[PATCH net-next 0/3] net: mvpp2: software TSO support

[PATCH net-next 0/3] net: mvpp2: software TSO support

[Antoine Tenart]

Hi all,

This series adds the s/w TSO support in the PPv2 driver, in addition to
two cosmetic commits. As stated in patch 3/3:

Using iperf and 10G ports, using TSO shows a significant performance
improvement by a factor 2 to reach around 9.5Gbps in TX; as well as a
significant CPU usage drop (from 25% to 15%).

Thanks!
Antoine

Antoine Tenart (3):
  net: define the TSO header size in net/tso.h
  net: mvpp2: unify the txq size define use
  net: mvpp2: software tso support

 drivers/net/ethernet/cavium/thunder/nicvf_queues.h |   1 -
 drivers/net/ethernet/freescale/fec_main.c          |   1 -
 drivers/net/ethernet/marvell/mv643xx_eth.c         |   2 -
 drivers/net/ethernet/marvell/mvneta.c              |   3 -
 drivers/net/ethernet/marvell/mvpp2.c               | 182 ++++++++++++++++++---
 include/net/tso.h                                  |   2 +
 6 files changed, 164 insertions(+), 27 deletions(-)

-- 
2.13.5

[PATCH net-next 1/3] net: define the TSO header size in net/tso.h

[Antoine Tenart]

The TSO header size was defined in many drivers. Factorize the code and
define its size in net/tso.h.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
---
 drivers/net/ethernet/cavium/thunder/nicvf_queues.h | 1 -
 drivers/net/ethernet/freescale/fec_main.c          | 1 -
 drivers/net/ethernet/marvell/mv643xx_eth.c         | 2 --
 drivers/net/ethernet/marvell/mvneta.c              | 3 ---
 include/net/tso.h                                  | 2 ++
 5 files changed, 2 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/cavium/thunder/nicvf_queues.h b/drivers/net/ethernet/cavium/thunder/nicvf_queues.h
index 57858522c33c..67d1a3230773 100644
--- a/drivers/net/ethernet/cavium/thunder/nicvf_queues.h
+++ b/drivers/net/ethernet/cavium/thunder/nicvf_queues.h
@@ -277,7 +277,6 @@ struct snd_queue {
 	u16		xdp_free_cnt;
 	bool		is_xdp;
 
-#define	TSO_HEADER_SIZE	128
 	/* For TSO segment's header */
 	char		*tso_hdrs;
 	dma_addr_t	tso_hdrs_phys;
diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index df09b254553d..56f56d6ada9c 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -226,7 +226,6 @@ MODULE_PARM_DESC(macaddr, "FEC Ethernet MAC address");
 
 #define COPYBREAK_DEFAULT	256
 
-#define TSO_HEADER_SIZE		128
 /* Max number of allowed TCP segments for software TSO */
 #define FEC_MAX_TSO_SEGS	100
 #define FEC_MAX_SKB_DESCS	(FEC_MAX_TSO_SEGS * 2 + MAX_SKB_FRAGS)
diff --git a/drivers/net/ethernet/marvell/mv643xx_eth.c b/drivers/net/ethernet/marvell/mv643xx_eth.c
index 9c94ea9b2b80..fb2d533ae4ef 100644
--- a/drivers/net/ethernet/marvell/mv643xx_eth.c
+++ b/drivers/net/ethernet/marvell/mv643xx_eth.c
@@ -183,8 +183,6 @@ static char mv643xx_eth_driver_version[] = "1.4";
 #define DEFAULT_TX_QUEUE_SIZE	512
 #define SKB_DMA_REALIGN		((PAGE_SIZE - NET_SKB_PAD) % SMP_CACHE_BYTES)
 
-#define TSO_HEADER_SIZE		128
-
 /* Max number of allowed TCP segments for software TSO */
 #define MV643XX_MAX_TSO_SEGS 100
 #define MV643XX_MAX_SKB_DESCS (MV643XX_MAX_TSO_SEGS * 2 + MAX_SKB_FRAGS)
diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
index 0aab74c2a209..35ff1ecfcff0 100644
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -281,9 +281,6 @@
  */
 #define MVNETA_RSS_LU_TABLE_SIZE	1
 
