net/octeontx: support multi segment

Adding multi segment support to the octeontx PMD. Also adding the logic to share rx/tx ofloads with the eventdev code. Signed-off-by: Harman Kalra <hkalra@marvell.com>
2020-03-16 15:03:37 +05:30 · 2020-03-16 15:03:37 +05:30 · 85221a0c7c
commit 85221a0c7c
parent 41fe7a3a11
9 changed files with 250 additions and 28 deletions
--- a/doc/guides/nics/features/octeontx.ini
+++ b/doc/guides/nics/features/octeontx.ini
@ -9,6 +9,7 @@ Link status          = Y
 Lock-free Tx queue   = Y
 Queue start/stop     = P
 Jumbo frame          = Y
+Scattered Rx         = Y
 Promiscuous mode     = Y
 Unicast MAC filter   = Y
 CRC offload          = Y
--- a/doc/guides/nics/octeontx.rst
+++ b/doc/guides/nics/octeontx.rst
@ -20,6 +20,7 @@ Features of the OCTEON TX Ethdev PMD are:
 - Promiscuous mode
 - Port hardware statistics
 - Jumbo frames
+- Scatter-Gather IO support
 - Link state information
 - SR-IOV VF
 - Multiple queues for TX
--- a/drivers/event/octeontx/ssovf_worker.c
+++ b/drivers/event/octeontx/ssovf_worker.c
@ -300,7 +300,7 @@ sso_event_tx_adapter_enqueue(void *port,
 	dq = &txq->dq;

 	if (__octeontx_xmit_pkts(dq->lmtline_va, dq->ioreg_va, dq->fc_status_va,
-				m) < 0)
+				m, OCCTX_TX_OFFLOAD_NONE) < 0)
 		return 0;

 	return 1;
--- a/drivers/net/octeontx/base/octeontx_pki_var.h
+++ b/drivers/net/octeontx/base/octeontx_pki_var.h
@ -215,4 +215,36 @@ enum lf_type_e {
 	LF_UDP_VXLAN	= OCCTX_PKI_LTYPE_UDP_VXLAN,
 	LF_NVGRE	= OCCTX_PKI_LTYPE_NVGRE,
 };
+
+/* Word 0 of HW segment buflink structure */
+typedef union octtx_pki_buflink_w0_u {
+	uint64_t v;
+	struct {
+		uint64_t        size:16;
+		uint64_t        rsvd1:15;
+		uint64_t        invfree:1;
+		/** Aura number of the next segment */
+		uint64_t        aura:16;
+		uint64_t        sw:9;
+		uint64_t        later_invfree:1;
+		uint64_t        rsvd2:5;
+		/** 1 if aura number is set */
+		uint64_t        has_aura:1;
+	} s;
+} octtx_pki_buflink_w0_t;
+
+/* Word 1 of HW segment buflink structure */
+typedef union octtx_pki_buflink_w1_u {
+	uint64_t v;
+	struct {
+		uint64_t        addr;
+	} s;
+} octtx_pki_buflink_w1_t;
+
+/* HW structure linking packet segments into singly linked list */
+typedef struct octtx_pki_buflink_s {
+	octtx_pki_buflink_w0_t    w0; /* Word 0 of the buflink */
+	octtx_pki_buflink_w1_t    w1; /* Word 1 of the buflink */
+} octtx_pki_buflink_t;
+
 #endif /* __OCTEONTX_PKI_VAR_H__ */
--- a/drivers/net/octeontx/base/octeontx_pkovf.h
+++ b/drivers/net/octeontx/base/octeontx_pkovf.h
@ -36,7 +36,10 @@

 /* pko_send_hdr_s + pko_send_link */
 #define PKO_CMD_SZ			(2 << 1)
-#define PKO_SEND_GATHER_SUBDC		(0x0ull << 60)
+#define PKO_SEND_BUFLINK_SUBDC		(0x0ull << 60)
+#define PKO_SEND_BUFLINK_LDTYPE(x)	((x) << 58)
+#define PKO_SEND_BUFLINK_GAUAR(x)	((x) << 24)
+#define PKO_SEND_GATHER_SUBDC		(0x2ull << 60)
 #define PKO_SEND_GATHER_LDTYPE(x)	((x) << 58)
 #define PKO_SEND_GATHER_GAUAR(x)	((x) << 24)

