Merge branch 'virtio-net-dynamic-coalescing-moderation'

Heng Qi says:

====================
virtio-net: support dynamic coalescing moderation

Now, virtio-net already supports per-queue moderation parameter
setting. Based on this, we use the linux dimlib to support
dynamic coalescing moderation for virtio-net.

Due to some scheduling issues, we only support and test the rx dim.

Some test results:

I. Sockperf UDP
=================================================
1. Env
rxq_0 with affinity to cpu_0.

2. Cmd
client: taskset -c 0 sockperf tp -p 8989 -i $IP -t 10 -m 16B
server: taskset -c 0 sockperf sr -p 8989

3. Result
dim off: 1143277.00 rxpps, throughput 17.844 MBps, cpu is 100%.
dim on: 1124161.00 rxpps, throughput 17.610 MBps, cpu is 83.5%.
=================================================

II. Redis
=================================================
1. Env
There are 8 rxqs, and rxq_i with affinity to cpu_i.

2. Result
When all cpus are 100%, ops/sec of memtier_benchmark client is
dim off: 978437.23
dim on: 1143638.28
=================================================

III. Nginx
=================================================
1. Env
There are 8 rxqs and rxq_i with affinity to cpu_i.

2. Result
When all cpus are 100%, requests/sec of wrk client is
dim off: 877931.67
dim on: 1019160.31
=================================================

IV. Latency of sockperf udp
=================================================
1. Rx cmd
taskset -c 0 sockperf sr -p 8989

2. Tx cmd
taskset -c 0 sockperf pp -i ${ip} -p 8989 -t 10

After running this cmd 5 times and averaging the results,

3. Result
dim off: 17.7735 usec
dim on: 18.0110 usec
=================================================

Changelog:
v7->v8:
- Add select DIMLIB.

v6->v7:
- Drop the patch titled "spin lock for ctrl cmd access"
- Use rtnl_trylock to avoid the deadlock.

v5->v6:
- Add patch(4/5): spin lock for ctrl cmd access
- Patch(5/5):
- Use spin lock and cancel_work_sync to synchronize

v4->v5:
- Patch(4/4):
- Fix possible synchronization issues with cancel_work_sync.
- Reduce if/else nesting levels

v3->v4:
- Patch(5/5): drop.

v2->v3:
- Patch(4/5): some minor modifications.

v1->v2:
- Patch(2/5): a minor fix.
- Patch(4/5):
- improve the judgment of dim switch conditions.
- Cancel the work when vq reset.
- Patch(5/5): drop the tx dim implementation.
====================

Signed-off-by: David S. Miller <[e-mail]>

drivers/net/Kconfig

···
 	tristate "Virtio network driver"
 	depends on VIRTIO
 	select NET_FAILOVER
+	select DIMLIB
 	help
 	  This is the virtual network driver for virtio.  It can be used with
 	  QEMU based VMMs (like KVM or Xen).  Say Y or M.

drivers/net/virtio_net.c

···
 #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>
···
 
 	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. */
···
 	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;
···
 	}
 }
 
-static void virtqueue_napi_complete(struct napi_struct *napi,
+static bool virtqueue_napi_complete(struct napi_struct *napi,
 				    struct virtqueue *vq, int processed)
 {
 	int opaque;
···
 	if (napi_complete_done(napi, processed)) {
 		if (unlikely(virtqueue_poll(vq, opaque)))
 			virtqueue_napi_schedule(napi, vq);
+		else
+			return true;
 	} else {
 		virtqueue_disable_cb(vq);
 	}
+
+	return false;
 }
 
 static void skb_xmit_done(struct virtqueue *vq)
···
 	struct virtnet_info *vi = rvq->vdev->priv;
 	struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
 
+	rq->calls++;
 	virtqueue_napi_schedule(&rq->napi, rvq);
 }
 
···
 	}
 }
 
+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 =
···
 	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);
···
 	disable_delayed_refill(vi);
 	cancel_delayed_work_sync(&vi->refill);
 
-	for (i--; i >= 0; i--)
+	for (i--; i >= 0; i--) {
 		virtnet_disable_queue_pair(vi, i);
+		cancel_work_sync(&vi->rq[i].dim.work);
+	}
+
 	return err;
 }
 
···
 
 	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)
···
 	/* Make sure refill_work doesn't re-enable napi! */
 	cancel_delayed_work_sync(&vi->refill);
 
