1/* 2 * Copyright (c) 2006 Oracle. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 * 32 */ 33#include <linux/kernel.h> 34#include <linux/list.h> 35#include <linux/slab.h> 36#include <linux/export.h> 37#include <net/inet_hashtables.h> 38 39#include "rds.h" 40#include "loop.h" 41 42#define RDS_CONNECTION_HASH_BITS 12 43#define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS) 44#define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1) 45 46/* converting this to RCU is a chore for another day.. */ 47static DEFINE_SPINLOCK(rds_conn_lock); 48static unsigned long rds_conn_count; 49static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES]; 50static struct kmem_cache *rds_conn_slab; 51 52static struct hlist_head *rds_conn_bucket(__be32 laddr, __be32 faddr) 53{ 54 static u32 rds_hash_secret __read_mostly; 55 56 unsigned long hash; 57 58 net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret)); 59 60 /* Pass NULL, don't need struct net for hash */ 61 hash = __inet_ehashfn(be32_to_cpu(laddr), 0, 62 be32_to_cpu(faddr), 0, 63 rds_hash_secret); 64 return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK]; 65} 66 67#define rds_conn_info_set(var, test, suffix) do { \ 68 if (test) \ 69 var |= RDS_INFO_CONNECTION_FLAG_##suffix; \ 70} while (0) 71 72/* rcu read lock must be held or the connection spinlock */ 73static struct rds_connection *rds_conn_lookup(struct hlist_head *head, 74 __be32 laddr, __be32 faddr, 75 struct rds_transport *trans) 76{ 77 struct rds_connection *conn, *ret = NULL; 78 79 hlist_for_each_entry_rcu(conn, head, c_hash_node) { 80 if (conn->c_faddr == faddr && conn->c_laddr == laddr && 81 conn->c_trans == trans) { 82 ret = conn; 83 break; 84 } 85 } 86 rdsdebug("returning conn %p for %pI4 -> %pI4\n", ret, 87 &laddr, &faddr); 88 return ret; 89} 90 91/* 92 * This is called by transports as they're bringing down a connection. 93 * It clears partial message state so that the transport can start sending 94 * and receiving over this connection again in the future. It is up to 95 * the transport to have serialized this call with its send and recv. 96 */ 97static void rds_conn_reset(struct rds_connection *conn) 98{ 99 rdsdebug("connection %pI4 to %pI4 reset\n", 100 &conn->c_laddr, &conn->c_faddr); 101 102 rds_stats_inc(s_conn_reset); 103 rds_send_reset(conn); 104 conn->c_flags = 0; 105 106 /* Do not clear next_rx_seq here, else we cannot distinguish 107 * retransmitted packets from new packets, and will hand all 108 * of them to the application. That is not consistent with the 109 * reliability guarantees of RDS. */ 110} 111 112/* 113 * There is only every one 'conn' for a given pair of addresses in the 114 * system at a time. They contain messages to be retransmitted and so 115 * span the lifetime of the actual underlying transport connections. 116 * 117 * For now they are not garbage collected once they're created. They 118 * are torn down as the module is removed, if ever. 119 */ 120static struct rds_connection *__rds_conn_create(__be32 laddr, __be32 faddr, 121 struct rds_transport *trans, gfp_t gfp, 122 int is_outgoing) 123{ 124 struct rds_connection *conn, *parent = NULL; 125 struct hlist_head *head = rds_conn_bucket(laddr, faddr); 126 struct rds_transport *loop_trans; 127 unsigned long flags; 128 int ret; 129 struct rds_transport *otrans = trans; 130 131 if (!is_outgoing && otrans->t_type == RDS_TRANS_TCP) 132 goto new_conn; 133 rcu_read_lock(); 134 conn = rds_conn_lookup(head, laddr, faddr, trans); 135 if (conn && conn->c_loopback && conn->c_trans != &rds_loop_transport && 136 laddr == faddr && !is_outgoing) { 137 /* This is a looped back IB connection, and we're 138 * called by the code handling the incoming connect. 