[PATCH v9 10/14] net/idpf: add support for RSS

[Andrew Rybchenko]

On 10/21/22 08:18, Junfeng Guo wrote:
> Add RSS support.
> 
> Signed-off-by: Beilei Xing <beilei.xing at intel.com>
> Signed-off-by: Junfeng Guo <junfeng.guo at intel.com>
> ---
>   drivers/net/idpf/idpf_ethdev.c | 123 ++++++++++++++++++++++++++++++++-
>   drivers/net/idpf/idpf_ethdev.h |  26 +++++++
>   drivers/net/idpf/idpf_rxtx.c   |  27 ++++++++
>   drivers/net/idpf/idpf_vchnl.c  |  96 +++++++++++++++++++++++++
>   4 files changed, 269 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/net/idpf/idpf_ethdev.c b/drivers/net/idpf/idpf_ethdev.c
> index ffde2ce7d1..cc1cfe402b 100644
> --- a/drivers/net/idpf/idpf_ethdev.c
> +++ b/drivers/net/idpf/idpf_ethdev.c
> @@ -83,6 +83,7 @@ idpf_dev_info_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
>   	dev_info->min_rx_bufsize = IDPF_MIN_BUF_SIZE;
>   	dev_info->max_rx_pktlen = IDPF_MAX_FRAME_SIZE;
>   
> +	dev_info->flow_type_rss_offloads = IDPF_RSS_OFFLOAD_ALL;
>   	dev_info->max_mac_addrs = IDPF_NUM_MACADDR_MAX;
>   	dev_info->dev_capa = RTE_ETH_DEV_CAPA_RUNTIME_RX_QUEUE_SETUP |
>   		RTE_ETH_DEV_CAPA_RUNTIME_TX_QUEUE_SETUP;
> @@ -90,7 +91,8 @@ idpf_dev_info_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
>   		RTE_ETH_RX_OFFLOAD_IPV4_CKSUM		|
>   		RTE_ETH_RX_OFFLOAD_UDP_CKSUM		|
>   		RTE_ETH_RX_OFFLOAD_TCP_CKSUM		|
> -		RTE_ETH_RX_OFFLOAD_OUTER_IPV4_CKSUM;
> +		RTE_ETH_RX_OFFLOAD_OUTER_IPV4_CKSUM	|
> +		RTE_ETH_RX_OFFLOAD_RSS_HASH;
>   
>   	dev_info->tx_offload_capa =
>   		RTE_ETH_TX_OFFLOAD_TCP_TSO		|
> @@ -197,6 +199,10 @@ idpf_parse_devarg_id(char *name)
>   	return val;
>   }
>   
> +#ifndef IDPF_RSS_KEY_LEN

Why do we ned ifndef here?

