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 available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/igmp.h>
  42 #include <linux/xarray.h>
  43 #include <linux/inetdevice.h>
  44 #include <linux/slab.h>
  45 #include <linux/module.h>
  46 #include <net/route.h>
  47
  48 #include <net/net_namespace.h>
  49 #include <net/netns/generic.h>
  50 #include <net/tcp.h>
  51 #include <net/ipv6.h>
  52 #include <net/ip_fib.h>
  53 #include <net/ip6_route.h>
  54
  55 #include <rdma/rdma_cm.h>
  56 #include <rdma/rdma_cm_ib.h>
  57 #include <rdma/rdma_netlink.h>
  58 #include <rdma/ib.h>
  59 #include <rdma/ib_cache.h>
  60 #include <rdma/ib_cm.h>
  61 #include <rdma/ib_sa.h>
  62 #include <rdma/iw_cm.h>
  63
  64 #include "core_priv.h"
  65 #include "cma_priv.h"
  66
  67 MODULE_AUTHOR("Sean Hefty");
  68 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  69 MODULE_LICENSE("Dual BSD/GPL");
  70
  71 #define CMA_CM_RESPONSE_TIMEOUT 20
  72 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  73 #define CMA_MAX_CM_RETRIES 15
  74 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  75 #define CMA_IBOE_PACKET_LIFETIME 18
  76 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  77
  78 static const char * const cma_events[] = {
  79         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  80         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  81         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  82         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  83         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  84         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  85         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  86         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  87         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  88         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  89         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  90         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  91         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  92         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  93         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  94         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  95 };
  96
  97 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  98 {
  99         size_t index = event;
 100
 101         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
 102                         cma_events[index] : "unrecognized event";
 103 }
 104 EXPORT_SYMBOL(rdma_event_msg);
 105
 106 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
 107                                                 int reason)
 108 {
 109         if (rdma_ib_or_roce(id->device, id->port_num))
 110                 return ibcm_reject_msg(reason);
 111
 112         if (rdma_protocol_iwarp(id->device, id->port_num))
 113                 return iwcm_reject_msg(reason);
 114
 115         WARN_ON_ONCE(1);
 116         return "unrecognized transport";
 117 }
 118 EXPORT_SYMBOL(rdma_reject_msg);
 119
 120 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 121 {
 122         if (rdma_ib_or_roce(id->device, id->port_num))
 123                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 124
 125         if (rdma_protocol_iwarp(id->device, id->port_num))
 126                 return reason == -ECONNREFUSED;
 127
 128         WARN_ON_ONCE(1);
 129         return false;
 130 }
 131 EXPORT_SYMBOL(rdma_is_consumer_reject);
 132
 133 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 134                                       struct rdma_cm_event *ev, u8 *data_len)
 135 {
 136         const void *p;
 137
 138         if (rdma_is_consumer_reject(id, ev->status)) {
 139                 *data_len = ev->param.conn.private_data_len;
 140                 p = ev->param.conn.private_data;
 141         } else {
 142                 *data_len = 0;
 143                 p = NULL;
 144         }
 145         return p;
 146 }
 147 EXPORT_SYMBOL(rdma_consumer_reject_data);
 148
 149 /**
 150  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 151  * @id: Communication Identifier
 152  */
 153 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 154 {
 155         struct rdma_id_private *id_priv;
 156
 157         id_priv = container_of(id, struct rdma_id_private, id);
 158         if (id->device->node_type == RDMA_NODE_RNIC)
 159                 return id_priv->cm_id.iw;
 160         return NULL;
 161 }
 162 EXPORT_SYMBOL(rdma_iw_cm_id);
 163
 164 /**
 165  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 166  * @res: rdma resource tracking entry pointer
 167  */
 168 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 169 {
 170         struct rdma_id_private *id_priv =
 171                 container_of(res, struct rdma_id_private, res);
 172
 173         return &id_priv->id;
 174 }
 175 EXPORT_SYMBOL(rdma_res_to_id);
 176
 177 static void cma_add_one(struct ib_device *device);
 178 static void cma_remove_one(struct ib_device *device, void *client_data);
 179
 180 static struct ib_client cma_client = {
 181         .name   = "cma",
 182         .add    = cma_add_one,
 183         .remove = cma_remove_one
 184 };
 185
 186 static struct ib_sa_client sa_client;
 187 static LIST_HEAD(dev_list);
 188 static LIST_HEAD(listen_any_list);
 189 static DEFINE_MUTEX(lock);
 190 static struct workqueue_struct *cma_wq;
 191 static unsigned int cma_pernet_id;
 192
 193 struct cma_pernet {
 194         struct xarray tcp_ps;
 195         struct xarray udp_ps;
 196         struct xarray ipoib_ps;
 197         struct xarray ib_ps;
 198 };
 199
 200 static struct cma_pernet *cma_pernet(struct net *net)
 201 {
 202         return net_generic(net, cma_pernet_id);
 203 }
 204
 205 static
 206 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 207 {
 208         struct cma_pernet *pernet = cma_pernet(net);
 209
 210         switch (ps) {
 211         case RDMA_PS_TCP:
 212                 return &pernet->tcp_ps;
 213         case RDMA_PS_UDP:
 214                 return &pernet->udp_ps;
 215         case RDMA_PS_IPOIB:
 216                 return &pernet->ipoib_ps;
 217         case RDMA_PS_IB:
 218                 return &pernet->ib_ps;
 219         default:
 220                 return NULL;
 221         }
 222 }
 223
 224 struct cma_device {
 225         struct list_head        list;
 226         struct ib_device        *device;
 227         struct completion       comp;
 228         atomic_t                refcount;
 229         struct list_head        id_list;
 230         enum ib_gid_type        *default_gid_type;
 231         u8                      *default_roce_tos;
 232 };
 233
 234 struct rdma_bind_list {
 235         enum rdma_ucm_port_space ps;
 236         struct hlist_head       owners;
 237         unsigned short          port;
 238 };
 239
 240 struct class_port_info_context {
 241         struct ib_class_port_info       *class_port_info;
 242         struct ib_device                *device;
 243         struct completion               done;
 244         struct ib_sa_query              *sa_query;
 245         u8                              port_num;
 246 };
 247
 248 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 249                         struct rdma_bind_list *bind_list, int snum)
 250 {
 251         struct xarray *xa = cma_pernet_xa(net, ps);
 252
 253         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 254 }
 255
 256 static struct rdma_bind_list *cma_ps_find(struct net *net,
 257                                           enum rdma_ucm_port_space ps, int snum)
 258 {
 259         struct xarray *xa = cma_pernet_xa(net, ps);
 260
 261         return xa_load(xa, snum);
 262 }
 263
 264 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 265                           int snum)
 266 {
 267         struct xarray *xa = cma_pernet_xa(net, ps);
 268
 269         xa_erase(xa, snum);
 270 }
 271
 272 enum {
 273         CMA_OPTION_AFONLY,
 274 };
 275
 276 void cma_ref_dev(struct cma_device *cma_dev)
 277 {
 278         atomic_inc(&cma_dev->refcount);
 279 }
 280
 281 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 282                                              void               *cookie)
 283 {
 284         struct cma_device *cma_dev;
 285         struct cma_device *found_cma_dev = NULL;
 286
 287         mutex_lock(&lock);
 288
 289         list_for_each_entry(cma_dev, &dev_list, list)
 290                 if (filter(cma_dev->device, cookie)) {
 291                         found_cma_dev = cma_dev;
 292                         break;
 293                 }
 294
 295         if (found_cma_dev)
 296                 cma_ref_dev(found_cma_dev);
 297         mutex_unlock(&lock);
 298         return found_cma_dev;
 299 }
 300
 301 int cma_get_default_gid_type(struct cma_device *cma_dev,
 302                              unsigned int port)
 303 {
 304         if (!rdma_is_port_valid(cma_dev->device, port))
 305                 return -EINVAL;
 306
 307         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 308 }
 309
 310 int cma_set_default_gid_type(struct cma_device *cma_dev,
 311                              unsigned int port,
 312                              enum ib_gid_type default_gid_type)
 313 {
 314         unsigned long supported_gids;
 315
 316         if (!rdma_is_port_valid(cma_dev->device, port))
 317                 return -EINVAL;
 318
 319         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 320
 321         if (!(supported_gids & 1 << default_gid_type))
 322                 return -EINVAL;
 323
 324         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 325                 default_gid_type;
 326
 327         return 0;
 328 }
 329
 330 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 331 {
 332         if (!rdma_is_port_valid(cma_dev->device, port))
 333                 return -EINVAL;
 334
 335         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 336 }
 337
 338 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 339                              u8 default_roce_tos)
 340 {
 341         if (!rdma_is_port_valid(cma_dev->device, port))
 342                 return -EINVAL;
 343
 344         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 345                  default_roce_tos;
 346
 347         return 0;
 348 }
 349 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 350 {
 351         return cma_dev->device;
 352 }
 353
 354 /*
 355  * Device removal can occur at anytime, so we need extra handling to
 356  * serialize notifying the user of device removal with other callbacks.
 357  * We do this by disabling removal notification while a callback is in process,
 358  * and reporting it after the callback completes.
 359  */
 360
 361 struct cma_multicast {
 362         struct rdma_id_private *id_priv;
 363         union {
 364                 struct ib_sa_multicast *ib;
 365         } multicast;
 366         struct list_head        list;
 367         void                    *context;
 368         struct sockaddr_storage addr;
 369         struct kref             mcref;
 370         u8                      join_state;
 371 };
 372
 373 struct cma_work {
 374         struct work_struct      work;
 375         struct rdma_id_private  *id;
 376         enum rdma_cm_state      old_state;
 377         enum rdma_cm_state      new_state;
 378         struct rdma_cm_event    event;
 379 };
 380
 381 struct cma_ndev_work {
 382         struct work_struct      work;
 383         struct rdma_id_private  *id;
 384         struct rdma_cm_event    event;
 385 };
 386
 387 struct iboe_mcast_work {
 388         struct work_struct       work;
 389         struct rdma_id_private  *id;
 390         struct cma_multicast    *mc;
 391 };
 392
 393 union cma_ip_addr {
 394         struct in6_addr ip6;
 395         struct {
 396                 __be32 pad[3];
 397                 __be32 addr;
 398         } ip4;
 399 };
 400
 401 struct cma_hdr {
 402         u8 cma_version;
 403         u8 ip_version;  /* IP version: 7:4 */
 404         __be16 port;
 405         union cma_ip_addr src_addr;
 406         union cma_ip_addr dst_addr;
 407 };
 408
 409 #define CMA_VERSION 0x00
 410
 411 struct cma_req_info {
 412         struct sockaddr_storage listen_addr_storage;
 413         struct sockaddr_storage src_addr_storage;
 414         struct ib_device *device;
 415         union ib_gid local_gid;
 416         __be64 service_id;
 417         int port;
 418         bool has_gid;
 419         u16 pkey;
 420 };
 421
 422 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 423 {
 424         unsigned long flags;
 425         int ret;
 426
 427         spin_lock_irqsave(&id_priv->lock, flags);
 428         ret = (id_priv->state == comp);
 429         spin_unlock_irqrestore(&id_priv->lock, flags);
 430         return ret;
 431 }
 432
 433 static int cma_comp_exch(struct rdma_id_private *id_priv,
 434                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 435 {
 436         unsigned long flags;
 437         int ret;
 438
 439         spin_lock_irqsave(&id_priv->lock, flags);
 440         if ((ret = (id_priv->state == comp)))
 441                 id_priv->state = exch;
 442         spin_unlock_irqrestore(&id_priv->lock, flags);
 443         return ret;
 444 }
 445
 446 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 447                                    enum rdma_cm_state exch)
 448 {
 449         unsigned long flags;
 450         enum rdma_cm_state old;
 451
 452         spin_lock_irqsave(&id_priv->lock, flags);
 453         old = id_priv->state;
 454         id_priv->state = exch;
 455         spin_unlock_irqrestore(&id_priv->lock, flags);
 456         return old;
 457 }
 458
 459 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 460 {
 461         return hdr->ip_version >> 4;
 462 }
 463
 464 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 465 {
 466         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 467 }
 468
 469 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 470 {
 471         struct in_device *in_dev = NULL;
 472
 473         if (ndev) {
 474                 rtnl_lock();
 475                 in_dev = __in_dev_get_rtnl(ndev);
 476                 if (in_dev) {
 477                         if (join)
 478                                 ip_mc_inc_group(in_dev,
 479                                                 *(__be32 *)(mgid->raw + 12));
 480                         else
 481                                 ip_mc_dec_group(in_dev,
 482                                                 *(__be32 *)(mgid->raw + 12));
 483                 }
 484                 rtnl_unlock();
 485         }
 486         return (in_dev) ? 0 : -ENODEV;
 487 }
 488
 489 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 490                                struct cma_device *cma_dev)
 491 {
 492         cma_ref_dev(cma_dev);
 493         id_priv->cma_dev = cma_dev;
 494         id_priv->id.device = cma_dev->device;
 495         id_priv->id.route.addr.dev_addr.transport =
 496                 rdma_node_get_transport(cma_dev->device->node_type);
 497         list_add_tail(&id_priv->list, &cma_dev->id_list);
 498         if (id_priv->res.kern_name)
 499                 rdma_restrack_kadd(&id_priv->res);
 500         else
 501                 rdma_restrack_uadd(&id_priv->res);
 502 }
 503
 504 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 505                               struct cma_device *cma_dev)
 506 {
 507         _cma_attach_to_dev(id_priv, cma_dev);
 508         id_priv->gid_type =
 509                 cma_dev->default_gid_type[id_priv->id.port_num -
 510                                           rdma_start_port(cma_dev->device)];
 511 }
 512
 513 void cma_deref_dev(struct cma_device *cma_dev)
 514 {
 515         if (atomic_dec_and_test(&cma_dev->refcount))
 516                 complete(&cma_dev->comp);
 517 }
 518
 519 static inline void release_mc(struct kref *kref)
 520 {
 521         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 522
 523         kfree(mc->multicast.ib);
 524         kfree(mc);
 525 }
 526
 527 static void cma_release_dev(struct rdma_id_private *id_priv)
 528 {
 529         mutex_lock(&lock);
 530         list_del(&id_priv->list);
 531         cma_deref_dev(id_priv->cma_dev);
 532         id_priv->cma_dev = NULL;
 533         mutex_unlock(&lock);
 534 }
 535
 536 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 537 {
 538         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 539 }
 540
 541 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 542 {
 543         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 544 }
 545
 546 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 547 {
 548         return id_priv->id.route.addr.src_addr.ss_family;
 549 }
 550
 551 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 552 {
 553         struct ib_sa_mcmember_rec rec;
 554         int ret = 0;
 555
 556         if (id_priv->qkey) {
 557                 if (qkey && id_priv->qkey != qkey)
 558                         return -EINVAL;
 559                 return 0;
 560         }
 561
 562         if (qkey) {
 563                 id_priv->qkey = qkey;
 564                 return 0;
 565         }
 566
 567         switch (id_priv->id.ps) {
 568         case RDMA_PS_UDP:
 569         case RDMA_PS_IB:
 570                 id_priv->qkey = RDMA_UDP_QKEY;
 571                 break;
 572         case RDMA_PS_IPOIB:
 573                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 574                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 575                                              id_priv->id.port_num, &rec.mgid,
 576                                              &rec);
 577                 if (!ret)
 578                         id_priv->qkey = be32_to_cpu(rec.qkey);
 579                 break;
 580         default:
 581                 break;
 582         }
 583         return ret;
 584 }
 585
 586 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 587 {
 588         dev_addr->dev_type = ARPHRD_INFINIBAND;
 589         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 590         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 591 }
 592
 593 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 594 {
 595         int ret;
 596
 597         if (addr->sa_family != AF_IB) {
 598                 ret = rdma_translate_ip(addr, dev_addr);
 599         } else {
 600                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 601                 ret = 0;
 602         }
 603
 604         return ret;
 605 }
 606
 607 static const struct ib_gid_attr *
 608 cma_validate_port(struct ib_device *device, u8 port,
 609                   enum ib_gid_type gid_type,
 610                   union ib_gid *gid,
 611                   struct rdma_id_private *id_priv)
 612 {
 613         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 614         int bound_if_index = dev_addr->bound_dev_if;
 615         const struct ib_gid_attr *sgid_attr;
 616         int dev_type = dev_addr->dev_type;
 617         struct net_device *ndev = NULL;
 618
 619         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 620                 return ERR_PTR(-ENODEV);
 621
 622         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 623                 return ERR_PTR(-ENODEV);
 624
 625         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 626                 return ERR_PTR(-ENODEV);
 627
 628         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 629                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 630                 if (!ndev)
 631                         return ERR_PTR(-ENODEV);
 632         } else {
 633                 gid_type = IB_GID_TYPE_IB;
 634         }
 635
 636         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 637         if (ndev)
 638                 dev_put(ndev);
 639         return sgid_attr;
 640 }
 641
 642 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 643                                const struct ib_gid_attr *sgid_attr)
 644 {
 645         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 646         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 647 }
 648
 649 /**
 650  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 651  * based on source ip address.
 652  * @id_priv:    cm_id which should be bound to cma device
 653  *
 654  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 655  * based on source IP address. It returns 0 on success or error code otherwise.
 656  * It is applicable to active and passive side cm_id.
