From 2b52ce02e1d17986ad7da8bc6e96c632dfa73219 Mon Sep 17 00:00:00 2001 From: Ido Schimmel Date: Mon, 22 Jan 2018 09:17:42 +0100 Subject: [PATCH] mlxsw: spectrum_router: Remove unnecessary prefix lengths from LPM tree In commit fc922bb0dd94 ("mlxsw: spectrum_router: Use one LPM tree for all virtual routers") I tried to make sure only used prefix lengths are present in the LPM tree shared between all virtual routers. However, this optimization had to be removed in commit a69518cf0b4c ("mlxsw: spectrum_router: Avoid expensive lookup during route removal"), since determining the used prefix lengths required us to traverse all the active virtual routers, which could result in a hung task depending on the number of VRFs and whether routes were removed due to abort or not. Re-introduce the optimization by moving the prefix usage accounting from the virtual routers to the LPM tree, as this accounting is only used in order to determine the tree's structure. To make the sharing of the trees more explicit, the two trees (for IPv4 and IPv6) are stored in the shared router struct and upon the creation of a virtual router it is immediately bound to both. Signed-off-by: Ido Schimmel Signed-off-by: Jiri Pirko Signed-off-by: David S. Miller --- .../ethernet/mellanox/mlxsw/spectrum_router.c | 226 ++++++++++-------- 1 file changed, 125 insertions(+), 101 deletions(-) diff --git a/drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c b/drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c index a590c0ab7524..91170d940268 100644 --- a/drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c +++ b/drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c @@ -71,6 +71,7 @@ #include "spectrum_mr_tcam.h" #include "spectrum_router.h" +struct mlxsw_sp_fib; struct mlxsw_sp_vr; struct mlxsw_sp_lpm_tree; struct mlxsw_sp_rif_ops; @@ -84,6 +85,8 @@ struct mlxsw_sp_router { struct rhashtable nexthop_ht; struct list_head nexthop_list; struct { + /* One tree for each protocol: IPv4 and IPv6 */ + struct mlxsw_sp_lpm_tree *proto_trees[2]; struct mlxsw_sp_lpm_tree *trees; unsigned int tree_count; } lpm; @@ -162,6 +165,15 @@ struct mlxsw_sp_rif_ops { struct mlxsw_sp_fid * (*fid_get)(struct mlxsw_sp_rif *rif); }; +static void mlxsw_sp_lpm_tree_hold(struct mlxsw_sp_lpm_tree *lpm_tree); +static void mlxsw_sp_lpm_tree_put(struct mlxsw_sp *mlxsw_sp, + struct mlxsw_sp_lpm_tree *lpm_tree); +static int mlxsw_sp_vr_lpm_tree_bind(struct mlxsw_sp *mlxsw_sp, + const struct mlxsw_sp_fib *fib, + u8 tree_id); +static int mlxsw_sp_vr_lpm_tree_unbind(struct mlxsw_sp *mlxsw_sp, + const struct mlxsw_sp_fib *fib); + static unsigned int * mlxsw_sp_rif_p_counter_get(struct mlxsw_sp_rif *rif, enum mlxsw_sp_rif_counter_dir dir) @@ -390,7 +402,6 @@ enum mlxsw_sp_fib_entry_type { }; struct