> +#define IDPF_RSS_KEY_LEN 52
> +#endif
> +
>   static int
>   idpf_init_vport(struct rte_eth_dev *dev)
>   {
> @@ -217,6 +223,10 @@ idpf_init_vport(struct rte_eth_dev *dev)
>   	vport->max_mtu = vport_info->max_mtu;
>   	rte_memcpy(vport->default_mac_addr,
>   		   vport_info->default_mac_addr, ETH_ALEN);
> +	vport->rss_algorithm = vport_info->rss_algorithm;
> +	vport->rss_key_size = RTE_MIN(IDPF_RSS_KEY_LEN,
> +				     vport_info->rss_key_size);
> +	vport->rss_lut_size = vport_info->rss_lut_size;
>   	vport->sw_idx = idx;
>   
>   	for (i = 0; i < vport_info->chunks.num_chunks; i++) {
> @@ -265,15 +275,116 @@ idpf_init_vport(struct rte_eth_dev *dev)
>   }
>   
>   static int
> -idpf_dev_configure(__rte_unused struct rte_eth_dev *dev)
> +idpf_config_rss(struct idpf_vport *vport)
> +{
> +	int ret;
> +
> +	ret = idpf_vc_set_rss_key(vport);
> +	if (ret) {
> +		PMD_INIT_LOG(ERR, "Failed to configure RSS key");
> +		return ret;
> +	}
> +
> +	ret = idpf_vc_set_rss_lut(vport);
> +	if (ret) {
> +		PMD_INIT_LOG(ERR, "Failed to configure RSS lut");
> +		return ret;
> +	}
> +
> +	ret = idpf_vc_set_rss_hash(vport);
> +	if (ret) {
> +		PMD_INIT_LOG(ERR, "Failed to configure RSS hash");
> +		return ret;
> +	}
> +
> +	return ret;
> +}
> +
> +static int
> +idpf_init_rss(struct idpf_vport *vport)
> +{
> +	struct rte_eth_rss_conf *rss_conf;
> +	uint16_t i, nb_q, lut_size;
> +	int ret = 0;
> +
> +	rss_conf = &vport->dev_data->dev_conf.rx_adv_conf.rss_conf;
> +	nb_q = vport->dev_data->nb_rx_queues;
> +
> +	vport->rss_key = (uint8_t *)rte_zmalloc("rss_key",
> +					     vport->rss_key_size, 0);
> +	if (!vport->rss_key) {
> +		PMD_INIT_LOG(ERR, "Failed to allocate RSS key");
> +		ret = -ENOMEM;
> +		goto err_key;
> +	}
> +
> +	lut_size = vport->rss_lut_size;
> +	vport->rss_lut = (uint32_t *)rte_zmalloc("rss_lut",
> +					      sizeof(uint32_t) * lut_size, 0);
> +	if (!vport->rss_lut) {
> +		PMD_INIT_LOG(ERR, "Failed to allocate RSS lut");
> +		ret = -ENOMEM;
> +		goto err_lut;
> +	}
> +
> +	if (!rss_conf->rss_key) {
> +		for (i = 0; i < vport->rss_key_size; i++)
> +			vport->rss_key[i] = (uint8_t)rte_rand();
> +	} else {
> +		rte_memcpy(vport->rss_key, rss_conf->rss_key,
> +			   RTE_MIN(rss_conf->rss_key_len,
> +				   vport->rss_key_size));

If provides length is smaller than required, shouldn't we
fill in the rest using rte_rand() as above?

Since the RSS key length is reported in device info,
IMHO it is better to require from user to provide RSS
key with correct legnth.

> +	}
> +
> +	for (i = 0; i < lut_size; i++)
> +		vport->rss_lut[i] = i % nb_q;

Are you sure that nb_q is not 0? In theory it is possible
to configure and start with Tx queues only.

> +
> +	vport->rss_hf = IDPF_DEFAULT_RSS_HASH_EXPANDED;
> +
> +	ret = idpf_config_rss(vport);
> +	if (ret) {
> +		PMD_INIT_LOG(ERR, "Failed to configure RSS");
> +		goto err_cfg;
> +	}
> +
> +	return ret;
> +
> +err_cfg:
> +	rte_free(vport->rss_lut);
> +	vport->rss_lut = NULL;
> +err_lut:
> +	rte_free(vport->rss_key);
> +	vport->rss_key = NULL;
> +err_key:
> +	return ret;
> +}
> +
> +static int
> +idpf_dev_configure(struct rte_eth_dev *dev)
>   {
> +	struct idpf_vport *vport = dev->data->dev_private;
> +	struct idpf_adapter *adapter = vport->adapter;
> +	int ret = 0;
> +
>   	if (dev->data->nb_tx_queues > IDPF_DEFAULT_TXQ_NUM ||
>   	    dev->data->nb_rx_queues > IDPF_DEFAULT_RXQ_NUM) {
>   		PMD_INIT_LOG(ERR, "Invalid queue number.");
>   		return -EINVAL;
>   	}
>   
> -	return 0;
> +	if (dev->data->dev_conf.rxmode.mq_mode & RTE_ETH_MQ_RX_RSS_FLAG)
> +		dev->data->dev_conf.rxmode.offloads |=
> +			RTE_ETH_RX_OFFLOAD_RSS_HASH;

Why? User controls offloads.

