1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB 2 /* 3 * Copyright (c) 2004-2007 Intel Corporation. All rights reserved. 4 * Copyright (c) 2004 Topspin Corporation. All rights reserved. 5 * Copyright (c) 2004, 2005 Voltaire Corporation. All rights reserved. 6 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 7 * Copyright (c) 2019, Mellanox Technologies inc. All rights reserved. 8 */ 9 10 #include <linux/completion.h> 11 #include <linux/dma-mapping.h> 12 #include <linux/device.h> 13 #include <linux/module.h> 14 #include <linux/err.h> 15 #include <linux/idr.h> 16 #include <linux/interrupt.h> 17 #include <linux/random.h> 18 #include <linux/rbtree.h> 19 #include <linux/spinlock.h> 20 #include <linux/slab.h> 21 #include <linux/sysfs.h> 22 #include <linux/workqueue.h> 23 #include <linux/kdev_t.h> 24 #include <linux/etherdevice.h> 25 26 #include <rdma/ib_cache.h> 27 #include <rdma/ib_cm.h> 28 #include <rdma/ib_sysfs.h> 29 #include "cm_msgs.h" 30 #include "core_priv.h" 31 #include "cm_trace.h" 32 33 MODULE_AUTHOR("Sean Hefty"); 34 MODULE_DESCRIPTION("InfiniBand CM"); 35 MODULE_LICENSE("Dual BSD/GPL"); 36 37 #define CM_DESTROY_ID_WAIT_TIMEOUT 10000 /* msecs */ 38 #define CM_DIRECT_RETRY_CTX ((void *) 1UL) 39 40 static const char * const ibcm_rej_reason_strs[] = { 41 [IB_CM_REJ_NO_QP] = "no QP", 42 [IB_CM_REJ_NO_EEC] = "no EEC", 43 [IB_CM_REJ_NO_RESOURCES] = "no resources", 44 [IB_CM_REJ_TIMEOUT] = "timeout", 45 [IB_CM_REJ_UNSUPPORTED] = "unsupported", 46 [IB_CM_REJ_INVALID_COMM_ID] = "invalid comm ID", 47 [IB_CM_REJ_INVALID_COMM_INSTANCE] = "invalid comm instance", 48 [IB_CM_REJ_INVALID_SERVICE_ID] = "invalid service ID", 49 [IB_CM_REJ_INVALID_TRANSPORT_TYPE] = "invalid transport type", 50 [IB_CM_REJ_STALE_CONN] = "stale conn", 51 [IB_CM_REJ_RDC_NOT_EXIST] = "RDC not exist", 52 [IB_CM_REJ_INVALID_GID] = "invalid GID", 53 [IB_CM_REJ_INVALID_LID] = "invalid LID", 54 [IB_CM_REJ_INVALID_SL] = "invalid SL", 55 [IB_CM_REJ_INVALID_TRAFFIC_CLASS] = "invalid traffic class", 56 [IB_CM_REJ_INVALID_HOP_LIMIT] = "invalid hop limit", 57 [IB_CM_REJ_INVALID_PACKET_RATE] = "invalid packet rate", 58 [IB_CM_REJ_INVALID_ALT_GID] = "invalid alt GID", 59 [IB_CM_REJ_INVALID_ALT_LID] = "invalid alt LID", 60 [IB_CM_REJ_INVALID_ALT_SL] = "invalid alt SL", 61 [IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS] = "invalid alt traffic class", 62 [IB_CM_REJ_INVALID_ALT_HOP_LIMIT] = "invalid alt hop limit", 63 [IB_CM_REJ_INVALID_ALT_PACKET_RATE] = "invalid alt packet rate", 64 [IB_CM_REJ_PORT_CM_REDIRECT] = "port CM redirect", 65 [IB_CM_REJ_PORT_REDIRECT] = "port redirect", 66 [IB_CM_REJ_INVALID_MTU] = "invalid MTU", 67 [IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES] = "insufficient resp resources", 68 [IB_CM_REJ_CONSUMER_DEFINED] = "consumer defined", 69 [IB_CM_REJ_INVALID_RNR_RETRY] = "invalid RNR retry", 70 [IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID] = "duplicate local comm ID", 71 [IB_CM_REJ_INVALID_CLASS_VERSION] = "invalid class version", 72 [IB_CM_REJ_INVALID_FLOW_LABEL] = "invalid flow label", 73 [IB_CM_REJ_INVALID_ALT_FLOW_LABEL] = "invalid alt flow label", 74 [IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED] = 75 "vendor option is not supported", 76 }; 77 78 const char *__attribute_const__ ibcm_reject_msg(int reason) 79 { 80 size_t index = reason; 81 82 if (index < ARRAY_SIZE(ibcm_rej_reason_strs) && 83 ibcm_rej_reason_strs[index]) 84 return ibcm_rej_reason_strs[index]; 85 else 86 return "unrecognized reason"; 87 } 88 EXPORT_SYMBOL(ibcm_reject_msg); 89 90 struct cm_id_private; 91 struct cm_work; 92 static int cm_add_one(struct ib_device *device); 93 static void cm_remove_one(struct ib_device *device, void *client_data); 94 static void cm_process_work(struct cm_id_private *cm_id_priv, 95 struct cm_work *work); 96 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv, 97 struct ib_cm_sidr_rep_param *param); 98 static void cm_issue_dreq(struct cm_id_private *cm_id_priv); 99 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv, 100 void *private_data, u8 private_data_len); 101 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv, 102 enum ib_cm_rej_reason reason, void *ari, 103 u8 ari_length, const void *private_data, 104 u8 private_data_len); 105 106 static struct ib_client cm_client = { 107 .name = "cm", 108 .add = cm_add_one, 109 .remove = cm_remove_one 110 }; 111 112 static struct ib_cm { 113 spinlock_t lock; 114 struct list_head device_list; 115 rwlock_t device_lock; 116 struct rb_root listen_service_table; 117 u64 listen_service_id; 118 /* struct rb_root peer_service_table; todo: fix peer to peer */ 119 struct rb_root remote_qp_table; 120 struct rb_root remote_id_table; 121 struct rb_root remote_sidr_table; 122 struct xarray local_id_table; 123 u32 local_id_next; 124 __be32 random_id_operand; 125 struct list_head timewait_list; 126 struct workqueue_struct *wq; 127 } cm; 128 129 /* Counter indexes ordered by attribute ID */ 130 enum { 131 CM_REQ_COUNTER, 132 CM_MRA_COUNTER, 133 CM_REJ_COUNTER, 134 CM_REP_COUNTER, 135 CM_RTU_COUNTER, 136 CM_DREQ_COUNTER, 137 CM_DREP_COUNTER, 138 CM_SIDR_REQ_COUNTER, 139 CM_SIDR_REP_COUNTER, 140 CM_LAP_COUNTER, 141 CM_APR_COUNTER, 142 CM_ATTR_COUNT, 143 CM_ATTR_ID_OFFSET = 0x0010, 144 }; 145 146 enum { 147 CM_XMIT, 148 CM_XMIT_RETRIES, 149 CM_RECV, 150 CM_RECV_DUPLICATES, 151 CM_COUNTER_GROUPS 152 }; 153 154 struct cm_counter_attribute { 155 struct ib_port_attribute attr; 156 unsigned short group; 157 unsigned short index; 158 }; 159 160 struct cm_port { 161 struct cm_device *cm_dev; 162 struct ib_mad_agent *mad_agent; 163 u32 port_num; 164 atomic_long_t counters[CM_COUNTER_GROUPS][CM_ATTR_COUNT]; 165 }; 166 167 struct cm_device { 168 struct kref kref; 169 struct list_head list; 170 spinlock_t mad_agent_lock; 171 struct ib_device *ib_device; 172 u8 ack_delay; 173 int going_down; 174 struct cm_port *port[]; 175 }; 176 177 struct cm_av { 178 struct cm_port *port; 179 struct rdma_ah_attr ah_attr; 180 u16 dlid_datapath; 181 u16 pkey_index; 182 u8 timeout; 183 }; 184 185 struct cm_work { 186 struct delayed_work work; 187 struct list_head list; 188 struct cm_port *port; 189 struct ib_mad_recv_wc *mad_recv_wc; /* Received MADs */ 190 __be32 local_id; /* Established / timewait */ 191 __be32 remote_id; 192 struct ib_cm_event cm_event; 193 struct sa_path_rec path[]; 194 }; 195 196 struct cm_timewait_info { 197 struct cm_work work; 198 struct list_head list; 199 struct rb_node remote_qp_node; 200 struct rb_node remote_id_node; 201 __be64 remote_ca_guid; 202 __be32 remote_qpn; 203 u8 inserted_remote_qp; 204 u8 inserted_remote_id; 205 }; 206 207 struct cm_id_private { 208 struct ib_cm_id id; 209 210 struct rb_node service_node; 211 struct rb_node sidr_id_node; 212 u32 sidr_slid; 213 spinlock_t lock; /* Do not acquire inside cm.lock */ 214 struct completion comp; 215 refcount_t refcount; 216 /* Number of clients sharing this ib_cm_id. Only valid for listeners. 217 * Protected by the cm.lock spinlock. 218 */ 219 int listen_sharecount; 220 struct rcu_head rcu; 221 222 struct ib_mad_send_buf *msg; 223 struct cm_timewait_info *timewait_info; 224 /* todo: use alternate port on send failure */ 225 struct cm_av av; 226 struct cm_av alt_av; 227 228 void *private_data; 229 __be64 tid; 230 __be32 local_qpn; 231 __be32 remote_qpn; 232 enum ib_qp_type qp_type; 233 __be32 sq_psn; 234 __be32 rq_psn; 235 int timeout_ms; 236 enum ib_mtu path_mtu; 237 __be16 pkey; 238 u8 private_data_len; 239 u8 max_cm_retries; 240 u8 responder_resources; 241 u8 initiator_depth; 242 u8 retry_count; 243 u8 rnr_retry_count; 244 u8 service_timeout; 245 u8 target_ack_delay; 246 247 struct list_head work_list; 248 atomic_t work_count; 249 250 struct rdma_ucm_ece ece; 251 }; 252 253 static void cm_dev_release(struct kref *kref) 254 { 255 struct cm_device *cm_dev = container_of(kref, struct cm_device, kref); 256 u32 i; 257 258 rdma_for_each_port(cm_dev->ib_device, i) 259 kfree(cm_dev->port[i - 1]); 260 261 kfree(cm_dev); 262 } 263 264 static void cm_device_put(struct cm_device *cm_dev) 265 { 266 kref_put(&cm_dev->kref, cm_dev_release); 267 } 268 269 static void cm_work_handler(struct work_struct *work); 270 271 static inline void cm_deref_id(struct cm_id_private *cm_id_priv) 272 { 273 if (refcount_dec_and_test(&cm_id_priv->refcount)) 274 complete(&cm_id_priv->comp); 275 } 276 277 static struct ib_mad_send_buf *cm_alloc_msg(struct cm_id_private *cm_id_priv) 278 { 279 struct ib_mad_agent *mad_agent; 280 struct ib_mad_send_buf *m; 281 struct ib_ah *ah; 282 283 lockdep_assert_held(&cm_id_priv->lock); 284 285 if (!cm_id_priv->av.port) 286 return ERR_PTR(-EINVAL); 287 288 spin_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 289 mad_agent = cm_id_priv->av.port->mad_agent; 290 if (!mad_agent) { 291 m = ERR_PTR(-EINVAL); 292 goto out; 293 } 294 295 ah = rdma_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr, 0); 296 if (IS_ERR(ah)) { 297 m = ERR_CAST(ah); 298 goto out; 299 } 300 301 m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn, 302 cm_id_priv->av.pkey_index, 303 0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA, 304 GFP_ATOMIC, 305 IB_MGMT_BASE_VERSION); 306 if (IS_ERR(m)) { 307 rdma_destroy_ah(ah, 0); 308 goto out; 309 } 310 311 m->ah = ah; 312 313 out: 314 spin_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 315 return m; 316 } 317 318 static void cm_free_msg(struct ib_mad_send_buf *msg) 319 { 320 if (msg->ah) 321 rdma_destroy_ah(msg->ah, 0); 322 ib_free_send_mad(msg); 323 } 324 325 static struct ib_mad_send_buf * 326 cm_alloc_priv_msg(struct cm_id_private *cm_id_priv, enum ib_cm_state state) 327 { 328 struct ib_mad_send_buf *msg; 329 330 lockdep_assert_held(&cm_id_priv->lock); 331 332 msg = cm_alloc_msg(cm_id_priv); 333 if (IS_ERR(msg)) 334 return msg; 335 336 cm_id_priv->msg = msg; 337 refcount_inc(&cm_id_priv->refcount); 338 msg->context[0] = cm_id_priv; 339 msg->context[1] = (void *) (unsigned long) state; 340 341 msg->retries = cm_id_priv->max_cm_retries; 342 msg->timeout_ms = cm_id_priv->timeout_ms; 343 344 return msg; 345 } 346 347 static void cm_free_priv_msg(struct ib_mad_send_buf *msg) 348 { 349 struct cm_id_private *cm_id_priv = msg->context[0]; 350 351 lockdep_assert_held(&cm_id_priv->lock); 352 353 if (!WARN_ON(cm_id_priv->msg != msg)) 354 cm_id_priv->msg = NULL; 355 356 if (msg->ah) 357 rdma_destroy_ah(msg->ah, 0); 358 cm_deref_id(cm_id_priv); 359 ib_free_send_mad(msg); 360 } 361 362 static struct ib_mad_send_buf * 363 cm_alloc_response_msg_no_ah(struct cm_port *port, 364 struct ib_mad_recv_wc *mad_recv_wc, 365 bool direct_retry) 366 { 367 struct ib_mad_send_buf *m; 368 369 m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index, 370 0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA, 371 GFP_ATOMIC, IB_MGMT_BASE_VERSION); 372 if (!IS_ERR(m)) 373 m->context[0] = direct_retry ? CM_DIRECT_RETRY_CTX : NULL; 374 375 return m; 376 } 377 378 static int cm_create_response_msg_ah(struct cm_port *port, 379 struct ib_mad_recv_wc *mad_recv_wc, 380 struct ib_mad_send_buf *msg) 381 { 382 struct ib_ah *ah; 383 384 ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc, 385 mad_recv_wc->recv_buf.grh, port->port_num); 386 if (IS_ERR(ah)) 387 return PTR_ERR(ah); 388 389 msg->ah = ah; 390 return 0; 391 } 392 393 static int cm_alloc_response_msg(struct cm_port *port, 394 struct ib_mad_recv_wc *mad_recv_wc, 395 bool direct_retry, 396 struct ib_mad_send_buf **msg) 397 { 398 struct ib_mad_send_buf *m; 399 int ret; 400 401 m = cm_alloc_response_msg_no_ah(port, mad_recv_wc, direct_retry); 402 if (IS_ERR(m)) 403 return PTR_ERR(m); 404 405 ret = cm_create_response_msg_ah(port, mad_recv_wc, m); 406 if (ret) { 407 ib_free_send_mad(m); 408 return ret; 409 } 410 411 *msg = m; 412 return 0; 413 } 414 415 static void *cm_copy_private_data(const void *private_data, u8 private_data_len) 416 { 417 void *data; 418 419 if (!private_data || !private_data_len) 420 return NULL; 421 422 data = kmemdup(private_data, private_data_len, GFP_KERNEL); 423 if (!data) 424 return ERR_PTR(-ENOMEM); 425 426 return data; 427 } 428 429 static void cm_set_private_data(struct cm_id_private *cm_id_priv, 430 void *private_data, u8 private_data_len) 431 { 432 if (cm_id_priv->private_data && cm_id_priv->private_data_len) 433 kfree(cm_id_priv->private_data); 434 435 cm_id_priv->private_data = private_data; 436 cm_id_priv->private_data_len = private_data_len; 437 } 438 439 static void cm_set_av_port(struct cm_av *av, struct cm_port *port) 440 { 441 struct cm_port *old_port = av->port; 442 443 if (old_port == port) 444 return; 445 446 av->port = port; 447 if (old_port) 448 cm_device_put(old_port->cm_dev); 449 if (port) 450 kref_get(&port->cm_dev->kref); 451 } 452 453 static void cm_init_av_for_lap(struct cm_port *port, struct ib_wc *wc, 454 struct rdma_ah_attr *ah_attr, struct cm_av *av) 455 { 456 cm_set_av_port(av, port); 457 av->pkey_index = wc->pkey_index; 458 rdma_move_ah_attr(&av->ah_attr, ah_attr); 459 } 460 461 static int cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc, 462 struct ib_grh *grh, struct cm_av *av) 463 { 464 cm_set_av_port(av, port); 465 av->pkey_index = wc->pkey_index; 466 return ib_init_ah_attr_from_wc(port->cm_dev->ib_device, 467 port->port_num, wc, 468 grh, &av->ah_attr); 469 } 470 471 static struct cm_port * 472 get_cm_port_from_path(struct sa_path_rec *path, const struct ib_gid_attr *attr) 473 { 474 struct cm_device *cm_dev; 475 struct cm_port *port = NULL; 476 unsigned long flags; 477 478 if (attr) { 479 read_lock_irqsave(&cm.device_lock, flags); 480 list_for_each_entry(cm_dev, &cm.device_list, list) { 481 if (cm_dev->ib_device == attr->device) { 482 port = cm_dev->port[attr->port_num - 1]; 483 break; 484 } 485 } 486 read_unlock_irqrestore(&cm.device_lock, flags); 487 } else { 488 /* SGID attribute can be NULL in following 489 * conditions. 490 * (a) Alternative path 491 * (b) IB link layer without GRH 492 * (c) LAP send messages 493 */ 494 read_lock_irqsave(&cm.device_lock, flags); 495 list_for_each_entry(cm_dev, &cm.device_list, list) { 496 attr = rdma_find_gid(cm_dev->ib_device, 497 &path->sgid, 498 sa_conv_pathrec_to_gid_type(path), 499 NULL); 500 if (!IS_ERR(attr)) { 501 port = cm_dev->port[attr->port_num - 1]; 502 break; 503 } 504 } 505 read_unlock_irqrestore(&cm.device_lock, flags); 506 if (port) 507 rdma_put_gid_attr(attr); 508 } 509 return port; 510 } 511 512 static int cm_init_av_by_path(struct sa_path_rec *path, 513 const struct ib_gid_attr *sgid_attr, 514 struct cm_av *av) 515 { 516 struct rdma_ah_attr new_ah_attr; 517 struct cm_device *cm_dev; 518 struct cm_port *port; 519 int ret; 520 521 port = get_cm_port_from_path(path, sgid_attr); 522 if (!port) 523 return -EINVAL; 524 cm_dev = port->cm_dev; 525 526 ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num, 527 be16_to_cpu(path->pkey), &av->pkey_index); 528 if (ret) 529 return ret; 530 531 cm_set_av_port(av, port); 532 533 /* 534 * av->ah_attr might be initialized based on wc or during 535 * request processing time which might have reference to sgid_attr. 536 * So initialize a new ah_attr on stack. 537 * If initialization fails, old ah_attr is used for sending any 538 * responses. If initialization is successful, than new ah_attr 539 * is used by overwriting the old one. So that right ah_attr 540 * can be used to return an error response. 541 */ 542 ret = ib_init_ah_attr_from_path(cm_dev->ib_device, port->port_num, path, 543 &new_ah_attr, sgid_attr); 544 if (ret) 545 return ret; 546 547 av->timeout = path->packet_life_time + 1; 548 rdma_move_ah_attr(&av->ah_attr, &new_ah_attr); 549 return 0; 550 } 551 552 /* Move av created by cm_init_av_by_path(), so av.dgid is not moved */ 553 static void cm_move_av_from_path(struct cm_av *dest, struct cm_av *src) 554 { 555 cm_set_av_port(dest, src->port); 556 cm_set_av_port(src, NULL); 557 dest->pkey_index = src->pkey_index; 558 rdma_move_ah_attr(&dest->ah_attr, &src->ah_attr); 559 dest->timeout = src->timeout; 560 } 561 562 static void cm_destroy_av(struct cm_av *av) 563 { 564 rdma_destroy_ah_attr(&av->ah_attr); 565 cm_set_av_port(av, NULL); 566 } 567 568 static u32 cm_local_id(__be32 local_id) 569 { 570 return (__force u32) (local_id ^ cm.random_id_operand); 571 } 572 573 static struct cm_id_private *cm_acquire_id(__be32 local_id, __be32 remote_id) 574 { 575 struct cm_id_private *cm_id_priv; 576 577 rcu_read_lock(); 578 cm_id_priv = xa_load(&cm.local_id_table, cm_local_id(local_id)); 579 if (!cm_id_priv || cm_id_priv->id.remote_id != remote_id || 580 !refcount_inc_not_zero(&cm_id_priv->refcount)) 581 cm_id_priv = NULL; 582 rcu_read_unlock(); 583 584 return cm_id_priv; 585 } 586 587 /* 588 * Trivial helpers to strip endian annotation and compare; the 589 * endianness doesn't actually matter since we just need a stable 590 * order for the RB tree. 591 */ 592 static int be32_lt(__be32 a, __be32 b) 593 { 594 return (__force u32) a < (__force u32) b; 595 } 596 597 static int be32_gt(__be32 a, __be32 b) 598 { 599 return (__force u32) a > (__force u32) b; 600 } 601 602 static int be64_lt(__be64 a, __be64 b) 603 { 604 return (__force u64) a < (__force u64) b; 605 } 606 607 static int be64_gt(__be64 a, __be64 b) 608 { 609 return (__force u64) a > (__force u64) b; 610 } 611 612 /* 613 * Inserts a new cm_id_priv into the listen_service_table. Returns cm_id_priv 614 * if the new ID was inserted, NULL if it could not be inserted due to a 615 * collision, or the existing cm_id_priv ready for shared usage. 