net/mlx5e: Add accelerated RFS support

Implement ndo_rx_flow_steer ndo.
A new flow steering rule will be composed from the
skb 4-tuple and added to the hardware aRFS flow table.

Each rule is stored in an internal hash table, if such
skb 4-tuple rule already exists we update the corresponding
hardware steering rule with the new destination.

For garbage collection rps_may_expire_flow will be
invoked for a limited amount of old rules upon any
ndo_rx_flow_steer invocation.
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/mellanox/mlx5/core/en.h

@@ -448,9 +448,12 @@ struct mlx5e_ttc_table {
 	struct mlx5_flow_rule	 *rules[MLX5E_NUM_TT];
 };
 
+#define ARFS_HASH_SHIFT BITS_PER_BYTE
+#define ARFS_HASH_SIZE BIT(BITS_PER_BYTE)
 struct arfs_table {
 	struct mlx5e_flow_table  ft;
 	struct mlx5_flow_rule    *default_rule;
+	struct hlist_head	 rules_hash[ARFS_HASH_SIZE];
 };
 
 enum  arfs_type {
@@ -463,6 +466,11 @@ enum  arfs_type {
 
 struct mlx5e_arfs_tables {
 	struct arfs_table arfs_tables[ARFS_NUM_TYPES];
+	/* Protect aRFS rules list */
+	spinlock_t                     arfs_lock;
+	struct list_head               rules;
+	int                            last_filter_id;
+	struct workqueue_struct        *wq;
 };
 
 /* NIC prio FTS */
@@ -685,6 +693,8 @@ static inline void mlx5e_arfs_destroy_tables(struct mlx5e_priv *priv) {}
 #else
 int mlx5e_arfs_create_tables(struct mlx5e_priv *priv);
 void mlx5e_arfs_destroy_tables(struct mlx5e_priv *priv);
+int mlx5e_rx_flow_steer(struct net_device *dev, const struct sk_buff *skb,
+			u16 rxq_index, u32 flow_id);
 #endif
 
 u16 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev);

drivers/net/ethernet/mellanox/mlx5/core/en_arfs.c

@@ -30,8 +30,47 @@
  * SOFTWARE.
  */
 
-#include "en.h"
+#include <linux/hash.h>
 #include <linux/mlx5/fs.h>
+#include <linux/ip.h>
+#include <linux/ipv6.h>
+#include "en.h"
+
+struct arfs_tuple {
+	__be16 etype;
+	u8     ip_proto;
+	union {
+		__be32 src_ipv4;
+		struct in6_addr src_ipv6;
+	};
+	union {
+		__be32 dst_ipv4;
+		struct in6_addr dst_ipv6;
+	};
+	__be16 src_port;
+	__be16 dst_port;
+};
+
+struct arfs_rule {
+	struct mlx5e_priv	*priv;
+	struct work_struct      arfs_work;
+	struct mlx5_flow_rule   *rule;
+	struct hlist_node	hlist;
+	int			rxq;
+	/* Flow ID passed to ndo_rx_flow_steer */
+	int			flow_id;
+	/* Filter ID returned by ndo_rx_flow_steer */
+	int			filter_id;
+	struct arfs_tuple	tuple;
+};
+
+#define mlx5e_for_each_arfs_rule(hn, tmp, arfs_tables, i, j) \
+	for (i = 0; i < ARFS_NUM_TYPES; i++) \
+		mlx5e_for_each_hash_arfs_rule(hn, tmp, arfs_tables[i].rules_hash, j)
+
+#define mlx5e_for_each_hash_arfs_rule(hn, tmp, hash, j) \
+	for (j = 0; j < ARFS_HASH_SIZE; j++) \
+		hlist_for_each_entry_safe(hn, tmp, &hash[j], hlist)
 
 static void arfs_destroy_table(struct arfs_table *arfs_t)
 {
@@ -39,12 +78,17 @@ static void arfs_destroy_table(struct arfs_table *arfs_t)
 	mlx5e_destroy_flow_table(&arfs_t->ft);
 }
 
