[PATCH v5 0/6] net: fec: Enable Software TSO to improve the tx performance

[PATCH v5 0/6] net: fec: Enable Software TSO to improve the tx performance

[Fugang Duan]

Add SG and software TSO support for FEC.
This feature allows to improve outbound throughput performance.
Tested on imx6dl sabresd board, running iperf tcp tests shows:
        * 82% improvement comparing with NO SG & TSO patch

$ ethtool -K eth0 sg on
$ ethtool -K eth0 tso on
[  3] local 10.192.242.108 port 35388 connected with 10.192.242.167 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0- 3.0 sec   181 MBytes   506 Mbits/sec
* cpu loading is 30%

$ ethtool -K eth0 sg off
$ ethtool -K eth0 tso off
[  3] local 10.192.242.108 port 52618 connected with 10.192.242.167 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0- 3.0 sec  99.5 MBytes   278 Mbits/sec

FEC HW support IP header and TCP/UDP hw checksum, support multi buffer descriptor transfer
one frame, but don't support HW TSO. And imx6q/dl SOC FEC Gbps speed has HW bus Bandwidth
limitation (400Mbps ~ 700Mbps), imx6sx SOC FEC Gbps speed has no HW bandwidth limitation.

The patch set just enable TSO feature, which is done following the mv643xx_eth driver.

Test result analyze:
imx6dl sabresd board: there have 82% improvement, since imx6dl FEC HW has bandwidth limitation,
                      the performance with SW TSO is a milestone.

Addition test:
imx6sx sdb board:
upstream still don't support imx6sx due to some patches being upstream... they use same FEC IP.
Use the SW TSO patches test imx6sx sdb board in internal kernel tree:
No SW TSO patch: tx bandwidth 840Mbps, cpu loading is 100%.
SW TSO patch:    tx bandwidth 942Mbps, cpu loading is 65%.
It means the patch set have great improvement for imx6sx FEC performance.

V2:
* From Frank Li's suggestion:
	Change the API "fec_enet_txdesc_entry_free" name to "fec_enet_get_free_txdesc_num".
* Summary David Laight and Eric Dumazet's thoughts:
	RX BD entry number change to 256.
* From ezequiel's suggestion:
	Follow the latest TSO fixes from his solution to rework the queue stop/wake-up.
	Avoid unmapping the TSO header buffers.
* From Eric Dumazet's suggestion:
	Avoid more bytes copy, just copying the unaligned part of the payload into first
	descriptor. The suggestion will bring more complex for the driver, and imx6dl FEC
	DMA need 16 bytes alignment, but cpu loading is not problem that cpu loading is
	30%, the current performance is so better. Later chip like imx6sx Gigbit FEC DMA
	support byte alignment, so there don't exist memory copy. So, the V2 version drop
	the suggestion.
	Anyway, thanks for Eric's response and suggestion.

V3:
* From David Laight's feedback:
	Decide to drop RX BD entry number change for the SW TSO patch set.
	I will generate one separate patch to increase RX BDs entry for interrupt coalescing feature which
	will be supported in my later patch set.

V4:
* From David Laight's feedback:
	Remove the conditional in .fec_enet_get_bd_index().

V5:
* Patch #4 update:
  From David Laight's feedback:
	"expect fec_enet_get_free_txdesc_num() to return one less than it does currently."
	Change the function:
	Return space available, 0..size-1.  it always leave one free entry. Which is same as linux circ_buf.


Thanks for Eric and ezequiel's help and idea.


Fugang Duan (6):
  net: fec: Factorize the .xmit transmit function
  net: fec: Enable IP header hardware checksum
  net: fec: Factorize feature setting
  net: fec: Increase buffer descriptor entry number
  net: fec: Add Scatter/gather support
  net: fec: Add software TSO support

 drivers/net/ethernet/freescale/fec.h      |   12 +-
 drivers/net/ethernet/freescale/fec_main.c |  548 ++++++++++++++++++++++++-----
 2 files changed, 460 insertions(+), 100 deletions(-)

-- 
1.7.8

[PATCH v5 1/6] net: fec: Factorize the .xmit transmit function

[Fugang Duan]

Make the code more readable and easy to support other features like
SG, TSO, moving the common transmit function to one api.

And the patch also factorize the getting BD index to it own function.

CC: David Laight <David.Laight@ACULAB.COM>
Signed-off-by: Fugang Duan <B38611@freescale.com>
---
 drivers/net/ethernet/freescale/fec.h      |    1 +
 drivers/net/ethernet/freescale/fec_main.c |   81 ++++++++++++++++------------
 2 files changed, 47 insertions(+), 35 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec.h b/drivers/net/ethernet/freescale/fec.h
index 3b8d6d1..db967a0 100644
--- a/drivers/net/ethernet/freescale/fec.h
+++ b/drivers/net/ethernet/freescale/fec.h
@@ -296,6 +296,7 @@ struct fec_enet_private {
 	/* The ring entries to be free()ed */
 	struct bufdesc	*dirty_tx;
 
+	unsigned short bufdesc_size;
 	unsigned short tx_ring_size;
 	unsigned short rx_ring_size;
 
diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index 802be17..ee8e547 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -287,6 +287,12 @@ struct bufdesc *fec_enet_get_prevdesc(struct bufdesc *bdp, struct fec_enet_priva
 		return (new_bd < base) ? (new_bd + ring_size) : new_bd;
 }
 
+static int fec_enet_get_bd_index(struct bufdesc *base, struct bufdesc *bdp,
+				struct fec_enet_private *fep)
+{
+	return ((const char *)bdp - (const char *)base) / fep->bufdesc_size;
+}
+
 static void *swap_buffer(void *bufaddr, int len)
 {
 	int i;
@@ -313,8 +319,7 @@ fec_enet_clear_csum(struct sk_buff *skb, struct net_device *ndev)
 	return 0;
 }
 
-static netdev_tx_t
-fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
+static int txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
 {
 	struct fec_enet_private *fep = netdev_priv(ndev);
 	const struct platform_device_id *id_entry =
@@ -329,14 +334,6 @@ fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 
 	status = bdp->cbd_sc;
 
-	if (status & BD_ENET_TX_READY) {
-		/* Ooops.  All transmit buffers are full.  Bail out.
-		 * This should not happen, since ndev->tbusy should be set.
-		 */
-		netdev_err(ndev, "tx queue full!\n");
-		return NETDEV_TX_BUSY;
-	}
-
 	/* Protocol checksum off-load for TCP and UDP. */
 	if (fec_enet_clear_csum(skb, ndev)) {
 		dev_kfree_skb_any(skb);
@@ -350,16 +347,7 @@ fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 	bufaddr = skb->data;
 	bdp->cbd_datlen = skb->len;
 
