net: enetc: add LSO support for i.MX95 ENETC PF

ENETC rev 4.1 supports large send offload (LSO), segmenting large TCP
and UDP transmit units into multiple Ethernet frames. To support LSO,
software needs to fill some auxiliary information in Tx BD, such as LSO
header length, frame length, LSO maximum segment size, etc.

At 1Gbps link rate, TCP segmentation was tested using iperf3, and the
CPU performance before and after applying the patch was compared through
the top command. It can be seen that LSO saves a significant amount of
CPU cycles compared to software TSO.

Before applying the patch:
%Cpu(s):  0.1 us,  4.1 sy,  0.0 ni, 85.7 id,  0.0 wa,  0.5 hi,  9.7 si

After applying the patch:
%Cpu(s):  0.1 us,  2.3 sy,  0.0 ni, 94.5 id,  0.0 wa,  0.4 hi,  2.6 si

Signed-off-by: Wei Fang <[email protected]>
Reviewed-by: Frank Li <[email protected]>
Reviewed-by: Claudiu Manoil <[email protected]>
Link: https://patch.msgid.link/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>

Author: Wei Fang

drivers/net/ethernet/freescale/enetc/enetc.c

@@ -532,6 +532,230 @@ static void enetc_tso_complete_csum(struct enetc_bdr *tx_ring, struct tso_t *tso
 	}
 }
 
