sfc: offload conntrack flow entries (match only) from CT zones

No handling yet for FLOW_ACTION_MANGLE (NAT or NAPT) actions.
Reviewed-by: Pieter Jansen van Vuuren <pieter.jansen-van-vuuren@amd.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Edward Cree <ecree.xilinx@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/sfc/tc.h

@@ -18,12 +18,10 @@
 
 #define IS_ALL_ONES(v)	(!(typeof (v))~(v))
 
-#ifdef CONFIG_IPV6
 static inline bool efx_ipv6_addr_all_ones(struct in6_addr *addr)
 {
 	return !memchr_inv(addr, 0xff, sizeof(*addr));
 }
-#endif
 
 struct efx_tc_encap_action; /* see tc_encap_actions.h */
 
@@ -197,6 +195,7 @@ struct efx_tc_table_ct { /* TABLE_ID_CONNTRACK_TABLE */
  * @encap_match_ht: Hashtable of TC encap matches
  * @match_action_ht: Hashtable of TC match-action rules
  * @ct_zone_ht: Hashtable of TC conntrack flowtable bindings
+ * @ct_ht: Hashtable of TC conntrack flow entries
  * @neigh_ht: Hashtable of neighbour watches (&struct efx_neigh_binder)
  * @meta_ct: MAE table layout for conntrack table
  * @reps_mport_id: MAE port allocated for representor RX
@@ -230,6 +229,7 @@ struct efx_tc_state {
 	struct rhashtable encap_match_ht;
 	struct rhashtable match_action_ht;
 	struct rhashtable ct_zone_ht;
+	struct rhashtable ct_ht;
 	struct rhashtable neigh_ht;
 	struct efx_tc_table_ct meta_ct;
 	u32 reps_mport_id, reps_mport_vport_id;

drivers/net/ethernet/sfc/tc_conntrack.c

@@ -21,6 +21,12 @@ static const struct rhashtable_params efx_tc_ct_zone_ht_params = {
 	.head_offset	= offsetof(struct efx_tc_ct_zone, linkage),
 };
 
+static const struct rhashtable_params efx_tc_ct_ht_params = {
+	.key_len	= offsetof(struct efx_tc_ct_entry, linkage),
+	.key_offset	= 0,
+	.head_offset	= offsetof(struct efx_tc_ct_entry, linkage),
+};
+
 static void efx_tc_ct_zone_free(void *ptr, void *arg)
 {
 	struct efx_tc_ct_zone *zone = ptr;
@@ -34,24 +40,420 @@ static void efx_tc_ct_zone_free(void *ptr, void *arg)
 	kfree(zone);
 }
 
+static void efx_tc_ct_free(void *ptr, void *arg)
+{
+	struct efx_tc_ct_entry *conn = ptr;
+	struct efx_nic *efx = arg;
+
+	netif_err(efx, drv, efx->net_dev,
+		  "tc ct_entry %lx still present at teardown\n",
+		  conn->cookie);
+
+	/* We can release the counter, but we can't remove the CT itself
+	 * from hardware because the table meta is already gone.
+	 */
+	efx_tc_flower_release_counter(efx, conn->cnt);
+	kfree(conn);
+}
+
 int efx_tc_init_conntrack(struct efx_nic *efx)
 {
 	int rc;
 
 	rc = rhashtable_init(&efx->tc->ct_zone_ht, &efx_tc_ct_zone_ht_params);
 	if (rc < 0)
-		return rc;
+		goto fail_ct_zone_ht;
+	rc = rhashtable_init(&efx->tc->ct_ht, &efx_tc_ct_ht_params);
+	if (rc < 0)
+		goto fail_ct_ht;
 	return 0;
+fail_ct_ht:
+	rhashtable_destroy(&efx->tc->ct_zone_ht);
+fail_ct_zone_ht:
+	return rc;
 }
 