-/* TSO header size */
-#define TSO_HEADER_SIZE 128
-
 /* Max number of Rx descriptors */
 #define MVNETA_MAX_RXD 128
 
diff --git a/include/net/tso.h b/include/net/tso.h
index b7be852bfe9d..9a56c39e6d0a 100644
--- a/include/net/tso.h
+++ b/include/net/tso.h
@@ -3,6 +3,8 @@
 
 #include <net/ip.h>
 
+#define TSO_HEADER_SIZE		128
+
 struct tso_t {
 	int next_frag_idx;
 	void *data;
-- 
2.13.5

[PATCH net-next 2/3] net: mvpp2: unify the txq size define use

[Antoine Tenart]

The txq size is defined by MVPP2_AGGR_TXQ_SIZE, which is sometime not
used directly but through variables. As it is a fixed value use the
define everywhere in the driver.

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
---
 drivers/net/ethernet/marvell/mvpp2.c | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

diff --git a/drivers/net/ethernet/marvell/mvpp2.c b/drivers/net/ethernet/marvell/mvpp2.c
index 4a2ed5c66bb4..0f5a2a4f9ddf 100644
--- a/drivers/net/ethernet/marvell/mvpp2.c
+++ b/drivers/net/ethernet/marvell/mvpp2.c
@@ -5413,15 +5413,14 @@ static unsigned int mvpp2_tx_done(struct mvpp2_port *port, u32 cause,
 
 /* Allocate and initialize descriptors for aggr TXQ */
 static int mvpp2_aggr_txq_init(struct platform_device *pdev,
-			       struct mvpp2_tx_queue *aggr_txq,
-			       int desc_num, int cpu,
+			       struct mvpp2_tx_queue *aggr_txq, int cpu,
 			       struct mvpp2 *priv)
 {
 	u32 txq_dma;
 
 	/* Allocate memory for TX descriptors */
 	aggr_txq->descs = dma_alloc_coherent(&pdev->dev,
-				desc_num * MVPP2_DESC_ALIGNED_SIZE,
+				MVPP2_AGGR_TXQ_SIZE * MVPP2_DESC_ALIGNED_SIZE,
 				&aggr_txq->descs_dma, GFP_KERNEL);
 	if (!aggr_txq->descs)
 		return -ENOMEM;
@@ -5442,7 +5441,8 @@ static int mvpp2_aggr_txq_init(struct platform_device *pdev,
 			MVPP22_AGGR_TXQ_DESC_ADDR_OFFS;
 
 	mvpp2_write(priv, MVPP2_AGGR_TXQ_DESC_ADDR_REG(cpu), txq_dma);
-	mvpp2_write(priv, MVPP2_AGGR_TXQ_DESC_SIZE_REG(cpu), desc_num);
+	mvpp2_write(priv, MVPP2_AGGR_TXQ_DESC_SIZE_REG(cpu),
+		    MVPP2_AGGR_TXQ_SIZE);
 
 	return 0;
 }
@@ -7691,8 +7691,7 @@ static int mvpp2_init(struct platform_device *pdev, struct mvpp2 *priv)
 	for_each_present_cpu(i) {
 		priv->aggr_txqs[i].id = i;
 		priv->aggr_txqs[i].size = MVPP2_AGGR_TXQ_SIZE;
-		err = mvpp2_aggr_txq_init(pdev, &priv->aggr_txqs[i],
-					  MVPP2_AGGR_TXQ_SIZE, i, priv);
+		err = mvpp2_aggr_txq_init(pdev, &priv->aggr_txqs[i], i, priv);
 		if (err < 0)
 			return err;
 	}
-- 
2.13.5

[PATCH net-next 3/3] net: mvpp2: software tso support

[Antoine Tenart]

The patch uses the tso API to implement the tso functionality in Marvell
PPv2 driver.