 657  */
 658 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 659 {
 660         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 661         const struct ib_gid_attr *sgid_attr;
 662         union ib_gid gid, iboe_gid, *gidp;
 663         struct cma_device *cma_dev;
 664         enum ib_gid_type gid_type;
 665         int ret = -ENODEV;
 666         unsigned int port;
 667
 668         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 669             id_priv->id.ps == RDMA_PS_IPOIB)
 670                 return -EINVAL;
 671
 672         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 673                     &iboe_gid);
 674
 675         memcpy(&gid, dev_addr->src_dev_addr +
 676                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 677
 678         mutex_lock(&lock);
 679         list_for_each_entry(cma_dev, &dev_list, list) {
 680                 rdma_for_each_port (cma_dev->device, port) {
 681                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 682                                &iboe_gid : &gid;
 683                         gid_type = cma_dev->default_gid_type[port - 1];
 684                         sgid_attr = cma_validate_port(cma_dev->device, port,
 685                                                       gid_type, gidp, id_priv);
 686                         if (!IS_ERR(sgid_attr)) {
 687                                 id_priv->id.port_num = port;
 688                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 689                                 cma_attach_to_dev(id_priv, cma_dev);
 690                                 ret = 0;
 691                                 goto out;
 692                         }
 693                 }
 694         }
 695 out:
 696         mutex_unlock(&lock);
 697         return ret;
 698 }
 699
 700 /**
 701  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 702  * @id_priv:            cm id to bind to cma device
 703  * @listen_id_priv:     listener cm id to match against
 704  * @req:                Pointer to req structure containaining incoming
 705  *                      request information
 706  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 707  * rdma device matches for listen_id and incoming request. It also verifies
 708  * that a GID table entry is present for the source address.
 709  * Returns 0 on success, or returns error code otherwise.
 710  */
 711 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 712                               const struct rdma_id_private *listen_id_priv,
 713                               struct cma_req_info *req)
 714 {
 715         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 716         const struct ib_gid_attr *sgid_attr;
 717         enum ib_gid_type gid_type;
 718         union ib_gid gid;
 719
 720         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 721             id_priv->id.ps == RDMA_PS_IPOIB)
 722                 return -EINVAL;
 723
 724         if (rdma_protocol_roce(req->device, req->port))
 725                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 726                             &gid);
 727         else
 728                 memcpy(&gid, dev_addr->src_dev_addr +
 729                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 730
 731         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 732         sgid_attr = cma_validate_port(req->device, req->port,
 733                                       gid_type, &gid, id_priv);
 734         if (IS_ERR(sgid_attr))
 735                 return PTR_ERR(sgid_attr);
 736
 737         id_priv->id.port_num = req->port;
 738         cma_bind_sgid_attr(id_priv, sgid_attr);
 739         /* Need to acquire lock to protect against reader
 740          * of cma_dev->id_list such as cma_netdev_callback() and
 741          * cma_process_remove().
 742          */
 743         mutex_lock(&lock);
 744         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 745         mutex_unlock(&lock);
 746         return 0;
 747 }
 748
 749 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 750                               const struct rdma_id_private *listen_id_priv)
 751 {
 752         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 753         const struct ib_gid_attr *sgid_attr;
 754         struct cma_device *cma_dev;
 755         enum ib_gid_type gid_type;
 756         int ret = -ENODEV;
 757         union ib_gid gid;
 758         u8 port;
 759
 760         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 761             id_priv->id.ps == RDMA_PS_IPOIB)
 762                 return -EINVAL;
 763
 764         memcpy(&gid, dev_addr->src_dev_addr +
 765                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 766
 767         mutex_lock(&lock);
 768
 769         cma_dev = listen_id_priv->cma_dev;
 770         port = listen_id_priv->id.port_num;
 771         gid_type = listen_id_priv->gid_type;
 772         sgid_attr = cma_validate_port(cma_dev->device, port,
 773                                       gid_type, &gid, id_priv);
 774         if (!IS_ERR(sgid_attr)) {
 775                 id_priv->id.port_num = port;
 776                 cma_bind_sgid_attr(id_priv, sgid_attr);
 777                 ret = 0;
 778                 goto out;
 779         }
 780
 781         list_for_each_entry(cma_dev, &dev_list, list) {
 782                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 783                         if (listen_id_priv->cma_dev == cma_dev &&
 784                             listen_id_priv->id.port_num == port)
 785                                 continue;
 786
 787                         gid_type = cma_dev->default_gid_type[port - 1];
 788                         sgid_attr = cma_validate_port(cma_dev->device, port,
 789                                                       gid_type, &gid, id_priv);
 790                         if (!IS_ERR(sgid_attr)) {
 791                                 id_priv->id.port_num = port;
 792                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 793                                 ret = 0;
 794                                 goto out;
 795                         }
 796                 }
 797         }
 798
 799 out:
 800         if (!ret)
 801                 cma_attach_to_dev(id_priv, cma_dev);
 802
 803         mutex_unlock(&lock);
 804         return ret;
 805 }
 806
 807 /*
 808  * Select the source IB device and address to reach the destination IB address.
 809  */
 810 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 811 {
 812         struct cma_device *cma_dev, *cur_dev;
 813         struct sockaddr_ib *addr;
 814         union ib_gid gid, sgid, *dgid;
 815         u16 pkey, index;
 816         u8 p;
 817         enum ib_port_state port_state;
 818         int i;
 819
 820         cma_dev = NULL;
 821         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 822         dgid = (union ib_gid *) &addr->sib_addr;
 823         pkey = ntohs(addr->sib_pkey);
 824
 825         mutex_lock(&lock);
 826         list_for_each_entry(cur_dev, &dev_list, list) {
 827                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 828                         if (!rdma_cap_af_ib(cur_dev->device, p))
 829                                 continue;
 830
 831                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 832                                 continue;
 833
 834                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 835                                 continue;
 836                         for (i = 0; !rdma_query_gid(cur_dev->device,
 837                                                     p, i, &gid);
 838                              i++) {
 839                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 840                                         cma_dev = cur_dev;
 841                                         sgid = gid;
 842                                         id_priv->id.port_num = p;
 843                                         goto found;
 844                                 }
 845
 846                                 if (!cma_dev && (gid.global.subnet_prefix ==
 847                                     dgid->global.subnet_prefix) &&
 848                                     port_state == IB_PORT_ACTIVE) {
 849                                         cma_dev = cur_dev;
 850                                         sgid = gid;
 851                                         id_priv->id.port_num = p;
 852                                         goto found;
 853                                 }
 854                         }
 855                 }
 856         }
 857         mutex_unlock(&lock);
 858         return -ENODEV;
 859
 860 found:
 861         cma_attach_to_dev(id_priv, cma_dev);
 862         mutex_unlock(&lock);
 863         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 864         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 865         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 866         return 0;
 867 }
 868
 869 static void cma_deref_id(struct rdma_id_private *id_priv)
 870 {
 871         if (atomic_dec_and_test(&id_priv->refcount))
 872                 complete(&id_priv->comp);
 873 }
 874
 875 struct rdma_cm_id *__rdma_create_id(struct net *net,
 876                                     rdma_cm_event_handler event_handler,
 877                                     void *context, enum rdma_ucm_port_space ps,
 878                                     enum ib_qp_type qp_type, const char *caller)
 879 {
 880         struct rdma_id_private *id_priv;
 881
 882         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 883         if (!id_priv)
 884                 return ERR_PTR(-ENOMEM);
 885
 886         rdma_restrack_set_task(&id_priv->res, caller);
 887         id_priv->res.type = RDMA_RESTRACK_CM_ID;
 888         id_priv->state = RDMA_CM_IDLE;
 889         id_priv->id.context = context;
 890         id_priv->id.event_handler = event_handler;
 891         id_priv->id.ps = ps;
 892         id_priv->id.qp_type = qp_type;
 893         id_priv->tos_set = false;
 894         id_priv->timeout_set = false;
 895         id_priv->gid_type = IB_GID_TYPE_IB;
 896         spin_lock_init(&id_priv->lock);
 897         mutex_init(&id_priv->qp_mutex);
 898         init_completion(&id_priv->comp);
 899         atomic_set(&id_priv->refcount, 1);
 900         mutex_init(&id_priv->handler_mutex);
 901         INIT_LIST_HEAD(&id_priv->listen_list);
 902         INIT_LIST_HEAD(&id_priv->mc_list);
 903         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 904         id_priv->id.route.addr.dev_addr.net = get_net(net);
 905         id_priv->seq_num &= 0x00ffffff;
 906
 907         return &id_priv->id;
 908 }
 909 EXPORT_SYMBOL(__rdma_create_id);
 910
 911 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 912 {
 913         struct ib_qp_attr qp_attr;
 914         int qp_attr_mask, ret;
 915
 916         qp_attr.qp_state = IB_QPS_INIT;
 917         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 918         if (ret)
 919                 return ret;
 920
 921         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 922         if (ret)
 923                 return ret;
 924
 925         qp_attr.qp_state = IB_QPS_RTR;
 926         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 927         if (ret)
 928                 return ret;
 929
 930         qp_attr.qp_state = IB_QPS_RTS;
 931         qp_attr.sq_psn = 0;
 932         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 933
 934         return ret;
 935 }
 936
 937 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 938 {
 939         struct ib_qp_attr qp_attr;
 940         int qp_attr_mask, ret;
 941
 942         qp_attr.qp_state = IB_QPS_INIT;
 943         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 944         if (ret)
 945                 return ret;
 946
 947         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 948 }
 949
 950 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 951                    struct ib_qp_init_attr *qp_init_attr)
 952 {
 953         struct rdma_id_private *id_priv;
 954         struct ib_qp *qp;
 955         int ret;
 956
 957         id_priv = container_of(id, struct rdma_id_private, id);
 958         if (id->device != pd->device)
 959                 return -EINVAL;
 960
 961         qp_init_attr->port_num = id->port_num;
 962         qp = ib_create_qp(pd, qp_init_attr);
 963         if (IS_ERR(qp))
 964                 return PTR_ERR(qp);
 965
 966         if (id->qp_type == IB_QPT_UD)
 967                 ret = cma_init_ud_qp(id_priv, qp);
 968         else
 969                 ret = cma_init_conn_qp(id_priv, qp);
 970         if (ret)
 971                 goto err;
 972
 973         id->qp = qp;
 974         id_priv->qp_num = qp->qp_num;
 975         id_priv->srq = (qp->srq != NULL);
 976         return 0;
 977 err:
 978         ib_destroy_qp(qp);
 979         return ret;
 980 }
 981 EXPORT_SYMBOL(rdma_create_qp);
 982
 983 void rdma_destroy_qp(struct rdma_cm_id *id)
 984 {
 985         struct rdma_id_private *id_priv;
 986
 987         id_priv = container_of(id, struct rdma_id_private, id);
 988         mutex_lock(&id_priv->qp_mutex);
 989         ib_destroy_qp(id_priv->id.qp);
 990         id_priv->id.qp = NULL;
 991         mutex_unlock(&id_priv->qp_mutex);
 992 }
 993 EXPORT_SYMBOL(rdma_destroy_qp);
 994
 995 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 996                              struct rdma_conn_param *conn_param)
 997 {
 998         struct ib_qp_attr qp_attr;
 999         int qp_attr_mask, ret;
1000
1001         mutex_lock(&id_priv->qp_mutex);
1002         if (!id_priv->id.qp) {
1003                 ret = 0;
1004                 goto out;
1005         }
1006
1007         /* Need to update QP attributes from default values. */
1008         qp_attr.qp_state = IB_QPS_INIT;
1009         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1010         if (ret)
1011                 goto out;
1012
1013         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1014         if (ret)
1015                 goto out;
1016
1017         qp_attr.qp_state = IB_QPS_RTR;
1018         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1019         if (ret)
1020                 goto out;
1021
1022         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1023
1024         if (conn_param)
1025                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1026         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1027 out:
1028         mutex_unlock(&id_priv->qp_mutex);
1029         return ret;
1030 }
1031
1032 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1033                              struct rdma_conn_param *conn_param)
1034 {
1035         struct ib_qp_attr qp_attr;
1036         int qp_attr_mask, ret;
1037
1038         mutex_lock(&id_priv->qp_mutex);
1039         if (!id_priv->id.qp) {
1040                 ret = 0;
1041                 goto out;
1042         }
1043
1044         qp_attr.qp_state = IB_QPS_RTS;
1045         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1046         if (ret)
1047                 goto out;
1048
1049         if (conn_param)
1050                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1051         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1052 out:
1053         mutex_unlock(&id_priv->qp_mutex);
1054         return ret;
1055 }
1056
1057 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1058 {
1059         struct ib_qp_attr qp_attr;
1060         int ret;
1061
1062         mutex_lock(&id_priv->qp_mutex);
1063         if (!id_priv->id.qp) {
1064                 ret = 0;
1065                 goto out;
1066         }
1067
1068         qp_attr.qp_state = IB_QPS_ERR;
1069         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1070 out:
1071         mutex_unlock(&id_priv->qp_mutex);
1072         return ret;
1073 }
1074
1075 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1076                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1077 {
1078         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1079         int ret;
1080         u16 pkey;
1081
1082         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1083                 pkey = 0xffff;
1084         else
1085                 pkey = ib_addr_get_pkey(dev_addr);
1086
1087         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1088                                   pkey, &qp_attr->pkey_index);
1089         if (ret)
1090                 return ret;
1091
1092         qp_attr->port_num = id_priv->id.port_num;
1093         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1094
1095         if (id_priv->id.qp_type == IB_QPT_UD) {
1096                 ret = cma_set_qkey(id_priv, 0);
1097                 if (ret)
1098                         return ret;
1099
1100                 qp_attr->qkey = id_priv->qkey;
1101                 *qp_attr_mask |= IB_QP_QKEY;
1102         } else {
1103                 qp_attr->qp_access_flags = 0;
1104                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1105         }
1106         return 0;
1107 }
1108
1109 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1110                        int *qp_attr_mask)
1111 {
1112         struct rdma_id_private *id_priv;
1113         int ret = 0;
1114
1115         id_priv = container_of(id, struct rdma_id_private, id);
1116         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1117                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1118                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1119                 else
1120                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1121                                                  qp_attr_mask);
1122
1123                 if (qp_attr->qp_state == IB_QPS_RTR)
1124                         qp_attr->rq_psn = id_priv->seq_num;
1125         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1126                 if (!id_priv->cm_id.iw) {
1127                         qp_attr->qp_access_flags = 0;
1128                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1129                 } else
1130                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1131                                                  qp_attr_mask);
1132                 qp_attr->port_num = id_priv->id.port_num;
1133                 *qp_attr_mask |= IB_QP_PORT;
1134         } else
1135                 ret = -ENOSYS;
1136
1137         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1138                 qp_attr->timeout = id_priv->timeout;
1139
1140         return ret;
1141 }
1142 EXPORT_SYMBOL(rdma_init_qp_attr);
1143
1144 static inline bool cma_zero_addr(const struct sockaddr *addr)
1145 {
1146         switch (addr->sa_family) {
1147         case AF_INET:
1148                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1149         case AF_INET6:
1150                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1151         case AF_IB:
1152                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1153         default:
1154                 return false;
1155         }
1156 }
1157
1158 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1159 {
1160         switch (addr->sa_family) {
1161         case AF_INET:
1162                 return ipv4_is_loopback(
1163                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1164         case AF_INET6:
1165                 return ipv6_addr_loopback(
1166                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1167         case AF_IB:
1168                 return ib_addr_loopback(
1169                         &((struct sockaddr_ib *)addr)->sib_addr);
1170         default:
1171                 return false;
1172         }
1173 }
1174
1175 static inline bool cma_any_addr(const struct sockaddr *addr)
1176 {
1177         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1178 }
1179
1180 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1181 {
1182         if (src->sa_family != dst->sa_family)
1183                 return -1;
1184
1185         switch (src->sa_family) {
1186         case AF_INET:
1187                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1188                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1189         case AF_INET6: {
1190                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1191                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1192                 bool link_local;
1193
1194                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1195                                           &dst_addr6->sin6_addr))
1196                         return 1;
1197                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1198                              IPV6_ADDR_LINKLOCAL;
1199                 /* Link local must match their scope_ids */
1200                 return link_local ? (src_addr6->sin6_scope_id !=
1201                                      dst_addr6->sin6_scope_id) :
1202                                     0;
1203         }
1204
1205         default:
1206                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1207                                    &((struct sockaddr_ib *) dst)->sib_addr);
1208         }
1209 }
1210
1211 static __be16 cma_port(const struct sockaddr *addr)
1212 {
1213         struct sockaddr_ib *sib;
1214
1215         switch (addr->sa_family) {
1216         case AF_INET:
1217                 return ((struct sockaddr_in *) addr)->sin_port;
1218         case AF_INET6:
1219                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1220         case AF_IB:
1221                 sib = (struct sockaddr_ib *) addr;
1222                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1223                                     be64_to_cpu(sib->sib_sid_mask)));
1224         default:
1225                 return 0;
1226         }
1227 }
1228
1229 static inline int cma_any_port(const struct sockaddr *addr)
1230 {
1231         return !cma_port(addr);
1232 }
1233
1234 static void cma_save_ib_info(struct sockaddr *src_addr,
1235                              struct sockaddr *dst_addr,
1236                              const struct rdma_cm_id *listen_id,
1237                              const struct sa_path_rec *path)
1238 {
1239         struct sockaddr_ib *listen_ib, *ib;
1240
1241         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1242         if (src_addr) {
1243                 ib = (struct sockaddr_ib *)src_addr;
1244                 ib->sib_family = AF_IB;
1245                 if (path) {
1246                         ib->sib_pkey = path->pkey;
1247                         ib->sib_flowinfo = path->flow_label;
1248                         memcpy(&ib->sib_addr, &path->sgid, 16);
1249                         ib->sib_sid = path->service_id;
1250                         ib->sib_scope_id = 0;
1251                 } else {
1252                         ib->sib_pkey = listen_ib->sib_pkey;
1253                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1254                         ib->sib_addr = listen_ib->sib_addr;
1255                         ib->sib_sid = listen_ib->sib_sid;
1256                         ib->sib_scope_id = listen_ib->sib_scope_id;
1257                 }
1258                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1259         }
1260         if (dst_addr) {
