Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1 | /* |
Florin Coras | 222e1f41 | 2019-02-16 20:47:32 -0800 | [diff] [blame] | 2 | * Copyright (c) 2016-2019 Cisco and/or its affiliates. |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 3 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | * you may not use this file except in compliance with the License. |
| 5 | * You may obtain a copy of the License at: |
| 6 | * |
| 7 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | * |
| 9 | * Unless required by applicable law or agreed to in writing, software |
| 10 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | * See the License for the specific language governing permissions and |
| 13 | * limitations under the License. |
| 14 | */ |
| 15 | |
Dave Barach | 3bbcfab | 2017-08-15 19:03:44 -0400 | [diff] [blame] | 16 | /** |
| 17 | * @file |
| 18 | * @brief TCP host stack utilities |
| 19 | */ |
| 20 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 21 | #include <vnet/tcp/tcp.h> |
Florin Coras | 999840c | 2020-03-18 20:31:34 +0000 | [diff] [blame] | 22 | #include <vnet/tcp/tcp_inlines.h> |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 23 | #include <vnet/session/session.h> |
| 24 | #include <vnet/fib/fib.h> |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 25 | #include <vnet/dpo/load_balance.h> |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 26 | #include <math.h> |
| 27 | |
| 28 | tcp_main_t tcp_main; |
| 29 | |
Florin Coras | 1c8ff63 | 2018-05-17 13:28:34 -0700 | [diff] [blame] | 30 | typedef struct |
| 31 | { |
| 32 | fib_protocol_t nh_proto; |
| 33 | vnet_link_t link_type; |
| 34 | ip46_address_t ip; |
| 35 | u32 sw_if_index; |
| 36 | u8 is_add; |
| 37 | } tcp_add_del_adj_args_t; |
| 38 | |
| 39 | static void |
| 40 | tcp_add_del_adj_cb (tcp_add_del_adj_args_t * args) |
| 41 | { |
| 42 | u32 ai; |
| 43 | if (args->is_add) |
| 44 | { |
| 45 | adj_nbr_add_or_lock (args->nh_proto, args->link_type, &args->ip, |
| 46 | args->sw_if_index); |
| 47 | } |
| 48 | else |
| 49 | { |
| 50 | ai = adj_nbr_find (FIB_PROTOCOL_IP6, VNET_LINK_IP6, &args->ip, |
| 51 | args->sw_if_index); |
| 52 | if (ai != ADJ_INDEX_INVALID) |
| 53 | adj_unlock (ai); |
| 54 | } |
| 55 | } |
| 56 | |
| 57 | static void |
| 58 | tcp_add_del_adjacency (tcp_connection_t * tc, u8 is_add) |
| 59 | { |
| 60 | tcp_add_del_adj_args_t args = { |
| 61 | .nh_proto = FIB_PROTOCOL_IP6, |
| 62 | .link_type = VNET_LINK_IP6, |
| 63 | .ip = tc->c_rmt_ip, |
| 64 | .sw_if_index = tc->sw_if_index, |
| 65 | .is_add = is_add |
| 66 | }; |
| 67 | vlib_rpc_call_main_thread (tcp_add_del_adj_cb, (u8 *) & args, |
| 68 | sizeof (args)); |
| 69 | } |
| 70 | |
Florin Coras | d9a145f | 2019-06-07 09:35:20 -0700 | [diff] [blame] | 71 | static void |
| 72 | tcp_cc_init (tcp_connection_t * tc) |
| 73 | { |
Florin Coras | d9a145f | 2019-06-07 09:35:20 -0700 | [diff] [blame] | 74 | tc->cc_algo->init (tc); |
| 75 | } |
| 76 | |
| 77 | static void |
| 78 | tcp_cc_cleanup (tcp_connection_t * tc) |
| 79 | { |
| 80 | if (tc->cc_algo->cleanup) |
| 81 | tc->cc_algo->cleanup (tc); |
| 82 | } |
| 83 | |
| 84 | void |
| 85 | tcp_cc_algo_register (tcp_cc_algorithm_type_e type, |
| 86 | const tcp_cc_algorithm_t * vft) |
| 87 | { |
| 88 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 89 | vec_validate (tm->cc_algos, type); |
| 90 | |
| 91 | tm->cc_algos[type] = *vft; |
| 92 | hash_set_mem (tm->cc_algo_by_name, vft->name, type); |
| 93 | } |
| 94 | |
| 95 | tcp_cc_algorithm_t * |
| 96 | tcp_cc_algo_get (tcp_cc_algorithm_type_e type) |
| 97 | { |
| 98 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 99 | return &tm->cc_algos[type]; |
| 100 | } |
| 101 | |
Florin Coras | 4e116fb | 2019-06-10 08:33:50 -0700 | [diff] [blame] | 102 | tcp_cc_algorithm_type_e |
| 103 | tcp_cc_algo_new_type (const tcp_cc_algorithm_t * vft) |
| 104 | { |
| 105 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 106 | tcp_cc_algo_register (++tm->cc_last_type, vft); |
| 107 | return tm->cc_last_type; |
| 108 | } |
| 109 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 110 | static u32 |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 111 | tcp_connection_bind (u32 session_index, transport_endpoint_t * lcl) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 112 | { |
| 113 | tcp_main_t *tm = &tcp_main; |
| 114 | tcp_connection_t *listener; |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 115 | void *iface_ip; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 116 | |
| 117 | pool_get (tm->listener_pool, listener); |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 118 | clib_memset (listener, 0, sizeof (*listener)); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 119 | |
| 120 | listener->c_c_index = listener - tm->listener_pool; |
Florin Coras | 0e49568 | 2017-09-19 22:27:18 -0700 | [diff] [blame] | 121 | listener->c_lcl_port = lcl->port; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 122 | |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 123 | /* If we are provided a sw_if_index, bind using one of its ips */ |
| 124 | if (ip_is_zero (&lcl->ip, 1) && lcl->sw_if_index != ENDPOINT_INVALID_INDEX) |
Florin Coras | 6cf30ad | 2017-04-04 23:08:23 -0700 | [diff] [blame] | 125 | { |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 126 | if ((iface_ip = ip_interface_get_first_ip (lcl->sw_if_index, |
| 127 | lcl->is_ip4))) |
| 128 | ip_set (&lcl->ip, iface_ip, lcl->is_ip4); |
Florin Coras | 6cf30ad | 2017-04-04 23:08:23 -0700 | [diff] [blame] | 129 | } |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 130 | ip_copy (&listener->c_lcl_ip, &lcl->ip, lcl->is_ip4); |
| 131 | listener->c_is_ip4 = lcl->is_ip4; |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 132 | listener->c_proto = TRANSPORT_PROTO_TCP; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 133 | listener->c_s_index = session_index; |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 134 | listener->c_fib_index = lcl->fib_index; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 135 | listener->state = TCP_STATE_LISTEN; |
Florin Coras | 12f6936 | 2019-08-16 09:44:00 -0700 | [diff] [blame] | 136 | listener->cc_algo = tcp_cc_algo_get (tcp_cfg.cc_algo); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 137 | |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 138 | tcp_connection_timers_init (listener); |
| 139 | |
Florin Coras | a436a42 | 2019-08-20 07:09:31 -0700 | [diff] [blame] | 140 | TCP_EVT (TCP_EVT_BIND, listener); |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 141 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 142 | return listener->c_c_index; |
| 143 | } |
| 144 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 145 | static u32 |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 146 | tcp_session_bind (u32 session_index, transport_endpoint_t * tep) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 147 | { |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 148 | return tcp_connection_bind (session_index, tep); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 149 | } |
| 150 | |
| 151 | static void |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 152 | tcp_connection_unbind (u32 listener_index) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 153 | { |
| 154 | tcp_main_t *tm = vnet_get_tcp_main (); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 155 | tcp_connection_t *tc; |
| 156 | |
| 157 | tc = pool_elt_at_index (tm->listener_pool, listener_index); |
| 158 | |
Florin Coras | a436a42 | 2019-08-20 07:09:31 -0700 | [diff] [blame] | 159 | TCP_EVT (TCP_EVT_UNBIND, tc); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 160 | |
| 161 | /* Poison the entry */ |
| 162 | if (CLIB_DEBUG > 0) |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 163 | clib_memset (tc, 0xFA, sizeof (*tc)); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 164 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 165 | pool_put_index (tm->listener_pool, listener_index); |
| 166 | } |
| 167 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 168 | static u32 |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 169 | tcp_session_unbind (u32 listener_index) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 170 | { |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 171 | tcp_connection_unbind (listener_index); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 172 | return 0; |
| 173 | } |
| 174 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 175 | static transport_connection_t * |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 176 | tcp_session_get_listener (u32 listener_index) |
| 177 | { |
| 178 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 179 | tcp_connection_t *tc; |
| 180 | tc = pool_elt_at_index (tm->listener_pool, listener_index); |
| 181 | return &tc->connection; |
| 182 | } |
| 183 | |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 184 | /** |
| 185 | * Cleanup half-open connection |
| 186 | * |
| 187 | */ |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 188 | static void |
Florin Coras | d50ff7f | 2020-04-16 04:30:22 +0000 | [diff] [blame] | 189 | tcp_half_open_connection_free (tcp_connection_t * tc) |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 190 | { |
| 191 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 192 | clib_spinlock_lock_if_init (&tm->half_open_lock); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 193 | if (CLIB_DEBUG) |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 194 | clib_memset (tc, 0xFA, sizeof (*tc)); |
Benoît Ganne | d4aeb84 | 2019-07-18 18:38:42 +0200 | [diff] [blame] | 195 | pool_put (tm->half_open_connections, tc); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 196 | clib_spinlock_unlock_if_init (&tm->half_open_lock); |
| 197 | } |
| 198 | |
| 199 | /** |
| 200 | * Try to cleanup half-open connection |
| 201 | * |
| 202 | * If called from a thread that doesn't own tc, the call won't have any |
| 203 | * effect. |
| 204 | * |
| 205 | * @param tc - connection to be cleaned up |
| 206 | * @return non-zero if cleanup failed. |
| 207 | */ |
| 208 | int |
| 209 | tcp_half_open_connection_cleanup (tcp_connection_t * tc) |
| 210 | { |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 211 | tcp_worker_ctx_t *wrk; |
| 212 | |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 213 | /* Make sure this is the owning thread */ |
| 214 | if (tc->c_thread_index != vlib_get_thread_index ()) |
| 215 | return 1; |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 216 | |
Florin Coras | d50ff7f | 2020-04-16 04:30:22 +0000 | [diff] [blame] | 217 | session_half_open_delete_notify (TRANSPORT_PROTO_TCP, tc->c_s_ho_handle); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 218 | wrk = tcp_get_worker (tc->c_thread_index); |
| 219 | tcp_timer_reset (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT_SYN); |
Florin Coras | d50ff7f | 2020-04-16 04:30:22 +0000 | [diff] [blame] | 220 | tcp_half_open_connection_free (tc); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 221 | return 0; |
| 222 | } |
| 223 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 224 | static tcp_connection_t * |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 225 | tcp_half_open_connection_new (void) |
| 226 | { |
