git.oblomov.eu Git - 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 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  56
  57 static void cma_add_one(struct ib_device *device);
  58 static void cma_remove_one(struct ib_device *device);
  59
  60 static struct ib_client cma_client = {
  61         .name   = "cma",
  62         .add    = cma_add_one,
  63         .remove = cma_remove_one
  64 };
  65
  66 static struct ib_sa_client sa_client;
  67 static struct rdma_addr_client addr_client;
  68 static LIST_HEAD(dev_list);
  69 static LIST_HEAD(listen_any_list);
  70 static DEFINE_MUTEX(lock);
  71 static struct workqueue_struct *cma_wq;
  72 static DEFINE_IDR(sdp_ps);
  73 static DEFINE_IDR(tcp_ps);
  74 static DEFINE_IDR(udp_ps);
  75 static DEFINE_IDR(ipoib_ps);
  76 static int next_port;
  77
  78 struct cma_device {
  79         struct list_head        list;
  80         struct ib_device        *device;
  81         struct completion       comp;
  82         atomic_t                refcount;
  83         struct list_head        id_list;
  84 };
  85
  86 enum cma_state {
  87         CMA_IDLE,
  88         CMA_ADDR_QUERY,
  89         CMA_ADDR_RESOLVED,
  90         CMA_ROUTE_QUERY,
  91         CMA_ROUTE_RESOLVED,
  92         CMA_CONNECT,
  93         CMA_DISCONNECT,
  94         CMA_ADDR_BOUND,
  95         CMA_LISTEN,
  96         CMA_DEVICE_REMOVAL,
  97         CMA_DESTROYING
  98 };
  99
 100 struct rdma_bind_list {
 101         struct idr              *ps;
 102         struct hlist_head       owners;
 103         unsigned short          port;
 104 };
 105
 106 /*
 107  * Device removal can occur at anytime, so we need extra handling to
 108  * serialize notifying the user of device removal with other callbacks.
 109  * We do this by disabling removal notification while a callback is in process,
 110  * and reporting it after the callback completes.
 111  */
 112 struct rdma_id_private {
 113         struct rdma_cm_id       id;
 114
 115         struct rdma_bind_list   *bind_list;
 116         struct hlist_node       node;
 117         struct list_head        list;
 118         struct list_head        listen_list;
 119         struct cma_device       *cma_dev;
 120         struct list_head        mc_list;
 121
 122         enum cma_state          state;
 123         spinlock_t              lock;
 124         struct mutex            qp_mutex;
 125
 126         struct completion       comp;
 127         atomic_t                refcount;
 128         wait_queue_head_t       wait_remove;
 129         atomic_t                dev_remove;
 130
 131         int                     backlog;
 132         int                     timeout_ms;
 133         struct ib_sa_query      *query;
 134         int                     query_id;
 135         union {
 136                 struct ib_cm_id *ib;
 137                 struct iw_cm_id *iw;
 138         } cm_id;
 139
 140         u32                     seq_num;
 141         u32                     qkey;
 142         u32                     qp_num;
 143         u8                      srq;
 144         u8                      tos;
 145 };
 146
 147 struct cma_multicast {
 148         struct rdma_id_private *id_priv;
 149         union {
 150                 struct ib_sa_multicast *ib;
 151         } multicast;
 152         struct list_head        list;
 153         void                    *context;
 154         struct sockaddr         addr;
 155         u8                      pad[sizeof(struct sockaddr_in6) -
 156                                     sizeof(struct sockaddr)];
 157 };
 158
 159 struct cma_work {
 160         struct work_struct      work;
 161         struct rdma_id_private  *id;
 162         enum cma_state          old_state;
 163         enum cma_state          new_state;
 164         struct rdma_cm_event    event;
 165 };
 166
 167 union cma_ip_addr {
 168         struct in6_addr ip6;
 169         struct {
 170                 __u32 pad[3];
 171                 __u32 addr;
 172         } ip4;
 173 };
 174
 175 struct cma_hdr {
 176         u8 cma_version;
 177         u8 ip_version;  /* IP version: 7:4 */
 178         __u16 port;
 179         union cma_ip_addr src_addr;
 180         union cma_ip_addr dst_addr;
 181 };
 182
 183 struct sdp_hh {
 184         u8 bsdh[16];
 185         u8 sdp_version; /* Major version: 7:4 */
 186         u8 ip_version;  /* IP version: 7:4 */
 187         u8 sdp_specific1[10];
 188         __u16 port;
 189         __u16 sdp_specific2;
 190         union cma_ip_addr src_addr;
 191         union cma_ip_addr dst_addr;
 192 };
 193
 194 struct sdp_hah {
 195         u8 bsdh[16];
 196         u8 sdp_version;
 197 };
 198
 199 #define CMA_VERSION 0x00
 200 #define SDP_MAJ_VERSION 0x2
 201
 202 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
 203 {
 204         unsigned long flags;
 205         int ret;
 206
 207         spin_lock_irqsave(&id_priv->lock, flags);
 208         ret = (id_priv->state == comp);
 209         spin_unlock_irqrestore(&id_priv->lock, flags);
 210         return ret;
 211 }
 212
 213 static int cma_comp_exch(struct rdma_id_private *id_priv,
 214                          enum cma_state comp, enum cma_state exch)
 215 {
 216         unsigned long flags;
 217         int ret;
 218
 219         spin_lock_irqsave(&id_priv->lock, flags);
 220         if ((ret = (id_priv->state == comp)))
 221                 id_priv->state = exch;
 222         spin_unlock_irqrestore(&id_priv->lock, flags);
 223         return ret;
 224 }
 225
 226 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
 227                                enum cma_state exch)
 228 {
 229         unsigned long flags;
 230         enum cma_state old;
 231
 232         spin_lock_irqsave(&id_priv->lock, flags);
 233         old = id_priv->state;
 234         id_priv->state = exch;
 235         spin_unlock_irqrestore(&id_priv->lock, flags);
 236         return old;
 237 }
 238
 239 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
 240 {
 241         return hdr->ip_version >> 4;
 242 }
 243
 244 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 245 {
 246         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 247 }
 248
 249 static inline u8 sdp_get_majv(u8 sdp_version)
 250 {
 251         return sdp_version >> 4;
 252 }
 253
 254 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
 255 {
 256         return hh->ip_version >> 4;
 257 }
 258
 259 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
 260 {
 261         hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
 262 }
 263
 264 static inline int cma_is_ud_ps(enum rdma_port_space ps)
 265 {
 266         return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
 267 }
 268
 269 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 270                               struct cma_device *cma_dev)
 271 {
 272         atomic_inc(&cma_dev->refcount);
 273         id_priv->cma_dev = cma_dev;
 274         id_priv->id.device = cma_dev->device;
 275         list_add_tail(&id_priv->list, &cma_dev->id_list);
 276 }
 277
 278 static inline void cma_deref_dev(struct cma_device *cma_dev)
 279 {
 280         if (atomic_dec_and_test(&cma_dev->refcount))
 281                 complete(&cma_dev->comp);
 282 }
 283
 284 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
 285 {
 286         list_del(&id_priv->list);
 287         cma_deref_dev(id_priv->cma_dev);
 288         id_priv->cma_dev = NULL;
 289 }
 290
 291 static int cma_set_qkey(struct ib_device *device, u8 port_num,
 292                         enum rdma_port_space ps,
 293                         struct rdma_dev_addr *dev_addr, u32 *qkey)
 294 {
 295         struct ib_sa_mcmember_rec rec;
 296         int ret = 0;
 297
 298         switch (ps) {
 299         case RDMA_PS_UDP:
 300                 *qkey = RDMA_UDP_QKEY;
 301                 break;
 302         case RDMA_PS_IPOIB:
 303                 ib_addr_get_mgid(dev_addr, &rec.mgid);
 304                 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec);
 305                 *qkey = be32_to_cpu(rec.qkey);
 306                 break;
 307         default:
 308                 break;
 309         }
 310         return ret;
 311 }
 312
 313 static int cma_acquire_dev(struct rdma_id_private *id_priv)
 314 {
 315         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 316         struct cma_device *cma_dev;
 317         union ib_gid gid;
 318         int ret = -ENODEV;
 319
 320         switch (rdma_node_get_transport(dev_addr->dev_type)) {
 321         case RDMA_TRANSPORT_IB:
 322                 ib_addr_get_sgid(dev_addr, &gid);
 323                 break;
 324         case RDMA_TRANSPORT_IWARP:
 325                 iw_addr_get_sgid(dev_addr, &gid);
 326                 break;
 327         default:
 328                 return -ENODEV;
 329         }
 330
 331         list_for_each_entry(cma_dev, &dev_list, list) {
 332                 ret = ib_find_cached_gid(cma_dev->device, &gid,
 333                                          &id_priv->id.port_num, NULL);
 334                 if (!ret) {
 335                         ret = cma_set_qkey(cma_dev->device,
 336                                            id_priv->id.port_num,
 337                                            id_priv->id.ps, dev_addr,
 338                                            &id_priv->qkey);
 339                         if (!ret)
 340                                 cma_attach_to_dev(id_priv, cma_dev);
 341                         break;
 342                 }
 343         }
 344         return ret;
 345 }
 346
 347 static void cma_deref_id(struct rdma_id_private *id_priv)
 348 {
 349         if (atomic_dec_and_test(&id_priv->refcount))
 350                 complete(&id_priv->comp);
 351 }
 352
 353 static int cma_disable_remove(struct rdma_id_private *id_priv,
 354                               enum cma_state state)
 355 {
 356         unsigned long flags;
 357         int ret;
 358
 359         spin_lock_irqsave(&id_priv->lock, flags);
 360         if (id_priv->state == state) {
 361                 atomic_inc(&id_priv->dev_remove);
 362                 ret = 0;
 363         } else
 364                 ret = -EINVAL;
 365         spin_unlock_irqrestore(&id_priv->lock, flags);
 366         return ret;
 367 }
 368
 369 static void cma_enable_remove(struct rdma_id_private *id_priv)
 370 {
 371         if (atomic_dec_and_test(&id_priv->dev_remove))
 372                 wake_up(&id_priv->wait_remove);
 373 }
 374
 375 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
 376 {
 377         return (id_priv->id.device && id_priv->cm_id.ib);
 378 }
 379
 380 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
 381                                   void *context, enum rdma_port_space ps)
 382 {
 383         struct rdma_id_private *id_priv;
 384
 385         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 386         if (!id_priv)
 387                 return ERR_PTR(-ENOMEM);
 388
 389         id_priv->state = CMA_IDLE;
 390         id_priv->id.context = context;
 391         id_priv->id.event_handler = event_handler;
 392         id_priv->id.ps = ps;
 393         spin_lock_init(&id_priv->lock);
 394         mutex_init(&id_priv->qp_mutex);
 395         init_completion(&id_priv->comp);
 396         atomic_set(&id_priv->refcount, 1);
 397         init_waitqueue_head(&id_priv->wait_remove);
 398         atomic_set(&id_priv->dev_remove, 0);
 399         INIT_LIST_HEAD(&id_priv->listen_list);
 400         INIT_LIST_HEAD(&id_priv->mc_list);
 401         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 402
 403         return &id_priv->id;
 404 }
 405 EXPORT_SYMBOL(rdma_create_id);
 406
 407 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 408 {
 409         struct ib_qp_attr qp_attr;
 410         int qp_attr_mask, ret;
 411
 412         qp_attr.qp_state = IB_QPS_INIT;
 413         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 414         if (ret)
 415                 return ret;
 416
 417         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 418         if (ret)
 419                 return ret;
 420
 421         qp_attr.qp_state = IB_QPS_RTR;
 422         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 423         if (ret)
 424                 return ret;
 425
 426         qp_attr.qp_state = IB_QPS_RTS;
 427         qp_attr.sq_psn = 0;
