drivers/infiniband/core/cma.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   4  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   5  * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
   6  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   7  */
   8
   9 #include <linux/completion.h>
  10 #include <linux/in.h>
  11 #include <linux/in6.h>
  12 #include <linux/mutex.h>
  13 #include <linux/random.h>
  14 #include <linux/igmp.h>
  15 #include <linux/xarray.h>
  16 #include <linux/inetdevice.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <net/route.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/netns/generic.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/ip_fib.h>
  26 #include <net/ip6_route.h>
  27
  28 #include <rdma/rdma_cm.h>
  29 #include <rdma/rdma_cm_ib.h>
  30 #include <rdma/rdma_netlink.h>
  31 #include <rdma/ib.h>
  32 #include <rdma/ib_cache.h>
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_sa.h>
  35 #include <rdma/iw_cm.h>
  36
  37 #include "core_priv.h"
  38 #include "cma_priv.h"
  39 #include "cma_trace.h"
  40
  41 MODULE_AUTHOR("Sean Hefty");
  42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  43 MODULE_LICENSE("Dual BSD/GPL");
  44
  45 #define CMA_CM_RESPONSE_TIMEOUT 20
  46 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  47 #define CMA_MAX_CM_RETRIES 15
  48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  49 #define CMA_IBOE_PACKET_LIFETIME 18
  50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  51
  52 static const char * const cma_events[] = {
  53         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  54         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  55         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  56         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  57         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  58         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  59         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  60         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  61         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  62         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  63         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  64         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  65         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  66         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  67         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  68         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  69 };
  70
  71 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  72 {
  73         size_t index = event;
  74
  75         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  76                         cma_events[index] : "unrecognized event";
  77 }
  78 EXPORT_SYMBOL(rdma_event_msg);
  79
  80 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  81                                                 int reason)
  82 {
  83         if (rdma_ib_or_roce(id->device, id->port_num))
  84                 return ibcm_reject_msg(reason);
  85
  86         if (rdma_protocol_iwarp(id->device, id->port_num))
  87                 return iwcm_reject_msg(reason);
  88
  89         WARN_ON_ONCE(1);
  90         return "unrecognized transport";
  91 }
  92 EXPORT_SYMBOL(rdma_reject_msg);
  93
  94 /**
  95  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  96  *                           request.
  97  * @id: Communication identifier that received the REJECT event.
  98  * @reason: Value returned in the REJECT event status field.
  99  */
 100 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 101 {
 102         if (rdma_ib_or_roce(id->device, id->port_num))
 103                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 104
 105         if (rdma_protocol_iwarp(id->device, id->port_num))
 106                 return reason == -ECONNREFUSED;
 107
 108         WARN_ON_ONCE(1);
 109         return false;
 110 }
 111
 112 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 113                                       struct rdma_cm_event *ev, u8 *data_len)
 114 {
 115         const void *p;
 116
 117         if (rdma_is_consumer_reject(id, ev->status)) {
 118                 *data_len = ev->param.conn.private_data_len;
 119                 p = ev->param.conn.private_data;
 120         } else {
 121                 *data_len = 0;
 122                 p = NULL;
 123         }
 124         return p;
 125 }
 126 EXPORT_SYMBOL(rdma_consumer_reject_data);
 127
 128 /**
 129  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 130  * @id: Communication Identifier
 131  */
 132 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 133 {
 134         struct rdma_id_private *id_priv;
 135
 136         id_priv = container_of(id, struct rdma_id_private, id);
 137         if (id->device->node_type == RDMA_NODE_RNIC)
 138                 return id_priv->cm_id.iw;
 139         return NULL;
 140 }
 141 EXPORT_SYMBOL(rdma_iw_cm_id);
 142
 143 /**
 144  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 145  * @res: rdma resource tracking entry pointer
 146  */
 147 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 148 {
 149         struct rdma_id_private *id_priv =
 150                 container_of(res, struct rdma_id_private, res);
 151
 152         return &id_priv->id;
 153 }
 154 EXPORT_SYMBOL(rdma_res_to_id);
 155
 156 static int cma_add_one(struct ib_device *device);
 157 static void cma_remove_one(struct ib_device *device, void *client_data);
 158
 159 static struct ib_client cma_client = {
 160         .name   = "cma",
 161         .add    = cma_add_one,
 162         .remove = cma_remove_one
 163 };
 164
 165 static struct ib_sa_client sa_client;
 166 static LIST_HEAD(dev_list);
 167 static LIST_HEAD(listen_any_list);
 168 static DEFINE_MUTEX(lock);
 169 static struct workqueue_struct *cma_wq;
 170 static unsigned int cma_pernet_id;
 171
 172 struct cma_pernet {
 173         struct xarray tcp_ps;
 174         struct xarray udp_ps;
 175         struct xarray ipoib_ps;
 176         struct xarray ib_ps;
 177 };
 178
 179 static struct cma_pernet *cma_pernet(struct net *net)
 180 {
 181         return net_generic(net, cma_pernet_id);
 182 }
 183
 184 static
 185 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 186 {
 187         struct cma_pernet *pernet = cma_pernet(net);
 188
 189         switch (ps) {
 190         case RDMA_PS_TCP:
 191                 return &pernet->tcp_ps;
 192         case RDMA_PS_UDP:
 193                 return &pernet->udp_ps;
 194         case RDMA_PS_IPOIB:
 195                 return &pernet->ipoib_ps;
 196         case RDMA_PS_IB:
 197                 return &pernet->ib_ps;
 198         default:
 199                 return NULL;
 200         }
 201 }
 202
 203 struct cma_device {
 204         struct list_head        list;
 205         struct ib_device        *device;
 206         struct completion       comp;
 207         refcount_t refcount;
 208         struct list_head        id_list;
 209         enum ib_gid_type        *default_gid_type;
 210         u8                      *default_roce_tos;
 211 };
 212
 213 struct rdma_bind_list {
 214         enum rdma_ucm_port_space ps;
 215         struct hlist_head       owners;
 216         unsigned short          port;
 217 };
 218
 219 struct class_port_info_context {
 220         struct ib_class_port_info       *class_port_info;
 221         struct ib_device                *device;
 222         struct completion               done;
 223         struct ib_sa_query              *sa_query;
 224         u8                              port_num;
 225 };
 226
 227 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 228                         struct rdma_bind_list *bind_list, int snum)
 229 {
 230         struct xarray *xa = cma_pernet_xa(net, ps);
 231
 232         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 233 }
 234
 235 static struct rdma_bind_list *cma_ps_find(struct net *net,
 236                                           enum rdma_ucm_port_space ps, int snum)
 237 {
 238         struct xarray *xa = cma_pernet_xa(net, ps);
 239
 240         return xa_load(xa, snum);
 241 }
 242
 243 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 244                           int snum)
 245 {
 246         struct xarray *xa = cma_pernet_xa(net, ps);
 247
 248         xa_erase(xa, snum);
 249 }
 250
 251 enum {
 252         CMA_OPTION_AFONLY,
 253 };
 254
 255 void cma_dev_get(struct cma_device *cma_dev)
 256 {
 257         refcount_inc(&cma_dev->refcount);
 258 }
 259
 260 void cma_dev_put(struct cma_device *cma_dev)
 261 {
 262         if (refcount_dec_and_test(&cma_dev->refcount))
 263                 complete(&cma_dev->comp);
 264 }
 265
 266 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 267                                              void               *cookie)
 268 {
 269         struct cma_device *cma_dev;
 270         struct cma_device *found_cma_dev = NULL;
 271
 272         mutex_lock(&lock);
 273
 274         list_for_each_entry(cma_dev, &dev_list, list)
 275                 if (filter(cma_dev->device, cookie)) {
 276                         found_cma_dev = cma_dev;
 277                         break;
 278                 }
 279
 280         if (found_cma_dev)
 281                 cma_dev_get(found_cma_dev);
 282         mutex_unlock(&lock);
 283         return found_cma_dev;
 284 }
 285
 286 int cma_get_default_gid_type(struct cma_device *cma_dev,
 287                              unsigned int port)
 288 {
 289         if (!rdma_is_port_valid(cma_dev->device, port))
 290                 return -EINVAL;
 291
 292         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 293 }
 294
 295 int cma_set_default_gid_type(struct cma_device *cma_dev,
 296                              unsigned int port,
 297                              enum ib_gid_type default_gid_type)
 298 {
 299         unsigned long supported_gids;
 300
 301         if (!rdma_is_port_valid(cma_dev->device, port))
 302                 return -EINVAL;
 303
 304         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 305
 306         if (!(supported_gids & 1 << default_gid_type))
 307                 return -EINVAL;
 308
 309         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 310                 default_gid_type;
 311
 312         return 0;
 313 }
 314
 315 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 316 {
 317         if (!rdma_is_port_valid(cma_dev->device, port))
 318                 return -EINVAL;
 319
 320         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 321 }
 322
 323 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 324                              u8 default_roce_tos)
 325 {
 326         if (!rdma_is_port_valid(cma_dev->device, port))
 327                 return -EINVAL;
 328
 329         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 330                  default_roce_tos;
 331
 332         return 0;
 333 }
 334 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 335 {
 336         return cma_dev->device;
 337 }
 338
 339 /*
 340  * Device removal can occur at anytime, so we need extra handling to
 341  * serialize notifying the user of device removal with other callbacks.
 342  * We do this by disabling removal notification while a callback is in process,
 343  * and reporting it after the callback completes.
 344  */
 345
 346 struct cma_multicast {
 347         struct rdma_id_private *id_priv;
 348         union {
 349                 struct ib_sa_multicast *ib;
 350         } multicast;
 351         struct list_head        list;
 352         void                    *context;
 353         struct sockaddr_storage addr;
 354         struct kref             mcref;
 355         u8                      join_state;
 356 };
 357
 358 struct cma_work {
 359         struct work_struct      work;
 360         struct rdma_id_private  *id;
 361         enum rdma_cm_state      old_state;
 362         enum rdma_cm_state      new_state;
 363         struct rdma_cm_event    event;
 364 };
 365
 366 struct cma_ndev_work {
 367         struct work_struct      work;
 368         struct rdma_id_private  *id;
 369         struct rdma_cm_event    event;
 370 };
 371
 372 struct iboe_mcast_work {
 373         struct work_struct       work;
 374         struct rdma_id_private  *id;
 375         struct cma_multicast    *mc;
 376 };
 377
 378 union cma_ip_addr {
 379         struct in6_addr ip6;
 380         struct {
 381                 __be32 pad[3];
 382                 __be32 addr;
 383         } ip4;
 384 };
 385
 386 struct cma_hdr {
 387         u8 cma_version;
 388         u8 ip_version;  /* IP version: 7:4 */
 389         __be16 port;
 390         union cma_ip_addr src_addr;
 391         union cma_ip_addr dst_addr;
 392 };
 393
 394 #define CMA_VERSION 0x00
 395
 396 struct cma_req_info {
 397         struct sockaddr_storage listen_addr_storage;
 398         struct sockaddr_storage src_addr_storage;
 399         struct ib_device *device;
 400         union ib_gid local_gid;
 401         __be64 service_id;
 402         int port;
 403         bool has_gid;
 404         u16 pkey;
 405 };
 406
 407 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 408 {
 409         unsigned long flags;
 410         int ret;
 411
 412         spin_lock_irqsave(&id_priv->lock, flags);
 413         ret = (id_priv->state == comp);
 414         spin_unlock_irqrestore(&id_priv->lock, flags);
 415         return ret;
 416 }
 417
 418 static int cma_comp_exch(struct rdma_id_private *id_priv,
 419                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 420 {
 421         unsigned long flags;
 422         int ret;
 423
 424         spin_lock_irqsave(&id_priv->lock, flags);
 425         if ((ret = (id_priv->state == comp)))
 426                 id_priv->state = exch;
 427         spin_unlock_irqrestore(&id_priv->lock, flags);
 428         return ret;
 429 }
 430
 431 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 432                                    enum rdma_cm_state exch)
 433 {
 434         unsigned long flags;
 435         enum rdma_cm_state old;
 436
 437         spin_lock_irqsave(&id_priv->lock, flags);
 438         old = id_priv->state;
 439         id_priv->state = exch;
 440         spin_unlock_irqrestore(&id_priv->lock, flags);
 441         return old;
 442 }
 443
 444 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 445 {
 446         return hdr->ip_version >> 4;
 447 }
 448
 449 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 450 {
 451         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 452 }
 453
 454 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 455 {
 456         struct in_device *in_dev = NULL;
 457
 458         if (ndev) {
 459                 rtnl_lock();
 460                 in_dev = __in_dev_get_rtnl(ndev);
 461                 if (in_dev) {
 462                         if (join)
 463                                 ip_mc_inc_group(in_dev,
 464                                                 *(__be32 *)(mgid->raw + 12));
 465                         else
 466                                 ip_mc_dec_group(in_dev,
 467                                                 *(__be32 *)(mgid->raw + 12));
 468                 }
 469                 rtnl_unlock();
 470         }
 471         return (in_dev) ? 0 : -ENODEV;
 472 }
 473
 474 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 475                                struct cma_device *cma_dev)
 476 {
 477         cma_dev_get(cma_dev);
 478         id_priv->cma_dev = cma_dev;
 479         id_priv->id.device = cma_dev->device;
 480         id_priv->id.route.addr.dev_addr.transport =
 481                 rdma_node_get_transport(cma_dev->device->node_type);
 482         list_add_tail(&id_priv->list, &cma_dev->id_list);
 483         if (id_priv->res.kern_name)
 484                 rdma_restrack_kadd(&id_priv->res);
 485         else
 486                 rdma_restrack_uadd(&id_priv->res);
 487 }
 488
 489 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 490                               struct cma_device *cma_dev)
 491 {
 492         _cma_attach_to_dev(id_priv, cma_dev);
 493         id_priv->gid_type =
 494                 cma_dev->default_gid_type[id_priv->id.port_num -
 495                                           rdma_start_port(cma_dev->device)];
 496 }
 497
 498 static inline void release_mc(struct kref *kref)
 499 {
 500         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 501
 502         kfree(mc->multicast.ib);
 503         kfree(mc);
 504 }
 505
 506 static void cma_release_dev(struct rdma_id_private *id_priv)
 507 {
 508         mutex_lock(&lock);
 509         list_del(&id_priv->list);
 510         cma_dev_put(id_priv->cma_dev);
 511         id_priv->cma_dev = NULL;
 512         mutex_unlock(&lock);
 513 }
 514
 515 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 516 {
 517         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 518 }
 519
 520 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 521 {
 522         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 523 }
 524
 525 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 526 {
 527         return id_priv->id.route.addr.src_addr.ss_family;
 528 }
 529
 530 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 531 {
 532         struct ib_sa_mcmember_rec rec;
 533         int ret = 0;
 534
 535         if (id_priv->qkey) {
 536                 if (qkey && id_priv->qkey != qkey)
 537                         return -EINVAL;
 538                 return 0;
 539         }
 540
 541         if (qkey) {
 542                 id_priv->qkey = qkey;
 543                 return 0;
 544         }
 545
 546         switch (id_priv->id.ps) {
 547         case RDMA_PS_UDP:
 548         case RDMA_PS_IB:
 549                 id_priv->qkey = RDMA_UDP_QKEY;
 550                 break;
 551         case RDMA_PS_IPOIB:
 552                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 553                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 554                                              id_priv->id.port_num, &rec.mgid,
 555                                              &rec);
 556                 if (!ret)
 557                         id_priv->qkey = be32_to_cpu(rec.qkey);
 558                 break;
 559         default:
 560                 break;
 561         }
 562         return ret;
 563 }
 564
 565 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 566 {
 567         dev_addr->dev_type = ARPHRD_INFINIBAND;
 568         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 569         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 570 }
 571
 572 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 573 {
 574         int ret;
 575
 576         if (addr->sa_family != AF_IB) {
 577                 ret = rdma_translate_ip(addr, dev_addr);
 578         } else {
 579                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 580                 ret = 0;
 581         }
 582
 583         return ret;
 584 }
 585
 586 static const struct ib_gid_attr *
 587 cma_validate_port(struct ib_device *device, u8 port,
 588                   enum ib_gid_type gid_type,
 589                   union ib_gid *gid,
 590                   struct rdma_id_private *id_priv)
 591 {
 592         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 593         int bound_if_index = dev_addr->bound_dev_if;
 594         const struct ib_gid_attr *sgid_attr;
 595         int dev_type = dev_addr->dev_type;
 596         struct net_device *ndev = NULL;
 597
 598         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 599                 return ERR_PTR(-ENODEV);
 600
 601         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 602                 return ERR_PTR(-ENODEV);
 603
 604         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 605                 return ERR_PTR(-ENODEV);
 606
 607         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 608                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 609                 if (!ndev)
 610                         return ERR_PTR(-ENODEV);
 611         } else {
 612                 gid_type = IB_GID_TYPE_IB;
 613         }
 614
 615         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 616         if (ndev)
 617                 dev_put(ndev);
 618         return sgid_attr;
 619 }
 620
 621 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 622                                const struct ib_gid_attr *sgid_attr)
 623 {
 624         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 625         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 626 }
 627
 628 /**
 629  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 630  * based on source ip address.
 631  * @id_priv:    cm_id which should be bound to cma device
 632  *
 633  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 634  * based on source IP address. It returns 0 on success or error code otherwise.
 635  * It is applicable to active and passive side cm_id.