| 227 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 228 | tcp_connection_t *tc = 0; |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 229 | ASSERT (vlib_get_thread_index () == 0); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 230 | pool_get (tm->half_open_connections, tc); |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 231 | clib_memset (tc, 0, sizeof (*tc)); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 232 | tc->c_c_index = tc - tm->half_open_connections; |
| 233 | return tc; |
| 234 | } |
| 235 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 236 | /** |
| 237 | * Cleans up connection state. |
| 238 | * |
| 239 | * No notifications. |
| 240 | */ |
| 241 | void |
| 242 | tcp_connection_cleanup (tcp_connection_t * tc) |
| 243 | { |
Florin Coras | a436a42 | 2019-08-20 07:09:31 -0700 | [diff] [blame] | 244 | TCP_EVT (TCP_EVT_DELETE, tc); |
Florin Coras | 070fd4b | 2019-04-02 19:03:23 -0700 | [diff] [blame] | 245 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 246 | /* Cleanup local endpoint if this was an active connect */ |
Florin Coras | 43818c1 | 2020-02-09 23:09:26 +0000 | [diff] [blame] | 247 | if (!(tc->cfg_flags & TCP_CFG_F_NO_ENDPOINT)) |
| 248 | transport_endpoint_cleanup (TRANSPORT_PROTO_TCP, &tc->c_lcl_ip, |
| 249 | tc->c_lcl_port); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 250 | |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 251 | /* Check if connection is not yet fully established */ |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 252 | if (tc->state == TCP_STATE_SYN_SENT) |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 253 | { |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 254 | /* Try to remove the half-open connection. If this is not the owning |
| 255 | * thread, tc won't be removed. Retransmit or establish timers will |
| 256 | * eventually expire and call again cleanup on the right thread. */ |
Florin Coras | c5347d9 | 2018-10-17 10:41:28 -0700 | [diff] [blame] | 257 | if (tcp_half_open_connection_cleanup (tc)) |
| 258 | tc->flags |= TCP_CONN_HALF_OPEN_DONE; |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 259 | } |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 260 | else |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 261 | { |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 262 | /* Make sure all timers are cleared */ |
| 263 | tcp_connection_timers_reset (tc); |
| 264 | |
Florin Coras | 1c8ff63 | 2018-05-17 13:28:34 -0700 | [diff] [blame] | 265 | if (!tc->c_is_ip4 && ip6_address_is_link_local_unicast (&tc->c_rmt_ip6)) |
| 266 | tcp_add_del_adjacency (tc, 0); |
| 267 | |
Florin Coras | d9a145f | 2019-06-07 09:35:20 -0700 | [diff] [blame] | 268 | tcp_cc_cleanup (tc); |
Florin Coras | b691f76 | 2019-02-22 09:07:20 -0800 | [diff] [blame] | 269 | vec_free (tc->snd_sacks); |
| 270 | vec_free (tc->snd_sacks_fl); |
Florin Coras | 558e3e0 | 2019-09-06 12:56:58 -0700 | [diff] [blame] | 271 | vec_free (tc->rcv_opts.sacks); |
| 272 | pool_free (tc->sack_sb.holes); |
Florin Coras | b691f76 | 2019-02-22 09:07:20 -0800 | [diff] [blame] | 273 | |
Florin Coras | bbcfaac | 2019-10-10 13:52:04 -0700 | [diff] [blame] | 274 | if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE) |
Florin Coras | 5281473 | 2019-06-12 15:38:19 -0700 | [diff] [blame] | 275 | tcp_bt_cleanup (tc); |
| 276 | |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 277 | tcp_connection_free (tc); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 278 | } |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 279 | } |
| 280 | |
| 281 | /** |
| 282 | * Connection removal. |
| 283 | * |
| 284 | * This should be called only once connection enters CLOSED state. Note |
| 285 | * that it notifies the session of the removal event, so if the goal is to |
| 286 | * just remove the connection, call tcp_connection_cleanup instead. |
| 287 | */ |
| 288 | void |
| 289 | tcp_connection_del (tcp_connection_t * tc) |
| 290 | { |
Florin Coras | 5a2ec8f | 2018-12-27 11:53:11 -0800 | [diff] [blame] | 291 | session_transport_delete_notify (&tc->connection); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 292 | tcp_connection_cleanup (tc); |
| 293 | } |
| 294 | |
Florin Coras | 6534b7a | 2017-07-18 05:38:03 -0400 | [diff] [blame] | 295 | tcp_connection_t * |
Florin Coras | 8124cb7 | 2018-12-16 20:57:29 -0800 | [diff] [blame] | 296 | tcp_connection_alloc (u8 thread_index) |
Florin Coras | 6534b7a | 2017-07-18 05:38:03 -0400 | [diff] [blame] | 297 | { |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 298 | tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index); |
Florin Coras | 6534b7a | 2017-07-18 05:38:03 -0400 | [diff] [blame] | 299 | tcp_connection_t *tc; |
| 300 | |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 301 | pool_get (wrk->connections, tc); |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 302 | clib_memset (tc, 0, sizeof (*tc)); |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 303 | tc->c_c_index = tc - wrk->connections; |
Florin Coras | 6534b7a | 2017-07-18 05:38:03 -0400 | [diff] [blame] | 304 | tc->c_thread_index = thread_index; |
| 305 | return tc; |
| 306 | } |
| 307 | |
Florin Coras | 12f6936 | 2019-08-16 09:44:00 -0700 | [diff] [blame] | 308 | tcp_connection_t * |
| 309 | tcp_connection_alloc_w_base (u8 thread_index, tcp_connection_t * base) |
| 310 | { |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 311 | tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index); |
Florin Coras | 12f6936 | 2019-08-16 09:44:00 -0700 | [diff] [blame] | 312 | tcp_connection_t *tc; |
| 313 | |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 314 | pool_get (wrk->connections, tc); |
Florin Coras | 12f6936 | 2019-08-16 09:44:00 -0700 | [diff] [blame] | 315 | clib_memcpy_fast (tc, base, sizeof (*tc)); |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 316 | tc->c_c_index = tc - wrk->connections; |
Florin Coras | 12f6936 | 2019-08-16 09:44:00 -0700 | [diff] [blame] | 317 | tc->c_thread_index = thread_index; |
| 318 | return tc; |
| 319 | } |
| 320 | |
Florin Coras | 8124cb7 | 2018-12-16 20:57:29 -0800 | [diff] [blame] | 321 | void |
| 322 | tcp_connection_free (tcp_connection_t * tc) |
| 323 | { |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 324 | tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index); |
Florin Coras | 6416e62 | 2019-04-03 17:52:43 -0700 | [diff] [blame] | 325 | if (CLIB_DEBUG) |
| 326 | { |
Florin Coras | 6416e62 | 2019-04-03 17:52:43 -0700 | [diff] [blame] | 327 | clib_memset (tc, 0xFA, sizeof (*tc)); |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 328 | pool_put (wrk->connections, tc); |
Florin Coras | 6416e62 | 2019-04-03 17:52:43 -0700 | [diff] [blame] | 329 | return; |
| 330 | } |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 331 | pool_put (wrk->connections, tc); |
Florin Coras | 8124cb7 | 2018-12-16 20:57:29 -0800 | [diff] [blame] | 332 | } |
| 333 | |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 334 | void |
| 335 | tcp_program_cleanup (tcp_worker_ctx_t * wrk, tcp_connection_t * tc) |
| 336 | { |
| 337 | tcp_cleanup_req_t *req; |
| 338 | clib_time_type_t now; |
| 339 | |
Florin Coras | 8f10b90 | 2021-04-02 18:32:00 -0700 | [diff] [blame] | 340 | now = tcp_time_now_us (tc->c_thread_index); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 341 | clib_fifo_add2 (wrk->pending_cleanups, req); |
| 342 | req->connection_index = tc->c_c_index; |
| 343 | req->free_time = now + tcp_cfg.cleanup_time; |
| 344 | } |
| 345 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 346 | /** |
| 347 | * Begin connection closing procedure. |
| 348 | * |
| 349 | * If at the end the connection is not in CLOSED state, it is not removed. |
| 350 | * Instead, we rely on on TCP to advance through state machine to either |
| 351 | * 1) LAST_ACK (passive close) whereby when the last ACK is received |
| 352 | * tcp_connection_del is called. This notifies session of the delete and |
| 353 | * calls cleanup. |
| 354 | * 2) TIME_WAIT (active close) whereby after 2MSL the 2MSL timer triggers |
| 355 | * and cleanup is called. |
Florin Coras | d79b41e | 2017-03-04 05:37:52 -0800 | [diff] [blame] | 356 | * |
| 357 | * N.B. Half-close connections are not supported |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 358 | */ |
| 359 | void |
| 360 | tcp_connection_close (tcp_connection_t * tc) |
| 361 | { |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 362 | tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index); |
| 363 | |
Florin Coras | a436a42 | 2019-08-20 07:09:31 -0700 | [diff] [blame] | 364 | TCP_EVT (TCP_EVT_CLOSE, tc); |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 365 | |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 366 | /* Send/Program FIN if needed and switch state */ |
| 367 | switch (tc->state) |
| 368 | { |
| 369 | case TCP_STATE_SYN_SENT: |
Florin Coras | 85a3ddd | 2018-12-24 16:54:34 -0800 | [diff] [blame] | 370 | /* Try to cleanup. If not on the right thread, mark as half-open done. |
| 371 | * Connection will be cleaned up when establish timer pops */ |
| 372 | tcp_connection_cleanup (tc); |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 373 | break; |
| 374 | case TCP_STATE_SYN_RCVD: |
Florin Coras | c5347d9 | 2018-10-17 10:41:28 -0700 | [diff] [blame] | 375 | tcp_connection_timers_reset (tc); |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 376 | tcp_send_fin (tc); |
Florin Coras | 3c514d5 | 2018-12-22 11:39:33 -0800 | [diff] [blame] | 377 | tcp_connection_set_state (tc, TCP_STATE_FIN_WAIT_1); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 378 | tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_WAITCLOSE, |
| 379 | tcp_cfg.finwait1_time); |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 380 | break; |
| 381 | case TCP_STATE_ESTABLISHED: |
Florin Coras | f65074e | 2019-03-31 17:17:11 -0700 | [diff] [blame] | 382 | /* If closing with unread data, reset the connection */ |
| 383 | if (transport_max_rx_dequeue (&tc->connection)) |
| 384 | { |
| 385 | tcp_send_reset (tc); |
| 386 | tcp_connection_timers_reset (tc); |
| 387 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | a0904f0 | 2019-07-22 20:55:11 -0700 | [diff] [blame] | 388 | session_transport_closed_notify (&tc->connection); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 389 | tcp_program_cleanup (tcp_get_worker (tc->c_thread_index), tc); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 390 | tcp_worker_stats_inc (wrk, rst_unread, 1); |
Florin Coras | f65074e | 2019-03-31 17:17:11 -0700 | [diff] [blame] | 391 | break; |
| 392 | } |
Florin Coras | 31c9955 | 2019-03-01 13:00:58 -0800 | [diff] [blame] | 393 | if (!transport_max_tx_dequeue (&tc->connection)) |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 394 | tcp_send_fin (tc); |
| 395 | else |
| 396 | tc->flags |= TCP_CONN_FINPNDG; |
Florin Coras | 3c514d5 | 2018-12-22 11:39:33 -0800 | [diff] [blame] | 397 | tcp_connection_set_state (tc, TCP_STATE_FIN_WAIT_1); |
Florin Coras | e96bf63 | 2018-12-18 22:44:27 -0800 | [diff] [blame] | 398 | /* Set a timer in case the peer stops responding. Otherwise the |
| 399 | * connection will be stuck here forever. */ |
Florin Coras | 85a3ddd | 2018-12-24 16:54:34 -0800 | [diff] [blame] | 400 | ASSERT (tc->timers[TCP_TIMER_WAITCLOSE] == TCP_TIMER_HANDLE_INVALID); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 401 | tcp_timer_set (&wrk->timer_wheel, tc, TCP_TIMER_WAITCLOSE, |