 428         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 429
 430         return ret;
 431 }
 432
 433 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 434 {
 435         struct ib_qp_attr qp_attr;
 436         int qp_attr_mask, ret;
 437
 438         qp_attr.qp_state = IB_QPS_INIT;
 439         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 440         if (ret)
 441                 return ret;
 442
 443         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 444 }
 445
 446 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 447                    struct ib_qp_init_attr *qp_init_attr)
 448 {
 449         struct rdma_id_private *id_priv;
 450         struct ib_qp *qp;
 451         int ret;
 452
 453         id_priv = container_of(id, struct rdma_id_private, id);
 454         if (id->device != pd->device)
 455                 return -EINVAL;
 456
 457         qp = ib_create_qp(pd, qp_init_attr);
 458         if (IS_ERR(qp))
 459                 return PTR_ERR(qp);
 460
 461         if (cma_is_ud_ps(id_priv->id.ps))
 462                 ret = cma_init_ud_qp(id_priv, qp);
 463         else
 464                 ret = cma_init_conn_qp(id_priv, qp);
 465         if (ret)
 466                 goto err;
 467
 468         id->qp = qp;
 469         id_priv->qp_num = qp->qp_num;
 470         id_priv->srq = (qp->srq != NULL);
 471         return 0;
 472 err:
 473         ib_destroy_qp(qp);
 474         return ret;
 475 }
 476 EXPORT_SYMBOL(rdma_create_qp);
 477
 478 void rdma_destroy_qp(struct rdma_cm_id *id)
 479 {
 480         struct rdma_id_private *id_priv;
 481
 482         id_priv = container_of(id, struct rdma_id_private, id);
 483         mutex_lock(&id_priv->qp_mutex);
 484         ib_destroy_qp(id_priv->id.qp);
 485         id_priv->id.qp = NULL;
 486         mutex_unlock(&id_priv->qp_mutex);
 487 }
 488 EXPORT_SYMBOL(rdma_destroy_qp);
 489
 490 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv)
 491 {
 492         struct ib_qp_attr qp_attr;
 493         int qp_attr_mask, ret;
 494
 495         mutex_lock(&id_priv->qp_mutex);
 496         if (!id_priv->id.qp) {
 497                 ret = 0;
 498                 goto out;
 499         }
 500
 501         /* Need to update QP attributes from default values. */
 502         qp_attr.qp_state = IB_QPS_INIT;
 503         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 504         if (ret)
 505                 goto out;
 506
 507         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 508         if (ret)
 509                 goto out;
 510
 511         qp_attr.qp_state = IB_QPS_RTR;
 512         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 513         if (ret)
 514                 goto out;
 515
 516         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 517 out:
 518         mutex_unlock(&id_priv->qp_mutex);
 519         return ret;
 520 }
 521
 522 static int cma_modify_qp_rts(struct rdma_id_private *id_priv)
 523 {
 524         struct ib_qp_attr qp_attr;
 525         int qp_attr_mask, ret;
 526
 527         mutex_lock(&id_priv->qp_mutex);
 528         if (!id_priv->id.qp) {
 529                 ret = 0;
 530                 goto out;
 531         }
 532
 533         qp_attr.qp_state = IB_QPS_RTS;
 534         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 535         if (ret)
 536                 goto out;
 537
 538         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 539 out:
 540         mutex_unlock(&id_priv->qp_mutex);
 541         return ret;
 542 }
 543
 544 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 545 {
 546         struct ib_qp_attr qp_attr;
 547         int ret;
 548
 549         mutex_lock(&id_priv->qp_mutex);
 550         if (!id_priv->id.qp) {
 551                 ret = 0;
 552                 goto out;
 553         }
 554
 555         qp_attr.qp_state = IB_QPS_ERR;
 556         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 557 out:
 558         mutex_unlock(&id_priv->qp_mutex);
 559         return ret;
 560 }
 561
 562 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 563                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 564 {
 565         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 566         int ret;
 567
 568         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 569                                   ib_addr_get_pkey(dev_addr),
 570                                   &qp_attr->pkey_index);
 571         if (ret)
 572                 return ret;
 573
 574         qp_attr->port_num = id_priv->id.port_num;
 575         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 576
 577         if (cma_is_ud_ps(id_priv->id.ps)) {
 578                 qp_attr->qkey = id_priv->qkey;
 579                 *qp_attr_mask |= IB_QP_QKEY;
 580         } else {
 581                 qp_attr->qp_access_flags = 0;
 582                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 583         }
 584         return 0;
 585 }
 586
 587 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 588                        int *qp_attr_mask)
 589 {
 590         struct rdma_id_private *id_priv;
 591         int ret = 0;
 592
 593         id_priv = container_of(id, struct rdma_id_private, id);
 594         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 595         case RDMA_TRANSPORT_IB:
 596                 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
 597                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 598                 else
 599                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 600                                                  qp_attr_mask);
 601                 if (qp_attr->qp_state == IB_QPS_RTR)
 602                         qp_attr->rq_psn = id_priv->seq_num;
 603                 break;
 604         case RDMA_TRANSPORT_IWARP:
 605                 if (!id_priv->cm_id.iw) {
 606                         qp_attr->qp_access_flags = 0;
 607                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 608                 } else
 609                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 610                                                  qp_attr_mask);
 611                 break;
 612         default:
 613                 ret = -ENOSYS;
 614                 break;
 615         }
 616
 617         return ret;
 618 }
 619 EXPORT_SYMBOL(rdma_init_qp_attr);
 620
 621 static inline int cma_zero_addr(struct sockaddr *addr)
 622 {
 623         struct in6_addr *ip6;
 624
 625         if (addr->sa_family == AF_INET)
 626                 return ZERONET(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 627         else {
 628                 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
 629                 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
 630                         ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
 631         }
 632 }
 633
 634 static inline int cma_loopback_addr(struct sockaddr *addr)
 635 {
 636         return LOOPBACK(((struct sockaddr_in *) addr)->sin_addr.s_addr);
 637 }
 638
 639 static inline int cma_any_addr(struct sockaddr *addr)
 640 {
 641         return cma_zero_addr(addr) || cma_loopback_addr(addr);
 642 }
 643
 644 static inline __be16 cma_port(struct sockaddr *addr)
 645 {
 646         if (addr->sa_family == AF_INET)
 647                 return ((struct sockaddr_in *) addr)->sin_port;
 648         else
 649                 return ((struct sockaddr_in6 *) addr)->sin6_port;
 650 }
 651
 652 static inline int cma_any_port(struct sockaddr *addr)
 653 {
 654         return !cma_port(addr);
 655 }
 656
 657 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
 658                             u8 *ip_ver, __u16 *port,
 659                             union cma_ip_addr **src, union cma_ip_addr **dst)
 660 {
 661         switch (ps) {
 662         case RDMA_PS_SDP:
 663                 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
 664                     SDP_MAJ_VERSION)
 665                         return -EINVAL;
 666
 667                 *ip_ver = sdp_get_ip_ver(hdr);
 668                 *port   = ((struct sdp_hh *) hdr)->port;
 669                 *src    = &((struct sdp_hh *) hdr)->src_addr;
 670                 *dst    = &((struct sdp_hh *) hdr)->dst_addr;
 671                 break;
 672         default:
 673                 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
 674                         return -EINVAL;
 675
 676                 *ip_ver = cma_get_ip_ver(hdr);
 677                 *port   = ((struct cma_hdr *) hdr)->port;
 678                 *src    = &((struct cma_hdr *) hdr)->src_addr;
 679                 *dst    = &((struct cma_hdr *) hdr)->dst_addr;
 680                 break;
 681         }
 682
 683         if (*ip_ver != 4 && *ip_ver != 6)
 684                 return -EINVAL;
 685         return 0;
 686 }
 687
 688 static void cma_save_net_info(struct rdma_addr *addr,
 689                               struct rdma_addr *listen_addr,
 690                               u8 ip_ver, __u16 port,
 691                               union cma_ip_addr *src, union cma_ip_addr *dst)
 692 {
 693         struct sockaddr_in *listen4, *ip4;
 694         struct sockaddr_in6 *listen6, *ip6;
 695
 696         switch (ip_ver) {
 697         case 4:
 698                 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
 699                 ip4 = (struct sockaddr_in *) &addr->src_addr;
 700                 ip4->sin_family = listen4->sin_family;
 701                 ip4->sin_addr.s_addr = dst->ip4.addr;
 702                 ip4->sin_port = listen4->sin_port;
 703
 704                 ip4 = (struct sockaddr_in *) &addr->dst_addr;
 705                 ip4->sin_family = listen4->sin_family;
 706                 ip4->sin_addr.s_addr = src->ip4.addr;
 707                 ip4->sin_port = port;
 708                 break;
 709         case 6:
 710                 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
 711                 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
 712                 ip6->sin6_family = listen6->sin6_family;
 713                 ip6->sin6_addr = dst->ip6;
 714                 ip6->sin6_port = listen6->sin6_port;
 715
 716                 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
 717                 ip6->sin6_family = listen6->sin6_family;
 718                 ip6->sin6_addr = src->ip6;
 719                 ip6->sin6_port = port;
 720                 break;
 721         default:
 722                 break;
 723         }
 724 }
 725
 726 static inline int cma_user_data_offset(enum rdma_port_space ps)
 727 {
 728         switch (ps) {
 729         case RDMA_PS_SDP:
 730                 return 0;
 731         default:
 732                 return sizeof(struct cma_hdr);
 733         }
 734 }
 735
 736 static void cma_cancel_route(struct rdma_id_private *id_priv)
 737 {
 738         switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 739         case RDMA_TRANSPORT_IB:
 740                 if (id_priv->query)
 741                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
 742                 break;
 743         default:
 744                 break;
 745         }
 746 }
 747
 748 static inline int cma_internal_listen(struct rdma_id_private *id_priv)
 749 {
 750         return (id_priv->state == CMA_LISTEN) && id_priv->cma_dev &&
 751                cma_any_addr(&id_priv->id.route.addr.src_addr);
 752 }
 753
 754 static void cma_destroy_listen(struct rdma_id_private *id_priv)
 755 {
 756         cma_exch(id_priv, CMA_DESTROYING);
 757
 758         if (id_priv->cma_dev) {
 759                 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
 760                 case RDMA_TRANSPORT_IB:
 761                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 762                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 763                         break;
 764                 case RDMA_TRANSPORT_IWARP:
 765                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 766                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 767                         break;
 768                 default:
 769                         break;
 770                 }
 771                 cma_detach_from_dev(id_priv);
 772         }
 773         list_del(&id_priv->listen_list);