> +
> +	if (adapter->caps->rss_caps) {
> +		ret = idpf_init_rss(vport);
> +		if (ret) {
> +			PMD_INIT_LOG(ERR, "Failed to init rss");
> +			return ret;
> +		}

If RSS is not supported but requested, it must be an error, not
simply ignoring of the RSS request.

> +	}
> +
> +	return ret;
>   }
>   
>   static int
> @@ -372,6 +483,12 @@ idpf_dev_close(struct rte_eth_dev *dev)
>   	idpf_dev_stop(dev);
>   	idpf_vc_destroy_vport(vport);
>   
> +	rte_free(vport->rss_lut);
> +	vport->rss_lut = NULL;
> +
> +	rte_free(vport->rss_key);
> +	vport->rss_key = NULL;
> +
>   	adapter->cur_vports &= ~BIT(vport->devarg_id);
>   
>   	rte_free(vport);
> diff --git a/drivers/net/idpf/idpf_ethdev.h b/drivers/net/idpf/idpf_ethdev.h
> index d414efc917..c25fa2e3ce 100644
> --- a/drivers/net/idpf/idpf_ethdev.h
> +++ b/drivers/net/idpf/idpf_ethdev.h
> @@ -56,6 +56,20 @@
>   #define IDPF_ETH_OVERHEAD \
>   	(RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN + IDPF_VLAN_TAG_SIZE * 2)
>   
> +#define IDPF_RSS_OFFLOAD_ALL ( \
> +	RTE_ETH_RSS_IPV4		| \
> +	RTE_ETH_RSS_FRAG_IPV4		| \
> +	RTE_ETH_RSS_NONFRAG_IPV4_TCP	| \
> +	RTE_ETH_RSS_NONFRAG_IPV4_UDP	| \
> +	RTE_ETH_RSS_NONFRAG_IPV4_SCTP	| \
> +	RTE_ETH_RSS_NONFRAG_IPV4_OTHER	| \
> +	RTE_ETH_RSS_IPV6		| \
> +	RTE_ETH_RSS_FRAG_IPV6		| \
> +	RTE_ETH_RSS_NONFRAG_IPV6_TCP	| \
> +	RTE_ETH_RSS_NONFRAG_IPV6_UDP	| \
> +	RTE_ETH_RSS_NONFRAG_IPV6_SCTP	| \
> +	RTE_ETH_RSS_NONFRAG_IPV6_OTHER)
> +
>   #ifndef ETH_ADDR_LEN
>   #define ETH_ADDR_LEN		6
>   #endif
> @@ -102,11 +116,20 @@ struct idpf_vport {
>   	uint16_t max_mtu;
>   	uint8_t default_mac_addr[VIRTCHNL_ETH_LENGTH_OF_ADDRESS];
>   
> +	enum virtchnl_rss_algorithm rss_algorithm;
> +	uint16_t rss_key_size;
> +	uint16_t rss_lut_size;
> +
>   	uint16_t sw_idx; /* SW idx */
>   
>   	struct rte_eth_dev_data *dev_data; /* Pointer to the device data */
>   	uint16_t max_pkt_len; /* Maximum packet length */
>   
> +	/* RSS info */
> +	uint32_t *rss_lut;
> +	uint8_t *rss_key;
> +	uint64_t rss_hf;
> +
>   	/* Chunk info */
>   	struct idpf_chunks_info chunks_info;
>   
> @@ -216,6 +239,9 @@ int idpf_get_pkt_type(struct idpf_adapter *adapter);
>   int idpf_vc_get_caps(struct idpf_adapter *adapter);
>   int idpf_vc_create_vport(struct rte_eth_dev *dev);
>   int idpf_vc_destroy_vport(struct idpf_vport *vport);
> +int idpf_vc_set_rss_key(struct idpf_vport *vport);
> +int idpf_vc_set_rss_lut(struct idpf_vport *vport);