 636  */
 637 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 638 {
 639         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 640         const struct ib_gid_attr *sgid_attr;
 641         union ib_gid gid, iboe_gid, *gidp;
 642         struct cma_device *cma_dev;
 643         enum ib_gid_type gid_type;
 644         int ret = -ENODEV;
 645         unsigned int port;
 646
 647         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 648             id_priv->id.ps == RDMA_PS_IPOIB)
 649                 return -EINVAL;
 650
 651         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 652                     &iboe_gid);
 653
 654         memcpy(&gid, dev_addr->src_dev_addr +
 655                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 656
 657         mutex_lock(&lock);
 658         list_for_each_entry(cma_dev, &dev_list, list) {
 659                 rdma_for_each_port (cma_dev->device, port) {
 660                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 661                                &iboe_gid : &gid;
 662                         gid_type = cma_dev->default_gid_type[port - 1];
 663                         sgid_attr = cma_validate_port(cma_dev->device, port,
 664                                                       gid_type, gidp, id_priv);
 665                         if (!IS_ERR(sgid_attr)) {
 666                                 id_priv->id.port_num = port;
 667                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 668                                 cma_attach_to_dev(id_priv, cma_dev);
 669                                 ret = 0;
 670                                 goto out;
 671                         }
 672                 }
 673         }
 674 out:
 675         mutex_unlock(&lock);
 676         return ret;
 677 }
 678
 679 /**
 680  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 681  * @id_priv:            cm id to bind to cma device
 682  * @listen_id_priv:     listener cm id to match against
 683  * @req:                Pointer to req structure containaining incoming
 684  *                      request information
 685  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 686  * rdma device matches for listen_id and incoming request. It also verifies
 687  * that a GID table entry is present for the source address.
 688  * Returns 0 on success, or returns error code otherwise.
 689  */
 690 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 691                               const struct rdma_id_private *listen_id_priv,
 692                               struct cma_req_info *req)
 693 {
 694         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 695         const struct ib_gid_attr *sgid_attr;
 696         enum ib_gid_type gid_type;
 697         union ib_gid gid;
 698
 699         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 700             id_priv->id.ps == RDMA_PS_IPOIB)
 701                 return -EINVAL;
 702
 703         if (rdma_protocol_roce(req->device, req->port))
 704                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 705                             &gid);
 706         else
 707                 memcpy(&gid, dev_addr->src_dev_addr +
 708                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 709
 710         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 711         sgid_attr = cma_validate_port(req->device, req->port,
 712                                       gid_type, &gid, id_priv);
 713         if (IS_ERR(sgid_attr))
 714                 return PTR_ERR(sgid_attr);
 715
 716         id_priv->id.port_num = req->port;
 717         cma_bind_sgid_attr(id_priv, sgid_attr);
 718         /* Need to acquire lock to protect against reader
 719          * of cma_dev->id_list such as cma_netdev_callback() and
 720          * cma_process_remove().
 721          */
 722         mutex_lock(&lock);
 723         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 724         mutex_unlock(&lock);
 725         return 0;
 726 }
 727
 728 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 729                               const struct rdma_id_private *listen_id_priv)
 730 {
 731         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 732         const struct ib_gid_attr *sgid_attr;
 733         struct cma_device *cma_dev;
 734         enum ib_gid_type gid_type;
 735         int ret = -ENODEV;
 736         unsigned int port;
 737         union ib_gid gid;
 738
 739         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 740             id_priv->id.ps == RDMA_PS_IPOIB)
 741                 return -EINVAL;
 742
 743         memcpy(&gid, dev_addr->src_dev_addr +
 744                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 745
 746         mutex_lock(&lock);
 747
 748         cma_dev = listen_id_priv->cma_dev;
 749         port = listen_id_priv->id.port_num;
 750         gid_type = listen_id_priv->gid_type;
 751         sgid_attr = cma_validate_port(cma_dev->device, port,
 752                                       gid_type, &gid, id_priv);
 753         if (!IS_ERR(sgid_attr)) {
 754                 id_priv->id.port_num = port;
 755                 cma_bind_sgid_attr(id_priv, sgid_attr);
 756                 ret = 0;
 757                 goto out;
 758         }
 759
 760         list_for_each_entry(cma_dev, &dev_list, list) {
 761                 rdma_for_each_port (cma_dev->device, port) {
 762                         if (listen_id_priv->cma_dev == cma_dev &&
 763                             listen_id_priv->id.port_num == port)
 764                                 continue;
 765
 766                         gid_type = cma_dev->default_gid_type[port - 1];
 767                         sgid_attr = cma_validate_port(cma_dev->device, port,
 768                                                       gid_type, &gid, id_priv);
 769                         if (!IS_ERR(sgid_attr)) {
 770                                 id_priv->id.port_num = port;
 771                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 772                                 ret = 0;
 773                                 goto out;
 774                         }
 775                 }
 776         }
 777
 778 out:
 779         if (!ret)
 780                 cma_attach_to_dev(id_priv, cma_dev);
 781
 782         mutex_unlock(&lock);
 783         return ret;
 784 }
 785
 786 /*
 787  * Select the source IB device and address to reach the destination IB address.
 788  */
 789 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 790 {
 791         struct cma_device *cma_dev, *cur_dev;
 792         struct sockaddr_ib *addr;
 793         union ib_gid gid, sgid, *dgid;
 794         unsigned int p;
 795         u16 pkey, index;
 796         enum ib_port_state port_state;
 797         int i;
 798
 799         cma_dev = NULL;
 800         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 801         dgid = (union ib_gid *) &addr->sib_addr;
 802         pkey = ntohs(addr->sib_pkey);
 803
 804         mutex_lock(&lock);
 805         list_for_each_entry(cur_dev, &dev_list, list) {
 806                 rdma_for_each_port (cur_dev->device, p) {
 807                         if (!rdma_cap_af_ib(cur_dev->device, p))
 808                                 continue;
 809
 810                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 811                                 continue;
 812
 813                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 814                                 continue;
 815                         for (i = 0; !rdma_query_gid(cur_dev->device,
 816                                                     p, i, &gid);
 817                              i++) {
 818                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 819                                         cma_dev = cur_dev;
 820                                         sgid = gid;
 821                                         id_priv->id.port_num = p;
 822                                         goto found;
 823                                 }
 824
 825                                 if (!cma_dev && (gid.global.subnet_prefix ==
 826                                     dgid->global.subnet_prefix) &&
 827                                     port_state == IB_PORT_ACTIVE) {
 828                                         cma_dev = cur_dev;
 829                                         sgid = gid;
 830                                         id_priv->id.port_num = p;
 831                                         goto found;
 832                                 }
 833                         }
 834                 }
 835         }
 836         mutex_unlock(&lock);
 837         return -ENODEV;
 838
 839 found:
 840         cma_attach_to_dev(id_priv, cma_dev);
 841         mutex_unlock(&lock);
 842         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 843         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 844         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 845         return 0;
 846 }
 847
 848 static void cma_id_get(struct rdma_id_private *id_priv)
 849 {
 850         refcount_inc(&id_priv->refcount);
 851 }
 852
 853 static void cma_id_put(struct rdma_id_private *id_priv)
 854 {
 855         if (refcount_dec_and_test(&id_priv->refcount))
 856                 complete(&id_priv->comp);
 857 }
 858
 859 struct rdma_cm_id *__rdma_create_id(struct net *net,
 860                                     rdma_cm_event_handler event_handler,
 861                                     void *context, enum rdma_ucm_port_space ps,
 862                                     enum ib_qp_type qp_type, const char *caller)
 863 {
 864         struct rdma_id_private *id_priv;
 865
 866         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 867         if (!id_priv)
 868                 return ERR_PTR(-ENOMEM);
 869
 870         rdma_restrack_set_task(&id_priv->res, caller);
 871         id_priv->res.type = RDMA_RESTRACK_CM_ID;
 872         id_priv->state = RDMA_CM_IDLE;
 873         id_priv->id.context = context;
 874         id_priv->id.event_handler = event_handler;
 875         id_priv->id.ps = ps;
 876         id_priv->id.qp_type = qp_type;
 877         id_priv->tos_set = false;
 878         id_priv->timeout_set = false;
 879         id_priv->gid_type = IB_GID_TYPE_IB;
 880         spin_lock_init(&id_priv->lock);
 881         mutex_init(&id_priv->qp_mutex);
 882         init_completion(&id_priv->comp);
 883         refcount_set(&id_priv->refcount, 1);
 884         mutex_init(&id_priv->handler_mutex);
 885         INIT_LIST_HEAD(&id_priv->listen_list);
 886         INIT_LIST_HEAD(&id_priv->mc_list);
 887         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 888         id_priv->id.route.addr.dev_addr.net = get_net(net);
 889         id_priv->seq_num &= 0x00ffffff;
 890
 891         trace_cm_id_create(id_priv);
 892         return &id_priv->id;
 893 }
 894 EXPORT_SYMBOL(__rdma_create_id);
 895
 896 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 897 {
 898         struct ib_qp_attr qp_attr;
 899         int qp_attr_mask, ret;
 900
 901         qp_attr.qp_state = IB_QPS_INIT;
 902         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 903         if (ret)
 904                 return ret;
 905
 906         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 907         if (ret)
 908                 return ret;
 909
 910         qp_attr.qp_state = IB_QPS_RTR;
 911         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 912         if (ret)
 913                 return ret;
 914
 915         qp_attr.qp_state = IB_QPS_RTS;
 916         qp_attr.sq_psn = 0;
 917         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 918
 919         return ret;
 920 }
 921
 922 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 923 {
 924         struct ib_qp_attr qp_attr;
 925         int qp_attr_mask, ret;
 926
 927         qp_attr.qp_state = IB_QPS_INIT;
 928         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 929         if (ret)
 930                 return ret;
 931
 932         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 933 }
 934
 935 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 936                    struct ib_qp_init_attr *qp_init_attr)
 937 {
 938         struct rdma_id_private *id_priv;
 939         struct ib_qp *qp;
 940         int ret;
 941
 942         id_priv = container_of(id, struct rdma_id_private, id);
 943         if (id->device != pd->device) {
 944                 ret = -EINVAL;
 945                 goto out_err;
 946         }
 947
 948         qp_init_attr->port_num = id->port_num;
 949         qp = ib_create_qp(pd, qp_init_attr);
 950         if (IS_ERR(qp)) {
 951                 ret = PTR_ERR(qp);
 952                 goto out_err;
 953         }
 954
 955         if (id->qp_type == IB_QPT_UD)
 956                 ret = cma_init_ud_qp(id_priv, qp);
 957         else
 958                 ret = cma_init_conn_qp(id_priv, qp);
 959         if (ret)
 960                 goto out_destroy;
 961
 962         id->qp = qp;
 963         id_priv->qp_num = qp->qp_num;
 964         id_priv->srq = (qp->srq != NULL);
 965         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 966         return 0;
 967 out_destroy:
 968         ib_destroy_qp(qp);
 969 out_err:
 970         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 971         return ret;
 972 }
 973 EXPORT_SYMBOL(rdma_create_qp);
 974
 975 void rdma_destroy_qp(struct rdma_cm_id *id)
 976 {
 977         struct rdma_id_private *id_priv;
 978
 979         id_priv = container_of(id, struct rdma_id_private, id);
 980         trace_cm_qp_destroy(id_priv);
 981         mutex_lock(&id_priv->qp_mutex);
 982         ib_destroy_qp(id_priv->id.qp);
 983         id_priv->id.qp = NULL;
 984         mutex_unlock(&id_priv->qp_mutex);
 985 }
 986 EXPORT_SYMBOL(rdma_destroy_qp);
 987
 988 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 989                              struct rdma_conn_param *conn_param)
 990 {
 991         struct ib_qp_attr qp_attr;
 992         int qp_attr_mask, ret;
 993
 994         mutex_lock(&id_priv->qp_mutex);
 995         if (!id_priv->id.qp) {
 996                 ret = 0;
 997                 goto out;
 998         }
 999
1000         /* Need to update QP attributes from default values. */
1001         qp_attr.qp_state = IB_QPS_INIT;
1002         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1003         if (ret)
1004                 goto out;
1005
1006         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1007         if (ret)
1008                 goto out;
1009
1010         qp_attr.qp_state = IB_QPS_RTR;
1011         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1012         if (ret)
1013                 goto out;
1014
1015         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1016
1017         if (conn_param)
1018                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1019         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1020 out:
1021         mutex_unlock(&id_priv->qp_mutex);
1022         return ret;
1023 }
1024
1025 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1026                              struct rdma_conn_param *conn_param)
1027 {
1028         struct ib_qp_attr qp_attr;
1029         int qp_attr_mask, ret;
1030
1031         mutex_lock(&id_priv->qp_mutex);
1032         if (!id_priv->id.qp) {
1033                 ret = 0;
1034                 goto out;
1035         }
1036
1037         qp_attr.qp_state = IB_QPS_RTS;
1038         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1039         if (ret)
1040                 goto out;
1041
1042         if (conn_param)
1043                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1044         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1045 out:
1046         mutex_unlock(&id_priv->qp_mutex);
1047         return ret;
1048 }
1049
1050 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1051 {
1052         struct ib_qp_attr qp_attr;
1053         int ret;
1054
1055         mutex_lock(&id_priv->qp_mutex);
1056         if (!id_priv->id.qp) {
1057                 ret = 0;
1058                 goto out;
1059         }
1060
1061         qp_attr.qp_state = IB_QPS_ERR;
1062         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1063 out:
1064         mutex_unlock(&id_priv->qp_mutex);
1065         return ret;
1066 }
1067
1068 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1069                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1070 {
1071         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1072         int ret;
1073         u16 pkey;
1074
1075         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1076                 pkey = 0xffff;
1077         else
1078                 pkey = ib_addr_get_pkey(dev_addr);
1079
1080         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1081                                   pkey, &qp_attr->pkey_index);
1082         if (ret)
1083                 return ret;
1084
1085         qp_attr->port_num = id_priv->id.port_num;
1086         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1087
1088         if (id_priv->id.qp_type == IB_QPT_UD) {
1089                 ret = cma_set_qkey(id_priv, 0);
1090                 if (ret)
1091                         return ret;
1092
1093                 qp_attr->qkey = id_priv->qkey;
1094                 *qp_attr_mask |= IB_QP_QKEY;
1095         } else {
1096                 qp_attr->qp_access_flags = 0;
1097                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1098         }
1099         return 0;
1100 }
1101
1102 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1103                        int *qp_attr_mask)
1104 {
1105         struct rdma_id_private *id_priv;
1106         int ret = 0;
1107
1108         id_priv = container_of(id, struct rdma_id_private, id);
1109         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1110                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1111                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1112                 else
1113                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1114                                                  qp_attr_mask);
1115
1116                 if (qp_attr->qp_state == IB_QPS_RTR)
1117                         qp_attr->rq_psn = id_priv->seq_num;
1118         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1119                 if (!id_priv->cm_id.iw) {
1120                         qp_attr->qp_access_flags = 0;
1121                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1122                 } else
1123                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1124                                                  qp_attr_mask);
1125                 qp_attr->port_num = id_priv->id.port_num;
1126                 *qp_attr_mask |= IB_QP_PORT;
1127         } else
1128                 ret = -ENOSYS;
1129
1130         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1131                 qp_attr->timeout = id_priv->timeout;
1132
1133         return ret;
1134 }
1135 EXPORT_SYMBOL(rdma_init_qp_attr);
1136
1137 static inline bool cma_zero_addr(const struct sockaddr *addr)
1138 {
1139         switch (addr->sa_family) {
1140         case AF_INET:
1141                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1142         case AF_INET6:
1143                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1144         case AF_IB:
1145                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1146         default:
1147                 return false;
1148         }
1149 }
1150
1151 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1152 {
1153         switch (addr->sa_family) {
1154         case AF_INET:
1155                 return ipv4_is_loopback(
1156                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1157         case AF_INET6:
1158                 return ipv6_addr_loopback(
1159                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1160         case AF_IB:
1161                 return ib_addr_loopback(
1162                         &((struct sockaddr_ib *)addr)->sib_addr);
1163         default:
1164                 return false;
1165         }
1166 }
1167
1168 static inline bool cma_any_addr(const struct sockaddr *addr)
1169 {
1170         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1171 }
1172
1173 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1174 {
1175         if (src->sa_family != dst->sa_family)
1176                 return -1;
1177
1178         switch (src->sa_family) {
1179         case AF_INET:
1180                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1181                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1182         case AF_INET6: {
1183                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1184                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1185                 bool link_local;
1186
1187                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1188                                           &dst_addr6->sin6_addr))
1189                         return 1;
1190                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1191                              IPV6_ADDR_LINKLOCAL;
1192                 /* Link local must match their scope_ids */
1193                 return link_local ? (src_addr6->sin6_scope_id !=
1194                                      dst_addr6->sin6_scope_id) :
1195                                     0;
1196         }
1197
1198         default:
1199                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1200                                    &((struct sockaddr_ib *) dst)->sib_addr);
1201         }
1202 }
1203
1204 static __be16 cma_port(const struct sockaddr *addr)
1205 {
1206         struct sockaddr_ib *sib;
1207
1208         switch (addr->sa_family) {
1209         case AF_INET:
1210                 return ((struct sockaddr_in *) addr)->sin_port;
1211         case AF_INET6:
1212                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1213         case AF_IB:
1214                 sib = (struct sockaddr_ib *) addr;
1215                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1216                                     be64_to_cpu(sib->sib_sid_mask)));
1217         default:
1218                 return 0;
1219         }
1220 }
1221
1222 static inline int cma_any_port(const struct sockaddr *addr)
1223 {
1224         return !cma_port(addr);
1225 }
1226
1227 static void cma_save_ib_info(struct sockaddr *src_addr,
1228                              struct sockaddr *dst_addr,
1229                              const struct rdma_cm_id *listen_id,
1230                              const struct sa_path_rec *path)
1231 {
1232         struct sockaddr_ib *listen_ib, *ib;
1233
1234         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1235         if (src_addr) {
1236                 ib = (struct sockaddr_ib *)src_addr;
1237                 ib->sib_family = AF_IB;
1238                 if (path) {
1239                         ib->sib_pkey = path->pkey;
1240                         ib->sib_flowinfo = path->flow_label;
1241                         memcpy(&ib->sib_addr, &path->sgid, 16);
1242                         ib->sib_sid = path->service_id;
1243                         ib->sib_scope_id = 0;
1244                 } else {
1245                         ib->sib_pkey = listen_ib->sib_pkey;
1246                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1247                         ib->sib_addr = listen_ib->sib_addr;
1248                         ib->sib_sid = listen_ib->sib_sid;
