Re: [dpdk-dev] [PATCH v2 07/15] net/mlx5: support tunnel RSS level

["Nélio Laranjeiro" <nelio.laranjeiro@6wind.com>]

On Tue, Apr 10, 2018 at 09:34:07PM +0800, Xueming Li wrote:
> Tunnel RSS level of flow RSS action offers user a choice to do RSS hash
> calculation on inner or outer RSS fields. Testpmd flow command examples:
> 
> GRE flow inner RSS:
>   flow create 0 ingress pattern eth / ipv4 proto is 47 / gre / end
> actions rss queues 1 2 end level 1 / end
> 
> GRE tunnel flow outer RSS:
>   flow create 0 ingress pattern eth  / ipv4 proto is 47 / gre / end
> actions rss queues 1 2 end level 0 / end
> 
> Signed-off-by: Xueming Li <xuemingl@mellanox.com>
> ---
>  drivers/net/mlx5/Makefile    |   2 +-
>  drivers/net/mlx5/mlx5_flow.c | 249 ++++++++++++++++++++++++++++++-------------
>  drivers/net/mlx5/mlx5_glue.c |  16 +++
>  drivers/net/mlx5/mlx5_glue.h |   8 ++
>  drivers/net/mlx5/mlx5_rxq.c  |  46 +++++++-
>  drivers/net/mlx5/mlx5_rxtx.h |   5 +-
>  6 files changed, 246 insertions(+), 80 deletions(-)
> 
> diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
> index ae118ad33..f9a6c460b 100644
> --- a/drivers/net/mlx5/Makefile
> +++ b/drivers/net/mlx5/Makefile
> @@ -35,7 +35,7 @@ include $(RTE_SDK)/mk/rte.vars.mk
>  LIB = librte_pmd_mlx5.a
>  LIB_GLUE = $(LIB_GLUE_BASE).$(LIB_GLUE_VERSION)
>  LIB_GLUE_BASE = librte_pmd_mlx5_glue.so
> -LIB_GLUE_VERSION = 18.02.0
> +LIB_GLUE_VERSION = 18.05.0
>  
>  # Sources.
>  SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD) += mlx5.c
> diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
> index 64658bc0e..66c7d7993 100644
> --- a/drivers/net/mlx5/mlx5_flow.c
> +++ b/drivers/net/mlx5/mlx5_flow.c
> @@ -113,6 +113,7 @@ enum hash_rxq_type {
>  	HASH_RXQ_UDPV6,
>  	HASH_RXQ_IPV6,
>  	HASH_RXQ_ETH,
> +	HASH_RXQ_TUNNEL,
>  };
>  
>  /* Initialization data for hash RX queue. */
> @@ -451,6 +452,7 @@ struct mlx5_flow_parse {
>  	uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
>  	uint8_t rss_key[40]; /**< copy of the RSS key. */
>  	enum hash_rxq_type layer; /**< Last pattern layer detected. */
> +	enum hash_rxq_type out_layer; /**< Last outer pattern layer detected. */
>  	uint32_t tunnel; /**< Tunnel type of RTE_PTYPE_TUNNEL_XXX. */
>  	struct ibv_counter_set *cs; /**< Holds the counter set for the rule */
>  	struct {
> @@ -458,6 +460,7 @@ struct mlx5_flow_parse {
>  		/**< Pointer to Verbs attributes. */
>  		unsigned int offset;
>  		/**< Current position or total size of the attribute. */
> +		uint64_t hash_fields; /**< Verbs hash fields. */
>  	} queue[RTE_DIM(hash_rxq_init)];
>  };
>  
> @@ -698,7 +701,8 @@ mlx5_flow_convert_actions(struct rte_eth_dev *dev,
>  						   " function is Toeplitz");
>  				return -rte_errno;
>  			}
> -			if (rss->level) {
> +#ifndef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
> +			if (parser->rss_conf.level > 0) {

According to Adrien's API level 0 means do whatever you want and 1 means outer.
This is removing the outer RSS support.

>  				rte_flow_error_set(error, EINVAL,
>  						   RTE_FLOW_ERROR_TYPE_ACTION,
>  						   actions,
> @@ -706,6 +710,15 @@ mlx5_flow_convert_actions(struct rte_eth_dev *dev,
>  						   " level is not supported");
>  				return -rte_errno;
>  			}
> +#endif
> +			if (parser->rss_conf.level > 1) {
> +				rte_flow_error_set(error, EINVAL,
> +						   RTE_FLOW_ERROR_TYPE_ACTION,
> +						   actions,
> +						   "RSS encapsulation level"
> +						   " > 1 is not supported");
> +				return -rte_errno;
> +			}

Seems the levels are wrongly used.

