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 void cma_cancel_route(struct rdma_id_private *id_priv)
 596 {
 597         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 598         case RDMA_TRANSPORT_IB:
 599                 if (id_priv->query)
 600                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 601                 break;
 602         default:
 603                 break;
 604         }
 605 }
 606
 607 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 608 {
 609         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 610                cma_any_addr(&id_priv->id.route.addr.src_addr);
 611 }
 612
 613 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 614 {
 615         cma_exch(id_priv, CMA_DESTROYING);
 616
 617         if (id_priv->cma_dev) {
 618                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 619                 case RDMA_TRANSPORT_IB:
 620                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 621                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 622                         break;
 623                 case RDMA_TRANSPORT_IWARP:
 624                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 625                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 626                         break;
 627                 default:
 628                         break;
 629                 }
 630                 cma_detach_from_dev(id_priv);
 631         }
 632         list_del(&id_priv->listen_list);
 633
 634         cma_deref_id(id_priv);
 635         wait_for_completion(&id_priv->comp);
 636
 637         kfree(id_priv);
 638 }
 639
 640 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 641 {
 642         struct rdma_id_private *dev_id_priv;
 643
 644         mutex_lock(&lock);
 645         list_del(&id_priv->list);
 646
 647         while (!list_empty(&id_priv->listen_list)) {
 648                 dev_id_priv = list_entry(id_priv->listen_list.next,
 649                                          struct rdma_id_private, listen_list);
 650                 cma_destroy_listen(dev_id_priv);
 651         }
 652         mutex_unlock(&lock);
 653 }
 654
 655 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 656                                  enum cma_state state)
 657 {
 658         switch (state) {
 659         case CMA_ADDR_QUERY:
 660                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 661                 break;
 662         case CMA_ROUTE_QUERY:
 663                 cma_cancel_route(id_priv);
 664                 break;
 665         case CMA_LISTEN:
 666                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 667                     !id_priv->cma_dev)
 668                         cma_cancel_listens(id_priv);
 669                 break;
 670         default:
 671                 break;
 672         }
 673 }
 674
 675 static void cma_release_port(struct rdma_id_private *id_priv)
 676 {
 677         struct rdma_bind_list *bind_list = id_priv->bind_list;
 678
 679         if (!bind_list)
 680                 return;
 681
 682         mutex_lock(&lock);
 683         hlist_del(&id_priv->node);
 684         if (hlist_empty(&bind_list->owners)) {
 685                 idr_remove(bind_list->ps, bind_list->port);
 686                 kfree(bind_list);
 687         }
 688         mutex_unlock(&lock);
 689 }
 690
 691 void rdma_destroy_id(struct rdma_cm_id *id)
 692 {
 693         struct rdma_id_private *id_priv;
 694         enum cma_state state;
 695
 696         id_priv = container_of(id, struct rdma_id_private, id);
 697         state = cma_exch(id_priv, CMA_DESTROYING);
 698         cma_cancel_operation(id_priv, state);
 699
 700         mutex_lock(&lock);
 701         if (id_priv->cma_dev) {
 702                 mutex_unlock(&lock);
 703                 switch (rdma_node_get_transport(id->device->node_type)) {
 704                 case RDMA_TRANSPORT_IB:
 705                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 706                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 707                         break;
 708                 case RDMA_TRANSPORT_IWARP:
 709                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 710                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 711                         break;
 712                 default:
 713                         break;
 714                 }
 715                 mutex_lock(&lock);
 716                 cma_detach_from_dev(id_priv);
 717         }
 718         mutex_unlock(&lock);
 719
 720         cma_release_port(id_priv);
 721         cma_deref_id(id_priv);
 722         wait_for_completion(&id_priv->comp);
 723
 724         kfree(id_priv->id.route.path_rec);
 725         kfree(id_priv);
 726 }
 727 EXPORT_SYMBOL(rdma_destroy_id);
 728
 729 static int cma_rep_recv(struct rdma_id_private *id_priv)
 730 {
 731         int ret;
 732
 733         ret = cma_modify_qp_rtr(&id_priv->id);
 734         if (ret)
 735                 goto reject;
 736
 737         ret = cma_modify_qp_rts(&id_priv->id);
 738         if (ret)
 739                 goto reject;
 740
 741         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 742         if (ret)
 743                 goto reject;
 744
 745         return 0;
 746 reject:
 747         cma_modify_qp_err(&id_priv->id);
 748         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 749                        NULL, 0, NULL, 0);
 750         return ret;
 751 }
 752
 753 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 754 {
 755         if (id_priv->id.ps == RDMA_PS_SDP &&
 756             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 757             SDP_MAJ_VERSION)
 758                 return -EINVAL;
 759
 760         return 0;
 761 }
 762
 763 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 764                                    struct ib_cm_rep_event_param *rep_data,
 765                                    void *private_data)
 766 {
 767         event->param.conn.private_data = private_data;
 768         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 769         event->param.conn.responder_resources = rep_data->responder_resources;
 770         event->param.conn.initiator_depth = rep_data->initiator_depth;
 771         event->param.conn.flow_control = rep_data->flow_control;
 772         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 773         event->param.conn.srq = rep_data->srq;
 774         event->param.conn.qp_num = rep_data->remote_qpn;
 775 }
 776
 777 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 778 {
 779         struct rdma_id_private *id_priv = cm_id->context;
 780         struct rdma_cm_event event;
 781         int ret = 0;
 782
 783         atomic_inc(&id_priv->dev_remove);
 784         if (!cma_comp(id_priv, CMA_CONNECT))
 785                 goto out;
 786
 787         memset(&event, 0, sizeof event);
 788         switch (ib_event->event) {
 789         case IB_CM_REQ_ERROR:
 790         case IB_CM_REP_ERROR:
 791                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 792                 event.status = -ETIMEDOUT;
 793                 break;
 794         case IB_CM_REP_RECEIVED:
 795                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 796                 if (event.status)
 797                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 798                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 799                         event.status = cma_rep_recv(id_priv);
 800                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 801                                                      RDMA_CM_EVENT_ESTABLISHED;
 802                 } else
 803                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 804                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 805                                        ib_event->private_data);
 806                 break;
 807         case IB_CM_RTU_RECEIVED:
 808         case IB_CM_USER_ESTABLISHED:
 809                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 810                 break;