| 402 | tcp_cfg.finwait1_time); |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 403 | break; |
| 404 | case TCP_STATE_CLOSE_WAIT: |
Florin Coras | 31c9955 | 2019-03-01 13:00:58 -0800 | [diff] [blame] | 405 | if (!transport_max_tx_dequeue (&tc->connection)) |
Florin Coras | 25579b4 | 2018-06-06 17:55:02 -0700 | [diff] [blame] | 406 | { |
| 407 | tcp_send_fin (tc); |
| 408 | tcp_connection_timers_reset (tc); |
Florin Coras | 3c514d5 | 2018-12-22 11:39:33 -0800 | [diff] [blame] | 409 | tcp_connection_set_state (tc, TCP_STATE_LAST_ACK); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 410 | tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_WAITCLOSE, |
| 411 | tcp_cfg.lastack_time); |
Florin Coras | 25579b4 | 2018-06-06 17:55:02 -0700 | [diff] [blame] | 412 | } |
| 413 | else |
| 414 | tc->flags |= TCP_CONN_FINPNDG; |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 415 | break; |
Florin Coras | c87c91d | 2017-08-16 19:55:49 -0700 | [diff] [blame] | 416 | case TCP_STATE_FIN_WAIT_1: |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 417 | tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_WAITCLOSE, |
| 418 | tcp_cfg.finwait1_time); |
Florin Coras | c87c91d | 2017-08-16 19:55:49 -0700 | [diff] [blame] | 419 | break; |
Florin Coras | 4af830c | 2018-12-04 09:21:36 -0800 | [diff] [blame] | 420 | case TCP_STATE_CLOSED: |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 421 | /* Cleanup should've been programmed already */ |
Florin Coras | 4af830c | 2018-12-04 09:21:36 -0800 | [diff] [blame] | 422 | break; |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 423 | default: |
Florin Coras | a096f2d | 2017-09-28 23:49:42 -0400 | [diff] [blame] | 424 | TCP_DBG ("state: %u", tc->state); |
Florin Coras | b2215d6 | 2017-08-01 16:56:58 -0700 | [diff] [blame] | 425 | } |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 426 | } |
| 427 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 428 | static void |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 429 | tcp_session_close (u32 conn_index, u32 thread_index) |
| 430 | { |
| 431 | tcp_connection_t *tc; |
| 432 | tc = tcp_connection_get (conn_index, thread_index); |
| 433 | tcp_connection_close (tc); |
| 434 | } |
| 435 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 436 | static void |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 437 | tcp_session_cleanup (u32 conn_index, u32 thread_index) |
| 438 | { |
| 439 | tcp_connection_t *tc; |
| 440 | tc = tcp_connection_get (conn_index, thread_index); |
Florin Coras | 54c93cf | 2019-09-24 12:45:14 -0700 | [diff] [blame] | 441 | if (!tc) |
| 442 | return; |
Florin Coras | 85a3ddd | 2018-12-24 16:54:34 -0800 | [diff] [blame] | 443 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | 25579b4 | 2018-06-06 17:55:02 -0700 | [diff] [blame] | 444 | tcp_connection_cleanup (tc); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 445 | } |
| 446 | |
Florin Coras | dfb3b87 | 2019-08-16 17:48:44 -0700 | [diff] [blame] | 447 | static void |
Florin Coras | d50ff7f | 2020-04-16 04:30:22 +0000 | [diff] [blame] | 448 | tcp_session_cleanup_ho (u32 conn_index) |
| 449 | { |
| 450 | tcp_worker_ctx_t *wrk; |
| 451 | tcp_connection_t *tc; |
| 452 | |
| 453 | tc = tcp_half_open_connection_get (conn_index); |
| 454 | wrk = tcp_get_worker (tc->c_thread_index); |
| 455 | tcp_timer_reset (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT_SYN); |
| 456 | tcp_half_open_connection_free (tc); |
| 457 | } |
| 458 | |
| 459 | static void |
Florin Coras | dfb3b87 | 2019-08-16 17:48:44 -0700 | [diff] [blame] | 460 | tcp_session_reset (u32 conn_index, u32 thread_index) |
| 461 | { |
| 462 | tcp_connection_t *tc; |
| 463 | tc = tcp_connection_get (conn_index, thread_index); |
Florin Coras | dfb3b87 | 2019-08-16 17:48:44 -0700 | [diff] [blame] | 464 | tcp_send_reset (tc); |
| 465 | tcp_connection_timers_reset (tc); |
Florin Coras | fd4c3fe | 2019-11-07 12:33:12 -0800 | [diff] [blame] | 466 | tcp_cong_recovery_off (tc); |
Florin Coras | dfb3b87 | 2019-08-16 17:48:44 -0700 | [diff] [blame] | 467 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 468 | session_transport_closed_notify (&tc->connection); |
| 469 | tcp_program_cleanup (tcp_get_worker (thread_index), tc); |
Florin Coras | dfb3b87 | 2019-08-16 17:48:44 -0700 | [diff] [blame] | 470 | } |
| 471 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 472 | /** |
| 473 | * Initialize all connection timers as invalid |
| 474 | */ |
| 475 | void |
| 476 | tcp_connection_timers_init (tcp_connection_t * tc) |
| 477 | { |
| 478 | int i; |
| 479 | |
| 480 | /* Set all to invalid */ |
| 481 | for (i = 0; i < TCP_N_TIMERS; i++) |
| 482 | { |
| 483 | tc->timers[i] = TCP_TIMER_HANDLE_INVALID; |
| 484 | } |
| 485 | |
| 486 | tc->rto = TCP_RTO_INIT; |
| 487 | } |
| 488 | |
| 489 | /** |
| 490 | * Stop all connection timers |
| 491 | */ |
| 492 | void |
| 493 | tcp_connection_timers_reset (tcp_connection_t * tc) |
| 494 | { |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 495 | tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 496 | int i; |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 497 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 498 | for (i = 0; i < TCP_N_TIMERS; i++) |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 499 | tcp_timer_reset (&wrk->timer_wheel, tc, i); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 500 | } |
| 501 | |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 502 | #if 0 |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 503 | typedef struct ip4_tcp_hdr |
| 504 | { |
| 505 | ip4_header_t ip; |
| 506 | tcp_header_t tcp; |
| 507 | } ip4_tcp_hdr_t; |
| 508 | |
| 509 | typedef struct ip6_tcp_hdr |
| 510 | { |
| 511 | ip6_header_t ip; |
| 512 | tcp_header_t tcp; |
| 513 | } ip6_tcp_hdr_t; |
| 514 | |
| 515 | static void |
| 516 | tcp_connection_select_lb_bucket (tcp_connection_t * tc, const dpo_id_t * dpo, |
| 517 | dpo_id_t * result) |
| 518 | { |
| 519 | const dpo_id_t *choice; |
| 520 | load_balance_t *lb; |
| 521 | int hash; |
| 522 | |
| 523 | lb = load_balance_get (dpo->dpoi_index); |
| 524 | if (tc->c_is_ip4) |
| 525 | { |
| 526 | ip4_tcp_hdr_t hdr; |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 527 | clib_memset (&hdr, 0, sizeof (hdr)); |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 528 | hdr.ip.protocol = IP_PROTOCOL_TCP; |
| 529 | hdr.ip.address_pair.src.as_u32 = tc->c_lcl_ip.ip4.as_u32; |
| 530 | hdr.ip.address_pair.dst.as_u32 = tc->c_rmt_ip.ip4.as_u32; |
| 531 | hdr.tcp.src_port = tc->c_lcl_port; |
| 532 | hdr.tcp.dst_port = tc->c_rmt_port; |
| 533 | hash = ip4_compute_flow_hash (&hdr.ip, lb->lb_hash_config); |
| 534 | } |
| 535 | else |
| 536 | { |
| 537 | ip6_tcp_hdr_t hdr; |
Dave Barach | b7b9299 | 2018-10-17 10:38:51 -0400 | [diff] [blame] | 538 | clib_memset (&hdr, 0, sizeof (hdr)); |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 539 | hdr.ip.protocol = IP_PROTOCOL_TCP; |
Dave Barach | 178cf49 | 2018-11-13 16:34:13 -0500 | [diff] [blame] | 540 | clib_memcpy_fast (&hdr.ip.src_address, &tc->c_lcl_ip.ip6, |
| 541 | sizeof (ip6_address_t)); |
| 542 | clib_memcpy_fast (&hdr.ip.dst_address, &tc->c_rmt_ip.ip6, |
| 543 | sizeof (ip6_address_t)); |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 544 | hdr.tcp.src_port = tc->c_lcl_port; |
| 545 | hdr.tcp.dst_port = tc->c_rmt_port; |
| 546 | hash = ip6_compute_flow_hash (&hdr.ip, lb->lb_hash_config); |
| 547 | } |
| 548 | choice = load_balance_get_bucket_i (lb, hash & lb->lb_n_buckets_minus_1); |
| 549 | dpo_copy (result, choice); |
| 550 | } |
| 551 | |
| 552 | fib_node_index_t |
| 553 | tcp_lookup_rmt_in_fib (tcp_connection_t * tc) |
| 554 | { |
| 555 | fib_prefix_t prefix; |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 556 | u32 fib_index; |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 557 | |
Dave Barach | 178cf49 | 2018-11-13 16:34:13 -0500 | [diff] [blame] | 558 | clib_memcpy_fast (&prefix.fp_addr, &tc->c_rmt_ip, sizeof (prefix.fp_addr)); |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 559 | prefix.fp_proto = tc->c_is_ip4 ? FIB_PROTOCOL_IP4 : FIB_PROTOCOL_IP6; |
| 560 | prefix.fp_len = tc->c_is_ip4 ? 32 : 128; |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 561 | fib_index = fib_table_find (prefix.fp_proto, tc->c_fib_index); |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 562 | return fib_table_lookup (fib_index, &prefix); |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 563 | } |
| 564 | |
| 565 | static int |
| 566 | tcp_connection_stack_on_fib_entry (tcp_connection_t * tc) |
| 567 | { |
| 568 | dpo_id_t choice = DPO_INVALID; |
| 569 | u32 output_node_index; |
| 570 | fib_entry_t *fe; |
| 571 | |
| 572 | fe = fib_entry_get (tc->c_rmt_fei); |
| 573 | if (fe->fe_lb.dpoi_type != DPO_LOAD_BALANCE) |
| 574 | return -1; |
| 575 | |
| 576 | tcp_connection_select_lb_bucket (tc, &fe->fe_lb, &choice); |
| 577 | |
| 578 | output_node_index = |
| 579 | tc->c_is_ip4 ? tcp4_output_node.index : tcp6_output_node.index; |
| 580 | dpo_stack_from_node (output_node_index, &tc->c_rmt_dpo, &choice); |
| 581 | return 0; |
| 582 | } |
| 583 | |
| 584 | /** Stack tcp connection on peer's fib entry. |
| 585 | * |
| 586 | * This ultimately populates the dpo the connection will use to send packets. |
| 587 | */ |
| 588 | static void |
| 589 | tcp_connection_fib_attach (tcp_connection_t * tc) |
| 590 | { |
| 591 | tc->c_rmt_fei = tcp_lookup_rmt_in_fib (tc); |
| 592 | |
| 593 | ASSERT (tc->c_rmt_fei != FIB_NODE_INDEX_INVALID); |
| 594 | |
| 595 | tcp_connection_stack_on_fib_entry (tc); |
| 596 | } |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 597 | #endif /* 0 */ |
Florin Coras | f6359c8 | 2017-06-19 12:26:09 -0400 | [diff] [blame] | 598 | |
Florin Coras | 18e0d4f | 2019-01-02 12:22:02 -0800 | [diff] [blame] | 599 | /** |
| 600 | * Generate random iss as per rfc6528 |
| 601 | */ |
| 602 | static u32 |
| 603 | tcp_generate_random_iss (tcp_connection_t * tc) |
| 604 | { |
| 605 | tcp_main_t *tm = &tcp_main; |
| 606 | u64 tmp; |
| 607 | |
| 608 | if (tc->c_is_ip4) |
| 609 | tmp = (u64) tc->c_lcl_ip.ip4.as_u32 << 32 | (u64) tc->c_rmt_ip.ip4.as_u32; |
| 610 | else |
| 611 | tmp = tc->c_lcl_ip.ip6.as_u64[0] ^ tc->c_lcl_ip.ip6.as_u64[1] |
| 612 | ^ tc->c_rmt_ip.ip6.as_u64[0] ^ tc->c_rmt_ip.ip6.as_u64[1]; |
| 613 | |
| 614 | tmp ^= tm->iss_seed.first | ((u64) tc->c_lcl_port << 16 | tc->c_rmt_port); |
| 615 | tmp ^= tm->iss_seed.second; |
| 616 | tmp = clib_xxhash (tmp) + clib_cpu_time_now (); |
| 617 | return ((tmp >> 32) ^ (tmp & 0xffffffff)); |
| 618 | } |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 619 | |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 620 | /** |
Florin Coras | cedcf60 | 2019-08-27 12:15:43 -0700 | [diff] [blame] | 621 | * Initialize max segment size we're able to process. |
| 622 | * |
| 623 | * The value is constrained by the output interface's MTU and by the size |
| 624 | * of the IP and TCP headers (see RFC6691). It is also what we advertise |
| 625 | * to our peer. |
| 626 | */ |
| 627 | static void |
| 628 | tcp_init_rcv_mss (tcp_connection_t * tc) |
| 629 | { |
| 630 | u8 ip_hdr_len; |
| 631 | |
Florin Coras | ff19e3b | 2020-02-10 17:44:13 +0000 | [diff] [blame] | 632 | /* Already provided at connection init time */ |
| 633 | if (tc->mss) |