1261                 ib = (struct sockaddr_ib *)dst_addr;
1262                 ib->sib_family = AF_IB;
1263                 if (path) {
1264                         ib->sib_pkey = path->pkey;
1265                         ib->sib_flowinfo = path->flow_label;
1266                         memcpy(&ib->sib_addr, &path->dgid, 16);
1267                 }
1268         }
1269 }
1270
1271 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1272                               struct sockaddr_in *dst_addr,
1273                               struct cma_hdr *hdr,
1274                               __be16 local_port)
1275 {
1276         if (src_addr) {
1277                 *src_addr = (struct sockaddr_in) {
1278                         .sin_family = AF_INET,
1279                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1280                         .sin_port = local_port,
1281                 };
1282         }
1283
1284         if (dst_addr) {
1285                 *dst_addr = (struct sockaddr_in) {
1286                         .sin_family = AF_INET,
1287                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1288                         .sin_port = hdr->port,
1289                 };
1290         }
1291 }
1292
1293 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1294                               struct sockaddr_in6 *dst_addr,
1295                               struct cma_hdr *hdr,
1296                               __be16 local_port)
1297 {
1298         if (src_addr) {
1299                 *src_addr = (struct sockaddr_in6) {
1300                         .sin6_family = AF_INET6,
1301                         .sin6_addr = hdr->dst_addr.ip6,
1302                         .sin6_port = local_port,
1303                 };
1304         }
1305
1306         if (dst_addr) {
1307                 *dst_addr = (struct sockaddr_in6) {
1308                         .sin6_family = AF_INET6,
1309                         .sin6_addr = hdr->src_addr.ip6,
1310                         .sin6_port = hdr->port,
1311                 };
1312         }
1313 }
1314
1315 static u16 cma_port_from_service_id(__be64 service_id)
1316 {
1317         return (u16)be64_to_cpu(service_id);
1318 }
1319
1320 static int cma_save_ip_info(struct sockaddr *src_addr,
1321                             struct sockaddr *dst_addr,
1322                             const struct ib_cm_event *ib_event,
1323                             __be64 service_id)
1324 {
1325         struct cma_hdr *hdr;
1326         __be16 port;
1327
1328         hdr = ib_event->private_data;
1329         if (hdr->cma_version != CMA_VERSION)
1330                 return -EINVAL;
1331
1332         port = htons(cma_port_from_service_id(service_id));
1333
1334         switch (cma_get_ip_ver(hdr)) {
1335         case 4:
1336                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1337                                   (struct sockaddr_in *)dst_addr, hdr, port);
1338                 break;
1339         case 6:
1340                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1341                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1342                 break;
1343         default:
1344                 return -EAFNOSUPPORT;
1345         }
1346
1347         return 0;
1348 }
1349
1350 static int cma_save_net_info(struct sockaddr *src_addr,
1351                              struct sockaddr *dst_addr,
1352                              const struct rdma_cm_id *listen_id,
1353                              const struct ib_cm_event *ib_event,
1354                              sa_family_t sa_family, __be64 service_id)
1355 {
1356         if (sa_family == AF_IB) {
1357                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1358                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1359                                          ib_event->param.req_rcvd.primary_path);
1360                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1361                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1362                 return 0;
1363         }
1364
1365         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1366 }
1367
1368 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1369                              struct cma_req_info *req)
1370 {
1371         const struct ib_cm_req_event_param *req_param =
1372                 &ib_event->param.req_rcvd;
1373         const struct ib_cm_sidr_req_event_param *sidr_param =
1374                 &ib_event->param.sidr_req_rcvd;
1375
1376         switch (ib_event->event) {
1377         case IB_CM_REQ_RECEIVED:
1378                 req->device     = req_param->listen_id->device;
1379                 req->port       = req_param->port;
1380                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1381                        sizeof(req->local_gid));
1382                 req->has_gid    = true;
1383                 req->service_id = req_param->primary_path->service_id;
1384                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1385                 if (req->pkey != req_param->bth_pkey)
1386                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1387                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1388                                             req_param->bth_pkey, req->pkey);
1389                 break;
1390         case IB_CM_SIDR_REQ_RECEIVED:
1391                 req->device     = sidr_param->listen_id->device;
1392                 req->port       = sidr_param->port;
1393                 req->has_gid    = false;
1394                 req->service_id = sidr_param->service_id;
1395                 req->pkey       = sidr_param->pkey;
1396                 if (req->pkey != sidr_param->bth_pkey)
1397                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1398                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1399                                             sidr_param->bth_pkey, req->pkey);
1400                 break;
1401         default:
1402                 return -EINVAL;
1403         }
1404
1405         return 0;
1406 }
1407
1408 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1409                                   const struct sockaddr_in *dst_addr,
1410                                   const struct sockaddr_in *src_addr)
1411 {
1412         __be32 daddr = dst_addr->sin_addr.s_addr,
1413                saddr = src_addr->sin_addr.s_addr;
1414         struct fib_result res;
1415         struct flowi4 fl4;
1416         int err;
1417         bool ret;
1418
1419         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1420             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1421             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1422             ipv4_is_loopback(saddr))
1423                 return false;
1424
1425         memset(&fl4, 0, sizeof(fl4));
1426         fl4.flowi4_iif = net_dev->ifindex;
1427         fl4.daddr = daddr;
1428         fl4.saddr = saddr;
1429
1430         rcu_read_lock();
1431         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1432         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1433         rcu_read_unlock();
1434
1435         return ret;
1436 }
1437
1438 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1439                                   const struct sockaddr_in6 *dst_addr,
1440                                   const struct sockaddr_in6 *src_addr)
1441 {
1442 #if IS_ENABLED(CONFIG_IPV6)
1443         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1444                            IPV6_ADDR_LINKLOCAL;
1445         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1446                                          &src_addr->sin6_addr, net_dev->ifindex,
1447                                          NULL, strict);
1448         bool ret;
1449
1450         if (!rt)
1451                 return false;
1452
1453         ret = rt->rt6i_idev->dev == net_dev;
1454         ip6_rt_put(rt);
1455
1456         return ret;
1457 #else
1458         return false;
1459 #endif
1460 }
1461
1462 static bool validate_net_dev(struct net_device *net_dev,
1463                              const struct sockaddr *daddr,
1464                              const struct sockaddr *saddr)
1465 {
1466         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1467         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1468         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1469         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1470
1471         switch (daddr->sa_family) {
1472         case AF_INET:
1473                 return saddr->sa_family == AF_INET &&
1474                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1475
1476         case AF_INET6:
1477                 return saddr->sa_family == AF_INET6 &&
1478                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1479
1480         default:
1481                 return false;
1482         }
1483 }
1484
1485 static struct net_device *
1486 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1487 {
1488         const struct ib_gid_attr *sgid_attr = NULL;
1489         struct net_device *ndev;
1490
1491         if (ib_event->event == IB_CM_REQ_RECEIVED)
1492                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1493         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1494                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1495
1496         if (!sgid_attr)
1497                 return NULL;
1498
1499         rcu_read_lock();
1500         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1501         if (IS_ERR(ndev))
1502                 ndev = NULL;
1503         else
1504                 dev_hold(ndev);
1505         rcu_read_unlock();
1506         return ndev;
1507 }
1508
1509 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1510                                           struct cma_req_info *req)
1511 {
1512         struct sockaddr *listen_addr =
1513                         (struct sockaddr *)&req->listen_addr_storage;
1514         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1515         struct net_device *net_dev;
1516         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1517         int err;
1518
1519         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1520                                req->service_id);
1521         if (err)
1522                 return ERR_PTR(err);
1523
1524         if (rdma_protocol_roce(req->device, req->port))
1525                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1526         else
1527                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1528                                                    req->pkey,
1529                                                    gid, listen_addr);
1530         if (!net_dev)
1531                 return ERR_PTR(-ENODEV);
1532
1533         return net_dev;
1534 }
1535
1536 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1537 {
1538         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1539 }
1540
1541 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1542                                    const struct cma_hdr *hdr)
1543 {
1544         struct sockaddr *addr = cma_src_addr(id_priv);
1545         __be32 ip4_addr;
1546         struct in6_addr ip6_addr;
1547
1548         if (cma_any_addr(addr) && !id_priv->afonly)
1549                 return true;
1550
1551         switch (addr->sa_family) {
1552         case AF_INET:
1553                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1554                 if (cma_get_ip_ver(hdr) != 4)
1555                         return false;
1556                 if (!cma_any_addr(addr) &&
1557                     hdr->dst_addr.ip4.addr != ip4_addr)
1558                         return false;
1559                 break;
1560         case AF_INET6:
1561                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1562                 if (cma_get_ip_ver(hdr) != 6)
1563                         return false;
1564                 if (!cma_any_addr(addr) &&
1565                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1566                         return false;
1567                 break;
1568         case AF_IB:
1569                 return true;
1570         default:
1571                 return false;
1572         }
1573
1574         return true;
1575 }
1576
1577 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1578 {
1579         struct ib_device *device = id->device;
1580         const int port_num = id->port_num ?: rdma_start_port(device);
1581
1582         return rdma_protocol_roce(device, port_num);
1583 }
1584
1585 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1586 {
1587         const struct sockaddr *daddr =
1588                         (const struct sockaddr *)&req->listen_addr_storage;
1589         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1590
1591         /* Returns true if the req is for IPv6 link local */
1592         return (daddr->sa_family == AF_INET6 &&
1593                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1594 }
1595
1596 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1597                               const struct net_device *net_dev,
1598                               const struct cma_req_info *req)
1599 {
1600         const struct rdma_addr *addr = &id->route.addr;
1601
1602         if (!net_dev)
1603                 /* This request is an AF_IB request */
1604                 return (!id->port_num || id->port_num == req->port) &&
1605                        (addr->src_addr.ss_family == AF_IB);
1606
1607         /*
1608          * If the request is not for IPv6 link local, allow matching
1609          * request to any netdevice of the one or multiport rdma device.
1610          */
1611         if (!cma_is_req_ipv6_ll(req))
1612                 return true;
1613         /*
1614          * Net namespaces must match, and if the listner is listening
1615          * on a specific netdevice than netdevice must match as well.
1616          */
1617         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1618             (!!addr->dev_addr.bound_dev_if ==
1619              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1620                 return true;
1621         else
1622                 return false;
1623 }
1624
1625 static struct rdma_id_private *cma_find_listener(
1626                 const struct rdma_bind_list *bind_list,
1627                 const struct ib_cm_id *cm_id,
1628                 const struct ib_cm_event *ib_event,
1629                 const struct cma_req_info *req,
1630                 const struct net_device *net_dev)
1631 {
1632         struct rdma_id_private *id_priv, *id_priv_dev;
1633
1634         if (!bind_list)
1635                 return ERR_PTR(-EINVAL);
1636
1637         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1638                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1639                         if (id_priv->id.device == cm_id->device &&
1640                             cma_match_net_dev(&id_priv->id, net_dev, req))
1641                                 return id_priv;
1642                         list_for_each_entry(id_priv_dev,
1643                                             &id_priv->listen_list,
1644                                             listen_list) {
1645                                 if (id_priv_dev->id.device == cm_id->device &&
1646                                     cma_match_net_dev(&id_priv_dev->id,
1647                                                       net_dev, req))
1648                                         return id_priv_dev;
1649                         }
1650                 }
1651         }
1652
1653         return ERR_PTR(-EINVAL);
1654 }
1655
1656 static struct rdma_id_private *
1657 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1658                      const struct ib_cm_event *ib_event,
1659                      struct cma_req_info *req,
1660                      struct net_device **net_dev)
1661 {
1662         struct rdma_bind_list *bind_list;
1663         struct rdma_id_private *id_priv;
1664         int err;
1665
1666         err = cma_save_req_info(ib_event, req);
1667         if (err)
1668                 return ERR_PTR(err);
1669
1670         *net_dev = cma_get_net_dev(ib_event, req);
1671         if (IS_ERR(*net_dev)) {
1672                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1673                         /* Assuming the protocol is AF_IB */
1674                         *net_dev = NULL;
1675                 } else {
1676                         return ERR_CAST(*net_dev);
1677                 }
1678         }
1679
1680         /*
1681          * Net namespace might be getting deleted while route lookup,
1682          * cm_id lookup is in progress. Therefore, perform netdevice
1683          * validation, cm_id lookup under rcu lock.
1684          * RCU lock along with netdevice state check, synchronizes with
1685          * netdevice migrating to different net namespace and also avoids
1686          * case where net namespace doesn't get deleted while lookup is in
1687          * progress.
1688          * If the device state is not IFF_UP, its properties such as ifindex
1689          * and nd_net cannot be trusted to remain valid without rcu lock.
1690          * net/core/dev.c change_net_namespace() ensures to synchronize with
1691          * ongoing operations on net device after device is closed using
1692          * synchronize_net().
1693          */
1694         rcu_read_lock();
1695         if (*net_dev) {
1696                 /*
1697                  * If netdevice is down, it is likely that it is administratively
1698                  * down or it might be migrating to different namespace.
1699                  * In that case avoid further processing, as the net namespace
1700                  * or ifindex may change.
1701                  */
1702                 if (((*net_dev)->flags & IFF_UP) == 0) {
1703                         id_priv = ERR_PTR(-EHOSTUNREACH);
1704                         goto err;
1705                 }
1706
1707                 if (!validate_net_dev(*net_dev,
1708                                  (struct sockaddr *)&req->listen_addr_storage,
1709                                  (struct sockaddr *)&req->src_addr_storage)) {
1710                         id_priv = ERR_PTR(-EHOSTUNREACH);
1711                         goto err;
1712                 }
1713         }
1714
1715         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1716                                 rdma_ps_from_service_id(req->service_id),
1717                                 cma_port_from_service_id(req->service_id));
1718         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1719 err:
1720         rcu_read_unlock();
1721         if (IS_ERR(id_priv) && *net_dev) {
1722                 dev_put(*net_dev);
1723                 *net_dev = NULL;
1724         }
1725         return id_priv;
1726 }
1727
1728 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1729 {
1730         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1731 }
1732
1733 static void cma_cancel_route(struct rdma_id_private *id_priv)
1734 {
1735         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1736                 if (id_priv->query)
1737                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1738         }
1739 }
1740
1741 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1742 {
1743         struct rdma_id_private *dev_id_priv;
1744
1745         /*
1746          * Remove from listen_any_list to prevent added devices from spawning
1747          * additional listen requests.
1748          */
1749         mutex_lock(&lock);
1750         list_del(&id_priv->list);
1751
1752         while (!list_empty(&id_priv->listen_list)) {
1753                 dev_id_priv = list_entry(id_priv->listen_list.next,
1754                                          struct rdma_id_private, listen_list);
1755                 /* sync with device removal to avoid duplicate destruction */
1756                 list_del_init(&dev_id_priv->list);
1757                 list_del(&dev_id_priv->listen_list);
1758                 mutex_unlock(&lock);
1759
1760                 rdma_destroy_id(&dev_id_priv->id);
1761                 mutex_lock(&lock);
1762         }
1763         mutex_unlock(&lock);
1764 }
1765
1766 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1767                                  enum rdma_cm_state state)
1768 {
1769         switch (state) {
1770         case RDMA_CM_ADDR_QUERY:
1771                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1772                 break;
1773         case RDMA_CM_ROUTE_QUERY:
1774                 cma_cancel_route(id_priv);
1775                 break;
1776         case RDMA_CM_LISTEN:
1777                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1778                         cma_cancel_listens(id_priv);
1779                 break;
1780         default:
1781                 break;
1782         }
1783 }
1784
1785 static void cma_release_port(struct rdma_id_private *id_priv)
1786 {
1787         struct rdma_bind_list *bind_list = id_priv->bind_list;
1788         struct net *net = id_priv->id.route.addr.dev_addr.net;
1789
1790         if (!bind_list)
1791                 return;
1792
1793         mutex_lock(&lock);
1794         hlist_del(&id_priv->node);
1795         if (hlist_empty(&bind_list->owners)) {
1796                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1797                 kfree(bind_list);
1798         }
1799         mutex_unlock(&lock);
1800 }
1801
1802 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv,
1803                                     struct cma_multicast *mc)
1804 {
1805         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1806         struct net_device *ndev = NULL;
1807
1808         if (dev_addr->bound_dev_if)
1809                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
1810         if (ndev) {
1811                 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1812                 dev_put(ndev);
1813         }
1814         kref_put(&mc->mcref, release_mc);
1815 }
1816
1817 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1818 {
1819         struct cma_multicast *mc;
1820
1821         while (!list_empty(&id_priv->mc_list)) {
1822                 mc = container_of(id_priv->mc_list.next,
1823                                   struct cma_multicast, list);
1824                 list_del(&mc->list);
1825                 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1826                                       id_priv->id.port_num)) {
1827                         ib_sa_free_multicast(mc->multicast.ib);
1828                         kfree(mc);
1829                 } else {
1830                         cma_leave_roce_mc_group(id_priv, mc);
1831                 }
1832         }
1833 }
1834
1835 void rdma_destroy_id(struct rdma_cm_id *id)
1836 {
1837         struct rdma_id_private *id_priv;
1838         enum rdma_cm_state state;
1839
1840         id_priv = container_of(id, struct rdma_id_private, id);
1841         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1842         cma_cancel_operation(id_priv, state);
1843
1844         /*
1845          * Wait for any active callback to finish.  New callbacks will find
1846          * the id_priv state set to destroying and abort.