-	for (i = 0; i < vi->max_queue_pairs; i++)
+	for (i = 0; i < vi->max_queue_pairs; i++) {
 		virtnet_disable_queue_pair(vi, i);
+		cancel_work_sync(&vi->rq[i].dim.work);
+	}
 
 	return 0;
 }
···
 					    &vi->node_dead);
 }
 
+static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
+					 u16 vqn, u32 max_usecs, u32 max_packets)
+{
+	struct scatterlist sgs;
+
+	vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
+	vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
+	vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
+	sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
+
+	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
+				  VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
+				  &sgs))
+		return -EINVAL;
+
+	return 0;
+}
+
+static int virtnet_send_rx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
+					    u16 queue, u32 max_usecs,
+					    u32 max_packets)
+{
+	int err;
+
+	err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
+					    max_usecs, max_packets);
+	if (err)
+		return err;
+
+	vi->rq[queue].intr_coal.max_usecs = max_usecs;
+	vi->rq[queue].intr_coal.max_packets = max_packets;
+
+	return 0;
+}
+
+static int virtnet_send_tx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
+					    u16 queue, u32 max_usecs,
+					    u32 max_packets)
+{
+	int err;
+
+	err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
+					    max_usecs, max_packets);
+	if (err)
+		return err;
+
+	vi->sq[queue].intr_coal.max_usecs = max_usecs;
+	vi->sq[queue].intr_coal.max_packets = max_packets;
+
+	return 0;
+}
+
 static void virtnet_get_ringparam(struct net_device *dev,
 				  struct ethtool_ringparam *ring,
 				  struct kernel_ethtool_ringparam *kernel_ring,
···
 	ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
 	ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
 }
-
-static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
-					 u16 vqn, u32 max_usecs, u32 max_packets);
 
 static int virtnet_set_ringparam(struct net_device *dev,
 				 struct ethtool_ringparam *ring,
···
 			 * through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
 			 * did not set any TX coalescing parameters, to 0.
 			 */
-			err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(i),
-							    vi->intr_coal_tx.max_usecs,
-							    vi->intr_coal_tx.max_packets);
+			err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, i,
+							       vi->intr_coal_tx.max_usecs,
+							       vi->intr_coal_tx.max_packets);
 			if (err)
 				return err;
-
-			vi->sq[i].intr_coal.max_usecs = vi->intr_coal_tx.max_usecs;
-			vi->sq[i].intr_coal.max_packets = vi->intr_coal_tx.max_packets;
 		}
 
 		if (ring->rx_pending != rx_pending) {
···
 				return err;
 
 			/* The reason is same as the transmit virtqueue reset */
-			err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(i),
-							    vi->intr_coal_rx.max_usecs,
-							    vi->intr_coal_rx.max_packets);
+			err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, i,
+							       vi->intr_coal_rx.max_usecs,
+							       vi->intr_coal_rx.max_packets);
 			if (err)
 				return err;
-
-			vi->rq[i].intr_coal.max_usecs = vi->intr_coal_rx.max_usecs;
-			vi->rq[i].intr_coal.max_packets = vi->intr_coal_rx.max_packets;
 		}
 	}
 
···
 	return 0;
 }
 
-static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
-				       struct ethtool_coalesce *ec)
+static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
+					  struct ethtool_coalesce *ec)
 {
-	struct scatterlist sgs_tx, sgs_rx;
+	struct scatterlist sgs_tx;
 	int i;
 
 	vi->ctrl->coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
···
 				  &sgs_tx))
 		return -EINVAL;
 
-	/* Save parameters */
 	vi->intr_coal_tx.max_usecs = ec->tx_coalesce_usecs;
 	vi->intr_coal_tx.max_packets = ec->tx_max_coalesced_frames;
 	for (i = 0; i < vi->max_queue_pairs; i++) {
···
 		vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
 	}
 
+	return 0;
+}
+
+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;
+	struct scatterlist sgs_rx;
+	int i;
+
+	if (rx_ctrl_dim_on && !virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
+		return -EOPNOTSUPP;
+
+	if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != vi->intr_coal_rx.max_usecs ||
+			       ec->rx_max_coalesced_frames != vi->intr_coal_rx.max_packets))
+		return -EINVAL;
+
+	if (rx_ctrl_dim_on && !vi->rx_dim_enabled) {
+		vi->rx_dim_enabled = true;
+		for (i = 0; i < vi->max_queue_pairs; i++)
+			vi->rq[i].dim_enabled = true;
+		return 0;
+	}
+
+	if (!rx_ctrl_dim_on && vi->rx_dim_enabled) {
+		vi->rx_dim_enabled = false;
+		for (i = 0; i < vi->max_queue_pairs; i++)
+			vi->rq[i].dim_enabled = false;
+	}
+
+	/* Since the per-queue coalescing params can be set,
+	 * we need apply the global new params even if they
+	 * are not updated.
+	 */
 	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));
···
 				  &sgs_rx))
 		return -EINVAL;
 