| 634 | return; |
| 635 | |
Florin Coras | cedcf60 | 2019-08-27 12:15:43 -0700 | [diff] [blame] | 636 | ip_hdr_len = tc->c_is_ip4 ? sizeof (ip4_header_t) : sizeof (ip6_header_t); |
| 637 | tc->mss = tcp_cfg.default_mtu - sizeof (tcp_header_t) - ip_hdr_len; |
| 638 | } |
| 639 | |
| 640 | static void |
| 641 | tcp_init_mss (tcp_connection_t * tc) |
| 642 | { |
| 643 | u16 default_min_mss = 536; |
| 644 | |
| 645 | tcp_init_rcv_mss (tc); |
| 646 | |
| 647 | /* TODO consider PMTU discovery */ |
| 648 | tc->snd_mss = clib_min (tc->rcv_opts.mss, tc->mss); |
| 649 | |
| 650 | if (tc->snd_mss < 45) |
| 651 | { |
| 652 | /* Assume that at least the min default mss works */ |
| 653 | tc->snd_mss = default_min_mss; |
| 654 | tc->rcv_opts.mss = default_min_mss; |
| 655 | } |
| 656 | |
| 657 | /* We should have enough space for 40 bytes of options */ |
| 658 | ASSERT (tc->snd_mss > 45); |
| 659 | |
Florin Coras | 7a588fc | 2021-03-05 13:30:18 -0800 | [diff] [blame] | 660 | /* If we use timestamp option, account for it and make sure |
| 661 | * the options are 4-byte aligned */ |
Florin Coras | cedcf60 | 2019-08-27 12:15:43 -0700 | [diff] [blame] | 662 | if (tcp_opts_tstamp (&tc->rcv_opts)) |
Florin Coras | 7a588fc | 2021-03-05 13:30:18 -0800 | [diff] [blame] | 663 | tc->snd_mss -= TCP_OPTION_LEN_TIMESTAMP + 2 /* alignment */; |
Florin Coras | cedcf60 | 2019-08-27 12:15:43 -0700 | [diff] [blame] | 664 | } |
| 665 | |
| 666 | /** |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 667 | * Initialize connection send variables. |
| 668 | */ |
| 669 | void |
| 670 | tcp_init_snd_vars (tcp_connection_t * tc) |
| 671 | { |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 672 | /* |
| 673 | * We use the time to randomize iss and for setting up the initial |
| 674 | * timestamp. Make sure it's updated otherwise syn and ack in the |
| 675 | * handshake may make it look as if time has flown in the opposite |
| 676 | * direction for us. |
| 677 | */ |
Florin Coras | 8f10b90 | 2021-04-02 18:32:00 -0700 | [diff] [blame] | 678 | tcp_update_time_now (tcp_get_worker (vlib_get_thread_index ())); |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 679 | |
Florin Coras | cedcf60 | 2019-08-27 12:15:43 -0700 | [diff] [blame] | 680 | tcp_init_rcv_mss (tc); |
Florin Coras | 18e0d4f | 2019-01-02 12:22:02 -0800 | [diff] [blame] | 681 | tc->iss = tcp_generate_random_iss (tc); |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 682 | tc->snd_una = tc->iss; |
| 683 | tc->snd_nxt = tc->iss + 1; |
Florin Coras | eedc74b | 2020-07-31 12:32:40 -0700 | [diff] [blame] | 684 | tc->srtt = 0.1 * THZ; /* 100 ms */ |
Florin Coras | f4ce6ba | 2019-11-20 18:34:58 -0800 | [diff] [blame] | 685 | |
| 686 | if (!tcp_cfg.csum_offload) |
| 687 | tc->cfg_flags |= TCP_CFG_F_NO_CSUM_OFFLOAD; |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 688 | } |
| 689 | |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 690 | void |
| 691 | tcp_enable_pacing (tcp_connection_t * tc) |
| 692 | { |
Florin Coras | 36ebcff | 2019-09-12 18:36:44 -0700 | [diff] [blame] | 693 | u32 byte_rate; |
| 694 | byte_rate = tc->cwnd / (tc->srtt * TCP_TICK); |
| 695 | transport_connection_tx_pacer_init (&tc->connection, byte_rate, tc->cwnd); |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 696 | tc->mrtt_us = (u32) ~ 0; |
| 697 | } |
| 698 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 699 | /** Initialize tcp connection variables |
| 700 | * |
| 701 | * Should be called after having received a msg from the peer, i.e., a SYN or |
| 702 | * a SYNACK, such that connection options have already been exchanged. */ |
| 703 | void |
| 704 | tcp_connection_init_vars (tcp_connection_t * tc) |
| 705 | { |
| 706 | tcp_connection_timers_init (tc); |
Florin Coras | c834341 | 2017-05-04 14:25:50 -0700 | [diff] [blame] | 707 | tcp_init_mss (tc); |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 708 | scoreboard_init (&tc->sack_sb); |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 709 | if (tc->state == TCP_STATE_SYN_RCVD) |
| 710 | tcp_init_snd_vars (tc); |
| 711 | |
Florin Coras | 36ebcff | 2019-09-12 18:36:44 -0700 | [diff] [blame] | 712 | tcp_cc_init (tc); |
| 713 | |
Florin Coras | 1c8ff63 | 2018-05-17 13:28:34 -0700 | [diff] [blame] | 714 | if (!tc->c_is_ip4 && ip6_address_is_link_local_unicast (&tc->c_rmt_ip6)) |
| 715 | tcp_add_del_adjacency (tc, 1); |
| 716 | |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 717 | /* tcp_connection_fib_attach (tc); */ |
| 718 | |
| 719 | if (transport_connection_is_tx_paced (&tc->connection) |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 720 | || tcp_cfg.enable_tx_pacing) |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 721 | tcp_enable_pacing (tc); |
Florin Coras | 5281473 | 2019-06-12 15:38:19 -0700 | [diff] [blame] | 722 | |
Florin Coras | bbcfaac | 2019-10-10 13:52:04 -0700 | [diff] [blame] | 723 | if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE) |
Florin Coras | 5281473 | 2019-06-12 15:38:19 -0700 | [diff] [blame] | 724 | tcp_bt_init (tc); |
Florin Coras | edfe0ee | 2019-07-29 18:13:25 -0700 | [diff] [blame] | 725 | |
Florin Coras | bbcfaac | 2019-10-10 13:52:04 -0700 | [diff] [blame] | 726 | if (!tcp_cfg.allow_tso) |
| 727 | tc->cfg_flags |= TCP_CFG_F_NO_TSO; |
| 728 | |
Florin Coras | edfe0ee | 2019-07-29 18:13:25 -0700 | [diff] [blame] | 729 | tc->start_ts = tcp_time_now_us (tc->c_thread_index); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 730 | } |
| 731 | |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 732 | static int |
| 733 | tcp_alloc_custom_local_endpoint (tcp_main_t * tm, ip46_address_t * lcl_addr, |
| 734 | u16 * lcl_port, u8 is_ip4) |
| 735 | { |
| 736 | int index, port; |
| 737 | if (is_ip4) |
| 738 | { |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 739 | index = tm->last_v4_addr_rotor++; |
| 740 | if (tm->last_v4_addr_rotor >= vec_len (tcp_cfg.ip4_src_addrs)) |
| 741 | tm->last_v4_addr_rotor = 0; |
| 742 | lcl_addr->ip4.as_u32 = tcp_cfg.ip4_src_addrs[index].as_u32; |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 743 | } |
| 744 | else |
| 745 | { |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 746 | index = tm->last_v6_addr_rotor++; |
| 747 | if (tm->last_v6_addr_rotor >= vec_len (tcp_cfg.ip6_src_addrs)) |
| 748 | tm->last_v6_addr_rotor = 0; |
| 749 | clib_memcpy_fast (&lcl_addr->ip6, &tcp_cfg.ip6_src_addrs[index], |
Dave Barach | 178cf49 | 2018-11-13 16:34:13 -0500 | [diff] [blame] | 750 | sizeof (ip6_address_t)); |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 751 | } |
| 752 | port = transport_alloc_local_port (TRANSPORT_PROTO_TCP, lcl_addr); |
| 753 | if (port < 1) |
Florin Coras | 00e01d3 | 2019-10-21 16:07:46 -0700 | [diff] [blame] | 754 | return SESSION_E_NOPORT; |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 755 | *lcl_port = port; |
| 756 | return 0; |
| 757 | } |
| 758 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 759 | static int |
Florin Coras | 5665ced | 2018-10-25 18:03:45 -0700 | [diff] [blame] | 760 | tcp_session_open (transport_endpoint_cfg_t * rmt) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 761 | { |
| 762 | tcp_main_t *tm = vnet_get_tcp_main (); |
| 763 | tcp_connection_t *tc; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 764 | ip46_address_t lcl_addr; |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 765 | u16 lcl_port; |
| 766 | int rv; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 767 | |
| 768 | /* |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 769 | * Allocate local endpoint |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 770 | */ |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 771 | if ((rmt->is_ip4 && vec_len (tcp_cfg.ip4_src_addrs)) |
| 772 | || (!rmt->is_ip4 && vec_len (tcp_cfg.ip6_src_addrs))) |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 773 | rv = tcp_alloc_custom_local_endpoint (tm, &lcl_addr, &lcl_port, |
| 774 | rmt->is_ip4); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 775 | else |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 776 | rv = transport_alloc_local_endpoint (TRANSPORT_PROTO_TCP, |
| 777 | rmt, &lcl_addr, &lcl_port); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 778 | |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 779 | if (rv) |
Florin Coras | 1e8d949 | 2020-04-10 14:51:46 +0000 | [diff] [blame] | 780 | { |
| 781 | if (rv != SESSION_E_PORTINUSE) |
| 782 | return rv; |
| 783 | |
| 784 | if (session_lookup_connection (rmt->fib_index, &lcl_addr, &rmt->ip, |
| 785 | lcl_port, rmt->port, TRANSPORT_PROTO_UDP, |
| 786 | rmt->is_ip4)) |
| 787 | return SESSION_E_PORTINUSE; |
| 788 | |
| 789 | /* 5-tuple is available so increase lcl endpoint refcount and proceed |
| 790 | * with connection allocation */ |
| 791 | transport_share_local_endpoint (TRANSPORT_PROTO_UDP, &lcl_addr, |
| 792 | lcl_port); |
| 793 | } |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 794 | |
| 795 | /* |
| 796 | * Create connection and send SYN |
| 797 | */ |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 798 | clib_spinlock_lock_if_init (&tm->half_open_lock); |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 799 | tc = tcp_half_open_connection_new (); |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 800 | ip_copy (&tc->c_rmt_ip, &rmt->ip, rmt->is_ip4); |
| 801 | ip_copy (&tc->c_lcl_ip, &lcl_addr, rmt->is_ip4); |
Florin Coras | 0e49568 | 2017-09-19 22:27:18 -0700 | [diff] [blame] | 802 | tc->c_rmt_port = rmt->port; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 803 | tc->c_lcl_port = clib_host_to_net_u16 (lcl_port); |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 804 | tc->c_is_ip4 = rmt->is_ip4; |
Florin Coras | 3cbc04b | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 805 | tc->c_proto = TRANSPORT_PROTO_TCP; |
Florin Coras | cea194d | 2017-10-02 00:18:51 -0700 | [diff] [blame] | 806 | tc->c_fib_index = rmt->fib_index; |
Florin Coras | 12f6936 | 2019-08-16 09:44:00 -0700 | [diff] [blame] | 807 | tc->cc_algo = tcp_cc_algo_get (tcp_cfg.cc_algo); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 808 | /* The other connection vars will be initialized after SYN ACK */ |
| 809 | tcp_connection_timers_init (tc); |
Florin Coras | ff19e3b | 2020-02-10 17:44:13 +0000 | [diff] [blame] | 810 | tc->mss = rmt->mss; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 811 | |
Florin Coras | a436a42 | 2019-08-20 07:09:31 -0700 | [diff] [blame] | 812 | TCP_EVT (TCP_EVT_OPEN, tc); |
Florin Coras | 6534b7a | 2017-07-18 05:38:03 -0400 | [diff] [blame] | 813 | tc->state = TCP_STATE_SYN_SENT; |
Florin Coras | 4eeeaaf | 2017-09-05 14:03:37 -0400 | [diff] [blame] | 814 | tcp_init_snd_vars (tc); |
Florin Coras | 6534b7a | 2017-07-18 05:38:03 -0400 | [diff] [blame] | 815 | tcp_send_syn (tc); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 816 | clib_spinlock_unlock_if_init (&tm->half_open_lock); |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 817 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 818 | return tc->c_c_index; |
| 819 | } |
| 820 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 821 | static u8 * |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 822 | format_tcp_session (u8 * s, va_list * args) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 823 | { |
| 824 | u32 tci = va_arg (*args, u32); |