>  			if (rss->types & MLX5_RSS_HF_MASK) {
>  				rte_flow_error_set(error, EINVAL,
>  						   RTE_FLOW_ERROR_TYPE_ACTION,
> @@ -756,7 +769,7 @@ mlx5_flow_convert_actions(struct rte_eth_dev *dev,
>  			}
>  			parser->rss_conf = (struct rte_flow_action_rss){
>  				.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
> -				.level = 0,
> +				.level = rss->level,
>  				.types = rss->types,
>  				.key_len = rss_key_len,
>  				.queue_num = rss->queue_num,
> @@ -842,11 +855,12 @@ mlx5_flow_convert_actions(struct rte_eth_dev *dev,
>   *   0 on success, a negative errno value otherwise and rte_errno is set.
>   */
>  static int
> -mlx5_flow_convert_items_validate(struct rte_eth_dev *dev __rte_unused,
> +mlx5_flow_convert_items_validate(struct rte_eth_dev *dev,
>  				 const struct rte_flow_item items[],
>  				 struct rte_flow_error *error,
>  				 struct mlx5_flow_parse *parser)
>  {
> +	struct priv *priv = dev->data->dev_private;
>  	const struct mlx5_flow_items *cur_item = mlx5_flow_items;
>  	unsigned int i;
>  	int ret = 0;
> @@ -886,6 +900,14 @@ mlx5_flow_convert_items_validate(struct rte_eth_dev *dev __rte_unused,
>  						   " tunnel encapsulations.");
>  				return -rte_errno;
>  			}
> +			if (!priv->config.tunnel_en &&
> +			    parser->rss_conf.level) {
> +				rte_flow_error_set(error, ENOTSUP,
> +					RTE_FLOW_ERROR_TYPE_ITEM,
> +					items,
> +					"Tunnel offloading not enabled");

I would suggest "RSS on tunnel is not supported".

> +				return -rte_errno;
> +			}
>  			parser->inner = IBV_FLOW_SPEC_INNER;
>  			parser->tunnel = flow_ptype[items->type];
>  		}
> @@ -993,7 +1015,11 @@ static void
>  mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
>  {
>  	unsigned int i;
> +	uint32_t inner = parser->inner;
>  
> +	/* Don't create extra flows for outer RSS. */
> +	if (parser->tunnel && !parser->rss_conf.level)
> +		return;
>  	/* Remove any other flow not matching the pattern. */
>  	if (parser->rss_conf.queue_num == 1 && !parser->rss_conf.types) {
>  		for (i = 0; i != hash_rxq_init_n; ++i) {
> @@ -1014,23 +1040,25 @@ mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
>  			struct ibv_flow_spec_ipv4_ext ipv4;
>  			struct ibv_flow_spec_ipv6 ipv6;
>  			struct ibv_flow_spec_tcp_udp udp_tcp;
> +			struct ibv_flow_spec_eth eth;
>  		} specs;
>  		void *dst;
>  		uint16_t size;
>  
>  		if (i == parser->layer)
>  			continue;
> -		if (parser->layer == HASH_RXQ_ETH) {
> +		if (parser->layer == HASH_RXQ_ETH ||
> +		    parser->layer == HASH_RXQ_TUNNEL) {
>  			if (hash_rxq_init[i].ip_version == MLX5_IPV4) {
>  				size = sizeof(struct ibv_flow_spec_ipv4_ext);
>  				specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){
> -					.type = IBV_FLOW_SPEC_IPV4_EXT,
> +					.type = inner | IBV_FLOW_SPEC_IPV4_EXT,
>  					.size = size,
>  				};
>  			} else {
>  				size = sizeof(struct ibv_flow_spec_ipv6);
>  				specs.ipv6 = (struct ibv_flow_spec_ipv6){
> -					.type = IBV_FLOW_SPEC_IPV6,
> +					.type = inner | IBV_FLOW_SPEC_IPV6,
>  					.size = size,
>  				};
>  			}
> @@ -1047,7 +1075,7 @@ mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
>  		    (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
>  			size = sizeof(struct ibv_flow_spec_tcp_udp);
>  			specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
> -				.type = ((i == HASH_RXQ_UDPV4 ||
> +				.type = inner | ((i == HASH_RXQ_UDPV4 ||
>  					  i == HASH_RXQ_UDPV6) ?
>  					 IBV_FLOW_SPEC_UDP :
>  					 IBV_FLOW_SPEC_TCP),
> @@ -1068,6 +1096,8 @@ mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
>  /**
>   * Update flows according to pattern and RSS hash fields.
>   *
> + * @param dev
> + *   Pointer to Ethernet device.
>   * @param[in, out] parser
>   *   Internal parser structure.
>   *
> @@ -1075,20 +1105,63 @@ mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
>   *   0 on success, a negative errno value otherwise and rte_errno is set.
>   */
>  static int
> -mlx5_flow_convert_rss(struct mlx5_flow_parse *parser)
> +mlx5_flow_convert_rss(struct rte_eth_dev *dev, struct mlx5_flow_parse *parser)
>  {
> -	const unsigned int ipv4 =
> +	unsigned int ipv4 =
>  		hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
>  	const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
>  	const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
>  	const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
>  	const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
> -	const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
> +	enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
>  	unsigned int i;
> +	int found = 0;
>  
> -	if (parser->layer == HASH_RXQ_ETH)
> +	/*
> +	 * Outer RSS.
> +	 * HASH_RXQ_ETH is the only rule since tunnel packet match this
> +	 * rule must match outer pattern.
> +	 */
> +	if (parser->tunnel && !parser->rss_conf.level) {
> +		/* Remove flows other than default. */
> +		for (i = 0; i != hash_rxq_init_n - 1; ++i) {
> +			rte_free(parser->queue[i].ibv_attr);
> +			parser->queue[i].ibv_attr = NULL;
> +		}
> +		ipv4 = hash_rxq_init[parser->out_layer].ip_version == MLX5_IPV4;
> +		ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
> +		if (hash_rxq_init[parser->out_layer].dpdk_rss_hf &
> +		    parser->rss_conf.types) {
> +			parser->queue[HASH_RXQ_ETH].hash_fields =
> +				hash_rxq_init[parser->out_layer].hash_fields;
> +		} else if (ip && (hash_rxq_init[ip].dpdk_rss_hf &
> +		    parser->rss_conf.types)) {
> +			parser->queue[HASH_RXQ_ETH].hash_fields =
> +				hash_rxq_init[ip].hash_fields;
> +		} else if (parser->rss_conf.types) {
> +			DRV_LOG(WARNING,
> +				"port %u rss outer hash function doesn't match"
> +				" pattern", dev->data->port_id);