--- a/drivers/net/octeontx/octeontx_ethdev.c
+++ b/drivers/net/octeontx/octeontx_ethdev.c
@ -24,6 +24,10 @@
 #include "octeontx_rxtx.h"
 #include "octeontx_logs.h"

+struct evdev_priv_data {
+	OFFLOAD_FLAGS; /*Sequence should not be changed */
+} __rte_cache_aligned;
+
 struct octeontx_vdev_init_params {
 	uint8_t	nr_port;
 };
@ -257,6 +261,43 @@ devconf_set_default_sane_values(struct rte_event_dev_config *dev_conf,
 			info->max_num_events;
 }

+static uint16_t
+octeontx_tx_offload_flags(struct rte_eth_dev *eth_dev)
+{
+	struct octeontx_nic *nic = octeontx_pmd_priv(eth_dev);
+	uint16_t flags = 0;
+
+	/* Created function for supoorting future offloads */
+	if (nic->tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS)
+		flags |= OCCTX_TX_MULTI_SEG_F;
+
+	return flags;
+}
+
+static uint16_t
+octeontx_rx_offload_flags(struct rte_eth_dev *eth_dev)
+{
+	struct octeontx_nic *nic = octeontx_pmd_priv(eth_dev);
+	struct rte_eth_dev_data *data = eth_dev->data;
+	struct rte_eth_conf *conf = &data->dev_conf;
+	struct rte_eth_rxmode *rxmode = &conf->rxmode;
+	uint16_t flags = 0;
+
+	if (rxmode->mq_mode == ETH_MQ_RX_RSS)
+		flags |= OCCTX_RX_OFFLOAD_RSS_F;
+
+	if (nic->rx_offloads & DEV_RX_OFFLOAD_SCATTER) {
+		flags |= OCCTX_RX_MULTI_SEG_F;
+		eth_dev->data->scattered_rx = 1;
+		/* If scatter mode is enabled, TX should also be in multi
+		 * seg mode, else memory leak will occur
+		 */
+		nic->tx_offloads |= DEV_TX_OFFLOAD_MULTI_SEGS;
+	}
+
+	return flags;
+}
+
 static int
 octeontx_dev_configure(struct rte_eth_dev *dev)
 {
@ -321,6 +362,11 @@ octeontx_dev_configure(struct rte_eth_dev *dev)
 	nic->pki.hash_enable = true;
 	nic->pki.initialized = false;

+	nic->rx_offloads |= rxmode->offloads;
+	nic->tx_offloads |= txmode->offloads;
+	nic->rx_offload_flags |= octeontx_rx_offload_flags(dev);
+	nic->tx_offload_flags |= octeontx_tx_offload_flags(dev);
+
 	return 0;
 }

@ -359,6 +405,51 @@ octeontx_dev_close(struct rte_eth_dev *dev)
 	dev->rx_pkt_burst = NULL;
 }