-	/*
-	 * On some FEC implementations data must be aligned on
-	 * 4-byte boundaries. Use bounce buffers to copy data
-	 * and get it aligned. Ugh.
-	 */
-	if (fep->bufdesc_ex)
-		index = (struct bufdesc_ex *)bdp -
-			(struct bufdesc_ex *)fep->tx_bd_base;
-	else
-		index = bdp - fep->tx_bd_base;
+	index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
 
 	if (((unsigned long) bufaddr) & FEC_ALIGNMENT) {
 		memcpy(fep->tx_bounce[index], skb->data, skb->len);
@@ -433,15 +421,43 @@ fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 
 	fep->cur_tx = bdp;
 
-	if (fep->cur_tx == fep->dirty_tx)
-		netif_stop_queue(ndev);
-
 	/* Trigger transmission start */
 	writel(0, fep->hwp + FEC_X_DES_ACTIVE);
 
 	return NETDEV_TX_OK;
 }
 
+static netdev_tx_t
+fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
+{
+	struct fec_enet_private *fep = netdev_priv(ndev);
+	struct bufdesc *bdp;
+	unsigned short	status;
+	int ret;
+
+	/* Fill in a Tx ring entry */
+	bdp = fep->cur_tx;
+
+	status = bdp->cbd_sc;
+
+	if (status & BD_ENET_TX_READY) {
+		/* Ooops.  All transmit buffers are full.  Bail out.
+		 * This should not happen, since ndev->tbusy should be set.
+		 */
+		netdev_err(ndev, "tx queue full!\n");
+		return NETDEV_TX_BUSY;
+	}
+
+	ret = txq_submit_skb(skb, ndev);
+	if (ret == -EBUSY)
+		return NETDEV_TX_BUSY;
+
+	if (fep->cur_tx == fep->dirty_tx)
+		netif_stop_queue(ndev);
+
+	return NETDEV_TX_OK;
+}
+
 /* Init RX & TX buffer descriptors
  */
 static void fec_enet_bd_init(struct net_device *dev)
@@ -770,11 +786,7 @@ fec_enet_tx(struct net_device *ndev)
 		if (bdp == fep->cur_tx)
 			break;
 
-		if (fep->bufdesc_ex)
-			index = (struct bufdesc_ex *)bdp -
-				(struct bufdesc_ex *)fep->tx_bd_base;
-		else
-			index = bdp - fep->tx_bd_base;
+		index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
 
 		skb = fep->tx_skbuff[index];
 		dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr, skb->len,
@@ -921,11 +933,7 @@ fec_enet_rx(struct net_device *ndev, int budget)
 		pkt_len = bdp->cbd_datlen;
 		ndev->stats.rx_bytes += pkt_len;
 
-		if (fep->bufdesc_ex)
-			index = (struct bufdesc_ex *)bdp -
-				(struct bufdesc_ex *)fep->rx_bd_base;
-		else
-			index = bdp - fep->rx_bd_base;
+		index = fec_enet_get_bd_index(fep->rx_bd_base, bdp, fep);
 		data = fep->rx_skbuff[index]->data;
 		dma_sync_single_for_cpu(&fep->pdev->dev, bdp->cbd_bufaddr,
 					FEC_ENET_RX_FRSIZE, DMA_FROM_DEVICE);
@@ -2061,11 +2069,14 @@ static int fec_enet_init(struct net_device *ndev)
 
 	/* Set receive and transmit descriptor base. */
 	fep->rx_bd_base = cbd_base;
-	if (fep->bufdesc_ex)
+	if (fep->bufdesc_ex) {
 		fep->tx_bd_base = (struct bufdesc *)
 			(((struct bufdesc_ex *)cbd_base) + fep->rx_ring_size);
-	else
+		fep->bufdesc_size = sizeof(struct bufdesc_ex);
+	} else {
 		fep->tx_bd_base = cbd_base + fep->rx_ring_size;
+		fep->bufdesc_size = sizeof(struct bufdesc);
+	}
 
 	/* The FEC Ethernet specific entries in the device structure */
 	ndev->watchdog_timeo = TX_TIMEOUT;
-- 
1.7.8

[PATCH v5 2/6] net: fec: Enable IP header hardware checksum

[Fugang Duan]

IP header checksum is calcalated by network layer in default.
To support software TSO, it is better to use HW calculate the
IP header checksum.

FEC hw checksum feature request the checksum field in frame
is zero, otherwise the calculative CRC is not correct.

For segmentated TCP packet, HW calculate the IP header checksum again,
it doesn't bring any impact. For SW TSO, HW calculated checksum bring
better performance.

Signed-off-by: Fugang Duan <B38611@freescale.com>
---
 drivers/net/ethernet/freescale/fec_main.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index ee8e547..55b55d2 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -314,6 +314,7 @@ fec_enet_clear_csum(struct sk_buff *skb, struct net_device *ndev)
 	if (unlikely(skb_cow_head(skb, 0)))
 		return -1;
 
+	ip_hdr(skb)->check = 0;
 	*(__sum16 *)(skb->head + skb->csum_start + skb->csum_offset) = 0;
 
 	return 0;
@@ -395,7 +396,7 @@ static int txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
 			 * are done by the kernel
 			 */
 			if (skb->ip_summed == CHECKSUM_PARTIAL)
-				ebdp->cbd_esc |= BD_ENET_TX_PINS;
+				ebdp->cbd_esc |= BD_ENET_TX_PINS | BD_ENET_TX_IINS;
 		}
 	}
 
-- 
1.7.8

[PATCH v5 3/6] net: fec: Factorize feature setting

[Fugang Duan]

In order to enhance the code readable, let's factorize the
feature list.