| 825 | u32 thread_index = va_arg (*args, u32); |
Florin Coras | bb292f4 | 2017-05-19 09:49:19 -0700 | [diff] [blame] | 826 | u32 verbose = va_arg (*args, u32); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 827 | tcp_connection_t *tc; |
| 828 | |
| 829 | tc = tcp_connection_get (tci, thread_index); |
Florin Coras | 6cf30ad | 2017-04-04 23:08:23 -0700 | [diff] [blame] | 830 | if (tc) |
Florin Coras | 93992a9 | 2017-05-24 18:03:56 -0700 | [diff] [blame] | 831 | s = format (s, "%U", format_tcp_connection, tc, verbose); |
Florin Coras | 6cf30ad | 2017-04-04 23:08:23 -0700 | [diff] [blame] | 832 | else |
Florin Coras | 1f152cd | 2017-08-18 19:28:03 -0700 | [diff] [blame] | 833 | s = format (s, "empty\n"); |
Florin Coras | 93992a9 | 2017-05-24 18:03:56 -0700 | [diff] [blame] | 834 | return s; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 835 | } |
| 836 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 837 | static u8 * |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 838 | format_tcp_listener_session (u8 * s, va_list * args) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 839 | { |
| 840 | u32 tci = va_arg (*args, u32); |
Aloys Augustin | a0abbff | 2019-07-12 12:16:16 +0200 | [diff] [blame] | 841 | u32 __clib_unused thread_index = va_arg (*args, u32); |
Florin Coras | 3389dd2 | 2019-02-01 18:00:05 -0800 | [diff] [blame] | 842 | u32 verbose = va_arg (*args, u32); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 843 | tcp_connection_t *tc = tcp_listener_get (tci); |
Florin Coras | 7bf6ed6 | 2020-09-23 12:02:08 -0700 | [diff] [blame] | 844 | s = format (s, "%-" SESSION_CLI_ID_LEN "U", format_tcp_connection_id, tc); |
Florin Coras | 3389dd2 | 2019-02-01 18:00:05 -0800 | [diff] [blame] | 845 | if (verbose) |
Florin Coras | 7bf6ed6 | 2020-09-23 12:02:08 -0700 | [diff] [blame] | 846 | s = format (s, "%-" SESSION_CLI_STATE_LEN "U", format_tcp_state, |
| 847 | tc->state); |
Florin Coras | 3389dd2 | 2019-02-01 18:00:05 -0800 | [diff] [blame] | 848 | return s; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 849 | } |
| 850 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 851 | static u8 * |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 852 | format_tcp_half_open_session (u8 * s, va_list * args) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 853 | { |
| 854 | u32 tci = va_arg (*args, u32); |
Aloys Augustin | a0abbff | 2019-07-12 12:16:16 +0200 | [diff] [blame] | 855 | u32 __clib_unused thread_index = va_arg (*args, u32); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 856 | tcp_connection_t *tc = tcp_half_open_connection_get (tci); |
Florin Coras | bb292f4 | 2017-05-19 09:49:19 -0700 | [diff] [blame] | 857 | return format (s, "%U", format_tcp_connection_id, tc); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 858 | } |
| 859 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 860 | static transport_connection_t * |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 861 | tcp_session_get_transport (u32 conn_index, u32 thread_index) |
| 862 | { |
| 863 | tcp_connection_t *tc = tcp_connection_get (conn_index, thread_index); |
Florin Coras | 5bb23ec | 2019-08-31 09:45:13 -0700 | [diff] [blame] | 864 | if (PREDICT_FALSE (!tc)) |
| 865 | return 0; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 866 | return &tc->connection; |
| 867 | } |
| 868 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 869 | static transport_connection_t * |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 870 | tcp_half_open_session_get_transport (u32 conn_index) |
| 871 | { |
| 872 | tcp_connection_t *tc = tcp_half_open_connection_get (conn_index); |
| 873 | return &tc->connection; |
| 874 | } |
| 875 | |
Florin Coras | 04ae827 | 2021-04-12 19:55:37 -0700 | [diff] [blame] | 876 | static int |
| 877 | tcp_set_attribute (tcp_connection_t *tc, transport_endpt_attr_t *attr) |
| 878 | { |
| 879 | int rv = 0; |
| 880 | |
| 881 | switch (attr->type) |
| 882 | { |
| 883 | case TRANSPORT_ENDPT_ATTR_NEXT_OUTPUT_NODE: |
| 884 | tc->next_node_index = attr->next_output_node & 0xffffffff; |
| 885 | tc->next_node_opaque = attr->next_output_node >> 32; |
| 886 | break; |
| 887 | case TRANSPORT_ENDPT_ATTR_MSS: |
| 888 | tc->mss = attr->mss; |
| 889 | tc->snd_mss = clib_min (tc->snd_mss, tc->mss); |
| 890 | break; |
| 891 | case TRANSPORT_ENDPT_ATTR_FLAGS: |
| 892 | if (attr->flags & TRANSPORT_ENDPT_ATTR_F_CSUM_OFFLOAD) |
| 893 | tc->cfg_flags |= TCP_CFG_F_NO_CSUM_OFFLOAD; |
| 894 | else |
| 895 | tc->cfg_flags &= ~TCP_CFG_F_NO_CSUM_OFFLOAD; |
| 896 | if (attr->flags & TRANSPORT_ENDPT_ATTR_F_GSO) |
| 897 | { |
| 898 | if (!(tc->cfg_flags & TCP_CFG_F_TSO)) |
| 899 | tcp_check_gso (tc); |
| 900 | tc->cfg_flags &= ~TCP_CFG_F_NO_TSO; |
| 901 | } |
| 902 | else |
| 903 | { |
| 904 | tc->cfg_flags |= TCP_CFG_F_NO_TSO; |
| 905 | tc->cfg_flags &= ~TCP_CFG_F_TSO; |
| 906 | } |
Florin Coras | 7fdf8b2 | 2021-04-15 08:50:00 -0700 | [diff] [blame] | 907 | if (attr->flags & TRANSPORT_ENDPT_ATTR_F_RATE_SAMPLING) |
| 908 | { |
| 909 | if (!(tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)) |
| 910 | tcp_bt_init (tc); |
| 911 | tc->cfg_flags |= TCP_CFG_F_RATE_SAMPLE; |
| 912 | } |
| 913 | else |
| 914 | { |
| 915 | if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE) |
| 916 | tcp_bt_cleanup (tc); |
| 917 | tc->cfg_flags &= ~TCP_CFG_F_RATE_SAMPLE; |
| 918 | } |
Florin Coras | 04ae827 | 2021-04-12 19:55:37 -0700 | [diff] [blame] | 919 | break; |
| 920 | case TRANSPORT_ENDPT_ATTR_CC_ALGO: |
| 921 | if (tc->cc_algo == tcp_cc_algo_get (attr->cc_algo)) |
| 922 | break; |
| 923 | tcp_cc_cleanup (tc); |
| 924 | tc->cc_algo = tcp_cc_algo_get (attr->cc_algo); |
| 925 | tcp_cc_init (tc); |
| 926 | break; |
| 927 | default: |
| 928 | rv = -1; |
| 929 | break; |
| 930 | } |
| 931 | |
| 932 | return rv; |
| 933 | } |
| 934 | |
| 935 | static int |
| 936 | tcp_get_attribute (tcp_connection_t *tc, transport_endpt_attr_t *attr) |
| 937 | { |
| 938 | int rv = 0; |
| 939 | u64 non; |
| 940 | |
| 941 | switch (attr->type) |
| 942 | { |
| 943 | case TRANSPORT_ENDPT_ATTR_NEXT_OUTPUT_NODE: |
| 944 | non = (u64) tc->next_node_opaque << 32 | tc->next_node_index; |
| 945 | attr->next_output_node = non; |
| 946 | break; |
| 947 | case TRANSPORT_ENDPT_ATTR_MSS: |
| 948 | attr->mss = tc->snd_mss; |
| 949 | break; |
| 950 | case TRANSPORT_ENDPT_ATTR_FLAGS: |
| 951 | attr->flags = 0; |
| 952 | if (!(tc->cfg_flags & TCP_CFG_F_NO_CSUM_OFFLOAD)) |
| 953 | attr->flags |= TRANSPORT_ENDPT_ATTR_F_CSUM_OFFLOAD; |
| 954 | if (tc->cfg_flags & TCP_CFG_F_TSO) |
| 955 | attr->flags |= TRANSPORT_ENDPT_ATTR_F_GSO; |
Florin Coras | 7fdf8b2 | 2021-04-15 08:50:00 -0700 | [diff] [blame] | 956 | if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE) |
| 957 | attr->flags |= TRANSPORT_ENDPT_ATTR_F_RATE_SAMPLING; |
Florin Coras | 04ae827 | 2021-04-12 19:55:37 -0700 | [diff] [blame] | 958 | break; |
| 959 | case TRANSPORT_ENDPT_ATTR_CC_ALGO: |
| 960 | attr->cc_algo = tc->cc_algo - tcp_main.cc_algos; |
| 961 | break; |
| 962 | default: |
| 963 | rv = -1; |
| 964 | break; |
| 965 | } |
| 966 | |
| 967 | return rv; |
| 968 | } |
| 969 | |
| 970 | static int |
| 971 | tcp_session_attribute (u32 conn_index, u32 thread_index, u8 is_get, |
| 972 | transport_endpt_attr_t *attr) |
| 973 | { |
| 974 | tcp_connection_t *tc = tcp_connection_get (conn_index, thread_index); |
| 975 | |
| 976 | if (PREDICT_FALSE (!tc)) |
| 977 | return -1; |
| 978 | |
| 979 | if (is_get) |
| 980 | return tcp_get_attribute (tc, attr); |
| 981 | else |
| 982 | return tcp_set_attribute (tc, attr); |
| 983 | } |
| 984 | |
Simon Zhang | 1146ff4 | 2019-09-02 22:54:00 +0800 | [diff] [blame] | 985 | static u16 |
| 986 | tcp_session_cal_goal_size (tcp_connection_t * tc) |
| 987 | { |
| 988 | u16 goal_size = tc->snd_mss; |
| 989 | |
Simon Zhang | 23c3d34 | 2020-09-15 23:40:28 +0800 | [diff] [blame] | 990 | goal_size = tcp_cfg.max_gso_size - tc->snd_mss % tcp_cfg.max_gso_size; |
Simon Zhang | 1146ff4 | 2019-09-02 22:54:00 +0800 | [diff] [blame] | 991 | goal_size = clib_min (goal_size, tc->snd_wnd / 2); |
| 992 | |
Simon Zhang | 8a047ed | 2019-09-24 21:16:56 +0800 | [diff] [blame] | 993 | return goal_size > tc->snd_mss ? goal_size : tc->snd_mss; |
Simon Zhang | 1146ff4 | 2019-09-02 22:54:00 +0800 | [diff] [blame] | 994 | } |
| 995 | |
Florin Coras | 3af90fc | 2017-05-03 21:09:42 -0700 | [diff] [blame] | 996 | always_inline u32 |
| 997 | tcp_round_snd_space (tcp_connection_t * tc, u32 snd_space) |
| 998 | { |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 999 | if (PREDICT_FALSE (tc->snd_wnd < tc->snd_mss)) |
Florin Coras | 3af90fc | 2017-05-03 21:09:42 -0700 | [diff] [blame] | 1000 | { |
Florin Coras | db84e57 | 2017-05-09 18:54:52 -0700 | [diff] [blame] | 1001 | return tc->snd_wnd <= snd_space ? tc->snd_wnd : 0; |
Florin Coras | 3af90fc | 2017-05-03 21:09:42 -0700 | [diff] [blame] | 1002 | } |
| 1003 | |
Florin Coras | 1f152cd | 2017-08-18 19:28:03 -0700 | [diff] [blame] | 1004 | /* If not snd_wnd constrained and we can't write at least a segment, |
| 1005 | * don't try at all */ |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 1006 | if (PREDICT_FALSE (snd_space < tc->snd_mss)) |
Florin Coras | 9d06304 | 2017-09-14 03:08:00 -0400 | [diff] [blame] | 1007 | return snd_space < tc->cwnd ? 0 : snd_space; |
Florin Coras | 3af90fc | 2017-05-03 21:09:42 -0700 | [diff] [blame] | 1008 | |
| 1009 | /* round down to mss multiple */ |
| 1010 | return snd_space - (snd_space % tc->snd_mss); |
| 1011 | } |
| 1012 | |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 1013 | /** |
| 1014 | * Compute tx window session is allowed to fill. |
Florin Coras | bb292f4 | 2017-05-19 09:49:19 -0700 | [diff] [blame] | 1015 | * |
| 1016 | * Takes into account available send space, snd_mss and the congestion |
| 1017 | * state of the connection. If possible, the value returned is a multiple |
| 1018 | * of snd_mss. |
| 1019 | * |
| 1020 | * @param tc tcp connection |
| 1021 | * @return number of bytes session is allowed to write |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 1022 | */ |
Florin Coras | 45ca73f | 2018-09-27 09:19:29 -0700 | [diff] [blame] | 1023 | static inline u32 |
| 1024 | tcp_snd_space_inline (tcp_connection_t * tc) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1025 | { |
Florin Coras | 36ebcff | 2019-09-12 18:36:44 -0700 | [diff] [blame] | 1026 | int snd_space; |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 1027 | |
Florin Coras | 7d8100f | 2020-10-21 23:24:53 -0700 | [diff] [blame] | 1028 | /* Fast path is disabled when recovery is on. @ref tcp_session_custom_tx |
| 1029 | * controls both retransmits and the sending of new data while congested |
| 1030 | */ |
| 1031 | if (PREDICT_FALSE (tcp_in_cong_recovery (tc) |
Florin Coras | 4af830c | 2018-12-04 09:21:36 -0800 | [diff] [blame] | 1032 | || tc->state == TCP_STATE_CLOSED)) |
Florin Coras | 36ee9f1 | 2018-11-02 12:52:10 -0700 | [diff] [blame] | 1033 | return 0; |
| 1034 | |
| 1035 | snd_space = tcp_available_output_snd_space (tc); |
| 1036 | |
Florin Coras | 36ebcff | 2019-09-12 18:36:44 -0700 | [diff] [blame] | 1037 | /* If we got dupacks or sacked bytes but we're not yet in recovery, try |
| 1038 | * to force the peer to send enough dupacks to start retransmitting as |
| 1039 | * per Limited Transmit (RFC3042) |
| 1040 | */ |
