sys/contrib/rdma/rdma_cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This Software is licensed under one of the following licenses:
   8  *
   9  * 1) under the terms of the "Common Public License 1.0" a copy of which is
  10  *    available from the Open Source Initiative, see
  11  *    http://www.opensource.org/licenses/cpl.php.
  12  *
  13  * 2) under the terms of the "The BSD License" a copy of which is
  14  *    available from the Open Source Initiative, see
  15  *    http://www.opensource.org/licenses/bsd-license.php.
  16  *
  17  * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
  18  *    copy of which is available from the Open Source Initiative, see
  19  *    http://www.opensource.org/licenses/gpl-license.php.
  20  *
  21  * Licensee has the right to choose one of the above licenses.
  22  *
  23  * Redistributions of source code must retain the above copyright
  24  * notice and one of the license notices.
  25  *
  26  * Redistributions in binary form must reproduce both the above copyright
  27  * notice, one of the license notices in the documentation
  28  * and/or other materials provided with the distribution.
  29  *
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/condvar.h>
  37 #include <sys/systm.h>
  38 #include <sys/kernel.h>
  39 #include <sys/libkern.h>
  40 #include <sys/socket.h>
  41 #include <sys/module.h>
  42 #include <sys/lock.h>
  43 #include <sys/mutex.h>
  44 #include <sys/rwlock.h>
  45 #include <sys/queue.h>
  46 #include <sys/taskqueue.h>
  47 #include <sys/priv.h>
  48 #include <sys/syslog.h>
  49
  50 #include <net/if.h>
  51 #include <netinet/in.h>
  52 #include <netinet/in_pcb.h>
  53
  54 #include <contrib/rdma/rdma_cm.h>
  55 #include <contrib/rdma/ib_cache.h>
  56 #include <contrib/rdma/ib_cm.h>
  57 #include <contrib/rdma/ib_sa.h>
  58 #include <contrib/rdma/iw_cm.h>
  59
  60 #define CMA_CM_RESPONSE_TIMEOUT 20
  61 #define CMA_MAX_CM_RETRIES 15
  62
  63 static void cma_add_one(struct ib_device *device);
  64 static void cma_remove_one(struct ib_device *device);
  65
  66 static struct ib_client cma_client = {
  67         .name   = "cma",
  68         .add    = cma_add_one,
  69         .remove = cma_remove_one
  70 };
  71
  72 #ifdef IB_SUPPORTED
  73 static struct ib_sa_client sa_client;
  74 #endif
  75 static struct rdma_addr_client addr_client;
  76 static TAILQ_HEAD(, cma_device) dev_list;
  77 static LIST_HEAD(, rdma_id_private) listen_any_list;
  78 static struct mtx lock;
  79 static struct taskqueue *cma_wq;
  80 static DEFINE_KVL(sdp_ps);
  81 static DEFINE_KVL(tcp_ps);
  82 static DEFINE_KVL(udp_ps);
  83 static DEFINE_KVL(ipoib_ps);
  84 static int next_port;
  85
  86 struct cma_device {
  87         struct ib_device        *device;
  88         struct mtx              lock;
  89         struct cv               comp;
  90         int                     refcount;
  91
  92         LIST_HEAD(, rdma_id_private) id_list;
  93         TAILQ_ENTRY(cma_device) list;
  94 };
  95
  96 enum cma_state {
  97         CMA_IDLE,
  98         CMA_ADDR_QUERY,
  99         CMA_ADDR_RESOLVED,
 100         CMA_ROUTE_QUERY,
 101         CMA_ROUTE_RESOLVED,
 102         CMA_CONNECT,
 103         CMA_DISCONNECT,
 104         CMA_ADDR_BOUND,
 105         CMA_LISTEN,
 106         CMA_DEVICE_REMOVAL,
 107         CMA_DESTROYING
 108 };
 109
 110 struct rdma_bind_list {
 111         struct kvl              *ps;
 112         TAILQ_HEAD(, rdma_id_private) owners;
 113         unsigned short          port;
 114 };
 115
 116 /*
 117  * Device removal can occur at anytime, so we need extra handling to
 118  * serialize notifying the user of device removal with other callbacks.
 119  * We do this by disabling removal notification while a callback is in process,
 120  * and reporting it after the callback completes.
 121  */
 122 struct rdma_id_private {
 123         struct rdma_cm_id       id;
 124
 125         struct rdma_bind_list   *bind_list;
 126         struct socket           *so;
 127         TAILQ_ENTRY(rdma_id_private) node;
 128         LIST_ENTRY(rdma_id_private) list; /* listen_any_list or cma_dev.list */
 129         LIST_HEAD(, rdma_id_private) listen_list; /* per-device listens */
 130         LIST_ENTRY(rdma_id_private) listen_entry;
 131         struct cma_device       *cma_dev;
 132 #ifdef IB_SUPPORTED
 133         LIST_HEAD(, cma_multicast) mc_list;
 134 #endif
 135         enum cma_state          state;
 136         struct mtx              lock;
 137         struct cv               comp;
 138         int                     refcount;
 139         struct cv               wait_remove;
 140         int                     dev_remove;
 141
 142         int                     backlog;
 143         int                     timeout_ms;
 144         struct ib_sa_query      *query;
 145         int                     query_id;
 146         union {
 147                 struct ib_cm_id *ib;
 148                 struct iw_cm_id *iw;
 149         } cm_id;
 150
 151         u32                     seq_num;
 152         u32                     qkey;
 153         u32                     qp_num;
 154         u8                      srq;
 155 };
 156
 157 #ifdef IB_SUPPORTED
 158 struct cma_multicast {
 159         struct rdma_id_private *id_priv;
 160         union {
 161                 struct ib_sa_multicast *ib;
 162         } multicast;
 163         struct list_head        list;
 164         void                    *context;
 165         struct sockaddr         addr;
 166         u8                      pad[sizeof(struct sockaddr_in6) -
 167                                     sizeof(struct sockaddr)];
 168 };
 169 #endif
 170
 171 struct cma_work {
 172         struct task             task;
 173         struct rdma_id_private  *id;
 174         enum cma_state          old_state;
 175         enum cma_state          new_state;
 176         struct rdma_cm_event    event;
 177 };
 178
 179 union cma_ip_addr {
 180         struct in6_addr ip6;
 181         struct {
 182                 __u32 pad[3];
 183                 __u32 addr;
 184         } ip4;
 185 };
 186
 187 struct cma_hdr {
 188         u8 cma_version;
 189         u8 ip_version;  /* IP version: 7:4 */
 190         __u16 port;
 191         union cma_ip_addr src_addr;
 192         union cma_ip_addr dst_addr;
 193 };
 194
 195 struct sdp_hh {
 196         u8 bsdh[16];
 197         u8 sdp_version; /* Major version: 7:4 */
 198         u8 ip_version;  /* IP version: 7:4 */
 199         u8 sdp_specific1[10];
 200         __u16 port;
 201         __u16 sdp_specific2;
 202         union cma_ip_addr src_addr;
 203         union cma_ip_addr dst_addr;
 204 };
 205
 206 struct sdp_hah {
 207         u8 bsdh[16];
 208         u8 sdp_version;
 209 };
 210
 211 #define CMA_VERSION 0x00
 212 #define SDP_MAJ_VERSION 0x2
 213
 214 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 215 {
 216         int ret;
 217
 218         mtx_lock(&id_priv->lock);
 219         ret = (id_priv->state == comp);
 220         mtx_unlock(&id_priv->lock);
 221         return ret;
 222 }
 223
 224 static int cma_comp_exch(struct rdma_id_private *id_priv,
 225                          enum cma_state comp, enum cma_state exch)
 226 {
 227         int ret;
 228
 229         mtx_lock(&id_priv->lock);
 230         if ((ret = (id_priv->state == comp)))
 231                 id_priv->state = exch;
 232         mtx_unlock(&id_priv->lock);
 233         return ret;
 234 }
 235
 236 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 237                                enum cma_state exch)
 238 {
 239         enum cma_state old;
 240
 241         mtx_lock(&id_priv->lock);
 242         old = id_priv->state;
 243         id_priv->state = exch;
 244         mtx_unlock(&id_priv->lock);
 245         return old;
 246 }
 247
 248 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 249 {
 250         return hdr->ip_version >> 4;
 251 }
 252
 253 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 254 {
 255         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 256 }
 257
 258 static inline u8 sdp_get_majv(u8 sdp_version)
 259 {
 260         return sdp_version >> 4;
 261 }
 262
 263 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 264 {
 265         return hh->ip_version >> 4;
 266 }
 267
 268 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 269 {
 270         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 271 }
 272
 273 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 274 {
 275         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 276 }
 277
 278 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 279                               struct cma_device *cma_dev)
 280 {
 281         mtx_lock(&cma_dev->lock);
 282         cma_dev->refcount++;
 283         mtx_unlock(&cma_dev->lock);
 284         id_priv->cma_dev = cma_dev;
 285         id_priv->id.device = cma_dev->device;
 286         LIST_INSERT_HEAD(&cma_dev->id_list, id_priv, list);
 287 }
 288
 289 static inline void cma_deref_dev(struct cma_device *cma_dev)
 290 {
 291         mtx_lock(&cma_dev->lock);
 292         if (--cma_dev->refcount == 0)
 293                 cv_broadcast(&cma_dev->comp);
 294         mtx_unlock(&cma_dev->lock);
 295 }
 296
 297 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 298 {
 299         LIST_REMOVE(id_priv, list);
 300         cma_deref_dev(id_priv->cma_dev);
 301         id_priv->cma_dev = NULL;
 302 }
 303
 304 #ifdef IB_SUPPORTED
 305 static int cma_set_qkey(struct ib_device *device, u8 port_num,
 306                         enum rdma_port_space ps,
 307                         struct rdma_dev_addr *dev_addr, u32 *qkey)
 308 {
 309         struct ib_sa_mcmember_rec rec;
 310         int ret = 0;
 311
 312         switch (ps) {
 313         case RDMA_PS_UDP:
 314                 *qkey = RDMA_UDP_QKEY;
 315                 break;
 316         case RDMA_PS_IPOIB:
 317                 ib_addr_get_mgid(dev_addr, &rec.mgid);
 318                 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec);
 319                 *qkey = be32_to_cpu(rec.qkey);
 320                 break;
 321         default:
 322                 break;
 323         }
 324         return ret;
 325 }
 326 #endif
 327
 328 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 329 {
 330         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 331         struct cma_device *cma_dev;
 332         union ib_gid gid;
 333         int ret = ENODEV;
 334
 335         switch (rdma_node_get_transport(dev_addr->dev_type)) {
 336 #ifdef IB_SUPPORTED
 337         case RDMA_TRANSPORT_IB:
 338                 ib_addr_get_sgid(dev_addr, &gid);
 339                 break;
 340 #endif
 341         case RDMA_TRANSPORT_IWARP:
 342                 iw_addr_get_sgid(dev_addr, &gid);
 343                 break;
 344         default:
 345                 return (ENODEV);
 346         }
 347
 348         TAILQ_FOREACH(cma_dev, &dev_list, list) {
 349                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 350                                          &id_priv->id.port_num, NULL);
 351                 if (!ret) {
 352 #ifdef IB_SUPPORTED
 353                         ret = cma_set_qkey(cma_dev->device,
 354                                            id_priv->id.port_num,
 355                                            id_priv->id.ps, dev_addr,
 356                                            &id_priv->qkey);
 357                         if (!ret)
 358 #endif
 359                                 cma_attach_to_dev(id_priv, cma_dev);
 360                         break;
 361                 }
 362         }
 363         return ret;
 364 }
 365
 366 static void cma_deref_id(struct rdma_id_private *id_priv)
 367 {
 368         mtx_lock(&id_priv->lock);
 369         if (--id_priv->refcount == 0) {
 370                 cv_broadcast(&id_priv->comp);
 371         }
 372         mtx_unlock(&id_priv->lock);
 373 }
 374
 375 static int cma_disable_remove(struct rdma_id_private *id_priv,
 376                               enum cma_state state)
 377 {
 378         int ret;
 379
 380         mtx_lock(&id_priv->lock);
 381         if (id_priv->state == state) {
 382                 id_priv->dev_remove++;
 383                 ret = 0;
 384         } else
 385                 ret = EINVAL;
 386         mtx_unlock(&id_priv->lock);
 387         return ret;
 388 }
 389
 390 static void cma_enable_remove(struct rdma_id_private *id_priv)
 391 {
 392         mtx_lock(&id_priv->lock);
 393         if (--id_priv->dev_remove == 0)
 394                 cv_broadcast(&id_priv->wait_remove);
 395         mtx_unlock(&id_priv->lock);
 396 }
 397
 398 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 399 {
 400         return (id_priv->id.device && id_priv->cm_id.ib);
 401 }
 402
 403 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 404                                   void *context, enum rdma_port_space ps)
 405 {
 406         struct rdma_id_private *id_priv;
 407
 408         id_priv = malloc(sizeof *id_priv, M_DEVBUF, M_NOWAIT);
 409         if (!id_priv)
 410                 return ERR_PTR(-ENOMEM);
 411         bzero(id_priv, sizeof *id_priv);
 412
 413         id_priv->state = CMA_IDLE;
 414         id_priv->id.context = context;
 415         id_priv->id.event_handler = event_handler;
 416         id_priv->id.ps = ps;
 417         mtx_init(&id_priv->lock, "rdma_cm_id_priv", NULL, MTX_DUPOK|MTX_DEF);
 418         cv_init(&id_priv->comp, "rdma_cm_id_priv");
 419         id_priv->refcount = 1;
 420         cv_init(&id_priv->wait_remove, "id priv wait remove");
 421         LIST_INIT(&id_priv->listen_list);
 422         arc4rand(&id_priv->seq_num, sizeof id_priv->seq_num, 0);
 423
 424         return &id_priv->id;
 425 }
 426
 427 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 428 {
 429         struct ib_qp_attr qp_attr;
 430         int qp_attr_mask, ret;
 431
 432         qp_attr.qp_state = IB_QPS_INIT;
 433         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 434         if (ret)
 435                 return ret;
 436
 437         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 438         if (ret)
 439                 return ret;
 440
 441         qp_attr.qp_state = IB_QPS_RTR;
 442         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 443         if (ret)
