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 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
  72                               enum ib_gid_type gid_type);
  73
  74 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  75 {
  76         size_t index = event;
  77
  78         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  79                         cma_events[index] : "unrecognized event";
  80 }
  81 EXPORT_SYMBOL(rdma_event_msg);
  82
  83 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  84                                                 int reason)
  85 {
  86         if (rdma_ib_or_roce(id->device, id->port_num))
  87                 return ibcm_reject_msg(reason);
  88
  89         if (rdma_protocol_iwarp(id->device, id->port_num))
  90                 return iwcm_reject_msg(reason);
  91
  92         WARN_ON_ONCE(1);
  93         return "unrecognized transport";
  94 }
  95 EXPORT_SYMBOL(rdma_reject_msg);
  96
  97 /**
  98  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  99  *                           request.
 100  * @id: Communication identifier that received the REJECT event.
 101  * @reason: Value returned in the REJECT event status field.
 102  */
 103 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 104 {
 105         if (rdma_ib_or_roce(id->device, id->port_num))
 106                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 107
 108         if (rdma_protocol_iwarp(id->device, id->port_num))
 109                 return reason == -ECONNREFUSED;
 110
 111         WARN_ON_ONCE(1);
 112         return false;
 113 }
 114
 115 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 116                                       struct rdma_cm_event *ev, u8 *data_len)
 117 {
 118         const void *p;
 119
 120         if (rdma_is_consumer_reject(id, ev->status)) {
 121                 *data_len = ev->param.conn.private_data_len;
 122                 p = ev->param.conn.private_data;
 123         } else {
 124                 *data_len = 0;
 125                 p = NULL;
 126         }
 127         return p;
 128 }
 129 EXPORT_SYMBOL(rdma_consumer_reject_data);
 130
 131 /**
 132  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 133  * @id: Communication Identifier
 134  */
 135 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 136 {
 137         struct rdma_id_private *id_priv;
 138
 139         id_priv = container_of(id, struct rdma_id_private, id);
 140         if (id->device->node_type == RDMA_NODE_RNIC)
 141                 return id_priv->cm_id.iw;
 142         return NULL;
 143 }
 144 EXPORT_SYMBOL(rdma_iw_cm_id);
 145
 146 /**
 147  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 148  * @res: rdma resource tracking entry pointer
 149  */
 150 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 151 {
 152         struct rdma_id_private *id_priv =
 153                 container_of(res, struct rdma_id_private, res);
 154
 155         return &id_priv->id;
 156 }
 157 EXPORT_SYMBOL(rdma_res_to_id);
 158
 159 static int cma_add_one(struct ib_device *device);
 160 static void cma_remove_one(struct ib_device *device, void *client_data);
 161
 162 static struct ib_client cma_client = {
 163         .name   = "cma",
 164         .add    = cma_add_one,
 165         .remove = cma_remove_one
 166 };
 167
 168 static struct ib_sa_client sa_client;
 169 static LIST_HEAD(dev_list);
 170 static LIST_HEAD(listen_any_list);
 171 static DEFINE_MUTEX(lock);
 172 static struct workqueue_struct *cma_wq;
 173 static unsigned int cma_pernet_id;
 174
 175 struct cma_pernet {
 176         struct xarray tcp_ps;
 177         struct xarray udp_ps;
 178         struct xarray ipoib_ps;
 179         struct xarray ib_ps;
 180 };
 181
 182 static struct cma_pernet *cma_pernet(struct net *net)
 183 {
 184         return net_generic(net, cma_pernet_id);
 185 }
 186
 187 static
 188 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 189 {
 190         struct cma_pernet *pernet = cma_pernet(net);
 191
 192         switch (ps) {
 193         case RDMA_PS_TCP:
 194                 return &pernet->tcp_ps;
 195         case RDMA_PS_UDP:
 196                 return &pernet->udp_ps;
 197         case RDMA_PS_IPOIB:
 198                 return &pernet->ipoib_ps;
 199         case RDMA_PS_IB:
 200                 return &pernet->ib_ps;
 201         default:
 202                 return NULL;
 203         }
 204 }
 205
 206 struct cma_device {
 207         struct list_head        list;
 208         struct ib_device        *device;
 209         struct completion       comp;
 210         refcount_t refcount;
 211         struct list_head        id_list;
 212         enum ib_gid_type        *default_gid_type;
 213         u8                      *default_roce_tos;
 214 };
 215
 216 struct rdma_bind_list {
 217         enum rdma_ucm_port_space ps;
 218         struct hlist_head       owners;
 219         unsigned short          port;
 220 };
 221
 222 struct class_port_info_context {
 223         struct ib_class_port_info       *class_port_info;
 224         struct ib_device                *device;
 225         struct completion               done;
 226         struct ib_sa_query              *sa_query;
 227         u8                              port_num;
 228 };
 229
 230 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 231                         struct rdma_bind_list *bind_list, int snum)
 232 {
 233         struct xarray *xa = cma_pernet_xa(net, ps);
 234
 235         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 236 }
 237
 238 static struct rdma_bind_list *cma_ps_find(struct net *net,
 239                                           enum rdma_ucm_port_space ps, int snum)
 240 {
 241         struct xarray *xa = cma_pernet_xa(net, ps);
 242
 243         return xa_load(xa, snum);
 244 }
 245
 246 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 247                           int snum)
 248 {
 249         struct xarray *xa = cma_pernet_xa(net, ps);
 250
 251         xa_erase(xa, snum);
 252 }
 253
 254 enum {
 255         CMA_OPTION_AFONLY,
 256 };
 257
 258 void cma_dev_get(struct cma_device *cma_dev)
 259 {
 260         refcount_inc(&cma_dev->refcount);
 261 }
 262
 263 void cma_dev_put(struct cma_device *cma_dev)
 264 {
 265         if (refcount_dec_and_test(&cma_dev->refcount))
 266                 complete(&cma_dev->comp);
 267 }
 268
 269 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 270                                              void               *cookie)
 271 {
 272         struct cma_device *cma_dev;
 273         struct cma_device *found_cma_dev = NULL;
 274
 275         mutex_lock(&lock);
 276
 277         list_for_each_entry(cma_dev, &dev_list, list)
 278                 if (filter(cma_dev->device, cookie)) {
 279                         found_cma_dev = cma_dev;
 280                         break;
 281                 }
 282
 283         if (found_cma_dev)
 284                 cma_dev_get(found_cma_dev);
 285         mutex_unlock(&lock);
 286         return found_cma_dev;
 287 }
 288
 289 int cma_get_default_gid_type(struct cma_device *cma_dev,
 290                              unsigned int port)
 291 {
 292         if (!rdma_is_port_valid(cma_dev->device, port))
 293                 return -EINVAL;
 294
 295         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 296 }
 297
 298 int cma_set_default_gid_type(struct cma_device *cma_dev,
 299                              unsigned int port,
 300                              enum ib_gid_type default_gid_type)
 301 {
 302         unsigned long supported_gids;
 303
 304         if (!rdma_is_port_valid(cma_dev->device, port))
 305                 return -EINVAL;
 306
 307         if (default_gid_type == IB_GID_TYPE_IB &&
 308             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 309                 default_gid_type = IB_GID_TYPE_ROCE;
 310
 311         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 312
 313         if (!(supported_gids & 1 << default_gid_type))
 314                 return -EINVAL;
 315
 316         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 317                 default_gid_type;
 318
 319         return 0;
 320 }
 321
 322 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 323 {
 324         if (!rdma_is_port_valid(cma_dev->device, port))
 325                 return -EINVAL;
 326
 327         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 328 }
 329
 330 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 331                              u8 default_roce_tos)
 332 {
 333         if (!rdma_is_port_valid(cma_dev->device, port))
 334                 return -EINVAL;
 335
 336         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 337                  default_roce_tos;
 338
 339         return 0;
 340 }
 341 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 342 {
 343         return cma_dev->device;
 344 }
 345
 346 /*
 347  * Device removal can occur at anytime, so we need extra handling to
 348  * serialize notifying the user of device removal with other callbacks.
 349  * We do this by disabling removal notification while a callback is in process,
 350  * and reporting it after the callback completes.
 351  */
 352
 353 struct cma_multicast {
 354         struct rdma_id_private *id_priv;
 355         union {
 356                 struct ib_sa_multicast *sa_mc;
 357                 struct {
 358                         struct work_struct work;
 359                         struct rdma_cm_event event;
 360                 } iboe_join;
 361         };
 362         struct list_head        list;
 363         void                    *context;
 364         struct sockaddr_storage addr;
 365         u8                      join_state;
 366 };
 367
 368 struct cma_work {
 369         struct work_struct      work;
 370         struct rdma_id_private  *id;
 371         enum rdma_cm_state      old_state;
 372         enum rdma_cm_state      new_state;
 373         struct rdma_cm_event    event;
 374 };
 375
 376 union cma_ip_addr {
 377         struct in6_addr ip6;
 378         struct {
 379                 __be32 pad[3];
 380                 __be32 addr;
 381         } ip4;
 382 };
 383
 384 struct cma_hdr {
 385         u8 cma_version;
 386         u8 ip_version;  /* IP version: 7:4 */
 387         __be16 port;
 388         union cma_ip_addr src_addr;
 389         union cma_ip_addr dst_addr;
 390 };
 391
 392 #define CMA_VERSION 0x00
 393
 394 struct cma_req_info {
 395         struct sockaddr_storage listen_addr_storage;
 396         struct sockaddr_storage src_addr_storage;
 397         struct ib_device *device;
 398         union ib_gid local_gid;
 399         __be64 service_id;
 400         int port;
 401         bool has_gid;
 402         u16 pkey;
 403 };
 404
 405 static int cma_comp_exch(struct rdma_id_private *id_priv,
 406                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 407 {
 408         unsigned long flags;
 409         int ret;
 410
 411         /*
 412          * The FSM uses a funny double locking where state is protected by both
 413          * the handler_mutex and the spinlock. State is not allowed to change
 414          * to/from a handler_mutex protected value without also holding
 415          * handler_mutex.
 416          */
 417         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 418                 lockdep_assert_held(&id_priv->handler_mutex);
 419
 420         spin_lock_irqsave(&id_priv->lock, flags);
 421         if ((ret = (id_priv->state == comp)))
 422                 id_priv->state = exch;
 423         spin_unlock_irqrestore(&id_priv->lock, flags);
 424         return ret;
 425 }
 426
 427 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 428 {
 429         return hdr->ip_version >> 4;
 430 }
 431
 432 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 433 {
 434         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 435 }
 436
 437 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 438 {
 439         struct in_device *in_dev = NULL;
 440
 441         if (ndev) {
 442                 rtnl_lock();
 443                 in_dev = __in_dev_get_rtnl(ndev);
 444                 if (in_dev) {
 445                         if (join)
 446                                 ip_mc_inc_group(in_dev,
 447                                                 *(__be32 *)(mgid->raw + 12));
 448                         else
 449                                 ip_mc_dec_group(in_dev,
 450                                                 *(__be32 *)(mgid->raw + 12));
 451                 }
 452                 rtnl_unlock();
 453         }
 454         return (in_dev) ? 0 : -ENODEV;
 455 }
 456
 457 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 458                                struct cma_device *cma_dev)
 459 {
 460         cma_dev_get(cma_dev);
 461         id_priv->cma_dev = cma_dev;
 462         id_priv->id.device = cma_dev->device;
 463         id_priv->id.route.addr.dev_addr.transport =
 464                 rdma_node_get_transport(cma_dev->device->node_type);
 465         list_add_tail(&id_priv->list, &cma_dev->id_list);
 466
 467         trace_cm_id_attach(id_priv, cma_dev->device);
 468 }
 469
 470 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 471                               struct cma_device *cma_dev)
 472 {
 473         _cma_attach_to_dev(id_priv, cma_dev);
 474         id_priv->gid_type =
 475                 cma_dev->default_gid_type[id_priv->id.port_num -
 476                                           rdma_start_port(cma_dev->device)];
 477 }
 478
 479 static void cma_release_dev(struct rdma_id_private *id_priv)
 480 {
 481         mutex_lock(&lock);
 482         list_del(&id_priv->list);
 483         cma_dev_put(id_priv->cma_dev);
 484         id_priv->cma_dev = NULL;
 485         id_priv->id.device = NULL;
 486         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 487                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 488                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 489         }
 490         mutex_unlock(&lock);
 491 }
 492
 493 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 494 {
 495         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 496 }
 497
 498 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 499 {
 500         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 501 }
 502
 503 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 504 {
 505         return id_priv->id.route.addr.src_addr.ss_family;
 506 }
 507
 508 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 509 {
 510         struct ib_sa_mcmember_rec rec;
 511         int ret = 0;
 512
 513         if (id_priv->qkey) {
 514                 if (qkey && id_priv->qkey != qkey)
 515                         return -EINVAL;
 516                 return 0;
 517         }
 518
 519         if (qkey) {
 520                 id_priv->qkey = qkey;
 521                 return 0;
 522         }
 523
 524         switch (id_priv->id.ps) {
 525         case RDMA_PS_UDP:
 526         case RDMA_PS_IB:
 527                 id_priv->qkey = RDMA_UDP_QKEY;
 528                 break;
 529         case RDMA_PS_IPOIB:
 530                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 531                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 532                                              id_priv->id.port_num, &rec.mgid,
 533                                              &rec);
 534                 if (!ret)
 535                         id_priv->qkey = be32_to_cpu(rec.qkey);
 536                 break;
 537         default:
 538                 break;
 539         }
 540         return ret;
 541 }
 542
 543 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 544 {
 545         dev_addr->dev_type = ARPHRD_INFINIBAND;
 546         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 547         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 548 }
 549
 550 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 551 {
 552         int ret;
 553
 554         if (addr->sa_family != AF_IB) {
 555                 ret = rdma_translate_ip(addr, dev_addr);
 556         } else {
 557                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 558                 ret = 0;
 559         }
 560
 561         return ret;
 562 }
 563
 564 static const struct ib_gid_attr *
 565 cma_validate_port(struct ib_device *device, u8 port,
 566                   enum ib_gid_type gid_type,
 567                   union ib_gid *gid,
 568                   struct rdma_id_private *id_priv)
 569 {
 570         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 571         int bound_if_index = dev_addr->bound_dev_if;
 572         const struct ib_gid_attr *sgid_attr;
 573         int dev_type = dev_addr->dev_type;
 574         struct net_device *ndev = NULL;
 575
 576         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 577                 return ERR_PTR(-ENODEV);
 578
 579         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 580                 return ERR_PTR(-ENODEV);
 581
 582         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 583                 return ERR_PTR(-ENODEV);
 584
 585         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 586                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 587                 if (!ndev)
 588                         return ERR_PTR(-ENODEV);
 589         } else {
 590                 gid_type = IB_GID_TYPE_IB;
 591         }
 592
 593         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 594         if (ndev)
 595                 dev_put(ndev);
 596         return sgid_attr;
 597 }
 598
 599 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 600                                const struct ib_gid_attr *sgid_attr)
 601 {
 602         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 603         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 604 }
 605
 606 /**
 607  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 608  * based on source ip address.
 609  * @id_priv:    cm_id which should be bound to cma device
 610  *
 611  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 612  * based on source IP address. It returns 0 on success or error code otherwise.
 613  * It is applicable to active and passive side cm_id.
 614  */
 615 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 616 {
 617         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 618         const struct ib_gid_attr *sgid_attr;
 619         union ib_gid gid, iboe_gid, *gidp;
 620         struct cma_device *cma_dev;
 621         enum ib_gid_type gid_type;
 622         int ret = -ENODEV;
 623         unsigned int port;
 624
 625         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 626             id_priv->id.ps == RDMA_PS_IPOIB)
 627                 return -EINVAL;
 628
 629         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 630                     &iboe_gid);
 631
 632         memcpy(&gid, dev_addr->src_dev_addr +
 633                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 634
 635         mutex_lock(&lock);
 636         list_for_each_entry(cma_dev, &dev_list, list) {
 637                 rdma_for_each_port (cma_dev->device, port) {
 638                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 639                                &iboe_gid : &gid;
 640                         gid_type = cma_dev->default_gid_type[port - 1];
 641                         sgid_attr = cma_validate_port(cma_dev->device, port,
 642                                                       gid_type, gidp, id_priv);
 643                         if (!IS_ERR(sgid_attr)) {
 644                                 id_priv->id.port_num = port;
 645                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 646                                 cma_attach_to_dev(id_priv, cma_dev);
 647                                 ret = 0;
 648                                 goto out;
 649                         }
 650                 }
 651         }
 652 out:
 653         mutex_unlock(&lock);
 654         return ret;
 655 }
 656
 657 /**
 658  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 659  * @id_priv:            cm id to bind to cma device
 660  * @listen_id_priv:     listener cm id to match against
 661  * @req:                Pointer to req structure containaining incoming
 662  *                      request information
 663  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 664  * rdma device matches for listen_id and incoming request. It also verifies
 665  * that a GID table entry is present for the source address.
 666  * Returns 0 on success, or returns error code otherwise.
 667  */
 668 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 669                               const struct rdma_id_private *listen_id_priv,
 670                               struct cma_req_info *req)
 671 {
 672         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 673         const struct ib_gid_attr *sgid_attr;
 674         enum ib_gid_type gid_type;
 675         union ib_gid gid;
 676
 677         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 678             id_priv->id.ps == RDMA_PS_IPOIB)
 679                 return -EINVAL;
 680
 681         if (rdma_protocol_roce(req->device, req->port))
 682                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 683                             &gid);
 684         else
 685                 memcpy(&gid, dev_addr->src_dev_addr +
 686                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 687
 688         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 689         sgid_attr = cma_validate_port(req->device, req->port,
 690                                       gid_type, &gid, id_priv);
 691         if (IS_ERR(sgid_attr))
 692                 return PTR_ERR(sgid_attr);
 693
 694         id_priv->id.port_num = req->port;
 695         cma_bind_sgid_attr(id_priv, sgid_attr);
 696         /* Need to acquire lock to protect against reader
 697          * of cma_dev->id_list such as cma_netdev_callback() and
 698          * cma_process_remove().