139 * We need a second connection object into which we 140 * can stick the other QP. */ 141 parent = conn; 142 conn = parent->c_passive; 143 } 144 rcu_read_unlock(); 145 if (conn) 146 goto out; 147 148new_conn: 149 conn = kmem_cache_zalloc(rds_conn_slab, gfp); 150 if (!conn) { 151 conn = ERR_PTR(-ENOMEM); 152 goto out; 153 } 154 155 INIT_HLIST_NODE(&conn->c_hash_node); 156 conn->c_laddr = laddr; 157 conn->c_faddr = faddr; 158 spin_lock_init(&conn->c_lock); 159 conn->c_next_tx_seq = 1; 160 161 init_waitqueue_head(&conn->c_waitq); 162 INIT_LIST_HEAD(&conn->c_send_queue); 163 INIT_LIST_HEAD(&conn->c_retrans); 164 165 ret = rds_cong_get_maps(conn); 166 if (ret) { 167 kmem_cache_free(rds_conn_slab, conn); 168 conn = ERR_PTR(ret); 169 goto out; 170 } 171 172 /* 173 * This is where a connection becomes loopback. If *any* RDS sockets 174 * can bind to the destination address then we'd rather the messages 175 * flow through loopback rather than either transport. 176 */ 177 loop_trans = rds_trans_get_preferred(faddr); 178 if (loop_trans) { 179 rds_trans_put(loop_trans); 180 conn->c_loopback = 1; 181 if (is_outgoing && trans->t_prefer_loopback) { 182 /* "outgoing" connection - and the transport 183 * says it wants the connection handled by the 184 * loopback transport. This is what TCP does. 185 */ 186 trans = &rds_loop_transport; 187 } 188 } 189 190 conn->c_trans = trans; 191 192 ret = trans->conn_alloc(conn, gfp); 193 if (ret) { 194 kmem_cache_free(rds_conn_slab, conn); 195 conn = ERR_PTR(ret); 196 goto out; 197 } 198 199 atomic_set(&conn->c_state, RDS_CONN_DOWN); 200 conn->c_send_gen = 0; 201 conn->c_reconnect_jiffies = 0; 202 INIT_DELAYED_WORK(&conn->c_send_w, rds_send_worker); 203 INIT_DELAYED_WORK(&conn->c_recv_w, rds_recv_worker); 204 INIT_DELAYED_WORK(&conn->c_conn_w, rds_connect_worker); 205 INIT_WORK(&conn->c_down_w, rds_shutdown_worker); 206 mutex_init(&conn->c_cm_lock); 207 conn->c_flags = 0; 208 209 rdsdebug("allocated conn %p for %pI4 -> %pI4 over %s %s\n", 210 conn, &laddr, &faddr, 211 trans->t_name ? trans->t_name : "[unknown]", 212 is_outgoing ? "(outgoing)" : ""); 213 214 /* 215 * Since we ran without holding the conn lock, someone could 216 * have created the same conn (either normal or passive) in the 217 * interim. We check while holding the lock. If we won, we complete 218 * init and return our conn. If we lost, we rollback and return the 219 * other one. 220 */ 221 spin_lock_irqsave(&rds_conn_lock, flags); 222 if (parent) { 223 /* Creating passive conn */ 224 if (parent->c_passive) { 225 trans->conn_free(conn->c_transport_data); 226 kmem_cache_free(rds_conn_slab, conn); 227 conn = parent->c_passive; 228 } else { 229 parent->c_passive = conn; 230 rds_cong_add_conn(conn); 231 rds_conn_count++; 232 } 233 } else { 234 /* Creating normal conn */ 235 struct rds_connection *found; 236 237 if (!is_outgoing && otrans->t_type == RDS_TRANS_TCP) 238 found = NULL; 239 else 240 found = rds_conn_lookup(head, laddr, faddr, trans); 241 if (found) { 242 