616 */ 617 static struct cm_id_private *cm_insert_listen(struct cm_id_private *cm_id_priv, 618 ib_cm_handler shared_handler) 619 { 620 struct rb_node **link = &cm.listen_service_table.rb_node; 621 struct rb_node *parent = NULL; 622 struct cm_id_private *cur_cm_id_priv; 623 __be64 service_id = cm_id_priv->id.service_id; 624 unsigned long flags; 625 626 spin_lock_irqsave(&cm.lock, flags); 627 while (*link) { 628 parent = *link; 629 cur_cm_id_priv = rb_entry(parent, struct cm_id_private, 630 service_node); 631 632 if (cm_id_priv->id.device < cur_cm_id_priv->id.device) 633 link = &(*link)->rb_left; 634 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device) 635 link = &(*link)->rb_right; 636 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id)) 637 link = &(*link)->rb_left; 638 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id)) 639 link = &(*link)->rb_right; 640 else { 641 /* 642 * Sharing an ib_cm_id with different handlers is not 643 * supported 644 */ 645 if (cur_cm_id_priv->id.cm_handler != shared_handler || 646 cur_cm_id_priv->id.context || 647 WARN_ON(!cur_cm_id_priv->id.cm_handler)) { 648 spin_unlock_irqrestore(&cm.lock, flags); 649 return NULL; 650 } 651 refcount_inc(&cur_cm_id_priv->refcount); 652 cur_cm_id_priv->listen_sharecount++; 653 spin_unlock_irqrestore(&cm.lock, flags); 654 return cur_cm_id_priv; 655 } 656 } 657 cm_id_priv->listen_sharecount++; 658 rb_link_node(&cm_id_priv->service_node, parent, link); 659 rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table); 660 spin_unlock_irqrestore(&cm.lock, flags); 661 return cm_id_priv; 662 } 663 664 static struct cm_id_private *cm_find_listen(struct ib_device *device, 665 __be64 service_id) 666 { 667 struct rb_node *node = cm.listen_service_table.rb_node; 668 struct cm_id_private *cm_id_priv; 669 670 while (node) { 671 cm_id_priv = rb_entry(node, struct cm_id_private, service_node); 672 673 if (device < cm_id_priv->id.device) 674 node = node->rb_left; 675 else if (device > cm_id_priv->id.device) 676 node = node->rb_right; 677 else if (be64_lt(service_id, cm_id_priv->id.service_id)) 678 node = node->rb_left; 679 else if (be64_gt(service_id, cm_id_priv->id.service_id)) 680 node = node->rb_right; 681 else { 682 refcount_inc(&cm_id_priv->refcount); 683 return cm_id_priv; 684 } 685 } 686 return NULL; 687 } 688 689 static struct cm_timewait_info * 690 cm_insert_remote_id(struct cm_timewait_info *timewait_info) 691 { 692 struct rb_node **link = &cm.remote_id_table.rb_node; 693 struct rb_node *parent = NULL; 694 struct cm_timewait_info *cur_timewait_info; 695 __be64 remote_ca_guid = timewait_info->remote_ca_guid; 696 __be32 remote_id = timewait_info->work.remote_id; 697 698 while (*link) { 699 parent = *link; 700 cur_timewait_info = rb_entry(parent, struct cm_timewait_info, 701 remote_id_node); 702 if (be32_lt(remote_id, cur_timewait_info->work.remote_id)) 703 link = &(*link)->rb_left; 704 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id)) 705 link = &(*link)->rb_right; 706 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 707 link = &(*link)->rb_left; 708 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 709 link = &(*link)->rb_right; 710 else 711 return cur_timewait_info; 712 } 713 timewait_info->inserted_remote_id = 1; 714 rb_link_node(&timewait_info->remote_id_node, parent, link); 715 rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table); 716 return NULL; 717 } 718 719 static struct cm_id_private *cm_find_remote_id(__be64 remote_ca_guid, 720 __be32 remote_id) 721 { 722 struct rb_node *node = cm.remote_id_table.rb_node; 723 struct cm_timewait_info *timewait_info; 724 struct cm_id_private *res = NULL; 725 726 spin_lock_irq(&cm.lock); 727 while (node) { 728 timewait_info = rb_entry(node, struct cm_timewait_info, 729 remote_id_node); 730 if (be32_lt(remote_id, timewait_info->work.remote_id)) 731 node = node->rb_left; 732 else if (be32_gt(remote_id, timewait_info->work.remote_id)) 733 node = node->rb_right; 734 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid)) 735 node = node->rb_left; 736 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid)) 737 node = node->rb_right; 738 else { 739 res = cm_acquire_id(timewait_info->work.local_id, 740 timewait_info->work.remote_id); 741 break; 742 } 743 } 744 spin_unlock_irq(&cm.lock); 745 return res; 746 } 747 748 static struct cm_timewait_info * 749 cm_insert_remote_qpn(struct cm_timewait_info *timewait_info) 750 { 751 struct rb_node **link = &cm.remote_qp_table.rb_node; 752 struct rb_node *parent = NULL; 753 struct cm_timewait_info *cur_timewait_info; 754 __be64 remote_ca_guid = timewait_info->remote_ca_guid; 755 __be32 remote_qpn = timewait_info->remote_qpn; 756 757 while (*link) { 758 parent = *link; 759 cur_timewait_info = rb_entry(parent, struct cm_timewait_info, 760 remote_qp_node); 761 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn)) 762 link = &(*link)->rb_left; 763 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn)) 764 link = &(*link)->rb_right; 765 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 766 link = &(*link)->rb_left; 767 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 768 link = &(*link)->rb_right; 769 else 770 return cur_timewait_info; 771 } 772 timewait_info->inserted_remote_qp = 1; 773 rb_link_node(&timewait_info->remote_qp_node, parent, link); 774 rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table); 775 return NULL; 776 } 777 778 static struct cm_id_private * 779 cm_insert_remote_sidr(struct cm_id_private *cm_id_priv) 780 { 781 struct rb_node **link = &cm.remote_sidr_table.rb_node; 782 struct rb_node *parent = NULL; 783 struct cm_id_private *cur_cm_id_priv; 784 __be32 remote_id = cm_id_priv->id.remote_id; 785 786 while (*link) { 787 parent = *link; 788 cur_cm_id_priv = rb_entry(parent, struct cm_id_private, 789 sidr_id_node); 790 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id)) 791 link = &(*link)->rb_left; 792 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id)) 793 link = &(*link)->rb_right; 794 else { 795 if (cur_cm_id_priv->sidr_slid < cm_id_priv->sidr_slid) 796 link = &(*link)->rb_left; 797 else if (cur_cm_id_priv->sidr_slid > cm_id_priv->sidr_slid) 798 link = &(*link)->rb_right; 799 else 800 return cur_cm_id_priv; 801 } 802 } 803 rb_link_node(&cm_id_priv->sidr_id_node, parent, link); 804 rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table); 805 return NULL; 806 } 807 808 static struct cm_id_private *cm_alloc_id_priv(struct ib_device *device, 809 ib_cm_handler cm_handler, 810 void *context) 811 { 812 struct cm_id_private *cm_id_priv; 813 u32 id; 814 int ret; 815 816 cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL); 817 if (!cm_id_priv) 818 return ERR_PTR(-ENOMEM); 819 820 cm_id_priv->id.state = IB_CM_IDLE; 821 cm_id_priv->id.device = device; 822 cm_id_priv->id.cm_handler = cm_handler; 823 cm_id_priv->id.context = context; 824 cm_id_priv->id.remote_cm_qpn = 1; 825 826 RB_CLEAR_NODE(&cm_id_priv->service_node); 827 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node); 828 spin_lock_init(&cm_id_priv->lock); 829 init_completion(&cm_id_priv->comp); 830 INIT_LIST_HEAD(&cm_id_priv->work_list); 831 atomic_set(&cm_id_priv->work_count, -1); 832 refcount_set(&cm_id_priv->refcount, 1); 833 834 ret = xa_alloc_cyclic(&cm.local_id_table, &id, NULL, xa_limit_32b, 835 &cm.local_id_next, GFP_KERNEL); 836 if (ret < 0) 837 goto error; 838 cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand; 839 840 return cm_id_priv; 841 842 error: 843 kfree(cm_id_priv); 844 return ERR_PTR(ret); 845 } 846 847 /* 848 * Make the ID visible to the MAD handlers and other threads that use the 849 * xarray. 850 */ 851 static void cm_finalize_id(struct cm_id_private *cm_id_priv) 852 { 853 xa_store(&cm.local_id_table, cm_local_id(cm_id_priv->id.local_id), 854 cm_id_priv, GFP_ATOMIC); 855 } 856 857 struct ib_cm_id *ib_create_cm_id(struct ib_device *device, 858 ib_cm_handler cm_handler, 859 void *context) 860 { 861 struct cm_id_private *cm_id_priv; 862 863 cm_id_priv = cm_alloc_id_priv(device, cm_handler, context); 864 if (IS_ERR(cm_id_priv)) 865 return ERR_CAST(cm_id_priv); 866 867 cm_finalize_id(cm_id_priv); 868 return &cm_id_priv->id; 869 } 870 EXPORT_SYMBOL(ib_create_cm_id); 871 872 static struct cm_work *cm_dequeue_work(struct cm_id_private *cm_id_priv) 873 { 874 struct cm_work *work; 875 876 if (list_empty(&cm_id_priv->work_list)) 877 return NULL; 878 879 work = list_entry(cm_id_priv->work_list.next, struct cm_work, list); 880 list_del(&work->list); 881 return work; 882 } 883 884 static void cm_free_work(struct cm_work *work) 885 { 886 if (work->mad_recv_wc) 887 ib_free_recv_mad(work->mad_recv_wc); 888 kfree(work); 889 } 890 891 static void cm_queue_work_unlock(struct cm_id_private *cm_id_priv, 892 struct cm_work *work) 893 __releases(&cm_id_priv->lock) 894 { 895 bool immediate; 896 897 /* 898 * To deliver the event to the user callback we have the drop the 899 * spinlock, however, we need to ensure that the user callback is single 900 * threaded and receives events in the temporal order. If there are 901 * already events being processed then thread new events onto a list, 902 * the thread currently processing will pick them up. 903 */ 904 immediate = atomic_inc_and_test(&cm_id_priv->work_count); 905 if (!immediate) { 906 list_add_tail(&work->list, &cm_id_priv->work_list); 907 /* 908 * This routine always consumes incoming reference. Once queued 909 * to the work_list then a reference is held by the thread 910 * currently running cm_process_work() and this reference is not 911 * needed. 912 */ 913 cm_deref_id(cm_id_priv); 914 } 915 spin_unlock_irq(&cm_id_priv->lock); 916 917 if (immediate) 918 cm_process_work(cm_id_priv, work); 919 } 920 921 static inline int cm_convert_to_ms(int iba_time) 922 { 923 /* approximate conversion to ms from 4.096us x 2^iba_time */ 924 return 1 << max(iba_time - 8, 0); 925 } 926 927 /* 928 * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time 929 * Because of how ack_timeout is stored, adding one doubles the timeout. 930 * To avoid large timeouts, select the max(ack_delay, life_time + 1), and 931 * increment it (round up) only if the other is within 50%. 932 */ 933 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time) 934 { 935 int ack_timeout = packet_life_time + 1; 936 937 if (ack_timeout >= ca_ack_delay) 938 ack_timeout += (ca_ack_delay >= (ack_timeout - 1)); 939 else 940 ack_timeout = ca_ack_delay + 941 (ack_timeout >= (ca_ack_delay - 1)); 942 943 return min(31, ack_timeout); 944 } 945 946 static void cm_remove_remote(struct cm_id_private *cm_id_priv) 947 { 948 struct cm_timewait_info *timewait_info = cm_id_priv->timewait_info; 949 950 if (timewait_info->inserted_remote_id) { 951 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table); 952 timewait_info->inserted_remote_id = 0; 953 } 954 955 if (timewait_info->inserted_remote_qp) { 956 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table); 957 timewait_info->inserted_remote_qp = 0; 958 } 959 } 960 961 static struct cm_timewait_info *cm_create_timewait_info(__be32 local_id) 962 { 963 struct cm_timewait_info *timewait_info; 964 965 timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL); 966 if (!timewait_info) 967 return ERR_PTR(-ENOMEM); 968 969 timewait_info->work.local_id = local_id; 970 INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler); 971 timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT; 972 return timewait_info; 973 } 974 975 static void cm_enter_timewait(struct cm_id_private *cm_id_priv) 976 { 977 int wait_time; 978 unsigned long flags; 979 struct cm_device *cm_dev; 980 981 lockdep_assert_held(&cm_id_priv->lock); 982 983 cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client); 984 if (!cm_dev) 985 return; 986 987 spin_lock_irqsave(&cm.lock, flags); 988 cm_remove_remote(cm_id_priv); 989 list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list); 990 spin_unlock_irqrestore(&cm.lock, flags); 991 992 /* 993 * The cm_id could be destroyed by the user before we exit timewait. 994 * To protect against this, we search for the cm_id after exiting 995 * timewait before notifying the user that we've exited timewait. 996 */ 997 cm_id_priv->id.state = IB_CM_TIMEWAIT; 998 wait_time = cm_convert_to_ms(cm_id_priv->av.timeout); 999 1000 /* Check if the device started its remove_one */ 1001 spin_lock_irqsave(&cm.lock, flags); 1002 if (!cm_dev->going_down) 1003 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work, 1004 msecs_to_jiffies(wait_time)); 1005 spin_unlock_irqrestore(&cm.lock, flags); 1006 1007 /* 1008 * The timewait_info is converted into a work and gets freed during 1009 * cm_free_work() in cm_timewait_handler(). 1010 */ 1011 BUILD_BUG_ON(offsetof(struct cm_timewait_info, work) != 0); 1012 cm_id_priv->timewait_info = NULL; 1013 } 1014 1015 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv) 1016 { 1017 unsigned long flags; 1018 1019 lockdep_assert_held(&cm_id_priv->lock); 1020 1021 cm_id_priv->id.state = IB_CM_IDLE; 1022 if (cm_id_priv->timewait_info) { 1023 spin_lock_irqsave(&cm.lock, flags); 1024 cm_remove_remote(cm_id_priv); 1025 spin_unlock_irqrestore(&cm.lock, flags); 1026 kfree(cm_id_priv->timewait_info); 1027 cm_id_priv->timewait_info = NULL; 1028 } 1029 } 1030 1031 static noinline void cm_destroy_id_wait_timeout(struct ib_cm_id *cm_id, 1032 enum ib_cm_state old_state) 1033 { 1034 struct cm_id_private *cm_id_priv; 1035 1036 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 1037 pr_err("%s: cm_id=%p timed out. state %d -> %d, refcnt=%d\n", __func__, 1038 cm_id, old_state, cm_id->state, refcount_read(&cm_id_priv->refcount)); 1039 } 1040 1041 static void cm_destroy_id(struct ib_cm_id *cm_id, int err) 1042 { 1043 struct cm_id_private *cm_id_priv; 1044 enum ib_cm_state old_state; 1045 struct cm_work *work; 1046 int ret; 1047 1048 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 1049 spin_lock_irq(&cm_id_priv->lock); 1050 old_state = cm_id->state; 1051 retest: 1052 switch (cm_id->state) { 1053 case IB_CM_LISTEN: 1054 spin_lock(&cm.lock); 1055 if (--cm_id_priv->listen_sharecount > 0) { 1056 /* The id is still shared. */ 1057 WARN_ON(refcount_read(&cm_id_priv->refcount) == 1); 1058 spin_unlock(&cm.lock); 1059 spin_unlock_irq(&cm_id_priv->lock); 1060 cm_deref_id(cm_id_priv); 1061 return; 1062 } 1063 cm_id->state = IB_CM_IDLE; 1064 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table); 1065 RB_CLEAR_NODE(&cm_id_priv->service_node); 1066 spin_unlock(&cm.lock); 1067 break; 1068 case IB_CM_SIDR_REQ_SENT: 1069 cm_id->state = IB_CM_IDLE; 1070 ib_cancel_mad(cm_id_priv->msg); 1071 break; 1072 case IB_CM_SIDR_REQ_RCVD: 1073 cm_send_sidr_rep_locked(cm_id_priv, 1074 &(struct ib_cm_sidr_rep_param){ 1075 .status = IB_SIDR_REJECT }); 1076 /* cm_send_sidr_rep_locked will not move to IDLE if it fails */ 1077 cm_id->state = IB_CM_IDLE; 1078 break; 1079 case IB_CM_REQ_SENT: 1080 case IB_CM_MRA_REQ_RCVD: 1081 ib_cancel_mad(cm_id_priv->msg); 1082 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_TIMEOUT, 1083 &cm_id_priv->id.device->node_guid, 1084 sizeof(cm_id_priv->id.device->node_guid), 1085 NULL, 0); 1086 break; 1087 case IB_CM_REQ_RCVD: 1088 if (err == -ENOMEM) { 1089 /* Do not reject to allow future retries. */ 1090 cm_reset_to_idle(cm_id_priv); 1091 } else { 1092 cm_send_rej_locked(cm_id_priv, 1093 IB_CM_REJ_CONSUMER_DEFINED, NULL, 0, 1094 NULL, 0); 1095 } 1096 break; 1097 case IB_CM_REP_SENT: 1098 case IB_CM_MRA_REP_RCVD: 1099 ib_cancel_mad(cm_id_priv->msg); 1100 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL, 1101 0, NULL, 0); 1102 goto retest; 1103 case IB_CM_MRA_REQ_SENT: 1104 case IB_CM_REP_RCVD: 1105 case IB_CM_MRA_REP_SENT: 1106 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL, 1107 0, NULL, 0); 1108 break; 1109 case IB_CM_ESTABLISHED: 1110 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT) { 1111 cm_id->state = IB_CM_IDLE; 1112 break; 1113 } 1114 cm_issue_dreq(cm_id_priv); 1115 cm_enter_timewait(cm_id_priv); 1116 goto retest; 1117 case IB_CM_DREQ_SENT: 1118 ib_cancel_mad(cm_id_priv->msg); 1119 cm_enter_timewait(cm_id_priv); 1120 goto retest; 1121 case IB_CM_DREQ_RCVD: 1122 cm_send_drep_locked(cm_id_priv, NULL, 0); 1123 WARN_ON(cm_id->state != IB_CM_TIMEWAIT); 1124 goto retest; 1125 case IB_CM_TIMEWAIT: 1126 /* 1127 * The cm_acquire_id in cm_timewait_handler will stop working 1128 * once we do xa_erase below, so just move to idle here for 1129 * consistency. 1130 */ 1131 cm_id->state = IB_CM_IDLE; 1132 break; 1133 case IB_CM_IDLE: 1134 break; 1135 } 1136 WARN_ON(cm_id->state != IB_CM_IDLE); 1137 1138 spin_lock(&cm.lock); 1139 /* Required for cleanup paths related cm_req_handler() */ 1140 if (cm_id_priv->timewait_info) { 1141 cm_remove_remote(cm_id_priv); 1142 kfree(cm_id_priv->timewait_info); 1143 cm_id_priv->timewait_info = NULL; 1144 } 1145 1146 WARN_ON(cm_id_priv->listen_sharecount); 1147 WARN_ON(!RB_EMPTY_NODE(&cm_id_priv->service_node)); 1148 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) 1149 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table); 1150 spin_unlock(&cm.lock); 1151 spin_unlock_irq(&cm_id_priv->lock); 1152 1153 xa_erase(&cm.local_id_table, cm_local_id(cm_id->local_id)); 1154 cm_deref_id(cm_id_priv); 1155 do { 1156 ret = wait_for_completion_timeout(&cm_id_priv->comp, 1157 msecs_to_jiffies( 1158 CM_DESTROY_ID_WAIT_TIMEOUT)); 1159 if (!ret) /* timeout happened */ 1160 cm_destroy_id_wait_timeout(cm_id, old_state); 1161 } while (!ret); 1162 1163 while ((work = cm_dequeue_work(cm_id_priv)) != NULL) 1164 cm_free_work(work); 1165 1166 cm_destroy_av(&cm_id_priv->av); 1167 cm_destroy_av(&cm_id_priv->alt_av); 1168 kfree(cm_id_priv->private_data); 1169 kfree_rcu(cm_id_priv, rcu); 1170 } 1171 1172 void ib_destroy_cm_id(struct ib_cm_id *cm_id) 1173 { 1174 cm_destroy_id(cm_id, 0); 1175 } 1176 EXPORT_SYMBOL(ib_destroy_cm_id); 1177 1178 static int cm_init_listen(struct cm_id_private *cm_id_priv, __be64 service_id) 1179 { 1180 if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID && 1181 (service_id != IB_CM_ASSIGN_SERVICE_ID)) 1182 return -EINVAL; 1183 1184 if (service_id == IB_CM_ASSIGN_SERVICE_ID) 1185 cm_id_priv->id.service_id = cpu_to_be64(cm.listen_service_id++); 1186 else 1187 cm_id_priv->id.service_id = service_id; 1188 1189 return 0; 1190 } 1191 1192 /** 1193 * ib_cm_listen - Initiates listening on the specified service ID for 1194 * connection and service ID resolution requests. 1195 * @cm_id: Connection identifier associated with the listen request. 1196 * @service_id: Service identifier matched against incoming connection 1197 * and service ID resolution requests. The service ID should be specified 1198 * network-byte order. If set to IB_CM_ASSIGN_SERVICE_ID, the CM will 1199 * assign a service ID to the caller. 