+static int enetc_lso_count_descs(const struct sk_buff *skb)
+{
+	/* 4 BDs: 1 BD for LSO header + 1 BD for extended BD + 1 BD
+	 * for linear area data but not include LSO header, namely
+	 * skb_headlen(skb) - lso_hdr_len (it may be 0, but that's
+	 * okay, we only need to consider the worst case). And 1 BD
+	 * for gap.
+	 */
+	return skb_shinfo(skb)->nr_frags + 4;
+}
+
+static int enetc_lso_get_hdr_len(const struct sk_buff *skb)
+{
+	int hdr_len, tlen;
+
+	tlen = skb_is_gso_tcp(skb) ? tcp_hdrlen(skb) : sizeof(struct udphdr);
+	hdr_len = skb_transport_offset(skb) + tlen;
+
+	return hdr_len;
+}
+
+static void enetc_lso_start(struct sk_buff *skb, struct enetc_lso_t *lso)
+{
+	lso->lso_seg_size = skb_shinfo(skb)->gso_size;
+	lso->ipv6 = enetc_skb_is_ipv6(skb);
+	lso->tcp = skb_is_gso_tcp(skb);
+	lso->l3_hdr_len = skb_network_header_len(skb);
+	lso->l3_start = skb_network_offset(skb);
+	lso->hdr_len = enetc_lso_get_hdr_len(skb);
+	lso->total_len = skb->len - lso->hdr_len;
+}
+
+static void enetc_lso_map_hdr(struct enetc_bdr *tx_ring, struct sk_buff *skb,
+			      int *i, struct enetc_lso_t *lso)
+{
+	union enetc_tx_bd txbd_tmp, *txbd;
+	struct enetc_tx_swbd *tx_swbd;
+	u16 frm_len, frm_len_ext;
+	u8 flags, e_flags = 0;
+	dma_addr_t addr;
+	char *hdr;
+
+	/* Get the first BD of the LSO BDs chain */
+	txbd = ENETC_TXBD(*tx_ring, *i);
+	tx_swbd = &tx_ring->tx_swbd[*i];
+	prefetchw(txbd);
+
+	/* Prepare LSO header: MAC + IP + TCP/UDP */
+	hdr = tx_ring->tso_headers + *i * TSO_HEADER_SIZE;
+	memcpy(hdr, skb->data, lso->hdr_len);
+	addr = tx_ring->tso_headers_dma + *i * TSO_HEADER_SIZE;
+
+	/* {frm_len_ext, frm_len} indicates the total length of
+	 * large transmit data unit. frm_len contains the 16 least
+	 * significant bits and frm_len_ext contains the 4 most
+	 * significant bits.
+	 */
+	frm_len = lso->total_len & 0xffff;
+	frm_len_ext = (lso->total_len >> 16) & 0xf;
+
+	/* Set the flags of the first BD */
+	flags = ENETC_TXBD_FLAGS_EX | ENETC_TXBD_FLAGS_CSUM_LSO |
+		ENETC_TXBD_FLAGS_LSO | ENETC_TXBD_FLAGS_L4CS;
+
+	enetc_clear_tx_bd(&txbd_tmp);
+	txbd_tmp.addr = cpu_to_le64(addr);
+	txbd_tmp.hdr_len = cpu_to_le16(lso->hdr_len);
+
+	/* first BD needs frm_len and offload flags set */
+	txbd_tmp.frm_len = cpu_to_le16(frm_len);
+	txbd_tmp.flags = flags;
+
+	txbd_tmp.l3_aux0 = FIELD_PREP(ENETC_TX_BD_L3_START, lso->l3_start);
+	/* l3_hdr_size in 32-bits (4 bytes) */
+	txbd_tmp.l3_aux1 = FIELD_PREP(ENETC_TX_BD_L3_HDR_LEN,
+				      lso->l3_hdr_len / 4);
+	if (lso->ipv6)
+		txbd_tmp.l3_aux1 |= ENETC_TX_BD_L3T;
+	else
+		txbd_tmp.l3_aux0 |= ENETC_TX_BD_IPCS;
+
+	txbd_tmp.l4_aux = FIELD_PREP(ENETC_TX_BD_L4T, lso->tcp ?
+				     ENETC_TXBD_L4T_TCP : ENETC_TXBD_L4T_UDP);
+
+	/* For the LSO header we do not set the dma address since
+	 * we do not want it unmapped when we do cleanup. We still
+	 * set len so that we count the bytes sent.
+	 */
+	tx_swbd->len = lso->hdr_len;
+	tx_swbd->do_twostep_tstamp = false;
+	tx_swbd->check_wb = false;
+
+	/* Actually write the header in the BD */
+	*txbd = txbd_tmp;
+
+	/* Get the next BD, and the next BD is extended BD */
+	enetc_bdr_idx_inc(tx_ring, i);
+	txbd = ENETC_TXBD(*tx_ring, *i);
+	tx_swbd = &tx_ring->tx_swbd[*i];
+	prefetchw(txbd);
+
+	enetc_clear_tx_bd(&txbd_tmp);
+	if (skb_vlan_tag_present(skb)) {
+		/* Setup the VLAN fields */
+		txbd_tmp.ext.vid = cpu_to_le16(skb_vlan_tag_get(skb));
+		txbd_tmp.ext.tpid = ENETC_TPID_8021Q;
+		e_flags = ENETC_TXBD_E_FLAGS_VLAN_INS;
+	}
+
+	/* Write the BD */
+	txbd_tmp.ext.e_flags = e_flags;
+	txbd_tmp.ext.lso_sg_size = cpu_to_le16(lso->lso_seg_size);
+	txbd_tmp.ext.frm_len_ext = cpu_to_le16(frm_len_ext);
+	*txbd = txbd_tmp;
+}
+
+static int enetc_lso_map_data(struct enetc_bdr *tx_ring, struct sk_buff *skb,
+			      int *i, struct enetc_lso_t *lso, int *count)
+{
+	union enetc_tx_bd txbd_tmp, *txbd = NULL;
+	struct enetc_tx_swbd *tx_swbd;
+	skb_frag_t *frag;
+	dma_addr_t dma;
+	u8 flags = 0;
+	int len, f;
+
+	len = skb_headlen(skb) - lso->hdr_len;
+	if (len > 0) {
+		dma = dma_map_single(tx_ring->dev, skb->data + lso->hdr_len,
+				     len, DMA_TO_DEVICE);
+		if (dma_mapping_error(tx_ring->dev, dma))
+			return -ENOMEM;
+
+		enetc_bdr_idx_inc(tx_ring, i);
+		txbd = ENETC_TXBD(*tx_ring, *i);
+		tx_swbd = &tx_ring->tx_swbd[*i];
+		prefetchw(txbd);
+		*count += 1;
+
+		enetc_clear_tx_bd(&txbd_tmp);
+		txbd_tmp.addr = cpu_to_le64(dma);
+		txbd_tmp.buf_len = cpu_to_le16(len);
+
+		tx_swbd->dma = dma;
+		tx_swbd->len = len;
+		tx_swbd->is_dma_page = 0;
+		tx_swbd->dir = DMA_TO_DEVICE;
+	}
+
+	frag = &skb_shinfo(skb)->frags[0];
+	for (f = 0; f < skb_shinfo(skb)->nr_frags; f++, frag++) {
+		if (txbd)
+			*txbd = txbd_tmp;
+
+		len = skb_frag_size(frag);
+		dma = skb_frag_dma_map(tx_ring->dev, frag);
+		if (dma_mapping_error(tx_ring->dev, dma))
+			return -ENOMEM;
+
+		/* Get the next BD */
+		enetc_bdr_idx_inc(tx_ring, i);
+		txbd = ENETC_TXBD(*tx_ring, *i);
+		tx_swbd = &tx_ring->tx_swbd[*i];
+		prefetchw(txbd);
+		*count += 1;
+
+		enetc_clear_tx_bd(&txbd_tmp);
+		txbd_tmp.addr = cpu_to_le64(dma);
+		txbd_tmp.buf_len = cpu_to_le16(len);
+
+		tx_swbd->dma = dma;
+		tx_swbd->len = len;
+		tx_swbd->is_dma_page = 1;
+		tx_swbd->dir = DMA_TO_DEVICE;
+	}
+
+	/* Last BD needs 'F' bit set */
+	flags |= ENETC_TXBD_FLAGS_F;
+	txbd_tmp.flags = flags;
+	*txbd = txbd_tmp;
+
+	tx_swbd->is_eof = 1;
+	tx_swbd->skb = skb;
+
+	return 0;
+}
+
+static int enetc_lso_hw_offload(struct enetc_bdr *tx_ring, struct sk_buff *skb)
+{
+	struct enetc_tx_swbd *tx_swbd;
+	struct enetc_lso_t lso = {0};
+	int err, i, count = 0;
+
+	/* Initialize the LSO handler */
+	enetc_lso_start(skb, &lso);
+	i = tx_ring->next_to_use;
+
+	enetc_lso_map_hdr(tx_ring, skb, &i, &lso);
+	/* First BD and an extend BD */
+	count += 2;
+
+	err = enetc_lso_map_data(tx_ring, skb, &i, &lso, &count);
+	if (err)
+		goto dma_err;
+
+	/* Go to the next BD */
+	enetc_bdr_idx_inc(tx_ring, &i);
+	tx_ring->next_to_use = i;
+	enetc_update_tx_ring_tail(tx_ring);
+
+	return count;
+
+dma_err:
+	do {
+		tx_swbd = &tx_ring->tx_swbd[i];
+		enetc_free_tx_frame(tx_ring, tx_swbd);
+		if (i == 0)
+			i = tx_ring->bd_count;
+		i--;
+	} while (--count);
+
+	return 0;
+}
+
 static int enetc_map_tx_tso_buffs(struct enetc_bdr *tx_ring, struct sk_buff *skb)
 {
 	struct enetc_ndev_priv *priv = netdev_priv(tx_ring->ndev);
@@ -652,14 +876,26 @@ static netdev_tx_t enetc_start_xmit(struct sk_buff *skb,
 	tx_ring = priv->tx_ring[skb->queue_mapping];
 