trans->conn_free(conn->c_transport_data); 243 kmem_cache_free(rds_conn_slab, conn); 244 conn = found; 245 } else { 246 if ((is_outgoing && otrans->t_type == RDS_TRANS_TCP) || 247 (otrans->t_type != RDS_TRANS_TCP)) { 248 /* Only the active side should be added to 249 * reconnect list for TCP. 250 */ 251 hlist_add_head_rcu(&conn->c_hash_node, head); 252 } 253 rds_cong_add_conn(conn); 254 rds_conn_count++; 255 } 256 } 257 spin_unlock_irqrestore(&rds_conn_lock, flags); 258 259out: 260 return conn; 261} 262 263struct rds_connection *rds_conn_create(__be32 laddr, __be32 faddr, 264 struct rds_transport *trans, gfp_t gfp) 265{ 266 return __rds_conn_create(laddr, faddr, trans, gfp, 0); 267} 268EXPORT_SYMBOL_GPL(rds_conn_create); 269 270struct rds_connection *rds_conn_create_outgoing(__be32 laddr, __be32 faddr, 271 struct rds_transport *trans, gfp_t gfp) 272{ 273 return __rds_conn_create(laddr, faddr, trans, gfp, 1); 274} 275EXPORT_SYMBOL_GPL(rds_conn_create_outgoing); 276 277void rds_conn_shutdown(struct rds_connection *conn) 278{ 279 /* shut it down unless it's down already */ 280 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_DOWN)) { 281 /* 282 * Quiesce the connection mgmt handlers before we start tearing 283 * things down. We don't hold the mutex for the entire 284 * duration of the shutdown operation, else we may be 285 * deadlocking with the CM handler. Instead, the CM event 286 * handler is supposed to check for state DISCONNECTING 287 */ 288 mutex_lock(&conn->c_cm_lock); 289 if (!rds_conn_transition(conn, RDS_CONN_UP, RDS_CONN_DISCONNECTING) 290 && !rds_conn_transition(conn, RDS_CONN_ERROR, RDS_CONN_DISCONNECTING)) { 291 rds_conn_error(conn, "shutdown called in state %d\n", 292 atomic_read(&conn->c_state)); 293 mutex_unlock(&conn->c_cm_lock); 294 return; 295 } 296 mutex_unlock(&conn->c_cm_lock); 297 298 wait_event(conn->c_waitq, 299 !test_bit(RDS_IN_XMIT, &conn->c_flags)); 300 301 conn->c_trans->conn_shutdown(conn); 302 rds_conn_reset(conn); 303 304 if (!rds_conn_transition(conn, RDS_CONN_DISCONNECTING, RDS_CONN_DOWN)) { 305 /* This can happen - eg when we're in the middle of tearing 306 * down the connection, and someone unloads the rds module. 307 * Quite reproduceable with loopback connections. 308 * Mostly harmless. 309 */ 310 rds_conn_error(conn, 311 "%s: failed to transition to state DOWN, " 312 "current state is %d\n", 313 __func__, 314 atomic_read(&conn->c_state)); 315 return; 316 } 317 } 318 319 /* Then reconnect if it's still live. 320 * The passive side of an IB loopback connection is never added 321 * to the conn hash, so we never trigger a reconnect on this 322 * conn - the reconnect is always triggered by the active peer. */ 323 cancel_delayed_work_sync(&conn->c_conn_w); 324 rcu_read_lock(); 325 if (!hlist_unhashed(&conn->c_hash_node)) { 326 rcu_read_unlock(); 327 rds_queue_reconnect(conn); 328 } else { 329 rcu_read_unlock(); 330 } 331} 332 333/* 334 * Stop and free a connection. 335 * 336 * This can only be used in very limited circumstances. It assumes that once 337 * the conn has been shutdown that no one else is referencing the connection. 