enic: Add Accelerated RFS support

This patch adds supports for Accelerated Receive Flow Steering.

When the desired rx is different from current rq, for a flow, kernel calls the
driver function enic_rx_flow_steer(). enic_rx_flow_steer adds a IP-TCP/UDP
hardware filter.

Driver registers a timer function enic_flow_may_expire. This function is called
every HZ/4 seconds. In this function we check if the added filter has expired
by calling rps_may_expire_flow(). If the flow has expired, it removes the hw
filter.

As of now adaptor supports only IPv4 - TCP/UDP filters.

Signed-off-by: Govindarajulu Varadarajan <[email protected]>
Signed-off-by: David S. Miller <[email protected]>

Author: Govindarajulu Varadarajan

drivers/net/ethernet/cisco/enic/enic.h

@@ -99,6 +99,44 @@ struct enic_port_profile {
 	u8 mac_addr[ETH_ALEN];
 };
 
+#ifdef CONFIG_RFS_ACCEL
+/* enic_rfs_fltr_node - rfs filter node in hash table
+ *	@@keys: IPv4 5 tuple
+ *	@flow_id: flow_id of clsf filter provided by kernel
+ *	@fltr_id: filter id of clsf filter returned by adaptor
+ *	@rq_id: desired rq index
+ *	@node: hlist_node
+ */
+struct enic_rfs_fltr_node {
+	struct flow_keys keys;
+	u32 flow_id;
+	u16 fltr_id;
+	u16 rq_id;
+	struct hlist_node node;
+};
+
+/* enic_rfs_flw_tbl - rfs flow table
+ *	@max: Maximum number of filters vNIC supports
+ *	@free: Number of free filters available
+ *	@toclean: hash table index to clean next
+ *	@ht_head: hash table list head
+ *	@lock: spin lock
+ *	@rfs_may_expire: timer function for enic_rps_may_expire_flow
+ */
+struct enic_rfs_flw_tbl {
+	u16 max;
+	int free;
+
+#define ENIC_RFS_FLW_BITSHIFT	(10)
+#define ENIC_RFS_FLW_MASK	((1 << ENIC_RFS_FLW_BITSHIFT) - 1)
+	u16 toclean:ENIC_RFS_FLW_BITSHIFT;
+	struct hlist_head ht_head[1 << ENIC_RFS_FLW_BITSHIFT];
+	spinlock_t lock;
+	struct timer_list rfs_may_expire;
+};
+
+#endif /* CONFIG_RFS_ACCEL */
+
 /* Per-instance private data structure */
 struct enic {
 	struct net_device *netdev;
@@ -150,6 +188,9 @@ struct enic {
 	/* completion queue cache line section */
 	____cacheline_aligned struct vnic_cq cq[ENIC_CQ_MAX];
 	unsigned int cq_count;
+#ifdef CONFIG_RFS_ACCEL
+	struct enic_rfs_flw_tbl rfs_h;
+#endif
 };
 
 static inline struct device *enic_get_dev(struct enic *enic)

drivers/net/ethernet/cisco/enic/enic_clsf.c

@@ -64,3 +64,216 @@ int enic_delfltr(struct enic *enic, u16 filter_id)
 
