[dpdk-dev] [PATCH 25/31] net/sfc: add basic stubs for RSS support on driver attach

[Andrew Rybchenko <arybchenko@solarflare.com>]

From: Ivan Malov <ivan.malov@oktetlabs.ru>

Reviewed-by: Andrew Lee <alee@solarflare.com>
Reviewed-by: Robert Stonehouse <rstonehouse@solarflare.com>
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
---
 doc/guides/nics/sfc_efx.rst  |  2 ++
 drivers/net/sfc/efsys.h      |  2 +-
 drivers/net/sfc/sfc.c        | 76 +++++++++++++++++++++++++++++++++++++++++
 drivers/net/sfc/sfc.h        | 17 ++++++++++
 drivers/net/sfc/sfc_ethdev.c |  8 +++++
 drivers/net/sfc/sfc_rx.c     | 81 +++++++++++++++++++++++++++++++++++++++++++-
 drivers/net/sfc/sfc_rx.h     |  8 +++++
 7 files changed, 192 insertions(+), 2 deletions(-)

diff --git a/doc/guides/nics/sfc_efx.rst b/doc/guides/nics/sfc_efx.rst
index ee2ba27..4f674c0 100644
--- a/doc/guides/nics/sfc_efx.rst
+++ b/doc/guides/nics/sfc_efx.rst
@@ -66,6 +66,8 @@ SFC EFX PMD has support for:
 
 - Received packet type information
 
+- Receive side scaling (RSS)
+
 - Scattered Rx DMA for packet that are larger that a single Rx descriptor
 
 - Deferred receive and transmit queue start
diff --git a/drivers/net/sfc/efsys.h b/drivers/net/sfc/efsys.h
index 0f941e6..fb2f3b5 100644
--- a/drivers/net/sfc/efsys.h
+++ b/drivers/net/sfc/efsys.h
@@ -195,7 +195,7 @@ prefetch_read_once(const volatile void *addr)
 #define EFSYS_OPT_BOOTCFG 0
 
 #define EFSYS_OPT_DIAG 0
-#define EFSYS_OPT_RX_SCALE 0
+#define EFSYS_OPT_RX_SCALE 1
 #define EFSYS_OPT_QSTATS 0
 /* Filters support is required for SFN7xxx and SFN8xx */
 #define EFSYS_OPT_FILTER 1
diff --git a/drivers/net/sfc/sfc.c b/drivers/net/sfc/sfc.c
index e2e6c9e..e79367d 100644
--- a/drivers/net/sfc/sfc.c
+++ b/drivers/net/sfc/sfc.c
@@ -484,6 +484,73 @@ sfc_mem_bar_fini(struct sfc_adapter *sa)
 	memset(ebp, 0, sizeof(*ebp));
 }
 
+#if EFSYS_OPT_RX_SCALE
+/*
+ * A fixed RSS key which has a property of being symmetric
+ * (symmetrical flows are distributed to the same CPU)
+ * and also known to give a uniform distribution
+ * (a good distribution of traffic between different CPUs)
+ */
+static const uint8_t default_rss_key[SFC_RSS_KEY_SIZE] = {
+	0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a,
+	0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a,
+	0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a,
+	0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a,
+	0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a, 0x6d, 0x5a,
+};
+#endif
+
+static int
+sfc_set_rss_defaults(struct sfc_adapter *sa)
+{
+#if EFSYS_OPT_RX_SCALE
+	int rc;
+
+	rc = efx_intr_init(sa->nic, sa->intr.type, NULL);
+	if (rc != 0)
+		goto fail_intr_init;
+
+	rc = efx_ev_init(sa->nic);
+	if (rc != 0)
+		goto fail_ev_init;
+
+	rc = efx_rx_init(sa->nic);
+	if (rc != 0)
+		goto fail_rx_init;
+
+	rc = efx_rx_scale_support_get(sa->nic, &sa->rss_support);
+	if (rc != 0)
+		goto fail_scale_support_get;
+
+	rc = efx_rx_hash_support_get(sa->nic, &sa->hash_support);
+	if (rc != 0)
+		goto fail_hash_support_get;
+
+	efx_rx_fini(sa->nic);
+	efx_ev_fini(sa->nic);
+	efx_intr_fini(sa->nic);
+
+	sa->rss_hash_types = sfc_rte_to_efx_hash_type(SFC_RSS_OFFLOADS);
+
+	rte_memcpy(sa->rss_key, default_rss_key, sizeof(sa->rss_key));
+
+	return 0;
+
+fail_hash_support_get:
+fail_scale_support_get:
+fail_rx_init:
+	efx_ev_fini(sa->nic);
+
+fail_ev_init:
+	efx_intr_fini(sa->nic);
+
+fail_intr_init:
+	return rc;
+#else
+	return 0;
+#endif
+}
+
 int
 sfc_attach(struct sfc_adapter *sa)
 {
@@ -550,6 +617,10 @@ sfc_attach(struct sfc_adapter *sa)
 	efx_phy_adv_cap_get(sa->nic, EFX_PHY_CAP_PERM,
 			    &sa->port.phy_adv_cap_mask);
 