 void efx_tc_fini_conntrack(struct efx_nic *efx)
 {
 	rhashtable_free_and_destroy(&efx->tc->ct_zone_ht, efx_tc_ct_zone_free, NULL);
+	rhashtable_free_and_destroy(&efx->tc->ct_ht, efx_tc_ct_free, efx);
+}
+
+#define EFX_NF_TCP_FLAG(flg)	cpu_to_be16(be32_to_cpu(TCP_FLAG_##flg) >> 16)
+
+static int efx_tc_ct_parse_match(struct efx_nic *efx, struct flow_rule *fr,
+				 struct efx_tc_ct_entry *conn)
+{
+	struct flow_dissector *dissector = fr->match.dissector;
+	unsigned char ipv = 0;
+	bool tcp = false;
+
+	if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_CONTROL)) {
+		struct flow_match_control fm;
+
+		flow_rule_match_control(fr, &fm);
+		if (IS_ALL_ONES(fm.mask->addr_type))
+			switch (fm.key->addr_type) {
+			case FLOW_DISSECTOR_KEY_IPV4_ADDRS:
+				ipv = 4;
+				break;
+			case FLOW_DISSECTOR_KEY_IPV6_ADDRS:
+				ipv = 6;
+				break;
+			default:
+				break;
+			}
+	}
+
+	if (!ipv) {
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Conntrack missing ipv specification\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (dissector->used_keys &
+	    ~(BIT_ULL(FLOW_DISSECTOR_KEY_CONTROL) |
+	      BIT_ULL(FLOW_DISSECTOR_KEY_BASIC) |
+	      BIT_ULL(FLOW_DISSECTOR_KEY_IPV4_ADDRS) |
+	      BIT_ULL(FLOW_DISSECTOR_KEY_IPV6_ADDRS) |
+	      BIT_ULL(FLOW_DISSECTOR_KEY_PORTS) |
+	      BIT_ULL(FLOW_DISSECTOR_KEY_TCP) |
+	      BIT_ULL(FLOW_DISSECTOR_KEY_META))) {
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Unsupported conntrack keys %#llx\n",
+			  dissector->used_keys);
+		return -EOPNOTSUPP;
+	}
+
+	if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_BASIC)) {
+		struct flow_match_basic fm;
+
+		flow_rule_match_basic(fr, &fm);
+		if (!IS_ALL_ONES(fm.mask->n_proto)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack eth_proto is not exact-match; mask %04x\n",
+				   ntohs(fm.mask->n_proto));
+			return -EOPNOTSUPP;
+		}
+		conn->eth_proto = fm.key->n_proto;
+		if (conn->eth_proto != (ipv == 4 ? htons(ETH_P_IP)
+						 : htons(ETH_P_IPV6))) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack eth_proto is not IPv%u, is %04x\n",
+				   ipv, ntohs(conn->eth_proto));
+			return -EOPNOTSUPP;
+		}
+		if (!IS_ALL_ONES(fm.mask->ip_proto)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ip_proto is not exact-match; mask %02x\n",
+				   fm.mask->ip_proto);
+			return -EOPNOTSUPP;
+		}
+		conn->ip_proto = fm.key->ip_proto;
+		switch (conn->ip_proto) {
+		case IPPROTO_TCP:
+			tcp = true;
+			break;
+		case IPPROTO_UDP:
+			break;
+		default:
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ip_proto not TCP or UDP, is %02x\n",
+				   conn->ip_proto);
+			return -EOPNOTSUPP;
+		}
+	} else {
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Conntrack missing eth_proto, ip_proto\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (ipv == 4 && flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_IPV4_ADDRS)) {
+		struct flow_match_ipv4_addrs fm;
+
+		flow_rule_match_ipv4_addrs(fr, &fm);
+		if (!IS_ALL_ONES(fm.mask->src)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ipv4.src is not exact-match; mask %08x\n",
+				   ntohl(fm.mask->src));
+			return -EOPNOTSUPP;
+		}
+		conn->src_ip = fm.key->src;
+		if (!IS_ALL_ONES(fm.mask->dst)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ipv4.dst is not exact-match; mask %08x\n",
+				   ntohl(fm.mask->dst));
+			return -EOPNOTSUPP;
+		}
+		conn->dst_ip = fm.key->dst;
+	} else if (ipv == 6 && flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_IPV6_ADDRS)) {
+		struct flow_match_ipv6_addrs fm;
+
+		flow_rule_match_ipv6_addrs(fr, &fm);
+		if (!efx_ipv6_addr_all_ones(&fm.mask->src)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ipv6.src is not exact-match; mask %pI6\n",
+				   &fm.mask->src);
+			return -EOPNOTSUPP;
+		}
+		conn->src_ip6 = fm.key->src;
+		if (!efx_ipv6_addr_all_ones(&fm.mask->dst)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ipv6.dst is not exact-match; mask %pI6\n",
+				   &fm.mask->dst);
+			return -EOPNOTSUPP;
+		}
+		conn->dst_ip6 = fm.key->dst;
+	} else {
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Conntrack missing IPv%u addrs\n", ipv);
+		return -EOPNOTSUPP;
+	}
+
+	if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_PORTS)) {
+		struct flow_match_ports fm;
+
+		flow_rule_match_ports(fr, &fm);
+		if (!IS_ALL_ONES(fm.mask->src)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ports.src is not exact-match; mask %04x\n",
+				   ntohs(fm.mask->src));
+			return -EOPNOTSUPP;
+		}
+		conn->l4_sport = fm.key->src;
+		if (!IS_ALL_ONES(fm.mask->dst)) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack ports.dst is not exact-match; mask %04x\n",
+				   ntohs(fm.mask->dst));
+			return -EOPNOTSUPP;
+		}
+		conn->l4_dport = fm.key->dst;
+	} else {
+		netif_dbg(efx, drv, efx->net_dev, "Conntrack missing L4 ports\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (flow_rule_match_key(fr, FLOW_DISSECTOR_KEY_TCP)) {
+		__be16 tcp_interesting_flags;
+		struct flow_match_tcp fm;
+
+		if (!tcp) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Conntrack matching on TCP keys but ipproto is not tcp\n");
+			return -EOPNOTSUPP;
+		}
+		flow_rule_match_tcp(fr, &fm);
+		tcp_interesting_flags = EFX_NF_TCP_FLAG(SYN) |
+					EFX_NF_TCP_FLAG(RST) |