 	if (skb_is_gso(skb)) {
-		if (enetc_bd_unused(tx_ring) < tso_count_descs(skb)) {
-			netif_stop_subqueue(ndev, tx_ring->index);
-			return NETDEV_TX_BUSY;
-		}
+		/* LSO data unit lengths of up to 256KB are supported */
+		if (priv->active_offloads & ENETC_F_LSO &&
+		    (skb->len - enetc_lso_get_hdr_len(skb)) <=
+		    ENETC_LSO_MAX_DATA_LEN) {
+			if (enetc_bd_unused(tx_ring) < enetc_lso_count_descs(skb)) {
+				netif_stop_subqueue(ndev, tx_ring->index);
+				return NETDEV_TX_BUSY;
+			}
 
-		enetc_lock_mdio();
-		count = enetc_map_tx_tso_buffs(tx_ring, skb);
-		enetc_unlock_mdio();
+			count = enetc_lso_hw_offload(tx_ring, skb);
+		} else {
+			if (enetc_bd_unused(tx_ring) < tso_count_descs(skb)) {
+				netif_stop_subqueue(ndev, tx_ring->index);
+				return NETDEV_TX_BUSY;
+			}
+
+			enetc_lock_mdio();
+			count = enetc_map_tx_tso_buffs(tx_ring, skb);
+			enetc_unlock_mdio();
+		}
 	} else {
 		if (unlikely(skb_shinfo(skb)->nr_frags > priv->max_frags))
 			if (unlikely(skb_linearize(skb)))
@@ -1799,6 +2035,9 @@ void enetc_get_si_caps(struct enetc_si *si)
 		rss = enetc_rd(hw, ENETC_SIRSSCAPR);
 		si->num_rss = ENETC_SIRSSCAPR_GET_NUM_RSS(rss);
 	}
+
+	if (val & ENETC_SIPCAPR0_LSO)
+		si->hw_features |= ENETC_SI_F_LSO;
 }
 EXPORT_SYMBOL_GPL(enetc_get_si_caps);
 
@@ -2095,6 +2334,14 @@ static int enetc_setup_default_rss_table(struct enetc_si *si, int num_groups)
 	return 0;
 }
 
