[PATCH net-next 0/3] dpaa2-eth: Add support for MQPRIO offloading

[PATCH net-next 0/3] dpaa2-eth: Add support for MQPRIO offloading

[Ioana Radulescu]

Add support for adding multiple TX traffic classes with mqprio. We have
one netdev queue and one hardware frame queue per TC.

Ioana Radulescu (3):
  dpaa2-eth: Refactor xps code
  dpaa2-eth: Support multiple traffic classes on Tx
  dpaa2-eth: Add mqprio support

 drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c | 129 ++++++++++++++++++-----
 drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h |   9 +-
 2 files changed, 112 insertions(+), 26 deletions(-)

-- 
2.7.4

[PATCH net-next 1/3] dpaa2-eth: Refactor xps code

[Ioana Radulescu]

Move the code configuring xps on the netdev TX queues to a
separate function. A subsequent patch will need to call
this in another context as well.

Signed-off-by: Ioana Radulescu <ruxandra.radulescu@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c | 45 +++++++++++++++++-------
 1 file changed, 32 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index 753957e..a12fc45 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -1872,6 +1872,35 @@ static int dpaa2_eth_xdp_xmit(struct net_device *net_dev, int n,
 	return n - drops;
 }
 
+static int update_xps(struct dpaa2_eth_priv *priv)
+{
+	struct net_device *net_dev = priv->net_dev;
+	struct cpumask xps_mask;
+	struct dpaa2_eth_fq *fq;
+	int i, num_queues;
+	int err = 0;
+
+	num_queues = dpaa2_eth_queue_count(priv);
+
+	/* The first <num_queues> entries in priv->fq array are Tx/Tx conf
+	 * queues, so only process those
+	 */
+	for (i = 0; i < num_queues; i++) {
+		fq = &priv->fq[i];
+
+		cpumask_clear(&xps_mask);
+		cpumask_set_cpu(fq->target_cpu, &xps_mask);
+
+		err = netif_set_xps_queue(net_dev, &xps_mask, i);
+		if (err) {
+			netdev_warn_once(net_dev, "Error setting XPS queue\n");
+			break;
+		}
+	}
+
+	return err;
+}
+
 static const struct net_device_ops dpaa2_eth_ops = {
 	.ndo_open = dpaa2_eth_open,
 	.ndo_start_xmit = dpaa2_eth_tx,
@@ -2138,10 +2167,9 @@ static struct dpaa2_eth_channel *get_affine_channel(struct dpaa2_eth_priv *priv,
 static void set_fq_affinity(struct dpaa2_eth_priv *priv)
 {
 	struct device *dev = priv->net_dev->dev.parent;
-	struct cpumask xps_mask;
 	struct dpaa2_eth_fq *fq;
 	int rx_cpu, txc_cpu;
-	int i, err;
+	int i;
 
 	/* For each FQ, pick one channel/CPU to deliver frames to.
 	 * This may well change at runtime, either through irqbalance or
@@ -2160,17 +2188,6 @@ static void set_fq_affinity(struct dpaa2_eth_priv *priv)
 			break;
 		case DPAA2_TX_CONF_FQ:
 			fq->target_cpu = txc_cpu;
-
-			/* Tell the stack to affine to txc_cpu the Tx queue
-			 * associated with the confirmation one
-			 */
-			cpumask_clear(&xps_mask);
-			cpumask_set_cpu(txc_cpu, &xps_mask);
-			err = netif_set_xps_queue(priv->net_dev, &xps_mask,
-						  fq->flowid);
-			if (err)
-				dev_err(dev, "Error setting XPS queue\n");
-
 			txc_cpu = cpumask_next(txc_cpu, &priv->dpio_cpumask);
 			if (txc_cpu >= nr_cpu_ids)
 				txc_cpu = cpumask_first(&priv->dpio_cpumask);
@@ -2180,6 +2197,8 @@ static void set_fq_affinity(struct dpaa2_eth_priv *priv)
 		}
 		fq->channel = get_affine_channel(priv, fq->target_cpu);
 	}
+
+	update_xps(priv);
 }
 
 static void setup_fqs(struct dpaa2_eth_priv *priv)
-- 
2.7.4

[PATCH net-next 2/3] dpaa2-eth: Support multiple traffic classes on Tx

[Ioana Radulescu]

DPNI objects can have multiple traffic classes, as reflected by
the num_tc attribute. Until now we ignored its value and only
used traffic class 0.