> +int idpf_vc_set_rss_hash(struct idpf_vport *vport);
>   int idpf_vc_config_rxqs(struct idpf_vport *vport);
>   int idpf_vc_config_rxq(struct idpf_vport *vport, uint16_t rxq_id);
>   int idpf_vc_config_txqs(struct idpf_vport *vport);
> diff --git a/drivers/net/idpf/idpf_rxtx.c b/drivers/net/idpf/idpf_rxtx.c
> index 31e266bbec..05976acf7f 100644
> --- a/drivers/net/idpf/idpf_rxtx.c
> +++ b/drivers/net/idpf/idpf_rxtx.c
> @@ -1310,6 +1310,32 @@ idpf_splitq_rx_csum_offload(uint8_t err)
>   	return flags;
>   }
>   
> +#define IDPF_RX_FLEX_DESC_ADV_HASH1_S	0
> +#define IDPF_RX_FLEX_DESC_ADV_HASH2_S	16
> +#define IDPF_RX_FLEX_DESC_ADV_HASH3_S	24
> +
> +static inline uint64_t
> +idpf_splitq_rx_rss_offload(struct rte_mbuf *mb,
> +			   volatile struct virtchnl2_rx_flex_desc_adv_nic_3 *rx_desc)
> +{
> +	uint8_t status_err0_qw0;
> +	uint64_t flags = 0;
> +
> +	status_err0_qw0 = rx_desc->status_err0_qw0;
> +
> +	if (status_err0_qw0 & BIT(VIRTCHNL2_RX_FLEX_DESC_ADV_STATUS0_RSS_VALID_S)) {
> +		flags |= RTE_MBUF_F_RX_RSS_HASH;
> +		mb->hash.rss = (rte_le_to_cpu_16(rx_desc->hash1) <<
> +				IDPF_RX_FLEX_DESC_ADV_HASH1_S) |
> +				((uint32_t)(rx_desc->ff2_mirrid_hash2.hash2) <<
> +				IDPF_RX_FLEX_DESC_ADV_HASH2_S) |
> +				((uint32_t)(rx_desc->hash3) <<
> +				IDPF_RX_FLEX_DESC_ADV_HASH3_S);
> +	}
> +
> +	return flags;
> +}
> +
>   static void
>   idpf_split_rx_bufq_refill(struct idpf_rx_queue *rx_bufq)
>   {
> @@ -1465,6 +1491,7 @@ idpf_splitq_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
>   
>   		status_err0_qw1 = rx_desc->status_err0_qw1;
>   		pkt_flags = idpf_splitq_rx_csum_offload(status_err0_qw1);
> +		pkt_flags |= idpf_splitq_rx_rss_offload(rxm, rx_desc);
>   
>   		rxm->ol_flags |= pkt_flags;
>   
> diff --git a/drivers/net/idpf/idpf_vchnl.c b/drivers/net/idpf/idpf_vchnl.c
> index 7389128712..8f447046b1 100644
> --- a/drivers/net/idpf/idpf_vchnl.c
> +++ b/drivers/net/idpf/idpf_vchnl.c
> @@ -679,6 +679,102 @@ idpf_vc_destroy_vport(struct idpf_vport *vport)
>   	return err;
>   }
>   
> +int
> +idpf_vc_set_rss_key(struct idpf_vport *vport)
> +{
> +	struct idpf_adapter *adapter = vport->adapter;
> +	struct virtchnl2_rss_key *rss_key;
> +	struct idpf_cmd_info args;
> +	int len, err;
> +
> +	len = sizeof(*rss_key) + sizeof(rss_key->key[0]) *
> +		(vport->rss_key_size - 1);
> +	rss_key = rte_zmalloc("rss_key", len, 0);
> +	if (!rss_key)
> +		return -ENOMEM;
> +
> +	rss_key->vport_id = vport->vport_id;
