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_DIRECT_RETRY_CTX ((void *) 1UL) 38 #define CM_MRA_SETTING 24 /* 4.096us * 2^24 = ~68.7 seconds */ 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 struct ib_mad_agent *rep_agent; 164 u32 port_num; 165 atomic_long_t counters[CM_COUNTER_GROUPS][CM_ATTR_COUNT]; 166 }; 167 168 struct cm_device { 169 struct kref kref; 170 struct list_head list; 171 rwlock_t mad_agent_lock; 172 struct ib_device *ib_device; 173 u8 ack_delay; 174 int going_down; 175 struct cm_port *port[]; 176 }; 177 178 struct cm_av { 179 struct cm_port *port; 180 struct rdma_ah_attr ah_attr; 181 u16 dlid_datapath; 182 u16 pkey_index; 183 u8 timeout; 184 }; 185 186 struct cm_work { 187 struct delayed_work work; 188 struct list_head list; 189 struct cm_port *port; 190 struct ib_mad_recv_wc *mad_recv_wc; /* Received MADs */ 191 __be32 local_id; /* Established / timewait */ 192 __be32 remote_id; 193 struct ib_cm_event cm_event; 194 struct sa_path_rec path[]; 195 }; 196 197 struct cm_timewait_info { 198 struct cm_work work; 199 struct list_head list; 200 struct rb_node remote_qp_node; 201 struct rb_node remote_id_node; 202 __be64 remote_ca_guid; 203 __be32 remote_qpn; 204 u8 inserted_remote_qp; 205 u8 inserted_remote_id; 206 }; 207 208 struct cm_id_private { 209 struct ib_cm_id id; 210 211 struct rb_node service_node; 212 struct rb_node sidr_id_node; 213 u32 sidr_slid; 214 spinlock_t lock; /* Do not acquire inside cm.lock */ 215 struct completion comp; 216 refcount_t refcount; 217 /* Number of clients sharing this ib_cm_id. Only valid for listeners. 218 * Protected by the cm.lock spinlock. 219 */ 220 int listen_sharecount; 221 struct rcu_head rcu; 222 223 struct ib_mad_send_buf *msg; 224 struct cm_timewait_info *timewait_info; 225 /* todo: use alternate port on send failure */ 226 struct cm_av av; 227 struct cm_av alt_av; 228 229 void *private_data; 230 __be64 tid; 231 __be32 local_qpn; 232 __be32 remote_qpn; 233 enum ib_qp_type qp_type; 234 __be32 sq_psn; 235 __be32 rq_psn; 236 int timeout_ms; 237 enum ib_mtu path_mtu; 238 __be16 pkey; 239 u8 private_data_len; 240 u8 max_cm_retries; 241 u8 responder_resources; 242 u8 initiator_depth; 243 u8 retry_count; 244 u8 rnr_retry_count; 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 * 278 cm_alloc_msg_agent(struct cm_id_private *cm_id_priv, bool rep_agent) 279 { 280 struct ib_mad_agent *mad_agent; 281 struct ib_mad_send_buf *m; 282 struct ib_ah *ah; 283 284 lockdep_assert_held(&cm_id_priv->lock); 285 286 if (!cm_id_priv->av.port) 287 return ERR_PTR(-EINVAL); 288 289 read_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 290 mad_agent = rep_agent ? cm_id_priv->av.port->rep_agent : 291 cm_id_priv->av.port->mad_agent; 292 if (!mad_agent) { 293 m = ERR_PTR(-EINVAL); 294 goto out; 295 } 296 297 ah = rdma_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr, 0); 298 if (IS_ERR(ah)) { 299 m = ERR_CAST(ah); 300 goto out; 301 } 302 303 m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn, 304 cm_id_priv->av.pkey_index, 305 0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA, 306 GFP_ATOMIC, 307 IB_MGMT_BASE_VERSION); 308 if (IS_ERR(m)) { 309 rdma_destroy_ah(ah, 0); 310 goto out; 311 } 312 313 m->ah = ah; 314 315 out: 316 read_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 317 return m; 318 } 319 320 static struct ib_mad_send_buf *cm_alloc_msg(struct cm_id_private *cm_id_priv) 321 { 322 return cm_alloc_msg_agent(cm_id_priv, false); 323 } 324 325 static void cm_free_msg(struct ib_mad_send_buf *msg) 326 { 327 if (msg->ah) 328 rdma_destroy_ah(msg->ah, 0); 329 ib_free_send_mad(msg); 330 } 331 332 static struct ib_mad_send_buf * 333 cm_alloc_priv_msg_rep(struct cm_id_private *cm_id_priv, enum ib_cm_state state, 334 bool rep_agent) 335 { 336 struct ib_mad_send_buf *msg; 337 338 lockdep_assert_held(&cm_id_priv->lock); 339 340 msg = cm_alloc_msg_agent(cm_id_priv, rep_agent); 341 if (IS_ERR(msg)) 342 return msg; 343 344 cm_id_priv->msg = msg; 345 refcount_inc(&cm_id_priv->refcount); 346 msg->context[0] = cm_id_priv; 347 msg->context[1] = (void *) (unsigned long) state; 348 349 msg->retries = cm_id_priv->max_cm_retries; 350 msg->timeout_ms = cm_id_priv->timeout_ms; 351 352 return msg; 353 } 354 355 static struct ib_mad_send_buf * 356 cm_alloc_priv_msg(struct cm_id_private *cm_id_priv, enum ib_cm_state state) 357 { 358 return cm_alloc_priv_msg_rep(cm_id_priv, state, false); 359 } 360 361 static void cm_free_priv_msg(struct ib_mad_send_buf *msg) 362 { 363 struct cm_id_private *cm_id_priv = msg->context[0]; 364 365 lockdep_assert_held(&cm_id_priv->lock); 366 367 if (!WARN_ON(cm_id_priv->msg != msg)) 368 cm_id_priv->msg = NULL; 369 370 if (msg->ah) 371 rdma_destroy_ah(msg->ah, 0); 372 cm_deref_id(cm_id_priv); 373 ib_free_send_mad(msg); 374 } 375 376 static struct ib_mad_send_buf * 377 cm_alloc_response_msg_no_ah(struct cm_port *port, 378 struct ib_mad_recv_wc *mad_recv_wc, 379 bool direct_retry) 380 { 381 struct ib_mad_send_buf *m; 382 383 m = ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index, 384 0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA, 385 GFP_ATOMIC, IB_MGMT_BASE_VERSION); 386 if (!IS_ERR(m)) 387 m->context[0] = direct_retry ? CM_DIRECT_RETRY_CTX : NULL; 388 389 return m; 390 } 391 392 static int cm_create_response_msg_ah(struct cm_port *port, 393 struct ib_mad_recv_wc *mad_recv_wc, 394 struct ib_mad_send_buf *msg) 395 { 396 struct ib_ah *ah; 397 398 ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc, 399 mad_recv_wc->recv_buf.grh, port->port_num); 400 if (IS_ERR(ah)) 401 return PTR_ERR(ah); 402 403 msg->ah = ah; 404 return 0; 405 } 406 407 static int cm_alloc_response_msg(struct cm_port *port, 408 struct ib_mad_recv_wc *mad_recv_wc, 409 bool direct_retry, 410 struct ib_mad_send_buf **msg) 411 { 412 struct ib_mad_send_buf *m; 413 int ret; 414 415 m = cm_alloc_response_msg_no_ah(port, mad_recv_wc, direct_retry); 416 if (IS_ERR(m)) 417 return PTR_ERR(m); 418 419 ret = cm_create_response_msg_ah(port, mad_recv_wc, m); 420 if (ret) { 421 ib_free_send_mad(m); 422 return ret; 423 } 424 425 *msg = m; 426 return 0; 427 } 428 429 static void *cm_copy_private_data(const void *private_data, u8 private_data_len) 430 { 431 void *data; 432 433 if (!private_data || !private_data_len) 434 return NULL; 435 436 data = kmemdup(private_data, private_data_len, GFP_KERNEL); 437 if (!data) 438 return ERR_PTR(-ENOMEM); 439 440 return data; 441 } 442 443 static void cm_set_private_data(struct cm_id_private *cm_id_priv, 444 void *private_data, u8 private_data_len) 445 { 446 if (cm_id_priv->private_data && cm_id_priv->private_data_len) 447 kfree(cm_id_priv->private_data); 448 449 cm_id_priv->private_data = private_data; 450 cm_id_priv->private_data_len = private_data_len; 451 } 452 453 static void cm_set_av_port(struct cm_av *av, struct cm_port *port) 454 { 455 struct cm_port *old_port = av->port; 456 457 if (old_port == port) 458 return; 459 460 av->port = port; 461 if (old_port) 462 cm_device_put(old_port->cm_dev); 463 if (port) 464 kref_get(&port->cm_dev->kref); 465 } 466 467 static void cm_init_av_for_lap(struct cm_port *port, struct ib_wc *wc, 468 struct rdma_ah_attr *ah_attr, struct cm_av *av) 469 { 470 cm_set_av_port(av, port); 471 av->pkey_index = wc->pkey_index; 472 rdma_move_ah_attr(&av->ah_attr, ah_attr); 473 } 474 475 static int cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc, 476 struct ib_grh *grh, struct cm_av *av) 477 { 478 cm_set_av_port(av, port); 479 av->pkey_index = wc->pkey_index; 480 return ib_init_ah_attr_from_wc(port->cm_dev->ib_device, 481 port->port_num, wc, 482 grh, &av->ah_attr); 483 } 484 485 static struct cm_port * 486 get_cm_port_from_path(struct sa_path_rec *path, const struct ib_gid_attr *attr) 487 { 488 struct cm_device *cm_dev; 489 struct cm_port *port = NULL; 490 unsigned long flags; 491 492 if (attr) { 493 read_lock_irqsave(&cm.device_lock, flags); 494 list_for_each_entry(cm_dev, &cm.device_list, list) { 495 if (cm_dev->ib_device == attr->device) { 496 port = cm_dev->port[attr->port_num - 1]; 497 break; 498 } 499 } 500 read_unlock_irqrestore(&cm.device_lock, flags); 501 } else { 502 /* SGID attribute can be NULL in following 503 * conditions. 504 * (a) Alternative path 505 * (b) IB link layer without GRH 506 * (c) LAP send messages 507 */ 508 read_lock_irqsave(&cm.device_lock, flags); 509 list_for_each_entry(cm_dev, &cm.device_list, list) { 510 attr = rdma_find_gid(cm_dev->ib_device, 511 &path->sgid, 512 sa_conv_pathrec_to_gid_type(path), 513 NULL); 514 if (!IS_ERR(attr)) { 515 port = cm_dev->port[attr->port_num - 1]; 516 break; 517 } 518 } 519 read_unlock_irqrestore(&cm.device_lock, flags); 520 if (port) 521 rdma_put_gid_attr(attr); 522 } 523 return port; 524 } 525 526 static int cm_init_av_by_path(struct sa_path_rec *path, 527 const struct ib_gid_attr *sgid_attr, 528 struct cm_av *av) 529 { 530 struct rdma_ah_attr new_ah_attr; 531 struct cm_device *cm_dev; 532 struct cm_port *port; 533 u16 pkey_index; 534 int ret; 535 536 port = get_cm_port_from_path(path, sgid_attr); 537 if (!port) 538 return -EINVAL; 539 cm_dev = port->cm_dev; 540 541 ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num, 542 be16_to_cpu(path->pkey), &pkey_index); 543 if (ret) 544 return ret; 545 546 /* 547 * av->ah_attr might be initialized based on wc or during 548 * request processing time which might have reference to sgid_attr. 549 * So initialize a new ah_attr on stack. 550 * If initialization fails, old ah_attr is used for sending any 551 * responses. If initialization is successful, than new ah_attr 552 * is used by overwriting the old one. So that right ah_attr 553 * can be used to return an error response. 554 */ 555 ret = ib_init_ah_attr_from_path(cm_dev->ib_device, port->port_num, path, 556 &new_ah_attr, sgid_attr); 557 if (ret) 558 return ret; 559 560 av->pkey_index = pkey_index; 561 cm_set_av_port(av, port); 562 av->timeout = path->packet_life_time + 1; 563 rdma_move_ah_attr(&av->ah_attr, &new_ah_attr); 564 return 0; 565 } 566 567 /* Move av created by cm_init_av_by_path(), so av.dgid is not moved */ 568 static void cm_move_av_from_path(struct cm_av *dest, struct cm_av *src) 569 { 570 cm_set_av_port(dest, src->port); 571 cm_set_av_port(src, NULL); 572 dest->pkey_index = src->pkey_index; 573 rdma_move_ah_attr(&dest->ah_attr, &src->ah_attr); 574 dest->timeout = src->timeout; 575 } 576 577 static void cm_destroy_av(struct cm_av *av) 578 { 579 rdma_destroy_ah_attr(&av->ah_attr); 580 cm_set_av_port(av, NULL); 581 } 582 583 static u32 cm_local_id(__be32 local_id) 584 { 585 return (__force u32) (local_id ^ cm.random_id_operand); 586 } 587 588 static struct cm_id_private *cm_acquire_id(__be32 local_id, __be32 remote_id) 589 { 590 struct cm_id_private *cm_id_priv; 591 592 rcu_read_lock(); 593 cm_id_priv = xa_load(&cm.local_id_table, cm_local_id(local_id)); 594 if (!cm_id_priv || cm_id_priv->id.remote_id != remote_id || 595 !refcount_inc_not_zero(&cm_id_priv->refcount)) 596 cm_id_priv = NULL; 597 rcu_read_unlock(); 598 599 return cm_id_priv; 600 } 601 602 /* 603 * Trivial helpers to strip endian annotation and compare; the 604 * endianness doesn't actually matter since we just need a stable 605 * order for the RB tree. 606 */ 607 static int be32_lt(__be32 a, __be32 b) 608 { 609 return (__force u32) a < (__force u32) b; 610 } 611 612 static int be32_gt(__be32 a, __be32 b) 613 { 614 return (__force u32) a > (__force u32) b; 615 } 616 617 static int be64_lt(__be64 a, __be64 b) 618 { 619 return (__force u64) a < (__force u64) b; 620 } 621 622 static int be64_gt(__be64 a, __be64 b) 623 { 624 return (__force u64) a > (__force u64) b; 625 } 626 627 /* 628 * Inserts a new cm_id_priv into the listen_service_table. Returns cm_id_priv 629 * if the new ID was inserted, NULL if it could not be inserted due to a 630 * collision, or the existing cm_id_priv ready for shared usage. 631 */ 632 static struct cm_id_private *cm_insert_listen(struct cm_id_private *cm_id_priv, 633 ib_cm_handler shared_handler) 634 { 635 struct rb_node **link = &cm.listen_service_table.rb_node; 636 struct rb_node *parent = NULL; 637 struct cm_id_private *cur_cm_id_priv; 638 __be64 service_id = cm_id_priv->id.service_id; 639 unsigned long flags; 640 641 spin_lock_irqsave(&cm.lock, flags); 642 while (*link) { 643 parent = *link; 644 cur_cm_id_priv = rb_entry(parent, struct cm_id_private, 645 service_node); 646 647 if (cm_id_priv->id.device < cur_cm_id_priv->id.device) 648 link = &(*link)->rb_left; 649 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device) 650 link = &(*link)->rb_right; 651 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id)) 652 link = &(*link)->rb_left; 653 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id)) 654 link = &(*link)->rb_right; 655 else { 656 /* 657 * Sharing an ib_cm_id with different handlers is not 658 * supported 659 */ 660 if (cur_cm_id_priv->id.cm_handler != shared_handler || 661 cur_cm_id_priv->id.context || 662 WARN_ON(!cur_cm_id_priv->id.cm_handler)) { 663 spin_unlock_irqrestore(&cm.lock, flags); 664 return NULL; 665 } 666 refcount_inc(&cur_cm_id_priv->refcount); 667 cur_cm_id_priv->listen_sharecount++; 668 spin_unlock_irqrestore(&cm.lock, flags); 669 return cur_cm_id_priv; 670 } 671 } 672 cm_id_priv->listen_sharecount++; 673 rb_link_node(&cm_id_priv->service_node, parent, link); 674 rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table); 675 spin_unlock_irqrestore(&cm.lock, flags); 676 return cm_id_priv; 677 } 678 679 static struct cm_id_private *cm_find_listen(struct ib_device *device, 680 __be64 service_id) 681 { 682 struct rb_node *node = cm.listen_service_table.rb_node; 683 struct cm_id_private *cm_id_priv; 684 685 while (node) { 686 cm_id_priv = rb_entry(node, struct cm_id_private, service_node); 687 688 if (device < cm_id_priv->id.device) 689 node = node->rb_left; 690 else if (device > cm_id_priv->id.device) 691 node = node->rb_right; 692 else if (be64_lt(service_id, cm_id_priv->id.service_id)) 693 node = node->rb_left; 694 else if (be64_gt(service_id, cm_id_priv->id.service_id)) 695 node = node->rb_right; 696 else { 697 refcount_inc(&cm_id_priv->refcount); 698 return cm_id_priv; 699 } 700 } 701 return NULL; 702 } 703 704 static struct cm_timewait_info * 705 cm_insert_remote_id(struct cm_timewait_info *timewait_info) 706 { 707 struct rb_node **link = &cm.remote_id_table.rb_node; 708 struct rb_node *parent = NULL; 709 struct cm_timewait_info *cur_timewait_info; 710 __be64 remote_ca_guid = timewait_info->remote_ca_guid; 711 __be32 remote_id = timewait_info->work.remote_id; 712 713 while (*link) { 714 parent = *link; 715 cur_timewait_info = rb_entry(parent, struct cm_timewait_info, 716 remote_id_node); 717 if (be32_lt(remote_id, cur_timewait_info->work.remote_id)) 718 link = &(*link)->rb_left; 719 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id)) 720 link = &(*link)->rb_right; 721 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 722 link = &(*link)->rb_left; 723 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 724 link = &(*link)->rb_right; 725 else 726 return cur_timewait_info; 727 } 728 timewait_info->inserted_remote_id = 1; 729 rb_link_node(&timewait_info->remote_id_node, parent, link); 730 rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table); 731 return NULL; 732 } 733 734 static struct cm_id_private *cm_find_remote_id(__be64 remote_ca_guid, 735 __be32 remote_id) 736 { 737 struct rb_node *node = cm.remote_id_table.rb_node; 738 struct cm_timewait_info *timewait_info; 739 struct cm_id_private *res = NULL; 740 741 spin_lock_irq(&cm.lock); 742 while (node) { 743 timewait_info = rb_entry(node, struct cm_timewait_info, 744 remote_id_node); 745 if (be32_lt(remote_id, timewait_info->work.remote_id)) 746 node = node->rb_left; 747 else if (be32_gt(remote_id, timewait_info->work.remote_id)) 748 node = node->rb_right; 749 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid)) 750 node = node->rb_left; 751 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid)) 752 node = node->rb_right; 753 else { 754 res = cm_acquire_id(timewait_info->work.local_id, 755 timewait_info->work.remote_id); 756 break; 757 } 758 } 759 spin_unlock_irq(&cm.lock); 760 return res; 761 } 762 763 static struct cm_timewait_info * 764 cm_insert_remote_qpn(struct cm_timewait_info *timewait_info) 765 { 766 struct rb_node **link = &cm.remote_qp_table.rb_node; 767 struct rb_node *parent = NULL; 768 struct cm_timewait_info *cur_timewait_info; 769 __be64 remote_ca_guid = timewait_info->remote_ca_guid; 770 __be32 remote_qpn = timewait_info->remote_qpn; 771 772 while (*link) { 773 parent = *link; 774 cur_timewait_info = rb_entry(parent, struct cm_timewait_info, 775 remote_qp_node); 776 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn)) 777 link = &(*link)->rb_left; 778 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn)) 779 link = &(*link)->rb_right; 780 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 781 link = &(*link)->rb_left; 782 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid)) 783 link = &(*link)->rb_right; 784 else 785 return cur_timewait_info; 786 } 787 timewait_info->inserted_remote_qp = 1; 788 rb_link_node(&timewait_info->remote_qp_node, parent, link); 789 rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table); 790 return NULL; 791 } 792 793 static struct cm_id_private * 794 cm_insert_remote_sidr(struct cm_id_private *cm_id_priv) 795 { 796 struct rb_node **link = &cm.remote_sidr_table.rb_node; 797 struct rb_node *parent = NULL; 798 struct cm_id_private *cur_cm_id_priv; 799 __be32 remote_id = cm_id_priv->id.remote_id; 800 801 while (*link) { 802 parent = *link; 803 cur_cm_id_priv = rb_entry(parent, struct cm_id_private, 804 sidr_id_node); 805 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id)) 806 link = &(*link)->rb_left; 807 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id)) 808 link = &(*link)->rb_right; 809 else { 810 if (cur_cm_id_priv->sidr_slid < cm_id_priv->sidr_slid) 811 link = &(*link)->rb_left; 812 else if (cur_cm_id_priv->sidr_slid > cm_id_priv->sidr_slid) 813 link = &(*link)->rb_right; 814 else 815 return cur_cm_id_priv; 816 } 817 } 818 rb_link_node(&cm_id_priv->sidr_id_node, parent, link); 819 rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table); 820 return NULL; 821 } 822 823 static struct cm_id_private *cm_alloc_id_priv(struct ib_device *device, 824 ib_cm_handler cm_handler, 825 void *context) 826 { 827 struct cm_id_private *cm_id_priv; 828 u32 id; 829 int ret; 830 831 cm_id_priv = kzalloc_obj(*cm_id_priv); 832 if (!cm_id_priv) 833 return ERR_PTR(-ENOMEM); 834 835 cm_id_priv->id.state = IB_CM_IDLE; 836 cm_id_priv->id.device = device; 837 cm_id_priv->id.cm_handler = cm_handler; 838 cm_id_priv->id.context = context; 839 cm_id_priv->id.remote_cm_qpn = 1; 840 841 RB_CLEAR_NODE(&cm_id_priv->service_node); 842 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node); 843 spin_lock_init(&cm_id_priv->lock); 844 init_completion(&cm_id_priv->comp); 845 INIT_LIST_HEAD(&cm_id_priv->work_list); 846 atomic_set(&cm_id_priv->work_count, -1); 847 refcount_set(&cm_id_priv->refcount, 1); 848 849 ret = xa_alloc_cyclic(&cm.local_id_table, &id, NULL, xa_limit_32b, 850 &cm.local_id_next, GFP_KERNEL); 851 if (ret < 0) 852 goto error; 853 cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand; 854 855 return cm_id_priv; 856 857 error: 858 kfree(cm_id_priv); 859 return ERR_PTR(ret); 860 } 861 862 /* 863 * Make the ID visible to the MAD handlers and other threads that use the 864 * xarray. 