Hash function, what do you mean ?  It seems to be the layers on the ones
the RSS is configured which does not match the Pattern.

Sincerely, I see such warning happening I will fully doubt on the fact
the rule has been taken and applied.
"port 0 rss outer hash function doesn't match pattern" --> what will
happen to the packets matching such flow?  Will they be dropped?
This is not helping at all, so please remove it.

> +		}
> +		return 0;
> +	}
> +	if (parser->layer == HASH_RXQ_ETH || parser->layer == HASH_RXQ_TUNNEL) {
> +		/* Remove unused flows according to hash function. */
> +		for (i = 0; i != hash_rxq_init_n - 1; ++i) {
> +			if (!parser->queue[i].ibv_attr)
> +				continue;
> +			if (hash_rxq_init[i].dpdk_rss_hf &
> +			    parser->rss_conf.types) {
> +				parser->queue[i].hash_fields =
> +					hash_rxq_init[i].hash_fields;
> +				continue;
> +			}
> +			rte_free(parser->queue[i].ibv_attr);
> +			parser->queue[i].ibv_attr = NULL;
> +		}
>  		return 0;
> -	/* This layer becomes useless as the pattern define under layers. */
> +	}
> +	/* Remove ETH layer flow. */
>  	rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr);
>  	parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
>  	/* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */
> @@ -1098,9 +1171,52 @@ mlx5_flow_convert_rss(struct mlx5_flow_parse *parser)
>  		rte_free(parser->queue[i].ibv_attr);
>  		parser->queue[i].ibv_attr = NULL;
>  	}
> -	/* Remove impossible flow according to the RSS configuration. */
> -	if (hash_rxq_init[parser->layer].dpdk_rss_hf &
> -	    parser->rss_conf.types) {
> +	/*
> +	 * Keep L4 flows as IP pattern has to support L4 RSS.
> +	 * Otherwise, only keep the flow that match the pattern.
> +	 */

This comment is not clear, please re-word it.