 699          */
 700         mutex_lock(&lock);
 701         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 702         mutex_unlock(&lock);
 703         rdma_restrack_add(&id_priv->res);
 704         return 0;
 705 }
 706
 707 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 708                               const struct rdma_id_private *listen_id_priv)
 709 {
 710         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 711         const struct ib_gid_attr *sgid_attr;
 712         struct cma_device *cma_dev;
 713         enum ib_gid_type gid_type;
 714         int ret = -ENODEV;
 715         unsigned int port;
 716         union ib_gid gid;
 717
 718         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 719             id_priv->id.ps == RDMA_PS_IPOIB)
 720                 return -EINVAL;
 721
 722         memcpy(&gid, dev_addr->src_dev_addr +
 723                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 724
 725         mutex_lock(&lock);
 726
 727         cma_dev = listen_id_priv->cma_dev;
 728         port = listen_id_priv->id.port_num;
 729         gid_type = listen_id_priv->gid_type;
 730         sgid_attr = cma_validate_port(cma_dev->device, port,
 731                                       gid_type, &gid, id_priv);
 732         if (!IS_ERR(sgid_attr)) {
 733                 id_priv->id.port_num = port;
 734                 cma_bind_sgid_attr(id_priv, sgid_attr);
 735                 ret = 0;
 736                 goto out;
 737         }
 738
 739         list_for_each_entry(cma_dev, &dev_list, list) {
 740                 rdma_for_each_port (cma_dev->device, port) {
 741                         if (listen_id_priv->cma_dev == cma_dev &&
 742                             listen_id_priv->id.port_num == port)
 743                                 continue;
 744
 745                         gid_type = cma_dev->default_gid_type[port - 1];
 746                         sgid_attr = cma_validate_port(cma_dev->device, port,
 747                                                       gid_type, &gid, id_priv);
 748                         if (!IS_ERR(sgid_attr)) {
 749                                 id_priv->id.port_num = port;
 750                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 751                                 ret = 0;
 752                                 goto out;
 753                         }
 754                 }
 755         }
 756
 757 out:
 758         if (!ret) {
 759                 cma_attach_to_dev(id_priv, cma_dev);
 760                 rdma_restrack_add(&id_priv->res);
 761         }
 762
 763         mutex_unlock(&lock);
 764         return ret;
 765 }
 766
 767 /*
 768  * Select the source IB device and address to reach the destination IB address.
 769  */
 770 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 771 {
 772         struct cma_device *cma_dev, *cur_dev;
 773         struct sockaddr_ib *addr;
 774         union ib_gid gid, sgid, *dgid;
 775         unsigned int p;
 776         u16 pkey, index;
 777         enum ib_port_state port_state;
 778         int ret;
 779         int i;
 780
 781         cma_dev = NULL;
 782         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 783         dgid = (union ib_gid *) &addr->sib_addr;
 784         pkey = ntohs(addr->sib_pkey);
 785
 786         mutex_lock(&lock);
 787         list_for_each_entry(cur_dev, &dev_list, list) {
 788                 rdma_for_each_port (cur_dev->device, p) {
 789                         if (!rdma_cap_af_ib(cur_dev->device, p))
 790                                 continue;
 791
 792                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 793                                 continue;
 794
 795                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 796                                 continue;
 797
 798                         for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
 799                              ++i) {
 800                                 ret = rdma_query_gid(cur_dev->device, p, i,
 801                                                      &gid);
 802                                 if (ret)
 803                                         continue;
 804
 805                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 806                                         cma_dev = cur_dev;
 807                                         sgid = gid;
 808                                         id_priv->id.port_num = p;
 809                                         goto found;
 810                                 }
 811
 812                                 if (!cma_dev && (gid.global.subnet_prefix ==
 813                                     dgid->global.subnet_prefix) &&
 814                                     port_state == IB_PORT_ACTIVE) {
 815                                         cma_dev = cur_dev;
 816                                         sgid = gid;
 817                                         id_priv->id.port_num = p;
 818                                         goto found;
 819                                 }
 820                         }
 821                 }
 822         }
 823         mutex_unlock(&lock);
 824         return -ENODEV;
 825
 826 found:
 827         cma_attach_to_dev(id_priv, cma_dev);
 828         rdma_restrack_add(&id_priv->res);
 829         mutex_unlock(&lock);
 830         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 831         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 832         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 833         return 0;
 834 }
 835
 836 static void cma_id_get(struct rdma_id_private *id_priv)
 837 {
 838         refcount_inc(&id_priv->refcount);
 839 }
 840
 841 static void cma_id_put(struct rdma_id_private *id_priv)
 842 {
 843         if (refcount_dec_and_test(&id_priv->refcount))
 844                 complete(&id_priv->comp);
 845 }
 846
 847 static struct rdma_id_private *
 848 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 849                  void *context, enum rdma_ucm_port_space ps,
 850                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 851 {
 852         struct rdma_id_private *id_priv;
 853
 854         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 855         if (!id_priv)
 856                 return ERR_PTR(-ENOMEM);
 857
 858         id_priv->state = RDMA_CM_IDLE;
 859         id_priv->id.context = context;
 860         id_priv->id.event_handler = event_handler;
 861         id_priv->id.ps = ps;
 862         id_priv->id.qp_type = qp_type;
 863         id_priv->tos_set = false;
 864         id_priv->timeout_set = false;
 865         id_priv->gid_type = IB_GID_TYPE_IB;
 866         spin_lock_init(&id_priv->lock);
 867         mutex_init(&id_priv->qp_mutex);
 868         init_completion(&id_priv->comp);
 869         refcount_set(&id_priv->refcount, 1);
 870         mutex_init(&id_priv->handler_mutex);
 871         INIT_LIST_HEAD(&id_priv->listen_list);
 872         INIT_LIST_HEAD(&id_priv->mc_list);
 873         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 874         id_priv->id.route.addr.dev_addr.net = get_net(net);
 875         id_priv->seq_num &= 0x00ffffff;
 876
 877         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 878         if (parent)
 879                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 880
 881         return id_priv;
 882 }
 883
 884 struct rdma_cm_id *
 885 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
 886                         void *context, enum rdma_ucm_port_space ps,
 887                         enum ib_qp_type qp_type, const char *caller)
 888 {
 889         struct rdma_id_private *ret;
 890
 891         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
 892         if (IS_ERR(ret))
 893                 return ERR_CAST(ret);
 894
 895         rdma_restrack_set_name(&ret->res, caller);
 896         return &ret->id;
 897 }
 898 EXPORT_SYMBOL(__rdma_create_kernel_id);
 899
 900 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
 901                                        void *context,
 902                                        enum rdma_ucm_port_space ps,
 903                                        enum ib_qp_type qp_type)
 904 {
 905         struct rdma_id_private *ret;
 906
 907         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
 908                                ps, qp_type, NULL);
 909         if (IS_ERR(ret))
 910                 return ERR_CAST(ret);
 911
 912         rdma_restrack_set_name(&ret->res, NULL);
 913         return &ret->id;
 914 }
 915 EXPORT_SYMBOL(rdma_create_user_id);
 916
 917 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 918 {
 919         struct ib_qp_attr qp_attr;
 920         int qp_attr_mask, ret;
 921
 922         qp_attr.qp_state = IB_QPS_INIT;
 923         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 924         if (ret)
 925                 return ret;
 926
 927         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 928         if (ret)
 929                 return ret;
 930
 931         qp_attr.qp_state = IB_QPS_RTR;
 932         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 933         if (ret)
 934                 return ret;
 935
 936         qp_attr.qp_state = IB_QPS_RTS;
 937         qp_attr.sq_psn = 0;
 938         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 939
 940         return ret;
 941 }
 942
 943 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 944 {
 945         struct ib_qp_attr qp_attr;
 946         int qp_attr_mask, ret;
 947
 948         qp_attr.qp_state = IB_QPS_INIT;
 949         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 950         if (ret)
 951                 return ret;
 952
 953         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 954 }
 955
 956 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 957                    struct ib_qp_init_attr *qp_init_attr)
 958 {
 959         struct rdma_id_private *id_priv;
 960         struct ib_qp *qp;
 961         int ret;
 962
 963         id_priv = container_of(id, struct rdma_id_private, id);
 964         if (id->device != pd->device) {
 965                 ret = -EINVAL;
 966                 goto out_err;
 967         }
 968
 969         qp_init_attr->port_num = id->port_num;
 970         qp = ib_create_qp(pd, qp_init_attr);
 971         if (IS_ERR(qp)) {
 972                 ret = PTR_ERR(qp);
 973                 goto out_err;
 974         }
 975
 976         if (id->qp_type == IB_QPT_UD)
 977                 ret = cma_init_ud_qp(id_priv, qp);
 978         else
 979                 ret = cma_init_conn_qp(id_priv, qp);
 980         if (ret)
 981                 goto out_destroy;
 982
 983         id->qp = qp;
 984         id_priv->qp_num = qp->qp_num;
 985         id_priv->srq = (qp->srq != NULL);
 986         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 987         return 0;
 988 out_destroy:
 989         ib_destroy_qp(qp);
 990 out_err:
 991         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 992         return ret;
 993 }
 994 EXPORT_SYMBOL(rdma_create_qp);
 995
 996 void rdma_destroy_qp(struct rdma_cm_id *id)
 997 {
 998         struct rdma_id_private *id_priv;
 999
1000         id_priv = container_of(id, struct rdma_id_private, id);
1001         trace_cm_qp_destroy(id_priv);
1002         mutex_lock(&id_priv->qp_mutex);
1003         ib_destroy_qp(id_priv->id.qp);
1004         id_priv->id.qp = NULL;
1005         mutex_unlock(&id_priv->qp_mutex);
1006 }
1007 EXPORT_SYMBOL(rdma_destroy_qp);
1008
1009 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1010                              struct rdma_conn_param *conn_param)
1011 {
1012         struct ib_qp_attr qp_attr;
1013         int qp_attr_mask, ret;
1014
1015         mutex_lock(&id_priv->qp_mutex);
1016         if (!id_priv->id.qp) {
1017                 ret = 0;
1018                 goto out;
1019         }
1020
1021         /* Need to update QP attributes from default values. */
1022         qp_attr.qp_state = IB_QPS_INIT;
1023         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1024         if (ret)
1025                 goto out;
1026
1027         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1028         if (ret)
1029                 goto out;
1030
1031         qp_attr.qp_state = IB_QPS_RTR;
1032         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1033         if (ret)
1034                 goto out;
1035
1036         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1037
1038         if (conn_param)
1039                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1040         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1041 out:
1042         mutex_unlock(&id_priv->qp_mutex);
1043         return ret;
1044 }
1045
1046 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1047                              struct rdma_conn_param *conn_param)
1048 {
1049         struct ib_qp_attr qp_attr;
1050         int qp_attr_mask, ret;
1051
1052         mutex_lock(&id_priv->qp_mutex);
1053         if (!id_priv->id.qp) {
1054                 ret = 0;
1055                 goto out;
1056         }
1057
1058         qp_attr.qp_state = IB_QPS_RTS;
1059         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1060         if (ret)
1061                 goto out;
1062
1063         if (conn_param)
1064                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1065         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1066 out:
1067         mutex_unlock(&id_priv->qp_mutex);
1068         return ret;
1069 }
1070
1071 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1072 {
1073         struct ib_qp_attr qp_attr;
1074         int ret;
1075
1076         mutex_lock(&id_priv->qp_mutex);
1077         if (!id_priv->id.qp) {
1078                 ret = 0;
1079                 goto out;
1080         }
1081
1082         qp_attr.qp_state = IB_QPS_ERR;
1083         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1084 out:
1085         mutex_unlock(&id_priv->qp_mutex);
1086         return ret;
1087 }
1088
1089 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1090                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1091 {
1092         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1093         int ret;
1094         u16 pkey;
1095
1096         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1097                 pkey = 0xffff;
1098         else
1099                 pkey = ib_addr_get_pkey(dev_addr);
1100
1101         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1102                                   pkey, &qp_attr->pkey_index);
1103         if (ret)
1104                 return ret;
1105
1106         qp_attr->port_num = id_priv->id.port_num;
1107         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1108
1109         if (id_priv->id.qp_type == IB_QPT_UD) {
1110                 ret = cma_set_qkey(id_priv, 0);
1111                 if (ret)
1112                         return ret;
1113
1114                 qp_attr->qkey = id_priv->qkey;
1115                 *qp_attr_mask |= IB_QP_QKEY;
1116         } else {
1117                 qp_attr->qp_access_flags = 0;
1118                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1119         }
1120         return 0;
1121 }
1122
1123 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1124                        int *qp_attr_mask)
1125 {
1126         struct rdma_id_private *id_priv;
1127         int ret = 0;
1128
1129         id_priv = container_of(id, struct rdma_id_private, id);
1130         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1131                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1132                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1133                 else
1134                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1135                                                  qp_attr_mask);
1136
1137                 if (qp_attr->qp_state == IB_QPS_RTR)
1138                         qp_attr->rq_psn = id_priv->seq_num;
1139         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1140                 if (!id_priv->cm_id.iw) {
1141                         qp_attr->qp_access_flags = 0;
1142                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1143                 } else
1144                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1145                                                  qp_attr_mask);
1146                 qp_attr->port_num = id_priv->id.port_num;
1147                 *qp_attr_mask |= IB_QP_PORT;
1148         } else
1149                 ret = -ENOSYS;
1150
1151         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1152                 qp_attr->timeout = id_priv->timeout;
1153
1154         return ret;
1155 }
1156 EXPORT_SYMBOL(rdma_init_qp_attr);
1157
1158 static inline bool cma_zero_addr(const struct sockaddr *addr)
1159 {
1160         switch (addr->sa_family) {
1161         case AF_INET:
1162                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1163         case AF_INET6:
1164                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1165         case AF_IB:
1166                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1167         default:
1168                 return false;
1169         }
1170 }
1171
1172 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1173 {
1174         switch (addr->sa_family) {
1175         case AF_INET:
1176                 return ipv4_is_loopback(
1177                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1178         case AF_INET6:
1179                 return ipv6_addr_loopback(
1180                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1181         case AF_IB:
1182                 return ib_addr_loopback(
1183                         &((struct sockaddr_ib *)addr)->sib_addr);
1184         default:
1185                 return false;
1186         }
1187 }
1188
1189 static inline bool cma_any_addr(const struct sockaddr *addr)
1190 {
1191         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1192 }
1193
1194 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1195 {
1196         if (src->sa_family != dst->sa_family)
1197                 return -1;
1198
1199         switch (src->sa_family) {
1200         case AF_INET:
1201                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1202                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1203         case AF_INET6: {
1204                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1205                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1206                 bool link_local;
1207
1208                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1209                                           &dst_addr6->sin6_addr))
1210                         return 1;
1211                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1212                              IPV6_ADDR_LINKLOCAL;
1213                 /* Link local must match their scope_ids */
1214                 return link_local ? (src_addr6->sin6_scope_id !=
1215                                      dst_addr6->sin6_scope_id) :
1216                                     0;
1217         }
1218
1219         default:
1220                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1221                                    &((struct sockaddr_ib *) dst)->sib_addr);
1222         }
1223 }
1224
1225 static __be16 cma_port(const struct sockaddr *addr)
1226 {
1227         struct sockaddr_ib *sib;
1228
1229         switch (addr->sa_family) {
1230         case AF_INET:
1231                 return ((struct sockaddr_in *) addr)->sin_port;
1232         case AF_INET6:
1233                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1234         case AF_IB:
1235                 sib = (struct sockaddr_ib *) addr;
1236                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1237                                     be64_to_cpu(sib->sib_sid_mask)));
1238         default:
1239                 return 0;
1240         }
1241 }
1242
1243 static inline int cma_any_port(const struct sockaddr *addr)
1244 {
1245         return !cma_port(addr);
1246 }
1247
1248 static void cma_save_ib_info(struct sockaddr *src_addr,
1249                              struct sockaddr *dst_addr,
1250                              const struct rdma_cm_id *listen_id,
1251                              const struct sa_path_rec *path)
1252 {
1253         struct sockaddr_ib *listen_ib, *ib;
1254
1255         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1256         if (src_addr) {
1257                 ib = (struct sockaddr_ib *)src_addr;
1258                 ib->sib_family = AF_IB;
1259                 if (path) {
1260                         ib->sib_pkey = path->pkey;
1261                         ib->sib_flowinfo = path->flow_label;
1262                         memcpy(&ib->sib_addr, &path->sgid, 16);
1263                         ib->sib_sid = path->service_id;
1264                         ib->sib_scope_id = 0;
1265                 } else {
1266                         ib->sib_pkey = listen_ib->sib_pkey;
1267                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1268                         ib->sib_addr = listen_ib->sib_addr;
1269                         ib->sib_sid = listen_ib->sib_sid;
1270                         ib->sib_scope_id = listen_ib->sib_scope_id;
1271                 }
1272                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1273         }
1274         if (dst_addr) {
1275                 ib = (struct sockaddr_ib *)dst_addr;
1276                 ib->sib_family = AF_IB;
1277                 if (path) {
1278                         ib->sib_pkey = path->pkey;
1279                         ib->sib_flowinfo = path->flow_label;
1280                         memcpy(&ib->sib_addr, &path->dgid, 16);
1281                 }
1282         }
1283 }
1284
1285 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1286                               struct sockaddr_in *dst_addr,
1287                               struct cma_hdr *hdr,
1288                               __be16 local_port)
1289 {
1290         if (src_addr) {
1291                 *src_addr = (struct sockaddr_in) {
1292                         .sin_family = AF_INET,
1293                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1294                         .sin_port = local_port,
1295                 };
1296         }
1297
1298         if (dst_addr) {
1299                 *dst_addr = (struct sockaddr_in) {
1300                         .sin_family = AF_INET,
1301                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1302                         .sin_port = hdr->port,
1303                 };
1304         }
1305 }
1306
1307 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1308                               struct sockaddr_in6 *dst_addr,
1309                               struct cma_hdr *hdr,
1310                               __be16 local_port)
1311 {
1312         if (src_addr) {
1313                 *src_addr = (struct sockaddr_in6) {
1314                         .sin6_family = AF_INET6,
1315                         .sin6_addr = hdr->dst_addr.ip6,
1316                         .sin6_port = local_port,
1317                 };
1318         }
1319
1320         if (dst_addr) {
1321                 *dst_addr = (struct sockaddr_in6) {
1322                         .sin6_family = AF_INET6,
1323                         .sin6_addr = hdr->src_addr.ip6,
1324                         .sin6_port = hdr->port,
1325                 };
1326         }
1327 }
1328