1249                         ib->sib_scope_id = listen_ib->sib_scope_id;
1250                 }
1251                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1252         }
1253         if (dst_addr) {
1254                 ib = (struct sockaddr_ib *)dst_addr;
1255                 ib->sib_family = AF_IB;
1256                 if (path) {
1257                         ib->sib_pkey = path->pkey;
1258                         ib->sib_flowinfo = path->flow_label;
1259                         memcpy(&ib->sib_addr, &path->dgid, 16);
1260                 }
1261         }
1262 }
1263
1264 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1265                               struct sockaddr_in *dst_addr,
1266                               struct cma_hdr *hdr,
1267                               __be16 local_port)
1268 {
1269         if (src_addr) {
1270                 *src_addr = (struct sockaddr_in) {
1271                         .sin_family = AF_INET,
1272                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1273                         .sin_port = local_port,
1274                 };
1275         }
1276
1277         if (dst_addr) {
1278                 *dst_addr = (struct sockaddr_in) {
1279                         .sin_family = AF_INET,
1280                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1281                         .sin_port = hdr->port,
1282                 };
1283         }
1284 }
1285
1286 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1287                               struct sockaddr_in6 *dst_addr,
1288                               struct cma_hdr *hdr,
1289                               __be16 local_port)
1290 {
1291         if (src_addr) {
1292                 *src_addr = (struct sockaddr_in6) {
1293                         .sin6_family = AF_INET6,
1294                         .sin6_addr = hdr->dst_addr.ip6,
1295                         .sin6_port = local_port,
1296                 };
1297         }
1298
1299         if (dst_addr) {
1300                 *dst_addr = (struct sockaddr_in6) {
1301                         .sin6_family = AF_INET6,
1302                         .sin6_addr = hdr->src_addr.ip6,
1303                         .sin6_port = hdr->port,
1304                 };
1305         }
1306 }
1307
1308 static u16 cma_port_from_service_id(__be64 service_id)
1309 {
1310         return (u16)be64_to_cpu(service_id);
1311 }
1312
1313 static int cma_save_ip_info(struct sockaddr *src_addr,
1314                             struct sockaddr *dst_addr,
1315                             const struct ib_cm_event *ib_event,
1316                             __be64 service_id)
1317 {
1318         struct cma_hdr *hdr;
1319         __be16 port;
1320
1321         hdr = ib_event->private_data;
1322         if (hdr->cma_version != CMA_VERSION)
1323                 return -EINVAL;
1324
1325         port = htons(cma_port_from_service_id(service_id));
1326
1327         switch (cma_get_ip_ver(hdr)) {
1328         case 4:
1329                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1330                                   (struct sockaddr_in *)dst_addr, hdr, port);
1331                 break;
1332         case 6:
1333                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1334                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1335                 break;
1336         default:
1337                 return -EAFNOSUPPORT;
1338         }
1339
1340         return 0;
1341 }
1342
1343 static int cma_save_net_info(struct sockaddr *src_addr,
1344                              struct sockaddr *dst_addr,
1345                              const struct rdma_cm_id *listen_id,
1346                              const struct ib_cm_event *ib_event,
1347                              sa_family_t sa_family, __be64 service_id)
1348 {
1349         if (sa_family == AF_IB) {
1350                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1351                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1352                                          ib_event->param.req_rcvd.primary_path);
1353                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1354                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1355                 return 0;
1356         }
1357
1358         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1359 }
1360
1361 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1362                              struct cma_req_info *req)
1363 {
1364         const struct ib_cm_req_event_param *req_param =
1365                 &ib_event->param.req_rcvd;
1366         const struct ib_cm_sidr_req_event_param *sidr_param =
1367                 &ib_event->param.sidr_req_rcvd;
1368
1369         switch (ib_event->event) {
1370         case IB_CM_REQ_RECEIVED:
1371                 req->device     = req_param->listen_id->device;
1372                 req->port       = req_param->port;
1373                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1374                        sizeof(req->local_gid));
1375                 req->has_gid    = true;
1376                 req->service_id = req_param->primary_path->service_id;
1377                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1378                 if (req->pkey != req_param->bth_pkey)
1379                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1380                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1381                                             req_param->bth_pkey, req->pkey);
1382                 break;
1383         case IB_CM_SIDR_REQ_RECEIVED:
1384                 req->device     = sidr_param->listen_id->device;
1385                 req->port       = sidr_param->port;
1386                 req->has_gid    = false;
1387                 req->service_id = sidr_param->service_id;
1388                 req->pkey       = sidr_param->pkey;
1389                 if (req->pkey != sidr_param->bth_pkey)
1390                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1391                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1392                                             sidr_param->bth_pkey, req->pkey);
1393                 break;
1394         default:
1395                 return -EINVAL;
1396         }
1397
1398         return 0;
1399 }
1400
1401 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1402                                   const struct sockaddr_in *dst_addr,
1403                                   const struct sockaddr_in *src_addr)
1404 {
1405         __be32 daddr = dst_addr->sin_addr.s_addr,
1406                saddr = src_addr->sin_addr.s_addr;
1407         struct fib_result res;
1408         struct flowi4 fl4;
1409         int err;
1410         bool ret;
1411
1412         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1413             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1414             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1415             ipv4_is_loopback(saddr))
1416                 return false;
1417
1418         memset(&fl4, 0, sizeof(fl4));
1419         fl4.flowi4_iif = net_dev->ifindex;
1420         fl4.daddr = daddr;
1421         fl4.saddr = saddr;
1422
1423         rcu_read_lock();
1424         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1425         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1426         rcu_read_unlock();
1427
1428         return ret;
1429 }
1430
1431 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1432                                   const struct sockaddr_in6 *dst_addr,
1433                                   const struct sockaddr_in6 *src_addr)
1434 {
1435 #if IS_ENABLED(CONFIG_IPV6)
1436         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1437                            IPV6_ADDR_LINKLOCAL;
1438         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1439                                          &src_addr->sin6_addr, net_dev->ifindex,
1440                                          NULL, strict);
1441         bool ret;
1442
1443         if (!rt)
1444                 return false;
1445
1446         ret = rt->rt6i_idev->dev == net_dev;
1447         ip6_rt_put(rt);
1448
1449         return ret;
1450 #else
1451         return false;
1452 #endif
1453 }
1454
1455 static bool validate_net_dev(struct net_device *net_dev,
1456                              const struct sockaddr *daddr,
1457                              const struct sockaddr *saddr)
1458 {
1459         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1460         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1461         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1462         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1463
1464         switch (daddr->sa_family) {
1465         case AF_INET:
1466                 return saddr->sa_family == AF_INET &&
1467                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1468
1469         case AF_INET6:
1470                 return saddr->sa_family == AF_INET6 &&
1471                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1472
1473         default:
1474                 return false;
1475         }
1476 }
1477
1478 static struct net_device *
1479 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1480 {
1481         const struct ib_gid_attr *sgid_attr = NULL;
1482         struct net_device *ndev;
1483
1484         if (ib_event->event == IB_CM_REQ_RECEIVED)
1485                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1486         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1487                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1488
1489         if (!sgid_attr)
1490                 return NULL;
1491
1492         rcu_read_lock();
1493         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1494         if (IS_ERR(ndev))
1495                 ndev = NULL;
1496         else
1497                 dev_hold(ndev);
1498         rcu_read_unlock();
1499         return ndev;
1500 }
1501
1502 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1503                                           struct cma_req_info *req)
1504 {
1505         struct sockaddr *listen_addr =
1506                         (struct sockaddr *)&req->listen_addr_storage;
1507         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1508         struct net_device *net_dev;
1509         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1510         int err;
1511
1512         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1513                                req->service_id);
1514         if (err)
1515                 return ERR_PTR(err);
1516
1517         if (rdma_protocol_roce(req->device, req->port))
1518                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1519         else
1520                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1521                                                    req->pkey,
1522                                                    gid, listen_addr);
1523         if (!net_dev)
1524                 return ERR_PTR(-ENODEV);
1525
1526         return net_dev;
1527 }
1528
1529 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1530 {
1531         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1532 }
1533
1534 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1535                                    const struct cma_hdr *hdr)
1536 {
1537         struct sockaddr *addr = cma_src_addr(id_priv);
1538         __be32 ip4_addr;
1539         struct in6_addr ip6_addr;
1540
1541         if (cma_any_addr(addr) && !id_priv->afonly)
1542                 return true;
1543
1544         switch (addr->sa_family) {
1545         case AF_INET:
1546                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1547                 if (cma_get_ip_ver(hdr) != 4)
1548                         return false;
1549                 if (!cma_any_addr(addr) &&
1550                     hdr->dst_addr.ip4.addr != ip4_addr)
1551                         return false;
1552                 break;
1553         case AF_INET6:
1554                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1555                 if (cma_get_ip_ver(hdr) != 6)
1556                         return false;
1557                 if (!cma_any_addr(addr) &&
1558                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1559                         return false;
1560                 break;
1561         case AF_IB:
1562                 return true;
1563         default:
1564                 return false;
1565         }
1566
1567         return true;
1568 }
1569
1570 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1571 {
1572         struct ib_device *device = id->device;
1573         const int port_num = id->port_num ?: rdma_start_port(device);
1574
1575         return rdma_protocol_roce(device, port_num);
1576 }
1577
1578 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1579 {
1580         const struct sockaddr *daddr =
1581                         (const struct sockaddr *)&req->listen_addr_storage;
1582         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1583
1584         /* Returns true if the req is for IPv6 link local */
1585         return (daddr->sa_family == AF_INET6 &&
1586                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1587 }
1588
1589 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1590                               const struct net_device *net_dev,
1591                               const struct cma_req_info *req)
1592 {
1593         const struct rdma_addr *addr = &id->route.addr;
1594
1595         if (!net_dev)
1596                 /* This request is an AF_IB request */
1597                 return (!id->port_num || id->port_num == req->port) &&
1598                        (addr->src_addr.ss_family == AF_IB);
1599
1600         /*
1601          * If the request is not for IPv6 link local, allow matching
1602          * request to any netdevice of the one or multiport rdma device.
1603          */
1604         if (!cma_is_req_ipv6_ll(req))
1605                 return true;
1606         /*
1607          * Net namespaces must match, and if the listner is listening
1608          * on a specific netdevice than netdevice must match as well.
1609          */
1610         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1611             (!!addr->dev_addr.bound_dev_if ==
1612              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1613                 return true;
1614         else
1615                 return false;
1616 }
1617
1618 static struct rdma_id_private *cma_find_listener(
1619                 const struct rdma_bind_list *bind_list,
1620                 const struct ib_cm_id *cm_id,
1621                 const struct ib_cm_event *ib_event,
1622                 const struct cma_req_info *req,
1623                 const struct net_device *net_dev)
1624 {
1625         struct rdma_id_private *id_priv, *id_priv_dev;
1626
1627         if (!bind_list)
1628                 return ERR_PTR(-EINVAL);
1629
1630         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1631                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1632                         if (id_priv->id.device == cm_id->device &&
1633                             cma_match_net_dev(&id_priv->id, net_dev, req))
1634                                 return id_priv;
1635                         list_for_each_entry(id_priv_dev,
1636                                             &id_priv->listen_list,
1637                                             listen_list) {
1638                                 if (id_priv_dev->id.device == cm_id->device &&
1639                                     cma_match_net_dev(&id_priv_dev->id,
1640                                                       net_dev, req))
1641                                         return id_priv_dev;
1642                         }
1643                 }
1644         }
1645
1646         return ERR_PTR(-EINVAL);
1647 }
1648
1649 static struct rdma_id_private *
1650 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1651                      const struct ib_cm_event *ib_event,
1652                      struct cma_req_info *req,
1653                      struct net_device **net_dev)
1654 {
1655         struct rdma_bind_list *bind_list;
1656         struct rdma_id_private *id_priv;
1657         int err;
1658
1659         err = cma_save_req_info(ib_event, req);
1660         if (err)
1661                 return ERR_PTR(err);
1662
1663         *net_dev = cma_get_net_dev(ib_event, req);
1664         if (IS_ERR(*net_dev)) {
1665                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1666                         /* Assuming the protocol is AF_IB */
1667                         *net_dev = NULL;
1668                 } else {
1669                         return ERR_CAST(*net_dev);
1670                 }
1671         }
1672
1673         /*
1674          * Net namespace might be getting deleted while route lookup,
1675          * cm_id lookup is in progress. Therefore, perform netdevice
1676          * validation, cm_id lookup under rcu lock.
1677          * RCU lock along with netdevice state check, synchronizes with
1678          * netdevice migrating to different net namespace and also avoids
1679          * case where net namespace doesn't get deleted while lookup is in
1680          * progress.
1681          * If the device state is not IFF_UP, its properties such as ifindex
1682          * and nd_net cannot be trusted to remain valid without rcu lock.
1683          * net/core/dev.c change_net_namespace() ensures to synchronize with
1684          * ongoing operations on net device after device is closed using
1685          * synchronize_net().
1686          */
1687         rcu_read_lock();
1688         if (*net_dev) {
1689                 /*
1690                  * If netdevice is down, it is likely that it is administratively
1691                  * down or it might be migrating to different namespace.
1692                  * In that case avoid further processing, as the net namespace
1693                  * or ifindex may change.
1694                  */
1695                 if (((*net_dev)->flags & IFF_UP) == 0) {
1696                         id_priv = ERR_PTR(-EHOSTUNREACH);
1697                         goto err;
1698                 }
1699
1700                 if (!validate_net_dev(*net_dev,
1701                                  (struct sockaddr *)&req->listen_addr_storage,
1702                                  (struct sockaddr *)&req->src_addr_storage)) {
1703                         id_priv = ERR_PTR(-EHOSTUNREACH);
1704                         goto err;
1705                 }
1706         }
1707
1708         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1709                                 rdma_ps_from_service_id(req->service_id),
1710                                 cma_port_from_service_id(req->service_id));
1711         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1712 err:
1713         rcu_read_unlock();
1714         if (IS_ERR(id_priv) && *net_dev) {
1715                 dev_put(*net_dev);
1716                 *net_dev = NULL;
1717         }
1718         return id_priv;
1719 }
1720
1721 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1722 {
1723         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1724 }
1725
1726 static void cma_cancel_route(struct rdma_id_private *id_priv)
1727 {
1728         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1729                 if (id_priv->query)
1730                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1731         }
1732 }
1733
1734 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1735 {
1736         struct rdma_id_private *dev_id_priv;
1737
1738         /*
1739          * Remove from listen_any_list to prevent added devices from spawning
1740          * additional listen requests.
1741          */
1742         mutex_lock(&lock);
1743         list_del(&id_priv->list);
1744
1745         while (!list_empty(&id_priv->listen_list)) {
1746                 dev_id_priv = list_entry(id_priv->listen_list.next,
1747                                          struct rdma_id_private, listen_list);
1748                 /* sync with device removal to avoid duplicate destruction */
1749                 list_del_init(&dev_id_priv->list);
1750                 list_del(&dev_id_priv->listen_list);
1751                 mutex_unlock(&lock);
1752
1753                 rdma_destroy_id(&dev_id_priv->id);
1754                 mutex_lock(&lock);
1755         }
1756         mutex_unlock(&lock);
1757 }
1758
1759 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1760                                  enum rdma_cm_state state)
1761 {
1762         switch (state) {
1763         case RDMA_CM_ADDR_QUERY:
1764                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1765                 break;
1766         case RDMA_CM_ROUTE_QUERY:
1767                 cma_cancel_route(id_priv);
1768                 break;
1769         case RDMA_CM_LISTEN:
1770                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1771                         cma_cancel_listens(id_priv);
1772                 break;
1773         default:
1774                 break;
1775         }
1776 }
1777
1778 static void cma_release_port(struct rdma_id_private *id_priv)
1779 {
1780         struct rdma_bind_list *bind_list = id_priv->bind_list;
1781         struct net *net = id_priv->id.route.addr.dev_addr.net;
1782
1783         if (!bind_list)
1784                 return;
1785
1786         mutex_lock(&lock);
1787         hlist_del(&id_priv->node);
1788         if (hlist_empty(&bind_list->owners)) {
1789                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1790                 kfree(bind_list);
1791         }
1792         mutex_unlock(&lock);
1793 }
1794
1795 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv,
1796                                     struct cma_multicast *mc)
1797 {
1798         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1799         struct net_device *ndev = NULL;
1800
1801         if (dev_addr->bound_dev_if)
1802                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
1803         if (ndev) {
1804                 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1805                 dev_put(ndev);
1806         }
1807         kref_put(&mc->mcref, release_mc);
1808 }
1809
1810 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1811 {
1812         struct cma_multicast *mc;
1813
1814         while (!list_empty(&id_priv->mc_list)) {
1815                 mc = container_of(id_priv->mc_list.next,
1816                                   struct cma_multicast, list);
1817                 list_del(&mc->list);
1818                 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1819                                       id_priv->id.port_num)) {
1820                         ib_sa_free_multicast(mc->multicast.ib);
1821                         kfree(mc);
1822                 } else {
1823                         cma_leave_roce_mc_group(id_priv, mc);
1824                 }
1825         }
1826 }
1827
1828 void rdma_destroy_id(struct rdma_cm_id *id)
1829 {
1830         struct rdma_id_private *id_priv;
1831         enum rdma_cm_state state;
1832
1833         id_priv = container_of(id, struct rdma_id_private, id);
1834         trace_cm_id_destroy(id_priv);
1835         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1836         cma_cancel_operation(id_priv, state);
1837
1838         /*
1839          * Wait for any active callback to finish.  New callbacks will find
1840          * the id_priv state set to destroying and abort.