> +	if (parser->layer != ip) {
> +		/* Only keep the flow that match the pattern. */
> +		for (i = hmin; i != (hmax + 1); ++i) {
> +			if (i == parser->layer)
> +				continue;
> +			rte_free(parser->queue[i].ibv_attr);
> +			parser->queue[i].ibv_attr = NULL;
> +		}
> +	}
> +	if (parser->rss_conf.types) {
> +		/* Remove impossible flow according to the RSS configuration. */
> +		for (i = hmin; i != (hmax + 1); ++i) {
> +			if (!parser->queue[i].ibv_attr)
> +				continue;
> +			if (parser->rss_conf.types &
> +			    hash_rxq_init[i].dpdk_rss_hf) {
> +				parser->queue[i].hash_fields =
> +					hash_rxq_init[i].hash_fields;
> +				found = 1;
> +				continue;
> +			}
> +			/* L4 flow could be used for L3 RSS. */
> +			if (i == parser->layer && i < ip &&
> +			    (hash_rxq_init[ip].dpdk_rss_hf &
> +			     parser->rss_conf.types)) {
> +				parser->queue[i].hash_fields =
> +					hash_rxq_init[ip].hash_fields;
> +				found = 1;
> +				continue;
> +			}
> +			/* L3 flow and L4 hash: non-rss L3 flow. */
> +			if (i == parser->layer && i == ip && found)
> +				/* IP pattern and L4 HF. */
> +				continue;
> +			rte_free(parser->queue[i].ibv_attr);
> +			parser->queue[i].ibv_attr = NULL;
> +		}
> +		if (!found)
> +			DRV_LOG(WARNING,
> +				"port %u rss hash function doesn't match "
> +				"pattern", dev->data->port_id);

Dito.

> +	} else {
>  		/* Remove any other flow. */
>  		for (i = hmin; i != (hmax + 1); ++i) {
>  			if (i == parser->layer || !parser->queue[i].ibv_attr)
> @@ -1108,8 +1224,6 @@ mlx5_flow_convert_rss(struct mlx5_flow_parse *parser)
>  			rte_free(parser->queue[i].ibv_attr);
>  			parser->queue[i].ibv_attr = NULL;
>  		}
> -	} else if (!parser->queue[ip].ibv_attr) {
> -		/* no RSS possible with the current configuration. */
>  		parser->rss_conf.queue_num = 1;
>  	}
>  	return 0;
> @@ -1179,10 +1293,6 @@ mlx5_flow_convert(struct rte_eth_dev *dev,
>  		for (i = 0; i != hash_rxq_init_n; ++i) {
>  			unsigned int offset;
>  
> -			if (!(parser->rss_conf.types &
> -			      hash_rxq_init[i].dpdk_rss_hf) &&
> -			    (i != HASH_RXQ_ETH))
> -				continue;
>  			offset = parser->queue[i].offset;
>  			parser->queue[i].ibv_attr =
>  				mlx5_flow_convert_allocate(offset, error);
> @@ -1194,6 +1304,7 @@ mlx5_flow_convert(struct rte_eth_dev *dev,
>  	/* Third step. Conversion parse, fill the specifications. */
>  	parser->inner = 0;
>  	parser->tunnel = 0;
> +	parser->layer = HASH_RXQ_ETH;
>  	for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
>  		struct mlx5_flow_data data = {
>  			.parser = parser,
> @@ -1211,23 +1322,23 @@ mlx5_flow_convert(struct rte_eth_dev *dev,
>  		if (ret)
>  			goto exit_free;
>  	}
> -	if (parser->mark)
> -		mlx5_flow_create_flag_mark(parser, parser->mark_id);
> -	if (parser->count && parser->create) {
> -		mlx5_flow_create_count(dev, parser);
> -		if (!parser->cs)
> -			goto exit_count_error;
> -	}
>  	/*
>  	 * Last step. Complete missing specification to reach the RSS
>  	 * configuration.
>  	 */
>  	if (!parser->drop)
> -		ret = mlx5_flow_convert_rss(parser);
> +		ret = mlx5_flow_convert_rss(dev, parser);
>  		if (ret)
>  			goto exit_free;
>  		mlx5_flow_convert_finalise(parser);
>  	mlx5_flow_update_priority(dev, parser, attr);
> +	if (parser->mark)
> +		mlx5_flow_create_flag_mark(parser, parser->mark_id);
> +	if (parser->count && parser->create) {
> +		mlx5_flow_create_count(dev, parser);
> +		if (!parser->cs)
> +			goto exit_count_error;
> +	}

Why do you need to move this code?