+	rc = sfc_set_rss_defaults(sa);
+	if (rc != 0)
+		goto fail_set_rss_defaults;
+
 	sfc_log_init(sa, "fini nic");
 	efx_nic_fini(enp);
 
@@ -558,7 +629,12 @@ sfc_attach(struct sfc_adapter *sa)
 	sfc_log_init(sa, "done");
 	return 0;
 
+fail_set_rss_defaults:
+	sfc_intr_detach(sa);
+
 fail_intr_attach:
+	efx_nic_fini(sa->nic);
+
 fail_estimate_rsrc_limits:
 fail_nic_reset:
 	sfc_log_init(sa, "unprobe nic");
diff --git a/drivers/net/sfc/sfc.h b/drivers/net/sfc/sfc.h
index f0bbaf8..01dbfb6 100644
--- a/drivers/net/sfc/sfc.h
+++ b/drivers/net/sfc/sfc.h
@@ -42,6 +42,13 @@
 extern "C" {
 #endif
 
+#if EFSYS_OPT_RX_SCALE
+/** RSS key length (bytes) */
+#define	SFC_RSS_KEY_SIZE	40
+/** RSS hash offloads mask */
+#define	SFC_RSS_OFFLOADS	(ETH_RSS_IP | ETH_RSS_TCP)
+#endif
+
 /*
  * +---------------+
  * | UNINITIALIZED |<-----------+
@@ -187,6 +194,16 @@ struct sfc_adapter {
 
 	unsigned int			txq_count;
 	struct sfc_txq_info		*txq_info;
+
+	unsigned int			rss_channels;
+
+#if EFSYS_OPT_RX_SCALE
+	efx_rx_scale_support_t		rss_support;
+	efx_rx_hash_support_t		hash_support;
+	unsigned int			rss_hash_types;
+	unsigned int			rss_tbl[EFX_RSS_TBL_SIZE];
+	uint8_t				rss_key[SFC_RSS_KEY_SIZE];
+#endif
 };
 
 /*
diff --git a/drivers/net/sfc/sfc_ethdev.c b/drivers/net/sfc/sfc_ethdev.c
index 0de17ca..b17607f 100644
--- a/drivers/net/sfc/sfc_ethdev.c
+++ b/drivers/net/sfc/sfc_ethdev.c
@@ -85,6 +85,14 @@ sfc_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 	else
 		dev_info->tx_offload_capa |= DEV_TX_OFFLOAD_VLAN_INSERT;
 
+#if EFSYS_OPT_RX_SCALE
+	if (sa->rss_support != EFX_RX_SCALE_UNAVAILABLE) {
+		dev_info->reta_size = EFX_RSS_TBL_SIZE;
+		dev_info->hash_key_size = SFC_RSS_KEY_SIZE;
+		dev_info->flow_type_rss_offloads = SFC_RSS_OFFLOADS;
+	}
+#endif
+
 	dev_info->rx_desc_lim.nb_max = EFX_RXQ_MAXNDESCS;
 	dev_info->rx_desc_lim.nb_min = EFX_RXQ_MINNDESCS;
 	/* The RXQ hardware requires that the descriptor count is a power
diff --git a/drivers/net/sfc/sfc_rx.c b/drivers/net/sfc/sfc_rx.c
index 3bfce1c..f0fe1b8 100644
--- a/drivers/net/sfc/sfc_rx.c
+++ b/drivers/net/sfc/sfc_rx.c
@@ -411,7 +411,8 @@ sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 
 	if (sw_index == 0) {
 		rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
-						    B_FALSE);
+						    (sa->rss_channels > 1) ?
+						    B_TRUE : B_FALSE);
 		if (rc != 0)
 			goto fail_mac_filter_default_rxq_set;
 	}
@@ -683,6 +684,11 @@ sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 	rxq->batch_max = encp->enc_rx_batch_max;
 	rxq->prefix_size = encp->enc_rx_prefix_size;
 
+#if EFSYS_OPT_RX_SCALE
+	if (sa->hash_support == EFX_RX_HASH_AVAILABLE)
+		rxq->flags |= SFC_RXQ_RSS_HASH;
+#endif
+
 	rxq->state = SFC_RXQ_INITIALIZED;
 
 	rxq_info->rxq = rxq;
@@ -728,6 +734,56 @@ sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 	rte_free(rxq);
 }
 