865 */ 866 static void cm_finalize_id(struct cm_id_private *cm_id_priv) 867 { 868 xa_store(&cm.local_id_table, cm_local_id(cm_id_priv->id.local_id), 869 cm_id_priv, GFP_ATOMIC); 870 } 871 872 struct ib_cm_id *ib_create_cm_id(struct ib_device *device, 873 ib_cm_handler cm_handler, 874 void *context) 875 { 876 struct cm_id_private *cm_id_priv; 877 878 cm_id_priv = cm_alloc_id_priv(device, cm_handler, context); 879 if (IS_ERR(cm_id_priv)) 880 return ERR_CAST(cm_id_priv); 881 882 cm_finalize_id(cm_id_priv); 883 return &cm_id_priv->id; 884 } 885 EXPORT_SYMBOL(ib_create_cm_id); 886 887 static struct cm_work *cm_dequeue_work(struct cm_id_private *cm_id_priv) 888 { 889 struct cm_work *work; 890 891 if (list_empty(&cm_id_priv->work_list)) 892 return NULL; 893 894 work = list_entry(cm_id_priv->work_list.next, struct cm_work, list); 895 list_del(&work->list); 896 return work; 897 } 898 899 static void cm_free_work(struct cm_work *work) 900 { 901 if (work->mad_recv_wc) 902 ib_free_recv_mad(work->mad_recv_wc); 903 kfree(work); 904 } 905 906 static void cm_queue_work_unlock(struct cm_id_private *cm_id_priv, 907 struct cm_work *work) 908 __releases(&cm_id_priv->lock) 909 { 910 bool immediate; 911 912 /* 913 * To deliver the event to the user callback we have the drop the 914 * spinlock, however, we need to ensure that the user callback is single 915 * threaded and receives events in the temporal order. If there are 916 * already events being processed then thread new events onto a list, 917 * the thread currently processing will pick them up. 918 */ 919 immediate = atomic_inc_and_test(&cm_id_priv->work_count); 920 if (!immediate) { 921 list_add_tail(&work->list, &cm_id_priv->work_list); 922 /* 923 * This routine always consumes incoming reference. Once queued 924 * to the work_list then a reference is held by the thread 925 * currently running cm_process_work() and this reference is not 926 * needed. 927 */ 928 cm_deref_id(cm_id_priv); 929 } 930 spin_unlock_irq(&cm_id_priv->lock); 931 932 if (immediate) 933 cm_process_work(cm_id_priv, work); 934 } 935 936 static inline int cm_convert_to_ms(int iba_time) 937 { 938 /* approximate conversion to ms from 4.096us x 2^iba_time */ 939 return 1 << max(iba_time - 8, 0); 940 } 941 942 /* 943 * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time 944 * Because of how ack_timeout is stored, adding one doubles the timeout. 945 * To avoid large timeouts, select the max(ack_delay, life_time + 1), and 946 * increment it (round up) only if the other is within 50%. 947 */ 948 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time) 949 { 950 int ack_timeout = packet_life_time + 1; 951 952 if (ack_timeout >= ca_ack_delay) 953 ack_timeout += (ca_ack_delay >= (ack_timeout - 1)); 954 else 955 ack_timeout = ca_ack_delay + 956 (ack_timeout >= (ca_ack_delay - 1)); 957 958 return min(31, ack_timeout); 959 } 960 961 static void cm_remove_remote(struct cm_id_private *cm_id_priv) 962 { 963 struct cm_timewait_info *timewait_info = cm_id_priv->timewait_info; 964 965 if (timewait_info->inserted_remote_id) { 966 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table); 967 timewait_info->inserted_remote_id = 0; 968 } 969 970 if (timewait_info->inserted_remote_qp) { 971 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table); 972 timewait_info->inserted_remote_qp = 0; 973 } 974 } 975 976 static struct cm_timewait_info *cm_create_timewait_info(__be32 local_id) 977 { 978 struct cm_timewait_info *timewait_info; 979 980 timewait_info = kzalloc_obj(*timewait_info); 981 if (!timewait_info) 982 return ERR_PTR(-ENOMEM); 983 984 timewait_info->work.local_id = local_id; 985 INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler); 986 timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT; 987 return timewait_info; 988 } 989 990 static void cm_enter_timewait(struct cm_id_private *cm_id_priv) 991 { 992 int wait_time; 993 unsigned long flags; 994 struct cm_device *cm_dev; 995 996 lockdep_assert_held(&cm_id_priv->lock); 997 998 cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client); 999 if (!cm_dev) 1000 return; 1001 1002 spin_lock_irqsave(&cm.lock, flags); 1003 cm_remove_remote(cm_id_priv); 1004 list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list); 1005 spin_unlock_irqrestore(&cm.lock, flags); 1006 1007 /* 1008 * The cm_id could be destroyed by the user before we exit timewait. 1009 * To protect against this, we search for the cm_id after exiting 1010 * timewait before notifying the user that we've exited timewait. 1011 */ 1012 cm_id_priv->id.state = IB_CM_TIMEWAIT; 1013 wait_time = cm_convert_to_ms(cm_id_priv->av.timeout); 1014 1015 /* Check if the device started its remove_one */ 1016 spin_lock_irqsave(&cm.lock, flags); 1017 if (!cm_dev->going_down) 1018 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work, 1019 msecs_to_jiffies(wait_time)); 1020 spin_unlock_irqrestore(&cm.lock, flags); 1021 1022 /* 1023 * The timewait_info is converted into a work and gets freed during 1024 * cm_free_work() in cm_timewait_handler(). 1025 */ 1026 BUILD_BUG_ON(offsetof(struct cm_timewait_info, work) != 0); 1027 cm_id_priv->timewait_info = NULL; 1028 } 1029 1030 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv) 1031 { 1032 unsigned long flags; 1033 1034 lockdep_assert_held(&cm_id_priv->lock); 1035 1036 cm_id_priv->id.state = IB_CM_IDLE; 1037 if (cm_id_priv->timewait_info) { 1038 spin_lock_irqsave(&cm.lock, flags); 1039 cm_remove_remote(cm_id_priv); 1040 spin_unlock_irqrestore(&cm.lock, flags); 1041 kfree(cm_id_priv->timewait_info); 1042 cm_id_priv->timewait_info = NULL; 1043 } 1044 } 1045 1046 static noinline void cm_destroy_id_wait_timeout(struct ib_cm_id *cm_id, 1047 enum ib_cm_state old_state) 1048 { 1049 struct cm_id_private *cm_id_priv; 1050 1051 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 1052 pr_err_ratelimited("%s: cm_id=%p timed out. state %d -> %d, refcnt=%d\n", __func__, 1053 cm_id, old_state, cm_id->state, refcount_read(&cm_id_priv->refcount)); 1054 } 1055 1056 static void cm_destroy_id(struct ib_cm_id *cm_id, int err) 1057 { 1058 struct cm_id_private *cm_id_priv; 1059 enum ib_cm_state old_state; 1060 unsigned long timeout; 1061 struct cm_work *work; 1062 int ret; 1063 1064 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 1065 spin_lock_irq(&cm_id_priv->lock); 1066 old_state = cm_id->state; 1067 retest: 1068 switch (cm_id->state) { 1069 case IB_CM_LISTEN: 1070 spin_lock(&cm.lock); 1071 if (--cm_id_priv->listen_sharecount > 0) { 1072 /* The id is still shared. */ 1073 WARN_ON(refcount_read(&cm_id_priv->refcount) == 1); 1074 spin_unlock(&cm.lock); 1075 spin_unlock_irq(&cm_id_priv->lock); 1076 cm_deref_id(cm_id_priv); 1077 return; 1078 } 1079 cm_id->state = IB_CM_IDLE; 1080 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table); 1081 RB_CLEAR_NODE(&cm_id_priv->service_node); 1082 spin_unlock(&cm.lock); 1083 break; 1084 case IB_CM_SIDR_REQ_SENT: 1085 cm_id->state = IB_CM_IDLE; 1086 ib_cancel_mad(cm_id_priv->msg); 1087 break; 1088 case IB_CM_SIDR_REQ_RCVD: 1089 cm_send_sidr_rep_locked(cm_id_priv, 1090 &(struct ib_cm_sidr_rep_param){ 1091 .status = IB_SIDR_REJECT }); 1092 /* cm_send_sidr_rep_locked will not move to IDLE if it fails */ 1093 cm_id->state = IB_CM_IDLE; 1094 break; 1095 case IB_CM_REQ_SENT: 1096 case IB_CM_MRA_REQ_RCVD: 1097 ib_cancel_mad(cm_id_priv->msg); 1098 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_TIMEOUT, 1099 &cm_id_priv->id.device->node_guid, 1100 sizeof(cm_id_priv->id.device->node_guid), 1101 NULL, 0); 1102 break; 1103 case IB_CM_REQ_RCVD: 1104 if (err == -ENOMEM) { 1105 /* Do not reject to allow future retries. */ 1106 cm_reset_to_idle(cm_id_priv); 1107 } else { 1108 cm_send_rej_locked(cm_id_priv, 1109 IB_CM_REJ_CONSUMER_DEFINED, NULL, 0, 1110 NULL, 0); 1111 } 1112 break; 1113 case IB_CM_REP_SENT: 1114 case IB_CM_MRA_REP_RCVD: 1115 ib_cancel_mad(cm_id_priv->msg); 1116 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL, 1117 0, NULL, 0); 1118 goto retest; 1119 case IB_CM_MRA_REQ_SENT: 1120 case IB_CM_REP_RCVD: 1121 case IB_CM_MRA_REP_SENT: 1122 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL, 1123 0, NULL, 0); 1124 break; 1125 case IB_CM_ESTABLISHED: 1126 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT) { 1127 cm_id->state = IB_CM_IDLE; 1128 break; 1129 } 1130 cm_issue_dreq(cm_id_priv); 1131 cm_enter_timewait(cm_id_priv); 1132 goto retest; 1133 case IB_CM_DREQ_SENT: 1134 ib_cancel_mad(cm_id_priv->msg); 1135 cm_enter_timewait(cm_id_priv); 1136 goto retest; 1137 case IB_CM_DREQ_RCVD: 1138 cm_send_drep_locked(cm_id_priv, NULL, 0); 1139 WARN_ON(cm_id->state != IB_CM_TIMEWAIT); 1140 goto retest; 1141 case IB_CM_TIMEWAIT: 1142 /* 1143 * The cm_acquire_id in cm_timewait_handler will stop working 1144 * once we do xa_erase below, so just move to idle here for 1145 * consistency. 1146 */ 1147 cm_id->state = IB_CM_IDLE; 1148 break; 1149 case IB_CM_IDLE: 1150 break; 1151 } 1152 WARN_ON(cm_id->state != IB_CM_IDLE); 1153 1154 spin_lock(&cm.lock); 1155 /* Required for cleanup paths related cm_req_handler() */ 1156 if (cm_id_priv->timewait_info) { 1157 cm_remove_remote(cm_id_priv); 1158 kfree(cm_id_priv->timewait_info); 1159 cm_id_priv->timewait_info = NULL; 1160 } 1161 1162 WARN_ON(cm_id_priv->listen_sharecount); 1163 WARN_ON(!RB_EMPTY_NODE(&cm_id_priv->service_node)); 1164 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) 1165 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table); 1166 spin_unlock(&cm.lock); 1167 spin_unlock_irq(&cm_id_priv->lock); 1168 1169 xa_erase(&cm.local_id_table, cm_local_id(cm_id->local_id)); 1170 cm_deref_id(cm_id_priv); 1171 timeout = msecs_to_jiffies((cm_id_priv->max_cm_retries * cm_id_priv->timeout_ms * 5) / 4); 1172 do { 1173 ret = wait_for_completion_timeout(&cm_id_priv->comp, timeout); 1174 if (!ret) /* timeout happened */ 1175 cm_destroy_id_wait_timeout(cm_id, old_state); 1176 } while (!ret); 1177 1178 while ((work = cm_dequeue_work(cm_id_priv)) != NULL) 1179 cm_free_work(work); 1180 1181 cm_destroy_av(&cm_id_priv->av); 1182 cm_destroy_av(&cm_id_priv->alt_av); 1183 kfree(cm_id_priv->private_data); 1184 kfree_rcu(cm_id_priv, rcu); 1185 } 1186 1187 void ib_destroy_cm_id(struct ib_cm_id *cm_id) 1188 { 1189 cm_destroy_id(cm_id, 0); 1190 } 1191 EXPORT_SYMBOL(ib_destroy_cm_id); 1192 1193 static int cm_init_listen(struct cm_id_private *cm_id_priv, __be64 service_id) 1194 { 1195 if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID && 1196 (service_id != IB_CM_ASSIGN_SERVICE_ID)) 1197 return -EINVAL; 1198 1199 if (service_id == IB_CM_ASSIGN_SERVICE_ID) 1200 cm_id_priv->id.service_id = cpu_to_be64(cm.listen_service_id++); 1201 else 1202 cm_id_priv->id.service_id = service_id; 1203 1204 return 0; 1205 } 1206 1207 /** 1208 * ib_cm_listen - Initiates listening on the specified service ID for 1209 * connection and service ID resolution requests. 1210 * @cm_id: Connection identifier associated with the listen request. 1211 * @service_id: Service identifier matched against incoming connection 1212 * and service ID resolution requests. The service ID should be specified 1213 * network-byte order. If set to IB_CM_ASSIGN_SERVICE_ID, the CM will 1214 * assign a service ID to the caller. 1215 */ 1216 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id) 1217 { 1218 struct cm_id_private *cm_id_priv = 1219 container_of(cm_id, struct cm_id_private, id); 1220 unsigned long flags; 1221 int ret; 1222 1223 spin_lock_irqsave(&cm_id_priv->lock, flags); 1224 if (cm_id_priv->id.state != IB_CM_IDLE) { 1225 ret = -EINVAL; 1226 goto out; 1227 } 1228 1229 ret = cm_init_listen(cm_id_priv, service_id); 1230 if (ret) 1231 goto out; 1232 1233 if (!cm_insert_listen(cm_id_priv, NULL)) { 1234 ret = -EBUSY; 1235 goto out; 1236 } 1237 1238 cm_id_priv->id.state = IB_CM_LISTEN; 1239 ret = 0; 1240 1241 out: 1242 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1243 return ret; 1244 } 1245 EXPORT_SYMBOL(ib_cm_listen); 1246 1247 /** 1248 * ib_cm_insert_listen - Create a new listening ib_cm_id and listen on 1249 * the given service ID. 1250 * 1251 * If there's an existing ID listening on that same device and service ID, 1252 * return it. 1253 * 1254 * @device: Device associated with the cm_id. All related communication will 1255 * be associated with the specified device. 1256 * @cm_handler: Callback invoked to notify the user of CM events. 1257 * @service_id: Service identifier matched against incoming connection 1258 * and service ID resolution requests. The service ID should be specified 1259 * network-byte order. If set to IB_CM_ASSIGN_SERVICE_ID, the CM will 1260 * assign a service ID to the caller. 1261 * 1262 * Callers should call ib_destroy_cm_id when done with the listener ID. 1263 */ 1264 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device, 1265 ib_cm_handler cm_handler, 1266 __be64 service_id) 1267 { 1268 struct cm_id_private *listen_id_priv; 1269 struct cm_id_private *cm_id_priv; 1270 int err = 0; 1271 1272 /* Create an ID in advance, since the creation may sleep */ 1273 cm_id_priv = cm_alloc_id_priv(device, cm_handler, NULL); 1274 if (IS_ERR(cm_id_priv)) 1275 return ERR_CAST(cm_id_priv); 1276 1277 err = cm_init_listen(cm_id_priv, service_id); 1278 if (err) { 1279 ib_destroy_cm_id(&cm_id_priv->id); 1280 return ERR_PTR(err); 1281 } 1282 1283 spin_lock_irq(&cm_id_priv->lock); 1284 listen_id_priv = cm_insert_listen(cm_id_priv, cm_handler); 1285 if (listen_id_priv != cm_id_priv) { 1286 spin_unlock_irq(&cm_id_priv->lock); 1287 ib_destroy_cm_id(&cm_id_priv->id); 1288 if (!listen_id_priv) 1289 return ERR_PTR(-EINVAL); 1290 return &listen_id_priv->id; 1291 } 1292 cm_id_priv->id.state = IB_CM_LISTEN; 1293 spin_unlock_irq(&cm_id_priv->lock); 1294 1295 /* 1296 * A listen ID does not need to be in the xarray since it does not 1297 * receive mads, is not placed in the remote_id or remote_qpn rbtree, 1298 * and does not enter timewait. 