SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/infiniband/core/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 <linux/completion.h>
  33 #include <linux/in.h>
  34 #include <linux/in6.h>
  35 #include <linux/mutex.h>
  36 #include <linux/random.h>
  37 #include <linux/idr.h>
  38 #include <linux/inetdevice.h>
  39
  40 #include <net/tcp.h>
  41
  42 #include <rdma/rdma_cm.h>
  43 #include <rdma/rdma_cm_ib.h>
  44 #include <rdma/ib_cache.h>
  45 #include <rdma/ib_cm.h>
  46 #include <rdma/ib_sa.h>
  47 #include <rdma/iw_cm.h>
  48
  49 MODULE_AUTHOR("Sean Hefty");
  50 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  51 MODULE_LICENSE("Dual BSD/GPL");
  52
  53 #define CMA_CM_RESPONSE_TIMEOUT 20
  54 #define CMA_MAX_CM_RETRIES 15
  55
  56 static void cma_add_one(struct ib_device *device);
  57 static void cma_remove_one(struct ib_device *device);
  58
  59 static struct ib_client cma_client = {
  60         .name   = "cma",
  61         .add    = cma_add_one,
  62         .remove = cma_remove_one
  63 };
  64
  65 static struct ib_sa_client sa_client;
  66 static struct rdma_addr_client addr_client;
  67 static LIST_HEAD(dev_list);
  68 static LIST_HEAD(listen_any_list);
  69 static DEFINE_MUTEX(lock);
  70 static struct workqueue_struct *cma_wq;
  71 static DEFINE_IDR(sdp_ps);
  72 static DEFINE_IDR(tcp_ps);
  73
  74 struct cma_device {
  75         struct list_head        list;
  76         struct ib_device        *device;
  77         __be64                  node_guid;
  78         struct completion       comp;
  79         atomic_t                refcount;
  80         struct list_head        id_list;
  81 };
  82
  83 enum cma_state {
  84         CMA_IDLE,
  85         CMA_ADDR_QUERY,
  86         CMA_ADDR_RESOLVED,
  87         CMA_ROUTE_QUERY,
  88         CMA_ROUTE_RESOLVED,
  89         CMA_CONNECT,
  90         CMA_DISCONNECT,
  91         CMA_ADDR_BOUND,
  92         CMA_LISTEN,
  93         CMA_DEVICE_REMOVAL,
  94         CMA_DESTROYING
  95 };
  96
  97 struct rdma_bind_list {
  98         struct idr              *ps;
  99         struct hlist_head       owners;
 100         unsigned short          port;
 101 };
 102
 103 /*
 104  * Device removal can occur at anytime, so we need extra handling to
 105  * serialize notifying the user of device removal with other callbacks.
 106  * We do this by disabling removal notification while a callback is in process,
 107  * and reporting it after the callback completes.
 108  */
 109 struct rdma_id_private {
 110         struct rdma_cm_id       id;
 111
 112         struct rdma_bind_list   *bind_list;
 113         struct hlist_node       node;
 114         struct list_head        list;
 115         struct list_head        listen_list;
 116         struct cma_device       *cma_dev;
 117
 118         enum cma_state          state;
 119         spinlock_t              lock;
 120         struct completion       comp;
 121         atomic_t                refcount;
 122         wait_queue_head_t       wait_remove;
 123         atomic_t                dev_remove;
 124
 125         int                     backlog;
 126         int                     timeout_ms;
 127         struct ib_sa_query      *query;
 128         int                     query_id;
 129         union {
 130                 struct ib_cm_id *ib;
 131                 struct iw_cm_id *iw;
 132         } cm_id;
 133
 134         u32                     seq_num;
 135         u32                     qp_num;
 136         u8                      srq;
 137 };
 138
 139 struct cma_work {
 140         struct work_struct      work;
 141         struct rdma_id_private  *id;
 142         enum cma_state          old_state;
 143         enum cma_state          new_state;
 144         struct rdma_cm_event    event;
 145 };
 146
 147 union cma_ip_addr {
 148         struct in6_addr ip6;
 149         struct {
 150                 __u32 pad[3];
 151                 __u32 addr;
 152         } ip4;
 153 };
 154
 155 struct cma_hdr {
 156         u8 cma_version;
 157         u8 ip_version;  /* IP version: 7:4 */
 158         __u16 port;
 159         union cma_ip_addr src_addr;
 160         union cma_ip_addr dst_addr;
 161 };
 162
 163 struct sdp_hh {
 164         u8 bsdh[16];
 165         u8 sdp_version; /* Major version: 7:4 */
 166         u8 ip_version;  /* IP version: 7:4 */
 167         u8 sdp_specific1[10];
 168         __u16 port;
 169         __u16 sdp_specific2;
 170         union cma_ip_addr src_addr;
 171         union cma_ip_addr dst_addr;
 172 };
 173
 174 struct sdp_hah {
 175         u8 bsdh[16];
 176         u8 sdp_version;
 177 };
 178
 179 #define CMA_VERSION 0x00
 180 #define SDP_MAJ_VERSION 0x2
 181
 182 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 183 {
 184         unsigned long flags;
 185         int ret;
 186
 187         spin_lock_irqsave(&id_priv->lock, flags);
 188         ret = (id_priv->state == comp);
 189         spin_unlock_irqrestore(&id_priv->lock, flags);
 190         return ret;
 191 }
 192
 193 static int cma_comp_exch(struct rdma_id_private *id_priv,
 194                          enum cma_state comp, enum cma_state exch)
 195 {
 196         unsigned long flags;
 197         int ret;
 198
 199         spin_lock_irqsave(&id_priv->lock, flags);
 200         if ((ret = (id_priv->state == comp)))
 201                 id_priv->state = exch;
 202         spin_unlock_irqrestore(&id_priv->lock, flags);
 203         return ret;
 204 }
 205
 206 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 207                                enum cma_state exch)
 208 {
 209         unsigned long flags;
 210         enum cma_state old;
 211
 212         spin_lock_irqsave(&id_priv->lock, flags);
 213         old = id_priv->state;
 214         id_priv->state = exch;
 215         spin_unlock_irqrestore(&id_priv->lock, flags);
 216         return old;
 217 }
 218
 219 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 220 {
 221         return hdr->ip_version >> 4;
 222 }
 223
 224 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 225 {
 226         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 227 }
 228
 229 static inline u8 sdp_get_majv(u8 sdp_version)
 230 {
 231         return sdp_version >> 4;
 232 }
 233
 234 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 235 {
 236         return hh->ip_version >> 4;
 237 }
 238
 239 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 240 {
 241         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 242 }
 243
 244 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 245                               struct cma_device *cma_dev)
 246 {
 247         atomic_inc(&cma_dev->refcount);
 248         id_priv->cma_dev = cma_dev;
 249         id_priv->id.device = cma_dev->device;
 250         list_add_tail(&id_priv->list, &cma_dev->id_list);
 251 }
 252
 253 static inline void cma_deref_dev(struct cma_device *cma_dev)
 254 {
 255         if (atomic_dec_and_test(&cma_dev->refcount))
 256                 complete(&cma_dev->comp);
 257 }
 258
 259 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 260 {
 261         list_del(&id_priv->list);
 262         cma_deref_dev(id_priv->cma_dev);
 263         id_priv->cma_dev = NULL;
 264 }
 265
 266 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 267 {
 268         enum rdma_node_type dev_type = id_priv->id.route.addr.dev_addr.dev_type;
 269         struct cma_device *cma_dev;
 270         union ib_gid gid;
 271         int ret = -ENODEV;
 272
 273         switch (rdma_node_get_transport(dev_type)) {
 274         case RDMA_TRANSPORT_IB:
 275                 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 276                 break;
 277         case RDMA_TRANSPORT_IWARP:
 278                 iw_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
 279                 break;
 280         default:
 281                 return -ENODEV;
 282         }
 283
 284         list_for_each_entry(cma_dev, &dev_list, list) {
 285                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 286                                          &id_priv->id.port_num, NULL);
 287                 if (!ret) {
 288                         cma_attach_to_dev(id_priv, cma_dev);
 289                         break;
 290                 }
 291         }
 292         return ret;
 293 }
 294
 295 static void cma_deref_id(struct rdma_id_private *id_priv)
 296 {
 297         if (atomic_dec_and_test(&id_priv->refcount))
 298                 complete(&id_priv->comp);
 299 }
 300
 301 static void cma_release_remove(struct rdma_id_private *id_priv)
 302 {
 303         if (atomic_dec_and_test(&id_priv->dev_remove))
 304                 wake_up(&id_priv->wait_remove);
 305 }
 306
 307 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 308                                   void *context, enum rdma_port_space ps)
 309 {
 310         struct rdma_id_private *id_priv;
 311
 312         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 313         if (!id_priv)
 314                 return ERR_PTR(-ENOMEM);
 315
 316         id_priv->state = CMA_IDLE;
 317         id_priv->id.context = context;
 318         id_priv->id.event_handler = event_handler;
 319         id_priv->id.ps = ps;
 320         spin_lock_init(&id_priv->lock);
 321         init_completion(&id_priv->comp);
 322         atomic_set(&id_priv->refcount, 1);
 323         init_waitqueue_head(&id_priv->wait_remove);
 324         atomic_set(&id_priv->dev_remove, 0);
 325         INIT_LIST_HEAD(&id_priv->listen_list);
 326         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 327
 328         return &id_priv->id;
 329 }
 330 EXPORT_SYMBOL(rdma_create_id);
 331
 332 static int cma_init_ib_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 333 {
 334         struct ib_qp_attr qp_attr;
 335         struct rdma_dev_addr *dev_addr;