1200 */ 1201 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id) 1202 { 1203 struct cm_id_private *cm_id_priv = 1204 container_of(cm_id, struct cm_id_private, id); 1205 unsigned long flags; 1206 int ret; 1207 1208 spin_lock_irqsave(&cm_id_priv->lock, flags); 1209 if (cm_id_priv->id.state != IB_CM_IDLE) { 1210 ret = -EINVAL; 1211 goto out; 1212 } 1213 1214 ret = cm_init_listen(cm_id_priv, service_id); 1215 if (ret) 1216 goto out; 1217 1218 if (!cm_insert_listen(cm_id_priv, NULL)) { 1219 ret = -EBUSY; 1220 goto out; 1221 } 1222 1223 cm_id_priv->id.state = IB_CM_LISTEN; 1224 ret = 0; 1225 1226 out: 1227 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1228 return ret; 1229 } 1230 EXPORT_SYMBOL(ib_cm_listen); 1231 1232 /** 1233 * ib_cm_insert_listen - Create a new listening ib_cm_id and listen on 1234 * the given service ID. 1235 * 1236 * If there's an existing ID listening on that same device and service ID, 1237 * return it. 1238 * 1239 * @device: Device associated with the cm_id. All related communication will 1240 * be associated with the specified device. 1241 * @cm_handler: Callback invoked to notify the user of CM events. 1242 * @service_id: Service identifier matched against incoming connection 1243 * and service ID resolution requests. The service ID should be specified 1244 * network-byte order. If set to IB_CM_ASSIGN_SERVICE_ID, the CM will 1245 * assign a service ID to the caller. 1246 * 1247 * Callers should call ib_destroy_cm_id when done with the listener ID. 1248 */ 1249 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device, 1250 ib_cm_handler cm_handler, 1251 __be64 service_id) 1252 { 1253 struct cm_id_private *listen_id_priv; 1254 struct cm_id_private *cm_id_priv; 1255 int err = 0; 1256 1257 /* Create an ID in advance, since the creation may sleep */ 1258 cm_id_priv = cm_alloc_id_priv(device, cm_handler, NULL); 1259 if (IS_ERR(cm_id_priv)) 1260 return ERR_CAST(cm_id_priv); 1261 1262 err = cm_init_listen(cm_id_priv, service_id); 1263 if (err) { 1264 ib_destroy_cm_id(&cm_id_priv->id); 1265 return ERR_PTR(err); 1266 } 1267 1268 spin_lock_irq(&cm_id_priv->lock); 1269 listen_id_priv = cm_insert_listen(cm_id_priv, cm_handler); 1270 if (listen_id_priv != cm_id_priv) { 1271 spin_unlock_irq(&cm_id_priv->lock); 1272 ib_destroy_cm_id(&cm_id_priv->id); 1273 if (!listen_id_priv) 1274 return ERR_PTR(-EINVAL); 1275 return &listen_id_priv->id; 1276 } 1277 cm_id_priv->id.state = IB_CM_LISTEN; 1278 spin_unlock_irq(&cm_id_priv->lock); 1279 1280 /* 1281 * A listen ID does not need to be in the xarray since it does not 1282 * receive mads, is not placed in the remote_id or remote_qpn rbtree, 1283 * and does not enter timewait. 1284 */ 1285 1286 return &cm_id_priv->id; 1287 } 1288 EXPORT_SYMBOL(ib_cm_insert_listen); 1289 1290 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv) 1291 { 1292 u64 hi_tid = 0, low_tid; 1293 1294 lockdep_assert_held(&cm_id_priv->lock); 1295 1296 low_tid = (u64)cm_id_priv->id.local_id; 1297 if (!cm_id_priv->av.port) 1298 return cpu_to_be64(low_tid); 1299 1300 spin_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 1301 if (cm_id_priv->av.port->mad_agent) 1302 hi_tid = ((u64)cm_id_priv->av.port->mad_agent->hi_tid) << 32; 1303 spin_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 1304 return cpu_to_be64(hi_tid | low_tid); 1305 } 1306 1307 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr, 1308 __be16 attr_id, __be64 tid) 1309 { 1310 hdr->base_version = IB_MGMT_BASE_VERSION; 1311 hdr->mgmt_class = IB_MGMT_CLASS_CM; 1312 hdr->class_version = IB_CM_CLASS_VERSION; 1313 hdr->method = IB_MGMT_METHOD_SEND; 1314 hdr->attr_id = attr_id; 1315 hdr->tid = tid; 1316 } 1317 1318 static void cm_format_mad_ece_hdr(struct ib_mad_hdr *hdr, __be16 attr_id, 1319 __be64 tid, u32 attr_mod) 1320 { 1321 cm_format_mad_hdr(hdr, attr_id, tid); 1322 hdr->attr_mod = cpu_to_be32(attr_mod); 1323 } 1324 1325 static void cm_format_req(struct cm_req_msg *req_msg, 1326 struct cm_id_private *cm_id_priv, 1327 struct ib_cm_req_param *param) 1328 { 1329 struct sa_path_rec *pri_path = param->primary_path; 1330 struct sa_path_rec *alt_path = param->alternate_path; 1331 bool pri_ext = false; 1332 __be16 lid; 1333 1334 if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA) 1335 pri_ext = opa_is_extended_lid(pri_path->opa.dlid, 1336 pri_path->opa.slid); 1337 1338 cm_format_mad_ece_hdr(&req_msg->hdr, CM_REQ_ATTR_ID, 1339 cm_form_tid(cm_id_priv), param->ece.attr_mod); 1340 1341 IBA_SET(CM_REQ_LOCAL_COMM_ID, req_msg, 1342 be32_to_cpu(cm_id_priv->id.local_id)); 1343 IBA_SET(CM_REQ_SERVICE_ID, req_msg, be64_to_cpu(param->service_id)); 1344 IBA_SET(CM_REQ_LOCAL_CA_GUID, req_msg, 1345 be64_to_cpu(cm_id_priv->id.device->node_guid)); 1346 IBA_SET(CM_REQ_LOCAL_QPN, req_msg, param->qp_num); 1347 IBA_SET(CM_REQ_INITIATOR_DEPTH, req_msg, param->initiator_depth); 1348 IBA_SET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg, 1349 param->remote_cm_response_timeout); 1350 cm_req_set_qp_type(req_msg, param->qp_type); 1351 IBA_SET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg, param->flow_control); 1352 IBA_SET(CM_REQ_STARTING_PSN, req_msg, param->starting_psn); 1353 IBA_SET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg, 1354 param->local_cm_response_timeout); 1355 IBA_SET(CM_REQ_PARTITION_KEY, req_msg, 1356 be16_to_cpu(param->primary_path->pkey)); 1357 IBA_SET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg, 1358 param->primary_path->mtu); 1359 IBA_SET(CM_REQ_MAX_CM_RETRIES, req_msg, param->max_cm_retries); 1360 1361 if (param->qp_type != IB_QPT_XRC_INI) { 1362 IBA_SET(CM_REQ_RESPONDER_RESOURCES, req_msg, 1363 param->responder_resources); 1364 IBA_SET(CM_REQ_RETRY_COUNT, req_msg, param->retry_count); 1365 IBA_SET(CM_REQ_RNR_RETRY_COUNT, req_msg, 1366 param->rnr_retry_count); 1367 IBA_SET(CM_REQ_SRQ, req_msg, param->srq); 1368 } 1369 1370 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) = 1371 pri_path->sgid; 1372 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) = 1373 pri_path->dgid; 1374 if (pri_ext) { 1375 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) 1376 ->global.interface_id = 1377 OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid)); 1378 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) 1379 ->global.interface_id = 1380 OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid)); 1381 } 1382 if (pri_path->hop_limit <= 1) { 1383 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 1384 be16_to_cpu(pri_ext ? 0 : 1385 htons(ntohl(sa_path_get_slid( 1386 pri_path))))); 1387 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, 1388 be16_to_cpu(pri_ext ? 0 : 1389 htons(ntohl(sa_path_get_dlid( 1390 pri_path))))); 1391 } else { 1392 1393 if (param->primary_path_inbound) { 1394 lid = param->primary_path_inbound->ib.dlid; 1395 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 1396 be16_to_cpu(lid)); 1397 } else 1398 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 1399 be16_to_cpu(IB_LID_PERMISSIVE)); 1400 1401 /* Work-around until there's a way to obtain remote LID info */ 1402 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, 1403 be16_to_cpu(IB_LID_PERMISSIVE)); 1404 } 1405 IBA_SET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg, 1406 be32_to_cpu(pri_path->flow_label)); 1407 IBA_SET(CM_REQ_PRIMARY_PACKET_RATE, req_msg, pri_path->rate); 1408 IBA_SET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg, pri_path->traffic_class); 1409 IBA_SET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg, pri_path->hop_limit); 1410 IBA_SET(CM_REQ_PRIMARY_SL, req_msg, pri_path->sl); 1411 IBA_SET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg, 1412 (pri_path->hop_limit <= 1)); 1413 IBA_SET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg, 1414 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay, 1415 pri_path->packet_life_time)); 1416 1417 if (alt_path) { 1418 bool alt_ext = false; 1419 1420 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA) 1421 alt_ext = opa_is_extended_lid(alt_path->opa.dlid, 1422 alt_path->opa.slid); 1423 1424 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg) = 1425 alt_path->sgid; 1426 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg) = 1427 alt_path->dgid; 1428 if (alt_ext) { 1429 IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, 1430 req_msg) 1431 ->global.interface_id = 1432 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid)); 1433 IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, 1434 req_msg) 1435 ->global.interface_id = 1436 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid)); 1437 } 1438 if (alt_path->hop_limit <= 1) { 1439 IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg, 1440 be16_to_cpu( 1441 alt_ext ? 0 : 1442 htons(ntohl(sa_path_get_slid( 1443 alt_path))))); 1444 IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg, 1445 be16_to_cpu( 1446 alt_ext ? 0 : 1447 htons(ntohl(sa_path_get_dlid( 1448 alt_path))))); 1449 } else { 1450 IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg, 1451 be16_to_cpu(IB_LID_PERMISSIVE)); 1452 IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg, 1453 be16_to_cpu(IB_LID_PERMISSIVE)); 1454 } 1455 IBA_SET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg, 1456 be32_to_cpu(alt_path->flow_label)); 1457 IBA_SET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg, alt_path->rate); 1458 IBA_SET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg, 1459 alt_path->traffic_class); 1460 IBA_SET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg, 1461 alt_path->hop_limit); 1462 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, alt_path->sl); 1463 IBA_SET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg, 1464 (alt_path->hop_limit <= 1)); 1465 IBA_SET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg, 1466 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay, 1467 alt_path->packet_life_time)); 1468 } 1469 IBA_SET(CM_REQ_VENDOR_ID, req_msg, param->ece.vendor_id); 1470 1471 if (param->private_data && param->private_data_len) 1472 IBA_SET_MEM(CM_REQ_PRIVATE_DATA, req_msg, param->private_data, 1473 param->private_data_len); 1474 } 1475 1476 static int cm_validate_req_param(struct ib_cm_req_param *param) 1477 { 1478 if (!param->primary_path) 1479 return -EINVAL; 1480 1481 if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC && 1482 param->qp_type != IB_QPT_XRC_INI) 1483 return -EINVAL; 1484 1485 if (param->private_data && 1486 param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE) 1487 return -EINVAL; 1488 1489 if (param->alternate_path && 1490 (param->alternate_path->pkey != param->primary_path->pkey || 1491 param->alternate_path->mtu != param->primary_path->mtu)) 1492 return -EINVAL; 1493 1494 return 0; 1495 } 1496 1497 int ib_send_cm_req(struct ib_cm_id *cm_id, 1498 struct ib_cm_req_param *param) 1499 { 1500 struct cm_av av = {}, alt_av = {}; 1501 struct cm_id_private *cm_id_priv; 1502 struct ib_mad_send_buf *msg; 1503 struct cm_req_msg *req_msg; 1504 unsigned long flags; 1505 int ret; 1506 1507 ret = cm_validate_req_param(param); 1508 if (ret) 1509 return ret; 1510 1511 /* Verify that we're not in timewait. */ 1512 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 1513 spin_lock_irqsave(&cm_id_priv->lock, flags); 1514 if (cm_id->state != IB_CM_IDLE || WARN_ON(cm_id_priv->timewait_info)) { 1515 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1516 return -EINVAL; 1517 } 1518 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1519 1520 cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv-> 1521 id.local_id); 1522 if (IS_ERR(cm_id_priv->timewait_info)) { 1523 ret = PTR_ERR(cm_id_priv->timewait_info); 1524 cm_id_priv->timewait_info = NULL; 1525 return ret; 1526 } 1527 1528 ret = cm_init_av_by_path(param->primary_path, 1529 param->ppath_sgid_attr, &av); 1530 if (ret) 1531 return ret; 1532 if (param->alternate_path) { 1533 ret = cm_init_av_by_path(param->alternate_path, NULL, 1534 &alt_av); 1535 if (ret) { 1536 cm_destroy_av(&av); 1537 return ret; 1538 } 1539 } 1540 cm_id->service_id = param->service_id; 1541 cm_id_priv->timeout_ms = cm_convert_to_ms( 1542 param->primary_path->packet_life_time) * 2 + 1543 cm_convert_to_ms( 1544 param->remote_cm_response_timeout); 1545 cm_id_priv->max_cm_retries = param->max_cm_retries; 1546 cm_id_priv->initiator_depth = param->initiator_depth; 1547 cm_id_priv->responder_resources = param->responder_resources; 1548 cm_id_priv->retry_count = param->retry_count; 1549 cm_id_priv->path_mtu = param->primary_path->mtu; 1550 cm_id_priv->pkey = param->primary_path->pkey; 1551 cm_id_priv->qp_type = param->qp_type; 1552 1553 spin_lock_irqsave(&cm_id_priv->lock, flags); 1554 1555 cm_move_av_from_path(&cm_id_priv->av, &av); 1556 if (param->primary_path_outbound) 1557 cm_id_priv->av.dlid_datapath = 1558 be16_to_cpu(param->primary_path_outbound->ib.dlid); 1559 1560 if (param->alternate_path) 1561 cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av); 1562 1563 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_REQ_SENT); 1564 if (IS_ERR(msg)) { 1565 ret = PTR_ERR(msg); 1566 goto out_unlock; 1567 } 1568 1569 req_msg = (struct cm_req_msg *)msg->mad; 1570 cm_format_req(req_msg, cm_id_priv, param); 1571 cm_id_priv->tid = req_msg->hdr.tid; 1572 1573 cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg)); 1574 cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg)); 1575 1576 trace_icm_send_req(&cm_id_priv->id); 1577 ret = ib_post_send_mad(msg, NULL); 1578 if (ret) 1579 goto out_free; 1580 BUG_ON(cm_id->state != IB_CM_IDLE); 1581 cm_id->state = IB_CM_REQ_SENT; 1582 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1583 return 0; 1584 out_free: 1585 cm_free_priv_msg(msg); 1586 out_unlock: 1587 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1588 return ret; 1589 } 1590 EXPORT_SYMBOL(ib_send_cm_req); 1591 1592 static int cm_issue_rej(struct cm_port *port, 1593 struct ib_mad_recv_wc *mad_recv_wc, 1594 enum ib_cm_rej_reason reason, 1595 enum cm_msg_response msg_rejected, 1596 void *ari, u8 ari_length) 1597 { 1598 struct ib_mad_send_buf *msg = NULL; 1599 struct cm_rej_msg *rej_msg, *rcv_msg; 1600 int ret; 1601 1602 ret = cm_alloc_response_msg(port, mad_recv_wc, false, &msg); 1603 if (ret) 1604 return ret; 1605 1606 /* We just need common CM header information. Cast to any message. */ 1607 rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad; 1608 rej_msg = (struct cm_rej_msg *) msg->mad; 1609 1610 cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid); 1611 IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg, 1612 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg)); 1613 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1614 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg)); 1615 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected); 1616 IBA_SET(CM_REJ_REASON, rej_msg, reason); 1617 1618 if (ari && ari_length) { 1619 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length); 1620 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length); 1621 } 1622 1623 trace_icm_issue_rej( 1624 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg), 1625 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg)); 1626 ret = ib_post_send_mad(msg, NULL); 1627 if (ret) 1628 cm_free_msg(msg); 1629 1630 return ret; 1631 } 1632 1633 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg) 1634 { 1635 return ((cpu_to_be16( 1636 IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) || 1637 (ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, 1638 req_msg)))); 1639 } 1640 1641 static void cm_path_set_rec_type(struct ib_device *ib_device, u32 port_num, 1642 struct sa_path_rec *path, union ib_gid *gid) 1643 { 1644 if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num)) 1645 path->rec_type = SA_PATH_REC_TYPE_OPA; 1646 else 1647 path->rec_type = SA_PATH_REC_TYPE_IB; 1648 } 1649 1650 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg, 1651 struct sa_path_rec *primary_path, 1652 struct sa_path_rec *alt_path, 1653 struct ib_wc *wc) 1654 { 1655 u32 lid; 1656 1657 if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) { 1658 sa_path_set_dlid(primary_path, wc->slid); 1659 sa_path_set_slid(primary_path, 1660 IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID, 1661 req_msg)); 1662 } else { 1663 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1664 CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg)); 1665 sa_path_set_dlid(primary_path, lid); 1666 1667 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1668 CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg)); 1669 sa_path_set_slid(primary_path, lid); 1670 } 1671 1672 if (!cm_req_has_alt_path(req_msg)) 1673 return; 1674 1675 if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) { 1676 sa_path_set_dlid(alt_path, 1677 IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, 1678 req_msg)); 1679 sa_path_set_slid(alt_path, 1680 IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, 1681 req_msg)); 1682 } else { 1683 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1684 CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg)); 1685 sa_path_set_dlid(alt_path, lid); 1686 1687 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1688 CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg)); 1689 sa_path_set_slid(alt_path, lid); 1690 } 1691 } 1692 1693 static void cm_format_paths_from_req(struct cm_req_msg *req_msg, 1694 struct sa_path_rec *primary_path, 1695 struct sa_path_rec *alt_path, 1696 struct ib_wc *wc) 1697 { 1698 primary_path->dgid = 1699 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg); 1700 primary_path->sgid = 1701 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg); 1702 primary_path->flow_label = 1703 cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg)); 1704 primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg); 1705 primary_path->traffic_class = 1706 IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg); 1707 primary_path->reversible = 1; 1708 primary_path->pkey = 1709 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg)); 1710 primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg); 1711 primary_path->mtu_selector = IB_SA_EQ; 1712 primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg); 1713 primary_path->rate_selector = IB_SA_EQ; 1714 primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg); 1715 primary_path->packet_life_time_selector = IB_SA_EQ; 1716 primary_path->packet_life_time = 1717 IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg); 1718 primary_path->packet_life_time -= (primary_path->packet_life_time > 0); 1719 primary_path->service_id = 1720 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)); 1721 if (sa_path_is_roce(primary_path)) 1722 primary_path->roce.route_resolved = false; 1723 1724 if (cm_req_has_alt_path(req_msg)) { 1725 alt_path->dgid = *IBA_GET_MEM_PTR( 1726 CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg); 1727 alt_path->sgid = *IBA_GET_MEM_PTR( 1728 CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg); 1729 alt_path->flow_label = cpu_to_be32( 1730 IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg)); 1731 alt_path->hop_limit = 1732 IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg); 1733 alt_path->traffic_class = 1734 IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg); 1735 alt_path->reversible = 1; 1736 alt_path->pkey = 1737 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg)); 1738 alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg); 1739 alt_path->mtu_selector = IB_SA_EQ; 1740 alt_path->mtu = 1741 IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg); 1742 alt_path->rate_selector = IB_SA_EQ; 1743 alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg); 1744 alt_path->packet_life_time_selector = IB_SA_EQ; 1745 alt_path->packet_life_time = 1746 IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg); 1747 alt_path->packet_life_time -= (alt_path->packet_life_time > 0); 1748 alt_path->service_id = 1749 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)); 1750 1751 if (sa_path_is_roce(alt_path)) 1752 alt_path->roce.route_resolved = false; 1753 } 1754 cm_format_path_lid_from_req(req_msg, primary_path, alt_path, wc); 1755 } 1756 1757 static u16 cm_get_bth_pkey(struct cm_work *work) 1758 { 1759 struct ib_device *ib_dev = work->port->cm_dev->ib_device; 1760 u32 port_num = work->port->port_num; 1761 u16 pkey_index = work->mad_recv_wc->wc->pkey_index; 1762 u16 pkey; 1763 int ret; 1764 1765 ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey); 1766 if (ret) { 1767 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %u, pkey index %u). %d\n", 1768 port_num, pkey_index, ret); 1769 return 0; 1770 } 1771 1772 return pkey; 1773 } 1774 1775 /** 1776 * cm_opa_to_ib_sgid - Convert OPA SGID to IB SGID 1777 * ULPs (such as IPoIB) do not understand OPA GIDs and will 1778 * reject them as the local_gid will not match the sgid. Therefore, 1779 * change the pathrec's SGID to an IB SGID. 1780 * 1781 * @work: Work completion 1782 * @path: Path record 1783 */ 1784 static void cm_opa_to_ib_sgid(struct cm_work *work, 1785 struct sa_path_rec *path) 1786 { 1787 struct ib_device *dev = work->port->cm_dev->ib_device; 1788 u32 port_num = work->port->port_num; 1789 1790 if (rdma_cap_opa_ah(dev, port_num) && 1791 (ib_is_opa_gid(&path->sgid))) { 1792 union ib_gid sgid; 1793 1794 if (rdma_query_gid(dev, port_num, 0, &sgid)) { 1795 dev_warn(&dev->dev, 1796 "Error updating sgid in CM request\n"); 1797 return; 1798 } 1799 1800 path->sgid = sgid; 1801 } 1802 } 1803 1804 static void cm_format_req_event(struct cm_work *work, 1805 struct cm_id_private *cm_id_priv, 1806 struct ib_cm_id *listen_id) 1807 { 1808 struct cm_req_msg *req_msg; 1809 struct ib_cm_req_event_param *param; 1810 1811 req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; 1812 param = &work->cm_event.param.req_rcvd; 1813 param->listen_id = listen_id; 1814 param->bth_pkey = cm_get_bth_pkey(work); 1815 param->port = cm_id_priv->av.port->port_num; 1816 param->primary_path = &work->path[0]; 1817 cm_opa_to_ib_sgid(work, param->primary_path); 1818 if (cm_req_has_alt_path(req_msg)) { 1819 param->alternate_path = &work->path[1]; 1820 cm_opa_to_ib_sgid(work, param->alternate_path); 1821 } else { 1822 param->alternate_path = NULL; 1823 } 1824 param->remote_ca_guid = 1825 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg)); 1826 param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg); 1827 param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg); 1828 param->qp_type = cm_req_get_qp_type(req_msg); 1829 param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg); 1830 param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg); 1831 param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg); 1832 param->local_cm_response_timeout = 1833 IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg); 1834 param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg); 1835 param->remote_cm_response_timeout = 1836 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg); 1837 param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg); 1838 param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg); 1839 param->srq = IBA_GET(CM_REQ_SRQ, req_msg); 1840 param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr; 1841 param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg); 1842 param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod); 1843 1844 work->cm_event.private_data = 1845 IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg); 1846 } 1847 1848 static void cm_process_work(struct cm_id_private *cm_id_priv, 1849 struct cm_work *work) 1850 { 1851 int ret; 1852 1853 /* We will typically only have the current event to report. */ 1854 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event); 1855 cm_free_work(work); 1856 1857 while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) { 1858 spin_lock_irq(&cm_id_priv->lock); 1859 work = cm_dequeue_work(cm_id_priv); 1860 spin_unlock_irq(&cm_id_priv->lock); 1861 if (!work) 1862 return; 1863 1864 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, 1865 &work->cm_event); 1866 cm_free_work(work); 1867 } 1868 cm_deref_id(cm_id_priv); 1869 if (ret) 1870 cm_destroy_id(&cm_id_priv->id, ret); 1871 } 1872 1873 static void cm_format_mra(struct cm_mra_msg *mra_msg, 1874 struct cm_id_private *cm_id_priv, 1875 enum cm_msg_response msg_mraed, u8 service_timeout, 1876 const void *private_data, u8 private_data_len) 1877 { 1878 cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid); 1879 IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed); 1880 IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg, 1881 be32_to_cpu(cm_id_priv->id.local_id)); 1882 IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg, 1883 be32_to_cpu(cm_id_priv->id.remote_id)); 1884 IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, service_timeout); 1885 1886 if (private_data && private_data_len) 1887 IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data, 1888 private_data_len); 1889 } 1890 1891 static void cm_format_rej(struct cm_rej_msg *rej_msg, 1892 struct cm_id_private *cm_id_priv, 1893 enum ib_cm_rej_reason reason, void *ari, 1894 u8 ari_length, const void *private_data, 1895 u8 private_data_len, enum ib_cm_state state) 1896 { 1897 lockdep_assert_held(&cm_id_priv->lock); 1898 1899 cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid); 1900 IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg, 1901 be32_to_cpu(cm_id_priv->id.remote_id)); 1902 1903 switch (state) { 1904 case IB_CM_REQ_RCVD: 1905 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0)); 1906 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ); 1907 break; 1908 case IB_CM_MRA_REQ_SENT: 1909 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1910 be32_to_cpu(cm_id_priv->id.local_id)); 1911 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ); 1912 break; 1913 case IB_CM_REP_RCVD: 1914 case IB_CM_MRA_REP_SENT: 1915 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1916 be32_to_cpu(cm_id_priv->id.local_id)); 1917 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP); 1918 break; 1919 default: 1920 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1921 be32_to_cpu(cm_id_priv->id.local_id)); 1922 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, 1923 CM_MSG_RESPONSE_OTHER); 1924 break; 1925 } 1926 1927 IBA_SET(CM_REJ_REASON, rej_msg, reason); 1928 if (ari && ari_length) { 1929 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length); 1930 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length); 1931 } 1932 1933 if (private_data && private_data_len) 1934 IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data, 1935 private_data_len); 1936 } 1937 1938 static void cm_dup_req_handler(struct cm_work *work, 1939 struct cm_id_private *cm_id_priv) 1940 { 1941 struct ib_mad_send_buf *msg = NULL; 1942 int ret; 1943 1944 atomic_long_inc( 1945 &work->port->counters[CM_RECV_DUPLICATES][CM_REQ_COUNTER]); 1946 1947 /* Quick state check to discard duplicate REQs. */ 1948 spin_lock_irq(&cm_id_priv->lock); 1949 if (cm_id_priv->id.state == IB_CM_REQ_RCVD) { 1950 spin_unlock_irq(&cm_id_priv->lock); 1951 return; 1952 } 1953 spin_unlock_irq(&cm_id_priv->lock); 1954 1955 ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, true, &msg); 1956 if (ret) 1957 return; 1958 1959 spin_lock_irq(&cm_id_priv->lock); 1960 switch (cm_id_priv->id.state) { 1961 case IB_CM_MRA_REQ_SENT: 1962 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 1963 CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout, 1964 cm_id_priv->private_data, 1965 cm_id_priv->private_data_len); 1966 break; 1967 case IB_CM_TIMEWAIT: 1968 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, 1969 IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0, 1970 IB_CM_TIMEWAIT); 1971 break; 1972 default: 1973 goto unlock; 1974 } 1975 spin_unlock_irq(&cm_id_priv->lock); 1976 1977 trace_icm_send_dup_req(&cm_id_priv->id); 1978 ret = ib_post_send_mad(msg, NULL); 1979 if (ret) 1980 goto free; 1981 return; 1982 1983 unlock: spin_unlock_irq(&cm_id_priv->lock); 1984 free: cm_free_msg(msg); 1985 } 1986 1987 static struct cm_id_private *cm_match_req(struct cm_work *work, 1988 struct cm_id_private *cm_id_priv) 1989 { 1990 struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv; 1991 struct cm_timewait_info *timewait_info; 1992 struct cm_req_msg *req_msg; 1993 1994 req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; 1995 1996 /* Check for possible duplicate REQ. */ 1997 spin_lock_irq(&cm.lock); 1998 timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info); 1999 if (timewait_info) { 2000 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 2001 timewait_info->work.remote_id); 2002 spin_unlock_irq(&cm.lock); 2003 if (cur_cm_id_priv) { 2004 cm_dup_req_handler(work, cur_cm_id_priv); 2005 cm_deref_id(cur_cm_id_priv); 2006 } 2007 return NULL; 2008 } 2009 2010 /* Check for stale connections. */ 2011 timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info); 2012 if (timewait_info) { 2013 cm_remove_remote(cm_id_priv); 2014 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 2015 timewait_info->work.remote_id); 2016 2017 spin_unlock_irq(&cm.lock); 2018 cm_issue_rej(work->port, work->mad_recv_wc, 2019 IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ, 2020 NULL, 0); 2021 if (cur_cm_id_priv) { 2022 ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0); 2023 cm_deref_id(cur_cm_id_priv); 2024 } 2025 return NULL; 2026 } 2027 2028 /* Find matching listen request. */ 2029 listen_cm_id_priv = cm_find_listen( 2030 cm_id_priv->id.device, 2031 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg))); 2032 if (!listen_cm_id_priv) { 2033 cm_remove_remote(cm_id_priv); 2034 spin_unlock_irq(&cm.lock); 2035 cm_issue_rej(work->port, work->mad_recv_wc, 2036 IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ, 2037 NULL, 0); 2038 return NULL; 2039 } 2040 spin_unlock_irq(&cm.lock); 2041 return listen_cm_id_priv; 2042 } 2043 2044 /* 2045 * Work-around for inter-subnet connections. If the LIDs are permissive, 2046 * we need to override the LID/SL data in the REQ with the LID information 2047 * in the work completion. 2048 */ 2049 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc) 2050 { 2051 if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) { 2052 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, 2053 req_msg)) == IB_LID_PERMISSIVE) { 2054 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 2055 be16_to_cpu(ib_lid_be16(wc->slid))); 2056 IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl); 2057 } 2058 2059 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID, 2060 req_msg)) == IB_LID_PERMISSIVE) 2061 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, 2062 wc->dlid_path_bits); 2063 } 2064 2065 if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) { 2066 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, 2067 req_msg)) == IB_LID_PERMISSIVE) { 2068 IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg, 2069 be16_to_cpu(ib_lid_be16(wc->slid))); 2070 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl); 2071 } 2072 2073 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, 2074 req_msg)) == IB_LID_PERMISSIVE) 2075 IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg, 2076 wc->dlid_path_bits); 2077 } 2078 } 2079 2080 static int cm_req_handler(struct cm_work *work) 2081 { 2082 struct cm_id_private *cm_id_priv, *listen_cm_id_priv; 2083 struct cm_req_msg *req_msg; 2084 const struct ib_global_route *grh; 2085 const struct ib_gid_attr *gid_attr; 2086 int ret; 2087 2088 req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; 2089 2090 cm_id_priv = 2091 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL); 2092 if (IS_ERR(cm_id_priv)) 2093 return PTR_ERR(cm_id_priv); 2094 2095 cm_id_priv->id.remote_id = 2096 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg)); 2097 cm_id_priv->id.service_id = 2098 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)); 2099 cm_id_priv->tid = req_msg->hdr.tid; 2100 cm_id_priv->timeout_ms = cm_convert_to_ms( 2101 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg)); 2102 cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg); 2103 cm_id_priv->remote_qpn = 2104 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg)); 2105 cm_id_priv->initiator_depth = 2106 IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg); 2107 cm_id_priv->responder_resources = 2108 IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg); 2109 cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg); 2110 cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg)); 2111 cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg)); 2112 cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg); 2113 cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg); 2114 cm_id_priv->qp_type = cm_req_get_qp_type(req_msg); 2115 2116 ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc, 2117 work->mad_recv_wc->recv_buf.grh, 2118 &cm_id_priv->av); 2119 if (ret) 2120 goto destroy; 2121 cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv-> 2122 id.local_id); 2123 if (IS_ERR(cm_id_priv->timewait_info)) { 2124 ret = PTR_ERR(cm_id_priv->timewait_info); 2125 cm_id_priv->timewait_info = NULL; 2126 goto destroy; 2127 } 2128 cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id; 2129 cm_id_priv->timewait_info->remote_ca_guid = 2130 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg)); 2131 cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn; 2132 2133 /* 2134 * Note that the ID pointer is not in the xarray at this point, 2135 * so this set is only visible to the local thread. 2136 */ 2137 cm_id_priv->id.state = IB_CM_REQ_RCVD; 2138 2139 listen_cm_id_priv = cm_match_req(work, cm_id_priv); 2140 if (!listen_cm_id_priv) { 2141 trace_icm_no_listener_err(&cm_id_priv->id); 2142 cm_id_priv->id.state = IB_CM_IDLE; 2143 ret = -EINVAL; 2144 goto destroy; 2145 } 2146 2147 memset(&work->path[0], 0, sizeof(work->path[0])); 2148 if (cm_req_has_alt_path(req_msg)) 2149 memset(&work->path[1], 0, sizeof(work->path[1])); 2150 grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr); 2151 gid_attr = grh->sgid_attr; 2152 2153 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE) { 2154 work->path[0].rec_type = 2155 sa_conv_gid_to_pathrec_type(gid_attr->gid_type); 2156 } else { 2157 cm_process_routed_req(req_msg, work->mad_recv_wc->wc); 2158 cm_path_set_rec_type( 2159 work->port->cm_dev->ib_device, work->port->port_num, 2160 &work->path[0], 2161 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, 2162 req_msg)); 2163 } 2164 if (cm_req_has_alt_path(req_msg)) 2165 work->path[1].rec_type = work->path[0].rec_type; 2166 cm_format_paths_from_req(req_msg, &work->path[0], 2167 &work->path[1], work->mad_recv_wc->wc); 2168 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE) 2169 sa_path_set_dmac(&work->path[0], 2170 cm_id_priv->av.ah_attr.roce.dmac); 2171 work->path[0].hop_limit = grh->hop_limit; 2172 2173 /* This destroy call is needed to pair with cm_init_av_for_response */ 2174 cm_destroy_av(&cm_id_priv->av); 2175 ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av); 2176 if (ret) { 2177 int err; 2178 2179 err = rdma_query_gid(work->port->cm_dev->ib_device, 2180 work->port->port_num, 0, 2181 &work->path[0].sgid); 2182 if (err) 2183 ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID, 2184 NULL, 0, NULL, 0); 2185 else 2186 ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID, 2187 &work->path[0].sgid, 2188 sizeof(work->path[0].sgid), 2189 NULL, 0); 2190 goto rejected; 2191 } 2192 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_IB) 2193 cm_id_priv->av.dlid_datapath = 2194 IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg); 2195 2196 if (cm_req_has_alt_path(req_msg)) { 2197 ret = cm_init_av_by_path(&work->path[1], NULL, 2198 &cm_id_priv->alt_av); 2199 if (ret) { 2200 ib_send_cm_rej(&cm_id_priv->id, 2201 IB_CM_REJ_INVALID_ALT_GID, 2202 &work->path[0].sgid, 2203 sizeof(work->path[0].sgid), NULL, 0); 2204 goto rejected; 2205 } 2206 } 2207 2208 cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler; 2209 cm_id_priv->id.context = listen_cm_id_priv->id.context; 2210 cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id); 2211 2212 /* Now MAD handlers can see the new ID */ 2213 spin_lock_irq(&cm_id_priv->lock); 2214 cm_finalize_id(cm_id_priv); 2215 2216 /* Refcount belongs to the event, pairs with cm_process_work() */ 2217 refcount_inc(&cm_id_priv->refcount); 2218 cm_queue_work_unlock(cm_id_priv, work); 2219 /* 2220 * Since this ID was just created and was not made visible to other MAD 2221 * handlers until the cm_finalize_id() above we know that the 2222 * cm_process_work() will deliver the event and the listen_cm_id 2223 * embedded in the event can be derefed here. 2224 */ 2225 cm_deref_id(listen_cm_id_priv); 2226 return 0; 2227 2228 rejected: 2229 cm_deref_id(listen_cm_id_priv); 2230 destroy: 2231 ib_destroy_cm_id(&cm_id_priv->id); 2232 return ret; 2233 } 2234 2235 static void cm_format_rep(struct cm_rep_msg *rep_msg, 2236 