>  exit_free:
>  	/* Only verification is expected, all resources should be released. */
>  	if (!parser->create) {
> @@ -1275,17 +1386,11 @@ mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
>  	for (i = 0; i != hash_rxq_init_n; ++i) {
>  		if (!parser->queue[i].ibv_attr)
>  			continue;
> -		/* Specification must be the same l3 type or none. */
> -		if (parser->layer == HASH_RXQ_ETH ||
> -		    (hash_rxq_init[parser->layer].ip_version ==
> -		     hash_rxq_init[i].ip_version) ||
> -		    (hash_rxq_init[i].ip_version == 0)) {
> -			dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
> -					parser->queue[i].offset);
> -			memcpy(dst, src, size);
> -			++parser->queue[i].ibv_attr->num_of_specs;
> -			parser->queue[i].offset += size;
> -		}
> +		dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
> +				parser->queue[i].offset);
> +		memcpy(dst, src, size);
> +		++parser->queue[i].ibv_attr->num_of_specs;
> +		parser->queue[i].offset += size;
>  	}
>  }
>  
> @@ -1316,9 +1421,7 @@ mlx5_flow_create_eth(const struct rte_flow_item *item,
>  		.size = eth_size,
>  	};
>  
> -	/* Don't update layer for the inner pattern. */
> -	if (!parser->inner)
> -		parser->layer = HASH_RXQ_ETH;
> +	parser->layer = HASH_RXQ_ETH;
>  	if (spec) {
>  		unsigned int i;
>  
> @@ -1431,9 +1534,7 @@ mlx5_flow_create_ipv4(const struct rte_flow_item *item,
>  		.size = ipv4_size,
>  	};
>  
> -	/* Don't update layer for the inner pattern. */
> -	if (!parser->inner)
> -		parser->layer = HASH_RXQ_IPV4;
> +	parser->layer = HASH_RXQ_IPV4;
>  	if (spec) {
>  		if (!mask)
>  			mask = default_mask;
> @@ -1486,9 +1587,7 @@ mlx5_flow_create_ipv6(const struct rte_flow_item *item,
>  		.size = ipv6_size,
>  	};
>  
> -	/* Don't update layer for the inner pattern. */
> -	if (!parser->inner)
> -		parser->layer = HASH_RXQ_IPV6;
> +	parser->layer = HASH_RXQ_IPV6;
>  	if (spec) {
>  		unsigned int i;
>  		uint32_t vtc_flow_val;
> @@ -1561,13 +1660,10 @@ mlx5_flow_create_udp(const struct rte_flow_item *item,
>  		.size = udp_size,
>  	};
>  
> -	/* Don't update layer for the inner pattern. */
> -	if (!parser->inner) {
> -		if (parser->layer == HASH_RXQ_IPV4)
> -			parser->layer = HASH_RXQ_UDPV4;
> -		else
> -			parser->layer = HASH_RXQ_UDPV6;
> -	}
> +	if (parser->layer == HASH_RXQ_IPV4)
> +		parser->layer = HASH_RXQ_UDPV4;
> +	else
> +		parser->layer = HASH_RXQ_UDPV6;
>  	if (spec) {
>  		if (!mask)
>  			mask = default_mask;
> @@ -1610,13 +1706,10 @@ mlx5_flow_create_tcp(const struct rte_flow_item *item,
>  		.size = tcp_size,
>  	};
>  
> -	/* Don't update layer for the inner pattern. */
> -	if (!parser->inner) {
> -		if (parser->layer == HASH_RXQ_IPV4)
> -			parser->layer = HASH_RXQ_TCPV4;
> -		else
> -			parser->layer = HASH_RXQ_TCPV6;
> -	}
> +	if (parser->layer == HASH_RXQ_IPV4)
> +		parser->layer = HASH_RXQ_TCPV4;
> +	else
> +		parser->layer = HASH_RXQ_TCPV6;
>  	if (spec) {
>  		if (!mask)
>  			mask = default_mask;
> @@ -1666,6 +1759,8 @@ mlx5_flow_create_vxlan(const struct rte_flow_item *item,
>  	id.vni[0] = 0;
>  	parser->inner = IBV_FLOW_SPEC_INNER;
>  	parser->tunnel = ptype_ext[PTYPE_IDX(RTE_PTYPE_TUNNEL_VXLAN)];
> +	parser->out_layer = parser->layer;
> +	parser->layer = HASH_RXQ_TUNNEL;
>  	if (spec) {
>  		if (!mask)
>  			mask = default_mask;
> @@ -1720,6 +1815,8 @@ mlx5_flow_create_gre(const struct rte_flow_item *item __rte_unused,
>  
>  	parser->inner = IBV_FLOW_SPEC_INNER;
>  	parser->tunnel = ptype_ext[PTYPE_IDX(RTE_PTYPE_TUNNEL_GRE)];
> +	parser->out_layer = parser->layer;
> +	parser->layer = HASH_RXQ_TUNNEL;
>  	mlx5_flow_create_copy(parser, &tunnel, size);
>  	return 0;
>  }
> @@ -1883,33 +1980,33 @@ mlx5_flow_create_action_queue_rss(struct rte_eth_dev *dev,
>  	unsigned int i;
>  
>  	for (i = 0; i != hash_rxq_init_n; ++i) {
> -		uint64_t hash_fields;
> -
>  		if (!parser->queue[i].ibv_attr)
>  			continue;
>  		flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
>  		parser->queue[i].ibv_attr = NULL;
> -		hash_fields = hash_rxq_init[i].hash_fields;
> +		flow->frxq[i].hash_fields = parser->queue[i].hash_fields;
>  		if (!priv->dev->data->dev_started)
>  			continue;
>  		flow->frxq[i].hrxq =
>  			mlx5_hrxq_get(dev,
>  				      parser->rss_conf.key,
>  				      parser->rss_conf.key_len,
> -				      hash_fields,
> +				      flow->frxq[i].hash_fields,
>  				      parser->rss_conf.queue,
>  				      parser->rss_conf.queue_num,
> -				      parser->tunnel);
> +				      parser->tunnel,
> +				      parser->rss_conf.level);
>  		if (flow->frxq[i].hrxq)
>  			continue;
>  		flow->frxq[i].hrxq =
>  			mlx5_hrxq_new(dev,
>  				      parser->rss_conf.key,