 336         int ret;
 337
 338         dev_addr = &id_priv->id.route.addr.dev_addr;
 339         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 340                                   ib_addr_get_pkey(dev_addr),
 341                                   &qp_attr.pkey_index);
 342         if (ret)
 343                 return ret;
 344
 345         qp_attr.qp_state = IB_QPS_INIT;
 346         qp_attr.qp_access_flags = 0;
 347         qp_attr.port_num = id_priv->id.port_num;
 348         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS |
 349                                           IB_QP_PKEY_INDEX | IB_QP_PORT);
 350 }
 351
 352 static int cma_init_iw_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 353 {
 354         struct ib_qp_attr qp_attr;
 355
 356         qp_attr.qp_state = IB_QPS_INIT;
 357         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 358
 359         return ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_ACCESS_FLAGS);
 360 }
 361
 362 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 363                    struct ib_qp_init_attr *qp_init_attr)
 364 {
 365         struct rdma_id_private *id_priv;
 366         struct ib_qp *qp;
 367         int ret;
 368
 369         id_priv = container_of(id, struct rdma_id_private, id);
 370         if (id->device != pd->device)
 371                 return -EINVAL;
 372
 373         qp = ib_create_qp(pd, qp_init_attr);
 374         if (IS_ERR(qp))
 375                 return PTR_ERR(qp);
 376
 377         switch (rdma_node_get_transport(id->device->node_type)) {
 378         case RDMA_TRANSPORT_IB:
 379                 ret = cma_init_ib_qp(id_priv, qp);
 380                 break;
 381         case RDMA_TRANSPORT_IWARP:
 382                 ret = cma_init_iw_qp(id_priv, qp);
 383                 break;
 384         default:
 385                 ret = -ENOSYS;
 386                 break;
 387         }
 388
 389         if (ret)
 390                 goto err;
 391
 392         id->qp = qp;
 393         id_priv->qp_num = qp->qp_num;
 394         id_priv->srq = (qp->srq != NULL);
 395         return 0;
 396 err:
 397         ib_destroy_qp(qp);
 398         return ret;
 399 }
 400 EXPORT_SYMBOL(rdma_create_qp);
 401
 402 void rdma_destroy_qp(struct rdma_cm_id *id)
 403 {
 404         ib_destroy_qp(id->qp);
 405 }
 406 EXPORT_SYMBOL(rdma_destroy_qp);
 407
 408 static int cma_modify_qp_rtr(struct rdma_cm_id *id)
 409 {
 410         struct ib_qp_attr qp_attr;
 411         int qp_attr_mask, ret;
 412
 413         if (!id->qp)
 414                 return 0;
 415
 416         /* Need to update QP attributes from default values. */
 417         qp_attr.qp_state = IB_QPS_INIT;
 418         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 419         if (ret)
 420                 return ret;
 421
 422         ret = ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 423         if (ret)
 424                 return ret;
 425
 426         qp_attr.qp_state = IB_QPS_RTR;
 427         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 428         if (ret)
 429                 return ret;
 430
 431         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 432 }
 433
 434 static int cma_modify_qp_rts(struct rdma_cm_id *id)
 435 {
 436         struct ib_qp_attr qp_attr;
 437         int qp_attr_mask, ret;
 438
 439         if (!id->qp)
 440                 return 0;
 441
 442         qp_attr.qp_state = IB_QPS_RTS;
 443         ret = rdma_init_qp_attr(id, &qp_attr, &qp_attr_mask);
 444         if (ret)
 445                 return ret;
 446
 447         return ib_modify_qp(id->qp, &qp_attr, qp_attr_mask);
 448 }
 449
 450 static int cma_modify_qp_err(struct rdma_cm_id *id)
 451 {
 452         struct ib_qp_attr qp_attr;
 453
 454         if (!id->qp)
 455                 return 0;
 456
 457         qp_attr.qp_state = IB_QPS_ERR;
 458         return ib_modify_qp(id->qp, &qp_attr, IB_QP_STATE);
 459 }
 460
 461 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 462                        int *qp_attr_mask)
 463 {
 464         struct rdma_id_private *id_priv;
 465         int ret;
 466
 467         id_priv = container_of(id, struct rdma_id_private, id);
 468         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 469         case RDMA_TRANSPORT_IB:
 470                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 471                                          qp_attr_mask);
 472                 if (qp_attr->qp_state == IB_QPS_RTR)
 473                         qp_attr->rq_psn = id_priv->seq_num;
 474                 break;
 475         case RDMA_TRANSPORT_IWARP:
 476                 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 477                                         qp_attr_mask);
 478                 break;
 479         default:
 480                 ret = -ENOSYS;
 481                 break;
 482         }
 483
 484         return ret;
 485 }
 486 EXPORT_SYMBOL(rdma_init_qp_attr);
 487
 488 static inline int cma_zero_addr(struct sockaddr *addr)
 489 {
 490         struct in6_addr *ip6;
 491
 492         if (addr->sa_family == AF_INET)
 493                 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 494         else {
 495                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 496                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 497                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 498         }
 499 }
 500
 501 static inline int cma_loopback_addr(struct sockaddr *addr)
 502 {
 503         return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 504 }
 505
 506 static inline int cma_any_addr(struct sockaddr *addr)
 507 {
 508         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 509 }
 510
 511 static inline int cma_any_port(struct sockaddr *addr)
 512 {
 513         return !((struct sockaddr_in *) addr)->sin_port;
 514 }
 515
 516 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 517                             u8 *ip_ver, __u16 *port,
 518                             union cma_ip_addr **src, union cma_ip_addr **dst)
 519 {
 520         switch (ps) {
 521         case RDMA_PS_SDP:
 522                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 523                     SDP_MAJ_VERSION)
 524                         return -EINVAL;
 525
 526                 *ip_ver = sdp_get_ip_ver(hdr);
 527                 *port   = ((struct sdp_hh *) hdr)->port;
 528                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 529                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 530                 break;
 531         default:
 532                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 533                         return -EINVAL;
 534
 535                 *ip_ver = cma_get_ip_ver(hdr);
 536                 *port   = ((struct cma_hdr *) hdr)->port;
 537                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 538                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 539                 break;
 540         }
 541
 542         if (*ip_ver != 4 && *ip_ver != 6)
 543                 return -EINVAL;
 544         return 0;
 545 }
 546
 547 static void cma_save_net_info(struct rdma_addr *addr,
 548                               struct rdma_addr *listen_addr,
 549                               u8 ip_ver, __u16 port,
 550                               union cma_ip_addr *src, union cma_ip_addr *dst)
 551 {
 552         struct sockaddr_in *listen4, *ip4;
 553         struct sockaddr_in6 *listen6, *ip6;
 554
 555         switch (ip_ver) {
 556         case 4:
 557                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 558                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 559                 ip4->sin_family = listen4->sin_family;
 560                 ip4->sin_addr.s_addr = dst->ip4.addr;
 561                 ip4->sin_port = listen4->sin_port;
 562
 563                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 564                 ip4->sin_family = listen4->sin_family;
 565                 ip4->sin_addr.s_addr = src->ip4.addr;
 566                 ip4->sin_port = port;
 567                 break;
 568         case 6:
 569                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 570                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 571                 ip6->sin6_family = listen6->sin6_family;
 572                 ip6->sin6_addr = dst->ip6;
 573                 ip6->sin6_port = listen6->sin6_port;
 574
 575                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 576                 ip6->sin6_family = listen6->sin6_family;
 577                 ip6->sin6_addr = src->ip6;
 578                 ip6->sin6_port = port;
 579                 break;
 580         default:
 581                 break;
 582         }
 583 }
 584
 585 static inline int cma_user_data_offset(enum rdma_port_space ps)
 586 {
 587         switch (ps) {
 588         case RDMA_PS_SDP:
 589                 return 0;
 590         default:
 591                 return sizeof(struct cma_hdr);
 592         }
 593 }
 594
 595 static int cma_notify_user(struct rdma_id_private *id_priv,
 596                            enum rdma_cm_event_type type, int status,
 597                            void *data, u8 data_len)
 598 {
 599         struct rdma_cm_event event;
 600
 601         event.event = type;
 602         event.status = status;
 603         event.private_data = data;
 604         event.private_data_len = data_len;
 605
 606         return id_priv->id.event_handler(&id_priv->id, &event);
 607 }
 608
 609 static void cma_cancel_route(struct rdma_id_private *id_priv)
 610 {
 611         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 612         case RDMA_TRANSPORT_IB:
 613                 if (id_priv->query)
 614                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 615                 break;
