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