 444                 return ret;
 445
 446         qp_attr.qp_state = IB_QPS_RTS;
 447         qp_attr.sq_psn = 0;
 448         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 449
 450         return ret;
 451 }
 452
 453 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 454 {
 455         struct ib_qp_attr qp_attr;
 456         int qp_attr_mask, ret;
 457
 458         qp_attr.qp_state = IB_QPS_INIT;
 459         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 460         if (ret)
 461                 return ret;
 462
 463         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 464 }
 465
 466 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 467                    struct ib_qp_init_attr *qp_init_attr)
 468 {
 469         struct rdma_id_private *id_priv;
 470         struct ib_qp *qp;
 471         int ret;
 472
 473         id_priv = container_of(id, struct rdma_id_private, id);
 474         if (id->device != pd->device)
 475                 return (EINVAL);
 476
 477         qp = ib_create_qp(pd, qp_init_attr);
 478         if (IS_ERR(qp))
 479                 return PTR_ERR(qp);
 480         if (cma_is_ud_ps(id_priv->id.ps))
 481                 ret = cma_init_ud_qp(id_priv, qp);
 482         else
 483                 ret = cma_init_conn_qp(id_priv, qp);
 484         if (ret)
 485                 goto err;
 486
 487         id->qp = qp;
 488         id_priv->qp_num = qp->qp_num;
 489         id_priv->srq = (qp->srq != NULL);
 490         return 0;
 491 err:
 492         ib_destroy_qp(qp);
 493         return ret;
 494 }
 495
 496 void rdma_destroy_qp(struct rdma_cm_id *id)
 497 {
 498         ib_destroy_qp(id->qp);
 499 }
 500
 501 static int cma_modify_qp_rtr(struct rdma_cm_id *id)
 502 {
 503         struct ib_qp_attr qp_attr;
 504         int qp_attr_mask, ret;
 505
 506         if (!id->qp)
 507                 return 0;
 508
 509         /* Need to update QP attributes from default values. */
 510         qp_attr.qp_state = IB_QPS_INIT;
 511         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 512         if (ret)
 513                 return ret;
 514
 515         ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 516         if (ret)
 517                 return ret;
 518
 519         qp_attr.qp_state = IB_QPS_RTR;
 520         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 521         if (ret)
 522                 return ret;
 523
 524         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 525 }
 526
 527 #ifdef IB_SUPPORTED
 528 static int cma_modify_qp_rts(struct rdma_cm_id *id)
 529 {
 530         struct ib_qp_attr qp_attr;
 531         int qp_attr_mask, ret;
 532
 533         if (!id->qp)
 534                 return 0;
 535
 536         qp_attr.qp_state = IB_QPS_RTS;
 537         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 538         if (ret)
 539                 return ret;
 540
 541         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 542 }
 543 #endif
 544
 545 static int cma_modify_qp_err(struct rdma_cm_id *id)
 546 {
 547         struct ib_qp_attr qp_attr;
 548
 549         if (!id->qp)
 550                 return 0;
 551
 552         qp_attr.qp_state = IB_QPS_ERR;
 553         return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE);
 554 }
 555
 556 #ifdef IB_SUPPORTED
 557 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 558                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 559 {
 560         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 561         int ret;
 562
 563         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 564                                   ib_addr_get_pkey(dev_addr),
 565                                   &qp_attr->pkey_index);
 566         if (ret)
 567                 return ret;
 568
 569         qp_attr->port_num = id_priv->id.port_num;
 570         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 571
 572         if (cma_is_ud_ps(id_priv->id.ps)) {
 573                 qp_attr->qkey = id_priv->qkey;
 574                 *qp_attr_mask |= IB_QP_QKEY;
 575         } else {
 576                 qp_attr->qp_access_flags = 0;
 577                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 578         }
 579         return 0;
 580 }
 581 #endif
 582
 583 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 584                        int *qp_attr_mask)
 585 {
 586         struct rdma_id_private *id_priv;
 587         int ret = 0;
 588
 589         id_priv = container_of(id, struct rdma_id_private, id);
 590 #ifdef IB_SUPPORTED
 591         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 592         case RDMA_TRANSPORT_IB:
 593                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 594                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 595                 else
 596                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 597                                                  qp_attr_mask);
 598                 if (qp_attr->qp_state == IB_QPS_RTR)
 599                         qp_attr->rq_psn = id_priv->seq_num;
 600                 break;
 601         case RDMA_TRANSPORT_IWARP:
 602 #endif
 603                 if (!id_priv->cm_id.iw) {
 604                         qp_attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 605                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 606                 } else
 607                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 608                                                  qp_attr_mask);
 609 #ifdef IB_SUPPORTED
 610                 break;
 611         default:
 612                 ret = ENOSYS;
 613                 break;
 614         }
 615 #endif
 616
 617         return ret;
 618 }
 619
 620 static inline int cma_zero_addr(struct sockaddr *addr)
 621 {
 622         struct in6_addr *ip6;
 623
 624         if (addr->sa_family == AF_INET)
 625                 return in_nullhost(((struct sockaddr_in *) addr)->sin_addr);
 626         else {
 627                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 628                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 629                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 630         }
 631 }
 632
 633 static inline int cma_loopback_addr(struct sockaddr *addr)
 634 {
 635         return ((struct sockaddr_in *)addr)->sin_addr.s_addr == INADDR_LOOPBACK;
 636 }
 637
 638 static inline int cma_any_addr(struct sockaddr *addr)
 639 {
 640         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 641 }
 642
 643 static inline __be16 cma_port(struct sockaddr *addr)
 644 {
 645         if (addr->sa_family == AF_INET)
 646                 return ((struct sockaddr_in *) addr)->sin_port;
 647         else
 648                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 649 }
 650
 651 static inline int cma_any_port(struct sockaddr *addr)
 652 {
 653         return !cma_port(addr);
 654 }
 655
 656 #ifdef IB_SUPPORTED
 657 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 658                             u8 *ip_ver, __u16 *port,
 659                             union cma_ip_addr **src, union cma_ip_addr **dst)
 660 {
 661         switch (ps) {
 662         case RDMA_PS_SDP:
 663                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 664                     SDP_MAJ_VERSION)
 665                         return (EINVAL);
 666
 667                 *ip_ver = sdp_get_ip_ver(hdr);
 668                 *port   = ((struct sdp_hh *) hdr)->port;
 669                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 670                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 671                 break;
 672         default:
 673                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 674                         return (EINVAL);
 675
 676                 *ip_ver = cma_get_ip_ver(hdr);
 677                 *port   = ((struct cma_hdr *) hdr)->port;
 678                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 679                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 680                 break;
 681         }
 682
 683         if (*ip_ver != 4 && *ip_ver != 6)
 684                 return (EINVAL);
 685         return 0;
 686 }
 687
 688 static void cma_save_net_info(struct rdma_addr *addr,
 689                               struct rdma_addr *listen_addr,
 690                               u8 ip_ver, __u16 port,
 691                               union cma_ip_addr *src, union cma_ip_addr *dst)
 692 {
 693         struct sockaddr_in *listen4, *ip4;
 694         struct sockaddr_in6 *listen6, *ip6;
 695
 696         switch (ip_ver) {
 697         case 4:
 698                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 699                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 700                 ip4->sin_family = listen4->sin_family;
 701                 ip4->sin_addr.s_addr = dst->ip4.addr;
 702                 ip4->sin_port = listen4->sin_port;
 703
 704                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 705                 ip4->sin_family = listen4->sin_family;
 706                 ip4->sin_addr.s_addr = src->ip4.addr;
 707                 ip4->sin_port = port;
 708                 break;
 709         case 6:
 710                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 711                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 712                 ip6->sin6_family = listen6->sin6_family;
 713                 ip6->sin6_addr = dst->ip6;
 714                 ip6->sin6_port = listen6->sin6_port;
 715
 716                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 717                 ip6->sin6_family = listen6->sin6_family;
 718                 ip6->sin6_addr = src->ip6;
 719                 ip6->sin6_port = port;
 720                 break;
 721         default:
 722                 break;
 723         }
 724 }
 725 #endif
 726
 727 static inline int cma_user_data_offset(enum rdma_port_space ps)
 728 {
 729         switch (ps) {
 730         case RDMA_PS_SDP:
 731                 return 0;
 732         default:
 733                 return sizeof(struct cma_hdr);
 734         }
 735 }
 736
 737 static void cma_cancel_route(struct rdma_id_private *id_priv)
 738 {
 739 #ifdef IB_SUPPORTED
 740         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 741         case RDMA_TRANSPORT_IB:
 742                 if (id_priv->query)
 743                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 744                 break;
 745         default:
 746                 break;
 747         }
 748 #endif
 749 }
 750
 751 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 752 {
 753         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 754                cma_any_addr(&id_priv->id.route.addr.src_addr);
 755 }
 756
 757 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 758 {
 759         cma_exch(id_priv, CMA_DESTROYING);
 760
 761         if (id_priv->cma_dev) {
 762 #ifdef IB_SUPPORTED
 763                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 764                 case RDMA_TRANSPORT_IB:
 765                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 766                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 767                         break;
 768                 case RDMA_TRANSPORT_IWARP:
 769 #endif
 770                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 771                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 772 #ifdef IB_SUPPORTED
 773                         break;
 774                 default:
 775                         break;
 776                 }
 777 #endif
 778                 cma_detach_from_dev(id_priv);
 779         }
 780         LIST_REMOVE(id_priv, listen_entry);
 781
 782         cma_deref_id(id_priv);
 783         mtx_lock(&id_priv->lock);
 784         if (id_priv->refcount)
 785                 cv_wait(&id_priv->comp, &id_priv->lock);
 786         mtx_unlock(&id_priv->lock);
 787
 788         free(id_priv, M_DEVBUF);
 789 }
 790
 791 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 792 {
 793         struct rdma_id_private *dev_id_priv;
 794
 795         mtx_lock(&lock);
 796         LIST_REMOVE(id_priv, list);
 797
 798         while (!LIST_EMPTY(&id_priv->listen_list)) {
 799                 dev_id_priv = LIST_FIRST(&id_priv->listen_list);
 800                 cma_destroy_listen(dev_id_priv);
 801         }
 802         mtx_unlock(&lock);
 803 }
 804
 805 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 806                                  enum cma_state state)
 807 {
 808         switch (state) {
 809         case CMA_ADDR_QUERY:
 810                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 811                 break;
 812         case CMA_ROUTE_QUERY:
 813                 cma_cancel_route(id_priv);
 814                 break;
 815         case CMA_LISTEN:
 816                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 817                     !id_priv->cma_dev)
 818                         cma_cancel_listens(id_priv);
 819                 break;
 820         default:
 821                 break;
 822         }
 823 }
 824
 825 static void cma_release_port(struct rdma_id_private *id_priv)
 826 {
 827         struct rdma_bind_list *bind_list = id_priv->bind_list;
 828
 829         if (!bind_list)
 830                 return;
 831
 832         mtx_lock(&lock);
 833         TAILQ_REMOVE(&bind_list->owners, id_priv, node);
 834         if (TAILQ_EMPTY(&bind_list->owners)) {
 835                 kvl_delete(bind_list->ps, bind_list->port);
 836                 free(bind_list, M_DEVBUF);
 837         }
 838         mtx_unlock(&lock);
 839         if (id_priv->so)
 840                 soclose(id_priv->so);
 841 }
 842
 843 #ifdef IB_SUPPORTED
 844 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 845 {
 846         struct cma_multicast *mc;
 847
 848         while (!LIST_EMPTY(&id_priv->mc_list)) {
 849                 mc = LIST_FIRST(&id_priv->mc_list);
 850                 LIST_REMOVE(mc, list);
 851                 ib_sa_free_multicast(mc->multicast.ib);
 852                 free(mc, M_DEVBUF);
 853         }
 854 }