+					EFX_NF_TCP_FLAG(FIN);
+		/* If any of the tcp_interesting_flags is set, we always
+		 * inhibit CT lookup in LHS (so SW can update CT table).
+		 */
+		if (fm.key->flags & tcp_interesting_flags) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Unsupported conntrack tcp.flags %04x/%04x\n",
+				   ntohs(fm.key->flags), ntohs(fm.mask->flags));
+			return -EOPNOTSUPP;
+		}
+		/* Other TCP flags cannot be filtered at CT */
+		if (fm.mask->flags & ~tcp_interesting_flags) {
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Unsupported conntrack tcp.flags %04x/%04x\n",
+				   ntohs(fm.key->flags), ntohs(fm.mask->flags));
+			return -EOPNOTSUPP;
+		}
+	}
+
+	return 0;
+}
+
+static int efx_tc_ct_replace(struct efx_tc_ct_zone *ct_zone,
+			     struct flow_cls_offload *tc)
+{
+	struct flow_rule *fr = flow_cls_offload_flow_rule(tc);
+	struct efx_tc_ct_entry *conn, *old;
+	struct efx_nic *efx = ct_zone->efx;
+	const struct flow_action_entry *fa;
+	struct efx_tc_counter *cnt;
+	int rc, i;
+
+	if (WARN_ON(!efx->tc))
+		return -ENETDOWN;
+	if (WARN_ON(!efx->tc->up))
+		return -ENETDOWN;
+
+	conn = kzalloc(sizeof(*conn), GFP_USER);
+	if (!conn)
+		return -ENOMEM;
+	conn->cookie = tc->cookie;
+	old = rhashtable_lookup_get_insert_fast(&efx->tc->ct_ht,
+						&conn->linkage,
+						efx_tc_ct_ht_params);
+	if (old) {
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Already offloaded conntrack (cookie %lx)\n", tc->cookie);
+		rc = -EEXIST;
+		goto release;
+	}
+
+	/* Parse match */
+	conn->zone = ct_zone;
+	rc = efx_tc_ct_parse_match(efx, fr, conn);
+	if (rc)
+		goto release;
+
+	/* Parse actions */
+	flow_action_for_each(i, fa, &fr->action) {
+		switch (fa->id) {
+		case FLOW_ACTION_CT_METADATA:
+			conn->mark = fa->ct_metadata.mark;
+			if (memchr_inv(fa->ct_metadata.labels, 0, sizeof(fa->ct_metadata.labels))) {
+				netif_dbg(efx, drv, efx->net_dev,
+					  "Setting CT label not supported\n");
+				rc = -EOPNOTSUPP;
+				goto release;
+			}
+			break;
+		default:
+			netif_dbg(efx, drv, efx->net_dev,
+				  "Unhandled action %u for conntrack\n", fa->id);
+			rc = -EOPNOTSUPP;
+			goto release;
+		}
+	}
+
+	/* fill in defaults for unmangled values */
+	conn->nat_ip = conn->dnat ? conn->dst_ip : conn->src_ip;
+	conn->l4_natport = conn->dnat ? conn->l4_dport : conn->l4_sport;
+
+	cnt = efx_tc_flower_allocate_counter(efx, EFX_TC_COUNTER_TYPE_CT);
+	if (IS_ERR(cnt)) {
+		rc = PTR_ERR(cnt);
+		goto release;
+	}
+	conn->cnt = cnt;
+
+	rc = efx_mae_insert_ct(efx, conn);
+	if (rc) {
+		netif_dbg(efx, drv, efx->net_dev,
+			  "Failed to insert conntrack, %d\n", rc);
+		goto release;
+	}
+	mutex_lock(&ct_zone->mutex);
+	list_add_tail(&conn->list, &ct_zone->cts);
+	mutex_unlock(&ct_zone->mutex);
+	return 0;
+release:
+	if (conn->cnt)
+		efx_tc_flower_release_counter(efx, conn->cnt);
+	if (!old)
+		rhashtable_remove_fast(&efx->tc->ct_ht, &conn->linkage,
+				       efx_tc_ct_ht_params);
+	kfree(conn);
+	return rc;
+}
+
+/* Caller must follow with efx_tc_ct_remove_finish() after RCU grace period! */
+static void efx_tc_ct_remove(struct efx_nic *efx, struct efx_tc_ct_entry *conn)
+{
+	int rc;
+
+	/* Remove it from HW */
+	rc = efx_mae_remove_ct(efx, conn);
+	/* Delete it from SW */
+	rhashtable_remove_fast(&efx->tc->ct_ht, &conn->linkage,
+			       efx_tc_ct_ht_params);
+	if (rc) {
+		netif_err(efx, drv, efx->net_dev,
+			  "Failed to remove conntrack %lx from hw, rc %d\n",
+			  conn->cookie, rc);
+	} else {
+		netif_dbg(efx, drv, efx->net_dev, "Removed conntrack %lx\n",
+			  conn->cookie);
+	}
+}
+
+static void efx_tc_ct_remove_finish(struct efx_nic *efx, struct efx_tc_ct_entry *conn)
+{
+	/* Remove related CT counter.  This is delayed after the conn object we
+	 * are working with has been successfully removed.  This protects the
+	 * counter from being used-after-free inside efx_tc_ct_stats.
+	 */
+	efx_tc_flower_release_counter(efx, conn->cnt);
+	kfree(conn);
+}
+
+static int efx_tc_ct_destroy(struct efx_tc_ct_zone *ct_zone,
+			     struct flow_cls_offload *tc)
+{
+	struct efx_nic *efx = ct_zone->efx;
+	struct efx_tc_ct_entry *conn;
+
+	conn = rhashtable_lookup_fast(&efx->tc->ct_ht, &tc->cookie,
+				      efx_tc_ct_ht_params);
+	if (!conn) {
+		netif_warn(efx, drv, efx->net_dev,
+			   "Conntrack %lx not found to remove\n", tc->cookie);
+		return -ENOENT;
+	}
+
+	mutex_lock(&ct_zone->mutex);
+	list_del(&conn->list);
+	efx_tc_ct_remove(efx, conn);
+	mutex_unlock(&ct_zone->mutex);
+	synchronize_rcu();
+	efx_tc_ct_remove_finish(efx, conn);
+	return 0;
+}
+
+static int efx_tc_ct_stats(struct efx_tc_ct_zone *ct_zone,
+			   struct flow_cls_offload *tc)
+{
+	struct efx_nic *efx = ct_zone->efx;
+	struct efx_tc_ct_entry *conn;
+	struct efx_tc_counter *cnt;
+
+	rcu_read_lock();
+	conn = rhashtable_lookup_fast(&efx->tc->ct_ht, &tc->cookie,
+				      efx_tc_ct_ht_params);
+	if (!conn) {
+		netif_warn(efx, drv, efx->net_dev,
+			   "Conntrack %lx not found for stats\n", tc->cookie);
+		rcu_read_unlock();
+		return -ENOENT;
+	}
+
+	cnt = conn->cnt;
+	spin_lock_bh(&cnt->lock);
+	/* Report only last use */
+	flow_stats_update(&tc->stats, 0, 0, 0, cnt->touched,
+			  FLOW_ACTION_HW_STATS_DELAYED);
+	spin_unlock_bh(&cnt->lock);
+	rcu_read_unlock();
+
+	return 0;
 }
 
 static int efx_tc_flow_block(enum tc_setup_type type, void *type_data,
 			     void *cb_priv)
 {
+	struct flow_cls_offload *tcb = type_data;
+	struct efx_tc_ct_zone *ct_zone = cb_priv;
+
+	if (type != TC_SETUP_CLSFLOWER)
+		return -EOPNOTSUPP;
+
+	switch (tcb->command) {
+	case FLOW_CLS_REPLACE:
+		return efx_tc_ct_replace(ct_zone, tcb);
+	case FLOW_CLS_DESTROY:
+		return efx_tc_ct_destroy(ct_zone, tcb);
+	case FLOW_CLS_STATS:
+		return efx_tc_ct_stats(ct_zone, tcb);
+	default:
+		break;
+	};
+
 	return -EOPNOTSUPP;
 }
 
@@ -81,6 +483,8 @@ struct efx_tc_ct_zone *efx_tc_ct_register_zone(struct efx_nic *efx, u16 zone,
 	}
 	ct_zone->nf_ft = ct_ft;
 	ct_zone->efx = efx;
+	INIT_LIST_HEAD(&ct_zone->cts);
+	mutex_init(&ct_zone->mutex);
 	rc = nf_flow_table_offload_add_cb(ct_ft, efx_tc_flow_block, ct_zone);
 	netif_dbg(efx, drv, efx->net_dev, "Adding new ct_zone for %u, rc %d\n",
 		  zone, rc);
@@ -98,11 +502,22 @@ struct efx_tc_ct_zone *efx_tc_ct_register_zone(struct efx_nic *efx, u16 zone,
 void efx_tc_ct_unregister_zone(struct efx_nic *efx,
 			       struct efx_tc_ct_zone *ct_zone)
 {
+	struct efx_tc_ct_entry *conn, *next;
+
 	if (!refcount_dec_and_test(&ct_zone->ref))
 		return; /* still in use */
 	nf_flow_table_offload_del_cb(ct_zone->nf_ft, efx_tc_flow_block, ct_zone);
 	rhashtable_remove_fast(&efx->tc->ct_zone_ht, &ct_zone->linkage,
 			       efx_tc_ct_zone_ht_params);
+	mutex_lock(&ct_zone->mutex);
+	list_for_each_entry(conn, &ct_zone->cts, list)
+		efx_tc_ct_remove(efx, conn);
+	synchronize_rcu();
+	/* need to use _safe because efx_tc_ct_remove_finish() frees conn */
+	list_for_each_entry_safe(conn, next, &ct_zone->cts, list)
+		efx_tc_ct_remove_finish(efx, conn);
+	mutex_unlock(&ct_zone->mutex);
+	mutex_destroy(&ct_zone->mutex);
 	netif_dbg(efx, drv, efx->net_dev, "Removed ct_zone for %u\n",
 		  ct_zone->zone);
 	kfree(ct_zone);

drivers/net/ethernet/sfc/tc_conntrack.h

@@ -22,6 +22,8 @@ struct efx_tc_ct_zone {
 	refcount_t ref;
 	struct nf_flowtable *nf_ft;
 	struct efx_nic *efx;
+	struct mutex mutex; /* protects cts list */
+	struct list_head cts; /* list of efx_tc_ct_entry in this zone */
 };
 
 /* create/teardown hashtables */
@@ -45,6 +47,7 @@ struct efx_tc_ct_entry {
 	struct efx_tc_ct_zone *zone;
 	u32 mark;
 	struct efx_tc_counter *cnt;
+	struct list_head list; /* entry on zone->cts */
 };
 
 #endif /* CONFIG_SFC_SRIOV */

drivers/net/ethernet/sfc/tc_counters.c

@@ -129,8 +129,8 @@ static void efx_tc_counter_work(struct work_struct *work)
 
 /* Counter allocation */
 
-static struct efx_tc_counter *efx_tc_flower_allocate_counter(struct efx_nic *efx,
-							     int type)
+struct efx_tc_counter *efx_tc_flower_allocate_counter(struct efx_nic *efx,
+						      int type)
 {
 	struct efx_tc_counter *cnt;
 	int rc, rc2;
@@ -169,8 +169,8 @@ static struct efx_tc_counter *efx_tc_flower_allocate_counter(struct efx_nic *efx
 	return ERR_PTR(rc > 0 ? -EIO : rc);
 }
 
-static void efx_tc_flower_release_counter(struct efx_nic *efx,
-					  struct efx_tc_counter *cnt)
+void efx_tc_flower_release_counter(struct efx_nic *efx,
+				   struct efx_tc_counter *cnt)
 {
 	int rc;
 

drivers/net/ethernet/sfc/tc_counters.h

@@ -49,6 +49,10 @@ int efx_tc_init_counters(struct efx_nic *efx);
 void efx_tc_destroy_counters(struct efx_nic *efx);
 void efx_tc_fini_counters(struct efx_nic *efx);
 
+struct efx_tc_counter *efx_tc_flower_allocate_counter(struct efx_nic *efx,
+						      int type);
+void efx_tc_flower_release_counter(struct efx_nic *efx,
+				   struct efx_tc_counter *cnt);
 struct efx_tc_counter_index *efx_tc_flower_get_counter_index(
 				struct efx_nic *efx, unsigned long cookie,
 				enum efx_tc_counter_type type);