1329 static u16 cma_port_from_service_id(__be64 service_id)
1330 {
1331         return (u16)be64_to_cpu(service_id);
1332 }
1333
1334 static int cma_save_ip_info(struct sockaddr *src_addr,
1335                             struct sockaddr *dst_addr,
1336                             const struct ib_cm_event *ib_event,
1337                             __be64 service_id)
1338 {
1339         struct cma_hdr *hdr;
1340         __be16 port;
1341
1342         hdr = ib_event->private_data;
1343         if (hdr->cma_version != CMA_VERSION)
1344                 return -EINVAL;
1345
1346         port = htons(cma_port_from_service_id(service_id));
1347
1348         switch (cma_get_ip_ver(hdr)) {
1349         case 4:
1350                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1351                                   (struct sockaddr_in *)dst_addr, hdr, port);
1352                 break;
1353         case 6:
1354                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1355                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1356                 break;
1357         default:
1358                 return -EAFNOSUPPORT;
1359         }
1360
1361         return 0;
1362 }
1363
1364 static int cma_save_net_info(struct sockaddr *src_addr,
1365                              struct sockaddr *dst_addr,
1366                              const struct rdma_cm_id *listen_id,
1367                              const struct ib_cm_event *ib_event,
1368                              sa_family_t sa_family, __be64 service_id)
1369 {
1370         if (sa_family == AF_IB) {
1371                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1372                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1373                                          ib_event->param.req_rcvd.primary_path);
1374                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1375                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1376                 return 0;
1377         }
1378
1379         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1380 }
1381
1382 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1383                              struct cma_req_info *req)
1384 {
1385         const struct ib_cm_req_event_param *req_param =
1386                 &ib_event->param.req_rcvd;
1387         const struct ib_cm_sidr_req_event_param *sidr_param =
1388                 &ib_event->param.sidr_req_rcvd;
1389
1390         switch (ib_event->event) {
1391         case IB_CM_REQ_RECEIVED:
1392                 req->device     = req_param->listen_id->device;
1393                 req->port       = req_param->port;
1394                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1395                        sizeof(req->local_gid));
1396                 req->has_gid    = true;
1397                 req->service_id = req_param->primary_path->service_id;
1398                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1399                 if (req->pkey != req_param->bth_pkey)
1400                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1401                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1402                                             req_param->bth_pkey, req->pkey);
1403                 break;
1404         case IB_CM_SIDR_REQ_RECEIVED:
1405                 req->device     = sidr_param->listen_id->device;
1406                 req->port       = sidr_param->port;
1407                 req->has_gid    = false;
1408                 req->service_id = sidr_param->service_id;
1409                 req->pkey       = sidr_param->pkey;
1410                 if (req->pkey != sidr_param->bth_pkey)
1411                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1412                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1413                                             sidr_param->bth_pkey, req->pkey);
1414                 break;
1415         default:
1416                 return -EINVAL;
1417         }
1418
1419         return 0;
1420 }
1421
1422 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1423                                   const struct sockaddr_in *dst_addr,
1424                                   const struct sockaddr_in *src_addr)
1425 {
1426         __be32 daddr = dst_addr->sin_addr.s_addr,
1427                saddr = src_addr->sin_addr.s_addr;
1428         struct fib_result res;
1429         struct flowi4 fl4;
1430         int err;
1431         bool ret;
1432
1433         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1434             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1435             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1436             ipv4_is_loopback(saddr))
1437                 return false;
1438
1439         memset(&fl4, 0, sizeof(fl4));
1440         fl4.flowi4_iif = net_dev->ifindex;
1441         fl4.daddr = daddr;
1442         fl4.saddr = saddr;
1443
1444         rcu_read_lock();
1445         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1446         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1447         rcu_read_unlock();
1448
1449         return ret;
1450 }
1451
1452 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1453                                   const struct sockaddr_in6 *dst_addr,
1454                                   const struct sockaddr_in6 *src_addr)
1455 {
1456 #if IS_ENABLED(CONFIG_IPV6)
1457         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1458                            IPV6_ADDR_LINKLOCAL;
1459         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1460                                          &src_addr->sin6_addr, net_dev->ifindex,
1461                                          NULL, strict);
1462         bool ret;
1463
1464         if (!rt)
1465                 return false;
1466
1467         ret = rt->rt6i_idev->dev == net_dev;
1468         ip6_rt_put(rt);
1469
1470         return ret;
1471 #else
1472         return false;
1473 #endif
1474 }
1475
1476 static bool validate_net_dev(struct net_device *net_dev,
1477                              const struct sockaddr *daddr,
1478                              const struct sockaddr *saddr)
1479 {
1480         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1481         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1482         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1483         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1484
1485         switch (daddr->sa_family) {
1486         case AF_INET:
1487                 return saddr->sa_family == AF_INET &&
1488                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1489
1490         case AF_INET6:
1491                 return saddr->sa_family == AF_INET6 &&
1492                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1493
1494         default:
1495                 return false;
1496         }
1497 }
1498
1499 static struct net_device *
1500 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1501 {
1502         const struct ib_gid_attr *sgid_attr = NULL;
1503         struct net_device *ndev;
1504
1505         if (ib_event->event == IB_CM_REQ_RECEIVED)
1506                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1507         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1508                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1509
1510         if (!sgid_attr)
1511                 return NULL;
1512
1513         rcu_read_lock();
1514         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1515         if (IS_ERR(ndev))
1516                 ndev = NULL;
1517         else
1518                 dev_hold(ndev);
1519         rcu_read_unlock();
1520         return ndev;
1521 }
1522
1523 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1524                                           struct cma_req_info *req)
1525 {
1526         struct sockaddr *listen_addr =
1527                         (struct sockaddr *)&req->listen_addr_storage;
1528         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1529         struct net_device *net_dev;
1530         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1531         int err;
1532
1533         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1534                                req->service_id);
1535         if (err)
1536                 return ERR_PTR(err);
1537
1538         if (rdma_protocol_roce(req->device, req->port))
1539                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1540         else
1541                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1542                                                    req->pkey,
1543                                                    gid, listen_addr);
1544         if (!net_dev)
1545                 return ERR_PTR(-ENODEV);
1546
1547         return net_dev;
1548 }
1549
1550 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1551 {
1552         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1553 }
1554
1555 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1556                                    const struct cma_hdr *hdr)
1557 {
1558         struct sockaddr *addr = cma_src_addr(id_priv);
1559         __be32 ip4_addr;
1560         struct in6_addr ip6_addr;
1561
1562         if (cma_any_addr(addr) && !id_priv->afonly)
1563                 return true;
1564
1565         switch (addr->sa_family) {
1566         case AF_INET:
1567                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1568                 if (cma_get_ip_ver(hdr) != 4)
1569                         return false;
1570                 if (!cma_any_addr(addr) &&
1571                     hdr->dst_addr.ip4.addr != ip4_addr)
1572                         return false;
1573                 break;
1574         case AF_INET6:
1575                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1576                 if (cma_get_ip_ver(hdr) != 6)
1577                         return false;
1578                 if (!cma_any_addr(addr) &&
1579                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1580                         return false;
1581                 break;
1582         case AF_IB:
1583                 return true;
1584         default:
1585                 return false;
1586         }
1587
1588         return true;
1589 }
1590
1591 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1592 {
1593         struct ib_device *device = id->device;
1594         const int port_num = id->port_num ?: rdma_start_port(device);
1595
1596         return rdma_protocol_roce(device, port_num);
1597 }
1598
1599 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1600 {
1601         const struct sockaddr *daddr =
1602                         (const struct sockaddr *)&req->listen_addr_storage;
1603         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1604
1605         /* Returns true if the req is for IPv6 link local */
1606         return (daddr->sa_family == AF_INET6 &&
1607                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1608 }
1609
1610 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1611                               const struct net_device *net_dev,
1612                               const struct cma_req_info *req)
1613 {
1614         const struct rdma_addr *addr = &id->route.addr;
1615
1616         if (!net_dev)
1617                 /* This request is an AF_IB request */
1618                 return (!id->port_num || id->port_num == req->port) &&
1619                        (addr->src_addr.ss_family == AF_IB);
1620
1621         /*
1622          * If the request is not for IPv6 link local, allow matching
1623          * request to any netdevice of the one or multiport rdma device.
1624          */
1625         if (!cma_is_req_ipv6_ll(req))
1626                 return true;
1627         /*
1628          * Net namespaces must match, and if the listner is listening
1629          * on a specific netdevice than netdevice must match as well.
1630          */
1631         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1632             (!!addr->dev_addr.bound_dev_if ==
1633              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1634                 return true;
1635         else
1636                 return false;
1637 }
1638
1639 static struct rdma_id_private *cma_find_listener(
1640                 const struct rdma_bind_list *bind_list,
1641                 const struct ib_cm_id *cm_id,
1642                 const struct ib_cm_event *ib_event,
1643                 const struct cma_req_info *req,
1644                 const struct net_device *net_dev)
1645 {
1646         struct rdma_id_private *id_priv, *id_priv_dev;
1647
1648         lockdep_assert_held(&lock);
1649
1650         if (!bind_list)
1651                 return ERR_PTR(-EINVAL);
1652
1653         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1654                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1655                         if (id_priv->id.device == cm_id->device &&
1656                             cma_match_net_dev(&id_priv->id, net_dev, req))
1657                                 return id_priv;
1658                         list_for_each_entry(id_priv_dev,
1659                                             &id_priv->listen_list,
1660                                             listen_list) {
1661                                 if (id_priv_dev->id.device == cm_id->device &&
1662                                     cma_match_net_dev(&id_priv_dev->id,
1663                                                       net_dev, req))
1664                                         return id_priv_dev;
1665                         }
1666                 }
1667         }
1668
1669         return ERR_PTR(-EINVAL);
1670 }
1671
1672 static struct rdma_id_private *
1673 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1674                      const struct ib_cm_event *ib_event,
1675                      struct cma_req_info *req,
1676                      struct net_device **net_dev)
1677 {
1678         struct rdma_bind_list *bind_list;
1679         struct rdma_id_private *id_priv;
1680         int err;
1681
1682         err = cma_save_req_info(ib_event, req);
1683         if (err)
1684                 return ERR_PTR(err);
1685
1686         *net_dev = cma_get_net_dev(ib_event, req);
1687         if (IS_ERR(*net_dev)) {
1688                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1689                         /* Assuming the protocol is AF_IB */
1690                         *net_dev = NULL;
1691                 } else {
1692                         return ERR_CAST(*net_dev);
1693                 }
1694         }
1695
1696         mutex_lock(&lock);
1697         /*
1698          * Net namespace might be getting deleted while route lookup,
1699          * cm_id lookup is in progress. Therefore, perform netdevice
1700          * validation, cm_id lookup under rcu lock.
1701          * RCU lock along with netdevice state check, synchronizes with
1702          * netdevice migrating to different net namespace and also avoids
1703          * case where net namespace doesn't get deleted while lookup is in
1704          * progress.
1705          * If the device state is not IFF_UP, its properties such as ifindex
1706          * and nd_net cannot be trusted to remain valid without rcu lock.
1707          * net/core/dev.c change_net_namespace() ensures to synchronize with
1708          * ongoing operations on net device after device is closed using
1709          * synchronize_net().
1710          */
1711         rcu_read_lock();
1712         if (*net_dev) {
1713                 /*
1714                  * If netdevice is down, it is likely that it is administratively
1715                  * down or it might be migrating to different namespace.
1716                  * In that case avoid further processing, as the net namespace
1717                  * or ifindex may change.
1718                  */
1719                 if (((*net_dev)->flags & IFF_UP) == 0) {
1720                         id_priv = ERR_PTR(-EHOSTUNREACH);
1721                         goto err;
1722                 }
1723
1724                 if (!validate_net_dev(*net_dev,
1725                                  (struct sockaddr *)&req->src_addr_storage,
1726                                  (struct sockaddr *)&req->listen_addr_storage)) {
1727                         id_priv = ERR_PTR(-EHOSTUNREACH);
1728                         goto err;
1729                 }
1730         }
1731
1732         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1733                                 rdma_ps_from_service_id(req->service_id),
1734                                 cma_port_from_service_id(req->service_id));
1735         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1736 err:
1737         rcu_read_unlock();
1738         mutex_unlock(&lock);
1739         if (IS_ERR(id_priv) && *net_dev) {
1740                 dev_put(*net_dev);
1741                 *net_dev = NULL;
1742         }
1743         return id_priv;
1744 }
1745
1746 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1747 {
1748         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1749 }
1750
1751 static void cma_cancel_route(struct rdma_id_private *id_priv)
1752 {
1753         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1754                 if (id_priv->query)
1755                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1756         }
1757 }
1758
1759 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1760 {
1761         struct rdma_id_private *dev_id_priv;
1762
1763         lockdep_assert_held(&lock);
1764
1765         /*
1766          * Remove from listen_any_list to prevent added devices from spawning
1767          * additional listen requests.
1768          */
1769         list_del(&id_priv->list);
1770
1771         while (!list_empty(&id_priv->listen_list)) {
1772                 dev_id_priv = list_entry(id_priv->listen_list.next,
1773                                          struct rdma_id_private, listen_list);
1774                 /* sync with device removal to avoid duplicate destruction */
1775                 list_del_init(&dev_id_priv->list);
1776                 list_del(&dev_id_priv->listen_list);
1777                 mutex_unlock(&lock);
1778
1779                 rdma_destroy_id(&dev_id_priv->id);
1780                 mutex_lock(&lock);
1781         }
1782 }
1783
1784 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1785 {
1786         mutex_lock(&lock);
1787         _cma_cancel_listens(id_priv);
1788         mutex_unlock(&lock);
1789 }
1790
1791 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1792                                  enum rdma_cm_state state)
1793 {
1794         switch (state) {
1795         case RDMA_CM_ADDR_QUERY:
1796                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1797                 break;
1798         case RDMA_CM_ROUTE_QUERY:
1799                 cma_cancel_route(id_priv);
1800                 break;
1801         case RDMA_CM_LISTEN:
1802                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1803                         cma_cancel_listens(id_priv);
1804                 break;
1805         default:
1806                 break;
1807         }
1808 }
1809
1810 static void cma_release_port(struct rdma_id_private *id_priv)
1811 {
1812         struct rdma_bind_list *bind_list = id_priv->bind_list;
1813         struct net *net = id_priv->id.route.addr.dev_addr.net;
1814
1815         if (!bind_list)
1816                 return;
1817
1818         mutex_lock(&lock);
1819         hlist_del(&id_priv->node);
1820         if (hlist_empty(&bind_list->owners)) {
1821                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1822                 kfree(bind_list);
1823         }
1824         mutex_unlock(&lock);
1825 }
1826
1827 static void destroy_mc(struct rdma_id_private *id_priv,
1828                        struct cma_multicast *mc)
1829 {
1830         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1831
1832         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1833                 ib_sa_free_multicast(mc->sa_mc);
1834
1835         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1836                 struct rdma_dev_addr *dev_addr =
1837                         &id_priv->id.route.addr.dev_addr;
1838                 struct net_device *ndev = NULL;
1839
1840                 if (dev_addr->bound_dev_if)
1841                         ndev = dev_get_by_index(dev_addr->net,
1842                                                 dev_addr->bound_dev_if);
1843                 if (ndev && !send_only) {
1844                         enum ib_gid_type gid_type;
1845                         union ib_gid mgid;
1846
1847                         gid_type = id_priv->cma_dev->default_gid_type
1848                                            [id_priv->id.port_num -
1849                                             rdma_start_port(
1850                                                     id_priv->cma_dev->device)];
1851                         cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
1852                                           gid_type);
1853                         cma_igmp_send(ndev, &mgid, false);
1854                 }
1855                 dev_put(ndev);
1856
1857                 cancel_work_sync(&mc->iboe_join.work);
1858         }
1859         kfree(mc);
1860 }
1861
1862 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1863 {
1864         struct cma_multicast *mc;
1865
1866         while (!list_empty(&id_priv->mc_list)) {
1867                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1868                                       list);
1869                 list_del(&mc->list);
1870                 destroy_mc(id_priv, mc);
1871         }
1872 }
1873
1874 static void _destroy_id(struct rdma_id_private *id_priv,
1875                         enum rdma_cm_state state)
1876 {
1877         cma_cancel_operation(id_priv, state);
1878
1879         rdma_restrack_del(&id_priv->res);
1880         if (id_priv->cma_dev) {
1881                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1882                         if (id_priv->cm_id.ib)
1883                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1884                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1885                         if (id_priv->cm_id.iw)
1886                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1887                 }
1888                 cma_leave_mc_groups(id_priv);
1889                 cma_release_dev(id_priv);
1890         }
1891
1892         cma_release_port(id_priv);
1893         cma_id_put(id_priv);
1894         wait_for_completion(&id_priv->comp);
1895
1896         if (id_priv->internal_id)
1897                 cma_id_put(id_priv->id.context);
1898
1899         kfree(id_priv->id.route.path_rec);
1900
1901         put_net(id_priv->id.route.addr.dev_addr.net);
1902         kfree(id_priv);
1903 }
1904
1905 /*
1906  * destroy an ID from within the handler_mutex. This ensures that no other
1907  * handlers can start running concurrently.
1908  */
1909 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1910         __releases(&idprv->handler_mutex)
1911 {
1912         enum rdma_cm_state state;
1913         unsigned long flags;
1914
1915         trace_cm_id_destroy(id_priv);
1916
1917         /*
1918          * Setting the state to destroyed under the handler mutex provides a
1919          * fence against calling handler callbacks. If this is invoked due to
1920          * the failure of a handler callback then it guarentees that no future
1921          * handlers will be called.