+static void arfs_del_rules(struct mlx5e_priv *priv);
+
 void mlx5e_arfs_destroy_tables(struct mlx5e_priv *priv)
 {
 	int i;
 
 	if (!(priv->netdev->hw_features & NETIF_F_NTUPLE))
 		return;
+
+	arfs_del_rules(priv);
+	destroy_workqueue(priv->fs.arfs.wq);
 	for (i = 0; i < ARFS_NUM_TYPES; i++) {
 		if (!IS_ERR_OR_NULL(priv->fs.arfs.arfs_tables[i].ft.t))
 			arfs_destroy_table(&priv->fs.arfs.arfs_tables[i]);
@@ -239,6 +283,12 @@ int mlx5e_arfs_create_tables(struct mlx5e_priv *priv)
 	if (!(priv->netdev->hw_features & NETIF_F_NTUPLE))
 		return 0;
 
+	spin_lock_init(&priv->fs.arfs.arfs_lock);
+	INIT_LIST_HEAD(&priv->fs.arfs.rules);
+	priv->fs.arfs.wq = create_singlethread_workqueue("mlx5e_arfs");
+	if (!priv->fs.arfs.wq)
+		return -ENOMEM;
+
 	for (i = 0; i < ARFS_NUM_TYPES; i++) {
 		err = arfs_create_table(priv, i);
 		if (err)
@@ -249,3 +299,378 @@ int mlx5e_arfs_create_tables(struct mlx5e_priv *priv)
 	mlx5e_arfs_destroy_tables(priv);
 	return err;
 }
+
+#define MLX5E_ARFS_EXPIRY_QUOTA 60
+
+static void arfs_may_expire_flow(struct mlx5e_priv *priv)
+{
+	struct arfs_rule *arfs_rule;
+	struct hlist_node *htmp;
+	int quota = 0;
+	int i;
+	int j;
+
+	HLIST_HEAD(del_list);
+	spin_lock_bh(&priv->fs.arfs.arfs_lock);
+	mlx5e_for_each_arfs_rule(arfs_rule, htmp, priv->fs.arfs.arfs_tables, i, j) {
+		if (quota++ > MLX5E_ARFS_EXPIRY_QUOTA)
+			break;
+		if (!work_pending(&arfs_rule->arfs_work) &&
+		    rps_may_expire_flow(priv->netdev,
+					arfs_rule->rxq, arfs_rule->flow_id,
+					arfs_rule->filter_id)) {
+			hlist_del_init(&arfs_rule->hlist);
+			hlist_add_head(&arfs_rule->hlist, &del_list);
+		}
+	}
+	spin_unlock_bh(&priv->fs.arfs.arfs_lock);
+	hlist_for_each_entry_safe(arfs_rule, htmp, &del_list, hlist) {
+		if (arfs_rule->rule)
+			mlx5_del_flow_rule(arfs_rule->rule);
+		hlist_del(&arfs_rule->hlist);
+		kfree(arfs_rule);
+	}
+}
+
+static void arfs_del_rules(struct mlx5e_priv *priv)
+{
+	struct hlist_node *htmp;
+	struct arfs_rule *rule;
+	int i;
+	int j;
+
+	HLIST_HEAD(del_list);
+	spin_lock_bh(&priv->fs.arfs.arfs_lock);
+	mlx5e_for_each_arfs_rule(rule, htmp, priv->fs.arfs.arfs_tables, i, j) {
+		hlist_del_init(&rule->hlist);
+		hlist_add_head(&rule->hlist, &del_list);
+	}
+	spin_unlock_bh(&priv->fs.arfs.arfs_lock);
+
+	hlist_for_each_entry_safe(rule, htmp, &del_list, hlist) {
+		cancel_work_sync(&rule->arfs_work);
+		if (rule->rule)
+			mlx5_del_flow_rule(rule->rule);
+		hlist_del(&rule->hlist);
+		kfree(rule);
+	}
+}
+
+static struct hlist_head *
+arfs_hash_bucket(struct arfs_table *arfs_t, __be16 src_port,
+		 __be16 dst_port)
+{
+	unsigned long l;
+	int bucket_idx;
+
+	l = (__force unsigned long)src_port |
+	    ((__force unsigned long)dst_port << 2);
+
+	bucket_idx = hash_long(l, ARFS_HASH_SHIFT);
+
+	return &arfs_t->rules_hash[bucket_idx];
+}
+
+static u8 arfs_get_ip_proto(const struct sk_buff *skb)
+{
+	return (skb->protocol == htons(ETH_P_IP)) ?
+		ip_hdr(skb)->protocol : ipv6_hdr(skb)->nexthdr;
+}
+
+static struct arfs_table *arfs_get_table(struct mlx5e_arfs_tables *arfs,
+					 u8 ip_proto, __be16 etype)
+{