1847          */
1848         mutex_lock(&id_priv->handler_mutex);
1849         mutex_unlock(&id_priv->handler_mutex);
1850
1851         rdma_restrack_del(&id_priv->res);
1852         if (id_priv->cma_dev) {
1853                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1854                         if (id_priv->cm_id.ib)
1855                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1856                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1857                         if (id_priv->cm_id.iw)
1858                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1859                 }
1860                 cma_leave_mc_groups(id_priv);
1861                 cma_release_dev(id_priv);
1862         }
1863
1864         cma_release_port(id_priv);
1865         cma_deref_id(id_priv);
1866         wait_for_completion(&id_priv->comp);
1867
1868         if (id_priv->internal_id)
1869                 cma_deref_id(id_priv->id.context);
1870
1871         kfree(id_priv->id.route.path_rec);
1872
1873         if (id_priv->id.route.addr.dev_addr.sgid_attr)
1874                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1875
1876         put_net(id_priv->id.route.addr.dev_addr.net);
1877         kfree(id_priv);
1878 }
1879 EXPORT_SYMBOL(rdma_destroy_id);
1880
1881 static int cma_rep_recv(struct rdma_id_private *id_priv)
1882 {
1883         int ret;
1884
1885         ret = cma_modify_qp_rtr(id_priv, NULL);
1886         if (ret)
1887                 goto reject;
1888
1889         ret = cma_modify_qp_rts(id_priv, NULL);
1890         if (ret)
1891                 goto reject;
1892
1893         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1894         if (ret)
1895                 goto reject;
1896
1897         return 0;
1898 reject:
1899         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1900         cma_modify_qp_err(id_priv);
1901         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1902                        NULL, 0, NULL, 0);
1903         return ret;
1904 }
1905
1906 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1907                                    const struct ib_cm_rep_event_param *rep_data,
1908                                    void *private_data)
1909 {
1910         event->param.conn.private_data = private_data;
1911         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1912         event->param.conn.responder_resources = rep_data->responder_resources;
1913         event->param.conn.initiator_depth = rep_data->initiator_depth;
1914         event->param.conn.flow_control = rep_data->flow_control;
1915         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1916         event->param.conn.srq = rep_data->srq;
1917         event->param.conn.qp_num = rep_data->remote_qpn;
1918 }
1919
1920 static int cma_ib_handler(struct ib_cm_id *cm_id,
1921                           const struct ib_cm_event *ib_event)
1922 {
1923         struct rdma_id_private *id_priv = cm_id->context;
1924         struct rdma_cm_event event = {};
1925         int ret = 0;
1926
1927         mutex_lock(&id_priv->handler_mutex);
1928         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1929              id_priv->state != RDMA_CM_CONNECT) ||
1930             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1931              id_priv->state != RDMA_CM_DISCONNECT))
1932                 goto out;
1933
1934         switch (ib_event->event) {
1935         case IB_CM_REQ_ERROR:
1936         case IB_CM_REP_ERROR:
1937                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1938                 event.status = -ETIMEDOUT;
1939                 break;
1940         case IB_CM_REP_RECEIVED:
1941                 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1942                     (id_priv->id.qp_type != IB_QPT_UD))
1943                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1944                 if (id_priv->id.qp) {
1945                         event.status = cma_rep_recv(id_priv);
1946                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1947                                                      RDMA_CM_EVENT_ESTABLISHED;
1948                 } else {
1949                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1950                 }
1951                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1952                                        ib_event->private_data);
1953                 break;
1954         case IB_CM_RTU_RECEIVED:
1955         case IB_CM_USER_ESTABLISHED:
1956                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1957                 break;
1958         case IB_CM_DREQ_ERROR:
1959                 event.status = -ETIMEDOUT; /* fall through */
1960         case IB_CM_DREQ_RECEIVED:
1961         case IB_CM_DREP_RECEIVED:
1962                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1963                                    RDMA_CM_DISCONNECT))
1964                         goto out;
1965                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1966                 break;
1967         case IB_CM_TIMEWAIT_EXIT:
1968                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1969                 break;
1970         case IB_CM_MRA_RECEIVED:
1971                 /* ignore event */
1972                 goto out;
1973         case IB_CM_REJ_RECEIVED:
1974                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1975                                                                                 ib_event->param.rej_rcvd.reason));
1976                 cma_modify_qp_err(id_priv);
1977                 event.status = ib_event->param.rej_rcvd.reason;
1978                 event.event = RDMA_CM_EVENT_REJECTED;
1979                 event.param.conn.private_data = ib_event->private_data;
1980                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1981                 break;
1982         default:
1983                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1984                        ib_event->event);
1985                 goto out;
1986         }
1987
1988         ret = id_priv->id.event_handler(&id_priv->id, &event);
1989         if (ret) {
1990                 /* Destroy the CM ID by returning a non-zero value. */
1991                 id_priv->cm_id.ib = NULL;
1992                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1993                 mutex_unlock(&id_priv->handler_mutex);
1994                 rdma_destroy_id(&id_priv->id);
1995                 return ret;
1996         }
1997 out:
1998         mutex_unlock(&id_priv->handler_mutex);
1999         return ret;
2000 }
2001
2002 static struct rdma_id_private *
2003 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2004                    const struct ib_cm_event *ib_event,
2005                    struct net_device *net_dev)
2006 {
2007         struct rdma_id_private *listen_id_priv;
2008         struct rdma_id_private *id_priv;
2009         struct rdma_cm_id *id;
2010         struct rdma_route *rt;
2011         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2012         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2013         const __be64 service_id =
2014                 ib_event->param.req_rcvd.primary_path->service_id;
2015         int ret;
2016
2017         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2018         id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2019                             listen_id->event_handler, listen_id->context,
2020                             listen_id->ps, ib_event->param.req_rcvd.qp_type,
2021                             listen_id_priv->res.kern_name);
2022         if (IS_ERR(id))
2023                 return NULL;
2024
2025         id_priv = container_of(id, struct rdma_id_private, id);
2026         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2027                               (struct sockaddr *)&id->route.addr.dst_addr,
2028                               listen_id, ib_event, ss_family, service_id))
2029                 goto err;
2030
2031         rt = &id->route;
2032         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2033         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2034                                      GFP_KERNEL);
2035         if (!rt->path_rec)
2036                 goto err;
2037
2038         rt->path_rec[0] = *path;
2039         if (rt->num_paths == 2)
2040                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2041
2042         if (net_dev) {
2043                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2044         } else {
2045                 if (!cma_protocol_roce(listen_id) &&
2046                     cma_any_addr(cma_src_addr(id_priv))) {
2047                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2048                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2049                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2050                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2051                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2052                         if (ret)
2053                                 goto err;
2054                 }
2055         }
2056         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2057
2058         id_priv->state = RDMA_CM_CONNECT;
2059         return id_priv;
2060
2061 err:
2062         rdma_destroy_id(id);
2063         return NULL;
2064 }
2065
2066 static struct rdma_id_private *
2067 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2068                   const struct ib_cm_event *ib_event,
2069                   struct net_device *net_dev)
2070 {
2071         const struct rdma_id_private *listen_id_priv;
2072         struct rdma_id_private *id_priv;
2073         struct rdma_cm_id *id;
2074         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2075         struct net *net = listen_id->route.addr.dev_addr.net;
2076         int ret;
2077
2078         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2079         id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
2080                               listen_id->ps, IB_QPT_UD,
2081                               listen_id_priv->res.kern_name);
2082         if (IS_ERR(id))
2083                 return NULL;
2084
2085         id_priv = container_of(id, struct rdma_id_private, id);
2086         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2087                               (struct sockaddr *)&id->route.addr.dst_addr,
2088                               listen_id, ib_event, ss_family,
2089                               ib_event->param.sidr_req_rcvd.service_id))
2090                 goto err;
2091
2092         if (net_dev) {
2093                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2094         } else {
2095                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2096                         ret = cma_translate_addr(cma_src_addr(id_priv),
2097                                                  &id->route.addr.dev_addr);
2098                         if (ret)
2099                                 goto err;
2100                 }
2101         }
2102
2103         id_priv->state = RDMA_CM_CONNECT;
2104         return id_priv;
2105 err:
2106         rdma_destroy_id(id);
2107         return NULL;
2108 }
2109
2110 static void cma_set_req_event_data(struct rdma_cm_event *event,
2111                                    const struct ib_cm_req_event_param *req_data,
2112                                    void *private_data, int offset)
2113 {
2114         event->param.conn.private_data = private_data + offset;
2115         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2116         event->param.conn.responder_resources = req_data->responder_resources;
2117         event->param.conn.initiator_depth = req_data->initiator_depth;
2118         event->param.conn.flow_control = req_data->flow_control;
2119         event->param.conn.retry_count = req_data->retry_count;
2120         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2121         event->param.conn.srq = req_data->srq;
2122         event->param.conn.qp_num = req_data->remote_qpn;
2123 }
2124
2125 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2126                                     const struct ib_cm_event *ib_event)
2127 {
2128         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2129                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2130                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2131                  (id->qp_type == IB_QPT_UD)) ||
2132                 (!id->qp_type));
2133 }
2134
2135 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2136                               const struct ib_cm_event *ib_event)
2137 {
2138         struct rdma_id_private *listen_id, *conn_id = NULL;
2139         struct rdma_cm_event event = {};
2140         struct cma_req_info req = {};
2141         struct net_device *net_dev;
2142         u8 offset;
2143         int ret;
2144
2145         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2146         if (IS_ERR(listen_id))
2147                 return PTR_ERR(listen_id);
2148
2149         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2150                 ret = -EINVAL;
2151                 goto net_dev_put;
2152         }
2153
2154         mutex_lock(&listen_id->handler_mutex);
2155         if (listen_id->state != RDMA_CM_LISTEN) {
2156                 ret = -ECONNABORTED;
2157                 goto err1;
2158         }
2159
2160         offset = cma_user_data_offset(listen_id);
2161         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2162         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2163                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2164                 event.param.ud.private_data = ib_event->private_data + offset;
2165                 event.param.ud.private_data_len =
2166                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2167         } else {
2168                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2169                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2170                                        ib_event->private_data, offset);
2171         }
2172         if (!conn_id) {
2173                 ret = -ENOMEM;
2174                 goto err1;
2175         }
2176
2177         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2178         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2179         if (ret)
2180                 goto err2;
2181
2182         conn_id->cm_id.ib = cm_id;
2183         cm_id->context = conn_id;
2184         cm_id->cm_handler = cma_ib_handler;
2185
2186         /*
2187          * Protect against the user destroying conn_id from another thread
2188          * until we're done accessing it.
2189          */
2190         atomic_inc(&conn_id->refcount);
2191         ret = conn_id->id.event_handler(&conn_id->id, &event);
2192         if (ret)
2193                 goto err3;
2194         /*
2195          * Acquire mutex to prevent user executing rdma_destroy_id()
2196          * while we're accessing the cm_id.
2197          */
2198         mutex_lock(&lock);
2199         if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2200             (conn_id->id.qp_type != IB_QPT_UD))
2201                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2202         mutex_unlock(&lock);
2203         mutex_unlock(&conn_id->handler_mutex);
2204         mutex_unlock(&listen_id->handler_mutex);
2205         cma_deref_id(conn_id);
2206         if (net_dev)
2207                 dev_put(net_dev);
2208         return 0;
2209
2210 err3:
2211         cma_deref_id(conn_id);
2212         /* Destroy the CM ID by returning a non-zero value. */
2213         conn_id->cm_id.ib = NULL;
2214 err2:
2215         cma_exch(conn_id, RDMA_CM_DESTROYING);
2216         mutex_unlock(&conn_id->handler_mutex);
2217 err1:
2218         mutex_unlock(&listen_id->handler_mutex);
2219         if (conn_id)
2220                 rdma_destroy_id(&conn_id->id);
2221
2222 net_dev_put:
2223         if (net_dev)
2224                 dev_put(net_dev);
2225
2226         return ret;
2227 }
2228
2229 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2230 {
2231         if (addr->sa_family == AF_IB)
2232                 return ((struct sockaddr_ib *) addr)->sib_sid;
2233
2234         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2235 }
2236 EXPORT_SYMBOL(rdma_get_service_id);
2237
2238 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2239                     union ib_gid *dgid)
2240 {
2241         struct rdma_addr *addr = &cm_id->route.addr;
2242
2243         if (!cm_id->device) {
2244                 if (sgid)
2245                         memset(sgid, 0, sizeof(*sgid));
2246                 if (dgid)
2247                         memset(dgid, 0, sizeof(*dgid));
2248                 return;
2249         }
2250
2251         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2252                 if (sgid)
2253                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2254                 if (dgid)
2255                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2256         } else {
2257                 if (sgid)
2258                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2259                 if (dgid)
2260                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2261         }
2262 }
2263 EXPORT_SYMBOL(rdma_read_gids);
2264
2265 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2266 {
2267         struct rdma_id_private *id_priv = iw_id->context;
2268         struct rdma_cm_event event = {};
2269         int ret = 0;
2270         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2271         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2272
2273         mutex_lock(&id_priv->handler_mutex);
2274         if (id_priv->state != RDMA_CM_CONNECT)
2275                 goto out;
2276
2277         switch (iw_event->event) {
2278         case IW_CM_EVENT_CLOSE:
2279                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2280                 break;
2281         case IW_CM_EVENT_CONNECT_REPLY:
2282                 memcpy(cma_src_addr(id_priv), laddr,
2283                        rdma_addr_size(laddr));
2284                 memcpy(cma_dst_addr(id_priv), raddr,
2285                        rdma_addr_size(raddr));
2286                 switch (iw_event->status) {
2287                 case 0:
2288                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2289                         event.param.conn.initiator_depth = iw_event->ird;
2290                         event.param.conn.responder_resources = iw_event->ord;
2291                         break;
2292                 case -ECONNRESET:
2293                 case -ECONNREFUSED:
2294                         event.event = RDMA_CM_EVENT_REJECTED;
2295                         break;
2296                 case -ETIMEDOUT:
2297                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2298                         break;
2299                 default:
2300                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2301                         break;
2302                 }
2303                 break;
2304         case IW_CM_EVENT_ESTABLISHED:
2305                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2306                 event.param.conn.initiator_depth = iw_event->ird;
2307                 event.param.conn.responder_resources = iw_event->ord;
2308                 break;
2309         default:
2310                 goto out;
2311         }
2312
2313         event.status = iw_event->status;
2314         event.param.conn.private_data = iw_event->private_data;
2315         event.param.conn.private_data_len = iw_event->private_data_len;
2316         ret = id_priv->id.event_handler(&id_priv->id, &event);
2317         if (ret) {
2318                 /* Destroy the CM ID by returning a non-zero value. */
2319                 id_priv->cm_id.iw = NULL;
2320                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2321                 mutex_unlock(&id_priv->handler_mutex);
2322                 rdma_destroy_id(&id_priv->id);
2323                 return ret;
2324         }
2325
2326 out:
2327         mutex_unlock(&id_priv->handler_mutex);
2328         return ret;
2329 }
2330
2331 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2332                                struct iw_cm_event *iw_event)
2333 {
2334         struct rdma_cm_id *new_cm_id;
2335         struct rdma_id_private *listen_id, *conn_id;
2336         struct rdma_cm_event event = {};
2337         int ret = -ECONNABORTED;
2338         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2339         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2340
2341         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2342         event.param.conn.private_data = iw_event->private_data;
2343         event.param.conn.private_data_len = iw_event->private_data_len;
2344         event.param.conn.initiator_depth = iw_event->ird;
2345         event.param.conn.responder_resources = iw_event->ord;
2346
2347         listen_id = cm_id->context;
2348
2349         mutex_lock(&listen_id->handler_mutex);
2350         if (listen_id->state != RDMA_CM_LISTEN)
2351                 goto out;
2352
2353         /* Create a new RDMA id for the new IW CM ID */
2354         new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2355                                      listen_id->id.event_handler,
2356                                      listen_id->id.context,
2357                                      RDMA_PS_TCP, IB_QPT_RC,
2358                                      listen_id->res.kern_name);
2359         if (IS_ERR(new_cm_id)) {
2360                 ret = -ENOMEM;
2361                 goto out;
2362         }
2363         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2364         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2365         conn_id->state = RDMA_CM_CONNECT;
2366
2367         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2368         if (ret) {
2369                 mutex_unlock(&conn_id->handler_mutex);
2370                 rdma_destroy_id(new_cm_id);
2371                 goto out;
2372         }
2373
2374         ret = cma_iw_acquire_dev(conn_id, listen_id);
2375         if (ret) {
2376                 mutex_unlock(&conn_id->handler_mutex);
2377                 rdma_destroy_id(new_cm_id);
2378                 goto out;
2379         }
2380
2381         conn_id->cm_id.iw = cm_id;
2382         cm_id->context = conn_id;
2383         cm_id->cm_handler = cma_iw_handler;
2384
2385         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2386         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2387
2388         /*
2389          * Protect against the user destroying conn_id from another thread
2390          * until we're done accessing it.