Signed-off-by: Fugang Duan <B38611@freescale.com>
---
 drivers/net/ethernet/freescale/fec_main.c |    8 +++-----
 1 files changed, 3 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index 55b55d2..f943cd1 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -2087,21 +2087,19 @@ static int fec_enet_init(struct net_device *ndev)
 	writel(FEC_RX_DISABLED_IMASK, fep->hwp + FEC_IMASK);
 	netif_napi_add(ndev, &fep->napi, fec_enet_rx_napi, NAPI_POLL_WEIGHT);
 
-	if (id_entry->driver_data & FEC_QUIRK_HAS_VLAN) {
+	if (id_entry->driver_data & FEC_QUIRK_HAS_VLAN)
 		/* enable hw VLAN support */
 		ndev->features |= NETIF_F_HW_VLAN_CTAG_RX;
-		ndev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
-	}
 
 	if (id_entry->driver_data & FEC_QUIRK_HAS_CSUM) {
 		/* enable hw accelerator */
 		ndev->features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM
 				| NETIF_F_RXCSUM);
-		ndev->hw_features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM
-				| NETIF_F_RXCSUM);
 		fep->csum_flags |= FLAG_RX_CSUM_ENABLED;
 	}
 
+	ndev->hw_features = ndev->features;
+
 	fec_restart(ndev, 0);
 
 	return 0;
-- 
1.7.8

[PATCH v5 4/6] net: fec: Increase buffer descriptor entry number

[Fugang Duan]

In order to support SG, software TSO, let's increase BD entry number.

CC: Ezequiel Garcia <ezequiel.garcia@free-electrons.com>
CC: Eric Dumazet <eric.dumazet@gmail.com>
CC: David Laight <David.Laight@ACULAB.COM>
Signed-off-by: Fugang Duan <B38611@freescale.com>
---
 drivers/net/ethernet/freescale/fec.h      |    4 ++--
 drivers/net/ethernet/freescale/fec_main.c |   29 +++++++++++++++--------------
 2 files changed, 17 insertions(+), 16 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec.h b/drivers/net/ethernet/freescale/fec.h
index db967a0..5ffd323 100644
--- a/drivers/net/ethernet/freescale/fec.h
+++ b/drivers/net/ethernet/freescale/fec.h
@@ -246,8 +246,8 @@ struct bufdesc_ex {
 #define RX_RING_SIZE		(FEC_ENET_RX_FRPPG * FEC_ENET_RX_PAGES)
 #define FEC_ENET_TX_FRSIZE	2048
 #define FEC_ENET_TX_FRPPG	(PAGE_SIZE / FEC_ENET_TX_FRSIZE)
-#define TX_RING_SIZE		16	/* Must be power of two */
-#define TX_RING_MOD_MASK	15	/*   for this to work */
+#define TX_RING_SIZE		512	/* Must be power of two */
+#define TX_RING_MOD_MASK	511	/*   for this to work */
 
 #define BD_ENET_RX_INT          0x00800000
 #define BD_ENET_RX_PTP          ((ushort)0x0400)
diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index f943cd1..b27a729 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -173,10 +173,6 @@ MODULE_PARM_DESC(macaddr, "FEC Ethernet MAC address");
 #endif
 #endif /* CONFIG_M5272 */
 
-#if (((RX_RING_SIZE + TX_RING_SIZE) * 32) > PAGE_SIZE)
-#error "FEC: descriptor ring size constants too large"
-#endif
-
 /* Interrupt events/masks. */
 #define FEC_ENET_HBERR	((uint)0x80000000)	/* Heartbeat error */
 #define FEC_ENET_BABR	((uint)0x40000000)	/* Babbling receiver */
@@ -2048,9 +2044,21 @@ static int fec_enet_init(struct net_device *ndev)
 	const struct platform_device_id *id_entry =
 				platform_get_device_id(fep->pdev);
 	struct bufdesc *cbd_base;
+	int bd_size;
+
+	/* init the tx & rx ring size */
+	fep->tx_ring_size = TX_RING_SIZE;
+	fep->rx_ring_size = RX_RING_SIZE;
+
+	if (fep->bufdesc_ex)
+		fep->bufdesc_size = sizeof(struct bufdesc_ex);
+	else
+		fep->bufdesc_size = sizeof(struct bufdesc);
+	bd_size = (fep->tx_ring_size + fep->rx_ring_size) *
+			fep->bufdesc_size;
 
 	/* Allocate memory for buffer descriptors. */
-	cbd_base = dma_alloc_coherent(NULL, PAGE_SIZE, &fep->bd_dma,
+	cbd_base = dma_alloc_coherent(NULL, bd_size, &fep->bd_dma,
 				      GFP_KERNEL);
 	if (!cbd_base)
 		return -ENOMEM;
@@ -2064,20 +2072,13 @@ static int fec_enet_init(struct net_device *ndev)
 	/* make sure MAC we just acquired is programmed into the hw */
 	fec_set_mac_address(ndev, NULL);
 
-	/* init the tx & rx ring size */
-	fep->tx_ring_size = TX_RING_SIZE;
-	fep->rx_ring_size = RX_RING_SIZE;
-
 	/* Set receive and transmit descriptor base. */
 	fep->rx_bd_base = cbd_base;
-	if (fep->bufdesc_ex) {
+	if (fep->bufdesc_ex)
 		fep->tx_bd_base = (struct bufdesc *)
 			(((struct bufdesc_ex *)cbd_base) + fep->rx_ring_size);
-		fep->bufdesc_size = sizeof(struct bufdesc_ex);
-	} else {
+	else
 		fep->tx_bd_base = cbd_base + fep->rx_ring_size;
-		fep->bufdesc_size = sizeof(struct bufdesc);
-	}
 
 	/* The FEC Ethernet specific entries in the device structure */
 	ndev->watchdog_timeo = TX_TIMEOUT;
-- 
1.7.8

[PATCH v5 5/6] net: fec: Add Scatter/gather support

[Fugang Duan]

Add Scatter/gather support for FEC.
This feature allows to improve outbound throughput performance.

Tested on imx6dl sabresd board:
Running iperf tests shows a 55.4% improvement.