1922          */
1923         lockdep_assert_held(&id_priv->handler_mutex);
1924         spin_lock_irqsave(&id_priv->lock, flags);
1925         state = id_priv->state;
1926         id_priv->state = RDMA_CM_DESTROYING;
1927         spin_unlock_irqrestore(&id_priv->lock, flags);
1928         mutex_unlock(&id_priv->handler_mutex);
1929         _destroy_id(id_priv, state);
1930 }
1931
1932 void rdma_destroy_id(struct rdma_cm_id *id)
1933 {
1934         struct rdma_id_private *id_priv =
1935                 container_of(id, struct rdma_id_private, id);
1936
1937         mutex_lock(&id_priv->handler_mutex);
1938         destroy_id_handler_unlock(id_priv);
1939 }
1940 EXPORT_SYMBOL(rdma_destroy_id);
1941
1942 static int cma_rep_recv(struct rdma_id_private *id_priv)
1943 {
1944         int ret;
1945
1946         ret = cma_modify_qp_rtr(id_priv, NULL);
1947         if (ret)
1948                 goto reject;
1949
1950         ret = cma_modify_qp_rts(id_priv, NULL);
1951         if (ret)
1952                 goto reject;
1953
1954         trace_cm_send_rtu(id_priv);
1955         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1956         if (ret)
1957                 goto reject;
1958
1959         return 0;
1960 reject:
1961         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1962         cma_modify_qp_err(id_priv);
1963         trace_cm_send_rej(id_priv);
1964         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1965                        NULL, 0, NULL, 0);
1966         return ret;
1967 }
1968
1969 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1970                                    const struct ib_cm_rep_event_param *rep_data,
1971                                    void *private_data)
1972 {
1973         event->param.conn.private_data = private_data;
1974         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1975         event->param.conn.responder_resources = rep_data->responder_resources;
1976         event->param.conn.initiator_depth = rep_data->initiator_depth;
1977         event->param.conn.flow_control = rep_data->flow_control;
1978         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1979         event->param.conn.srq = rep_data->srq;
1980         event->param.conn.qp_num = rep_data->remote_qpn;
1981
1982         event->ece.vendor_id = rep_data->ece.vendor_id;
1983         event->ece.attr_mod = rep_data->ece.attr_mod;
1984 }
1985
1986 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1987                                 struct rdma_cm_event *event)
1988 {
1989         int ret;
1990
1991         lockdep_assert_held(&id_priv->handler_mutex);
1992
1993         trace_cm_event_handler(id_priv, event);
1994         ret = id_priv->id.event_handler(&id_priv->id, event);
1995         trace_cm_event_done(id_priv, event, ret);
1996         return ret;
1997 }
1998
1999 static int cma_ib_handler(struct ib_cm_id *cm_id,
2000                           const struct ib_cm_event *ib_event)
2001 {
2002         struct rdma_id_private *id_priv = cm_id->context;
2003         struct rdma_cm_event event = {};
2004         enum rdma_cm_state state;
2005         int ret;
2006
2007         mutex_lock(&id_priv->handler_mutex);
2008         state = READ_ONCE(id_priv->state);
2009         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2010              state != RDMA_CM_CONNECT) ||
2011             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2012              state != RDMA_CM_DISCONNECT))
2013                 goto out;
2014
2015         switch (ib_event->event) {
2016         case IB_CM_REQ_ERROR:
2017         case IB_CM_REP_ERROR:
2018                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2019                 event.status = -ETIMEDOUT;
2020                 break;
2021         case IB_CM_REP_RECEIVED:
2022                 if (state == RDMA_CM_CONNECT &&
2023                     (id_priv->id.qp_type != IB_QPT_UD)) {
2024                         trace_cm_send_mra(id_priv);
2025                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2026                 }
2027                 if (id_priv->id.qp) {
2028                         event.status = cma_rep_recv(id_priv);
2029                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2030                                                      RDMA_CM_EVENT_ESTABLISHED;
2031                 } else {
2032                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2033                 }
2034                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2035                                        ib_event->private_data);
2036                 break;
2037         case IB_CM_RTU_RECEIVED:
2038         case IB_CM_USER_ESTABLISHED:
2039                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2040                 break;
2041         case IB_CM_DREQ_ERROR:
2042                 event.status = -ETIMEDOUT;
2043                 fallthrough;
2044         case IB_CM_DREQ_RECEIVED:
2045         case IB_CM_DREP_RECEIVED:
2046                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2047                                    RDMA_CM_DISCONNECT))
2048                         goto out;
2049                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2050                 break;
2051         case IB_CM_TIMEWAIT_EXIT:
2052                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2053                 break;
2054         case IB_CM_MRA_RECEIVED:
2055                 /* ignore event */
2056                 goto out;
2057         case IB_CM_REJ_RECEIVED:
2058                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2059                                                                                 ib_event->param.rej_rcvd.reason));
2060                 cma_modify_qp_err(id_priv);
2061                 event.status = ib_event->param.rej_rcvd.reason;
2062                 event.event = RDMA_CM_EVENT_REJECTED;
2063                 event.param.conn.private_data = ib_event->private_data;
2064                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2065                 break;
2066         default:
2067                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2068                        ib_event->event);
2069                 goto out;
2070         }
2071
2072         ret = cma_cm_event_handler(id_priv, &event);
2073         if (ret) {
2074                 /* Destroy the CM ID by returning a non-zero value. */
2075                 id_priv->cm_id.ib = NULL;
2076                 destroy_id_handler_unlock(id_priv);
2077                 return ret;
2078         }
2079 out:
2080         mutex_unlock(&id_priv->handler_mutex);
2081         return 0;
2082 }
2083
2084 static struct rdma_id_private *
2085 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2086                    const struct ib_cm_event *ib_event,
2087                    struct net_device *net_dev)
2088 {
2089         struct rdma_id_private *listen_id_priv;
2090         struct rdma_id_private *id_priv;
2091         struct rdma_cm_id *id;
2092         struct rdma_route *rt;
2093         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2094         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2095         const __be64 service_id =
2096                 ib_event->param.req_rcvd.primary_path->service_id;
2097         int ret;
2098
2099         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2100         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2101                                    listen_id->event_handler, listen_id->context,
2102                                    listen_id->ps,
2103                                    ib_event->param.req_rcvd.qp_type,
2104                                    listen_id_priv);
2105         if (IS_ERR(id_priv))
2106                 return NULL;
2107
2108         id = &id_priv->id;
2109         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2110                               (struct sockaddr *)&id->route.addr.dst_addr,
2111                               listen_id, ib_event, ss_family, service_id))
2112                 goto err;
2113
2114         rt = &id->route;
2115         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2116         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2117                                      GFP_KERNEL);
2118         if (!rt->path_rec)
2119                 goto err;
2120
2121         rt->path_rec[0] = *path;
2122         if (rt->num_paths == 2)
2123                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2124
2125         if (net_dev) {
2126                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2127         } else {
2128                 if (!cma_protocol_roce(listen_id) &&
2129                     cma_any_addr(cma_src_addr(id_priv))) {
2130                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2131                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2132                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2133                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2134                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2135                         if (ret)
2136                                 goto err;
2137                 }
2138         }
2139         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2140
2141         id_priv->state = RDMA_CM_CONNECT;
2142         return id_priv;
2143
2144 err:
2145         rdma_destroy_id(id);
2146         return NULL;
2147 }
2148
2149 static struct rdma_id_private *
2150 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2151                   const struct ib_cm_event *ib_event,
2152                   struct net_device *net_dev)
2153 {
2154         const struct rdma_id_private *listen_id_priv;
2155         struct rdma_id_private *id_priv;
2156         struct rdma_cm_id *id;
2157         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2158         struct net *net = listen_id->route.addr.dev_addr.net;
2159         int ret;
2160
2161         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2162         id_priv = __rdma_create_id(net, listen_id->event_handler,
2163                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2164                                    listen_id_priv);
2165         if (IS_ERR(id_priv))
2166                 return NULL;
2167
2168         id = &id_priv->id;
2169         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2170                               (struct sockaddr *)&id->route.addr.dst_addr,
2171                               listen_id, ib_event, ss_family,
2172                               ib_event->param.sidr_req_rcvd.service_id))
2173                 goto err;
2174
2175         if (net_dev) {
2176                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2177         } else {
2178                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2179                         ret = cma_translate_addr(cma_src_addr(id_priv),
2180                                                  &id->route.addr.dev_addr);
2181                         if (ret)
2182                                 goto err;
2183                 }
2184         }
2185
2186         id_priv->state = RDMA_CM_CONNECT;
2187         return id_priv;
2188 err:
2189         rdma_destroy_id(id);
2190         return NULL;
2191 }
2192
2193 static void cma_set_req_event_data(struct rdma_cm_event *event,
2194                                    const struct ib_cm_req_event_param *req_data,
2195                                    void *private_data, int offset)
2196 {
2197         event->param.conn.private_data = private_data + offset;
2198         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2199         event->param.conn.responder_resources = req_data->responder_resources;
2200         event->param.conn.initiator_depth = req_data->initiator_depth;
2201         event->param.conn.flow_control = req_data->flow_control;
2202         event->param.conn.retry_count = req_data->retry_count;
2203         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2204         event->param.conn.srq = req_data->srq;
2205         event->param.conn.qp_num = req_data->remote_qpn;
2206
2207         event->ece.vendor_id = req_data->ece.vendor_id;
2208         event->ece.attr_mod = req_data->ece.attr_mod;
2209 }
2210
2211 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2212                                     const struct ib_cm_event *ib_event)
2213 {
2214         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2215                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2216                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2217                  (id->qp_type == IB_QPT_UD)) ||
2218                 (!id->qp_type));
2219 }
2220
2221 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2222                               const struct ib_cm_event *ib_event)
2223 {
2224         struct rdma_id_private *listen_id, *conn_id = NULL;
2225         struct rdma_cm_event event = {};
2226         struct cma_req_info req = {};
2227         struct net_device *net_dev;
2228         u8 offset;
2229         int ret;
2230
2231         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2232         if (IS_ERR(listen_id))
2233                 return PTR_ERR(listen_id);
2234
2235         trace_cm_req_handler(listen_id, ib_event->event);
2236         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2237                 ret = -EINVAL;
2238                 goto net_dev_put;
2239         }
2240
2241         mutex_lock(&listen_id->handler_mutex);
2242         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2243                 ret = -ECONNABORTED;
2244                 goto err_unlock;
2245         }
2246
2247         offset = cma_user_data_offset(listen_id);
2248         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2249         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2250                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2251                 event.param.ud.private_data = ib_event->private_data + offset;
2252                 event.param.ud.private_data_len =
2253                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2254         } else {
2255                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2256                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2257                                        ib_event->private_data, offset);
2258         }
2259         if (!conn_id) {
2260                 ret = -ENOMEM;
2261                 goto err_unlock;
2262         }
2263
2264         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2265         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2266         if (ret) {
2267                 destroy_id_handler_unlock(conn_id);
2268                 goto err_unlock;
2269         }
2270
2271         conn_id->cm_id.ib = cm_id;
2272         cm_id->context = conn_id;
2273         cm_id->cm_handler = cma_ib_handler;
2274
2275         ret = cma_cm_event_handler(conn_id, &event);
2276         if (ret) {
2277                 /* Destroy the CM ID by returning a non-zero value. */
2278                 conn_id->cm_id.ib = NULL;
2279                 mutex_unlock(&listen_id->handler_mutex);
2280                 destroy_id_handler_unlock(conn_id);
2281                 goto net_dev_put;
2282         }
2283
2284         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2285             conn_id->id.qp_type != IB_QPT_UD) {
2286                 trace_cm_send_mra(cm_id->context);
2287                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2288         }
2289         mutex_unlock(&conn_id->handler_mutex);
2290
2291 err_unlock:
2292         mutex_unlock(&listen_id->handler_mutex);
2293
2294 net_dev_put:
2295         if (net_dev)
2296                 dev_put(net_dev);
2297
2298         return ret;
2299 }
2300
2301 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2302 {
2303         if (addr->sa_family == AF_IB)
2304                 return ((struct sockaddr_ib *) addr)->sib_sid;
2305
2306         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2307 }
2308 EXPORT_SYMBOL(rdma_get_service_id);
2309
2310 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2311                     union ib_gid *dgid)
2312 {
2313         struct rdma_addr *addr = &cm_id->route.addr;
2314
2315         if (!cm_id->device) {
2316                 if (sgid)
2317                         memset(sgid, 0, sizeof(*sgid));
2318                 if (dgid)
2319                         memset(dgid, 0, sizeof(*dgid));
2320                 return;
2321         }
2322
2323         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2324                 if (sgid)
2325                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2326                 if (dgid)
2327                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2328         } else {
2329                 if (sgid)
2330                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2331                 if (dgid)
2332                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2333         }
2334 }
2335 EXPORT_SYMBOL(rdma_read_gids);
2336
2337 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2338 {
2339         struct rdma_id_private *id_priv = iw_id->context;
2340         struct rdma_cm_event event = {};
2341         int ret = 0;
2342         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2343         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2344
2345         mutex_lock(&id_priv->handler_mutex);
2346         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2347                 goto out;
2348
2349         switch (iw_event->event) {
2350         case IW_CM_EVENT_CLOSE:
2351                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2352                 break;
2353         case IW_CM_EVENT_CONNECT_REPLY:
2354                 memcpy(cma_src_addr(id_priv), laddr,
2355                        rdma_addr_size(laddr));
2356                 memcpy(cma_dst_addr(id_priv), raddr,
2357                        rdma_addr_size(raddr));
2358                 switch (iw_event->status) {
2359                 case 0:
2360                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2361                         event.param.conn.initiator_depth = iw_event->ird;
2362                         event.param.conn.responder_resources = iw_event->ord;
2363                         break;
2364                 case -ECONNRESET:
2365                 case -ECONNREFUSED:
2366                         event.event = RDMA_CM_EVENT_REJECTED;
2367                         break;
2368                 case -ETIMEDOUT:
2369                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2370                         break;
2371                 default:
2372                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2373                         break;
2374                 }
2375                 break;
2376         case IW_CM_EVENT_ESTABLISHED:
2377                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2378                 event.param.conn.initiator_depth = iw_event->ird;
2379                 event.param.conn.responder_resources = iw_event->ord;
2380                 break;
2381         default:
2382                 goto out;
2383         }
2384
2385         event.status = iw_event->status;
2386         event.param.conn.private_data = iw_event->private_data;
2387         event.param.conn.private_data_len = iw_event->private_data_len;
2388         ret = cma_cm_event_handler(id_priv, &event);
2389         if (ret) {
2390                 /* Destroy the CM ID by returning a non-zero value. */
2391                 id_priv->cm_id.iw = NULL;
2392                 destroy_id_handler_unlock(id_priv);
2393                 return ret;
2394         }
2395
2396 out:
2397         mutex_unlock(&id_priv->handler_mutex);
2398         return ret;
2399 }
2400
2401 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2402                                struct iw_cm_event *iw_event)
2403 {
2404         struct rdma_id_private *listen_id, *conn_id;
2405         struct rdma_cm_event event = {};
2406         int ret = -ECONNABORTED;
2407         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2408         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2409
2410         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2411         event.param.conn.private_data = iw_event->private_data;
2412         event.param.conn.private_data_len = iw_event->private_data_len;
2413         event.param.conn.initiator_depth = iw_event->ird;
2414         event.param.conn.responder_resources = iw_event->ord;
2415
2416         listen_id = cm_id->context;
2417
2418         mutex_lock(&listen_id->handler_mutex);
2419         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2420                 goto out;
2421
2422         /* Create a new RDMA id for the new IW CM ID */
2423         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2424                                    listen_id->id.event_handler,
2425                                    listen_id->id.context, RDMA_PS_TCP,
2426                                    IB_QPT_RC, listen_id);
2427         if (IS_ERR(conn_id)) {
2428                 ret = -ENOMEM;
2429                 goto out;
2430         }
2431         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2432         conn_id->state = RDMA_CM_CONNECT;
2433
2434         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2435         if (ret) {
2436                 mutex_unlock(&listen_id->handler_mutex);
2437                 destroy_id_handler_unlock(conn_id);
2438                 return ret;
2439         }
2440
2441         ret = cma_iw_acquire_dev(conn_id, listen_id);
2442         if (ret) {
2443                 mutex_unlock(&listen_id->handler_mutex);
2444                 destroy_id_handler_unlock(conn_id);
2445                 return ret;
2446         }
2447
2448         conn_id->cm_id.iw = cm_id;
2449         cm_id->context = conn_id;
2450         cm_id->cm_handler = cma_iw_handler;
2451
2452         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2453         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2454
2455         ret = cma_cm_event_handler(conn_id, &event);
2456         if (ret) {
2457                 /* User wants to destroy the CM ID */
2458                 conn_id->cm_id.iw = NULL;
2459                 mutex_unlock(&listen_id->handler_mutex);
2460                 destroy_id_handler_unlock(conn_id);
2461                 return ret;
2462         }
2463
2464         mutex_unlock(&conn_id->handler_mutex);
2465
2466 out:
2467         mutex_unlock(&listen_id->handler_mutex);
2468         return ret;
2469 }
2470
2471 static int cma_ib_listen(struct rdma_id_private *id_priv)
2472 {
2473         struct sockaddr *addr;
2474         struct ib_cm_id *id;
2475         __be64 svc_id;
2476
2477         addr = cma_src_addr(id_priv);
2478         svc_id = rdma_get_service_id(&id_priv->id, addr);
2479         id = ib_cm_insert_listen(id_priv->id.device,
2480                                  cma_ib_req_handler, svc_id);
2481         if (IS_ERR(id))
2482                 return PTR_ERR(id);
2483         id_priv->cm_id.ib = id;
2484
2485         return 0;
2486 }
2487
2488 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2489 {
2490         int ret;
2491         struct iw_cm_id *id;
2492
2493         id = iw_create_cm_id(id_priv->id.device,
2494                              iw_conn_req_handler,
2495                              id_priv);
2496         if (IS_ERR(id))
2497                 return PTR_ERR(id);
2498
2499         mutex_lock(&id_priv->qp_mutex);
2500         id->tos = id_priv->tos;
2501         id->tos_set = id_priv->tos_set;
2502         mutex_unlock(&id_priv->qp_mutex);
2503         id_priv->cm_id.iw = id;
2504
2505         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2506                rdma_addr_size(cma_src_addr(id_priv)));
2507
2508         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2509
2510         if (ret) {
2511                 iw_destroy_cm_id(id_priv->cm_id.iw);
2512                 id_priv->cm_id.iw = NULL;
2513         }
2514
2515         return ret;
2516 }
2517
2518 static int cma_listen_handler(struct rdma_cm_id *id,
2519                               struct rdma_cm_event *event)
2520 {
2521         struct rdma_id_private *id_priv = id->context;
2522
2523         /* Listening IDs are always destroyed on removal */
2524         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2525                 return -1;
2526
2527         id->context = id_priv->id.context;
2528         id->event_handler = id_priv->id.event_handler;
2529         trace_cm_event_handler(id_priv, event);
2530         return id_priv->id.event_handler(id, event);
2531 }
2532
2533 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2534                              struct cma_device *cma_dev,
2535                              struct rdma_id_private **to_destroy)
2536 {
2537         struct rdma_id_private *dev_id_priv;
2538         struct net *net = id_priv->id.route.addr.dev_addr.net;
2539         int ret;
2540
2541         lockdep_assert_held(&lock);
2542
2543         *to_destroy = NULL;
2544         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2545                 return 0;
2546
2547         dev_id_priv =
2548                 __rdma_create_id(net, cma_listen_handler, id_priv,
2549                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2550         if (IS_ERR(dev_id_priv))
2551                 return PTR_ERR(dev_id_priv);
2552
2553         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2554         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2555                rdma_addr_size(cma_src_addr(id_priv)));
2556
2557         _cma_attach_to_dev(dev_id_priv, cma_dev);
2558         rdma_restrack_add(&dev_id_priv->res);
2559         cma_id_get(id_priv);
2560         dev_id_priv->internal_id = 1;
2561         dev_id_priv->afonly = id_priv->afonly;
2562         mutex_lock(&id_priv->qp_mutex);
2563         dev_id_priv->tos_set = id_priv->tos_set;
2564         dev_id_priv->tos = id_priv->tos;
2565         mutex_unlock(&id_priv->qp_mutex);
2566
2567         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2568         if (ret)
2569                 goto err_listen;
2570         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2571         return 0;
2572 err_listen:
2573         /* Caller must destroy this after releasing lock */
2574         *to_destroy = dev_id_priv;
2575         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2576         return ret;
2577 }
2578
2579 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2580 {
2581         struct rdma_id_private *to_destroy;
2582         struct cma_device *cma_dev;
2583         int ret;
2584
2585         mutex_lock(&lock);
2586         list_add_tail(&id_priv->list, &listen_any_list);
2587         list_for_each_entry(cma_dev, &dev_list, list) {
2588                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2589                 if (ret) {
2590                         /* Prevent racing with cma_process_remove() */
2591                         if (to_destroy)
2592                                 list_del_init(&to_destroy->list);
2593                         goto err_listen;
2594                 }
2595         }
2596         mutex_unlock(&lock);
2597         return 0;
2598
2599 err_listen:
2600         _cma_cancel_listens(id_priv);
2601         mutex_unlock(&lock);
2602         if (to_destroy)
2603                 rdma_destroy_id(&to_destroy->id);
2604         return ret;
2605 }
2606
2607 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2608 {
2609         struct rdma_id_private *id_priv;
2610
2611         id_priv = container_of(id, struct rdma_id_private, id);
2612         mutex_lock(&id_priv->qp_mutex);
2613         id_priv->tos = (u8) tos;
2614         id_priv->tos_set = true;
2615         mutex_unlock(&id_priv->qp_mutex);
2616 }
2617 EXPORT_SYMBOL(rdma_set_service_type);
2618
2619 /**
2620  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2621  *                          with a connection identifier.