struct cm_id_private *cm_id_priv, 2237 struct ib_cm_rep_param *param) 2238 { 2239 cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid, 2240 param->ece.attr_mod); 2241 IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg, 2242 be32_to_cpu(cm_id_priv->id.local_id)); 2243 IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg, 2244 be32_to_cpu(cm_id_priv->id.remote_id)); 2245 IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn); 2246 IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg, 2247 param->responder_resources); 2248 IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg, 2249 cm_id_priv->av.port->cm_dev->ack_delay); 2250 IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted); 2251 IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count); 2252 IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg, 2253 be64_to_cpu(cm_id_priv->id.device->node_guid)); 2254 2255 if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) { 2256 IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg, 2257 param->initiator_depth); 2258 IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg, 2259 param->flow_control); 2260 IBA_SET(CM_REP_SRQ, rep_msg, param->srq); 2261 IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num); 2262 } else { 2263 IBA_SET(CM_REP_SRQ, rep_msg, 1); 2264 IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num); 2265 } 2266 2267 IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id); 2268 IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8); 2269 IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16); 2270 2271 if (param->private_data && param->private_data_len) 2272 IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data, 2273 param->private_data_len); 2274 } 2275 2276 int ib_send_cm_rep(struct ib_cm_id *cm_id, 2277 struct ib_cm_rep_param *param) 2278 { 2279 struct cm_id_private *cm_id_priv; 2280 struct ib_mad_send_buf *msg; 2281 struct cm_rep_msg *rep_msg; 2282 unsigned long flags; 2283 int ret; 2284 2285 if (param->private_data && 2286 param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE) 2287 return -EINVAL; 2288 2289 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 2290 spin_lock_irqsave(&cm_id_priv->lock, flags); 2291 if (cm_id->state != IB_CM_REQ_RCVD && 2292 cm_id->state != IB_CM_MRA_REQ_SENT) { 2293 trace_icm_send_rep_err(cm_id_priv->id.local_id, cm_id->state); 2294 ret = -EINVAL; 2295 goto out; 2296 } 2297 2298 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_REP_SENT); 2299 if (IS_ERR(msg)) { 2300 ret = PTR_ERR(msg); 2301 goto out; 2302 } 2303 2304 rep_msg = (struct cm_rep_msg *) msg->mad; 2305 cm_format_rep(rep_msg, cm_id_priv, param); 2306 2307 trace_icm_send_rep(cm_id); 2308 ret = ib_post_send_mad(msg, NULL); 2309 if (ret) 2310 goto out_free; 2311 2312 cm_id->state = IB_CM_REP_SENT; 2313 cm_id_priv->initiator_depth = param->initiator_depth; 2314 cm_id_priv->responder_resources = param->responder_resources; 2315 cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg)); 2316 WARN_ONCE(param->qp_num & 0xFF000000, 2317 "IBTA declares QPN to be 24 bits, but it is 0x%X\n", 2318 param->qp_num); 2319 cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF); 2320 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2321 return 0; 2322 2323 out_free: 2324 cm_free_priv_msg(msg); 2325 out: 2326 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2327 return ret; 2328 } 2329 EXPORT_SYMBOL(ib_send_cm_rep); 2330 2331 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg, 2332 struct cm_id_private *cm_id_priv, 2333 const void *private_data, 2334 u8 private_data_len) 2335 { 2336 cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid); 2337 IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg, 2338 be32_to_cpu(cm_id_priv->id.local_id)); 2339 IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg, 2340 be32_to_cpu(cm_id_priv->id.remote_id)); 2341 2342 if (private_data && private_data_len) 2343 IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data, 2344 private_data_len); 2345 } 2346 2347 int ib_send_cm_rtu(struct ib_cm_id *cm_id, 2348 const void *private_data, 2349 u8 private_data_len) 2350 { 2351 struct cm_id_private *cm_id_priv; 2352 struct ib_mad_send_buf *msg; 2353 unsigned long flags; 2354 void *data; 2355 int ret; 2356 2357 if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE) 2358 return -EINVAL; 2359 2360 data = cm_copy_private_data(private_data, private_data_len); 2361 if (IS_ERR(data)) 2362 return PTR_ERR(data); 2363 2364 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 2365 spin_lock_irqsave(&cm_id_priv->lock, flags); 2366 if (cm_id->state != IB_CM_REP_RCVD && 2367 cm_id->state != IB_CM_MRA_REP_SENT) { 2368 trace_icm_send_cm_rtu_err(cm_id); 2369 ret = -EINVAL; 2370 goto error; 2371 } 2372 2373 msg = cm_alloc_msg(cm_id_priv); 2374 if (IS_ERR(msg)) { 2375 ret = PTR_ERR(msg); 2376 goto error; 2377 } 2378 2379 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv, 2380 private_data, private_data_len); 2381 2382 trace_icm_send_rtu(cm_id); 2383 ret = ib_post_send_mad(msg, NULL); 2384 if (ret) { 2385 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2386 cm_free_msg(msg); 2387 kfree(data); 2388 return ret; 2389 } 2390 2391 cm_id->state = IB_CM_ESTABLISHED; 2392 cm_set_private_data(cm_id_priv, data, private_data_len); 2393 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2394 return 0; 2395 2396 error: spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2397 kfree(data); 2398 return ret; 2399 } 2400 EXPORT_SYMBOL(ib_send_cm_rtu); 2401 2402 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type) 2403 { 2404 struct cm_rep_msg *rep_msg; 2405 struct ib_cm_rep_event_param *param; 2406 2407 rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad; 2408 param = &work->cm_event.param.rep_rcvd; 2409 param->remote_ca_guid = 2410 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg)); 2411 param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg); 2412 param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type)); 2413 param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg); 2414 param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg); 2415 param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg); 2416 param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg); 2417 param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg); 2418 param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg); 2419 param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg); 2420 param->srq = IBA_GET(CM_REP_SRQ, rep_msg); 2421 param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16; 2422 param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8; 2423 param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg); 2424 param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod); 2425 2426 work->cm_event.private_data = 2427 IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg); 2428 } 2429 2430 static void cm_dup_rep_handler(struct cm_work *work) 2431 { 2432 struct cm_id_private *cm_id_priv; 2433 struct cm_rep_msg *rep_msg; 2434 struct ib_mad_send_buf *msg = NULL; 2435 int ret; 2436 2437 rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad; 2438 cm_id_priv = cm_acquire_id( 2439 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 2440 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg))); 2441 if (!cm_id_priv) 2442 return; 2443 2444 atomic_long_inc( 2445 &work->port->counters[CM_RECV_DUPLICATES][CM_REP_COUNTER]); 2446 ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, true, &msg); 2447 if (ret) 2448 goto deref; 2449 2450 spin_lock_irq(&cm_id_priv->lock); 2451 if (cm_id_priv->id.state == IB_CM_ESTABLISHED) 2452 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv, 2453 cm_id_priv->private_data, 2454 cm_id_priv->private_data_len); 2455 else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT) 2456 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 2457 CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout, 2458 cm_id_priv->private_data, 2459 cm_id_priv->private_data_len); 2460 else 2461 goto unlock; 2462 spin_unlock_irq(&cm_id_priv->lock); 2463 2464 trace_icm_send_dup_rep(&cm_id_priv->id); 2465 ret = ib_post_send_mad(msg, NULL); 2466 if (ret) 2467 goto free; 2468 goto deref; 2469 2470 unlock: spin_unlock_irq(&cm_id_priv->lock); 2471 free: cm_free_msg(msg); 2472 deref: cm_deref_id(cm_id_priv); 2473 } 2474 2475 static int cm_rep_handler(struct cm_work *work) 2476 { 2477 struct cm_id_private *cm_id_priv; 2478 struct cm_rep_msg *rep_msg; 2479 int ret; 2480 struct cm_id_private *cur_cm_id_priv; 2481 struct cm_timewait_info *timewait_info; 2482 2483 rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad; 2484 cm_id_priv = cm_acquire_id( 2485 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0); 2486 if (!cm_id_priv) { 2487 cm_dup_rep_handler(work); 2488 trace_icm_remote_no_priv_err( 2489 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)); 2490 return -EINVAL; 2491 } 2492 2493 cm_format_rep_event(work, cm_id_priv->qp_type); 2494 2495 spin_lock_irq(&cm_id_priv->lock); 2496 switch (cm_id_priv->id.state) { 2497 case IB_CM_REQ_SENT: 2498 case IB_CM_MRA_REQ_RCVD: 2499 break; 2500 default: 2501 ret = -EINVAL; 2502 trace_icm_rep_unknown_err( 2503 IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg), 2504 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg), 2505 cm_id_priv->id.state); 2506 spin_unlock_irq(&cm_id_priv->lock); 2507 goto error; 2508 } 2509 2510 cm_id_priv->timewait_info->work.remote_id = 2511 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)); 2512 cm_id_priv->timewait_info->remote_ca_guid = 2513 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg)); 2514 cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type); 2515 2516 spin_lock(&cm.lock); 2517 /* Check for duplicate REP. */ 2518 if (cm_insert_remote_id(cm_id_priv->timewait_info)) { 2519 spin_unlock(&cm.lock); 2520 spin_unlock_irq(&cm_id_priv->lock); 2521 ret = -EINVAL; 2522 trace_icm_insert_failed_err( 2523 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)); 2524 goto error; 2525 } 2526 /* Check for a stale connection. */ 2527 timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info); 2528 if (timewait_info) { 2529 cm_remove_remote(cm_id_priv); 2530 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 2531 timewait_info->work.remote_id); 2532 2533 spin_unlock(&cm.lock); 2534 spin_unlock_irq(&cm_id_priv->lock); 2535 cm_issue_rej(work->port, work->mad_recv_wc, 2536 IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP, 2537 NULL, 0); 2538 ret = -EINVAL; 2539 trace_icm_staleconn_err( 2540 IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg), 2541 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)); 2542 2543 if (cur_cm_id_priv) { 2544 ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0); 2545 cm_deref_id(cur_cm_id_priv); 2546 } 2547 2548 goto error; 2549 } 2550 spin_unlock(&cm.lock); 2551 2552 cm_id_priv->id.state = IB_CM_REP_RCVD; 2553 cm_id_priv->id.remote_id = 2554 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)); 2555 cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type); 2556 cm_id_priv->initiator_depth = 2557 IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg); 2558 cm_id_priv->responder_resources = 2559 IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg); 2560 cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg)); 2561 cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg); 2562 cm_id_priv->target_ack_delay = 2563 IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg); 2564 cm_id_priv->av.timeout = 2565 cm_ack_timeout(cm_id_priv->target_ack_delay, 2566 cm_id_priv->av.timeout - 1); 2567 cm_id_priv->alt_av.timeout = 2568 cm_ack_timeout(cm_id_priv->target_ack_delay, 2569 cm_id_priv->alt_av.timeout - 1); 2570 2571 ib_cancel_mad(cm_id_priv->msg); 2572 cm_queue_work_unlock(cm_id_priv, work); 2573 return 0; 2574 2575 error: 2576 cm_deref_id(cm_id_priv); 2577 return ret; 2578 } 2579 2580 static int cm_establish_handler(struct cm_work *work) 2581 { 2582 struct cm_id_private *cm_id_priv; 2583 2584 /* See comment in cm_establish about lookup. */ 2585 cm_id_priv = cm_acquire_id(work->local_id, work->remote_id); 2586 if (!cm_id_priv) 2587 return -EINVAL; 2588 2589 spin_lock_irq(&cm_id_priv->lock); 2590 if (cm_id_priv->id.state != IB_CM_ESTABLISHED) { 2591 spin_unlock_irq(&cm_id_priv->lock); 2592 goto out; 2593 } 2594 2595 ib_cancel_mad(cm_id_priv->msg); 2596 cm_queue_work_unlock(cm_id_priv, work); 2597 return 0; 2598 out: 2599 cm_deref_id(cm_id_priv); 2600 return -EINVAL; 2601 } 2602 2603 static int cm_rtu_handler(struct cm_work *work) 2604 { 2605 struct cm_id_private *cm_id_priv; 2606 struct cm_rtu_msg *rtu_msg; 2607 2608 rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad; 2609 cm_id_priv = cm_acquire_id( 2610 cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)), 2611 cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg))); 2612 if (!cm_id_priv) 2613 return -EINVAL; 2614 2615 work->cm_event.private_data = 2616 IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg); 2617 2618 spin_lock_irq(&cm_id_priv->lock); 2619 if (cm_id_priv->id.state != IB_CM_REP_SENT && 2620 cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) { 2621 spin_unlock_irq(&cm_id_priv->lock); 2622 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2623 [CM_RTU_COUNTER]); 2624 goto out; 2625 } 2626 cm_id_priv->id.state = IB_CM_ESTABLISHED; 2627 2628 ib_cancel_mad(cm_id_priv->msg); 2629 cm_queue_work_unlock(cm_id_priv, work); 2630 return 0; 2631 out: 2632 cm_deref_id(cm_id_priv); 2633 return -EINVAL; 2634 } 2635 2636 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg, 2637 struct cm_id_private *cm_id_priv, 2638 const void *private_data, 2639 u8 private_data_len) 2640 { 2641 cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID, 2642 cm_form_tid(cm_id_priv)); 2643 IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg, 2644 be32_to_cpu(cm_id_priv->id.local_id)); 2645 IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg, 2646 be32_to_cpu(cm_id_priv->id.remote_id)); 2647 IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg, 2648 be32_to_cpu(cm_id_priv->remote_qpn)); 2649 2650 if (private_data && private_data_len) 2651 IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data, 2652 private_data_len); 2653 } 2654 2655 static void cm_issue_dreq(struct cm_id_private *cm_id_priv) 2656 { 2657 struct ib_mad_send_buf *msg; 2658 int ret; 2659 2660 lockdep_assert_held(&cm_id_priv->lock); 2661 2662 msg = cm_alloc_msg(cm_id_priv); 2663 if (IS_ERR(msg)) 2664 return; 2665 2666 cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv, NULL, 0); 2667 2668 trace_icm_send_dreq(&cm_id_priv->id); 2669 ret = ib_post_send_mad(msg, NULL); 2670 if (ret) 2671 cm_free_msg(msg); 2672 } 2673 2674 int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data, 2675 u8 private_data_len) 2676 { 2677 struct cm_id_private *cm_id_priv = 2678 container_of(cm_id, struct cm_id_private, id); 2679 struct ib_mad_send_buf *msg; 2680 unsigned long flags; 2681 int ret; 2682 2683 if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE) 2684 return -EINVAL; 2685 2686 spin_lock_irqsave(&cm_id_priv->lock, flags); 2687 if (cm_id_priv->id.state != IB_CM_ESTABLISHED) { 2688 trace_icm_dreq_skipped(&cm_id_priv->id); 2689 ret = -EINVAL; 2690 goto unlock; 2691 } 2692 2693 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT || 2694 cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD) 2695 ib_cancel_mad(cm_id_priv->msg); 2696 2697 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_DREQ_SENT); 2698 if (IS_ERR(msg)) { 2699 cm_enter_timewait(cm_id_priv); 2700 ret = PTR_ERR(msg); 2701 goto unlock; 2702 } 2703 2704 cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv, 2705 private_data, private_data_len); 2706 2707 trace_icm_send_dreq(&cm_id_priv->id); 2708 ret = ib_post_send_mad(msg, NULL); 2709 if (ret) { 2710 cm_enter_timewait(cm_id_priv); 2711 cm_free_priv_msg(msg); 2712 goto unlock; 2713 } 2714 2715 cm_id_priv->id.state = IB_CM_DREQ_SENT; 2716 unlock: 2717 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2718 return ret; 2719 } 2720 EXPORT_SYMBOL(ib_send_cm_dreq); 2721 2722 static void cm_format_drep(struct cm_drep_msg *drep_msg, 2723 struct cm_id_private *cm_id_priv, 2724 const void *private_data, 2725 u8 private_data_len) 2726 { 2727 cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid); 2728 IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg, 2729 be32_to_cpu(cm_id_priv->id.local_id)); 2730 IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg, 2731 be32_to_cpu(cm_id_priv->id.remote_id)); 2732 2733 if (private_data && private_data_len) 2734 IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data, 2735 private_data_len); 2736 } 2737 2738 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv, 2739 void *private_data, u8 private_data_len) 2740 { 2741 struct ib_mad_send_buf *msg; 2742 int ret; 2743 2744 lockdep_assert_held(&cm_id_priv->lock); 2745 2746 if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE) 2747 return -EINVAL; 2748 2749 if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) { 2750 trace_icm_send_drep_err(&cm_id_priv->id); 2751 kfree(private_data); 2752 return -EINVAL; 2753 } 2754 2755 cm_set_private_data(cm_id_priv, private_data, private_data_len); 2756 cm_enter_timewait(cm_id_priv); 2757 2758 msg = cm_alloc_msg(cm_id_priv); 2759 if (IS_ERR(msg)) 2760 return PTR_ERR(msg); 2761 2762 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv, 2763 private_data, private_data_len); 2764 2765 trace_icm_send_drep(&cm_id_priv->id); 2766 ret = ib_post_send_mad(msg, NULL); 2767 if (ret) { 2768 cm_free_msg(msg); 2769 return ret; 2770 } 2771 return 0; 2772 } 2773 2774 int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data, 2775 u8 private_data_len) 2776 { 2777 struct cm_id_private *cm_id_priv = 2778 container_of(cm_id, struct cm_id_private, id); 2779 unsigned long flags; 2780 void *data; 2781 int ret; 2782 2783 data = cm_copy_private_data(private_data, private_data_len); 2784 if (IS_ERR(data)) 2785 return PTR_ERR(data); 2786 2787 spin_lock_irqsave(&cm_id_priv->lock, flags); 2788 ret = cm_send_drep_locked(cm_id_priv, data, private_data_len); 2789 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2790 return ret; 2791 } 2792 