$ ethtool -K eth0 sg off
$ iperf -c 10.192.242.167 -t 3 &
[  3] local 10.192.242.108 port 52618 connected with 10.192.242.167 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0- 3.0 sec  99.5 MBytes   278 Mbits/sec

$ ethtool -K eth0 sg on
$ iperf -c 10.192.242.167 -t 3 &
[  3] local 10.192.242.108 port 52617 connected with 10.192.242.167 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0- 3.0 sec   154 MBytes   432 Mbits/sec

CC: Li Frank <B20596@freescale.com>
Signed-off-by: Fugang Duan <B38611@freescale.com>
---
 drivers/net/ethernet/freescale/fec.h      |    2 +-
 drivers/net/ethernet/freescale/fec_main.c |  238 +++++++++++++++++++++--------
 2 files changed, 178 insertions(+), 62 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec.h b/drivers/net/ethernet/freescale/fec.h
index 5ffd323..e7ce14d 100644
--- a/drivers/net/ethernet/freescale/fec.h
+++ b/drivers/net/ethernet/freescale/fec.h
@@ -221,7 +221,7 @@ struct bufdesc_ex {
 #define BD_ENET_TX_RCMASK       ((ushort)0x003c)
 #define BD_ENET_TX_UN           ((ushort)0x0002)
 #define BD_ENET_TX_CSL          ((ushort)0x0001)
-#define BD_ENET_TX_STATS        ((ushort)0x03ff)        /* All status bits */
+#define BD_ENET_TX_STATS        ((ushort)0x0fff)        /* All status bits */
 
 /*enhanced buffer descriptor control/status used by Ethernet transmit*/
 #define BD_ENET_TX_INT          0x40000000
diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index b27a729..bea00a8 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -289,6 +289,16 @@ static int fec_enet_get_bd_index(struct bufdesc *base, struct bufdesc *bdp,
 	return ((const char *)bdp - (const char *)base) / fep->bufdesc_size;
 }
 
+static int fec_enet_get_free_txdesc_num(struct fec_enet_private *fep)
+{
+	int entries;
+
+	entries = ((const char *)fep->dirty_tx -
+			(const char *)fep->cur_tx) / fep->bufdesc_size - 1;
+
+	return entries > 0 ? entries : entries + fep->tx_ring_size;
+}
+
 static void *swap_buffer(void *bufaddr, int len)
 {
 	int i;
@@ -316,20 +326,119 @@ fec_enet_clear_csum(struct sk_buff *skb, struct net_device *ndev)
 	return 0;
 }
 
-static int txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
+static void
+fec_enet_submit_work(struct bufdesc *bdp, struct fec_enet_private *fep)
+{
+	const struct platform_device_id *id_entry =
+				platform_get_device_id(fep->pdev);
+	struct bufdesc *bdp_pre;
+
+	bdp_pre = fec_enet_get_prevdesc(bdp, fep);
+	if ((id_entry->driver_data & FEC_QUIRK_ERR006358) &&
+	    !(bdp_pre->cbd_sc & BD_ENET_TX_READY)) {
+		fep->delay_work.trig_tx = true;
+		schedule_delayed_work(&(fep->delay_work.delay_work),
+					msecs_to_jiffies(1));
+	}
+}
+
+static int
+fec_enet_txq_submit_frag_skb(struct sk_buff *skb, struct net_device *ndev)
 {
 	struct fec_enet_private *fep = netdev_priv(ndev);
 	const struct platform_device_id *id_entry =
 				platform_get_device_id(fep->pdev);
-	struct bufdesc *bdp, *bdp_pre;
-	void *bufaddr;
-	unsigned short	status;
+	struct bufdesc *bdp = fep->cur_tx;
+	struct bufdesc_ex *ebdp;
+	int nr_frags = skb_shinfo(skb)->nr_frags;
+	int frag, frag_len;
+	unsigned short status;
+	unsigned int estatus = 0;
+	skb_frag_t *this_frag;
 	unsigned int index;
+	void *bufaddr;
+	int i;
 
-	/* Fill in a Tx ring entry */
+	for (frag = 0; frag < nr_frags; frag++) {
+		this_frag = &skb_shinfo(skb)->frags[frag];
+		bdp = fec_enet_get_nextdesc(bdp, fep);
+		ebdp = (struct bufdesc_ex *)bdp;
+
+		status = bdp->cbd_sc;
+		status &= ~BD_ENET_TX_STATS;
+		status |= (BD_ENET_TX_TC | BD_ENET_TX_READY);
+		frag_len = skb_shinfo(skb)->frags[frag].size;
+
+		/* Handle the last BD specially */
+		if (frag == nr_frags - 1) {
+			status |= (BD_ENET_TX_INTR | BD_ENET_TX_LAST);
+			if (fep->bufdesc_ex) {
+				estatus |= BD_ENET_TX_INT;
+				if (unlikely(skb_shinfo(skb)->tx_flags &
+					SKBTX_HW_TSTAMP && fep->hwts_tx_en))
+					estatus |= BD_ENET_TX_TS;
+			}
+		}
+
+		if (fep->bufdesc_ex) {
+			if (skb->ip_summed == CHECKSUM_PARTIAL)
+				estatus |= BD_ENET_TX_PINS | BD_ENET_TX_IINS;
+			ebdp->cbd_bdu = 0;
+			ebdp->cbd_esc = estatus;
+		}
+
+		bufaddr = page_address(this_frag->page.p) + this_frag->page_offset;
+
+		index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
+		if (((unsigned long) bufaddr) & FEC_ALIGNMENT ||
+			id_entry->driver_data & FEC_QUIRK_SWAP_FRAME) {
+			memcpy(fep->tx_bounce[index], bufaddr, frag_len);
+			bufaddr = fep->tx_bounce[index];
+
+			if (id_entry->driver_data & FEC_QUIRK_SWAP_FRAME)
+				swap_buffer(bufaddr, frag_len);
+		}
+
+		bdp->cbd_bufaddr = dma_map_single(&fep->pdev->dev, bufaddr,
+						frag_len, DMA_TO_DEVICE);
+		if (dma_mapping_error(&fep->pdev->dev, bdp->cbd_bufaddr)) {
+			dev_kfree_skb_any(skb);
+			if (net_ratelimit())
+				netdev_err(ndev, "Tx DMA memory map failed\n");
+			goto dma_mapping_error;
+		}
+
+		bdp->cbd_datlen = frag_len;
+		bdp->cbd_sc = status;
+	}
+
+	fep->cur_tx = bdp;
+
+	return 0;
+
+dma_mapping_error:
 	bdp = fep->cur_tx;
+	for (i = 0; i < frag; i++) {
+		bdp = fec_enet_get_nextdesc(bdp, fep);
+		dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr,
+				bdp->cbd_datlen, DMA_TO_DEVICE);
+	}
+	return NETDEV_TX_OK;
+}
 
-	status = bdp->cbd_sc;
+static int fec_enet_txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
+{
+	struct fec_enet_private *fep = netdev_priv(ndev);
+	const struct platform_device_id *id_entry =
+				platform_get_device_id(fep->pdev);
+	int nr_frags = skb_shinfo(skb)->nr_frags;
+	struct bufdesc *bdp, *last_bdp;
+	void *bufaddr;
+	unsigned short status;
+	unsigned short buflen;
+	unsigned int estatus = 0;
+	unsigned int index;
+	int ret;
 
