[PATCH net-next v3 0/9] net: rswitch: Add jumbo frames support

[PATCH net-next v3 0/9] net: rswitch: Add jumbo frames support

[Yoshihiro Shimoda]

This patch series is based on the latest net-next.git / main branch.

Changes from v2:
https://lore.kernel.org/all/20231201054655.3731772-1-yoshihiro.shimoda.uh@renesas.com/
 - Based on the latest net-next.git / main branch.
 - Fix using a variable in the patch 8/9.
 - Add Reviewed-by tag in the patch 1/9.

Changes from v1:
https://lore.kernel.org/all/20231127115334.3670790-1-yoshihiro.shimoda.uh@renesas.com/
 - Based on the latest net-next.git / main branch.
 - Fix commit descriptions (s/near the future/the near future/).

Yoshihiro Shimoda (9):
  net: rswitch: Drop unused argument/return value
  net: rswitch: Use unsigned int for desc related array index
  net: rswitch: Use build_skb() for RX
  net: rswitch: Add unmap_addrs instead of dma address in each desc
  net: rswitch: Add a setting ext descriptor function
  net: rswitch: Set GWMDNC register
  net: rswitch: Add jumbo frames handling for RX
  net: rswitch: Add jumbo frames handling for TX
  net: rswitch: Allow jumbo frames

 drivers/net/ethernet/renesas/Makefile  |   1 -
 drivers/net/ethernet/renesas/rswitch.c | 376 +++++++++++++++++--------
 drivers/net/ethernet/renesas/rswitch.h |  43 ++-
 3 files changed, 294 insertions(+), 126 deletions(-)

-- 
2.34.1

[PATCH net-next v3 1/9] net: rswitch: Drop unused argument/return value

[Yoshihiro Shimoda]

Drop unused argument and return value of rswitch_tx_free() to
simplify the code.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
Reviewed-by: Geert Uytterhoeven <geert+renesas@glider.be>
---
 drivers/net/ethernet/renesas/rswitch.c | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index 89e7a6551c64..8c03c0e18003 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -761,20 +761,19 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 	return 0;
 }
 
-static int rswitch_tx_free(struct net_device *ndev, bool free_txed_only)
+static void rswitch_tx_free(struct net_device *ndev)
 {
 	struct rswitch_device *rdev = netdev_priv(ndev);
 	struct rswitch_gwca_queue *gq = rdev->tx_queue;
 	struct rswitch_ext_desc *desc;
 	dma_addr_t dma_addr;
 	struct sk_buff *skb;
-	int free_num = 0;
 	int size;
 
 	for (; rswitch_get_num_cur_queues(gq) > 0;
 	     gq->dirty = rswitch_next_queue_index(gq, false, 1)) {
 		desc = &gq->tx_ring[gq->dirty];
-		if (free_txed_only && (desc->desc.die_dt & DT_MASK) != DT_FEMPTY)
+		if ((desc->desc.die_dt & DT_MASK) != DT_FEMPTY)
 			break;
 
 		dma_rmb();
@@ -786,14 +785,11 @@ static int rswitch_tx_free(struct net_device *ndev, bool free_txed_only)
 					 size, DMA_TO_DEVICE);
 			dev_kfree_skb_any(gq->skbs[gq->dirty]);
 			gq->skbs[gq->dirty] = NULL;
-			free_num++;
 		}
 		desc->desc.die_dt = DT_EEMPTY;
 		rdev->ndev->stats.tx_packets++;
 		rdev->ndev->stats.tx_bytes += size;
 	}
-
-	return free_num;
 }
 
 static int rswitch_poll(struct napi_struct *napi, int budget)
@@ -808,7 +804,7 @@ static int rswitch_poll(struct napi_struct *napi, int budget)
 	priv = rdev->priv;
 
 retry:
-	rswitch_tx_free(ndev, true);
+	rswitch_tx_free(ndev);
 
 	if (rswitch_rx(ndev, &quota))
 		goto out;
-- 
2.34.1

[PATCH net-next v3 2/9] net: rswitch: Use unsigned int for desc related array index

[Yoshihiro Shimoda]

Array index should not be negative, so use unsigned int for
descriptors related array index.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/Makefile  |  1 -
 drivers/net/ethernet/renesas/rswitch.c | 88 ++++++++++++++------------
 drivers/net/ethernet/renesas/rswitch.h | 14 ++--
 3 files changed, 56 insertions(+), 47 deletions(-)

diff --git a/drivers/net/ethernet/renesas/Makefile b/drivers/net/ethernet/renesas/Makefile
index 9070acfd6aaf..f9cfd9750c1c 100644
--- a/drivers/net/ethernet/renesas/Makefile
+++ b/drivers/net/ethernet/renesas/Makefile
@@ -2,7 +2,6 @@
 #
 # Makefile for the Renesas network device drivers
 #
-
 obj-$(CONFIG_SH_ETH) += sh_eth.o
 
 ravb-objs := ravb_main.o ravb_ptp.o
diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index 8c03c0e18003..0928003ad245 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -56,7 +56,8 @@ static void rswitch_clock_disable(struct rswitch_private *priv)
 	iowrite32(RCDC_RCD, priv->addr + RCDC);
 }
 
-static bool rswitch_agent_clock_is_enabled(void __iomem *coma_addr, int port)
+static bool rswitch_agent_clock_is_enabled(void __iomem *coma_addr,
+					   unsigned int port)
 {
 	u32 val = ioread32(coma_addr + RCEC);
 
@@ -66,7 +67,8 @@ static bool rswitch_agent_clock_is_enabled(void __iomem *coma_addr, int port)
 		return false;
 }
 
-static void rswitch_agent_clock_ctrl(void __iomem *coma_addr, int port, int enable)
+static void rswitch_agent_clock_ctrl(void __iomem *coma_addr, unsigned int port,
+				     int enable)
 {
 	u32 val;
 
@@ -100,7 +102,7 @@ static void rswitch_coma_init(struct rswitch_private *priv)
 /* R-Switch-2 block (TOP) */
 static void rswitch_top_init(struct rswitch_private *priv)
 {
-	int i;
+	unsigned int i;
 
 	for (i = 0; i < RSWITCH_MAX_NUM_QUEUES; i++)
 		iowrite32((i / 16) << (GWCA_INDEX * 8), priv->addr + TPEMIMC7(i));
@@ -109,7 +111,7 @@ static void rswitch_top_init(struct rswitch_private *priv)
 /* Forwarding engine block (MFWD) */
 static void rswitch_fwd_init(struct rswitch_private *priv)
 {
-	int i;
+	unsigned int i;
 
 	/* For ETHA */
 	for (i = 0; i < RSWITCH_NUM_PORTS; i++) {
@@ -166,7 +168,7 @@ static int rswitch_gwca_axi_ram_reset(struct rswitch_private *priv)
 static bool rswitch_is_any_data_irq(struct rswitch_private *priv, u32 *dis, bool tx)
 {
 	u32 *mask = tx ? priv->gwca.tx_irq_bits : priv->gwca.rx_irq_bits;
-	int i;
+	unsigned int i;
 
 	for (i = 0; i < RSWITCH_NUM_IRQ_REGS; i++) {
 		if (dis[i] & mask[i])
@@ -178,7 +180,7 @@ static bool rswitch_is_any_data_irq(struct rswitch_private *priv, u32 *dis, bool
 
 static void rswitch_get_data_irq_status(struct rswitch_private *priv, u32 *dis)
 {
-	int i;
+	unsigned int i;
 
 	for (i = 0; i < RSWITCH_NUM_IRQ_REGS; i++) {
 		dis[i] = ioread32(priv->addr + GWDIS(i));
@@ -186,23 +188,26 @@ static void rswitch_get_data_irq_status(struct rswitch_private *priv, u32 *dis)
 	}
 }
 