1299 */ 1300 1301 return &cm_id_priv->id; 1302 } 1303 EXPORT_SYMBOL(ib_cm_insert_listen); 1304 1305 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv) 1306 { 1307 u64 hi_tid = 0, low_tid; 1308 1309 lockdep_assert_held(&cm_id_priv->lock); 1310 1311 low_tid = (u64)cm_id_priv->id.local_id; 1312 if (!cm_id_priv->av.port) 1313 return cpu_to_be64(low_tid); 1314 1315 read_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 1316 if (cm_id_priv->av.port->mad_agent) 1317 hi_tid = ((u64)cm_id_priv->av.port->mad_agent->hi_tid) << 32; 1318 read_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock); 1319 return cpu_to_be64(hi_tid | low_tid); 1320 } 1321 1322 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr, 1323 __be16 attr_id, __be64 tid) 1324 { 1325 hdr->base_version = IB_MGMT_BASE_VERSION; 1326 hdr->mgmt_class = IB_MGMT_CLASS_CM; 1327 hdr->class_version = IB_CM_CLASS_VERSION; 1328 hdr->method = IB_MGMT_METHOD_SEND; 1329 hdr->attr_id = attr_id; 1330 hdr->tid = tid; 1331 } 1332 1333 static void cm_format_mad_ece_hdr(struct ib_mad_hdr *hdr, __be16 attr_id, 1334 __be64 tid, u32 attr_mod) 1335 { 1336 cm_format_mad_hdr(hdr, attr_id, tid); 1337 hdr->attr_mod = cpu_to_be32(attr_mod); 1338 } 1339 1340 static void cm_format_req(struct cm_req_msg *req_msg, 1341 struct cm_id_private *cm_id_priv, 1342 struct ib_cm_req_param *param) 1343 { 1344 struct sa_path_rec *pri_path = param->primary_path; 1345 struct sa_path_rec *alt_path = param->alternate_path; 1346 bool pri_ext = false; 1347 __be16 lid; 1348 1349 if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA) 1350 pri_ext = opa_is_extended_lid(pri_path->opa.dlid, 1351 pri_path->opa.slid); 1352 1353 cm_format_mad_ece_hdr(&req_msg->hdr, CM_REQ_ATTR_ID, 1354 cm_form_tid(cm_id_priv), param->ece.attr_mod); 1355 1356 IBA_SET(CM_REQ_LOCAL_COMM_ID, req_msg, 1357 be32_to_cpu(cm_id_priv->id.local_id)); 1358 IBA_SET(CM_REQ_SERVICE_ID, req_msg, be64_to_cpu(param->service_id)); 1359 IBA_SET(CM_REQ_LOCAL_CA_GUID, req_msg, 1360 be64_to_cpu(cm_id_priv->id.device->node_guid)); 1361 IBA_SET(CM_REQ_LOCAL_QPN, req_msg, param->qp_num); 1362 IBA_SET(CM_REQ_INITIATOR_DEPTH, req_msg, param->initiator_depth); 1363 IBA_SET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg, 1364 param->remote_cm_response_timeout); 1365 cm_req_set_qp_type(req_msg, param->qp_type); 1366 IBA_SET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg, param->flow_control); 1367 IBA_SET(CM_REQ_STARTING_PSN, req_msg, param->starting_psn); 1368 IBA_SET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg, 1369 param->local_cm_response_timeout); 1370 IBA_SET(CM_REQ_PARTITION_KEY, req_msg, 1371 be16_to_cpu(param->primary_path->pkey)); 1372 IBA_SET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg, 1373 param->primary_path->mtu); 1374 IBA_SET(CM_REQ_MAX_CM_RETRIES, req_msg, param->max_cm_retries); 1375 1376 if (param->qp_type != IB_QPT_XRC_INI) { 1377 IBA_SET(CM_REQ_RESPONDER_RESOURCES, req_msg, 1378 param->responder_resources); 1379 IBA_SET(CM_REQ_RETRY_COUNT, req_msg, param->retry_count); 1380 IBA_SET(CM_REQ_RNR_RETRY_COUNT, req_msg, 1381 param->rnr_retry_count); 1382 IBA_SET(CM_REQ_SRQ, req_msg, param->srq); 1383 } 1384 1385 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) = 1386 pri_path->sgid; 1387 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) = 1388 pri_path->dgid; 1389 if (pri_ext) { 1390 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) 1391 ->global.interface_id = 1392 OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid)); 1393 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) 1394 ->global.interface_id = 1395 OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid)); 1396 } 1397 if (pri_path->hop_limit <= 1) { 1398 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 1399 be16_to_cpu(pri_ext ? 0 : 1400 htons(ntohl(sa_path_get_slid( 1401 pri_path))))); 1402 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, 1403 be16_to_cpu(pri_ext ? 0 : 1404 htons(ntohl(sa_path_get_dlid( 1405 pri_path))))); 1406 } else { 1407 1408 if (param->primary_path_inbound) { 1409 lid = param->primary_path_inbound->ib.dlid; 1410 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 1411 be16_to_cpu(lid)); 1412 } else 1413 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 1414 be16_to_cpu(IB_LID_PERMISSIVE)); 1415 1416 /* Work-around until there's a way to obtain remote LID info */ 1417 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, 1418 be16_to_cpu(IB_LID_PERMISSIVE)); 1419 } 1420 IBA_SET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg, 1421 be32_to_cpu(pri_path->flow_label)); 1422 IBA_SET(CM_REQ_PRIMARY_PACKET_RATE, req_msg, pri_path->rate); 1423 IBA_SET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg, pri_path->traffic_class); 1424 IBA_SET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg, pri_path->hop_limit); 1425 IBA_SET(CM_REQ_PRIMARY_SL, req_msg, pri_path->sl); 1426 IBA_SET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg, 1427 (pri_path->hop_limit <= 1)); 1428 IBA_SET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg, 1429 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay, 1430 pri_path->packet_life_time)); 1431 1432 if (alt_path) { 1433 bool alt_ext = false; 1434 1435 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA) 1436 alt_ext = opa_is_extended_lid(alt_path->opa.dlid, 1437 alt_path->opa.slid); 1438 1439 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg) = 1440 alt_path->sgid; 1441 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg) = 1442 alt_path->dgid; 1443 if (alt_ext) { 1444 IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, 1445 req_msg) 1446 ->global.interface_id = 1447 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid)); 1448 IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, 1449 req_msg) 1450 ->global.interface_id = 1451 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid)); 1452 } 1453 if (alt_path->hop_limit <= 1) { 1454 IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg, 1455 be16_to_cpu( 1456 alt_ext ? 0 : 1457 htons(ntohl(sa_path_get_slid( 1458 alt_path))))); 1459 IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg, 1460 be16_to_cpu( 1461 alt_ext ? 0 : 1462 htons(ntohl(sa_path_get_dlid( 1463 alt_path))))); 1464 } else { 1465 IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg, 1466 be16_to_cpu(IB_LID_PERMISSIVE)); 1467 IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg, 1468 be16_to_cpu(IB_LID_PERMISSIVE)); 1469 } 1470 IBA_SET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg, 1471 be32_to_cpu(alt_path->flow_label)); 1472 IBA_SET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg, alt_path->rate); 1473 IBA_SET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg, 1474 alt_path->traffic_class); 1475 IBA_SET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg, 1476 alt_path->hop_limit); 1477 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, alt_path->sl); 1478 IBA_SET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg, 1479 (alt_path->hop_limit <= 1)); 1480 IBA_SET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg, 1481 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay, 1482 alt_path->packet_life_time)); 1483 } 1484 IBA_SET(CM_REQ_VENDOR_ID, req_msg, param->ece.vendor_id); 1485 1486 if (param->private_data && param->private_data_len) 1487 IBA_SET_MEM(CM_REQ_PRIVATE_DATA, req_msg, param->private_data, 1488 param->private_data_len); 1489 } 1490 1491 static int cm_validate_req_param(struct ib_cm_req_param *param) 1492 { 1493 if (!param->primary_path) 1494 return -EINVAL; 1495 1496 if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC && 1497 param->qp_type != IB_QPT_XRC_INI) 1498 return -EINVAL; 1499 1500 if (param->private_data && 1501 param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE) 1502 return -EINVAL; 1503 1504 if (param->alternate_path && 1505 (param->alternate_path->pkey != param->primary_path->pkey || 1506 param->alternate_path->mtu != param->primary_path->mtu)) 1507 return -EINVAL; 1508 1509 return 0; 1510 } 1511 1512 int ib_send_cm_req(struct ib_cm_id *cm_id, 1513 struct ib_cm_req_param *param) 1514 { 1515 struct cm_av av = {}, alt_av = {}; 1516 struct cm_id_private *cm_id_priv; 1517 struct ib_mad_send_buf *msg; 1518 struct cm_req_msg *req_msg; 1519 unsigned long flags; 1520 int ret; 1521 1522 ret = cm_validate_req_param(param); 1523 if (ret) 1524 return ret; 1525 1526 /* Verify that we're not in timewait. */ 1527 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 1528 spin_lock_irqsave(&cm_id_priv->lock, flags); 1529 if (cm_id->state != IB_CM_IDLE || WARN_ON(cm_id_priv->timewait_info)) { 1530 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1531 return -EINVAL; 1532 } 1533 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1534 1535 cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv-> 1536 id.local_id); 1537 if (IS_ERR(cm_id_priv->timewait_info)) { 1538 ret = PTR_ERR(cm_id_priv->timewait_info); 1539 cm_id_priv->timewait_info = NULL; 1540 return ret; 1541 } 1542 1543 ret = cm_init_av_by_path(param->primary_path, 1544 param->ppath_sgid_attr, &av); 1545 if (ret) 1546 return ret; 1547 if (param->alternate_path) { 1548 ret = cm_init_av_by_path(param->alternate_path, NULL, 1549 &alt_av); 1550 if (ret) { 1551 cm_destroy_av(&av); 1552 return ret; 1553 } 1554 } 1555 cm_id->service_id = param->service_id; 1556 cm_id_priv->timeout_ms = cm_convert_to_ms( 1557 param->primary_path->packet_life_time) * 2 + 1558 cm_convert_to_ms( 1559 param->remote_cm_response_timeout); 1560 cm_id_priv->max_cm_retries = param->max_cm_retries; 1561 cm_id_priv->initiator_depth = param->initiator_depth; 1562 cm_id_priv->responder_resources = param->responder_resources; 1563 cm_id_priv->retry_count = param->retry_count; 1564 cm_id_priv->path_mtu = param->primary_path->mtu; 1565 cm_id_priv->pkey = param->primary_path->pkey; 1566 cm_id_priv->qp_type = param->qp_type; 1567 1568 spin_lock_irqsave(&cm_id_priv->lock, flags); 1569 1570 cm_move_av_from_path(&cm_id_priv->av, &av); 1571 if (param->primary_path_outbound) 1572 cm_id_priv->av.dlid_datapath = 1573 be16_to_cpu(param->primary_path_outbound->ib.dlid); 1574 1575 if (param->alternate_path) 1576 cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av); 1577 1578 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_REQ_SENT); 1579 if (IS_ERR(msg)) { 1580 ret = PTR_ERR(msg); 1581 goto out_unlock; 1582 } 1583 1584 req_msg = (struct cm_req_msg *)msg->mad; 1585 cm_format_req(req_msg, cm_id_priv, param); 1586 cm_id_priv->tid = req_msg->hdr.tid; 1587 1588 cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg)); 1589 cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg)); 1590 1591 trace_icm_send_req(&cm_id_priv->id); 1592 ret = ib_post_send_mad(msg, NULL); 1593 if (ret) 1594 goto out_free; 1595 BUG_ON(cm_id->state != IB_CM_IDLE); 1596 cm_id->state = IB_CM_REQ_SENT; 1597 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1598 return 0; 1599 out_free: 1600 cm_free_priv_msg(msg); 1601 out_unlock: 1602 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 1603 return ret; 1604 } 1605 EXPORT_SYMBOL(ib_send_cm_req); 1606 1607 static int cm_issue_rej(struct cm_port *port, 1608 struct ib_mad_recv_wc *mad_recv_wc, 1609 enum ib_cm_rej_reason reason, 1610 enum cm_msg_response msg_rejected, 1611 void *ari, u8 ari_length) 1612 { 1613 struct ib_mad_send_buf *msg = NULL; 1614 struct cm_rej_msg *rej_msg, *rcv_msg; 1615 int ret; 1616 1617 ret = cm_alloc_response_msg(port, mad_recv_wc, false, &msg); 1618 if (ret) 1619 return ret; 1620 1621 /* We just need common CM header information. Cast to any message. */ 1622 rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad; 1623 rej_msg = (struct cm_rej_msg *) msg->mad; 1624 1625 cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid); 1626 IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg, 1627 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg)); 1628 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1629 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg)); 1630 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected); 1631 IBA_SET(CM_REJ_REASON, rej_msg, reason); 1632 1633 if (ari && ari_length) { 1634 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length); 1635 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length); 1636 } 1637 1638 trace_icm_issue_rej( 1639 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg), 1640 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg)); 1641 ret = ib_post_send_mad(msg, NULL); 1642 if (ret) 1643 cm_free_msg(msg); 1644 1645 return ret; 1646 } 1647 1648 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg) 1649 { 1650 return ((cpu_to_be16( 1651 IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) || 1652 (ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, 1653 req_msg)))); 1654 } 1655 1656 static void cm_path_set_rec_type(struct ib_device *ib_device, u32 port_num, 1657 struct sa_path_rec *path, union ib_gid *gid) 1658 { 1659 if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num)) 1660 path->rec_type = SA_PATH_REC_TYPE_OPA; 1661 else 1662 path->rec_type = SA_PATH_REC_TYPE_IB; 1663 } 1664 1665 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg, 1666 struct sa_path_rec *primary_path, 1667 struct sa_path_rec *alt_path, 1668 struct ib_wc *wc) 1669 { 1670 u32 lid; 1671 1672 if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) { 1673 sa_path_set_dlid(primary_path, wc->slid); 1674 sa_path_set_slid(primary_path, 1675 IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID, 1676 req_msg)); 1677 } else { 1678 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1679 CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg)); 1680 sa_path_set_dlid(primary_path, lid); 1681 1682 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1683 CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg)); 1684 sa_path_set_slid(primary_path, lid); 1685 } 1686 1687 if (!cm_req_has_alt_path(req_msg)) 1688 return; 1689 1690 if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) { 1691 sa_path_set_dlid(alt_path, 1692 IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, 1693 req_msg)); 1694 sa_path_set_slid(alt_path, 1695 IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, 1696 req_msg)); 1697 } else { 1698 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1699 CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg)); 1700 sa_path_set_dlid(alt_path, lid); 1701 1702 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 1703 CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg)); 1704 sa_path_set_slid(alt_path, lid); 1705 } 1706 } 1707 1708 static void cm_format_paths_from_req(struct cm_req_msg *req_msg, 1709 struct sa_path_rec *primary_path, 1710 struct sa_path_rec *alt_path, 1711 struct ib_wc *wc) 1712 { 1713 primary_path->dgid = 1714 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg); 1715 primary_path->sgid = 1716 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg); 1717 primary_path->flow_label = 1718 cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg)); 1719 primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg); 1720 primary_path->traffic_class = 1721 IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg); 1722 primary_path->reversible = 1; 1723 primary_path->pkey = 1724 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg)); 1725 primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg); 1726 primary_path->mtu_selector = IB_SA_EQ; 1727 primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg); 1728 primary_path->rate_selector = IB_SA_EQ; 1729 primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg); 1730 primary_path->packet_life_time_selector = IB_SA_EQ; 1731 primary_path->packet_life_time = 1732 IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg); 1733 primary_path->packet_life_time -= (primary_path->packet_life_time > 0); 1734 primary_path->service_id = 1735 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)); 1736 if (sa_path_is_roce(primary_path)) 1737 primary_path->roce.route_resolved = false; 1738 1739 if (cm_req_has_alt_path(req_msg)) { 1740 alt_path->dgid = *IBA_GET_MEM_PTR( 1741 CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg); 1742 alt_path->sgid = *IBA_GET_MEM_PTR( 1743 CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg); 1744 alt_path->flow_label = cpu_to_be32( 1745 IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg)); 1746 alt_path->hop_limit = 1747 IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg); 1748 alt_path->traffic_class = 1749 IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg); 1750 alt_path->reversible = 1; 1751 alt_path->pkey = 1752 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg)); 1753 alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg); 1754 alt_path->mtu_selector = IB_SA_EQ; 1755 alt_path->mtu = 1756 IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg); 1757 alt_path->rate_selector = IB_SA_EQ; 1758 alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg); 1759 alt_path->packet_life_time_selector = IB_SA_EQ; 1760 alt_path->packet_life_time = 1761 IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg); 1762 alt_path->packet_life_time -= (alt_path->packet_life_time > 0); 1763 alt_path->service_id = 1764 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)); 1765 1766 if (sa_path_is_roce(alt_path)) 1767 alt_path->roce.route_resolved = false; 1768 } 1769 cm_format_path_lid_from_req(req_msg, primary_path, alt_path, wc); 1770 } 1771 1772 static u16 cm_get_bth_pkey(struct cm_work *work) 1773 { 1774 struct ib_device *ib_dev = work->port->cm_dev->ib_device; 1775 u32 port_num = work->port->port_num; 1776 u16 pkey_index = work->mad_recv_wc->wc->pkey_index; 1777 u16 pkey; 1778 int ret; 1779 1780 ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey); 1781 if (ret) { 1782 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %u, pkey index %u). %d\n", 1783 port_num, pkey_index, ret); 1784 return 0; 1785 } 1786 1787 return pkey; 1788 } 1789 1790 /** 1791 * cm_opa_to_ib_sgid - Convert OPA SGID to IB SGID 1792 * ULPs (such as IPoIB) do not understand OPA GIDs and will 1793 * reject them as the local_gid will not match the sgid. Therefore, 1794 * change the pathrec's SGID to an IB SGID. 1795 * 1796 * @work: Work completion 1797 * @path: Path record 1798 */ 1799 static void cm_opa_to_ib_sgid(struct cm_work *work, 1800 struct sa_path_rec *path) 1801 { 1802 struct ib_device *dev = work->port->cm_dev->ib_device; 1803 u32 port_num = work->port->port_num; 1804 1805 if (rdma_cap_opa_ah(dev, port_num) && 1806 (ib_is_opa_gid(&path->sgid))) { 1807 union ib_gid sgid; 1808 1809 if (rdma_query_gid(dev, port_num, 0, &sgid)) { 1810 dev_warn(&dev->dev, 1811 "Error updating sgid in CM request\n"); 1812 return; 1813 } 1814 1815 path->sgid = sgid; 1816 } 1817 } 1818 1819 static void cm_format_req_event(struct cm_work *work, 1820 struct cm_id_private *cm_id_priv, 1821 struct ib_cm_id *listen_id) 1822 { 1823 struct cm_req_msg *req_msg; 1824 struct ib_cm_req_event_param *param; 1825 1826 req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; 1827 param = &work->cm_event.param.req_rcvd; 1828 param->listen_id = listen_id; 1829 