This patch adds support for multiple Tx traffic classes; the skb
priority information received from the stack is used to select the
hardware Tx queue on which to enqueue the frame.

Signed-off-by: Ioana Radulescu <ruxandra.radulescu@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c | 47 ++++++++++++++++--------
 drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h |  9 ++++-
 2 files changed, 40 insertions(+), 16 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index a12fc45..0f6d15c 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -757,6 +757,7 @@ static netdev_tx_t dpaa2_eth_tx(struct sk_buff *skb, struct net_device *net_dev)
 	u16 queue_mapping;
 	unsigned int needed_headroom;
 	u32 fd_len;
+	u8 prio;
 	int err, i;
 
 	percpu_stats = this_cpu_ptr(priv->percpu_stats);
@@ -814,6 +815,18 @@ static netdev_tx_t dpaa2_eth_tx(struct sk_buff *skb, struct net_device *net_dev)
 	 * a queue affined to the same core that processed the Rx frame
 	 */
 	queue_mapping = skb_get_queue_mapping(skb);
+	prio = netdev_txq_to_tc(net_dev, queue_mapping);
+
+	/* Hardware interprets priority level 0 as being the highest,
+	 * so we need to do a reverse mapping to the netdev tc index
+	 */
+	if (net_dev->num_tc)
+		prio = net_dev->num_tc - prio - 1;
+
+	/* We have only one FQ array entry for all Tx hardware queues
+	 * with the same flow id (but different priority levels)
+	 */
+	queue_mapping %= dpaa2_eth_queue_count(priv);
 	fq = &priv->fq[queue_mapping];
 
 	fd_len = dpaa2_fd_get_len(&fd);
@@ -824,7 +837,7 @@ static netdev_tx_t dpaa2_eth_tx(struct sk_buff *skb, struct net_device *net_dev)
 	 * the Tx confirmation callback for this frame
 	 */
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, &fd, 0);
+		err = priv->enqueue(priv, fq, &fd, prio);
 		if (err != -EBUSY)
 			break;
 	}
@@ -1877,16 +1890,17 @@ static int update_xps(struct dpaa2_eth_priv *priv)
 	struct net_device *net_dev = priv->net_dev;
 	struct cpumask xps_mask;
 	struct dpaa2_eth_fq *fq;
-	int i, num_queues;
+	int i, num_queues, netdev_queues;
 	int err = 0;
 
 	num_queues = dpaa2_eth_queue_count(priv);
+	netdev_queues = (net_dev->num_tc ? : 1) * num_queues;
 
 	/* The first <num_queues> entries in priv->fq array are Tx/Tx conf
 	 * queues, so only process those
 	 */
-	for (i = 0; i < num_queues; i++) {
-		fq = &priv->fq[i];
+	for (i = 0; i < netdev_queues; i++) {
+		fq = &priv->fq[i % num_queues];
 
 		cpumask_clear(&xps_mask);
 		cpumask_set_cpu(fq->target_cpu, &xps_mask);
@@ -2380,11 +2394,10 @@ static inline int dpaa2_eth_enqueue_qd(struct dpaa2_eth_priv *priv,
 
 static inline int dpaa2_eth_enqueue_fq(struct dpaa2_eth_priv *priv,
 				       struct dpaa2_eth_fq *fq,
-				       struct dpaa2_fd *fd,
-				       u8 prio __always_unused)
+				       struct dpaa2_fd *fd, u8 prio)
 {
 	return dpaa2_io_service_enqueue_fq(fq->channel->dpio,
-					   fq->tx_fqid, fd);
+					   fq->tx_fqid[prio], fd);
 }
 
 static void set_enqueue_mode(struct dpaa2_eth_priv *priv)
@@ -2540,17 +2553,21 @@ static int setup_tx_flow(struct dpaa2_eth_priv *priv,
 	struct device *dev = priv->net_dev->dev.parent;
 	struct dpni_queue queue;
 	struct dpni_queue_id qid;
-	int err;
+	int i, err;
 