Florin Coras | cc4d6d0 | 2020-07-29 23:03:39 -0700 | [diff] [blame] | 1041 | if (PREDICT_FALSE (tc->rcv_dupacks || tc->sack_sb.sacked_bytes)) |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 1042 | { |
Florin Coras | cc4d6d0 | 2020-07-29 23:03:39 -0700 | [diff] [blame] | 1043 | int snt_limited, n_pkts; |
| 1044 | |
| 1045 | n_pkts = tcp_opts_sack_permitted (&tc->rcv_opts) |
| 1046 | ? tc->sack_sb.reorder - 1 : 2; |
| 1047 | |
| 1048 | if ((seq_lt (tc->limited_transmit, tc->snd_nxt - n_pkts * tc->snd_mss) |
| 1049 | || seq_gt (tc->limited_transmit, tc->snd_nxt))) |
Florin Coras | 36ee9f1 | 2018-11-02 12:52:10 -0700 | [diff] [blame] | 1050 | tc->limited_transmit = tc->snd_nxt; |
Florin Coras | 36ebcff | 2019-09-12 18:36:44 -0700 | [diff] [blame] | 1051 | |
Florin Coras | 36ee9f1 | 2018-11-02 12:52:10 -0700 | [diff] [blame] | 1052 | ASSERT (seq_leq (tc->limited_transmit, tc->snd_nxt)); |
Florin Coras | f03a59a | 2017-06-09 21:07:32 -0700 | [diff] [blame] | 1053 | |
Florin Coras | cc4d6d0 | 2020-07-29 23:03:39 -0700 | [diff] [blame] | 1054 | snt_limited = tc->snd_nxt - tc->limited_transmit; |
| 1055 | snd_space = clib_max (n_pkts * tc->snd_mss - snt_limited, 0); |
Florin Coras | 3af90fc | 2017-05-03 21:09:42 -0700 | [diff] [blame] | 1056 | } |
Florin Coras | 36ee9f1 | 2018-11-02 12:52:10 -0700 | [diff] [blame] | 1057 | return tcp_round_snd_space (tc, snd_space); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1058 | } |
| 1059 | |
Florin Coras | 45ca73f | 2018-09-27 09:19:29 -0700 | [diff] [blame] | 1060 | u32 |
| 1061 | tcp_snd_space (tcp_connection_t * tc) |
| 1062 | { |
| 1063 | return tcp_snd_space_inline (tc); |
| 1064 | } |
| 1065 | |
Florin Coras | 70f879d | 2020-03-13 17:54:42 +0000 | [diff] [blame] | 1066 | static int |
| 1067 | tcp_session_send_params (transport_connection_t * trans_conn, |
| 1068 | transport_send_params_t * sp) |
Florin Coras | bb292f4 | 2017-05-19 09:49:19 -0700 | [diff] [blame] | 1069 | { |
| 1070 | tcp_connection_t *tc = (tcp_connection_t *) trans_conn; |
Florin Coras | bb292f4 | 2017-05-19 09:49:19 -0700 | [diff] [blame] | 1071 | |
Florin Coras | 70f879d | 2020-03-13 17:54:42 +0000 | [diff] [blame] | 1072 | /* Ensure snd_mss does accurately reflect the amount of data we can push |
| 1073 | * in a segment. This also makes sure that options are updated according to |
| 1074 | * the current state of the connection. */ |
| 1075 | tcp_update_burst_snd_vars (tc); |
| 1076 | |
| 1077 | if (PREDICT_FALSE (tc->cfg_flags & TCP_CFG_F_TSO)) |
| 1078 | sp->snd_mss = tcp_session_cal_goal_size (tc); |
| 1079 | else |
| 1080 | sp->snd_mss = tc->snd_mss; |
| 1081 | |
| 1082 | sp->snd_space = clib_min (tcp_snd_space_inline (tc), |
| 1083 | tc->snd_wnd - (tc->snd_nxt - tc->snd_una)); |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 1084 | |
| 1085 | ASSERT (seq_geq (tc->snd_nxt, tc->snd_una)); |
Florin Coras | 6792ec0 | 2017-03-13 03:49:51 -0700 | [diff] [blame] | 1086 | /* This still works if fast retransmit is on */ |
Florin Coras | 70f879d | 2020-03-13 17:54:42 +0000 | [diff] [blame] | 1087 | sp->tx_offset = tc->snd_nxt - tc->snd_una; |
| 1088 | |
Florin Coras | 6080e0d | 2020-03-13 20:39:43 +0000 | [diff] [blame] | 1089 | sp->flags = sp->snd_space ? 0 : TRANSPORT_SND_F_DESCHED; |
Florin Coras | 70f879d | 2020-03-13 17:54:42 +0000 | [diff] [blame] | 1090 | |
| 1091 | return 0; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1092 | } |
| 1093 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 1094 | static void |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1095 | tcp_timer_waitclose_handler (tcp_connection_t * tc) |
| 1096 | { |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1097 | tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index); |
| 1098 | |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1099 | switch (tc->state) |
| 1100 | { |
| 1101 | case TCP_STATE_CLOSE_WAIT: |
| 1102 | tcp_connection_timers_reset (tc); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1103 | /* App never returned with a close */ |
| 1104 | if (!(tc->flags & TCP_CONN_FINPNDG)) |
| 1105 | { |
| 1106 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1107 | session_transport_closed_notify (&tc->connection); |
| 1108 | tcp_program_cleanup (wrk, tc); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1109 | tcp_worker_stats_inc (wrk, to_closewait, 1); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1110 | break; |
| 1111 | } |
| 1112 | |
| 1113 | /* Send FIN either way and switch to LAST_ACK. */ |
| 1114 | tcp_cong_recovery_off (tc); |
| 1115 | /* Make sure we don't try to send unsent data */ |
| 1116 | tc->snd_nxt = tc->snd_una; |
| 1117 | tcp_send_fin (tc); |
| 1118 | tcp_connection_set_state (tc, TCP_STATE_LAST_ACK); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1119 | session_transport_closed_notify (&tc->connection); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1120 | |
| 1121 | /* Make sure we don't wait in LAST ACK forever */ |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1122 | tcp_timer_set (&wrk->timer_wheel, tc, TCP_TIMER_WAITCLOSE, |
| 1123 | tcp_cfg.lastack_time); |
| 1124 | tcp_worker_stats_inc (wrk, to_closewait2, 1); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1125 | |
| 1126 | /* Don't delete the connection yet */ |
| 1127 | break; |
| 1128 | case TCP_STATE_FIN_WAIT_1: |
| 1129 | tcp_connection_timers_reset (tc); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1130 | if (tc->flags & TCP_CONN_FINPNDG) |
| 1131 | { |
| 1132 | /* If FIN pending, we haven't sent everything, but we did try. |
| 1133 | * Notify session layer that transport is closed. */ |
| 1134 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
| 1135 | tcp_send_reset (tc); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1136 | tcp_program_cleanup (wrk, tc); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1137 | } |
| 1138 | else |
| 1139 | { |
| 1140 | /* We've sent the fin but no progress. Close the connection and |
| 1141 | * to make sure everything is flushed, setup a cleanup timer */ |
| 1142 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1143 | tcp_program_cleanup (wrk, tc); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1144 | } |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1145 | session_transport_closed_notify (&tc->connection); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1146 | tcp_worker_stats_inc (wrk, to_finwait1, 1); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1147 | break; |
| 1148 | case TCP_STATE_LAST_ACK: |
| 1149 | tcp_connection_timers_reset (tc); |
| 1150 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1151 | session_transport_closed_notify (&tc->connection); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1152 | tcp_program_cleanup (wrk, tc); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1153 | tcp_worker_stats_inc (wrk, to_lastack, 1); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1154 | break; |
| 1155 | case TCP_STATE_CLOSING: |
| 1156 | tcp_connection_timers_reset (tc); |
| 1157 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1158 | session_transport_closed_notify (&tc->connection); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1159 | tcp_program_cleanup (wrk, tc); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1160 | tcp_worker_stats_inc (wrk, to_closing, 1); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1161 | break; |
| 1162 | case TCP_STATE_FIN_WAIT_2: |
| 1163 | tcp_send_reset (tc); |
| 1164 | tcp_connection_timers_reset (tc); |
| 1165 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
| 1166 | session_transport_closed_notify (&tc->connection); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1167 | tcp_program_cleanup (wrk, tc); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1168 | tcp_worker_stats_inc (wrk, to_finwait2, 1); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1169 | break; |
| 1170 | case TCP_STATE_TIME_WAIT: |
| 1171 | tcp_connection_set_state (tc, TCP_STATE_CLOSED); |
| 1172 | tcp_program_cleanup (wrk, tc); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1173 | break; |
| 1174 | default: |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1175 | clib_warning ("waitclose in state: %U", format_tcp_state, tc->state); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1176 | break; |
| 1177 | } |
| 1178 | } |
| 1179 | |
| 1180 | /* *INDENT-OFF* */ |
| 1181 | static timer_expiration_handler *timer_expiration_handlers[TCP_N_TIMERS] = |
| 1182 | { |
| 1183 | tcp_timer_retransmit_handler, |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1184 | tcp_timer_persist_handler, |
| 1185 | tcp_timer_waitclose_handler, |
| 1186 | tcp_timer_retransmit_syn_handler, |
| 1187 | }; |
| 1188 | /* *INDENT-ON* */ |
| 1189 | |
| 1190 | static void |
| 1191 | tcp_dispatch_pending_timers (tcp_worker_ctx_t * wrk) |
| 1192 | { |
| 1193 | u32 n_timers, connection_index, timer_id, thread_index, timer_handle; |
| 1194 | tcp_connection_t *tc; |
| 1195 | int i; |
| 1196 | |
| 1197 | if (!(n_timers = clib_fifo_elts (wrk->pending_timers))) |
| 1198 | return; |
| 1199 | |
| 1200 | thread_index = wrk->vm->thread_index; |
Florin Coras | a9d8cb4 | 2020-02-20 05:45:31 +0000 | [diff] [blame] | 1201 | for (i = 0; i < clib_min (n_timers, wrk->max_timers_per_loop); i++) |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1202 | { |
| 1203 | clib_fifo_sub1 (wrk->pending_timers, timer_handle); |
| 1204 | connection_index = timer_handle & 0x0FFFFFFF; |
| 1205 | timer_id = timer_handle >> 28; |
| 1206 | |
| 1207 | if (PREDICT_TRUE (timer_id != TCP_TIMER_RETRANSMIT_SYN)) |
| 1208 | tc = tcp_connection_get (connection_index, thread_index); |
| 1209 | else |
| 1210 | tc = tcp_half_open_connection_get (connection_index); |
| 1211 | |
| 1212 | if (PREDICT_FALSE (!tc)) |
| 1213 | continue; |
| 1214 | |
Florin Coras | 1caf7f1 | 2020-07-16 10:05:02 -0700 | [diff] [blame] | 1215 | /* Skip if the timer is not pending. Probably it was reset while |
Florin Coras | aa04395 | 2020-10-08 13:33:20 -0700 | [diff] [blame] | 1216 | * waiting for dispatch */ |
Florin Coras | 1caf7f1 | 2020-07-16 10:05:02 -0700 | [diff] [blame] | 1217 | if (PREDICT_FALSE (!(tc->pending_timers & (1 << timer_id)))) |
| 1218 | continue; |
| 1219 | |
| 1220 | tc->pending_timers &= ~(1 << timer_id); |
| 1221 | |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1222 | /* Skip timer if it was rearmed while pending dispatch */ |
| 1223 | if (PREDICT_FALSE (tc->timers[timer_id] != TCP_TIMER_HANDLE_INVALID)) |
| 1224 | continue; |
| 1225 | |
| 1226 | (*timer_expiration_handlers[timer_id]) (tc); |
| 1227 | } |
Florin Coras | a9d8cb4 | 2020-02-20 05:45:31 +0000 | [diff] [blame] | 1228 | |
| 1229 | if (thread_index == 0 && clib_fifo_elts (wrk->pending_timers)) |
Florin Coras | 5484daa | 2020-03-27 23:55:06 +0000 | [diff] [blame] | 1230 | session_queue_run_on_main_thread (wrk->vm); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1231 | } |
| 1232 | |
| 1233 | static void |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1234 | tcp_handle_cleanups (tcp_worker_ctx_t * wrk, clib_time_type_t now) |
| 1235 | { |
| 1236 | u32 thread_index = wrk->vm->thread_index; |
| 1237 | tcp_cleanup_req_t *req; |
| 1238 | tcp_connection_t *tc; |
| 1239 | |
| 1240 | while (clib_fifo_elts (wrk->pending_cleanups)) |
| 1241 | { |
| 1242 | req = clib_fifo_head (wrk->pending_cleanups); |