param->bth_pkey = cm_get_bth_pkey(work); 1830 param->port = cm_id_priv->av.port->port_num; 1831 param->primary_path = &work->path[0]; 1832 cm_opa_to_ib_sgid(work, param->primary_path); 1833 if (cm_req_has_alt_path(req_msg)) { 1834 param->alternate_path = &work->path[1]; 1835 cm_opa_to_ib_sgid(work, param->alternate_path); 1836 } else { 1837 param->alternate_path = NULL; 1838 } 1839 param->remote_ca_guid = 1840 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg)); 1841 param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg); 1842 param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg); 1843 param->qp_type = cm_req_get_qp_type(req_msg); 1844 param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg); 1845 param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg); 1846 param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg); 1847 param->local_cm_response_timeout = 1848 IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg); 1849 param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg); 1850 param->remote_cm_response_timeout = 1851 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg); 1852 param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg); 1853 param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg); 1854 param->srq = IBA_GET(CM_REQ_SRQ, req_msg); 1855 param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr; 1856 param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg); 1857 param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod); 1858 1859 work->cm_event.private_data = 1860 IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg); 1861 } 1862 1863 static void cm_process_work(struct cm_id_private *cm_id_priv, 1864 struct cm_work *work) 1865 { 1866 int ret; 1867 1868 /* We will typically only have the current event to report. */ 1869 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event); 1870 cm_free_work(work); 1871 1872 while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) { 1873 spin_lock_irq(&cm_id_priv->lock); 1874 work = cm_dequeue_work(cm_id_priv); 1875 spin_unlock_irq(&cm_id_priv->lock); 1876 if (!work) 1877 return; 1878 1879 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, 1880 &work->cm_event); 1881 cm_free_work(work); 1882 } 1883 cm_deref_id(cm_id_priv); 1884 if (ret) 1885 cm_destroy_id(&cm_id_priv->id, ret); 1886 } 1887 1888 static void cm_format_mra(struct cm_mra_msg *mra_msg, 1889 struct cm_id_private *cm_id_priv, 1890 enum cm_msg_response msg_mraed, 1891 const void *private_data, u8 private_data_len) 1892 { 1893 cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid); 1894 IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed); 1895 IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg, 1896 be32_to_cpu(cm_id_priv->id.local_id)); 1897 IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg, 1898 be32_to_cpu(cm_id_priv->id.remote_id)); 1899 IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, CM_MRA_SETTING); 1900 1901 if (private_data && private_data_len) 1902 IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data, 1903 private_data_len); 1904 } 1905 1906 static void cm_format_rej(struct cm_rej_msg *rej_msg, 1907 struct cm_id_private *cm_id_priv, 1908 enum ib_cm_rej_reason reason, void *ari, 1909 u8 ari_length, const void *private_data, 1910 u8 private_data_len, enum ib_cm_state state) 1911 { 1912 lockdep_assert_held(&cm_id_priv->lock); 1913 1914 cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid); 1915 IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg, 1916 be32_to_cpu(cm_id_priv->id.remote_id)); 1917 1918 switch (state) { 1919 case IB_CM_REQ_RCVD: 1920 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0)); 1921 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ); 1922 break; 1923 case IB_CM_MRA_REQ_SENT: 1924 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1925 be32_to_cpu(cm_id_priv->id.local_id)); 1926 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ); 1927 break; 1928 case IB_CM_REP_RCVD: 1929 case IB_CM_MRA_REP_SENT: 1930 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1931 be32_to_cpu(cm_id_priv->id.local_id)); 1932 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP); 1933 break; 1934 default: 1935 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, 1936 be32_to_cpu(cm_id_priv->id.local_id)); 1937 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, 1938 CM_MSG_RESPONSE_OTHER); 1939 break; 1940 } 1941 1942 IBA_SET(CM_REJ_REASON, rej_msg, reason); 1943 if (ari && ari_length) { 1944 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length); 1945 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length); 1946 } 1947 1948 if (private_data && private_data_len) 1949 IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data, 1950 private_data_len); 1951 } 1952 1953 static void cm_dup_req_handler(struct cm_work *work, 1954 struct cm_id_private *cm_id_priv) 1955 { 1956 struct ib_mad_send_buf *msg = NULL; 1957 int ret; 1958 1959 atomic_long_inc( 1960 &work->port->counters[CM_RECV_DUPLICATES][CM_REQ_COUNTER]); 1961 1962 /* Quick state check to discard duplicate REQs. */ 1963 spin_lock_irq(&cm_id_priv->lock); 1964 if (cm_id_priv->id.state == IB_CM_REQ_RCVD) { 1965 spin_unlock_irq(&cm_id_priv->lock); 1966 return; 1967 } 1968 spin_unlock_irq(&cm_id_priv->lock); 1969 1970 ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, true, &msg); 1971 if (ret) 1972 return; 1973 1974 spin_lock_irq(&cm_id_priv->lock); 1975 switch (cm_id_priv->id.state) { 1976 case IB_CM_MRA_REQ_SENT: 1977 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 1978 CM_MSG_RESPONSE_REQ, 1979 cm_id_priv->private_data, 1980 cm_id_priv->private_data_len); 1981 break; 1982 case IB_CM_TIMEWAIT: 1983 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, 1984 IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0, 1985 IB_CM_TIMEWAIT); 1986 break; 1987 default: 1988 goto unlock; 1989 } 1990 spin_unlock_irq(&cm_id_priv->lock); 1991 1992 trace_icm_send_dup_req(&cm_id_priv->id); 1993 ret = ib_post_send_mad(msg, NULL); 1994 if (ret) 1995 goto free; 1996 return; 1997 1998 unlock: spin_unlock_irq(&cm_id_priv->lock); 1999 free: cm_free_msg(msg); 2000 } 2001 2002 static struct cm_id_private *cm_match_req(struct cm_work *work, 2003 struct cm_id_private *cm_id_priv) 2004 { 2005 struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv; 2006 struct cm_timewait_info *timewait_info; 2007 struct cm_req_msg *req_msg; 2008 2009 req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; 2010 2011 /* Check for possible duplicate REQ. */ 2012 spin_lock_irq(&cm.lock); 2013 timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info); 2014 if (timewait_info) { 2015 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 2016 timewait_info->work.remote_id); 2017 spin_unlock_irq(&cm.lock); 2018 if (cur_cm_id_priv) { 2019 cm_dup_req_handler(work, cur_cm_id_priv); 2020 cm_deref_id(cur_cm_id_priv); 2021 } 2022 return NULL; 2023 } 2024 2025 /* Check for stale connections. */ 2026 timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info); 2027 if (timewait_info) { 2028 cm_remove_remote(cm_id_priv); 2029 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 2030 timewait_info->work.remote_id); 2031 2032 spin_unlock_irq(&cm.lock); 2033 cm_issue_rej(work->port, work->mad_recv_wc, 2034 IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ, 2035 NULL, 0); 2036 if (cur_cm_id_priv) { 2037 ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0); 2038 cm_deref_id(cur_cm_id_priv); 2039 } 2040 return NULL; 2041 } 2042 2043 /* Find matching listen request. */ 2044 listen_cm_id_priv = cm_find_listen( 2045 cm_id_priv->id.device, 2046 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg))); 2047 if (!listen_cm_id_priv) { 2048 cm_remove_remote(cm_id_priv); 2049 spin_unlock_irq(&cm.lock); 2050 cm_issue_rej(work->port, work->mad_recv_wc, 2051 IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ, 2052 NULL, 0); 2053 return NULL; 2054 } 2055 spin_unlock_irq(&cm.lock); 2056 return listen_cm_id_priv; 2057 } 2058 2059 /* 2060 * Work-around for inter-subnet connections. If the LIDs are permissive, 2061 * we need to override the LID/SL data in the REQ with the LID information 2062 * in the work completion. 2063 */ 2064 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc) 2065 { 2066 if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) { 2067 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, 2068 req_msg)) == IB_LID_PERMISSIVE) { 2069 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg, 2070 be16_to_cpu(ib_lid_be16(wc->slid))); 2071 IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl); 2072 } 2073 2074 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID, 2075 req_msg)) == IB_LID_PERMISSIVE) 2076 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg, 2077 wc->dlid_path_bits); 2078 } 2079 2080 if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) { 2081 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, 2082 req_msg)) == IB_LID_PERMISSIVE) { 2083 IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg, 2084 be16_to_cpu(ib_lid_be16(wc->slid))); 2085 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl); 2086 } 2087 2088 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, 2089 req_msg)) == IB_LID_PERMISSIVE) 2090 IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg, 2091 wc->dlid_path_bits); 2092 } 2093 } 2094 2095 static int cm_req_handler(struct cm_work *work) 2096 { 2097 struct cm_id_private *cm_id_priv, *listen_cm_id_priv; 2098 struct cm_req_msg *req_msg; 2099 const struct ib_global_route *grh; 2100 const struct ib_gid_attr *gid_attr; 2101 int ret; 2102 2103 req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad; 2104 2105 cm_id_priv = 2106 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL); 2107 if (IS_ERR(cm_id_priv)) 2108 return PTR_ERR(cm_id_priv); 2109 2110 cm_id_priv->id.remote_id = 2111 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg)); 2112 cm_id_priv->id.service_id = 2113 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)); 2114 cm_id_priv->tid = req_msg->hdr.tid; 2115 cm_id_priv->timeout_ms = cm_convert_to_ms( 2116 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg)); 2117 cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg); 2118 cm_id_priv->remote_qpn = 2119 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg)); 2120 cm_id_priv->initiator_depth = 2121 IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg); 2122 cm_id_priv->responder_resources = 2123 IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg); 2124 cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg); 2125 cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg)); 2126 cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg)); 2127 cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg); 2128 cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg); 2129 cm_id_priv->qp_type = cm_req_get_qp_type(req_msg); 2130 2131 ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc, 2132 work->mad_recv_wc->recv_buf.grh, 2133 &cm_id_priv->av); 2134 if (ret) 2135 goto destroy; 2136 cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv-> 2137 id.local_id); 2138 if (IS_ERR(cm_id_priv->timewait_info)) { 2139 ret = PTR_ERR(cm_id_priv->timewait_info); 2140 cm_id_priv->timewait_info = NULL; 2141 goto destroy; 2142 } 2143 cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id; 2144 cm_id_priv->timewait_info->remote_ca_guid = 2145 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg)); 2146 cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn; 2147 2148 /* 2149 * Note that the ID pointer is not in the xarray at this point, 2150 * so this set is only visible to the local thread. 2151 */ 2152 cm_id_priv->id.state = IB_CM_REQ_RCVD; 2153 2154 listen_cm_id_priv = cm_match_req(work, cm_id_priv); 2155 if (!listen_cm_id_priv) { 2156 trace_icm_no_listener_err(&cm_id_priv->id); 2157 cm_id_priv->id.state = IB_CM_IDLE; 2158 ret = -EINVAL; 2159 goto destroy; 2160 } 2161 2162 memset(&work->path[0], 0, sizeof(work->path[0])); 2163 if (cm_req_has_alt_path(req_msg)) 2164 memset(&work->path[1], 0, sizeof(work->path[1])); 2165 grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr); 2166 gid_attr = grh->sgid_attr; 2167 2168 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE) { 2169 work->path[0].rec_type = 2170 sa_conv_gid_to_pathrec_type(gid_attr->gid_type); 2171 } else { 2172 cm_process_routed_req(req_msg, work->mad_recv_wc->wc); 2173 cm_path_set_rec_type( 2174 work->port->cm_dev->ib_device, work->port->port_num, 2175 &work->path[0], 2176 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, 2177 req_msg)); 2178 } 2179 if (cm_req_has_alt_path(req_msg)) 2180 work->path[1].rec_type = work->path[0].rec_type; 2181 cm_format_paths_from_req(req_msg, &work->path[0], 2182 &work->path[1], work->mad_recv_wc->wc); 2183 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE) 2184 sa_path_set_dmac(&work->path[0], 2185 cm_id_priv->av.ah_attr.roce.dmac); 2186 work->path[0].hop_limit = grh->hop_limit; 2187 2188 /* 2189 * cm_init_av_by_path() will internally pair with the above 2190 * cm_init_av_for_response() if it succeeds. 2191 */ 2192 ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av); 2193 if (ret) { 2194 int err; 2195 2196 err = rdma_query_gid(work->port->cm_dev->ib_device, 2197 work->port->port_num, 0, 2198 &work->path[0].sgid); 2199 if (err) 2200 ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID, 2201 NULL, 0, NULL, 0); 2202 else 2203 ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID, 2204 &work->path[0].sgid, 2205 sizeof(work->path[0].sgid), 2206 NULL, 0); 2207 goto rejected; 2208 } 2209 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_IB) 2210 cm_id_priv->av.dlid_datapath = 2211 IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg); 2212 2213 if (cm_req_has_alt_path(req_msg)) { 2214 ret = cm_init_av_by_path(&work->path[1], NULL, 2215 &cm_id_priv->alt_av); 2216 if (ret) { 2217 ib_send_cm_rej(&cm_id_priv->id, 2218 IB_CM_REJ_INVALID_ALT_GID, 2219 &work->path[0].sgid, 2220 sizeof(work->path[0].sgid), NULL, 0); 2221 goto rejected; 2222 } 2223 } 2224 2225 cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler; 2226 cm_id_priv->id.context = listen_cm_id_priv->id.context; 2227 cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id); 2228 2229 /* Now MAD handlers can see the new ID */ 2230 spin_lock_irq(&cm_id_priv->lock); 2231 cm_finalize_id(cm_id_priv); 2232 2233 /* Refcount belongs to the event, pairs with cm_process_work() */ 2234 refcount_inc(&cm_id_priv->refcount); 2235 cm_queue_work_unlock(cm_id_priv, work); 2236 /* 2237 * Since this ID was just created and was not made visible to other MAD 2238 * handlers until the cm_finalize_id() above we know that the 2239 * cm_process_work() will deliver the event and the listen_cm_id 2240 * embedded in the event can be derefed here. 2241 */ 2242 cm_deref_id(listen_cm_id_priv); 2243 return 0; 2244 2245 rejected: 2246 cm_deref_id(listen_cm_id_priv); 2247 destroy: 2248 ib_destroy_cm_id(&cm_id_priv->id); 2249 return ret; 2250 } 2251 2252 static void cm_format_rep(struct cm_rep_msg *rep_msg, 2253 struct cm_id_private *cm_id_priv, 2254 struct ib_cm_rep_param *param) 2255 { 2256 cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid, 2257 param->ece.attr_mod); 2258 IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg, 2259 be32_to_cpu(cm_id_priv->id.local_id)); 2260 IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg, 2261 be32_to_cpu(cm_id_priv->id.remote_id)); 2262 IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn); 2263 IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg, 2264 param->responder_resources); 2265 IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg, 2266 cm_id_priv->av.port->cm_dev->ack_delay); 2267 IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted); 2268 IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count); 2269 IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg, 2270 be64_to_cpu(cm_id_priv->id.device->node_guid)); 2271 2272 if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) { 2273 IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg, 2274 param->initiator_depth); 2275 IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg, 2276 param->flow_control); 2277 IBA_SET(CM_REP_SRQ, rep_msg, param->srq); 2278 IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num); 2279 } else { 2280 IBA_SET(CM_REP_SRQ, rep_msg, 1); 2281 IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num); 2282 } 2283 2284 IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id); 2285 IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8); 2286 IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16); 2287 2288 if (param->private_data && param->private_data_len) 2289 IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data, 2290 param->private_data_len); 2291 } 2292 2293 int ib_send_cm_rep(struct ib_cm_id *cm_id, 2294 struct ib_cm_rep_param *param) 2295 { 2296 struct cm_id_private *cm_id_priv; 2297 struct ib_mad_send_buf *msg; 2298 struct cm_rep_msg *rep_msg; 2299 unsigned long flags; 2300 int ret; 2301 2302 if (param->private_data && 2303 param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE) 2304 return -EINVAL; 2305 2306 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 2307 spin_lock_irqsave(&cm_id_priv->lock, flags); 2308 if (cm_id->state != IB_CM_REQ_RCVD && 2309 cm_id->state != IB_CM_MRA_REQ_SENT) { 2310 trace_icm_send_rep_err(cm_id_priv->id.local_id, cm_id->state); 2311 ret = -EINVAL; 2312 goto out; 