-static void rswitch_enadis_data_irq(struct rswitch_private *priv, int index, bool enable)
+static void rswitch_enadis_data_irq(struct rswitch_private *priv,
+				    unsigned int index, bool enable)
 {
 	u32 offs = enable ? GWDIE(index / 32) : GWDID(index / 32);
 
 	iowrite32(BIT(index % 32), priv->addr + offs);
 }
 
-static void rswitch_ack_data_irq(struct rswitch_private *priv, int index)
+static void rswitch_ack_data_irq(struct rswitch_private *priv,
+				 unsigned int index)
 {
 	u32 offs = GWDIS(index / 32);
 
 	iowrite32(BIT(index % 32), priv->addr + offs);
 }
 
-static int rswitch_next_queue_index(struct rswitch_gwca_queue *gq, bool cur, int num)
+static unsigned int rswitch_next_queue_index(struct rswitch_gwca_queue *gq,
+					     bool cur, unsigned int num)
 {
-	int index = cur ? gq->cur : gq->dirty;
+	unsigned int index = cur ? gq->cur : gq->dirty;
 
 	if (index + num >= gq->ring_size)
 		index = (index + num) % gq->ring_size;
@@ -212,7 +217,7 @@ static int rswitch_next_queue_index(struct rswitch_gwca_queue *gq, bool cur, int
 	return index;
 }
 
-static int rswitch_get_num_cur_queues(struct rswitch_gwca_queue *gq)
+static unsigned int rswitch_get_num_cur_queues(struct rswitch_gwca_queue *gq)
 {
 	if (gq->cur >= gq->dirty)
 		return gq->cur - gq->dirty;
@@ -231,9 +236,10 @@ static bool rswitch_is_queue_rxed(struct rswitch_gwca_queue *gq)
 }
 
 static int rswitch_gwca_queue_alloc_skb(struct rswitch_gwca_queue *gq,
-					int start_index, int num)
+					unsigned int start_index,
+					unsigned int num)
 {
-	int i, index;
+	unsigned int i, index;
 
 	for (i = 0; i < num; i++) {
 		index = (i + start_index) % gq->ring_size;
@@ -248,7 +254,7 @@ static int rswitch_gwca_queue_alloc_skb(struct rswitch_gwca_queue *gq,
 	return 0;
 
 err:
-	for (i--; i >= 0; i--) {
+	for (; i-- > 0; ) {
 		index = (i + start_index) % gq->ring_size;
 		dev_kfree_skb(gq->skbs[index]);
 		gq->skbs[index] = NULL;
@@ -260,7 +266,7 @@ static int rswitch_gwca_queue_alloc_skb(struct rswitch_gwca_queue *gq,
 static void rswitch_gwca_queue_free(struct net_device *ndev,
 				    struct rswitch_gwca_queue *gq)
 {
-	int i;
+	unsigned int i;
 
 	if (!gq->dir_tx) {
 		dma_free_coherent(ndev->dev.parent,
@@ -294,9 +300,9 @@ static void rswitch_gwca_ts_queue_free(struct rswitch_private *priv)
 static int rswitch_gwca_queue_alloc(struct net_device *ndev,
 				    struct rswitch_private *priv,
 				    struct rswitch_gwca_queue *gq,
-				    bool dir_tx, int ring_size)
+				    bool dir_tx, unsigned int ring_size)
 {
-	int i, bit;
+	unsigned int i, bit;
 
 	gq->dir_tx = dir_tx;
 	gq->ring_size = ring_size;
@@ -351,11 +357,11 @@ static int rswitch_gwca_queue_format(struct net_device *ndev,
 				     struct rswitch_private *priv,
 				     struct rswitch_gwca_queue *gq)
 {
-	int ring_size = sizeof(struct rswitch_ext_desc) * gq->ring_size;
+	unsigned int ring_size = sizeof(struct rswitch_ext_desc) * gq->ring_size;
 	struct rswitch_ext_desc *desc;
 	struct rswitch_desc *linkfix;
 	dma_addr_t dma_addr;
-	int i;
+	unsigned int i;
 
 	memset(gq->tx_ring, 0, ring_size);
 	for (i = 0, desc = gq->tx_ring; i < gq->ring_size; i++, desc++) {
@@ -387,7 +393,7 @@ static int rswitch_gwca_queue_format(struct net_device *ndev,
 
 err:
 	if (!gq->dir_tx) {
-		for (i--, desc = gq->tx_ring; i >= 0; i--, desc++) {
+		for (desc = gq->tx_ring; i-- > 0; desc++) {
 			dma_addr = rswitch_desc_get_dptr(&desc->desc);
 			dma_unmap_single(ndev->dev.parent, dma_addr, PKT_BUF_SZ,
 					 DMA_FROM_DEVICE);
@@ -398,11 +404,12 @@ static int rswitch_gwca_queue_format(struct net_device *ndev,
 }
 
 static void rswitch_gwca_ts_queue_fill(struct rswitch_private *priv,
-				       int start_index, int num)
+				       unsigned int start_index,
+				       unsigned int num)
 {
 	struct rswitch_gwca_queue *gq = &priv->gwca.ts_queue;
 	struct rswitch_ts_desc *desc;
-	int i, index;
+	unsigned int i, index;
 
 	for (i = 0; i < num; i++) {
 		index = (i + start_index) % gq->ring_size;
@@ -413,12 +420,13 @@ static void rswitch_gwca_ts_queue_fill(struct rswitch_private *priv,
 
 static int rswitch_gwca_queue_ext_ts_fill(struct net_device *ndev,
 					  struct rswitch_gwca_queue *gq,
-					  int start_index, int num)
+					  unsigned int start_index,
+					  unsigned int num)
 {
 	struct rswitch_device *rdev = netdev_priv(ndev);
 	struct rswitch_ext_ts_desc *desc;
+	unsigned int i, index;
 	dma_addr_t dma_addr;
-	int i, index;
 
 	for (i = 0; i < num; i++) {
 		index = (i + start_index) % gq->ring_size;
@@ -444,7 +452,7 @@ static int rswitch_gwca_queue_ext_ts_fill(struct net_device *ndev,
 
 err:
 	if (!gq->dir_tx) {
-		for (i--; i >= 0; i--) {
+		for (; i-- > 0; ) {
 			index = (i + start_index) % gq->ring_size;
 			desc = &gq->rx_ring[index];
 			dma_addr = rswitch_desc_get_dptr(&desc->desc);
@@ -460,7 +468,7 @@ static int rswitch_gwca_queue_ext_ts_format(struct net_device *ndev,
 					    struct rswitch_private *priv,
 					    struct rswitch_gwca_queue *gq)
 {
-	int ring_size = sizeof(struct rswitch_ext_ts_desc) * gq->ring_size;
+	unsigned int ring_size = sizeof(struct rswitch_ext_ts_desc) * gq->ring_size;
 	struct rswitch_ext_ts_desc *desc;
 	struct rswitch_desc *linkfix;
 	int err;
@@ -487,7 +495,7 @@ static int rswitch_gwca_queue_ext_ts_format(struct net_device *ndev,
 
 static int rswitch_gwca_linkfix_alloc(struct rswitch_private *priv)
 {
-	int i, num_queues = priv->gwca.num_queues;
+	unsigned int i, num_queues = priv->gwca.num_queues;
 	struct rswitch_gwca *gwca = &priv->gwca;
 	struct device *dev = &priv->pdev->dev;
 
@@ -537,7 +545,7 @@ static int rswitch_gwca_ts_queue_alloc(struct rswitch_private *priv)
 static struct rswitch_gwca_queue *rswitch_gwca_get(struct rswitch_private *priv)
 {
 	struct rswitch_gwca_queue *gq;
-	int index;
+	unsigned int index;
 
 	index = find_first_zero_bit(priv->gwca.used, priv->gwca.num_queues);
 	if (index >= priv->gwca.num_queues)
@@ -583,7 +591,7 @@ static void rswitch_txdmac_free(struct net_device *ndev)
 	rswitch_gwca_put(rdev->priv, rdev->tx_queue);
 }
 
-static int rswitch_txdmac_init(struct rswitch_private *priv, int index)
+static int rswitch_txdmac_init(struct rswitch_private *priv, unsigned int index)
 {
 	struct rswitch_device *rdev = priv->rdev[index];
 