 616         default:
 617                 break;
 618         }
 619 }
 620
 621 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 622 {
 623         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 624                cma_any_addr(&id_priv->id.route.addr.src_addr);
 625 }
 626
 627 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 628 {
 629         cma_exch(id_priv, CMA_DESTROYING);
 630
 631         if (id_priv->cma_dev) {
 632                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 633                 case RDMA_TRANSPORT_IB:
 634                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 635                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 636                         break;
 637                 case RDMA_TRANSPORT_IWARP:
 638                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 639                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 640                         break;
 641                 default:
 642                         break;
 643                 }
 644                 cma_detach_from_dev(id_priv);
 645         }
 646         list_del(&id_priv->listen_list);
 647
 648         cma_deref_id(id_priv);
 649         wait_for_completion(&id_priv->comp);
 650
 651         kfree(id_priv);
 652 }
 653
 654 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 655 {
 656         struct rdma_id_private *dev_id_priv;
 657
 658         mutex_lock(&lock);
 659         list_del(&id_priv->list);
 660
 661         while (!list_empty(&id_priv->listen_list)) {
 662                 dev_id_priv = list_entry(id_priv->listen_list.next,
 663                                          struct rdma_id_private, listen_list);
 664                 cma_destroy_listen(dev_id_priv);
 665         }
 666         mutex_unlock(&lock);
 667 }
 668
 669 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 670                                  enum cma_state state)
 671 {
 672         switch (state) {
 673         case CMA_ADDR_QUERY:
 674                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 675                 break;
 676         case CMA_ROUTE_QUERY:
 677                 cma_cancel_route(id_priv);
 678                 break;
 679         case CMA_LISTEN:
 680                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 681                     !id_priv->cma_dev)
 682                         cma_cancel_listens(id_priv);
 683                 break;
 684         default:
 685                 break;
 686         }
 687 }
 688
 689 static void cma_release_port(struct rdma_id_private *id_priv)
 690 {
 691         struct rdma_bind_list *bind_list = id_priv->bind_list;
 692
 693         if (!bind_list)
 694                 return;
 695
 696         mutex_lock(&lock);
 697         hlist_del(&id_priv->node);
 698         if (hlist_empty(&bind_list->owners)) {
 699                 idr_remove(bind_list->ps, bind_list->port);
 700                 kfree(bind_list);
 701         }
 702         mutex_unlock(&lock);
 703 }
 704
 705 void rdma_destroy_id(struct rdma_cm_id *id)
 706 {
 707         struct rdma_id_private *id_priv;
 708         enum cma_state state;
 709
 710         id_priv = container_of(id, struct rdma_id_private, id);
 711         state = cma_exch(id_priv, CMA_DESTROYING);
 712         cma_cancel_operation(id_priv, state);
 713
 714         mutex_lock(&lock);
 715         if (id_priv->cma_dev) {
 716                 mutex_unlock(&lock);
 717                 switch (rdma_node_get_transport(id->device->node_type)) {
 718                 case RDMA_TRANSPORT_IB:
 719                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 720                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 721                         break;
 722                 case RDMA_TRANSPORT_IWARP:
 723                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 724                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 725                         break;
 726                 default:
 727                         break;
 728                 }
 729                 mutex_lock(&lock);
 730                 cma_detach_from_dev(id_priv);
 731         }
 732         mutex_unlock(&lock);
 733
 734         cma_release_port(id_priv);
 735         cma_deref_id(id_priv);
 736         wait_for_completion(&id_priv->comp);
 737
 738         kfree(id_priv->id.route.path_rec);
 739         kfree(id_priv);
 740 }
 741 EXPORT_SYMBOL(rdma_destroy_id);
 742
 743 static int cma_rep_recv(struct rdma_id_private *id_priv)
 744 {
 745         int ret;
 746
 747         ret = cma_modify_qp_rtr(&id_priv->id);
 748         if (ret)
 749                 goto reject;
 750
 751         ret = cma_modify_qp_rts(&id_priv->id);
 752         if (ret)
 753                 goto reject;
 754
 755         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 756         if (ret)
 757                 goto reject;
 758
 759         return 0;
 760 reject:
 761         cma_modify_qp_err(&id_priv->id);
 762         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 763                        NULL, 0, NULL, 0);
 764         return ret;
 765 }
 766
 767 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 768 {
 769         if (id_priv->id.ps == RDMA_PS_SDP &&
 770             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 771             SDP_MAJ_VERSION)
 772                 return -EINVAL;
 773
 774         return 0;
 775 }
 776
 777 static int cma_rtu_recv(struct rdma_id_private *id_priv)
 778 {
 779         int ret;
 780
 781         ret = cma_modify_qp_rts(&id_priv->id);
 782         if (ret)
 783                 goto reject;
 784
 785         return 0;
 786 reject:
 787         cma_modify_qp_err(&id_priv->id);
 788         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 789                        NULL, 0, NULL, 0);
 790         return ret;
 791 }
 792
 793 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 794 {
 795         struct rdma_id_private *id_priv = cm_id->context;
 796         enum rdma_cm_event_type event;
 797         u8 private_data_len = 0;
 798         int ret = 0, status = 0;
 799
 800         atomic_inc(&id_priv->dev_remove);
 801         if (!cma_comp(id_priv, CMA_CONNECT))
 802                 goto out;
 803
 804         switch (ib_event->event) {
 805         case IB_CM_REQ_ERROR:
 806         case IB_CM_REP_ERROR:
 807                 event = RDMA_CM_EVENT_UNREACHABLE;
 808                 status = -ETIMEDOUT;
 809                 break;
 810         case IB_CM_REP_RECEIVED:
 811                 status = cma_verify_rep(id_priv, ib_event->private_data);
 812                 if (status)
 813                         event = RDMA_CM_EVENT_CONNECT_ERROR;
 814                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 815                         status = cma_rep_recv(id_priv);
 816                         event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 817                                          RDMA_CM_EVENT_ESTABLISHED;
 818                 } else
 819                         event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 820                 private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 821                 break;
 822         case IB_CM_RTU_RECEIVED:
 823                 status = cma_rtu_recv(id_priv);
 824                 event = status ? RDMA_CM_EVENT_CONNECT_ERROR :
 825                                  RDMA_CM_EVENT_ESTABLISHED;
 826                 break;
 827         case IB_CM_DREQ_ERROR:
 828                 status = -ETIMEDOUT; /* fall through */
 829         case IB_CM_DREQ_RECEIVED:
 830         case IB_CM_DREP_RECEIVED:
 831                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 832                         goto out;
 833                 event = RDMA_CM_EVENT_DISCONNECTED;
 834                 break;
 835         case IB_CM_TIMEWAIT_EXIT:
 836         case IB_CM_MRA_RECEIVED:
 837                 /* ignore event */
 838                 goto out;
 839         case IB_CM_REJ_RECEIVED:
 840                 cma_modify_qp_err(&id_priv->id);
 841                 status = ib_event->param.rej_rcvd.reason;
 842                 event = RDMA_CM_EVENT_REJECTED;
 843                 private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 844                 break;
 845         default:
 846                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 847                        ib_event->event);
 848                 goto out;
 849         }
 850
 851         ret = cma_notify_user(id_priv, event, status, ib_event->private_data,
 852                               private_data_len);
 853         if (ret) {
 854                 /* Destroy the CM ID by returning a non-zero value. */
 855                 id_priv->cm_id.ib = NULL;
 856                 cma_exch(id_priv, CMA_DESTROYING);
 857                 cma_release_remove(id_priv);
 858                 rdma_destroy_id(&id_priv->id);
 859                 return ret;
 860         }
 861 out:
 862         cma_release_remove(id_priv);
 863         return ret;
 864 }
 865
 866 static struct rdma_id_private *cma_new_id(struct rdma_cm_id *listen_id,
 867                                           struct ib_cm_event *ib_event)
 868 {
 869         struct rdma_id_private *id_priv;
 870         struct rdma_cm_id *id;
 871         struct rdma_route *rt;
 872         union cma_ip_addr *src, *dst;
 873         __u16 port;
 874         u8 ip_ver;
 875
 876         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 877                              &ip_ver, &port, &src, &dst))
 878                 goto err;
 879
 880         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 881                             listen_id->ps);
 882         if (IS_ERR(id))
 883                 goto err;