2313 } 2314 2315 msg = cm_alloc_priv_msg_rep(cm_id_priv, IB_CM_REP_SENT, true); 2316 if (IS_ERR(msg)) { 2317 ret = PTR_ERR(msg); 2318 goto out; 2319 } 2320 2321 rep_msg = (struct cm_rep_msg *) msg->mad; 2322 cm_format_rep(rep_msg, cm_id_priv, param); 2323 2324 trace_icm_send_rep(cm_id); 2325 ret = ib_post_send_mad(msg, NULL); 2326 if (ret) 2327 goto out_free; 2328 2329 cm_id->state = IB_CM_REP_SENT; 2330 cm_id_priv->initiator_depth = param->initiator_depth; 2331 cm_id_priv->responder_resources = param->responder_resources; 2332 cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg)); 2333 WARN_ONCE(param->qp_num & 0xFF000000, 2334 "IBTA declares QPN to be 24 bits, but it is 0x%X\n", 2335 param->qp_num); 2336 cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF); 2337 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2338 return 0; 2339 2340 out_free: 2341 cm_free_priv_msg(msg); 2342 out: 2343 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2344 return ret; 2345 } 2346 EXPORT_SYMBOL(ib_send_cm_rep); 2347 2348 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg, 2349 struct cm_id_private *cm_id_priv, 2350 const void *private_data, 2351 u8 private_data_len) 2352 { 2353 cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid); 2354 IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg, 2355 be32_to_cpu(cm_id_priv->id.local_id)); 2356 IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg, 2357 be32_to_cpu(cm_id_priv->id.remote_id)); 2358 2359 if (private_data && private_data_len) 2360 IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data, 2361 private_data_len); 2362 } 2363 2364 int ib_send_cm_rtu(struct ib_cm_id *cm_id, 2365 const void *private_data, 2366 u8 private_data_len) 2367 { 2368 struct cm_id_private *cm_id_priv; 2369 struct ib_mad_send_buf *msg; 2370 unsigned long flags; 2371 void *data; 2372 int ret; 2373 2374 if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE) 2375 return -EINVAL; 2376 2377 data = cm_copy_private_data(private_data, private_data_len); 2378 if (IS_ERR(data)) 2379 return PTR_ERR(data); 2380 2381 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 2382 spin_lock_irqsave(&cm_id_priv->lock, flags); 2383 if (cm_id->state != IB_CM_REP_RCVD && 2384 cm_id->state != IB_CM_MRA_REP_SENT) { 2385 trace_icm_send_cm_rtu_err(cm_id); 2386 ret = -EINVAL; 2387 goto error; 2388 } 2389 2390 msg = cm_alloc_msg(cm_id_priv); 2391 if (IS_ERR(msg)) { 2392 ret = PTR_ERR(msg); 2393 goto error; 2394 } 2395 2396 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv, 2397 private_data, private_data_len); 2398 2399 trace_icm_send_rtu(cm_id); 2400 ret = ib_post_send_mad(msg, NULL); 2401 if (ret) { 2402 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2403 cm_free_msg(msg); 2404 kfree(data); 2405 return ret; 2406 } 2407 2408 cm_id->state = IB_CM_ESTABLISHED; 2409 cm_set_private_data(cm_id_priv, data, private_data_len); 2410 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2411 return 0; 2412 2413 error: spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2414 kfree(data); 2415 return ret; 2416 } 2417 EXPORT_SYMBOL(ib_send_cm_rtu); 2418 2419 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type) 2420 { 2421 struct cm_rep_msg *rep_msg; 2422 struct ib_cm_rep_event_param *param; 2423 2424 rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad; 2425 param = &work->cm_event.param.rep_rcvd; 2426 param->remote_ca_guid = 2427 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg)); 2428 param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg); 2429 param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type)); 2430 param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg); 2431 param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg); 2432 param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg); 2433 param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg); 2434 param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg); 2435 param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg); 2436 param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg); 2437 param->srq = IBA_GET(CM_REP_SRQ, rep_msg); 2438 param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16; 2439 param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8; 2440 param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg); 2441 param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod); 2442 2443 work->cm_event.private_data = 2444 IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg); 2445 } 2446 2447 static void cm_dup_rep_handler(struct cm_work *work) 2448 { 2449 struct cm_id_private *cm_id_priv; 2450 struct cm_rep_msg *rep_msg; 2451 struct ib_mad_send_buf *msg = NULL; 2452 int ret; 2453 2454 rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad; 2455 cm_id_priv = cm_acquire_id( 2456 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 2457 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg))); 2458 if (!cm_id_priv) 2459 return; 2460 2461 atomic_long_inc( 2462 &work->port->counters[CM_RECV_DUPLICATES][CM_REP_COUNTER]); 2463 ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, true, &msg); 2464 if (ret) 2465 goto deref; 2466 2467 spin_lock_irq(&cm_id_priv->lock); 2468 if (cm_id_priv->id.state == IB_CM_ESTABLISHED) 2469 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv, 2470 cm_id_priv->private_data, 2471 cm_id_priv->private_data_len); 2472 else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT) 2473 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 2474 CM_MSG_RESPONSE_REP, 2475 cm_id_priv->private_data, 2476 cm_id_priv->private_data_len); 2477 else 2478 goto unlock; 2479 spin_unlock_irq(&cm_id_priv->lock); 2480 2481 trace_icm_send_dup_rep(&cm_id_priv->id); 2482 ret = ib_post_send_mad(msg, NULL); 2483 if (ret) 2484 goto free; 2485 goto deref; 2486 2487 unlock: spin_unlock_irq(&cm_id_priv->lock); 2488 free: cm_free_msg(msg); 2489 deref: cm_deref_id(cm_id_priv); 2490 } 2491 2492 static int cm_rep_handler(struct cm_work *work) 2493 { 2494 struct cm_id_private *cm_id_priv; 2495 struct cm_rep_msg *rep_msg; 2496 int ret; 2497 struct cm_id_private *cur_cm_id_priv; 2498 struct cm_timewait_info *timewait_info; 2499 2500 rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad; 2501 cm_id_priv = cm_acquire_id( 2502 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0); 2503 if (!cm_id_priv) { 2504 cm_dup_rep_handler(work); 2505 trace_icm_remote_no_priv_err( 2506 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)); 2507 return -EINVAL; 2508 } 2509 2510 cm_format_rep_event(work, cm_id_priv->qp_type); 2511 2512 spin_lock_irq(&cm_id_priv->lock); 2513 switch (cm_id_priv->id.state) { 2514 case IB_CM_REQ_SENT: 2515 case IB_CM_MRA_REQ_RCVD: 2516 break; 2517 default: 2518 ret = -EINVAL; 2519 trace_icm_rep_unknown_err( 2520 IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg), 2521 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg), 2522 cm_id_priv->id.state); 2523 spin_unlock_irq(&cm_id_priv->lock); 2524 goto error; 2525 } 2526 2527 cm_id_priv->timewait_info->work.remote_id = 2528 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)); 2529 cm_id_priv->timewait_info->remote_ca_guid = 2530 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg)); 2531 cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type); 2532 2533 spin_lock(&cm.lock); 2534 /* Check for duplicate REP. */ 2535 if (cm_insert_remote_id(cm_id_priv->timewait_info)) { 2536 spin_unlock(&cm.lock); 2537 spin_unlock_irq(&cm_id_priv->lock); 2538 ret = -EINVAL; 2539 trace_icm_insert_failed_err( 2540 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)); 2541 goto error; 2542 } 2543 /* Check for a stale connection. */ 2544 timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info); 2545 if (timewait_info) { 2546 cm_remove_remote(cm_id_priv); 2547 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 2548 timewait_info->work.remote_id); 2549 2550 spin_unlock(&cm.lock); 2551 spin_unlock_irq(&cm_id_priv->lock); 2552 cm_issue_rej(work->port, work->mad_recv_wc, 2553 IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP, 2554 NULL, 0); 2555 ret = -EINVAL; 2556 trace_icm_staleconn_err( 2557 IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg), 2558 IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)); 2559 2560 if (cur_cm_id_priv) { 2561 ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0); 2562 cm_deref_id(cur_cm_id_priv); 2563 } 2564 2565 goto error; 2566 } 2567 spin_unlock(&cm.lock); 2568 2569 cm_id_priv->id.state = IB_CM_REP_RCVD; 2570 cm_id_priv->id.remote_id = 2571 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)); 2572 cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type); 2573 cm_id_priv->initiator_depth = 2574 IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg); 2575 cm_id_priv->responder_resources = 2576 IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg); 2577 cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg)); 2578 cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg); 2579 cm_id_priv->target_ack_delay = 2580 IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg); 2581 cm_id_priv->av.timeout = 2582 cm_ack_timeout(cm_id_priv->target_ack_delay, 2583 cm_id_priv->av.timeout - 1); 2584 cm_id_priv->alt_av.timeout = 2585 cm_ack_timeout(cm_id_priv->target_ack_delay, 2586 cm_id_priv->alt_av.timeout - 1); 2587 2588 ib_cancel_mad(cm_id_priv->msg); 2589 cm_queue_work_unlock(cm_id_priv, work); 2590 return 0; 2591 2592 error: 2593 cm_deref_id(cm_id_priv); 2594 return ret; 2595 } 2596 2597 static int cm_establish_handler(struct cm_work *work) 2598 { 2599 struct cm_id_private *cm_id_priv; 2600 2601 /* See comment in cm_establish about lookup. */ 2602 cm_id_priv = cm_acquire_id(work->local_id, work->remote_id); 2603 if (!cm_id_priv) 2604 return -EINVAL; 2605 2606 spin_lock_irq(&cm_id_priv->lock); 2607 if (cm_id_priv->id.state != IB_CM_ESTABLISHED) { 2608 spin_unlock_irq(&cm_id_priv->lock); 2609 goto out; 2610 } 2611 2612 ib_cancel_mad(cm_id_priv->msg); 2613 cm_queue_work_unlock(cm_id_priv, work); 2614 return 0; 2615 out: 2616 cm_deref_id(cm_id_priv); 2617 return -EINVAL; 2618 } 2619 2620 static int cm_rtu_handler(struct cm_work *work) 2621 { 2622 struct cm_id_private *cm_id_priv; 2623 struct cm_rtu_msg *rtu_msg; 2624 2625 rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad; 2626 cm_id_priv = cm_acquire_id( 2627 cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)), 2628 cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg))); 2629 if (!cm_id_priv) 2630 return -EINVAL; 2631 2632 work->cm_event.private_data = 2633 IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg); 2634 2635 spin_lock_irq(&cm_id_priv->lock); 2636 if (cm_id_priv->id.state != IB_CM_REP_SENT && 2637 cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) { 2638 spin_unlock_irq(&cm_id_priv->lock); 2639 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2640 [CM_RTU_COUNTER]); 2641 goto out; 2642 } 2643 cm_id_priv->id.state = IB_CM_ESTABLISHED; 2644 2645 ib_cancel_mad(cm_id_priv->msg); 2646 cm_queue_work_unlock(cm_id_priv, work); 2647 return 0; 2648 out: 2649 cm_deref_id(cm_id_priv); 2650 return -EINVAL; 2651 } 2652 2653 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg, 2654 struct cm_id_private *cm_id_priv, 2655 const void *private_data, 2656 u8 private_data_len) 2657 { 2658 cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID, 2659 cm_form_tid(cm_id_priv)); 2660 IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg, 2661 be32_to_cpu(cm_id_priv->id.local_id)); 2662 IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg, 2663 be32_to_cpu(cm_id_priv->id.remote_id)); 2664 IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg, 2665 be32_to_cpu(cm_id_priv->remote_qpn)); 2666 2667 if (private_data && private_data_len) 2668 IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data, 2669 private_data_len); 2670 } 2671 2672 static void cm_issue_dreq(struct cm_id_private *cm_id_priv) 2673 { 2674 struct ib_mad_send_buf *msg; 2675 int ret; 2676 2677 lockdep_assert_held(&cm_id_priv->lock); 2678 2679 msg = cm_alloc_msg(cm_id_priv); 2680 if (IS_ERR(msg)) 2681 return; 2682 2683 cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv, NULL, 0); 2684 2685 trace_icm_send_dreq(&cm_id_priv->id); 2686 ret = ib_post_send_mad(msg, NULL); 2687 if (ret) 2688 cm_free_msg(msg); 2689 } 2690 2691 int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data, 2692 u8 private_data_len) 2693 { 2694 struct cm_id_private *cm_id_priv = 2695 container_of(cm_id, struct cm_id_private, id); 2696 struct ib_mad_send_buf *msg; 2697 unsigned long flags; 2698 int ret; 2699 2700 if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE) 2701 return -EINVAL; 2702 2703 spin_lock_irqsave(&cm_id_priv->lock, flags); 2704 if (cm_id_priv->id.state != IB_CM_ESTABLISHED) { 2705 trace_icm_dreq_skipped(&cm_id_priv->id); 2706 ret = -EINVAL; 2707 goto unlock; 2708 } 2709 2710 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT || 2711 cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD) 2712 ib_cancel_mad(cm_id_priv->msg); 2713 2714 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_DREQ_SENT); 2715 if (IS_ERR(msg)) { 2716 cm_enter_timewait(cm_id_priv); 2717 ret = PTR_ERR(msg); 2718 goto unlock; 2719 } 2720 2721 cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv, 2722 private_data, private_data_len); 2723 2724 trace_icm_send_dreq(&cm_id_priv->id); 2725 ret = ib_post_send_mad(msg, NULL); 2726 if (ret) { 2727 cm_enter_timewait(cm_id_priv); 2728 cm_free_priv_msg(msg); 2729 goto unlock; 2730 } 2731 2732 cm_id_priv->id.state = IB_CM_DREQ_SENT; 2733 unlock: 2734 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2735 return ret; 2736 } 2737 EXPORT_SYMBOL(ib_send_cm_dreq); 2738 2739 static void cm_format_drep(struct cm_drep_msg *drep_msg, 2740 struct cm_id_private *cm_id_priv, 2741 const void *private_data, 2742 u8 private_data_len) 2743 { 2744 cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid); 2745 IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg, 2746 be32_to_cpu(cm_id_priv->id.local_id)); 2747 IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg, 2748 be32_to_cpu(cm_id_priv->id.remote_id)); 2749 2750 if (private_data && private_data_len) 2751 IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data, 2752 private_data_len); 2753 } 2754 2755 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv, 2756 void *private_data, u8 private_data_len) 2757 { 2758 struct ib_mad_send_buf *msg; 2759 int ret; 2760 2761 lockdep_assert_held(&cm_id_priv->lock); 2762 2763 if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE) 2764 return -EINVAL; 2765 2766 if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) { 2767 trace_icm_send_drep_err(&cm_id_priv->id); 2768 kfree(private_data); 2769 return -EINVAL; 2770 } 2771 2772 cm_set_private_data(cm_id_priv, private_data, private_data_len); 2773 cm_enter_timewait(cm_id_priv); 2774 2775 msg = cm_alloc_msg(cm_id_priv); 2776 if (IS_ERR(msg)) 2777 return PTR_ERR(msg); 2778 2779 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv, 2780 private_data, private_data_len); 2781 2782 trace_icm_send_drep(&cm_id_priv->id); 2783 ret = ib_post_send_mad(msg, NULL); 2784 if (ret) { 2785 cm_free_msg(msg); 2786 return ret; 2787 } 2788 return 0; 2789 } 2790 2791 int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data, 2792 u8 private_data_len) 2793 { 2794 struct cm_id_private *cm_id_priv = 2795 container_of(cm_id, struct cm_id_private, id); 2796 unsigned long flags; 2797 void *data; 2798 int ret; 2799 2800 data = cm_copy_private_data(private_data, private_data_len); 2801 if (IS_ERR(data)) 2802 return PTR_ERR(data); 2803 2804 spin_lock_irqsave(&cm_id_priv->lock, flags); 2805 ret = cm_send_drep_locked(cm_id_priv, data, private_data_len); 2806 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 2807 return ret; 2808 } 2809 EXPORT_SYMBOL(ib_send_cm_drep); 2810 2811 static int cm_issue_drep(struct cm_port *port, 2812 struct ib_mad_recv_wc *mad_recv_wc) 2813 { 2814 struct ib_mad_send_buf *msg = NULL; 2815 struct cm_dreq_msg *dreq_msg; 2816 struct cm_drep_msg *drep_msg; 2817 int ret; 2818 2819 ret = cm_alloc_response_msg(port, mad_recv_wc, true, &msg); 2820 if (ret) 2821 return ret; 2822 2823 dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad; 2824 drep_msg = (struct cm_drep_msg *) msg->mad; 2825 2826 cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid); 2827 IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg, 2828 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)); 2829 IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg, 2830 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)); 2831 2832 trace_icm_issue_drep( 2833 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg), 2834 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)); 2835 ret = ib_post_send_mad(msg, NULL); 2836 if (ret) 2837 cm_free_msg(msg); 2838 2839 return ret; 2840 } 2841 2842 static int cm_dreq_handler(struct cm_work *work) 2843 { 2844 struct cm_id_private *cm_id_priv; 2845 struct cm_dreq_msg *dreq_msg; 2846 struct ib_mad_send_buf *msg = NULL; 2847 2848 dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad; 2849 cm_id_priv = cm_acquire_id( 2850 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)), 2851 cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg))); 2852 if (!cm_id_priv) { 2853 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2854 [CM_DREQ_COUNTER]); 2855 cm_issue_drep(work->port, work->mad_recv_wc); 2856 trace_icm_no_priv_err( 2857 