-	err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
-			     DPNI_QUEUE_TX, 0, fq->flowid, &queue, &qid);
-	if (err) {
-		dev_err(dev, "dpni_get_queue(TX) failed\n");
-		return err;
+	for (i = 0; i < dpaa2_eth_tc_count(priv); i++) {
+		err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
+				     DPNI_QUEUE_TX, i, fq->flowid,
+				     &queue, &qid);
+		if (err) {
+			dev_err(dev, "dpni_get_queue(TX) failed\n");
+			return err;
+		}
+		fq->tx_fqid[i] = qid.fqid;
 	}
 
+	/* All Tx queues belonging to the same flowid have the same qdbin */
 	fq->tx_qdbin = qid.qdbin;
-	fq->tx_fqid = qid.fqid;
 
 	err = dpni_get_queue(priv->mc_io, 0, priv->mc_token,
 			     DPNI_QUEUE_TX_CONFIRM, 0, fq->flowid,
@@ -3250,7 +3267,7 @@ static int dpaa2_eth_probe(struct fsl_mc_device *dpni_dev)
 	dev = &dpni_dev->dev;
 
 	/* Net device */
-	net_dev = alloc_etherdev_mq(sizeof(*priv), DPAA2_ETH_MAX_TX_QUEUES);
+	net_dev = alloc_etherdev_mq(sizeof(*priv), DPAA2_ETH_MAX_NETDEV_QUEUES);
 	if (!net_dev) {
 		dev_err(dev, "alloc_etherdev_mq() failed\n");
 		return -ENOMEM;
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
index e180d5a..9af18c2 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
@@ -282,10 +282,13 @@ struct dpaa2_eth_ch_stats {
 };
 
 /* Maximum number of queues associated with a DPNI */
+#define DPAA2_ETH_MAX_TCS		8
 #define DPAA2_ETH_MAX_RX_QUEUES		16
 #define DPAA2_ETH_MAX_TX_QUEUES		16
 #define DPAA2_ETH_MAX_QUEUES		(DPAA2_ETH_MAX_RX_QUEUES + \
 					DPAA2_ETH_MAX_TX_QUEUES)
+#define DPAA2_ETH_MAX_NETDEV_QUEUES	\
+	(DPAA2_ETH_MAX_TX_QUEUES * DPAA2_ETH_MAX_TCS)
 
 #define DPAA2_ETH_MAX_DPCONS		16
 
@@ -299,8 +302,9 @@ struct dpaa2_eth_priv;
 struct dpaa2_eth_fq {
 	u32 fqid;
 	u32 tx_qdbin;
-	u32 tx_fqid;
+	u32 tx_fqid[DPAA2_ETH_MAX_TCS];
 	u16 flowid;
+	u8 tc;
 	int target_cpu;
 	u32 dq_frames;
 	u32 dq_bytes;
@@ -448,6 +452,9 @@ static inline int dpaa2_eth_cmp_dpni_ver(struct dpaa2_eth_priv *priv,
 #define dpaa2_eth_fs_count(priv)        \
 	((priv)->dpni_attrs.fs_entries)
 
+#define dpaa2_eth_tc_count(priv)	\
+	((priv)->dpni_attrs.num_tcs)
+
 /* We have exactly one {Rx, Tx conf} queue per channel */
 #define dpaa2_eth_queue_count(priv)     \
 	((priv)->num_channels)
-- 
2.7.4

[PATCH net-next 3/3] dpaa2-eth: Add mqprio support

[Ioana Radulescu]

Implement mqprio qdisc support by mapping traffic classes to
different hardware enqueue priorities. The maximum number of
supported traffic classes is an attribute of each DPNI object.

The traffic classes map to hardware priorities from highest (0)
to lowest (highest prio number). The driver assigns num_queues
to each traffic class, for a total of num_queues x num_tcs
hardware frame queues.