1841          */
1842         mutex_lock(&id_priv->handler_mutex);
1843         mutex_unlock(&id_priv->handler_mutex);
1844
1845         rdma_restrack_del(&id_priv->res);
1846         if (id_priv->cma_dev) {
1847                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1848                         if (id_priv->cm_id.ib)
1849                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1850                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1851                         if (id_priv->cm_id.iw)
1852                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1853                 }
1854                 cma_leave_mc_groups(id_priv);
1855                 cma_release_dev(id_priv);
1856         }
1857
1858         cma_release_port(id_priv);
1859         cma_id_put(id_priv);
1860         wait_for_completion(&id_priv->comp);
1861
1862         if (id_priv->internal_id)
1863                 cma_id_put(id_priv->id.context);
1864
1865         kfree(id_priv->id.route.path_rec);
1866
1867         if (id_priv->id.route.addr.dev_addr.sgid_attr)
1868                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1869
1870         put_net(id_priv->id.route.addr.dev_addr.net);
1871         kfree(id_priv);
1872 }
1873 EXPORT_SYMBOL(rdma_destroy_id);
1874
1875 static int cma_rep_recv(struct rdma_id_private *id_priv)
1876 {
1877         int ret;
1878
1879         ret = cma_modify_qp_rtr(id_priv, NULL);
1880         if (ret)
1881                 goto reject;
1882
1883         ret = cma_modify_qp_rts(id_priv, NULL);
1884         if (ret)
1885                 goto reject;
1886
1887         trace_cm_send_rtu(id_priv);
1888         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1889         if (ret)
1890                 goto reject;
1891
1892         return 0;
1893 reject:
1894         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1895         cma_modify_qp_err(id_priv);
1896         trace_cm_send_rej(id_priv);
1897         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1898                        NULL, 0, NULL, 0);
1899         return ret;
1900 }
1901
1902 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1903                                    const struct ib_cm_rep_event_param *rep_data,
1904                                    void *private_data)
1905 {
1906         event->param.conn.private_data = private_data;
1907         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1908         event->param.conn.responder_resources = rep_data->responder_resources;
1909         event->param.conn.initiator_depth = rep_data->initiator_depth;
1910         event->param.conn.flow_control = rep_data->flow_control;
1911         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1912         event->param.conn.srq = rep_data->srq;
1913         event->param.conn.qp_num = rep_data->remote_qpn;
1914 }
1915
1916 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1917                                 struct rdma_cm_event *event)
1918 {
1919         int ret;
1920
1921         trace_cm_event_handler(id_priv, event);
1922         ret = id_priv->id.event_handler(&id_priv->id, event);
1923         trace_cm_event_done(id_priv, event, ret);
1924         return ret;
1925 }
1926
1927 static int cma_ib_handler(struct ib_cm_id *cm_id,
1928                           const struct ib_cm_event *ib_event)
1929 {
1930         struct rdma_id_private *id_priv = cm_id->context;
1931         struct rdma_cm_event event = {};
1932         int ret = 0;
1933
1934         mutex_lock(&id_priv->handler_mutex);
1935         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1936              id_priv->state != RDMA_CM_CONNECT) ||
1937             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1938              id_priv->state != RDMA_CM_DISCONNECT))
1939                 goto out;
1940
1941         switch (ib_event->event) {
1942         case IB_CM_REQ_ERROR:
1943         case IB_CM_REP_ERROR:
1944                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1945                 event.status = -ETIMEDOUT;
1946                 break;
1947         case IB_CM_REP_RECEIVED:
1948                 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1949                     (id_priv->id.qp_type != IB_QPT_UD)) {
1950                         trace_cm_send_mra(id_priv);
1951                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1952                 }
1953                 if (id_priv->id.qp) {
1954                         event.status = cma_rep_recv(id_priv);
1955                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1956                                                      RDMA_CM_EVENT_ESTABLISHED;
1957                 } else {
1958                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1959                 }
1960                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1961                                        ib_event->private_data);
1962                 break;
1963         case IB_CM_RTU_RECEIVED:
1964         case IB_CM_USER_ESTABLISHED:
1965                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1966                 break;
1967         case IB_CM_DREQ_ERROR:
1968                 event.status = -ETIMEDOUT; /* fall through */
1969         case IB_CM_DREQ_RECEIVED:
1970         case IB_CM_DREP_RECEIVED:
1971                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1972                                    RDMA_CM_DISCONNECT))
1973                         goto out;
1974                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1975                 break;
1976         case IB_CM_TIMEWAIT_EXIT:
1977                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1978                 break;
1979         case IB_CM_MRA_RECEIVED:
1980                 /* ignore event */
1981                 goto out;
1982         case IB_CM_REJ_RECEIVED:
1983                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1984                                                                                 ib_event->param.rej_rcvd.reason));
1985                 cma_modify_qp_err(id_priv);
1986                 event.status = ib_event->param.rej_rcvd.reason;
1987                 event.event = RDMA_CM_EVENT_REJECTED;
1988                 event.param.conn.private_data = ib_event->private_data;
1989                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1990                 break;
1991         default:
1992                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1993                        ib_event->event);
1994                 goto out;
1995         }
1996
1997         ret = cma_cm_event_handler(id_priv, &event);
1998         if (ret) {
1999                 /* Destroy the CM ID by returning a non-zero value. */
2000                 id_priv->cm_id.ib = NULL;
2001                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2002                 mutex_unlock(&id_priv->handler_mutex);
2003                 rdma_destroy_id(&id_priv->id);
2004                 return ret;
2005         }
2006 out:
2007         mutex_unlock(&id_priv->handler_mutex);
2008         return ret;
2009 }
2010
2011 static struct rdma_id_private *
2012 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2013                    const struct ib_cm_event *ib_event,
2014                    struct net_device *net_dev)
2015 {
2016         struct rdma_id_private *listen_id_priv;
2017         struct rdma_id_private *id_priv;
2018         struct rdma_cm_id *id;
2019         struct rdma_route *rt;
2020         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2021         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2022         const __be64 service_id =
2023                 ib_event->param.req_rcvd.primary_path->service_id;
2024         int ret;
2025
2026         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2027         id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2028                             listen_id->event_handler, listen_id->context,
2029                             listen_id->ps, ib_event->param.req_rcvd.qp_type,
2030                             listen_id_priv->res.kern_name);
2031         if (IS_ERR(id))
2032                 return NULL;
2033
2034         id_priv = container_of(id, struct rdma_id_private, id);
2035         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2036                               (struct sockaddr *)&id->route.addr.dst_addr,
2037                               listen_id, ib_event, ss_family, service_id))
2038                 goto err;
2039
2040         rt = &id->route;
2041         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2042         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2043                                      GFP_KERNEL);
2044         if (!rt->path_rec)
2045                 goto err;
2046
2047         rt->path_rec[0] = *path;
2048         if (rt->num_paths == 2)
2049                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2050
2051         if (net_dev) {
2052                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2053         } else {
2054                 if (!cma_protocol_roce(listen_id) &&
2055                     cma_any_addr(cma_src_addr(id_priv))) {
2056                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2057                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2058                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2059                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2060                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2061                         if (ret)
2062                                 goto err;
2063                 }
2064         }
2065         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2066
2067         id_priv->state = RDMA_CM_CONNECT;
2068         return id_priv;
2069
2070 err:
2071         rdma_destroy_id(id);
2072         return NULL;
2073 }
2074
2075 static struct rdma_id_private *
2076 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2077                   const struct ib_cm_event *ib_event,
2078                   struct net_device *net_dev)
2079 {
2080         const struct rdma_id_private *listen_id_priv;
2081         struct rdma_id_private *id_priv;
2082         struct rdma_cm_id *id;
2083         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2084         struct net *net = listen_id->route.addr.dev_addr.net;
2085         int ret;
2086
2087         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2088         id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
2089                               listen_id->ps, IB_QPT_UD,
2090                               listen_id_priv->res.kern_name);
2091         if (IS_ERR(id))
2092                 return NULL;
2093
2094         id_priv = container_of(id, struct rdma_id_private, id);
2095         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2096                               (struct sockaddr *)&id->route.addr.dst_addr,
2097                               listen_id, ib_event, ss_family,
2098                               ib_event->param.sidr_req_rcvd.service_id))
2099                 goto err;
2100
2101         if (net_dev) {
2102                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2103         } else {
2104                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2105                         ret = cma_translate_addr(cma_src_addr(id_priv),
2106                                                  &id->route.addr.dev_addr);
2107                         if (ret)
2108                                 goto err;
2109                 }
2110         }
2111
2112         id_priv->state = RDMA_CM_CONNECT;
2113         return id_priv;
2114 err:
2115         rdma_destroy_id(id);
2116         return NULL;
2117 }
2118
2119 static void cma_set_req_event_data(struct rdma_cm_event *event,
2120                                    const struct ib_cm_req_event_param *req_data,
2121                                    void *private_data, int offset)
2122 {
2123         event->param.conn.private_data = private_data + offset;
2124         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2125         event->param.conn.responder_resources = req_data->responder_resources;
2126         event->param.conn.initiator_depth = req_data->initiator_depth;
2127         event->param.conn.flow_control = req_data->flow_control;
2128         event->param.conn.retry_count = req_data->retry_count;
2129         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2130         event->param.conn.srq = req_data->srq;
2131         event->param.conn.qp_num = req_data->remote_qpn;
2132 }
2133
2134 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2135                                     const struct ib_cm_event *ib_event)
2136 {
2137         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2138                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2139                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2140                  (id->qp_type == IB_QPT_UD)) ||
2141                 (!id->qp_type));
2142 }
2143
2144 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2145                               const struct ib_cm_event *ib_event)
2146 {
2147         struct rdma_id_private *listen_id, *conn_id = NULL;
2148         struct rdma_cm_event event = {};
2149         struct cma_req_info req = {};
2150         struct net_device *net_dev;
2151         u8 offset;
2152         int ret;
2153
2154         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2155         if (IS_ERR(listen_id))
2156                 return PTR_ERR(listen_id);
2157
2158         trace_cm_req_handler(listen_id, ib_event->event);
2159         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2160                 ret = -EINVAL;
2161                 goto net_dev_put;
2162         }
2163
2164         mutex_lock(&listen_id->handler_mutex);
2165         if (listen_id->state != RDMA_CM_LISTEN) {
2166                 ret = -ECONNABORTED;
2167                 goto err1;
2168         }
2169
2170         offset = cma_user_data_offset(listen_id);
2171         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2172         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2173                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2174                 event.param.ud.private_data = ib_event->private_data + offset;
2175                 event.param.ud.private_data_len =
2176                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2177         } else {
2178                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2179                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2180                                        ib_event->private_data, offset);
2181         }
2182         if (!conn_id) {
2183                 ret = -ENOMEM;
2184                 goto err1;
2185         }
2186
2187         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2188         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2189         if (ret)
2190                 goto err2;
2191
2192         conn_id->cm_id.ib = cm_id;
2193         cm_id->context = conn_id;
2194         cm_id->cm_handler = cma_ib_handler;
2195
2196         /*
2197          * Protect against the user destroying conn_id from another thread
2198          * until we're done accessing it.
2199          */
2200         cma_id_get(conn_id);
2201         ret = cma_cm_event_handler(conn_id, &event);
2202         if (ret)
2203                 goto err3;
2204         /*
2205          * Acquire mutex to prevent user executing rdma_destroy_id()
2206          * while we're accessing the cm_id.
2207          */
2208         mutex_lock(&lock);
2209         if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2210             (conn_id->id.qp_type != IB_QPT_UD)) {
2211                 trace_cm_send_mra(cm_id->context);
2212                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2213         }
2214         mutex_unlock(&lock);
2215         mutex_unlock(&conn_id->handler_mutex);
2216         mutex_unlock(&listen_id->handler_mutex);
2217         cma_id_put(conn_id);
2218         if (net_dev)
2219                 dev_put(net_dev);
2220         return 0;
2221
2222 err3:
2223         cma_id_put(conn_id);
2224         /* Destroy the CM ID by returning a non-zero value. */
2225         conn_id->cm_id.ib = NULL;
2226 err2:
2227         cma_exch(conn_id, RDMA_CM_DESTROYING);
2228         mutex_unlock(&conn_id->handler_mutex);
2229 err1:
2230         mutex_unlock(&listen_id->handler_mutex);
2231         if (conn_id)
2232                 rdma_destroy_id(&conn_id->id);
2233
2234 net_dev_put:
2235         if (net_dev)
2236                 dev_put(net_dev);
2237
2238         return ret;
2239 }
2240
2241 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2242 {
2243         if (addr->sa_family == AF_IB)
2244                 return ((struct sockaddr_ib *) addr)->sib_sid;
2245
2246         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2247 }
2248 EXPORT_SYMBOL(rdma_get_service_id);
2249
2250 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2251                     union ib_gid *dgid)
2252 {
2253         struct rdma_addr *addr = &cm_id->route.addr;
2254
2255         if (!cm_id->device) {
2256                 if (sgid)
2257                         memset(sgid, 0, sizeof(*sgid));
2258                 if (dgid)
2259                         memset(dgid, 0, sizeof(*dgid));
2260                 return;
2261         }
2262
2263         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2264                 if (sgid)
2265                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2266                 if (dgid)
2267                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2268         } else {
2269                 if (sgid)
2270                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2271                 if (dgid)
2272                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2273         }
2274 }
2275 EXPORT_SYMBOL(rdma_read_gids);
2276
2277 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2278 {
2279         struct rdma_id_private *id_priv = iw_id->context;
2280         struct rdma_cm_event event = {};
2281         int ret = 0;
2282         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2283         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2284
2285         mutex_lock(&id_priv->handler_mutex);
2286         if (id_priv->state != RDMA_CM_CONNECT)
2287                 goto out;
2288
2289         switch (iw_event->event) {
2290         case IW_CM_EVENT_CLOSE:
2291                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2292                 break;
2293         case IW_CM_EVENT_CONNECT_REPLY:
2294                 memcpy(cma_src_addr(id_priv), laddr,
2295                        rdma_addr_size(laddr));
2296                 memcpy(cma_dst_addr(id_priv), raddr,
2297                        rdma_addr_size(raddr));
2298                 switch (iw_event->status) {
2299                 case 0:
2300                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2301                         event.param.conn.initiator_depth = iw_event->ird;
2302                         event.param.conn.responder_resources = iw_event->ord;
2303                         break;
2304                 case -ECONNRESET:
2305                 case -ECONNREFUSED:
2306                         event.event = RDMA_CM_EVENT_REJECTED;
2307                         break;
2308                 case -ETIMEDOUT:
2309                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2310                         break;
2311                 default:
2312                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2313                         break;
2314                 }
2315                 break;
2316         case IW_CM_EVENT_ESTABLISHED:
2317                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2318                 event.param.conn.initiator_depth = iw_event->ird;
2319                 event.param.conn.responder_resources = iw_event->ord;
2320                 break;
2321         default:
2322                 goto out;
2323         }
2324
2325         event.status = iw_event->status;
2326         event.param.conn.private_data = iw_event->private_data;
2327         event.param.conn.private_data_len = iw_event->private_data_len;
2328         ret = cma_cm_event_handler(id_priv, &event);
2329         if (ret) {
2330                 /* Destroy the CM ID by returning a non-zero value. */
2331                 id_priv->cm_id.iw = NULL;
2332                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2333                 mutex_unlock(&id_priv->handler_mutex);
2334                 rdma_destroy_id(&id_priv->id);
2335                 return ret;
2336         }
2337
2338 out:
2339         mutex_unlock(&id_priv->handler_mutex);
2340         return ret;
2341 }
2342
2343 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2344                                struct iw_cm_event *iw_event)
2345 {
2346         struct rdma_cm_id *new_cm_id;
2347         struct rdma_id_private *listen_id, *conn_id;
2348         struct rdma_cm_event event = {};
2349         int ret = -ECONNABORTED;
2350         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2351         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2352
2353         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2354         event.param.conn.private_data = iw_event->private_data;
2355         event.param.conn.private_data_len = iw_event->private_data_len;
2356         event.param.conn.initiator_depth = iw_event->ird;
2357         event.param.conn.responder_resources = iw_event->ord;
2358
2359         listen_id = cm_id->context;
2360
2361         mutex_lock(&listen_id->handler_mutex);
2362         if (listen_id->state != RDMA_CM_LISTEN)
2363                 goto out;
2364
2365         /* Create a new RDMA id for the new IW CM ID */
2366         new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2367                                      listen_id->id.event_handler,
2368                                      listen_id->id.context,
2369                                      RDMA_PS_TCP, IB_QPT_RC,
2370                                      listen_id->res.kern_name);
2371         if (IS_ERR(new_cm_id)) {
2372                 ret = -ENOMEM;
2373                 goto out;
2374         }
2375         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2376         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2377         conn_id->state = RDMA_CM_CONNECT;
2378
2379         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2380         if (ret) {
2381                 mutex_unlock(&conn_id->handler_mutex);
2382                 rdma_destroy_id(new_cm_id);
2383                 goto out;
2384         }
2385
2386         ret = cma_iw_acquire_dev(conn_id, listen_id);
2387         if (ret) {
2388                 mutex_unlock(&conn_id->handler_mutex);
2389                 rdma_destroy_id(new_cm_id);
2390                 goto out;
2391         }
2392
2393         conn_id->cm_id.iw = cm_id;
2394         cm_id->context = conn_id;
2395         cm_id->cm_handler = cma_iw_handler;
2396
2397         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2398         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2399
2400         /*
2401          * Protect against the user destroying conn_id from another thread
2402          * until we're done accessing it.