 	/* Protocol checksum off-load for TCP and UDP. */
 	if (fec_enet_clear_csum(skb, ndev)) {
@@ -337,82 +446,83 @@ static int txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
 		return NETDEV_TX_OK;
 	}
 
-	/* Clear all of the status flags */
+	/* Fill in a Tx ring entry */
+	bdp = fep->cur_tx;
+	status = bdp->cbd_sc;
 	status &= ~BD_ENET_TX_STATS;
 
 	/* Set buffer length and buffer pointer */
 	bufaddr = skb->data;
-	bdp->cbd_datlen = skb->len;
+	buflen = skb_headlen(skb);
 
 	index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
-
-	if (((unsigned long) bufaddr) & FEC_ALIGNMENT) {
-		memcpy(fep->tx_bounce[index], skb->data, skb->len);
+	if (((unsigned long) bufaddr) & FEC_ALIGNMENT ||
+		id_entry->driver_data & FEC_QUIRK_SWAP_FRAME) {
+		memcpy(fep->tx_bounce[index], skb->data, buflen);
 		bufaddr = fep->tx_bounce[index];
-	}
 
-	/*
-	 * Some design made an incorrect assumption on endian mode of
-	 * the system that it's running on. As the result, driver has to
-	 * swap every frame going to and coming from the controller.
-	 */
-	if (id_entry->driver_data & FEC_QUIRK_SWAP_FRAME)
-		swap_buffer(bufaddr, skb->len);
-
-	/* Save skb pointer */
-	fep->tx_skbuff[index] = skb;
+		if (id_entry->driver_data & FEC_QUIRK_SWAP_FRAME)
+			swap_buffer(bufaddr, buflen);
+	}
 
 	/* Push the data cache so the CPM does not get stale memory
 	 * data.
 	 */
 	bdp->cbd_bufaddr = dma_map_single(&fep->pdev->dev, bufaddr,
-			skb->len, DMA_TO_DEVICE);
+					buflen, DMA_TO_DEVICE);
 	if (dma_mapping_error(&fep->pdev->dev, bdp->cbd_bufaddr)) {
-		bdp->cbd_bufaddr = 0;
-		fep->tx_skbuff[index] = NULL;
 		dev_kfree_skb_any(skb);
 		if (net_ratelimit())
 			netdev_err(ndev, "Tx DMA memory map failed\n");
 		return NETDEV_TX_OK;
 	}
 
+	if (nr_frags) {
+		ret = fec_enet_txq_submit_frag_skb(skb, ndev);
+		if (ret)
+			return ret;
+	} else {
+		status |= (BD_ENET_TX_INTR | BD_ENET_TX_LAST);
+		if (fep->bufdesc_ex) {
+			estatus = BD_ENET_TX_INT;
+			if (unlikely(skb_shinfo(skb)->tx_flags &
+				SKBTX_HW_TSTAMP && fep->hwts_tx_en))
+				estatus |= BD_ENET_TX_TS;
+		}
+	}
+
 	if (fep->bufdesc_ex) {
 
 		struct bufdesc_ex *ebdp = (struct bufdesc_ex *)bdp;
-		ebdp->cbd_bdu = 0;
+
 		if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP &&
-			fep->hwts_tx_en)) {
-			ebdp->cbd_esc = (BD_ENET_TX_TS | BD_ENET_TX_INT);
+			fep->hwts_tx_en))
 			skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
-		} else {
-			ebdp->cbd_esc = BD_ENET_TX_INT;
 
-			/* Enable protocol checksum flags
-			 * We do not bother with the IP Checksum bits as they
-			 * are done by the kernel
-			 */
-			if (skb->ip_summed == CHECKSUM_PARTIAL)
-				ebdp->cbd_esc |= BD_ENET_TX_PINS | BD_ENET_TX_IINS;
-		}
+		if (skb->ip_summed == CHECKSUM_PARTIAL)
+			estatus |= BD_ENET_TX_PINS | BD_ENET_TX_IINS;
+
+		ebdp->cbd_bdu = 0;
+		ebdp->cbd_esc = estatus;
 	}
 
+	last_bdp = fep->cur_tx;
+	index = fec_enet_get_bd_index(fep->tx_bd_base, last_bdp, fep);
+	/* Save skb pointer */
+	fep->tx_skbuff[index] = skb;
+
+	bdp->cbd_datlen = buflen;
+
 	/* Send it on its way.  Tell FEC it's ready, interrupt when done,
 	 * it's the last BD of the frame, and to put the CRC on the end.
 	 */
-	status |= (BD_ENET_TX_READY | BD_ENET_TX_INTR
-			| BD_ENET_TX_LAST | BD_ENET_TX_TC);
+	status |= (BD_ENET_TX_READY | BD_ENET_TX_TC);
 	bdp->cbd_sc = status;
 
-	bdp_pre = fec_enet_get_prevdesc(bdp, fep);
-	if ((id_entry->driver_data & FEC_QUIRK_ERR006358) &&
-	    !(bdp_pre->cbd_sc & BD_ENET_TX_READY)) {
-		fep->delay_work.trig_tx = true;
-		schedule_delayed_work(&(fep->delay_work.delay_work),
-					msecs_to_jiffies(1));
-	}
+	fec_enet_submit_work(bdp, fep);
 
 	/* If this was the last BD in the ring, start at the beginning again. */
-	bdp = fec_enet_get_nextdesc(bdp, fep);
+	bdp = fec_enet_get_nextdesc(last_bdp, fep);
 
 	skb_tx_timestamp(skb);
 
@@ -421,7 +531,7 @@ static int txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
 	/* Trigger transmission start */
 	writel(0, fep->hwp + FEC_X_DES_ACTIVE);
 
-	return NETDEV_TX_OK;
+	return 0;
 }
 
 static netdev_tx_t
@@ -430,6 +540,7 @@ fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 	struct fec_enet_private *fep = netdev_priv(ndev);
 	struct bufdesc *bdp;
 	unsigned short	status;
+	int entries_free;
 	int ret;
 
 	/* Fill in a Tx ring entry */
@@ -441,15 +552,17 @@ fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 		/* Ooops.  All transmit buffers are full.  Bail out.
 		 * This should not happen, since ndev->tbusy should be set.
 		 */
-		netdev_err(ndev, "tx queue full!\n");
+		if (net_ratelimit())
+			netdev_err(ndev, "tx queue full!\n");
 		return NETDEV_TX_BUSY;
 	}
 