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg), 2858 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)); 2859 return -EINVAL; 2860 } 2861 2862 work->cm_event.private_data = 2863 IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg); 2864 2865 spin_lock_irq(&cm_id_priv->lock); 2866 if (cm_id_priv->local_qpn != 2867 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg))) 2868 goto unlock; 2869 2870 switch (cm_id_priv->id.state) { 2871 case IB_CM_REP_SENT: 2872 case IB_CM_DREQ_SENT: 2873 case IB_CM_MRA_REP_RCVD: 2874 ib_cancel_mad(cm_id_priv->msg); 2875 break; 2876 case IB_CM_ESTABLISHED: 2877 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT || 2878 cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD) 2879 ib_cancel_mad(cm_id_priv->msg); 2880 break; 2881 case IB_CM_TIMEWAIT: 2882 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2883 [CM_DREQ_COUNTER]); 2884 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc, 2885 true); 2886 if (IS_ERR(msg)) 2887 goto unlock; 2888 2889 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv, 2890 cm_id_priv->private_data, 2891 cm_id_priv->private_data_len); 2892 spin_unlock_irq(&cm_id_priv->lock); 2893 2894 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) || 2895 ib_post_send_mad(msg, NULL)) 2896 cm_free_msg(msg); 2897 goto deref; 2898 case IB_CM_DREQ_RCVD: 2899 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 2900 [CM_DREQ_COUNTER]); 2901 goto unlock; 2902 default: 2903 trace_icm_dreq_unknown_err(&cm_id_priv->id); 2904 goto unlock; 2905 } 2906 cm_id_priv->id.state = IB_CM_DREQ_RCVD; 2907 cm_id_priv->tid = dreq_msg->hdr.tid; 2908 cm_queue_work_unlock(cm_id_priv, work); 2909 return 0; 2910 2911 unlock: spin_unlock_irq(&cm_id_priv->lock); 2912 deref: cm_deref_id(cm_id_priv); 2913 return -EINVAL; 2914 } 2915 2916 static int cm_drep_handler(struct cm_work *work) 2917 { 2918 struct cm_id_private *cm_id_priv; 2919 struct cm_drep_msg *drep_msg; 2920 2921 drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad; 2922 cm_id_priv = cm_acquire_id( 2923 cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)), 2924 cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg))); 2925 if (!cm_id_priv) 2926 return -EINVAL; 2927 2928 work->cm_event.private_data = 2929 IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg); 2930 2931 spin_lock_irq(&cm_id_priv->lock); 2932 if (cm_id_priv->id.state != IB_CM_DREQ_SENT && 2933 cm_id_priv->id.state != IB_CM_DREQ_RCVD) { 2934 spin_unlock_irq(&cm_id_priv->lock); 2935 goto out; 2936 } 2937 cm_enter_timewait(cm_id_priv); 2938 2939 ib_cancel_mad(cm_id_priv->msg); 2940 cm_queue_work_unlock(cm_id_priv, work); 2941 return 0; 2942 out: 2943 cm_deref_id(cm_id_priv); 2944 return -EINVAL; 2945 } 2946 2947 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv, 2948 enum ib_cm_rej_reason reason, void *ari, 2949 u8 ari_length, const void *private_data, 2950 u8 private_data_len) 2951 { 2952 enum ib_cm_state state = cm_id_priv->id.state; 2953 struct ib_mad_send_buf *msg; 2954 int ret; 2955 2956 lockdep_assert_held(&cm_id_priv->lock); 2957 2958 if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) || 2959 (ari && ari_length > IB_CM_REJ_ARI_LENGTH)) 2960 return -EINVAL; 2961 2962 trace_icm_send_rej(&cm_id_priv->id, reason); 2963 2964 switch (state) { 2965 case IB_CM_REQ_SENT: 2966 case IB_CM_MRA_REQ_RCVD: 2967 case IB_CM_REQ_RCVD: 2968 case IB_CM_MRA_REQ_SENT: 2969 case IB_CM_REP_RCVD: 2970 case IB_CM_MRA_REP_SENT: 2971 cm_reset_to_idle(cm_id_priv); 2972 msg = cm_alloc_msg(cm_id_priv); 2973 if (IS_ERR(msg)) 2974 return PTR_ERR(msg); 2975 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason, 2976 ari, ari_length, private_data, private_data_len, 2977 state); 2978 break; 2979 case IB_CM_REP_SENT: 2980 case IB_CM_MRA_REP_RCVD: 2981 cm_enter_timewait(cm_id_priv); 2982 msg = cm_alloc_msg(cm_id_priv); 2983 if (IS_ERR(msg)) 2984 return PTR_ERR(msg); 2985 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason, 2986 ari, ari_length, private_data, private_data_len, 2987 state); 2988 break; 2989 default: 2990 trace_icm_send_unknown_rej_err(&cm_id_priv->id); 2991 return -EINVAL; 2992 } 2993 2994 ret = ib_post_send_mad(msg, NULL); 2995 if (ret) { 2996 cm_free_msg(msg); 2997 return ret; 2998 } 2999 3000 return 0; 3001 } 3002 3003 int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason, 3004 void *ari, u8 ari_length, const void *private_data, 3005 u8 private_data_len) 3006 { 3007 struct cm_id_private *cm_id_priv = 3008 container_of(cm_id, struct cm_id_private, id); 3009 unsigned long flags; 3010 int ret; 3011 3012 spin_lock_irqsave(&cm_id_priv->lock, flags); 3013 ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length, 3014 private_data, private_data_len); 3015 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3016 return ret; 3017 } 3018 EXPORT_SYMBOL(ib_send_cm_rej); 3019 3020 static void cm_format_rej_event(struct cm_work *work) 3021 { 3022 struct cm_rej_msg *rej_msg; 3023 struct ib_cm_rej_event_param *param; 3024 3025 rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad; 3026 param = &work->cm_event.param.rej_rcvd; 3027 param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg); 3028 param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg); 3029 param->reason = IBA_GET(CM_REJ_REASON, rej_msg); 3030 work->cm_event.private_data = 3031 IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg); 3032 } 3033 3034 static struct cm_id_private *cm_acquire_rejected_id(struct cm_rej_msg *rej_msg) 3035 { 3036 struct cm_id_private *cm_id_priv; 3037 __be32 remote_id; 3038 3039 remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg)); 3040 3041 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) { 3042 cm_id_priv = cm_find_remote_id( 3043 *((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)), 3044 remote_id); 3045 } else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) == 3046 CM_MSG_RESPONSE_REQ) 3047 cm_id_priv = cm_acquire_id( 3048 cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)), 3049 0); 3050 else 3051 cm_id_priv = cm_acquire_id( 3052 cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)), 3053 remote_id); 3054 3055 return cm_id_priv; 3056 } 3057 3058 static int cm_rej_handler(struct cm_work *work) 3059 { 3060 struct cm_id_private *cm_id_priv; 3061 struct cm_rej_msg *rej_msg; 3062 3063 rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad; 3064 cm_id_priv = cm_acquire_rejected_id(rej_msg); 3065 if (!cm_id_priv) 3066 return -EINVAL; 3067 3068 cm_format_rej_event(work); 3069 3070 spin_lock_irq(&cm_id_priv->lock); 3071 switch (cm_id_priv->id.state) { 3072 case IB_CM_REQ_SENT: 3073 case IB_CM_MRA_REQ_RCVD: 3074 case IB_CM_REP_SENT: 3075 case IB_CM_MRA_REP_RCVD: 3076 ib_cancel_mad(cm_id_priv->msg); 3077 fallthrough; 3078 case IB_CM_REQ_RCVD: 3079 case IB_CM_MRA_REQ_SENT: 3080 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN) 3081 cm_enter_timewait(cm_id_priv); 3082 else 3083 cm_reset_to_idle(cm_id_priv); 3084 break; 3085 case IB_CM_DREQ_SENT: 3086 ib_cancel_mad(cm_id_priv->msg); 3087 fallthrough; 3088 case IB_CM_REP_RCVD: 3089 case IB_CM_MRA_REP_SENT: 3090 cm_enter_timewait(cm_id_priv); 3091 break; 3092 case IB_CM_ESTABLISHED: 3093 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT || 3094 cm_id_priv->id.lap_state == IB_CM_LAP_SENT) { 3095 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT) 3096 ib_cancel_mad(cm_id_priv->msg); 3097 cm_enter_timewait(cm_id_priv); 3098 break; 3099 } 3100 fallthrough; 3101 default: 3102 trace_icm_rej_unknown_err(&cm_id_priv->id); 3103 spin_unlock_irq(&cm_id_priv->lock); 3104 goto out; 3105 } 3106 3107 cm_queue_work_unlock(cm_id_priv, work); 3108 return 0; 3109 out: 3110 cm_deref_id(cm_id_priv); 3111 return -EINVAL; 3112 } 3113 3114 int ib_prepare_cm_mra(struct ib_cm_id *cm_id) 3115 { 3116 struct cm_id_private *cm_id_priv; 3117 enum ib_cm_state cm_state; 3118 enum ib_cm_lap_state lap_state; 3119 unsigned long flags; 3120 int ret = 0; 3121 3122 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3123 3124 spin_lock_irqsave(&cm_id_priv->lock, flags); 3125 switch (cm_id_priv->id.state) { 3126 case IB_CM_REQ_RCVD: 3127 cm_state = IB_CM_MRA_REQ_SENT; 3128 lap_state = cm_id->lap_state; 3129 break; 3130 case IB_CM_REP_RCVD: 3131 cm_state = IB_CM_MRA_REP_SENT; 3132 lap_state = cm_id->lap_state; 3133 break; 3134 case IB_CM_ESTABLISHED: 3135 if (cm_id->lap_state == IB_CM_LAP_RCVD) { 3136 cm_state = cm_id->state; 3137 lap_state = IB_CM_MRA_LAP_SENT; 3138 break; 3139 } 3140 fallthrough; 3141 default: 3142 trace_icm_prepare_mra_unknown_err(&cm_id_priv->id); 3143 ret = -EINVAL; 3144 goto error_unlock; 3145 } 3146 3147 cm_id->state = cm_state; 3148 cm_id->lap_state = lap_state; 3149 cm_set_private_data(cm_id_priv, NULL, 0); 3150 3151 error_unlock: 3152 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3153 return ret; 3154 } 3155 EXPORT_SYMBOL(ib_prepare_cm_mra); 3156 3157 static struct cm_id_private *cm_acquire_mraed_id(struct cm_mra_msg *mra_msg) 3158 { 3159 switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) { 3160 case CM_MSG_RESPONSE_REQ: 3161 return cm_acquire_id( 3162 cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)), 3163 0); 3164 case CM_MSG_RESPONSE_REP: 3165 case CM_MSG_RESPONSE_OTHER: 3166 return cm_acquire_id( 3167 cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)), 3168 cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg))); 3169 default: 3170 return NULL; 3171 } 3172 } 3173 3174 static int cm_mra_handler(struct cm_work *work) 3175 { 3176 struct cm_id_private *cm_id_priv; 3177 struct cm_mra_msg *mra_msg; 3178 int timeout; 3179 3180 mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad; 3181 cm_id_priv = cm_acquire_mraed_id(mra_msg); 3182 if (!cm_id_priv) 3183 return -EINVAL; 3184 3185 work->cm_event.private_data = 3186 IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg); 3187 work->cm_event.param.mra_rcvd.service_timeout = 3188 IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg); 3189 timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) + 3190 cm_convert_to_ms(cm_id_priv->av.timeout); 3191 3192 spin_lock_irq(&cm_id_priv->lock); 3193 switch (cm_id_priv->id.state) { 3194 case IB_CM_REQ_SENT: 3195 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) != 3196 CM_MSG_RESPONSE_REQ || 3197 ib_modify_mad(cm_id_priv->msg, timeout)) 3198 goto out; 3199 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD; 3200 break; 3201 case IB_CM_REP_SENT: 3202 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) != 3203 CM_MSG_RESPONSE_REP || 3204 ib_modify_mad(cm_id_priv->msg, timeout)) 3205 goto out; 3206 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD; 3207 break; 3208 case IB_CM_ESTABLISHED: 3209 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) != 3210 CM_MSG_RESPONSE_OTHER || 3211 cm_id_priv->id.lap_state != IB_CM_LAP_SENT || 3212 ib_modify_mad(cm_id_priv->msg, timeout)) { 3213 if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD) 3214 atomic_long_inc( 3215 &work->port->counters[CM_RECV_DUPLICATES] 3216 [CM_MRA_COUNTER]); 3217 goto out; 3218 } 3219 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD; 3220 break; 3221 case IB_CM_MRA_REQ_RCVD: 3222 case IB_CM_MRA_REP_RCVD: 3223 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3224 [CM_MRA_COUNTER]); 3225 fallthrough; 3226 default: 3227 trace_icm_mra_unknown_err(&cm_id_priv->id); 3228 goto out; 3229 } 3230 3231 cm_id_priv->msg->context[1] = (void *) (unsigned long) 3232 cm_id_priv->id.state; 3233 cm_queue_work_unlock(cm_id_priv, work); 3234 return 0; 3235 out: 3236 spin_unlock_irq(&cm_id_priv->lock); 3237 cm_deref_id(cm_id_priv); 3238 return -EINVAL; 3239 } 3240 3241 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg, 3242 struct sa_path_rec *path) 3243 { 3244 u32 lid; 3245 3246 if (path->rec_type != SA_PATH_REC_TYPE_OPA) { 3247 sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID, 3248 lap_msg)); 3249 sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID, 3250 lap_msg)); 3251 } else { 3252 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 3253 CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg)); 3254 sa_path_set_dlid(path, lid); 3255 3256 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR( 3257 CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg)); 3258 sa_path_set_slid(path, lid); 3259 } 3260 } 3261 3262 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv, 3263 struct sa_path_rec *path, 3264 struct cm_lap_msg *lap_msg) 3265 { 3266 path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg); 3267 path->sgid = 3268 *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg); 3269 path->flow_label = 3270 cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg)); 3271 path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg); 3272 path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg); 3273 path->reversible = 1; 3274 path->pkey = cm_id_priv->pkey; 3275 path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg); 3276 path->mtu_selector = IB_SA_EQ; 3277 path->mtu = cm_id_priv->path_mtu; 3278 path->rate_selector = IB_SA_EQ; 3279 path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg); 3280 path->packet_life_time_selector = IB_SA_EQ; 3281 path->packet_life_time = 3282 IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg); 3283 path->packet_life_time -= (path->packet_life_time > 0); 3284 cm_format_path_lid_from_lap(lap_msg, path); 3285 } 3286 3287 static int cm_lap_handler(struct cm_work *work) 3288 { 3289 struct cm_id_private *cm_id_priv; 3290 struct cm_lap_msg *lap_msg; 3291 struct ib_cm_lap_event_param *param; 3292 struct ib_mad_send_buf *msg = NULL; 3293 struct rdma_ah_attr ah_attr; 3294 struct cm_av alt_av = {}; 3295 int ret; 3296 3297 /* Currently Alternate path messages are not supported for 3298 * RoCE link layer. 3299 */ 3300 if (rdma_protocol_roce(work->port->cm_dev->ib_device, 3301 work->port->port_num)) 3302 return -EINVAL; 3303 3304 /* todo: verify LAP request and send reject APR if invalid. */ 3305 lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad; 3306 cm_id_priv = cm_acquire_id( 3307 cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)), 3308 cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg))); 3309 if (!cm_id_priv) 3310 return -EINVAL; 3311 3312 param = &work->cm_event.param.lap_rcvd; 3313 memset(&work->path[0], 0, sizeof(work->path[1])); 3314 cm_path_set_rec_type(work->port->cm_dev->ib_device, 3315 work->port->port_num, &work->path[0], 3316 IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, 3317 lap_msg)); 3318 param->alternate_path = &work->path[0]; 3319 cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg); 3320 work->cm_event.private_data = 3321 IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg); 3322 3323 ret = ib_init_ah_attr_from_wc(work->port->cm_dev->ib_device, 3324 work->port->port_num, 3325 work->mad_recv_wc->wc, 3326 work->mad_recv_wc->recv_buf.grh, 3327 &ah_attr); 3328 if (ret) 3329 goto deref; 3330 3331 ret = cm_init_av_by_path(param->alternate_path, NULL, &alt_av); 3332 if (ret) { 3333 rdma_destroy_ah_attr(&ah_attr); 3334 goto deref; 3335 } 3336 3337 spin_lock_irq(&cm_id_priv->lock); 3338 cm_init_av_for_lap(work->port, work->mad_recv_wc->wc, 3339 &ah_attr, &cm_id_priv->av); 3340 cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av); 3341 3342 if (cm_id_priv->id.state != IB_CM_ESTABLISHED) 3343 goto unlock; 3344 3345 switch (cm_id_priv->id.lap_state) { 3346 case IB_CM_LAP_UNINIT: 3347 case IB_CM_LAP_IDLE: 3348 break; 3349 case IB_CM_MRA_LAP_SENT: 3350 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3351 [CM_LAP_COUNTER]); 3352 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc, 3353 true); 3354 if (IS_ERR(msg)) 3355 goto unlock; 3356 3357 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv, 3358 CM_MSG_RESPONSE_OTHER, 3359 cm_id_priv->private_data, 3360 cm_id_priv->private_data_len); 3361 spin_unlock_irq(&cm_id_priv->lock); 3362 3363 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) || 3364 ib_post_send_mad(msg, NULL)) 3365 cm_free_msg(msg); 3366 goto deref; 3367 case IB_CM_LAP_RCVD: 3368 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3369 [CM_LAP_COUNTER]); 3370 goto unlock; 3371 default: 3372 goto unlock; 3373 } 3374 3375 cm_id_priv->id.lap_state = IB_CM_LAP_RCVD; 3376 cm_id_priv->tid = lap_msg->hdr.tid; 3377 cm_queue_work_unlock(cm_id_priv, work); 3378 return 0; 3379 3380 unlock: spin_unlock_irq(&cm_id_priv->lock); 3381 deref: cm_deref_id(cm_id_priv); 3382 return -EINVAL; 3383 } 3384 3385 static int cm_apr_handler(struct cm_work *work) 3386 { 3387 struct cm_id_private *cm_id_priv; 3388 struct cm_apr_msg *apr_msg; 3389 3390 /* Currently Alternate path messages are not supported for 3391 * RoCE link layer. 