2403          */
2404         cma_id_get(conn_id);
2405         ret = cma_cm_event_handler(conn_id, &event);
2406         if (ret) {
2407                 /* User wants to destroy the CM ID */
2408                 conn_id->cm_id.iw = NULL;
2409                 cma_exch(conn_id, RDMA_CM_DESTROYING);
2410                 mutex_unlock(&conn_id->handler_mutex);
2411                 mutex_unlock(&listen_id->handler_mutex);
2412                 cma_id_put(conn_id);
2413                 rdma_destroy_id(&conn_id->id);
2414                 return ret;
2415         }
2416
2417         mutex_unlock(&conn_id->handler_mutex);
2418         cma_id_put(conn_id);
2419
2420 out:
2421         mutex_unlock(&listen_id->handler_mutex);
2422         return ret;
2423 }
2424
2425 static int cma_ib_listen(struct rdma_id_private *id_priv)
2426 {
2427         struct sockaddr *addr;
2428         struct ib_cm_id *id;
2429         __be64 svc_id;
2430
2431         addr = cma_src_addr(id_priv);
2432         svc_id = rdma_get_service_id(&id_priv->id, addr);
2433         id = ib_cm_insert_listen(id_priv->id.device,
2434                                  cma_ib_req_handler, svc_id);
2435         if (IS_ERR(id))
2436                 return PTR_ERR(id);
2437         id_priv->cm_id.ib = id;
2438
2439         return 0;
2440 }
2441
2442 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2443 {
2444         int ret;
2445         struct iw_cm_id *id;
2446
2447         id = iw_create_cm_id(id_priv->id.device,
2448                              iw_conn_req_handler,
2449                              id_priv);
2450         if (IS_ERR(id))
2451                 return PTR_ERR(id);
2452
2453         id->tos = id_priv->tos;
2454         id->tos_set = id_priv->tos_set;
2455         id_priv->cm_id.iw = id;
2456
2457         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2458                rdma_addr_size(cma_src_addr(id_priv)));
2459
2460         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2461
2462         if (ret) {
2463                 iw_destroy_cm_id(id_priv->cm_id.iw);
2464                 id_priv->cm_id.iw = NULL;
2465         }
2466
2467         return ret;
2468 }
2469
2470 static int cma_listen_handler(struct rdma_cm_id *id,
2471                               struct rdma_cm_event *event)
2472 {
2473         struct rdma_id_private *id_priv = id->context;
2474
2475         id->context = id_priv->id.context;
2476         id->event_handler = id_priv->id.event_handler;
2477         trace_cm_event_handler(id_priv, event);
2478         return id_priv->id.event_handler(id, event);
2479 }
2480
2481 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2482                               struct cma_device *cma_dev)
2483 {
2484         struct rdma_id_private *dev_id_priv;
2485         struct rdma_cm_id *id;
2486         struct net *net = id_priv->id.route.addr.dev_addr.net;
2487         int ret;
2488
2489         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2490                 return;
2491
2492         id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2493                               id_priv->id.qp_type, id_priv->res.kern_name);
2494         if (IS_ERR(id))
2495                 return;
2496
2497         dev_id_priv = container_of(id, struct rdma_id_private, id);
2498
2499         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2500         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2501                rdma_addr_size(cma_src_addr(id_priv)));
2502
2503         _cma_attach_to_dev(dev_id_priv, cma_dev);
2504         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2505         cma_id_get(id_priv);
2506         dev_id_priv->internal_id = 1;
2507         dev_id_priv->afonly = id_priv->afonly;
2508         dev_id_priv->tos_set = id_priv->tos_set;
2509         dev_id_priv->tos = id_priv->tos;
2510
2511         ret = rdma_listen(id, id_priv->backlog);
2512         if (ret)
2513                 dev_warn(&cma_dev->device->dev,
2514                          "RDMA CMA: cma_listen_on_dev, error %d\n", ret);
2515 }
2516
2517 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2518 {
2519         struct cma_device *cma_dev;
2520
2521         mutex_lock(&lock);
2522         list_add_tail(&id_priv->list, &listen_any_list);
2523         list_for_each_entry(cma_dev, &dev_list, list)
2524                 cma_listen_on_dev(id_priv, cma_dev);
2525         mutex_unlock(&lock);
2526 }
2527
2528 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2529 {
2530         struct rdma_id_private *id_priv;
2531
2532         id_priv = container_of(id, struct rdma_id_private, id);
2533         id_priv->tos = (u8) tos;
2534         id_priv->tos_set = true;
2535 }
2536 EXPORT_SYMBOL(rdma_set_service_type);
2537
2538 /**
2539  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2540  *                          with a connection identifier.
2541  * @id: Communication identifier to associated with service type.
2542  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2543  *
2544  * This function should be called before rdma_connect() on active side,
2545  * and on passive side before rdma_accept(). It is applicable to primary
2546  * path only. The timeout will affect the local side of the QP, it is not
2547  * negotiated with remote side and zero disables the timer. In case it is
2548  * set before rdma_resolve_route, the value will also be used to determine
2549  * PacketLifeTime for RoCE.
2550  *
2551  * Return: 0 for success
2552  */
2553 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2554 {
2555         struct rdma_id_private *id_priv;
2556
2557         if (id->qp_type != IB_QPT_RC)
2558                 return -EINVAL;
2559
2560         id_priv = container_of(id, struct rdma_id_private, id);
2561         id_priv->timeout = timeout;
2562         id_priv->timeout_set = true;
2563
2564         return 0;
2565 }
2566 EXPORT_SYMBOL(rdma_set_ack_timeout);
2567
2568 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2569                               void *context)
2570 {
2571         struct cma_work *work = context;
2572         struct rdma_route *route;
2573
2574         route = &work->id->id.route;
2575
2576         if (!status) {
2577                 route->num_paths = 1;
2578                 *route->path_rec = *path_rec;
2579         } else {
2580                 work->old_state = RDMA_CM_ROUTE_QUERY;
2581                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2582                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2583                 work->event.status = status;
2584                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2585                                      status);
2586         }
2587
2588         queue_work(cma_wq, &work->work);
2589 }
2590
2591 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2592                               unsigned long timeout_ms, struct cma_work *work)
2593 {
2594         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2595         struct sa_path_rec path_rec;
2596         ib_sa_comp_mask comp_mask;
2597         struct sockaddr_in6 *sin6;
2598         struct sockaddr_ib *sib;
2599
2600         memset(&path_rec, 0, sizeof path_rec);
2601
2602         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2603                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2604         else
2605                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2606         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2607         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2608         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2609         path_rec.numb_path = 1;
2610         path_rec.reversible = 1;
2611         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2612                                                   cma_dst_addr(id_priv));
2613
2614         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2615                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2616                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2617
2618         switch (cma_family(id_priv)) {
2619         case AF_INET:
2620                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2621                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2622                 break;
2623         case AF_INET6:
2624                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2625                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2626                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2627                 break;
2628         case AF_IB:
2629                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2630                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2631                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2632                 break;
2633         }
2634
2635         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2636                                                id_priv->id.port_num, &path_rec,
2637                                                comp_mask, timeout_ms,
2638                                                GFP_KERNEL, cma_query_handler,
2639                                                work, &id_priv->query);
2640
2641         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2642 }
2643
2644 static void cma_work_handler(struct work_struct *_work)
2645 {
2646         struct cma_work *work = container_of(_work, struct cma_work, work);
2647         struct rdma_id_private *id_priv = work->id;
2648         int destroy = 0;
2649
2650         mutex_lock(&id_priv->handler_mutex);
2651         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2652                 goto out;
2653
2654         if (cma_cm_event_handler(id_priv, &work->event)) {
2655                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2656                 destroy = 1;
2657         }
2658 out:
2659         mutex_unlock(&id_priv->handler_mutex);
2660         cma_id_put(id_priv);
2661         if (destroy)
2662                 rdma_destroy_id(&id_priv->id);
2663         kfree(work);
2664 }
2665
2666 static void cma_ndev_work_handler(struct work_struct *_work)
2667 {
2668         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2669         struct rdma_id_private *id_priv = work->id;
2670         int destroy = 0;
2671
2672         mutex_lock(&id_priv->handler_mutex);
2673         if (id_priv->state == RDMA_CM_DESTROYING ||
2674             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2675                 goto out;
2676
2677         if (cma_cm_event_handler(id_priv, &work->event)) {
2678                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2679                 destroy = 1;
2680         }
2681
2682 out:
2683         mutex_unlock(&id_priv->handler_mutex);
2684         cma_id_put(id_priv);
2685         if (destroy)
2686                 rdma_destroy_id(&id_priv->id);
2687         kfree(work);
2688 }
2689
2690 static void cma_init_resolve_route_work(struct cma_work *work,
2691                                         struct rdma_id_private *id_priv)
2692 {
2693         work->id = id_priv;
2694         INIT_WORK(&work->work, cma_work_handler);
2695         work->old_state = RDMA_CM_ROUTE_QUERY;
2696         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2697         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2698 }
2699
2700 static void enqueue_resolve_addr_work(struct cma_work *work,
2701                                       struct rdma_id_private *id_priv)
2702 {
2703         /* Balances with cma_id_put() in cma_work_handler */
2704         cma_id_get(id_priv);
2705
2706         work->id = id_priv;
2707         INIT_WORK(&work->work, cma_work_handler);
2708         work->old_state = RDMA_CM_ADDR_QUERY;
2709         work->new_state = RDMA_CM_ADDR_RESOLVED;
2710         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2711
2712         queue_work(cma_wq, &work->work);
2713 }
2714
2715 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2716                                 unsigned long timeout_ms)
2717 {
2718         struct rdma_route *route = &id_priv->id.route;
2719         struct cma_work *work;
2720         int ret;
2721
2722         work = kzalloc(sizeof *work, GFP_KERNEL);
2723         if (!work)
2724                 return -ENOMEM;
2725
2726         cma_init_resolve_route_work(work, id_priv);
2727
2728         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2729         if (!route->path_rec) {
2730                 ret = -ENOMEM;
2731                 goto err1;
2732         }
2733
2734         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2735         if (ret)
2736                 goto err2;
2737
2738         return 0;
2739 err2:
2740         kfree(route->path_rec);
2741         route->path_rec = NULL;
2742 err1:
2743         kfree(work);
2744         return ret;
2745 }
2746
2747 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2748                                            unsigned long supported_gids,
2749                                            enum ib_gid_type default_gid)
2750 {
2751         if ((network_type == RDMA_NETWORK_IPV4 ||
2752              network_type == RDMA_NETWORK_IPV6) &&
2753             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2754                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2755
2756         return default_gid;
2757 }
2758
2759 /*
2760  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2761  * path record type based on GID type.
2762  * It also sets up other L2 fields which includes destination mac address
2763  * netdev ifindex, of the path record.
2764  * It returns the netdev of the bound interface for this path record entry.
2765  */
2766 static struct net_device *
2767 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2768 {
2769         struct rdma_route *route = &id_priv->id.route;
2770         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2771         struct rdma_addr *addr = &route->addr;
2772         unsigned long supported_gids;
2773         struct net_device *ndev;
2774
2775         if (!addr->dev_addr.bound_dev_if)
2776                 return NULL;
2777
2778         ndev = dev_get_by_index(addr->dev_addr.net,
2779                                 addr->dev_addr.bound_dev_if);
2780         if (!ndev)
2781                 return NULL;
2782
2783         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2784                                                     id_priv->id.port_num);
2785         gid_type = cma_route_gid_type(addr->dev_addr.network,
2786                                       supported_gids,
2787                                       id_priv->gid_type);
2788         /* Use the hint from IP Stack to select GID Type */
2789         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2790                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2791         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2792
2793         route->path_rec->roce.route_resolved = true;
2794         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2795         return ndev;
2796 }
2797
2798 int rdma_set_ib_path(struct rdma_cm_id *id,
2799                      struct sa_path_rec *path_rec)
2800 {
2801         struct rdma_id_private *id_priv;
2802         struct net_device *ndev;
2803         int ret;
2804
2805         id_priv = container_of(id, struct rdma_id_private, id);
2806         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2807                            RDMA_CM_ROUTE_RESOLVED))
2808                 return -EINVAL;
2809
2810         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2811                                      GFP_KERNEL);
2812         if (!id->route.path_rec) {
2813                 ret = -ENOMEM;
2814                 goto err;
2815         }
2816
2817         if (rdma_protocol_roce(id->device, id->port_num)) {
2818                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2819                 if (!ndev) {
2820                         ret = -ENODEV;
2821                         goto err_free;
2822                 }
2823                 dev_put(ndev);
2824         }
2825
2826         id->route.num_paths = 1;
2827         return 0;
2828
2829 err_free:
2830         kfree(id->route.path_rec);
2831         id->route.path_rec = NULL;
2832 err:
2833         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2834         return ret;
2835 }
2836 EXPORT_SYMBOL(rdma_set_ib_path);
2837
2838 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2839 {
2840         struct cma_work *work;
2841
2842         work = kzalloc(sizeof *work, GFP_KERNEL);
2843         if (!work)
2844                 return -ENOMEM;
2845
2846         cma_init_resolve_route_work(work, id_priv);
2847         queue_work(cma_wq, &work->work);
2848         return 0;
2849 }
2850
2851 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2852 {
2853         struct net_device *dev;
2854
2855         dev = vlan_dev_real_dev(vlan_ndev);
2856         if (dev->num_tc)
2857                 return netdev_get_prio_tc_map(dev, prio);
2858
2859         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2860                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2861 }
2862
2863 struct iboe_prio_tc_map {
2864         int input_prio;
2865         int output_tc;
2866         bool found;
2867 };
2868
2869 static int get_lower_vlan_dev_tc(struct net_device *dev, void *data)
2870 {
2871         struct iboe_prio_tc_map *map = data;
2872
2873         if (is_vlan_dev(dev))
2874                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2875         else if (dev->num_tc)
2876                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2877         else
2878                 map->output_tc = 0;
2879         /* We are interested only in first level VLAN device, so always
2880          * return 1 to stop iterating over next level devices.
2881          */
2882         map->found = true;
2883         return 1;
2884 }
2885
2886 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2887 {
2888         struct iboe_prio_tc_map prio_tc_map = {};
2889         int prio = rt_tos2priority(tos);
2890
2891         /* If VLAN device, get it directly from the VLAN netdev */
2892         if (is_vlan_dev(ndev))
2893                 return get_vlan_ndev_tc(ndev, prio);
2894
2895         prio_tc_map.input_prio = prio;
2896         rcu_read_lock();
2897         netdev_walk_all_lower_dev_rcu(ndev,
2898                                       get_lower_vlan_dev_tc,
2899                                       &prio_tc_map);
2900         rcu_read_unlock();
2901         /* If map is found from lower device, use it; Otherwise
2902          * continue with the current netdevice to get priority to tc map.
2903          */
2904         if (prio_tc_map.found)
2905                 return prio_tc_map.output_tc;
2906         else if (ndev->num_tc)
2907                 return netdev_get_prio_tc_map(ndev, prio);
2908         else
2909                 return 0;
2910 }
2911
2912 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
2913 {
2914         struct sockaddr_in6 *addr6;
2915         u16 dport, sport;
2916         u32 hash, fl;
2917
2918         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
2919         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
2920         if ((cma_family(id_priv) != AF_INET6) || !fl) {
2921                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
2922                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
2923                 hash = (u32)sport * 31 + dport;
2924                 fl = hash & IB_GRH_FLOWLABEL_MASK;
2925         }
2926
2927         return cpu_to_be32(fl);
2928 }
2929
2930 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2931 {
2932         struct rdma_route *route = &id_priv->id.route;
2933         struct rdma_addr *addr = &route->addr;
2934         struct cma_work *work;
2935         int ret;
2936         struct net_device *ndev;
2937
2938         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2939                                         rdma_start_port(id_priv->cma_dev->device)];
2940         u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2941
2942
2943         work = kzalloc(sizeof *work, GFP_KERNEL);
2944         if (!work)
2945                 return -ENOMEM;
2946
2947         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2948         if (!route->path_rec) {
2949                 ret = -ENOMEM;
2950                 goto err1;
2951         }
2952
2953         route->num_paths = 1;
2954
2955         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2956         if (!ndev) {
2957                 ret = -ENODEV;
2958                 goto err2;
2959         }
2960
2961         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2962                     &route->path_rec->sgid);
2963         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2964                     &route->path_rec->dgid);
2965
2966         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2967                 /* TODO: get the hoplimit from the inet/inet6 device */
2968                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2969         else
2970                 route->path_rec->hop_limit = 1;
2971         route->path_rec->reversible = 1;
2972         route->path_rec->pkey = cpu_to_be16(0xffff);
2973         route->path_rec->mtu_selector = IB_SA_EQ;
2974         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2975         route->path_rec->traffic_class = tos;
2976         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2977         route->path_rec->rate_selector = IB_SA_EQ;
2978         route->path_rec->rate = iboe_get_rate(ndev);
2979         dev_put(ndev);
2980         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2981         /* In case ACK timeout is set, use this value to calculate
2982          * PacketLifeTime.  As per IBTA 12.7.34,
2983          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
2984          * Assuming a negligible local ACK delay, we can use
2985          * PacketLifeTime = local ACK timeout/2
2986          * as a reasonable approximation for RoCE networks.
2987          */
2988         route->path_rec->packet_life_time = id_priv->timeout_set ?
2989                 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
2990
2991         if (!route->path_rec->mtu) {
2992                 ret = -EINVAL;
2993                 goto err2;
2994         }
2995
2996         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
2997                                          id_priv->id.port_num))
2998                 route->path_rec->flow_label =
2999                         cma_get_roce_udp_flow_label(id_priv);
3000
3001         cma_init_resolve_route_work(work, id_priv);
3002         queue_work(cma_wq, &work->work);
3003
3004         return 0;
3005
3006 err2:
3007         kfree(route->path_rec);
3008         route->path_rec = NULL;
3009         route->num_paths = 0;
3010 err1:
3011         kfree(work);
3012         return ret;
3013 }
3014
3015 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3016 {
3017         struct rdma_id_private *id_priv;
3018         int ret;
3019
3020         id_priv = container_of(id, struct rdma_id_private, id);
3021         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3022                 return -EINVAL;
3023
3024         cma_id_get(id_priv);
3025         if (rdma_cap_ib_sa(id->device, id->port_num))
3026                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3027         else if (rdma_protocol_roce(id->device, id->port_num))
3028                 ret = cma_resolve_iboe_route(id_priv);
3029         else if (rdma_protocol_iwarp(id->device, id->port_num))
3030                 ret = cma_resolve_iw_route(id_priv);
3031         else
3032                 ret = -ENOSYS;
3033
3034         if (ret)
3035                 goto err;
3036
3037         return 0;
3038 err:
3039         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3040         cma_id_put(id_priv);
3041         return ret;
3042 }
3043 EXPORT_SYMBOL(rdma_resolve_route);
3044
3045 static void cma_set_loopback(struct sockaddr *addr)
3046 {
3047         switch (addr->sa_family) {
3048         case AF_INET:
3049                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3050                 break;
3051         case AF_INET6:
3052                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3053                               0, 0, 0, htonl(1));
3054                 break;
3055         default:
3056                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3057                             0, 0, 0, htonl(1));
3058                 break;
3059         }
3060 }
3061
3062 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3063 {
3064         struct cma_device *cma_dev, *cur_dev;
3065         union ib_gid gid;
3066         enum ib_port_state port_state;
3067         unsigned int p;
3068         u16 pkey;
3069         int ret;
3070
3071         cma_dev = NULL;
3072         mutex_lock(&lock);
3073         list_for_each_entry(cur_dev, &dev_list, list) {
3074                 if (cma_family(id_priv) == AF_IB &&
3075                     !rdma_cap_ib_cm(cur_dev->device, 1))
3076                         continue;
3077
3078                 if (!cma_dev)
3079                         cma_dev = cur_dev;
3080
3081                 rdma_for_each_port (cur_dev->device, p) {
3082                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3083                             port_state == IB_PORT_ACTIVE) {
3084                                 cma_dev = cur_dev;
3085                                 goto port_found;
3086                         }
3087                 }
3088         }
3089
3090         if (!cma_dev) {
3091                 ret = -ENODEV;
3092                 goto out;
3093         }
3094
3095         p = 1;
3096
3097 port_found:
3098         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3099         if (ret)
3100                 goto out;
3101
3102         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3103         if (ret)
3104                 goto out;
3105
3106         id_priv->id.route.addr.dev_addr.dev_type =
3107                 (rdma_protocol_ib(cma_dev->device, p)) ?