+static int
+octeontx_recheck_rx_offloads(struct octeontx_rxq *rxq)
+{
+	struct rte_eth_dev *eth_dev = rxq->eth_dev;
+	struct octeontx_nic *nic = octeontx_pmd_priv(eth_dev);
+	struct rte_eth_dev_data *data = eth_dev->data;
+	struct rte_pktmbuf_pool_private *mbp_priv;
+	struct evdev_priv_data *evdev_priv;
+	struct rte_eventdev *dev;
+	uint32_t buffsz;
+
+	/* Get rx buffer size */
+	mbp_priv = rte_mempool_get_priv(rxq->pool);
+	buffsz = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
+
+	/* Setup scatter mode if needed by jumbo */
+	if (data->dev_conf.rxmode.max_rx_pkt_len > buffsz) {
+		nic->rx_offloads |= DEV_RX_OFFLOAD_SCATTER;
+		nic->rx_offload_flags |= octeontx_rx_offload_flags(eth_dev);
+		nic->tx_offload_flags |= octeontx_tx_offload_flags(eth_dev);
+	}
+
+	/* Sharing offload flags via eventdev priv region */
+	dev = &rte_eventdevs[rxq->evdev];
+	evdev_priv = dev->data->dev_private;
+	evdev_priv->rx_offload_flags = nic->rx_offload_flags;
+	evdev_priv->tx_offload_flags = nic->tx_offload_flags;
+
+	return 0;
+}
+
+static void
+octeontx_set_tx_function(struct rte_eth_dev *dev)
+{
+	struct octeontx_nic *nic = octeontx_pmd_priv(dev);
+
+	const eth_tx_burst_t tx_burst_func[2] = {
+		[0] = octeontx_xmit_pkts,
+		[1] = octeontx_xmit_pkts_mseg,
+	};
+
+	dev->tx_pkt_burst =
+		tx_burst_func[!!(nic->tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS)];
+}
+
 static int
 octeontx_dev_start(struct rte_eth_dev *dev)
 {
@ -371,7 +462,7 @@ octeontx_dev_start(struct rte_eth_dev *dev)
 	/*
 	 * Tx start
 	 */
-	dev->tx_pkt_burst = octeontx_xmit_pkts;
+	octeontx_set_tx_function(dev);
 	ret = octeontx_pko_channel_start(nic->base_ochan);
 	if (ret < 0) {
 		octeontx_log_err("fail to conf VF%d no. txq %d chan %d ret %d",
@ -599,10 +690,8 @@ octeontx_dev_default_mac_addr_set(struct rte_eth_dev *dev,
 					struct rte_ether_addr *addr)
 {
 	struct octeontx_nic *nic = octeontx_pmd_priv(dev);
-	uint8_t prom_mode = dev->data->promiscuous;
 	int ret;

-	dev->data->promiscuous = 0;
 	ret = octeontx_bgx_port_mac_set(nic->port_id, addr->addr_bytes);
 	if (ret == 0) {
 		/* Update same mac address to BGX CAM table */
@ -610,7 +699,6 @@ octeontx_dev_default_mac_addr_set(struct rte_eth_dev *dev,
 						0);
 	}
 	if (ret < 0) {
-		dev->data->promiscuous = prom_mode;
 		octeontx_log_err("failed to set MAC address on port %d",
 				 nic->port_id);
 	}
@ -977,7 +1065,9 @@ octeontx_dev_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qidx,
 	rxq->evdev = nic->evdev;
 	rxq->ev_queues = ev_queues;
 	rxq->ev_ports = ev_ports;
+	rxq->pool = mb_pool;

+	octeontx_recheck_rx_offloads(rxq);
 	dev->data->rx_queues[qidx] = rxq;
 	dev->data->rx_queue_state[qidx] = RTE_ETH_QUEUE_STATE_STOPPED;
 	return 0;
--- a/drivers/net/octeontx/octeontx_ethdev.h
+++ b/drivers/net/octeontx/octeontx_ethdev.h
@ -29,8 +29,12 @@
 #define OCTEONTX_MAX_BGX_PORTS			4
 #define OCTEONTX_MAX_LMAC_PER_BGX		4

-#define OCTEONTX_RX_OFFLOADS			DEV_RX_OFFLOAD_CHECKSUM
-#define OCTEONTX_TX_OFFLOADS			DEV_TX_OFFLOAD_MT_LOCKFREE
+#define OCTEONTX_RX_OFFLOADS			(DEV_RX_OFFLOAD_CHECKSUM     | \
+						 DEV_RX_OFFLOAD_SCATTER	     | \
+						 DEV_RX_OFFLOAD_JUMBO_FRAME)
+
+#define OCTEONTX_TX_OFFLOADS			(DEV_TX_OFFLOAD_MT_LOCKFREE  | \
+						 DEV_TX_OFFLOAD_MULTI_SEGS)

 static inline struct octeontx_nic *
 octeontx_pmd_priv(struct rte_eth_dev *dev)
@ -73,6 +77,10 @@ struct octeontx_nic {

 	uint16_t ev_queues;
 	uint16_t ev_ports;
+	uint64_t rx_offloads;
+	uint16_t rx_offload_flags;
+	uint64_t tx_offloads;
+	uint16_t tx_offload_flags;
 } __rte_cache_aligned;

 struct octeontx_txq {
@ -88,6 +96,7 @@ struct octeontx_rxq {
 	struct rte_eth_dev *eth_dev;
 	uint16_t ev_queues;
 	uint16_t ev_ports;
+	struct rte_mempool *pool;
 } __rte_cache_aligned;