 	return ret;
 }
+
+#ifdef CONFIG_RFS_ACCEL
+void enic_flow_may_expire(unsigned long data)
+{
+	struct enic *enic = (struct enic *)data;
+	bool res;
+	int j;
+
+	spin_lock(&enic->rfs_h.lock);
+	for (j = 0; j < ENIC_CLSF_EXPIRE_COUNT; j++) {
+		struct hlist_head *hhead;
+		struct hlist_node *tmp;
+		struct enic_rfs_fltr_node *n;
+
+		hhead = &enic->rfs_h.ht_head[enic->rfs_h.toclean++];
+		hlist_for_each_entry_safe(n, tmp, hhead, node) {
+			res = rps_may_expire_flow(enic->netdev, n->rq_id,
+						  n->flow_id, n->fltr_id);
+			if (res) {
+				res = enic_delfltr(enic, n->fltr_id);
+				if (unlikely(res))
+					continue;
+				hlist_del(&n->node);
+				kfree(n);
+				enic->rfs_h.free++;
+			}
+		}
+	}
+	spin_unlock(&enic->rfs_h.lock);
+	mod_timer(&enic->rfs_h.rfs_may_expire, jiffies + HZ/4);
+}
+
+/* enic_rfs_flw_tbl_init - initialize enic->rfs_h members
+ *	@enic: enic data
+ */
+void enic_rfs_flw_tbl_init(struct enic *enic)
+{
+	int i;
+
+	spin_lock_init(&enic->rfs_h.lock);
+	for (i = 0; i <= ENIC_RFS_FLW_MASK; i++)
+		INIT_HLIST_HEAD(&enic->rfs_h.ht_head[i]);
+	enic->rfs_h.max = enic->config.num_arfs;
+	enic->rfs_h.free = enic->rfs_h.max;
+	enic->rfs_h.toclean = 0;
+	init_timer(&enic->rfs_h.rfs_may_expire);
+	enic->rfs_h.rfs_may_expire.function = enic_flow_may_expire;
+	enic->rfs_h.rfs_may_expire.data = (unsigned long)enic;
+	mod_timer(&enic->rfs_h.rfs_may_expire, jiffies + HZ/4);
+}
+
+void enic_rfs_flw_tbl_free(struct enic *enic)
+{
+	int i, res;
+
+	del_timer_sync(&enic->rfs_h.rfs_may_expire);
+	spin_lock(&enic->rfs_h.lock);
+	enic->rfs_h.free = 0;
+	for (i = 0; i < (1 << ENIC_RFS_FLW_BITSHIFT); i++) {
+		struct hlist_head *hhead;
+		struct hlist_node *tmp;
+		struct enic_rfs_fltr_node *n;
+
+		hhead = &enic->rfs_h.ht_head[i];
+		hlist_for_each_entry_safe(n, tmp, hhead, node) {
+			enic_delfltr(enic, n->fltr_id);
+			hlist_del(&n->node);
+			kfree(n);
+		}
+	}
+	spin_unlock(&enic->rfs_h.lock);
+}
+
+static struct enic_rfs_fltr_node *htbl_key_search(struct hlist_head *h,
+						  struct flow_keys *k)
+{
+	struct enic_rfs_fltr_node *tpos;
+
+	hlist_for_each_entry(tpos, h, node)
+		if (tpos->keys.src == k->src &&
+		    tpos->keys.dst == k->dst &&
+		    tpos->keys.ports == k->ports &&
+		    tpos->keys.ip_proto == k->ip_proto &&
+		    tpos->keys.n_proto == k->n_proto)
+			return tpos;
+	return NULL;
+}
+
+int enic_rx_flow_steer(struct net_device *dev, const struct sk_buff *skb,
+		       u16 rxq_index, u32 flow_id)
+{
+	struct flow_keys keys;
+	struct enic_rfs_fltr_node *n;
+	struct enic *enic;
+	u16 tbl_idx;
+	int res, i;
+
+	enic = netdev_priv(dev);
+	res = skb_flow_dissect(skb, &keys);
+	if (!res || keys.n_proto != htons(ETH_P_IP) ||
+	    (keys.ip_proto != IPPROTO_TCP && keys.ip_proto != IPPROTO_UDP))
+		return -EPROTONOSUPPORT;
+
+	tbl_idx = skb_get_hash_raw(skb) & ENIC_RFS_FLW_MASK;
+	spin_lock(&enic->rfs_h.lock);
+	n = htbl_key_search(&enic->rfs_h.ht_head[tbl_idx], &keys);
+
+	if (n) { /* entry already present  */
+		if (rxq_index == n->rq_id) {
+			res = -EEXIST;
+			goto ret_unlock;
+		}
+
+		/* desired rq changed for the flow, we need to delete
+		 * old fltr and add new one
+		 *
+		 * The moment we delete the fltr, the upcoming pkts
+		 * are put it default rq based on rss. When we add
+		 * new filter, upcoming pkts are put in desired queue.
+		 * This could cause ooo pkts.
+		 *
+		 * Lets 1st try adding new fltr and then del old one.
+		 */
+		i = --enic->rfs_h.free;
+		/* clsf tbl is full, we have to del old fltr first*/
+		if (unlikely(i < 0)) {
+			enic->rfs_h.free++;
+			res = enic_delfltr(enic, n->fltr_id);
+			if (unlikely(res < 0))
+				goto ret_unlock;
+			res = enic_addfltr_5t(enic, &keys, rxq_index);
+			if (res < 0) {
+				hlist_del(&n->node);
+				enic->rfs_h.free++;
+				goto ret_unlock;
+			}
+		/* add new fltr 1st then del old fltr */
+		} else {
+			int ret;
+
+			res = enic_addfltr_5t(enic, &keys, rxq_index);
+			if (res < 0) {
+				enic->rfs_h.free++;
+				goto ret_unlock;
+			}
+			ret = enic_delfltr(enic, n->fltr_id);
+			/* deleting old fltr failed. Add old fltr to list.
+			 * enic_flow_may_expire() will try to delete it later.
+			 */
+			if (unlikely(ret < 0)) {
+				struct enic_rfs_fltr_node *d;
+				struct hlist_head *head;
+
+				head = &enic->rfs_h.ht_head[tbl_idx];
+				d = kmalloc(sizeof(*d), GFP_ATOMIC);
+				if (d) {
+					d->fltr_id = n->fltr_id;
+					INIT_HLIST_NODE(&d->node);
+					hlist_add_head(&d->node, head);
+				}
+			} else {
+				enic->rfs_h.free++;
+			}
+		}
+		n->rq_id = rxq_index;
+		n->fltr_id = res;
+		n->flow_id = flow_id;
+	/* entry not present */
+	} else {
+		i = --enic->rfs_h.free;
+		if (i <= 0) {
+			enic->rfs_h.free++;
+			res = -EBUSY;
+			goto ret_unlock;
+		}
+
+		n = kmalloc(sizeof(*n), GFP_ATOMIC);
+		if (!n) {
+			res = -ENOMEM;
+			enic->rfs_h.free++;
+			goto ret_unlock;
+		}
+
+		res = enic_addfltr_5t(enic, &keys, rxq_index);
+		if (res < 0) {
+			kfree(n);
+			enic->rfs_h.free++;
+			goto ret_unlock;
+		}
+		n->rq_id = rxq_index;
+		n->fltr_id = res;
+		n->flow_id = flow_id;
+		n->keys = keys;
+		INIT_HLIST_NODE(&n->node);
+		hlist_add_head(&n->node, &enic->rfs_h.ht_head[tbl_idx]);
+	}
+
+ret_unlock:
+	spin_unlock(&enic->rfs_h.lock);
+	return res;
+}
+
+#else
+
+void enic_rfs_flw_tbl_init(struct enic *enic)
+{
+}
+
+void enic_rfs_flw_tbl_free(struct enic *enic)
+{
+}
+
+#endif /* CONFIG_RFS_ACCEL */