@@ -617,7 +625,7 @@ static void rswitch_rxdmac_free(struct net_device *ndev)
 	rswitch_gwca_put(rdev->priv, rdev->rx_queue);
 }
 
-static int rswitch_rxdmac_init(struct rswitch_private *priv, int index)
+static int rswitch_rxdmac_init(struct rswitch_private *priv, unsigned int index)
 {
 	struct rswitch_device *rdev = priv->rdev[index];
 	struct net_device *ndev = rdev->ndev;
@@ -627,7 +635,8 @@ static int rswitch_rxdmac_init(struct rswitch_private *priv, int index)
 
 static int rswitch_gwca_hw_init(struct rswitch_private *priv)
 {
-	int i, err;
+	unsigned int i;
+	int err;
 
 	err = rswitch_gwca_change_mode(priv, GWMC_OPC_DISABLE);
 	if (err < 0)
@@ -698,9 +707,10 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 	struct rswitch_device *rdev = netdev_priv(ndev);
 	struct rswitch_gwca_queue *gq = rdev->rx_queue;
 	struct rswitch_ext_ts_desc *desc;
-	int limit, boguscnt, num, ret;
+	int limit, boguscnt, ret;
 	struct sk_buff *skb;
 	dma_addr_t dma_addr;
+	unsigned int num;
 	u16 pkt_len;
 	u32 get_ts;
 
@@ -768,7 +778,7 @@ static void rswitch_tx_free(struct net_device *ndev)
 	struct rswitch_ext_desc *desc;
 	dma_addr_t dma_addr;
 	struct sk_buff *skb;
-	int size;
+	unsigned int size;
 
 	for (; rswitch_get_num_cur_queues(gq) > 0;
 	     gq->dirty = rswitch_next_queue_index(gq, false, 1)) {
@@ -847,7 +857,7 @@ static void rswitch_queue_interrupt(struct net_device *ndev)
 static irqreturn_t rswitch_data_irq(struct rswitch_private *priv, u32 *dis)
 {
 	struct rswitch_gwca_queue *gq;
-	int i, index, bit;
+	unsigned int i, index, bit;
 
 	for (i = 0; i < priv->gwca.num_queues; i++) {
 		gq = &priv->gwca.queues[i];
@@ -914,8 +924,8 @@ static void rswitch_ts(struct rswitch_private *priv)
 	struct skb_shared_hwtstamps shhwtstamps;
 	struct rswitch_ts_desc *desc;
 	struct timespec64 ts;
+	unsigned int num;
 	u32 tag, port;
-	int num;
 
 	desc = &gq->ts_ring[gq->cur];
 	while ((desc->desc.die_dt & DT_MASK) != DT_FEMPTY_ND) {
@@ -1434,7 +1444,7 @@ static int rswitch_ether_port_init_all(struct rswitch_private *priv)
 
 static void rswitch_ether_port_deinit_all(struct rswitch_private *priv)
 {
-	int i;
+	unsigned int i;
 
 	for (i = 0; i < RSWITCH_NUM_PORTS; i++) {
 		phy_exit(priv->rdev[i]->serdes);
@@ -1689,7 +1699,7 @@ static const struct of_device_id renesas_eth_sw_of_table[] = {
 };
 MODULE_DEVICE_TABLE(of, renesas_eth_sw_of_table);
 
-static void rswitch_etha_init(struct rswitch_private *priv, int index)
+static void rswitch_etha_init(struct rswitch_private *priv, unsigned int index)
 {
 	struct rswitch_etha *etha = &priv->etha[index];
 
@@ -1705,7 +1715,7 @@ static void rswitch_etha_init(struct rswitch_private *priv, int index)
 	etha->psmcs = clk_get_rate(priv->clk) / 100000 / (25 * 2) - 1;
 }
 
-static int rswitch_device_alloc(struct rswitch_private *priv, int index)
+static int rswitch_device_alloc(struct rswitch_private *priv, unsigned int index)
 {
 	struct platform_device *pdev = priv->pdev;
 	struct rswitch_device *rdev;
@@ -1776,7 +1786,7 @@ static int rswitch_device_alloc(struct rswitch_private *priv, int index)
 	return err;
 }
 
-static void rswitch_device_free(struct rswitch_private *priv, int index)
+static void rswitch_device_free(struct rswitch_private *priv, unsigned int index)
 {
 	struct rswitch_device *rdev = priv->rdev[index];
 	struct net_device *ndev = rdev->ndev;
diff --git a/drivers/net/ethernet/renesas/rswitch.h b/drivers/net/ethernet/renesas/rswitch.h
index 27c9d3872c0e..f1c7f7cdefc2 100644
--- a/drivers/net/ethernet/renesas/rswitch.h
+++ b/drivers/net/ethernet/renesas/rswitch.h
@@ -909,7 +909,7 @@ struct rswitch_ext_ts_desc {
 } __packed;
 
 struct rswitch_etha {
-	int index;
+	unsigned int index;
 	void __iomem *addr;
 	void __iomem *coma_addr;
 	bool external_phy;
@@ -938,12 +938,12 @@ struct rswitch_gwca_queue {
 
 	/* Common */
 	dma_addr_t ring_dma;
-	int ring_size;
-	int cur;
-	int dirty;
+	unsigned int ring_size;
+	unsigned int cur;
+	unsigned int dirty;
 
-	/* For [rt]_ring */
-	int index;
+	/* For [rt]x_ring */
+	unsigned int index;
 	bool dir_tx;
 	struct sk_buff **skbs;
 	struct net_device *ndev;	/* queue to ndev for irq */
@@ -959,7 +959,7 @@ struct rswitch_gwca_ts_info {
 
 #define RSWITCH_NUM_IRQ_REGS	(RSWITCH_MAX_NUM_QUEUES / BITS_PER_TYPE(u32))
 struct rswitch_gwca {
-	int index;
+	unsigned int index;
 	struct rswitch_desc *linkfix_table;
 	dma_addr_t linkfix_table_dma;
 	u32 linkfix_table_size;
-- 
2.34.1

[PATCH net-next v3 3/9] net: rswitch: Use build_skb() for RX

[Yoshihiro Shimoda]

If this hardware receives a jumbo frame like 2KiB or more, it will be
split into multiple queues. In the near future, to support this, use
build_skb() instead of netdev_alloc_skb_ip_align().

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 72 +++++++++++++++-----------
 drivers/net/ethernet/renesas/rswitch.h | 19 ++++++-
 2 files changed, 58 insertions(+), 33 deletions(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index 0928003ad245..8337c860eb9a 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -235,19 +235,18 @@ static bool rswitch_is_queue_rxed(struct rswitch_gwca_queue *gq)
 	return false;
 }
 
-static int rswitch_gwca_queue_alloc_skb(struct rswitch_gwca_queue *gq,
-					unsigned int start_index,
-					unsigned int num)
+static int rswitch_gwca_queue_alloc_rx_buf(struct rswitch_gwca_queue *gq,
+					   unsigned int start_index,
+					   unsigned int num)
 {
 	unsigned int i, index;
 
 	for (i = 0; i < num; i++) {
 		index = (i + start_index) % gq->ring_size;
-		if (gq->skbs[index])
+		if (gq->rx_bufs[index])
 			continue;
-		gq->skbs[index] = netdev_alloc_skb_ip_align(gq->ndev,
-							    PKT_BUF_SZ + RSWITCH_ALIGN - 1);
-		if (!gq->skbs[index])
+		gq->rx_bufs[index] = netdev_alloc_frag(RSWITCH_BUF_SIZE);
+		if (!gq->rx_bufs[index])
 			goto err;
 	}
 