2622  * @id: Communication identifier to associated with service type.
2623  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2624  *
2625  * This function should be called before rdma_connect() on active side,
2626  * and on passive side before rdma_accept(). It is applicable to primary
2627  * path only. The timeout will affect the local side of the QP, it is not
2628  * negotiated with remote side and zero disables the timer. In case it is
2629  * set before rdma_resolve_route, the value will also be used to determine
2630  * PacketLifeTime for RoCE.
2631  *
2632  * Return: 0 for success
2633  */
2634 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2635 {
2636         struct rdma_id_private *id_priv;
2637
2638         if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2639                 return -EINVAL;
2640
2641         id_priv = container_of(id, struct rdma_id_private, id);
2642         mutex_lock(&id_priv->qp_mutex);
2643         id_priv->timeout = timeout;
2644         id_priv->timeout_set = true;
2645         mutex_unlock(&id_priv->qp_mutex);
2646
2647         return 0;
2648 }
2649 EXPORT_SYMBOL(rdma_set_ack_timeout);
2650
2651 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2652                               void *context)
2653 {
2654         struct cma_work *work = context;
2655         struct rdma_route *route;
2656
2657         route = &work->id->id.route;
2658
2659         if (!status) {
2660                 route->num_paths = 1;
2661                 *route->path_rec = *path_rec;
2662         } else {
2663                 work->old_state = RDMA_CM_ROUTE_QUERY;
2664                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2665                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2666                 work->event.status = status;
2667                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2668                                      status);
2669         }
2670
2671         queue_work(cma_wq, &work->work);
2672 }
2673
2674 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2675                               unsigned long timeout_ms, struct cma_work *work)
2676 {
2677         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2678         struct sa_path_rec path_rec;
2679         ib_sa_comp_mask comp_mask;
2680         struct sockaddr_in6 *sin6;
2681         struct sockaddr_ib *sib;
2682
2683         memset(&path_rec, 0, sizeof path_rec);
2684
2685         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2686                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2687         else
2688                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2689         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2690         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2691         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2692         path_rec.numb_path = 1;
2693         path_rec.reversible = 1;
2694         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2695                                                   cma_dst_addr(id_priv));
2696
2697         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2698                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2699                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2700
2701         switch (cma_family(id_priv)) {
2702         case AF_INET:
2703                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2704                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2705                 break;
2706         case AF_INET6:
2707                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2708                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2709                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2710                 break;
2711         case AF_IB:
2712                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2713                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2714                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2715                 break;
2716         }
2717
2718         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2719                                                id_priv->id.port_num, &path_rec,
2720                                                comp_mask, timeout_ms,
2721                                                GFP_KERNEL, cma_query_handler,
2722                                                work, &id_priv->query);
2723
2724         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2725 }
2726
2727 static void cma_iboe_join_work_handler(struct work_struct *work)
2728 {
2729         struct cma_multicast *mc =
2730                 container_of(work, struct cma_multicast, iboe_join.work);
2731         struct rdma_cm_event *event = &mc->iboe_join.event;
2732         struct rdma_id_private *id_priv = mc->id_priv;
2733         int ret;
2734
2735         mutex_lock(&id_priv->handler_mutex);
2736         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2737             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2738                 goto out_unlock;
2739
2740         ret = cma_cm_event_handler(id_priv, event);
2741         WARN_ON(ret);
2742
2743 out_unlock:
2744         mutex_unlock(&id_priv->handler_mutex);
2745         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2746                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2747 }
2748
2749 static void cma_work_handler(struct work_struct *_work)
2750 {
2751         struct cma_work *work = container_of(_work, struct cma_work, work);
2752         struct rdma_id_private *id_priv = work->id;
2753
2754         mutex_lock(&id_priv->handler_mutex);
2755         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2756             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2757                 goto out_unlock;
2758         if (work->old_state != 0 || work->new_state != 0) {
2759                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2760                         goto out_unlock;
2761         }
2762
2763         if (cma_cm_event_handler(id_priv, &work->event)) {
2764                 cma_id_put(id_priv);
2765                 destroy_id_handler_unlock(id_priv);
2766                 goto out_free;
2767         }
2768
2769 out_unlock:
2770         mutex_unlock(&id_priv->handler_mutex);
2771         cma_id_put(id_priv);
2772 out_free:
2773         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2774                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2775         kfree(work);
2776 }
2777
2778 static void cma_init_resolve_route_work(struct cma_work *work,
2779                                         struct rdma_id_private *id_priv)
2780 {
2781         work->id = id_priv;
2782         INIT_WORK(&work->work, cma_work_handler);
2783         work->old_state = RDMA_CM_ROUTE_QUERY;
2784         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2785         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2786 }
2787
2788 static void enqueue_resolve_addr_work(struct cma_work *work,
2789                                       struct rdma_id_private *id_priv)
2790 {
2791         /* Balances with cma_id_put() in cma_work_handler */
2792         cma_id_get(id_priv);
2793
2794         work->id = id_priv;
2795         INIT_WORK(&work->work, cma_work_handler);
2796         work->old_state = RDMA_CM_ADDR_QUERY;
2797         work->new_state = RDMA_CM_ADDR_RESOLVED;
2798         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2799
2800         queue_work(cma_wq, &work->work);
2801 }
2802
2803 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2804                                 unsigned long timeout_ms)
2805 {
2806         struct rdma_route *route = &id_priv->id.route;
2807         struct cma_work *work;
2808         int ret;
2809
2810         work = kzalloc(sizeof *work, GFP_KERNEL);
2811         if (!work)
2812                 return -ENOMEM;
2813
2814         cma_init_resolve_route_work(work, id_priv);
2815
2816         if (!route->path_rec)
2817                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2818         if (!route->path_rec) {
2819                 ret = -ENOMEM;
2820                 goto err1;
2821         }
2822
2823         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2824         if (ret)
2825                 goto err2;
2826
2827         return 0;
2828 err2:
2829         kfree(route->path_rec);
2830         route->path_rec = NULL;
2831 err1:
2832         kfree(work);
2833         return ret;
2834 }
2835
2836 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2837                                            unsigned long supported_gids,
2838                                            enum ib_gid_type default_gid)
2839 {
2840         if ((network_type == RDMA_NETWORK_IPV4 ||
2841              network_type == RDMA_NETWORK_IPV6) &&
2842             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2843                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2844
2845         return default_gid;
2846 }
2847
2848 /*
2849  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2850  * path record type based on GID type.
2851  * It also sets up other L2 fields which includes destination mac address
2852  * netdev ifindex, of the path record.
2853  * It returns the netdev of the bound interface for this path record entry.
2854  */
2855 static struct net_device *
2856 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2857 {
2858         struct rdma_route *route = &id_priv->id.route;
2859         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2860         struct rdma_addr *addr = &route->addr;
2861         unsigned long supported_gids;
2862         struct net_device *ndev;
2863
2864         if (!addr->dev_addr.bound_dev_if)
2865                 return NULL;
2866
2867         ndev = dev_get_by_index(addr->dev_addr.net,
2868                                 addr->dev_addr.bound_dev_if);
2869         if (!ndev)
2870                 return NULL;
2871
2872         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2873                                                     id_priv->id.port_num);
2874         gid_type = cma_route_gid_type(addr->dev_addr.network,
2875                                       supported_gids,
2876                                       id_priv->gid_type);
2877         /* Use the hint from IP Stack to select GID Type */
2878         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2879                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2880         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2881
2882         route->path_rec->roce.route_resolved = true;
2883         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2884         return ndev;
2885 }
2886
2887 int rdma_set_ib_path(struct rdma_cm_id *id,
2888                      struct sa_path_rec *path_rec)
2889 {
2890         struct rdma_id_private *id_priv;
2891         struct net_device *ndev;
2892         int ret;
2893
2894         id_priv = container_of(id, struct rdma_id_private, id);
2895         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2896                            RDMA_CM_ROUTE_RESOLVED))
2897                 return -EINVAL;
2898
2899         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2900                                      GFP_KERNEL);
2901         if (!id->route.path_rec) {
2902                 ret = -ENOMEM;
2903                 goto err;
2904         }
2905
2906         if (rdma_protocol_roce(id->device, id->port_num)) {
2907                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2908                 if (!ndev) {
2909                         ret = -ENODEV;
2910                         goto err_free;
2911                 }
2912                 dev_put(ndev);
2913         }
2914
2915         id->route.num_paths = 1;
2916         return 0;
2917
2918 err_free:
2919         kfree(id->route.path_rec);
2920         id->route.path_rec = NULL;
2921 err:
2922         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2923         return ret;
2924 }
2925 EXPORT_SYMBOL(rdma_set_ib_path);
2926
2927 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2928 {
2929         struct cma_work *work;
2930
2931         work = kzalloc(sizeof *work, GFP_KERNEL);
2932         if (!work)
2933                 return -ENOMEM;
2934
2935         cma_init_resolve_route_work(work, id_priv);
2936         queue_work(cma_wq, &work->work);
2937         return 0;
2938 }
2939
2940 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2941 {
2942         struct net_device *dev;
2943
2944         dev = vlan_dev_real_dev(vlan_ndev);
2945         if (dev->num_tc)
2946                 return netdev_get_prio_tc_map(dev, prio);
2947
2948         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2949                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2950 }
2951
2952 struct iboe_prio_tc_map {
2953         int input_prio;
2954         int output_tc;
2955         bool found;
2956 };
2957
2958 static int get_lower_vlan_dev_tc(struct net_device *dev,
2959                                  struct netdev_nested_priv *priv)
2960 {
2961         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2962
2963         if (is_vlan_dev(dev))
2964                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2965         else if (dev->num_tc)
2966                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2967         else
2968                 map->output_tc = 0;
2969         /* We are interested only in first level VLAN device, so always
2970          * return 1 to stop iterating over next level devices.
2971          */
2972         map->found = true;
2973         return 1;
2974 }
2975
2976 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2977 {
2978         struct iboe_prio_tc_map prio_tc_map = {};
2979         int prio = rt_tos2priority(tos);
2980         struct netdev_nested_priv priv;
2981
2982         /* If VLAN device, get it directly from the VLAN netdev */
2983         if (is_vlan_dev(ndev))
2984                 return get_vlan_ndev_tc(ndev, prio);
2985
2986         prio_tc_map.input_prio = prio;
2987         priv.data = (void *)&prio_tc_map;
2988         rcu_read_lock();
2989         netdev_walk_all_lower_dev_rcu(ndev,
2990                                       get_lower_vlan_dev_tc,
2991                                       &priv);
2992         rcu_read_unlock();
2993         /* If map is found from lower device, use it; Otherwise
2994          * continue with the current netdevice to get priority to tc map.
2995          */
2996         if (prio_tc_map.found)
2997                 return prio_tc_map.output_tc;
2998         else if (ndev->num_tc)
2999                 return netdev_get_prio_tc_map(ndev, prio);
3000         else
3001                 return 0;
3002 }
3003
3004 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3005 {
3006         struct sockaddr_in6 *addr6;
3007         u16 dport, sport;
3008         u32 hash, fl;
3009
3010         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3011         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3012         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3013                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3014                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3015                 hash = (u32)sport * 31 + dport;
3016                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3017         }
3018
3019         return cpu_to_be32(fl);
3020 }
3021
3022 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3023 {
3024         struct rdma_route *route = &id_priv->id.route;
3025         struct rdma_addr *addr = &route->addr;
3026         struct cma_work *work;
3027         int ret;
3028         struct net_device *ndev;
3029
3030         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3031                                         rdma_start_port(id_priv->cma_dev->device)];
3032         u8 tos;
3033
3034         mutex_lock(&id_priv->qp_mutex);
3035         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3036         mutex_unlock(&id_priv->qp_mutex);
3037
3038         work = kzalloc(sizeof *work, GFP_KERNEL);
3039         if (!work)
3040                 return -ENOMEM;
3041
3042         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3043         if (!route->path_rec) {
3044                 ret = -ENOMEM;
3045                 goto err1;
3046         }
3047
3048         route->num_paths = 1;
3049
3050         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3051         if (!ndev) {
3052                 ret = -ENODEV;
3053                 goto err2;
3054         }
3055
3056         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3057                     &route->path_rec->sgid);
3058         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3059                     &route->path_rec->dgid);
3060
3061         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3062                 /* TODO: get the hoplimit from the inet/inet6 device */
3063                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3064         else
3065                 route->path_rec->hop_limit = 1;
3066         route->path_rec->reversible = 1;
3067         route->path_rec->pkey = cpu_to_be16(0xffff);
3068         route->path_rec->mtu_selector = IB_SA_EQ;
3069         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3070         route->path_rec->traffic_class = tos;
3071         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3072         route->path_rec->rate_selector = IB_SA_EQ;
3073         route->path_rec->rate = iboe_get_rate(ndev);
3074         dev_put(ndev);
3075         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3076         /* In case ACK timeout is set, use this value to calculate
3077          * PacketLifeTime.  As per IBTA 12.7.34,
3078          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3079          * Assuming a negligible local ACK delay, we can use
3080          * PacketLifeTime = local ACK timeout/2
3081          * as a reasonable approximation for RoCE networks.
3082          */
3083         mutex_lock(&id_priv->qp_mutex);
3084         if (id_priv->timeout_set && id_priv->timeout)
3085                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3086         else
3087                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3088         mutex_unlock(&id_priv->qp_mutex);
3089
3090         if (!route->path_rec->mtu) {
3091                 ret = -EINVAL;
3092                 goto err2;
3093         }
3094
3095         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3096                                          id_priv->id.port_num))
3097                 route->path_rec->flow_label =
3098                         cma_get_roce_udp_flow_label(id_priv);
3099
3100         cma_init_resolve_route_work(work, id_priv);
3101         queue_work(cma_wq, &work->work);
3102
3103         return 0;
3104
3105 err2:
3106         kfree(route->path_rec);
3107         route->path_rec = NULL;
3108         route->num_paths = 0;
3109 err1:
3110         kfree(work);
3111         return ret;
3112 }
3113
3114 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3115 {
3116         struct rdma_id_private *id_priv;
3117         int ret;
3118
3119         id_priv = container_of(id, struct rdma_id_private, id);
3120         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3121                 return -EINVAL;
3122
3123         cma_id_get(id_priv);
3124         if (rdma_cap_ib_sa(id->device, id->port_num))
3125                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3126         else if (rdma_protocol_roce(id->device, id->port_num))
3127                 ret = cma_resolve_iboe_route(id_priv);
3128         else if (rdma_protocol_iwarp(id->device, id->port_num))
3129                 ret = cma_resolve_iw_route(id_priv);
3130         else
3131                 ret = -ENOSYS;
3132
3133         if (ret)
3134                 goto err;
3135
3136         return 0;
3137 err:
3138         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3139         cma_id_put(id_priv);
3140         return ret;
3141 }
3142 EXPORT_SYMBOL(rdma_resolve_route);
3143
3144 static void cma_set_loopback(struct sockaddr *addr)
3145 {
3146         switch (addr->sa_family) {
3147         case AF_INET:
3148                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3149                 break;
3150         case AF_INET6:
3151                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3152                               0, 0, 0, htonl(1));
3153                 break;
3154         default:
3155                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3156                             0, 0, 0, htonl(1));
3157                 break;
3158         }
3159 }
3160
3161 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3162 {
3163         struct cma_device *cma_dev, *cur_dev;
3164         union ib_gid gid;
3165         enum ib_port_state port_state;
3166         unsigned int p;
3167         u16 pkey;
3168         int ret;
3169
3170         cma_dev = NULL;
3171         mutex_lock(&lock);
3172         list_for_each_entry(cur_dev, &dev_list, list) {
3173                 if (cma_family(id_priv) == AF_IB &&
3174                     !rdma_cap_ib_cm(cur_dev->device, 1))
3175                         continue;
3176
3177                 if (!cma_dev)
3178                         cma_dev = cur_dev;
3179
3180                 rdma_for_each_port (cur_dev->device, p) {
3181                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3182                             port_state == IB_PORT_ACTIVE) {
3183                                 cma_dev = cur_dev;
3184                                 goto port_found;
3185                         }
3186                 }
3187         }
3188
3189         if (!cma_dev) {
3190                 ret = -ENODEV;
3191                 goto out;
3192         }
3193
3194         p = 1;
3195
3196 port_found:
3197         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3198         if (ret)
3199                 goto out;
3200
3201         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3202         if (ret)
3203                 goto out;
3204
3205         id_priv->id.route.addr.dev_addr.dev_type =
3206                 (rdma_protocol_ib(cma_dev->device, p)) ?
3207                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3208
3209         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3210         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3211         id_priv->id.port_num = p;
3212         cma_attach_to_dev(id_priv, cma_dev);
3213         rdma_restrack_add(&id_priv->res);
3214         cma_set_loopback(cma_src_addr(id_priv));
3215 out:
3216         mutex_unlock(&lock);
3217         return ret;
3218 }
3219
3220 static void addr_handler(int status, struct sockaddr *src_addr,
3221                          struct rdma_dev_addr *dev_addr, void *context)
3222 {
3223         struct rdma_id_private *id_priv = context;
3224         struct rdma_cm_event event = {};
3225         struct sockaddr *addr;
3226         struct sockaddr_storage old_addr;
3227
3228         mutex_lock(&id_priv->handler_mutex);
3229         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3230                            RDMA_CM_ADDR_RESOLVED))
3231                 goto out;
3232
3233         /*
3234          * Store the previous src address, so that if we fail to acquire
3235          * matching rdma device, old address can be restored back, which helps
3236          * to cancel the cma listen operation correctly.