Signed-off-by: Ioana Radulescu <ruxandra.radulescu@nxp.com>
Signed-off-by: Bogdan Purcareata <bogdan.purcareata@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c | 43 ++++++++++++++++++++++++
 1 file changed, 43 insertions(+)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index 0f6d15c..cd34898 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -1915,6 +1915,48 @@ static int update_xps(struct dpaa2_eth_priv *priv)
 	return err;
 }
 
+static int dpaa2_eth_setup_tc(struct net_device *net_dev,
+			      enum tc_setup_type type, void *type_data)
+{
+	struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
+	struct tc_mqprio_qopt *mqprio = type_data;
+	u8 num_tc, num_queues;
+	int i;
+
+	if (type != TC_SETUP_QDISC_MQPRIO)
+		return -EINVAL;
+
+	mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
+	num_queues = dpaa2_eth_queue_count(priv);
+	num_tc = mqprio->num_tc;
+
+	if (num_tc == net_dev->num_tc)
+		return 0;
+
+	if (num_tc  > dpaa2_eth_tc_count(priv)) {
+		netdev_err(net_dev, "Max %d traffic classes supported\n",
+			   dpaa2_eth_tc_count(priv));
+		return -EINVAL;
+	}
+
+	if (!num_tc) {
+		netdev_reset_tc(net_dev);
+		netif_set_real_num_tx_queues(net_dev, num_queues);
+		goto out;
+	}
+
+	netdev_set_num_tc(net_dev, num_tc);
+	netif_set_real_num_tx_queues(net_dev, num_tc * num_queues);
+
+	for (i = 0; i < num_tc; i++)
+		netdev_set_tc_queue(net_dev, i, num_queues, i * num_queues);
+
+out:
+	update_xps(priv);
+
+	return 0;
+}
+
 static const struct net_device_ops dpaa2_eth_ops = {
 	.ndo_open = dpaa2_eth_open,
 	.ndo_start_xmit = dpaa2_eth_tx,
@@ -1927,6 +1969,7 @@ static const struct net_device_ops dpaa2_eth_ops = {
 	.ndo_change_mtu = dpaa2_eth_change_mtu,
 	.ndo_bpf = dpaa2_eth_xdp,
 	.ndo_xdp_xmit = dpaa2_eth_xdp_xmit,
+	.ndo_setup_tc = dpaa2_eth_setup_tc,
 };
 
 static void cdan_cb(struct dpaa2_io_notification_ctx *ctx)
-- 
2.7.4

Re: [PATCH net-next 2/3] dpaa2-eth: Support multiple traffic classes on Tx

[David Miller]

From: Ioana Radulescu <ruxandra.radulescu@nxp.com>
Date: Wed,  5 Jun 2019 12:57:25 +0300

> +	queue_mapping %= dpaa2_eth_queue_count(priv);

You are now performing a very expensive modulus operation every single TX
packet, regardless of whether TC is in use or not.

The whole reason we store the queue mapping pre-cooked in the SKB is so
that you don't have to do this.

RE: [PATCH net-next 2/3] dpaa2-eth: Support multiple traffic classes on Tx

[Ioana Ciocoi Radulescu]

> -----Original Message-----
> From: David Miller <davem@davemloft.net>
> Sent: Thursday, June 6, 2019 4:46 AM
> To: Ioana Ciocoi Radulescu <ruxandra.radulescu@nxp.com>
> Cc: netdev@vger.kernel.org; Ioana Ciornei <ioana.ciornei@nxp.com>
> Subject: Re: [PATCH net-next 2/3] dpaa2-eth: Support multiple traffic classes
> on Tx
> 
> From: Ioana Radulescu <ruxandra.radulescu@nxp.com>
> Date: Wed,  5 Jun 2019 12:57:25 +0300
> 
> > +	queue_mapping %= dpaa2_eth_queue_count(priv);
> 
> You are now performing a very expensive modulus operation every single TX
> packet, regardless of whether TC is in use or not.
> 
> The whole reason we store the queue mapping pre-cooked in the SKB is so
> that you don't have to do this.

Good point, thanks. Will refactor and send a v2.

Ioana