 811         case IB_CM_DREQ_ERROR:
 812                 event.status = -ETIMEDOUT; /* fall through */
 813         case IB_CM_DREQ_RECEIVED:
 814         case IB_CM_DREP_RECEIVED:
 815                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 816                         goto out;
 817                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 818                 break;
 819         case IB_CM_TIMEWAIT_EXIT:
 820         case IB_CM_MRA_RECEIVED:
 821                 /* ignore event */
 822                 goto out;
 823         case IB_CM_REJ_RECEIVED:
 824                 cma_modify_qp_err(&id_priv->id);
 825                 event.status = ib_event->param.rej_rcvd.reason;
 826                 event.event = RDMA_CM_EVENT_REJECTED;
 827                 event.param.conn.private_data = ib_event->private_data;
 828                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 829                 break;
 830         default:
 831                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 832                        ib_event->event);
 833                 goto out;
 834         }
 835
 836         ret = id_priv->id.event_handler(&id_priv->id, &event);
 837         if (ret) {
 838                 /* Destroy the CM ID by returning a non-zero value. */
 839                 id_priv->cm_id.ib = NULL;
 840                 cma_exch(id_priv, CMA_DESTROYING);
 841                 cma_release_remove(id_priv);
 842                 rdma_destroy_id(&id_priv->id);
 843                 return ret;
 844         }
 845 out:
 846         cma_release_remove(id_priv);
 847         return ret;
 848 }
 849
 850 static struct rdma_id_private *cma_new_id(struct rdma_cm_id *listen_id,
 851                                           struct ib_cm_event *ib_event)
 852 {
 853         struct rdma_id_private *id_priv;
 854         struct rdma_cm_id *id;
 855         struct rdma_route *rt;
 856         union cma_ip_addr *src, *dst;
 857         __u16 port;
 858         u8 ip_ver;
 859
 860         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
 861                              &ip_ver, &port, &src, &dst))
 862                 goto err;
 863
 864         id = rdma_create_id(listen_id->event_handler, listen_id->context,
 865                             listen_id->ps);
 866         if (IS_ERR(id))
 867                 goto err;
 868
 869         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
 870                           ip_ver, port, src, dst);
 871
 872         rt = &id->route;
 873         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
 874         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
 875                                GFP_KERNEL);
 876         if (!rt->path_rec)
 877                 goto destroy_id;
 878
 879         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
 880         if (rt->num_paths == 2)
 881                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
 882
 883         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
 884         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
 885         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
 886         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
 887
 888         id_priv = container_of(id, struct rdma_id_private, id);
 889         id_priv->state = CMA_CONNECT;
 890         return id_priv;
 891
 892 destroy_id:
 893         rdma_destroy_id(id);
 894 err:
 895         return NULL;
 896 }
 897
 898 static void cma_set_req_event_data(struct rdma_cm_event *event,
 899                                    struct ib_cm_req_event_param *req_data,
 900                                    void *private_data, int offset)
 901 {
 902         event->param.conn.private_data = private_data + offset;
 903         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
 904         event->param.conn.responder_resources = req_data->responder_resources;
 905         event->param.conn.initiator_depth = req_data->initiator_depth;
 906         event->param.conn.flow_control = req_data->flow_control;
 907         event->param.conn.retry_count = req_data->retry_count;
 908         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
 909         event->param.conn.srq = req_data->srq;
 910         event->param.conn.qp_num = req_data->remote_qpn;
 911 }
 912
 913 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 914 {
 915         struct rdma_id_private *listen_id, *conn_id;
 916         struct rdma_cm_event event;
 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         memset(&event, 0, sizeof event);
 945         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
 946         cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
 947                                ib_event->private_data, offset);
 948         ret = conn_id->id.event_handler(&conn_id->id, &event);
 949         if (!ret)
 950                 goto out;
 951
 952         /* Destroy the CM ID by returning a non-zero value. */
 953         conn_id->cm_id.ib = NULL;
 954
 955 release_conn_id:
 956         cma_exch(conn_id, CMA_DESTROYING);
 957         cma_release_remove(conn_id);
 958         rdma_destroy_id(&conn_id->id);
 959
 960 out:
 961         cma_release_remove(listen_id);
 962         return ret;
 963 }
 964
 965 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
 966 {
 967         return cpu_to_be64(((u64)ps << 16) +
 968                be16_to_cpu(((struct sockaddr_in *) addr)->sin_port));
 969 }
 970
 971 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
 972                                  struct ib_cm_compare_data *compare)
 973 {
 974         struct cma_hdr *cma_data, *cma_mask;
 975         struct sdp_hh *sdp_data, *sdp_mask;
 976         __u32 ip4_addr;
 977         struct in6_addr ip6_addr;
 978
 979         memset(compare, 0, sizeof *compare);
 980         cma_data = (void *) compare->data;
 981         cma_mask = (void *) compare->mask;
 982         sdp_data = (void *) compare->data;
 983         sdp_mask = (void *) compare->mask;
 984
 985         switch (addr->sa_family) {
 986         case AF_INET:
 987                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
 988                 if (ps == RDMA_PS_SDP) {
 989                         sdp_set_ip_ver(sdp_data, 4);
 990                         sdp_set_ip_ver(sdp_mask, 0xF);
 991                         sdp_data->dst_addr.ip4.addr = ip4_addr;
 992                         sdp_mask->dst_addr.ip4.addr = ~0;
 993                 } else {
 994                         cma_set_ip_ver(cma_data, 4);
 995                         cma_set_ip_ver(cma_mask, 0xF);
 996                         cma_data->dst_addr.ip4.addr = ip4_addr;
 997                         cma_mask->dst_addr.ip4.addr = ~0;
 998                 }
 999                 break;
1000         case AF_INET6:
1001                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1002                 if (ps == RDMA_PS_SDP) {
1003                         sdp_set_ip_ver(sdp_data, 6);
1004                         sdp_set_ip_ver(sdp_mask, 0xF);
1005                         sdp_data->dst_addr.ip6 = ip6_addr;
1006                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1007                                sizeof sdp_mask->dst_addr.ip6);
1008                 } else {
1009                         cma_set_ip_ver(cma_data, 6);
1010                         cma_set_ip_ver(cma_mask, 0xF);
1011                         cma_data->dst_addr.ip6 = ip6_addr;
1012                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1013                                sizeof cma_mask->dst_addr.ip6);
1014                 }
1015                 break;
1016         default:
1017                 break;
1018         }
1019 }
1020
1021 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1022 {
1023         struct rdma_id_private *id_priv = iw_id->context;
1024         struct rdma_cm_event event;
1025         struct sockaddr_in *sin;
1026         int ret = 0;
1027
1028         memset(&event, 0, sizeof event);
1029         atomic_inc(&id_priv->dev_remove);
1030
1031         switch (iw_event->event) {
1032         case IW_CM_EVENT_CLOSE:
1033                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1034                 break;
1035         case IW_CM_EVENT_CONNECT_REPLY:
1036                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1037                 *sin = iw_event->local_addr;
1038                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1039                 *sin = iw_event->remote_addr;
1040                 if (iw_event->status)
1041                         event.event = RDMA_CM_EVENT_REJECTED;
1042                 else
1043                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1044                 break;
1045         case IW_CM_EVENT_ESTABLISHED:
1046                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1047                 break;
1048         default:
1049                 BUG_ON(1);
1050         }
1051
1052         event.status = iw_event->status;
1053         event.param.conn.private_data = iw_event->private_data;
1054         event.param.conn.private_data_len = iw_event->private_data_len;
1055         ret = id_priv->id.event_handler(&id_priv->id, &event);
1056         if (ret) {
1057                 /* Destroy the CM ID by returning a non-zero value. */
1058                 id_priv->cm_id.iw = NULL;
1059                 cma_exch(id_priv, CMA_DESTROYING);
1060                 cma_release_remove(id_priv);
1061                 rdma_destroy_id(&id_priv->id);
1062                 return ret;
1063         }
1064
1065         cma_release_remove(id_priv);
1066         return ret;
1067 }
1068
1069 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1070                                struct iw_cm_event *iw_event)
1071 {
1072         struct rdma_cm_id *new_cm_id;
1073         struct rdma_id_private *listen_id, *conn_id;
1074         struct sockaddr_in *sin;
1075         struct net_device *dev = NULL;
1076         struct rdma_cm_event event;
1077         int ret;
1078
1079         listen_id = cm_id->context;
1080         atomic_inc(&listen_id->dev_remove);
1081         if (!cma_comp(listen_id, CMA_LISTEN)) {
1082                 ret = -ECONNABORTED;
1083                 goto out;
1084         }
1085
1086         /* Create a new RDMA id for the new IW CM ID */
1087         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1088                                    listen_id->id.context,
1089                                    RDMA_PS_TCP);
1090         if (!new_cm_id) {
1091                 ret = -ENOMEM;
1092                 goto out;
1093         }
1094         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1095         atomic_inc(&conn_id->dev_remove);
1096         conn_id->state = CMA_CONNECT;
1097
1098         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1099         if (!dev) {
1100                 ret = -EADDRNOTAVAIL;
1101                 cma_release_remove(conn_id);
1102                 rdma_destroy_id(new_cm_id);
1103                 goto out;
1104         }
1105         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1106         if (ret) {
1107                 cma_release_remove(conn_id);
1108                 rdma_destroy_id(new_cm_id);
1109                 goto out;
1110         }
1111
1112         mutex_lock(&lock);
1113         ret = cma_acquire_dev(conn_id);
1114         mutex_unlock(&lock);
1115         if (ret) {
1116                 cma_release_remove(conn_id);
1117                 rdma_destroy_id(new_cm_id);
1118                 goto out;
1119         }
1120
1121         conn_id->cm_id.iw = cm_id;
1122         cm_id->context = conn_id;
1123         cm_id->cm_handler = cma_iw_handler;
1124
1125         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1126         *sin = iw_event->local_addr;
1127         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1128         *sin = iw_event->remote_addr;
1129
1130         memset(&event, 0, sizeof event);
1131         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1132         event.param.conn.private_data = iw_event->private_data;
1133         event.param.conn.private_data_len = iw_event->private_data_len;
1134         ret = conn_id->id.event_handler(&conn_id->id, &event);
1135         if (ret) {
1136                 /* User wants to destroy the CM ID */
1137                 conn_id->cm_id.iw = NULL;
1138                 cma_exch(conn_id, CMA_DESTROYING);
1139                 cma_release_remove(conn_id);
1140                 rdma_destroy_id(&conn_id->id);
1141         }
1142
1143 out:
1144         if (dev)
1145                 dev_put(dev);
1146         cma_release_remove(listen_id);
1147         return ret;
1148 }
1149
1150 static int cma_ib_listen(struct rdma_id_private *id_priv)
1151 {
1152         struct ib_cm_compare_data compare_data;
1153         struct sockaddr *addr;
1154         __be64 svc_id;
1155         int ret;
1156
1157         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1158                                             id_priv);
1159         if (IS_ERR(id_priv->cm_id.ib))
1160                 return PTR_ERR(id_priv->cm_id.ib);
1161
1162         addr = &id_priv->id.route.addr.src_addr;
1163         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1164         if (cma_any_addr(addr))
1165                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1166         else {
1167                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1168                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1169         }
1170
1171         if (ret) {
1172                 ib_destroy_cm_id(id_priv->cm_id.ib);
1173                 id_priv->cm_id.ib = NULL;
1174         }
1175
1176         return ret;
1177 }
1178
1179 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1180 {
1181         int ret;
1182         struct sockaddr_in *sin;
1183
1184         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1185                                             iw_conn_req_handler,
1186                                             id_priv);
1187         if (IS_ERR(id_priv->cm_id.iw))
1188                 return PTR_ERR(id_priv->cm_id.iw);
1189
1190         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1191         id_priv->cm_id.iw->local_addr = *sin;
1192
1193         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1194
1195         if (ret) {
1196                 iw_destroy_cm_id(id_priv->cm_id.iw);
1197                 id_priv->cm_id.iw = NULL;
1198         }
1199
1200         return ret;
1201 }
1202
1203 static int cma_listen_handler(struct rdma_cm_id *id,
1204                               struct rdma_cm_event *event)
1205 {
1206         struct rdma_id_private *id_priv = id->context;
1207
1208         id->context = id_priv->id.context;
1209         id->event_handler = id_priv->id.event_handler;
1210         return id_priv->id.event_handler(id, event);
1211 }
1212
1213 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1214                               struct cma_device *cma_dev)
1215 {
1216         struct rdma_id_private *dev_id_priv;
1217         struct rdma_cm_id *id;
1218         int ret;
1219
1220         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1221         if (IS_ERR(id))
1222                 return;
1223
1224         dev_id_priv = container_of(id, struct rdma_id_private, id);
1225
1226         dev_id_priv->state = CMA_ADDR_BOUND;
1227         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1228                ip_addr_size(&id_priv->id.route.addr.src_addr));
1229
1230         cma_attach_to_dev(dev_id_priv, cma_dev);
1231         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1232
1233         ret = rdma_listen(id, id_priv->backlog);
1234         if (ret)
1235                 goto err;
1236
1237         return;
1238 err:
1239         cma_destroy_listen(dev_id_priv);
1240 }
1241
1242 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1243 {
1244         struct cma_device *cma_dev;
1245
1246         mutex_lock(&lock);
1247         list_add_tail(&id_priv->list, &listen_any_list);
1248         list_for_each_entry(cma_dev, &dev_list, list)
1249                 cma_listen_on_dev(id_priv, cma_dev);
1250         mutex_unlock(&lock);
1251 }
1252
1253 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1254 {
1255         struct sockaddr_in addr_in;
1256
1257         memset(&addr_in, 0, sizeof addr_in);
1258         addr_in.sin_family = af;
1259         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1260 }
1261
1262 int rdma_listen(struct rdma_cm_id *id, int backlog)
1263 {
1264         struct rdma_id_private *id_priv;
1265         int ret;
1266
1267         id_priv = container_of(id, struct rdma_id_private, id);
1268         if (id_priv->state == CMA_IDLE) {
1269                 ret = cma_bind_any(id, AF_INET);
1270                 if (ret)
1271                         return ret;
1272         }
1273
1274         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1275                 return -EINVAL;
1276
1277         id_priv->backlog = backlog;
1278         if (id->device) {
1279                 switch (rdma_node_get_transport(id->device->node_type)) {
1280                 case RDMA_TRANSPORT_IB:
1281                         ret = cma_ib_listen(id_priv);
1282                         if (ret)
1283                                 goto err;
1284                         break;
1285                 case RDMA_TRANSPORT_IWARP:
1286                         ret = cma_iw_listen(id_priv, backlog);
1287                         if (ret)
1288                                 goto err;
1289                         break;
1290                 default:
1291                         ret = -ENOSYS;
1292                         goto err;
1293                 }
1294         } else
1295                 cma_listen_on_all(id_priv);
1296
1297         return 0;
1298 err:
1299         id_priv->backlog = 0;
1300         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1301         return ret;
1302 }
1303 EXPORT_SYMBOL(rdma_listen);
1304
1305 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1306                               void *context)
1307 {
1308         struct cma_work *work = context;
1309         struct rdma_route *route;
1310
1311         route = &work->id->id.route;
1312
1313         if (!status) {
1314                 route->num_paths = 1;
1315                 *route->path_rec = *path_rec;
1316         } else {
1317                 work->old_state = CMA_ROUTE_QUERY;
1318                 work->new_state = CMA_ADDR_RESOLVED;
1319                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1320                 work->event.status = status;
1321         }
1322
1323         queue_work(cma_wq, &work->work);
1324 }
1325
1326 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1327                               struct cma_work *work)
1328 {
1329         struct rdma_dev_addr *addr = &id_priv->id.route.addr.dev_addr;
1330         struct ib_sa_path_rec path_rec;
1331
1332         memset(&path_rec, 0, sizeof path_rec);
1333         ib_addr_get_sgid(addr, &path_rec.sgid);
1334         ib_addr_get_dgid(addr, &path_rec.dgid);
1335         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(addr));
1336         path_rec.numb_path = 1;
1337
1338         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1339                                 id_priv->id.port_num, &path_rec,
1340                                 IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1341                                 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH,
1342                                 timeout_ms, GFP_KERNEL,
1343                                 cma_query_handler, work, &id_priv->query);
1344
1345         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1346 }
1347
1348 static void cma_work_handler(struct work_struct *_work)
1349 {
1350         struct cma_work *work = container_of(_work, struct cma_work, work);
1351         struct rdma_id_private *id_priv = work->id;
1352         int destroy = 0;
1353
1354         atomic_inc(&id_priv->dev_remove);
1355         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1356                 goto out;
1357
1358         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1359                 cma_exch(id_priv, CMA_DESTROYING);
1360                 destroy = 1;
1361         }
1362 out:
1363         cma_release_remove(id_priv);
1364         cma_deref_id(id_priv);
1365         if (destroy)
1366                 rdma_destroy_id(&id_priv->id);
1367         kfree(work);
1368 }
1369
1370 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1371 {
1372         struct rdma_route *route = &id_priv->id.route;
1373         struct cma_work *work;
1374         int ret;
1375
1376         work = kzalloc(sizeof *work, GFP_KERNEL);
1377         if (!work)
1378                 return -ENOMEM;
1379
1380         work->id = id_priv;
1381         INIT_WORK(&work->work, cma_work_handler);
1382         work->old_state = CMA_ROUTE_QUERY;
1383         work->new_state = CMA_ROUTE_RESOLVED;
1384         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1385
1386         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1387         if (!route->path_rec) {
1388                 ret = -ENOMEM;
1389                 goto err1;
1390         }
1391
1392         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1393         if (ret)
1394                 goto err2;
1395
1396         return 0;
1397 err2:
1398         kfree(route->path_rec);
1399         route->path_rec = NULL;
1400 err1:
1401         kfree(work);
1402         return ret;
1403 }
1404
1405 int rdma_set_ib_paths(struct rdma_cm_id *id,
1406                       struct ib_sa_path_rec *path_rec, int num_paths)
1407 {
1408         struct rdma_id_private *id_priv;
1409         int ret;
1410
1411         id_priv = container_of(id, struct rdma_id_private, id);
1412         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1413                 return -EINVAL;
1414
1415         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1416         if (!id->route.path_rec) {
1417                 ret = -ENOMEM;
1418                 goto err;
1419         }
1420
1421         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1422         return 0;
1423 err:
1424         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1425         return ret;
1426 }
1427 EXPORT_SYMBOL(rdma_set_ib_paths);
1428
1429 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1430 {
1431         struct cma_work *work;
1432
1433         work = kzalloc(sizeof *work, GFP_KERNEL);
1434         if (!work)
1435                 return -ENOMEM;
1436
1437         work->id = id_priv;
1438         INIT_WORK(&work->work, cma_work_handler);
1439         work->old_state = CMA_ROUTE_QUERY;
1440         work->new_state = CMA_ROUTE_RESOLVED;
1441         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1442         queue_work(cma_wq, &work->work);
1443         return 0;
1444 }
1445
1446 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1447 {
1448         struct rdma_id_private *id_priv;
1449         int ret;
1450
1451         id_priv = container_of(id, struct rdma_id_private, id);
1452         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1453                 return -EINVAL;
1454
1455         atomic_inc(&id_priv->refcount);
1456         switch (rdma_node_get_transport(id->device->node_type)) {
1457         case RDMA_TRANSPORT_IB:
1458                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1459                 break;
1460         case RDMA_TRANSPORT_IWARP:
1461                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1462                 break;
1463         default:
1464                 ret = -ENOSYS;
1465                 break;
1466         }
1467         if (ret)
1468                 goto err;
1469
1470         return 0;
1471 err:
1472         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1473         cma_deref_id(id_priv);
1474         return ret;
1475 }
1476 EXPORT_SYMBOL(rdma_resolve_route);
1477
1478 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1479 {
1480         struct cma_device *cma_dev;
1481         struct ib_port_attr port_attr;
1482         union ib_gid gid;
1483         u16 pkey;
1484         int ret;
1485         u8 p;
1486
1487         mutex_lock(&lock);
1488         if (list_empty(&dev_list)) {
1489                 ret = -ENODEV;
1490                 goto out;
1491         }
1492         list_for_each_entry(cma_dev, &dev_list, list)
1493                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1494                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1495                             port_attr.state == IB_PORT_ACTIVE)
1496                                 goto port_found;
1497
1498         p = 1;
1499         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1500
1501 port_found:
1502         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1503         if (ret)
1504                 goto out;
1505
1506         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1507         if (ret)
1508                 goto out;
1509
1510         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1511         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1512         id_priv->id.port_num = p;
1513         cma_attach_to_dev(id_priv, cma_dev);
1514 out:
1515         mutex_unlock(&lock);
1516         return ret;
1517 }
1518
1519 static void addr_handler(int status, struct sockaddr *src_addr,
1520                          struct rdma_dev_addr *dev_addr, void *context)
1521 {
1522         struct rdma_id_private *id_priv = context;
1523         struct rdma_cm_event event;
1524
1525         memset(&event, 0, sizeof event);
1526         atomic_inc(&id_priv->dev_remove);
1527
1528         /*
1529          * Grab mutex to block rdma_destroy_id() from removing the device while
1530          * we're trying to acquire it.
1531          */
1532         mutex_lock(&lock);
1533         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1534                 mutex_unlock(&lock);
1535                 goto out;
1536         }
1537
1538         if (!status && !id_priv->cma_dev)
1539                 status = cma_acquire_dev(id_priv);
1540         mutex_unlock(&lock);
1541
1542         if (status) {
1543                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1544                         goto out;
1545                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1546                 event.status = status;
1547         } else {
1548                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1549                        ip_addr_size(src_addr));
1550                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1551         }
1552
1553         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1554                 cma_exch(id_priv, CMA_DESTROYING);
1555                 cma_release_remove(id_priv);
1556                 cma_deref_id(id_priv);
1557                 rdma_destroy_id(&id_priv->id);
1558                 return;
1559         }
1560 out:
1561         cma_release_remove(id_priv);
1562         cma_deref_id(id_priv);
1563 }
1564
1565 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1566 {
1567         struct cma_work *work;
1568         struct sockaddr_in *src_in, *dst_in;
1569         union ib_gid gid;
1570         int ret;
1571
1572         work = kzalloc(sizeof *work, GFP_KERNEL);
1573         if (!work)
1574                 return -ENOMEM;
1575
1576         if (!id_priv->cma_dev) {
1577                 ret = cma_bind_loopback(id_priv);
1578                 if (ret)
1579                         goto err;
1580         }
1581
1582         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1583         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1584
1585         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1586                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1587                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1588                 src_in->sin_family = dst_in->sin_family;
1589                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1590         }
1591
1592         work->id = id_priv;
1593         INIT_WORK(&work->work, cma_work_handler);
1594         work->old_state = CMA_ADDR_QUERY;
1595         work->new_state = CMA_ADDR_RESOLVED;
1596         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1597         queue_work(cma_wq, &work->work);
1598         return 0;
1599 err:
1600         kfree(work);
1601         return ret;
1602 }
1603
1604 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1605                          struct sockaddr *dst_addr)
1606 {
1607         if (src_addr && src_addr->sa_family)
1608                 return rdma_bind_addr(id, src_addr);
1609         else
1610                 return cma_bind_any(id, dst_addr->sa_family);
1611 }
1612
1613 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1614                       struct sockaddr *dst_addr, int timeout_ms)
1615 {
1616         struct rdma_id_private *id_priv;
1617         int ret;
1618
1619         id_priv = container_of(id, struct rdma_id_private, id);
1620         if (id_priv->state == CMA_IDLE) {
1621                 ret = cma_bind_addr(id, src_addr, dst_addr);
1622                 if (ret)
1623                         return ret;
1624         }
1625
1626         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1627                 return -EINVAL;
1628
1629         atomic_inc(&id_priv->refcount);
1630         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1631         if (cma_any_addr(dst_addr))
1632                 ret = cma_resolve_loopback(id_priv);
1633         else
1634                 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1635                                       dst_addr, &id->route.addr.dev_addr,
1636                                       timeout_ms, addr_handler, id_priv);
1637         if (ret)
1638                 goto err;
1639
1640         return 0;
1641 err:
1642         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1643         cma_deref_id(id_priv);
1644         return ret;
1645 }
1646 EXPORT_SYMBOL(rdma_resolve_addr);
1647
1648 static void cma_bind_port(struct rdma_bind_list *bind_list,
1649                           struct rdma_id_private *id_priv)
1650 {
1651         struct sockaddr_in *sin;
1652
1653         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1654         sin->sin_port = htons(bind_list->port);
1655         id_priv->bind_list = bind_list;
1656         hlist_add_head(&id_priv->node, &bind_list->owners);
1657 }
1658
1659 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1660                           unsigned short snum)
1661 {
1662         struct rdma_bind_list *bind_list;
1663         int port, start, ret;
1664
1665         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1666         if (!bind_list)
1667                 return -ENOMEM;
1668
1669         start = snum ? snum : sysctl_local_port_range[0];
1670
1671         do {
1672                 ret = idr_get_new_above(ps, bind_list, start, &port);
1673         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1674
1675         if (ret)
1676                 goto err;
1677
1678         if ((snum && port != snum) ||
1679             (!snum && port > sysctl_local_port_range[1])) {
1680                 idr_remove(ps, port);
1681                 ret = -EADDRNOTAVAIL;
1682                 goto err;
1683         }
1684
1685         bind_list->ps = ps;
1686         bind_list->port = (unsigned short) port;
1687         cma_bind_port(bind_list, id_priv);
1688         return 0;
1689 err:
1690         kfree(bind_list);
1691         return ret;
1692 }
1693
1694 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1695 {
1696         struct rdma_id_private *cur_id;
1697         struct sockaddr_in *sin, *cur_sin;
1698         struct rdma_bind_list *bind_list;
1699         struct hlist_node *node;
1700         unsigned short snum;
1701
1702         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1703         snum = ntohs(sin->sin_port);
1704         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1705                 return -EACCES;
1706
1707         bind_list = idr_find(ps, snum);
1708         if (!bind_list)
1709                 return cma_alloc_port(ps, id_priv, snum);
1710
1711         /*
1712          * We don't support binding to any address if anyone is bound to
1713          * a specific address on the same port.
1714          */
1715         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1716                 return -EADDRNOTAVAIL;
1717
1718         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1719                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1720                         return -EADDRNOTAVAIL;
1721
1722                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1723                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1724                         return -EADDRINUSE;
1725         }
1726
1727         cma_bind_port(bind_list, id_priv);
1728         return 0;
1729 }
1730
1731 static int cma_get_port(struct rdma_id_private *id_priv)
1732 {
1733         struct idr *ps;
1734         int ret;
1735
1736         switch (id_priv->id.ps) {
1737         case RDMA_PS_SDP:
1738                 ps = &sdp_ps;
1739                 break;
1740         case RDMA_PS_TCP:
1741                 ps = &tcp_ps;
1742                 break;
1743         default:
1744                 return -EPROTONOSUPPORT;
1745         }
1746
1747         mutex_lock(&lock);
1748         if (cma_any_port(&id_priv->id.route.addr.src_addr))
1749                 ret = cma_alloc_port(ps, id_priv, 0);
1750         else
1751                 ret = cma_use_port(ps, id_priv);
1752         mutex_unlock(&lock);
1753
1754         return ret;
1755 }
1756
1757 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
1758 {
1759         struct rdma_id_private *id_priv;
1760         int ret;
1761
1762         if (addr->sa_family != AF_INET)
1763                 return -EAFNOSUPPORT;
1764
1765         id_priv = container_of(id, struct rdma_id_private, id);
1766         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
1767                 return -EINVAL;
1768
1769         if (!cma_any_addr(addr)) {
1770                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
1771                 if (ret)
1772                         goto err1;
1773
1774                 mutex_lock(&lock);
1775                 ret = cma_acquire_dev(id_priv);
1776                 mutex_unlock(&lock);
1777                 if (ret)
1778                         goto err1;
1779         }
1780
1781         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
1782         ret = cma_get_port(id_priv);
1783         if (ret)
1784                 goto err2;
1785
1786         return 0;
1787 err2:
1788         if (!cma_any_addr(addr)) {
1789                 mutex_lock(&lock);
1790                 cma_detach_from_dev(id_priv);
1791                 mutex_unlock(&lock);
1792         }
1793 err1:
1794         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
1795         return ret;
1796 }
1797 EXPORT_SYMBOL(rdma_bind_addr);
1798
1799 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
1800                           struct rdma_route *route)
1801 {
1802         struct sockaddr_in *src4, *dst4;
1803         struct cma_hdr *cma_hdr;
1804         struct sdp_hh *sdp_hdr;
1805
1806         src4 = (struct sockaddr_in *) &route->addr.src_addr;
1807         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
1808
1809         switch (ps) {
1810         case RDMA_PS_SDP:
1811                 sdp_hdr = hdr;
1812                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
1813                         return -EINVAL;
1814                 sdp_set_ip_ver(sdp_hdr, 4);
1815                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1816                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1817                 sdp_hdr->port = src4->sin_port;
1818                 break;
1819         default:
1820                 cma_hdr = hdr;
1821                 cma_hdr->cma_version = CMA_VERSION;
1822                 cma_set_ip_ver(cma_hdr, 4);
1823                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
1824                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
1825                 cma_hdr->port = src4->sin_port;
1826                 break;
1827         }
1828         return 0;
1829 }
1830
1831 static int cma_connect_ib(struct rdma_id_private *id_priv,
1832                           struct rdma_conn_param *conn_param)
1833 {
1834         struct ib_cm_req_param req;
1835         struct rdma_route *route;
1836         void *private_data;
1837         int offset, ret;
1838
1839         memset(&req, 0, sizeof req);
1840         offset = cma_user_data_offset(id_priv->id.ps);
1841         req.private_data_len = offset + conn_param->private_data_len;
1842         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
1843         if (!private_data)
1844                 return -ENOMEM;
1845
1846         if (conn_param->private_data && conn_param->private_data_len)
1847                 memcpy(private_data + offset, conn_param->private_data,
1848                        conn_param->private_data_len);
1849
1850         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
1851                                             id_priv);
1852         if (IS_ERR(id_priv->cm_id.ib)) {
1853                 ret = PTR_ERR(id_priv->cm_id.ib);
1854                 goto out;
1855         }
1856
1857         route = &id_priv->id.route;
1858         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
1859         if (ret)
1860                 goto out;
1861         req.private_data = private_data;
1862
1863         req.primary_path = &route->path_rec[0];
1864         if (route->num_paths == 2)
1865                 req.alternate_path = &route->path_rec[1];
1866
1867         req.service_id = cma_get_service_id(id_priv->id.ps,
1868                                             &route->addr.dst_addr);
1869         req.qp_num = id_priv->qp_num;
1870         req.qp_type = IB_QPT_RC;
1871         req.starting_psn = id_priv->seq_num;
1872         req.responder_resources = conn_param->responder_resources;
1873         req.initiator_depth = conn_param->initiator_depth;
1874         req.flow_control = conn_param->flow_control;
1875         req.retry_count = conn_param->retry_count;
1876         req.rnr_retry_count = conn_param->rnr_retry_count;
1877         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1878         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
1879         req.max_cm_retries = CMA_MAX_CM_RETRIES;
1880         req.srq = id_priv->srq ? 1 : 0;
1881
1882         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
1883 out:
1884         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
1885                 ib_destroy_cm_id(id_priv->cm_id.ib);
1886                 id_priv->cm_id.ib = NULL;
1887         }
1888
1889         kfree(private_data);
1890         return ret;
1891 }
1892
1893 static int cma_connect_iw(struct rdma_id_private *id_priv,
1894                           struct rdma_conn_param *conn_param)
1895 {
1896         struct iw_cm_id *cm_id;
1897         struct sockaddr_in* sin;
1898         int ret;
1899         struct iw_cm_conn_param iw_param;
1900
1901         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
1902         if (IS_ERR(cm_id)) {
1903                 ret = PTR_ERR(cm_id);
1904                 goto out;
1905         }
1906
1907         id_priv->cm_id.iw = cm_id;
1908
1909         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
1910         cm_id->local_addr = *sin;
1911
1912         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
1913         cm_id->remote_addr = *sin;
1914
1915         ret = cma_modify_qp_rtr(&id_priv->id);
1916         if (ret)
1917                 goto out;
1918
1919         iw_param.ord = conn_param->initiator_depth;
1920         iw_param.ird = conn_param->responder_resources;
1921         iw_param.private_data = conn_param->private_data;
1922         iw_param.private_data_len = conn_param->private_data_len;
1923         if (id_priv->id.qp)
1924                 iw_param.qpn = id_priv->qp_num;
1925         else
1926                 iw_param.qpn = conn_param->qp_num;
1927         ret = iw_cm_connect(cm_id, &iw_param);
1928 out:
1929         if (ret && !IS_ERR(cm_id)) {
1930                 iw_destroy_cm_id(cm_id);
1931                 id_priv->cm_id.iw = NULL;
1932         }
1933         return ret;
1934 }
1935
1936 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
1937 {
1938         struct rdma_id_private *id_priv;
1939         int ret;
1940
1941         id_priv = container_of(id, struct rdma_id_private, id);
1942         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
1943                 return -EINVAL;
1944
1945         if (!id->qp) {
1946                 id_priv->qp_num = conn_param->qp_num;
1947                 id_priv->srq = conn_param->srq;
1948         }
1949
1950         switch (rdma_node_get_transport(id->device->node_type)) {
1951         case RDMA_TRANSPORT_IB:
1952                 ret = cma_connect_ib(id_priv, conn_param);
1953                 break;
1954         case RDMA_TRANSPORT_IWARP:
1955                 ret = cma_connect_iw(id_priv, conn_param);
1956                 break;
1957         default:
1958                 ret = -ENOSYS;
1959                 break;
1960         }
1961         if (ret)
1962                 goto err;
1963
1964         return 0;
1965 err:
1966         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
1967         return ret;
1968 }
1969 EXPORT_SYMBOL(rdma_connect);
1970
1971 static int cma_accept_ib(struct rdma_id_private *id_priv,
1972                          struct rdma_conn_param *conn_param)
1973 {
1974         struct ib_cm_rep_param rep;
1975         struct ib_qp_attr qp_attr;
1976         int qp_attr_mask, ret;
1977
1978         if (id_priv->id.qp) {
1979                 ret = cma_modify_qp_rtr(&id_priv->id);
1980                 if (ret)
1981                         goto out;
1982
1983                 qp_attr.qp_state = IB_QPS_RTS;
1984                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr,
1985                                          &qp_attr_mask);
1986                 if (ret)
1987                         goto out;
1988
1989                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1990                 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1991                 if (ret)
1992                         goto out;
1993         }
1994
1995         memset(&rep, 0, sizeof rep);
1996         rep.qp_num = id_priv->qp_num;
1997         rep.starting_psn = id_priv->seq_num;
1998         rep.private_data = conn_param->private_data;
1999         rep.private_data_len = conn_param->private_data_len;
2000         rep.responder_resources = conn_param->responder_resources;
2001         rep.initiator_depth = conn_param->initiator_depth;
2002         rep.target_ack_delay = CMA_CM_RESPONSE_TIMEOUT;
2003         rep.failover_accepted = 0;
2004         rep.flow_control = conn_param->flow_control;
2005         rep.rnr_retry_count = conn_param->rnr_retry_count;
2006         rep.srq = id_priv->srq ? 1 : 0;
2007
2008         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2009 out:
2010         return ret;
2011 }
2012
2013 static int cma_accept_iw(struct rdma_id_private *id_priv,
2014                   struct rdma_conn_param *conn_param)
2015 {
2016         struct iw_cm_conn_param iw_param;
2017         int ret;
2018
2019         ret = cma_modify_qp_rtr(&id_priv->id);
2020         if (ret)
2021                 return ret;
2022
2023         iw_param.ord = conn_param->initiator_depth;
2024         iw_param.ird = conn_param->responder_resources;
2025         iw_param.private_data = conn_param->private_data;
2026         iw_param.private_data_len = conn_param->private_data_len;
2027         if (id_priv->id.qp) {
2028                 iw_param.qpn = id_priv->qp_num;
2029         } else
2030                 iw_param.qpn = conn_param->qp_num;
2031
2032         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2033 }
2034
2035 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2036 {
2037         struct rdma_id_private *id_priv;
2038         int ret;
2039
2040         id_priv = container_of(id, struct rdma_id_private, id);
2041         if (!cma_comp(id_priv, CMA_CONNECT))
2042                 return -EINVAL;
2043
2044         if (!id->qp && conn_param) {
2045                 id_priv->qp_num = conn_param->qp_num;
2046                 id_priv->srq = conn_param->srq;
2047         }
2048
2049         switch (rdma_node_get_transport(id->device->node_type)) {
2050         case RDMA_TRANSPORT_IB:
2051                 if (conn_param)
2052                         ret = cma_accept_ib(id_priv, conn_param);
2053                 else
2054                         ret = cma_rep_recv(id_priv);
2055                 break;
2056         case RDMA_TRANSPORT_IWARP:
2057                 ret = cma_accept_iw(id_priv, conn_param);
2058                 break;
2059         default:
2060                 ret = -ENOSYS;
2061                 break;
2062         }
2063
2064         if (ret)
2065                 goto reject;
2066
2067         return 0;
2068 reject:
2069         cma_modify_qp_err(id);
2070         rdma_reject(id, NULL, 0);
2071         return ret;
2072 }
2073 EXPORT_SYMBOL(rdma_accept);
2074
2075 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2076 {
2077         struct rdma_id_private *id_priv;
2078         int ret;
2079
2080         id_priv = container_of(id, struct rdma_id_private, id);
2081         if (!cma_comp(id_priv, CMA_CONNECT))
2082                 return -EINVAL;
2083
2084         switch (id->device->node_type) {
2085         case RDMA_NODE_IB_CA:
2086                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2087                 break;
2088         default:
2089                 ret = 0;
2090                 break;
2091         }
2092         return ret;
2093 }
2094 EXPORT_SYMBOL(rdma_notify);
2095
2096 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2097                 u8 private_data_len)
2098 {
2099         struct rdma_id_private *id_priv;
2100         int ret;
2101
2102         id_priv = container_of(id, struct rdma_id_private, id);
2103         if (!cma_comp(id_priv, CMA_CONNECT))
2104                 return -EINVAL;
2105
2106         switch (rdma_node_get_transport(id->device->node_type)) {
2107         case RDMA_TRANSPORT_IB:
2108                 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2109                                      IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2110                                      private_data, private_data_len);
2111                 break;
2112         case RDMA_TRANSPORT_IWARP:
2113                 ret = iw_cm_reject(id_priv->cm_id.iw,
2114                                    private_data, private_data_len);
2115                 break;
2116         default:
2117                 ret = -ENOSYS;
2118                 break;
2119         }
2120         return ret;
2121 }
2122 EXPORT_SYMBOL(rdma_reject);
2123
2124 int rdma_disconnect(struct rdma_cm_id *id)
2125 {
2126         struct rdma_id_private *id_priv;
2127         int ret;
2128
2129         id_priv = container_of(id, struct rdma_id_private, id);
2130         if (!cma_comp(id_priv, CMA_CONNECT) &&
2131             !cma_comp(id_priv, CMA_DISCONNECT))
2132                 return -EINVAL;
2133
2134         switch (rdma_node_get_transport(id->device->node_type)) {
2135         case RDMA_TRANSPORT_IB:
2136                 ret = cma_modify_qp_err(id);
2137                 if (ret)
2138                         goto out;
2139                 /* Initiate or respond to a disconnect. */
2140                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2141                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2142                 break;
2143         case RDMA_TRANSPORT_IWARP:
2144                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2145                 break;
2146         default:
2147                 ret = -EINVAL;
2148                 break;
2149         }
2150 out:
2151         return ret;
2152 }
2153 EXPORT_SYMBOL(rdma_disconnect);
2154
2155 static void cma_add_one(struct ib_device *device)
2156 {
2157         struct cma_device *cma_dev;
2158         struct rdma_id_private *id_priv;
2159
2160         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2161         if (!cma_dev)
2162                 return;
2163
2164         cma_dev->device = device;
2165         cma_dev->node_guid = device->node_guid;
2166
2167         init_completion(&cma_dev->comp);
2168         atomic_set(&cma_dev->refcount, 1);
2169         INIT_LIST_HEAD(&cma_dev->id_list);
2170         ib_set_client_data(device, &cma_client, cma_dev);
2171
2172         mutex_lock(&lock);
2173         list_add_tail(&cma_dev->list, &dev_list);
2174         list_for_each_entry(id_priv, &listen_any_list, list)
2175                 cma_listen_on_dev(id_priv, cma_dev);
2176         mutex_unlock(&lock);
2177 }
2178
2179 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2180 {
2181         struct rdma_cm_event event;
2182         enum cma_state state;
2183
2184         /* Record that we want to remove the device */
2185         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2186         if (state == CMA_DESTROYING)
2187                 return 0;
2188
2189         cma_cancel_operation(id_priv, state);
2190         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2191
2192         /* Check for destruction from another callback. */
2193         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2194                 return 0;
2195
2196         memset(&event, 0, sizeof event);
2197         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2198         return id_priv->id.event_handler(&id_priv->id, &event);
2199 }
2200
2201 static void cma_process_remove(struct cma_device *cma_dev)
2202 {
2203         struct rdma_id_private *id_priv;
2204         int ret;
2205
2206         mutex_lock(&lock);
2207         while (!list_empty(&cma_dev->id_list)) {
2208                 id_priv = list_entry(cma_dev->id_list.next,
2209                                      struct rdma_id_private, list);
2210
2211                 if (cma_internal_listen(id_priv)) {
2212                         cma_destroy_listen(id_priv);
2213                         continue;
2214                 }
2215
2216                 list_del_init(&id_priv->list);
2217                 atomic_inc(&id_priv->refcount);
2218                 mutex_unlock(&lock);
2219
2220                 ret = cma_remove_id_dev(id_priv);
2221                 cma_deref_id(id_priv);
2222                 if (ret)
2223                         rdma_destroy_id(&id_priv->id);
2224
2225                 mutex_lock(&lock);
2226         }
2227         mutex_unlock(&lock);
2228
2229         cma_deref_dev(cma_dev);
2230         wait_for_completion(&cma_dev->comp);
2231 }
2232
2233 static void cma_remove_one(struct ib_device *device)
2234 {
2235         struct cma_device *cma_dev;
2236
2237         cma_dev = ib_get_client_data(device, &cma_client);
2238         if (!cma_dev)
2239                 return;
2240
2241         mutex_lock(&lock);
2242         list_del(&cma_dev->list);
2243         mutex_unlock(&lock);
2244
2245         cma_process_remove(cma_dev);
2246         kfree(cma_dev);
2247 }
2248
2249 static int cma_init(void)
2250 {
2251         int ret;
2252
2253         cma_wq = create_singlethread_workqueue("rdma_cm_wq");
2254         if (!cma_wq)
2255                 return -ENOMEM;
2256
2257         ib_sa_register_client(&sa_client);
2258         rdma_addr_register_client(&addr_client);
2259
2260         ret = ib_register_client(&cma_client);
2261         if (ret)
2262                 goto err;
2263         return 0;
2264
2265 err:
2266         rdma_addr_unregister_client(&addr_client);
2267         ib_sa_unregister_client(&sa_client);
2268         destroy_workqueue(cma_wq);
2269         return ret;
2270 }
2271
2272 static void cma_cleanup(void)
2273 {
2274         ib_unregister_client(&cma_client);
2275         rdma_addr_unregister_client(&addr_client);
2276         ib_sa_unregister_client(&sa_client);
2277         destroy_workqueue(cma_wq);
2278         idr_destroy(&sdp_ps);
2279         idr_destroy(&tcp_ps);
2280 }
2281
2282 module_init(cma_init);
2283 module_exit(cma_cleanup);