mlxsw_sp_nexthop_group; -struct mlxsw_sp_fib; struct mlxsw_sp_fib_node { struct list_head entry_list; @@ -437,6 +448,7 @@ struct mlxsw_sp_lpm_tree { u8 id; /* tree ID */ unsigned int ref_count; enum mlxsw_sp_l3proto proto; + unsigned long prefix_ref_count[MLXSW_SP_PREFIX_COUNT]; struct mlxsw_sp_prefix_usage prefix_usage; }; @@ -445,8 +457,6 @@ struct mlxsw_sp_fib { struct list_head node_list; struct mlxsw_sp_vr *vr; struct mlxsw_sp_lpm_tree *lpm_tree; - unsigned long prefix_ref_count[MLXSW_SP_PREFIX_COUNT]; - struct mlxsw_sp_prefix_usage prefix_usage; enum mlxsw_sp_l3proto proto; }; @@ -461,12 +471,15 @@ struct mlxsw_sp_vr { static const struct rhashtable_params mlxsw_sp_fib_ht_params; -static struct mlxsw_sp_fib *mlxsw_sp_fib_create(struct mlxsw_sp_vr *vr, +static struct mlxsw_sp_fib *mlxsw_sp_fib_create(struct mlxsw_sp *mlxsw_sp, + struct mlxsw_sp_vr *vr, enum mlxsw_sp_l3proto proto) { + struct mlxsw_sp_lpm_tree *lpm_tree; struct mlxsw_sp_fib *fib; int err; + lpm_tree = mlxsw_sp->router->lpm.proto_trees[proto]; fib = kzalloc(sizeof(*fib), GFP_KERNEL); if (!fib) return ERR_PTR(-ENOMEM); @@ -476,17 +489,26 @@ static struct mlxsw_sp_fib *mlxsw_sp_fib_create(struct mlxsw_sp_vr *vr, INIT_LIST_HEAD(&fib->node_list); fib->proto = proto; fib->vr = vr; + fib->lpm_tree = lpm_tree; + mlxsw_sp_lpm_tree_hold(lpm_tree); + err = mlxsw_sp_vr_lpm_tree_bind(mlxsw_sp, fib, lpm_tree->id); + if (err) + goto err_lpm_tree_bind; return fib; +err_lpm_tree_bind: + mlxsw_sp_lpm_tree_put(mlxsw_sp, lpm_tree); err_rhashtable_init: kfree(fib); return ERR_PTR(err); } -static void mlxsw_sp_fib_destroy(struct mlxsw_sp_fib *fib) +static void mlxsw_sp_fib_destroy(struct mlxsw_sp *mlxsw_sp, + struct mlxsw_sp_fib *fib) { + mlxsw_sp_vr_lpm_tree_unbind(mlxsw_sp, fib); + mlxsw_sp_lpm_tree_put(mlxsw_sp, fib->lpm_tree); WARN_ON(!list_empty(&fib->node_list)); - WARN_ON(fib->lpm_tree); rhashtable_destroy(&fib->ht); kfree(fib); } @@ -573,6 +595,9 @@ mlxsw_sp_lpm_tree_create(struct mlxsw_sp *mlxsw_sp, goto err_left_struct_set; memcpy(&lpm_tree->prefix_usage, prefix_usage, sizeof(lpm_tree->prefix_usage)); + memset(&lpm_tree->prefix_ref_count, 0, + sizeof(lpm_tree->prefix_ref_count)); + lpm_tree->ref_count = 1; return lpm_tree; err_left_struct_set: @@ -599,8 +624,10 @@ mlxsw_sp_lpm_tree_get(struct mlxsw_sp *mlxsw_sp, if (lpm_tree->ref_count != 0 && lpm_tree->proto == proto && mlxsw_sp_prefix_usage_eq(&lpm_tree->prefix_usage, - prefix_usage)) + prefix_usage)) { + mlxsw_sp_lpm_tree_hold(lpm_tree); return lpm_tree; + } } return mlxsw_sp_lpm_tree_create(mlxsw_sp, prefix_usage, proto); } @@ -621,9 +648,10 @@ static void mlxsw_sp_lpm_tree_put(struct mlxsw_sp *mlxsw_sp, static int mlxsw_sp_lpm_init(struct mlxsw_sp *mlxsw_sp) { + struct mlxsw_sp_prefix_usage req_prefix_usage = {{ 0 } }; struct mlxsw_sp_lpm_tree *lpm_tree; u64 max_trees; - int i; + int err, i; if (!MLXSW_CORE_RES_VALID(mlxsw_sp->core, MAX_LPM_TREES)) return -EIO; @@ -641,11 +669,42 @@ static int mlxsw_sp_lpm_init(struct mlxsw_sp *mlxsw_sp) lpm_tree->id = i + MLXSW_SP_LPM_TREE_MIN; } + lpm_tree = mlxsw_sp_lpm_tree_get(mlxsw_sp, &req_prefix_usage, + MLXSW_SP_L3_PROTO_IPV4); + if (IS_ERR(lpm_tree)) { + err = PTR_ERR(lpm_tree); + goto err_ipv4_tree_get; + } + mlxsw_sp->router->lpm.proto_trees[MLXSW_SP_L3_PROTO_IPV4] = lpm_tree; + + lpm_tree = mlxsw_sp_lpm_tree_get(mlxsw_sp, &req_prefix_usage, + MLXSW_SP_L3_PROTO_IPV6); + if (IS_ERR(lpm_tree)) { + err = PTR_ERR(lpm_tree); + goto err_ipv6_tree_get; + } + mlxsw_sp->router->lpm.proto_trees[MLXSW_SP_L3_PROTO_IPV6] = lpm_tree; + return 0; + +err_ipv6_tree_get: + lpm_tree = mlxsw_sp->router->lpm.proto_trees[MLXSW_SP_L3_PROTO_IPV4]; + mlxsw_sp_lpm_tree_put(mlxsw_sp, lpm_tree); +err_ipv4_tree_get: + kfree(mlxsw_sp->router->lpm.trees); + return err; } static void mlxsw_sp_lpm_fini(struct mlxsw_sp *mlxsw_sp) { + struct mlxsw_sp_lpm_tree *lpm_tree; + + lpm_tree = mlxsw_sp->router->lpm.proto_trees[MLXSW_SP_L3_PROTO_IPV6]; + mlxsw_sp_lpm_tree_put(mlxsw_sp, lpm_tree); + + lpm_tree = mlxsw_sp->router->lpm.proto_trees[MLXSW_SP_L3_PROTO_IPV4]; + mlxsw_sp_lpm_tree_put(mlxsw_sp, lpm_tree); + kfree(mlxsw_sp->router->lpm.trees); } @@ -737,10 +796,10 @@ static struct mlxsw_sp_vr *mlxsw_sp_vr_create(struct mlxsw_sp *mlxsw_sp, NL_SET_ERR_MSG(extack, "spectrum: Exceeded number of supported virtual routers"); return ERR_PTR(-EBUSY); } - vr->fib4 = mlxsw_sp_fib_create(vr, MLXSW_SP_L3_PROTO_IPV4); + vr->fib4 = mlxsw_sp_fib_create(mlxsw_sp, vr, MLXSW_SP_L3_PROTO_IPV4); if (IS_ERR(vr->fib4)) return ERR_CAST(vr->fib4); - vr->fib6 = mlxsw_sp_fib_create(vr, MLXSW_SP_L3_PROTO_IPV6); + vr->fib6 = mlxsw_sp_fib_create(mlxsw_sp, vr, MLXSW_SP_L3_PROTO_IPV6); if (IS_ERR(vr->fib6)) { err = PTR_ERR(vr->fib6); goto err_fib6_create; @@ -755,21 +814,22 @@ static struct mlxsw_sp_vr *mlxsw_sp_vr_create(struct mlxsw_sp *mlxsw_sp, return vr; err_mr_table_create: - mlxsw_sp_fib_destroy(vr->fib6); + mlxsw_sp_fib_destroy(mlxsw_sp, vr->fib6); vr->fib6 = NULL; err_fib6_create: - mlxsw_sp_fib_destroy(vr->fib4); + mlxsw_sp_fib_destroy(mlxsw_sp, vr->fib4); vr->fib4 = NULL; return ERR_PTR(err); } -static void mlxsw_sp_vr_destroy(struct mlxsw_sp_vr *vr) +static void mlxsw_sp_vr_destroy(struct mlxsw_sp *mlxsw_sp, + struct mlxsw_sp_vr *vr) { mlxsw_sp_mr_table_destroy(vr->mr4_table); vr->mr4_table = NULL; - mlxsw_sp_fib_destroy(vr->fib6); + mlxsw_sp_fib_destroy(mlxsw_sp, vr->fib6); vr->fib6 = NULL; - mlxsw_sp_fib_destroy(vr->fib4); + mlxsw_sp_fib_destroy(mlxsw_sp, vr->fib4); vr->fib4 = NULL; } @@ -785,12 +845,12 @@ static struct mlxsw_sp_vr *mlxsw_sp_vr_get(struct