 #endif /* __OCTEONTX_ETHDEV_H__ */
--- a/drivers/net/octeontx/octeontx_rxtx.c
+++ b/drivers/net/octeontx/octeontx_rxtx.c
@ -32,8 +32,34 @@ octeontx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	rte_cio_wmb();
 	while (count < nb_pkts) {
 		res = __octeontx_xmit_pkts(dq->lmtline_va, dq->ioreg_va,
-					   dq->fc_status_va,
-					   tx_pkts[count]);
+					   dq->fc_status_va, tx_pkts[count],
+					   OCCTX_TX_OFFLOAD_NONE);
+		if (res < 0)
+			break;
+
+		count++;
+	}
+
+	return count; /* return number of pkts transmitted */
+}
+
+uint16_t __hot
+octeontx_xmit_pkts_mseg(void *tx_queue, struct rte_mbuf **tx_pkts,
+			uint16_t nb_pkts)
+{
+	int count;
+	struct octeontx_txq *txq = tx_queue;
+	octeontx_dq_t *dq = &txq->dq;
+	int res;
+
+	count = 0;
+
+	rte_cio_wmb();
+	while (count < nb_pkts) {
+		res = __octeontx_xmit_pkts(dq->lmtline_va, dq->ioreg_va,
+					   dq->fc_status_va, tx_pkts[count],
+					   OCCTX_TX_OFFLOAD_NONE |
+					   OCCTX_TX_MULTI_SEG_F);
 		if (res < 0)
 			break;

--- a/drivers/net/octeontx/octeontx_rxtx.h
+++ b/drivers/net/octeontx/octeontx_rxtx.h
@ -7,6 +7,19 @@

 #include <rte_ethdev_driver.h>

+#define OFFLOAD_FLAGS					\
+	uint16_t rx_offload_flags;			\
+	uint16_t tx_offload_flags
+
+#define BIT(nr) (1UL << (nr))
+
+#define OCCTX_RX_OFFLOAD_NONE		(0)
+#define OCCTX_RX_OFFLOAD_RSS_F          BIT(0)
+#define OCCTX_RX_MULTI_SEG_F		BIT(15)
+
+#define OCCTX_TX_OFFLOAD_NONE		(0)
+
+#define OCCTX_TX_MULTI_SEG_F		BIT(15)
 /* Packet type table */
 #define PTYPE_SIZE	OCCTX_PKI_LTYPE_LAST

@ -98,33 +111,76 @@ ptype_table[PTYPE_SIZE][PTYPE_SIZE][PTYPE_SIZE] = {

 static __rte_always_inline int
 __octeontx_xmit_pkts(void *lmtline_va, void *ioreg_va, int64_t *fc_status_va,
-			struct rte_mbuf *tx_pkt)
+			struct rte_mbuf *tx_pkt, const uint16_t flag)
 {
-	uint64_t cmd_buf[4] __rte_cache_aligned;
-	uint16_t gaura_id;
+	uint8_t sz = (4 + (!!(flag & OCCTX_TX_MULTI_SEG_F) * 10));
+	/* Max size of PKO SEND desc is 112 bytes*/
+	uint64_t cmd_buf[sz] __rte_cache_aligned;
+	uint8_t nb_segs, nb_desc = 0;
+	uint16_t gaura_id, len = 0;
+	struct rte_mbuf *m_next = NULL;

 	if (unlikely(*((volatile int64_t *)fc_status_va) < 0))
 		return -ENOSPC;

-	/* Get the gaura Id */
-	gaura_id = octeontx_fpa_bufpool_gpool((uintptr_t)tx_pkt->pool->pool_id);