-	ret = txq_submit_skb(skb, ndev);
-	if (ret == -EBUSY)
-		return NETDEV_TX_BUSY;
+	ret = fec_enet_txq_submit_skb(skb, ndev);
+	if (ret)
+		return ret;
 
-	if (fep->cur_tx == fep->dirty_tx)
+	entries_free = fec_enet_get_free_txdesc_num(fep);
+	if (entries_free < MAX_SKB_FRAGS + 1)
 		netif_stop_queue(ndev);
 
 	return NETDEV_TX_OK;
@@ -770,6 +883,7 @@ fec_enet_tx(struct net_device *ndev)
 	unsigned short status;
 	struct	sk_buff	*skb;
 	int	index = 0;
+	int	entries;
 
 	fep = netdev_priv(ndev);
 	bdp = fep->dirty_tx;
@@ -786,9 +900,13 @@ fec_enet_tx(struct net_device *ndev)
 		index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
 
 		skb = fep->tx_skbuff[index];
-		dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr, skb->len,
+		dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr, bdp->cbd_datlen,
 				DMA_TO_DEVICE);
 		bdp->cbd_bufaddr = 0;
+		if (!skb) {
+			bdp = fec_enet_get_nextdesc(bdp, fep);
+			continue;
+		}
 
 		/* Check for errors. */
 		if (status & (BD_ENET_TX_HB | BD_ENET_TX_LC |
@@ -807,7 +925,7 @@ fec_enet_tx(struct net_device *ndev)
 				ndev->stats.tx_carrier_errors++;
 		} else {
 			ndev->stats.tx_packets++;
-			ndev->stats.tx_bytes += bdp->cbd_datlen;
+			ndev->stats.tx_bytes += skb->len;
 		}
 
 		if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_IN_PROGRESS) &&
@@ -844,15 +962,13 @@ fec_enet_tx(struct net_device *ndev)
 
 		/* Since we have freed up a buffer, the ring is no longer full
 		 */
-		if (fep->dirty_tx != fep->cur_tx) {
-			if (netif_queue_stopped(ndev))
-				netif_wake_queue(ndev);
-		}
+		entries = fec_enet_get_free_txdesc_num(fep);
+		if (entries >= MAX_SKB_FRAGS + 1 && netif_queue_stopped(ndev))
+			netif_wake_queue(ndev);
 	}
 	return;
 }
 
-
 /* During a receive, the cur_rx points to the current incoming buffer.
  * When we update through the ring, if the next incoming buffer has
  * not been given to the system, we just set the empty indicator,
@@ -2095,7 +2211,7 @@ static int fec_enet_init(struct net_device *ndev)
 	if (id_entry->driver_data & FEC_QUIRK_HAS_CSUM) {
 		/* enable hw accelerator */
 		ndev->features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM
-				| NETIF_F_RXCSUM);
+				| NETIF_F_RXCSUM | NETIF_F_SG);
 		fep->csum_flags |= FLAG_RX_CSUM_ENABLED;
 	}
 
-- 
1.7.8

[PATCH v5 6/6] net: fec: Add software TSO support

[Fugang Duan]

Add software TSO support for FEC.
This feature allows to improve outbound throughput performance.

Tested on imx6dl sabresd board, running iperf tcp tests shows:
- 16.2% improvement comparing with FEC SG patch
- 82% improvement comparing with NO SG & TSO patch

$ ethtool -K eth0 tso on
$ iperf -c 10.192.242.167 -t 3 &
[  3] local 10.192.242.108 port 35388 connected with 10.192.242.167 port 5001
[ ID] Interval       Transfer     Bandwidth
[  3]  0.0- 3.0 sec   181 MBytes   506 Mbits/sec

During the testing, CPU loading is 30%.
Since imx6dl FEC Bandwidth is limited to SOC system bus bandwidth, the
performance with SW TSO is a milestone.

CC: Ezequiel Garcia <ezequiel.garcia@free-electrons.com>
CC: Eric Dumazet <eric.dumazet@gmail.com>
CC: David Laight <David.Laight@ACULAB.COM>
CC: Li Frank <B20596@freescale.com>
Signed-off-by: Fugang Duan <B38611@freescale.com>
---
 drivers/net/ethernet/freescale/fec.h      |    6 +
 drivers/net/ethernet/freescale/fec_main.c |  255 ++++++++++++++++++++++++++---
 2 files changed, 238 insertions(+), 23 deletions(-)

diff --git a/drivers/net/ethernet/freescale/fec.h b/drivers/net/ethernet/freescale/fec.h
index e7ce14d..671d080 100644
--- a/drivers/net/ethernet/freescale/fec.h
+++ b/drivers/net/ethernet/freescale/fec.h
@@ -299,6 +299,12 @@ struct fec_enet_private {
 	unsigned short bufdesc_size;
 	unsigned short tx_ring_size;
 	unsigned short rx_ring_size;
+	unsigned short tx_stop_threshold;
+	unsigned short tx_wake_threshold;
+
+	/* Software TSO */
+	char *tso_hdrs;
+	dma_addr_t tso_hdrs_dma;
 
 	struct	platform_device *pdev;
 
diff --git a/drivers/net/ethernet/freescale/fec_main.c b/drivers/net/ethernet/freescale/fec_main.c
index 6bf8d47..d76099d 100644
--- a/drivers/net/ethernet/freescale/fec_main.c
+++ b/drivers/net/ethernet/freescale/fec_main.c
@@ -36,6 +36,7 @@
 #include <linux/in.h>
 #include <linux/ip.h>
 #include <net/ip.h>
+#include <net/tso.h>
 #include <linux/tcp.h>
 #include <linux/udp.h>
 #include <linux/icmp.h>
@@ -228,6 +229,15 @@ MODULE_PARM_DESC(macaddr, "FEC Ethernet MAC address");
 #define FEC_PAUSE_FLAG_AUTONEG	0x1
 #define FEC_PAUSE_FLAG_ENABLE	0x2
 
+#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)
+
+#define IS_TSO_HEADER(txq, addr) \
+	((addr >= txq->tso_hdrs_dma) && \
+	(addr < txq->tso_hdrs_dma + txq->tx_ring_size * TSO_HEADER_SIZE))
+
 static int mii_cnt;
 
 static inline
@@ -439,8 +449,17 @@ static int fec_enet_txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
 	unsigned short buflen;
 	unsigned int estatus = 0;
 	unsigned int index;
+	int entries_free;
 	int ret;
 