3392 */ 3393 if (rdma_protocol_roce(work->port->cm_dev->ib_device, 3394 work->port->port_num)) 3395 return -EINVAL; 3396 3397 apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad; 3398 cm_id_priv = cm_acquire_id( 3399 cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)), 3400 cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg))); 3401 if (!cm_id_priv) 3402 return -EINVAL; /* Unmatched reply. */ 3403 3404 work->cm_event.param.apr_rcvd.ap_status = 3405 IBA_GET(CM_APR_AR_STATUS, apr_msg); 3406 work->cm_event.param.apr_rcvd.apr_info = 3407 IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg); 3408 work->cm_event.param.apr_rcvd.info_len = 3409 IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg); 3410 work->cm_event.private_data = 3411 IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg); 3412 3413 spin_lock_irq(&cm_id_priv->lock); 3414 if (cm_id_priv->id.state != IB_CM_ESTABLISHED || 3415 (cm_id_priv->id.lap_state != IB_CM_LAP_SENT && 3416 cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) { 3417 spin_unlock_irq(&cm_id_priv->lock); 3418 goto out; 3419 } 3420 cm_id_priv->id.lap_state = IB_CM_LAP_IDLE; 3421 ib_cancel_mad(cm_id_priv->msg); 3422 cm_queue_work_unlock(cm_id_priv, work); 3423 return 0; 3424 out: 3425 cm_deref_id(cm_id_priv); 3426 return -EINVAL; 3427 } 3428 3429 static int cm_timewait_handler(struct cm_work *work) 3430 { 3431 struct cm_timewait_info *timewait_info; 3432 struct cm_id_private *cm_id_priv; 3433 3434 timewait_info = container_of(work, struct cm_timewait_info, work); 3435 spin_lock_irq(&cm.lock); 3436 list_del(&timewait_info->list); 3437 spin_unlock_irq(&cm.lock); 3438 3439 cm_id_priv = cm_acquire_id(timewait_info->work.local_id, 3440 timewait_info->work.remote_id); 3441 if (!cm_id_priv) 3442 return -EINVAL; 3443 3444 spin_lock_irq(&cm_id_priv->lock); 3445 if (cm_id_priv->id.state != IB_CM_TIMEWAIT || 3446 cm_id_priv->remote_qpn != timewait_info->remote_qpn) { 3447 spin_unlock_irq(&cm_id_priv->lock); 3448 goto out; 3449 } 3450 cm_id_priv->id.state = IB_CM_IDLE; 3451 cm_queue_work_unlock(cm_id_priv, work); 3452 return 0; 3453 out: 3454 cm_deref_id(cm_id_priv); 3455 return -EINVAL; 3456 } 3457 3458 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg, 3459 struct cm_id_private *cm_id_priv, 3460 struct ib_cm_sidr_req_param *param) 3461 { 3462 cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID, 3463 cm_form_tid(cm_id_priv)); 3464 IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg, 3465 be32_to_cpu(cm_id_priv->id.local_id)); 3466 IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg, 3467 be16_to_cpu(param->path->pkey)); 3468 IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg, 3469 be64_to_cpu(param->service_id)); 3470 3471 if (param->private_data && param->private_data_len) 3472 IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg, 3473 param->private_data, param->private_data_len); 3474 } 3475 3476 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id, 3477 struct ib_cm_sidr_req_param *param) 3478 { 3479 struct cm_id_private *cm_id_priv; 3480 struct ib_mad_send_buf *msg; 3481 struct cm_av av = {}; 3482 unsigned long flags; 3483 int ret; 3484 3485 if (!param->path || (param->private_data && 3486 param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE)) 3487 return -EINVAL; 3488 3489 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3490 ret = cm_init_av_by_path(param->path, param->sgid_attr, &av); 3491 if (ret) 3492 return ret; 3493 3494 spin_lock_irqsave(&cm_id_priv->lock, flags); 3495 cm_move_av_from_path(&cm_id_priv->av, &av); 3496 cm_id->service_id = param->service_id; 3497 cm_id_priv->timeout_ms = param->timeout_ms; 3498 cm_id_priv->max_cm_retries = param->max_cm_retries; 3499 if (cm_id->state != IB_CM_IDLE) { 3500 ret = -EINVAL; 3501 goto out_unlock; 3502 } 3503 3504 msg = cm_alloc_priv_msg(cm_id_priv, IB_CM_SIDR_REQ_SENT); 3505 if (IS_ERR(msg)) { 3506 ret = PTR_ERR(msg); 3507 goto out_unlock; 3508 } 3509 3510 cm_format_sidr_req((struct cm_sidr_req_msg *)msg->mad, cm_id_priv, 3511 param); 3512 3513 trace_icm_send_sidr_req(&cm_id_priv->id); 3514 ret = ib_post_send_mad(msg, NULL); 3515 if (ret) 3516 goto out_free; 3517 cm_id->state = IB_CM_SIDR_REQ_SENT; 3518 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3519 return 0; 3520 out_free: 3521 cm_free_priv_msg(msg); 3522 out_unlock: 3523 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3524 return ret; 3525 } 3526 EXPORT_SYMBOL(ib_send_cm_sidr_req); 3527 3528 static void cm_format_sidr_req_event(struct cm_work *work, 3529 const struct cm_id_private *rx_cm_id, 3530 struct ib_cm_id *listen_id) 3531 { 3532 struct cm_sidr_req_msg *sidr_req_msg; 3533 struct ib_cm_sidr_req_event_param *param; 3534 3535 sidr_req_msg = (struct cm_sidr_req_msg *) 3536 work->mad_recv_wc->recv_buf.mad; 3537 param = &work->cm_event.param.sidr_req_rcvd; 3538 param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg); 3539 param->listen_id = listen_id; 3540 param->service_id = 3541 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg)); 3542 param->bth_pkey = cm_get_bth_pkey(work); 3543 param->port = work->port->port_num; 3544 param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr; 3545 work->cm_event.private_data = 3546 IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg); 3547 } 3548 3549 static int cm_sidr_req_handler(struct cm_work *work) 3550 { 3551 struct cm_id_private *cm_id_priv, *listen_cm_id_priv; 3552 struct cm_sidr_req_msg *sidr_req_msg; 3553 struct ib_wc *wc; 3554 int ret; 3555 3556 cm_id_priv = 3557 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL); 3558 if (IS_ERR(cm_id_priv)) 3559 return PTR_ERR(cm_id_priv); 3560 3561 /* Record SGID/SLID and request ID for lookup. */ 3562 sidr_req_msg = (struct cm_sidr_req_msg *) 3563 work->mad_recv_wc->recv_buf.mad; 3564 3565 cm_id_priv->id.remote_id = 3566 cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg)); 3567 cm_id_priv->id.service_id = 3568 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg)); 3569 cm_id_priv->tid = sidr_req_msg->hdr.tid; 3570 3571 wc = work->mad_recv_wc->wc; 3572 cm_id_priv->sidr_slid = wc->slid; 3573 ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc, 3574 work->mad_recv_wc->recv_buf.grh, 3575 &cm_id_priv->av); 3576 if (ret) 3577 goto out; 3578 3579 spin_lock_irq(&cm.lock); 3580 listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv); 3581 if (listen_cm_id_priv) { 3582 spin_unlock_irq(&cm.lock); 3583 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES] 3584 [CM_SIDR_REQ_COUNTER]); 3585 goto out; /* Duplicate message. */ 3586 } 3587 cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD; 3588 listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device, 3589 cm_id_priv->id.service_id); 3590 if (!listen_cm_id_priv) { 3591 spin_unlock_irq(&cm.lock); 3592 ib_send_cm_sidr_rep(&cm_id_priv->id, 3593 &(struct ib_cm_sidr_rep_param){ 3594 .status = IB_SIDR_UNSUPPORTED }); 3595 goto out; /* No match. */ 3596 } 3597 spin_unlock_irq(&cm.lock); 3598 3599 cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler; 3600 cm_id_priv->id.context = listen_cm_id_priv->id.context; 3601 3602 /* 3603 * A SIDR ID does not need to be in the xarray since it does not receive 3604 * mads, is not placed in the remote_id or remote_qpn rbtree, and does 3605 * not enter timewait. 3606 */ 3607 3608 cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id); 3609 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event); 3610 cm_free_work(work); 3611 /* 3612 * A pointer to the listen_cm_id is held in the event, so this deref 3613 * must be after the event is delivered above. 3614 */ 3615 cm_deref_id(listen_cm_id_priv); 3616 if (ret) 3617 cm_destroy_id(&cm_id_priv->id, ret); 3618 return 0; 3619 out: 3620 ib_destroy_cm_id(&cm_id_priv->id); 3621 return -EINVAL; 3622 } 3623 3624 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg, 3625 struct cm_id_private *cm_id_priv, 3626 struct ib_cm_sidr_rep_param *param) 3627 { 3628 cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID, 3629 cm_id_priv->tid, param->ece.attr_mod); 3630 IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg, 3631 be32_to_cpu(cm_id_priv->id.remote_id)); 3632 IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status); 3633 IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num); 3634 IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg, 3635 be64_to_cpu(cm_id_priv->id.service_id)); 3636 IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey); 3637 IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg, 3638 param->ece.vendor_id & 0xFF); 3639 IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg, 3640 (param->ece.vendor_id >> 8) & 0xFF); 3641 3642 if (param->info && param->info_length) 3643 IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg, 3644 param->info, param->info_length); 3645 3646 if (param->private_data && param->private_data_len) 3647 IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg, 3648 param->private_data, param->private_data_len); 3649 } 3650 3651 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv, 3652 struct ib_cm_sidr_rep_param *param) 3653 { 3654 struct ib_mad_send_buf *msg; 3655 unsigned long flags; 3656 int ret; 3657 3658 lockdep_assert_held(&cm_id_priv->lock); 3659 3660 if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) || 3661 (param->private_data && 3662 param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE)) 3663 return -EINVAL; 3664 3665 if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD) 3666 return -EINVAL; 3667 3668 msg = cm_alloc_msg(cm_id_priv); 3669 if (IS_ERR(msg)) 3670 return PTR_ERR(msg); 3671 3672 cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv, 3673 param); 3674 trace_icm_send_sidr_rep(&cm_id_priv->id); 3675 ret = ib_post_send_mad(msg, NULL); 3676 if (ret) { 3677 cm_free_msg(msg); 3678 return ret; 3679 } 3680 cm_id_priv->id.state = IB_CM_IDLE; 3681 spin_lock_irqsave(&cm.lock, flags); 3682 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) { 3683 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table); 3684 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node); 3685 } 3686 spin_unlock_irqrestore(&cm.lock, flags); 3687 return 0; 3688 } 3689 3690 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id, 3691 struct ib_cm_sidr_rep_param *param) 3692 { 3693 struct cm_id_private *cm_id_priv = 3694 container_of(cm_id, struct cm_id_private, id); 3695 unsigned long flags; 3696 int ret; 3697 3698 spin_lock_irqsave(&cm_id_priv->lock, flags); 3699 ret = cm_send_sidr_rep_locked(cm_id_priv, param); 3700 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3701 return ret; 3702 } 3703 EXPORT_SYMBOL(ib_send_cm_sidr_rep); 3704 3705 static void cm_format_sidr_rep_event(struct cm_work *work, 3706 const struct cm_id_private *cm_id_priv) 3707 { 3708 struct cm_sidr_rep_msg *sidr_rep_msg; 3709 struct ib_cm_sidr_rep_event_param *param; 3710 3711 sidr_rep_msg = (struct cm_sidr_rep_msg *) 3712 work->mad_recv_wc->recv_buf.mad; 3713 param = &work->cm_event.param.sidr_rep_rcvd; 3714 param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg); 3715 param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg); 3716 param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg); 3717 param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION, 3718 sidr_rep_msg); 3719 param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH, 3720 sidr_rep_msg); 3721 param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr; 3722 work->cm_event.private_data = 3723 IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg); 3724 } 3725 3726 static int cm_sidr_rep_handler(struct cm_work *work) 3727 { 3728 struct cm_sidr_rep_msg *sidr_rep_msg; 3729 struct cm_id_private *cm_id_priv; 3730 3731 sidr_rep_msg = (struct cm_sidr_rep_msg *) 3732 work->mad_recv_wc->recv_buf.mad; 3733 cm_id_priv = cm_acquire_id( 3734 cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0); 3735 if (!cm_id_priv) 3736 return -EINVAL; /* Unmatched reply. */ 3737 3738 spin_lock_irq(&cm_id_priv->lock); 3739 if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) { 3740 spin_unlock_irq(&cm_id_priv->lock); 3741 goto out; 3742 } 3743 cm_id_priv->id.state = IB_CM_IDLE; 3744 ib_cancel_mad(cm_id_priv->msg); 3745 spin_unlock_irq(&cm_id_priv->lock); 3746 3747 cm_format_sidr_rep_event(work, cm_id_priv); 3748 cm_process_work(cm_id_priv, work); 3749 return 0; 3750 out: 3751 cm_deref_id(cm_id_priv); 3752 return -EINVAL; 3753 } 3754 3755 static void cm_process_send_error(struct cm_id_private *cm_id_priv, 3756 struct ib_mad_send_buf *msg, 3757 enum ib_wc_status wc_status) 3758 { 3759 enum ib_cm_state state = (unsigned long) msg->context[1]; 3760 struct ib_cm_event cm_event = {}; 3761 int ret; 3762 3763 /* Discard old sends. */ 3764 spin_lock_irq(&cm_id_priv->lock); 3765 if (msg != cm_id_priv->msg) { 3766 spin_unlock_irq(&cm_id_priv->lock); 3767 cm_free_msg(msg); 3768 cm_deref_id(cm_id_priv); 3769 return; 3770 } 3771 cm_free_priv_msg(msg); 3772 3773 if (state != cm_id_priv->id.state || wc_status == IB_WC_SUCCESS || 3774 wc_status == IB_WC_WR_FLUSH_ERR) 3775 goto out_unlock; 3776 3777 trace_icm_mad_send_err(state, wc_status); 3778 switch (state) { 3779 case IB_CM_REQ_SENT: 3780 case IB_CM_MRA_REQ_RCVD: 3781 cm_reset_to_idle(cm_id_priv); 3782 cm_event.event = IB_CM_REQ_ERROR; 3783 break; 3784 case IB_CM_REP_SENT: 3785 case IB_CM_MRA_REP_RCVD: 3786 cm_reset_to_idle(cm_id_priv); 3787 cm_event.event = IB_CM_REP_ERROR; 3788 break; 3789 case IB_CM_DREQ_SENT: 3790 cm_enter_timewait(cm_id_priv); 3791 cm_event.event = IB_CM_DREQ_ERROR; 3792 break; 3793 case IB_CM_SIDR_REQ_SENT: 3794 cm_id_priv->id.state = IB_CM_IDLE; 3795 cm_event.event = IB_CM_SIDR_REQ_ERROR; 3796 break; 3797 default: 3798 goto out_unlock; 3799 } 3800 spin_unlock_irq(&cm_id_priv->lock); 3801 cm_event.param.send_status = wc_status; 3802 3803 /* No other events can occur on the cm_id at this point. */ 3804 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event); 3805 if (ret) 3806 ib_destroy_cm_id(&cm_id_priv->id); 3807 return; 3808 out_unlock: 3809 spin_unlock_irq(&cm_id_priv->lock); 3810 } 3811 3812 static void cm_send_handler(struct ib_mad_agent *mad_agent, 3813 struct ib_mad_send_wc *mad_send_wc) 3814 { 3815 struct ib_mad_send_buf *msg = mad_send_wc->send_buf; 3816 struct cm_id_private *cm_id_priv; 3817 struct cm_port *port; 3818 u16 attr_index; 3819 3820 port = mad_agent->context; 3821 attr_index = be16_to_cpu(((struct ib_mad_hdr *) 3822 msg->mad)->attr_id) - CM_ATTR_ID_OFFSET; 3823 3824 if (msg->context[0] == CM_DIRECT_RETRY_CTX) { 3825 msg->retries = 1; 3826 cm_id_priv = NULL; 3827 } else { 3828 cm_id_priv = msg->context[0]; 3829 } 3830 3831 atomic_long_add(1 + msg->retries, &port->counters[CM_XMIT][attr_index]); 3832 if (msg->retries) 3833 atomic_long_add(msg->retries, 3834 &port->counters[CM_XMIT_RETRIES][attr_index]); 3835 3836 if (cm_id_priv) 3837 cm_process_send_error(cm_id_priv, msg, mad_send_wc->status); 3838 else 3839 cm_free_msg(msg); 3840 } 3841 3842 static void cm_work_handler(struct work_struct *_work) 3843 { 3844 struct cm_work *work = container_of(_work, struct cm_work, work.work); 3845 int ret; 3846 3847 switch (work->cm_event.event) { 3848 case IB_CM_REQ_RECEIVED: 3849 ret = cm_req_handler(work); 3850 break; 3851 case IB_CM_MRA_RECEIVED: 3852 ret = cm_mra_handler(work); 3853 break; 3854 case IB_CM_REJ_RECEIVED: 3855 ret = cm_rej_handler(work); 3856 break; 3857 case IB_CM_REP_RECEIVED: 3858 ret = cm_rep_handler(work); 3859 break; 3860 case IB_CM_RTU_RECEIVED: 3861 ret = cm_rtu_handler(work); 3862 break; 3863 case IB_CM_USER_ESTABLISHED: 3864 ret = cm_establish_handler(work); 3865 break; 3866 case IB_CM_DREQ_RECEIVED: 3867 ret = cm_dreq_handler(work); 3868 break; 3869 case IB_CM_DREP_RECEIVED: 3870 ret = cm_drep_handler(work); 3871 break; 3872 case IB_CM_SIDR_REQ_RECEIVED: 3873 ret = cm_sidr_req_handler(work); 3874 break; 3875 case IB_CM_SIDR_REP_RECEIVED: 3876 ret = cm_sidr_rep_handler(work); 3877 break; 3878 case IB_CM_LAP_RECEIVED: 3879 ret = cm_lap_handler(work); 3880 break; 3881 case IB_CM_APR_RECEIVED: 3882 ret = cm_apr_handler(work); 3883 break; 3884 case IB_CM_TIMEWAIT_EXIT: 3885 ret = cm_timewait_handler(work); 3886 break; 3887 default: 3888 trace_icm_handler_err(work->cm_event.event); 3889 ret = -EINVAL; 3890 break; 3891 } 3892 if (ret) 3893 cm_free_work(work); 3894 } 3895 3896 static int cm_establish(struct ib_cm_id *cm_id) 3897 { 3898 struct cm_id_private *cm_id_priv; 3899 struct cm_work *work; 3900 unsigned long flags; 3901 int ret = 0; 3902 struct cm_device *cm_dev; 3903 3904 cm_dev = ib_get_client_data(cm_id->device, &cm_client); 3905 if (!cm_dev) 3906 return -ENODEV; 3907 3908 work = kmalloc_obj(*work, GFP_ATOMIC); 3909 if (!work) 3910 return -ENOMEM; 3911 3912 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3913 spin_lock_irqsave(&cm_id_priv->lock, flags); 3914 switch (cm_id->state) { 3915 case IB_CM_REP_SENT: 3916 case IB_CM_MRA_REP_RCVD: 3917 cm_id->state = IB_CM_ESTABLISHED; 3918 break; 3919 case IB_CM_ESTABLISHED: 3920 ret = -EISCONN; 3921 break; 3922 default: 3923 trace_icm_establish_err(cm_id); 3924 ret = -EINVAL; 3925 break; 3926 } 3927 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3928 3929 if (ret) { 3930 kfree(work); 3931 goto out; 3932 } 3933 3934 /* 3935 * The CM worker thread may try to destroy the cm_id before it 3936 * can execute this work item. To prevent potential deadlock, 3937 * we need to find the cm_id once we're in the context of the 3938 * worker thread, rather than holding a reference on it. 3939 */ 3940 INIT_DELAYED_WORK(&work->work, cm_work_handler); 3941 work->local_id = cm_id->local_id; 3942 work->remote_id = cm_id->remote_id; 3943 work->mad_recv_wc = NULL; 3944 work->cm_event.event = IB_CM_USER_ESTABLISHED; 3945 3946 /* Check if the device started its remove_one */ 3947 spin_lock_irqsave(&cm.lock, flags); 3948 if (!cm_dev->going_down) { 3949 queue_delayed_work(cm.wq, &work->work, 0); 3950 } else { 3951 kfree(work); 3952 ret = -ENODEV; 3953 } 3954 spin_unlock_irqrestore(&cm.lock, flags); 3955 3956 out: 3957 return ret; 3958 } 3959 3960 static int cm_migrate(struct ib_cm_id *cm_id) 3961 { 3962 struct cm_id_private *cm_id_priv; 3963 unsigned long flags; 3964 int ret = 0; 3965 3966 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 3967 spin_lock_irqsave(&cm_id_priv->lock, flags); 3968 if (cm_id->state == IB_CM_ESTABLISHED && 3969 (cm_id->lap_state == IB_CM_LAP_UNINIT || 3970 cm_id->lap_state == IB_CM_LAP_IDLE)) { 3971 cm_id->lap_state = IB_CM_LAP_IDLE; 3972 cm_id_priv->av = cm_id_priv->alt_av; 3973 } else 3974 ret = -EINVAL; 3975 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 3976 3977 return ret; 3978 } 3979 3980 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event) 3981 { 3982 int ret; 3983 3984 switch (event) { 3985 case IB_EVENT_COMM_EST: 3986 ret = cm_establish(cm_id); 3987 break; 3988 case IB_EVENT_PATH_MIG: 3989 ret = cm_migrate(cm_id); 3990 break; 3991 default: 3992 ret = -EINVAL; 3993 } 3994 return ret; 3995 } 3996 EXPORT_SYMBOL(ib_cm_notify); 3997 3998 static void cm_recv_handler(struct ib_mad_agent *mad_agent, 3999 struct ib_mad_send_buf *send_buf, 4000 struct ib_mad_recv_wc *mad_recv_wc) 4001 { 4002 struct cm_port *port = mad_agent->context; 4003 struct cm_work *work; 4004 enum ib_cm_event_type event; 4005 bool alt_path = false; 4006 u16 attr_id; 4007 int paths = 0; 4008 int going_down = 0; 4009 4010 switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) { 4011 case CM_REQ_ATTR_ID: 4012 alt_path = cm_req_has_alt_path((struct cm_req_msg *) 4013 mad_recv_wc->recv_buf.mad); 4014 paths = 1 + (alt_path != 0); 4015 event = IB_CM_REQ_RECEIVED; 4016 break; 4017 case CM_MRA_ATTR_ID: 4018 event = IB_CM_MRA_RECEIVED; 4019 break; 4020 case CM_REJ_ATTR_ID: 4021 event = IB_CM_REJ_RECEIVED; 4022 break; 4023 case CM_REP_ATTR_ID: 4024 event = IB_CM_REP_RECEIVED; 4025 break; 4026 case CM_RTU_ATTR_ID: 4027 event = IB_CM_RTU_RECEIVED; 4028 break; 4029 case CM_DREQ_ATTR_ID: 4030 event = IB_CM_DREQ_RECEIVED; 4031 break; 4032 case CM_DREP_ATTR_ID: 4033 event = IB_CM_DREP_RECEIVED; 4034 break; 4035 case CM_SIDR_REQ_ATTR_ID: 4036 event = IB_CM_SIDR_REQ_RECEIVED; 4037 break; 4038 case CM_SIDR_REP_ATTR_ID: 4039 event = IB_CM_SIDR_REP_RECEIVED; 4040 break; 4041 case CM_LAP_ATTR_ID: 4042 paths = 1; 4043 event = IB_CM_LAP_RECEIVED; 4044 break; 4045 case CM_APR_ATTR_ID: 4046 event = IB_CM_APR_RECEIVED; 4047 break; 4048 default: 4049 ib_free_recv_mad(mad_recv_wc); 4050 return; 4051 } 4052 4053 attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id); 4054 atomic_long_inc(&port->counters[CM_RECV][attr_id - CM_ATTR_ID_OFFSET]); 4055 4056 work = kmalloc_flex(*work, path, paths); 4057 if (!work) { 4058 ib_free_recv_mad(mad_recv_wc); 4059 return; 4060 } 4061 4062 INIT_DELAYED_WORK(&work->work, cm_work_handler); 4063 work->cm_event.event = event; 4064 work->mad_recv_wc = mad_recv_wc; 4065 work->port = port; 4066 4067 /* Check if the device started its remove_one */ 4068 spin_lock_irq(&cm.lock); 4069 if (!port->cm_dev->going_down) 4070 queue_delayed_work(cm.wq, &work->work, 0); 4071 else 4072 going_down = 1; 4073 spin_unlock_irq(&cm.lock); 4074 4075 if (going_down) { 4076 kfree(work); 4077 ib_free_recv_mad(mad_recv_wc); 4078 } 4079 } 4080 4081 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv, 4082 struct ib_qp_attr *qp_attr, 4083 int *qp_attr_mask) 4084 { 4085 unsigned long flags; 4086 int ret; 4087 4088 spin_lock_irqsave(&cm_id_priv->lock, flags); 4089 switch (cm_id_priv->id.state) { 4090 case IB_CM_REQ_SENT: 4091 case IB_CM_MRA_REQ_RCVD: 4092 case IB_CM_REQ_RCVD: 4093 case IB_CM_MRA_REQ_SENT: 4094 case IB_CM_REP_RCVD: 4095 case IB_CM_MRA_REP_SENT: 4096 case IB_CM_REP_SENT: 4097 case IB_CM_MRA_REP_RCVD: 4098 case IB_CM_ESTABLISHED: 4099 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | 4100 IB_QP_PKEY_INDEX | IB_QP_PORT; 4101 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE; 4102 if (cm_id_priv->responder_resources) { 4103 struct ib_device *ib_dev = cm_id_priv->id.device; 4104 u64 support_flush = ib_dev->attrs.device_cap_flags & 4105 (IB_DEVICE_FLUSH_GLOBAL | IB_DEVICE_FLUSH_PERSISTENT); 4106 u32 flushable = support_flush ? 4107 (IB_ACCESS_FLUSH_GLOBAL | 4108 IB_ACCESS_FLUSH_PERSISTENT) : 0; 4109 4110 qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ | 4111 IB_ACCESS_REMOTE_ATOMIC | 4112 flushable; 4113 } 4114 qp_attr->pkey_index = cm_id_priv->av.pkey_index; 4115 if (cm_id_priv->av.port) 4116 qp_attr->port_num = cm_id_priv->av.port->port_num; 4117 ret = 0; 4118 break; 4119 default: 4120 trace_icm_qp_init_err(&cm_id_priv->id); 4121 ret = -EINVAL; 4122 break; 4123 } 4124 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 4125 return ret; 4126 } 4127 4128 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv, 4129 struct ib_qp_attr *qp_attr, 4130 int *qp_attr_mask) 4131 { 4132 unsigned long flags; 4133 int ret; 4134 4135 spin_lock_irqsave(&cm_id_priv->lock, flags); 4136 switch (cm_id_priv->id.state) { 4137 case IB_CM_REQ_RCVD: 4138 case IB_CM_MRA_REQ_SENT: 4139 case IB_CM_REP_RCVD: 4140 case IB_CM_MRA_REP_SENT: 4141 case IB_CM_REP_SENT: 4142 case IB_CM_MRA_REP_RCVD: 4143 case IB_CM_ESTABLISHED: 4144 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU | 4145 IB_QP_DEST_QPN | IB_QP_RQ_PSN; 4146 qp_attr->ah_attr = cm_id_priv->av.ah_attr; 4147 if ((qp_attr->ah_attr.type == RDMA_AH_ATTR_TYPE_IB) && 4148 cm_id_priv->av.dlid_datapath && 4149 (cm_id_priv->av.dlid_datapath != 0xffff)) 4150 qp_attr->ah_attr.ib.dlid = cm_id_priv->av.dlid_datapath; 4151 qp_attr->path_mtu = cm_id_priv->path_mtu; 4152 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn); 4153 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn); 4154 if (cm_id_priv->qp_type == IB_QPT_RC || 4155 cm_id_priv->qp_type == IB_QPT_XRC_TGT) { 4156 *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC | 4157 IB_QP_MIN_RNR_TIMER; 4158 qp_attr->max_dest_rd_atomic = 4159 cm_id_priv->responder_resources; 4160 qp_attr->min_rnr_timer = 0; 4161 } 4162 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr) && 4163 cm_id_priv->alt_av.port) { 4164 *qp_attr_mask |= IB_QP_ALT_PATH; 4165 qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num; 4166 qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index; 4167 qp_attr->alt_timeout = cm_id_priv->alt_av.timeout; 4168 qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr; 4169 } 4170 ret = 0; 4171 break; 4172 default: 4173 trace_icm_qp_rtr_err(&cm_id_priv->id); 4174 ret = -EINVAL; 4175 break; 4176 } 4177 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 4178 return ret; 4179 } 4180 4181 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv, 4182 struct ib_qp_attr *qp_attr, 4183 int *qp_attr_mask) 4184 { 4185 unsigned long flags; 4186 int ret; 4187 4188 spin_lock_irqsave(&cm_id_priv->lock, flags); 4189 switch (cm_id_priv->id.state) { 4190 /* Allow transition to RTS before sending REP */ 4191 case IB_CM_REQ_RCVD: 4192 case IB_CM_MRA_REQ_SENT: 4193 4194 case IB_CM_REP_RCVD: 4195 case IB_CM_MRA_REP_SENT: 4196 case IB_CM_REP_SENT: 4197 case IB_CM_MRA_REP_RCVD: 4198 case IB_CM_ESTABLISHED: 4199 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) { 4200 *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN; 4201 qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn); 4202 switch (cm_id_priv->qp_type) { 4203 case IB_QPT_RC: 4204 case IB_QPT_XRC_INI: 4205 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY | 4206 IB_QP_MAX_QP_RD_ATOMIC; 4207 qp_attr->retry_cnt = cm_id_priv->retry_count; 4208 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count; 4209 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth; 4210 fallthrough; 4211 case IB_QPT_XRC_TGT: 4212 *qp_attr_mask |= IB_QP_TIMEOUT; 4213 qp_attr->timeout = cm_id_priv->av.timeout; 4214 break; 4215 default: 4216 break; 4217 } 4218 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) { 4219 *qp_attr_mask |= IB_QP_PATH_MIG_STATE; 4220 qp_attr->path_mig_state = IB_MIG_REARM; 4221 } 4222 } else { 4223 *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE; 4224 if (cm_id_priv->alt_av.port) 4225 qp_attr->alt_port_num = 4226 cm_id_priv->alt_av.port->port_num; 4227 qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index; 4228 qp_attr->alt_timeout = cm_id_priv->alt_av.timeout; 4229 qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr; 4230 qp_attr->path_mig_state = IB_MIG_REARM; 4231 } 4232 ret = 0; 4233 break; 4234 default: 4235 trace_icm_qp_rts_err(&cm_id_priv->id); 4236 ret = -EINVAL; 4237 break; 4238 } 4239 spin_unlock_irqrestore(&cm_id_priv->lock, flags); 4240 return ret; 4241 } 4242 4243 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id, 4244 struct ib_qp_attr *qp_attr, 4245 int *qp_attr_mask) 4246 { 4247 struct cm_id_private *cm_id_priv; 4248 int ret; 4249 4250 cm_id_priv = container_of(cm_id, struct cm_id_private, id); 4251 switch (qp_attr->qp_state) { 4252 case IB_QPS_INIT: 4253 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask); 4254 break; 4255 case IB_QPS_RTR: 4256 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask); 4257 break; 4258 case IB_QPS_RTS: 4259 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask); 4260 break; 4261 default: 4262 ret = -EINVAL; 4263 break; 4264 } 4265 return ret; 4266 } 4267 EXPORT_SYMBOL(ib_cm_init_qp_attr); 4268 4269 static ssize_t cm_show_counter(struct ib_device *ibdev, u32 port_num, 4270 struct ib_port_attribute *attr, char *buf) 4271 { 4272 struct cm_counter_attribute *cm_attr = 4273 container_of(attr, struct cm_counter_attribute, attr); 4274 struct cm_device *cm_dev = ib_get_client_data(ibdev, &cm_client); 4275 4276 if (WARN_ON(!cm_dev)) 4277 return -EINVAL; 4278 4279 return sysfs_emit( 4280 buf, "%ld\n", 4281 atomic_long_read( 4282 &cm_dev->port[port_num - 1] 4283 ->counters[cm_attr->group][cm_attr->index])); 4284 } 4285 4286 #define CM_COUNTER_ATTR(_name, _group, _index) \ 4287 { \ 4288 .attr = __ATTR(_name, 0444, cm_show_counter, NULL), \ 4289 .group = _group, .index = _index \ 4290 } 4291 4292 #define CM_COUNTER_GROUP(_group, _name) \ 4293 static struct cm_counter_attribute cm_counter_attr_##_group[] = { \ 4294 CM_COUNTER_ATTR(req, _group, CM_REQ_COUNTER), \ 4295 CM_COUNTER_ATTR(mra, _group, CM_MRA_COUNTER), \ 4296 CM_COUNTER_ATTR(rej, _group, CM_REJ_COUNTER), \ 4297 CM_COUNTER_ATTR(rep, _group, CM_REP_COUNTER), \ 4298 CM_COUNTER_ATTR(rtu, _group, CM_RTU_COUNTER), \ 4299 CM_COUNTER_ATTR(dreq, _group, CM_DREQ_COUNTER), \ 4300 CM_COUNTER_ATTR(drep, _group, CM_DREP_COUNTER), \ 4301 CM_COUNTER_ATTR(sidr_req, _group, CM_SIDR_REQ_COUNTER), \ 4302 CM_COUNTER_ATTR(sidr_rep, _group, CM_SIDR_REP_COUNTER), \ 4303 CM_COUNTER_ATTR(lap, _group, CM_LAP_COUNTER), \ 4304 CM_COUNTER_ATTR(apr, _group, CM_APR_COUNTER), \ 4305 }; \ 4306 static struct attribute *cm_counter_attrs_##_group[] = { \ 4307 &cm_counter_attr_##_group[0].attr.attr, \ 4308 &cm_counter_attr_##_group[1].attr.attr, \ 4309 &cm_counter_attr_##_group[2].attr.attr, \ 4310 &cm_counter_attr_##_group[3].attr.attr, \ 4311 &cm_counter_attr_##_group[4].attr.attr, \ 4312 &cm_counter_attr_##_group[5].attr.attr, \ 4313 &cm_counter_attr_##_group[6].attr.attr, \ 4314 &cm_counter_attr_##_group[7].attr.attr, \ 4315 &cm_counter_attr_##_group[8].attr.attr, \ 4316 &cm_counter_attr_##_group[9].attr.attr, \ 4317 &cm_counter_attr_##_group[10].attr.attr, \ 4318 NULL, \ 4319 }; \ 4320 static const struct attribute_group cm_counter_group_##_group = { \ 4321 .name = _name, \ 4322 .attrs = cm_counter_attrs_##_group, \ 4323 }; 4324 4325 CM_COUNTER_GROUP(CM_XMIT, "cm_tx_msgs") 4326 CM_COUNTER_GROUP(CM_XMIT_RETRIES, "cm_tx_retries") 4327 CM_COUNTER_GROUP(CM_RECV, "cm_rx_msgs") 4328 CM_COUNTER_GROUP(CM_RECV_DUPLICATES, "cm_rx_duplicates") 4329 4330 static const struct attribute_group *cm_counter_groups[] = { 4331 &cm_counter_group_CM_XMIT, 4332 &cm_counter_group_CM_XMIT_RETRIES, 4333 &cm_counter_group_CM_RECV, 4334 &cm_counter_group_CM_RECV_DUPLICATES, 4335 NULL, 4336 }; 4337 4338 static int cm_add_one(struct ib_device *ib_device) 4339 { 4340 struct cm_device *cm_dev; 4341 struct cm_port *port; 4342 struct ib_mad_reg_req reg_req = { 4343 .mgmt_class = IB_MGMT_CLASS_CM, 4344 .mgmt_class_version = IB_CM_CLASS_VERSION, 4345 }; 4346 struct ib_port_modify port_modify = { 4347 .set_port_cap_mask = IB_PORT_CM_SUP 4348 }; 4349 unsigned long flags; 4350 int ret; 4351 int count = 0; 4352 u32 i; 4353 4354 cm_dev = kzalloc_flex(*cm_dev, port, ib_device->phys_port_cnt); 4355 if (!cm_dev) 4356 return -ENOMEM; 4357 4358 kref_init(&cm_dev->kref); 4359 rwlock_init(&cm_dev->mad_agent_lock); 4360 cm_dev->ib_device = ib_device; 4361 cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay; 4362 cm_dev->going_down = 0; 4363 4364 ib_set_client_data(ib_device, &cm_client, cm_dev); 4365 4366 set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask); 4367 rdma_for_each_port (ib_device, i) { 4368 if (!rdma_cap_ib_cm(ib_device, i)) 4369 continue; 4370 4371 port = kzalloc_obj(*port); 4372 if (!port) { 4373 ret = -ENOMEM; 4374 goto error1; 4375 } 4376 4377 cm_dev->port[i-1] = port; 4378 port->cm_dev = cm_dev; 4379 port->port_num = i; 4380 4381 ret = ib_port_register_client_groups(ib_device, i, 4382 cm_counter_groups); 4383 if (ret) 4384 goto error1; 4385 4386 port->mad_agent = ib_register_mad_agent(ib_device, i, 4387 IB_QPT_GSI, 4388 ®_req, 4389 0, 4390 cm_send_handler, 4391 cm_recv_handler, 4392 port, 4393 0); 4394 if (IS_ERR(port->mad_agent)) { 4395 ret = PTR_ERR(port->mad_agent); 4396 goto error2; 4397 } 4398 4399 port->rep_agent = ib_register_mad_agent(ib_device, i, 4400 IB_QPT_GSI, 4401 NULL, 4402 0, 4403 cm_send_handler, 4404 NULL, 4405 port, 4406 0); 4407 if (IS_ERR(port->rep_agent)) { 4408 ret = PTR_ERR(port->rep_agent); 4409 goto error3; 4410 } 4411 4412 ret = ib_modify_port(ib_device, i, 0, &port_modify); 4413 if (ret) 4414 goto error4; 4415 4416 count++; 4417 } 4418 4419 if (!count) { 4420 ret = -EOPNOTSUPP; 4421 goto free; 4422 } 4423 4424 write_lock_irqsave(&cm.device_lock, flags); 4425 list_add_tail(&cm_dev->list, &cm.device_list); 4426 write_unlock_irqrestore(&cm.device_lock, flags); 4427 return 0; 4428 4429 error4: 4430 ib_unregister_mad_agent(port->rep_agent); 4431 error3: 4432 ib_unregister_mad_agent(port->mad_agent); 4433 error2: 4434 ib_port_unregister_client_groups(ib_device, i, cm_counter_groups); 4435 error1: 4436 port_modify.set_port_cap_mask = 0; 4437 port_modify.clr_port_cap_mask = IB_PORT_CM_SUP; 4438 while (--i) { 4439 if (!rdma_cap_ib_cm(ib_device, i)) 4440 continue; 4441 4442 port = cm_dev->port[i-1]; 4443 ib_modify_port(ib_device, port->port_num, 0, &port_modify); 4444 ib_unregister_mad_agent(port->rep_agent); 4445 ib_unregister_mad_agent(port->mad_agent); 4446 ib_port_unregister_client_groups(ib_device, i, 4447 cm_counter_groups); 4448 } 4449 free: 4450 cm_device_put(cm_dev); 4451 return ret; 4452 } 4453 4454 static void cm_remove_one(struct ib_device *ib_device, void *client_data) 4455 { 4456 struct cm_device *cm_dev = client_data; 4457 struct cm_port *port; 4458 struct ib_port_modify port_modify = { 4459 .clr_port_cap_mask = IB_PORT_CM_SUP 4460 }; 4461 unsigned long flags; 4462 u32 i; 4463 4464 write_lock_irqsave(&cm.device_lock, flags); 4465 list_del(&cm_dev->list); 4466 write_unlock_irqrestore(&cm.device_lock, flags); 4467 4468 spin_lock_irq(&cm.lock); 4469 cm_dev->going_down = 1; 4470 spin_unlock_irq(&cm.lock); 4471 4472 rdma_for_each_port (ib_device, i) { 4473 struct ib_mad_agent *mad_agent; 4474 struct ib_mad_agent *rep_agent; 4475 4476 if (!rdma_cap_ib_cm(ib_device, i)) 4477 continue; 4478 4479 port = cm_dev->port[i-1]; 4480 mad_agent = port->mad_agent; 4481 rep_agent = port->rep_agent; 4482 ib_modify_port(ib_device, port->port_num, 0, &port_modify); 4483 /* 4484 * We flush the queue here after the going_down set, this 4485 * verify that no new works will be queued in the recv handler, 4486 * after that we can call the unregister_mad_agent 4487 */ 4488 flush_workqueue(cm.wq); 4489 /* 4490 * The above ensures no call paths from the work are running, 4491 * the remaining paths all take the mad_agent_lock. 4492 */ 4493 write_lock(&cm_dev->mad_agent_lock); 4494 port->mad_agent = NULL; 4495 port->rep_agent = NULL; 4496 write_unlock(&cm_dev->mad_agent_lock); 4497 ib_unregister_mad_agent(mad_agent); 4498 ib_unregister_mad_agent(rep_agent); 4499 ib_port_unregister_client_groups(ib_device, i, 4500 cm_counter_groups); 4501 } 4502 4503 cm_device_put(cm_dev); 4504 } 4505 4506 static int __init ib_cm_init(void) 4507 { 4508 int ret; 4509 4510 INIT_LIST_HEAD(&cm.device_list); 4511 rwlock_init(&cm.device_lock); 4512 spin_lock_init(&cm.lock); 4513 cm.listen_service_table = RB_ROOT; 4514 cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID); 4515 cm.remote_id_table = RB_ROOT; 4516 cm.remote_qp_table = RB_ROOT; 4517 cm.remote_sidr_table = RB_ROOT; 4518 xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC); 4519 get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand); 4520 INIT_LIST_HEAD(&cm.timewait_list); 4521 4522 cm.wq = alloc_workqueue("ib_cm", WQ_PERCPU, 1); 4523 if (!cm.wq) { 4524 ret = -ENOMEM; 4525 goto error2; 4526 } 4527 4528 ret = ib_register_client(&cm_client); 4529 if (ret) 4530 goto error3; 4531 4532 return 0; 4533 error3: 4534 destroy_workqueue(cm.wq); 4535 error2: 4536 return ret; 4537 } 4538 4539 static void __exit ib_cm_cleanup(void) 4540 { 4541 struct cm_timewait_info *timewait_info, *tmp; 4542 4543 spin_lock_irq(&cm.lock); 4544 list_for_each_entry(timewait_info, &cm.timewait_list, list) 4545 cancel_delayed_work(&timewait_info->work.work); 4546 spin_unlock_irq(&cm.lock); 4547 4548 ib_unregister_client(&cm_client); 4549 destroy_workqueue(cm.wq); 4550 4551 list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) { 4552 list_del(&timewait_info->list); 4553 kfree(timewait_info); 4554 } 4555 4556 WARN_ON(!xa_empty(&cm.local_id_table)); 4557 } 4558 4559 module_init(ib_cm_init); 4560 module_exit(ib_cm_cleanup); 4561