[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CACGkMEvdbUzxMKoz+=TO8C5H06TvdL-nFPpe8qd==5PiXRjpTA@mail.gmail.com>
Date: Tue, 18 Jun 2024 09:29:48 +0800
From: Jason Wang <jasowang@...hat.com>
To: Heng Qi <hengqi@...ux.alibaba.com>
Cc: netdev@...r.kernel.org, virtualization@...ts.linux.dev,
"Michael S. Tsirkin" <mst@...hat.com>, Xuan Zhuo <xuanzhuo@...ux.alibaba.com>,
Eugenio Pérez <eperezma@...hat.com>,
Eric Dumazet <edumazet@...gle.com>, "David S. Miller" <davem@...emloft.net>,
Jakub Kicinski <kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>
Subject: Re: [PATCH net-next v3 4/4] virtio_net: improve dim command request efficiency
On Mon, Jun 17, 2024 at 4:08 PM Heng Qi <hengqi@...ux.alibaba.com> wrote:
>
> On Mon, 17 Jun 2024 12:05:30 +0800, Jason Wang <jasowang@...hat.com> wrote:
> > On Thu, Jun 6, 2024 at 2:15 PM Heng Qi <hengqi@...ux.alibaba.com> wrote:
> > >
> > > Currently, control vq handles commands synchronously,
> > > leading to increased delays for dim commands during multi-queue
> > > VM configuration and directly impacting dim performance.
> > >
> > > To address this, we are shifting to asynchronous processing of
> > > ctrlq's dim commands.
> > >
> > > Signed-off-by: Heng Qi <hengqi@...ux.alibaba.com>
> > > ---
> > > drivers/net/virtio_net.c | 233 ++++++++++++++++++++++++++++++++++-----
> > > 1 file changed, 208 insertions(+), 25 deletions(-)
> > >
> > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > > index e59e12bb7601..0338528993ab 100644
> > > --- a/drivers/net/virtio_net.c
> > > +++ b/drivers/net/virtio_net.c
> > > @@ -376,6 +376,13 @@ struct control_buf {
> > > struct completion completion;
> > > };
> > >
> > > +struct virtnet_coal_node {
> > > + struct control_buf ctrl;
> > > + struct virtio_net_ctrl_coal_vq coal_vqs;
> > > + bool is_coal_wait;
> > > + struct list_head list;
> > > +};
> > > +
> > > struct virtnet_info {
> > > struct virtio_device *vdev;
> > > struct virtqueue *cvq;
> > > @@ -420,6 +427,9 @@ struct virtnet_info {
> > > /* Lock to protect the control VQ */
> > > struct mutex cvq_lock;
> > >
> > > + /* Work struct for acquisition of cvq processing results. */
> > > + struct work_struct get_cvq;
> > > +
> > > /* Host can handle any s/g split between our header and packet data */
> > > bool any_header_sg;
> > >
> > > @@ -464,6 +474,14 @@ struct virtnet_info {
> > > struct virtnet_interrupt_coalesce intr_coal_tx;
> > > struct virtnet_interrupt_coalesce intr_coal_rx;
> > >
> > > + /* Free nodes used for concurrent delivery */
> > > + struct mutex coal_free_lock;
> > > + struct list_head coal_free_list;
> > > +
> > > + /* Filled when there are no free nodes or cvq buffers */
> > > + struct mutex coal_wait_lock;
> > > + struct list_head coal_wait_list;
> > > +
> > > unsigned long guest_offloads;
> > > unsigned long guest_offloads_capable;
> > >
> > > @@ -670,7 +688,7 @@ static void virtnet_cvq_done(struct virtqueue *cvq)
> > > {
> > > struct virtnet_info *vi = cvq->vdev->priv;
> > >
> > > - complete(&vi->ctrl->completion);
> > > + schedule_work(&vi->get_cvq);
> > > }
> > >
> > > static void skb_xmit_done(struct virtqueue *vq)