@@ -256,8 +255,8 @@ static int rswitch_gwca_queue_alloc_skb(struct rswitch_gwca_queue *gq,
 err:
 	for (; i-- > 0; ) {
 		index = (i + start_index) % gq->ring_size;
-		dev_kfree_skb(gq->skbs[index]);
-		gq->skbs[index] = NULL;
+		skb_free_frag(gq->rx_bufs[index]);
+		gq->rx_bufs[index] = NULL;
 	}
 
 	return -ENOMEM;
@@ -275,16 +274,17 @@ static void rswitch_gwca_queue_free(struct net_device *ndev,
 		gq->rx_ring = NULL;
 
 		for (i = 0; i < gq->ring_size; i++)
-			dev_kfree_skb(gq->skbs[i]);
+			skb_free_frag(gq->rx_bufs[i]);
+		kfree(gq->rx_bufs);
+		gq->rx_bufs = NULL;
 	} else {
 		dma_free_coherent(ndev->dev.parent,
 				  sizeof(struct rswitch_ext_desc) *
 				  (gq->ring_size + 1), gq->tx_ring, gq->ring_dma);
 		gq->tx_ring = NULL;
+		kfree(gq->skbs);
+		gq->skbs = NULL;
 	}
-
-	kfree(gq->skbs);
-	gq->skbs = NULL;
 }
 
 static void rswitch_gwca_ts_queue_free(struct rswitch_private *priv)
@@ -308,17 +308,19 @@ static int rswitch_gwca_queue_alloc(struct net_device *ndev,
 	gq->ring_size = ring_size;
 	gq->ndev = ndev;
 
-	gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
-	if (!gq->skbs)
-		return -ENOMEM;
-
 	if (!dir_tx) {
-		rswitch_gwca_queue_alloc_skb(gq, 0, gq->ring_size);
+		gq->rx_bufs = kcalloc(gq->ring_size, sizeof(*gq->rx_bufs), GFP_KERNEL);
+		if (!gq->rx_bufs)
+			goto out;
+		rswitch_gwca_queue_alloc_rx_buf(gq, 0, gq->ring_size);
 
 		gq->rx_ring = dma_alloc_coherent(ndev->dev.parent,
 						 sizeof(struct rswitch_ext_ts_desc) *
 						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);
 	} else {
+		gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
+		if (!gq->skbs)
+			return -ENOMEM;
 		gq->tx_ring = dma_alloc_coherent(ndev->dev.parent,
 						 sizeof(struct rswitch_ext_desc) *
 						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);
@@ -367,12 +369,13 @@ static int rswitch_gwca_queue_format(struct net_device *ndev,
 	for (i = 0, desc = gq->tx_ring; i < gq->ring_size; i++, desc++) {
 		if (!gq->dir_tx) {
 			dma_addr = dma_map_single(ndev->dev.parent,
-						  gq->skbs[i]->data, PKT_BUF_SZ,
+						  gq->rx_bufs[i] + RSWITCH_HEADROOM,
+						  RSWITCH_MAP_BUF_SIZE,
 						  DMA_FROM_DEVICE);
 			if (dma_mapping_error(ndev->dev.parent, dma_addr))
 				goto err;
 
-			desc->desc.info_ds = cpu_to_le16(PKT_BUF_SZ);
+			desc->desc.info_ds = cpu_to_le16(RSWITCH_DESC_BUF_SIZE);
 			rswitch_desc_set_dptr(&desc->desc, dma_addr);
 			desc->desc.die_dt = DT_FEMPTY | DIE;
 		} else {
@@ -395,8 +398,8 @@ static int rswitch_gwca_queue_format(struct net_device *ndev,
 	if (!gq->dir_tx) {
 		for (desc = gq->tx_ring; i-- > 0; desc++) {
 			dma_addr = rswitch_desc_get_dptr(&desc->desc);
-			dma_unmap_single(ndev->dev.parent, dma_addr, PKT_BUF_SZ,
-					 DMA_FROM_DEVICE);
+			dma_unmap_single(ndev->dev.parent, dma_addr,
+					 RSWITCH_MAP_BUF_SIZE, DMA_FROM_DEVICE);
 		}
 	}
 
@@ -433,12 +436,13 @@ static int rswitch_gwca_queue_ext_ts_fill(struct net_device *ndev,
 		desc = &gq->rx_ring[index];
 		if (!gq->dir_tx) {
 			dma_addr = dma_map_single(ndev->dev.parent,
-						  gq->skbs[index]->data, PKT_BUF_SZ,
+						  gq->rx_bufs[index] + RSWITCH_HEADROOM,
+						  RSWITCH_MAP_BUF_SIZE,
 						  DMA_FROM_DEVICE);
 			if (dma_mapping_error(ndev->dev.parent, dma_addr))
 				goto err;
 
-			desc->desc.info_ds = cpu_to_le16(PKT_BUF_SZ);
+			desc->desc.info_ds = cpu_to_le16(RSWITCH_DESC_BUF_SIZE);
 			rswitch_desc_set_dptr(&desc->desc, dma_addr);
 			dma_wmb();
 			desc->desc.die_dt = DT_FEMPTY | DIE;
@@ -456,8 +460,8 @@ static int rswitch_gwca_queue_ext_ts_fill(struct net_device *ndev,
 			index = (i + start_index) % gq->ring_size;
 			desc = &gq->rx_ring[index];
 			dma_addr = rswitch_desc_get_dptr(&desc->desc);
-			dma_unmap_single(ndev->dev.parent, dma_addr, PKT_BUF_SZ,
-					 DMA_FROM_DEVICE);
+			dma_unmap_single(ndev->dev.parent, dma_addr,
+					 RSWITCH_MAP_BUF_SIZE, DMA_FROM_DEVICE);
 		}
 	}
 
@@ -724,10 +728,15 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 	while ((desc->desc.die_dt & DT_MASK) != DT_FEMPTY) {
 		dma_rmb();
 		pkt_len = le16_to_cpu(desc->desc.info_ds) & RX_DS;
-		skb = gq->skbs[gq->cur];
-		gq->skbs[gq->cur] = NULL;
 		dma_addr = rswitch_desc_get_dptr(&desc->desc);
-		dma_unmap_single(ndev->dev.parent, dma_addr, PKT_BUF_SZ, DMA_FROM_DEVICE);
+		dma_unmap_single(ndev->dev.parent, dma_addr,
+				 RSWITCH_MAP_BUF_SIZE, DMA_FROM_DEVICE);
+		skb = build_skb(gq->rx_bufs[gq->cur], RSWITCH_BUF_SIZE);
+		if (!skb)
+			goto out;
+		skb_reserve(skb, RSWITCH_HEADROOM);
+		skb_put(skb, pkt_len);
+
 		get_ts = rdev->priv->ptp_priv->tstamp_rx_ctrl & RCAR_GEN4_RXTSTAMP_TYPE_V2_L2_EVENT;
 		if (get_ts) {
 			struct skb_shared_hwtstamps *shhwtstamps;
@@ -739,12 +748,13 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 			ts.tv_nsec = __le32_to_cpu(desc->ts_nsec & cpu_to_le32(0x3fffffff));
 			shhwtstamps->hwtstamp = timespec64_to_ktime(ts);
 		}
-		skb_put(skb, pkt_len);
 		skb->protocol = eth_type_trans(skb, ndev);
 		napi_gro_receive(&rdev->napi, skb);
 		rdev->ndev->stats.rx_packets++;
 		rdev->ndev->stats.rx_bytes += pkt_len;
 
+out:
+		gq->rx_bufs[gq->cur] = NULL;
 		gq->cur = rswitch_next_queue_index(gq, true, 1);
 		desc = &gq->rx_ring[gq->cur];
 
@@ -753,7 +763,7 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 	}
 
 	num = rswitch_get_num_cur_queues(gq);
-	ret = rswitch_gwca_queue_alloc_skb(gq, gq->dirty, num);
+	ret = rswitch_gwca_queue_alloc_rx_buf(gq, gq->dirty, num);
 	if (ret < 0)
 		goto err;
 	ret = rswitch_gwca_queue_ext_ts_fill(ndev, gq, gq->dirty, num);
diff --git a/drivers/net/ethernet/renesas/rswitch.h b/drivers/net/ethernet/renesas/rswitch.h
index f1c7f7cdefc2..a407d85dcfad 100644
--- a/drivers/net/ethernet/renesas/rswitch.h
+++ b/drivers/net/ethernet/renesas/rswitch.h
@@ -29,8 +29,13 @@
 #define RX_RING_SIZE		1024
 #define TS_RING_SIZE		(TX_RING_SIZE * RSWITCH_NUM_PORTS)
 