2391          */
2392         atomic_inc(&conn_id->refcount);
2393         ret = conn_id->id.event_handler(&conn_id->id, &event);
2394         if (ret) {
2395                 /* User wants to destroy the CM ID */
2396                 conn_id->cm_id.iw = NULL;
2397                 cma_exch(conn_id, RDMA_CM_DESTROYING);
2398                 mutex_unlock(&conn_id->handler_mutex);
2399                 mutex_unlock(&listen_id->handler_mutex);
2400                 cma_deref_id(conn_id);
2401                 rdma_destroy_id(&conn_id->id);
2402                 return ret;
2403         }
2404
2405         mutex_unlock(&conn_id->handler_mutex);
2406         cma_deref_id(conn_id);
2407
2408 out:
2409         mutex_unlock(&listen_id->handler_mutex);
2410         return ret;
2411 }
2412
2413 static int cma_ib_listen(struct rdma_id_private *id_priv)
2414 {
2415         struct sockaddr *addr;
2416         struct ib_cm_id *id;
2417         __be64 svc_id;
2418
2419         addr = cma_src_addr(id_priv);
2420         svc_id = rdma_get_service_id(&id_priv->id, addr);
2421         id = ib_cm_insert_listen(id_priv->id.device,
2422                                  cma_ib_req_handler, svc_id);
2423         if (IS_ERR(id))
2424                 return PTR_ERR(id);
2425         id_priv->cm_id.ib = id;
2426
2427         return 0;
2428 }
2429
2430 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2431 {
2432         int ret;
2433         struct iw_cm_id *id;
2434
2435         id = iw_create_cm_id(id_priv->id.device,
2436                              iw_conn_req_handler,
2437                              id_priv);
2438         if (IS_ERR(id))
2439                 return PTR_ERR(id);
2440
2441         id->tos = id_priv->tos;
2442         id->tos_set = id_priv->tos_set;
2443         id_priv->cm_id.iw = id;
2444
2445         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2446                rdma_addr_size(cma_src_addr(id_priv)));
2447
2448         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2449
2450         if (ret) {
2451                 iw_destroy_cm_id(id_priv->cm_id.iw);
2452                 id_priv->cm_id.iw = NULL;
2453         }
2454
2455         return ret;
2456 }
2457
2458 static int cma_listen_handler(struct rdma_cm_id *id,
2459                               struct rdma_cm_event *event)
2460 {
2461         struct rdma_id_private *id_priv = id->context;
2462
2463         id->context = id_priv->id.context;
2464         id->event_handler = id_priv->id.event_handler;
2465         return id_priv->id.event_handler(id, event);
2466 }
2467
2468 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2469                               struct cma_device *cma_dev)
2470 {
2471         struct rdma_id_private *dev_id_priv;
2472         struct rdma_cm_id *id;
2473         struct net *net = id_priv->id.route.addr.dev_addr.net;
2474         int ret;
2475
2476         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2477                 return;
2478
2479         id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2480                               id_priv->id.qp_type, id_priv->res.kern_name);
2481         if (IS_ERR(id))
2482                 return;
2483
2484         dev_id_priv = container_of(id, struct rdma_id_private, id);
2485
2486         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2487         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2488                rdma_addr_size(cma_src_addr(id_priv)));
2489
2490         _cma_attach_to_dev(dev_id_priv, cma_dev);
2491         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2492         atomic_inc(&id_priv->refcount);
2493         dev_id_priv->internal_id = 1;
2494         dev_id_priv->afonly = id_priv->afonly;
2495         dev_id_priv->tos_set = id_priv->tos_set;
2496         dev_id_priv->tos = id_priv->tos;
2497
2498         ret = rdma_listen(id, id_priv->backlog);
2499         if (ret)
2500                 dev_warn(&cma_dev->device->dev,
2501                          "RDMA CMA: cma_listen_on_dev, error %d\n", ret);
2502 }
2503
2504 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2505 {
2506         struct cma_device *cma_dev;
2507
2508         mutex_lock(&lock);
2509         list_add_tail(&id_priv->list, &listen_any_list);
2510         list_for_each_entry(cma_dev, &dev_list, list)
2511                 cma_listen_on_dev(id_priv, cma_dev);
2512         mutex_unlock(&lock);
2513 }
2514
2515 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2516 {
2517         struct rdma_id_private *id_priv;
2518
2519         id_priv = container_of(id, struct rdma_id_private, id);
2520         id_priv->tos = (u8) tos;
2521         id_priv->tos_set = true;
2522 }
2523 EXPORT_SYMBOL(rdma_set_service_type);
2524
2525 /**
2526  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2527  *                          with a connection identifier.
2528  * @id: Communication identifier to associated with service type.
2529  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2530  *
2531  * This function should be called before rdma_connect() on active side,
2532  * and on passive side before rdma_accept(). It is applicable to primary
2533  * path only. The timeout will affect the local side of the QP, it is not
2534  * negotiated with remote side and zero disables the timer.
2535  *
2536  * Return: 0 for success
2537  */
2538 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2539 {
2540         struct rdma_id_private *id_priv;
2541
2542         if (id->qp_type != IB_QPT_RC)
2543                 return -EINVAL;
2544
2545         id_priv = container_of(id, struct rdma_id_private, id);
2546         id_priv->timeout = timeout;
2547         id_priv->timeout_set = true;
2548
2549         return 0;
2550 }
2551 EXPORT_SYMBOL(rdma_set_ack_timeout);
2552
2553 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2554                               void *context)
2555 {
2556         struct cma_work *work = context;
2557         struct rdma_route *route;
2558
2559         route = &work->id->id.route;
2560
2561         if (!status) {
2562                 route->num_paths = 1;
2563                 *route->path_rec = *path_rec;
2564         } else {
2565                 work->old_state = RDMA_CM_ROUTE_QUERY;
2566                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2567                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2568                 work->event.status = status;
2569                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2570                                      status);
2571         }
2572
2573         queue_work(cma_wq, &work->work);
2574 }
2575
2576 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2577                               unsigned long timeout_ms, struct cma_work *work)
2578 {
2579         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2580         struct sa_path_rec path_rec;
2581         ib_sa_comp_mask comp_mask;
2582         struct sockaddr_in6 *sin6;
2583         struct sockaddr_ib *sib;
2584
2585         memset(&path_rec, 0, sizeof path_rec);
2586
2587         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2588                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2589         else
2590                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2591         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2592         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2593         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2594         path_rec.numb_path = 1;
2595         path_rec.reversible = 1;
2596         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2597                                                   cma_dst_addr(id_priv));
2598
2599         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2600                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2601                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2602
2603         switch (cma_family(id_priv)) {
2604         case AF_INET:
2605                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2606                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2607                 break;
2608         case AF_INET6:
2609                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2610                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2611                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2612                 break;
2613         case AF_IB:
2614                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2615                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2616                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2617                 break;
2618         }
2619
2620         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2621                                                id_priv->id.port_num, &path_rec,
2622                                                comp_mask, timeout_ms,
2623                                                GFP_KERNEL, cma_query_handler,
2624                                                work, &id_priv->query);
2625
2626         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2627 }
2628
2629 static void cma_work_handler(struct work_struct *_work)
2630 {
2631         struct cma_work *work = container_of(_work, struct cma_work, work);
2632         struct rdma_id_private *id_priv = work->id;
2633         int destroy = 0;
2634
2635         mutex_lock(&id_priv->handler_mutex);
2636         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2637                 goto out;
2638
2639         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2640                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2641                 destroy = 1;
2642         }
2643 out:
2644         mutex_unlock(&id_priv->handler_mutex);
2645         cma_deref_id(id_priv);
2646         if (destroy)
2647                 rdma_destroy_id(&id_priv->id);
2648         kfree(work);
2649 }
2650
2651 static void cma_ndev_work_handler(struct work_struct *_work)
2652 {
2653         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2654         struct rdma_id_private *id_priv = work->id;
2655         int destroy = 0;
2656
2657         mutex_lock(&id_priv->handler_mutex);
2658         if (id_priv->state == RDMA_CM_DESTROYING ||
2659             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2660                 goto out;
2661
2662         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2663                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2664                 destroy = 1;
2665         }
2666
2667 out:
2668         mutex_unlock(&id_priv->handler_mutex);
2669         cma_deref_id(id_priv);
2670         if (destroy)
2671                 rdma_destroy_id(&id_priv->id);
2672         kfree(work);
2673 }
2674
2675 static void cma_init_resolve_route_work(struct cma_work *work,
2676                                         struct rdma_id_private *id_priv)
2677 {
2678         work->id = id_priv;
2679         INIT_WORK(&work->work, cma_work_handler);
2680         work->old_state = RDMA_CM_ROUTE_QUERY;
2681         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2682         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2683 }
2684
2685 static void cma_init_resolve_addr_work(struct cma_work *work,
2686                                        struct rdma_id_private *id_priv)
2687 {
2688         work->id = id_priv;
2689         INIT_WORK(&work->work, cma_work_handler);
2690         work->old_state = RDMA_CM_ADDR_QUERY;
2691         work->new_state = RDMA_CM_ADDR_RESOLVED;
2692         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2693 }
2694
2695 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2696                                 unsigned long timeout_ms)
2697 {
2698         struct rdma_route *route = &id_priv->id.route;
2699         struct cma_work *work;
2700         int ret;
2701
2702         work = kzalloc(sizeof *work, GFP_KERNEL);
2703         if (!work)
2704                 return -ENOMEM;
2705
2706         cma_init_resolve_route_work(work, id_priv);
2707
2708         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2709         if (!route->path_rec) {
2710                 ret = -ENOMEM;
2711                 goto err1;
2712         }
2713
2714         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2715         if (ret)
2716                 goto err2;
2717
2718         return 0;
2719 err2:
2720         kfree(route->path_rec);
2721         route->path_rec = NULL;
2722 err1:
2723         kfree(work);
2724         return ret;
2725 }
2726
2727 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2728                                            unsigned long supported_gids,
2729                                            enum ib_gid_type default_gid)
2730 {
2731         if ((network_type == RDMA_NETWORK_IPV4 ||
2732              network_type == RDMA_NETWORK_IPV6) &&
2733             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2734                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2735
2736         return default_gid;
2737 }
2738
2739 /*
2740  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2741  * path record type based on GID type.
2742  * It also sets up other L2 fields which includes destination mac address
2743  * netdev ifindex, of the path record.
2744  * It returns the netdev of the bound interface for this path record entry.
2745  */
2746 static struct net_device *
2747 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2748 {
2749         struct rdma_route *route = &id_priv->id.route;
2750         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2751         struct rdma_addr *addr = &route->addr;
2752         unsigned long supported_gids;
2753         struct net_device *ndev;
2754
2755         if (!addr->dev_addr.bound_dev_if)
2756                 return NULL;
2757
2758         ndev = dev_get_by_index(addr->dev_addr.net,
2759                                 addr->dev_addr.bound_dev_if);
2760         if (!ndev)
2761                 return NULL;
2762
2763         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2764                                                     id_priv->id.port_num);
2765         gid_type = cma_route_gid_type(addr->dev_addr.network,
2766                                       supported_gids,
2767                                       id_priv->gid_type);
2768         /* Use the hint from IP Stack to select GID Type */
2769         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2770                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2771         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2772
2773         route->path_rec->roce.route_resolved = true;
2774         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2775         return ndev;
2776 }
2777
2778 int rdma_set_ib_path(struct rdma_cm_id *id,
2779                      struct sa_path_rec *path_rec)
2780 {
2781         struct rdma_id_private *id_priv;
2782         struct net_device *ndev;
2783         int ret;
2784
2785         id_priv = container_of(id, struct rdma_id_private, id);
2786         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2787                            RDMA_CM_ROUTE_RESOLVED))
2788                 return -EINVAL;
2789
2790         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2791                                      GFP_KERNEL);
2792         if (!id->route.path_rec) {
2793                 ret = -ENOMEM;
2794                 goto err;
2795         }
2796
2797         if (rdma_protocol_roce(id->device, id->port_num)) {
2798                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2799                 if (!ndev) {
2800                         ret = -ENODEV;
2801                         goto err_free;
2802                 }
2803                 dev_put(ndev);
2804         }
2805
2806         id->route.num_paths = 1;
2807         return 0;
2808
2809 err_free:
2810         kfree(id->route.path_rec);
2811         id->route.path_rec = NULL;
2812 err:
2813         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2814         return ret;
2815 }
2816 EXPORT_SYMBOL(rdma_set_ib_path);
2817
2818 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2819 {
2820         struct cma_work *work;
2821
2822         work = kzalloc(sizeof *work, GFP_KERNEL);
2823         if (!work)
2824                 return -ENOMEM;
2825
2826         cma_init_resolve_route_work(work, id_priv);
2827         queue_work(cma_wq, &work->work);
2828         return 0;
2829 }
2830
2831 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2832 {
2833         int prio;
2834         struct net_device *dev;
2835
2836         prio = rt_tos2priority(tos);
2837         dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
2838         if (dev->num_tc)
2839                 return netdev_get_prio_tc_map(dev, prio);
2840
2841 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2842         if (is_vlan_dev(ndev))
2843                 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2844                         VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2845 #endif
2846         return 0;
2847 }
2848
2849 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2850 {
2851         struct rdma_route *route = &id_priv->id.route;
2852         struct rdma_addr *addr = &route->addr;
2853         struct cma_work *work;
2854         int ret;
2855         struct net_device *ndev;
2856
2857         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2858                                         rdma_start_port(id_priv->cma_dev->device)];
2859         u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2860
2861
2862         work = kzalloc(sizeof *work, GFP_KERNEL);
2863         if (!work)
2864                 return -ENOMEM;
2865
2866         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2867         if (!route->path_rec) {
2868                 ret = -ENOMEM;
2869                 goto err1;
2870         }
2871
2872         route->num_paths = 1;
2873
2874         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2875         if (!ndev) {
2876                 ret = -ENODEV;
2877                 goto err2;
2878         }
2879
2880         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2881                     &route->path_rec->sgid);
2882         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2883                     &route->path_rec->dgid);
2884
2885         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2886                 /* TODO: get the hoplimit from the inet/inet6 device */
2887                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2888         else
2889                 route->path_rec->hop_limit = 1;
2890         route->path_rec->reversible = 1;
2891         route->path_rec->pkey = cpu_to_be16(0xffff);
2892         route->path_rec->mtu_selector = IB_SA_EQ;
2893         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2894         route->path_rec->traffic_class = tos;
2895         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2896         route->path_rec->rate_selector = IB_SA_EQ;
2897         route->path_rec->rate = iboe_get_rate(ndev);
2898         dev_put(ndev);
2899         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2900         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2901         if (!route->path_rec->mtu) {
2902                 ret = -EINVAL;
2903                 goto err2;
2904         }
2905
2906         cma_init_resolve_route_work(work, id_priv);
2907         queue_work(cma_wq, &work->work);
2908
2909         return 0;
2910
2911 err2:
2912         kfree(route->path_rec);
2913         route->path_rec = NULL;
2914 err1:
2915         kfree(work);
2916         return ret;
2917 }
2918
2919 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
2920 {
2921         struct rdma_id_private *id_priv;
2922         int ret;
2923
2924         id_priv = container_of(id, struct rdma_id_private, id);
2925         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2926                 return -EINVAL;
2927
2928         atomic_inc(&id_priv->refcount);
2929         if (rdma_cap_ib_sa(id->device, id->port_num))
2930                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2931         else if (rdma_protocol_roce(id->device, id->port_num))
2932                 ret = cma_resolve_iboe_route(id_priv);
2933         else if (rdma_protocol_iwarp(id->device, id->port_num))
2934                 ret = cma_resolve_iw_route(id_priv);
2935         else
2936                 ret = -ENOSYS;
2937
2938         if (ret)
2939                 goto err;
2940
2941         return 0;
2942 err:
2943         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2944         cma_deref_id(id_priv);
2945         return ret;
2946 }
2947 EXPORT_SYMBOL(rdma_resolve_route);
2948
2949 static void cma_set_loopback(struct sockaddr *addr)
2950 {
2951         switch (addr->sa_family) {
2952         case AF_INET:
2953                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2954                 break;
2955         case AF_INET6:
2956                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2957                               0, 0, 0, htonl(1));
2958                 break;
2959         default:
2960                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2961                             0, 0, 0, htonl(1));
2962                 break;
2963         }
2964 }
2965
2966 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2967 {
2968         struct cma_device *cma_dev, *cur_dev;
2969         union ib_gid gid;
2970         enum ib_port_state port_state;
2971         u16 pkey;
2972         int ret;
2973         u8 p;
2974
2975         cma_dev = NULL;
2976         mutex_lock(&lock);
2977         list_for_each_entry(cur_dev, &dev_list, list) {
2978                 if (cma_family(id_priv) == AF_IB &&
2979                     !rdma_cap_ib_cm(cur_dev->device, 1))
2980                         continue;
2981
2982                 if (!cma_dev)
2983                         cma_dev = cur_dev;
2984
2985                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2986                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
2987                             port_state == IB_PORT_ACTIVE) {
2988                                 cma_dev = cur_dev;
2989                                 goto port_found;
2990                         }
2991                 }
2992         }
2993
2994         if (!cma_dev) {
2995                 ret = -ENODEV;
2996                 goto out;
2997         }
2998
2999         p = 1;
3000
3001 port_found:
3002         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3003         if (ret)
3004                 goto out;
3005
3006         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3007         if (ret)
3008                 goto out;
3009
3010         id_priv->id.route.addr.dev_addr.dev_type =
3011                 (rdma_protocol_ib(cma_dev->device, p)) ?