| 1243 | if (req->free_time > now) |
| 1244 | break; |
| 1245 | clib_fifo_sub2 (wrk->pending_cleanups, req); |
| 1246 | tc = tcp_connection_get (req->connection_index, thread_index); |
Florin Coras | 9f4db3c | 2020-03-10 19:34:28 +0000 | [diff] [blame] | 1247 | if (PREDICT_FALSE (!tc)) |
| 1248 | continue; |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1249 | session_transport_delete_notify (&tc->connection); |
| 1250 | tcp_connection_cleanup (tc); |
| 1251 | } |
| 1252 | } |
| 1253 | |
| 1254 | static void |
Florin Coras | 561af9b | 2017-12-09 10:19:43 -0800 | [diff] [blame] | 1255 | tcp_update_time (f64 now, u8 thread_index) |
| 1256 | { |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1257 | tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index); |
| 1258 | |
Florin Coras | 8f10b90 | 2021-04-02 18:32:00 -0700 | [diff] [blame] | 1259 | tcp_set_time_now (wrk, now); |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1260 | tcp_handle_cleanups (wrk, now); |
Florin Coras | 49036a5 | 2020-10-08 09:28:32 -0700 | [diff] [blame] | 1261 | tcp_timer_expire_timers (&wrk->timer_wheel, now); |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1262 | tcp_dispatch_pending_timers (wrk); |
Florin Coras | 561af9b | 2017-12-09 10:19:43 -0800 | [diff] [blame] | 1263 | } |
| 1264 | |
Florin Coras | 42ceddb | 2018-12-12 10:56:01 -0800 | [diff] [blame] | 1265 | static void |
| 1266 | tcp_session_flush_data (transport_connection_t * tconn) |
| 1267 | { |
| 1268 | tcp_connection_t *tc = (tcp_connection_t *) tconn; |
| 1269 | if (tc->flags & TCP_CONN_PSH_PENDING) |
| 1270 | return; |
| 1271 | tc->flags |= TCP_CONN_PSH_PENDING; |
Florin Coras | 4759683 | 2019-03-12 18:58:54 -0700 | [diff] [blame] | 1272 | tc->psh_seq = tc->snd_una + transport_max_tx_dequeue (tconn) - 1; |
Florin Coras | 42ceddb | 2018-12-12 10:56:01 -0800 | [diff] [blame] | 1273 | } |
| 1274 | |
Florin Coras | 5a41fd5 | 2021-04-19 10:17:26 -0700 | [diff] [blame^] | 1275 | static int |
| 1276 | tcp_session_app_rx_evt (transport_connection_t *conn) |
| 1277 | { |
| 1278 | tcp_connection_t *tc = (tcp_connection_t *) conn; |
| 1279 | u32 min_free, lo = 4 << 10, hi = 128 << 10; |
| 1280 | |
| 1281 | if (!(tc->flags & TCP_CONN_ZERO_RWND_SENT)) |
| 1282 | return 0; |
| 1283 | |
| 1284 | min_free = clib_clamp (transport_rx_fifo_size (conn) >> 3, lo, hi); |
| 1285 | if (transport_max_rx_enqueue (conn) < min_free) |
| 1286 | { |
| 1287 | transport_rx_fifo_req_deq_ntf (conn); |
| 1288 | return 0; |
| 1289 | } |
| 1290 | |
| 1291 | tcp_send_ack (tc); |
| 1292 | |
| 1293 | return 0; |
| 1294 | } |
| 1295 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1296 | /* *INDENT-OFF* */ |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 1297 | const static transport_proto_vft_t tcp_proto = { |
Florin Coras | 561af9b | 2017-12-09 10:19:43 -0800 | [diff] [blame] | 1298 | .enable = vnet_tcp_enable_disable, |
Florin Coras | 1ee7830 | 2019-02-05 15:51:15 -0800 | [diff] [blame] | 1299 | .start_listen = tcp_session_bind, |
| 1300 | .stop_listen = tcp_session_unbind, |
Florin Coras | 8b20bf5 | 2018-06-14 14:55:50 -0700 | [diff] [blame] | 1301 | .push_header = tcp_session_push_header, |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1302 | .get_connection = tcp_session_get_transport, |
| 1303 | .get_listener = tcp_session_get_listener, |
| 1304 | .get_half_open = tcp_half_open_session_get_transport, |
Florin Coras | 04ae827 | 2021-04-12 19:55:37 -0700 | [diff] [blame] | 1305 | .attribute = tcp_session_attribute, |
Florin Coras | 1ee7830 | 2019-02-05 15:51:15 -0800 | [diff] [blame] | 1306 | .connect = tcp_session_open, |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1307 | .close = tcp_session_close, |
| 1308 | .cleanup = tcp_session_cleanup, |
Florin Coras | d50ff7f | 2020-04-16 04:30:22 +0000 | [diff] [blame] | 1309 | .cleanup_ho = tcp_session_cleanup_ho, |
Florin Coras | dfb3b87 | 2019-08-16 17:48:44 -0700 | [diff] [blame] | 1310 | .reset = tcp_session_reset, |
Florin Coras | 70f879d | 2020-03-13 17:54:42 +0000 | [diff] [blame] | 1311 | .send_params = tcp_session_send_params, |
Florin Coras | 561af9b | 2017-12-09 10:19:43 -0800 | [diff] [blame] | 1312 | .update_time = tcp_update_time, |
Florin Coras | 42ceddb | 2018-12-12 10:56:01 -0800 | [diff] [blame] | 1313 | .flush_data = tcp_session_flush_data, |
Florin Coras | 26dd6de | 2019-07-23 23:54:47 -0700 | [diff] [blame] | 1314 | .custom_tx = tcp_session_custom_tx, |
Florin Coras | 5a41fd5 | 2021-04-19 10:17:26 -0700 | [diff] [blame^] | 1315 | .app_rx_evt = tcp_session_app_rx_evt, |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 1316 | .format_connection = format_tcp_session, |
| 1317 | .format_listener = format_tcp_listener_session, |
| 1318 | .format_half_open = format_tcp_half_open_session, |
Nathan Skrzypczak | e971bc9 | 2019-06-19 13:42:37 +0200 | [diff] [blame] | 1319 | .transport_options = { |
Florin Coras | 07063b8 | 2020-03-13 04:44:51 +0000 | [diff] [blame] | 1320 | .name = "tcp", |
| 1321 | .short_name = "T", |
Nathan Skrzypczak | e971bc9 | 2019-06-19 13:42:37 +0200 | [diff] [blame] | 1322 | .tx_type = TRANSPORT_TX_PEEK, |
| 1323 | .service_type = TRANSPORT_SERVICE_VC, |
| 1324 | }, |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1325 | }; |
| 1326 | /* *INDENT-ON* */ |
| 1327 | |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 1328 | void |
Florin Coras | c44a558 | 2018-11-01 16:30:54 -0700 | [diff] [blame] | 1329 | tcp_connection_tx_pacer_update (tcp_connection_t * tc) |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 1330 | { |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 1331 | if (!transport_connection_is_tx_paced (&tc->connection)) |
| 1332 | return; |
| 1333 | |
Florin Coras | 11e9e35 | 2019-11-13 19:09:47 -0800 | [diff] [blame] | 1334 | f64 srtt = clib_min ((f64) tc->srtt * TCP_TICK, tc->mrtt_us); |
| 1335 | |
Florin Coras | d206724 | 2019-08-16 10:33:49 -0700 | [diff] [blame] | 1336 | transport_connection_tx_pacer_update (&tc->connection, |
Florin Coras | 11e9e35 | 2019-11-13 19:09:47 -0800 | [diff] [blame] | 1337 | tcp_cc_get_pacing_rate (tc), |
| 1338 | srtt * CLIB_US_TIME_FREQ); |
Florin Coras | d67f112 | 2018-05-21 17:47:40 -0700 | [diff] [blame] | 1339 | } |
| 1340 | |
Florin Coras | c44a558 | 2018-11-01 16:30:54 -0700 | [diff] [blame] | 1341 | void |
| 1342 | tcp_connection_tx_pacer_reset (tcp_connection_t * tc, u32 window, |
| 1343 | u32 start_bucket) |
| 1344 | { |
Florin Coras | 36ebcff | 2019-09-12 18:36:44 -0700 | [diff] [blame] | 1345 | f64 srtt = clib_min ((f64) tc->srtt * TCP_TICK, tc->mrtt_us); |
Florin Coras | 599db9e | 2019-11-25 09:41:37 -0800 | [diff] [blame] | 1346 | transport_connection_tx_pacer_reset (&tc->connection, |
| 1347 | tcp_cc_get_pacing_rate (tc), |
| 1348 | start_bucket, |
Florin Coras | 11e9e35 | 2019-11-13 19:09:47 -0800 | [diff] [blame] | 1349 | srtt * CLIB_US_TIME_FREQ); |
Florin Coras | c44a558 | 2018-11-01 16:30:54 -0700 | [diff] [blame] | 1350 | } |
| 1351 | |
Florin Coras | 6080e0d | 2020-03-13 20:39:43 +0000 | [diff] [blame] | 1352 | void |
| 1353 | tcp_reschedule (tcp_connection_t * tc) |
| 1354 | { |
| 1355 | if (tcp_in_cong_recovery (tc) || tcp_snd_space_inline (tc)) |
| 1356 | transport_connection_reschedule (&tc->connection); |
| 1357 | } |
| 1358 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 1359 | static void |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1360 | tcp_expired_timers_dispatch (u32 * expired_timers) |
| 1361 | { |
Florin Coras | a9d8cb4 | 2020-02-20 05:45:31 +0000 | [diff] [blame] | 1362 | u32 thread_index = vlib_get_thread_index (), n_left, max_per_loop; |
| 1363 | u32 connection_index, timer_id, n_expired, max_loops; |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1364 | tcp_worker_ctx_t *wrk; |
Florin Coras | b72a0ff | 2019-11-22 17:38:25 -0800 | [diff] [blame] | 1365 | tcp_connection_t *tc; |
| 1366 | int i; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1367 | |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1368 | wrk = tcp_get_worker (thread_index); |
Florin Coras | 5e6305f | 2020-02-12 07:42:01 +0000 | [diff] [blame] | 1369 | n_expired = vec_len (expired_timers); |
Florin Coras | 0765d97 | 2020-03-18 21:26:41 +0000 | [diff] [blame] | 1370 | tcp_worker_stats_inc (wrk, timer_expirations, n_expired); |
Florin Coras | a9d8cb4 | 2020-02-20 05:45:31 +0000 | [diff] [blame] | 1371 | n_left = clib_fifo_elts (wrk->pending_timers); |
Florin Coras | 5e6305f | 2020-02-12 07:42:01 +0000 | [diff] [blame] | 1372 | |
Florin Coras | b72a0ff | 2019-11-22 17:38:25 -0800 | [diff] [blame] | 1373 | /* |
| 1374 | * Invalidate all timer handles before dispatching. This avoids dangling |
| 1375 | * index references to timer wheel pool entries that have been freed. |
| 1376 | */ |
Florin Coras | 5e6305f | 2020-02-12 07:42:01 +0000 | [diff] [blame] | 1377 | for (i = 0; i < n_expired; i++) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1378 | { |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1379 | connection_index = expired_timers[i] & 0x0FFFFFFF; |
| 1380 | timer_id = expired_timers[i] >> 28; |
| 1381 | |
Florin Coras | b72a0ff | 2019-11-22 17:38:25 -0800 | [diff] [blame] | 1382 | if (timer_id != TCP_TIMER_RETRANSMIT_SYN) |
| 1383 | tc = tcp_connection_get (connection_index, thread_index); |
| 1384 | else |
| 1385 | tc = tcp_half_open_connection_get (connection_index); |
| 1386 | |
Florin Coras | a436a42 | 2019-08-20 07:09:31 -0700 | [diff] [blame] | 1387 | TCP_EVT (TCP_EVT_TIMER_POP, connection_index, timer_id); |
Florin Coras | e69f495 | 2017-03-07 10:06:24 -0800 | [diff] [blame] | 1388 | |
Florin Coras | b72a0ff | 2019-11-22 17:38:25 -0800 | [diff] [blame] | 1389 | tc->timers[timer_id] = TCP_TIMER_HANDLE_INVALID; |
Florin Coras | 1caf7f1 | 2020-07-16 10:05:02 -0700 | [diff] [blame] | 1390 | tc->pending_timers |= (1 << timer_id); |
Florin Coras | b72a0ff | 2019-11-22 17:38:25 -0800 | [diff] [blame] | 1391 | } |
| 1392 | |
Florin Coras | aa38869 | 2020-02-14 23:41:25 +0000 | [diff] [blame] | 1393 | clib_fifo_add (wrk->pending_timers, expired_timers, n_expired); |
Florin Coras | a9d8cb4 | 2020-02-20 05:45:31 +0000 | [diff] [blame] | 1394 | |
| 1395 | max_loops = clib_max (1, 0.5 * TCP_TIMER_TICK * wrk->vm->loops_per_second); |
| 1396 | max_per_loop = clib_max ((n_left + n_expired) / max_loops, 10); |
| 1397 | max_per_loop = clib_min (max_per_loop, VLIB_FRAME_SIZE); |
| 1398 | wrk->max_timers_per_loop = clib_max (n_left ? wrk->max_timers_per_loop : 0, |
| 1399 | max_per_loop); |
| 1400 | |
| 1401 | if (thread_index == 0) |
Florin Coras | 5484daa | 2020-03-27 23:55:06 +0000 | [diff] [blame] | 1402 | session_queue_run_on_main_thread (wrk->vm); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1403 | } |
| 1404 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 1405 | static void |
Florin Coras | 18e0d4f | 2019-01-02 12:22:02 -0800 | [diff] [blame] | 1406 | tcp_initialize_iss_seed (tcp_main_t * tm) |
| 1407 | { |
| 1408 | u32 default_seed = random_default_seed (); |
| 1409 | u64 time_now = clib_cpu_time_now (); |
| 1410 | |
| 1411 | tm->iss_seed.first = (u64) random_u32 (&default_seed) << 32; |
| 1412 | tm->iss_seed.second = random_u64 (&time_now); |
| 1413 | } |
| 1414 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 1415 | static clib_error_t * |