 774
 775         cma_deref_id(id_priv);
 776         wait_for_completion(&id_priv->comp);
 777
 778         kfree(id_priv);
 779 }
 780
 781 static void cma_cancel_listens(struct rdma_id_private *id_priv)
 782 {
 783         struct rdma_id_private *dev_id_priv;
 784
 785         mutex_lock(&lock);
 786         list_del(&id_priv->list);
 787
 788         while (!list_empty(&id_priv->listen_list)) {
 789                 dev_id_priv = list_entry(id_priv->listen_list.next,
 790                                          struct rdma_id_private, listen_list);
 791                 cma_destroy_listen(dev_id_priv);
 792         }
 793         mutex_unlock(&lock);
 794 }
 795
 796 static void cma_cancel_operation(struct rdma_id_private *id_priv,
 797                                  enum cma_state state)
 798 {
 799         switch (state) {
 800         case CMA_ADDR_QUERY:
 801                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
 802                 break;
 803         case CMA_ROUTE_QUERY:
 804                 cma_cancel_route(id_priv);
 805                 break;
 806         case CMA_LISTEN:
 807                 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
 808                     !id_priv->cma_dev)
 809                         cma_cancel_listens(id_priv);
 810                 break;
 811         default:
 812                 break;
 813         }
 814 }
 815
 816 static void cma_release_port(struct rdma_id_private *id_priv)
 817 {
 818         struct rdma_bind_list *bind_list = id_priv->bind_list;
 819
 820         if (!bind_list)
 821                 return;
 822
 823         mutex_lock(&lock);
 824         hlist_del(&id_priv->node);
 825         if (hlist_empty(&bind_list->owners)) {
 826                 idr_remove(bind_list->ps, bind_list->port);
 827                 kfree(bind_list);
 828         }
 829         mutex_unlock(&lock);
 830 }
 831
 832 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
 833 {
 834         struct cma_multicast *mc;
 835
 836         while (!list_empty(&id_priv->mc_list)) {
 837                 mc = container_of(id_priv->mc_list.next,
 838                                   struct cma_multicast, list);
 839                 list_del(&mc->list);
 840                 ib_sa_free_multicast(mc->multicast.ib);
 841                 kfree(mc);
 842         }
 843 }
 844
 845 void rdma_destroy_id(struct rdma_cm_id *id)
 846 {
 847         struct rdma_id_private *id_priv;
 848         enum cma_state state;
 849
 850         id_priv = container_of(id, struct rdma_id_private, id);
 851         state = cma_exch(id_priv, CMA_DESTROYING);
 852         cma_cancel_operation(id_priv, state);
 853
 854         mutex_lock(&lock);
 855         if (id_priv->cma_dev) {
 856                 mutex_unlock(&lock);
 857                 switch (rdma_node_get_transport(id->device->node_type)) {
 858                 case RDMA_TRANSPORT_IB:
 859                         if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
 860                                 ib_destroy_cm_id(id_priv->cm_id.ib);
 861                         break;
 862                 case RDMA_TRANSPORT_IWARP:
 863                         if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
 864                                 iw_destroy_cm_id(id_priv->cm_id.iw);
 865                         break;
 866                 default:
 867                         break;
 868                 }
 869                 cma_leave_mc_groups(id_priv);
 870                 mutex_lock(&lock);
 871                 cma_detach_from_dev(id_priv);
 872         }
 873         mutex_unlock(&lock);
 874
 875         cma_release_port(id_priv);
 876         cma_deref_id(id_priv);
 877         wait_for_completion(&id_priv->comp);
 878
 879         kfree(id_priv->id.route.path_rec);
 880         kfree(id_priv);
 881 }
 882 EXPORT_SYMBOL(rdma_destroy_id);
 883
 884 static int cma_rep_recv(struct rdma_id_private *id_priv)
 885 {
 886         int ret;
 887
 888         ret = cma_modify_qp_rtr(id_priv);
 889         if (ret)
 890                 goto reject;
 891
 892         ret = cma_modify_qp_rts(id_priv);
 893         if (ret)
 894                 goto reject;
 895
 896         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
 897         if (ret)
 898                 goto reject;
 899
 900         return 0;
 901 reject:
 902         cma_modify_qp_err(id_priv);
 903         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
 904                        NULL, 0, NULL, 0);
 905         return ret;
 906 }
 907
 908 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
 909 {
 910         if (id_priv->id.ps == RDMA_PS_SDP &&
 911             sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
 912             SDP_MAJ_VERSION)
 913                 return -EINVAL;
 914
 915         return 0;
 916 }
 917
 918 static void cma_set_rep_event_data(struct rdma_cm_event *event,
 919                                    struct ib_cm_rep_event_param *rep_data,
 920                                    void *private_data)
 921 {
 922         event->param.conn.private_data = private_data;
 923         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
 924         event->param.conn.responder_resources = rep_data->responder_resources;
 925         event->param.conn.initiator_depth = rep_data->initiator_depth;
 926         event->param.conn.flow_control = rep_data->flow_control;
 927         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
 928         event->param.conn.srq = rep_data->srq;
 929         event->param.conn.qp_num = rep_data->remote_qpn;
 930 }
 931
 932 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
 933 {
 934         struct rdma_id_private *id_priv = cm_id->context;
 935         struct rdma_cm_event event;
 936         int ret = 0;
 937
 938         if (cma_disable_remove(id_priv, CMA_CONNECT))
 939                 return 0;
 940
 941         memset(&event, 0, sizeof event);
 942         switch (ib_event->event) {
 943         case IB_CM_REQ_ERROR:
 944         case IB_CM_REP_ERROR:
 945                 event.event = RDMA_CM_EVENT_UNREACHABLE;
 946                 event.status = -ETIMEDOUT;
 947                 break;
 948         case IB_CM_REP_RECEIVED:
 949                 event.status = cma_verify_rep(id_priv, ib_event->private_data);
 950                 if (event.status)
 951                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
 952                 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
 953                         event.status = cma_rep_recv(id_priv);
 954                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
 955                                                      RDMA_CM_EVENT_ESTABLISHED;
 956                 } else
 957                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
 958                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
 959                                        ib_event->private_data);
 960                 break;
 961         case IB_CM_RTU_RECEIVED:
 962         case IB_CM_USER_ESTABLISHED:
 963                 event.event = RDMA_CM_EVENT_ESTABLISHED;
 964                 break;
 965         case IB_CM_DREQ_ERROR:
 966                 event.status = -ETIMEDOUT; /* fall through */
 967         case IB_CM_DREQ_RECEIVED:
 968         case IB_CM_DREP_RECEIVED:
 969                 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
 970                         goto out;
 971                 event.event = RDMA_CM_EVENT_DISCONNECTED;
 972                 break;
 973         case IB_CM_TIMEWAIT_EXIT:
 974         case IB_CM_MRA_RECEIVED:
 975                 /* ignore event */
 976                 goto out;
 977         case IB_CM_REJ_RECEIVED:
 978                 cma_modify_qp_err(id_priv);
 979                 event.status = ib_event->param.rej_rcvd.reason;
 980                 event.event = RDMA_CM_EVENT_REJECTED;
 981                 event.param.conn.private_data = ib_event->private_data;
 982                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
 983                 break;
 984         default:
 985                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
 986                        ib_event->event);
 987                 goto out;
 988         }
 989
 990         ret = id_priv->id.event_handler(&id_priv->id, &event);
 991         if (ret) {
 992                 /* Destroy the CM ID by returning a non-zero value. */
 993                 id_priv->cm_id.ib = NULL;
 994                 cma_exch(id_priv, CMA_DESTROYING);
 995                 cma_enable_remove(id_priv);
 996                 rdma_destroy_id(&id_priv->id);
 997                 return ret;
 998         }
 999 out:
1000         cma_enable_remove(id_priv);
1001         return ret;
1002 }
1003
1004 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1005                                                struct ib_cm_event *ib_event)
1006 {
1007         struct rdma_id_private *id_priv;
1008         struct rdma_cm_id *id;
1009         struct rdma_route *rt;
1010         union cma_ip_addr *src, *dst;
1011         __u16 port;
1012         u8 ip_ver;
1013
1014         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1015                              &ip_ver, &port, &src, &dst))
1016                 goto err;
1017
1018         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1019                             listen_id->ps);
1020         if (IS_ERR(id))
1021                 goto err;
1022
1023         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1024                           ip_ver, port, src, dst);
1025
1026         rt = &id->route;
1027         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1028         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1029                                GFP_KERNEL);
1030         if (!rt->path_rec)
1031                 goto destroy_id;
1032
1033         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1034         if (rt->num_paths == 2)
1035                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1036
1037         ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1038         ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1039         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1040         rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
1041
1042         id_priv = container_of(id, struct rdma_id_private, id);
1043         id_priv->state = CMA_CONNECT;
1044         return id_priv;
1045
1046 destroy_id:
1047         rdma_destroy_id(id);
1048 err:
1049         return NULL;
1050 }
1051
1052 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1053                                               struct ib_cm_event *ib_event)
1054 {
1055         struct rdma_id_private *id_priv;
1056         struct rdma_cm_id *id;
1057         union cma_ip_addr *src, *dst;
1058         __u16 port;
1059         u8 ip_ver;
1060         int ret;
1061
1062         id = rdma_create_id(listen_id->event_handler, listen_id->context,
1063                             listen_id->ps);
1064         if (IS_ERR(id))
1065                 return NULL;
1066
1067
1068         if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1069                              &ip_ver, &port, &src, &dst))
1070                 goto err;
1071
1072         cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1073                           ip_ver, port, src, dst);
1074
1075         ret = rdma_translate_ip(&id->route.addr.src_addr,
1076                                 &id->route.addr.dev_addr);
1077         if (ret)
1078                 goto err;
1079
1080         id_priv = container_of(id, struct rdma_id_private, id);
1081         id_priv->state = CMA_CONNECT;
1082         return id_priv;
1083 err:
1084         rdma_destroy_id(id);
1085         return NULL;
1086 }
1087
1088 static void cma_set_req_event_data(struct rdma_cm_event *event,
1089                                    struct ib_cm_req_event_param *req_data,
1090                                    void *private_data, int offset)
1091 {
1092         event->param.conn.private_data = private_data + offset;
1093         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1094         event->param.conn.responder_resources = req_data->responder_resources;
1095         event->param.conn.initiator_depth = req_data->initiator_depth;