3012                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3013
3014         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3015         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3016         id_priv->id.port_num = p;
3017         cma_attach_to_dev(id_priv, cma_dev);
3018         cma_set_loopback(cma_src_addr(id_priv));
3019 out:
3020         mutex_unlock(&lock);
3021         return ret;
3022 }
3023
3024 static void addr_handler(int status, struct sockaddr *src_addr,
3025                          struct rdma_dev_addr *dev_addr, void *context)
3026 {
3027         struct rdma_id_private *id_priv = context;
3028         struct rdma_cm_event event = {};
3029         struct sockaddr *addr;
3030         struct sockaddr_storage old_addr;
3031
3032         mutex_lock(&id_priv->handler_mutex);
3033         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3034                            RDMA_CM_ADDR_RESOLVED))
3035                 goto out;
3036
3037         /*
3038          * Store the previous src address, so that if we fail to acquire
3039          * matching rdma device, old address can be restored back, which helps
3040          * to cancel the cma listen operation correctly.
3041          */
3042         addr = cma_src_addr(id_priv);
3043         memcpy(&old_addr, addr, rdma_addr_size(addr));
3044         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3045         if (!status && !id_priv->cma_dev) {
3046                 status = cma_acquire_dev_by_src_ip(id_priv);
3047                 if (status)
3048                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3049                                              status);
3050         } else if (status) {
3051                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3052         }
3053
3054         if (status) {
3055                 memcpy(addr, &old_addr,
3056                        rdma_addr_size((struct sockaddr *)&old_addr));
3057                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3058                                    RDMA_CM_ADDR_BOUND))
3059                         goto out;
3060                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3061                 event.status = status;
3062         } else
3063                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3064
3065         if (id_priv->id.event_handler(&id_priv->id, &event)) {
3066                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3067                 mutex_unlock(&id_priv->handler_mutex);
3068                 rdma_destroy_id(&id_priv->id);
3069                 return;
3070         }
3071 out:
3072         mutex_unlock(&id_priv->handler_mutex);
3073 }
3074
3075 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3076 {
3077         struct cma_work *work;
3078         union ib_gid gid;
3079         int ret;
3080
3081         work = kzalloc(sizeof *work, GFP_KERNEL);
3082         if (!work)
3083                 return -ENOMEM;
3084
3085         if (!id_priv->cma_dev) {
3086                 ret = cma_bind_loopback(id_priv);
3087                 if (ret)
3088                         goto err;
3089         }
3090
3091         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3092         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3093
3094         cma_init_resolve_addr_work(work, id_priv);
3095         queue_work(cma_wq, &work->work);
3096         return 0;
3097 err:
3098         kfree(work);
3099         return ret;
3100 }
3101
3102 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3103 {
3104         struct cma_work *work;
3105         int ret;
3106
3107         work = kzalloc(sizeof *work, GFP_KERNEL);
3108         if (!work)
3109                 return -ENOMEM;
3110
3111         if (!id_priv->cma_dev) {
3112                 ret = cma_resolve_ib_dev(id_priv);
3113                 if (ret)
3114                         goto err;
3115         }
3116
3117         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3118                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3119
3120         cma_init_resolve_addr_work(work, id_priv);
3121         queue_work(cma_wq, &work->work);
3122         return 0;
3123 err:
3124         kfree(work);
3125         return ret;
3126 }
3127
3128 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3129                          const struct sockaddr *dst_addr)
3130 {
3131         if (!src_addr || !src_addr->sa_family) {
3132                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3133                 src_addr->sa_family = dst_addr->sa_family;
3134                 if (IS_ENABLED(CONFIG_IPV6) &&
3135                     dst_addr->sa_family == AF_INET6) {
3136                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3137                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3138                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3139                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3140                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3141                 } else if (dst_addr->sa_family == AF_IB) {
3142                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3143                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3144                 }
3145         }
3146         return rdma_bind_addr(id, src_addr);
3147 }
3148
3149 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3150                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3151 {
3152         struct rdma_id_private *id_priv;
3153         int ret;
3154
3155         id_priv = container_of(id, struct rdma_id_private, id);
3156         if (id_priv->state == RDMA_CM_IDLE) {
3157                 ret = cma_bind_addr(id, src_addr, dst_addr);
3158                 if (ret)
3159                         return ret;
3160         }
3161
3162         if (cma_family(id_priv) != dst_addr->sa_family)
3163                 return -EINVAL;
3164
3165         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
3166                 return -EINVAL;
3167
3168         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3169         if (cma_any_addr(dst_addr)) {
3170                 ret = cma_resolve_loopback(id_priv);
3171         } else {
3172                 if (dst_addr->sa_family == AF_IB) {
3173                         ret = cma_resolve_ib_addr(id_priv);
3174                 } else {
3175                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3176                                               &id->route.addr.dev_addr,
3177                                               timeout_ms, addr_handler,
3178                                               false, id_priv);
3179                 }
3180         }
3181         if (ret)
3182                 goto err;
3183
3184         return 0;
3185 err:
3186         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3187         return ret;
3188 }
3189 EXPORT_SYMBOL(rdma_resolve_addr);
3190
3191 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3192 {
3193         struct rdma_id_private *id_priv;
3194         unsigned long flags;
3195         int ret;
3196
3197         id_priv = container_of(id, struct rdma_id_private, id);
3198         spin_lock_irqsave(&id_priv->lock, flags);
3199         if (reuse || id_priv->state == RDMA_CM_IDLE) {
3200                 id_priv->reuseaddr = reuse;
3201                 ret = 0;
3202         } else {
3203                 ret = -EINVAL;
3204         }
3205         spin_unlock_irqrestore(&id_priv->lock, flags);
3206         return ret;
3207 }
3208 EXPORT_SYMBOL(rdma_set_reuseaddr);
3209
3210 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3211 {
3212         struct rdma_id_private *id_priv;
3213         unsigned long flags;
3214         int ret;
3215
3216         id_priv = container_of(id, struct rdma_id_private, id);
3217         spin_lock_irqsave(&id_priv->lock, flags);
3218         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3219                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3220                 id_priv->afonly = afonly;
3221                 ret = 0;
3222         } else {
3223                 ret = -EINVAL;
3224         }
3225         spin_unlock_irqrestore(&id_priv->lock, flags);
3226         return ret;
3227 }
3228 EXPORT_SYMBOL(rdma_set_afonly);
3229
3230 static void cma_bind_port(struct rdma_bind_list *bind_list,
3231                           struct rdma_id_private *id_priv)
3232 {
3233         struct sockaddr *addr;
3234         struct sockaddr_ib *sib;
3235         u64 sid, mask;
3236         __be16 port;
3237
3238         addr = cma_src_addr(id_priv);
3239         port = htons(bind_list->port);
3240
3241         switch (addr->sa_family) {
3242         case AF_INET:
3243                 ((struct sockaddr_in *) addr)->sin_port = port;
3244                 break;
3245         case AF_INET6:
3246                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3247                 break;
3248         case AF_IB:
3249                 sib = (struct sockaddr_ib *) addr;
3250                 sid = be64_to_cpu(sib->sib_sid);
3251                 mask = be64_to_cpu(sib->sib_sid_mask);
3252                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3253                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3254                 break;
3255         }
3256         id_priv->bind_list = bind_list;
3257         hlist_add_head(&id_priv->node, &bind_list->owners);
3258 }
3259
3260 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3261                           struct rdma_id_private *id_priv, unsigned short snum)
3262 {
3263         struct rdma_bind_list *bind_list;
3264         int ret;
3265
3266         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3267         if (!bind_list)
3268                 return -ENOMEM;
3269
3270         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3271                            snum);
3272         if (ret < 0)
3273                 goto err;
3274
3275         bind_list->ps = ps;
3276         bind_list->port = snum;
3277         cma_bind_port(bind_list, id_priv);
3278         return 0;
3279 err:
3280         kfree(bind_list);
3281         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3282 }
3283
3284 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3285                               struct rdma_id_private *id_priv)
3286 {
3287         struct rdma_id_private *cur_id;
3288         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3289         struct sockaddr  *saddr = cma_src_addr(id_priv);
3290         __be16 dport = cma_port(daddr);
3291
3292         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3293                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3294                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3295                 __be16 cur_dport = cma_port(cur_daddr);
3296
3297                 if (id_priv == cur_id)
3298                         continue;
3299
3300                 /* different dest port -> unique */
3301                 if (!cma_any_port(daddr) &&
3302                     !cma_any_port(cur_daddr) &&
3303                     (dport != cur_dport))
3304                         continue;
3305
3306                 /* different src address -> unique */
3307                 if (!cma_any_addr(saddr) &&
3308                     !cma_any_addr(cur_saddr) &&
3309                     cma_addr_cmp(saddr, cur_saddr))
3310                         continue;
3311
3312                 /* different dst address -> unique */
3313                 if (!cma_any_addr(daddr) &&
3314                     !cma_any_addr(cur_daddr) &&
3315                     cma_addr_cmp(daddr, cur_daddr))
3316                         continue;
3317
3318                 return -EADDRNOTAVAIL;
3319         }
3320         return 0;
3321 }
3322
3323 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3324                               struct rdma_id_private *id_priv)
3325 {
3326         static unsigned int last_used_port;
3327         int low, high, remaining;
3328         unsigned int rover;
3329         struct net *net = id_priv->id.route.addr.dev_addr.net;
3330
3331         inet_get_local_port_range(net, &low, &high);
3332         remaining = (high - low) + 1;
3333         rover = prandom_u32() % remaining + low;
3334 retry:
3335         if (last_used_port != rover) {
3336                 struct rdma_bind_list *bind_list;
3337                 int ret;
3338
3339                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3340
3341                 if (!bind_list) {
3342                         ret = cma_alloc_port(ps, id_priv, rover);
3343                 } else {
3344                         ret = cma_port_is_unique(bind_list, id_priv);
3345                         if (!ret)
3346                                 cma_bind_port(bind_list, id_priv);
3347                 }
3348                 /*
3349                  * Remember previously used port number in order to avoid
3350                  * re-using same port immediately after it is closed.
3351                  */
3352                 if (!ret)
3353                         last_used_port = rover;
3354                 if (ret != -EADDRNOTAVAIL)
3355                         return ret;
3356         }
3357         if (--remaining) {
3358                 rover++;
3359                 if ((rover < low) || (rover > high))
3360                         rover = low;
3361                 goto retry;
3362         }
3363         return -EADDRNOTAVAIL;
3364 }
3365
3366 /*
3367  * Check that the requested port is available.  This is called when trying to
3368  * bind to a specific port, or when trying to listen on a bound port.  In
3369  * the latter case, the provided id_priv may already be on the bind_list, but
3370  * we still need to check that it's okay to start listening.
3371  */
3372 static int cma_check_port(struct rdma_bind_list *bind_list,
3373                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3374 {
3375         struct rdma_id_private *cur_id;
3376         struct sockaddr *addr, *cur_addr;
3377
3378         addr = cma_src_addr(id_priv);
3379         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3380                 if (id_priv == cur_id)
3381                         continue;
3382
3383                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3384                     cur_id->reuseaddr)
3385                         continue;
3386
3387                 cur_addr = cma_src_addr(cur_id);
3388                 if (id_priv->afonly && cur_id->afonly &&
3389                     (addr->sa_family != cur_addr->sa_family))
3390                         continue;
3391
3392                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3393                         return -EADDRNOTAVAIL;
3394
3395                 if (!cma_addr_cmp(addr, cur_addr))
3396                         return -EADDRINUSE;
3397         }
3398         return 0;
3399 }
3400
3401 static int cma_use_port(enum rdma_ucm_port_space ps,
3402                         struct rdma_id_private *id_priv)
3403 {
3404         struct rdma_bind_list *bind_list;
3405         unsigned short snum;
3406         int ret;
3407
3408         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3409         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3410                 return -EACCES;
3411
3412         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3413         if (!bind_list) {
3414                 ret = cma_alloc_port(ps, id_priv, snum);
3415         } else {
3416                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3417                 if (!ret)
3418                         cma_bind_port(bind_list, id_priv);
3419         }
3420         return ret;
3421 }
3422
3423 static int cma_bind_listen(struct rdma_id_private *id_priv)
3424 {
3425         struct rdma_bind_list *bind_list = id_priv->bind_list;
3426         int ret = 0;
3427
3428         mutex_lock(&lock);
3429         if (bind_list->owners.first->next)
3430                 ret = cma_check_port(bind_list, id_priv, 0);
3431         mutex_unlock(&lock);
3432         return ret;
3433 }
3434
3435 static enum rdma_ucm_port_space
3436 cma_select_inet_ps(struct rdma_id_private *id_priv)
3437 {
3438         switch (id_priv->id.ps) {
3439         case RDMA_PS_TCP:
3440         case RDMA_PS_UDP:
3441         case RDMA_PS_IPOIB:
3442         case RDMA_PS_IB:
3443                 return id_priv->id.ps;
3444         default:
3445
3446                 return 0;
3447         }
3448 }
3449
3450 static enum rdma_ucm_port_space
3451 cma_select_ib_ps(struct rdma_id_private *id_priv)
3452 {
3453         enum rdma_ucm_port_space ps = 0;
3454         struct sockaddr_ib *sib;
3455         u64 sid_ps, mask, sid;
3456
3457         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3458         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3459         sid = be64_to_cpu(sib->sib_sid) & mask;
3460
3461         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3462                 sid_ps = RDMA_IB_IP_PS_IB;
3463                 ps = RDMA_PS_IB;
3464         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3465                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3466                 sid_ps = RDMA_IB_IP_PS_TCP;
3467                 ps = RDMA_PS_TCP;
3468         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3469                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3470                 sid_ps = RDMA_IB_IP_PS_UDP;
3471                 ps = RDMA_PS_UDP;
3472         }
3473
3474         if (ps) {
3475                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3476                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3477                                                 be64_to_cpu(sib->sib_sid_mask));
3478         }
3479         return ps;
3480 }
3481
3482 static int cma_get_port(struct rdma_id_private *id_priv)
3483 {
3484         enum rdma_ucm_port_space ps;
3485         int ret;
3486
3487         if (cma_family(id_priv) != AF_IB)
3488                 ps = cma_select_inet_ps(id_priv);
3489         else
3490                 ps = cma_select_ib_ps(id_priv);
3491         if (!ps)
3492                 return -EPROTONOSUPPORT;
3493
3494         mutex_lock(&lock);
3495         if (cma_any_port(cma_src_addr(id_priv)))
3496                 ret = cma_alloc_any_port(ps, id_priv);
3497         else
3498                 ret = cma_use_port(ps, id_priv);
3499         mutex_unlock(&lock);
3500
3501         return ret;
3502 }
3503
3504 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3505                                struct sockaddr *addr)
3506 {
3507 #if IS_ENABLED(CONFIG_IPV6)
3508         struct sockaddr_in6 *sin6;
3509
3510         if (addr->sa_family != AF_INET6)
3511                 return 0;
3512
3513         sin6 = (struct sockaddr_in6 *) addr;
3514
3515         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3516                 return 0;
3517
3518         if (!sin6->sin6_scope_id)
3519                         return -EINVAL;
3520
3521         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3522 #endif
3523         return 0;
3524 }
3525
3526 int rdma_listen(struct rdma_cm_id *id, int backlog)
3527 {
3528         struct rdma_id_private *id_priv;
3529         int ret;
3530
3531         id_priv = container_of(id, struct rdma_id_private, id);
3532         if (id_priv->state == RDMA_CM_IDLE) {
3533                 id->route.addr.src_addr.ss_family = AF_INET;
3534                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3535                 if (ret)
3536                         return ret;
3537         }
3538
3539         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3540                 return -EINVAL;
3541
3542         if (id_priv->reuseaddr) {
3543                 ret = cma_bind_listen(id_priv);
3544                 if (ret)
3545                         goto err;
3546         }
3547
3548         id_priv->backlog = backlog;
3549         if (id->device) {
3550                 if (rdma_cap_ib_cm(id->device, 1)) {
3551                         ret = cma_ib_listen(id_priv);
3552                         if (ret)
3553                                 goto err;
3554                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3555                         ret = cma_iw_listen(id_priv, backlog);
3556                         if (ret)
3557                                 goto err;
3558                 } else {
3559                         ret = -ENOSYS;
3560                         goto err;
3561                 }
3562         } else
3563                 cma_listen_on_all(id_priv);
3564
3565         return 0;
3566 err:
3567         id_priv->backlog = 0;
3568         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3569         return ret;
3570 }
3571 EXPORT_SYMBOL(rdma_listen);
3572
3573 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3574 {
3575         struct rdma_id_private *id_priv;
3576         int ret;
3577         struct sockaddr  *daddr;
3578
3579         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3580             addr->sa_family != AF_IB)
3581                 return -EAFNOSUPPORT;
3582
3583         id_priv = container_of(id, struct rdma_id_private, id);
3584         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3585                 return -EINVAL;
3586
3587         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3588         if (ret)
3589                 goto err1;
3590
3591         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3592         if (!cma_any_addr(addr)) {
3593                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3594                 if (ret)
3595                         goto err1;
3596
3597                 ret = cma_acquire_dev_by_src_ip(id_priv);
3598                 if (ret)
3599                         goto err1;
3600         }
3601
3602         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3603                 if (addr->sa_family == AF_INET)
3604                         id_priv->afonly = 1;
3605 #if IS_ENABLED(CONFIG_IPV6)
3606                 else if (addr->sa_family == AF_INET6) {
3607                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3608
3609                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3610                 }