338 * We can only ensure this in the rmmod path in the current code. 339 */ 340void rds_conn_destroy(struct rds_connection *conn) 341{ 342 struct rds_message *rm, *rtmp; 343 unsigned long flags; 344 345 rdsdebug("freeing conn %p for %pI4 -> " 346 "%pI4\n", conn, &conn->c_laddr, 347 &conn->c_faddr); 348 349 /* Ensure conn will not be scheduled for reconnect */ 350 spin_lock_irq(&rds_conn_lock); 351 hlist_del_init_rcu(&conn->c_hash_node); 352 spin_unlock_irq(&rds_conn_lock); 353 synchronize_rcu(); 354 355 /* shut the connection down */ 356 rds_conn_drop(conn); 357 flush_work(&conn->c_down_w); 358 359 /* make sure lingering queued work won't try to ref the conn */ 360 cancel_delayed_work_sync(&conn->c_send_w); 361 cancel_delayed_work_sync(&conn->c_recv_w); 362 363 /* tear down queued messages */ 364 list_for_each_entry_safe(rm, rtmp, 365 &conn->c_send_queue, 366 m_conn_item) { 367 list_del_init(&rm->m_conn_item); 368 BUG_ON(!list_empty(&rm->m_sock_item)); 369 rds_message_put(rm); 370 } 371 if (conn->c_xmit_rm) 372 rds_message_put(conn->c_xmit_rm); 373 374 conn->c_trans->conn_free(conn->c_transport_data); 375 376 /* 377 * The congestion maps aren't freed up here. They're 378 * freed by rds_cong_exit() after all the connections 379 * have been freed. 380 */ 381 rds_cong_remove_conn(conn); 382 383 BUG_ON(!list_empty(&conn->c_retrans)); 384 kmem_cache_free(rds_conn_slab, conn); 385 386 spin_lock_irqsave(&rds_conn_lock, flags); 387 rds_conn_count--; 388 spin_unlock_irqrestore(&rds_conn_lock, flags); 389} 390EXPORT_SYMBOL_GPL(rds_conn_destroy); 391 392static void rds_conn_message_info(struct socket *sock, unsigned int len, 393 struct rds_info_iterator *iter, 394 struct rds_info_lengths *lens, 395 int want_send) 396{ 397 struct hlist_head *head; 398 struct list_head *list; 399 struct rds_connection *conn; 400 struct rds_message *rm; 401 unsigned int total = 0; 402 unsigned long flags; 403 size_t i; 404 405 len /= sizeof(struct rds_info_message); 406 407 rcu_read_lock(); 408 409 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); 410 i++, head++) { 411 hlist_for_each_entry_rcu(conn, head, c_hash_node) { 412 if (want_send) 413 list = &conn->c_send_queue; 414 else 415 list = &conn->c_retrans; 416 417 spin_lock_irqsave(&conn->c_lock, flags); 418 419 /* XXX too lazy to maintain counts.. */ 420 list_for_each_entry(rm, list, m_conn_item) { 421 total++; 422 if (total <= len) 423 rds_inc_info_copy(&rm->m_inc, iter, 424 conn->c_laddr, 425 conn->c_faddr, 0); 426 } 427 428 spin_unlock_irqrestore(&conn->c_lock, flags); 429 } 430 } 431 rcu_read_unlock(); 432 433 lens->nr = total; 434 lens->each = sizeof(struct rds_info_message); 435} 436 437static void rds_conn_message_info_send(struct socket *sock, unsigned int len, 438 struct rds_info_iterator *iter, 439 struct rds_info_lengths *lens) 440{ 441 rds_conn_message_info(sock, len, iter, lens, 1); 442} 443 444static void rds_conn_message_info_retrans(struct socket *sock, 445 unsigned int len, 446 struct rds_info_iterator *iter, 447 struct rds_info_lengths *lens) 448{ 449 rds_conn_message_info(sock, len, iter, lens, 0); 450} 451 452void rds_for_each_conn_info(struct socket *sock, unsigned int len, 453 struct rds_info_iterator *iter, 454 struct rds_info_lengths *lens, 455 int (*visitor)(struct rds_connection *, void *), 456 size_t item_len) 457{ 458 uint64_t buffer[(item_len + 7) / 8]; 459 struct hlist_head *head; 460 struct rds_connection *conn; 461 size_t i; 462 463 rcu_read_lock(); 464 465 lens->nr = 0; 466 lens->each = item_len; 467 468 for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); 469 