3237          */
3238         addr = cma_src_addr(id_priv);
3239         memcpy(&old_addr, addr, rdma_addr_size(addr));
3240         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3241         if (!status && !id_priv->cma_dev) {
3242                 status = cma_acquire_dev_by_src_ip(id_priv);
3243                 if (status)
3244                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3245                                              status);
3246                 rdma_restrack_add(&id_priv->res);
3247         } else if (status) {
3248                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3249         }
3250
3251         if (status) {
3252                 memcpy(addr, &old_addr,
3253                        rdma_addr_size((struct sockaddr *)&old_addr));
3254                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3255                                    RDMA_CM_ADDR_BOUND))
3256                         goto out;
3257                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3258                 event.status = status;
3259         } else
3260                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3261
3262         if (cma_cm_event_handler(id_priv, &event)) {
3263                 destroy_id_handler_unlock(id_priv);
3264                 return;
3265         }
3266 out:
3267         mutex_unlock(&id_priv->handler_mutex);
3268 }
3269
3270 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3271 {
3272         struct cma_work *work;
3273         union ib_gid gid;
3274         int ret;
3275
3276         work = kzalloc(sizeof *work, GFP_KERNEL);
3277         if (!work)
3278                 return -ENOMEM;
3279
3280         if (!id_priv->cma_dev) {
3281                 ret = cma_bind_loopback(id_priv);
3282                 if (ret)
3283                         goto err;
3284         }
3285
3286         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3287         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3288
3289         enqueue_resolve_addr_work(work, id_priv);
3290         return 0;
3291 err:
3292         kfree(work);
3293         return ret;
3294 }
3295
3296 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3297 {
3298         struct cma_work *work;
3299         int ret;
3300
3301         work = kzalloc(sizeof *work, GFP_KERNEL);
3302         if (!work)
3303                 return -ENOMEM;
3304
3305         if (!id_priv->cma_dev) {
3306                 ret = cma_resolve_ib_dev(id_priv);
3307                 if (ret)
3308                         goto err;
3309         }
3310
3311         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3312                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3313
3314         enqueue_resolve_addr_work(work, id_priv);
3315         return 0;
3316 err:
3317         kfree(work);
3318         return ret;
3319 }
3320
3321 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3322                          const struct sockaddr *dst_addr)
3323 {
3324         struct sockaddr_storage zero_sock = {};
3325
3326         if (src_addr && src_addr->sa_family)
3327                 return rdma_bind_addr(id, src_addr);
3328
3329         /*
3330          * When the src_addr is not specified, automatically supply an any addr
3331          */
3332         zero_sock.ss_family = dst_addr->sa_family;
3333         if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
3334                 struct sockaddr_in6 *src_addr6 =
3335                         (struct sockaddr_in6 *)&zero_sock;
3336                 struct sockaddr_in6 *dst_addr6 =
3337                         (struct sockaddr_in6 *)dst_addr;
3338
3339                 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3340                 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3341                         id->route.addr.dev_addr.bound_dev_if =
3342                                 dst_addr6->sin6_scope_id;
3343         } else if (dst_addr->sa_family == AF_IB) {
3344                 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
3345                         ((struct sockaddr_ib *)dst_addr)->sib_pkey;
3346         }
3347         return rdma_bind_addr(id, (struct sockaddr *)&zero_sock);
3348 }
3349
3350 /*
3351  * If required, resolve the source address for bind and leave the id_priv in
3352  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3353  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3354  * ignored.
3355  */
3356 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3357                                struct sockaddr *src_addr,
3358                                const struct sockaddr *dst_addr)
3359 {
3360         int ret;
3361
3362         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3363         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3364                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3365                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3366                 if (ret)
3367                         goto err_dst;
3368                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3369                                            RDMA_CM_ADDR_QUERY))) {
3370                         ret = -EINVAL;
3371                         goto err_dst;
3372                 }
3373         }
3374
3375         if (cma_family(id_priv) != dst_addr->sa_family) {
3376                 ret = -EINVAL;
3377                 goto err_state;
3378         }
3379         return 0;
3380
3381 err_state:
3382         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3383 err_dst:
3384         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3385         return ret;
3386 }
3387
3388 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3389                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3390 {
3391         struct rdma_id_private *id_priv =
3392                 container_of(id, struct rdma_id_private, id);
3393         int ret;
3394
3395         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3396         if (ret)
3397                 return ret;
3398
3399         if (cma_any_addr(dst_addr)) {
3400                 ret = cma_resolve_loopback(id_priv);
3401         } else {
3402                 if (dst_addr->sa_family == AF_IB) {
3403                         ret = cma_resolve_ib_addr(id_priv);
3404                 } else {
3405                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3406                                               &id->route.addr.dev_addr,
3407                                               timeout_ms, addr_handler,
3408                                               false, id_priv);
3409                 }
3410         }
3411         if (ret)
3412                 goto err;
3413
3414         return 0;
3415 err:
3416         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3417         return ret;
3418 }
3419 EXPORT_SYMBOL(rdma_resolve_addr);
3420
3421 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3422 {
3423         struct rdma_id_private *id_priv;
3424         unsigned long flags;
3425         int ret;
3426
3427         id_priv = container_of(id, struct rdma_id_private, id);
3428         spin_lock_irqsave(&id_priv->lock, flags);
3429         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3430             id_priv->state == RDMA_CM_IDLE) {
3431                 id_priv->reuseaddr = reuse;
3432                 ret = 0;
3433         } else {
3434                 ret = -EINVAL;
3435         }
3436         spin_unlock_irqrestore(&id_priv->lock, flags);
3437         return ret;
3438 }
3439 EXPORT_SYMBOL(rdma_set_reuseaddr);
3440
3441 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3442 {
3443         struct rdma_id_private *id_priv;
3444         unsigned long flags;
3445         int ret;
3446
3447         id_priv = container_of(id, struct rdma_id_private, id);
3448         spin_lock_irqsave(&id_priv->lock, flags);
3449         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3450                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3451                 id_priv->afonly = afonly;
3452                 ret = 0;
3453         } else {
3454                 ret = -EINVAL;
3455         }
3456         spin_unlock_irqrestore(&id_priv->lock, flags);
3457         return ret;
3458 }
3459 EXPORT_SYMBOL(rdma_set_afonly);
3460
3461 static void cma_bind_port(struct rdma_bind_list *bind_list,
3462                           struct rdma_id_private *id_priv)
3463 {
3464         struct sockaddr *addr;
3465         struct sockaddr_ib *sib;
3466         u64 sid, mask;
3467         __be16 port;
3468
3469         lockdep_assert_held(&lock);
3470
3471         addr = cma_src_addr(id_priv);
3472         port = htons(bind_list->port);
3473
3474         switch (addr->sa_family) {
3475         case AF_INET:
3476                 ((struct sockaddr_in *) addr)->sin_port = port;
3477                 break;
3478         case AF_INET6:
3479                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3480                 break;
3481         case AF_IB:
3482                 sib = (struct sockaddr_ib *) addr;
3483                 sid = be64_to_cpu(sib->sib_sid);
3484                 mask = be64_to_cpu(sib->sib_sid_mask);
3485                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3486                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3487                 break;
3488         }
3489         id_priv->bind_list = bind_list;
3490         hlist_add_head(&id_priv->node, &bind_list->owners);
3491 }
3492
3493 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3494                           struct rdma_id_private *id_priv, unsigned short snum)
3495 {
3496         struct rdma_bind_list *bind_list;
3497         int ret;
3498
3499         lockdep_assert_held(&lock);
3500
3501         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3502         if (!bind_list)
3503                 return -ENOMEM;
3504
3505         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3506                            snum);
3507         if (ret < 0)
3508                 goto err;
3509
3510         bind_list->ps = ps;
3511         bind_list->port = snum;
3512         cma_bind_port(bind_list, id_priv);
3513         return 0;
3514 err:
3515         kfree(bind_list);
3516         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3517 }
3518
3519 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3520                               struct rdma_id_private *id_priv)
3521 {
3522         struct rdma_id_private *cur_id;
3523         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3524         struct sockaddr  *saddr = cma_src_addr(id_priv);
3525         __be16 dport = cma_port(daddr);
3526
3527         lockdep_assert_held(&lock);
3528
3529         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3530                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3531                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3532                 __be16 cur_dport = cma_port(cur_daddr);
3533
3534                 if (id_priv == cur_id)
3535                         continue;
3536
3537                 /* different dest port -> unique */
3538                 if (!cma_any_port(daddr) &&
3539                     !cma_any_port(cur_daddr) &&
3540                     (dport != cur_dport))
3541                         continue;
3542
3543                 /* different src address -> unique */
3544                 if (!cma_any_addr(saddr) &&
3545                     !cma_any_addr(cur_saddr) &&
3546                     cma_addr_cmp(saddr, cur_saddr))
3547                         continue;
3548
3549                 /* different dst address -> unique */
3550                 if (!cma_any_addr(daddr) &&
3551                     !cma_any_addr(cur_daddr) &&
3552                     cma_addr_cmp(daddr, cur_daddr))
3553                         continue;
3554
3555                 return -EADDRNOTAVAIL;
3556         }
3557         return 0;
3558 }
3559
3560 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3561                               struct rdma_id_private *id_priv)
3562 {
3563         static unsigned int last_used_port;
3564         int low, high, remaining;
3565         unsigned int rover;
3566         struct net *net = id_priv->id.route.addr.dev_addr.net;
3567
3568         lockdep_assert_held(&lock);
3569
3570         inet_get_local_port_range(net, &low, &high);
3571         remaining = (high - low) + 1;
3572         rover = prandom_u32() % remaining + low;
3573 retry:
3574         if (last_used_port != rover) {
3575                 struct rdma_bind_list *bind_list;
3576                 int ret;
3577
3578                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3579
3580                 if (!bind_list) {
3581                         ret = cma_alloc_port(ps, id_priv, rover);
3582                 } else {
3583                         ret = cma_port_is_unique(bind_list, id_priv);
3584                         if (!ret)
3585                                 cma_bind_port(bind_list, id_priv);
3586                 }
3587                 /*
3588                  * Remember previously used port number in order to avoid
3589                  * re-using same port immediately after it is closed.
3590                  */
3591                 if (!ret)
3592                         last_used_port = rover;
3593                 if (ret != -EADDRNOTAVAIL)
3594                         return ret;
3595         }
3596         if (--remaining) {
3597                 rover++;
3598                 if ((rover < low) || (rover > high))
3599                         rover = low;
3600                 goto retry;
3601         }
3602         return -EADDRNOTAVAIL;
3603 }
3604
3605 /*
3606  * Check that the requested port is available.  This is called when trying to
3607  * bind to a specific port, or when trying to listen on a bound port.  In
3608  * the latter case, the provided id_priv may already be on the bind_list, but
3609  * we still need to check that it's okay to start listening.
3610  */
3611 static int cma_check_port(struct rdma_bind_list *bind_list,
3612                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3613 {
3614         struct rdma_id_private *cur_id;
3615         struct sockaddr *addr, *cur_addr;
3616
3617         lockdep_assert_held(&lock);
3618
3619         addr = cma_src_addr(id_priv);
3620         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3621                 if (id_priv == cur_id)
3622                         continue;
3623
3624                 if (reuseaddr && cur_id->reuseaddr)
3625                         continue;
3626
3627                 cur_addr = cma_src_addr(cur_id);
3628                 if (id_priv->afonly && cur_id->afonly &&
3629                     (addr->sa_family != cur_addr->sa_family))
3630                         continue;
3631
3632                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3633                         return -EADDRNOTAVAIL;
3634
3635                 if (!cma_addr_cmp(addr, cur_addr))
3636                         return -EADDRINUSE;
3637         }
3638         return 0;
3639 }
3640
3641 static int cma_use_port(enum rdma_ucm_port_space ps,
3642                         struct rdma_id_private *id_priv)
3643 {
3644         struct rdma_bind_list *bind_list;
3645         unsigned short snum;
3646         int ret;
3647
3648         lockdep_assert_held(&lock);
3649
3650         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3651         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3652                 return -EACCES;
3653
3654         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3655         if (!bind_list) {
3656                 ret = cma_alloc_port(ps, id_priv, snum);
3657         } else {
3658                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3659                 if (!ret)
3660                         cma_bind_port(bind_list, id_priv);
3661         }
3662         return ret;
3663 }
3664
3665 static enum rdma_ucm_port_space
3666 cma_select_inet_ps(struct rdma_id_private *id_priv)
3667 {
3668         switch (id_priv->id.ps) {
3669         case RDMA_PS_TCP:
3670         case RDMA_PS_UDP:
3671         case RDMA_PS_IPOIB:
3672         case RDMA_PS_IB:
3673                 return id_priv->id.ps;
3674         default:
3675
3676                 return 0;
3677         }
3678 }
3679
3680 static enum rdma_ucm_port_space
3681 cma_select_ib_ps(struct rdma_id_private *id_priv)
3682 {
3683         enum rdma_ucm_port_space ps = 0;
3684         struct sockaddr_ib *sib;
3685         u64 sid_ps, mask, sid;
3686
3687         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3688         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3689         sid = be64_to_cpu(sib->sib_sid) & mask;
3690
3691         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3692                 sid_ps = RDMA_IB_IP_PS_IB;
3693                 ps = RDMA_PS_IB;
3694         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3695                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3696                 sid_ps = RDMA_IB_IP_PS_TCP;
3697                 ps = RDMA_PS_TCP;
3698         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3699                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3700                 sid_ps = RDMA_IB_IP_PS_UDP;
3701                 ps = RDMA_PS_UDP;
3702         }
3703
3704         if (ps) {
3705                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3706                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3707                                                 be64_to_cpu(sib->sib_sid_mask));
3708         }
3709         return ps;
3710 }
3711
3712 static int cma_get_port(struct rdma_id_private *id_priv)
3713 {
3714         enum rdma_ucm_port_space ps;
3715         int ret;
3716
3717         if (cma_family(id_priv) != AF_IB)
3718                 ps = cma_select_inet_ps(id_priv);
3719         else
3720                 ps = cma_select_ib_ps(id_priv);
3721         if (!ps)
3722                 return -EPROTONOSUPPORT;
3723
3724         mutex_lock(&lock);
3725         if (cma_any_port(cma_src_addr(id_priv)))
3726                 ret = cma_alloc_any_port(ps, id_priv);
3727         else
3728                 ret = cma_use_port(ps, id_priv);
3729         mutex_unlock(&lock);
3730
3731         return ret;
3732 }
3733
3734 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3735                                struct sockaddr *addr)
3736 {
3737 #if IS_ENABLED(CONFIG_IPV6)
3738         struct sockaddr_in6 *sin6;
3739
3740         if (addr->sa_family != AF_INET6)
3741                 return 0;
3742
3743         sin6 = (struct sockaddr_in6 *) addr;
3744
3745         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3746                 return 0;
3747
3748         if (!sin6->sin6_scope_id)
3749                         return -EINVAL;
3750
3751         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3752 #endif
3753         return 0;
3754 }
3755
3756 int rdma_listen(struct rdma_cm_id *id, int backlog)
3757 {
3758         struct rdma_id_private *id_priv =
3759                 container_of(id, struct rdma_id_private, id);
3760         int ret;
3761
3762         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3763                 struct sockaddr_in any_in = {
3764                         .sin_family = AF_INET,
3765                         .sin_addr.s_addr = htonl(INADDR_ANY),
3766                 };
3767
3768                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3769                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3770                 if (ret)
3771                         return ret;
3772                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3773                                            RDMA_CM_LISTEN)))
3774                         return -EINVAL;
3775         }
3776
3777         /*
3778          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3779          * any more, and has to be unique in the bind list.
3780          */
3781         if (id_priv->reuseaddr) {
3782                 mutex_lock(&lock);
3783                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3784                 if (!ret)
3785                         id_priv->reuseaddr = 0;
3786                 mutex_unlock(&lock);
3787                 if (ret)
3788                         goto err;
3789         }
3790
3791         id_priv->backlog = backlog;
3792         if (id_priv->cma_dev) {
3793                 if (rdma_cap_ib_cm(id->device, 1)) {
3794                         ret = cma_ib_listen(id_priv);
3795                         if (ret)
3796                                 goto err;
3797                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3798                         ret = cma_iw_listen(id_priv, backlog);
3799                         if (ret)
3800                                 goto err;
3801                 } else {
3802                         ret = -ENOSYS;
3803                         goto err;
3804                 }
3805         } else {
3806                 ret = cma_listen_on_all(id_priv);
3807                 if (ret)
3808                         goto err;
3809         }
3810
3811         return 0;
3812 err:
3813         id_priv->backlog = 0;
3814         /*
3815          * All the failure paths that lead here will not allow the req_handler's
3816          * to have run.