mlxsw_sp *mlxsw_sp, u32 tb_id, return vr; } -static void mlxsw_sp_vr_put(struct mlxsw_sp_vr *vr) +static void mlxsw_sp_vr_put(struct mlxsw_sp *mlxsw_sp, struct mlxsw_sp_vr *vr) { if (!vr->rif_count && list_empty(&vr->fib4->node_list) && list_empty(&vr->fib6->node_list) && mlxsw_sp_mr_table_empty(vr->mr4_table)) - mlxsw_sp_vr_destroy(vr); + mlxsw_sp_vr_destroy(mlxsw_sp, vr); } static bool @@ -801,7 +861,7 @@ mlxsw_sp_vr_lpm_tree_should_replace(struct mlxsw_sp_vr *vr, if (!mlxsw_sp_vr_is_used(vr)) return false; - if (fib->lpm_tree && fib->lpm_tree->id == tree_id) + if (fib->lpm_tree->id == tree_id) return true; return false; } @@ -831,14 +891,13 @@ static int mlxsw_sp_vrs_lpm_tree_replace(struct mlxsw_sp *mlxsw_sp, struct mlxsw_sp_fib *fib, struct mlxsw_sp_lpm_tree *new_tree) { - struct mlxsw_sp_lpm_tree *old_tree = fib->lpm_tree; enum mlxsw_sp_l3proto proto = fib->proto; + struct mlxsw_sp_lpm_tree *old_tree; u8 old_id, new_id = new_tree->id; struct mlxsw_sp_vr *vr; int i, err; - if (!old_tree) - goto no_replace; + old_tree = mlxsw_sp->router->lpm.proto_trees[proto]; old_id = old_tree->id; for (i = 0; i < MLXSW_CORE_RES_GET(mlxsw_sp->core, MAX_VRS); i++) { @@ -852,6 +911,11 @@ static int mlxsw_sp_vrs_lpm_tree_replace(struct mlxsw_sp *mlxsw_sp, goto err_tree_replace; } + memcpy(new_tree->prefix_ref_count, old_tree->prefix_ref_count, + sizeof(new_tree->prefix_ref_count)); + mlxsw_sp->router->lpm.proto_trees[proto] = new_tree; + mlxsw_sp_lpm_tree_put(mlxsw_sp, old_tree); + return 0; err_tree_replace: @@ -863,36 +927,6 @@ err_tree_replace: old_tree); } return err; - -no_replace: - fib->lpm_tree = new_tree; - mlxsw_sp_lpm_tree_hold(new_tree); - err = mlxsw_sp_vr_lpm_tree_bind(mlxsw_sp, fib, new_tree->id); - if (err) { - mlxsw_sp_lpm_tree_put(mlxsw_sp, new_tree); - fib->lpm_tree = NULL; - return err; - } - return 0; -} - -static void -mlxsw_sp_vrs_prefixes(struct mlxsw_sp *mlxsw_sp, - enum mlxsw_sp_l3proto proto, - struct mlxsw_sp_prefix_usage *req_prefix_usage) -{ - int i; - - for (i = 0; i < MLXSW_CORE_RES_GET(mlxsw_sp->core, MAX_VRS); i++) { - struct mlxsw_sp_vr *vr = &mlxsw_sp->router->vrs[i]; - struct mlxsw_sp_fib *fib = mlxsw_sp_vr_fib(vr, proto); - unsigned char prefix; - - if (!mlxsw_sp_vr_is_used(vr)) - continue; - mlxsw_sp_prefix_usage_for_each(prefix, &fib->prefix_usage) - mlxsw_sp_prefix_usage_set(req_prefix_usage, prefix); - } } static int mlxsw_sp_vrs_init(struct mlxsw_sp *mlxsw_sp) @@ -4195,71 +4229,64 @@ mlxsw_sp_fib_node_entry_is_first(const struct mlxsw_sp_fib_node *fib_node, static int mlxsw_sp_fib_lpm_tree_link(struct mlxsw_sp *mlxsw_sp, struct mlxsw_sp_fib_node *fib_node) { - struct mlxsw_sp_prefix_usage req_prefix_usage = {{ 0 } }; + struct mlxsw_sp_prefix_usage req_prefix_usage; struct mlxsw_sp_fib *fib = fib_node->fib; struct