>  				      parser->rss_conf.key_len,
> -				      hash_fields,
> +				      flow->frxq[i].hash_fields,
>  				      parser->rss_conf.queue,
>  				      parser->rss_conf.queue_num,
> -				      parser->tunnel);
> +				      parser->tunnel,
> +				      parser->rss_conf.level);
>  		if (!flow->frxq[i].hrxq) {
>  			return rte_flow_error_set(error, ENOMEM,
>  						  RTE_FLOW_ERROR_TYPE_HANDLE,
> @@ -2006,7 +2103,7 @@ mlx5_flow_create_action_queue(struct rte_eth_dev *dev,
>  		DRV_LOG(DEBUG, "port %u %p type %d QP %p ibv_flow %p",
>  			dev->data->port_id,
>  			(void *)flow, i,
> -			(void *)flow->frxq[i].hrxq,
> +			(void *)flow->frxq[i].hrxq->qp,
>  			(void *)flow->frxq[i].ibv_flow);
>  	}
>  	if (!flows_n) {
> @@ -2532,19 +2629,21 @@ mlx5_flow_start(struct rte_eth_dev *dev, struct mlx5_flows *list)
>  			flow->frxq[i].hrxq =
>  				mlx5_hrxq_get(dev, flow->rss_conf.key,
>  					      flow->rss_conf.key_len,
> -					      hash_rxq_init[i].hash_fields,
> +					      flow->frxq[i].hash_fields,
>  					      flow->rss_conf.queue,
>  					      flow->rss_conf.queue_num,
> -					      flow->tunnel);
> +					      flow->tunnel,
> +					      flow->rss_conf.level);
>  			if (flow->frxq[i].hrxq)
>  				goto flow_create;
>  			flow->frxq[i].hrxq =
>  				mlx5_hrxq_new(dev, flow->rss_conf.key,
>  					      flow->rss_conf.key_len,
> -					      hash_rxq_init[i].hash_fields,
> +					      flow->frxq[i].hash_fields,
>  					      flow->rss_conf.queue,
>  					      flow->rss_conf.queue_num,
> -					      flow->tunnel);
> +					      flow->tunnel,
> +					      flow->rss_conf.level);
>  			if (!flow->frxq[i].hrxq) {
>  				DRV_LOG(DEBUG,
>  					"port %u flow %p cannot be applied",
> diff --git a/drivers/net/mlx5/mlx5_glue.c b/drivers/net/mlx5/mlx5_glue.c
> index be684d378..6874aa32a 100644
> --- a/drivers/net/mlx5/mlx5_glue.c
> +++ b/drivers/net/mlx5/mlx5_glue.c
> @@ -313,6 +313,21 @@ mlx5_glue_dv_init_obj(struct mlx5dv_obj *obj, uint64_t obj_type)
>  	return mlx5dv_init_obj(obj, obj_type);
>  }
>  
> +static struct ibv_qp *
> +mlx5_glue_dv_create_qp(struct ibv_context *context,
> +		       struct ibv_qp_init_attr_ex *qp_init_attr_ex,
> +		       struct mlx5dv_qp_init_attr *dv_qp_init_attr)
> +{
> +#ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
> +	return mlx5dv_create_qp(context, qp_init_attr_ex, dv_qp_init_attr);
> +#else
> +	(void)context;
> +	(void)qp_init_attr_ex;
> +	(void)dv_qp_init_attr;
> +	return NULL;
> +#endif
> +}
> +
>  const struct mlx5_glue *mlx5_glue = &(const struct mlx5_glue){
>  	.version = MLX5_GLUE_VERSION,
>  	.fork_init = mlx5_glue_fork_init,
> @@ -356,4 +371,5 @@ const struct mlx5_glue *mlx5_glue = &(const struct mlx5_glue){
>  	.dv_query_device = mlx5_glue_dv_query_device,
>  	.dv_set_context_attr = mlx5_glue_dv_set_context_attr,
>  	.dv_init_obj = mlx5_glue_dv_init_obj,
> +	.dv_create_qp = mlx5_glue_dv_create_qp,
>  };
> diff --git a/drivers/net/mlx5/mlx5_glue.h b/drivers/net/mlx5/mlx5_glue.h
> index b5efee3b6..841363872 100644
> --- a/drivers/net/mlx5/mlx5_glue.h
> +++ b/drivers/net/mlx5/mlx5_glue.h
> @@ -31,6 +31,10 @@ struct ibv_counter_set_init_attr;
>  struct ibv_query_counter_set_attr;
>  #endif
>  
> +#ifndef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
> +struct mlx5dv_qp_init_attr;
> +#endif
> +
>  /* LIB_GLUE_VERSION must be updated every time this structure is modified. */
>  struct mlx5_glue {
>  	const char *version;
> @@ -106,6 +110,10 @@ struct mlx5_glue {
>  				   enum mlx5dv_set_ctx_attr_type type,
>  				   void *attr);
>  	int (*dv_init_obj)(struct mlx5dv_obj *obj, uint64_t obj_type);
> +	struct ibv_qp *(*dv_create_qp)
> +		(struct ibv_context *context,
> +		 struct ibv_qp_init_attr_ex *qp_init_attr_ex,
> +		 struct mlx5dv_qp_init_attr *dv_qp_init_attr);
>  };
>  
>  const struct mlx5_glue *mlx5_glue;
> diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
> index 073732e16..6e5565fb2 100644
> --- a/drivers/net/mlx5/mlx5_rxq.c
> +++ b/drivers/net/mlx5/mlx5_rxq.c
> @@ -1386,6 +1386,8 @@ mlx5_ind_table_ibv_verify(struct rte_eth_dev *dev)
>   *   Number of queues.
>   * @param tunnel
>   *   Tunnel type.
> + * @param rss_level
> + *   RSS hash on tunnel level.
>   *
>   * @return
>   *   The Verbs object initialised, NULL otherwise and rte_errno is set.
> @@ -1394,13 +1396,17 @@ struct mlx5_hrxq *
>  mlx5_hrxq_new(struct rte_eth_dev *dev,
>  	      const uint8_t *rss_key, uint32_t rss_key_len,
>  	      uint64_t hash_fields,
> -	      const uint16_t *queues, uint32_t queues_n, uint32_t tunnel)
> +	      const uint16_t *queues, uint32_t queues_n,
> +	      uint32_t tunnel, uint32_t rss_level)