 855 #endif
 856
 857 void rdma_destroy_id(struct rdma_cm_id *id)
 858 {
 859         struct rdma_id_private *id_priv;
 860         enum cma_state state;
 861
 862         id_priv = container_of(id, struct rdma_id_private, id);
 863         state = cma_exch(id_priv, CMA_DESTROYING);
 864         cma_cancel_operation(id_priv, state);
 865
 866         mtx_lock(&lock);
 867         if (id_priv->cma_dev) {
 868                 mtx_unlock(&lock);
 869 #ifdef IB_SUPPORTED
 870                 switch (rdma_node_get_transport(id->device->node_type)) {
 871                 case RDMA_TRANSPORT_IB:
 872                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 873                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 874                         break;
 875                 case RDMA_TRANSPORT_IWARP:
 876 #endif
 877                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 878                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 879 #ifdef IB_SUPPORTED
 880                         break;
 881                 default:
 882                         break;
 883                 }
 884                 cma_leave_mc_groups(id_priv);
 885 #endif
 886                 mtx_lock(&lock);
 887                 cma_detach_from_dev(id_priv);
 888         }
 889         mtx_unlock(&lock);
 890         cma_release_port(id_priv);
 891         cma_deref_id(id_priv);
 892         mtx_lock(&id_priv->lock);
 893         PANIC_IF(id_priv->refcount < 0);
 894         if (id_priv->refcount)
 895                 cv_wait(&id_priv->comp, &id_priv->lock);
 896         mtx_unlock(&id_priv->lock);
 897         free(id_priv->id.route.path_rec, M_DEVBUF);
 898         free(id_priv, M_DEVBUF);
 899 }
 900
 901 #ifdef IB_SUPPORTED
 902 static int cma_rep_recv(struct rdma_id_private *id_priv)
 903 {
 904         int ret;
 905
 906         ret = cma_modify_qp_rtr(&id_priv->id);
 907         if (ret)
 908                 goto reject;
 909
 910         ret = cma_modify_qp_rts(&id_priv->id);
 911         if (ret)
 912                 goto reject;
 913
 914         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 915         if (ret)
 916                 goto reject;
 917
 918         return 0;
 919 reject:
 920         cma_modify_qp_err(&id_priv->id);
 921         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 922                        NULL, 0, NULL, 0);
 923         return ret;
 924 }
 925
 926 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 927 {
 928         if (id_priv->id.ps == RDMA_PS_SDP &&
 929             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 930             SDP_MAJ_VERSION)
 931                 return (EINVAL);
 932
 933         return 0;
 934 }
 935
 936 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 937                                    struct ib_cm_rep_event_param *rep_data,
 938                                    void *private_data)
 939 {
 940         event->param.conn.private_data = private_data;
 941         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 942         event->param.conn.responder_resources = rep_data->responder_resources;
 943         event->param.conn.initiator_depth = rep_data->initiator_depth;
 944         event->param.conn.flow_control = rep_data->flow_control;
 945         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 946         event->param.conn.srq = rep_data->srq;
 947         event->param.conn.qp_num = rep_data->remote_qpn;
 948 }
 949
 950 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 951 {
 952         struct rdma_id_private *id_priv = cm_id->context;
 953         struct rdma_cm_event event;
 954         int ret = 0;
 955
 956         if (cma_disable_remove(id_priv, CMA_CONNECT))
 957                 return 0;
 958
 959         memset(&event, 0, sizeof event);
 960         switch (ib_event->event) {
 961         case IB_CM_REQ_ERROR:
 962         case IB_CM_REP_ERROR:
 963                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 964                 event.status = ETIMEDOUT;
 965                 break;
 966         case IB_CM_REP_RECEIVED:
 967                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 968                 if (event.status)
 969                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 970                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 971                         event.status = cma_rep_recv(id_priv);
 972                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 973                                                      RDMA_CM_EVENT_ESTABLISHED;
 974                 } else
 975                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 976                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 977                                        ib_event->private_data);
 978                 break;
 979         case IB_CM_RTU_RECEIVED:
 980         case IB_CM_USER_ESTABLISHED:
 981                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 982                 break;
 983         case IB_CM_DREQ_ERROR:
 984                 event.status = ETIMEDOUT; /* fall through */
 985         case IB_CM_DREQ_RECEIVED:
 986         case IB_CM_DREP_RECEIVED:
 987                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 988                         goto out;
 989                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 990                 break;
 991         case IB_CM_TIMEWAIT_EXIT:
 992         case IB_CM_MRA_RECEIVED:
 993                 /* ignore event */
 994                 goto out;
 995         case IB_CM_REJ_RECEIVED:
 996                 cma_modify_qp_err(&id_priv->id);
 997                 event.status = ib_event->param.rej_rcvd.reason;
 998                 event.event = RDMA_CM_EVENT_REJECTED;
 999                 event.param.conn.private_data = ib_event->private_data;
1000                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1001                 break;
1002         default:
1003                 log(LOG_ERR, "RDMA CMA: unexpected IB CM event: %d",
1004                        ib_event->event);
1005                 goto out;
1006         }
1007
1008         ret = id_priv->id.event_handler(&id_priv->id, &event);
1009         if (ret) {
1010                 /* Destroy the CM ID by returning a non-zero value. */
1011                 id_priv->cm_id.ib = NULL;
1012                 cma_exch(id_priv, CMA_DESTROYING);
1013                 cma_enable_remove(id_priv);
1014                 rdma_destroy_id(&id_priv->id);
1015                 return ret;
1016         }
1017 out:
1018         cma_enable_remove(id_priv);
1019         return ret;
1020 }
1021
1022 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1023                                                struct ib_cm_event *ib_event)
1024 {
1025         struct rdma_id_private *id_priv;
1026         struct rdma_cm_id *id;
1027         struct rdma_route *rt;
1028         union cma_ip_addr *src, *dst;
1029         __u16 port;
1030         u8 ip_ver;
1031
1032         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1033                              &ip_ver, &port, &src, &dst))
1034                 goto err;
1035
1036         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1037                             listen_id->ps);
1038         if (IS_ERR(id))
1039                 goto err;
1040
1041         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1042                           ip_ver, port, src, dst);
1043
1044         rt = &id->route;
1045         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1046         rt->path_rec = malloc(sizeof *rt->path_rec * rt->num_paths,
1047                                M_DEVBUF, M_NOWAIT);
1048         if (!rt->path_rec)
1049                 goto destroy_id;
1050
1051         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1052         if (rt->num_paths == 2)
1053                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1054
1055         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1056         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1057         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1058         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
1059
1060         id_priv = container_of(id, struct rdma_id_private, id);
1061         id_priv->state = CMA_CONNECT;
1062         return id_priv;
1063
1064 destroy_id:
1065         rdma_destroy_id(id);
1066 err:
1067         return NULL;
1068 }
1069
1070 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1071                                               struct ib_cm_event *ib_event)
1072 {
1073         struct rdma_id_private *id_priv;
1074         struct rdma_cm_id *id;
1075         union cma_ip_addr *src, *dst;
1076         __u16 port;
1077         u8 ip_ver;
1078         int ret;
1079
1080         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1081                             listen_id->ps);
1082         if (IS_ERR(id))
1083                 return NULL;
1084
1085
1086         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1087                              &ip_ver, &port, &src, &dst))
1088                 goto err;
1089
1090         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1091                           ip_ver, port, src, dst);
1092
1093         ret = rdma_translate_ip(&id->route.addr.src_addr,
1094                                 &id->route.addr.dev_addr);
1095         if (ret)
1096                 goto err;
1097
1098         id_priv = container_of(id, struct rdma_id_private, id);
1099         id_priv->state = CMA_CONNECT;
1100         return id_priv;
1101 err:
1102         rdma_destroy_id(id);
1103         return NULL;
1104 }
1105
1106 static void cma_set_req_event_data(struct rdma_cm_event *event,
1107                                    struct ib_cm_req_event_param *req_data,
1108                                    void *private_data, int offset)
1109 {
1110         event->param.conn.private_data = private_data + offset;
1111         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1112         event->param.conn.responder_resources = req_data->responder_resources;
1113         event->param.conn.initiator_depth = req_data->initiator_depth;
1114         event->param.conn.flow_control = req_data->flow_control;
1115         event->param.conn.retry_count = req_data->retry_count;
1116         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1117         event->param.conn.srq = req_data->srq;
1118         event->param.conn.qp_num = req_data->remote_qpn;
1119 }
1120
1121 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1122 {
1123         struct rdma_id_private *listen_id, *conn_id;
1124         struct rdma_cm_event event;
1125         int offset, ret;
1126
1127         listen_id = cm_id->context;
1128         if (cma_disable_remove(listen_id, CMA_LISTEN))
1129                 return (ECONNABORTED);
1130
1131         memset(&event, 0, sizeof event);
1132         offset = cma_user_data_offset(listen_id->id.ps);
1133         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1134         if (cma_is_ud_ps(listen_id->id.ps)) {
1135                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1136                 event.param.ud.private_data = ib_event->private_data + offset;
1137                 event.param.ud.private_data_len =
1138                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1139         } else {
1140                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1141                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1142                                        ib_event->private_data, offset);
1143         }
1144         if (!conn_id) {
1145                 ret = ENOMEM;
1146                 goto out;
1147         }
1148
1149         mtx_lock(&conn_id->lock);
1150         conn_id->dev_remove++;
1151         mtx_unlock(&conn_id->lock);
1152         mtx_lock(&lock);
1153         ret = cma_acquire_dev(conn_id);
1154         mtx_unlock(&lock);
1155         if (ret)
1156                 goto release_conn_id;
1157
1158         conn_id->cm_id.ib = cm_id;
1159         cm_id->context = conn_id;
1160         cm_id->cm_handler = cma_ib_handler;
1161
1162         ret = conn_id->id.event_handler(&conn_id->id, &event);
1163         if (!ret)
1164                 goto out;
1165
1166         /* Destroy the CM ID by returning a non-zero value. */
1167         conn_id->cm_id.ib = NULL;
1168
1169 release_conn_id:
1170         cma_exch(conn_id, CMA_DESTROYING);
1171         cma_enable_remove(conn_id);
1172         rdma_destroy_id(&conn_id->id);
1173
1174 out:
1175         cma_enable_remove(listen_id);
1176         return ret;
1177 }
1178
1179 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1180 {
1181         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1182 }
1183
1184 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1185                                  struct ib_cm_compare_data *compare)
1186 {
1187         struct cma_hdr *cma_data, *cma_mask;
1188         struct sdp_hh *sdp_data, *sdp_mask;
1189         __u32 ip4_addr;
1190         struct in6_addr ip6_addr;
1191
1192         memset(compare, 0, sizeof *compare);
1193         cma_data = (void *) compare->data;
1194         cma_mask = (void *) compare->mask;
1195         sdp_data = (void *) compare->data;
1196         sdp_mask = (void *) compare->mask;
1197
1198         switch (addr->sa_family) {
1199         case AF_INET:
1200                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1201                 if (ps == RDMA_PS_SDP) {
1202                         sdp_set_ip_ver(sdp_data, 4);
1203                         sdp_set_ip_ver(sdp_mask, 0xF);
1204                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1205                         sdp_mask->dst_addr.ip4.addr = ~0;
1206                 } else {
1207                         cma_set_ip_ver(cma_data, 4);
1208                         cma_set_ip_ver(cma_mask, 0xF);
1209                         cma_data->dst_addr.ip4.addr = ip4_addr;
1210                         cma_mask->dst_addr.ip4.addr = ~0;
1211                 }
1212                 break;
1213         case AF_INET6:
1214                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1215                 if (ps == RDMA_PS_SDP) {