EXPORT_SYMBOL(ib_send_cm_drep); 2793 2794 static int cm_issue_drep(struct cm_port *port, 2795 struct ib_mad_recv_wc *mad_recv_wc) 2796 { 2797 struct ib_mad_send_buf *msg = NULL; 2798 struct cm_dreq_msg *dreq_msg; 2799 struct cm_drep_msg *drep_msg; 2800 int ret; 2801 2802 ret = cm_alloc_response_msg(port, mad_recv_wc, true, &msg); 2803 if (ret) 2804 return ret; 2805 2806 dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad; 2807 drep_msg = (struct cm_drep_msg *) msg->mad; 2808 2809 cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid); 2810 IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg, 2811 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)); 2812 IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg, 2813 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)); 2814 2815 trace_icm_issue_drep( 2816 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg), 2817 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)); 2818 ret = ib_post_send_mad(msg, NULL); 2819 if (ret) 2820 cm_free_msg(msg); 2821 2822 return ret; 2823 } 2824 2825 static int cm_dreq_handler(struct cm_work *work) 2826 { 2827 struct cm_id_private *cm_id_priv; 2828 struct cm_dreq_msg *dreq_msg; 2829 struct ib_mad_send_buf *msg = NULL; 2830 2831 dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad; 2832 cm_id_priv = cm_acquire_id( 2833 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)), 2834 cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg))); 2835 if (!cm_id_priv) { 2836 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2837 [CM_DREQ_COUNTER]); 2838 cm_issue_drep(work->port, work->mad_recv_wc); 2839 trace_icm_no_priv_err( 2840 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg), 2841 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)); 2842 return -EINVAL; 2843 } 2844 2845 work->cm_event.private_data = 2846 IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg); 2847 2848 spin_lock_irq(&cm_id_priv->lock); 2849 if (cm_id_priv->local_qpn != 2850 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg))) 2851 goto unlock; 2852 2853 switch (cm_id_priv->id.state) { 2854 case IB_CM_REP_SENT: 2855 case IB_CM_DREQ_SENT: 2856 case IB_CM_MRA_REP_RCVD: 2857 ib_cancel_mad(cm_id_priv->msg); 2858 break; 2859 case IB_CM_ESTABLISHED: 2860 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT || 2861 cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD) 2862 ib_cancel_mad(cm_id_priv->msg); 2863 break; 2864 case IB_CM_TIMEWAIT: 2865 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2866 [CM_DREQ_COUNTER]); 2867 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc, 2868 true); 2869 if (IS_ERR(msg)) 2870 goto unlock; 2871 2872 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv, 2873 cm_id_priv->private_data, 2874 cm_id_priv->private_data_len); 2875 spin_unlock_irq(&cm_id_priv->lock); 2876 2877 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) || 2878 ib_post_send_mad(msg, NULL)) 2879 cm_free_msg(msg); 2880 goto deref; 2881 case IB_CM_DREQ_RCVD: 2882 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2883 [CM_DREQ_COUNTER]); 2884 goto unlock; 2885 default: 2886 trace_icm_dreq_unknown_err(&cm_id_priv->id); 2887 goto unlock; 2888 } 2889 cm_id_priv->id.state = IB_CM_DREQ_RCVD; 2890 cm_id_priv->tid = dreq_msg->hdr.tid; 2891 cm_queue_work_unlock(cm_id_priv, work); 2892 return 0; 2893 2894 unlock: spin_unlock_irq(&cm_id_priv->lock); 2895 deref: cm_deref_id(cm_id_priv); 2896 return -EINVAL; 2897 } 2898 2899 static int cm_drep_handler(struct cm_work *work) 2900 { 2901 struct cm_id_private *cm_id_priv; 2902 struct cm_drep_msg *drep_msg; 2903 2904 drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad; 2905 cm_id_priv = cm_acquire_id( 2906 cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)), 2907 cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg))); 2908 if (!cm_id_priv) 2909 return -EINVAL; 2910 2911 work->cm_event.private_data = 2912 IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg); 2913 2914 spin_lock_irq(&cm_id_priv->lock); 2915 if (cm_id_priv->id.state != IB_CM_DREQ_SENT && 2916 cm_id_priv->id.state != IB_CM_DREQ_RCVD) { 2917 spin_unlock_irq(&cm_id_priv->lock); 2918 goto out; 2919 } 2920 cm_enter_timewait(cm_id_priv); 2921 2922 ib_cancel_mad(cm_id_priv->msg); 2923 cm_queue_work_unlock(cm_id_priv, work); 2924 return 0; 2925 out: 2926 cm_deref_id(cm_id_priv); 2927 return -EINVAL; 2928 } 2929 2930 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv, 2931 enum ib_cm_rej_reason reason, void *ari, 2932 u8 ari_length, const void *private_data, 2933 u8 private_data_len) 2934 { 2935 enum ib_cm_state state = cm_id_priv->id.state; 2936 struct ib_mad_send_buf *msg; 2937 int ret; 2938 2939 lockdep_assert_held(&cm_id_priv->lock); 2940 2941 if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) || 2942 (ari && ari_length > IB_CM_REJ_ARI_LENGTH)) 2943 return -EINVAL; 2944 2945 trace_icm_send_rej(&cm_id_priv->id, reason); 2946 2947 switch (state) { 2948 case IB_CM_REQ_SENT: 2949 case IB_CM_MRA_REQ_RCVD: 2950 case IB_CM_REQ_RCVD: 2951 case IB_CM_MRA_REQ_SENT: 2952 case IB_CM_REP_RCVD: 2953 case IB_CM_MRA_REP_SENT: 2954 cm_reset_to_idle(cm_id_priv); 2955 msg = cm_alloc_msg(cm_id_priv); 2956 if (IS_ERR(msg)) 2957 return PTR_ERR(msg); 2958 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason, 2959 ari, ari_length, private_data, private_data_len, 2960 state); 2961 break; 2962 case IB_CM_REP_SENT: 2963 case IB_CM_MRA_REP_RCVD: 2964 cm_enter_timewait(cm_id_priv); 2965 msg = cm_alloc_msg(cm_id_priv); 2966 if (IS_ERR(msg)) 2967 return PTR_ERR(msg); 2968 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason, 2969 ari, ari_length, private_data, private_data_len, 2970 state); 2971 break; 2972 default: 2973 trace_icm_send_unknown_rej_err(&cm_id_priv->id); 2974 return -EINVAL; 2975 } 2976 2977 ret = ib_post_send_mad(msg, NULL); 2978 if (ret) { 2979 cm_free_msg(msg); 2980 return ret; 2981 } 2982 2983 return 0; 2984 } 2985 2986 int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason, 2987 void *ari, u8 ari_length, const void *private_data, 2988 u8 private_data_len) 2989 { 2990 struct cm_id_private *cm_id_priv = 2991 container_of(cm_id, struct cm_id_private, id); 2992 unsigned long flags; 2993 int ret; 2994 2995 spin_lock_irqsave(&cm_id_priv->lock, flags); 2996 ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length, 2997 private_data, private_data_len); 2998 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2999 return ret; 3000 } 3001 EXPORT_SYMBOL(ib_send_cm_rej); 3002 3003 static void cm_format_rej_event(struct cm_work *work) 3004 { 3005 struct cm_rej_msg *rej_msg; 3006 struct ib_cm_rej_event_param *param; 3007 3008 rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad; 3009 param = &work->cm_event.param.rej_rcvd; 3010 param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg); 3011 param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg); 3012 param->reason = IBA_GET(CM_REJ_REASON, rej_msg); 3013 work->cm_event.private_data = 3014 IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg); 3015 } 3016 3017 static struct cm_id_private *cm_acquire_rejected_id(struct cm_rej_msg *rej_msg) 3018 { 3019 struct cm_id_private *cm_id_priv; 3020 __be32 remote_id; 3021 3022 remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg)); 3023 3024 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) { 3025 cm_id_priv = cm_find_remote_id( 3026 *((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)), 3027 remote_id); 3028 } else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) == 3029 CM_MSG_RESPONSE_REQ) 3030 cm_id_priv = cm_acquire_id( 3031 cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)), 3032 0); 3033 else 3034 cm_id_priv = cm_acquire_id( 3035 cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)), 3036 remote_id); 3037 3038 return cm_id_priv; 3039 } 3040 3041 static int cm_rej_handler(struct cm_work *work) 3042 { 3043 struct cm_id_private *cm_id_priv; 3044 struct cm_rej_msg *rej_msg; 3045 3046 rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad; 3047 cm_id_priv = cm_acquire_rejected_id(rej_msg); 3048 if (!cm_id_priv) 3049 return -EINVAL; 3050 3051 cm_format_rej_event(work); 3052 3053 spin_lock_irq(&cm_id_priv->lock); 3054 switch (cm_id_priv->id.state) { 3055 case IB_CM_REQ_SENT: 3056 case IB_CM_MRA_REQ_RCVD: 3057 case IB_CM_REP_SENT: 3058 case IB_CM_MRA_REP_RCVD: 3059 ib_cancel_mad(cm_id_priv->msg); 3060 fallthrough; 3061 case IB_CM_REQ_RCVD: 3062 case IB_CM_MRA_REQ_SENT: 3063 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN) 3064 cm_enter_timewait(cm_id_priv); 3065 else 3066 cm_reset_to_idle(cm_id_priv); 3067 break; 3068 case IB_CM_DREQ_SENT: 3069 ib_cancel_mad(cm_id_priv->msg); 3070 fallthrough; 3071 case IB_CM_REP_RCVD: 3072 case IB_CM_MRA_REP_SENT: 3073 cm_enter_timewait(cm_id_priv); 3074 break; 3075 case IB_CM_ESTABLISHED: 3076 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT || 3077 cm_id_priv->id.lap_state == IB_CM_LAP_SENT) { 3078 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT) 3079 ib_cancel_mad(cm_id_priv->msg); 3080 cm_enter_timewait(cm_id_priv); 3081 break; 3082 } 3083 fallthrough; 3084 default: 3085 trace_icm_rej_unknown_err(&cm_id_priv->id); 3086 spin_unlock_irq(&cm_id_priv->lock); 3087 goto out; 3088 } 3089 3090 cm_queue_work_unlock(cm_id_priv, work); 3091 return 0; 3092 out: 3093 cm_deref_id(cm_id_priv); 3094 return -EINVAL; 3095 } 3096 3097 int ib_send_cm_mra(struct ib_cm_id *cm_id, 3098 u8 service_timeout, 3099 const void *private_data, 3100 u8 private_data_len) 3101 { 3102 struct cm_id_private *cm_id_priv; 3103 struct ib_mad_send_buf *msg; 3104 enum ib_cm_state cm_state; 3105 enum ib_cm_lap_state lap_state; 3106 enum cm_msg_response msg_response; 3107 void *data; 3108 unsigned long flags; 3109 int ret; 3110 3111 if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE) 3112 return -EINVAL; 3113 3114 data = cm_copy_private_data(private_data, private_data_len); 3115 if (IS_ERR(data)) 3116 return PTR_ERR(data); 3117 3118 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3119 3120 spin_lock_irqsave(&cm_id_priv->lock, flags); 3121 switch (cm_id_priv->id.state) { 3122 case IB_CM_REQ_RCVD: 3123 cm_state = IB_CM_MRA_REQ_SENT; 3124 lap_state = cm_id->lap_state; 3125 msg_response = CM_MSG_RESPONSE_REQ; 3126 break; 3127 case IB_CM_REP_RCVD: 3128 cm_state = IB_CM_MRA_REP_SENT; 3129 lap_state = cm_id->lap_state; 3130 msg_response = CM_MSG_RESPONSE_REP; 3131 break; 3132 case IB_CM_ESTABLISHED: 3133 if (cm_id->lap_state == IB_CM_LAP_RCVD) { 3134 cm_state = cm_id->state; 3135 lap_state = IB_CM_MRA_LAP_SENT; 3136 msg_response = CM_MSG_RESPONSE_OTHER; 3137 break; 3138 } 3139 fallthrough; 3140 default: 3141 trace_icm_send_mra_unknown_err(&cm_id_priv->id); 3142 ret = -EINVAL; 3143 goto error_unlock; 3144 } 3145 3146 if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) { 3147 msg = cm_alloc_msg(cm_id_priv); 3148 if (IS_ERR(msg)) { 3149 ret = PTR_ERR(msg); 3150 goto error_unlock; 3151 } 3152 3153 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 3154 msg_response, service_timeout, 3155 private_data, private_data_len); 3156 trace_icm_send_mra(cm_id); 3157 ret = ib_post_send_mad(msg, NULL); 3158 if (ret) 3159 goto error_free_msg; 3160 } 3161 3162 cm_id->state = cm_state; 3163 cm_id->lap_state = lap_state; 3164 cm_id_priv->service_timeout = service_timeout; 3165 cm_set_private_data(cm_id_priv, data, private_data_len); 3166 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3167 return 0; 3168 3169 error_free_msg: 3170 cm_free_msg(msg); 3171 error_unlock: 3172 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3173 kfree(data); 3174 return ret; 3175 } 3176 EXPORT_SYMBOL(ib_send_cm_mra); 3177 3178 static struct cm_id_private *cm_acquire_mraed_id(struct cm_mra_msg *mra_msg) 3179 { 3180 switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) { 3181 case CM_MSG_RESPONSE_REQ: 3182 return cm_acquire_id( 3183 cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)), 3184 0); 3185 case CM_MSG_RESPONSE_REP: 3186 case CM_MSG_RESPONSE_OTHER: 3187 return cm_acquire_id( 3188 cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)), 3189 cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg))); 3190 default: 3191 return NULL; 3192 } 3193 } 3194 3195 static int cm_mra_handler(struct cm_work *work) 3196 { 3197 struct cm_id_private *cm_id_priv; 3198 struct cm_mra_msg *mra_msg; 3199 int timeout; 3200 3201 mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad; 3202 cm_id_priv = cm_acquire_mraed_id(mra_msg); 3203 if (!cm_id_priv) 3204 return -EINVAL; 3205 3206 work->cm_event.private_data = 3207 IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg); 3208 work->cm_event.param.mra_rcvd.service_timeout = 3209 IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg); 3210 timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) + 3211 cm_convert_to_ms(cm_id_priv->av.timeout); 3212 3213 spin_lock_irq(&cm_id_priv->lock); 3214 switch (cm_id_priv->id.state) { 3215 case IB_CM_REQ_SENT: 3216 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) != 3217 CM_MSG_RESPONSE_REQ || 3218 ib_modify_mad(cm_id_priv->msg, timeout)) 3219 goto out; 3220 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD; 3221 break; 3222 case IB_CM_REP_SENT: 3223 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) != 3224 CM_MSG_RESPONSE_REP || 3225 ib_modify_mad(cm_id_priv->msg, timeout)) 3226 goto out; 3227 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD; 3228 break; 3229 case IB_CM_ESTABLISHED: 3230 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) != 3231 CM_MSG_RESPONSE_OTHER || 3232 cm_id_priv->id.lap_state != IB_CM_LAP_SENT || 3233 ib_modify_mad(cm_id_priv->msg, timeout)) { 3234 if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD) 3235 atomic_long_inc( 3236 &work->port->counters[CM_RECV_DUPLICATES] 3237 [CM_MRA_COUNTER]); 3238 goto out; 3239 } 3240 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD; 3241 break; 3242 case IB_CM_MRA_REQ_RCVD: 3243 case IB_CM_MRA_REP_RCVD: 3244 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3245 [CM_MRA_COUNTER]); 3246 fallthrough; 3247 default: 3248 trace_icm_mra_unknown_err(&cm_id_priv->id); 3249 goto out; 3250 } 3251 3252 cm_id_priv->msg->context[1] = (void *) (unsigned long) 3253 cm_id_priv->id.state; 3254 cm_queue_work_unlock(cm_id_priv, work); 3255 return 0; 3256 out: 3257 spin_unlock_irq(&cm_id_priv->lock); 3258 cm_deref_id(cm_id_priv); 3259 return -EINVAL; 3260 } 3261 3262 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg, 3263 struct sa_path_rec *path) 3264 { 3265 u32 lid; 3266 3267 if (path->rec_type != SA_PATH_REC_TYPE_OPA) { 3268 sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID, 3269 lap_msg)); 3270 sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID, 3271 lap_msg)); 3272 } else { 3273 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 3274 CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg)); 3275 sa_path_set_dlid(path, lid); 3276 3277 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 3278 CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg)); 3279 sa_path_set_slid(path, lid); 3280 } 3281 } 3282 3283 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv, 3284 struct sa_path_rec *path, 3285 struct cm_lap_msg *lap_msg) 3286 { 3287 path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg); 3288 path->sgid = 3289 *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg); 3290 path->flow_label = 3291 cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg)); 3292 path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg); 3293 path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg); 3294 path->reversible = 1; 3295 path->pkey = cm_id_priv->pkey; 3296 path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg); 3297 path->mtu_selector = IB_SA_EQ; 3298 path->mtu = cm_id_priv->path_mtu; 3299 path->rate_selector = IB_SA_EQ; 3300 path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg); 3301 path->packet_life_time_selector = IB_SA_EQ; 3302 path->packet_life_time = 3303 IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg); 3304 path->packet_life_time -= (path->packet_life_time > 0); 3305 cm_format_path_lid_from_lap(lap_msg, path); 3306 } 3307 3308 static int cm_lap_handler(struct cm_work *work) 3309 { 3310 struct cm_id_private *cm_id_priv; 3311 struct cm_lap_msg *lap_msg; 3312 struct ib_cm_lap_event_param *param; 3313 struct ib_mad_send_buf *msg = NULL; 3314 struct rdma_ah_attr ah_attr; 3315 struct cm_av alt_av = {}; 3316 int ret; 3317 3318 /* Currently Alternate path messages are not supported for 3319 * RoCE link layer. 3320 */ 3321 if (rdma_protocol_roce(work->port->cm_dev->ib_device, 3322 work->port->port_num)) 3323 return -EINVAL; 3324 3325 /* todo: verify LAP request and send reject APR if invalid. */ 3326 lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad; 3327 cm_id_priv = cm_acquire_id( 3328 cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)), 3329 cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg))); 3330 if (!cm_id_priv) 3331 return -EINVAL; 3332 3333 param = &work->cm_event.param.lap_rcvd; 3334 memset(&work->path[0], 0, sizeof(work->path[1])); 3335 cm_path_set_rec_type(work->port->cm_dev->ib_device, 3336 work->port->port_num, &work->path[0], 3337 IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, 3338 lap_msg)); 3339 param->alternate_path = &work->path[0]; 3340 cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg); 3341 work->cm_event.private_data = 3342 IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg); 3343 3344 ret = ib_init_ah_attr_from_wc(work->port->cm_dev->ib_device, 3345 work->port->port_num, 3346 work->mad_recv_wc->wc, 3347 work->mad_recv_wc->recv_buf.grh, 3348 &ah_attr); 3349 if (ret) 3350 goto deref; 3351 3352 ret = cm_init_av_by_path(param->alternate_path, NULL, &alt_av); 3353 if (ret) { 3354 rdma_destroy_ah_attr(&ah_attr); 3355 goto deref; 3356 } 3357 3358 spin_lock_irq(&cm_id_priv->lock); 3359 cm_init_av_for_lap(work->port, work->mad_recv_wc->wc, 3360 &ah_attr, &cm_id_priv->av); 3361 cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av); 3362 3363 if (cm_id_priv->id.state != IB_CM_ESTABLISHED) 3364 goto unlock; 3365 3366 switch (cm_id_priv->id.lap_state) { 3367 case IB_CM_LAP_UNINIT: 3368 case IB_CM_LAP_IDLE: 3369 break; 3370 case IB_CM_MRA_LAP_SENT: 3371 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3372 [CM_LAP_COUNTER]); 3373 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc, 3374 true); 3375 if (IS_ERR(msg)) 3376 goto unlock; 3377 3378 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 3379 CM_MSG_RESPONSE_OTHER, 3380 cm_id_priv->service_timeout, 3381 cm_id_priv->private_data, 3382 cm_id_priv->private_data_len); 3383 spin_unlock_irq(&cm_id_priv->lock); 3384 3385 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) || 3386 ib_post_send_mad(msg, NULL)) 3387 cm_free_msg(msg); 3388 goto deref; 3389 case IB_CM_LAP_RCVD: 3390 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3391 [CM_LAP_COUNTER]); 3392 goto unlock; 3393 default: 3394 goto unlock; 3395 } 3396 3397 cm_id_priv->id.lap_state = IB_CM_LAP_RCVD; 3398 cm_id_priv->tid = lap_msg->hdr.tid; 3399 cm_queue_work_unlock(cm_id_priv, work); 3400 return 0; 3401 3402 unlock: spin_unlock_irq(&cm_id_priv->lock); 3403 deref: cm_deref_id(cm_id_priv); 3404 return -EINVAL; 3405 } 3406 3407 static int cm_apr_handler(struct cm_work *work) 3408 { 3409 struct cm_id_private *cm_id_priv; 3410 struct cm_apr_msg *apr_msg; 3411 3412 /* Currently Alternate path messages are not supported for 3413 * RoCE link layer. 