock(&vq_rx->mutex);
> > > +
> > >   	sock = vq->private_data;
> > >   	if (!sock)
> > >   		goto out;
> > > @@ -933,6 +936,8 @@ static void handle_tx(struct vhost_net *net)
> > >   		handle_tx_copy(net, sock);
> > >   out:
> > > +	if (vq->busyloop_timeout)
> > > +		mutex_unlock(&vq_rx->mutex);
> > >   	mutex_unlock(&vq->mutex);
> > >   }
> > So rx mutex taken on tx path now.  And tx mutex is on rc path ...  This
> > is just messed up. Why can't tx polling drop rx lock before
> > getting the tx lock and vice versa?
> 
> 
> Because we want to poll both tx and rx virtqueue at the same time
> (vhost_net_busy_poll()).
> 
> ÂÂÂ while (vhost_can_busy_poll(endtime)) {
> ÂÂ ÂÂÂ Âif (vhost_has_work(&net->dev)) {
> ÂÂ ÂÂÂ ÂÂÂ Â*busyloop_intr = true;
> ÂÂ ÂÂÂ ÂÂÂ Âbreak;
> ÂÂ ÂÂÂ Â}
> 
> ÂÂ ÂÂÂ Âif ((sock_has_rx_data(sock) &&
> ÂÂ ÂÂÂ ÂÂÂÂÂ !vhost_vq_avail_empty(&net->dev, rvq)) ||
> ÂÂ ÂÂÂ ÂÂÂÂ !vhost_vq_avail_empty(&net->dev, tvq))
> ÂÂ ÂÂÂ ÂÂÂ Âbreak;
> 
> ÂÂ ÂÂÂ Âcpu_relax();
> 
> ÂÂ Â}
> 
> 
> And we disable kicks and notification for better performance.

Right but it's all slow path - it happens when queue is
otherwise empty. So this is what I am saying: let's drop the locks
we hold around this.


> 
> > 
> > Or if we really wanted to force everything to be locked at
> > all times, let's just use a single mutex.
> > 
> > 
> > 
> 
> We could, but it might requires more changes which could be done for -next I
> believe.
> 
> 
> Thanks

I'd rather we kept the fine grained locking. E.g. people are
looking at splitting the tx and rx threads. But if not possible
let's fix it cleanly with a coarse-grained one. A mess here will
just create more trouble later.

-- 
MST