+	entries_free = fec_enet_get_free_txdesc_num(fep);
+	if (entries_free < MAX_SKB_FRAGS + 1) {
+		dev_kfree_skb_any(skb);
+		if (net_ratelimit())
+			netdev_err(ndev, "NOT enough BD for SG!\n");
+		return NETDEV_TX_OK;
+	}
+
 	/* Protocol checksum off-load for TCP and UDP. */
 	if (fec_enet_clear_csum(skb, ndev)) {
 		dev_kfree_skb_any(skb);
@@ -535,35 +554,210 @@ static int fec_enet_txq_submit_skb(struct sk_buff *skb, struct net_device *ndev)
 	return 0;
 }
 
-static netdev_tx_t
-fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
+static int
+fec_enet_txq_put_data_tso(struct sk_buff *skb, struct net_device *ndev,
+			struct bufdesc *bdp, int index, char *data,
+			int size, bool last_tcp, bool is_last)
 {
 	struct fec_enet_private *fep = netdev_priv(ndev);
-	struct bufdesc *bdp;
-	unsigned short	status;
-	int entries_free;
-	int ret;
-
-	/* Fill in a Tx ring entry */
-	bdp = fep->cur_tx;
+	const struct platform_device_id *id_entry =
+				platform_get_device_id(fep->pdev);
+	struct bufdesc_ex *ebdp = (struct bufdesc_ex *)bdp;
+	unsigned short status;
+	unsigned int estatus = 0;
 
 	status = bdp->cbd_sc;
+	status &= ~BD_ENET_TX_STATS;
 
-	if (status & BD_ENET_TX_READY) {
-		/* Ooops.  All transmit buffers are full.  Bail out.
-		 * This should not happen, since ndev->tbusy should be set.
-		 */
+	status |= (BD_ENET_TX_TC | BD_ENET_TX_READY);
+	bdp->cbd_datlen = size;
+
+	if (((unsigned long) data) & FEC_ALIGNMENT ||
+		id_entry->driver_data & FEC_QUIRK_SWAP_FRAME) {
+		memcpy(fep->tx_bounce[index], data, size);
+		data = fep->tx_bounce[index];
+
+		if (id_entry->driver_data & FEC_QUIRK_SWAP_FRAME)
+			swap_buffer(data, size);
+	}
+
+	bdp->cbd_bufaddr = dma_map_single(&fep->pdev->dev, data,
+					size, DMA_TO_DEVICE);
+	if (dma_mapping_error(&fep->pdev->dev, bdp->cbd_bufaddr)) {
+		dev_kfree_skb_any(skb);
 		if (net_ratelimit())
-			netdev_err(ndev, "tx queue full!\n");
+			netdev_err(ndev, "Tx DMA memory map failed\n");
 		return NETDEV_TX_BUSY;
 	}
 
-	ret = fec_enet_txq_submit_skb(skb, ndev);
+	if (fep->bufdesc_ex) {
+		if (skb->ip_summed == CHECKSUM_PARTIAL)
+			estatus |= BD_ENET_TX_PINS | BD_ENET_TX_IINS;
+		ebdp->cbd_bdu = 0;
+		ebdp->cbd_esc = estatus;
+	}
+
+	/* Handle the last BD specially */
+	if (last_tcp)
+		status |= (BD_ENET_TX_LAST | BD_ENET_TX_TC);
+	if (is_last) {
+		status |= BD_ENET_TX_INTR;
+		if (fep->bufdesc_ex)
+			ebdp->cbd_esc |= BD_ENET_TX_INT;
+	}
+
+	bdp->cbd_sc = status;
+
+	return 0;
+}
+
+static int
+fec_enet_txq_put_hdr_tso(struct sk_buff *skb, struct net_device *ndev,
+			struct bufdesc *bdp, int index)
+{
+	struct fec_enet_private *fep = netdev_priv(ndev);
+	const struct platform_device_id *id_entry =
+				platform_get_device_id(fep->pdev);
+	int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	struct bufdesc_ex *ebdp = (struct bufdesc_ex *)bdp;
+	void *bufaddr;
+	unsigned long dmabuf;
+	unsigned short status;
+	unsigned int estatus = 0;
+
+	status = bdp->cbd_sc;
+	status &= ~BD_ENET_TX_STATS;
+	status |= (BD_ENET_TX_TC | BD_ENET_TX_READY);
+
+	bufaddr = fep->tso_hdrs + index * TSO_HEADER_SIZE;
+	dmabuf = fep->tso_hdrs_dma + index * TSO_HEADER_SIZE;
+	if (((unsigned long) bufaddr) & FEC_ALIGNMENT ||
+		id_entry->driver_data & FEC_QUIRK_SWAP_FRAME) {
+		memcpy(fep->tx_bounce[index], skb->data, hdr_len);
+		bufaddr = fep->tx_bounce[index];
+
+		if (id_entry->driver_data & FEC_QUIRK_SWAP_FRAME)
+			swap_buffer(bufaddr, hdr_len);
+
+		dmabuf = dma_map_single(&fep->pdev->dev, bufaddr,
+					hdr_len, DMA_TO_DEVICE);
+		if (dma_mapping_error(&fep->pdev->dev, dmabuf)) {
+			dev_kfree_skb_any(skb);
+			if (net_ratelimit())
+				netdev_err(ndev, "Tx DMA memory map failed\n");
+			return NETDEV_TX_BUSY;
+		}
+	}
+
+	bdp->cbd_bufaddr = dmabuf;
+	bdp->cbd_datlen = hdr_len;
+
+	if (fep->bufdesc_ex) {
+		if (skb->ip_summed == CHECKSUM_PARTIAL)
+			estatus |= BD_ENET_TX_PINS | BD_ENET_TX_IINS;
+		ebdp->cbd_bdu = 0;
+		ebdp->cbd_esc = estatus;
+	}
+
+	bdp->cbd_sc = status;
+
+	return 0;
+}
+
+static int fec_enet_txq_submit_tso(struct sk_buff *skb, struct net_device *ndev)
+{
+	struct fec_enet_private *fep = netdev_priv(ndev);
+	int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	int total_len, data_left;
+	struct bufdesc *bdp = fep->cur_tx;
+	struct tso_t tso;
+	unsigned int index = 0;
+	int ret;
+
+	if (tso_count_descs(skb) >= fec_enet_get_free_txdesc_num(fep)) {
+		dev_kfree_skb_any(skb);
+		if (net_ratelimit())
+			netdev_err(ndev, "NOT enough BD for TSO!\n");
+		return NETDEV_TX_OK;
+	}
+
+	/* Protocol checksum off-load for TCP and UDP. */
+	if (fec_enet_clear_csum(skb, ndev)) {
+		dev_kfree_skb_any(skb);
+		return NETDEV_TX_OK;
+	}
+
+	/* Initialize the TSO handler, and prepare the first payload */
+	tso_start(skb, &tso);
+
+	total_len = skb->len - hdr_len;
+	while (total_len > 0) {
+		char *hdr;
+
+		index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
+		data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
+		total_len -= data_left;
+
+		/* prepare packet headers: MAC + IP + TCP */
+		hdr = fep->tso_hdrs + index * TSO_HEADER_SIZE;
+		tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0);
+		ret = fec_enet_txq_put_hdr_tso(skb, ndev, bdp, index);
+		if (ret)
+			goto err_release;
+
+		while (data_left > 0) {
+			int size;
+
+			size = min_t(int, tso.size, data_left);
+			bdp = fec_enet_get_nextdesc(bdp, fep);
+			index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
+			ret = fec_enet_txq_put_data_tso(skb, ndev, bdp, index, tso.data,
+							size, size == data_left,
+							total_len == 0);
+			if (ret)
+				goto err_release;
+
+			data_left -= size;
+			tso_build_data(skb, &tso, size);
+		}
+
+		bdp = fec_enet_get_nextdesc(bdp, fep);
+	}
+
+	/* Save skb pointer */
+	fep->tx_skbuff[index] = skb;
+
+	fec_enet_submit_work(bdp, fep);
+
+	skb_tx_timestamp(skb);
+	fep->cur_tx = bdp;
+
+	/* Trigger transmission start */
+	writel(0, fep->hwp + FEC_X_DES_ACTIVE);
+
+	return 0;
+
+err_release:
+	/* TODO: Release all used data descriptors for TSO */
+	return ret;
+}
+
+static netdev_tx_t
+fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
+{
+	struct fec_enet_private *fep = netdev_priv(ndev);
+	int entries_free;
+	int ret;
+
+	if (skb_is_gso(skb))
+		ret = fec_enet_txq_submit_tso(skb, ndev);
+	else
+		ret = fec_enet_txq_submit_skb(skb, ndev);
 	if (ret)
 		return ret;
 