1216                         sdp_set_ip_ver(sdp_data, 6);
1217                         sdp_set_ip_ver(sdp_mask, 0xF);
1218                         sdp_data->dst_addr.ip6 = ip6_addr;
1219                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1220                                sizeof sdp_mask->dst_addr.ip6);
1221                 } else {
1222                         cma_set_ip_ver(cma_data, 6);
1223                         cma_set_ip_ver(cma_mask, 0xF);
1224                         cma_data->dst_addr.ip6 = ip6_addr;
1225                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1226                                sizeof cma_mask->dst_addr.ip6);
1227                 }
1228                 break;
1229         default:
1230                 break;
1231         }
1232 }
1233 #endif /* IB_SUPPORTED */
1234
1235 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1236 {
1237         struct rdma_id_private *id_priv = iw_id->context;
1238         struct rdma_cm_event event;
1239         struct sockaddr_in *sin;
1240         int ret = 0;
1241
1242         if (cma_disable_remove(id_priv, CMA_CONNECT))
1243                 return 0;
1244
1245         memset(&event, 0, sizeof event);
1246         switch (iw_event->event) {
1247         case IW_CM_EVENT_CLOSE:
1248                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1249                 break;
1250         case IW_CM_EVENT_CONNECT_REPLY:
1251                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1252                 *sin = iw_event->local_addr;
1253                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1254                 *sin = iw_event->remote_addr;
1255                 switch ((int)iw_event->status) {
1256                 case 0:
1257                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1258                         break;
1259                 case ECONNRESET:
1260                 case ECONNREFUSED:
1261                         event.event = RDMA_CM_EVENT_REJECTED;
1262                         break;
1263                 case ETIMEDOUT:
1264                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1265                         break;
1266                 default:
1267                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1268                         break;
1269                 }
1270                 break;
1271         case IW_CM_EVENT_ESTABLISHED:
1272                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1273                 break;
1274         default:
1275                 panic("unknown event type %d", iw_event->event);
1276
1277         }
1278
1279         event.status = iw_event->status;
1280         event.param.conn.private_data = iw_event->private_data;
1281         event.param.conn.private_data_len = iw_event->private_data_len;
1282         ret = id_priv->id.event_handler(&id_priv->id, &event);
1283         if (ret) {
1284                 /* Destroy the CM ID by returning a non-zero value. */
1285                 id_priv->cm_id.iw = NULL;
1286                 cma_exch(id_priv, CMA_DESTROYING);
1287                 cma_enable_remove(id_priv);
1288                 rdma_destroy_id(&id_priv->id);
1289                 return ret;
1290         }
1291
1292         cma_enable_remove(id_priv);
1293         return ret;
1294 }
1295
1296 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1297                                struct iw_cm_event *iw_event)
1298 {
1299         struct rdma_cm_id *new_cm_id;
1300         struct rdma_id_private *listen_id, *conn_id;
1301         struct sockaddr_in *sin;
1302         struct ifnet *dev;
1303         struct rdma_cm_event event;
1304         int ret;
1305         struct ifaddr *ifa;
1306         uint16_t port;
1307
1308         listen_id = cm_id->context;
1309         if (cma_disable_remove(listen_id, CMA_LISTEN))
1310                 return (ECONNABORTED);
1311
1312         /* Create a new RDMA id for the new IW CM ID */
1313         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1314                                    listen_id->id.context,
1315                                    RDMA_PS_TCP);
1316         if (!new_cm_id) {
1317                 ret = ENOMEM;
1318                 goto out;
1319         }
1320         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1321         mtx_lock(&conn_id->lock);
1322         ++conn_id->dev_remove;
1323         mtx_unlock(&conn_id->lock);
1324         conn_id->state = CMA_CONNECT;
1325
1326         port = iw_event->local_addr.sin_port;
1327         iw_event->local_addr.sin_port = 0;
1328         ifa = ifa_ifwithaddr((struct sockaddr *)&iw_event->local_addr);
1329         iw_event->local_addr.sin_port = port;
1330         if (!ifa) {
1331                 ret = EADDRNOTAVAIL;
1332                 cma_enable_remove(conn_id);
1333                 rdma_destroy_id(new_cm_id);
1334                 goto out;
1335         }
1336         dev = ifa->ifa_ifp;
1337         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1338         ifa_free(ifa);
1339         if (ret) {
1340                 cma_enable_remove(conn_id);
1341                 rdma_destroy_id(new_cm_id);
1342                 goto out;
1343         }
1344
1345         mtx_lock(&lock);
1346         ret = cma_acquire_dev(conn_id);
1347         mtx_unlock(&lock);
1348         if (ret) {
1349                 cma_enable_remove(conn_id);
1350                 rdma_destroy_id(new_cm_id);
1351                 goto out;
1352         }
1353
1354         conn_id->cm_id.iw = cm_id;
1355         cm_id->context = conn_id;
1356         cm_id->cm_handler = cma_iw_handler;
1357
1358         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1359         *sin = iw_event->local_addr;
1360         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1361         *sin = iw_event->remote_addr;
1362         conn_id->so = cm_id->so;
1363
1364         memset(&event, 0, sizeof event);
1365         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1366         event.param.conn.private_data = iw_event->private_data;
1367         event.param.conn.private_data_len = iw_event->private_data_len;
1368         ret = conn_id->id.event_handler(&conn_id->id, &event);
1369         if (ret) {
1370                 /* User wants to destroy the CM ID */
1371                 conn_id->cm_id.iw = NULL;
1372                 cma_exch(conn_id, CMA_DESTROYING);
1373                 cma_enable_remove(conn_id);
1374                 rdma_destroy_id(&conn_id->id);
1375         }
1376
1377 out:
1378         cma_enable_remove(listen_id);
1379         return ret;
1380 }
1381
1382 #ifdef IB_SUPPORTED
1383 static int cma_ib_listen(struct rdma_id_private *id_priv)
1384 {
1385         struct ib_cm_compare_data compare_data;
1386         struct sockaddr *addr;
1387         __be64 svc_id;
1388         int ret;
1389
1390         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1391                                             id_priv);
1392         if (IS_ERR(id_priv->cm_id.ib))
1393                 return PTR_ERR(id_priv->cm_id.ib);
1394
1395         addr = &id_priv->id.route.addr.src_addr;
1396         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1397         if (cma_any_addr(addr))
1398                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1399         else {
1400                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1401                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1402         }
1403
1404         if (ret) {
1405                 ib_destroy_cm_id(id_priv->cm_id.ib);
1406                 id_priv->cm_id.ib = NULL;
1407         }
1408
1409         return ret;
1410 }
1411 #endif
1412
1413 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1414 {
1415         int ret;
1416         struct sockaddr_in *sin;
1417
1418         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device, id_priv->so,
1419                                             iw_conn_req_handler, id_priv);
1420         if (IS_ERR(id_priv->cm_id.iw))
1421                 return PTR_ERR(id_priv->cm_id.iw);
1422
1423         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1424         id_priv->cm_id.iw->local_addr = *sin;
1425
1426         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1427
1428         if (ret) {
1429                 iw_destroy_cm_id(id_priv->cm_id.iw);
1430                 id_priv->cm_id.iw = NULL;
1431         }
1432
1433         return ret;
1434 }
1435
1436 static int cma_listen_handler(struct rdma_cm_id *id,
1437                               struct rdma_cm_event *event)
1438 {
1439         struct rdma_id_private *id_priv = id->context;
1440
1441         id->context = id_priv->id.context;
1442         id->event_handler = id_priv->id.event_handler;
1443         return id_priv->id.event_handler(id, event);
1444 }
1445
1446 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1447                               struct cma_device *cma_dev)
1448 {
1449         struct rdma_id_private *dev_id_priv;
1450         struct rdma_cm_id *id;
1451         int ret;
1452
1453         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1454         if (IS_ERR(id))
1455                 return;
1456
1457         dev_id_priv = container_of(id, struct rdma_id_private, id);
1458
1459         dev_id_priv->state = CMA_ADDR_BOUND;
1460         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1461                ip_addr_size(&id_priv->id.route.addr.src_addr));
1462         dev_id_priv->so = id_priv->so; /* XXX */
1463
1464         cma_attach_to_dev(dev_id_priv, cma_dev);
1465         LIST_INSERT_HEAD(&id_priv->listen_list, dev_id_priv, listen_entry);
1466
1467         ret = rdma_listen(id, id_priv->backlog);
1468         if (ret)
1469                 goto err;
1470
1471         return;
1472 err:
1473         cma_destroy_listen(dev_id_priv);
1474 }
1475
1476 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1477 {
1478         struct cma_device *cma_dev;
1479
1480         mtx_lock(&lock);
1481         LIST_INSERT_HEAD(&listen_any_list, id_priv, list);
1482         TAILQ_FOREACH(cma_dev, &dev_list, list)
1483                 cma_listen_on_dev(id_priv, cma_dev);
1484         mtx_unlock(&lock);
1485 }
1486
1487 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1488 {
1489         struct sockaddr_in addr_in;
1490
1491         memset(&addr_in, 0, sizeof addr_in);
1492         addr_in.sin_family = af;
1493         addr_in.sin_len = sizeof addr_in;
1494         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1495 }
1496
1497 int rdma_listen(struct rdma_cm_id *id, int backlog)
1498 {
1499         struct rdma_id_private *id_priv;
1500         int ret;
1501
1502         id_priv = container_of(id, struct rdma_id_private, id);
1503         if (id_priv->state == CMA_IDLE) {
1504                 ret = cma_bind_any(id, AF_INET);
1505                 if (ret)
1506                         return ret;
1507         }
1508
1509         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1510                 return (EINVAL);
1511
1512         id_priv->backlog = backlog;
1513         if (id->device) {
1514 #ifdef IB_SUPPORTED
1515                 switch (rdma_node_get_transport(id->device->node_type)) {
1516                 case RDMA_TRANSPORT_IB:
1517                         ret = cma_ib_listen(id_priv);
1518                         if (ret)
1519                                 goto err;
1520                         break;
1521                 case RDMA_TRANSPORT_IWARP:
1522 #endif
1523                         ret = cma_iw_listen(id_priv, backlog);
1524                         if (ret)
1525                                 goto err;
1526 #ifdef IB_SUPPORTED
1527                         break;
1528                 default:
1529                         ret = ENOSYS;
1530                         goto err;
1531                 }
1532 #endif
1533         } else
1534                 cma_listen_on_all(id_priv);
1535
1536         return 0;
1537 err:
1538         id_priv->backlog = 0;
1539         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1540         return ret;
1541 }
1542
1543 #ifdef IB_SUPPORTED
1544 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1545                               void *context)
1546 {
1547         struct cma_work *work = context;
1548         struct rdma_route *route;
1549
1550         route = &work->id->id.route;
1551
1552         if (!status) {
1553                 route->num_paths = 1;
1554                 *route->path_rec = *path_rec;
1555         } else {
1556                 work->old_state = CMA_ROUTE_QUERY;
1557                 work->new_state = CMA_ADDR_RESOLVED;
1558                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1559                 work->event.status = status;
1560         }
1561
1562         taskqueue_enqueue(cma_wq, &work->task);
1563 }
1564
1565 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1566                               struct cma_work *work)
1567 {
1568         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1569         struct ib_sa_path_rec path_rec;
1570
1571         memset(&path_rec, 0, sizeof path_rec);
1572         ib_addr_get_sgid(addr, &path_rec.sgid);
1573         ib_addr_get_dgid(addr, &path_rec.dgid);
1574         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1575         path_rec.numb_path = 1;
1576         path_rec.reversible = 1;
1577
1578         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1579                                 id_priv->id.port_num, &path_rec,
1580                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1581                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1582                                 IB_SA_PATH_REC_REVERSIBLE,
1583                                 timeout_ms, M_NOWAIT,
1584                                 cma_query_handler, work, &id_priv->query);
1585
1586         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1587 }
1588 #endif
1589
1590 static void cma_work_handler(void *context, int pending)
1591 {
1592         struct cma_work *work = context;
1593         struct rdma_id_private *id_priv = work->id;
1594         int destroy = 0;
1595
1596         mtx_lock(&id_priv->lock);
1597         ++id_priv->dev_remove;