> +	rss_key->key_len = vport->rss_key_size;
> +	rte_memcpy(rss_key->key, vport->rss_key,
> +		   sizeof(rss_key->key[0]) * vport->rss_key_size);
> +
> +	memset(&args, 0, sizeof(args));
> +	args.ops = VIRTCHNL2_OP_SET_RSS_KEY;
> +	args.in_args = (uint8_t *)rss_key;
> +	args.in_args_size = len;
> +	args.out_buffer = adapter->mbx_resp;
> +	args.out_size = IDPF_DFLT_MBX_BUF_SIZE;
> +
> +	err = idpf_execute_vc_cmd(adapter, &args);
> +	if (err) {
> +		PMD_DRV_LOG(ERR, "Failed to execute command of VIRTCHNL2_OP_SET_RSS_KEY");
> +		return err;
> +	}
> +
> +	rte_free(rss_key);
> +	return err;
> +}
> +
> +int
> +idpf_vc_set_rss_lut(struct idpf_vport *vport)
> +{
> +	struct idpf_adapter *adapter = vport->adapter;
> +	struct virtchnl2_rss_lut *rss_lut;
> +	struct idpf_cmd_info args;
> +	int len, err;
> +
> +	len = sizeof(*rss_lut) + sizeof(rss_lut->lut[0]) *
> +		(vport->rss_lut_size - 1);
> +	rss_lut = rte_zmalloc("rss_lut", len, 0);
> +	if (!rss_lut)
> +		return -ENOMEM;
> +
> +	rss_lut->vport_id = vport->vport_id;
> +	rss_lut->lut_entries = vport->rss_lut_size;
> +	rte_memcpy(rss_lut->lut, vport->rss_lut,
> +		   sizeof(rss_lut->lut[0]) * vport->rss_lut_size);
> +
> +	memset(&args, 0, sizeof(args));
> +	args.ops = VIRTCHNL2_OP_SET_RSS_LUT;
> +	args.in_args = (uint8_t *)rss_lut;
> +	args.in_args_size = len;
> +	args.out_buffer = adapter->mbx_resp;
> +	args.out_size = IDPF_DFLT_MBX_BUF_SIZE;
> +
> +	err = idpf_execute_vc_cmd(adapter, &args);
> +	if (err)
> +		PMD_DRV_LOG(ERR, "Failed to execute command of VIRTCHNL2_OP_SET_RSS_LUT");
> +
> +	rte_free(rss_lut);
> +	return err;
> +}
> +
> +int
> +idpf_vc_set_rss_hash(struct idpf_vport *vport)
> +{
> +	struct idpf_adapter *adapter = vport->adapter;
> +	struct virtchnl2_rss_hash rss_hash;
> +	struct idpf_cmd_info args;
> +	int err;
> +
> +	memset(&rss_hash, 0, sizeof(rss_hash));
> +	rss_hash.ptype_groups = vport->rss_hf;
> +	rss_hash.vport_id = vport->vport_id;
> +
> +	memset(&args, 0, sizeof(args));
> +	args.ops = VIRTCHNL2_OP_SET_RSS_HASH;
> +	args.in_args = (uint8_t *)&rss_hash;
> +	args.in_args_size = sizeof(rss_hash);
> +	args.out_buffer = adapter->mbx_resp;
> +	args.out_size = IDPF_DFLT_MBX_BUF_SIZE;
> +
> +	err = idpf_execute_vc_cmd(adapter, &args);
> +	if (err)
> +		PMD_DRV_LOG(ERR, "Failed to execute command of OP_SET_RSS_HASH");
> +
> +	return err;
> +}
> +
>   #define IDPF_RX_BUF_STRIDE		64
>   int
>   idpf_vc_config_rxqs(struct idpf_vport *vport)