-#define PKT_BUF_SZ		1584
+#define RSWITCH_HEADROOM	(NET_SKB_PAD + NET_IP_ALIGN)
+#define RSWITCH_DESC_BUF_SIZE	2048
+#define RSWITCH_TAILROOM	SKB_DATA_ALIGN(sizeof(struct skb_shared_info))
 #define RSWITCH_ALIGN		128
+#define RSWITCH_BUF_SIZE	(RSWITCH_HEADROOM + RSWITCH_DESC_BUF_SIZE + \
+				 RSWITCH_TAILROOM + RSWITCH_ALIGN)
+#define RSWITCH_MAP_BUF_SIZE	(RSWITCH_BUF_SIZE - RSWITCH_HEADROOM)
 #define RSWITCH_MAX_CTAG_PCP	7
 
 #define RSWITCH_TIMEOUT_US	100000
@@ -945,8 +950,18 @@ struct rswitch_gwca_queue {
 	/* For [rt]x_ring */
 	unsigned int index;
 	bool dir_tx;
-	struct sk_buff **skbs;
 	struct net_device *ndev;	/* queue to ndev for irq */
+
+	union {
+		/* For TX */
+		struct {
+			struct sk_buff **skbs;
+		};
+		/* For RX */
+		struct {
+			void **rx_bufs;
+		};
+	};
 };
 
 struct rswitch_gwca_ts_info {
-- 
2.34.1

Re: [PATCH net-next v3 3/9] net: rswitch: Use build_skb() for RX

[Simon Horman]

On Mon, Dec 04, 2023 at 10:20:52AM +0900, Yoshihiro Shimoda wrote:
> If this hardware receives a jumbo frame like 2KiB or more, it will be
> split into multiple queues. In the near future, to support this, use
> build_skb() instead of netdev_alloc_skb_ip_align().
> 
> Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>

...

>  static void rswitch_gwca_ts_queue_free(struct rswitch_private *priv)
> @@ -308,17 +308,19 @@ static int rswitch_gwca_queue_alloc(struct net_device *ndev,
>  	gq->ring_size = ring_size;
>  	gq->ndev = ndev;
>  
> -	gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
> -	if (!gq->skbs)
> -		return -ENOMEM;
> -
>  	if (!dir_tx) {
> -		rswitch_gwca_queue_alloc_skb(gq, 0, gq->ring_size);
> +		gq->rx_bufs = kcalloc(gq->ring_size, sizeof(*gq->rx_bufs), GFP_KERNEL);
> +		if (!gq->rx_bufs)
> +			goto out;

Hi Shimoda-san,

there is no need to re-spin because of this,
but I have some commends on error handling.

I think that for consistency this can just return -ENOMEM.
Or alternatively, perhaps 'goto out' can be used for the (!gq->skbs)
condition below.

> +		rswitch_gwca_queue_alloc_rx_buf(gq, 0, gq->ring_size);

Not strictly related to this patch, but should
the return value of rswitch_gwca_queue_alloc_rx_buf be checked?

>  
>  		gq->rx_ring = dma_alloc_coherent(ndev->dev.parent,
>  						 sizeof(struct rswitch_ext_ts_desc) *
>  						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);
>  	} else {
> +		gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
> +		if (!gq->skbs)
> +			return -ENOMEM;
>  		gq->tx_ring = dma_alloc_coherent(ndev->dev.parent,
>  						 sizeof(struct rswitch_ext_desc) *
>  						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);

...

RE: [PATCH net-next v3 3/9] net: rswitch: Use build_skb() for RX

[Yoshihiro Shimoda]

Hello Simon-san,

> From: Simon Horman, Sent: Thursday, December 7, 2023 4:08 AM
> 
> On Mon, Dec 04, 2023 at 10:20:52AM +0900, Yoshihiro Shimoda wrote:
> > If this hardware receives a jumbo frame like 2KiB or more, it will be
> > split into multiple queues. In the near future, to support this, use
> > build_skb() instead of netdev_alloc_skb_ip_align().
> >
> > Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
> 
> ...
> 
> >  static void rswitch_gwca_ts_queue_free(struct rswitch_private *priv)
> > @@ -308,17 +308,19 @@ static int rswitch_gwca_queue_alloc(struct net_device *ndev,
> >  	gq->ring_size = ring_size;
> >  	gq->ndev = ndev;
> >
> > -	gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
> > -	if (!gq->skbs)
> > -		return -ENOMEM;
> > -
> >  	if (!dir_tx) {
> > -		rswitch_gwca_queue_alloc_skb(gq, 0, gq->ring_size);
> > +		gq->rx_bufs = kcalloc(gq->ring_size, sizeof(*gq->rx_bufs), GFP_KERNEL);
> > +		if (!gq->rx_bufs)
> > +			goto out;
> 
> Hi Shimoda-san,
> 
> there is no need to re-spin because of this,
> but I have some commends on error handling.
> 
> I think that for consistency this can just return -ENOMEM.

Thank you for your comments! I think so. So, I'll fix it on v4.

> Or alternatively, perhaps 'goto out' can be used for the (!gq->skbs)
> condition below.
> 
> > +		rswitch_gwca_queue_alloc_rx_buf(gq, 0, gq->ring_size);
> 
> Not strictly related to this patch, but should
> the return value of rswitch_gwca_queue_alloc_rx_buf be checked?

I think so. I'll fix it.

Best regards,
Yoshihiro Shimoda


> >
> >  		gq->rx_ring = dma_alloc_coherent(ndev->dev.parent,
> >  						 sizeof(struct rswitch_ext_ts_desc) *
> >  						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);
> >  	} else {
> > +		gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
> > +		if (!gq->skbs)
> > +			return -ENOMEM;
> >  		gq->tx_ring = dma_alloc_coherent(ndev->dev.parent,
> >  						 sizeof(struct rswitch_ext_desc) *
> >  						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);
> 
> ...

[PATCH net-next v3 4/9] net: rswitch: Add unmap_addrs instead of dma address in each desc

[Yoshihiro Shimoda]

If the driver would like to transmit a jumbo frame like 2KiB or more,
it should be split into multiple queues. In the near future, to support
this, add unmap_addrs array to unmap dma mapping address instead of dma
address in each TX descriptor because the descriptors may not have
the top dma address.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 19 +++++++++++--------
 drivers/net/ethernet/renesas/rswitch.h |  1 +
 2 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index 8337c860eb9a..4774130c417b 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -284,6 +284,8 @@ static void rswitch_gwca_queue_free(struct net_device *ndev,
 		gq->tx_ring = NULL;
 		kfree(gq->skbs);
 		gq->skbs = NULL;
+		kfree(gq->unmap_addrs);
+		gq->unmap_addrs = NULL;
 	}
 }
 