1598         mtx_unlock(&id_priv->lock);
1599         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1600                 goto out;
1601
1602         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1603                 cma_exch(id_priv, CMA_DESTROYING);
1604                 destroy = 1;
1605         }
1606 out:
1607         cma_enable_remove(id_priv);
1608         cma_deref_id(id_priv);
1609         if (destroy)
1610                 rdma_destroy_id(&id_priv->id);
1611         free(work, M_DEVBUF);
1612 }
1613
1614 #ifdef IB_SUPPORTED
1615 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1616 {
1617         struct rdma_route *route = &id_priv->id.route;
1618         struct cma_work *work;
1619         int ret;
1620
1621         work = malloc(sizeof *work, M_DEVBUF, M_NOWAIT);
1622         if (!work)
1623                 return (ENOMEM);
1624         bzero(work, sizeof *work);
1625
1626         work->id = id_priv;
1627         TASK_INIT(&work->task, 0, cma_work_handler, work);
1628         work->old_state = CMA_ROUTE_QUERY;
1629         work->new_state = CMA_ROUTE_RESOLVED;
1630         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1631
1632         route->path_rec = malloc(sizeof *route->path_rec, M_DEVBUF, M_NOWAIT);
1633         if (!route->path_rec) {
1634                 ret = ENOMEM;
1635                 goto err1;
1636         }
1637
1638         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1639         if (ret)
1640                 goto err2;
1641
1642         return 0;
1643 err2:
1644         free(route->path_rec, M_DEVBUF);
1645         route->path_rec = NULL;
1646 err1:
1647         free(work, M_DEVBUF);
1648         return ret;
1649 }
1650
1651 int rdma_set_ib_paths(struct rdma_cm_id *id,
1652                       struct ib_sa_path_rec *path_rec, int num_paths)
1653 {
1654         struct rdma_id_private *id_priv;
1655         int ret;
1656
1657         id_priv = container_of(id, struct rdma_id_private, id);
1658         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1659                 return (EINVAL);
1660
1661         id->route.path_rec = malloc(sizeof *path_rec * num_paths, M_DEVBUF, M_NOWAIT);
1662         if (!id->route.path_rec) {
1663                 ret = ENOMEM;
1664                 goto err;
1665         }
1666
1667         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1668         return 0;
1669 err:
1670         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1671         return ret;
1672 }
1673 #endif
1674
1675 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1676 {
1677         struct cma_work *work;
1678
1679         work = malloc(sizeof *work, M_DEVBUF, M_NOWAIT);
1680         if (!work)
1681                 return (ENOMEM);
1682         bzero(work, sizeof *work);
1683
1684         work->id = id_priv;
1685         TASK_INIT(&work->task, 0, cma_work_handler, work);
1686         work->old_state = CMA_ROUTE_QUERY;
1687         work->new_state = CMA_ROUTE_RESOLVED;
1688         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1689         taskqueue_enqueue(cma_wq, &work->task);
1690         return 0;
1691 }
1692
1693 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1694 {
1695         struct rdma_id_private *id_priv;
1696         int ret;
1697
1698         id_priv = container_of(id, struct rdma_id_private, id);
1699         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1700                 return (EINVAL);
1701
1702         mtx_lock(&id_priv->lock);
1703         id_priv->refcount++;
1704         mtx_unlock(&id_priv->lock);
1705 #ifdef IB_SUPPORTED
1706         switch (rdma_node_get_transport(id->device->node_type)) {
1707         case RDMA_TRANSPORT_IB:
1708                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1709                 break;
1710         case RDMA_TRANSPORT_IWARP:
1711 #endif
1712                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1713 #ifdef IB_SUPPORTED
1714                 break;
1715         default:
1716                 ret = ENOSYS;
1717                 break;
1718         }
1719 #endif
1720         if (ret)
1721                 goto err;
1722
1723         return 0;
1724 err:
1725         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1726         cma_deref_id(id_priv);
1727         return ret;
1728 }
1729
1730 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1731 {
1732         struct cma_device *cma_dev;
1733         struct ib_port_attr port_attr;
1734         union ib_gid gid;
1735         u16 pkey;
1736         int ret;
1737         u8 p;
1738
1739         mtx_lock(&lock);
1740         if (TAILQ_EMPTY(&dev_list)) {
1741                 ret = ENODEV;
1742                 goto out;
1743         }
1744         TAILQ_FOREACH(cma_dev, &dev_list, list)
1745                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1746                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1747                             port_attr.state == IB_PORT_ACTIVE)
1748                                 goto port_found;
1749
1750         p = 1;
1751         cma_dev = TAILQ_FIRST(&dev_list);
1752
1753 port_found:
1754         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1755         if (ret)
1756                 goto out;
1757
1758         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1759         if (ret)
1760                 goto out;
1761
1762         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1763         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1764         id_priv->id.port_num = p;
1765         cma_attach_to_dev(id_priv, cma_dev);
1766 out:
1767         mtx_unlock(&lock);
1768         return ret;
1769 }
1770
1771 static void addr_handler(int status, struct sockaddr *src_addr,
1772                          struct rdma_dev_addr *dev_addr, void *context)
1773 {
1774         struct rdma_id_private *id_priv = context;
1775         struct rdma_cm_event event;
1776
1777         memset(&event, 0, sizeof event);
1778         mtx_lock(&id_priv->lock);
1779         ++id_priv->dev_remove;
1780         mtx_unlock(&id_priv->lock);
1781
1782         /*
1783          * Grab mutex to block rdma_destroy_id() from removing the device while
1784          * we're trying to acquire it.
1785          */
1786         mtx_lock(&lock);
1787         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1788                 mtx_unlock(&lock);
1789                 goto out;
1790         }
1791
1792         if (!status && !id_priv->cma_dev)
1793                 status = cma_acquire_dev(id_priv);
1794         mtx_unlock(&lock);
1795
1796         if (status) {
1797                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1798                         goto out;
1799                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1800                 event.status = status;
1801         } else {
1802                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1803                        ip_addr_size(src_addr));
1804                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1805         }
1806
1807         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1808                 cma_exch(id_priv, CMA_DESTROYING);
1809                 cma_enable_remove(id_priv);
1810                 cma_deref_id(id_priv);
1811                 rdma_destroy_id(&id_priv->id);
1812                 return;
1813         }
1814 out:
1815         cma_enable_remove(id_priv);
1816         cma_deref_id(id_priv);
1817 }
1818
1819 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1820 {
1821         struct cma_work *work;
1822         struct sockaddr_in *src_in, *dst_in;
1823         union ib_gid gid;
1824         int ret;
1825
1826         work = malloc(sizeof *work, M_DEVBUF, M_NOWAIT);
1827         if (!work)
1828                 return (ENOMEM);
1829         bzero(work, sizeof *work);
1830
1831         if (!id_priv->cma_dev) {
1832                 ret = cma_bind_loopback(id_priv);
1833                 if (ret)
1834                         goto err;
1835         }
1836
1837         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1838         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1839
1840         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1841                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1842                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1843                 src_in->sin_family = dst_in->sin_family;
1844                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1845         }
1846
1847         work->id = id_priv;
1848         TASK_INIT(&work->task, 0, cma_work_handler, work);
1849         work->old_state = CMA_ADDR_QUERY;
1850         work->new_state = CMA_ADDR_RESOLVED;
1851         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1852         taskqueue_enqueue(cma_wq, &work->task);
1853         return 0;
1854 err:
1855         free(work, M_DEVBUF);
1856         return ret;
1857 }
1858
1859 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1860                          struct sockaddr *dst_addr)
1861 {
1862         if (src_addr && src_addr->sa_family)
1863                 return rdma_bind_addr(id, src_addr);
1864         else
1865                 return cma_bind_any(id, dst_addr->sa_family);
1866 }
1867
1868 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1869                       struct sockaddr *dst_addr, int timeout_ms)
1870 {
1871         struct rdma_id_private *id_priv;
1872         int ret;
1873
1874         id_priv = container_of(id, struct rdma_id_private, id);
1875         if (id_priv->state == CMA_IDLE) {
1876                 ret = cma_bind_addr(id, src_addr, dst_addr);
1877                 if (ret)
1878                         return ret;
1879         }
1880
1881         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1882                 return (EINVAL);
1883
1884         mtx_lock(&id_priv->lock);
1885         id_priv->refcount++;
1886         mtx_unlock(&id_priv->lock);
1887         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1888         if (cma_any_addr(dst_addr))
1889                 ret = cma_resolve_loopback(id_priv);
1890         else
1891                 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1892                                       dst_addr, &id->route.addr.dev_addr,
1893                                       timeout_ms, addr_handler, id_priv);
1894         if (ret)
1895                 goto err;
1896
1897         return 0;
1898 err:
1899         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1900         cma_deref_id(id_priv);
1901         return ret;
1902 }
1903
1904 static void cma_bind_port(struct rdma_bind_list *bind_list,
1905                           struct rdma_id_private *id_priv)
1906 {
1907         struct sockaddr_in *sin;
1908
1909         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1910         sin->sin_port = htons(bind_list->port);
1911         id_priv->bind_list = bind_list;
1912         TAILQ_INSERT_HEAD(&bind_list->owners, id_priv, node);
1913 }
1914
1915 static int cma_alloc_port(struct kvl *ps, struct rdma_id_private *id_priv,
1916                           unsigned short snum)
1917 {
1918         struct rdma_bind_list *bind_list;
1919         int port, ret;
1920
1921         bind_list = malloc(sizeof *bind_list, M_DEVBUF, M_NOWAIT);
1922         if (!bind_list)
1923                 return (ENOMEM);
1924         bzero(bind_list, sizeof *bind_list);
1925
1926         do {
1927                 ret = kvl_alloc_above(ps, bind_list, snum, &port);
1928         } while (ret == EAGAIN);
1929
1930         if (ret)
1931                 goto err1;
1932
1933         if (port != snum) {
1934                 ret = EADDRNOTAVAIL;
1935                 goto err2;
1936         }
1937
1938         bind_list->ps = ps;
1939         bind_list->port = (unsigned short) port;
1940         cma_bind_port(bind_list, id_priv);
1941         return 0;
1942 err2:
1943         kvl_delete(ps, port);
1944 err1:
1945         free(bind_list, M_DEVBUF);
1946         return ret;
1947 }
1948
1949 static int cma_alloc_any_port(struct kvl *ps, struct rdma_id_private *id_priv)
1950 {
1951         struct rdma_bind_list *bind_list;
1952         int port, ret;
1953
1954         bind_list = malloc(sizeof *bind_list, M_DEVBUF, M_NOWAIT);
1955         if (!bind_list)
1956                 return (ENOMEM);
1957         bzero(bind_list, sizeof *bind_list);
1958
1959 retry:
1960         do {
1961                 ret = kvl_alloc_above(ps, bind_list, next_port, &port);
1962         } while (ret == EAGAIN);
1963
1964         if (ret)
1965                 goto err1;
1966
1967         if (port > V_ipport_lastauto) {
1968                 if (next_port != V_ipport_firstauto) {
1969                         kvl_delete(ps, port);
1970                         next_port = V_ipport_firstauto;
1971                         goto retry;
1972                 }
1973                 ret = EADDRNOTAVAIL;
1974                 goto err2;
1975         }
1976
1977         if (port == V_ipport_lastauto)
1978                 next_port = V_ipport_firstauto;
1979         else
1980                 next_port = port + 1;
1981
1982         bind_list->ps = ps;
1983         bind_list->port = (unsigned short) port;
1984         cma_bind_port(bind_list, id_priv);
1985         return 0;
1986 err2:
1987         kvl_delete(ps, port);
1988 err1:
1989         free(bind_list, M_DEVBUF);
1990         return ret;
1991 }
1992
1993 static int cma_use_port(struct kvl *ps, struct rdma_id_private *id_priv)
1994 {
1995         struct rdma_id_private *cur_id;
1996         struct sockaddr_in *sin, *cur_sin;
1997         struct rdma_bind_list *bind_list;
1998         unsigned short snum;
1999
2000         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2001         snum = ntohs(sin->sin_port);
2002         if (snum <= V_ipport_reservedhigh && snum >= V_ipport_reservedlow &&
2003             priv_check(curthread, PRIV_NETINET_RESERVEDPORT))
2004                 return (EACCES);
2005
2006         bind_list = kvl_lookup(ps, snum);
2007         if (!bind_list)
2008                 return cma_alloc_port(ps, id_priv, snum);
2009
2010         /*
2011          * We don't support binding to any address if anyone is bound to
2012          * a specific address on the same port.