i++, head++) { 470 hlist_for_each_entry_rcu(conn, head, c_hash_node) { 471 472 /* XXX no c_lock usage.. */ 473 if (!visitor(conn, buffer)) 474 continue; 475 476 /* We copy as much as we can fit in the buffer, 477 * but we count all items so that the caller 478 * can resize the buffer. */ 479 if (len >= item_len) { 480 rds_info_copy(iter, buffer, item_len); 481 len -= item_len; 482 } 483 lens->nr++; 484 } 485 } 486 rcu_read_unlock(); 487} 488EXPORT_SYMBOL_GPL(rds_for_each_conn_info); 489 490static int rds_conn_info_visitor(struct rds_connection *conn, 491 void *buffer) 492{ 493 struct rds_info_connection *cinfo = buffer; 494 495 cinfo->next_tx_seq = conn->c_next_tx_seq; 496 cinfo->next_rx_seq = conn->c_next_rx_seq; 497 cinfo->laddr = conn->c_laddr; 498 cinfo->faddr = conn->c_faddr; 499 strncpy(cinfo->transport, conn->c_trans->t_name, 500 sizeof(cinfo->transport)); 501 cinfo->flags = 0; 502 503 rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &conn->c_flags), 504 SENDING); 505 /* XXX Future: return the state rather than these funky bits */ 506 rds_conn_info_set(cinfo->flags, 507 atomic_read(&conn->c_state) == RDS_CONN_CONNECTING, 508 CONNECTING); 509 rds_conn_info_set(cinfo->flags, 510 atomic_read(&conn->c_state) == RDS_CONN_UP, 511 CONNECTED); 512 return 1; 513} 514 515static void rds_conn_info(struct socket *sock, unsigned int len, 516 struct rds_info_iterator *iter, 517 struct rds_info_lengths *lens) 518{ 519 rds_for_each_conn_info(sock, len, iter, lens, 520 rds_conn_info_visitor, 521 sizeof(struct rds_info_connection)); 522} 523 524int rds_conn_init(void) 525{ 526 rds_conn_slab = kmem_cache_create("rds_connection", 527 sizeof(struct rds_connection), 528 0, 0, NULL); 529 if (!rds_conn_slab) 530 return -ENOMEM; 531 532 rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info); 533 rds_info_register_func(RDS_INFO_SEND_MESSAGES, 534 rds_conn_message_info_send); 535 rds_info_register_func(RDS_INFO_RETRANS_MESSAGES, 536 rds_conn_message_info_retrans); 537 538 return 0; 539} 540 541void rds_conn_exit(void) 542{ 543 rds_loop_exit(); 544 545 WARN_ON(!hlist_empty(rds_conn_hash)); 546 547 kmem_cache_destroy(rds_conn_slab); 548 549 rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info); 550 rds_info_deregister_func(RDS_INFO_SEND_MESSAGES, 551 rds_conn_message_info_send); 552 rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES, 553 rds_conn_message_info_retrans); 554} 555 556/* 557 * Force a disconnect 558 */ 559void rds_conn_drop(struct rds_connection *conn) 560{ 561 atomic_set(&conn->c_state, RDS_CONN_ERROR); 562 queue_work(rds_wq, &conn->c_down_w); 563} 564EXPORT_SYMBOL_GPL(rds_conn_drop); 565 566/* 567 * If the connection is down, trigger a connect. We may have scheduled a 568 * delayed reconnect however - in this case we should not interfere. 569 */ 570void rds_conn_connect_if_down(struct rds_connection *conn) 571{ 572 if (rds_conn_state(conn) == RDS_CONN_DOWN && 573 !test_and_set_bit(RDS_RECONNECT_PENDING, &conn->c_flags)) 574 queue_delayed_work(rds_wq, &conn->c_conn_w, 0); 575} 576EXPORT_SYMBOL_GPL(rds_conn_connect_if_down); 577 578/* 579 * An error occurred on the connection 580 */ 581void 582__rds_conn_error(struct rds_connection *conn, const char *fmt, ...) 583{ 584 va_list ap; 585 586 va_start(ap, fmt); 587 vprintk(fmt, ap); 588 va_end(ap); 589 590 rds_conn_drop(conn); 591} 592