3108                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3109
3110         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3111         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3112         id_priv->id.port_num = p;
3113         cma_attach_to_dev(id_priv, cma_dev);
3114         cma_set_loopback(cma_src_addr(id_priv));
3115 out:
3116         mutex_unlock(&lock);
3117         return ret;
3118 }
3119
3120 static void addr_handler(int status, struct sockaddr *src_addr,
3121                          struct rdma_dev_addr *dev_addr, void *context)
3122 {
3123         struct rdma_id_private *id_priv = context;
3124         struct rdma_cm_event event = {};
3125         struct sockaddr *addr;
3126         struct sockaddr_storage old_addr;
3127
3128         mutex_lock(&id_priv->handler_mutex);
3129         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3130                            RDMA_CM_ADDR_RESOLVED))
3131                 goto out;
3132
3133         /*
3134          * Store the previous src address, so that if we fail to acquire
3135          * matching rdma device, old address can be restored back, which helps
3136          * to cancel the cma listen operation correctly.
3137          */
3138         addr = cma_src_addr(id_priv);
3139         memcpy(&old_addr, addr, rdma_addr_size(addr));
3140         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3141         if (!status && !id_priv->cma_dev) {
3142                 status = cma_acquire_dev_by_src_ip(id_priv);
3143                 if (status)
3144                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3145                                              status);
3146         } else if (status) {
3147                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3148         }
3149
3150         if (status) {
3151                 memcpy(addr, &old_addr,
3152                        rdma_addr_size((struct sockaddr *)&old_addr));
3153                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3154                                    RDMA_CM_ADDR_BOUND))
3155                         goto out;
3156                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3157                 event.status = status;
3158         } else
3159                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3160
3161         if (cma_cm_event_handler(id_priv, &event)) {
3162                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3163                 mutex_unlock(&id_priv->handler_mutex);
3164                 rdma_destroy_id(&id_priv->id);
3165                 return;
3166         }
3167 out:
3168         mutex_unlock(&id_priv->handler_mutex);
3169 }
3170
3171 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3172 {
3173         struct cma_work *work;
3174         union ib_gid gid;
3175         int ret;
3176
3177         work = kzalloc(sizeof *work, GFP_KERNEL);
3178         if (!work)
3179                 return -ENOMEM;
3180
3181         if (!id_priv->cma_dev) {
3182                 ret = cma_bind_loopback(id_priv);
3183                 if (ret)
3184                         goto err;
3185         }
3186
3187         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3188         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3189
3190         enqueue_resolve_addr_work(work, id_priv);
3191         return 0;
3192 err:
3193         kfree(work);
3194         return ret;
3195 }
3196
3197 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3198 {
3199         struct cma_work *work;
3200         int ret;
3201
3202         work = kzalloc(sizeof *work, GFP_KERNEL);
3203         if (!work)
3204                 return -ENOMEM;
3205
3206         if (!id_priv->cma_dev) {
3207                 ret = cma_resolve_ib_dev(id_priv);
3208                 if (ret)
3209                         goto err;
3210         }
3211
3212         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3213                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3214
3215         enqueue_resolve_addr_work(work, id_priv);
3216         return 0;
3217 err:
3218         kfree(work);
3219         return ret;
3220 }
3221
3222 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3223                          const struct sockaddr *dst_addr)
3224 {
3225         if (!src_addr || !src_addr->sa_family) {
3226                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3227                 src_addr->sa_family = dst_addr->sa_family;
3228                 if (IS_ENABLED(CONFIG_IPV6) &&
3229                     dst_addr->sa_family == AF_INET6) {
3230                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3231                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3232                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3233                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3234                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3235                 } else if (dst_addr->sa_family == AF_IB) {
3236                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3237                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3238                 }
3239         }
3240         return rdma_bind_addr(id, src_addr);
3241 }
3242
3243 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3244                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3245 {
3246         struct rdma_id_private *id_priv;
3247         int ret;
3248
3249         id_priv = container_of(id, struct rdma_id_private, id);
3250         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3251         if (id_priv->state == RDMA_CM_IDLE) {
3252                 ret = cma_bind_addr(id, src_addr, dst_addr);
3253                 if (ret) {
3254                         memset(cma_dst_addr(id_priv), 0,
3255                                rdma_addr_size(dst_addr));
3256                         return ret;
3257                 }
3258         }
3259
3260         if (cma_family(id_priv) != dst_addr->sa_family) {
3261                 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3262                 return -EINVAL;
3263         }
3264
3265         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3266                 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3267                 return -EINVAL;
3268         }
3269
3270         if (cma_any_addr(dst_addr)) {
3271                 ret = cma_resolve_loopback(id_priv);
3272         } else {
3273                 if (dst_addr->sa_family == AF_IB) {
3274                         ret = cma_resolve_ib_addr(id_priv);
3275                 } else {
3276                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3277                                               &id->route.addr.dev_addr,
3278                                               timeout_ms, addr_handler,
3279                                               false, id_priv);
3280                 }
3281         }
3282         if (ret)
3283                 goto err;
3284
3285         return 0;
3286 err:
3287         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3288         return ret;
3289 }
3290 EXPORT_SYMBOL(rdma_resolve_addr);
3291
3292 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3293 {
3294         struct rdma_id_private *id_priv;
3295         unsigned long flags;
3296         int ret;
3297
3298         id_priv = container_of(id, struct rdma_id_private, id);
3299         spin_lock_irqsave(&id_priv->lock, flags);
3300         if (reuse || id_priv->state == RDMA_CM_IDLE) {
3301                 id_priv->reuseaddr = reuse;
3302                 ret = 0;
3303         } else {
3304                 ret = -EINVAL;
3305         }
3306         spin_unlock_irqrestore(&id_priv->lock, flags);
3307         return ret;
3308 }
3309 EXPORT_SYMBOL(rdma_set_reuseaddr);
3310
3311 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3312 {
3313         struct rdma_id_private *id_priv;
3314         unsigned long flags;
3315         int ret;
3316
3317         id_priv = container_of(id, struct rdma_id_private, id);
3318         spin_lock_irqsave(&id_priv->lock, flags);
3319         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3320                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3321                 id_priv->afonly = afonly;
3322                 ret = 0;
3323         } else {
3324                 ret = -EINVAL;
3325         }
3326         spin_unlock_irqrestore(&id_priv->lock, flags);
3327         return ret;
3328 }
3329 EXPORT_SYMBOL(rdma_set_afonly);
3330
3331 static void cma_bind_port(struct rdma_bind_list *bind_list,
3332                           struct rdma_id_private *id_priv)
3333 {
3334         struct sockaddr *addr;
3335         struct sockaddr_ib *sib;
3336         u64 sid, mask;
3337         __be16 port;
3338
3339         addr = cma_src_addr(id_priv);
3340         port = htons(bind_list->port);
3341
3342         switch (addr->sa_family) {
3343         case AF_INET:
3344                 ((struct sockaddr_in *) addr)->sin_port = port;
3345                 break;
3346         case AF_INET6:
3347                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3348                 break;
3349         case AF_IB:
3350                 sib = (struct sockaddr_ib *) addr;
3351                 sid = be64_to_cpu(sib->sib_sid);
3352                 mask = be64_to_cpu(sib->sib_sid_mask);
3353                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3354                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3355                 break;
3356         }
3357         id_priv->bind_list = bind_list;
3358         hlist_add_head(&id_priv->node, &bind_list->owners);
3359 }
3360
3361 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3362                           struct rdma_id_private *id_priv, unsigned short snum)
3363 {
3364         struct rdma_bind_list *bind_list;
3365         int ret;
3366
3367         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3368         if (!bind_list)
3369                 return -ENOMEM;
3370
3371         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3372                            snum);
3373         if (ret < 0)
3374                 goto err;
3375
3376         bind_list->ps = ps;
3377         bind_list->port = snum;
3378         cma_bind_port(bind_list, id_priv);
3379         return 0;
3380 err:
3381         kfree(bind_list);
3382         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3383 }
3384
3385 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3386                               struct rdma_id_private *id_priv)
3387 {
3388         struct rdma_id_private *cur_id;
3389         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3390         struct sockaddr  *saddr = cma_src_addr(id_priv);
3391         __be16 dport = cma_port(daddr);
3392
3393         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3394                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3395                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3396                 __be16 cur_dport = cma_port(cur_daddr);
3397
3398                 if (id_priv == cur_id)
3399                         continue;
3400
3401                 /* different dest port -> unique */
3402                 if (!cma_any_port(daddr) &&
3403                     !cma_any_port(cur_daddr) &&
3404                     (dport != cur_dport))
3405                         continue;
3406
3407                 /* different src address -> unique */
3408                 if (!cma_any_addr(saddr) &&
3409                     !cma_any_addr(cur_saddr) &&
3410                     cma_addr_cmp(saddr, cur_saddr))
3411                         continue;
3412
3413                 /* different dst address -> unique */
3414                 if (!cma_any_addr(daddr) &&
3415                     !cma_any_addr(cur_daddr) &&
3416                     cma_addr_cmp(daddr, cur_daddr))
3417                         continue;
3418
3419                 return -EADDRNOTAVAIL;
3420         }
3421         return 0;
3422 }
3423
3424 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3425                               struct rdma_id_private *id_priv)
3426 {
3427         static unsigned int last_used_port;
3428         int low, high, remaining;
3429         unsigned int rover;
3430         struct net *net = id_priv->id.route.addr.dev_addr.net;
3431
3432         inet_get_local_port_range(net, &low, &high);
3433         remaining = (high - low) + 1;
3434         rover = prandom_u32() % remaining + low;
3435 retry:
3436         if (last_used_port != rover) {
3437                 struct rdma_bind_list *bind_list;
3438                 int ret;
3439
3440                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3441
3442                 if (!bind_list) {
3443                         ret = cma_alloc_port(ps, id_priv, rover);
3444                 } else {
3445                         ret = cma_port_is_unique(bind_list, id_priv);
3446                         if (!ret)
3447                                 cma_bind_port(bind_list, id_priv);
3448                 }
3449                 /*
3450                  * Remember previously used port number in order to avoid
3451                  * re-using same port immediately after it is closed.
3452                  */
3453                 if (!ret)
3454                         last_used_port = rover;
3455                 if (ret != -EADDRNOTAVAIL)
3456                         return ret;
3457         }
3458         if (--remaining) {
3459                 rover++;
3460                 if ((rover < low) || (rover > high))
3461                         rover = low;
3462                 goto retry;
3463         }
3464         return -EADDRNOTAVAIL;
3465 }
3466
3467 /*
3468  * Check that the requested port is available.  This is called when trying to
3469  * bind to a specific port, or when trying to listen on a bound port.  In
3470  * the latter case, the provided id_priv may already be on the bind_list, but
3471  * we still need to check that it's okay to start listening.
3472  */
3473 static int cma_check_port(struct rdma_bind_list *bind_list,
3474                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3475 {
3476         struct rdma_id_private *cur_id;
3477         struct sockaddr *addr, *cur_addr;
3478
3479         addr = cma_src_addr(id_priv);
3480         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3481                 if (id_priv == cur_id)
3482                         continue;
3483
3484                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3485                     cur_id->reuseaddr)
3486                         continue;
3487
3488                 cur_addr = cma_src_addr(cur_id);
3489                 if (id_priv->afonly && cur_id->afonly &&
3490                     (addr->sa_family != cur_addr->sa_family))
3491                         continue;
3492
3493                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3494                         return -EADDRNOTAVAIL;
3495
3496                 if (!cma_addr_cmp(addr, cur_addr))
3497                         return -EADDRINUSE;
3498         }
3499         return 0;
3500 }
3501
3502 static int cma_use_port(enum rdma_ucm_port_space ps,
3503                         struct rdma_id_private *id_priv)
3504 {
3505         struct rdma_bind_list *bind_list;
3506         unsigned short snum;
3507         int ret;
3508
3509         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3510         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3511                 return -EACCES;
3512
3513         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3514         if (!bind_list) {
3515                 ret = cma_alloc_port(ps, id_priv, snum);
3516         } else {
3517                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3518                 if (!ret)
3519                         cma_bind_port(bind_list, id_priv);
3520         }
3521         return ret;
3522 }
3523
3524 static int cma_bind_listen(struct rdma_id_private *id_priv)
3525 {
3526         struct rdma_bind_list *bind_list = id_priv->bind_list;
3527         int ret = 0;
3528
3529         mutex_lock(&lock);
3530         if (bind_list->owners.first->next)
3531                 ret = cma_check_port(bind_list, id_priv, 0);
3532         mutex_unlock(&lock);
3533         return ret;
3534 }
3535
3536 static enum rdma_ucm_port_space
3537 cma_select_inet_ps(struct rdma_id_private *id_priv)
3538 {
3539         switch (id_priv->id.ps) {
3540         case RDMA_PS_TCP:
3541         case RDMA_PS_UDP:
3542         case RDMA_PS_IPOIB:
3543         case RDMA_PS_IB:
3544                 return id_priv->id.ps;
3545         default:
3546
3547                 return 0;
3548         }
3549 }
3550
3551 static enum rdma_ucm_port_space
3552 cma_select_ib_ps(struct rdma_id_private *id_priv)
3553 {
3554         enum rdma_ucm_port_space ps = 0;
3555         struct sockaddr_ib *sib;
3556         u64 sid_ps, mask, sid;
3557
3558         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3559         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3560         sid = be64_to_cpu(sib->sib_sid) & mask;
3561
3562         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3563                 sid_ps = RDMA_IB_IP_PS_IB;
3564                 ps = RDMA_PS_IB;
3565         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3566                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3567                 sid_ps = RDMA_IB_IP_PS_TCP;
3568                 ps = RDMA_PS_TCP;
3569         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3570                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3571                 sid_ps = RDMA_IB_IP_PS_UDP;
3572                 ps = RDMA_PS_UDP;
3573         }
3574
3575         if (ps) {
3576                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3577                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3578                                                 be64_to_cpu(sib->sib_sid_mask));
3579         }
3580         return ps;
3581 }
3582
3583 static int cma_get_port(struct rdma_id_private *id_priv)
3584 {
3585         enum rdma_ucm_port_space ps;
3586         int ret;
3587
3588         if (cma_family(id_priv) != AF_IB)
3589                 ps = cma_select_inet_ps(id_priv);
3590         else
3591                 ps = cma_select_ib_ps(id_priv);
3592         if (!ps)
3593                 return -EPROTONOSUPPORT;
3594
3595         mutex_lock(&lock);
3596         if (cma_any_port(cma_src_addr(id_priv)))
3597                 ret = cma_alloc_any_port(ps, id_priv);
3598         else
3599                 ret = cma_use_port(ps, id_priv);
3600         mutex_unlock(&lock);
3601
3602         return ret;
3603 }
3604
3605 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3606                                struct sockaddr *addr)
3607 {
3608 #if IS_ENABLED(CONFIG_IPV6)
3609         struct sockaddr_in6 *sin6;
3610
3611         if (addr->sa_family != AF_INET6)
3612                 return 0;
3613
3614         sin6 = (struct sockaddr_in6 *) addr;
3615
3616         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3617                 return 0;
3618
3619         if (!sin6->sin6_scope_id)
3620                         return -EINVAL;
3621
3622         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3623 #endif
3624         return 0;
3625 }
3626
3627 int rdma_listen(struct rdma_cm_id *id, int backlog)
3628 {
3629         struct rdma_id_private *id_priv;
3630         int ret;
3631
3632         id_priv = container_of(id, struct rdma_id_private, id);
3633         if (id_priv->state == RDMA_CM_IDLE) {
3634                 id->route.addr.src_addr.ss_family = AF_INET;
3635                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3636                 if (ret)
3637                         return ret;
3638         }
3639
3640         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3641                 return -EINVAL;
3642
3643         if (id_priv->reuseaddr) {
3644                 ret = cma_bind_listen(id_priv);
3645                 if (ret)
3646                         goto err;
3647         }
3648
3649         id_priv->backlog = backlog;
3650         if (id->device) {
3651                 if (rdma_cap_ib_cm(id->device, 1)) {
3652                         ret = cma_ib_listen(id_priv);
3653                         if (ret)
3654                                 goto err;
3655                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3656                         ret = cma_iw_listen(id_priv, backlog);
3657                         if (ret)
3658                                 goto err;
3659                 } else {
3660                         ret = -ENOSYS;
3661                         goto err;
3662                 }
3663         } else
3664                 cma_listen_on_all(id_priv);
3665
3666         return 0;
3667 err:
3668         id_priv->backlog = 0;
3669         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3670         return ret;
3671 }
3672 EXPORT_SYMBOL(rdma_listen);
3673
3674 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3675 {
3676         struct rdma_id_private *id_priv;
3677         int ret;
3678         struct sockaddr  *daddr;
3679
3680         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3681             addr->sa_family != AF_IB)
3682                 return -EAFNOSUPPORT;
3683
3684         id_priv = container_of(id, struct rdma_id_private, id);
3685         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3686                 return -EINVAL;
3687
3688         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3689         if (ret)
3690                 goto err1;
3691
3692         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3693         if (!cma_any_addr(addr)) {
3694                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3695                 if (ret)
3696                         goto err1;
3697
3698                 ret = cma_acquire_dev_by_src_ip(id_priv);
3699                 if (ret)
3700                         goto err1;
3701         }
3702
3703         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3704                 if (addr->sa_family == AF_INET)
3705                         id_priv->afonly = 1;
3706 #if IS_ENABLED(CONFIG_IPV6)
3707                 else if (addr->sa_family == AF_INET6) {
3708                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3709
3710                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3711                 }
3712 #endif
3713         }
3714         daddr = cma_dst_addr(id_priv);
3715         daddr->sa_family = addr->sa_family;
3716
3717         ret = cma_get_port(id_priv);
3718         if (ret)