+#if EFSYS_OPT_RX_SCALE
+unsigned int
+sfc_rte_to_efx_hash_type(uint64_t rss_hf)
+{
+	unsigned int efx_hash_types = 0;
+
+	if ((rss_hf & (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
+		       ETH_RSS_NONFRAG_IPV4_OTHER)) != 0)
+		efx_hash_types |= (1 << EFX_RX_HASH_IPV4);
+
+	if ((rss_hf & ETH_RSS_NONFRAG_IPV4_TCP) != 0)
+		efx_hash_types |= (1 << EFX_RX_HASH_TCPIPV4);
+
+	if ((rss_hf & (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
+			ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX)) != 0)
+		efx_hash_types |= (1 << EFX_RX_HASH_IPV6);
+
+	if ((rss_hf & (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX)) != 0)
+		efx_hash_types |= (1 << EFX_RX_HASH_TCPIPV6);
+
+	return efx_hash_types;
+}
+#endif
+
+static int
+sfc_rx_rss_config(struct sfc_adapter *sa)
+{
+	int rc = 0;
+
+#if EFSYS_OPT_RX_SCALE
+	if (sa->rss_channels > 1) {
+		rc = efx_rx_scale_mode_set(sa->nic, EFX_RX_HASHALG_TOEPLITZ,
+					   sa->rss_hash_types, B_TRUE);
+		if (rc != 0)
+			goto finish;
+
+		rc = efx_rx_scale_key_set(sa->nic, sa->rss_key,
+					  sizeof(sa->rss_key));
+		if (rc != 0)
+			goto finish;
+
+		rc = efx_rx_scale_tbl_set(sa->nic, sa->rss_tbl,
+					  sizeof(sa->rss_tbl));
+	}
+
+finish:
+#endif
+	return rc;
+}
+
 int
 sfc_rx_start(struct sfc_adapter *sa)
 {
@@ -740,6 +796,10 @@ sfc_rx_start(struct sfc_adapter *sa)
 	if (rc != 0)
 		goto fail_rx_init;
 
+	rc = sfc_rx_rss_config(sa);
+	if (rc != 0)
+		goto fail_rss_config;
+
 	for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 		if ((!sa->rxq_info[sw_index].deferred_start ||
 		     sa->rxq_info[sw_index].deferred_started)) {
@@ -755,6 +815,7 @@ sfc_rx_start(struct sfc_adapter *sa)
 	while (sw_index-- > 0)
 		sfc_rx_qstop(sa, sw_index);
 
+fail_rss_config:
 	efx_rx_fini(sa->nic);
 
 fail_rx_init:
@@ -801,6 +862,14 @@ sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 	case ETH_MQ_RX_NONE:
 		/* No special checks are required */
 		break;
+#if EFSYS_OPT_RX_SCALE
+	case ETH_MQ_RX_RSS:
+		if (sa->rss_support == EFX_RX_SCALE_UNAVAILABLE) {
+			sfc_err(sa, "RSS is not available");
+			rc = EINVAL;
+		}
+		break;
+#endif
 	default:
 		sfc_err(sa, "Rx multi-queue mode %u not supported",
 			rxmode->mq_mode);
@@ -876,6 +945,16 @@ sfc_rx_init(struct sfc_adapter *sa)
 			goto fail_rx_qinit_info;
 	}
 
+#if EFSYS_OPT_RX_SCALE
+	sa->rss_channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
+			   MIN(sa->rxq_count, EFX_MAXRSS) : 1;
+
+	if (sa->rss_channels > 1) {
+		for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
+			sa->rss_tbl[sw_index] = sw_index % sa->rss_channels;
+	}
+#endif
+
 	return 0;
 
 fail_rx_qinit_info:
diff --git a/drivers/net/sfc/sfc_rx.h b/drivers/net/sfc/sfc_rx.h
index 4aa6aea..2b8b9eb 100644
--- a/drivers/net/sfc/sfc_rx.h
+++ b/drivers/net/sfc/sfc_rx.h
@@ -83,6 +83,10 @@ struct sfc_rxq {
 	unsigned int		completed;
 	uint16_t		batch_max;
 	uint16_t		prefix_size;
+#if EFSYS_OPT_RX_SCALE
+	unsigned int		flags;
+#define	SFC_RXQ_RSS_HASH	0x1
+#endif
 
 	/* Used on refill */
 	unsigned int		added;
@@ -146,6 +150,10 @@ unsigned int sfc_rx_qdesc_npending(struct sfc_adapter *sa,
 				   unsigned int sw_index);
 int sfc_rx_qdesc_done(struct sfc_rxq *rxq, unsigned int offset);
 
+#if EFSYS_OPT_RX_SCALE
+unsigned int sfc_rte_to_efx_hash_type(uint64_t rss_hf);
+#endif
+
 #ifdef __cplusplus
 }
 #endif
-- 
2.5.5