@@ -321,6 +323,9 @@ static int rswitch_gwca_queue_alloc(struct net_device *ndev,
 		gq->skbs = kcalloc(gq->ring_size, sizeof(*gq->skbs), GFP_KERNEL);
 		if (!gq->skbs)
 			return -ENOMEM;
+		gq->unmap_addrs = kcalloc(gq->ring_size, sizeof(*gq->unmap_addrs), GFP_KERNEL);
+		if (!gq->unmap_addrs)
+			goto out;
 		gq->tx_ring = dma_alloc_coherent(ndev->dev.parent,
 						 sizeof(struct rswitch_ext_desc) *
 						 (gq->ring_size + 1), &gq->ring_dma, GFP_KERNEL);
@@ -786,9 +791,7 @@ static void rswitch_tx_free(struct net_device *ndev)
 	struct rswitch_device *rdev = netdev_priv(ndev);
 	struct rswitch_gwca_queue *gq = rdev->tx_queue;
 	struct rswitch_ext_desc *desc;
-	dma_addr_t dma_addr;
 	struct sk_buff *skb;
-	unsigned int size;
 
 	for (; rswitch_get_num_cur_queues(gq) > 0;
 	     gq->dirty = rswitch_next_queue_index(gq, false, 1)) {
@@ -797,18 +800,17 @@ static void rswitch_tx_free(struct net_device *ndev)
 			break;
 
 		dma_rmb();
-		size = le16_to_cpu(desc->desc.info_ds) & TX_DS;
 		skb = gq->skbs[gq->dirty];
 		if (skb) {
-			dma_addr = rswitch_desc_get_dptr(&desc->desc);
-			dma_unmap_single(ndev->dev.parent, dma_addr,
-					 size, DMA_TO_DEVICE);
+			dma_unmap_single(ndev->dev.parent,
+					 gq->unmap_addrs[gq->dirty],
+					 skb->len, DMA_TO_DEVICE);
 			dev_kfree_skb_any(gq->skbs[gq->dirty]);
 			gq->skbs[gq->dirty] = NULL;
+			rdev->ndev->stats.tx_packets++;
+			rdev->ndev->stats.tx_bytes += skb->len;
 		}
 		desc->desc.die_dt = DT_EEMPTY;
-		rdev->ndev->stats.tx_packets++;
-		rdev->ndev->stats.tx_bytes += size;
 	}
 }
 
@@ -1537,6 +1539,7 @@ static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *nd
 		goto err_kfree;
 
 	gq->skbs[gq->cur] = skb;
+	gq->unmap_addrs[gq->cur] = dma_addr;
 	desc = &gq->tx_ring[gq->cur];
 	rswitch_desc_set_dptr(&desc->desc, dma_addr);
 	desc->desc.info_ds = cpu_to_le16(skb->len);
diff --git a/drivers/net/ethernet/renesas/rswitch.h b/drivers/net/ethernet/renesas/rswitch.h
index a407d85dcfad..2ac9a86b6238 100644
--- a/drivers/net/ethernet/renesas/rswitch.h
+++ b/drivers/net/ethernet/renesas/rswitch.h
@@ -956,6 +956,7 @@ struct rswitch_gwca_queue {
 		/* For TX */
 		struct {
 			struct sk_buff **skbs;
+			dma_addr_t *unmap_addrs;
 		};
 		/* For RX */
 		struct {
-- 
2.34.1

[PATCH net-next v3 5/9] net: rswitch: Add a setting ext descriptor function

[Yoshihiro Shimoda]

If the driver would like to transmit a jumbo frame like 2KiB or more,
it should be split into multiple queues. In the near future, to support
this, add a setting ext descriptor function to improve code readability.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 73 +++++++++++++++++---------
 1 file changed, 47 insertions(+), 26 deletions(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index 4774130c417b..f94c76161794 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -1518,6 +1518,51 @@ static int rswitch_stop(struct net_device *ndev)
 	return 0;
 };
 
+static bool rswitch_ext_desc_set_info1(struct rswitch_device *rdev,
+				       struct sk_buff *skb,
+				       struct rswitch_ext_desc *desc)
+{
+	desc->info1 = cpu_to_le64(INFO1_DV(BIT(rdev->etha->index)) |
+				  INFO1_IPV(GWCA_IPV_NUM) | INFO1_FMT);
+	if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
+		struct rswitch_gwca_ts_info *ts_info;
+
+		ts_info = kzalloc(sizeof(*ts_info), GFP_ATOMIC);
+		if (!ts_info)
+			return false;
+
+		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
+		rdev->ts_tag++;
+		desc->info1 |= cpu_to_le64(INFO1_TSUN(rdev->ts_tag) | INFO1_TXC);
+
+		ts_info->skb = skb_get(skb);
+		ts_info->port = rdev->port;
+		ts_info->tag = rdev->ts_tag;
+		list_add_tail(&ts_info->list, &rdev->priv->gwca.ts_info_list);
+
+		skb_tx_timestamp(skb);
+	}
+
+	return true;
+}
+
+static bool rswitch_ext_desc_set(struct rswitch_device *rdev,
+				 struct sk_buff *skb,
+				 struct rswitch_ext_desc *desc,
+				 dma_addr_t dma_addr, u16 len, u8 die_dt)
+{
+	rswitch_desc_set_dptr(&desc->desc, dma_addr);
+	desc->desc.info_ds = cpu_to_le16(len);
+	if (!rswitch_ext_desc_set_info1(rdev, skb, desc))
+		return false;
+
+	dma_wmb();
+
+	desc->desc.die_dt = die_dt;
+
+	return true;
+}
+
 static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 {
 	struct rswitch_device *rdev = netdev_priv(ndev);
@@ -1541,33 +1586,9 @@ static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *nd
 	gq->skbs[gq->cur] = skb;
 	gq->unmap_addrs[gq->cur] = dma_addr;
 	desc = &gq->tx_ring[gq->cur];
-	rswitch_desc_set_dptr(&desc->desc, dma_addr);
-	desc->desc.info_ds = cpu_to_le16(skb->len);
-
-	desc->info1 = cpu_to_le64(INFO1_DV(BIT(rdev->etha->index)) |
-				  INFO1_IPV(GWCA_IPV_NUM) | INFO1_FMT);
-	if (skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
-		struct rswitch_gwca_ts_info *ts_info;
-
-		ts_info = kzalloc(sizeof(*ts_info), GFP_ATOMIC);
-		if (!ts_info)
-			goto err_unmap;
-
-		skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
-		rdev->ts_tag++;
-		desc->info1 |= cpu_to_le64(INFO1_TSUN(rdev->ts_tag) | INFO1_TXC);
-
-		ts_info->skb = skb_get(skb);
-		ts_info->port = rdev->port;
-		ts_info->tag = rdev->ts_tag;
-		list_add_tail(&ts_info->list, &rdev->priv->gwca.ts_info_list);
-
-		skb_tx_timestamp(skb);
-	}
-
-	dma_wmb();
+	if (!rswitch_ext_desc_set(rdev, skb, desc, dma_addr, skb->len, DT_FSINGLE | DIE))
+		goto err_unmap;
 
-	desc->desc.die_dt = DT_FSINGLE | DIE;
 	wmb();	/* gq->cur must be incremented after die_dt was set */
 
 	gq->cur = rswitch_next_queue_index(gq, true, 1);
-- 
2.34.1

[PATCH net-next v3 6/9] net: rswitch: Set GWMDNC register

[Yoshihiro Shimoda]

To support jumbo frames, set GWMDNC register with acceptable maximum
values for TX and RX.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 2 ++
 drivers/net/ethernet/renesas/rswitch.h | 4 ++++
 2 files changed, 6 insertions(+)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index f94c76161794..d24ba2411d5e 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -667,6 +667,8 @@ static int rswitch_gwca_hw_init(struct rswitch_private *priv)
 	iowrite32(upper_32_bits(priv->gwca.linkfix_table_dma), priv->addr + GWDCBAC0);
 	iowrite32(lower_32_bits(priv->gwca.ts_queue.ring_dma), priv->addr + GWTDCAC10);
 	iowrite32(upper_32_bits(priv->gwca.ts_queue.ring_dma), priv->addr + GWTDCAC00);
+	iowrite32(GWMDNC_TSDMN(1) | GWMDNC_TXDMN(0x1e) | GWMDNC_RXDMN(0x1f),
+		  priv->addr + GWMDNC);
 	iowrite32(GWCA_TS_IRQ_BIT, priv->addr + GWTSDCC0);
 
 	iowrite32(GWTPC_PPPL(GWCA_IPV_NUM), priv->addr + GWTPC0);
diff --git a/drivers/net/ethernet/renesas/rswitch.h b/drivers/net/ethernet/renesas/rswitch.h
index 2ac9a86b6238..17e617ec4187 100644
--- a/drivers/net/ethernet/renesas/rswitch.h
+++ b/drivers/net/ethernet/renesas/rswitch.h
@@ -773,6 +773,10 @@ enum rswitch_gwca_mode {
 #define GWARIRM_ARIOG		BIT(0)
 #define GWARIRM_ARR		BIT(1)
 