3817          */
3818         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3819         return ret;
3820 }
3821 EXPORT_SYMBOL(rdma_listen);
3822
3823 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3824 {
3825         struct rdma_id_private *id_priv;
3826         int ret;
3827         struct sockaddr  *daddr;
3828
3829         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3830             addr->sa_family != AF_IB)
3831                 return -EAFNOSUPPORT;
3832
3833         id_priv = container_of(id, struct rdma_id_private, id);
3834         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3835                 return -EINVAL;
3836
3837         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3838         if (ret)
3839                 goto err1;
3840
3841         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3842         if (!cma_any_addr(addr)) {
3843                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3844                 if (ret)
3845                         goto err1;
3846
3847                 ret = cma_acquire_dev_by_src_ip(id_priv);
3848                 if (ret)
3849                         goto err1;
3850         }
3851
3852         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3853                 if (addr->sa_family == AF_INET)
3854                         id_priv->afonly = 1;
3855 #if IS_ENABLED(CONFIG_IPV6)
3856                 else if (addr->sa_family == AF_INET6) {
3857                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3858
3859                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3860                 }
3861 #endif
3862         }
3863         daddr = cma_dst_addr(id_priv);
3864         daddr->sa_family = addr->sa_family;
3865
3866         ret = cma_get_port(id_priv);
3867         if (ret)
3868                 goto err2;
3869
3870         if (!cma_any_addr(addr))
3871                 rdma_restrack_add(&id_priv->res);
3872         return 0;
3873 err2:
3874         if (id_priv->cma_dev)
3875                 cma_release_dev(id_priv);
3876 err1:
3877         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3878         return ret;
3879 }
3880 EXPORT_SYMBOL(rdma_bind_addr);
3881
3882 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3883 {
3884         struct cma_hdr *cma_hdr;
3885
3886         cma_hdr = hdr;
3887         cma_hdr->cma_version = CMA_VERSION;
3888         if (cma_family(id_priv) == AF_INET) {
3889                 struct sockaddr_in *src4, *dst4;
3890
3891                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3892                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3893
3894                 cma_set_ip_ver(cma_hdr, 4);
3895                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3896                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3897                 cma_hdr->port = src4->sin_port;
3898         } else if (cma_family(id_priv) == AF_INET6) {
3899                 struct sockaddr_in6 *src6, *dst6;
3900
3901                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3902                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3903
3904                 cma_set_ip_ver(cma_hdr, 6);
3905                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3906                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3907                 cma_hdr->port = src6->sin6_port;
3908         }
3909         return 0;
3910 }
3911
3912 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3913                                 const struct ib_cm_event *ib_event)
3914 {
3915         struct rdma_id_private *id_priv = cm_id->context;
3916         struct rdma_cm_event event = {};
3917         const struct ib_cm_sidr_rep_event_param *rep =
3918                                 &ib_event->param.sidr_rep_rcvd;
3919         int ret;
3920
3921         mutex_lock(&id_priv->handler_mutex);
3922         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3923                 goto out;
3924
3925         switch (ib_event->event) {
3926         case IB_CM_SIDR_REQ_ERROR:
3927                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3928                 event.status = -ETIMEDOUT;
3929                 break;
3930         case IB_CM_SIDR_REP_RECEIVED:
3931                 event.param.ud.private_data = ib_event->private_data;
3932                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3933                 if (rep->status != IB_SIDR_SUCCESS) {
3934                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3935                         event.status = ib_event->param.sidr_rep_rcvd.status;
3936                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3937                                              event.status);
3938                         break;
3939                 }
3940                 ret = cma_set_qkey(id_priv, rep->qkey);
3941                 if (ret) {
3942                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3943                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3944                         event.status = ret;
3945                         break;
3946                 }
3947                 ib_init_ah_attr_from_path(id_priv->id.device,
3948                                           id_priv->id.port_num,
3949                                           id_priv->id.route.path_rec,
3950                                           &event.param.ud.ah_attr,
3951                                           rep->sgid_attr);
3952                 event.param.ud.qp_num = rep->qpn;
3953                 event.param.ud.qkey = rep->qkey;
3954                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3955                 event.status = 0;
3956                 break;
3957         default:
3958                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3959                        ib_event->event);
3960                 goto out;
3961         }
3962
3963         ret = cma_cm_event_handler(id_priv, &event);
3964
3965         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3966         if (ret) {
3967                 /* Destroy the CM ID by returning a non-zero value. */
3968                 id_priv->cm_id.ib = NULL;
3969                 destroy_id_handler_unlock(id_priv);
3970                 return ret;
3971         }
3972 out:
3973         mutex_unlock(&id_priv->handler_mutex);
3974         return 0;
3975 }
3976
3977 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3978                               struct rdma_conn_param *conn_param)
3979 {
3980         struct ib_cm_sidr_req_param req;
3981         struct ib_cm_id *id;
3982         void *private_data;
3983         u8 offset;
3984         int ret;
3985
3986         memset(&req, 0, sizeof req);
3987         offset = cma_user_data_offset(id_priv);
3988         req.private_data_len = offset + conn_param->private_data_len;
3989         if (req.private_data_len < conn_param->private_data_len)
3990                 return -EINVAL;
3991
3992         if (req.private_data_len) {
3993                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3994                 if (!private_data)
3995                         return -ENOMEM;
3996         } else {
3997                 private_data = NULL;
3998         }
3999
4000         if (conn_param->private_data && conn_param->private_data_len)
4001                 memcpy(private_data + offset, conn_param->private_data,
4002                        conn_param->private_data_len);
4003
4004         if (private_data) {
4005                 ret = cma_format_hdr(private_data, id_priv);
4006                 if (ret)
4007                         goto out;
4008                 req.private_data = private_data;
4009         }
4010
4011         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4012                              id_priv);
4013         if (IS_ERR(id)) {
4014                 ret = PTR_ERR(id);
4015                 goto out;
4016         }
4017         id_priv->cm_id.ib = id;
4018
4019         req.path = id_priv->id.route.path_rec;
4020         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4021         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4022         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4023         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4024
4025         trace_cm_send_sidr_req(id_priv);
4026         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4027         if (ret) {
4028                 ib_destroy_cm_id(id_priv->cm_id.ib);
4029                 id_priv->cm_id.ib = NULL;
4030         }
4031 out:
4032         kfree(private_data);
4033         return ret;
4034 }
4035
4036 static int cma_connect_ib(struct rdma_id_private *id_priv,
4037                           struct rdma_conn_param *conn_param)
4038 {
4039         struct ib_cm_req_param req;
4040         struct rdma_route *route;
4041         void *private_data;
4042         struct ib_cm_id *id;
4043         u8 offset;
4044         int ret;
4045
4046         memset(&req, 0, sizeof req);
4047         offset = cma_user_data_offset(id_priv);
4048         req.private_data_len = offset + conn_param->private_data_len;
4049         if (req.private_data_len < conn_param->private_data_len)
4050                 return -EINVAL;
4051
4052         if (req.private_data_len) {
4053                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4054                 if (!private_data)
4055                         return -ENOMEM;
4056         } else {
4057                 private_data = NULL;
4058         }
4059
4060         if (conn_param->private_data && conn_param->private_data_len)
4061                 memcpy(private_data + offset, conn_param->private_data,
4062                        conn_param->private_data_len);
4063
4064         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4065         if (IS_ERR(id)) {
4066                 ret = PTR_ERR(id);
4067                 goto out;
4068         }
4069         id_priv->cm_id.ib = id;
4070
4071         route = &id_priv->id.route;
4072         if (private_data) {
4073                 ret = cma_format_hdr(private_data, id_priv);
4074                 if (ret)
4075                         goto out;
4076                 req.private_data = private_data;
4077         }
4078
4079         req.primary_path = &route->path_rec[0];
4080         if (route->num_paths == 2)
4081                 req.alternate_path = &route->path_rec[1];
4082
4083         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4084         /* Alternate path SGID attribute currently unsupported */
4085         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4086         req.qp_num = id_priv->qp_num;
4087         req.qp_type = id_priv->id.qp_type;
4088         req.starting_psn = id_priv->seq_num;
4089         req.responder_resources = conn_param->responder_resources;
4090         req.initiator_depth = conn_param->initiator_depth;
4091         req.flow_control = conn_param->flow_control;
4092         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4093         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4094         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4095         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4096         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4097         req.srq = id_priv->srq ? 1 : 0;
4098         req.ece.vendor_id = id_priv->ece.vendor_id;
4099         req.ece.attr_mod = id_priv->ece.attr_mod;
4100
4101         trace_cm_send_req(id_priv);
4102         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4103 out:
4104         if (ret && !IS_ERR(id)) {
4105                 ib_destroy_cm_id(id);
4106                 id_priv->cm_id.ib = NULL;
4107         }
4108
4109         kfree(private_data);
4110         return ret;
4111 }
4112
4113 static int cma_connect_iw(struct rdma_id_private *id_priv,
4114                           struct rdma_conn_param *conn_param)
4115 {
4116         struct iw_cm_id *cm_id;
4117         int ret;
4118         struct iw_cm_conn_param iw_param;
4119
4120         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4121         if (IS_ERR(cm_id))
4122                 return PTR_ERR(cm_id);
4123
4124         mutex_lock(&id_priv->qp_mutex);
4125         cm_id->tos = id_priv->tos;
4126         cm_id->tos_set = id_priv->tos_set;
4127         mutex_unlock(&id_priv->qp_mutex);
4128
4129         id_priv->cm_id.iw = cm_id;
4130
4131         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4132                rdma_addr_size(cma_src_addr(id_priv)));
4133         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4134                rdma_addr_size(cma_dst_addr(id_priv)));
4135
4136         ret = cma_modify_qp_rtr(id_priv, conn_param);
4137         if (ret)
4138                 goto out;
4139
4140         if (conn_param) {
4141                 iw_param.ord = conn_param->initiator_depth;
4142                 iw_param.ird = conn_param->responder_resources;
4143                 iw_param.private_data = conn_param->private_data;
4144                 iw_param.private_data_len = conn_param->private_data_len;
4145                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4146         } else {
4147                 memset(&iw_param, 0, sizeof iw_param);
4148                 iw_param.qpn = id_priv->qp_num;
4149         }
4150         ret = iw_cm_connect(cm_id, &iw_param);
4151 out:
4152         if (ret) {
4153                 iw_destroy_cm_id(cm_id);
4154                 id_priv->cm_id.iw = NULL;
4155         }
4156         return ret;
4157 }
4158
4159 /**
4160  * rdma_connect_locked - Initiate an active connection request.
4161  * @id: Connection identifier to connect.
4162  * @conn_param: Connection information used for connected QPs.
4163  *
4164  * Same as rdma_connect() but can only be called from the
4165  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4166  */
4167 int rdma_connect_locked(struct rdma_cm_id *id,
4168                         struct rdma_conn_param *conn_param)
4169 {
4170         struct rdma_id_private *id_priv =
4171                 container_of(id, struct rdma_id_private, id);
4172         int ret;
4173
4174         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4175                 return -EINVAL;
4176
4177         if (!id->qp) {
4178                 id_priv->qp_num = conn_param->qp_num;
4179                 id_priv->srq = conn_param->srq;
4180         }
4181
4182         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4183                 if (id->qp_type == IB_QPT_UD)
4184                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4185                 else
4186                         ret = cma_connect_ib(id_priv, conn_param);
4187         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4188                 ret = cma_connect_iw(id_priv, conn_param);
4189         else
4190                 ret = -ENOSYS;
4191         if (ret)
4192                 goto err_state;
4193         return 0;
4194 err_state:
4195         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4196         return ret;
4197 }
4198 EXPORT_SYMBOL(rdma_connect_locked);
4199
4200 /**
4201  * rdma_connect - Initiate an active connection request.
4202  * @id: Connection identifier to connect.
4203  * @conn_param: Connection information used for connected QPs.
4204  *
4205  * Users must have resolved a route for the rdma_cm_id to connect with by having
4206  * called rdma_resolve_route before calling this routine.
4207  *
4208  * This call will either connect to a remote QP or obtain remote QP information
4209  * for unconnected rdma_cm_id's.  The actual operation is based on the
4210  * rdma_cm_id's port space.
4211  */
4212 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4213 {
4214         struct rdma_id_private *id_priv =
4215                 container_of(id, struct rdma_id_private, id);
4216         int ret;
4217
4218         mutex_lock(&id_priv->handler_mutex);
4219         ret = rdma_connect_locked(id, conn_param);
4220         mutex_unlock(&id_priv->handler_mutex);
4221         return ret;
4222 }
4223 EXPORT_SYMBOL(rdma_connect);
4224
4225 /**
4226  * rdma_connect_ece - Initiate an active connection request with ECE data.
4227  * @id: Connection identifier to connect.
4228  * @conn_param: Connection information used for connected QPs.
4229  * @ece: ECE parameters
4230  *
4231  * See rdma_connect() explanation.
4232  */
4233 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4234                      struct rdma_ucm_ece *ece)
4235 {
4236         struct rdma_id_private *id_priv =
4237                 container_of(id, struct rdma_id_private, id);
4238
4239         id_priv->ece.vendor_id = ece->vendor_id;
4240         id_priv->ece.attr_mod = ece->attr_mod;
4241
4242         return rdma_connect(id, conn_param);
4243 }
4244 EXPORT_SYMBOL(rdma_connect_ece);
4245
4246 static int cma_accept_ib(struct rdma_id_private *id_priv,
4247                          struct rdma_conn_param *conn_param)
4248 {
4249         struct ib_cm_rep_param rep;
4250         int ret;
4251
4252         ret = cma_modify_qp_rtr(id_priv, conn_param);
4253         if (ret)
4254                 goto out;
4255
4256         ret = cma_modify_qp_rts(id_priv, conn_param);
4257         if (ret)
4258                 goto out;
4259
4260         memset(&rep, 0, sizeof rep);
4261         rep.qp_num = id_priv->qp_num;
4262         rep.starting_psn = id_priv->seq_num;
4263         rep.private_data = conn_param->private_data;
4264         rep.private_data_len = conn_param->private_data_len;
4265         rep.responder_resources = conn_param->responder_resources;
4266         rep.initiator_depth = conn_param->initiator_depth;
4267         rep.failover_accepted = 0;
4268         rep.flow_control = conn_param->flow_control;
4269         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4270         rep.srq = id_priv->srq ? 1 : 0;
4271         rep.ece.vendor_id = id_priv->ece.vendor_id;
4272         rep.ece.attr_mod = id_priv->ece.attr_mod;
4273
4274         trace_cm_send_rep(id_priv);
4275         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4276 out:
4277         return ret;
4278 }
4279
4280 static int cma_accept_iw(struct rdma_id_private *id_priv,
4281                   struct rdma_conn_param *conn_param)
4282 {
4283         struct iw_cm_conn_param iw_param;
4284         int ret;
4285
4286         if (!conn_param)
4287                 return -EINVAL;
4288
4289         ret = cma_modify_qp_rtr(id_priv, conn_param);
4290         if (ret)
4291                 return ret;
4292
4293         iw_param.ord = conn_param->initiator_depth;
4294         iw_param.ird = conn_param->responder_resources;
4295         iw_param.private_data = conn_param->private_data;
4296         iw_param.private_data_len = conn_param->private_data_len;
4297         if (id_priv->id.qp) {
4298                 iw_param.qpn = id_priv->qp_num;
4299         } else
4300                 iw_param.qpn = conn_param->qp_num;
4301
4302         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4303 }
4304
4305 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4306                              enum ib_cm_sidr_status status, u32 qkey,
4307                              const void *private_data, int private_data_len)
4308 {
4309         struct ib_cm_sidr_rep_param rep;
4310         int ret;
4311
4312         memset(&rep, 0, sizeof rep);
4313         rep.status = status;
4314         if (status == IB_SIDR_SUCCESS) {
4315                 ret = cma_set_qkey(id_priv, qkey);
4316                 if (ret)
4317                         return ret;
4318                 rep.qp_num = id_priv->qp_num;
4319                 rep.qkey = id_priv->qkey;
4320
4321                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4322                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4323         }
4324
4325         rep.private_data = private_data;
4326         rep.private_data_len = private_data_len;
4327
4328         trace_cm_send_sidr_rep(id_priv);
4329         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4330 }
4331
4332 /**
4333  * rdma_accept - Called to accept a connection request or response.
4334  * @id: Connection identifier associated with the request.
4335  * @conn_param: Information needed to establish the connection.  This must be
4336  *   provided if accepting a connection request.  If accepting a connection
4337  *   response, this parameter must be NULL.
4338  *
4339  * Typically, this routine is only called by the listener to accept a connection
4340  * request.  It must also be called on the active side of a connection if the
4341  * user is performing their own QP transitions.
4342  *
4343  * In the case of error, a reject message is sent to the remote side and the
4344  * state of the qp associated with the id is modified to error, such that any
4345  * previously posted receive buffers would be flushed.
4346  *
4347  * This function is for use by kernel ULPs and must be called from under the
4348  * handler callback.