1096         event->param.conn.flow_control = req_data->flow_control;
1097         event->param.conn.retry_count = req_data->retry_count;
1098         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1099         event->param.conn.srq = req_data->srq;
1100         event->param.conn.qp_num = req_data->remote_qpn;
1101 }
1102
1103 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1104 {
1105         struct rdma_id_private *listen_id, *conn_id;
1106         struct rdma_cm_event event;
1107         int offset, ret;
1108
1109         listen_id = cm_id->context;
1110         if (cma_disable_remove(listen_id, CMA_LISTEN))
1111                 return -ECONNABORTED;
1112
1113         memset(&event, 0, sizeof event);
1114         offset = cma_user_data_offset(listen_id->id.ps);
1115         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1116         if (cma_is_ud_ps(listen_id->id.ps)) {
1117                 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1118                 event.param.ud.private_data = ib_event->private_data + offset;
1119                 event.param.ud.private_data_len =
1120                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1121         } else {
1122                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1123                 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1124                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1125                                        ib_event->private_data, offset);
1126         }
1127         if (!conn_id) {
1128                 ret = -ENOMEM;
1129                 goto out;
1130         }
1131
1132         atomic_inc(&conn_id->dev_remove);
1133         mutex_lock(&lock);
1134         ret = cma_acquire_dev(conn_id);
1135         mutex_unlock(&lock);
1136         if (ret)
1137                 goto release_conn_id;
1138
1139         conn_id->cm_id.ib = cm_id;
1140         cm_id->context = conn_id;
1141         cm_id->cm_handler = cma_ib_handler;
1142
1143         ret = conn_id->id.event_handler(&conn_id->id, &event);
1144         if (!ret)
1145                 goto out;
1146
1147         /* Destroy the CM ID by returning a non-zero value. */
1148         conn_id->cm_id.ib = NULL;
1149
1150 release_conn_id:
1151         cma_exch(conn_id, CMA_DESTROYING);
1152         cma_enable_remove(conn_id);
1153         rdma_destroy_id(&conn_id->id);
1154
1155 out:
1156         cma_enable_remove(listen_id);
1157         return ret;
1158 }
1159
1160 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1161 {
1162         return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1163 }
1164
1165 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1166                                  struct ib_cm_compare_data *compare)
1167 {
1168         struct cma_hdr *cma_data, *cma_mask;
1169         struct sdp_hh *sdp_data, *sdp_mask;
1170         __u32 ip4_addr;
1171         struct in6_addr ip6_addr;
1172
1173         memset(compare, 0, sizeof *compare);
1174         cma_data = (void *) compare->data;
1175         cma_mask = (void *) compare->mask;
1176         sdp_data = (void *) compare->data;
1177         sdp_mask = (void *) compare->mask;
1178
1179         switch (addr->sa_family) {
1180         case AF_INET:
1181                 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1182                 if (ps == RDMA_PS_SDP) {
1183                         sdp_set_ip_ver(sdp_data, 4);
1184                         sdp_set_ip_ver(sdp_mask, 0xF);
1185                         sdp_data->dst_addr.ip4.addr = ip4_addr;
1186                         sdp_mask->dst_addr.ip4.addr = ~0;
1187                 } else {
1188                         cma_set_ip_ver(cma_data, 4);
1189                         cma_set_ip_ver(cma_mask, 0xF);
1190                         cma_data->dst_addr.ip4.addr = ip4_addr;
1191                         cma_mask->dst_addr.ip4.addr = ~0;
1192                 }
1193                 break;
1194         case AF_INET6:
1195                 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1196                 if (ps == RDMA_PS_SDP) {
1197                         sdp_set_ip_ver(sdp_data, 6);
1198                         sdp_set_ip_ver(sdp_mask, 0xF);
1199                         sdp_data->dst_addr.ip6 = ip6_addr;
1200                         memset(&sdp_mask->dst_addr.ip6, 0xFF,
1201                                sizeof sdp_mask->dst_addr.ip6);
1202                 } else {
1203                         cma_set_ip_ver(cma_data, 6);
1204                         cma_set_ip_ver(cma_mask, 0xF);
1205                         cma_data->dst_addr.ip6 = ip6_addr;
1206                         memset(&cma_mask->dst_addr.ip6, 0xFF,
1207                                sizeof cma_mask->dst_addr.ip6);
1208                 }
1209                 break;
1210         default:
1211                 break;
1212         }
1213 }
1214
1215 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1216 {
1217         struct rdma_id_private *id_priv = iw_id->context;
1218         struct rdma_cm_event event;
1219         struct sockaddr_in *sin;
1220         int ret = 0;
1221
1222         if (cma_disable_remove(id_priv, CMA_CONNECT))
1223                 return 0;
1224
1225         memset(&event, 0, sizeof event);
1226         switch (iw_event->event) {
1227         case IW_CM_EVENT_CLOSE:
1228                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1229                 break;
1230         case IW_CM_EVENT_CONNECT_REPLY:
1231                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1232                 *sin = iw_event->local_addr;
1233                 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1234                 *sin = iw_event->remote_addr;
1235                 switch (iw_event->status) {
1236                 case 0:
1237                         event.event = RDMA_CM_EVENT_ESTABLISHED;
1238                         break;
1239                 case -ECONNRESET:
1240                 case -ECONNREFUSED:
1241                         event.event = RDMA_CM_EVENT_REJECTED;
1242                         break;
1243                 case -ETIMEDOUT:
1244                         event.event = RDMA_CM_EVENT_UNREACHABLE;
1245                         break;
1246                 default:
1247                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1248                         break;
1249                 }
1250                 break;
1251         case IW_CM_EVENT_ESTABLISHED:
1252                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1253                 break;
1254         default:
1255                 BUG_ON(1);
1256         }
1257
1258         event.status = iw_event->status;
1259         event.param.conn.private_data = iw_event->private_data;
1260         event.param.conn.private_data_len = iw_event->private_data_len;
1261         ret = id_priv->id.event_handler(&id_priv->id, &event);
1262         if (ret) {
1263                 /* Destroy the CM ID by returning a non-zero value. */
1264                 id_priv->cm_id.iw = NULL;
1265                 cma_exch(id_priv, CMA_DESTROYING);
1266                 cma_enable_remove(id_priv);
1267                 rdma_destroy_id(&id_priv->id);
1268                 return ret;
1269         }
1270
1271         cma_enable_remove(id_priv);
1272         return ret;
1273 }
1274
1275 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1276                                struct iw_cm_event *iw_event)
1277 {
1278         struct rdma_cm_id *new_cm_id;
1279         struct rdma_id_private *listen_id, *conn_id;
1280         struct sockaddr_in *sin;
1281         struct net_device *dev = NULL;
1282         struct rdma_cm_event event;
1283         int ret;
1284
1285         listen_id = cm_id->context;
1286         if (cma_disable_remove(listen_id, CMA_LISTEN))
1287                 return -ECONNABORTED;
1288
1289         /* Create a new RDMA id for the new IW CM ID */
1290         new_cm_id = rdma_create_id(listen_id->id.event_handler,
1291                                    listen_id->id.context,
1292                                    RDMA_PS_TCP);
1293         if (!new_cm_id) {
1294                 ret = -ENOMEM;
1295                 goto out;
1296         }
1297         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1298         atomic_inc(&conn_id->dev_remove);
1299         conn_id->state = CMA_CONNECT;
1300
1301         dev = ip_dev_find(iw_event->local_addr.sin_addr.s_addr);
1302         if (!dev) {
1303                 ret = -EADDRNOTAVAIL;
1304                 cma_enable_remove(conn_id);
1305                 rdma_destroy_id(new_cm_id);
1306                 goto out;
1307         }
1308         ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1309         if (ret) {
1310                 cma_enable_remove(conn_id);
1311                 rdma_destroy_id(new_cm_id);
1312                 goto out;
1313         }
1314
1315         mutex_lock(&lock);
1316         ret = cma_acquire_dev(conn_id);
1317         mutex_unlock(&lock);
1318         if (ret) {
1319                 cma_enable_remove(conn_id);
1320                 rdma_destroy_id(new_cm_id);
1321                 goto out;
1322         }
1323
1324         conn_id->cm_id.iw = cm_id;
1325         cm_id->context = conn_id;
1326         cm_id->cm_handler = cma_iw_handler;
1327
1328         sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1329         *sin = iw_event->local_addr;
1330         sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1331         *sin = iw_event->remote_addr;
1332
1333         memset(&event, 0, sizeof event);
1334         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1335         event.param.conn.private_data = iw_event->private_data;
1336         event.param.conn.private_data_len = iw_event->private_data_len;
1337         ret = conn_id->id.event_handler(&conn_id->id, &event);
1338         if (ret) {
1339                 /* User wants to destroy the CM ID */
1340                 conn_id->cm_id.iw = NULL;
1341                 cma_exch(conn_id, CMA_DESTROYING);
1342                 cma_enable_remove(conn_id);
1343                 rdma_destroy_id(&conn_id->id);
1344         }
1345
1346 out:
1347         if (dev)
1348                 dev_put(dev);
1349         cma_enable_remove(listen_id);
1350         return ret;
1351 }
1352
1353 static int cma_ib_listen(struct rdma_id_private *id_priv)
1354 {
1355         struct ib_cm_compare_data compare_data;
1356         struct sockaddr *addr;
1357         __be64 svc_id;
1358         int ret;
1359
1360         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1361                                             id_priv);
1362         if (IS_ERR(id_priv->cm_id.ib))
1363                 return PTR_ERR(id_priv->cm_id.ib);
1364
1365         addr = &id_priv->id.route.addr.src_addr;
1366         svc_id = cma_get_service_id(id_priv->id.ps, addr);
1367         if (cma_any_addr(addr))
1368                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1369         else {
1370                 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1371                 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1372         }
1373
1374         if (ret) {
1375                 ib_destroy_cm_id(id_priv->cm_id.ib);
1376                 id_priv->cm_id.ib = NULL;
1377         }
1378
1379         return ret;
1380 }
1381
1382 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1383 {
1384         int ret;
1385         struct sockaddr_in *sin;
1386
1387         id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1388                                             iw_conn_req_handler,
1389                                             id_priv);
1390         if (IS_ERR(id_priv->cm_id.iw))
1391                 return PTR_ERR(id_priv->cm_id.iw);
1392
1393         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1394         id_priv->cm_id.iw->local_addr = *sin;
1395
1396         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1397
1398         if (ret) {
1399                 iw_destroy_cm_id(id_priv->cm_id.iw);
1400                 id_priv->cm_id.iw = NULL;
1401         }
1402
1403         return ret;
1404 }
1405
1406 static int cma_listen_handler(struct rdma_cm_id *id,
1407                               struct rdma_cm_event *event)
1408 {
1409         struct rdma_id_private *id_priv = id->context;
1410