tunnel and rss_level seems to be redundant here.

rss_level > 1 is equivalent to tunnel, there is no need to have both.

>  {
>  	struct priv *priv = dev->data->dev_private;
>  	struct mlx5_hrxq *hrxq;
>  	struct mlx5_ind_table_ibv *ind_tbl;
>  	struct ibv_qp *qp;
>  	int err;
> +#ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
> +	struct mlx5dv_qp_init_attr qp_init_attr = {0};
> +#endif
>  
>  	queues_n = hash_fields ? queues_n : 1;
>  	ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
> @@ -1410,6 +1416,33 @@ mlx5_hrxq_new(struct rte_eth_dev *dev,
>  		rte_errno = ENOMEM;
>  		return NULL;
>  	}
> +#ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
> +	if (tunnel) {
> +		qp_init_attr.comp_mask =
> +				MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
> +		qp_init_attr.create_flags = MLX5DV_QP_CREATE_TUNNEL_OFFLOADS;
> +	}
> +	qp = mlx5_glue->dv_create_qp(
> +		priv->ctx,
> +		&(struct ibv_qp_init_attr_ex){
> +			.qp_type = IBV_QPT_RAW_PACKET,
> +			.comp_mask =
> +				IBV_QP_INIT_ATTR_PD |
> +				IBV_QP_INIT_ATTR_IND_TABLE |
> +				IBV_QP_INIT_ATTR_RX_HASH,
> +			.rx_hash_conf = (struct ibv_rx_hash_conf){
> +				.rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
> +				.rx_hash_key_len = rss_key_len,
> +				.rx_hash_key = (void *)(uintptr_t)rss_key,
> +				.rx_hash_fields_mask = hash_fields |
> +					(tunnel && rss_level ?
> +					(uint32_t)IBV_RX_HASH_INNER : 0),
> +			},
> +			.rwq_ind_tbl = ind_tbl->ind_table,
> +			.pd = priv->pd,
> +		},
> +		&qp_init_attr);
> +#else
>  	qp = mlx5_glue->create_qp_ex
>  		(priv->ctx,
>  		 &(struct ibv_qp_init_attr_ex){
> @@ -1427,6 +1460,7 @@ mlx5_hrxq_new(struct rte_eth_dev *dev,
>  			.rwq_ind_tbl = ind_tbl->ind_table,
>  			.pd = priv->pd,
>  		 });
> +#endif
>  	if (!qp) {
>  		rte_errno = errno;
>  		goto error;
> @@ -1439,6 +1473,7 @@ mlx5_hrxq_new(struct rte_eth_dev *dev,
>  	hrxq->rss_key_len = rss_key_len;
>  	hrxq->hash_fields = hash_fields;
>  	hrxq->tunnel = tunnel;
> +	hrxq->rss_level = rss_level;
>  	memcpy(hrxq->rss_key, rss_key, rss_key_len);
>  	rte_atomic32_inc(&hrxq->refcnt);
>  	LIST_INSERT_HEAD(&priv->hrxqs, hrxq, next);
> @@ -1448,6 +1483,8 @@ mlx5_hrxq_new(struct rte_eth_dev *dev,
>  	return hrxq;
>  error:
>  	err = rte_errno; /* Save rte_errno before cleanup. */
> +	DRV_LOG(ERR, "port %u: Error creating Hash Rx queue",
> +		dev->data->port_id);