+	if (etype == htons(ETH_P_IP) && ip_proto == IPPROTO_TCP)
+		return &arfs->arfs_tables[ARFS_IPV4_TCP];
+	if (etype == htons(ETH_P_IP) && ip_proto == IPPROTO_UDP)
+		return &arfs->arfs_tables[ARFS_IPV4_UDP];
+	if (etype == htons(ETH_P_IPV6) && ip_proto == IPPROTO_TCP)
+		return &arfs->arfs_tables[ARFS_IPV6_TCP];
+	if (etype == htons(ETH_P_IPV6) && ip_proto == IPPROTO_UDP)
+		return &arfs->arfs_tables[ARFS_IPV6_UDP];
+
+	return NULL;
+}
+
+static struct mlx5_flow_rule *arfs_add_rule(struct mlx5e_priv *priv,
+					    struct arfs_rule *arfs_rule)
+{
+	struct mlx5e_arfs_tables *arfs = &priv->fs.arfs;
+	struct arfs_tuple *tuple = &arfs_rule->tuple;
+	struct mlx5_flow_rule *rule = NULL;
+	struct mlx5_flow_destination dest;
+	struct arfs_table *arfs_table;
+	u8 match_criteria_enable = 0;
+	struct mlx5_flow_table *ft;
+	u32 *match_criteria;
+	u32 *match_value;
+	int err = 0;
+
+	match_value	= mlx5_vzalloc(MLX5_ST_SZ_BYTES(fte_match_param));
+	match_criteria	= mlx5_vzalloc(MLX5_ST_SZ_BYTES(fte_match_param));
+	if (!match_value || !match_criteria) {
+		netdev_err(priv->netdev, "%s: alloc failed\n", __func__);
+		err = -ENOMEM;
+		goto out;
+	}
+	match_criteria_enable = MLX5_MATCH_OUTER_HEADERS;
+	MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+			 outer_headers.ethertype);
+	MLX5_SET(fte_match_param, match_value, outer_headers.ethertype,
+		 ntohs(tuple->etype));
+	arfs_table = arfs_get_table(arfs, tuple->ip_proto, tuple->etype);
+	if (!arfs_table) {
+		err = -EINVAL;
+		goto out;
+	}
+
+	ft = arfs_table->ft.t;
+	if (tuple->ip_proto == IPPROTO_TCP) {
+		MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+				 outer_headers.tcp_dport);
+		MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+				 outer_headers.tcp_sport);
+		MLX5_SET(fte_match_param, match_value, outer_headers.tcp_dport,
+			 ntohs(tuple->dst_port));
+		MLX5_SET(fte_match_param, match_value, outer_headers.tcp_sport,
+			 ntohs(tuple->src_port));
+	} else {
+		MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+				 outer_headers.udp_dport);
+		MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+				 outer_headers.udp_sport);
+		MLX5_SET(fte_match_param, match_value, outer_headers.udp_dport,
+			 ntohs(tuple->dst_port));
+		MLX5_SET(fte_match_param, match_value, outer_headers.udp_sport,
+			 ntohs(tuple->src_port));
+	}
+	if (tuple->etype == htons(ETH_P_IP)) {
+		memcpy(MLX5_ADDR_OF(fte_match_param, match_value,
+				    outer_headers.src_ipv4_src_ipv6.ipv4_layout.ipv4),
+		       &tuple->src_ipv4,
+		       4);
+		memcpy(MLX5_ADDR_OF(fte_match_param, match_value,
+				    outer_headers.dst_ipv4_dst_ipv6.ipv4_layout.ipv4),
+		       &tuple->dst_ipv4,
+		       4);
+		MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+				 outer_headers.src_ipv4_src_ipv6.ipv4_layout.ipv4);
+		MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+				 outer_headers.dst_ipv4_dst_ipv6.ipv4_layout.ipv4);
+	} else {
+		memcpy(MLX5_ADDR_OF(fte_match_param, match_value,
+				    outer_headers.src_ipv4_src_ipv6.ipv6_layout.ipv6),
+		       &tuple->src_ipv6,
+		       16);