2013          */
2014         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
2015                 return (EADDRNOTAVAIL);
2016
2017         TAILQ_FOREACH(cur_id, &bind_list->owners, node) {
2018                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
2019                         return (EADDRNOTAVAIL);
2020
2021                 cur_sin = (struct sockaddr_in *)&cur_id->id.route.addr.src_addr;
2022                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2023                         return (EADDRINUSE);
2024         }
2025
2026         cma_bind_port(bind_list, id_priv);
2027         return 0;
2028 }
2029
2030 static int cma_get_tcp_port(struct rdma_id_private *id_priv)
2031 {
2032         int ret;
2033         struct socket *so;
2034
2035         ret = socreate(AF_INET, &so, SOCK_STREAM, IPPROTO_TCP,
2036                                 curthread->td_ucred, curthread);
2037         if (ret) {
2038                 printf("%s socreate err %d\n", __FUNCTION__, ret);
2039                 return ret;
2040         }
2041
2042         ret = sobind(so, (struct sockaddr *)&id_priv->id.route.addr.src_addr,
2043                         curthread);
2044         if (ret) {
2045                 soclose(so);
2046                 return ret;
2047         }
2048         id_priv->so = so;
2049         return 0;
2050 }
2051
2052 static int cma_get_port(struct rdma_id_private *id_priv)
2053 {
2054         struct kvl *ps;
2055         int ret;
2056
2057         switch (id_priv->id.ps) {
2058         case RDMA_PS_SDP:
2059                 ps = &sdp_ps;
2060                 break;
2061         case RDMA_PS_TCP:
2062                 ps = &tcp_ps;
2063                 ret = cma_get_tcp_port(id_priv); /* Synch with native stack */
2064                 if (ret)
2065                         return ret;
2066                 break;
2067         case RDMA_PS_UDP:
2068                 ps = &udp_ps;
2069                 break;
2070         case RDMA_PS_IPOIB:
2071                 ps = &ipoib_ps;
2072                 break;
2073         default:
2074                 return (EPROTONOSUPPORT);
2075         }
2076
2077         mtx_lock(&lock);
2078         if (cma_any_port(&id_priv->id.route.addr.src_addr))
2079                 ret = cma_alloc_any_port(ps, id_priv);
2080         else
2081                 ret = cma_use_port(ps, id_priv);
2082         mtx_unlock(&lock);
2083
2084         return ret;
2085 }
2086
2087 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2088 {
2089         struct rdma_id_private *id_priv;
2090         int ret;
2091
2092         if (addr->sa_family != AF_INET)
2093                 return (EAFNOSUPPORT);
2094
2095         id_priv = container_of(id, struct rdma_id_private, id);
2096         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2097                 return (EINVAL);
2098
2099         if (!cma_any_addr(addr)) {
2100                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2101                 if (ret)
2102                         goto err1;
2103
2104                 mtx_lock(&lock);
2105                 ret = cma_acquire_dev(id_priv);
2106                 mtx_unlock(&lock);
2107                 if (ret)
2108                         goto err1;
2109         }
2110
2111         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2112         ret = cma_get_port(id_priv);
2113         if (ret)
2114                 goto err2;
2115
2116         return 0;
2117 err2:
2118         if (!cma_any_addr(addr)) {
2119                 mtx_lock(&lock);
2120                 cma_detach_from_dev(id_priv);
2121                 mtx_unlock(&lock);
2122         }
2123 err1:
2124         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2125         return ret;
2126 }
2127
2128 #ifdef IB_SUPPORTED
2129 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2130                           struct rdma_route *route)
2131 {
2132         struct sockaddr_in *src4, *dst4;
2133         struct cma_hdr *cma_hdr;
2134         struct sdp_hh *sdp_hdr;
2135
2136         src4 = (struct sockaddr_in *) &route->addr.src_addr;
2137         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2138
2139         switch (ps) {
2140         case RDMA_PS_SDP:
2141                 sdp_hdr = hdr;
2142                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2143                         return (EINVAL);
2144                 sdp_set_ip_ver(sdp_hdr, 4);
2145                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2146                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2147                 sdp_hdr->port = src4->sin_port;
2148                 break;
2149         default:
2150                 cma_hdr = hdr;
2151                 cma_hdr->cma_version = CMA_VERSION;
2152                 cma_set_ip_ver(cma_hdr, 4);
2153                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2154                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2155                 cma_hdr->port = src4->sin_port;
2156                 break;
2157         }
2158         return 0;
2159 }
2160
2161 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2162                                 struct ib_cm_event *ib_event)
2163 {
2164         struct rdma_id_private *id_priv = cm_id->context;
2165         struct rdma_cm_event event;
2166         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2167         int ret = 0;
2168
2169         if (cma_disable_remove(id_priv, CMA_CONNECT))
2170                 return 0;
2171
2172         memset(&event, 0, sizeof event);
2173         switch (ib_event->event) {
2174         case IB_CM_SIDR_REQ_ERROR:
2175                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2176                 event.status = ETIMEDOUT;
2177                 break;
2178         case IB_CM_SIDR_REP_RECEIVED:
2179                 event.param.ud.private_data = ib_event->private_data;
2180                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2181                 if (rep->status != IB_SIDR_SUCCESS) {
2182                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2183                         event.status = ib_event->param.sidr_rep_rcvd.status;
2184                         break;
2185                 }
2186                 if (id_priv->qkey != rep->qkey) {
2187                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2188                         event.status = EINVAL;
2189                         break;
2190                 }
2191                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2192                                      id_priv->id.route.path_rec,
2193                                      &event.param.ud.ah_attr);
2194                 event.param.ud.qp_num = rep->qpn;
2195                 event.param.ud.qkey = rep->qkey;
2196                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2197                 event.status = 0;
2198                 break;
2199         default:
2200                 log(LOG_ERR, "RDMA CMA: unexpected IB CM event: %d",
2201                        ib_event->event);
2202                 goto out;
2203         }
2204
2205         ret = id_priv->id.event_handler(&id_priv->id, &event);
2206         if (ret) {
2207                 /* Destroy the CM ID by returning a non-zero value. */
2208                 id_priv->cm_id.ib = NULL;
2209                 cma_exch(id_priv, CMA_DESTROYING);
2210                 cma_enable_remove(id_priv);
2211                 rdma_destroy_id(&id_priv->id);
2212                 return ret;
2213         }
2214 out:
2215         cma_enable_remove(id_priv);
2216         return ret;
2217 }
2218
2219 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2220                               struct rdma_conn_param *conn_param)
2221 {
2222         struct ib_cm_sidr_req_param req;
2223         struct rdma_route *route;
2224         int ret;
2225
2226         req.private_data_len = sizeof(struct cma_hdr) +
2227                                conn_param->private_data_len;
2228         req.private_data = malloc(req.private_data_len, M_DEVBUF, M_NOWAIT);
2229         if (!req.private_data)
2230                 return (ENOMEM);
2231         bzero((void *)req.private_data, req.private_data_len);
2232
2233         if (conn_param->private_data && conn_param->private_data_len)
2234                 memcpy((caddr_t) req.private_data + sizeof(struct cma_hdr),
2235                        conn_param->private_data, conn_param->private_data_len);
2236
2237         route = &id_priv->id.route;
2238         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2239         if (ret)
2240                 goto out;
2241
2242         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2243                                             cma_sidr_rep_handler, id_priv);
2244         if (IS_ERR(id_priv->cm_id.ib)) {
2245                 ret = PTR_ERR(id_priv->cm_id.ib);
2246                 goto out;
2247         }
2248
2249         req.path = route->path_rec;
2250         req.service_id = cma_get_service_id(id_priv->id.ps,
2251                                             &route->addr.dst_addr);
2252         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2253         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2254
2255         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2256         if (ret) {
2257                 ib_destroy_cm_id(id_priv->cm_id.ib);
2258                 id_priv->cm_id.ib = NULL;
2259         }
2260 out:
2261         free(req.private_data, M_DEVBUF);
2262         return ret;
2263 }
2264
2265 static int cma_connect_ib(struct rdma_id_private *id_priv,
2266                           struct rdma_conn_param *conn_param)
2267 {
2268         struct ib_cm_req_param req;
2269         struct rdma_route *route;
2270         void *private_data;
2271         int offset, ret;
2272
2273         memset(&req, 0, sizeof req);
2274         offset = cma_user_data_offset(id_priv->id.ps);
2275         req.private_data_len = offset + conn_param->private_data_len;
2276         private_data = malloc(req.private_data_len, M_DEVBUF, M_NOWAIT);
2277         if (!private_data)
2278                 return (ENOMEM);
2279         bzero(private_data, req.private_data_len);
2280
2281         if (conn_param->private_data && conn_param->private_data_len)
2282                 memcpy(private_data + offset, conn_param->private_data,
2283                        conn_param->private_data_len);
2284
2285         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2286                                             id_priv);
2287         if (IS_ERR(id_priv->cm_id.ib)) {
2288                 ret = PTR_ERR(id_priv->cm_id.ib);
2289                 goto out;
2290         }
2291
2292         route = &id_priv->id.route;
2293         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2294         if (ret)
2295                 goto out;
2296         req.private_data = private_data;
2297
2298         req.primary_path = &route->path_rec[0];
2299         if (route->num_paths == 2)
2300                 req.alternate_path = &route->path_rec[1];
2301
2302         req.service_id = cma_get_service_id(id_priv->id.ps,
2303                                             &route->addr.dst_addr);
2304         req.qp_num = id_priv->qp_num;
2305         req.qp_type = IB_QPT_RC;
2306         req.starting_psn = id_priv->seq_num;
2307         req.responder_resources = conn_param->responder_resources;
2308         req.initiator_depth = conn_param->initiator_depth;
2309         req.flow_control = conn_param->flow_control;
2310         req.retry_count = conn_param->retry_count;
2311         req.rnr_retry_count = conn_param->rnr_retry_count;
2312         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2313         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2314         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2315         req.srq = id_priv->srq ? 1 : 0;
2316
2317         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2318 out:
2319         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2320                 ib_destroy_cm_id(id_priv->cm_id.ib);
2321                 id_priv->cm_id.ib = NULL;
2322         }
2323
2324         free(private_data, M_DEVBUF);
2325         return ret;
2326 }
2327 #endif
2328
2329 static int cma_connect_iw(struct rdma_id_private *id_priv,
2330                           struct rdma_conn_param *conn_param)
2331 {
2332         struct iw_cm_id *cm_id;
2333         struct sockaddr_in* sin;
2334         int ret;
2335         struct iw_cm_conn_param iw_param;
2336
2337         cm_id = iw_create_cm_id(id_priv->id.device, id_priv->so,
2338                                  cma_iw_handler, id_priv);
2339         if (IS_ERR(cm_id)) {
2340                 ret = PTR_ERR(cm_id);
2341                 goto out;
2342         }
2343
2344         id_priv->cm_id.iw = cm_id;
2345
2346         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2347         cm_id->local_addr = *sin;
2348
2349         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2350         cm_id->remote_addr = *sin;
2351
2352         ret = cma_modify_qp_rtr(&id_priv->id);
2353         if (ret)
2354                 goto out;
2355
2356         iw_param.ord = conn_param->initiator_depth;
2357         iw_param.ird = conn_param->responder_resources;
2358         iw_param.private_data = conn_param->private_data;
2359         iw_param.private_data_len = conn_param->private_data_len;
2360         if (id_priv->id.qp)
2361                 iw_param.qpn = id_priv->qp_num;
2362         else
2363                 iw_param.qpn = conn_param->qp_num;
2364         ret = iw_cm_connect(cm_id, &iw_param);
2365 out:
2366         if (ret && !IS_ERR(cm_id)) {
2367                 iw_destroy_cm_id(cm_id);
2368                 id_priv->cm_id.iw = NULL;
2369         }
2370         return ret;
2371 }
2372
2373 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2374 {
2375         struct rdma_id_private *id_priv;
2376         int ret;
2377
2378         id_priv = container_of(id, struct rdma_id_private, id);
2379         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2380                 return (EINVAL);
2381
2382         if (!id->qp) {
2383                 id_priv->qp_num = conn_param->qp_num;
2384                 id_priv->srq = conn_param->srq;
2385         }
2386
2387 #ifdef IB_SUPPORTED
2388         switch (rdma_node_get_transport(id->device->node_type)) {
2389         case RDMA_TRANSPORT_IB:
2390                 if (cma_is_ud_ps(id->ps))
2391                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2392                 else
2393                         ret = cma_connect_ib(id_priv, conn_param);
2394                 break;
2395         case RDMA_TRANSPORT_IWARP:
2396 #endif
2397                 ret = cma_connect_iw(id_priv, conn_param);
2398 #ifdef IB_SUPPORTED
2399                 break;
2400         default:
2401                 ret = ENOSYS;
2402                 break;
2403         }
2404 #endif
2405         if (ret)
2406                 goto err;
2407
2408         return 0;
2409 err:
2410         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2411         return ret;
2412 }
2413
2414 #ifdef IB_SUPPORTED
2415 static int cma_accept_ib(struct rdma_id_private *id_priv,
2416                          struct rdma_conn_param *conn_param)
2417 {
2418         struct ib_cm_rep_param rep;
2419         struct ib_qp_attr qp_attr;
2420         int qp_attr_mask, ret;
2421
2422         if (id_priv->id.qp) {
2423                 ret = cma_modify_qp_rtr(&id_priv->id);
2424                 if (ret)
2425                         goto out;
2426
2427                 qp_attr.qp_state = IB_QPS_RTS;
2428                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr,
2429                                          &qp_attr_mask);
2430                 if (ret)
2431                         goto out;
2432
2433                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
2434                 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
2435                 if (ret)
2436                         goto out;
2437         }
2438
2439         memset(&rep, 0, sizeof rep);
2440         rep.qp_num = id_priv->qp_num;
2441         rep.starting_psn = id_priv->seq_num;
2442         rep.private_data = conn_param->private_data;
2443         rep.private_data_len = conn_param->private_data_len;
2444         rep.responder_resources = conn_param->responder_resources;
2445         rep.initiator_depth = conn_param->initiator_depth;
2446         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
2447         rep.failover_accepted = 0;
2448         rep.flow_control = conn_param->flow_control;
2449         rep.rnr_retry_count = conn_param->rnr_retry_count;
2450         rep.srq = id_priv->srq ? 1 : 0;
2451
2452         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2453 out:
2454         return ret;
2455 }
2456 #endif
2457
2458 static int cma_accept_iw(struct rdma_id_private *id_priv,
2459                   struct rdma_conn_param *conn_param)
2460 {
2461         struct iw_cm_conn_param iw_param;
2462         int ret;
2463
2464         ret = cma_modify_qp_rtr(&id_priv->id);
2465         if (ret)
2466                 return ret;
2467
2468         iw_param.ord = conn_param->initiator_depth;
2469         iw_param.ird = conn_param->responder_resources;
2470         iw_param.private_data = conn_param->private_data;
2471         iw_param.private_data_len = conn_param->private_data_len;
2472         if (id_priv->id.qp) {
2473                 iw_param.qpn = id_priv->qp_num;
2474         } else
2475                 iw_param.qpn = conn_param->qp_num;
2476
2477         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2478 }
2479
2480 #ifdef IB_SUPPORTED
2481 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2482                              enum ib_cm_sidr_status status,
2483                              const void *private_data, int private_data_len)
2484 {
2485         struct ib_cm_sidr_rep_param rep;
2486
2487         memset(&rep, 0, sizeof rep);
2488         rep.status = status;
2489         if (status == IB_SIDR_SUCCESS) {
2490                 rep.qp_num = id_priv->qp_num;
2491                 rep.qkey = id_priv->qkey;
2492         }
2493         rep.private_data = private_data;
2494         rep.private_data_len = private_data_len;
2495
2496         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2497 }
2498 #endif
2499
2500 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2501 {
2502         struct rdma_id_private *id_priv;
2503         int ret;
2504
2505         id_priv = container_of(id, struct rdma_id_private, id);
2506         if (!cma_comp(id_priv, CMA_CONNECT))
2507                 return (EINVAL);
2508
2509         if (!id->qp && conn_param) {
2510                 id_priv->qp_num = conn_param->qp_num;
2511                 id_priv->srq = conn_param->srq;
2512         }
2513
2514 #ifdef IB_SUPPORTED
2515         switch (rdma_node_get_transport(id->device->node_type)) {
2516         case RDMA_TRANSPORT_IB:
2517                 if (cma_is_ud_ps(id->ps))
2518                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2519                                                 conn_param->private_data,
2520                                                 conn_param->private_data_len);
2521                 else if (conn_param)
2522                         ret = cma_accept_ib(id_priv, conn_param);
2523                 else
2524                         ret = cma_rep_recv(id_priv);
2525                 break;
2526         case RDMA_TRANSPORT_IWARP:
2527 #endif
2528                 ret = cma_accept_iw(id_priv, conn_param);
2529 #ifdef IB_SUPPORTED
2530                 break;
2531         default:
2532                 ret = ENOSYS;
2533                 break;
2534         }
2535 #endif
2536
2537         if (ret)
2538                 goto reject;
2539
2540         return 0;
2541 reject:
2542         cma_modify_qp_err(id);
2543         rdma_reject(id, NULL, 0);
2544         return ret;
2545 }
2546
2547 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2548 {
2549         struct rdma_id_private *id_priv;
2550         int ret;
2551
2552         id_priv = container_of(id, struct rdma_id_private, id);
2553         if (!cma_has_cm_dev(id_priv))
2554                 return (EINVAL);
2555 #ifdef IB_SUPPORTED
2556         switch (id->device->node_type) {
2557         case RDMA_NODE_IB_CA:
2558                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2559                 break;
2560         default:
2561 #endif
2562                 ret = 0;
2563 #ifdef IB_SUPPORTED
2564                 break;
2565         }
2566 #endif
2567         return ret;
2568 }
2569
2570 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2571                 u8 private_data_len)
2572 {
2573         struct rdma_id_private *id_priv;
2574         int ret;
2575
2576         id_priv = container_of(id, struct rdma_id_private, id);
2577         if (!cma_has_cm_dev(id_priv))
2578                 return (EINVAL);
2579
2580 #ifdef IB_SUPPORTED
2581         switch (rdma_node_get_transport(id->device->node_type)) {
2582         case RDMA_TRANSPORT_IB:
2583                 if (cma_is_ud_ps(id->ps))
2584                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2585                                                 private_data, private_data_len);
2586                 else
2587                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2588                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2589                                              0, private_data, private_data_len);
2590                 break;
2591         case RDMA_TRANSPORT_IWARP:
2592 #endif
2593                 ret = iw_cm_reject(id_priv->cm_id.iw,
2594                                    private_data, private_data_len);
2595 #ifdef IB_SUPPORTED
2596                 break;
2597         default:
2598                 ret = ENOSYS;
2599                 break;
2600         }
2601 #endif
2602         return ret;
2603 }
2604
2605 int rdma_disconnect(struct rdma_cm_id *id)
2606 {
2607         struct rdma_id_private *id_priv;
2608         int ret;
2609
2610         id_priv = container_of(id, struct rdma_id_private, id);
2611         if (!cma_has_cm_dev(id_priv))
2612                 return (EINVAL);
2613
2614 #ifdef IB_SUPPORTED
2615         switch (rdma_node_get_transport(id->device->node_type)) {
2616         case RDMA_TRANSPORT_IB:
2617                 ret = cma_modify_qp_err(id);
2618                 if (ret)
2619                         goto out;
2620                 /* Initiate or respond to a disconnect. */
2621                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2622                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2623                 break;
2624         case RDMA_TRANSPORT_IWARP:
2625 #endif
2626                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2627 #ifdef IB_SUPPORTED
2628                 break;
2629         default:
2630                 ret = EINVAL;
2631                 break;
2632         }
2633 out:
2634 #endif
2635         return ret;
2636 }
2637
2638 #ifdef IB_SUPPORTED
2639 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2640 {
2641         struct rdma_id_private *id_priv;
2642         struct cma_multicast *mc = multicast->context;
2643         struct rdma_cm_event event;
2644         int ret;
2645
2646         id_priv = mc->id_priv;
2647         if (cma_disable_remove(id_priv, CMA_ADDR_BOUND) &&
2648             cma_disable_remove(id_priv, CMA_ADDR_RESOLVED))
2649                 return 0;
2650
2651         if (!status && id_priv->id.qp)
2652                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2653                                          multicast->rec.mlid);
2654
2655         memset(&event, 0, sizeof event);
2656         event.status = status;
2657         event.param.ud.private_data = mc->context;
2658         if (!status) {
2659                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2660                 ib_init_ah_from_mcmember(id_priv->id.device,
2661                                          id_priv->id.port_num, &multicast->rec,
2662                                          &event.param.ud.ah_attr);
2663                 event.param.ud.qp_num = 0xFFFFFF;