+static void enetc_set_lso_flags_mask(struct enetc_hw *hw)
+{
+	enetc_wr(hw, ENETC4_SILSOSFMR0,
+		 SILSOSFMR0_VAL_SET(ENETC4_TCP_NL_SEG_FLAGS_DMASK,
+				    ENETC4_TCP_NL_SEG_FLAGS_DMASK));
+	enetc_wr(hw, ENETC4_SILSOSFMR1, 0);
+}
+
 int enetc_configure_si(struct enetc_ndev_priv *priv)
 {
 	struct enetc_si *si = priv->si;
@@ -2108,6 +2355,9 @@ int enetc_configure_si(struct enetc_ndev_priv *priv)
 	/* enable SI */
 	enetc_wr(hw, ENETC_SIMR, ENETC_SIMR_EN);
 
+	if (si->hw_features & ENETC_SI_F_LSO)
+		enetc_set_lso_flags_mask(hw);
+
 	/* TODO: RSS support for i.MX95 will be supported later, and the
 	 * is_enetc_rev1() condition will be removed
 	 */

drivers/net/ethernet/freescale/enetc/enetc.h

@@ -41,6 +41,18 @@ struct enetc_tx_swbd {
 	u8 qbv_en:1;
 };
 
+struct enetc_lso_t {
+	bool	ipv6;
+	bool	tcp;
+	u8	l3_hdr_len;
+	u8	hdr_len; /* LSO header length */
+	u8	l3_start;
+	u16	lso_seg_size;
+	int	total_len; /* total data length, not include LSO header */
+};
+
+#define ENETC_LSO_MAX_DATA_LEN		SZ_256K
+
 #define ENETC_RX_MAXFRM_SIZE	ENETC_MAC_MAXFRM_SIZE
 #define ENETC_RXB_TRUESIZE	2048 /* PAGE_SIZE >> 1 */
 #define ENETC_RXB_PAD		NET_SKB_PAD /* add extra space if needed */
@@ -238,6 +250,7 @@ enum enetc_errata {
 #define ENETC_SI_F_PSFP BIT(0)
 #define ENETC_SI_F_QBV  BIT(1)
 #define ENETC_SI_F_QBU  BIT(2)
+#define ENETC_SI_F_LSO	BIT(3)
 
 struct enetc_drvdata {
 	u32 pmac_offset; /* Only valid for PSI which supports 802.1Qbu */
@@ -351,6 +364,7 @@ enum enetc_active_offloads {
 	ENETC_F_QCI			= BIT(10),
 	ENETC_F_QBU			= BIT(11),
 	ENETC_F_TXCSUM			= BIT(12),
+	ENETC_F_LSO			= BIT(13),
 };
 
 enum enetc_flags_bit {

drivers/net/ethernet/freescale/enetc/enetc4_hw.h

@@ -12,6 +12,29 @@
 #define NXP_ENETC_VENDOR_ID		0x1131
 #define NXP_ENETC_PF_DEV_ID		0xe101
 
+/**********************Station interface registers************************/
+/* Station interface LSO segmentation flag mask register 0/1 */
+#define ENETC4_SILSOSFMR0		0x1300
+#define  SILSOSFMR0_TCP_MID_SEG		GENMASK(27, 16)
+#define  SILSOSFMR0_TCP_1ST_SEG		GENMASK(11, 0)
+#define  SILSOSFMR0_VAL_SET(first, mid)	(FIELD_PREP(SILSOSFMR0_TCP_MID_SEG, mid) | \
+					 FIELD_PREP(SILSOSFMR0_TCP_1ST_SEG, first))
+
+#define ENETC4_SILSOSFMR1		0x1304
+#define  SILSOSFMR1_TCP_LAST_SEG	GENMASK(11, 0)
+#define   ENETC4_TCP_FLAGS_FIN		BIT(0)
+#define   ENETC4_TCP_FLAGS_SYN		BIT(1)
+#define   ENETC4_TCP_FLAGS_RST		BIT(2)
+#define   ENETC4_TCP_FLAGS_PSH		BIT(3)
+#define   ENETC4_TCP_FLAGS_ACK		BIT(4)
+#define   ENETC4_TCP_FLAGS_URG		BIT(5)
+#define   ENETC4_TCP_FLAGS_ECE		BIT(6)
+#define   ENETC4_TCP_FLAGS_CWR		BIT(7)
+#define   ENETC4_TCP_FLAGS_NS		BIT(8)
+/* According to tso_build_hdr(), clear all special flags for not last packet. */
+#define ENETC4_TCP_NL_SEG_FLAGS_DMASK	(ENETC4_TCP_FLAGS_FIN | \
+					 ENETC4_TCP_FLAGS_RST | ENETC4_TCP_FLAGS_PSH)
+
 /***************************ENETC port registers**************************/
 #define ENETC4_ECAPR0			0x0
 #define  ECAPR0_RFS			BIT(2)