3414 */ 3415 if (rdma_protocol_roce(work->port->cm_dev->ib_device, 3416 work->port->port_num)) 3417 return -EINVAL; 3418 3419 apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad; 3420 cm_id_priv = cm_acquire_id( 3421 cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)), 3422 cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg))); 3423 if (!cm_id_priv) 3424 return -EINVAL; /* Unmatched reply. */ 3425 3426 work->cm_event.param.apr_rcvd.ap_status = 3427 IBA_GET(CM_APR_AR_STATUS, apr_msg); 3428 work->cm_event.param.apr_rcvd.apr_info = 3429 IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg); 3430 work->cm_event.param.apr_rcvd.info_len = 3431 IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg); 3432 work->cm_event.private_data = 3433 IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg); 3434 3435 spin_lock_irq(&cm_id_priv->lock); 3436 if (cm_id_priv->id.state != IB_CM_ESTABLISHED || 3437 (cm_id_priv->id.lap_state != IB_CM_LAP_SENT && 3438 cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) { 3439 spin_unlock_irq(&cm_id_priv->lock); 3440 goto out; 3441 } 3442 cm_id_priv->id.lap_state = IB_CM_LAP_IDLE; 3443 ib_cancel_mad(cm_id_priv->msg); 3444 cm_queue_work_unlock(cm_id_priv, work); 3445 return 0; 3446 out: 3447 cm_deref_id(cm_id_priv); 3448 return -EINVAL; 3449 } 3450 3451 static int cm_timewait_handler(struct cm_work *work) 3452 { 3453 struct cm_timewait_info *timewait_info; 3454 struct cm_id_private *cm_id_priv; 3455 3456 timewait_info = container_of(work, struct cm_timewait_info, work); 3457 spin_lock_irq(&cm.lock); 3458 list_del(&timewait_info->list); 3459 spin_unlock_irq(&cm.lock); 3460 3461 cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 3462 timewait_info->work.remote_id); 3463 if (!cm_id_priv) 3464 return -EINVAL; 3465 3466 spin_lock_irq(&cm_id_priv->lock); 3467 if (cm_id_priv->id.state != IB_CM_TIMEWAIT || 3468 cm_id_priv->remote_qpn != timewait_info->remote_qpn) { 3469 spin_unlock_irq(&cm_id_priv->lock); 3470 goto out; 3471 } 3472 cm_id_priv->id.state = IB_CM_IDLE; 3473 cm_queue_work_unlock(cm_id_priv, work); 3474 return 0; 3475 out: 3476 cm_deref_id(cm_id_priv); 3477 return -EINVAL; 3478 } 3479 3480 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg, 3481 struct cm_id_private *cm_id_priv, 3482 struct ib_cm_sidr_req_param *param) 3483 { 3484 cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID, 3485 cm_form_tid(cm_id_priv)); 3486 IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg, 3487 be32_to_cpu(cm_id_priv->id.local_id)); 3488 IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg, 3489 be16_to_cpu(param->path->pkey)); 3490 IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg, 3491 be64_to_cpu(param->service_id)); 3492 3493 if (param->private_data && param->private_data_len) 3494 IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg, 3495 param->private_data, param->private_data_len); 3496 } 3497 3498 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id, 3499 struct ib_cm_sidr_req_param *param) 3500 { 3501 struct cm_id_private *cm_id_priv; 3502 struct ib_mad_send_buf *msg; 3503 struct cm_av av = {}; 3504 unsigned long flags; 3505 int ret; 3506 3507 if (!param->path || (param->private_data && 3508 param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE)) 3509 return -EINVAL; 3510 3511 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3512 ret = cm_init_av_by_path(param->path, param->sgid_attr, &av); 3513 if (ret) 3514 return ret; 3515 3516 spin_lock_irqsave(&cm_id_priv->lock, flags); 3517 cm_move_av_from_path(&cm_id_priv->av, &av); 3518 cm_id->service_id = param->service_id; 3519 cm_id_priv->timeout_ms = param->timeout_ms; 3520 cm_id_priv->max_cm_retries = param->max_cm_retries; 3521 if (cm_id->state != IB_CM_IDLE) { 3522 ret = -EINVAL; 3523 goto out_unlock; 3524 } 3525 3526 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_SIDR_REQ_SENT); 3527 if (IS_ERR(msg)) { 3528 ret = PTR_ERR(msg); 3529 goto out_unlock; 3530 } 3531 3532 cm_format_sidr_req((struct cm_sidr_req_msg *)msg->mad, cm_id_priv, 3533 param); 3534 3535 trace_icm_send_sidr_req(&cm_id_priv->id); 3536 ret = ib_post_send_mad(msg, NULL); 3537 if (ret) 3538 goto out_free; 3539 cm_id->state = IB_CM_SIDR_REQ_SENT; 3540 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3541 return 0; 3542 out_free: 3543 cm_free_priv_msg(msg); 3544 out_unlock: 3545 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3546 return ret; 3547 } 3548 EXPORT_SYMBOL(ib_send_cm_sidr_req); 3549 3550 static void cm_format_sidr_req_event(struct cm_work *work, 3551 const struct cm_id_private *rx_cm_id, 3552 struct ib_cm_id *listen_id) 3553 { 3554 struct cm_sidr_req_msg *sidr_req_msg; 3555 struct ib_cm_sidr_req_event_param *param; 3556 3557 sidr_req_msg = (struct cm_sidr_req_msg *) 3558 work->mad_recv_wc->recv_buf.mad; 3559 param = &work->cm_event.param.sidr_req_rcvd; 3560 param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg); 3561 param->listen_id = listen_id; 3562 param->service_id = 3563 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg)); 3564 param->bth_pkey = cm_get_bth_pkey(work); 3565 param->port = work->port->port_num; 3566 param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr; 3567 work->cm_event.private_data = 3568 IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg); 3569 } 3570 3571 static int cm_sidr_req_handler(struct cm_work *work) 3572 { 3573 struct cm_id_private *cm_id_priv, *listen_cm_id_priv; 3574 struct cm_sidr_req_msg *sidr_req_msg; 3575 struct ib_wc *wc; 3576 int ret; 3577 3578 cm_id_priv = 3579 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL); 3580 if (IS_ERR(cm_id_priv)) 3581 return PTR_ERR(cm_id_priv); 3582 3583 /* Record SGID/SLID and request ID for lookup. */ 3584 sidr_req_msg = (struct cm_sidr_req_msg *) 3585 work->mad_recv_wc->recv_buf.mad; 3586 3587 cm_id_priv->id.remote_id = 3588 cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg)); 3589 cm_id_priv->id.service_id = 3590 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg)); 3591 cm_id_priv->tid = sidr_req_msg->hdr.tid; 3592 3593 wc = work->mad_recv_wc->wc; 3594 cm_id_priv->sidr_slid = wc->slid; 3595 ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc, 3596 work->mad_recv_wc->recv_buf.grh, 3597 &cm_id_priv->av); 3598 if (ret) 3599 goto out; 3600 3601 spin_lock_irq(&cm.lock); 3602 listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv); 3603 if (listen_cm_id_priv) { 3604 spin_unlock_irq(&cm.lock); 3605 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3606 [CM_SIDR_REQ_COUNTER]); 3607 goto out; /* Duplicate message. */ 3608 } 3609 cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD; 3610 listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device, 3611 cm_id_priv->id.service_id); 3612 if (!listen_cm_id_priv) { 3613 spin_unlock_irq(&cm.lock); 3614 ib_send_cm_sidr_rep(&cm_id_priv->id, 3615 &(struct ib_cm_sidr_rep_param){ 3616 .status = IB_SIDR_UNSUPPORTED }); 3617 goto out; /* No match. */ 3618 } 3619 spin_unlock_irq(&cm.lock); 3620 3621 cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler; 3622 cm_id_priv->id.context = listen_cm_id_priv->id.context; 3623 3624 /* 3625 * A SIDR ID does not need to be in the xarray since it does not receive 3626 * mads, is not placed in the remote_id or remote_qpn rbtree, and does 3627 * not enter timewait. 3628 */ 3629 3630 cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id); 3631 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event); 3632 cm_free_work(work); 3633 /* 3634 * A pointer to the listen_cm_id is held in the event, so this deref 3635 * must be after the event is delivered above. 3636 */ 3637 cm_deref_id(listen_cm_id_priv); 3638 if (ret) 3639 cm_destroy_id(&cm_id_priv->id, ret); 3640 return 0; 3641 out: 3642 ib_destroy_cm_id(&cm_id_priv->id); 3643 return -EINVAL; 3644 } 3645 3646 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg, 3647 struct cm_id_private *cm_id_priv, 3648 struct ib_cm_sidr_rep_param *param) 3649 { 3650 cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID, 3651 cm_id_priv->tid, param->ece.attr_mod); 3652 IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg, 3653 be32_to_cpu(cm_id_priv->id.remote_id)); 3654 IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status); 3655 IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num); 3656 IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg, 3657 be64_to_cpu(cm_id_priv->id.service_id)); 3658 IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey); 3659 IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg, 3660 param->ece.vendor_id & 0xFF); 3661 IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg, 3662 (param->ece.vendor_id >> 8) & 0xFF); 3663 3664 if (param->info && param->info_length) 3665 IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg, 3666 param->info, param->info_length); 3667 3668 if (param->private_data && param->private_data_len) 3669 IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg, 3670 param->private_data, param->private_data_len); 3671 } 3672 3673 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv, 3674 struct ib_cm_sidr_rep_param *param) 3675 { 3676 struct ib_mad_send_buf *msg; 3677 unsigned long flags; 3678 int ret; 3679 3680 lockdep_assert_held(&cm_id_priv->lock); 3681 3682 if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) || 3683 (param->private_data && 3684 param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE)) 3685 return -EINVAL; 3686 3687 if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD) 3688 return -EINVAL; 3689 3690 msg = cm_alloc_msg(cm_id_priv); 3691 if (IS_ERR(msg)) 3692 return PTR_ERR(msg); 3693 3694 cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv, 3695 param); 3696 trace_icm_send_sidr_rep(&cm_id_priv->id); 3697 ret = ib_post_send_mad(msg, NULL); 3698 if (ret) { 3699 cm_free_msg(msg); 3700 return ret; 3701 } 3702 cm_id_priv->id.state = IB_CM_IDLE; 3703 spin_lock_irqsave(&cm.lock, flags); 3704 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) { 3705 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table); 3706 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node); 3707 } 3708 spin_unlock_irqrestore(&cm.lock, flags); 3709 return 0; 3710 } 3711 3712 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id, 3713 struct ib_cm_sidr_rep_param *param) 3714 { 3715 struct cm_id_private *cm_id_priv = 3716 container_of(cm_id, struct cm_id_private, id); 3717 unsigned long flags; 3718 int ret; 3719 3720 spin_lock_irqsave(&cm_id_priv->lock, flags); 3721 ret = cm_send_sidr_rep_locked(cm_id_priv, param); 3722 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3723 return ret; 3724 } 3725 EXPORT_SYMBOL(ib_send_cm_sidr_rep); 3726 3727 static void cm_format_sidr_rep_event(struct cm_work *work, 3728 const struct cm_id_private *cm_id_priv) 3729 { 3730 struct cm_sidr_rep_msg *sidr_rep_msg; 3731 struct ib_cm_sidr_rep_event_param *param; 3732 3733 sidr_rep_msg = (struct cm_sidr_rep_msg *) 3734 work->mad_recv_wc->recv_buf.mad; 3735 param = &work->cm_event.param.sidr_rep_rcvd; 3736 param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg); 3737 param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg); 3738 param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg); 3739 param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION, 3740 sidr_rep_msg); 3741 param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH, 3742 sidr_rep_msg); 3743 param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr; 3744 work->cm_event.private_data = 3745 IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg); 3746 } 3747 3748 static int cm_sidr_rep_handler(struct cm_work *work) 3749 { 3750 struct cm_sidr_rep_msg *sidr_rep_msg; 3751 struct cm_id_private *cm_id_priv; 3752 3753 sidr_rep_msg = (struct cm_sidr_rep_msg *) 3754 work->mad_recv_wc->recv_buf.mad; 3755 cm_id_priv = cm_acquire_id( 3756 cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0); 3757 if (!cm_id_priv) 3758 return -EINVAL; /* Unmatched reply. */ 3759 3760 spin_lock_irq(&cm_id_priv->lock); 3761 if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) { 3762 spin_unlock_irq(&cm_id_priv->lock); 3763 goto out; 3764 } 3765 cm_id_priv->id.state = IB_CM_IDLE; 3766 ib_cancel_mad(cm_id_priv->msg); 3767 spin_unlock_irq(&cm_id_priv->lock); 3768 3769 cm_format_sidr_rep_event(work, cm_id_priv); 3770 cm_process_work(cm_id_priv, work); 3771 return 0; 3772 out: 3773 cm_deref_id(cm_id_priv); 3774 return -EINVAL; 3775 } 3776 3777 static void cm_process_send_error(struct cm_id_private *cm_id_priv, 3778 struct ib_mad_send_buf *msg, 3779 enum ib_wc_status wc_status) 3780 { 3781 enum ib_cm_state state = (unsigned long) msg->context[1]; 3782 struct ib_cm_event cm_event = {}; 3783 int ret; 3784 3785 /* Discard old sends. */ 3786 spin_lock_irq(&cm_id_priv->lock); 3787 if (msg != cm_id_priv->msg) { 3788 spin_unlock_irq(&cm_id_priv->lock); 3789 cm_free_priv_msg(msg); 3790 return; 3791 } 3792 cm_free_priv_msg(msg); 3793 3794 if (state != cm_id_priv->id.state || wc_status == IB_WC_SUCCESS || 3795 wc_status == IB_WC_WR_FLUSH_ERR) 3796 goto out_unlock; 3797 3798 trace_icm_mad_send_err(state, wc_status); 3799 switch (state) { 3800 case IB_CM_REQ_SENT: 3801 case IB_CM_MRA_REQ_RCVD: 3802 cm_reset_to_idle(cm_id_priv); 3803 cm_event.event = IB_CM_REQ_ERROR; 3804 break; 3805 case IB_CM_REP_SENT: 3806 case IB_CM_MRA_REP_RCVD: 3807 cm_reset_to_idle(cm_id_priv); 3808 cm_event.event = IB_CM_REP_ERROR; 3809 break; 3810 case IB_CM_DREQ_SENT: 3811 cm_enter_timewait(cm_id_priv); 3812 cm_event.event = IB_CM_DREQ_ERROR; 3813 break; 3814 case IB_CM_SIDR_REQ_SENT: 3815 cm_id_priv->id.state = IB_CM_IDLE; 3816 cm_event.event = IB_CM_SIDR_REQ_ERROR; 3817 break; 3818 default: 3819 goto out_unlock; 3820 } 3821 spin_unlock_irq(&cm_id_priv->lock); 3822 cm_event.param.send_status = wc_status; 3823 3824 /* No other events can occur on the cm_id at this point. */ 3825 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event); 3826 if (ret) 3827 ib_destroy_cm_id(&cm_id_priv->id); 3828 return; 3829 out_unlock: 3830 spin_unlock_irq(&cm_id_priv->lock); 3831 } 3832 3833 static void cm_send_handler(struct ib_mad_agent *mad_agent, 3834 struct ib_mad_send_wc *mad_send_wc) 3835 { 3836 struct ib_mad_send_buf *msg = mad_send_wc->send_buf; 3837 struct cm_id_private *cm_id_priv; 3838 struct cm_port *port; 3839 u16 attr_index; 3840 3841 port = mad_agent->context; 3842 attr_index = be16_to_cpu(((struct ib_mad_hdr *) 3843 msg->mad)->attr_id) - CM_ATTR_ID_OFFSET; 3844 3845 if (msg->context[0] == CM_DIRECT_RETRY_CTX) { 3846 msg->retries = 1; 3847 cm_id_priv = NULL; 3848 } else { 3849 cm_id_priv = msg->context[0]; 3850 } 3851 3852 atomic_long_add(1 + msg->retries, &port->counters[CM_XMIT][attr_index]); 3853 if (msg->retries) 3854 atomic_long_add(msg->retries, 3855 &port->counters[CM_XMIT_RETRIES][attr_index]); 3856 3857 if (cm_id_priv) 3858 cm_process_send_error(cm_id_priv, msg, mad_send_wc->status); 3859 else 3860 cm_free_msg(msg); 3861 } 3862 3863 static void cm_work_handler(struct work_struct *_work) 3864 { 3865 struct cm_work *work = container_of(_work, struct cm_work, work.work); 3866 int ret; 3867 3868 switch (work->cm_event.event) { 3869 case IB_CM_REQ_RECEIVED: 3870 ret = cm_req_handler(work); 3871 break; 3872 case IB_CM_MRA_RECEIVED: 3873 ret = cm_mra_handler(work); 3874 break; 3875 case IB_CM_REJ_RECEIVED: 3876 ret = cm_rej_handler(work); 3877 break; 3878 case IB_CM_REP_RECEIVED: 3879 ret = cm_rep_handler(work); 3880 break; 3881 case IB_CM_RTU_RECEIVED: 3882 ret = cm_rtu_handler(work); 3883 break; 3884 case IB_CM_USER_ESTABLISHED: 3885 ret = cm_establish_handler(work); 3886 break; 3887 case IB_CM_DREQ_RECEIVED: 3888 ret = cm_dreq_handler(work); 3889 break; 3890 case IB_CM_DREP_RECEIVED: 3891 ret = cm_drep_handler(work); 3892 break; 3893 case IB_CM_SIDR_REQ_RECEIVED: 3894 ret = cm_sidr_req_handler(work); 3895 break; 3896 case IB_CM_SIDR_REP_RECEIVED: 3897 ret = cm_sidr_rep_handler(work); 3898 break; 3899 case IB_CM_LAP_RECEIVED: 3900 ret = cm_lap_handler(work); 3901 break; 3902 case IB_CM_APR_RECEIVED: 3903 ret = cm_apr_handler(work); 3904 break; 3905 case IB_CM_TIMEWAIT_EXIT: 3906 ret = cm_timewait_handler(work); 3907 break; 3908 default: 3909 trace_icm_handler_err(work->cm_event.event); 3910 ret = -EINVAL; 3911 break; 3912 } 3913 if (ret) 3914 cm_free_work(work); 3915 } 3916 3917 static int cm_establish(struct ib_cm_id *cm_id) 3918 { 3919 struct cm_id_private *cm_id_priv; 3920 struct cm_work *work; 3921 unsigned long flags; 3922 int ret = 0; 3923 struct cm_device *cm_dev; 3924 3925 cm_dev = ib_get_client_data(cm_id->device, &cm_client); 3926 if (!cm_dev) 3927 return -ENODEV; 3928 3929 work = kmalloc(sizeof *work, GFP_ATOMIC); 3930 if (!work) 3931 return -ENOMEM; 3932 3933 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3934 spin_lock_irqsave(&cm_id_priv->lock, flags); 3935 switch (cm_id->state) { 3936 case IB_CM_REP_SENT: 3937 case IB_CM_MRA_REP_RCVD: 3938 cm_id->state = IB_CM_ESTABLISHED; 3939 break; 3940 case IB_CM_ESTABLISHED: 3941 ret = -EISCONN; 3942 break; 3943 default: 3944 trace_icm_establish_err(cm_id); 3945 ret = -EINVAL; 