> > > @@ -2696,7 +2714,7 @@ static bool virtnet_send_command_reply(struct virtnet_info *vi,
> > > struct scatterlist *in)
> > > {
> > > struct scatterlist *sgs[5], hdr, stat;
> > > - u32 out_num = 0, tmp, in_num = 0;
> > > + u32 out_num = 0, in_num = 0;
> > > int ret;
> > >
> > > /* Caller should know better */
> > > @@ -2730,14 +2748,14 @@ static bool virtnet_send_command_reply(struct virtnet_info *vi,
> > > return false;
> > > }
> > >
> > > - if (unlikely(!virtqueue_kick(vi->cvq)))
> > > - goto unlock;
> > > + if (unlikely(!virtqueue_kick(vi->cvq))) {
> > > + mutex_unlock(&vi->cvq_lock);
> > > + return false;
> > > + }
> > > + mutex_unlock(&vi->cvq_lock);
> > >
> > > - wait_for_completion(&vi->ctrl->completion);
> > > - virtqueue_get_buf(vi->cvq, &tmp);
> > > + wait_for_completion(&ctrl->completion);
> > >
> > > -unlock:
> > > - mutex_unlock(&vi->cvq_lock);
> > > return ctrl->status == VIRTIO_NET_OK;
> > > }
> > >
> > > @@ -2747,6 +2765,86 @@ static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
> > > return virtnet_send_command_reply(vi, class, cmd, vi->ctrl, out, NULL);
> > > }
> > >
> > > +static void virtnet_process_dim_cmd(struct virtnet_info *vi,
> > > + struct virtnet_coal_node *node)
> > > +{
> > > + u16 qnum = le16_to_cpu(node->coal_vqs.vqn) / 2;
> > > +
> > > + mutex_lock(&vi->rq[qnum].dim_lock);
> > > + vi->rq[qnum].intr_coal.max_usecs =
> > > + le32_to_cpu(node->coal_vqs.coal.max_usecs);
> > > + vi->rq[qnum].intr_coal.max_packets =
> > > + le32_to_cpu(node->coal_vqs.coal.max_packets);
> > > + vi->rq[qnum].dim.state = DIM_START_MEASURE;
> > > + mutex_unlock(&vi->rq[qnum].dim_lock);
> > > +
> > > + if (node->is_coal_wait) {
> > > + mutex_lock(&vi->coal_wait_lock);
> > > + list_del(&node->list);
> > > + mutex_unlock(&vi->coal_wait_lock);
> > > + kfree(node);
> > > + } else {
> > > + mutex_lock(&vi->coal_free_lock);
> > > + list_add(&node->list, &vi->coal_free_list);
> > > + mutex_unlock(&vi->coal_free_lock);
> > > + }
> > > +}
> > > +
> > > +static int virtnet_add_dim_command(struct virtnet_info *vi,
> > > + struct virtnet_coal_node *coal_node)
> > > +{
> > > + struct scatterlist sg;
> > > + int ret;
> > > +
> > > + sg_init_one(&sg, &coal_node->coal_vqs, sizeof(coal_node->coal_vqs));
> > > + ret = virtnet_send_command_reply(vi, VIRTIO_NET_CTRL_NOTF_COAL,
> > > + VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
> > > + &coal_node->ctrl, &sg, NULL);
> > > + if (!ret) {
> > > + dev_warn(&vi->dev->dev,
> > > + "Failed to change coalescing params.\n");
> > > + return ret;
> > > + }
> > > +
> > > + virtnet_process_dim_cmd(vi, coal_node);
> > > +
> > > + return 0;
> > > +}
> > > +
> > > +static void virtnet_get_cvq_work(struct work_struct *work)
> > > +{
> > > + struct virtnet_info *vi =
> > > + container_of(work, struct virtnet_info, get_cvq);
> > > + struct virtnet_coal_node *wait_coal;
> > > + bool valid = false;
> > > + unsigned int tmp;
> > > + void *res;
> > > +
> > > + mutex_lock(&vi->cvq_lock);
> > > + while ((res = virtqueue_get_buf(vi->cvq, &tmp)) != NULL) {
> > > + complete((struct completion *)res);
> > > + valid = true;
> > > + }
> > > + mutex_unlock(&vi->cvq_lock);
> >
> > How could we synchronize with the device in this case?