Developer log, please remove it, for the user the flow won't be created
with the correct error reported.

>  	mlx5_ind_table_ibv_release(dev, ind_tbl);
>  	if (qp)
>  		claim_zero(mlx5_glue->destroy_qp(qp));
> @@ -1469,6 +1506,8 @@ mlx5_hrxq_new(struct rte_eth_dev *dev,
>   *   Number of queues.
>   * @param tunnel
>   *   Tunnel type.
> + * @param rss_level
> + *   RSS hash on tunnel level
>   *
>   * @return
>   *   An hash Rx queue on success.
> @@ -1477,7 +1516,8 @@ struct mlx5_hrxq *
>  mlx5_hrxq_get(struct rte_eth_dev *dev,
>  	      const uint8_t *rss_key, uint32_t rss_key_len,
>  	      uint64_t hash_fields,
> -	      const uint16_t *queues, uint32_t queues_n, uint32_t tunnel)
> +	      const uint16_t *queues, uint32_t queues_n,
> +	      uint32_t tunnel, uint32_t rss_level)

Dito.

>  {
>  	struct priv *priv = dev->data->dev_private;
>  	struct mlx5_hrxq *hrxq;
> @@ -1494,6 +1534,8 @@ mlx5_hrxq_get(struct rte_eth_dev *dev,
>  			continue;
>  		if (hrxq->tunnel != tunnel)
>  			continue;
> +		if (hrxq->rss_level != rss_level)
> +			continue;
>  		ind_tbl = mlx5_ind_table_ibv_get(dev, queues, queues_n);
>  		if (!ind_tbl)
>  			continue;
> diff --git a/drivers/net/mlx5/mlx5_rxtx.h b/drivers/net/mlx5/mlx5_rxtx.h
> index d35605b55..62cf55109 100644
> --- a/drivers/net/mlx5/mlx5_rxtx.h
> +++ b/drivers/net/mlx5/mlx5_rxtx.h
> @@ -147,6 +147,7 @@ struct mlx5_hrxq {
>  	struct ibv_qp *qp; /* Verbs queue pair. */
>  	uint64_t hash_fields; /* Verbs Hash fields. */
>  	uint32_t tunnel; /* Tunnel type. */
> +	uint32_t rss_level; /* RSS on tunnel level. */
>  	uint32_t rss_key_len; /* Hash key length in bytes. */
>  	uint8_t rss_key[]; /* Hash key. */
>  };
> @@ -251,12 +252,12 @@ struct mlx5_hrxq *mlx5_hrxq_new(struct rte_eth_dev *dev,
>  				const uint8_t *rss_key, uint32_t rss_key_len,
>  				uint64_t hash_fields,
>  				const uint16_t *queues, uint32_t queues_n,
> -				uint32_t tunnel);
> +				uint32_t tunnel, uint32_t rss_level);
>  struct mlx5_hrxq *mlx5_hrxq_get(struct rte_eth_dev *dev,
>  				const uint8_t *rss_key, uint32_t rss_key_len,
>  				uint64_t hash_fields,
>  				const uint16_t *queues, uint32_t queues_n,
> -				uint32_t tunnel);
> +				uint32_t tunnel, uint32_t rss_level);
>  int mlx5_hrxq_release(struct rte_eth_dev *dev, struct mlx5_hrxq *hxrq);
>  int mlx5_hrxq_ibv_verify(struct rte_eth_dev *dev);
>  uint64_t mlx5_get_rx_port_offloads(void);
> -- 
> 2.13.3
> 

Thanks,

-- 
Nélio Laranjeiro
6WIND