drivers/net/ethernet/cisco/enic/enic_clsf.h

@@ -4,7 +4,16 @@
 #include "vnic_dev.h"
 #include "enic.h"
 
+#define ENIC_CLSF_EXPIRE_COUNT 128
+
 int enic_addfltr_5t(struct enic *enic, struct flow_keys *keys, u16 rq);
 int enic_delfltr(struct enic *enic, u16 filter_id);
 
+#ifdef CONFIG_RFS_ACCEL
+void enic_rfs_flw_tbl_init(struct enic *enic);
+void enic_rfs_flw_tbl_free(struct enic *enic);
+int enic_rx_flow_steer(struct net_device *dev, const struct sk_buff *skb,
+		       u16 rxq_index, u32 flow_id);
+#endif /* CONFIG_RFS_ACCEL */
+
 #endif /* _ENIC_CLSF_H_ */

drivers/net/ethernet/cisco/enic/enic_main.c

@@ -52,6 +52,7 @@
 #include "enic.h"
 #include "enic_dev.h"
 #include "enic_pp.h"
+#include "enic_clsf.h"
 
 #define ENIC_NOTIFY_TIMER_PERIOD	(2 * HZ)
 #define WQ_ENET_MAX_DESC_LEN		(1 << WQ_ENET_LEN_BITS)
@@ -1546,6 +1547,7 @@ static int enic_open(struct net_device *netdev)
 		vnic_intr_unmask(&enic->intr[i]);
 
 	enic_notify_timer_start(enic);
+	enic_rfs_flw_tbl_init(enic);
 
 	return 0;
 
@@ -1572,6 +1574,7 @@ static int enic_stop(struct net_device *netdev)
 	enic_synchronize_irqs(enic);
 
 	del_timer_sync(&enic->notify_timer);
+	enic_rfs_flw_tbl_free(enic);
 
 	enic_dev_disable(enic);
 
@@ -2064,6 +2067,9 @@ static const struct net_device_ops enic_netdev_dynamic_ops = {
 #ifdef CONFIG_NET_POLL_CONTROLLER
 	.ndo_poll_controller	= enic_poll_controller,
 #endif
+#ifdef CONFIG_RFS_ACCEL
+	.ndo_rx_flow_steer	= enic_rx_flow_steer,
+#endif
 };
 
 static const struct net_device_ops enic_netdev_ops = {
@@ -2084,6 +2090,9 @@ static const struct net_device_ops enic_netdev_ops = {
 #ifdef CONFIG_NET_POLL_CONTROLLER
 	.ndo_poll_controller	= enic_poll_controller,
 #endif
+#ifdef CONFIG_RFS_ACCEL
+	.ndo_rx_flow_steer	= enic_rx_flow_steer,
+#endif
 };
 
 static void enic_dev_deinit(struct enic *enic)
@@ -2429,6 +2438,10 @@ static int enic_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
 
 	netdev->features |= netdev->hw_features;
 
+#ifdef CONFIG_RFS_ACCEL
+	netdev->hw_features |= NETIF_F_NTUPLE;
+#endif
+
 	if (using_dac)
 		netdev->features |= NETIF_F_HIGHDMA;
 

drivers/net/ethernet/cisco/enic/enic_res.c

@@ -71,6 +71,7 @@ int enic_get_vnic_config(struct enic *enic)
 	GET_CONFIG(intr_mode);
 	GET_CONFIG(intr_timer_usec);
 	GET_CONFIG(loop_tag);
+	GET_CONFIG(num_arfs);
 
 	c->wq_desc_count =
 		min_t(u32, ENIC_MAX_WQ_DESCS,

drivers/net/ethernet/cisco/enic/vnic_enet.h

@@ -32,6 +32,8 @@ struct vnic_enet_config {
 	char devname[16];
 	u32 intr_timer_usec;
 	u16 loop_tag;
+	u16 vf_rq_count;
+	u16 num_arfs;
 };
 
 #define VENETF_TSO		0x1	/* TSO enabled */