Using iperf and 10G ports, using TSO shows a significant performance
improvement by a factor 2 to reach around 9.5Gbps in TX; as well as a
significant CPU usage drop (from 25% to 15%).

Signed-off-by: Antoine Tenart <antoine.tenart@free-electrons.com>
---
 drivers/net/ethernet/marvell/mvpp2.c | 171 ++++++++++++++++++++++++++++++++---
 1 file changed, 157 insertions(+), 14 deletions(-)

diff --git a/drivers/net/ethernet/marvell/mvpp2.c b/drivers/net/ethernet/marvell/mvpp2.c
index 0f5a2a4f9ddf..90103bce6c66 100644
--- a/drivers/net/ethernet/marvell/mvpp2.c
+++ b/drivers/net/ethernet/marvell/mvpp2.c
@@ -37,6 +37,7 @@
 #include <uapi/linux/ppp_defs.h>
 #include <net/ip.h>
 #include <net/ipv6.h>
+#include <net/tso.h>
 
 /* RX Fifo Registers */
 #define MVPP2_RX_DATA_FIFO_SIZE_REG(port)	(0x00 + 4 * (port))
@@ -1033,6 +1034,10 @@ struct mvpp2_txq_pcpu {
 
 	/* Index of the TX DMA descriptor to be cleaned up */
 	int txq_get_index;
+
+	/* DMA buffer for TSO headers */
+	char *tso_headers;
+	dma_addr_t tso_headers_dma;
 };
 
 struct mvpp2_tx_queue {
@@ -5623,6 +5628,14 @@ static int mvpp2_txq_init(struct mvpp2_port *port,
 		txq_pcpu->reserved_num = 0;
 		txq_pcpu->txq_put_index = 0;
 		txq_pcpu->txq_get_index = 0;
+
+		txq_pcpu->tso_headers =
+			dma_alloc_coherent(port->dev->dev.parent,
+					   MVPP2_AGGR_TXQ_SIZE * TSO_HEADER_SIZE,
+					   &txq_pcpu->tso_headers_dma,
+					   GFP_KERNEL);
+		if (!txq_pcpu->tso_headers)
+			goto cleanup;
 	}
 
 	return 0;
@@ -5630,6 +5643,11 @@ static int mvpp2_txq_init(struct mvpp2_port *port,
 	for_each_present_cpu(cpu) {
 		txq_pcpu = per_cpu_ptr(txq->pcpu, cpu);
 		kfree(txq_pcpu->buffs);
+
+		dma_free_coherent(port->dev->dev.parent,
+				  MVPP2_AGGR_TXQ_SIZE * MVPP2_DESC_ALIGNED_SIZE,
+				  txq_pcpu->tso_headers,
+				  txq_pcpu->tso_headers_dma);
 	}
 
 	dma_free_coherent(port->dev->dev.parent,
@@ -5649,6 +5667,11 @@ static void mvpp2_txq_deinit(struct mvpp2_port *port,
 	for_each_present_cpu(cpu) {
 		txq_pcpu = per_cpu_ptr(txq->pcpu, cpu);
 		kfree(txq_pcpu->buffs);
+
+		dma_free_coherent(port->dev->dev.parent,
+				  MVPP2_AGGR_TXQ_SIZE * MVPP2_DESC_ALIGNED_SIZE,
+				  txq_pcpu->tso_headers,
+				  txq_pcpu->tso_headers_dma);
 	}
 
 	if (txq->descs)
@@ -6238,6 +6261,123 @@ static int mvpp2_tx_frag_process(struct mvpp2_port *port, struct sk_buff *skb,
 	return -ENOMEM;
 }
 