+#define GWMDNC_TSDMN(num)	(((num) << 16) & GENMASK(17, 16))
+#define GWMDNC_TXDMN(num)	(((num) << 8) & GENMASK(12, 8))
+#define GWMDNC_RXDMN(num)	((num) & GENMASK(4, 0))
+
 #define GWDCC_BALR		BIT(24)
 #define GWDCC_DCP_MASK		GENMASK(18, 16)
 #define GWDCC_DCP(prio)		FIELD_PREP(GWDCC_DCP_MASK, (prio))
-- 
2.34.1

[PATCH net-next v3 7/9] net: rswitch: Add jumbo frames handling for RX

[Yoshihiro Shimoda]

If this hardware receives a jumbo frame like 2KiB or more, it will be
split into multiple queues. In the near future, to support this,
add handling specific descriptor types F{START,MID,END}. However, such
jumbo frames will not happen yet because the maximum MTU size is still
default for now.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 86 +++++++++++++++++++++++---
 drivers/net/ethernet/renesas/rswitch.h |  2 +
 2 files changed, 78 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index d24ba2411d5e..009e6bfdad27 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -713,6 +713,80 @@ static int rswitch_gwca_halt(struct rswitch_private *priv)
 	return err;
 }
 
+static struct sk_buff *rswitch_rx_handle_desc(struct net_device *ndev,
+					      struct rswitch_gwca_queue *gq,
+					      struct rswitch_ext_ts_desc *desc)
+{
+	dma_addr_t dma_addr = rswitch_desc_get_dptr(&desc->desc);
+	u16 pkt_len = le16_to_cpu(desc->desc.info_ds) & RX_DS;
+	u8 die_dt = desc->desc.die_dt & DT_MASK;
+	struct sk_buff *skb = NULL;
+
+	dma_unmap_single(ndev->dev.parent, dma_addr, RSWITCH_MAP_BUF_SIZE,
+			 DMA_FROM_DEVICE);
+
+	/* The RX descriptor order will be one of the following:
+	 * - FSINGLE
+	 * - FSTART -> FEND
+	 * - FSTART -> FMID -> FEND
+	 */
+
+	/* Check whether the descriptor is unexpected order */
+	switch (die_dt) {
+	case DT_FSTART:
+	case DT_FSINGLE:
+		if (gq->skb_fstart) {
+			dev_kfree_skb_any(gq->skb_fstart);
+			gq->skb_fstart = NULL;
+			ndev->stats.rx_dropped++;
+		}
+		break;
+	case DT_FMID:
+	case DT_FEND:
+		if (!gq->skb_fstart) {
+			ndev->stats.rx_dropped++;
+			return NULL;
+		}
+		break;
+	default:
+		break;
+	}
+
+	/* Handle the descriptor */
+	switch (die_dt) {
+	case DT_FSTART:
+	case DT_FSINGLE:
+		skb = build_skb(gq->rx_bufs[gq->cur], RSWITCH_BUF_SIZE);
+		if (skb) {
+			skb_reserve(skb, RSWITCH_HEADROOM);
+			skb_put(skb, pkt_len);
+			gq->pkt_len = pkt_len;
+			if (die_dt == DT_FSTART) {
+				gq->skb_fstart = skb;
+				skb = NULL;
+			}
+		}
+		break;
+	case DT_FMID:
+	case DT_FEND:
+		skb_add_rx_frag(gq->skb_fstart, skb_shinfo(gq->skb_fstart)->nr_frags,
+				virt_to_page(gq->rx_bufs[gq->cur]),
+				offset_in_page(gq->rx_bufs[gq->cur]) + RSWITCH_HEADROOM,
+				pkt_len, RSWITCH_BUF_SIZE);
+		if (die_dt == DT_FEND) {
+			skb = gq->skb_fstart;
+			gq->skb_fstart = NULL;
+		}
+		gq->pkt_len += pkt_len;
+		break;
+	default:
+		netdev_err(ndev, "%s: unexpected value (%x)\n", __func__, die_dt);
+		break;
+	}
+
+	return skb;
+}
+
 static bool rswitch_rx(struct net_device *ndev, int *quota)
 {
 	struct rswitch_device *rdev = netdev_priv(ndev);
@@ -720,9 +794,7 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 	struct rswitch_ext_ts_desc *desc;
 	int limit, boguscnt, ret;
 	struct sk_buff *skb;
-	dma_addr_t dma_addr;
 	unsigned int num;
-	u16 pkt_len;
 	u32 get_ts;
 
 	if (*quota <= 0)
@@ -734,15 +806,9 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 	desc = &gq->rx_ring[gq->cur];
 	while ((desc->desc.die_dt & DT_MASK) != DT_FEMPTY) {
 		dma_rmb();
-		pkt_len = le16_to_cpu(desc->desc.info_ds) & RX_DS;
-		dma_addr = rswitch_desc_get_dptr(&desc->desc);
-		dma_unmap_single(ndev->dev.parent, dma_addr,
-				 RSWITCH_MAP_BUF_SIZE, DMA_FROM_DEVICE);
-		skb = build_skb(gq->rx_bufs[gq->cur], RSWITCH_BUF_SIZE);
+		skb = rswitch_rx_handle_desc(ndev, gq, desc);
 		if (!skb)
 			goto out;
-		skb_reserve(skb, RSWITCH_HEADROOM);
-		skb_put(skb, pkt_len);
 
 		get_ts = rdev->priv->ptp_priv->tstamp_rx_ctrl & RCAR_GEN4_RXTSTAMP_TYPE_V2_L2_EVENT;
 		if (get_ts) {
@@ -758,7 +824,7 @@ static bool rswitch_rx(struct net_device *ndev, int *quota)
 		skb->protocol = eth_type_trans(skb, ndev);
 		napi_gro_receive(&rdev->napi, skb);
 		rdev->ndev->stats.rx_packets++;
-		rdev->ndev->stats.rx_bytes += pkt_len;
+		rdev->ndev->stats.rx_bytes += gq->pkt_len;
 
 out:
 		gq->rx_bufs[gq->cur] = NULL;
diff --git a/drivers/net/ethernet/renesas/rswitch.h b/drivers/net/ethernet/renesas/rswitch.h
index 17e617ec4187..4252677e2a55 100644
--- a/drivers/net/ethernet/renesas/rswitch.h
+++ b/drivers/net/ethernet/renesas/rswitch.h
@@ -965,6 +965,8 @@ struct rswitch_gwca_queue {
 		/* For RX */
 		struct {
 			void **rx_bufs;
+			struct sk_buff *skb_fstart;
+			u16 pkt_len;
 		};
 	};
 };
-- 
2.34.1

[PATCH net-next v3 8/9] net: rswitch: Add jumbo frames handling for TX

[Yoshihiro Shimoda]

If the driver would like to transmit a jumbo frame like 2KiB or more,
it should be split into multiple queues. In the near future, to support
this, add handling specific descriptor types F{START,MID,END}. However,
such jumbo frames will not happen yet because the maximum MTU size is
still default for now.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 54 +++++++++++++++++++++-----
 1 file changed, 45 insertions(+), 9 deletions(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index 009e6bfdad27..df419112b67a 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -1631,15 +1631,44 @@ static bool rswitch_ext_desc_set(struct rswitch_device *rdev,
 	return true;
 }
 
+static u8 rswitch_ext_desc_get_die_dt(unsigned int nr_desc, unsigned int index)
+{
+	if (nr_desc == 1)
+		return DT_FSINGLE | DIE;
+	if (index == 0)
+		return DT_FSTART;
+	if (nr_desc - 1 == index)
+		return DT_FEND | DIE;
+	return DT_FMID;
+}
+
+static u16 rswitch_ext_desc_get_len(u8 die_dt, unsigned int orig_len)
+{
+	switch (die_dt & DT_MASK) {
+	case DT_FSINGLE:
+	case DT_FEND:
+		return (orig_len % RSWITCH_DESC_BUF_SIZE) ?: RSWITCH_DESC_BUF_SIZE;
+	case DT_FSTART:
+	case DT_FMID:
+		return RSWITCH_DESC_BUF_SIZE;
+	default:
+		return 0;
+	}
+}
+
 static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *ndev)
 {
 	struct rswitch_device *rdev = netdev_priv(ndev);
 	struct rswitch_gwca_queue *gq = rdev->tx_queue;
+	dma_addr_t dma_addr, dma_addr_orig;
 	netdev_tx_t ret = NETDEV_TX_OK;
 	struct rswitch_ext_desc *desc;
-	dma_addr_t dma_addr;
+	unsigned int i, nr_desc;
+	u8 die_dt;
+	u16 len;
 