mlxsw_sp_lpm_tree *lpm_tree; int err; - /* Since the tree is shared between all virtual routers we must - * make sure it contains all the required prefix lengths. This - * can be computed by either adding the new prefix length to the - * existing prefix usage of a bound tree, or by aggregating the - * prefix lengths across all virtual routers and adding the new - * one as well. - */ - if (fib->lpm_tree) - mlxsw_sp_prefix_usage_cpy(&req_prefix_usage, - &fib->lpm_tree->prefix_usage); - else - mlxsw_sp_vrs_prefixes(mlxsw_sp, fib->proto, &req_prefix_usage); - mlxsw_sp_prefix_usage_set(&req_prefix_usage, fib_node->key.prefix_len); + lpm_tree = mlxsw_sp->router->lpm.proto_trees[fib->proto]; + if (lpm_tree->prefix_ref_count[fib_node->key.prefix_len] != 0) + goto out; + mlxsw_sp_prefix_usage_cpy(&req_prefix_usage, &lpm_tree->prefix_usage); + mlxsw_sp_prefix_usage_set(&req_prefix_usage, fib_node->key.prefix_len); lpm_tree = mlxsw_sp_lpm_tree_get(mlxsw_sp, &req_prefix_usage, fib->proto); if (IS_ERR(lpm_tree)) return PTR_ERR(lpm_tree); - if (fib->lpm_tree && fib->lpm_tree->id == lpm_tree->id) - return 0; - err = mlxsw_sp_vrs_lpm_tree_replace(mlxsw_sp, fib, lpm_tree); if (err) - return err; + goto err_lpm_tree_replace; +out: + lpm_tree->prefix_ref_count[fib_node->key.prefix_len]++; return 0; + +err_lpm_tree_replace: + mlxsw_sp_lpm_tree_put(mlxsw_sp, lpm_tree); + return err; } static void mlxsw_sp_fib_lpm_tree_unlink(struct mlxsw_sp *mlxsw_sp, struct mlxsw_sp_fib_node *fib_node) { + struct mlxsw_sp_lpm_tree *lpm_tree = fib_node->fib->lpm_tree; + struct mlxsw_sp_prefix_usage req_prefix_usage; struct mlxsw_sp_fib *fib = fib_node->fib; + int err; - if (!list_is_singular(&fib->node_list)) + if (--lpm_tree->prefix_ref_count[fib_node->key.prefix_len] != 0) return; - /* Last node is being unlinked from the FIB. Unbind the - * tree and drop the reference. + /* Try to construct a new LPM tree from the current prefix usage + * minus the unused one. If we fail, continue using the old one. */ - mlxsw_sp_vr_lpm_tree_unbind(mlxsw_sp, fib); - mlxsw_sp_lpm_tree_put(mlxsw_sp, fib->lpm_tree); - fib->lpm_tree = NULL; -} - -static void mlxsw_sp_fib_node_prefix_inc(struct mlxsw_sp_fib_node *fib_node) -{ - unsigned char prefix_len = fib_node->key.prefix_len; - struct mlxsw_sp_fib *fib = fib_node->fib; + mlxsw_sp_prefix_usage_cpy(&req_prefix_usage, &lpm_tree->prefix_usage); + mlxsw_sp_prefix_usage_clear(&req_prefix_usage, + fib_node->key.prefix_len); + lpm_tree = mlxsw_sp_lpm_tree_get(mlxsw_sp, &req_prefix_usage, + fib->proto); + if (IS_ERR(lpm_tree)) + return; - if (fib->prefix_ref_count[prefix_len]++ == 0) - mlxsw_sp_prefix_usage_set(&fib->prefix_usage, prefix_len); -} + err = mlxsw_sp_vrs_lpm_tree_replace(mlxsw_sp, fib, lpm_tree); + if (err) + goto