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