> >
> > E.g what happens if the device finishes another buf here?
>
> That's a good question. I think we can solve it using the following snippet?
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index e59e12bb7601..5dc3e1244016 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
>
> @@ -420,6 +427,12 @@ struct virtnet_info {
> /* Lock to protect the control VQ */
> struct mutex cvq_lock;
>
> + /* Atomic to confirm whether the cvq work is scheduled. */
> + atomic_t scheduled;
> +
> + /* Work struct for acquisition of cvq processing results. */
> + struct work_struct get_cvq;
> +
>
>
> @@ -670,7 +691,9 @@ static void virtnet_cvq_done(struct virtqueue *cvq)
> {
> struct virtnet_info *vi = cvq->vdev->priv;
>
> - complete(&vi->ctrl->completion);
> + virtqueue_disable_cb(cvq);
> + if (!atomic_xchg(&vi->scheduled, 1))
> + schedule_work(&vi->get_cvq);
I think workqueue subsystem should already handle things like this.
> }
>
>
> +static void virtnet_get_cvq_work(struct work_struct *work)
> +{
> + struct virtnet_info *vi =
> + container_of(work, struct virtnet_info, get_cvq);
> + struct virtnet_coal_node *wait_coal;
> + bool valid = false;
> + unsigned int tmp;
> + void *res;
> +
> + mutex_lock(&vi->cvq_lock);
> + while ((res = virtqueue_get_buf(vi->cvq, &tmp)) != NULL) {
> + complete((struct completion *)res);
> + valid = true;
> + }
> + mutex_unlock(&vi->cvq_lock);
> +
> + atomic_set(&vi->scheduled, 0);
> + virtqueue_enable_cb_prepare(vi->cvq);
We have a brunch of examples in the current codes. Generally it should
be something like.
again:
disable_cb()
while(get_buf());
if (enable_cb())
disable_cb()
goto again;
> +}
>
> >
> > > +
> > > + if (!valid)
> > > + return;
> > > +
> > > + while (true) {
> > > + wait_coal = NULL;
> > > + mutex_lock(&vi->coal_wait_lock);
> > > + if (!list_empty(&vi->coal_wait_list))
> > > + wait_coal = list_first_entry(&vi->coal_wait_list,
> > > + struct virtnet_coal_node,
> > > + list);
> > > + mutex_unlock(&vi->coal_wait_lock);
> > > + if (wait_coal)
> > > + if (virtnet_add_dim_command(vi, wait_coal))
> > > + break;
> > > + else
> > > + break;
> > > + }
> >
> > This is still an ad-hoc optimization for dim in the general path here.
> >
> > Could we have a fn callback so for non dim it's just a completion and
> > for dim it would be a schedule_work()?
> >
>
> OK, I will try this.