+		memcpy(MLX5_ADDR_OF(fte_match_param, match_value,
+				    outer_headers.dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
+		       &tuple->dst_ipv6,
+		       16);
+		memset(MLX5_ADDR_OF(fte_match_param, match_criteria,
+				    outer_headers.src_ipv4_src_ipv6.ipv6_layout.ipv6),
+		       0xff,
+		       16);
+		memset(MLX5_ADDR_OF(fte_match_param, match_criteria,
+				    outer_headers.dst_ipv4_dst_ipv6.ipv6_layout.ipv6),
+		       0xff,
+		       16);
+	}
+	dest.type = MLX5_FLOW_DESTINATION_TYPE_TIR;
+	dest.tir_num = priv->direct_tir[arfs_rule->rxq].tirn;
+	rule = mlx5_add_flow_rule(ft, match_criteria_enable, match_criteria,
+				  match_value, MLX5_FLOW_CONTEXT_ACTION_FWD_DEST,
+				  MLX5_FS_DEFAULT_FLOW_TAG,
+				  &dest);
+	if (IS_ERR(rule)) {
+		err = PTR_ERR(rule);
+		netdev_err(priv->netdev, "%s: add rule(filter id=%d, rq idx=%d) failed, err=%d\n",
+			   __func__, arfs_rule->filter_id, arfs_rule->rxq, err);
+	}
+
+out:
+	kvfree(match_criteria);
+	kvfree(match_value);
+	return err ? ERR_PTR(err) : rule;
+}
+
+static void arfs_modify_rule_rq(struct mlx5e_priv *priv,
+				struct mlx5_flow_rule *rule, u16 rxq)
+{
+	struct mlx5_flow_destination dst;
+	int err = 0;
+
+	dst.type = MLX5_FLOW_DESTINATION_TYPE_TIR;
+	dst.tir_num = priv->direct_tir[rxq].tirn;
+	err =  mlx5_modify_rule_destination(rule, &dst);
+	if (err)
+		netdev_warn(priv->netdev,
+			    "Failed to modfiy aRFS rule destination to rq=%d\n", rxq);
+}
+
+static void arfs_handle_work(struct work_struct *work)
+{
+	struct arfs_rule *arfs_rule = container_of(work,
+						   struct arfs_rule,
+						   arfs_work);
+	struct mlx5e_priv *priv = arfs_rule->priv;
+	struct mlx5_flow_rule *rule;
+
+	mutex_lock(&priv->state_lock);
+	if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
+		spin_lock_bh(&priv->fs.arfs.arfs_lock);
+		hlist_del(&arfs_rule->hlist);
+		spin_unlock_bh(&priv->fs.arfs.arfs_lock);
+
+		mutex_unlock(&priv->state_lock);
+		kfree(arfs_rule);
+		goto out;
+	}
+	mutex_unlock(&priv->state_lock);
+
+	if (!arfs_rule->rule) {
+		rule = arfs_add_rule(priv, arfs_rule);
+		if (IS_ERR(rule))
+			goto out;
+		arfs_rule->rule = rule;
+	} else {
+		arfs_modify_rule_rq(priv, arfs_rule->rule,
+				    arfs_rule->rxq);
+	}
+out:
+	arfs_may_expire_flow(priv);
+}
+
+/* return L4 destination port from ip4/6 packets */
+static __be16 arfs_get_dst_port(const struct sk_buff *skb)
+{
+	char *transport_header;
+
+	transport_header = skb_transport_header(skb);
+	if (arfs_get_ip_proto(skb) == IPPROTO_TCP)
+		return ((struct tcphdr *)transport_header)->dest;
+	return ((struct udphdr *)transport_header)->dest;
+}
+
+/* return L4 source port from ip4/6 packets */
+static __be16 arfs_get_src_port(const struct sk_buff *skb)
+{
+	char *transport_header;
+
+	transport_header = skb_transport_header(skb);
+	if (arfs_get_ip_proto(skb) == IPPROTO_TCP)
+		return ((struct tcphdr *)transport_header)->source;
+	return ((struct udphdr *)transport_header)->source;
+}
+
+static struct arfs_rule *arfs_alloc_rule(struct mlx5e_priv *priv,
+					 struct arfs_table *arfs_t,
+					 const struct sk_buff *skb,