-	if (rswitch_get_num_cur_queues(gq) >= gq->ring_size - 1) {
+	nr_desc = (skb->len - 1) / RSWITCH_DESC_BUF_SIZE + 1;
+	if (rswitch_get_num_cur_queues(gq) >= gq->ring_size - nr_desc) {
 		netif_stop_subqueue(ndev, 0);
 		return NETDEV_TX_BUSY;
 	}
@@ -1647,25 +1676,32 @@ static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *nd
 	if (skb_put_padto(skb, ETH_ZLEN))
 		return ret;
 
-	dma_addr = dma_map_single(ndev->dev.parent, skb->data, skb->len, DMA_TO_DEVICE);
-	if (dma_mapping_error(ndev->dev.parent, dma_addr))
+	dma_addr_orig = dma_map_single(ndev->dev.parent, skb->data, skb->len, DMA_TO_DEVICE);
+	if (dma_mapping_error(ndev->dev.parent, dma_addr_orig))
 		goto err_kfree;
 
 	gq->skbs[gq->cur] = skb;
 	gq->unmap_addrs[gq->cur] = dma_addr;
-	desc = &gq->tx_ring[gq->cur];
-	if (!rswitch_ext_desc_set(rdev, skb, desc, dma_addr, skb->len, DT_FSINGLE | DIE))
-		goto err_unmap;
+
+	/* DT_FSTART should be set at last. So, this is reverse order. */
+	for (i = nr_desc; i-- > 0; ) {
+		desc = &gq->tx_ring[rswitch_next_queue_index(gq, true, i)];
+		die_dt = rswitch_ext_desc_get_die_dt(nr_desc, i);
+		dma_addr = dma_addr_orig + i * RSWITCH_DESC_BUF_SIZE;
+		len = rswitch_ext_desc_get_len(die_dt, skb->len);
+		if (!rswitch_ext_desc_set(rdev, skb, desc, dma_addr, len, die_dt))
+			goto err_unmap;
+	}
 
 	wmb();	/* gq->cur must be incremented after die_dt was set */
 
-	gq->cur = rswitch_next_queue_index(gq, true, 1);
+	gq->cur = rswitch_next_queue_index(gq, true, nr_desc);
 	rswitch_modify(rdev->addr, GWTRC(gq->index), 0, BIT(gq->index % 32));
 
 	return ret;
 
 err_unmap:
-	dma_unmap_single(ndev->dev.parent, dma_addr, skb->len, DMA_TO_DEVICE);
+	dma_unmap_single(ndev->dev.parent, dma_addr_orig, skb->len, DMA_TO_DEVICE);
 
 err_kfree:
 	dev_kfree_skb_any(skb);
-- 
2.34.1

Re: [PATCH net-next v3 8/9] net: rswitch: Add jumbo frames handling for TX

[Paolo Abeni]

On Mon, 2023-12-04 at 10:20 +0900, Yoshihiro Shimoda wrote:
> @@ -1647,25 +1676,32 @@ static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *nd
>  	if (skb_put_padto(skb, ETH_ZLEN))
>  		return ret;
>  
> -	dma_addr = dma_map_single(ndev->dev.parent, skb->data, skb->len, DMA_TO_DEVICE);
> -	if (dma_mapping_error(ndev->dev.parent, dma_addr))
> +	dma_addr_orig = dma_map_single(ndev->dev.parent, skb->data, skb->len, DMA_TO_DEVICE);
> +	if (dma_mapping_error(ndev->dev.parent, dma_addr_orig))
>  		goto err_kfree;
>  
>  	gq->skbs[gq->cur] = skb;
>  	gq->unmap_addrs[gq->cur] = dma_addr;

Here 'dma_addr' is still used uninitialized. I guess you should use
'dma_addr_orig' even here.

Cheers,

Paolo

RE: [PATCH net-next v3 8/9] net: rswitch: Add jumbo frames handling for TX

[Yoshihiro Shimoda]

Hello Paolo,

> From: Paolo Abeni, Sent: Wednesday, December 6, 2023 7:21 PM
> 
> On Mon, 2023-12-04 at 10:20 +0900, Yoshihiro Shimoda wrote:
> > @@ -1647,25 +1676,32 @@ static netdev_tx_t rswitch_start_xmit(struct sk_buff *skb, struct net_device *nd
> >  	if (skb_put_padto(skb, ETH_ZLEN))
> >  		return ret;
> >
> > -	dma_addr = dma_map_single(ndev->dev.parent, skb->data, skb->len, DMA_TO_DEVICE);
> > -	if (dma_mapping_error(ndev->dev.parent, dma_addr))
> > +	dma_addr_orig = dma_map_single(ndev->dev.parent, skb->data, skb->len, DMA_TO_DEVICE);
> > +	if (dma_mapping_error(ndev->dev.parent, dma_addr_orig))
> >  		goto err_kfree;
> >
> >  	gq->skbs[gq->cur] = skb;
> >  	gq->unmap_addrs[gq->cur] = dma_addr;
> 
> Here 'dma_addr' is still used uninitialized. I guess you should use
> 'dma_addr_orig' even here.

Oops! I should fix it. I'll make v4 patch. Thank you for your review!

Best regards,
Yoshihiro Shimoda

> Cheers,
> 
> Paolo

[PATCH net-next v3 9/9] net: rswitch: Allow jumbo frames

[Yoshihiro Shimoda]

Allow jumbo frames by changing maximum MTU size and number of RX queues.

Signed-off-by: Yoshihiro Shimoda <yoshihiro.shimoda.uh@renesas.com>
---
 drivers/net/ethernet/renesas/rswitch.c | 2 ++
 drivers/net/ethernet/renesas/rswitch.h | 3 ++-
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/renesas/rswitch.c b/drivers/net/ethernet/renesas/rswitch.c
index df419112b67a..9cdcac9c56b9 100644
--- a/drivers/net/ethernet/renesas/rswitch.c
+++ b/drivers/net/ethernet/renesas/rswitch.c
@@ -1882,6 +1882,8 @@ static int rswitch_device_alloc(struct rswitch_private *priv, unsigned int index
 	snprintf(ndev->name, IFNAMSIZ, "tsn%d", index);
 	ndev->netdev_ops = &rswitch_netdev_ops;
 	ndev->ethtool_ops = &rswitch_ethtool_ops;
+	ndev->max_mtu = RSWITCH_MAX_MTU;
+	ndev->min_mtu = ETH_MIN_MTU;
 
 	netif_napi_add(ndev, &rdev->napi, rswitch_poll);
 
diff --git a/drivers/net/ethernet/renesas/rswitch.h b/drivers/net/ethernet/renesas/rswitch.h
index 4252677e2a55..72e3ff596d31 100644
--- a/drivers/net/ethernet/renesas/rswitch.h
+++ b/drivers/net/ethernet/renesas/rswitch.h
@@ -26,9 +26,10 @@
 		else
 
 #define TX_RING_SIZE		1024
-#define RX_RING_SIZE		1024
+#define RX_RING_SIZE		4096
 #define TS_RING_SIZE		(TX_RING_SIZE * RSWITCH_NUM_PORTS)
 
+#define RSWITCH_MAX_MTU		9600
 #define RSWITCH_HEADROOM	(NET_SKB_PAD + NET_IP_ALIGN)
 #define RSWITCH_DESC_BUF_SIZE	2048
 #define RSWITCH_TAILROOM	SKB_DATA_ALIGN(sizeof(struct skb_shared_info))
-- 
2.34.1