1411         id->context = id_priv->id.context;
1412         id->event_handler = id_priv->id.event_handler;
1413         return id_priv->id.event_handler(id, event);
1414 }
1415
1416 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1417                               struct cma_device *cma_dev)
1418 {
1419         struct rdma_id_private *dev_id_priv;
1420         struct rdma_cm_id *id;
1421         int ret;
1422
1423         id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1424         if (IS_ERR(id))
1425                 return;
1426
1427         dev_id_priv = container_of(id, struct rdma_id_private, id);
1428
1429         dev_id_priv->state = CMA_ADDR_BOUND;
1430         memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1431                ip_addr_size(&id_priv->id.route.addr.src_addr));
1432
1433         cma_attach_to_dev(dev_id_priv, cma_dev);
1434         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1435
1436         ret = rdma_listen(id, id_priv->backlog);
1437         if (ret)
1438                 goto err;
1439
1440         return;
1441 err:
1442         cma_destroy_listen(dev_id_priv);
1443 }
1444
1445 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1446 {
1447         struct cma_device *cma_dev;
1448
1449         mutex_lock(&lock);
1450         list_add_tail(&id_priv->list, &listen_any_list);
1451         list_for_each_entry(cma_dev, &dev_list, list)
1452                 cma_listen_on_dev(id_priv, cma_dev);
1453         mutex_unlock(&lock);
1454 }
1455
1456 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1457 {
1458         struct sockaddr_in addr_in;
1459
1460         memset(&addr_in, 0, sizeof addr_in);
1461         addr_in.sin_family = af;
1462         return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1463 }
1464
1465 int rdma_listen(struct rdma_cm_id *id, int backlog)
1466 {
1467         struct rdma_id_private *id_priv;
1468         int ret;
1469
1470         id_priv = container_of(id, struct rdma_id_private, id);
1471         if (id_priv->state == CMA_IDLE) {
1472                 ret = cma_bind_any(id, AF_INET);
1473                 if (ret)
1474                         return ret;
1475         }
1476
1477         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1478                 return -EINVAL;
1479
1480         id_priv->backlog = backlog;
1481         if (id->device) {
1482                 switch (rdma_node_get_transport(id->device->node_type)) {
1483                 case RDMA_TRANSPORT_IB:
1484                         ret = cma_ib_listen(id_priv);
1485                         if (ret)
1486                                 goto err;
1487                         break;
1488                 case RDMA_TRANSPORT_IWARP:
1489                         ret = cma_iw_listen(id_priv, backlog);
1490                         if (ret)
1491                                 goto err;
1492                         break;
1493                 default:
1494                         ret = -ENOSYS;
1495                         goto err;
1496                 }
1497         } else
1498                 cma_listen_on_all(id_priv);
1499
1500         return 0;
1501 err:
1502         id_priv->backlog = 0;
1503         cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1504         return ret;
1505 }
1506 EXPORT_SYMBOL(rdma_listen);
1507
1508 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1509 {
1510         struct rdma_id_private *id_priv;
1511
1512         id_priv = container_of(id, struct rdma_id_private, id);
1513         id_priv->tos = (u8) tos;
1514 }
1515 EXPORT_SYMBOL(rdma_set_service_type);
1516
1517 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1518                               void *context)
1519 {
1520         struct cma_work *work = context;
1521         struct rdma_route *route;
1522
1523         route = &work->id->id.route;
1524
1525         if (!status) {
1526                 route->num_paths = 1;
1527                 *route->path_rec = *path_rec;
1528         } else {
1529                 work->old_state = CMA_ROUTE_QUERY;
1530                 work->new_state = CMA_ADDR_RESOLVED;
1531                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1532                 work->event.status = status;
1533         }
1534
1535         queue_work(cma_wq, &work->work);
1536 }
1537
1538 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1539                               struct cma_work *work)
1540 {
1541         struct rdma_addr *addr = &id_priv->id.route.addr;
1542         struct ib_sa_path_rec path_rec;
1543         ib_sa_comp_mask comp_mask;
1544         struct sockaddr_in6 *sin6;
1545
1546         memset(&path_rec, 0, sizeof path_rec);
1547         ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1548         ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1549         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1550         path_rec.numb_path = 1;
1551         path_rec.reversible = 1;
1552         path_rec.service_id = cma_get_service_id(id_priv->id.ps, &addr->dst_addr);
1553
1554         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1555                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1556                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1557
1558         if (addr->src_addr.sa_family == AF_INET) {
1559                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1560                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1561         } else {
1562                 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1563                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1564                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1565         }
1566
1567         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1568                                                id_priv->id.port_num, &path_rec,
1569                                                comp_mask, timeout_ms,
1570                                                GFP_KERNEL, cma_query_handler,
1571                                                work, &id_priv->query);
1572
1573         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1574 }
1575
1576 static void cma_work_handler(struct work_struct *_work)
1577 {
1578         struct cma_work *work = container_of(_work, struct cma_work, work);
1579         struct rdma_id_private *id_priv = work->id;
1580         int destroy = 0;
1581
1582         atomic_inc(&id_priv->dev_remove);
1583         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1584                 goto out;
1585
1586         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1587                 cma_exch(id_priv, CMA_DESTROYING);
1588                 destroy = 1;
1589         }
1590 out:
1591         cma_enable_remove(id_priv);
1592         cma_deref_id(id_priv);
1593         if (destroy)
1594                 rdma_destroy_id(&id_priv->id);
1595         kfree(work);
1596 }
1597
1598 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1599 {
1600         struct rdma_route *route = &id_priv->id.route;
1601         struct cma_work *work;
1602         int ret;
1603
1604         work = kzalloc(sizeof *work, GFP_KERNEL);
1605         if (!work)
1606                 return -ENOMEM;
1607
1608         work->id = id_priv;
1609         INIT_WORK(&work->work, cma_work_handler);
1610         work->old_state = CMA_ROUTE_QUERY;
1611         work->new_state = CMA_ROUTE_RESOLVED;
1612         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1613
1614         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1615         if (!route->path_rec) {
1616                 ret = -ENOMEM;
1617                 goto err1;
1618         }
1619
1620         ret = cma_query_ib_route(id_priv, timeout_ms, work);
1621         if (ret)
1622                 goto err2;
1623
1624         return 0;
1625 err2:
1626         kfree(route->path_rec);
1627         route->path_rec = NULL;
1628 err1:
1629         kfree(work);
1630         return ret;
1631 }
1632
1633 int rdma_set_ib_paths(struct rdma_cm_id *id,
1634                       struct ib_sa_path_rec *path_rec, int num_paths)
1635 {
1636         struct rdma_id_private *id_priv;
1637         int ret;
1638
1639         id_priv = container_of(id, struct rdma_id_private, id);
1640         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1641                 return -EINVAL;
1642
1643         id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1644         if (!id->route.path_rec) {
1645                 ret = -ENOMEM;
1646                 goto err;
1647         }
1648
1649         memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1650         return 0;
1651 err:
1652         cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1653         return ret;
1654 }
1655 EXPORT_SYMBOL(rdma_set_ib_paths);
1656
1657 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1658 {
1659         struct cma_work *work;
1660
1661         work = kzalloc(sizeof *work, GFP_KERNEL);
1662         if (!work)
1663                 return -ENOMEM;
1664
1665         work->id = id_priv;
1666         INIT_WORK(&work->work, cma_work_handler);
1667         work->old_state = CMA_ROUTE_QUERY;
1668         work->new_state = CMA_ROUTE_RESOLVED;
1669         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1670         queue_work(cma_wq, &work->work);
1671         return 0;
1672 }
1673
1674 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1675 {
1676         struct rdma_id_private *id_priv;
1677         int ret;
1678
1679         id_priv = container_of(id, struct rdma_id_private, id);
1680         if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1681                 return -EINVAL;
1682
1683         atomic_inc(&id_priv->refcount);
1684         switch (rdma_node_get_transport(id->device->node_type)) {
1685         case RDMA_TRANSPORT_IB:
1686                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1687                 break;
1688         case RDMA_TRANSPORT_IWARP:
1689                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1690                 break;
1691         default:
1692                 ret = -ENOSYS;
1693                 break;
1694         }
1695         if (ret)
1696                 goto err;
1697
1698         return 0;
1699 err:
1700         cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1701         cma_deref_id(id_priv);
1702         return ret;
1703 }
1704 EXPORT_SYMBOL(rdma_resolve_route);
1705
1706 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1707 {
1708         struct cma_device *cma_dev;
1709         struct ib_port_attr port_attr;
1710         union ib_gid gid;
1711         u16 pkey;
1712         int ret;
1713         u8 p;
1714
1715         mutex_lock(&lock);
1716         if (list_empty(&dev_list)) {
1717                 ret = -ENODEV;
1718                 goto out;
1719         }
1720         list_for_each_entry(cma_dev, &dev_list, list)
1721                 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1722                         if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1723                             port_attr.state == IB_PORT_ACTIVE)
1724                                 goto port_found;
1725
1726         p = 1;
1727         cma_dev = list_entry(dev_list.next, struct cma_device, list);
1728
1729 port_found:
1730         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1731         if (ret)
1732                 goto out;
1733
1734         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1735         if (ret)
1736                 goto out;
1737
1738         ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1739         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1740         id_priv->id.port_num = p;
1741         cma_attach_to_dev(id_priv, cma_dev);
1742 out:
1743         mutex_unlock(&lock);
1744         return ret;
1745 }
1746
1747 static void addr_handler(int status, struct sockaddr *src_addr,
1748                          struct rdma_dev_addr *dev_addr, void *context)
1749 {
1750         struct rdma_id_private *id_priv = context;
1751         struct rdma_cm_event event;
1752
1753         memset(&event, 0, sizeof event);
1754         atomic_inc(&id_priv->dev_remove);
1755
1756         /*
1757          * Grab mutex to block rdma_destroy_id() from removing the device while
1758          * we're trying to acquire it.
1759          */
1760         mutex_lock(&lock);
1761         if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1762                 mutex_unlock(&lock);
1763                 goto out;
1764         }
1765
1766         if (!status && !id_priv->cma_dev)
1767                 status = cma_acquire_dev(id_priv);
1768         mutex_unlock(&lock);
1769
1770         if (status) {
1771                 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1772                         goto out;
1773                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1774                 event.status = status;
1775         } else {
1776                 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1777                        ip_addr_size(src_addr));
1778                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1779         }
1780
1781         if (id_priv->id.event_handler(&id_priv->id, &event)) {
1782                 cma_exch(id_priv, CMA_DESTROYING);
1783                 cma_enable_remove(id_priv);
1784                 cma_deref_id(id_priv);
1785                 rdma_destroy_id(&id_priv->id);
1786                 return;
1787         }
1788 out:
1789         cma_enable_remove(id_priv);
1790         cma_deref_id(id_priv);
1791 }
1792
1793 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1794 {
1795         struct cma_work *work;
1796         struct sockaddr_in *src_in, *dst_in;
1797         union ib_gid gid;
1798         int ret;
1799
1800         work = kzalloc(sizeof *work, GFP_KERNEL);
1801         if (!work)
1802                 return -ENOMEM;
1803
1804         if (!id_priv->cma_dev) {
1805                 ret = cma_bind_loopback(id_priv);
1806                 if (ret)
1807                         goto err;
1808         }
1809
1810         ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1811         ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1812
1813         if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1814                 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1815                 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1816                 src_in->sin_family = dst_in->sin_family;
1817                 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1818         }
1819
1820         work->id = id_priv;
1821         INIT_WORK(&work->work, cma_work_handler);
1822         work->old_state = CMA_ADDR_QUERY;
1823         work->new_state = CMA_ADDR_RESOLVED;
1824         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1825         queue_work(cma_wq, &work->work);
1826         return 0;
1827 err:
1828         kfree(work);
1829         return ret;
1830 }
1831
1832 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1833                          struct sockaddr *dst_addr)
1834 {
1835         if (src_addr && src_addr->sa_family)
1836                 return rdma_bind_addr(id, src_addr);
1837         else
1838                 return cma_bind_any(id, dst_addr->sa_family);
1839 }
1840
1841 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1842                       struct sockaddr *dst_addr, int timeout_ms)
1843 {
1844         struct rdma_id_private *id_priv;
1845         int ret;
1846
1847         id_priv = container_of(id, struct rdma_id_private, id);
1848         if (id_priv->state == CMA_IDLE) {
1849                 ret = cma_bind_addr(id, src_addr, dst_addr);
1850                 if (ret)
1851                         return ret;
1852         }
1853
1854         if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1855                 return -EINVAL;
1856
1857         atomic_inc(&id_priv->refcount);
1858         memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1859         if (cma_any_addr(dst_addr))
1860                 ret = cma_resolve_loopback(id_priv);
1861         else
1862                 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1863                                       dst_addr, &id->route.addr.dev_addr,
1864                                       timeout_ms, addr_handler, id_priv);
1865         if (ret)
1866                 goto err;
1867
1868         return 0;
1869 err:
1870         cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1871         cma_deref_id(id_priv);
1872         return ret;
1873 }
1874 EXPORT_SYMBOL(rdma_resolve_addr);
1875
1876 static void cma_bind_port(struct rdma_bind_list *bind_list,
1877                           struct rdma_id_private *id_priv)
1878 {
1879         struct sockaddr_in *sin;
1880
1881         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1882         sin->sin_port = htons(bind_list->port);
1883         id_priv->bind_list = bind_list;
1884         hlist_add_head(&id_priv->node, &bind_list->owners);
1885 }
1886
1887 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1888                           unsigned short snum)
1889 {
1890         struct rdma_bind_list *bind_list;
1891         int port, ret;
1892
1893         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1894         if (!bind_list)
1895                 return -ENOMEM;
1896
1897         do {
1898                 ret = idr_get_new_above(ps, bind_list, snum, &port);
1899         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1900
1901         if (ret)
1902                 goto err1;
1903
1904         if (port != snum) {
1905                 ret = -EADDRNOTAVAIL;
1906                 goto err2;
1907         }
1908
1909         bind_list->ps = ps;
1910         bind_list->port = (unsigned short) port;
1911         cma_bind_port(bind_list, id_priv);
1912         return 0;
1913 err2:
1914         idr_remove(ps, port);
1915 err1:
1916         kfree(bind_list);
1917         return ret;
1918 }
1919
1920 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1921 {
1922         struct rdma_bind_list *bind_list;
1923         int port, ret, low, high;
1924
1925         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1926         if (!bind_list)
1927                 return -ENOMEM;
1928
1929 retry:
1930         /* FIXME: add proper port randomization per like inet_csk_get_port */
1931         do {
1932                 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1933         } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1934
1935         if (ret)
1936                 goto err1;
1937
1938         inet_get_local_port_range(&low, &high);
1939         if (port > high) {
1940                 if (next_port != low) {
1941                         idr_remove(ps, port);
1942                         next_port = low;
1943                         goto retry;
1944                 }
1945                 ret = -EADDRNOTAVAIL;
1946                 goto err2;
1947         }
1948
1949         if (port == high)
1950                 next_port = low;
1951         else
1952                 next_port = port + 1;
1953
1954         bind_list->ps = ps;
1955         bind_list->port = (unsigned short) port;
1956         cma_bind_port(bind_list, id_priv);
1957         return 0;
1958 err2:
1959         idr_remove(ps, port);
1960 err1:
1961         kfree(bind_list);
1962         return ret;
1963 }
1964
1965 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1966 {
1967         struct rdma_id_private *cur_id;
1968         struct sockaddr_in *sin, *cur_sin;
1969         struct rdma_bind_list *bind_list;
1970         struct hlist_node *node;
1971         unsigned short snum;
1972
1973         sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1974         snum = ntohs(sin->sin_port);
1975         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1976                 return -EACCES;
1977
1978         bind_list = idr_find(ps, snum);
1979         if (!bind_list)
1980                 return cma_alloc_port(ps, id_priv, snum);
1981
1982         /*
1983          * We don't support binding to any address if anyone is bound to
1984          * a specific address on the same port.
1985          */
1986         if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1987                 return -EADDRNOTAVAIL;
1988
1989         hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1990                 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1991                         return -EADDRNOTAVAIL;
1992
1993                 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1994                 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1995                         return -EADDRINUSE;
1996         }
1997
1998         cma_bind_port(bind_list, id_priv);
1999         return 0;
2000 }
2001
2002 static int cma_get_port(struct rdma_id_private *id_priv)
2003 {
2004         struct idr *ps;
2005         int ret;
2006
2007         switch (id_priv->id.ps) {
2008         case RDMA_PS_SDP:
2009                 ps = &sdp_ps;
2010                 break;
2011         case RDMA_PS_TCP:
2012                 ps = &tcp_ps;
2013                 break;
2014         case RDMA_PS_UDP:
2015                 ps = &udp_ps;
2016                 break;
2017         case RDMA_PS_IPOIB:
2018                 ps = &ipoib_ps;
2019                 break;
2020         default:
2021                 return -EPROTONOSUPPORT;
2022         }
2023
2024         mutex_lock(&lock);
2025         if (cma_any_port(&id_priv->id.route.addr.src_addr))
2026                 ret = cma_alloc_any_port(ps, id_priv);
2027         else
2028                 ret = cma_use_port(ps, id_priv);
2029         mutex_unlock(&lock);
2030
2031         return ret;
2032 }
2033
2034 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2035 {
2036         struct rdma_id_private *id_priv;
2037         int ret;
2038
2039         if (addr->sa_family != AF_INET)
2040                 return -EAFNOSUPPORT;
2041
2042         id_priv = container_of(id, struct rdma_id_private, id);
2043         if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2044                 return -EINVAL;
2045
2046         if (!cma_any_addr(addr)) {
2047                 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2048                 if (ret)
2049                         goto err1;
2050
2051                 mutex_lock(&lock);
2052                 ret = cma_acquire_dev(id_priv);
2053                 mutex_unlock(&lock);
2054                 if (ret)
2055                         goto err1;
2056         }
2057
2058         memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2059         ret = cma_get_port(id_priv);
2060         if (ret)
2061                 goto err2;
2062
2063         return 0;
2064 err2:
2065         if (!cma_any_addr(addr)) {
2066                 mutex_lock(&lock);
2067                 cma_detach_from_dev(id_priv);
2068                 mutex_unlock(&lock);
2069         }
2070 err1:
2071         cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2072         return ret;
2073 }
2074 EXPORT_SYMBOL(rdma_bind_addr);
2075
2076 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2077                           struct rdma_route *route)
2078 {
2079         struct sockaddr_in *src4, *dst4;
2080         struct cma_hdr *cma_hdr;
2081         struct sdp_hh *sdp_hdr;
2082
2083         src4 = (struct sockaddr_in *) &route->addr.src_addr;
2084         dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2085
2086         switch (ps) {
2087         case RDMA_PS_SDP:
2088                 sdp_hdr = hdr;
2089                 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2090                         return -EINVAL;
2091                 sdp_set_ip_ver(sdp_hdr, 4);
2092                 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2093                 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2094                 sdp_hdr->port = src4->sin_port;
2095                 break;
2096         default:
2097                 cma_hdr = hdr;
2098                 cma_hdr->cma_version = CMA_VERSION;
2099                 cma_set_ip_ver(cma_hdr, 4);
2100                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2101                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2102                 cma_hdr->port = src4->sin_port;
2103                 break;
2104         }
2105         return 0;
2106 }
2107
2108 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2109                                 struct ib_cm_event *ib_event)
2110 {
2111         struct rdma_id_private *id_priv = cm_id->context;
2112         struct rdma_cm_event event;
2113         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2114         int ret = 0;
2115
2116         if (cma_disable_remove(id_priv, CMA_CONNECT))
2117                 return 0;
2118
2119         memset(&event, 0, sizeof event);
2120         switch (ib_event->event) {
2121         case IB_CM_SIDR_REQ_ERROR:
2122                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2123                 event.status = -ETIMEDOUT;
2124                 break;
2125         case IB_CM_SIDR_REP_RECEIVED:
2126                 event.param.ud.private_data = ib_event->private_data;
2127                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2128                 if (rep->status != IB_SIDR_SUCCESS) {
2129                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2130                         event.status = ib_event->param.sidr_rep_rcvd.status;
2131                         break;
2132                 }
2133                 if (id_priv->qkey != rep->qkey) {
2134                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2135                         event.status = -EINVAL;
2136                         break;
2137                 }
2138                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2139                                      id_priv->id.route.path_rec,
2140                                      &event.param.ud.ah_attr);
2141                 event.param.ud.qp_num = rep->qpn;
2142                 event.param.ud.qkey = rep->qkey;
2143                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2144                 event.status = 0;
2145                 break;
2146         default:
2147                 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
2148                        ib_event->event);
2149                 goto out;
2150         }
2151
2152         ret = id_priv->id.event_handler(&id_priv->id, &event);
2153         if (ret) {
2154                 /* Destroy the CM ID by returning a non-zero value. */
2155                 id_priv->cm_id.ib = NULL;
2156                 cma_exch(id_priv, CMA_DESTROYING);
2157                 cma_enable_remove(id_priv);
2158                 rdma_destroy_id(&id_priv->id);
2159                 return ret;
2160         }
2161 out:
2162         cma_enable_remove(id_priv);
2163         return ret;
2164 }
2165
2166 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2167                               struct rdma_conn_param *conn_param)
2168 {
2169         struct ib_cm_sidr_req_param req;
2170         struct rdma_route *route;
2171         int ret;
2172
2173         req.private_data_len = sizeof(struct cma_hdr) +
2174                                conn_param->private_data_len;
2175         req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2176         if (!req.private_data)
2177                 return -ENOMEM;
2178
2179         if (conn_param->private_data && conn_param->private_data_len)
2180                 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2181                        conn_param->private_data, conn_param->private_data_len);
2182
2183         route = &id_priv->id.route;
2184         ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2185         if (ret)
2186                 goto out;
2187
2188         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2189                                             cma_sidr_rep_handler, id_priv);
2190         if (IS_ERR(id_priv->cm_id.ib)) {
2191                 ret = PTR_ERR(id_priv->cm_id.ib);
2192                 goto out;
2193         }
2194
2195         req.path = route->path_rec;
2196         req.service_id = cma_get_service_id(id_priv->id.ps,
2197                                             &route->addr.dst_addr);
2198         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2199         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2200
2201         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2202         if (ret) {
2203                 ib_destroy_cm_id(id_priv->cm_id.ib);
2204                 id_priv->cm_id.ib = NULL;
2205         }
2206 out:
2207         kfree(req.private_data);
2208         return ret;
2209 }
2210
2211 static int cma_connect_ib(struct rdma_id_private *id_priv,
2212                           struct rdma_conn_param *conn_param)
2213 {
2214         struct ib_cm_req_param req;
2215         struct rdma_route *route;
2216         void *private_data;
2217         int offset, ret;
2218
2219         memset(&req, 0, sizeof req);
2220         offset = cma_user_data_offset(id_priv->id.ps);
2221         req.private_data_len = offset + conn_param->private_data_len;
2222         private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2223         if (!private_data)
2224                 return -ENOMEM;
2225
2226         if (conn_param->private_data && conn_param->private_data_len)
2227                 memcpy(private_data + offset, conn_param->private_data,
2228                        conn_param->private_data_len);
2229
2230         id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2231                                             id_priv);
2232         if (IS_ERR(id_priv->cm_id.ib)) {
2233                 ret = PTR_ERR(id_priv->cm_id.ib);
2234                 goto out;
2235         }
2236
2237         route = &id_priv->id.route;
2238         ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2239         if (ret)
2240                 goto out;
2241         req.private_data = private_data;
2242
2243         req.primary_path = &route->path_rec[0];
2244         if (route->num_paths == 2)
2245                 req.alternate_path = &route->path_rec[1];
2246
2247         req.service_id = cma_get_service_id(id_priv->id.ps,
2248                                             &route->addr.dst_addr);
2249         req.qp_num = id_priv->qp_num;
2250         req.qp_type = IB_QPT_RC;
2251         req.starting_psn = id_priv->seq_num;
2252         req.responder_resources = conn_param->responder_resources;
2253         req.initiator_depth = conn_param->initiator_depth;
2254         req.flow_control = conn_param->flow_control;
2255         req.retry_count = conn_param->retry_count;
2256         req.rnr_retry_count = conn_param->rnr_retry_count;
2257         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2258         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2259         req.max_cm_retries = CMA_MAX_CM_RETRIES;
2260         req.srq = id_priv->srq ? 1 : 0;
2261
2262         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2263 out:
2264         if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2265                 ib_destroy_cm_id(id_priv->cm_id.ib);
2266                 id_priv->cm_id.ib = NULL;
2267         }
2268
2269         kfree(private_data);
2270         return ret;
2271 }
2272
2273 static int cma_connect_iw(struct rdma_id_private *id_priv,
2274                           struct rdma_conn_param *conn_param)
2275 {
2276         struct iw_cm_id *cm_id;
2277         struct sockaddr_in* sin;
2278         int ret;
2279         struct iw_cm_conn_param iw_param;
2280
2281         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2282         if (IS_ERR(cm_id)) {
2283                 ret = PTR_ERR(cm_id);
2284                 goto out;
2285         }
2286
2287         id_priv->cm_id.iw = cm_id;
2288
2289         sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2290         cm_id->local_addr = *sin;
2291
2292         sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2293         cm_id->remote_addr = *sin;
2294
2295         ret = cma_modify_qp_rtr(id_priv);
2296         if (ret)
2297                 goto out;
2298
2299         iw_param.ord = conn_param->initiator_depth;
2300         iw_param.ird = conn_param->responder_resources;
2301         iw_param.private_data = conn_param->private_data;
2302         iw_param.private_data_len = conn_param->private_data_len;
2303         if (id_priv->id.qp)
2304                 iw_param.qpn = id_priv->qp_num;
2305         else
2306                 iw_param.qpn = conn_param->qp_num;
2307         ret = iw_cm_connect(cm_id, &iw_param);
2308 out:
2309         if (ret && !IS_ERR(cm_id)) {
2310                 iw_destroy_cm_id(cm_id);
2311                 id_priv->cm_id.iw = NULL;
2312         }
2313         return ret;
2314 }
2315
2316 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2317 {
2318         struct rdma_id_private *id_priv;
2319         int ret;
2320
2321         id_priv = container_of(id, struct rdma_id_private, id);
2322         if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2323                 return -EINVAL;
2324
2325         if (!id->qp) {
2326                 id_priv->qp_num = conn_param->qp_num;
2327                 id_priv->srq = conn_param->srq;
2328         }
2329
2330         switch (rdma_node_get_transport(id->device->node_type)) {
2331         case RDMA_TRANSPORT_IB:
2332                 if (cma_is_ud_ps(id->ps))
2333                         ret = cma_resolve_ib_udp(id_priv, conn_param);
2334                 else
2335                         ret = cma_connect_ib(id_priv, conn_param);
2336                 break;
2337         case RDMA_TRANSPORT_IWARP:
2338                 ret = cma_connect_iw(id_priv, conn_param);
2339                 break;
2340         default:
2341                 ret = -ENOSYS;
2342                 break;
2343         }
2344         if (ret)
2345                 goto err;
2346
2347         return 0;
2348 err:
2349         cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2350         return ret;
2351 }
2352 EXPORT_SYMBOL(rdma_connect);
2353
2354 static int cma_accept_ib(struct rdma_id_private *id_priv,
2355                          struct rdma_conn_param *conn_param)
2356 {
2357         struct ib_cm_rep_param rep;
2358         struct ib_qp_attr qp_attr;
2359         int qp_attr_mask, ret;
2360
2361         if (id_priv->id.qp) {
2362                 ret = cma_modify_qp_rtr(id_priv);
2363                 if (ret)
2364                         goto out;
2365
2366                 qp_attr.qp_state = IB_QPS_RTS;
2367                 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, &qp_attr,
2368                                          &qp_attr_mask);
2369                 if (ret)
2370                         goto out;
2371
2372                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
2373                 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
2374                 if (ret)
2375                         goto out;
2376         }
2377
2378         memset(&rep, 0, sizeof rep);
2379         rep.qp_num = id_priv->qp_num;
2380         rep.starting_psn = id_priv->seq_num;
2381         rep.private_data = conn_param->private_data;
2382         rep.private_data_len = conn_param->private_data_len;
2383         rep.responder_resources = conn_param->responder_resources;
2384         rep.initiator_depth = conn_param->initiator_depth;
2385         rep.failover_accepted = 0;
2386         rep.flow_control = conn_param->flow_control;
2387         rep.rnr_retry_count = conn_param->rnr_retry_count;
2388         rep.srq = id_priv->srq ? 1 : 0;
2389
2390         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2391 out:
2392         return ret;
2393 }
2394
2395 static int cma_accept_iw(struct rdma_id_private *id_priv,
2396                   struct rdma_conn_param *conn_param)
2397 {
2398         struct iw_cm_conn_param iw_param;
2399         int ret;
2400
2401         ret = cma_modify_qp_rtr(id_priv);
2402         if (ret)
2403                 return ret;
2404
2405         iw_param.ord = conn_param->initiator_depth;
2406         iw_param.ird = conn_param->responder_resources;
2407         iw_param.private_data = conn_param->private_data;
2408         iw_param.private_data_len = conn_param->private_data_len;
2409         if (id_priv->id.qp) {
2410                 iw_param.qpn = id_priv->qp_num;
2411         } else
2412                 iw_param.qpn = conn_param->qp_num;
2413
2414         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2415 }
2416
2417 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2418                              enum ib_cm_sidr_status status,
2419                              const void *private_data, int private_data_len)
2420 {
2421         struct ib_cm_sidr_rep_param rep;
2422
2423         memset(&rep, 0, sizeof rep);
2424         rep.status = status;
2425         if (status == IB_SIDR_SUCCESS) {
2426                 rep.qp_num = id_priv->qp_num;
2427                 rep.qkey = id_priv->qkey;
2428         }
2429         rep.private_data = private_data;
2430         rep.private_data_len = private_data_len;
2431
2432         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2433 }
2434
2435 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2436 {
2437         struct rdma_id_private *id_priv;
2438         int ret;
2439
2440         id_priv = container_of(id, struct rdma_id_private, id);
2441         if (!cma_comp(id_priv, CMA_CONNECT))
2442                 return -EINVAL;
2443
2444         if (!id->qp && conn_param) {
2445                 id_priv->qp_num = conn_param->qp_num;
2446                 id_priv->srq = conn_param->srq;
2447         }
2448
2449         switch (rdma_node_get_transport(id->device->node_type)) {
2450         case RDMA_TRANSPORT_IB:
2451                 if (cma_is_ud_ps(id->ps))
2452                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2453                                                 conn_param->private_data,
2454                                                 conn_param->private_data_len);
2455                 else if (conn_param)
2456                         ret = cma_accept_ib(id_priv, conn_param);
2457                 else
2458                         ret = cma_rep_recv(id_priv);
2459                 break;
2460         case RDMA_TRANSPORT_IWARP:
2461                 ret = cma_accept_iw(id_priv, conn_param);
2462                 break;
2463         default:
2464                 ret = -ENOSYS;
2465                 break;
2466         }
2467
2468         if (ret)
2469                 goto reject;
2470
2471         return 0;
2472 reject:
2473         cma_modify_qp_err(id_priv);
2474         rdma_reject(id, NULL, 0);
2475         return ret;
2476 }
2477 EXPORT_SYMBOL(rdma_accept);
2478
2479 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2480 {
2481         struct rdma_id_private *id_priv;
2482         int ret;
2483
2484         id_priv = container_of(id, struct rdma_id_private, id);
2485         if (!cma_has_cm_dev(id_priv))
2486                 return -EINVAL;
2487
2488         switch (id->device->node_type) {
2489         case RDMA_NODE_IB_CA:
2490                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2491                 break;
2492         default:
2493                 ret = 0;
2494                 break;
2495         }
2496         return ret;
2497 }
2498 EXPORT_SYMBOL(rdma_notify);
2499
2500 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2501                 u8 private_data_len)
2502 {
2503         struct rdma_id_private *id_priv;
2504         int ret;
2505
2506         id_priv = container_of(id, struct rdma_id_private, id);
2507         if (!cma_has_cm_dev(id_priv))