+					 u16 rxq, u32 flow_id)
+{
+	struct arfs_rule *rule;
+	struct arfs_tuple *tuple;
+
+	rule = kzalloc(sizeof(*rule), GFP_ATOMIC);
+	if (!rule)
+		return NULL;
+
+	rule->priv = priv;
+	rule->rxq = rxq;
+	INIT_WORK(&rule->arfs_work, arfs_handle_work);
+
+	tuple = &rule->tuple;
+	tuple->etype = skb->protocol;
+	if (tuple->etype == htons(ETH_P_IP)) {
+		tuple->src_ipv4 = ip_hdr(skb)->saddr;
+		tuple->dst_ipv4 = ip_hdr(skb)->daddr;
+	} else {
+		memcpy(&tuple->src_ipv6, &ipv6_hdr(skb)->saddr,
+		       sizeof(struct in6_addr));
+		memcpy(&tuple->dst_ipv6, &ipv6_hdr(skb)->daddr,
+		       sizeof(struct in6_addr));
+	}
+	tuple->ip_proto = arfs_get_ip_proto(skb);
+	tuple->src_port = arfs_get_src_port(skb);
+	tuple->dst_port = arfs_get_dst_port(skb);
+
+	rule->flow_id = flow_id;
+	rule->filter_id = priv->fs.arfs.last_filter_id++ % RPS_NO_FILTER;
+
+	hlist_add_head(&rule->hlist,
+		       arfs_hash_bucket(arfs_t, tuple->src_port,
+					tuple->dst_port));
+	return rule;
+}
+
+static bool arfs_cmp_ips(struct arfs_tuple *tuple,
+			 const struct sk_buff *skb)
+{
+	if (tuple->etype == htons(ETH_P_IP) &&
+	    tuple->src_ipv4 == ip_hdr(skb)->saddr &&
+	    tuple->dst_ipv4 == ip_hdr(skb)->daddr)
+		return true;
+	if (tuple->etype == htons(ETH_P_IPV6) &&
+	    (!memcmp(&tuple->src_ipv6, &ipv6_hdr(skb)->saddr,
+		     sizeof(struct in6_addr))) &&
+	    (!memcmp(&tuple->dst_ipv6, &ipv6_hdr(skb)->daddr,
+		     sizeof(struct in6_addr))))
+		return true;
+	return false;
+}
+
+static struct arfs_rule *arfs_find_rule(struct arfs_table *arfs_t,
+					const struct sk_buff *skb)
+{
+	struct arfs_rule *arfs_rule;
+	struct hlist_head *head;
+	__be16 src_port = arfs_get_src_port(skb);
+	__be16 dst_port = arfs_get_dst_port(skb);
+
+	head = arfs_hash_bucket(arfs_t, src_port, dst_port);
+	hlist_for_each_entry(arfs_rule, head, hlist) {
+		if (arfs_rule->tuple.src_port == src_port &&
+		    arfs_rule->tuple.dst_port == dst_port &&
+		    arfs_cmp_ips(&arfs_rule->tuple, skb)) {
+			return arfs_rule;
+		}
+	}
+
+	return NULL;
+}
+
+int mlx5e_rx_flow_steer(struct net_device *dev, const struct sk_buff *skb,
+			u16 rxq_index, u32 flow_id)
+{
+	struct mlx5e_priv *priv = netdev_priv(dev);
+	struct mlx5e_arfs_tables *arfs = &priv->fs.arfs;
+	struct arfs_table *arfs_t;
+	struct arfs_rule *arfs_rule;
+
+	if (skb->protocol != htons(ETH_P_IP) &&
+	    skb->protocol != htons(ETH_P_IPV6))
+		return -EPROTONOSUPPORT;
+
+	arfs_t = arfs_get_table(arfs, arfs_get_ip_proto(skb), skb->protocol);
+	if (!arfs_t)
+		return -EPROTONOSUPPORT;
+
+	spin_lock_bh(&arfs->arfs_lock);
+	arfs_rule = arfs_find_rule(arfs_t, skb);
+	if (arfs_rule) {
+		if (arfs_rule->rxq == rxq_index) {
+			spin_unlock_bh(&arfs->arfs_lock);
+			return arfs_rule->filter_id;
+		}
+		arfs_rule->rxq = rxq_index;
+	} else {
+		arfs_rule = arfs_alloc_rule(priv, arfs_t, skb,
+					    rxq_index, flow_id);
+		if (!arfs_rule) {
+			spin_unlock_bh(&arfs->arfs_lock);
+			return -ENOMEM;
+		}
+	}
+	queue_work(priv->fs.arfs.wq, &arfs_rule->arfs_work);
+	spin_unlock_bh(&arfs->arfs_lock);
+	return arfs_rule->filter_id;
+}