3611 #endif
3612         }
3613         daddr = cma_dst_addr(id_priv);
3614         daddr->sa_family = addr->sa_family;
3615
3616         ret = cma_get_port(id_priv);
3617         if (ret)
3618                 goto err2;
3619
3620         return 0;
3621 err2:
3622         rdma_restrack_del(&id_priv->res);
3623         if (id_priv->cma_dev)
3624                 cma_release_dev(id_priv);
3625 err1:
3626         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3627         return ret;
3628 }
3629 EXPORT_SYMBOL(rdma_bind_addr);
3630
3631 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3632 {
3633         struct cma_hdr *cma_hdr;
3634
3635         cma_hdr = hdr;
3636         cma_hdr->cma_version = CMA_VERSION;
3637         if (cma_family(id_priv) == AF_INET) {
3638                 struct sockaddr_in *src4, *dst4;
3639
3640                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3641                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3642
3643                 cma_set_ip_ver(cma_hdr, 4);
3644                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3645                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3646                 cma_hdr->port = src4->sin_port;
3647         } else if (cma_family(id_priv) == AF_INET6) {
3648                 struct sockaddr_in6 *src6, *dst6;
3649
3650                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3651                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3652
3653                 cma_set_ip_ver(cma_hdr, 6);
3654                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3655                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3656                 cma_hdr->port = src6->sin6_port;
3657         }
3658         return 0;
3659 }
3660
3661 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3662                                 const struct ib_cm_event *ib_event)
3663 {
3664         struct rdma_id_private *id_priv = cm_id->context;
3665         struct rdma_cm_event event = {};
3666         const struct ib_cm_sidr_rep_event_param *rep =
3667                                 &ib_event->param.sidr_rep_rcvd;
3668         int ret = 0;
3669
3670         mutex_lock(&id_priv->handler_mutex);
3671         if (id_priv->state != RDMA_CM_CONNECT)
3672                 goto out;
3673
3674         switch (ib_event->event) {
3675         case IB_CM_SIDR_REQ_ERROR:
3676                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3677                 event.status = -ETIMEDOUT;
3678                 break;
3679         case IB_CM_SIDR_REP_RECEIVED:
3680                 event.param.ud.private_data = ib_event->private_data;
3681                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3682                 if (rep->status != IB_SIDR_SUCCESS) {
3683                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3684                         event.status = ib_event->param.sidr_rep_rcvd.status;
3685                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3686                                              event.status);
3687                         break;
3688                 }
3689                 ret = cma_set_qkey(id_priv, rep->qkey);
3690                 if (ret) {
3691                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3692                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3693                         event.status = ret;
3694                         break;
3695                 }
3696                 ib_init_ah_attr_from_path(id_priv->id.device,
3697                                           id_priv->id.port_num,
3698                                           id_priv->id.route.path_rec,
3699                                           &event.param.ud.ah_attr,
3700                                           rep->sgid_attr);
3701                 event.param.ud.qp_num = rep->qpn;
3702                 event.param.ud.qkey = rep->qkey;
3703                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3704                 event.status = 0;
3705                 break;
3706         default:
3707                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3708                        ib_event->event);
3709                 goto out;
3710         }
3711
3712         ret = id_priv->id.event_handler(&id_priv->id, &event);
3713
3714         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3715         if (ret) {
3716                 /* Destroy the CM ID by returning a non-zero value. */
3717                 id_priv->cm_id.ib = NULL;
3718                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3719                 mutex_unlock(&id_priv->handler_mutex);
3720                 rdma_destroy_id(&id_priv->id);
3721                 return ret;
3722         }
3723 out:
3724         mutex_unlock(&id_priv->handler_mutex);
3725         return ret;
3726 }
3727
3728 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3729                               struct rdma_conn_param *conn_param)
3730 {
3731         struct ib_cm_sidr_req_param req;
3732         struct ib_cm_id *id;
3733         void *private_data;
3734         u8 offset;
3735         int ret;
3736
3737         memset(&req, 0, sizeof req);
3738         offset = cma_user_data_offset(id_priv);
3739         req.private_data_len = offset + conn_param->private_data_len;
3740         if (req.private_data_len < conn_param->private_data_len)
3741                 return -EINVAL;
3742
3743         if (req.private_data_len) {
3744                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3745                 if (!private_data)
3746                         return -ENOMEM;
3747         } else {
3748                 private_data = NULL;
3749         }
3750
3751         if (conn_param->private_data && conn_param->private_data_len)
3752                 memcpy(private_data + offset, conn_param->private_data,
3753                        conn_param->private_data_len);
3754
3755         if (private_data) {
3756                 ret = cma_format_hdr(private_data, id_priv);
3757                 if (ret)
3758                         goto out;
3759                 req.private_data = private_data;
3760         }
3761
3762         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3763                              id_priv);
3764         if (IS_ERR(id)) {
3765                 ret = PTR_ERR(id);
3766                 goto out;
3767         }
3768         id_priv->cm_id.ib = id;
3769
3770         req.path = id_priv->id.route.path_rec;
3771         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3772         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3773         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3774         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3775
3776         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3777         if (ret) {
3778                 ib_destroy_cm_id(id_priv->cm_id.ib);
3779                 id_priv->cm_id.ib = NULL;
3780         }
3781 out:
3782         kfree(private_data);
3783         return ret;
3784 }
3785
3786 static int cma_connect_ib(struct rdma_id_private *id_priv,
3787                           struct rdma_conn_param *conn_param)
3788 {
3789         struct ib_cm_req_param req;
3790         struct rdma_route *route;
3791         void *private_data;
3792         struct ib_cm_id *id;
3793         u8 offset;
3794         int ret;
3795
3796         memset(&req, 0, sizeof req);
3797         offset = cma_user_data_offset(id_priv);
3798         req.private_data_len = offset + conn_param->private_data_len;
3799         if (req.private_data_len < conn_param->private_data_len)
3800                 return -EINVAL;
3801
3802         if (req.private_data_len) {
3803                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3804                 if (!private_data)
3805                         return -ENOMEM;
3806         } else {
3807                 private_data = NULL;
3808         }
3809
3810         if (conn_param->private_data && conn_param->private_data_len)
3811                 memcpy(private_data + offset, conn_param->private_data,
3812                        conn_param->private_data_len);
3813
3814         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3815         if (IS_ERR(id)) {
3816                 ret = PTR_ERR(id);
3817                 goto out;
3818         }
3819         id_priv->cm_id.ib = id;
3820
3821         route = &id_priv->id.route;
3822         if (private_data) {
3823                 ret = cma_format_hdr(private_data, id_priv);
3824                 if (ret)
3825                         goto out;
3826                 req.private_data = private_data;
3827         }
3828
3829         req.primary_path = &route->path_rec[0];
3830         if (route->num_paths == 2)
3831                 req.alternate_path = &route->path_rec[1];
3832
3833         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3834         /* Alternate path SGID attribute currently unsupported */
3835         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3836         req.qp_num = id_priv->qp_num;
3837         req.qp_type = id_priv->id.qp_type;
3838         req.starting_psn = id_priv->seq_num;
3839         req.responder_resources = conn_param->responder_resources;
3840         req.initiator_depth = conn_param->initiator_depth;
3841         req.flow_control = conn_param->flow_control;
3842         req.retry_count = min_t(u8, 7, conn_param->retry_count);
3843         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3844         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3845         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3846         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3847         req.srq = id_priv->srq ? 1 : 0;
3848
3849         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3850 out:
3851         if (ret && !IS_ERR(id)) {
3852                 ib_destroy_cm_id(id);
3853                 id_priv->cm_id.ib = NULL;
3854         }
3855
3856         kfree(private_data);
3857         return ret;
3858 }
3859
3860 static int cma_connect_iw(struct rdma_id_private *id_priv,
3861                           struct rdma_conn_param *conn_param)
3862 {
3863         struct iw_cm_id *cm_id;
3864         int ret;
3865         struct iw_cm_conn_param iw_param;
3866
3867         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3868         if (IS_ERR(cm_id))
3869                 return PTR_ERR(cm_id);
3870
3871         cm_id->tos = id_priv->tos;
3872         cm_id->tos_set = id_priv->tos_set;
3873         id_priv->cm_id.iw = cm_id;
3874
3875         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3876                rdma_addr_size(cma_src_addr(id_priv)));
3877         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3878                rdma_addr_size(cma_dst_addr(id_priv)));
3879
3880         ret = cma_modify_qp_rtr(id_priv, conn_param);
3881         if (ret)
3882                 goto out;
3883
3884         if (conn_param) {
3885                 iw_param.ord = conn_param->initiator_depth;
3886                 iw_param.ird = conn_param->responder_resources;
3887                 iw_param.private_data = conn_param->private_data;
3888                 iw_param.private_data_len = conn_param->private_data_len;
3889                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3890         } else {
3891                 memset(&iw_param, 0, sizeof iw_param);
3892                 iw_param.qpn = id_priv->qp_num;
3893         }
3894         ret = iw_cm_connect(cm_id, &iw_param);
3895 out:
3896         if (ret) {
3897                 iw_destroy_cm_id(cm_id);
3898                 id_priv->cm_id.iw = NULL;
3899         }
3900         return ret;
3901 }
3902
3903 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3904 {
3905         struct rdma_id_private *id_priv;
3906         int ret;
3907
3908         id_priv = container_of(id, struct rdma_id_private, id);
3909         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3910                 return -EINVAL;
3911
3912         if (!id->qp) {
3913                 id_priv->qp_num = conn_param->qp_num;
3914                 id_priv->srq = conn_param->srq;
3915         }
3916
3917         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3918                 if (id->qp_type == IB_QPT_UD)
3919                         ret = cma_resolve_ib_udp(id_priv, conn_param);
3920                 else
3921                         ret = cma_connect_ib(id_priv, conn_param);
3922         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3923                 ret = cma_connect_iw(id_priv, conn_param);
3924         else
3925                 ret = -ENOSYS;
3926         if (ret)
3927                 goto err;
3928
3929         return 0;
3930 err:
3931         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3932         return ret;
3933 }
3934 EXPORT_SYMBOL(rdma_connect);
3935
3936 static int cma_accept_ib(struct rdma_id_private *id_priv,
3937                          struct rdma_conn_param *conn_param)
3938 {
3939         struct ib_cm_rep_param rep;
3940         int ret;
3941
3942         ret = cma_modify_qp_rtr(id_priv, conn_param);
3943         if (ret)
3944                 goto out;
3945
3946         ret = cma_modify_qp_rts(id_priv, conn_param);
3947         if (ret)
3948                 goto out;
3949
3950         memset(&rep, 0, sizeof rep);
3951         rep.qp_num = id_priv->qp_num;
3952         rep.starting_psn = id_priv->seq_num;
3953         rep.private_data = conn_param->private_data;
3954         rep.private_data_len = conn_param->private_data_len;
3955         rep.responder_resources = conn_param->responder_resources;
3956         rep.initiator_depth = conn_param->initiator_depth;
3957         rep.failover_accepted = 0;
3958         rep.flow_control = conn_param->flow_control;
3959         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3960         rep.srq = id_priv->srq ? 1 : 0;
3961
3962         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3963 out:
3964         return ret;
3965 }
3966
3967 static int cma_accept_iw(struct rdma_id_private *id_priv,
3968                   struct rdma_conn_param *conn_param)
3969 {
3970         struct iw_cm_conn_param iw_param;
3971         int ret;
3972
3973         if (!conn_param)
3974                 return -EINVAL;
3975
3976         ret = cma_modify_qp_rtr(id_priv, conn_param);
3977         if (ret)
3978                 return ret;
3979
3980         iw_param.ord = conn_param->initiator_depth;
3981         iw_param.ird = conn_param->responder_resources;
3982         iw_param.private_data = conn_param->private_data;
3983         iw_param.private_data_len = conn_param->private_data_len;
3984         if (id_priv->id.qp) {
3985                 iw_param.qpn = id_priv->qp_num;
3986         } else
3987                 iw_param.qpn = conn_param->qp_num;
3988
3989         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3990 }
3991
3992 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3993                              enum ib_cm_sidr_status status, u32 qkey,
3994                              const void *private_data, int private_data_len)
3995 {
3996         struct ib_cm_sidr_rep_param rep;
3997         int ret;
3998
3999         memset(&rep, 0, sizeof rep);
4000         rep.status = status;
4001         if (status == IB_SIDR_SUCCESS) {
4002                 ret = cma_set_qkey(id_priv, qkey);
4003                 if (ret)
4004                         return ret;
4005                 rep.qp_num = id_priv->qp_num;
4006                 rep.qkey = id_priv->qkey;
4007         }
4008         rep.private_data = private_data;
4009         rep.private_data_len = private_data_len;
4010
4011         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4012 }
4013
4014 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4015                   const char *caller)
4016 {
4017         struct rdma_id_private *id_priv;
4018         int ret;
4019
4020         id_priv = container_of(id, struct rdma_id_private, id);
4021
4022         rdma_restrack_set_task(&id_priv->res, caller);
4023
4024         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
4025                 return -EINVAL;
4026
4027         if (!id->qp && conn_param) {
4028                 id_priv->qp_num = conn_param->qp_num;
4029                 id_priv->srq = conn_param->srq;
4030         }
4031
4032         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4033                 if (id->qp_type == IB_QPT_UD) {
4034                         if (conn_param)
4035                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4036                                                         conn_param->qkey,
4037                                                         conn_param->private_data,
4038                                                         conn_param->private_data_len);
4039                         else
4040                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4041                                                         0, NULL, 0);
4042                 } else {
4043                         if (conn_param)
4044                                 ret = cma_accept_ib(id_priv, conn_param);
4045                         else
4046                                 ret = cma_rep_recv(id_priv);
4047                 }
4048         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4049                 ret = cma_accept_iw(id_priv, conn_param);
4050         else
4051                 ret = -ENOSYS;
4052
4053         if (ret)
4054                 goto reject;
4055
4056         return 0;
4057 reject:
4058         cma_modify_qp_err(id_priv);
4059         rdma_reject(id, NULL, 0);
4060         return ret;
4061 }
4062 EXPORT_SYMBOL(__rdma_accept);
4063
4064 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4065 {
4066         struct rdma_id_private *id_priv;
4067         int ret;
4068
4069         id_priv = container_of(id, struct rdma_id_private, id);
4070         if (!id_priv->cm_id.ib)
4071                 return -EINVAL;
4072
4073         switch (id->device->node_type) {
4074         case RDMA_NODE_IB_CA:
4075                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4076                 break;
4077         default:
4078                 ret = 0;
4079                 break;
4080         }
4081         return ret;
4082 }
4083 EXPORT_SYMBOL(rdma_notify);
4084
4085 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4086                 u8 private_data_len)
4087 {
4088         struct rdma_id_private *id_priv;
4089         int ret;
4090
4091         id_priv = container_of(id, struct rdma_id_private, id);
4092         if (!id_priv->cm_id.ib)
4093                 return -EINVAL;
4094
4095         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4096                 if (id->qp_type == IB_QPT_UD)
4097                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4098                                                 private_data, private_data_len);
4099                 else
4100                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
4101                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
4102                                              0, private_data, private_data_len);
4103         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4104                 ret = iw_cm_reject(id_priv->cm_id.iw,
4105                                    private_data, private_data_len);
4106         } else
4107                 ret = -ENOSYS;
4108
4109         return ret;
4110 }
4111 EXPORT_SYMBOL(rdma_reject);
4112
4113 int rdma_disconnect(struct rdma_cm_id *id)
4114 {
4115         struct rdma_id_private *id_priv;
4116         int ret;
4117
4118         id_priv = container_of(id, struct rdma_id_private, id);
4119         if (!id_priv->cm_id.ib)
4120                 return -EINVAL;
4121
4122         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4123                 ret = cma_modify_qp_err(id_priv);
4124                 if (ret)
4125                         goto out;
4126                 /* Initiate or respond to a disconnect. */
4127                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
4128                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
4129         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4130                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4131         } else
4132                 ret = -EINVAL;
4133
4134 out:
4135         return ret;
4136 }
4137 EXPORT_SYMBOL(rdma_disconnect);
4138
4139 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4140 {
4141         struct rdma_id_private *id_priv;
4142         struct cma_multicast *mc = multicast->context;
4143         struct rdma_cm_event event = {};
4144         int ret = 0;
4145
4146         id_priv = mc->id_priv;
4147         mutex_lock(&id_priv->handler_mutex);
4148         if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4149             id_priv->state != RDMA_CM_ADDR_RESOLVED)
4150                 goto out;
4151
4152         if (!status)
4153                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4154         else
4155                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4156                                      status);
4157         mutex_lock(&id_priv->qp_mutex);
4158         if (!status && id_priv->id.qp) {
4159                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
4160                                          be16_to_cpu(multicast->rec.mlid));
4161                 if (status)
4162                         pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
4163                                              status);
4164         }
4165         mutex_unlock(&id_priv->qp_mutex);
4166
4167         event.status = status;
4168         event.param.ud.private_data = mc->context;
4169         if (!status) {
4170                 struct rdma_dev_addr *dev_addr =
4171                         &id_priv->id.route.addr.dev_addr;
4172                 struct net_device *ndev =
4173                         dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4174                 enum ib_gid_type gid_type =
4175                         id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4176                         rdma_start_port(id_priv->cma_dev->device)];