2508                 return -EINVAL;
2509
2510         switch (rdma_node_get_transport(id->device->node_type)) {
2511         case RDMA_TRANSPORT_IB:
2512                 if (cma_is_ud_ps(id->ps))
2513                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2514                                                 private_data, private_data_len);
2515                 else
2516                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
2517                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
2518                                              0, private_data, private_data_len);
2519                 break;
2520         case RDMA_TRANSPORT_IWARP:
2521                 ret = iw_cm_reject(id_priv->cm_id.iw,
2522                                    private_data, private_data_len);
2523                 break;
2524         default:
2525                 ret = -ENOSYS;
2526                 break;
2527         }
2528         return ret;
2529 }
2530 EXPORT_SYMBOL(rdma_reject);
2531
2532 int rdma_disconnect(struct rdma_cm_id *id)
2533 {
2534         struct rdma_id_private *id_priv;
2535         int ret;
2536
2537         id_priv = container_of(id, struct rdma_id_private, id);
2538         if (!cma_has_cm_dev(id_priv))
2539                 return -EINVAL;
2540
2541         switch (rdma_node_get_transport(id->device->node_type)) {
2542         case RDMA_TRANSPORT_IB:
2543                 ret = cma_modify_qp_err(id_priv);
2544                 if (ret)
2545                         goto out;
2546                 /* Initiate or respond to a disconnect. */
2547                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2548                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2549                 break;
2550         case RDMA_TRANSPORT_IWARP:
2551                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2552                 break;
2553         default:
2554                 ret = -EINVAL;
2555                 break;
2556         }
2557 out:
2558         return ret;
2559 }
2560 EXPORT_SYMBOL(rdma_disconnect);
2561
2562 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2563 {
2564         struct rdma_id_private *id_priv;
2565         struct cma_multicast *mc = multicast->context;
2566         struct rdma_cm_event event;
2567         int ret;
2568
2569         id_priv = mc->id_priv;
2570         if (cma_disable_remove(id_priv, CMA_ADDR_BOUND) &&
2571             cma_disable_remove(id_priv, CMA_ADDR_RESOLVED))
2572                 return 0;
2573
2574         mutex_lock(&id_priv->qp_mutex);
2575         if (!status && id_priv->id.qp)
2576                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2577                                          multicast->rec.mlid);
2578         mutex_unlock(&id_priv->qp_mutex);
2579
2580         memset(&event, 0, sizeof event);
2581         event.status = status;
2582         event.param.ud.private_data = mc->context;
2583         if (!status) {
2584                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2585                 ib_init_ah_from_mcmember(id_priv->id.device,
2586                                          id_priv->id.port_num, &multicast->rec,
2587                                          &event.param.ud.ah_attr);
2588                 event.param.ud.qp_num = 0xFFFFFF;
2589                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2590         } else
2591                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2592
2593         ret = id_priv->id.event_handler(&id_priv->id, &event);
2594         if (ret) {
2595                 cma_exch(id_priv, CMA_DESTROYING);
2596                 cma_enable_remove(id_priv);
2597                 rdma_destroy_id(&id_priv->id);
2598                 return 0;
2599         }
2600
2601         cma_enable_remove(id_priv);
2602         return 0;
2603 }
2604
2605 static void cma_set_mgid(struct rdma_id_private *id_priv,
2606                          struct sockaddr *addr, union ib_gid *mgid)
2607 {
2608         unsigned char mc_map[MAX_ADDR_LEN];
2609         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2610         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2611         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2612
2613         if (cma_any_addr(addr)) {
2614                 memset(mgid, 0, sizeof *mgid);
2615         } else if ((addr->sa_family == AF_INET6) &&
2616                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2617                                                                  0xFF10A01B)) {
2618                 /* IPv6 address is an SA assigned MGID. */
2619                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2620         } else {
2621                 ip_ib_mc_map(sin->sin_addr.s_addr, mc_map);
2622                 if (id_priv->id.ps == RDMA_PS_UDP)
2623                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
2624                 mc_map[8] = ib_addr_get_pkey(dev_addr) >> 8;
2625                 mc_map[9] = (unsigned char) ib_addr_get_pkey(dev_addr);
2626                 *mgid = *(union ib_gid *) (mc_map + 4);
2627         }
2628 }
2629
2630 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2631                                  struct cma_multicast *mc)
2632 {
2633         struct ib_sa_mcmember_rec rec;
2634         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2635         ib_sa_comp_mask comp_mask;
2636         int ret;
2637
2638         ib_addr_get_mgid(dev_addr, &rec.mgid);
2639         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2640                                      &rec.mgid, &rec);
2641         if (ret)
2642                 return ret;
2643
2644         cma_set_mgid(id_priv, &mc->addr, &rec.mgid);
2645         if (id_priv->id.ps == RDMA_PS_UDP)
2646                 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2647         ib_addr_get_sgid(dev_addr, &rec.port_gid);
2648         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2649         rec.join_state = 1;
2650
2651         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2652                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2653                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2654                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
2655                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2656
2657         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2658                                                 id_priv->id.port_num, &rec,
2659                                                 comp_mask, GFP_KERNEL,
2660                                                 cma_ib_mc_handler, mc);
2661         if (IS_ERR(mc->multicast.ib))
2662                 return PTR_ERR(mc->multicast.ib);
2663
2664         return 0;
2665 }
2666
2667 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2668                         void *context)
2669 {
2670         struct rdma_id_private *id_priv;
2671         struct cma_multicast *mc;
2672         int ret;
2673
2674         id_priv = container_of(id, struct rdma_id_private, id);
2675         if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2676             !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2677                 return -EINVAL;
2678
2679         mc = kmalloc(sizeof *mc, GFP_KERNEL);
2680         if (!mc)
2681                 return -ENOMEM;
2682
2683         memcpy(&mc->addr, addr, ip_addr_size(addr));
2684         mc->context = context;
2685         mc->id_priv = id_priv;
2686
2687         spin_lock(&id_priv->lock);
2688         list_add(&mc->list, &id_priv->mc_list);
2689         spin_unlock(&id_priv->lock);
2690
2691         switch (rdma_node_get_transport(id->device->node_type)) {
2692         case RDMA_TRANSPORT_IB:
2693                 ret = cma_join_ib_multicast(id_priv, mc);
2694                 break;
2695         default:
2696                 ret = -ENOSYS;
2697                 break;
2698         }
2699
2700         if (ret) {
2701                 spin_lock_irq(&id_priv->lock);
2702                 list_del(&mc->list);
2703                 spin_unlock_irq(&id_priv->lock);
2704                 kfree(mc);
2705         }
2706         return ret;
2707 }
2708 EXPORT_SYMBOL(rdma_join_multicast);
2709
2710 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2711 {
2712         struct rdma_id_private *id_priv;
2713         struct cma_multicast *mc;
2714
2715         id_priv = container_of(id, struct rdma_id_private, id);
2716         spin_lock_irq(&id_priv->lock);
2717         list_for_each_entry(mc, &id_priv->mc_list, list) {
2718                 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2719                         list_del(&mc->list);
2720                         spin_unlock_irq(&id_priv->lock);
2721
2722                         if (id->qp)
2723                                 ib_detach_mcast(id->qp,
2724                                                 &mc->multicast.ib->rec.mgid,
2725                                                 mc->multicast.ib->rec.mlid);
2726                         ib_sa_free_multicast(mc->multicast.ib);
2727                         kfree(mc);
2728                         return;
2729                 }
2730         }
2731         spin_unlock_irq(&id_priv->lock);
2732 }
2733 EXPORT_SYMBOL(rdma_leave_multicast);
2734
2735 static void cma_add_one(struct ib_device *device)
2736 {
2737         struct cma_device *cma_dev;
2738         struct rdma_id_private *id_priv;
2739
2740         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2741         if (!cma_dev)
2742                 return;
2743
2744         cma_dev->device = device;
2745
2746         init_completion(&cma_dev->comp);
2747         atomic_set(&cma_dev->refcount, 1);
2748         INIT_LIST_HEAD(&cma_dev->id_list);
2749         ib_set_client_data(device, &cma_client, cma_dev);
2750
2751         mutex_lock(&lock);
2752         list_add_tail(&cma_dev->list, &dev_list);
2753         list_for_each_entry(id_priv, &listen_any_list, list)
2754                 cma_listen_on_dev(id_priv, cma_dev);
2755         mutex_unlock(&lock);
2756 }
2757
2758 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2759 {
2760         struct rdma_cm_event event;
2761         enum cma_state state;
2762
2763         /* Record that we want to remove the device */
2764         state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2765         if (state == CMA_DESTROYING)
2766                 return 0;
2767
2768         cma_cancel_operation(id_priv, state);
2769         wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2770
2771         /* Check for destruction from another callback. */
2772         if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2773                 return 0;
2774
2775         memset(&event, 0, sizeof event);
2776         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2777         return id_priv->id.event_handler(&id_priv->id, &event);
2778 }
2779
2780 static void cma_process_remove(struct cma_device *cma_dev)
2781 {
2782         struct rdma_id_private *id_priv;
2783         int ret;
2784
2785         mutex_lock(&lock);
2786         while (!list_empty(&cma_dev->id_list)) {
2787                 id_priv = list_entry(cma_dev->id_list.next,
2788                                      struct rdma_id_private, list);
2789
2790                 if (cma_internal_listen(id_priv)) {
2791                         cma_destroy_listen(id_priv);
2792                         continue;
2793                 }
2794
2795                 list_del_init(&id_priv->list);
2796                 atomic_inc(&id_priv->refcount);
2797                 mutex_unlock(&lock);
2798
2799                 ret = cma_remove_id_dev(id_priv);
2800                 cma_deref_id(id_priv);
2801                 if (ret)
2802                         rdma_destroy_id(&id_priv->id);
2803
2804                 mutex_lock(&lock);
2805         }
2806         mutex_unlock(&lock);
2807
2808         cma_deref_dev(cma_dev);
2809         wait_for_completion(&cma_dev->comp);
2810 }
2811
2812 static void cma_remove_one(struct ib_device *device)
2813 {
2814         struct cma_device *cma_dev;
2815
2816         cma_dev = ib_get_client_data(device, &cma_client);
2817         if (!cma_dev)
2818                 return;
2819
2820         mutex_lock(&lock);
2821         list_del(&cma_dev->list);
2822         mutex_unlock(&lock);
2823
2824         cma_process_remove(cma_dev);
2825         kfree(cma_dev);
2826 }
2827
2828 static int cma_init(void)
2829 {
2830         int ret, low, high;
2831
2832         get_random_bytes(&next_port, sizeof next_port);
2833         inet_get_local_port_range(&low, &high);
2834         next_port = ((unsigned int) next_port % (high - low)) + low;
2835
2836         cma_wq = create_singlethread_workqueue("rdma_cm");
2837         if (!cma_wq)
2838                 return -ENOMEM;
2839
2840         ib_sa_register_client(&sa_client);
2841         rdma_addr_register_client(&addr_client);
2842
2843         ret = ib_register_client(&cma_client);
2844         if (ret)
2845                 goto err;
2846         return 0;
2847
2848 err:
2849         rdma_addr_unregister_client(&addr_client);
2850         ib_sa_unregister_client(&sa_client);
2851         destroy_workqueue(cma_wq);
2852         return ret;
2853 }
2854
2855 static void cma_cleanup(void)
2856 {
2857         ib_unregister_client(&cma_client);
2858         rdma_addr_unregister_client(&addr_client);
2859         ib_sa_unregister_client(&sa_client);
2860         destroy_workqueue(cma_wq);
2861         idr_destroy(&sdp_ps);
2862         idr_destroy(&tcp_ps);
2863         idr_destroy(&udp_ps);
2864         idr_destroy(&ipoib_ps);
2865 }
2866
2867 module_init(cma_init);
2868 module_exit(cma_cleanup);