-	/* Setup PKO_SEND_HDR_S */
-	cmd_buf[0] = tx_pkt->data_len & 0xffff;
-	cmd_buf[1] = 0x0;
+	if (flag & OCCTX_TX_MULTI_SEG_F) {
+		nb_segs = tx_pkt->nb_segs;
+		/* Setup PKO_SEND_HDR_S */
+		cmd_buf[nb_desc++] = tx_pkt->pkt_len & 0xffff;
+		cmd_buf[nb_desc++] = 0x0;

-	/* Set don't free bit if reference count > 1 */
-	if (rte_mbuf_refcnt_read(tx_pkt) > 1)
-		cmd_buf[0] |= (1ULL << 58); /* SET DF */
+		do {
+			m_next = tx_pkt->next;
+			/* To handle case where mbufs belong to diff pools, like
+			 * fragmentation
+			 */
+			gaura_id = octeontx_fpa_bufpool_gpool((uintptr_t)
+							tx_pkt->pool->pool_id);

-	/* Setup PKO_SEND_GATHER_S */
-	cmd_buf[(1 << 1) | 1] = rte_mbuf_data_iova(tx_pkt);
-	cmd_buf[(1 << 1) | 0] = PKO_SEND_GATHER_SUBDC |
-				PKO_SEND_GATHER_LDTYPE(0x1ull) |
-				PKO_SEND_GATHER_GAUAR((long)gaura_id) |
-				tx_pkt->data_len;
+			/* Setup PKO_SEND_GATHER_S */
+			cmd_buf[nb_desc] = PKO_SEND_GATHER_SUBDC           |
+					     PKO_SEND_GATHER_LDTYPE(0x1ull)  |
+					     PKO_SEND_GATHER_GAUAR((long)
+								   gaura_id) |
+					     tx_pkt->data_len;
+			/* Mark mempool object as "put" since it is freed by
+			 * PKO.
+			 */
+			if (!(cmd_buf[nb_desc] & (1ULL << 57))) {
+				tx_pkt->next = NULL;
+				__mempool_check_cookies(tx_pkt->pool,
+							(void **)&tx_pkt, 1, 0);
+			}
+			nb_desc++;

-	octeontx_reg_lmtst(lmtline_va, ioreg_va, cmd_buf, PKO_CMD_SZ);
+			cmd_buf[nb_desc++] = rte_mbuf_data_iova(tx_pkt);
+
+			nb_segs--;
+			len += tx_pkt->data_len;
+			tx_pkt = m_next;
+		} while (nb_segs);
+	} else {
+		/* Setup PKO_SEND_HDR_S */
+		cmd_buf[nb_desc++] = tx_pkt->data_len & 0xffff;
+		cmd_buf[nb_desc++] = 0x0;
+
+		/* Mark mempool object as "put" since it is freed by PKO */
+		if (!(cmd_buf[0] & (1ULL << 58)))
+			__mempool_check_cookies(tx_pkt->pool, (void **)&tx_pkt,
+						1, 0);
+		/* Get the gaura Id */
+		gaura_id = octeontx_fpa_bufpool_gpool((uintptr_t)
+						      tx_pkt->pool->pool_id);
+
+		/* Setup PKO_SEND_BUFLINK_S */
+		cmd_buf[nb_desc++] = PKO_SEND_BUFLINK_SUBDC |
+				     PKO_SEND_BUFLINK_LDTYPE(0x1ull) |
+				     PKO_SEND_BUFLINK_GAUAR((long)gaura_id) |
+				     tx_pkt->data_len;
+		cmd_buf[nb_desc++] = rte_mbuf_data_iova(tx_pkt);
+	}
+	octeontx_reg_lmtst(lmtline_va, ioreg_va, cmd_buf, nb_desc);

 	return 0;
 }
@ -132,6 +188,10 @@ __octeontx_xmit_pkts(void *lmtline_va, void *ioreg_va, int64_t *fc_status_va,
 uint16_t
 octeontx_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts);

+uint16_t
+octeontx_xmit_pkts_mseg(void *tx_queue, struct rte_mbuf **tx_pkts,
+			uint16_t nb_pkts);
+
 uint16_t
 octeontx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts);