/* * Copyright (c) 2016 Cisco and/or its affiliates. * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at: * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #include #include #include #include /* * Vector Hash tables of neighbour (traditional) adjacencies * Key: interface(for the vector index), address (and its proto), * link-type/ether-type. */ static BVT(clib_bihash) **adj_nbr_tables[FIB_PROTOCOL_MAX]; // FIXME SIZE APPROPRIATELY. ASK DAVEB. #define ADJ_NBR_DEFAULT_HASH_NUM_BUCKETS (64 * 64) #define ADJ_NBR_DEFAULT_HASH_MEMORY_SIZE (32<<20) #define ADJ_NBR_SET_KEY(_key, _lt, _nh) \ { \ _key.key[0] = (_nh)->as_u64[0]; \ _key.key[1] = (_nh)->as_u64[1]; \ _key.key[2] = (_lt); \ } #define ADJ_NBR_ITF_OK(_proto, _itf) \ (((_itf) < vec_len(adj_nbr_tables[_proto])) && \ (NULL != adj_nbr_tables[_proto][sw_if_index])) static void adj_nbr_insert (fib_protocol_t nh_proto, vnet_link_t link_type, const ip46_address_t *nh_addr, u32 sw_if_index, adj_index_t adj_index) { BVT(clib_bihash_kv) kv; if (sw_if_index >= vec_len(adj_nbr_tables[nh_proto])) { vec_validate(adj_nbr_tables[nh_proto], sw_if_index); } if (NULL == adj_nbr_tables[nh_proto][sw_if_index]) { adj_nbr_tables[nh_proto][sw_if_index] = clib_mem_alloc_aligned(sizeof(BVT(clib_bihash)), CLIB_CACHE_LINE_BYTES); memset(adj_nbr_tables[nh_proto][sw_if_index], 0, sizeof(BVT(clib_bihash))); BV(clib_bihash_init) (adj_nbr_tables[nh_proto][sw_if_index], "Adjacency Neighbour table", ADJ_NBR_DEFAULT_HASH_NUM_BUCKETS, ADJ_NBR_DEFAULT_HASH_MEMORY_SIZE); } ADJ_NBR_SET_KEY(kv, link_type, nh_addr); kv.value = adj_index; BV(clib_bihash_add_del) (adj_nbr_tables[nh_proto][sw_if_index], &kv, 1); } void adj_nbr_remove (adj_index_t ai, fib_protocol_t nh_proto, vnet_link_t link_type, const ip46_address_t *nh_addr, u32 sw_if_index) { BVT(clib_bihash_kv) kv; if (!ADJ_NBR_ITF_OK(nh_proto, sw_if_index)) return; ADJ_NBR_SET_KEY(kv, link_type, nh_addr); kv.value = ai; BV(clib_bihash_add_del) (adj_nbr_tables[nh_proto][sw_if_index], &kv, 0); } static adj_index_t adj_nbr_find (fib_protocol_t nh_proto, vnet_link_t link_type, const ip46_address_t *nh_addr, u32 sw_if_index) { BVT(clib_bihash_kv) kv; ADJ_NBR_SET_KEY(kv, link_type, nh_addr); if (!ADJ_NBR_ITF_OK(nh_proto, sw_if_index)) return (ADJ_INDEX_INVALID); if (BV(clib_bihash_search)(adj_nbr_tables[nh_proto][sw_if_index], &kv, &kv) < 0) { return (ADJ_INDEX_INVALID); } else { return (kv.value); } } static inline u32 adj_get_nd_node (fib_protocol_t proto) { switch (proto) { case FIB_PROTOCOL_IP4: return (ip4_arp_node.index); case FIB_PROTOCOL_IP6: return (ip6_discover_neighbor_node.index); case FIB_PROTOCOL_MPLS: break; } ASSERT(0); return (ip4_arp_node.index); } /** * @brief Check and set feature flags if o/p interface has any o/p features. */ static void adj_nbr_evaluate_feature (adj_index_t ai) { ip_adjacency_t *adj; vnet_feature_main_t *fm = &feature_main; i16 feature_count; u8 arc_index; u32 sw_if_index; adj = adj_get(ai); switch (adj->ia_link) { case VNET_LINK_IP4: arc_index = ip4_main.lookup_main.output_feature_arc_index; break; case VNET_LINK_IP6: arc_index = ip6_main.lookup_main.output_feature_arc_index; break; case VNET_LINK_MPLS: arc_index = mpls_main.output_feature_arc_index; break; default: return; } sw_if_index = adj->rewrite_header.sw_if_index; if (vec_len(fm->feature_count_by_sw_if_index[arc_index]) > sw_if_index) { feature_count = fm->feature_count_by_sw_if_index[arc_index][sw_if_index]; if (feature_count > 0) adj->rewrite_header.flags |= VNET_REWRITE_HAS_FEATURES; } return; } static ip_adjacency_t* adj_nbr_alloc (fib_protocol_t nh_proto, vnet_link_t link_type, const ip46_address_t *nh_addr, u32 sw_if_index) { ip_adjacency_t *adj; adj = adj_alloc(nh_proto); adj_nbr_insert(nh_proto, link_type, nh_addr, sw_if_index, adj_get_index(adj)); /* * since we just added the ADJ we have no rewrite string for it, * so its for ARP */ adj->lookup_next_index = IP_LOOKUP_NEXT_ARP; adj->sub_type.nbr.next_hop = *nh_addr; adj->ia_link = link_type; adj->ia_nh_proto = nh_proto; adj->rewrite_header.sw_if_index = sw_if_index; adj_nbr_evaluate_feature (adj_get_index(adj)); return (adj); } /* * adj_nbr_add_or_lock * * Add an adjacency for the neighbour requested. * * The key for an adj is: * - the Next-hops protocol (i.e. v4 or v6) * - the address of the next-hop * - the interface the next-hop is reachable through */ adj_index_t adj_nbr_add_or_lock (fib_protocol_t nh_proto, vnet_link_t link_type, const ip46_address_t *nh_addr, u32 sw_if_index) { adj_index_t adj_index; ip_adjacency_t *adj; adj_index = adj_nbr_find(nh_proto, link_type, nh_addr, sw_if_index); if (ADJ_INDEX_INVALID == adj_index) { vnet_main_t *vnm; vnm = vnet_get_main(); adj = adj_nbr_alloc(nh_proto, link_type, nh_addr, sw_if_index); adj_index = adj_get_index(adj); adj_lock(adj_index); vnet_rewrite_init(vnm, sw_if_index, adj_get_nd_node(nh_proto), vnet_tx_node_index_for_sw_interface(vnm, sw_if_index), &adj->rewrite_header); /* * we need a rewrite where the destination IP address is converted * to the appropriate link-layer address. This is interface specific. * So ask the interface to do it. */ vnet_update_adjacency_for_sw_interface(vnm, sw_if_index, adj_index); } else { adj_lock(adj_index); } return (adj_index); } adj_index_t adj_nbr_add_or_lock_w_rewrite (fib_protocol_t nh_proto, vnet_link_t link_type, const ip46_address_t *nh_addr, u32 sw_if_index, u8 *rewrite) { adj_index_t adj_index; ip_adjacency_t *adj; adj_index = adj_nbr_find(nh_proto, link_type, nh_addr, sw_if_index); if (ADJ_INDEX_INVALID == adj_index) { adj = adj_nbr_alloc(nh_proto, link_type, nh_addr, sw_if_index); adj->rewrite_header.sw_if_index = sw_if_index; } else { adj = adj_get(adj_index); } adj_lock(adj_get_index(adj)); adj_nbr_update_rewrite(adj_get_index(adj), ADJ_NBR_REWRITE_FLAG_COMPLETE, rewrite); return (adj_get_index(adj)); } /** * adj_nbr_update_rewrite * * Update the adjacency's rewrite string. A NULL string implies the * rewirte is reset (i.e. when ARP/ND etnry is gone). * NB: the adj being updated may be handling traffic in the DP. */ void adj_nbr_update_rewrite (adj_index_t adj_index, adj_nbr_rewrite_flag_t flags, u8 *rewrite) { ip_adjacency_t *adj; ASSERT(ADJ_INDEX_INVALID != adj_index); adj = adj_get(adj_index); if (flags & ADJ_NBR_REWRITE_FLAG_COMPLETE) { /* * update the adj's rewrite string and build the arc * from the rewrite node to the interface's TX node */ adj_nbr_update_rewrite_internal(adj, IP_LOOKUP_NEXT_REWRITE, adj_get_rewrite_node(adj->ia_link), vnet_tx_node_index_for_sw_interface( vnet_get_main(), adj->rewrite_header.sw_if_index), rewrite); } else { adj_nbr_update_rewrite_internal(adj, IP_LOOKUP_NEXT_ARP, adj_get_nd_node(adj->ia_nh_proto), vnet_tx_node_index_for_sw_interface( vnet_get_main(), adj->rewrite_header.sw_if_index), rewrite); } } /** * adj_nbr_update_rewrite_internal * * Update the adjacency's rewrite string. A NULL string implies the * rewirte is reset (i.e. when ARP/ND etnry is gone). * NB: the adj being updated may be handling traffic in the DP. */ void adj_nbr_update_rewrite_internal (ip_adjacency_t *adj, ip_lookup_next_t adj_next_index, u32 this_node, u32 next_node, u8 *rewrite) { ip_adjacency_t *walk_adj; adj_index_t walk_ai; vlib_main_t * vm; u32 old_next; int do_walk; vm = vlib_get_main(); old_next = adj->lookup_next_index; walk_ai = adj_get_index(adj); if (VNET_LINK_MPLS == adj->ia_link) { /* * The link type MPLS has no children in the control plane graph, it only * has children in the data-palne graph. The backwalk is up the former. * So we need to walk from its IP cousin. */ walk_ai = adj_nbr_find(adj->ia_nh_proto, fib_proto_to_link(adj->ia_nh_proto), &adj->sub_type.nbr.next_hop, adj->rewrite_header.sw_if_index); } /* * Don't call the walk re-entrantly */ if (ADJ_INDEX_INVALID != walk_ai) { walk_adj = adj_get(walk_ai); if (ADJ_FLAG_SYNC_WALK_ACTIVE & walk_adj->ia_flags) { do_walk = 0; } else { /* * Prevent re-entrant walk of the same adj */ walk_adj->ia_flags |= ADJ_FLAG_SYNC_WALK_ACTIVE; do_walk = 1; } } else { do_walk = 0; } /* * lock the adjacencies that are affected by updates this walk will provoke. * Since the aim of the walk is to update children to link to a different * DPO, this adj will no longer be in use and its lock count will drop to 0. * We don't want it to be deleted as part of this endevour. */ adj_lock(adj_get_index(adj)); adj_lock(walk_ai); /* * Updating a rewrite string is not atomic; * - the rewrite string is too long to write in one instruction * - when swapping from incomplete to complete, we also need to update * the VLIB graph next-index of the adj. * ideally we would only want to suspend forwarding via this adj whilst we * do this, but we do not have that level of granularity - it's suspend all * worker threads or nothing. * The other chioces are: * - to mark the adj down and back walk so child load-balances drop this adj * from the set. * - update the next_node index of this adj to point to error-drop * both of which will mean for MAC change we will drop for this adj * which is not acceptable. However, when the adj changes type (from * complete to incomplete and vice-versa) the child DPOs, which have the * VLIB graph next node index, will be sending packets to the wrong graph * node. So from the options above, updating the next_node of the adj to * be drop will work, but it relies on each graph node v4/v6/mpls, rewrite/ * arp/midchain always be valid w.r.t. a mis-match of adj type and node type * (i.e. a rewrite adj in the arp node). This is not enforcable. Getting it * wrong will lead to hard to find bugs since its a race condition. So we * choose the more reliable method of updating the children to use the drop, * then switching adj's type, then updating the children again. Did I mention * that this doesn't happen often... * So we need to distinguish between the two cases: * 1 - mac change * 2 - adj type change */ if (do_walk && old_next != adj_next_index && ADJ_INDEX_INVALID != walk_ai) { /* * the adj is changing type. we need to fix all children so that they * stack momentarily on a drop, while the adj changes. If we don't do * this the children will send packets to a VLIB graph node that does * not correspond to the adj's type - and it goes downhill from there. */ fib_node_back_walk_ctx_t bw_ctx = { .fnbw_reason = FIB_NODE_BW_REASON_FLAG_ADJ_DOWN, /* * force this walk to be synchrous. if we don't and a node in the graph * (a heavily shared path-list) chooses to back-ground the walk (make it * async) then it will pause and we will do the adj update below, before * all the children are updated. not good. */ .fnbw_flags = FIB_NODE_BW_FLAG_FORCE_SYNC, }; fib_walk_sync(FIB_NODE_TYPE_ADJ, walk_ai, &bw_ctx); } /* * If we are just updating the MAC string of the adj (which we also can't * do atomically), then we need to stop packets switching through the adj. * We can't do that on a per-adj basis, so it's all the packets. * If we are updating the type, and we walked back to the children above, * then this barrier serves to flush the queues/frames. */ vlib_worker_thread_barrier_sync(vm); adj->lookup_next_index = adj_next_index; if (NULL != rewrite) { /* * new rewrite provided. * fill in the adj's rewrite string, and build the VLIB graph arc. */ vnet_rewrite_set_data_internal(&adj->rewrite_header, sizeof(adj->rewrite_data), rewrite, vec_len(rewrite)); vec_free(rewrite); } else { vnet_rewrite_clear_data_internal(&adj->rewrite_header, sizeof(adj->rewrite_data)); } adj->rewrite_header.next_index = vlib_node_add_next(vlib_get_main(), this_node, next_node); /* * done with the rewirte update - let the workers loose. */ vlib_worker_thread_barrier_release(vm); if (do_walk && (old_next != adj->lookup_next_index) && (ADJ_INDEX_INVALID != walk_ai)) { /* * backwalk to the children so they can stack on the now updated * adjacency */ fib_node_back_walk_ctx_t bw_ctx = { .fnbw_reason = FIB_NODE_BW_REASON_FLAG_ADJ_UPDATE, }; fib_walk_sync(FIB_NODE_TYPE_ADJ, walk_ai, &bw_ctx); } /* * Prevent re-entrant walk of the same adj */ if (do_walk) { walk_adj->ia_flags &= ~ADJ_FLAG_SYNC_WALK_ACTIVE; } adj_unlock(adj_get_index(adj)); adj_unlock(walk_ai); } typedef struct adj_db_count_ctx_t_ { u64 count; } adj_db_count_ctx_t; static void adj_db_count (BVT(clib_bihash_kv) * kvp, void *arg) { adj_db_count_ctx_t * ctx = arg; ctx->count++; } u32 adj_nbr_db_size (void) { adj_db_count_ctx_t ctx = { .count = 0, }; fib_protocol_t proto; u32 sw_if_index = 0; for (proto = FIB_PROTOCOL_IP4; proto <= FIB_PROTOCOL_IP6; proto++) { vec_foreach_index(sw_if_index, adj_nbr_tables[proto]) { if (NULL != adj_nbr_tables[proto][sw_if_index]) { BV(clib_bihash_foreach_key_value_pair) ( adj_nbr_tables[proto][sw_if_index], adj_db_count, &ctx); } } } return (ctx.count); } /** * @brief Context for a walk of the adjacency neighbour DB */ typedef struct adj_walk_ctx_t_ { adj_walk_cb_t awc_cb; void *awc_ctx; } adj_walk_ctx_t; static void adj_nbr_walk_cb (BVT(clib_bihash_kv) * kvp, void *arg) { adj_walk_ctx_t *ctx = arg; // FIXME: can't stop early... ctx->awc_cb(kvp->value, ctx->awc_ctx); } void adj_nbr_walk (u32 sw_if_index, fib_protocol_t adj_nh_proto, adj_walk_cb_t cb, void *ctx) { if (!ADJ_NBR_ITF_OK(adj_nh_proto, sw_if_index)) return; adj_walk_ctx_t awc = { .awc_ctx = ctx, .awc_cb = cb, }; BV(clib_bihash_foreach_key_value_pair) ( adj_nbr_tables[adj_nh_proto][sw_if_index], adj_nbr_walk_cb, &awc); } /** * @brief Context for a walk of the adjacency neighbour DB */ typedef struct adj_walk_nh_ctx_t_ { adj_walk_cb_t awc_cb; void *awc_ctx; const ip46_address_t *awc_nh; } adj_walk_nh_ctx_t; static void adj_nbr_walk_nh_cb (BVT(clib_bihash_kv) * kvp, void *arg) { ip_adjacency_t *adj; adj_walk_nh_ctx_t *ctx = arg; adj = adj_get(kvp->value); if (!ip46_address_cmp(&adj->sub_type.nbr.next_hop, ctx->awc_nh)) ctx->awc_cb(kvp->value, ctx->awc_ctx); } /** * @brief Walk adjacencies on a link with a given v4 next-hop. * that is visit the adjacencies with different link types. */ void adj_nbr_walk_nh4 (u32 sw_if_index, const ip4_address_t *addr, adj_walk_cb_t cb, void *ctx) { if (!ADJ_NBR_ITF_OK(FIB_PROTOCOL_IP4, sw_if_index)) return; ip46_address_t nh = { .ip4 = *addr, }; adj_walk_nh_ctx_t awc = { .awc_ctx = ctx, .awc_cb = cb, .awc_nh = &nh, }; BV(clib_bihash_foreach_key_value_pair) ( adj_nbr_tables[FIB_PROTOCOL_IP4][sw_if_index], adj_nbr_walk_nh_cb, &awc); } /** * @brief Walk adjacencies on a link with a given v6 next-hop. * that is visit the adjacencies with different link types. */ void adj_nbr_walk_nh6 (u32 sw_if_index, const ip6_address_t *addr, adj_walk_cb_t cb, void *ctx) { if (!ADJ_NBR_ITF_OK(FIB_PROTOCOL_IP6, sw_if_index)) return; ip46_address_t nh = { .ip6 = *addr, }; adj_walk_nh_ctx_t awc = { .awc_ctx = ctx, .awc_cb = cb, .awc_nh = &nh, }; BV(clib_bihash_foreach_key_value_pair) ( adj_nbr_tables[FIB_PROTOCOL_IP6][sw_if_index], adj_nbr_walk_nh_cb, &awc); } /** * @brief Walk adjacencies on a link with a given next-hop. * that is visit the adjacencies with different link types. */ void adj_nbr_walk_nh (u32 sw_if_index, fib_protocol_t adj_nh_proto, const ip46_address_t *nh, adj_walk_cb_t cb, void *ctx) { if (!ADJ_NBR_ITF_OK(adj_nh_proto, sw_if_index)) return; adj_walk_nh_ctx_t awc = { .awc_ctx = ctx, .awc_cb = cb, .awc_nh = nh, }; BV(clib_bihash_foreach_key_value_pair) ( adj_nbr_tables[adj_nh_proto][sw_if_index], adj_nbr_walk_nh_cb, &awc); } /** * Flags associated with the interface state walks */ typedef enum adj_nbr_interface_flags_t_ { ADJ_NBR_INTERFACE_UP = (1 << 0), } adj_nbr_interface_flags_t; /** * Context for the state change walk of the DB */ typedef struct adj_nbr_interface_state_change_ctx_t_ { /** * Flags on the interface */ adj_nbr_interface_flags_t flags; } adj_nbr_interface_state_change_ctx_t; static adj_walk_rc_t adj_nbr_interface_state_change_one (adj_index_t ai, void *arg) { /* * Back walk the graph to inform the forwarding entries * that this interface state has changed. Do this synchronously * since this is the walk that provides convergence */ adj_nbr_interface_state_change_ctx_t *ctx = arg; fib_node_back_walk_ctx_t bw_ctx = { .fnbw_reason = ((ctx->flags & ADJ_NBR_INTERFACE_UP) ? FIB_NODE_BW_REASON_FLAG_INTERFACE_UP : FIB_NODE_BW_REASON_FLAG_INTERFACE_DOWN), /* * the force sync applies only as far as the first fib_entry. * And it's the fib_entry's we need to converge away from * the adjacencies on the now down link */ .fnbw_flags = (!(ctx->flags & ADJ_NBR_INTERFACE_UP) ? FIB_NODE_BW_FLAG_FORCE_SYNC : 0), }; fib_walk_sync(FIB_NODE_TYPE_ADJ, ai, &bw_ctx); return (ADJ_WALK_RC_CONTINUE); } /** * @brief Registered function for SW interface state changes */ static clib_error_t * adj_nbr_sw_interface_state_change (vnet_main_t * vnm, u32 sw_if_index, u32 flags) { fib_protocol_t proto; /* * walk each adj on the interface a
;;; Copyright (c) 2016 Cisco and/or its affiliates.
;;; Licensed under the Apache License, Version 2.0 (the "License");
;;; you may not use this file except in compliance with the License.