 884
 885         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 886                           ip_ver, port, src, dst);
 887
 888         rt = &id->route;
 889         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 890         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
 891                                GFP_KERNEL);
 892         if (!rt->path_rec)
 893                 goto destroy_id;
 894
 895         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 896         if (rt->num_paths == 2)
 897                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 898
 899         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 900         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 901         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 902         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
 903
 904         id_priv = container_of(id, struct rdma_id_private, id);
 905         id_priv->state = CMA_CONNECT;
 906         return id_priv;
 907
 908 destroy_id:
 909         rdma_destroy_id(id);
 910 err:
 911         return NULL;
 912 }
 913
 914 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 915 {
 916         struct rdma_id_private *listen_id, *conn_id;
 917         int offset, ret;
 918
 919         listen_id = cm_id->context;
 920         atomic_inc(&listen_id->dev_remove);
 921         if (!cma_comp(listen_id, CMA_LISTEN)) {
 922                 ret = -ECONNABORTED;
 923                 goto out;
 924         }
 925
 926         conn_id = cma_new_id(&listen_id->id, ib_event);
 927         if (!conn_id) {
 928                 ret = -ENOMEM;
 929                 goto out;
 930         }
 931
 932         atomic_inc(&conn_id->dev_remove);
 933         mutex_lock(&lock);
 934         ret = cma_acquire_dev(conn_id);
 935         mutex_unlock(&lock);
 936         if (ret)
 937                 goto release_conn_id;
 938
 939         conn_id->cm_id.ib = cm_id;
 940         cm_id->context = conn_id;
 941         cm_id->cm_handler = cma_ib_handler;
 942
 943         offset = cma_user_data_offset(listen_id->id.ps);
 944         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
 945                               ib_event->private_data + offset,
 946                               IB_CM_REQ_PRIVATE_DATA_SIZE - offset);
 947         if (!ret)
 948                 goto out;
 949
 950         /* Destroy the CM ID by returning a non-zero value. */
 951         conn_id->cm_id.ib = NULL;
 952
 953 release_conn_id:
 954         cma_exch(conn_id, CMA_DESTROYING);
 955         cma_release_remove(conn_id);
 956         rdma_destroy_id(&conn_id->id);
 957
 958 out:
 959         cma_release_remove(listen_id);
 960         return ret;
 961 }
 962
 963 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 964 {
 965         return cpu_to_be64(((u64)ps << 16) +
 966                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 967 }
 968
 969 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 970                                  struct ib_cm_compare_data *compare)
 971 {
 972         struct cma_hdr *cma_data, *cma_mask;
 973         struct sdp_hh *sdp_data, *sdp_mask;
 974         __u32 ip4_addr;
 975         struct in6_addr ip6_addr;
 976
 977         memset(compare, 0, sizeof *compare);
 978         cma_data = (void *) compare->data;
 979         cma_mask = (void *) compare->mask;
 980         sdp_data = (void *) compare->data;
 981         sdp_mask = (void *) compare->mask;
 982
 983         switch (addr->sa_family) {
 984         case AF_INET:
 985                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 986                 if (ps == RDMA_PS_SDP) {
 987                         sdp_set_ip_ver(sdp_data, 4);
 988                         sdp_set_ip_ver(sdp_mask, 0xF);
 989                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 990                         sdp_mask->dst_addr.ip4.addr = ~0;
 991                 } else {
 992                         cma_set_ip_ver(cma_data, 4);
 993                         cma_set_ip_ver(cma_mask, 0xF);
 994                         cma_data->dst_addr.ip4.addr = ip4_addr;
 995                         cma_mask->dst_addr.ip4.addr = ~0;
 996                 }
 997                 break;
 998         case AF_INET6:
 999                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1000                 if (ps == RDMA_PS_SDP) {
1001                         sdp_set_ip_ver(sdp_data, 6);
1002                         sdp_set_ip_ver(sdp_mask, 0xF);
1003                         sdp_data->dst_addr.ip6 = ip6_addr;
1004                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1005                                sizeof sdp_mask->dst_addr.ip6);
1006                 } else {
1007                         cma_set_ip_ver(cma_data, 6);
1008                         cma_set_ip_ver(cma_mask, 0xF);
1009                         cma_data->dst_addr.ip6 = ip6_addr;
1010                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1011                                sizeof cma_mask->dst_addr.ip6);
1012                 }
1013                 break;
1014         default:
1015                 break;
1016         }
1017 }
1018
1019 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1020 {
1021         struct rdma_id_private *id_priv = iw_id->context;
1022         enum rdma_cm_event_type event = 0;
1023         struct sockaddr_in *sin;
1024         int ret = 0;
1025
1026         atomic_inc(&id_priv->dev_remove);
1027
1028         switch (iw_event->event) {
1029         case IW_CM_EVENT_CLOSE:
1030                 event = RDMA_CM_EVENT_DISCONNECTED;
1031                 break;
1032         case IW_CM_EVENT_CONNECT_REPLY:
1033                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1034                 *sin = iw_event->local_addr;
1035                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1036                 *sin = iw_event->remote_addr;
1037                 if (iw_event->status)
1038                         event = RDMA_CM_EVENT_REJECTED;
1039                 else
1040                         event = RDMA_CM_EVENT_ESTABLISHED;
1041                 break;
1042         case IW_CM_EVENT_ESTABLISHED:
1043                 event = RDMA_CM_EVENT_ESTABLISHED;
1044                 break;
1045         default:
1046                 BUG_ON(1);
1047         }
1048
1049         ret = cma_notify_user(id_priv, event, iw_event->status,
1050                               iw_event->private_data,
1051                               iw_event->private_data_len);
1052         if (ret) {
1053                 /* Destroy the CM ID by returning a non-zero value. */
1054                 id_priv->cm_id.iw = NULL;
1055                 cma_exch(id_priv, CMA_DESTROYING);
1056                 cma_release_remove(id_priv);
1057                 rdma_destroy_id(&id_priv->id);
1058                 return ret;
1059         }
1060
1061         cma_release_remove(id_priv);
1062         return ret;
1063 }
1064
1065 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1066                                struct iw_cm_event *iw_event)
1067 {
1068         struct rdma_cm_id *new_cm_id;
1069         struct rdma_id_private *listen_id, *conn_id;
1070         struct sockaddr_in *sin;
1071         struct net_device *dev = NULL;
1072         int ret;
1073
1074         listen_id = cm_id->context;
1075         atomic_inc(&listen_id->dev_remove);
1076         if (!cma_comp(listen_id, CMA_LISTEN)) {
1077                 ret = -ECONNABORTED;
1078                 goto out;
1079         }
1080
1081         /* Create a new RDMA id for the new IW CM ID */
1082         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1083                                    listen_id->id.context,
1084                                    RDMA_PS_TCP);
1085         if (!new_cm_id) {
1086                 ret = -ENOMEM;
1087                 goto out;
1088         }
1089         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1090         atomic_inc(&conn_id->dev_remove);
1091         conn_id->state = CMA_CONNECT;
1092
1093         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1094         if (!dev) {
1095                 ret = -EADDRNOTAVAIL;
1096                 cma_release_remove(conn_id);
1097                 rdma_destroy_id(new_cm_id);
1098                 goto out;
1099         }
1100         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1101         if (ret) {
1102                 cma_release_remove(conn_id);
1103                 rdma_destroy_id(new_cm_id);
1104                 goto out;
1105         }
1106
1107         mutex_lock(&lock);
1108         ret = cma_acquire_dev(conn_id);
1109         mutex_unlock(&lock);
1110         if (ret) {
1111                 cma_release_remove(conn_id);
1112                 rdma_destroy_id(new_cm_id);
1113                 goto out;
1114         }
1115
1116         conn_id->cm_id.iw = cm_id;
1117         cm_id->context = conn_id;
1118         cm_id->cm_handler = cma_iw_handler;
1119
1120         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1121         *sin = iw_event->local_addr;
1122         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1123         *sin = iw_event->remote_addr;
1124
1125         ret = cma_notify_user(conn_id, RDMA_CM_EVENT_CONNECT_REQUEST, 0,
1126                               iw_event->private_data,
1127                               iw_event->private_data_len);
1128         if (ret) {
1129                 /* User wants to destroy the CM ID */
1130                 conn_id->cm_id.iw = NULL;
1131                 cma_exch(conn_id, CMA_DESTROYING);
1132                 cma_release_remove(conn_id);
1133                 rdma_destroy_id(&conn_id->id);
1134         }
1135
1136 out:
1137         if (dev)
1138                 dev_put(dev);
1139         cma_release_remove(listen_id);