+static inline void mvpp2_tso_put_hdr(struct sk_buff *skb,
+				     struct net_device *dev,
+				     struct mvpp2_tx_queue *txq,
+				     struct mvpp2_tx_queue *aggr_txq,
+				     struct mvpp2_txq_pcpu *txq_pcpu,
+				     int hdr_sz)
+{
+	struct mvpp2_port *port = netdev_priv(dev);
+	struct mvpp2_tx_desc *tx_desc = mvpp2_txq_next_desc_get(aggr_txq);
+	dma_addr_t addr;
+
+	mvpp2_txdesc_txq_set(port, tx_desc, txq->id);
+	mvpp2_txdesc_size_set(port, tx_desc, hdr_sz);
+
+	addr = txq_pcpu->tso_headers_dma +
+	       txq_pcpu->txq_put_index * TSO_HEADER_SIZE;
+	mvpp2_txdesc_offset_set(port, tx_desc, addr & MVPP2_TX_DESC_ALIGN);
+	mvpp2_txdesc_dma_addr_set(port, tx_desc, addr & ~MVPP2_TX_DESC_ALIGN);
+
+	mvpp2_txdesc_cmd_set(port, tx_desc, mvpp2_skb_tx_csum(port, skb) |
+					    MVPP2_TXD_F_DESC |
+					    MVPP2_TXD_PADDING_DISABLE);
+	mvpp2_txq_inc_put(port, txq_pcpu, NULL, tx_desc);
+}
+
+static inline int mvpp2_tso_put_data(struct sk_buff *skb,
+				     struct net_device *dev, struct tso_t *tso,
+				     struct mvpp2_tx_queue *txq,
+				     struct mvpp2_tx_queue *aggr_txq,
+				     struct mvpp2_txq_pcpu *txq_pcpu,
+				     int sz, bool left, bool last)
+{
+	struct mvpp2_port *port = netdev_priv(dev);
+	struct mvpp2_tx_desc *tx_desc = mvpp2_txq_next_desc_get(aggr_txq);
+	dma_addr_t buf_dma_addr;
+
+	mvpp2_txdesc_txq_set(port, tx_desc, txq->id);
+	mvpp2_txdesc_size_set(port, tx_desc, sz);
+
+	buf_dma_addr = dma_map_single(dev->dev.parent, tso->data, sz,
+				      DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(dev->dev.parent, buf_dma_addr))) {
+		mvpp2_txq_desc_put(txq);
+		return -ENOMEM;
+	}
+
+	mvpp2_txdesc_offset_set(port, tx_desc,
+				buf_dma_addr & MVPP2_TX_DESC_ALIGN);
+	mvpp2_txdesc_dma_addr_set(port, tx_desc,
+				  buf_dma_addr & ~MVPP2_TX_DESC_ALIGN);
+
+	if (!left) {
+		mvpp2_txdesc_cmd_set(port, tx_desc, MVPP2_TXD_L_DESC);
+		if (last) {
+			mvpp2_txq_inc_put(port, txq_pcpu, skb, tx_desc);
+			return 0;
+		}
+	} else {
+		mvpp2_txdesc_cmd_set(port, tx_desc, 0);
+	}
+
+	mvpp2_txq_inc_put(port, txq_pcpu, NULL, tx_desc);
+	return 0;
+}
+
+static int mvpp2_tx_tso(struct sk_buff *skb, struct net_device *dev,
+			struct mvpp2_tx_queue *txq,
+			struct mvpp2_tx_queue *aggr_txq,
+			struct mvpp2_txq_pcpu *txq_pcpu)
+{
+	struct mvpp2_port *port = netdev_priv(dev);
+	struct tso_t tso;
+	int hdr_sz = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	int i, len, descs = 0;
+
+	/* Check number of available descriptors */
+	if (mvpp2_aggr_desc_num_check(port->priv, aggr_txq,
+				      tso_count_descs(skb)) ||
+	    mvpp2_txq_reserved_desc_num_proc(port->priv, txq, txq_pcpu,
+					     tso_count_descs(skb)))
+		return 0;
+
+	tso_start(skb, &tso);
+	len = skb->len - hdr_sz;
+	while (len > 0) {
+		int left = min_t(int, skb_shinfo(skb)->gso_size, len);
+		char *hdr = txq_pcpu->tso_headers +
+			    txq_pcpu->txq_put_index * TSO_HEADER_SIZE;
+
+		len -= left;
+		descs++;
+
+		tso_build_hdr(skb, hdr, &tso, left, len == 0);
+		mvpp2_tso_put_hdr(skb, dev, txq, aggr_txq, txq_pcpu, hdr_sz);
+
+		while (left > 0) {
+			int sz = min_t(int, tso.size, left);
+			left -= sz;
+			descs++;
+
+			if (mvpp2_tso_put_data(skb, dev, &tso, txq, aggr_txq,
+					       txq_pcpu, sz, left, len == 0))
+				goto release;
+			tso_build_data(skb, &tso, sz);
+		}
+	}
+
+	return descs;
+
+release:
+	for (i = descs - 1; i >= 0; i--) {
+		struct mvpp2_tx_desc *tx_desc = txq->descs + i;
+		tx_desc_unmap_put(port, txq, tx_desc);
+	}
+	return 0;
+}
+
 /* Main tx processing */
 static int mvpp2_tx(struct sk_buff *skb, struct net_device *dev)
 {
@@ -6255,6 +6395,10 @@ static int mvpp2_tx(struct sk_buff *skb, struct net_device *dev)
 	txq_pcpu = this_cpu_ptr(txq->pcpu);
 	aggr_txq = &port->priv->aggr_txqs[smp_processor_id()];
 