err_lpm_tree_replace; -static void mlxsw_sp_fib_node_prefix_dec(struct mlxsw_sp_fib_node *fib_node) -{ - unsigned char prefix_len = fib_node->key.prefix_len; - struct mlxsw_sp_fib *fib = fib_node->fib; + return; - if (--fib->prefix_ref_count[prefix_len] == 0) - mlxsw_sp_prefix_usage_clear(&fib->prefix_usage, prefix_len); +err_lpm_tree_replace: + mlxsw_sp_lpm_tree_put(mlxsw_sp, lpm_tree); } static int mlxsw_sp_fib_node_init(struct mlxsw_sp *mlxsw_sp, @@ -4277,8 +4304,6 @@ static int mlxsw_sp_fib_node_init(struct mlxsw_sp *mlxsw_sp, if (err) goto err_fib_lpm_tree_link; - mlxsw_sp_fib_node_prefix_inc(fib_node); - return 0; err_fib_lpm_tree_link: @@ -4292,7 +4317,6 @@ static void mlxsw_sp_fib_node_fini(struct mlxsw_sp *mlxsw_sp, { struct mlxsw_sp_fib *fib = fib_node->fib; - mlxsw_sp_fib_node_prefix_dec(fib_node); mlxsw_sp_fib_lpm_tree_unlink(mlxsw_sp, fib_node); fib_node->fib = NULL; mlxsw_sp_fib_node_remove(fib, fib_node); @@ -4332,7 +4356,7 @@ mlxsw_sp_fib_node_get(struct mlxsw_sp *mlxsw_sp, u32 tb_id, const void *addr, err_fib_node_init: mlxsw_sp_fib_node_destroy(fib_node); err_fib_node_create: - mlxsw_sp_vr_put(vr); + mlxsw_sp_vr_put(mlxsw_sp, vr); return ERR_PTR(err); } @@ -4345,7 +4369,7 @@ static void mlxsw_sp_fib_node_put(struct mlxsw_sp *mlxsw_sp, return; mlxsw_sp_fib_node_fini(mlxsw_sp, fib_node); mlxsw_sp_fib_node_destroy(fib_node); - mlxsw_sp_vr_put(vr); + mlxsw_sp_vr_put(mlxsw_sp, vr); } static struct mlxsw_sp_fib4_entry * @@ -5368,7 +5392,7 @@ static void mlxsw_sp_router_fibmr_del(struct mlxsw_sp *mlxsw_sp, return; mlxsw_sp_mr_route4_del(vr->mr4_table, men_info->mfc); - mlxsw_sp_vr_put(vr); + mlxsw_sp_vr_put(mlxsw_sp, vr); } static int @@ -5405,7 +5429,7 @@ mlxsw_sp_router_fibmr_vif_del(struct mlxsw_sp *mlxsw_sp, return; mlxsw_sp_mr_vif_del(vr->mr4_table, ven_info->vif_index); - mlxsw_sp_vr_put(vr); + mlxsw_sp_vr_put(mlxsw_sp, vr); } static int mlxsw_sp_router_set_abort_trap(struct mlxsw_sp *mlxsw_sp) @@ -6054,7 +6078,7 @@ err_fid_get: err_rif_alloc: err_rif_index_alloc: vr->rif_count--; - mlxsw_sp_vr_put(vr); + mlxsw_sp_vr_put(mlxsw_sp, vr); return ERR_PTR(err); } @@ -6077,7 +6101,7 @@ void mlxsw_sp_rif_destroy(struct mlxsw_sp_rif *rif) mlxsw_sp_fid_put(fid); kfree(rif); vr->rif_count--; - mlxsw_sp_vr_put(vr); + mlxsw_sp_vr_put(mlxsw_sp, vr); } static void @@ -6867,7 +6891,7 @@ mlxsw_sp_rif_ipip_lb_configure(struct mlxsw_sp_rif *rif) return 0; err_loopback_op: - mlxsw_sp_vr_put(ul_vr); + mlxsw_sp_vr_put(mlxsw_sp, ul_vr); return err; } @@ -6881,7 +6905,7 @@ static void mlxsw_sp_rif_ipip_lb_deconfigure(struct mlxsw_sp_rif *rif) mlxsw_sp_rif_ipip_lb_op(lb_rif, ul_vr, false); --ul_vr->rif_count; - mlxsw_sp_vr_put(ul_vr); + mlxsw_sp_vr_put(mlxsw_sp, ul_vr); } static const struct mlxsw_sp_rif_ops mlxsw_sp_rif_ipip_lb_ops = { -- 2.30.2