3946 break; 3947 } 3948 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3949 3950 if (ret) { 3951 kfree(work); 3952 goto out; 3953 } 3954 3955 /* 3956 * The CM worker thread may try to destroy the cm_id before it 3957 * can execute this work item. To prevent potential deadlock, 3958 * we need to find the cm_id once we're in the context of the 3959 * worker thread, rather than holding a reference on it. 3960 */ 3961 INIT_DELAYED_WORK(&work->work, cm_work_handler); 3962 work->local_id = cm_id->local_id; 3963 work->remote_id = cm_id->remote_id; 3964 work->mad_recv_wc = NULL; 3965 work->cm_event.event = IB_CM_USER_ESTABLISHED; 3966 3967 /* Check if the device started its remove_one */ 3968 spin_lock_irqsave(&cm.lock, flags); 3969 if (!cm_dev->going_down) { 3970 queue_delayed_work(cm.wq, &work->work, 0); 3971 } else { 3972 kfree(work); 3973 ret = -ENODEV; 3974 } 3975 spin_unlock_irqrestore(&cm.lock, flags); 3976 3977 out: 3978 return ret; 3979 } 3980 3981 static int cm_migrate(struct ib_cm_id *cm_id) 3982 { 3983 struct cm_id_private *cm_id_priv; 3984 unsigned long flags; 3985 int ret = 0; 3986 3987 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3988 spin_lock_irqsave(&cm_id_priv->lock, flags); 3989 if (cm_id->state == IB_CM_ESTABLISHED && 3990 (cm_id->lap_state == IB_CM_LAP_UNINIT || 3991 cm_id->lap_state == IB_CM_LAP_IDLE)) { 3992 cm_id->lap_state = IB_CM_LAP_IDLE; 3993 cm_id_priv->av = cm_id_priv->alt_av; 3994 } else 3995 ret = -EINVAL; 3996 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3997 3998 return ret; 3999 } 4000 4001 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event) 4002 { 4003 int ret; 4004 4005 switch (event) { 4006 case IB_EVENT_COMM_EST: 4007 ret = cm_establish(cm_id); 4008 break; 4009 case IB_EVENT_PATH_MIG: 4010 ret = cm_migrate(cm_id); 4011 break; 4012 default: 4013 ret = -EINVAL; 4014 } 4015 return ret; 4016 } 4017 EXPORT_SYMBOL(ib_cm_notify); 4018 4019 static void cm_recv_handler(struct ib_mad_agent *mad_agent, 4020 struct ib_mad_send_buf *send_buf, 4021 struct ib_mad_recv_wc *mad_recv_wc) 4022 { 4023 struct cm_port *port = mad_agent->context; 4024 struct cm_work *work; 4025 enum ib_cm_event_type event; 4026 bool alt_path = false; 4027 u16 attr_id; 4028 int paths = 0; 4029 int going_down = 0; 4030 4031 switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) { 4032 case CM_REQ_ATTR_ID: 4033 alt_path = cm_req_has_alt_path((struct cm_req_msg *) 4034 mad_recv_wc->recv_buf.mad); 4035 paths = 1 + (alt_path != 0); 4036 event = IB_CM_REQ_RECEIVED; 4037 break; 4038 case CM_MRA_ATTR_ID: 4039 event = IB_CM_MRA_RECEIVED; 4040 break; 4041 case CM_REJ_ATTR_ID: 4042 event = IB_CM_REJ_RECEIVED; 4043 break; 4044 case CM_REP_ATTR_ID: 4045 event = IB_CM_REP_RECEIVED; 4046 break; 4047 case CM_RTU_ATTR_ID: 4048 event = IB_CM_RTU_RECEIVED; 4049 break; 4050 case CM_DREQ_ATTR_ID: 4051 event = IB_CM_DREQ_RECEIVED; 4052 break; 4053 case CM_DREP_ATTR_ID: 4054 event = IB_CM_DREP_RECEIVED; 4055 break; 4056 case CM_SIDR_REQ_ATTR_ID: 4057 event = IB_CM_SIDR_REQ_RECEIVED; 4058 break; 4059 case CM_SIDR_REP_ATTR_ID: 4060 event = IB_CM_SIDR_REP_RECEIVED; 4061 break; 4062 case CM_LAP_ATTR_ID: 4063 paths = 1; 4064 event = IB_CM_LAP_RECEIVED; 4065 break; 4066 case CM_APR_ATTR_ID: 4067 event = IB_CM_APR_RECEIVED; 4068 break; 4069 default: 4070 ib_free_recv_mad(mad_recv_wc); 4071 return; 4072 } 4073 4074 attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id); 4075 atomic_long_inc(&port->counters[CM_RECV][attr_id - CM_ATTR_ID_OFFSET]); 4076 4077 work = kmalloc(struct_size(work, path, paths), GFP_KERNEL); 4078 if (!work) { 4079 ib_free_recv_mad(mad_recv_wc); 4080 return; 4081 } 4082 4083 INIT_DELAYED_WORK(&work->work, cm_work_handler); 4084 work->cm_event.event = event; 4085 work->mad_recv_wc = mad_recv_wc; 4086 work->port = port; 4087 4088 /* Check if the device started its remove_one */ 4089 spin_lock_irq(&cm.lock); 4090 if (!port->cm_dev->going_down) 4091 queue_delayed_work(cm.wq, &work->work, 0); 4092 else 4093 going_down = 1; 4094 spin_unlock_irq(&cm.lock); 4095 4096 if (going_down) { 4097 kfree(work); 4098 ib_free_recv_mad(mad_recv_wc); 4099 } 4100 } 4101 4102 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv, 4103 struct ib_qp_attr *qp_attr, 4104 int *qp_attr_mask) 4105 { 4106 unsigned long flags; 4107 int ret; 4108 4109 spin_lock_irqsave(&cm_id_priv->lock, flags); 4110 switch (cm_id_priv->id.state) { 4111 case IB_CM_REQ_SENT: 4112 case IB_CM_MRA_REQ_RCVD: 4113 case IB_CM_REQ_RCVD: 4114 case IB_CM_MRA_REQ_SENT: 4115 case IB_CM_REP_RCVD: 4116 case IB_CM_MRA_REP_SENT: 4117 case IB_CM_REP_SENT: 4118 case IB_CM_MRA_REP_RCVD: 4119 case IB_CM_ESTABLISHED: 4120 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | 4121 IB_QP_PKEY_INDEX | IB_QP_PORT; 4122 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE; 4123 if (cm_id_priv->responder_resources) { 4124 struct ib_device *ib_dev = cm_id_priv->id.device; 4125 u64 support_flush = ib_dev->attrs.device_cap_flags & 4126 (IB_DEVICE_FLUSH_GLOBAL | IB_DEVICE_FLUSH_PERSISTENT); 4127 u32 flushable = support_flush ? 4128 (IB_ACCESS_FLUSH_GLOBAL | 4129 IB_ACCESS_FLUSH_PERSISTENT) : 0; 4130 4131 qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ | 4132 IB_ACCESS_REMOTE_ATOMIC | 4133 flushable; 4134 } 4135 qp_attr->pkey_index = cm_id_priv->av.pkey_index; 4136 if (cm_id_priv->av.port) 4137 qp_attr->port_num = cm_id_priv->av.port->port_num; 4138 ret = 0; 4139 break; 4140 default: 4141 trace_icm_qp_init_err(&cm_id_priv->id); 4142 ret = -EINVAL; 4143 break; 4144 } 4145 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 4146 return ret; 4147 } 4148 4149 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv, 4150 struct ib_qp_attr *qp_attr, 4151 int *qp_attr_mask) 4152 { 4153 unsigned long flags; 4154 int ret; 4155 4156 spin_lock_irqsave(&cm_id_priv->lock, flags); 4157 switch (cm_id_priv->id.state) { 4158 case IB_CM_REQ_RCVD: 4159 case IB_CM_MRA_REQ_SENT: 4160 case IB_CM_REP_RCVD: 4161 case IB_CM_MRA_REP_SENT: 4162 case IB_CM_REP_SENT: 4163 case IB_CM_MRA_REP_RCVD: 4164 case IB_CM_ESTABLISHED: 4165 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | 4166 IB_QP_DEST_QPN | IB_QP_RQ_PSN; 4167 qp_attr->ah_attr = cm_id_priv->av.ah_attr; 4168 if ((qp_attr->ah_attr.type == RDMA_AH_ATTR_TYPE_IB) && 4169 cm_id_priv->av.dlid_datapath && 4170 (cm_id_priv->av.dlid_datapath != 0xffff)) 4171 qp_attr->ah_attr.ib.dlid = cm_id_priv->av.dlid_datapath; 4172 qp_attr->path_mtu = cm_id_priv->path_mtu; 4173 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn); 4174 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn); 4175 if (cm_id_priv->qp_type == IB_QPT_RC || 4176 cm_id_priv->qp_type == IB_QPT_XRC_TGT) { 4177 *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC | 4178 IB_QP_MIN_RNR_TIMER; 4179 qp_attr->max_dest_rd_atomic = 4180 cm_id_priv->responder_resources; 4181 qp_attr->min_rnr_timer = 0; 4182 } 4183 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr) && 4184 cm_id_priv->alt_av.port) { 4185 *qp_attr_mask |= IB_QP_ALT_PATH; 4186 qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num; 4187 qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index; 4188 qp_attr->alt_timeout = cm_id_priv->alt_av.timeout; 4189 qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr; 4190 } 4191 ret = 0; 4192 break; 4193 default: 4194 trace_icm_qp_rtr_err(&cm_id_priv->id); 4195 ret = -EINVAL; 4196 break; 4197 } 4198 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 4199 return ret; 4200 } 4201 4202 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv, 4203 struct ib_qp_attr *qp_attr, 4204 int *qp_attr_mask) 4205 { 4206 unsigned long flags; 4207 int ret; 4208 4209 spin_lock_irqsave(&cm_id_priv->lock, flags); 4210 switch (cm_id_priv->id.state) { 4211 /* Allow transition to RTS before sending REP */ 4212 case IB_CM_REQ_RCVD: 4213 case IB_CM_MRA_REQ_SENT: 4214 4215 case IB_CM_REP_RCVD: 4216 case IB_CM_MRA_REP_SENT: 4217 case IB_CM_REP_SENT: 4218 case IB_CM_MRA_REP_RCVD: 4219 case IB_CM_ESTABLISHED: 4220 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) { 4221 *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN; 4222 qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn); 4223 switch (cm_id_priv->qp_type) { 4224 case IB_QPT_RC: 4225 case IB_QPT_XRC_INI: 4226 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY | 4227 IB_QP_MAX_QP_RD_ATOMIC; 4228 qp_attr->retry_cnt = cm_id_priv->retry_count; 4229 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count; 4230 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth; 4231 fallthrough; 4232 case IB_QPT_XRC_TGT: 4233 *qp_attr_mask |= IB_QP_TIMEOUT; 4234 qp_attr->timeout = cm_id_priv->av.timeout; 4235 break; 4236 default: 4237 break; 4238 } 4239 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) { 4240 *qp_attr_mask |= IB_QP_PATH_MIG_STATE; 4241 qp_attr->path_mig_state = IB_MIG_REARM; 4242 } 4243 } else { 4244 *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE; 4245 if (cm_id_priv->alt_av.port) 4246 qp_attr->alt_port_num = 4247 cm_id_priv->alt_av.port->port_num; 4248 qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index; 4249 qp_attr->alt_timeout = cm_id_priv->alt_av.timeout; 4250 qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr; 4251 qp_attr->path_mig_state = IB_MIG_REARM; 4252 } 4253 ret = 0; 4254 break; 4255 default: 4256 trace_icm_qp_rts_err(&cm_id_priv->id); 4257 ret = -EINVAL; 4258 break; 4259 } 4260 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 4261 return ret; 4262 } 4263 4264 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id, 4265 struct ib_qp_attr *qp_attr, 4266 int *qp_attr_mask) 4267 { 4268 struct cm_id_private *cm_id_priv; 4269 int ret; 4270 4271 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 4272 switch (qp_attr->qp_state) { 4273 case IB_QPS_INIT: 4274 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask); 4275 break; 4276 case IB_QPS_RTR: 4277 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask); 4278 break; 4279 case IB_QPS_RTS: 4280 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask); 4281 break; 4282 default: 4283 ret = -EINVAL; 4284 break; 4285 } 4286 return ret; 4287 } 4288 EXPORT_SYMBOL(ib_cm_init_qp_attr); 4289 4290 static ssize_t cm_show_counter(struct ib_device *ibdev, u32 port_num, 4291 struct ib_port_attribute *attr, char *buf) 4292 { 4293 struct cm_counter_attribute *cm_attr = 4294 container_of(attr, struct cm_counter_attribute, attr); 4295 struct cm_device *cm_dev = ib_get_client_data(ibdev, &cm_client); 4296 4297 if (WARN_ON(!cm_dev)) 4298 return -EINVAL; 4299 4300 return sysfs_emit( 4301 buf, "%ld\n", 4302 atomic_long_read( 4303 &cm_dev->port[port_num - 1] 4304 ->counters[cm_attr->group][cm_attr->index])); 4305 } 4306 4307 #define CM_COUNTER_ATTR(_name, _group, _index) \ 4308 { \ 4309 .attr = __ATTR(_name, 0444, cm_show_counter, NULL), \ 4310 .group = _group, .index = _index \ 4311 } 4312 4313 #define CM_COUNTER_GROUP(_group, _name) \ 4314 static struct cm_counter_attribute cm_counter_attr_##_group[] = { \ 4315 CM_COUNTER_ATTR(req, _group, CM_REQ_COUNTER), \ 4316 CM_COUNTER_ATTR(mra, _group, CM_MRA_COUNTER), \ 4317 CM_COUNTER_ATTR(rej, _group, CM_REJ_COUNTER), \ 4318 CM_COUNTER_ATTR(rep, _group, CM_REP_COUNTER), \ 4319 CM_COUNTER_ATTR(rtu, _group, CM_RTU_COUNTER), \ 4320 CM_COUNTER_ATTR(dreq, _group, CM_DREQ_COUNTER), \ 4321 CM_COUNTER_ATTR(drep, _group, CM_DREP_COUNTER), \ 4322 CM_COUNTER_ATTR(sidr_req, _group, CM_SIDR_REQ_COUNTER), \ 4323 CM_COUNTER_ATTR(sidr_rep, _group, CM_SIDR_REP_COUNTER), \ 4324 CM_COUNTER_ATTR(lap, _group, CM_LAP_COUNTER), \ 4325 CM_COUNTER_ATTR(apr, _group, CM_APR_COUNTER), \ 4326 }; \ 4327 static struct attribute *cm_counter_attrs_##_group[] = { \ 4328 &cm_counter_attr_##_group[0].attr.attr, \ 4329 &cm_counter_attr_##_group[1].attr.attr, \ 4330 &cm_counter_attr_##_group[2].attr.attr, \ 4331 &cm_counter_attr_##_group[3].attr.attr, \ 4332 &cm_counter_attr_##_group[4].attr.attr, \ 4333 &cm_counter_attr_##_group[5].attr.attr, \ 4334 &cm_counter_attr_##_group[6].attr.attr, \ 4335 &cm_counter_attr_##_group[7].attr.attr, \ 4336 &cm_counter_attr_##_group[8].attr.attr, \ 4337 &cm_counter_attr_##_group[9].attr.attr, \ 4338 &cm_counter_attr_##_group[10].attr.attr, \ 4339 NULL, \ 4340 }; \ 4341 static const struct attribute_group cm_counter_group_##_group = { \ 4342 .name = _name, \ 4343 .attrs = cm_counter_attrs_##_group, \ 4344 }; 4345 4346 CM_COUNTER_GROUP(CM_XMIT, "cm_tx_msgs") 4347 CM_COUNTER_GROUP(CM_XMIT_RETRIES, "cm_tx_retries") 4348 CM_COUNTER_GROUP(CM_RECV, "cm_rx_msgs") 4349 CM_COUNTER_GROUP(CM_RECV_DUPLICATES, "cm_rx_duplicates") 4350 4351 static const struct attribute_group *cm_counter_groups[] = { 4352 &cm_counter_group_CM_XMIT, 4353 &cm_counter_group_CM_XMIT_RETRIES, 4354 &cm_counter_group_CM_RECV, 4355 &cm_counter_group_CM_RECV_DUPLICATES, 4356 NULL, 4357 }; 4358 4359 static int cm_add_one(struct ib_device *ib_device) 4360 { 4361 struct cm_device *cm_dev; 4362 struct cm_port *port; 4363 struct ib_mad_reg_req reg_req = { 4364 .mgmt_class = IB_MGMT_CLASS_CM, 4365 .mgmt_class_version = IB_CM_CLASS_VERSION, 4366 }; 4367 struct ib_port_modify port_modify = { 4368 .set_port_cap_mask = IB_PORT_CM_SUP 4369 }; 4370 unsigned long flags; 4371 int ret; 4372 int count = 0; 4373 u32 i; 4374 4375 cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt), 4376 GFP_KERNEL); 4377 if (!cm_dev) 4378 return -ENOMEM; 4379 4380 kref_init(&cm_dev->kref); 4381 spin_lock_init(&cm_dev->mad_agent_lock); 4382 cm_dev->ib_device = ib_device; 4383 cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay; 4384 cm_dev->going_down = 0; 4385 4386 ib_set_client_data(ib_device, &cm_client, cm_dev); 4387 4388 set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask); 4389 rdma_for_each_port (ib_device, i) { 4390 if (!rdma_cap_ib_cm(ib_device, i)) 4391 continue; 4392 4393 port = kzalloc(sizeof *port, GFP_KERNEL); 4394 if (!port) { 4395 ret = -ENOMEM; 4396 goto error1; 4397 } 4398 4399 cm_dev->port[i-1] = port; 4400 port->cm_dev = cm_dev; 4401 port->port_num = i; 4402 4403 ret = ib_port_register_client_groups(ib_device, i, 4404 cm_counter_groups); 4405 if (ret) 4406 goto error1; 4407 4408 port->mad_agent = ib_register_mad_agent(ib_device, i, 4409 IB_QPT_GSI, 4410 ®_req, 4411 0, 4412 cm_send_handler, 4413 cm_recv_handler, 4414 port, 4415 0); 4416 if (IS_ERR(port->mad_agent)) { 4417 ret = PTR_ERR(port->mad_agent); 4418 goto error2; 4419 } 4420 4421 ret = ib_modify_port(ib_device, i, 0, &port_modify); 4422 if (ret) 4423 goto error3; 4424 4425 count++; 4426 } 4427 4428 if (!count) { 4429 ret = -EOPNOTSUPP; 4430 goto free; 4431 } 4432 4433 write_lock_irqsave(&cm.device_lock, flags); 4434 list_add_tail(&cm_dev->list, &cm.device_list); 4435 write_unlock_irqrestore(&cm.device_lock, flags); 4436 return 0; 4437 4438 error3: 4439 ib_unregister_mad_agent(port->mad_agent); 4440 error2: 4441 ib_port_unregister_client_groups(ib_device, i, cm_counter_groups); 4442 error1: 4443 port_modify.set_port_cap_mask = 0; 4444 port_modify.clr_port_cap_mask = IB_PORT_CM_SUP; 4445 while (--i) { 4446 if (!rdma_cap_ib_cm(ib_device, i)) 4447 continue; 4448 4449 port = cm_dev->port[i-1]; 4450 ib_modify_port(ib_device, port->port_num, 0, &port_modify); 4451 ib_unregister_mad_agent(port->mad_agent); 4452 ib_port_unregister_client_groups(ib_device, i, 4453 cm_counter_groups); 4454 } 4455 free: 4456 cm_device_put(cm_dev); 4457 return ret; 4458 } 4459 4460 static void cm_remove_one(struct ib_device *ib_device, void *client_data) 4461 { 4462 struct cm_device *cm_dev = client_data; 4463 struct cm_port *port; 4464 struct ib_port_modify port_modify = { 4465 .clr_port_cap_mask = IB_PORT_CM_SUP 4466 }; 4467 unsigned long flags; 4468 u32 i; 4469 4470 write_lock_irqsave(&cm.device_lock, flags); 4471 list_del(&cm_dev->list); 4472 write_unlock_irqrestore(&cm.device_lock, flags); 4473 4474 spin_lock_irq(&cm.lock); 4475 cm_dev->going_down = 1; 4476 spin_unlock_irq(&cm.lock); 4477 4478 rdma_for_each_port (ib_device, i) { 4479 struct ib_mad_agent *mad_agent; 4480 4481 if (!rdma_cap_ib_cm(ib_device, i)) 4482 continue; 4483 4484 port = cm_dev->port[i-1]; 4485 mad_agent = port->mad_agent; 4486 ib_modify_port(ib_device, port->port_num, 0, &port_modify); 4487 /* 4488 * We flush the queue here after the going_down set, this 4489 * verify that no new works will be queued in the recv handler, 4490 * after that we can call the unregister_mad_agent 4491 */ 4492 flush_workqueue(cm.wq); 4493 /* 4494 * The above ensures no call paths from the work are running, 4495 * the remaining paths all take the mad_agent_lock. 4496 */ 4497 spin_lock(&cm_dev->mad_agent_lock); 4498 port->mad_agent = NULL; 4499 spin_unlock(&cm_dev->mad_agent_lock); 4500 ib_unregister_mad_agent(mad_agent); 4501 ib_port_unregister_client_groups(ib_device, i, 4502 cm_counter_groups); 4503 } 4504 4505 cm_device_put(cm_dev); 4506 } 4507 4508 static int __init ib_cm_init(void) 4509 { 4510 int ret; 4511 4512 INIT_LIST_HEAD(&cm.device_list); 4513 rwlock_init(&cm.device_lock); 4514 spin_lock_init(&cm.lock); 4515 cm.listen_service_table = RB_ROOT; 4516 cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID); 4517 cm.remote_id_table = RB_ROOT; 4518 cm.remote_qp_table = RB_ROOT; 4519 cm.remote_sidr_table = RB_ROOT; 4520 xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC); 4521 get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand); 4522 INIT_LIST_HEAD(&cm.timewait_list); 4523 4524 cm.wq = alloc_workqueue("ib_cm", 0, 1); 4525 if (!cm.wq) { 4526 ret = -ENOMEM; 4527 goto error2; 4528 } 4529 4530 ret = ib_register_client(&cm_client); 4531 if (ret) 4532 goto error3; 4533 4534 return 0; 4535 error3: 4536 destroy_workqueue(cm.wq); 4537 error2: 4538 return ret; 4539 } 4540 4541 static void __exit ib_cm_cleanup(void) 4542 { 4543 struct cm_timewait_info *timewait_info, *tmp; 4544 4545 spin_lock_irq(&cm.lock); 4546 list_for_each_entry(timewait_info, &cm.timewait_list, list) 4547 cancel_delayed_work(&timewait_info->work.work); 4548 spin_unlock_irq(&cm.lock); 4549 4550 ib_unregister_client(&cm_client); 4551 destroy_workqueue(cm.wq); 4552 4553 list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) { 4554 list_del(&timewait_info->list); 4555 kfree(timewait_info); 4556 } 4557 4558 WARN_ON(!xa_empty(&cm.local_id_table)); 4559 } 4560 4561 module_init(ib_cm_init); 4562 module_exit(ib_cm_cleanup); 4563