3719                 goto err2;
3720
3721         return 0;
3722 err2:
3723         rdma_restrack_del(&id_priv->res);
3724         if (id_priv->cma_dev)
3725                 cma_release_dev(id_priv);
3726 err1:
3727         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3728         return ret;
3729 }
3730 EXPORT_SYMBOL(rdma_bind_addr);
3731
3732 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3733 {
3734         struct cma_hdr *cma_hdr;
3735
3736         cma_hdr = hdr;
3737         cma_hdr->cma_version = CMA_VERSION;
3738         if (cma_family(id_priv) == AF_INET) {
3739                 struct sockaddr_in *src4, *dst4;
3740
3741                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3742                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3743
3744                 cma_set_ip_ver(cma_hdr, 4);
3745                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3746                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3747                 cma_hdr->port = src4->sin_port;
3748         } else if (cma_family(id_priv) == AF_INET6) {
3749                 struct sockaddr_in6 *src6, *dst6;
3750
3751                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3752                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3753
3754                 cma_set_ip_ver(cma_hdr, 6);
3755                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3756                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3757                 cma_hdr->port = src6->sin6_port;
3758         }
3759         return 0;
3760 }
3761
3762 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3763                                 const struct ib_cm_event *ib_event)
3764 {
3765         struct rdma_id_private *id_priv = cm_id->context;
3766         struct rdma_cm_event event = {};
3767         const struct ib_cm_sidr_rep_event_param *rep =
3768                                 &ib_event->param.sidr_rep_rcvd;
3769         int ret = 0;
3770
3771         mutex_lock(&id_priv->handler_mutex);
3772         if (id_priv->state != RDMA_CM_CONNECT)
3773                 goto out;
3774
3775         switch (ib_event->event) {
3776         case IB_CM_SIDR_REQ_ERROR:
3777                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3778                 event.status = -ETIMEDOUT;
3779                 break;
3780         case IB_CM_SIDR_REP_RECEIVED:
3781                 event.param.ud.private_data = ib_event->private_data;
3782                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3783                 if (rep->status != IB_SIDR_SUCCESS) {
3784                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3785                         event.status = ib_event->param.sidr_rep_rcvd.status;
3786                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3787                                              event.status);
3788                         break;
3789                 }
3790                 ret = cma_set_qkey(id_priv, rep->qkey);
3791                 if (ret) {
3792                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3793                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3794                         event.status = ret;
3795                         break;
3796                 }
3797                 ib_init_ah_attr_from_path(id_priv->id.device,
3798                                           id_priv->id.port_num,
3799                                           id_priv->id.route.path_rec,
3800                                           &event.param.ud.ah_attr,
3801                                           rep->sgid_attr);
3802                 event.param.ud.qp_num = rep->qpn;
3803                 event.param.ud.qkey = rep->qkey;
3804                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3805                 event.status = 0;
3806                 break;
3807         default:
3808                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3809                        ib_event->event);
3810                 goto out;
3811         }
3812
3813         ret = cma_cm_event_handler(id_priv, &event);
3814
3815         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3816         if (ret) {
3817                 /* Destroy the CM ID by returning a non-zero value. */
3818                 id_priv->cm_id.ib = NULL;
3819                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3820                 mutex_unlock(&id_priv->handler_mutex);
3821                 rdma_destroy_id(&id_priv->id);
3822                 return ret;
3823         }
3824 out:
3825         mutex_unlock(&id_priv->handler_mutex);
3826         return ret;
3827 }
3828
3829 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3830                               struct rdma_conn_param *conn_param)
3831 {
3832         struct ib_cm_sidr_req_param req;
3833         struct ib_cm_id *id;
3834         void *private_data;
3835         u8 offset;
3836         int ret;
3837
3838         memset(&req, 0, sizeof req);
3839         offset = cma_user_data_offset(id_priv);
3840         req.private_data_len = offset + conn_param->private_data_len;
3841         if (req.private_data_len < conn_param->private_data_len)
3842                 return -EINVAL;
3843
3844         if (req.private_data_len) {
3845                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3846                 if (!private_data)
3847                         return -ENOMEM;
3848         } else {
3849                 private_data = NULL;
3850         }
3851
3852         if (conn_param->private_data && conn_param->private_data_len)
3853                 memcpy(private_data + offset, conn_param->private_data,
3854                        conn_param->private_data_len);
3855
3856         if (private_data) {
3857                 ret = cma_format_hdr(private_data, id_priv);
3858                 if (ret)
3859                         goto out;
3860                 req.private_data = private_data;
3861         }
3862
3863         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3864                              id_priv);
3865         if (IS_ERR(id)) {
3866                 ret = PTR_ERR(id);
3867                 goto out;
3868         }
3869         id_priv->cm_id.ib = id;
3870
3871         req.path = id_priv->id.route.path_rec;
3872         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3873         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3874         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3875         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3876
3877         trace_cm_send_sidr_req(id_priv);
3878         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3879         if (ret) {
3880                 ib_destroy_cm_id(id_priv->cm_id.ib);
3881                 id_priv->cm_id.ib = NULL;
3882         }
3883 out:
3884         kfree(private_data);
3885         return ret;
3886 }
3887
3888 static int cma_connect_ib(struct rdma_id_private *id_priv,
3889                           struct rdma_conn_param *conn_param)
3890 {
3891         struct ib_cm_req_param req;
3892         struct rdma_route *route;
3893         void *private_data;
3894         struct ib_cm_id *id;
3895         u8 offset;
3896         int ret;
3897
3898         memset(&req, 0, sizeof req);
3899         offset = cma_user_data_offset(id_priv);
3900         req.private_data_len = offset + conn_param->private_data_len;
3901         if (req.private_data_len < conn_param->private_data_len)
3902                 return -EINVAL;
3903
3904         if (req.private_data_len) {
3905                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3906                 if (!private_data)
3907                         return -ENOMEM;
3908         } else {
3909                 private_data = NULL;
3910         }
3911
3912         if (conn_param->private_data && conn_param->private_data_len)
3913                 memcpy(private_data + offset, conn_param->private_data,
3914                        conn_param->private_data_len);
3915
3916         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3917         if (IS_ERR(id)) {
3918                 ret = PTR_ERR(id);
3919                 goto out;
3920         }
3921         id_priv->cm_id.ib = id;
3922
3923         route = &id_priv->id.route;
3924         if (private_data) {
3925                 ret = cma_format_hdr(private_data, id_priv);
3926                 if (ret)
3927                         goto out;
3928                 req.private_data = private_data;
3929         }
3930
3931         req.primary_path = &route->path_rec[0];
3932         if (route->num_paths == 2)
3933                 req.alternate_path = &route->path_rec[1];
3934
3935         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3936         /* Alternate path SGID attribute currently unsupported */
3937         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3938         req.qp_num = id_priv->qp_num;
3939         req.qp_type = id_priv->id.qp_type;
3940         req.starting_psn = id_priv->seq_num;
3941         req.responder_resources = conn_param->responder_resources;
3942         req.initiator_depth = conn_param->initiator_depth;
3943         req.flow_control = conn_param->flow_control;
3944         req.retry_count = min_t(u8, 7, conn_param->retry_count);
3945         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3946         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3947         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3948         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3949         req.srq = id_priv->srq ? 1 : 0;
3950
3951         trace_cm_send_req(id_priv);
3952         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3953 out:
3954         if (ret && !IS_ERR(id)) {
3955                 ib_destroy_cm_id(id);
3956                 id_priv->cm_id.ib = NULL;
3957         }
3958
3959         kfree(private_data);
3960         return ret;
3961 }
3962
3963 static int cma_connect_iw(struct rdma_id_private *id_priv,
3964                           struct rdma_conn_param *conn_param)
3965 {
3966         struct iw_cm_id *cm_id;
3967         int ret;
3968         struct iw_cm_conn_param iw_param;
3969
3970         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3971         if (IS_ERR(cm_id))
3972                 return PTR_ERR(cm_id);
3973
3974         cm_id->tos = id_priv->tos;
3975         cm_id->tos_set = id_priv->tos_set;
3976         id_priv->cm_id.iw = cm_id;
3977
3978         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3979                rdma_addr_size(cma_src_addr(id_priv)));
3980         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3981                rdma_addr_size(cma_dst_addr(id_priv)));
3982
3983         ret = cma_modify_qp_rtr(id_priv, conn_param);
3984         if (ret)
3985                 goto out;
3986
3987         if (conn_param) {
3988                 iw_param.ord = conn_param->initiator_depth;
3989                 iw_param.ird = conn_param->responder_resources;
3990                 iw_param.private_data = conn_param->private_data;
3991                 iw_param.private_data_len = conn_param->private_data_len;
3992                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3993         } else {
3994                 memset(&iw_param, 0, sizeof iw_param);
3995                 iw_param.qpn = id_priv->qp_num;
3996         }
3997         ret = iw_cm_connect(cm_id, &iw_param);
3998 out:
3999         if (ret) {
4000                 iw_destroy_cm_id(cm_id);
4001                 id_priv->cm_id.iw = NULL;
4002         }
4003         return ret;
4004 }
4005
4006 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4007 {
4008         struct rdma_id_private *id_priv;
4009         int ret;
4010
4011         id_priv = container_of(id, struct rdma_id_private, id);
4012         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4013                 return -EINVAL;
4014
4015         if (!id->qp) {
4016                 id_priv->qp_num = conn_param->qp_num;
4017                 id_priv->srq = conn_param->srq;
4018         }
4019
4020         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4021                 if (id->qp_type == IB_QPT_UD)
4022                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4023                 else
4024                         ret = cma_connect_ib(id_priv, conn_param);
4025         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4026                 ret = cma_connect_iw(id_priv, conn_param);
4027         else
4028                 ret = -ENOSYS;
4029         if (ret)
4030                 goto err;
4031
4032         return 0;
4033 err:
4034         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4035         return ret;
4036 }
4037 EXPORT_SYMBOL(rdma_connect);
4038
4039 static int cma_accept_ib(struct rdma_id_private *id_priv,
4040                          struct rdma_conn_param *conn_param)
4041 {
4042         struct ib_cm_rep_param rep;
4043         int ret;
4044
4045         ret = cma_modify_qp_rtr(id_priv, conn_param);
4046         if (ret)
4047                 goto out;
4048
4049         ret = cma_modify_qp_rts(id_priv, conn_param);
4050         if (ret)
4051                 goto out;
4052
4053         memset(&rep, 0, sizeof rep);
4054         rep.qp_num = id_priv->qp_num;
4055         rep.starting_psn = id_priv->seq_num;
4056         rep.private_data = conn_param->private_data;
4057         rep.private_data_len = conn_param->private_data_len;
4058         rep.responder_resources = conn_param->responder_resources;
4059         rep.initiator_depth = conn_param->initiator_depth;
4060         rep.failover_accepted = 0;
4061         rep.flow_control = conn_param->flow_control;
4062         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4063         rep.srq = id_priv->srq ? 1 : 0;
4064
4065         trace_cm_send_rep(id_priv);
4066         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4067 out:
4068         return ret;
4069 }
4070
4071 static int cma_accept_iw(struct rdma_id_private *id_priv,
4072                   struct rdma_conn_param *conn_param)
4073 {
4074         struct iw_cm_conn_param iw_param;
4075         int ret;
4076
4077         if (!conn_param)
4078                 return -EINVAL;
4079
4080         ret = cma_modify_qp_rtr(id_priv, conn_param);
4081         if (ret)
4082                 return ret;
4083
4084         iw_param.ord = conn_param->initiator_depth;
4085         iw_param.ird = conn_param->responder_resources;
4086         iw_param.private_data = conn_param->private_data;
4087         iw_param.private_data_len = conn_param->private_data_len;
4088         if (id_priv->id.qp) {
4089                 iw_param.qpn = id_priv->qp_num;
4090         } else
4091                 iw_param.qpn = conn_param->qp_num;
4092
4093         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4094 }
4095
4096 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4097                              enum ib_cm_sidr_status status, u32 qkey,
4098                              const void *private_data, int private_data_len)
4099 {
4100         struct ib_cm_sidr_rep_param rep;
4101         int ret;
4102
4103         memset(&rep, 0, sizeof rep);
4104         rep.status = status;
4105         if (status == IB_SIDR_SUCCESS) {
4106                 ret = cma_set_qkey(id_priv, qkey);
4107                 if (ret)
4108                         return ret;
4109                 rep.qp_num = id_priv->qp_num;
4110                 rep.qkey = id_priv->qkey;
4111         }
4112         rep.private_data = private_data;
4113         rep.private_data_len = private_data_len;
4114
4115         trace_cm_send_sidr_rep(id_priv);
4116         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4117 }
4118
4119 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4120                   const char *caller)
4121 {
4122         struct rdma_id_private *id_priv;
4123         int ret;
4124
4125         id_priv = container_of(id, struct rdma_id_private, id);
4126
4127         rdma_restrack_set_task(&id_priv->res, caller);
4128
4129         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
4130                 return -EINVAL;
4131
4132         if (!id->qp && conn_param) {
4133                 id_priv->qp_num = conn_param->qp_num;
4134                 id_priv->srq = conn_param->srq;
4135         }
4136
4137         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4138                 if (id->qp_type == IB_QPT_UD) {
4139                         if (conn_param)
4140                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4141                                                         conn_param->qkey,
4142                                                         conn_param->private_data,
4143                                                         conn_param->private_data_len);
4144                         else
4145                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4146                                                         0, NULL, 0);
4147                 } else {
4148                         if (conn_param)
4149                                 ret = cma_accept_ib(id_priv, conn_param);
4150                         else
4151                                 ret = cma_rep_recv(id_priv);
4152                 }
4153         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4154                 ret = cma_accept_iw(id_priv, conn_param);
4155         else
4156                 ret = -ENOSYS;
4157
4158         if (ret)
4159                 goto reject;
4160
4161         return 0;
4162 reject:
4163         cma_modify_qp_err(id_priv);
4164         rdma_reject(id, NULL, 0);
4165         return ret;
4166 }
4167 EXPORT_SYMBOL(__rdma_accept);
4168
4169 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4170 {
4171         struct rdma_id_private *id_priv;
4172         int ret;
4173
4174         id_priv = container_of(id, struct rdma_id_private, id);
4175         if (!id_priv->cm_id.ib)
4176                 return -EINVAL;
4177
4178         switch (id->device->node_type) {
4179         case RDMA_NODE_IB_CA:
4180                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4181                 break;
4182         default:
4183                 ret = 0;
4184                 break;
4185         }
4186         return ret;
4187 }
4188 EXPORT_SYMBOL(rdma_notify);
4189
4190 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4191                 u8 private_data_len)
4192 {
4193         struct rdma_id_private *id_priv;
4194         int ret;
4195
4196         id_priv = container_of(id, struct rdma_id_private, id);
4197         if (!id_priv->cm_id.ib)
4198                 return -EINVAL;
4199
4200         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4201                 if (id->qp_type == IB_QPT_UD) {
4202                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4203                                                 private_data, private_data_len);
4204                 } else {
4205                         trace_cm_send_rej(id_priv);
4206                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
4207                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
4208                                              0, private_data, private_data_len);
4209                 }
4210         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4211                 ret = iw_cm_reject(id_priv->cm_id.iw,
4212                                    private_data, private_data_len);
4213         } else
4214                 ret = -ENOSYS;
4215
4216         return ret;
4217 }
4218 EXPORT_SYMBOL(rdma_reject);
4219
4220 int rdma_disconnect(struct rdma_cm_id *id)
4221 {
4222         struct rdma_id_private *id_priv;
4223         int ret;
4224
4225         id_priv = container_of(id, struct rdma_id_private, id);
4226         if (!id_priv->cm_id.ib)
4227                 return -EINVAL;
4228
4229         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4230                 ret = cma_modify_qp_err(id_priv);
4231                 if (ret)
4232                         goto out;
4233                 /* Initiate or respond to a disconnect. */
4234                 trace_cm_disconnect(id_priv);
4235                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4236                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4237                                 trace_cm_sent_drep(id_priv);
4238                 } else {
4239                         trace_cm_sent_dreq(id_priv);
4240                 }
4241         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4242                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4243         } else
4244                 ret = -EINVAL;
4245
4246 out:
4247         return ret;
4248 }
4249 EXPORT_SYMBOL(rdma_disconnect);
4250
4251 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4252 {
4253         struct rdma_id_private *id_priv;
4254         struct cma_multicast *mc = multicast->context;
4255         struct rdma_cm_event event = {};
4256         int ret = 0;
4257
4258         id_priv = mc->id_priv;
4259         mutex_lock(&id_priv->handler_mutex);
4260         if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4261             id_priv->state != RDMA_CM_ADDR_RESOLVED)
4262                 goto out;
4263
4264         if (!status)
4265                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4266         else
4267                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4268                                      status);
4269         mutex_lock(&id_priv->qp_mutex);
4270         if (!status && id_priv->id.qp) {
4271                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
4272                                          be16_to_cpu(multicast->rec.mlid));
4273                 if (status)
4274                         pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
4275                                              status);
4276         }
4277         mutex_unlock(&id_priv->qp_mutex);
4278
4279         event.status = status;
4280         event.param.ud.private_data = mc->context;
4281         if (!status) {
4282                 struct rdma_dev_addr *dev_addr =
4283                         &id_priv->id.route.addr.dev_addr;
4284                 struct net_device *ndev =
4285                         dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4286                 enum ib_gid_type gid_type =
4287                         id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4288                         rdma_start_port(id_priv->cma_dev->device)];
4289
4290                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4291                 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4292                                                id_priv->id.port_num,
