在 2023/11/22 下午1:52, Jason Wang 写道:

On Mon, Nov 20, 2023 at 8:38 PM Heng Qi [off-list ref] wrote:

quoted

By comparing the traffic information in the complete napi processes,
let the virtio-net driver automatically adjust the coalescing
moderation parameters of each receive queue.

Signed-off-by: Heng Qi <redacted>
---
v2->v3:
- Some minor modifications.

v1->v2:
- Improved the judgment of dim switch conditions.
- Cancel the work when vq reset.

  drivers/net/virtio_net.c | 191 ++++++++++++++++++++++++++++++++++-----
  1 file changed, 169 insertions(+), 22 deletions(-)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 69fe09e99b3c..bc32d5aae005 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c

@@ -19,6 +19,7 @@
  #include <linux/average.h>
  #include <linux/filter.h>
  #include <linux/kernel.h>
+#include <linux/dim.h>
  #include <net/route.h>
  #include <net/xdp.h>
  #include <net/net_failover.h>

@@ -172,6 +173,17 @@ struct receive_queue {

         struct virtnet_rq_stats stats;

+       /* The number of rx notifications */
+       u16 calls;
+
+       /* Is dynamic interrupt moderation enabled? */
+       bool dim_enabled;
+
+       /* Dynamic Interrupt Moderation */
+       struct dim dim;
+
+       u32 packets_in_napi;
+
         struct virtnet_interrupt_coalesce intr_coal;

         /* Chain pages by the private ptr. */

@@ -305,6 +317,9 @@ struct virtnet_info {
         u8 duplex;
         u32 speed;

+       /* Is rx dynamic interrupt moderation enabled? */
+       bool rx_dim_enabled;
+
         /* Interrupt coalescing settings */
         struct virtnet_interrupt_coalesce intr_coal_tx;
         struct virtnet_interrupt_coalesce intr_coal_rx;

@@ -2001,6 +2016,7 @@ static void skb_recv_done(struct virtqueue *rvq)
         struct virtnet_info *vi = rvq->vdev->priv;
         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];

+       rq->calls++;
         virtqueue_napi_schedule(&rq->napi, rvq);
  }

@@ -2141,6 +2157,26 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
         }
  }

+static void virtnet_rx_dim_work(struct work_struct *work);
+
+static void virtnet_rx_dim_update(struct virtnet_info *vi, struct receive_queue *rq)
+{
+       struct dim_sample cur_sample = {};
+
+       if (!rq->packets_in_napi)
+               return;
+
+       u64_stats_update_begin(&rq->stats.syncp);
+       dim_update_sample(rq->calls,
+                         u64_stats_read(&rq->stats.packets),
+                         u64_stats_read(&rq->stats.bytes),
+                         &cur_sample);
+       u64_stats_update_end(&rq->stats.syncp);
+
+       net_dim(&rq->dim, cur_sample);
+       rq->packets_in_napi = 0;
+}
+
  static int virtnet_poll(struct napi_struct *napi, int budget)
  {
         struct receive_queue *rq =

@@ -2149,17 +2185,22 @@ static int virtnet_poll(struct napi_struct *napi, int budget)
         struct send_queue *sq;
         unsigned int received;
         unsigned int xdp_xmit = 0;
+       bool napi_complete;

         virtnet_poll_cleantx(rq);

         received = virtnet_receive(rq, budget, &xdp_xmit);
+       rq->packets_in_napi += received;

         if (xdp_xmit & VIRTIO_XDP_REDIR)
                 xdp_do_flush();

         /* Out of packets? */
-       if (received < budget)
-               virtqueue_napi_complete(napi, rq->vq, received);
+       if (received < budget) {
+               napi_complete = virtqueue_napi_complete(napi, rq->vq, received);
+               if (napi_complete && rq->dim_enabled)
+                       virtnet_rx_dim_update(vi, rq);
+       }

         if (xdp_xmit & VIRTIO_XDP_TX) {
                 sq = virtnet_xdp_get_sq(vi);

@@ -2179,6 +2220,7 @@ static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
         virtnet_napi_tx_disable(&vi->sq[qp_index].napi);
         napi_disable(&vi->rq[qp_index].napi);
         xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
+       cancel_work_sync(&vi->rq[qp_index].dim.work);
  }

  static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)

@@ -2196,6 +2238,9 @@ static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
         if (err < 0)
                 goto err_xdp_reg_mem_model;

+       INIT_WORK(&vi->rq[qp_index].dim.work, virtnet_rx_dim_work);
+       vi->rq[qp_index].dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;

So in V2, you explained it can be done here but I want to know why it
must be done here.

It is not must be done here. We can place it in probe and make sure that 
INIT_WORK() is
done before napi is enabled for the first time.

So then I'll don't call INIT_WORK() every time after "napi_disable -> 
napi_enable" happens,
and use the following process:
Call INIT_WORK() in the driver probe and do cancel_work() after each 
related napi_disable.

This will be clearer.

Additionally, Yinjun Zhang found that cancel_work_sync() will cause a 
deadlock
when ndo_stop and virtnet_rx_dim_work occurs at the same time, so 
cancel_work() will be used.

For example, the refill_work is initialized in alloc_queues().

quoted

+
         virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
         virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);

@@ -2393,8 +2438,10 @@ static int virtnet_rx_resize(struct virtnet_info *vi,

         qindex = rq - vi->rq;

-       if (running)
+       if (running) {
                 napi_disable(&rq->napi);
+               cancel_work_sync(&rq->dim.work);
+       }

         err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
         if (err)

@@ -2403,8 +2450,10 @@ static int virtnet_rx_resize(struct virtnet_info *vi,
         if (!try_fill_recv(vi, rq, GFP_KERNEL))
                 schedule_delayed_work(&vi->refill, 0);

-       if (running)
+       if (running) {
+               INIT_WORK(&rq->dim.work, virtnet_rx_dim_work);
                 virtnet_napi_enable(rq->vq, &rq->napi);
+       }
         return err;
  }

@@ -3341,24 +3390,55 @@ static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
  static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
                                           struct ethtool_coalesce *ec)
  {
+       bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
+       bool update = false, switch_dim;
         struct scatterlist sgs_rx;
         int i;

-       vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
-       vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
-       sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
-
-       if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
-                                 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
-                                 &sgs_rx))
-               return -EINVAL;
+       switch_dim = rx_ctrl_dim_on != vi->rx_dim_enabled;
+       if (switch_dim) {
+               if (rx_ctrl_dim_on) {
+                       if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {

So basically, I'm asking why we need to duplicate the check here?

Here is VIRTIO_NET_F_*VQ*_NOTF_COAL.

E.g caller has done the check for us:

         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
                 ret = virtnet_send_notf_coal_cmds(vi, ec);
         else
                 ret = virtnet_coal_params_supported(ec);

?

Please also check other nested if/else, usually, too many levels of
nesting is a hint that the logic needs to be optimized.

I will optimize this.

Thanks!

Thanks