1140         return ret;
1141 }
1142
1143 static int cma_ib_listen(struct rdma_id_private *id_priv)
1144 {
1145         struct ib_cm_compare_data compare_data;
1146         struct sockaddr *addr;
1147         __be64 svc_id;
1148         int ret;
1149
1150         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1151                                             id_priv);
1152         if (IS_ERR(id_priv->cm_id.ib))
1153                 return PTR_ERR(id_priv->cm_id.ib);
1154
1155         addr = &id_priv->id.route.addr.src_addr;
1156         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1157         if (cma_any_addr(addr))
1158                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1159         else {
1160                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1161                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1162         }
1163
1164         if (ret) {
1165                 ib_destroy_cm_id(id_priv->cm_id.ib);
1166                 id_priv->cm_id.ib = NULL;
1167         }
1168
1169         return ret;
1170 }
1171
1172 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1173 {
1174         int ret;
1175         struct sockaddr_in *sin;
1176
1177         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1178                                             iw_conn_req_handler,
1179                                             id_priv);
1180         if (IS_ERR(id_priv->cm_id.iw))
1181                 return PTR_ERR(id_priv->cm_id.iw);
1182
1183         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1184         id_priv->cm_id.iw->local_addr = *sin;
1185
1186         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1187
1188         if (ret) {
1189                 iw_destroy_cm_id(id_priv->cm_id.iw);
1190                 id_priv->cm_id.iw = NULL;
1191         }
1192
1193         return ret;
1194 }
1195
1196 static int cma_listen_handler(struct rdma_cm_id *id,
1197                               struct rdma_cm_event *event)
1198 {
1199         struct rdma_id_private *id_priv = id->context;
1200
1201         id->context = id_priv->id.context;
1202         id->event_handler = id_priv->id.event_handler;
1203         return id_priv->id.event_handler(id, event);
1204 }
1205
1206 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1207                               struct cma_device *cma_dev)
1208 {
1209         struct rdma_id_private *dev_id_priv;
1210         struct rdma_cm_id *id;
1211         int ret;
1212
1213         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1214         if (IS_ERR(id))
1215                 return;
1216
1217         dev_id_priv = container_of(id, struct rdma_id_private, id);
1218
1219         dev_id_priv->state = CMA_ADDR_BOUND;
1220         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1221                ip_addr_size(&id_priv->id.route.addr.src_addr));
1222
1223         cma_attach_to_dev(dev_id_priv, cma_dev);
1224         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1225
1226         ret = rdma_listen(id, id_priv->backlog);
1227         if (ret)
1228                 goto err;
1229
1230         return;
1231 err:
1232         cma_destroy_listen(dev_id_priv);
1233 }
1234
1235 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1236 {
1237         struct cma_device *cma_dev;
1238
1239         mutex_lock(&lock);
1240         list_add_tail(&id_priv->list, &listen_any_list);
1241         list_for_each_entry(cma_dev, &dev_list, list)
1242                 cma_listen_on_dev(id_priv, cma_dev);
1243         mutex_unlock(&lock);
1244 }
1245
1246 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1247 {
1248         struct sockaddr_in addr_in;
1249
1250         memset(&addr_in, 0, sizeof addr_in);
1251         addr_in.sin_family = af;
1252         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1253 }
1254
1255 int rdma_listen(struct rdma_cm_id *id, int backlog)
1256 {
1257         struct rdma_id_private *id_priv;
1258         int ret;
1259
1260         id_priv = container_of(id, struct rdma_id_private, id);
1261         if (id_priv->state == CMA_IDLE) {
1262                 ret = cma_bind_any(id, AF_INET);
1263                 if (ret)
1264                         return ret;
1265         }
1266
1267         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1268                 return -EINVAL;
1269
1270         id_priv->backlog = backlog;
1271         if (id->device) {
1272                 switch (rdma_node_get_transport(id->device->node_type)) {
1273                 case RDMA_TRANSPORT_IB:
1274                         ret = cma_ib_listen(id_priv);
1275                         if (ret)
1276                                 goto err;
1277                         break;
1278                 case RDMA_TRANSPORT_IWARP:
1279                         ret = cma_iw_listen(id_priv, backlog);
1280                         if (ret)
1281                                 goto err;
1282                         break;
1283                 default:
1284                         ret = -ENOSYS;
1285                         goto err;
1286                 }
1287         } else
1288                 cma_listen_on_all(id_priv);
1289
1290         return 0;
1291 err:
1292         id_priv->backlog = 0;
1293         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1294         return ret;
1295 }
1296 EXPORT_SYMBOL(rdma_listen);
1297
1298 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1299                               void *context)
1300 {
1301         struct cma_work *work = context;
1302         struct rdma_route *route;
1303
1304         route = &work->id->id.route;
1305
1306         if (!status) {
1307                 route->num_paths = 1;
1308                 *route->path_rec = *path_rec;
1309         } else {
1310                 work->old_state = CMA_ROUTE_QUERY;
1311                 work->new_state = CMA_ADDR_RESOLVED;
1312                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1313                 work->event.status = status;
1314         }
1315
1316         queue_work(cma_wq, &work->work);
1317 }
1318
1319 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1320                               struct cma_work *work)
1321 {
1322         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1323         struct ib_sa_path_rec path_rec;
1324
1325         memset(&path_rec, 0, sizeof path_rec);
1326         ib_addr_get_sgid(addr, &path_rec.sgid);
1327         ib_addr_get_dgid(addr, &path_rec.dgid);
1328         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1329         path_rec.numb_path = 1;
1330
1331         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1332                                 id_priv->id.port_num, &path_rec,
1333                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1334                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1335                                 timeout_ms, GFP_KERNEL,
1336                                 cma_query_handler, work, &id_priv->query);
1337
1338         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1339 }
1340
1341 static void cma_work_handler(struct work_struct *_work)
1342 {
1343         struct cma_work *work = container_of(_work, struct cma_work, work);
1344         struct rdma_id_private *id_priv = work->id;
1345         int destroy = 0;
1346
1347         atomic_inc(&id_priv->dev_remove);
1348         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1349                 goto out;
1350
1351         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1352                 cma_exch(id_priv, CMA_DESTROYING);
1353                 destroy = 1;
1354         }
1355 out:
1356         cma_release_remove(id_priv);
1357         cma_deref_id(id_priv);
1358         if (destroy)
1359                 rdma_destroy_id(&id_priv->id);
1360         kfree(work);
1361 }
1362
1363 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1364 {
1365         struct rdma_route *route = &id_priv->id.route;
1366         struct cma_work *work;
1367         int ret;
1368
1369         work = kzalloc(sizeof *work, GFP_KERNEL);
1370         if (!work)
1371                 return -ENOMEM;
1372
1373         work->id = id_priv;
1374         INIT_WORK(&work->work, cma_work_handler);
1375         work->old_state = CMA_ROUTE_QUERY;
1376         work->new_state = CMA_ROUTE_RESOLVED;
1377         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1378
1379         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1380         if (!route->path_rec) {
1381                 ret = -ENOMEM;
1382                 goto err1;
1383         }
1384
1385         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1386         if (ret)
1387                 goto err2;
1388
1389         return 0;
1390 err2:
1391         kfree(route->path_rec);
1392         route->path_rec = NULL;
1393 err1:
1394         kfree(work);
1395         return ret;
1396 }
1397
1398 int rdma_set_ib_paths(struct rdma_cm_id *id,
1399                       struct ib_sa_path_rec *path_rec, int num_paths)
1400 {
1401         struct rdma_id_private *id_priv;
1402         int ret;
1403
1404         id_priv = container_of(id, struct rdma_id_private, id);
1405         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1406                 return -EINVAL;
1407
1408         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1409         if (!id->route.path_rec) {
1410                 ret = -ENOMEM;
1411                 goto err;
1412         }
1413
1414         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1415         return 0;
1416 err:
1417         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1418         return ret;
1419 }
1420 EXPORT_SYMBOL(rdma_set_ib_paths);
1421
1422 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1423 {
1424         struct cma_work *work;
1425
1426         work = kzalloc(sizeof *work, GFP_KERNEL);
1427         if (!work)
1428                 return -ENOMEM;
1429
1430         work->id = id_priv;
1431         INIT_WORK(&work->work, cma_work_handler);
1432         work->old_state = CMA_ROUTE_QUERY;
1433         work->new_state = CMA_ROUTE_RESOLVED;
1434         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1435         queue_work(cma_wq, &work->work);
1436         return 0;
1437 }
1438
1439 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1440 {
1441         struct rdma_id_private *id_priv;
1442         int ret;
1443
1444         id_priv = container_of(id, struct rdma_id_private, id);
1445         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1446                 return -EINVAL;
1447
1448         atomic_inc(&id_priv->refcount);
1449         switch (rdma_node_get_transport(id->device->node_type)) {
1450         case RDMA_TRANSPORT_IB:
1451                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1452                 break;
1453         case RDMA_TRANSPORT_IWARP:
1454                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1455                 break;
1456         default:
1457                 ret = -ENOSYS;
1458                 break;
1459         }
1460         if (ret)
1461                 goto err;
1462
1463         return 0;
1464 err:
1465         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1466         cma_deref_id(id_priv);
1467         return ret;
1468 }
1469 EXPORT_SYMBOL(rdma_resolve_route);
1470
1471 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1472 {
1473         struct cma_device *cma_dev;
1474         struct ib_port_attr port_attr;
1475         union ib_gid gid;
1476         u16 pkey;
1477         int ret;
1478         u8 p;
1479
1480         mutex_lock(&lock);
1481         if (list_empty(&dev_list)) {
1482                 ret = -ENODEV;
1483                 goto out;
1484         }
1485         list_for_each_entry(cma_dev, &dev_list, list)
1486                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1487                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1488                             port_attr.state == IB_PORT_ACTIVE)
1489                                 goto port_found;
1490
1491         p = 1;
1492         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1493
1494 port_found:
1495         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1496         if (ret)
1497                 goto out;
1498
1499         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1500         if (ret)
1501                 goto out;
1502
1503         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1504         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1505         id_priv->id.port_num = p;
1506         cma_attach_to_dev(id_priv, cma_dev);
1507 out:
1508         mutex_unlock(&lock);
1509         return ret;
1510 }
1511
1512 static void addr_handler(int status, struct sockaddr *src_addr,
1513                          struct rdma_dev_addr *dev_addr, void *context)
1514 {
1515         struct rdma_id_private *id_priv = context;
1516         enum rdma_cm_event_type event;
1517
1518         atomic_inc(&id_priv->dev_remove);
1519
1520         /*
1521          * Grab mutex to block rdma_destroy_id() from removing the device while
1522          * we're trying to acquire it.
1523          */
1524         mutex_lock(&lock);
1525         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1526                 mutex_unlock(&lock);
1527                 goto out;
1528         }
1529
1530         if (!status && !id_priv->cma_dev)
1531                 status = cma_acquire_dev(id_priv);
1532         mutex_unlock(&lock);
1533
1534         if (status) {
1535                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1536                         goto out;
1537                 event = RDMA_CM_EVENT_ADDR_ERROR;
1538         } else {
1539                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1540                        ip_addr_size(src_addr));
1541                 event = RDMA_CM_EVENT_ADDR_RESOLVED;
1542         }
1543
1544         if (cma_notify_user(id_priv, event, status, NULL, 0)) {
1545                 cma_exch(id_priv, CMA_DESTROYING);
1546                 cma_release_remove(id_priv);
1547                 cma_deref_id(id_priv);
1548                 rdma_destroy_id(&id_priv->id);
1549                 return;
1550         }
1551 out:
1552         cma_release_remove(id_priv);
1553         cma_deref_id(id_priv);
1554 }
1555
1556 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1557 {
1558         struct cma_work *work;
1559         struct sockaddr_in *src_in, *dst_in;
1560         union ib_gid gid;
1561         int ret;
1562
1563         work = kzalloc(sizeof *work, GFP_KERNEL);
1564         if (!work)
1565                 return -ENOMEM;
1566
1567         if (!id_priv->cma_dev) {
1568                 ret = cma_bind_loopback(id_priv);
1569                 if (ret)
1570                         goto err;
1571         }
1572
1573         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1574         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1575
1576         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1577                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1578                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1579                 src_in->sin_family = dst_in->sin_family;
1580                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1581         }
1582
1583         work->id = id_priv;
1584         INIT_WORK(&work->work, cma_work_handler);
1585         work->old_state = CMA_ADDR_QUERY;
1586         work->new_state = CMA_ADDR_RESOLVED;
1587         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1588         queue_work(cma_wq, &work->work);
1589         return 0;
1590 err:
1591         kfree(work);
1592         return ret;
1593 }
1594
1595 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1596                          struct sockaddr *dst_addr)
1597 {
1598         if (src_addr && src_addr->sa_family)
1599                 return rdma_bind_addr(id, src_addr);
1600         else
1601                 return cma_bind_any(id, dst_addr->sa_family);
1602 }
1603
1604 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1605                       struct sockaddr *dst_addr, int timeout_ms)
1606 {
1607         struct rdma_id_private *id_priv;
1608         int ret;
1609
1610         id_priv = container_of(id, struct rdma_id_private, id);
1611         if (id_priv->state == CMA_IDLE) {
1612                 ret = cma_bind_addr(id, src_addr, dst_addr);
1613                 if (ret)
1614                         return ret;
1615         }
1616
1617         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1618                 return -EINVAL;
1619
1620         atomic_inc(&id_priv->refcount);
1621         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1622         if (cma_any_addr(dst_addr))
1623                 ret = cma_resolve_loopback(id_priv);
1624         else
1625                 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1626                                       dst_addr, &id->route.addr.dev_addr,
1627                                       timeout_ms, addr_handler, id_priv);
1628         if (ret)
1629                 goto err;
1630
1631         return 0;
1632 err:
1633         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1634         cma_deref_id(id_priv);
1635         return ret;
1636 }
1637 EXPORT_SYMBOL(rdma_resolve_addr);
1638
1639 static void cma_bind_port(struct rdma_bind_list *bind_list,
1640                           struct rdma_id_private *id_priv)
1641 {
1642         struct sockaddr_in *sin;
1643
1644         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1645         sin->sin_port = htons(bind_list->port);
1646         id_priv->bind_list = bind_list;
1647         hlist_add_head(&id_priv->node, &bind_list->owners);
1648 }
1649
1650 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1651                           unsigned short snum)
1652 {
1653         struct rdma_bind_list *bind_list;
1654         int port, start, ret;
1655
1656         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1657         if (!bind_list)
1658                 return -ENOMEM;
1659
1660         start = snum ? snum : sysctl_local_port_range[0];
1661
1662         do {
1663                 ret = idr_get_new_above(ps, bind_list, start, &port);
1664         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1665
1666         if (ret)
1667                 goto err;
1668
1669         if ((snum && port != snum) ||
1670             (!snum && port > sysctl_local_port_range[1])) {
1671                 idr_remove(ps, port);
1672                 ret = -EADDRNOTAVAIL;
1673                 goto err;
1674         }
1675
1676         bind_list->ps = ps;
1677         bind_list->port = (unsigned short) port;
1678         cma_bind_port(bind_list, id_priv);
1679         return 0;
1680 err:
1681         kfree(bind_list);
1682         return ret;
1683 }
1684
1685 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1686 {
1687         struct rdma_id_private *cur_id;
1688         struct sockaddr_in *sin, *cur_sin;
1689         struct rdma_bind_list *bind_list;
1690         struct hlist_node *node;
1691         unsigned short snum;
1692
1693         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1694         snum = ntohs(sin->sin_port);
1695         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1696                 return -EACCES;
1697
1698         bind_list = idr_find(ps, snum);
1699         if (!bind_list)
1700                 return cma_alloc_port(ps, id_priv, snum);
1701
1702         /*
1703          * We don't support binding to any address if anyone is bound to
1704          * a specific address on the same port.
1705          */
1706         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1707                 return -EADDRNOTAVAIL;
1708
1709         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1710                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1711                         return -EADDRNOTAVAIL;
1712
1713                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1714                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1715                         return -EADDRINUSE;
1716         }
1717
1718         cma_bind_port(bind_list, id_priv);
1719         return 0;
1720 }
1721
1722 static int cma_get_port(struct rdma_id_private *id_priv)
1723 {
1724         struct idr *ps;
1725         int ret;
1726
1727         switch (id_priv->id.ps) {
1728         case RDMA_PS_SDP:
1729                 ps = &sdp_ps;
1730                 break;
1731         case RDMA_PS_TCP:
1732                 ps = &tcp_ps;
1733                 break;
1734         default:
1735                 return -EPROTONOSUPPORT;
1736         }
1737
1738         mutex_lock(&lock);
1739         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1740                 ret = cma_alloc_port(ps, id_priv, 0);
1741         else
1742                 ret = cma_use_port(ps, id_priv);
1743         mutex_unlock(&lock);
1744
1745         return ret;
1746 }
1747
1748 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1749 {
1750         struct rdma_id_private *id_priv;
1751         int ret;
1752
1753         if (addr->sa_family != AF_INET)
1754                 return -EAFNOSUPPORT;
1755
1756         id_priv = container_of(id, struct rdma_id_private, id);
1757         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1758                 return -EINVAL;
1759
1760         if (!cma_any_addr(addr)) {
1761                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1762                 if (ret)
1763                         goto err1;
1764
1765                 mutex_lock(&lock);
1766                 ret = cma_acquire_dev(id_priv);
1767                 mutex_unlock(&lock);
1768                 if (ret)
1769                         goto err1;
1770         }
1771
1772         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1773         ret = cma_get_port(id_priv);
1774         if (ret)
1775                 goto err2;
1776
1777         return 0;
1778 err2:
1779         if (!cma_any_addr(addr)) {
1780                 mutex_lock(&lock);
1781                 cma_detach_from_dev(id_priv);
1782                 mutex_unlock(&lock);
1783         }
1784 err1:
1785         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1786         return ret;
1787 }
1788 EXPORT_SYMBOL(rdma_bind_addr);
1789
1790 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1791                           struct rdma_route *route)
1792 {
1793         struct sockaddr_in *src4, *dst4;
1794         struct cma_hdr *cma_hdr;
1795         struct sdp_hh *sdp_hdr;
1796
1797         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1798         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1799
1800         switch (ps) {
1801         case RDMA_PS_SDP:
1802                 sdp_hdr = hdr;
1803                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1804                         return -EINVAL;
1805                 sdp_set_ip_ver(sdp_hdr, 4);
1806                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1807                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1808                 sdp_hdr->port = src4->sin_port;
1809                 break;
1810         default:
1811                 cma_hdr = hdr;
1812                 cma_hdr->cma_version = CMA_VERSION;
1813                 cma_set_ip_ver(cma_hdr, 4);
1814                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1815                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1816                 cma_hdr->port = src4->sin_port;
1817                 break;
1818         }
1819         return 0;
1820 }
1821
1822 static int cma_connect_ib(struct rdma_id_private *id_priv,
1823                           struct rdma_conn_param *conn_param)
1824 {
1825         struct ib_cm_req_param req;
1826         struct rdma_route *route;
1827         void *private_data;
1828         int offset, ret;
1829
1830         memset(&req, 0, sizeof req);
1831         offset = cma_user_data_offset(id_priv->id.ps);
1832         req.private_data_len = offset + conn_param->private_data_len;
1833         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1834         if (!private_data)
1835                 return -ENOMEM;
1836
1837         if (conn_param->private_data && conn_param->private_data_len)
1838                 memcpy(private_data + offset, conn_param->private_data,
1839                        conn_param->private_data_len);
1840
1841         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1842                                             id_priv);
1843         if (IS_ERR(id_priv->cm_id.ib)) {
1844                 ret = PTR_ERR(id_priv->cm_id.ib);
1845                 goto out;
1846         }
1847
1848         route = &id_priv->id.route;
1849         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1850         if (ret)
1851                 goto out;
1852         req.private_data = private_data;
1853
1854         req.primary_path = &route->path_rec[0];
1855         if (route->num_paths == 2)
1856                 req.alternate_path = &route->path_rec[1];
1857
1858         req.service_id = cma_get_service_id(id_priv->id.ps,
1859                                             &route->addr.dst_addr);
1860         req.qp_num = id_priv->qp_num;
1861         req.qp_type = IB_QPT_RC;
1862         req.starting_psn = id_priv->seq_num;
1863         req.responder_resources = conn_param->responder_resources;
1864         req.initiator_depth = conn_param->initiator_depth;
1865         req.flow_control = conn_param->flow_control;
1866         req.retry_count = conn_param->retry_count;
1867         req.rnr_retry_count = conn_param->rnr_retry_count;
1868         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1869         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1870         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1871         req.srq = id_priv->srq ? 1 : 0;
1872
1873         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1874 out:
1875         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
1876                 ib_destroy_cm_id(id_priv->cm_id.ib);
1877                 id_priv->cm_id.ib = NULL;
1878         }
1879
1880         kfree(private_data);
1881         return ret;
1882 }
1883
1884 static int cma_connect_iw(struct rdma_id_private *id_priv,
1885                           struct rdma_conn_param *conn_param)
1886 {
1887         struct iw_cm_id *cm_id;
1888         struct sockaddr_in* sin;
1889         int ret;
1890         struct iw_cm_conn_param iw_param;
1891
1892         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
1893         if (IS_ERR(cm_id)) {
1894                 ret = PTR_ERR(cm_id);
1895                 goto out;
1896         }
1897
1898         id_priv->cm_id.iw = cm_id;
1899
1900         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
1901         cm_id->local_addr = *sin;
1902
1903         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
1904         cm_id->remote_addr = *sin;
1905
1906         ret = cma_modify_qp_rtr(&id_priv->id);
1907         if (ret)
1908                 goto out;
1909
1910         iw_param.ord = conn_param->initiator_depth;
1911         iw_param.ird = conn_param->responder_resources;
1912         iw_param.private_data = conn_param->private_data;
1913         iw_param.private_data_len = conn_param->private_data_len;
1914         if (id_priv->id.qp)
1915                 iw_param.qpn = id_priv->qp_num;
1916         else
1917                 iw_param.qpn = conn_param->qp_num;
1918         ret = iw_cm_connect(cm_id, &iw_param);
1919 out:
1920         if (ret && !IS_ERR(cm_id)) {
1921                 iw_destroy_cm_id(cm_id);
1922                 id_priv->cm_id.iw = NULL;
1923         }
1924         return ret;
1925 }
1926
1927 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1928 {
1929         struct rdma_id_private *id_priv;
1930         int ret;
1931
1932         id_priv = container_of(id, struct rdma_id_private, id);
1933         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1934                 return -EINVAL;
1935
1936         if (!id->qp) {
1937                 id_priv->qp_num = conn_param->qp_num;
1938                 id_priv->srq = conn_param->srq;
1939         }
1940
1941         switch (rdma_node_get_transport(id->device->node_type)) {
1942         case RDMA_TRANSPORT_IB:
1943                 ret = cma_connect_ib(id_priv, conn_param);
1944                 break;
1945         case RDMA_TRANSPORT_IWARP:
1946                 ret = cma_connect_iw(id_priv, conn_param);
1947                 break;
1948         default:
1949                 ret = -ENOSYS;
1950                 break;
1951         }
1952         if (ret)
1953                 goto err;
1954
1955         return 0;
1956 err:
1957         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1958         return ret;
1959 }
1960 EXPORT_SYMBOL(rdma_connect);
1961
1962 static int cma_accept_ib(struct rdma_id_private *id_priv,
1963                          struct rdma_conn_param *conn_param)
1964 {
1965         struct ib_cm_rep_param rep;
1966         int ret;
1967
1968         ret = cma_modify_qp_rtr(&id_priv->id);
1969         if (ret)
1970                 return ret;
1971
1972         memset(&rep, 0, sizeof rep);
1973         rep.qp_num = id_priv->qp_num;
1974         rep.starting_psn = id_priv->seq_num;
1975         rep.private_data = conn_param->private_data;
1976         rep.private_data_len = conn_param->private_data_len;
1977         rep.responder_resources = conn_param->responder_resources;
1978         rep.initiator_depth = conn_param->initiator_depth;
1979         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
1980         rep.failover_accepted = 0;
1981         rep.flow_control = conn_param->flow_control;
1982         rep.rnr_retry_count = conn_param->rnr_retry_count;
1983         rep.srq = id_priv->srq ? 1 : 0;
1984
1985         return ib_send_cm_rep(id_priv->cm_id.ib, &rep);
1986 }
1987
1988 static int cma_accept_iw(struct rdma_id_private *id_priv,
1989                   struct rdma_conn_param *conn_param)
1990 {
1991         struct iw_cm_conn_param iw_param;
1992         int ret;
1993
1994         ret = cma_modify_qp_rtr(&id_priv->id);
1995         if (ret)
1996                 return ret;
1997
1998         iw_param.ord = conn_param->initiator_depth;
1999         iw_param.ird = conn_param->responder_resources;
2000         iw_param.private_data = conn_param->private_data;
2001         iw_param.private_data_len = conn_param->private_data_len;
2002         if (id_priv->id.qp) {
2003                 iw_param.qpn = id_priv->qp_num;
2004         } else
2005                 iw_param.qpn = conn_param->qp_num;
2006
2007         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2008 }
2009
2010 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2011 {
2012         struct rdma_id_private *id_priv;
2013         int ret;
2014
2015         id_priv = container_of(id, struct rdma_id_private, id);
2016         if (!cma_comp(id_priv, CMA_CONNECT))
2017                 return -EINVAL;
2018
2019         if (!id->qp && conn_param) {
2020                 id_priv->qp_num = conn_param->qp_num;
2021                 id_priv->srq = conn_param->srq;
2022         }
2023
2024         switch (rdma_node_get_transport(id->device->node_type)) {
2025         case RDMA_TRANSPORT_IB:
2026                 if (conn_param)
2027                         ret = cma_accept_ib(id_priv, conn_param);
2028                 else
2029                         ret = cma_rep_recv(id_priv);
2030                 break;
2031         case RDMA_TRANSPORT_IWARP:
2032                 ret = cma_accept_iw(id_priv, conn_param);
2033                 break;
2034         default:
2035                 ret = -ENOSYS;
2036                 break;
2037         }
2038
2039         if (ret)
2040                 goto reject;
2041
2042         return 0;
2043 reject:
2044         cma_modify_qp_err(id);
2045         rdma_reject(id, NULL, 0);
2046         return ret;
2047 }
2048 EXPORT_SYMBOL(rdma_accept);
2049
2050 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2051                 u8 private_data_len)
2052 {
2053         struct rdma_id_private *id_priv;
2054         int ret;
2055
2056         id_priv = container_of(id, struct rdma_id_private, id);
2057         if (!cma_comp(id_priv, CMA_CONNECT))
2058                 return -EINVAL;
2059
2060         switch (rdma_node_get_transport(id->device->node_type)) {
2061         case RDMA_TRANSPORT_IB:
2062                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2063                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2064                                      private_data, private_data_len);
2065                 break;
2066         case RDMA_TRANSPORT_IWARP:
2067                 ret = iw_cm_reject(id_priv->cm_id.iw,
2068                                    private_data, private_data_len);
2069                 break;
2070         default:
2071                 ret = -ENOSYS;
2072                 break;
2073         }
2074         return ret;
2075 }
2076 EXPORT_SYMBOL(rdma_reject);
2077
2078 int rdma_disconnect(struct rdma_cm_id *id)
2079 {
2080         struct rdma_id_private *id_priv;
2081         int ret;
2082
2083         id_priv = container_of(id, struct rdma_id_private, id);
2084         if (!cma_comp(id_priv, CMA_CONNECT) &&
2085             !cma_comp(id_priv, CMA_DISCONNECT))
2086                 return -EINVAL;
2087
2088         switch (rdma_node_get_transport(id->device->node_type)) {
2089         case RDMA_TRANSPORT_IB:
2090                 ret = cma_modify_qp_err(id);
2091                 if (ret)
2092                         goto out;
2093                 /* Initiate or respond to a disconnect. */
2094                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2095                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2096                 break;
2097         case RDMA_TRANSPORT_IWARP:
2098                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2099                 break;
2100         default:
2101                 ret = -EINVAL;
2102                 break;
2103         }
2104 out:
2105         return ret;
2106 }
2107 EXPORT_SYMBOL(rdma_disconnect);
2108
2109 static void cma_add_one(struct ib_device *device)
2110 {
2111         struct cma_device *cma_dev;
2112         struct rdma_id_private *id_priv;
2113
2114         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2115         if (!cma_dev)
2116                 return;
2117
2118         cma_dev->device = device;
2119         cma_dev->node_guid = device->node_guid;
2120
2121         init_completion(&cma_dev->comp);
2122         atomic_set(&cma_dev->refcount, 1);
2123         INIT_LIST_HEAD(&cma_dev->id_list);
2124         ib_set_client_data(device, &cma_client, cma_dev);
2125
2126         mutex_lock(&lock);
2127         list_add_tail(&cma_dev->list, &dev_list);
2128         list_for_each_entry(id_priv, &listen_any_list, list)
2129                 cma_listen_on_dev(id_priv, cma_dev);
2130         mutex_unlock(&lock);
2131 }
2132
2133 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2134 {
2135         enum cma_state state;
2136
2137         /* Record that we want to remove the device */
2138         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2139         if (state == CMA_DESTROYING)
2140                 return 0;
2141
2142         cma_cancel_operation(id_priv, state);
2143         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2144
2145         /* Check for destruction from another callback. */
2146         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2147                 return 0;
2148
2149         return cma_notify_user(id_priv, RDMA_CM_EVENT_DEVICE_REMOVAL,
2150                                0, NULL, 0);
2151 }
2152
2153 static void cma_process_remove(struct cma_device *cma_dev)
2154 {
2155         struct rdma_id_private *id_priv;
2156         int ret;
2157
2158         mutex_lock(&lock);
2159         while (!list_empty(&cma_dev->id_list)) {
2160                 id_priv = list_entry(cma_dev->id_list.next,
2161                                      struct rdma_id_private, list);
2162
2163                 if (cma_internal_listen(id_priv)) {
2164                         cma_destroy_listen(id_priv);
2165                         continue;
2166                 }
2167
2168                 list_del_init(&id_priv->list);
2169                 atomic_inc(&id_priv->refcount);
2170                 mutex_unlock(&lock);
2171
2172                 ret = cma_remove_id_dev(id_priv);
2173                 cma_deref_id(id_priv);
2174                 if (ret)
2175                         rdma_destroy_id(&id_priv->id);
2176
2177                 mutex_lock(&lock);
2178         }
2179         mutex_unlock(&lock);
2180
2181         cma_deref_dev(cma_dev);
2182         wait_for_completion(&cma_dev->comp);
2183 }
2184
2185 static void cma_remove_one(struct ib_device *device)
2186 {
2187         struct cma_device *cma_dev;
2188
2189         cma_dev = ib_get_client_data(device, &cma_client);
2190         if (!cma_dev)
2191                 return;
2192
2193         mutex_lock(&lock);
2194         list_del(&cma_dev->list);
2195         mutex_unlock(&lock);
2196
2197         cma_process_remove(cma_dev);
2198         kfree(cma_dev);
2199 }
2200
2201 static int cma_init(void)
2202 {
2203         int ret;
2204
2205         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2206         if (!cma_wq)
2207                 return -ENOMEM;
2208
2209         ib_sa_register_client(&sa_client);
2210         rdma_addr_register_client(&addr_client);
2211
2212         ret = ib_register_client(&cma_client);
2213         if (ret)
2214                 goto err;
2215         return 0;
2216
2217 err:
2218         rdma_addr_unregister_client(&addr_client);
2219         ib_sa_unregister_client(&sa_client);
2220         destroy_workqueue(cma_wq);
2221         return ret;
2222 }
2223
2224 static void cma_cleanup(void)
2225 {
2226         ib_unregister_client(&cma_client);
2227         rdma_addr_unregister_client(&addr_client);
2228         ib_sa_unregister_client(&sa_client);
2229         destroy_workqueue(cma_wq);
2230         idr_destroy(&sdp_ps);
2231         idr_destroy(&tcp_ps);
2232 }
2233
2234 module_init(cma_init);
2235 module_exit(cma_cleanup);