4349  */
4350 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4351 {
4352         struct rdma_id_private *id_priv =
4353                 container_of(id, struct rdma_id_private, id);
4354         int ret;
4355
4356         lockdep_assert_held(&id_priv->handler_mutex);
4357
4358         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4359                 return -EINVAL;
4360
4361         if (!id->qp && conn_param) {
4362                 id_priv->qp_num = conn_param->qp_num;
4363                 id_priv->srq = conn_param->srq;
4364         }
4365
4366         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4367                 if (id->qp_type == IB_QPT_UD) {
4368                         if (conn_param)
4369                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4370                                                         conn_param->qkey,
4371                                                         conn_param->private_data,
4372                                                         conn_param->private_data_len);
4373                         else
4374                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4375                                                         0, NULL, 0);
4376                 } else {
4377                         if (conn_param)
4378                                 ret = cma_accept_ib(id_priv, conn_param);
4379                         else
4380                                 ret = cma_rep_recv(id_priv);
4381                 }
4382         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4383                 ret = cma_accept_iw(id_priv, conn_param);
4384         else
4385                 ret = -ENOSYS;
4386
4387         if (ret)
4388                 goto reject;
4389
4390         return 0;
4391 reject:
4392         cma_modify_qp_err(id_priv);
4393         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4394         return ret;
4395 }
4396 EXPORT_SYMBOL(rdma_accept);
4397
4398 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4399                     struct rdma_ucm_ece *ece)
4400 {
4401         struct rdma_id_private *id_priv =
4402                 container_of(id, struct rdma_id_private, id);
4403
4404         id_priv->ece.vendor_id = ece->vendor_id;
4405         id_priv->ece.attr_mod = ece->attr_mod;
4406
4407         return rdma_accept(id, conn_param);
4408 }
4409 EXPORT_SYMBOL(rdma_accept_ece);
4410
4411 void rdma_lock_handler(struct rdma_cm_id *id)
4412 {
4413         struct rdma_id_private *id_priv =
4414                 container_of(id, struct rdma_id_private, id);
4415
4416         mutex_lock(&id_priv->handler_mutex);
4417 }
4418 EXPORT_SYMBOL(rdma_lock_handler);
4419
4420 void rdma_unlock_handler(struct rdma_cm_id *id)
4421 {
4422         struct rdma_id_private *id_priv =
4423                 container_of(id, struct rdma_id_private, id);
4424
4425         mutex_unlock(&id_priv->handler_mutex);
4426 }
4427 EXPORT_SYMBOL(rdma_unlock_handler);
4428
4429 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4430 {
4431         struct rdma_id_private *id_priv;
4432         int ret;
4433
4434         id_priv = container_of(id, struct rdma_id_private, id);
4435         if (!id_priv->cm_id.ib)
4436                 return -EINVAL;
4437
4438         switch (id->device->node_type) {
4439         case RDMA_NODE_IB_CA:
4440                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4441                 break;
4442         default:
4443                 ret = 0;
4444                 break;
4445         }
4446         return ret;
4447 }
4448 EXPORT_SYMBOL(rdma_notify);
4449
4450 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4451                 u8 private_data_len, u8 reason)
4452 {
4453         struct rdma_id_private *id_priv;
4454         int ret;
4455
4456         id_priv = container_of(id, struct rdma_id_private, id);
4457         if (!id_priv->cm_id.ib)
4458                 return -EINVAL;
4459
4460         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4461                 if (id->qp_type == IB_QPT_UD) {
4462                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4463                                                 private_data, private_data_len);
4464                 } else {
4465                         trace_cm_send_rej(id_priv);
4466                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4467                                              private_data, private_data_len);
4468                 }
4469         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4470                 ret = iw_cm_reject(id_priv->cm_id.iw,
4471                                    private_data, private_data_len);
4472         } else
4473                 ret = -ENOSYS;
4474
4475         return ret;
4476 }
4477 EXPORT_SYMBOL(rdma_reject);
4478
4479 int rdma_disconnect(struct rdma_cm_id *id)
4480 {
4481         struct rdma_id_private *id_priv;
4482         int ret;
4483
4484         id_priv = container_of(id, struct rdma_id_private, id);
4485         if (!id_priv->cm_id.ib)
4486                 return -EINVAL;
4487
4488         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4489                 ret = cma_modify_qp_err(id_priv);
4490                 if (ret)
4491                         goto out;
4492                 /* Initiate or respond to a disconnect. */
4493                 trace_cm_disconnect(id_priv);
4494                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4495                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4496                                 trace_cm_sent_drep(id_priv);
4497                 } else {
4498                         trace_cm_sent_dreq(id_priv);
4499                 }
4500         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4501                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4502         } else
4503                 ret = -EINVAL;
4504
4505 out:
4506         return ret;
4507 }
4508 EXPORT_SYMBOL(rdma_disconnect);
4509
4510 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4511                               struct ib_sa_multicast *multicast,
4512                               struct rdma_cm_event *event,
4513                               struct cma_multicast *mc)
4514 {
4515         struct rdma_dev_addr *dev_addr;
4516         enum ib_gid_type gid_type;
4517         struct net_device *ndev;
4518
4519         if (!status)
4520                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4521         else
4522                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4523                                      status);
4524
4525         event->status = status;
4526         event->param.ud.private_data = mc->context;
4527         if (status) {
4528                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4529                 return;
4530         }
4531
4532         dev_addr = &id_priv->id.route.addr.dev_addr;
4533         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4534         gid_type =
4535                 id_priv->cma_dev
4536                         ->default_gid_type[id_priv->id.port_num -
4537                                            rdma_start_port(
4538                                                    id_priv->cma_dev->device)];
4539
4540         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4541         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4542                                      &multicast->rec, ndev, gid_type,
4543                                      &event->param.ud.ah_attr)) {
4544                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4545                 goto out;
4546         }
4547
4548         event->param.ud.qp_num = 0xFFFFFF;
4549         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4550
4551 out:
4552         if (ndev)
4553                 dev_put(ndev);
4554 }
4555
4556 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4557 {
4558         struct cma_multicast *mc = multicast->context;
4559         struct rdma_id_private *id_priv = mc->id_priv;
4560         struct rdma_cm_event event = {};
4561         int ret = 0;
4562
4563         mutex_lock(&id_priv->handler_mutex);
4564         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4565             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4566                 goto out;
4567
4568         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4569         ret = cma_cm_event_handler(id_priv, &event);
4570         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4571         WARN_ON(ret);
4572
4573 out:
4574         mutex_unlock(&id_priv->handler_mutex);
4575         return 0;
4576 }
4577
4578 static void cma_set_mgid(struct rdma_id_private *id_priv,
4579                          struct sockaddr *addr, union ib_gid *mgid)
4580 {
4581         unsigned char mc_map[MAX_ADDR_LEN];
4582         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4583         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4584         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4585
4586         if (cma_any_addr(addr)) {
4587                 memset(mgid, 0, sizeof *mgid);
4588         } else if ((addr->sa_family == AF_INET6) &&
4589                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4590                                                                  0xFF10A01B)) {
4591                 /* IPv6 address is an SA assigned MGID. */
4592                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4593         } else if (addr->sa_family == AF_IB) {
4594                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4595         } else if (addr->sa_family == AF_INET6) {
4596                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4597                 if (id_priv->id.ps == RDMA_PS_UDP)
4598                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4599                 *mgid = *(union ib_gid *) (mc_map + 4);
4600         } else {
4601                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4602                 if (id_priv->id.ps == RDMA_PS_UDP)
4603                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4604                 *mgid = *(union ib_gid *) (mc_map + 4);
4605         }
4606 }
4607
4608 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4609                                  struct cma_multicast *mc)
4610 {
4611         struct ib_sa_mcmember_rec rec;
4612         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4613         ib_sa_comp_mask comp_mask;
4614         int ret;
4615
4616         ib_addr_get_mgid(dev_addr, &rec.mgid);
4617         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4618                                      &rec.mgid, &rec);
4619         if (ret)
4620                 return ret;
4621
4622         ret = cma_set_qkey(id_priv, 0);
4623         if (ret)
4624                 return ret;
4625
4626         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4627         rec.qkey = cpu_to_be32(id_priv->qkey);
4628         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4629         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4630         rec.join_state = mc->join_state;
4631
4632         if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4633             (!ib_sa_sendonly_fullmem_support(&sa_client,
4634                                              id_priv->id.device,
4635                                              id_priv->id.port_num))) {
4636                 dev_warn(
4637                         &id_priv->id.device->dev,
4638                         "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4639                         id_priv->id.port_num);
4640                 return -EOPNOTSUPP;
4641         }
4642
4643         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4644                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4645                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4646                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4647                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4648
4649         if (id_priv->id.ps == RDMA_PS_IPOIB)
4650                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4651                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4652                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4653                              IB_SA_MCMEMBER_REC_MTU |
4654                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4655
4656         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4657                                          id_priv->id.port_num, &rec, comp_mask,
4658                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4659         return PTR_ERR_OR_ZERO(mc->sa_mc);
4660 }
4661
4662 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4663                               enum ib_gid_type gid_type)
4664 {
4665         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4666         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4667
4668         if (cma_any_addr(addr)) {
4669                 memset(mgid, 0, sizeof *mgid);
4670         } else if (addr->sa_family == AF_INET6) {
4671                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4672         } else {
4673                 mgid->raw[0] =
4674                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4675                 mgid->raw[1] =
4676                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4677                 mgid->raw[2] = 0;
4678                 mgid->raw[3] = 0;
4679                 mgid->raw[4] = 0;
4680                 mgid->raw[5] = 0;
4681                 mgid->raw[6] = 0;
4682                 mgid->raw[7] = 0;
4683                 mgid->raw[8] = 0;
4684                 mgid->raw[9] = 0;
4685                 mgid->raw[10] = 0xff;
4686                 mgid->raw[11] = 0xff;
4687                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4688         }
4689 }
4690
4691 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4692                                    struct cma_multicast *mc)
4693 {
4694         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4695         int err = 0;
4696         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4697         struct net_device *ndev = NULL;
4698         struct ib_sa_multicast ib;
4699         enum ib_gid_type gid_type;
4700         bool send_only;
4701
4702         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4703
4704         if (cma_zero_addr(addr))
4705                 return -EINVAL;
4706
4707         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4708                    rdma_start_port(id_priv->cma_dev->device)];
4709         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4710
4711         ib.rec.pkey = cpu_to_be16(0xffff);
4712         if (id_priv->id.ps == RDMA_PS_UDP)
4713                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4714
4715         if (dev_addr->bound_dev_if)
4716                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4717         if (!ndev)
4718                 return -ENODEV;
4719
4720         ib.rec.rate = iboe_get_rate(ndev);
4721         ib.rec.hop_limit = 1;
4722         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4723
4724         if (addr->sa_family == AF_INET) {
4725                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4726                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4727                         if (!send_only) {
4728                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4729                                                     true);
4730                         }
4731                 }
4732         } else {
4733                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4734                         err = -ENOTSUPP;
4735         }
4736         dev_put(ndev);
4737         if (err || !ib.rec.mtu)
4738                 return err ?: -EINVAL;
4739
4740         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4741                     &ib.rec.port_gid);
4742         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4743         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4744         queue_work(cma_wq, &mc->iboe_join.work);
4745         return 0;
4746 }
4747
4748 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4749                         u8 join_state, void *context)
4750 {
4751         struct rdma_id_private *id_priv =
4752                 container_of(id, struct rdma_id_private, id);
4753         struct cma_multicast *mc;
4754         int ret;
4755
4756         /* Not supported for kernel QPs */
4757         if (WARN_ON(id->qp))
4758                 return -EINVAL;
4759
4760         /* ULP is calling this wrong. */
4761         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4762                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4763                 return -EINVAL;
4764
4765         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4766         if (!mc)
4767                 return -ENOMEM;
4768
4769         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4770         mc->context = context;
4771         mc->id_priv = id_priv;
4772         mc->join_state = join_state;
4773
4774         if (rdma_protocol_roce(id->device, id->port_num)) {
4775                 ret = cma_iboe_join_multicast(id_priv, mc);
4776                 if (ret)
4777                         goto out_err;
4778         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4779                 ret = cma_join_ib_multicast(id_priv, mc);
4780                 if (ret)
4781                         goto out_err;
4782         } else {
4783                 ret = -ENOSYS;
4784                 goto out_err;
4785         }
4786
4787         spin_lock(&id_priv->lock);
4788         list_add(&mc->list, &id_priv->mc_list);
4789         spin_unlock(&id_priv->lock);
4790
4791         return 0;
4792 out_err:
4793         kfree(mc);
4794         return ret;
4795 }
4796 EXPORT_SYMBOL(rdma_join_multicast);
4797
4798 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4799 {
4800         struct rdma_id_private *id_priv;
4801         struct cma_multicast *mc;
4802
4803         id_priv = container_of(id, struct rdma_id_private, id);
4804         spin_lock_irq(&id_priv->lock);
4805         list_for_each_entry(mc, &id_priv->mc_list, list) {
4806                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4807                         continue;
4808                 list_del(&mc->list);
4809                 spin_unlock_irq(&id_priv->lock);
4810
4811                 WARN_ON(id_priv->cma_dev->device != id->device);
4812                 destroy_mc(id_priv, mc);
4813                 return;
4814         }
4815         spin_unlock_irq(&id_priv->lock);
4816 }
4817 EXPORT_SYMBOL(rdma_leave_multicast);
4818
4819 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4820 {
4821         struct rdma_dev_addr *dev_addr;
4822         struct cma_work *work;
4823
4824         dev_addr = &id_priv->id.route.addr.dev_addr;
4825
4826         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4827             (net_eq(dev_net(ndev), dev_addr->net)) &&
4828             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4829                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4830                         ndev->name, &id_priv->id);
4831                 work = kzalloc(sizeof *work, GFP_KERNEL);
4832                 if (!work)
4833                         return -ENOMEM;
4834
4835                 INIT_WORK(&work->work, cma_work_handler);
4836                 work->id = id_priv;
4837                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4838                 cma_id_get(id_priv);
4839                 queue_work(cma_wq, &work->work);
4840         }
4841
4842         return 0;
4843 }
4844
4845 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4846                                void *ptr)
4847 {
4848         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4849         struct cma_device *cma_dev;
4850         struct rdma_id_private *id_priv;
4851         int ret = NOTIFY_DONE;
4852
4853         if (event != NETDEV_BONDING_FAILOVER)
4854                 return NOTIFY_DONE;
4855
4856         if (!netif_is_bond_master(ndev))
4857                 return NOTIFY_DONE;
4858
4859         mutex_lock(&lock);
4860         list_for_each_entry(cma_dev, &dev_list, list)
4861                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4862                         ret = cma_netdev_change(ndev, id_priv);
4863                         if (ret)
4864                                 goto out;
4865                 }
4866
4867 out:
4868         mutex_unlock(&lock);
4869         return ret;
4870 }
4871
4872 static struct notifier_block cma_nb = {
4873         .notifier_call = cma_netdev_callback
4874 };
4875
4876 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4877 {
4878         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4879         enum rdma_cm_state state;
4880         unsigned long flags;
4881
4882         mutex_lock(&id_priv->handler_mutex);
4883         /* Record that we want to remove the device */
4884         spin_lock_irqsave(&id_priv->lock, flags);
4885         state = id_priv->state;
4886         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4887                 spin_unlock_irqrestore(&id_priv->lock, flags);
4888                 mutex_unlock(&id_priv->handler_mutex);
4889                 cma_id_put(id_priv);
4890                 return;
4891         }
4892         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4893         spin_unlock_irqrestore(&id_priv->lock, flags);
4894
4895         if (cma_cm_event_handler(id_priv, &event)) {
4896                 /*
4897                  * At this point the ULP promises it won't call
4898                  * rdma_destroy_id() concurrently
4899                  */
4900                 cma_id_put(id_priv);
4901                 mutex_unlock(&id_priv->handler_mutex);
4902                 trace_cm_id_destroy(id_priv);
4903                 _destroy_id(id_priv, state);
4904                 return;
4905         }
4906         mutex_unlock(&id_priv->handler_mutex);
4907
4908         /*
4909          * If this races with destroy then the thread that first assigns state
4910          * to a destroying does the cancel.
4911          */
4912         cma_cancel_operation(id_priv, state);
4913         cma_id_put(id_priv);
4914 }
4915
4916 static void cma_process_remove(struct cma_device *cma_dev)
4917 {
4918         mutex_lock(&lock);
4919         while (!list_empty(&cma_dev->id_list)) {
4920                 struct rdma_id_private *id_priv = list_first_entry(
4921                         &cma_dev->id_list, struct rdma_id_private, list);
4922
4923                 list_del(&id_priv->listen_list);
4924                 list_del_init(&id_priv->list);
4925                 cma_id_get(id_priv);
4926                 mutex_unlock(&lock);
4927
4928                 cma_send_device_removal_put(id_priv);
4929
4930                 mutex_lock(&lock);
4931         }
4932         mutex_unlock(&lock);
4933
4934         cma_dev_put(cma_dev);
4935         wait_for_completion(&cma_dev->comp);
4936 }
4937
4938 static int cma_add_one(struct ib_device *device)
4939 {
4940         struct rdma_id_private *to_destroy;
4941         struct cma_device *cma_dev;
4942         struct rdma_id_private *id_priv;
4943         unsigned int i;
4944         unsigned long supported_gids = 0;
4945         int ret;
4946
4947         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4948         if (!cma_dev)
4949                 return -ENOMEM;
4950
4951         cma_dev->device = device;
4952         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4953                                             sizeof(*cma_dev->default_gid_type),
4954                                             GFP_KERNEL);
4955         if (!cma_dev->default_gid_type) {
4956                 ret = -ENOMEM;
4957                 goto free_cma_dev;
4958         }
4959
4960         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4961                                             sizeof(*cma_dev->default_roce_tos),
4962                                             GFP_KERNEL);
4963         if (!cma_dev->default_roce_tos) {
4964                 ret = -ENOMEM;
4965                 goto free_gid_type;
4966         }
4967
4968         rdma_for_each_port (device, i) {
4969                 supported_gids = roce_gid_type_mask_support(device, i);
4970                 WARN_ON(!supported_gids);
4971                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4972                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4973                                 CMA_PREFERRED_ROCE_GID_TYPE;
4974                 else
4975                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4976                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4977                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4978         }
4979
4980         init_completion(&cma_dev->comp);
4981         refcount_set(&cma_dev->refcount, 1);
4982         INIT_LIST_HEAD(&cma_dev->id_list);
4983         ib_set_client_data(device, &cma_client, cma_dev);
4984
4985         mutex_lock(&lock);
4986         list_add_tail(&cma_dev->list, &dev_list);
4987         list_for_each_entry(id_priv, &listen_any_list, list) {
4988                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
4989                 if (ret)
4990                         goto free_listen;
4991         }
4992         mutex_unlock(&lock);
4993
4994         trace_cm_add_one(device);
4995         return 0;
4996
4997 free_listen:
4998         list_del(&cma_dev->list);
4999         mutex_unlock(&lock);
5000
5001         /* cma_process_remove() will delete to_destroy */
5002         cma_process_remove(cma_dev);
5003         kfree(cma_dev->default_roce_tos);
5004 free_gid_type:
5005         kfree(cma_dev->default_gid_type);
5006
5007 free_cma_dev:
5008         kfree(cma_dev);
5009         return ret;
5010 }
5011
5012 static void cma_remove_one(struct ib_device *device, void *client_data)
5013 {
5014         struct cma_device *cma_dev = client_data;
5015
5016         trace_cm_remove_one(device);
5017
5018         mutex_lock(&lock);
5019         list_del(&cma_dev->list);
5020         mutex_unlock(&lock);
5021
5022         cma_process_remove(cma_dev);
5023         kfree(cma_dev->default_roce_tos);
5024         kfree(cma_dev->default_gid_type);
5025         kfree(cma_dev);
5026 }
5027
5028 static int cma_init_net(struct net *net)
5029 {
5030         struct cma_pernet *pernet = cma_pernet(net);
5031
5032         xa_init(&pernet->tcp_ps);
5033         xa_init(&pernet->udp_ps);
5034         xa_init(&pernet->ipoib_ps);
5035         xa_init(&pernet->ib_ps);
5036
5037         return 0;
5038 }
5039
5040 static void cma_exit_net(struct net *net)
5041 {
5042         struct cma_pernet *pernet = cma_pernet(net);
5043
5044         WARN_ON(!xa_empty(&pernet->tcp_ps));
5045         WARN_ON(!xa_empty(&pernet->udp_ps));
5046         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5047         WARN_ON(!xa_empty(&pernet->ib_ps));
5048 }
5049
5050 static struct pernet_operations cma_pernet_operations = {
5051         .init = cma_init_net,
5052         .exit = cma_exit_net,
5053         .id = &cma_pernet_id,
5054         .size = sizeof(struct cma_pernet),
5055 };
5056
5057 static int __init cma_init(void)
5058 {
5059         int ret;
5060
5061         /*
5062          * There is a rare lock ordering dependency in cma_netdev_callback()
5063          * that only happens when bonding is enabled. Teach lockdep that rtnl
5064          * must never be nested under lock so it can find these without having
5065          * to test with bonding.
5066          */
5067         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5068                 rtnl_lock();
5069                 mutex_lock(&lock);
5070                 mutex_unlock(&lock);
5071                 rtnl_unlock();
5072         }
5073
5074         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5075         if (!cma_wq)
5076                 return -ENOMEM;
5077
5078         ret = register_pernet_subsys(&cma_pernet_operations);
5079         if (ret)
5080                 goto err_wq;
5081
5082         ib_sa_register_client(&sa_client);
5083         register_netdevice_notifier(&cma_nb);
5084
5085         ret = ib_register_client(&cma_client);
5086         if (ret)
5087                 goto err;
5088
5089         ret = cma_configfs_init();
5090         if (ret)
5091                 goto err_ib;
5092
5093         return 0;
5094
5095 err_ib:
5096         ib_unregister_client(&cma_client);
5097 err:
5098         unregister_netdevice_notifier(&cma_nb);
5099         ib_sa_unregister_client(&sa_client);
5100         unregister_pernet_subsys(&cma_pernet_operations);
5101 err_wq:
5102         destroy_workqueue(cma_wq);
5103         return ret;
5104 }
5105
5106 static void __exit cma_cleanup(void)
5107 {
5108         cma_configfs_exit();
5109         ib_unregister_client(&cma_client);
5110         unregister_netdevice_notifier(&cma_nb);
5111         ib_sa_unregister_client(&sa_client);
5112         unregister_pernet_subsys(&cma_pernet_operations);
5113         destroy_workqueue(cma_wq);
5114 }
5115
5116 module_init(cma_init);
5117 module_exit(cma_cleanup);