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