4177
4178                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4179                 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4180                                                id_priv->id.port_num,
4181                                                &multicast->rec,
4182                                                ndev, gid_type,
4183                                                &event.param.ud.ah_attr);
4184                 if (ret)
4185                         event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4186
4187                 event.param.ud.qp_num = 0xFFFFFF;
4188                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4189                 if (ndev)
4190                         dev_put(ndev);
4191         } else
4192                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4193
4194         ret = id_priv->id.event_handler(&id_priv->id, &event);
4195
4196         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4197         if (ret) {
4198                 cma_exch(id_priv, RDMA_CM_DESTROYING);
4199                 mutex_unlock(&id_priv->handler_mutex);
4200                 rdma_destroy_id(&id_priv->id);
4201                 return 0;
4202         }
4203
4204 out:
4205         mutex_unlock(&id_priv->handler_mutex);
4206         return 0;
4207 }
4208
4209 static void cma_set_mgid(struct rdma_id_private *id_priv,
4210                          struct sockaddr *addr, union ib_gid *mgid)
4211 {
4212         unsigned char mc_map[MAX_ADDR_LEN];
4213         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4214         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4215         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4216
4217         if (cma_any_addr(addr)) {
4218                 memset(mgid, 0, sizeof *mgid);
4219         } else if ((addr->sa_family == AF_INET6) &&
4220                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4221                                                                  0xFF10A01B)) {
4222                 /* IPv6 address is an SA assigned MGID. */
4223                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4224         } else if (addr->sa_family == AF_IB) {
4225                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4226         } else if (addr->sa_family == AF_INET6) {
4227                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4228                 if (id_priv->id.ps == RDMA_PS_UDP)
4229                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4230                 *mgid = *(union ib_gid *) (mc_map + 4);
4231         } else {
4232                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4233                 if (id_priv->id.ps == RDMA_PS_UDP)
4234                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4235                 *mgid = *(union ib_gid *) (mc_map + 4);
4236         }
4237 }
4238
4239 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4240                                  struct cma_multicast *mc)
4241 {
4242         struct ib_sa_mcmember_rec rec;
4243         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4244         ib_sa_comp_mask comp_mask;
4245         int ret;
4246
4247         ib_addr_get_mgid(dev_addr, &rec.mgid);
4248         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4249                                      &rec.mgid, &rec);
4250         if (ret)
4251                 return ret;
4252
4253         ret = cma_set_qkey(id_priv, 0);
4254         if (ret)
4255                 return ret;
4256
4257         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4258         rec.qkey = cpu_to_be32(id_priv->qkey);
4259         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4260         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4261         rec.join_state = mc->join_state;
4262
4263         if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4264             (!ib_sa_sendonly_fullmem_support(&sa_client,
4265                                              id_priv->id.device,
4266                                              id_priv->id.port_num))) {
4267                 dev_warn(
4268                         &id_priv->id.device->dev,
4269                         "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4270                         id_priv->id.port_num);
4271                 return -EOPNOTSUPP;
4272         }
4273
4274         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4275                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4276                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4277                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4278                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4279
4280         if (id_priv->id.ps == RDMA_PS_IPOIB)
4281                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4282                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4283                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4284                              IB_SA_MCMEMBER_REC_MTU |
4285                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4286
4287         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4288                                                 id_priv->id.port_num, &rec,
4289                                                 comp_mask, GFP_KERNEL,
4290                                                 cma_ib_mc_handler, mc);
4291         return PTR_ERR_OR_ZERO(mc->multicast.ib);
4292 }
4293
4294 static void iboe_mcast_work_handler(struct work_struct *work)
4295 {
4296         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4297         struct cma_multicast *mc = mw->mc;
4298         struct ib_sa_multicast *m = mc->multicast.ib;
4299
4300         mc->multicast.ib->context = mc;
4301         cma_ib_mc_handler(0, m);
4302         kref_put(&mc->mcref, release_mc);
4303         kfree(mw);
4304 }
4305
4306 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4307                               enum ib_gid_type gid_type)
4308 {
4309         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4310         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4311
4312         if (cma_any_addr(addr)) {
4313                 memset(mgid, 0, sizeof *mgid);
4314         } else if (addr->sa_family == AF_INET6) {
4315                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4316         } else {
4317                 mgid->raw[0] =
4318                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4319                 mgid->raw[1] =
4320                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4321                 mgid->raw[2] = 0;
4322                 mgid->raw[3] = 0;
4323                 mgid->raw[4] = 0;
4324                 mgid->raw[5] = 0;
4325                 mgid->raw[6] = 0;
4326                 mgid->raw[7] = 0;
4327                 mgid->raw[8] = 0;
4328                 mgid->raw[9] = 0;
4329                 mgid->raw[10] = 0xff;
4330                 mgid->raw[11] = 0xff;
4331                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4332         }
4333 }
4334
4335 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4336                                    struct cma_multicast *mc)
4337 {
4338         struct iboe_mcast_work *work;
4339         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4340         int err = 0;
4341         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4342         struct net_device *ndev = NULL;
4343         enum ib_gid_type gid_type;
4344         bool send_only;
4345
4346         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4347
4348         if (cma_zero_addr((struct sockaddr *)&mc->addr))
4349                 return -EINVAL;
4350
4351         work = kzalloc(sizeof *work, GFP_KERNEL);
4352         if (!work)
4353                 return -ENOMEM;
4354
4355         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4356         if (!mc->multicast.ib) {
4357                 err = -ENOMEM;
4358                 goto out1;
4359         }
4360
4361         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4362                    rdma_start_port(id_priv->cma_dev->device)];
4363         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4364
4365         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4366         if (id_priv->id.ps == RDMA_PS_UDP)
4367                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4368
4369         if (dev_addr->bound_dev_if)
4370                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4371         if (!ndev) {
4372                 err = -ENODEV;
4373                 goto out2;
4374         }
4375         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4376         mc->multicast.ib->rec.hop_limit = 1;
4377         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4378
4379         if (addr->sa_family == AF_INET) {
4380                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4381                         mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4382                         if (!send_only) {
4383                                 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4384                                                     true);
4385                         }
4386                 }
4387         } else {
4388                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4389                         err = -ENOTSUPP;
4390         }
4391         dev_put(ndev);
4392         if (err || !mc->multicast.ib->rec.mtu) {
4393                 if (!err)
4394                         err = -EINVAL;
4395                 goto out2;
4396         }
4397         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4398                     &mc->multicast.ib->rec.port_gid);
4399         work->id = id_priv;
4400         work->mc = mc;
4401         INIT_WORK(&work->work, iboe_mcast_work_handler);
4402         kref_get(&mc->mcref);
4403         queue_work(cma_wq, &work->work);
4404
4405         return 0;
4406
4407 out2:
4408         kfree(mc->multicast.ib);
4409 out1:
4410         kfree(work);
4411         return err;
4412 }
4413
4414 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4415                         u8 join_state, void *context)
4416 {
4417         struct rdma_id_private *id_priv;
4418         struct cma_multicast *mc;
4419         int ret;
4420
4421         if (!id->device)
4422                 return -EINVAL;
4423
4424         id_priv = container_of(id, struct rdma_id_private, id);
4425         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4426             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4427                 return -EINVAL;
4428
4429         mc = kmalloc(sizeof *mc, GFP_KERNEL);
4430         if (!mc)
4431                 return -ENOMEM;
4432
4433         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4434         mc->context = context;
4435         mc->id_priv = id_priv;
4436         mc->join_state = join_state;
4437
4438         if (rdma_protocol_roce(id->device, id->port_num)) {
4439                 kref_init(&mc->mcref);
4440                 ret = cma_iboe_join_multicast(id_priv, mc);
4441                 if (ret)
4442                         goto out_err;
4443         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4444                 ret = cma_join_ib_multicast(id_priv, mc);
4445                 if (ret)
4446                         goto out_err;
4447         } else {
4448                 ret = -ENOSYS;
4449                 goto out_err;
4450         }
4451
4452         spin_lock(&id_priv->lock);
4453         list_add(&mc->list, &id_priv->mc_list);
4454         spin_unlock(&id_priv->lock);
4455
4456         return 0;
4457 out_err:
4458         kfree(mc);
4459         return ret;
4460 }
4461 EXPORT_SYMBOL(rdma_join_multicast);
4462
4463 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4464 {
4465         struct rdma_id_private *id_priv;
4466         struct cma_multicast *mc;
4467
4468         id_priv = container_of(id, struct rdma_id_private, id);
4469         spin_lock_irq(&id_priv->lock);
4470         list_for_each_entry(mc, &id_priv->mc_list, list) {
4471                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4472                         list_del(&mc->list);
4473                         spin_unlock_irq(&id_priv->lock);
4474
4475                         if (id->qp)
4476                                 ib_detach_mcast(id->qp,
4477                                                 &mc->multicast.ib->rec.mgid,
4478                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
4479
4480                         BUG_ON(id_priv->cma_dev->device != id->device);
4481
4482                         if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4483                                 ib_sa_free_multicast(mc->multicast.ib);
4484                                 kfree(mc);
4485                         } else if (rdma_protocol_roce(id->device, id->port_num)) {
4486                                 cma_leave_roce_mc_group(id_priv, mc);
4487                         }
4488                         return;
4489                 }
4490         }
4491         spin_unlock_irq(&id_priv->lock);
4492 }
4493 EXPORT_SYMBOL(rdma_leave_multicast);
4494
4495 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4496 {
4497         struct rdma_dev_addr *dev_addr;
4498         struct cma_ndev_work *work;
4499
4500         dev_addr = &id_priv->id.route.addr.dev_addr;
4501
4502         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4503             (net_eq(dev_net(ndev), dev_addr->net)) &&
4504             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4505                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4506                         ndev->name, &id_priv->id);
4507                 work = kzalloc(sizeof *work, GFP_KERNEL);
4508                 if (!work)
4509                         return -ENOMEM;
4510
4511                 INIT_WORK(&work->work, cma_ndev_work_handler);
4512                 work->id = id_priv;
4513                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4514                 atomic_inc(&id_priv->refcount);
4515                 queue_work(cma_wq, &work->work);
4516         }
4517
4518         return 0;
4519 }
4520
4521 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4522                                void *ptr)
4523 {
4524         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4525         struct cma_device *cma_dev;
4526         struct rdma_id_private *id_priv;
4527         int ret = NOTIFY_DONE;
4528
4529         if (event != NETDEV_BONDING_FAILOVER)
4530                 return NOTIFY_DONE;
4531
4532         if (!netif_is_bond_master(ndev))
4533                 return NOTIFY_DONE;
4534
4535         mutex_lock(&lock);
4536         list_for_each_entry(cma_dev, &dev_list, list)
4537                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4538                         ret = cma_netdev_change(ndev, id_priv);
4539                         if (ret)
4540                                 goto out;
4541                 }
4542
4543 out:
4544         mutex_unlock(&lock);
4545         return ret;
4546 }
4547
4548 static struct notifier_block cma_nb = {
4549         .notifier_call = cma_netdev_callback
4550 };
4551
4552 static void cma_add_one(struct ib_device *device)
4553 {
4554         struct cma_device *cma_dev;
4555         struct rdma_id_private *id_priv;
4556         unsigned int i;
4557         unsigned long supported_gids = 0;
4558
4559         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4560         if (!cma_dev)
4561                 return;
4562
4563         cma_dev->device = device;
4564         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4565                                             sizeof(*cma_dev->default_gid_type),
4566                                             GFP_KERNEL);
4567         if (!cma_dev->default_gid_type)
4568                 goto free_cma_dev;
4569
4570         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4571                                             sizeof(*cma_dev->default_roce_tos),
4572                                             GFP_KERNEL);
4573         if (!cma_dev->default_roce_tos)
4574                 goto free_gid_type;
4575
4576         rdma_for_each_port (device, i) {
4577                 supported_gids = roce_gid_type_mask_support(device, i);
4578                 WARN_ON(!supported_gids);
4579                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4580                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4581                                 CMA_PREFERRED_ROCE_GID_TYPE;
4582                 else
4583                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4584                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4585                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4586         }
4587
4588         init_completion(&cma_dev->comp);
4589         atomic_set(&cma_dev->refcount, 1);
4590         INIT_LIST_HEAD(&cma_dev->id_list);
4591         ib_set_client_data(device, &cma_client, cma_dev);
4592
4593         mutex_lock(&lock);
4594         list_add_tail(&cma_dev->list, &dev_list);
4595         list_for_each_entry(id_priv, &listen_any_list, list)
4596                 cma_listen_on_dev(id_priv, cma_dev);
4597         mutex_unlock(&lock);
4598
4599         return;
4600
4601 free_gid_type:
4602         kfree(cma_dev->default_gid_type);
4603
4604 free_cma_dev:
4605         kfree(cma_dev);
4606
4607         return;
4608 }
4609
4610 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4611 {
4612         struct rdma_cm_event event = {};
4613         enum rdma_cm_state state;
4614         int ret = 0;
4615
4616         /* Record that we want to remove the device */
4617         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4618         if (state == RDMA_CM_DESTROYING)
4619                 return 0;
4620
4621         cma_cancel_operation(id_priv, state);
4622         mutex_lock(&id_priv->handler_mutex);
4623
4624         /* Check for destruction from another callback. */
4625         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4626                 goto out;
4627
4628         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4629         ret = id_priv->id.event_handler(&id_priv->id, &event);
4630 out:
4631         mutex_unlock(&id_priv->handler_mutex);
4632         return ret;
4633 }
4634
4635 static void cma_process_remove(struct cma_device *cma_dev)
4636 {
4637         struct rdma_id_private *id_priv;
4638         int ret;
4639
4640         mutex_lock(&lock);
4641         while (!list_empty(&cma_dev->id_list)) {
4642                 id_priv = list_entry(cma_dev->id_list.next,
4643                                      struct rdma_id_private, list);
4644
4645                 list_del(&id_priv->listen_list);
4646                 list_del_init(&id_priv->list);
4647                 atomic_inc(&id_priv->refcount);
4648                 mutex_unlock(&lock);
4649
4650                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4651                 cma_deref_id(id_priv);
4652                 if (ret)
4653                         rdma_destroy_id(&id_priv->id);
4654
4655                 mutex_lock(&lock);
4656         }
4657         mutex_unlock(&lock);
4658
4659         cma_deref_dev(cma_dev);
4660         wait_for_completion(&cma_dev->comp);
4661 }
4662
4663 static void cma_remove_one(struct ib_device *device, void *client_data)
4664 {
4665         struct cma_device *cma_dev = client_data;
4666
4667         if (!cma_dev)
4668                 return;
4669
4670         mutex_lock(&lock);
4671         list_del(&cma_dev->list);
4672         mutex_unlock(&lock);
4673
4674         cma_process_remove(cma_dev);
4675         kfree(cma_dev->default_roce_tos);
4676         kfree(cma_dev->default_gid_type);
4677         kfree(cma_dev);
4678 }
4679
4680 static int cma_init_net(struct net *net)
4681 {
4682         struct cma_pernet *pernet = cma_pernet(net);
4683
4684         xa_init(&pernet->tcp_ps);
4685         xa_init(&pernet->udp_ps);
4686         xa_init(&pernet->ipoib_ps);
4687         xa_init(&pernet->ib_ps);
4688
4689         return 0;
4690 }
4691
4692 static void cma_exit_net(struct net *net)
4693 {
4694         struct cma_pernet *pernet = cma_pernet(net);
4695
4696         WARN_ON(!xa_empty(&pernet->tcp_ps));
4697         WARN_ON(!xa_empty(&pernet->udp_ps));
4698         WARN_ON(!xa_empty(&pernet->ipoib_ps));
4699         WARN_ON(!xa_empty(&pernet->ib_ps));
4700 }
4701
4702 static struct pernet_operations cma_pernet_operations = {
4703         .init = cma_init_net,
4704         .exit = cma_exit_net,
4705         .id = &cma_pernet_id,
4706         .size = sizeof(struct cma_pernet),
4707 };
4708
4709 static int __init cma_init(void)
4710 {
4711         int ret;
4712
4713         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4714         if (!cma_wq)
4715                 return -ENOMEM;
4716
4717         ret = register_pernet_subsys(&cma_pernet_operations);
4718         if (ret)
4719                 goto err_wq;
4720
4721         ib_sa_register_client(&sa_client);
4722         register_netdevice_notifier(&cma_nb);
4723
4724         ret = ib_register_client(&cma_client);
4725         if (ret)
4726                 goto err;
4727
4728         ret = cma_configfs_init();
4729         if (ret)
4730                 goto err_ib;
4731
4732         return 0;
4733
4734 err_ib:
4735         ib_unregister_client(&cma_client);
4736 err:
4737         unregister_netdevice_notifier(&cma_nb);
4738         ib_sa_unregister_client(&sa_client);
4739 err_wq:
4740         destroy_workqueue(cma_wq);
4741         return ret;
4742 }
4743
4744 static void __exit cma_cleanup(void)
4745 {
4746         cma_configfs_exit();
4747         ib_unregister_client(&cma_client);
4748         unregister_netdevice_notifier(&cma_nb);
4749         ib_sa_unregister_client(&sa_client);
4750         unregister_pernet_subsys(&cma_pernet_operations);
4751         destroy_workqueue(cma_wq);
4752 }
4753
4754 module_init(cma_init);
4755 module_exit(cma_cleanup);