-	/* Save parameters */
 	vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
 	vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
 	for (i = 0; i < vi->max_queue_pairs; i++) {
···
 	return 0;
 }
 
-static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
-					 u16 vqn, u32 max_usecs, u32 max_packets)
+static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
+				       struct ethtool_coalesce *ec)
 {
-	struct scatterlist sgs;
+	int err;
 
-	vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
-	vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
-	vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
-	sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
+	err = virtnet_send_tx_notf_coal_cmds(vi, ec);
+	if (err)
+		return err;
 
-	if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
-				  VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
-				  &sgs))
+	err = virtnet_send_rx_notf_coal_cmds(vi, ec);
+	if (err)
+		return err;
+
+	return 0;
+}
+
+static int virtnet_send_rx_notf_coal_vq_cmds(struct virtnet_info *vi,
+					     struct ethtool_coalesce *ec,
+					     u16 queue)
+{
+	bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
+	bool cur_rx_dim = vi->rq[queue].dim_enabled;
+	u32 max_usecs, max_packets;
+	int err;
+
+	max_usecs = vi->rq[queue].intr_coal.max_usecs;
+	max_packets = vi->rq[queue].intr_coal.max_packets;
+
+	if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != max_usecs ||
+			       ec->rx_max_coalesced_frames != max_packets))
 		return -EINVAL;
+
+	if (rx_ctrl_dim_on && !cur_rx_dim) {
+		vi->rq[queue].dim_enabled = true;
+		return 0;
+	}
+
+	if (!rx_ctrl_dim_on && cur_rx_dim)
+		vi->rq[queue].dim_enabled = false;
+
+	/* If no params are updated, userspace ethtool will
+	 * reject the modification.
+	 */
+	err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, queue,
+					       ec->rx_coalesce_usecs,
+					       ec->rx_max_coalesced_frames);
+	if (err)
+		return err;
 
 	return 0;
 }
···
 {
 	int err;
 
-	err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
-					    ec->rx_coalesce_usecs,
-					    ec->rx_max_coalesced_frames);
+	err = virtnet_send_rx_notf_coal_vq_cmds(vi, ec, queue);
 	if (err)
 		return err;
 
-	vi->rq[queue].intr_coal.max_usecs = ec->rx_coalesce_usecs;
-	vi->rq[queue].intr_coal.max_packets = ec->rx_max_coalesced_frames;
-
-	err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
-					    ec->tx_coalesce_usecs,
-					    ec->tx_max_coalesced_frames);
+	err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, queue,
+					       ec->tx_coalesce_usecs,
+					       ec->tx_max_coalesced_frames);
 	if (err)
 		return err;
-
-	vi->sq[queue].intr_coal.max_usecs = ec->tx_coalesce_usecs;
-	vi->sq[queue].intr_coal.max_packets = ec->tx_max_coalesced_frames;
 
 	return 0;
+}
+
+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 dim_cq_moder update_moder;
+	int i, qnum, err;
+
+	if (!rtnl_trylock())
+		return;
+
+	/* Each rxq's work is queued by "net_dim()->schedule_work()"
+	 * in response to NAPI traffic changes. Note that dim->profile_ix
+	 * for each rxq is updated prior to the queuing action.
+	 * So we only need to traverse and update profiles for all rxqs
+	 * in the work which is holding rtnl_lock.
+	 */
+	for (i = 0; i < vi->curr_queue_pairs; i++) {
+		rq = &vi->rq[i];
+		dim = &rq->dim;
+		qnum = rq - vi->rq;
+
+		if (!rq->dim_enabled)
+			continue;
+
+		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;
+		}
+	}
+
+	rtnl_unlock();
 }
 
 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
···
 		ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
 		ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
 		ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
+		ec->use_adaptive_rx_coalesce = vi->rx_dim_enabled;
 	} else {
 		ec->rx_max_coalesced_frames = 1;
 
···
 		ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
 		ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
 		ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
+		ec->use_adaptive_rx_coalesce = vi->rq[queue].dim_enabled;
 	} else {
 		ec->rx_max_coalesced_frames = 1;
 
···
 
 static const struct ethtool_ops virtnet_ethtool_ops = {
 	.supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
-		ETHTOOL_COALESCE_USECS,
+		ETHTOOL_COALESCE_USECS | ETHTOOL_COALESCE_USE_ADAPTIVE_RX,
 	.get_drvinfo = virtnet_get_drvinfo,
 	.get_link = ethtool_op_get_link,
 	.get_ringparam = virtnet_get_ringparam,
···
 		netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
 					 virtnet_poll_tx,
 					 napi_tx ? napi_weight : 0);
+
+		INIT_WORK(&vi->rq[i].dim.work, virtnet_rx_dim_work);
+		vi->rq[i].dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
 
 		sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
 		ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);