+	if (skb_is_gso(skb)) {
+		frags = mvpp2_tx_tso(skb, dev, txq, aggr_txq, txq_pcpu);
+		goto out;
+	}
 	frags = skb_shinfo(skb)->nr_frags + 1;
 
 	/* Check number of available descriptors */
@@ -6304,22 +6448,21 @@ static int mvpp2_tx(struct sk_buff *skb, struct net_device *dev)
 		}
 	}
 
-	txq_pcpu->reserved_num -= frags;
-	txq_pcpu->count += frags;
-	aggr_txq->count += frags;
-
-	/* Enable transmit */
-	wmb();
-	mvpp2_aggr_txq_pend_desc_add(port, frags);
-
-	if (txq_pcpu->size - txq_pcpu->count < MAX_SKB_FRAGS + 1) {
-		struct netdev_queue *nq = netdev_get_tx_queue(dev, txq_id);
-
-		netif_tx_stop_queue(nq);
-	}
 out:
 	if (frags > 0) {
 		struct mvpp2_pcpu_stats *stats = this_cpu_ptr(port->stats);
+		struct netdev_queue *nq = netdev_get_tx_queue(dev, txq_id);
+
+		txq_pcpu->reserved_num -= frags;
+		txq_pcpu->count += frags;
+		aggr_txq->count += frags;
+
+		/* Enable transmit */
+		wmb();
+		mvpp2_aggr_txq_pend_desc_add(port, frags);
+
+		if (txq_pcpu->size - txq_pcpu->count < MAX_SKB_FRAGS + 1)
+			netif_tx_stop_queue(nq);
 
 		u64_stats_update_begin(&stats->syncp);
 		stats->tx_packets++;
@@ -7496,7 +7639,7 @@ static int mvpp2_port_probe(struct platform_device *pdev,
 		}
 	}
 
-	features = NETIF_F_SG | NETIF_F_IP_CSUM;
+	features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_TSO;
 	dev->features = features | NETIF_F_RXCSUM;
 	dev->hw_features |= features | NETIF_F_RXCSUM | NETIF_F_GRO;
 	dev->vlan_features |= features;
-- 
2.13.5

Re: [PATCH net-next 0/3] net: mvpp2: software TSO support

[David Miller]

From: Antoine Tenart <antoine.tenart@free-electrons.com>
Date: Wed, 23 Aug 2017 09:46:53 +0200

> This series adds the s/w TSO support in the PPv2 driver, in addition to
> two cosmetic commits. As stated in patch 3/3:
> 
> Using iperf and 10G ports, using TSO shows a significant performance
> improvement by a factor 2 to reach around 9.5Gbps in TX; as well as a
> significant CPU usage drop (from 25% to 15%).

Series applied, thank you.