2664                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2665         } else
2666                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2667
2668         ret = id_priv->id.event_handler(&id_priv->id, &event);
2669         if (ret) {
2670                 cma_exch(id_priv, CMA_DESTROYING);
2671                 cma_enable_remove(id_priv);
2672                 rdma_destroy_id(&id_priv->id);
2673                 return 0;
2674         }
2675
2676         cma_enable_remove(id_priv);
2677         return 0;
2678 }
2679
2680 static void cma_set_mgid(struct rdma_id_private *id_priv,
2681                          struct sockaddr *addr, union ib_gid *mgid)
2682 {
2683         unsigned char mc_map[MAX_ADDR_LEN];
2684         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2685         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2686         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2687
2688         if (cma_any_addr(addr)) {
2689                 memset(mgid, 0, sizeof *mgid);
2690         } else if ((addr->sa_family == AF_INET6) &&
2691                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2692                                                                  0xFF10A01B)) {
2693                 /* IPv6 address is an SA assigned MGID. */
2694                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2695         } else {
2696                 ip_ib_mc_map(sin->sin_addr.s_addr, mc_map);
2697                 if (id_priv->id.ps == RDMA_PS_UDP)
2698                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2699                 mc_map[8] = ib_addr_get_pkey(dev_addr) >> 8;
2700                 mc_map[9] = (unsigned char) ib_addr_get_pkey(dev_addr);
2701                 *mgid = *(union ib_gid *) (mc_map + 4);
2702         }
2703 }
2704
2705 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2706                                  struct cma_multicast *mc)
2707 {
2708         struct ib_sa_mcmember_rec rec;
2709         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2710         ib_sa_comp_mask comp_mask;
2711         int ret;
2712
2713         ib_addr_get_mgid(dev_addr, &rec.mgid);
2714         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2715                                      &rec.mgid, &rec);
2716         if (ret)
2717                 return ret;
2718
2719         cma_set_mgid(id_priv, &mc->addr, &rec.mgid);
2720         if (id_priv->id.ps == RDMA_PS_UDP)
2721                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2722         ib_addr_get_sgid(dev_addr, &rec.port_gid);
2723         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2724         rec.join_state = 1;
2725
2726         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2727                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2728                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2729                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2730                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2731
2732         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2733                                                 id_priv->id.port_num, &rec,
2734                                                 comp_mask, M_NOWAIT,
2735                                                 cma_ib_mc_handler, mc);
2736         if (IS_ERR(mc->multicast.ib))
2737                 return PTR_ERR(mc->multicast.ib);
2738
2739         return 0;
2740 }
2741
2742 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2743                         void *context)
2744 {
2745         struct rdma_id_private *id_priv;
2746         struct cma_multicast *mc;
2747         int ret;
2748
2749         id_priv = container_of(id, struct rdma_id_private, id);
2750         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2751             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2752                 return (EINVAL);
2753
2754         mc = malloc(sizeof *mc, M_DEVBUF, M_NOWAIT);
2755         if (!mc)
2756                 return (ENOMEM);
2757
2758         memcpy(&mc->addr, addr, ip_addr_size(addr));
2759         mc->context = context;
2760         mc->id_priv = id_priv;
2761
2762         mtx_lock(&id_priv->lock);
2763         LIST_INSERT_HEAD(&id_priv->mc_list, mc, list);
2764         mtx_unlock(&id_priv->lock);
2765
2766         switch (rdma_node_get_transport(id->device->node_type)) {
2767         case RDMA_TRANSPORT_IB:
2768                 ret = cma_join_ib_multicast(id_priv, mc);
2769                 break;
2770         default:
2771                 ret = ENOSYS;
2772                 break;
2773         }
2774
2775         if (ret) {
2776                 mtx_lock(&id_priv->lock);
2777                 list_del(&mc->list);
2778                 mtx_unlock(&id_priv->lock);
2779                 free(mc, M_DEVBUF);
2780         }
2781         return ret;
2782 }
2783
2784 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2785 {
2786         struct rdma_id_private *id_priv;
2787         struct cma_multicast *mc;
2788
2789         id_priv = container_of(id, struct rdma_id_private, id);
2790         mtx_lock(&id_priv->lock);
2791         LIST_FOREACH(mc, &id_priv->mc_list, list) {
2792                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2793                         list_del(&mc->list);
2794                         mtx_unlock(&id_priv->lock);
2795
2796                         if (id->qp)
2797                                 ib_detach_mcast(id->qp,
2798                                                 &mc->multicast.ib->rec.mgid,
2799                                                 mc->multicast.ib->rec.mlid);
2800                         ib_sa_free_multicast(mc->multicast.ib, M_DEVBUF);
2801                         free(mc, M_DEVBUF);
2802                         return;
2803                 }
2804         }
2805         mtx_unlock(&id_priv->lock);
2806 }
2807 #endif
2808
2809 static void cma_add_one(struct ib_device *device)
2810 {
2811         struct cma_device *cma_dev;
2812         struct rdma_id_private *id_priv;
2813
2814         cma_dev = malloc(sizeof *cma_dev, M_DEVBUF, M_NOWAIT|M_ZERO);
2815         if (!cma_dev)
2816                 return;
2817
2818         cma_dev->device = device;
2819
2820         cv_init(&cma_dev->comp, "cma_device");
2821         mtx_init(&cma_dev->lock, "cma_device", NULL, MTX_DUPOK|MTX_DEF);
2822         cma_dev->refcount = 1;
2823         LIST_INIT(&cma_dev->id_list);
2824         ib_set_client_data(device, &cma_client, cma_dev);
2825
2826         mtx_lock(&lock);
2827         TAILQ_INSERT_TAIL(&dev_list, cma_dev, list);
2828         LIST_FOREACH(id_priv, &listen_any_list, list)
2829                 cma_listen_on_dev(id_priv, cma_dev);
2830         mtx_unlock(&lock);
2831 }
2832
2833 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2834 {
2835         struct rdma_cm_event event;
2836         enum cma_state state;
2837
2838         /* Record that we want to remove the device */
2839         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2840         if (state == CMA_DESTROYING)
2841                 return 0;
2842
2843         cma_cancel_operation(id_priv, state);
2844         mtx_lock(&id_priv->lock);
2845         PANIC_IF(id_priv->dev_remove < 0);
2846         if (id_priv->dev_remove)
2847                 cv_wait(&id_priv->wait_remove, &id_priv->lock);
2848         mtx_unlock(&id_priv->lock);
2849
2850         /* Check for destruction from another callback. */
2851         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2852                 return 0;
2853
2854         memset(&event, 0, sizeof event);
2855         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2856         return id_priv->id.event_handler(&id_priv->id, &event);
2857 }
2858
2859 static void cma_process_remove(struct cma_device *cma_dev)
2860 {
2861         struct rdma_id_private *id_priv;
2862         int ret;
2863
2864         mtx_lock(&lock);
2865         while (!LIST_EMPTY(&cma_dev->id_list)) {
2866                 id_priv = LIST_FIRST(&cma_dev->id_list);
2867
2868                 if (cma_internal_listen(id_priv)) {
2869                         cma_destroy_listen(id_priv);
2870                         continue;
2871                 }
2872
2873                 LIST_REMOVE(id_priv, list);
2874                 mtx_lock(&id_priv->lock);
2875                 id_priv->refcount++;
2876                 mtx_unlock(&id_priv->lock);
2877                 mtx_unlock(&lock);
2878
2879                 ret = cma_remove_id_dev(id_priv);
2880                 cma_deref_id(id_priv);
2881                 if (ret)
2882                         rdma_destroy_id(&id_priv->id);
2883
2884                 mtx_lock(&lock);
2885         }
2886         mtx_unlock(&lock);
2887
2888         cma_deref_dev(cma_dev);
2889         mtx_lock(&cma_dev->lock);
2890         PANIC_IF(cma_dev->refcount < 0);
2891         if (cma_dev->refcount)
2892                 cv_wait(&cma_dev->comp, &cma_dev->lock);
2893         mtx_unlock(&cma_dev->lock);
2894 }
2895
2896 static void cma_remove_one(struct ib_device *device)
2897 {
2898         struct cma_device *cma_dev;
2899
2900         cma_dev = ib_get_client_data(device, &cma_client);
2901         if (!cma_dev)
2902                 return;
2903
2904         mtx_lock(&lock);
2905         TAILQ_REMOVE(&dev_list, cma_dev, list);
2906         mtx_unlock(&lock);
2907
2908         cma_process_remove(cma_dev);
2909         free(cma_dev, M_DEVBUF);
2910 }
2911
2912 static int cma_init(void)
2913 {
2914         int ret;
2915
2916         LIST_INIT(&listen_any_list);
2917         TAILQ_INIT(&dev_list);
2918         mtx_init(&lock, "cma_device list", NULL, MTX_DEF);
2919
2920         arc4rand(&next_port, sizeof next_port, 0);
2921         next_port = ((unsigned int) next_port %
2922                     (V_ipport_lastauto - V_ipport_firstauto)) +
2923                     V_ipport_firstauto;
2924         cma_wq = taskqueue_create("rdma_cm", M_NOWAIT, taskqueue_thread_enqueue,
2925                 &cma_wq);
2926
2927         if (!cma_wq)
2928                 return (ENOMEM);
2929
2930         taskqueue_start_threads(&cma_wq, 1, PI_NET, "cma_wq thread");
2931 #ifdef IB_SUPPORTED
2932         ib_sa_register_client(&sa_client);
2933 #endif
2934         rdma_addr_register_client(&addr_client);
2935
2936         ret = ib_register_client(&cma_client);
2937         if (ret)
2938                 goto err;
2939         return 0;
2940
2941 err:
2942         rdma_addr_unregister_client(&addr_client);
2943 #ifdef IB_SUPPORTED
2944         ib_sa_unregister_client(&sa_client);
2945 #endif
2946         taskqueue_free(cma_wq);
2947         return ret;
2948 }
2949
2950 static void cma_cleanup(void)
2951 {
2952         ib_unregister_client(&cma_client);
2953         rdma_addr_unregister_client(&addr_client);
2954 #ifdef IB_SUPPORTED
2955         ib_sa_unregister_client(&sa_client);
2956 #endif
2957         taskqueue_free(cma_wq);
2958         kvl_free(&sdp_ps);
2959         kvl_free(&tcp_ps);
2960         kvl_free(&udp_ps);
2961         kvl_free(&ipoib_ps);
2962 }
2963
2964 static int
2965 cma_load(module_t mod, int cmd, void *arg)
2966 {
2967         int err = 0;
2968
2969         switch (cmd) {
2970         case MOD_LOAD:
2971                 printf("Loading rdma_cma.\n");
2972                 cma_init();
2973                 break;
2974         case MOD_QUIESCE:
2975                 break;
2976         case MOD_UNLOAD:
2977                 printf("Unloading rdma_cma.\n");
2978                 cma_cleanup();
2979                 break;
2980         case MOD_SHUTDOWN:
2981                 break;
2982         default:
2983                 err = EOPNOTSUPP;
2984                 break;
2985         }
2986
2987         return (err);
2988 }
2989
2990 static moduledata_t mod_data = {
2991         "rdma_cma",
2992         cma_load,
2993         0
2994 };
2995
2996 MODULE_VERSION(rdma_cma, 1);
2997 MODULE_DEPEND(rdma_cma, rdma_core, 1, 1, 1);
2998 MODULE_DEPEND(rdma_cma, rdma_addr, 1, 1, 1);
2999 MODULE_DEPEND(rdma_cma, rdma_iwcm, 1, 1, 1);
3000 DECLARE_MODULE(rdma_cma, mod_data, SI_SUB_EXEC, SI_ORDER_ANY);