>
> And how about this :
>
> +static void virtnet_cvq_work_sched(struct virtqueue *cvq)
> +{
> + struct virtnet_info *vi = cvq->vdev->priv;
> +
> + virtqueue_disable_cb(cvq);
> + if (!atomic_xchg(&vi->scheduled, 1))
> + schedule_work(&vi->get_cvq);
> +}
> +
> static void virtnet_cvq_done(struct virtqueue *cvq)
> {
> struct virtnet_info *vi = cvq->vdev->priv;
> + unsigned int tmp;
>
> + virtqueue_get_buf(vi->cvq, &tmp);
> complete(&vi->ctrl->completion);
> }
>
> @@ -5318,7 +5472,11 @@ static int virtnet_find_vqs(struct virtnet_info *vi)
>
> /* Parameters for control virtqueue, if any */
> if (vi->has_cvq) {
> - callbacks[total_vqs - 1] = virtnet_cvq_done;
> + if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
> + callbacks[total_vqs - 1] = virtnet_cvq_work_sched;
> + else
> + callbacks[total_vqs - 1] = virtnet_cvq_done;
> +
> names[total_vqs - 1] = "control";
> }
Just to clarify, I meant a callback function per control_buf. (I've
avoid touching virtqueue callback layers)
>
> > > +}
> > > static int virtnet_set_mac_address(struct net_device *dev, void *p)
> > > {
> > > struct virtnet_info *vi = netdev_priv(dev);
> > > @@ -4398,35 +4496,73 @@ static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
> > > return 0;
> > > }
> > >
> > > +static void virtnet_put_wait_coal(struct virtnet_info *vi,
> > > + struct receive_queue *rq,
> > > + struct dim_cq_moder moder)
> > > +{
> > > + struct virtnet_coal_node *wait_node;
> > > +
> > > + wait_node = kzalloc(sizeof(*wait_node), GFP_KERNEL);
> > > + if (!wait_node) {
> > > + rq->dim.state = DIM_START_MEASURE;
> > > + return;
> > > + }
> > > +
> > > + wait_node->is_coal_wait = true;
> > > + wait_node->coal_vqs.vqn = cpu_to_le16(rxq2vq(rq - vi->rq));
> > > + wait_node->coal_vqs.coal.max_usecs = cpu_to_le32(moder.usec);
> > > + wait_node->coal_vqs.coal.max_packets = cpu_to_le32(moder.pkts);
> > > + mutex_lock(&vi->coal_wait_lock);
> > > + list_add_tail(&wait_node->list, &vi->coal_wait_list);
> > > + mutex_unlock(&vi->coal_wait_lock);
> > > +}
> > > +
> > > static void virtnet_rx_dim_work(struct work_struct *work)
> > > {
> > > struct dim *dim = container_of(work, struct dim, work);
> > > struct receive_queue *rq = container_of(dim,
> > > struct receive_queue, dim);
> > > struct virtnet_info *vi = rq->vq->vdev->priv;
> > > - struct net_device *dev = vi->dev;
> > > + struct virtnet_coal_node *avail_coal;
> > > struct dim_cq_moder update_moder;
> > > - int qnum, err;
> > >
> > > - qnum = rq - vi->rq;
> > > + update_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
> > >
> > > mutex_lock(&rq->dim_lock);
> > > - if (!rq->dim_enabled)
> > > - goto out;
> > > -
> > > - update_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
> > > - if (update_moder.usec != rq->intr_coal.max_usecs ||
> > > - update_moder.pkts != rq->intr_coal.max_packets) {
> > > - err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, qnum,
> > > - update_moder.usec,
> > > - update_moder.pkts);
> > > - if (err)
> > > - pr_debug("%s: Failed to send dim parameters on rxq%d\n",
> > > - dev->name, qnum);
> > > - dim->state = DIM_START_MEASURE;
> > > + if (!rq->dim_enabled ||
> > > + (update_moder.usec == rq->intr_coal.max_usecs &&
> > > + update_moder.pkts == rq->intr_coal.max_packets)) {
> > > + rq->dim.state = DIM_START_MEASURE;
> > > + mutex_unlock(&rq->dim_lock);
> > > + return;
> > > }
> > > -out:
> > > mutex_unlock(&rq->dim_lock);
> > > +
> > > + mutex_lock(&vi->cvq_lock);
> > > + if (vi->cvq->num_free < 3) {
> > > + virtnet_put_wait_coal(vi, rq, update_moder);
> > > + mutex_unlock(&vi->cvq_lock);
> > > + return;
> > > + }
> >
> > Could we simply sleep instead of using a list here?
>
> Do you mean using a semaphore, or a waitqueue?
I meant sleep and wait for more space.
Thanks
Powered by blists - more mailing lists