4293                                                &multicast->rec,
4294                                                ndev, gid_type,
4295                                                &event.param.ud.ah_attr);
4296                 if (ret)
4297                         event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4298
4299                 event.param.ud.qp_num = 0xFFFFFF;
4300                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4301                 if (ndev)
4302                         dev_put(ndev);
4303         } else
4304                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4305
4306         ret = cma_cm_event_handler(id_priv, &event);
4307
4308         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4309         if (ret) {
4310                 cma_exch(id_priv, RDMA_CM_DESTROYING);
4311                 mutex_unlock(&id_priv->handler_mutex);
4312                 rdma_destroy_id(&id_priv->id);
4313                 return 0;
4314         }
4315
4316 out:
4317         mutex_unlock(&id_priv->handler_mutex);
4318         return 0;
4319 }
4320
4321 static void cma_set_mgid(struct rdma_id_private *id_priv,
4322                          struct sockaddr *addr, union ib_gid *mgid)
4323 {
4324         unsigned char mc_map[MAX_ADDR_LEN];
4325         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4326         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4327         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4328
4329         if (cma_any_addr(addr)) {
4330                 memset(mgid, 0, sizeof *mgid);
4331         } else if ((addr->sa_family == AF_INET6) &&
4332                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4333                                                                  0xFF10A01B)) {
4334                 /* IPv6 address is an SA assigned MGID. */
4335                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4336         } else if (addr->sa_family == AF_IB) {
4337                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4338         } else if (addr->sa_family == AF_INET6) {
4339                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4340                 if (id_priv->id.ps == RDMA_PS_UDP)
4341                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4342                 *mgid = *(union ib_gid *) (mc_map + 4);
4343         } else {
4344                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4345                 if (id_priv->id.ps == RDMA_PS_UDP)
4346                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4347                 *mgid = *(union ib_gid *) (mc_map + 4);
4348         }
4349 }
4350
4351 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4352                                  struct cma_multicast *mc)
4353 {
4354         struct ib_sa_mcmember_rec rec;
4355         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4356         ib_sa_comp_mask comp_mask;
4357         int ret;
4358
4359         ib_addr_get_mgid(dev_addr, &rec.mgid);
4360         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4361                                      &rec.mgid, &rec);
4362         if (ret)
4363                 return ret;
4364
4365         ret = cma_set_qkey(id_priv, 0);
4366         if (ret)
4367                 return ret;
4368
4369         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4370         rec.qkey = cpu_to_be32(id_priv->qkey);
4371         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4372         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4373         rec.join_state = mc->join_state;
4374
4375         if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4376             (!ib_sa_sendonly_fullmem_support(&sa_client,
4377                                              id_priv->id.device,
4378                                              id_priv->id.port_num))) {
4379                 dev_warn(
4380                         &id_priv->id.device->dev,
4381                         "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4382                         id_priv->id.port_num);
4383                 return -EOPNOTSUPP;
4384         }
4385
4386         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4387                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4388                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4389                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4390                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4391
4392         if (id_priv->id.ps == RDMA_PS_IPOIB)
4393                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4394                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4395                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4396                              IB_SA_MCMEMBER_REC_MTU |
4397                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4398
4399         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4400                                                 id_priv->id.port_num, &rec,
4401                                                 comp_mask, GFP_KERNEL,
4402                                                 cma_ib_mc_handler, mc);
4403         return PTR_ERR_OR_ZERO(mc->multicast.ib);
4404 }
4405
4406 static void iboe_mcast_work_handler(struct work_struct *work)
4407 {
4408         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4409         struct cma_multicast *mc = mw->mc;
4410         struct ib_sa_multicast *m = mc->multicast.ib;
4411
4412         mc->multicast.ib->context = mc;
4413         cma_ib_mc_handler(0, m);
4414         kref_put(&mc->mcref, release_mc);
4415         kfree(mw);
4416 }
4417
4418 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4419                               enum ib_gid_type gid_type)
4420 {
4421         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4422         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4423
4424         if (cma_any_addr(addr)) {
4425                 memset(mgid, 0, sizeof *mgid);
4426         } else if (addr->sa_family == AF_INET6) {
4427                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4428         } else {
4429                 mgid->raw[0] =
4430                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4431                 mgid->raw[1] =
4432                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4433                 mgid->raw[2] = 0;
4434                 mgid->raw[3] = 0;
4435                 mgid->raw[4] = 0;
4436                 mgid->raw[5] = 0;
4437                 mgid->raw[6] = 0;
4438                 mgid->raw[7] = 0;
4439                 mgid->raw[8] = 0;
4440                 mgid->raw[9] = 0;
4441                 mgid->raw[10] = 0xff;
4442                 mgid->raw[11] = 0xff;
4443                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4444         }
4445 }
4446
4447 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4448                                    struct cma_multicast *mc)
4449 {
4450         struct iboe_mcast_work *work;
4451         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4452         int err = 0;
4453         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4454         struct net_device *ndev = NULL;
4455         enum ib_gid_type gid_type;
4456         bool send_only;
4457
4458         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4459
4460         if (cma_zero_addr((struct sockaddr *)&mc->addr))
4461                 return -EINVAL;
4462
4463         work = kzalloc(sizeof *work, GFP_KERNEL);
4464         if (!work)
4465                 return -ENOMEM;
4466
4467         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4468         if (!mc->multicast.ib) {
4469                 err = -ENOMEM;
4470                 goto out1;
4471         }
4472
4473         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4474                    rdma_start_port(id_priv->cma_dev->device)];
4475         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4476
4477         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4478         if (id_priv->id.ps == RDMA_PS_UDP)
4479                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4480
4481         if (dev_addr->bound_dev_if)
4482                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4483         if (!ndev) {
4484                 err = -ENODEV;
4485                 goto out2;
4486         }
4487         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4488         mc->multicast.ib->rec.hop_limit = 1;
4489         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4490
4491         if (addr->sa_family == AF_INET) {
4492                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4493                         mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4494                         if (!send_only) {
4495                                 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4496                                                     true);
4497                         }
4498                 }
4499         } else {
4500                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4501                         err = -ENOTSUPP;
4502         }
4503         dev_put(ndev);
4504         if (err || !mc->multicast.ib->rec.mtu) {
4505                 if (!err)
4506                         err = -EINVAL;
4507                 goto out2;
4508         }
4509         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4510                     &mc->multicast.ib->rec.port_gid);
4511         work->id = id_priv;
4512         work->mc = mc;
4513         INIT_WORK(&work->work, iboe_mcast_work_handler);
4514         kref_get(&mc->mcref);
4515         queue_work(cma_wq, &work->work);
4516
4517         return 0;
4518
4519 out2:
4520         kfree(mc->multicast.ib);
4521 out1:
4522         kfree(work);
4523         return err;
4524 }
4525
4526 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4527                         u8 join_state, void *context)
4528 {
4529         struct rdma_id_private *id_priv;
4530         struct cma_multicast *mc;
4531         int ret;
4532
4533         if (!id->device)
4534                 return -EINVAL;
4535
4536         id_priv = container_of(id, struct rdma_id_private, id);
4537         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4538             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4539                 return -EINVAL;
4540
4541         mc = kmalloc(sizeof *mc, GFP_KERNEL);
4542         if (!mc)
4543                 return -ENOMEM;
4544
4545         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4546         mc->context = context;
4547         mc->id_priv = id_priv;
4548         mc->join_state = join_state;
4549
4550         if (rdma_protocol_roce(id->device, id->port_num)) {
4551                 kref_init(&mc->mcref);
4552                 ret = cma_iboe_join_multicast(id_priv, mc);
4553                 if (ret)
4554                         goto out_err;
4555         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4556                 ret = cma_join_ib_multicast(id_priv, mc);
4557                 if (ret)
4558                         goto out_err;
4559         } else {
4560                 ret = -ENOSYS;
4561                 goto out_err;
4562         }
4563
4564         spin_lock(&id_priv->lock);
4565         list_add(&mc->list, &id_priv->mc_list);
4566         spin_unlock(&id_priv->lock);
4567
4568         return 0;
4569 out_err:
4570         kfree(mc);
4571         return ret;
4572 }
4573 EXPORT_SYMBOL(rdma_join_multicast);
4574
4575 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4576 {
4577         struct rdma_id_private *id_priv;
4578         struct cma_multicast *mc;
4579
4580         id_priv = container_of(id, struct rdma_id_private, id);
4581         spin_lock_irq(&id_priv->lock);
4582         list_for_each_entry(mc, &id_priv->mc_list, list) {
4583                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4584                         list_del(&mc->list);
4585                         spin_unlock_irq(&id_priv->lock);
4586
4587                         if (id->qp)
4588                                 ib_detach_mcast(id->qp,
4589                                                 &mc->multicast.ib->rec.mgid,
4590                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
4591
4592                         BUG_ON(id_priv->cma_dev->device != id->device);
4593
4594                         if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4595                                 ib_sa_free_multicast(mc->multicast.ib);
4596                                 kfree(mc);
4597                         } else if (rdma_protocol_roce(id->device, id->port_num)) {
4598                                 cma_leave_roce_mc_group(id_priv, mc);
4599                         }
4600                         return;
4601                 }
4602         }
4603         spin_unlock_irq(&id_priv->lock);
4604 }
4605 EXPORT_SYMBOL(rdma_leave_multicast);
4606
4607 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4608 {
4609         struct rdma_dev_addr *dev_addr;
4610         struct cma_ndev_work *work;
4611
4612         dev_addr = &id_priv->id.route.addr.dev_addr;
4613
4614         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4615             (net_eq(dev_net(ndev), dev_addr->net)) &&
4616             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4617                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4618                         ndev->name, &id_priv->id);
4619                 work = kzalloc(sizeof *work, GFP_KERNEL);
4620                 if (!work)
4621                         return -ENOMEM;
4622
4623                 INIT_WORK(&work->work, cma_ndev_work_handler);
4624                 work->id = id_priv;
4625                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4626                 cma_id_get(id_priv);
4627                 queue_work(cma_wq, &work->work);
4628         }
4629
4630         return 0;
4631 }
4632
4633 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4634                                void *ptr)
4635 {
4636         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4637         struct cma_device *cma_dev;
4638         struct rdma_id_private *id_priv;
4639         int ret = NOTIFY_DONE;
4640
4641         if (event != NETDEV_BONDING_FAILOVER)
4642                 return NOTIFY_DONE;
4643
4644         if (!netif_is_bond_master(ndev))
4645                 return NOTIFY_DONE;
4646
4647         mutex_lock(&lock);
4648         list_for_each_entry(cma_dev, &dev_list, list)
4649                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4650                         ret = cma_netdev_change(ndev, id_priv);
4651                         if (ret)
4652                                 goto out;
4653                 }
4654
4655 out:
4656         mutex_unlock(&lock);
4657         return ret;
4658 }
4659
4660 static struct notifier_block cma_nb = {
4661         .notifier_call = cma_netdev_callback
4662 };
4663
4664 static int cma_add_one(struct ib_device *device)
4665 {
4666         struct cma_device *cma_dev;
4667         struct rdma_id_private *id_priv;
4668         unsigned int i;
4669         unsigned long supported_gids = 0;
4670         int ret;
4671
4672         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4673         if (!cma_dev)
4674                 return -ENOMEM;
4675
4676         cma_dev->device = device;
4677         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4678                                             sizeof(*cma_dev->default_gid_type),
4679                                             GFP_KERNEL);
4680         if (!cma_dev->default_gid_type) {
4681                 ret = -ENOMEM;
4682                 goto free_cma_dev;
4683         }
4684
4685         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4686                                             sizeof(*cma_dev->default_roce_tos),
4687                                             GFP_KERNEL);
4688         if (!cma_dev->default_roce_tos) {
4689                 ret = -ENOMEM;
4690                 goto free_gid_type;
4691         }
4692
4693         rdma_for_each_port (device, i) {
4694                 supported_gids = roce_gid_type_mask_support(device, i);
4695                 WARN_ON(!supported_gids);
4696                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4697                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4698                                 CMA_PREFERRED_ROCE_GID_TYPE;
4699                 else
4700                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4701                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4702                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4703         }
4704
4705         init_completion(&cma_dev->comp);
4706         refcount_set(&cma_dev->refcount, 1);
4707         INIT_LIST_HEAD(&cma_dev->id_list);
4708         ib_set_client_data(device, &cma_client, cma_dev);
4709
4710         mutex_lock(&lock);
4711         list_add_tail(&cma_dev->list, &dev_list);
4712         list_for_each_entry(id_priv, &listen_any_list, list)
4713                 cma_listen_on_dev(id_priv, cma_dev);
4714         mutex_unlock(&lock);
4715
4716         trace_cm_add_one(device);
4717         return 0;
4718
4719 free_gid_type:
4720         kfree(cma_dev->default_gid_type);
4721
4722 free_cma_dev:
4723         kfree(cma_dev);
4724         return ret;
4725 }
4726
4727 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4728 {
4729         struct rdma_cm_event event = {};
4730         enum rdma_cm_state state;
4731         int ret = 0;
4732
4733         /* Record that we want to remove the device */
4734         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4735         if (state == RDMA_CM_DESTROYING)
4736                 return 0;
4737
4738         cma_cancel_operation(id_priv, state);
4739         mutex_lock(&id_priv->handler_mutex);
4740
4741         /* Check for destruction from another callback. */
4742         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4743                 goto out;
4744
4745         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4746         ret = cma_cm_event_handler(id_priv, &event);
4747 out:
4748         mutex_unlock(&id_priv->handler_mutex);
4749         return ret;
4750 }
4751
4752 static void cma_process_remove(struct cma_device *cma_dev)
4753 {
4754         struct rdma_id_private *id_priv;
4755         int ret;
4756
4757         mutex_lock(&lock);
4758         while (!list_empty(&cma_dev->id_list)) {
4759                 id_priv = list_entry(cma_dev->id_list.next,
4760                                      struct rdma_id_private, list);
4761
4762                 list_del(&id_priv->listen_list);
4763                 list_del_init(&id_priv->list);
4764                 cma_id_get(id_priv);
4765                 mutex_unlock(&lock);
4766
4767                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4768                 cma_id_put(id_priv);
4769                 if (ret)
4770                         rdma_destroy_id(&id_priv->id);
4771
4772                 mutex_lock(&lock);
4773         }
4774         mutex_unlock(&lock);
4775
4776         cma_dev_put(cma_dev);
4777         wait_for_completion(&cma_dev->comp);
4778 }
4779
4780 static void cma_remove_one(struct ib_device *device, void *client_data)
4781 {
4782         struct cma_device *cma_dev = client_data;
4783
4784         trace_cm_remove_one(device);
4785
4786         mutex_lock(&lock);
4787         list_del(&cma_dev->list);
4788         mutex_unlock(&lock);
4789
4790         cma_process_remove(cma_dev);
4791         kfree(cma_dev->default_roce_tos);
4792         kfree(cma_dev->default_gid_type);
4793         kfree(cma_dev);
4794 }
4795
4796 static int cma_init_net(struct net *net)
4797 {
4798         struct cma_pernet *pernet = cma_pernet(net);
4799
4800         xa_init(&pernet->tcp_ps);
4801         xa_init(&pernet->udp_ps);
4802         xa_init(&pernet->ipoib_ps);
4803         xa_init(&pernet->ib_ps);
4804
4805         return 0;
4806 }
4807
4808 static void cma_exit_net(struct net *net)
4809 {
4810         struct cma_pernet *pernet = cma_pernet(net);
4811
4812         WARN_ON(!xa_empty(&pernet->tcp_ps));
4813         WARN_ON(!xa_empty(&pernet->udp_ps));
4814         WARN_ON(!xa_empty(&pernet->ipoib_ps));
4815         WARN_ON(!xa_empty(&pernet->ib_ps));
4816 }
4817
4818 static struct pernet_operations cma_pernet_operations = {
4819         .init = cma_init_net,
4820         .exit = cma_exit_net,
4821         .id = &cma_pernet_id,
4822         .size = sizeof(struct cma_pernet),
4823 };
4824
4825 static int __init cma_init(void)
4826 {
4827         int ret;
4828
4829         /*
4830          * There is a rare lock ordering dependency in cma_netdev_callback()
4831          * that only happens when bonding is enabled. Teach lockdep that rtnl
4832          * must never be nested under lock so it can find these without having
4833          * to test with bonding.
4834          */
4835         if (IS_ENABLED(CONFIG_LOCKDEP)) {
4836                 rtnl_lock();
4837                 mutex_lock(&lock);
4838                 mutex_unlock(&lock);
4839                 rtnl_unlock();
4840         }
4841
4842         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4843         if (!cma_wq)
4844                 return -ENOMEM;
4845
4846         ret = register_pernet_subsys(&cma_pernet_operations);
4847         if (ret)
4848                 goto err_wq;
4849
4850         ib_sa_register_client(&sa_client);
4851         register_netdevice_notifier(&cma_nb);
4852
4853         ret = ib_register_client(&cma_client);
4854         if (ret)
4855                 goto err;
4856
4857         ret = cma_configfs_init();
4858         if (ret)
4859                 goto err_ib;
4860
4861         return 0;
4862
4863 err_ib:
4864         ib_unregister_client(&cma_client);
4865 err:
4866         unregister_netdevice_notifier(&cma_nb);
4867         ib_sa_unregister_client(&sa_client);
4868         unregister_pernet_subsys(&cma_pernet_operations);
4869 err_wq:
4870         destroy_workqueue(cma_wq);
4871         return ret;
4872 }
4873
4874 static void __exit cma_cleanup(void)
4875 {
4876         cma_configfs_exit();
4877         ib_unregister_client(&cma_client);
4878         unregister_netdevice_notifier(&cma_nb);
4879         ib_sa_unregister_client(&sa_client);
4880         unregister_pernet_subsys(&cma_pernet_operations);
4881         destroy_workqueue(cma_wq);
4882 }
4883
4884 module_init(cma_init);
4885 module_exit(cma_cleanup);