Florin Coras | a0b34a7 | 2017-03-07 01:20:52 -0800 | [diff] [blame] | 1416 | tcp_main_enable (vlib_main_t * vm) |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1417 | { |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1418 | vlib_thread_main_t *vtm = vlib_get_thread_main (); |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1419 | u32 num_threads, n_workers, prealloc_conn_per_wrk; |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 1420 | tcp_connection_t *tc __attribute__ ((unused)); |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1421 | tcp_main_t *tm = vnet_get_tcp_main (); |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 1422 | tcp_worker_ctx_t *wrk; |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1423 | clib_error_t *error = 0; |
| 1424 | int thread; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1425 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1426 | if ((error = vlib_call_init_function (vm, ip_main_init))) |
| 1427 | return error; |
| 1428 | if ((error = vlib_call_init_function (vm, ip4_lookup_init))) |
| 1429 | return error; |
| 1430 | if ((error = vlib_call_init_function (vm, ip6_lookup_init))) |
| 1431 | return error; |
| 1432 | |
| 1433 | /* |
| 1434 | * Registrations |
| 1435 | */ |
| 1436 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1437 | ip4_register_protocol (IP_PROTOCOL_TCP, tcp4_input_node.index); |
root | c9d1c5b | 2017-08-15 12:58:31 -0400 | [diff] [blame] | 1438 | ip6_register_protocol (IP_PROTOCOL_TCP, tcp6_input_node.index); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1439 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1440 | /* |
| 1441 | * Initialize data structures |
| 1442 | */ |
| 1443 | |
| 1444 | num_threads = 1 /* main thread */ + vtm->n_threads; |
Florin Coras | e55a6d7 | 2018-10-31 23:09:22 -0700 | [diff] [blame] | 1445 | vec_validate (tm->wrk_ctx, num_threads - 1); |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1446 | n_workers = num_threads == 1 ? 1 : vtm->n_threads; |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1447 | prealloc_conn_per_wrk = tcp_cfg.preallocated_connections / n_workers; |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1448 | |
Florin Coras | 5484daa | 2020-03-27 23:55:06 +0000 | [diff] [blame] | 1449 | wrk = &tm->wrk_ctx[0]; |
| 1450 | wrk->tco_next_node[0] = vlib_node_get_next (vm, session_queue_node.index, |
| 1451 | tcp4_output_node.index); |
| 1452 | wrk->tco_next_node[1] = vlib_node_get_next (vm, session_queue_node.index, |
| 1453 | tcp6_output_node.index); |
| 1454 | |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1455 | for (thread = 0; thread < num_threads; thread++) |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 1456 | { |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 1457 | wrk = &tm->wrk_ctx[thread]; |
| 1458 | |
| 1459 | vec_validate (wrk->pending_deq_acked, 255); |
| 1460 | vec_validate (wrk->pending_disconnects, 255); |
| 1461 | vec_validate (wrk->pending_resets, 255); |
| 1462 | vec_reset_length (wrk->pending_deq_acked); |
| 1463 | vec_reset_length (wrk->pending_disconnects); |
| 1464 | vec_reset_length (wrk->pending_resets); |
Damjan Marion | 6ffb7c6 | 2021-03-26 13:06:13 +0100 | [diff] [blame] | 1465 | wrk->vm = vlib_get_main_by_index (thread); |
Florin Coras | a9d8cb4 | 2020-02-20 05:45:31 +0000 | [diff] [blame] | 1466 | wrk->max_timers_per_loop = 10; |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1467 | |
Florin Coras | 5484daa | 2020-03-27 23:55:06 +0000 | [diff] [blame] | 1468 | if (thread > 0) |
| 1469 | { |
| 1470 | wrk->tco_next_node[0] = tm->wrk_ctx[0].tco_next_node[0]; |
| 1471 | wrk->tco_next_node[1] = tm->wrk_ctx[0].tco_next_node[1]; |
| 1472 | } |
| 1473 | |
Florin Coras | be72ae6 | 2018-11-01 11:23:03 -0700 | [diff] [blame] | 1474 | /* |
| 1475 | * Preallocate connections. Assume that thread 0 won't |
| 1476 | * use preallocated threads when running multi-core |
| 1477 | */ |
| 1478 | if ((thread > 0 || num_threads == 1) && prealloc_conn_per_wrk) |
Florin Coras | 0426185 | 2020-02-12 01:24:29 +0000 | [diff] [blame] | 1479 | pool_init_fixed (wrk->connections, prealloc_conn_per_wrk); |
Florin Coras | 49036a5 | 2020-10-08 09:28:32 -0700 | [diff] [blame] | 1480 | |
| 1481 | tcp_timer_initialize_wheel (&wrk->timer_wheel, |
| 1482 | tcp_expired_timers_dispatch, |
| 1483 | vlib_time_now (vm)); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 1484 | } |
| 1485 | |
| 1486 | /* |
Dave Barach | b7f1faa | 2017-08-29 11:43:37 -0400 | [diff] [blame] | 1487 | * Use a preallocated half-open connection pool? |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 1488 | */ |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1489 | if (tcp_cfg.preallocated_half_open_connections) |
Dave Barach | b7f1faa | 2017-08-29 11:43:37 -0400 | [diff] [blame] | 1490 | pool_init_fixed (tm->half_open_connections, |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1491 | tcp_cfg.preallocated_half_open_connections); |
Dave Barach | 2c25a62 | 2017-06-26 11:35:07 -0400 | [diff] [blame] | 1492 | |
Florin Coras | 04e5344 | 2017-07-16 17:12:15 -0700 | [diff] [blame] | 1493 | if (num_threads > 1) |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 1494 | { |
| 1495 | clib_spinlock_init (&tm->half_open_lock); |
Florin Coras | 6881062 | 2017-07-24 17:40:28 -0700 | [diff] [blame] | 1496 | } |
Florin Coras | 66b1131 | 2017-07-31 17:18:03 -0700 | [diff] [blame] | 1497 | |
Florin Coras | 18e0d4f | 2019-01-02 12:22:02 -0800 | [diff] [blame] | 1498 | tcp_initialize_iss_seed (tm); |
Florin Coras | 66b1131 | 2017-07-31 17:18:03 -0700 | [diff] [blame] | 1499 | |
Damjan Marion | 8934a04 | 2019-02-09 23:29:26 +0100 | [diff] [blame] | 1500 | tm->bytes_per_buffer = vlib_buffer_get_default_data_size (vm); |
Florin Coras | 4e116fb | 2019-06-10 08:33:50 -0700 | [diff] [blame] | 1501 | tm->cc_last_type = TCP_CC_LAST; |
Florin Coras | 5484daa | 2020-03-27 23:55:06 +0000 | [diff] [blame] | 1502 | |
| 1503 | tm->ipl_next_node[0] = vlib_node_get_next (vm, session_queue_node.index, |
| 1504 | ip4_lookup_node.index); |
| 1505 | tm->ipl_next_node[1] = vlib_node_get_next (vm, session_queue_node.index, |
| 1506 | ip6_lookup_node.index); |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1507 | return error; |
| 1508 | } |
| 1509 | |
Florin Coras | a0b34a7 | 2017-03-07 01:20:52 -0800 | [diff] [blame] | 1510 | clib_error_t * |
| 1511 | vnet_tcp_enable_disable (vlib_main_t * vm, u8 is_en) |
| 1512 | { |
| 1513 | if (is_en) |
| 1514 | { |
| 1515 | if (tcp_main.is_enabled) |
| 1516 | return 0; |
| 1517 | |
| 1518 | return tcp_main_enable (vm); |
| 1519 | } |
| 1520 | else |
| 1521 | { |
| 1522 | tcp_main.is_enabled = 0; |
| 1523 | } |
| 1524 | |
| 1525 | return 0; |
| 1526 | } |
| 1527 | |
Pierre Pfister | 7fe51f3 | 2017-09-20 08:48:36 +0200 | [diff] [blame] | 1528 | void |
| 1529 | tcp_punt_unknown (vlib_main_t * vm, u8 is_ip4, u8 is_add) |
| 1530 | { |
| 1531 | tcp_main_t *tm = &tcp_main; |
| 1532 | if (is_ip4) |
| 1533 | tm->punt_unknown4 = is_add; |
| 1534 | else |
| 1535 | tm->punt_unknown6 = is_add; |
| 1536 | } |
| 1537 | |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1538 | /** |
| 1539 | * Initialize default values for tcp parameters |
| 1540 | */ |
| 1541 | static void |
| 1542 | tcp_configuration_init (void) |
| 1543 | { |
| 1544 | /* Initial wnd for SYN. Fifos are not allocated at that point so use some |
| 1545 | * predefined value. For SYN-ACK we still want the scale to be computed in |
| 1546 | * the same way */ |
| 1547 | tcp_cfg.max_rx_fifo = 32 << 20; |
| 1548 | tcp_cfg.min_rx_fifo = 4 << 10; |
| 1549 | |
Florin Coras | cedcf60 | 2019-08-27 12:15:43 -0700 | [diff] [blame] | 1550 | tcp_cfg.default_mtu = 1500; |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1551 | tcp_cfg.initial_cwnd_multiplier = 0; |
| 1552 | tcp_cfg.enable_tx_pacing = 1; |
Florin Coras | bbcfaac | 2019-10-10 13:52:04 -0700 | [diff] [blame] | 1553 | tcp_cfg.allow_tso = 0; |
Florin Coras | f4ce6ba | 2019-11-20 18:34:58 -0800 | [diff] [blame] | 1554 | tcp_cfg.csum_offload = 1; |
Florin Coras | e57df7c | 2020-04-17 16:10:51 +0000 | [diff] [blame] | 1555 | tcp_cfg.cc_algo = TCP_CC_CUBIC; |
Florin Coras | 017dc45 | 2019-08-30 11:06:35 -0700 | [diff] [blame] | 1556 | tcp_cfg.rwnd_min_update_ack = 1; |
Simon Zhang | 23c3d34 | 2020-09-15 23:40:28 +0800 | [diff] [blame] | 1557 | tcp_cfg.max_gso_size = TCP_MAX_GSO_SZ; |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1558 | |
Ryujiro Shibuya | 6fa0988 | 2020-10-22 05:26:32 +0000 | [diff] [blame] | 1559 | /* Time constants defined as timer tick (100us) multiples */ |
Ryujiro Shibuya | 6fa0988 | 2020-10-22 05:26:32 +0000 | [diff] [blame] | 1560 | tcp_cfg.closewait_time = 20000; /* 2s */ |
| 1561 | tcp_cfg.timewait_time = 100000; /* 10s */ |
| 1562 | tcp_cfg.finwait1_time = 600000; /* 60s */ |
| 1563 | tcp_cfg.lastack_time = 300000; /* 30s */ |
| 1564 | tcp_cfg.finwait2_time = 300000; /* 30s */ |
| 1565 | tcp_cfg.closing_time = 300000; /* 30s */ |
| 1566 | |
| 1567 | /* This value is seconds */ |
Florin Coras | 7a3a866 | 2020-02-22 02:27:21 +0000 | [diff] [blame] | 1568 | tcp_cfg.cleanup_time = 0.1; /* 100ms */ |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1569 | } |
| 1570 | |
Florin Coras | 0dbd517 | 2018-06-25 16:19:34 -0700 | [diff] [blame] | 1571 | static clib_error_t * |
Florin Coras | a0b34a7 | 2017-03-07 01:20:52 -0800 | [diff] [blame] | 1572 | tcp_init (vlib_main_t * vm) |
| 1573 | { |
| 1574 | tcp_main_t *tm = vnet_get_tcp_main (); |
Florin Coras | 4036427 | 2017-11-16 09:57:50 -0800 | [diff] [blame] | 1575 | ip_main_t *im = &ip_main; |
| 1576 | ip_protocol_info_t *pi; |
| 1577 | |
| 1578 | /* Session layer, and by implication tcp, are disabled by default */ |
Florin Coras | a0b34a7 | 2017-03-07 01:20:52 -0800 | [diff] [blame] | 1579 | tm->is_enabled = 0; |
Florin Coras | 4036427 | 2017-11-16 09:57:50 -0800 | [diff] [blame] | 1580 | |
| 1581 | /* Register with IP for header parsing */ |
| 1582 | pi = ip_get_protocol_info (im, IP_PROTOCOL_TCP); |
| 1583 | if (pi == 0) |
| 1584 | return clib_error_return (0, "TCP protocol info AWOL"); |
| 1585 | pi->format_header = format_tcp_header; |
| 1586 | pi->unformat_pg_edit = unformat_pg_tcp_header; |
| 1587 | |
Florin Coras | 561af9b | 2017-12-09 10:19:43 -0800 | [diff] [blame] | 1588 | /* Register as transport with session layer */ |
| 1589 | transport_register_protocol (TRANSPORT_PROTO_TCP, &tcp_proto, |
| 1590 | FIB_PROTOCOL_IP4, tcp4_output_node.index); |
| 1591 | transport_register_protocol (TRANSPORT_PROTO_TCP, &tcp_proto, |
| 1592 | FIB_PROTOCOL_IP6, tcp6_output_node.index); |
| 1593 | |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1594 | tcp_configuration_init (); |
| 1595 | |
Florin Coras | fbf278a | 2019-03-26 14:05:38 -0700 | [diff] [blame] | 1596 | tm->cc_algo_by_name = hash_create_string (0, sizeof (uword)); |
Florin Coras | 9094b5c | 2019-08-12 14:17:47 -0700 | [diff] [blame] | 1597 | |
Florin Coras | a0b34a7 | 2017-03-07 01:20:52 -0800 | [diff] [blame] | 1598 | return 0; |
| 1599 | } |
| 1600 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1601 | VLIB_INIT_FUNCTION (tcp_init); |
| 1602 | |
Dave Barach | 68b0fb0 | 2017-02-28 15:15:56 -0500 | [diff] [blame] | 1603 | /* |
| 1604 | * fd.io coding-style-patch-verification: ON |
| 1605 | * |
| 1606 | * Local Variables: |
| 1607 | * eval: (c-set-style "gnu") |
| 1608 | * End: |
| 1609 | */ |