 	entries_free = fec_enet_get_free_txdesc_num(fep);
-	if (entries_free < MAX_SKB_FRAGS + 1)
+	if (entries_free <= fep->tx_stop_threshold)
 		netif_stop_queue(ndev);
 
 	return NETDEV_TX_OK;
@@ -884,7 +1078,7 @@ fec_enet_tx(struct net_device *ndev)
 	unsigned short status;
 	struct	sk_buff	*skb;
 	int	index = 0;
-	int	entries;
+	int	entries_free;
 
 	fep = netdev_priv(ndev);
 	bdp = fep->dirty_tx;
@@ -901,8 +1095,9 @@ fec_enet_tx(struct net_device *ndev)
 		index = fec_enet_get_bd_index(fep->tx_bd_base, bdp, fep);
 
 		skb = fep->tx_skbuff[index];
-		dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr, bdp->cbd_datlen,
-				DMA_TO_DEVICE);
+		if (!IS_TSO_HEADER(fep, bdp->cbd_bufaddr))
+			dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr,
+					bdp->cbd_datlen, DMA_TO_DEVICE);
 		bdp->cbd_bufaddr = 0;
 		if (!skb) {
 			bdp = fec_enet_get_nextdesc(bdp, fep);
@@ -963,9 +1158,11 @@ fec_enet_tx(struct net_device *ndev)
 
 		/* Since we have freed up a buffer, the ring is no longer full
 		 */
-		entries = fec_enet_get_free_txdesc_num(fep);
-		if (entries >= MAX_SKB_FRAGS + 1 && netif_queue_stopped(ndev))
-			netif_wake_queue(ndev);
+		if (netif_queue_stopped(ndev)) {
+			entries_free = fec_enet_get_free_txdesc_num(fep);
+			if (entries_free >= fep->tx_wake_threshold)
+				netif_wake_queue(ndev);
+		}
 	}
 	return;
 }
@@ -2167,6 +2364,9 @@ static int fec_enet_init(struct net_device *ndev)
 	fep->tx_ring_size = TX_RING_SIZE;
 	fep->rx_ring_size = RX_RING_SIZE;
 
+	fep->tx_stop_threshold = FEC_MAX_SKB_DESCS;
+	fep->tx_wake_threshold = (fep->tx_ring_size - fep->tx_stop_threshold) / 2;
+
 	if (fep->bufdesc_ex)
 		fep->bufdesc_size = sizeof(struct bufdesc_ex);
 	else
@@ -2180,6 +2380,13 @@ static int fec_enet_init(struct net_device *ndev)
 	if (!cbd_base)
 		return -ENOMEM;
 
+	fep->tso_hdrs = dma_alloc_coherent(NULL, fep->tx_ring_size * TSO_HEADER_SIZE,
+						&fep->tso_hdrs_dma, GFP_KERNEL);
+	if (!fep->tso_hdrs) {
+		dma_free_coherent(NULL, bd_size, cbd_base, fep->bd_dma);
+		return -ENOMEM;
+	}
+
 	memset(cbd_base, 0, PAGE_SIZE);
 
 	fep->netdev = ndev;
@@ -2210,9 +2417,11 @@ static int fec_enet_init(struct net_device *ndev)
 		ndev->features |= NETIF_F_HW_VLAN_CTAG_RX;
 
 	if (id_entry->driver_data & FEC_QUIRK_HAS_CSUM) {
+		ndev->gso_max_segs = FEC_MAX_TSO_SEGS;
+
 		/* enable hw accelerator */
 		ndev->features |= (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM
-				| NETIF_F_RXCSUM | NETIF_F_SG);
+				| NETIF_F_RXCSUM | NETIF_F_SG | NETIF_F_TSO);
 		fep->csum_flags |= FLAG_RX_CSUM_ENABLED;
 	}
 
-- 
1.7.8

Re: [PATCH v5 0/6] net: fec: Enable Software TSO to improve the tx performance

[David Miller]

From: Fugang Duan <b38611@freescale.com>
Date: Thu, 12 Jun 2014 08:16:17 +0800

> Add SG and software TSO support for FEC.

Series applied, thanks.