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_MAX_CM_RETRIES 15
  47 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  48 #define CMA_IBOE_PACKET_LIFETIME 18
  49 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  50
  51 static const char * const cma_events[] = {
  52         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  53         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  54         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  55         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  56         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  57         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  58         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  59         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  60         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  61         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  62         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  63         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  64         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  65         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  66         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  67         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  68 };
  69
  70 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
  71                               enum ib_gid_type gid_type);
  72
  73 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  74 {
  75         size_t index = event;
  76
  77         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  78                         cma_events[index] : "unrecognized event";
  79 }
  80 EXPORT_SYMBOL(rdma_event_msg);
  81
  82 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  83                                                 int reason)
  84 {
  85         if (rdma_ib_or_roce(id->device, id->port_num))
  86                 return ibcm_reject_msg(reason);
  87
  88         if (rdma_protocol_iwarp(id->device, id->port_num))
  89                 return iwcm_reject_msg(reason);
  90
  91         WARN_ON_ONCE(1);
  92         return "unrecognized transport";
  93 }
  94 EXPORT_SYMBOL(rdma_reject_msg);
  95
  96 /**
  97  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  98  *                           request.
  99  * @id: Communication identifier that received the REJECT event.
 100  * @reason: Value returned in the REJECT event status field.
 101  */
 102 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 103 {
 104         if (rdma_ib_or_roce(id->device, id->port_num))
 105                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 106
 107         if (rdma_protocol_iwarp(id->device, id->port_num))
 108                 return reason == -ECONNREFUSED;
 109
 110         WARN_ON_ONCE(1);
 111         return false;
 112 }
 113
 114 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 115                                       struct rdma_cm_event *ev, u8 *data_len)
 116 {
 117         const void *p;
 118
 119         if (rdma_is_consumer_reject(id, ev->status)) {
 120                 *data_len = ev->param.conn.private_data_len;
 121                 p = ev->param.conn.private_data;
 122         } else {
 123                 *data_len = 0;
 124                 p = NULL;
 125         }
 126         return p;
 127 }
 128 EXPORT_SYMBOL(rdma_consumer_reject_data);
 129
 130 /**
 131  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 132  * @id: Communication Identifier
 133  */
 134 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 135 {
 136         struct rdma_id_private *id_priv;
 137
 138         id_priv = container_of(id, struct rdma_id_private, id);
 139         if (id->device->node_type == RDMA_NODE_RNIC)
 140                 return id_priv->cm_id.iw;
 141         return NULL;
 142 }
 143 EXPORT_SYMBOL(rdma_iw_cm_id);
 144
 145 /**
 146  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 147  * @res: rdma resource tracking entry pointer
 148  */
 149 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 150 {
 151         struct rdma_id_private *id_priv =
 152                 container_of(res, struct rdma_id_private, res);
 153
 154         return &id_priv->id;
 155 }
 156 EXPORT_SYMBOL(rdma_res_to_id);
 157
 158 static int cma_add_one(struct ib_device *device);
 159 static void cma_remove_one(struct ib_device *device, void *client_data);
 160
 161 static struct ib_client cma_client = {
 162         .name   = "cma",
 163         .add    = cma_add_one,
 164         .remove = cma_remove_one
 165 };
 166
 167 static struct ib_sa_client sa_client;
 168 static LIST_HEAD(dev_list);
 169 static LIST_HEAD(listen_any_list);
 170 static DEFINE_MUTEX(lock);
 171 static struct workqueue_struct *cma_wq;
 172 static unsigned int cma_pernet_id;
 173
 174 struct cma_pernet {
 175         struct xarray tcp_ps;
 176         struct xarray udp_ps;
 177         struct xarray ipoib_ps;
 178         struct xarray ib_ps;
 179 };
 180
 181 static struct cma_pernet *cma_pernet(struct net *net)
 182 {
 183         return net_generic(net, cma_pernet_id);
 184 }
 185
 186 static
 187 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 188 {
 189         struct cma_pernet *pernet = cma_pernet(net);
 190
 191         switch (ps) {
 192         case RDMA_PS_TCP:
 193                 return &pernet->tcp_ps;
 194         case RDMA_PS_UDP:
 195                 return &pernet->udp_ps;
 196         case RDMA_PS_IPOIB:
 197                 return &pernet->ipoib_ps;
 198         case RDMA_PS_IB:
 199                 return &pernet->ib_ps;
 200         default:
 201                 return NULL;
 202         }
 203 }
 204
 205 struct cma_device {
 206         struct list_head        list;
 207         struct ib_device        *device;
 208         struct completion       comp;
 209         refcount_t refcount;
 210         struct list_head        id_list;
 211         enum ib_gid_type        *default_gid_type;
 212         u8                      *default_roce_tos;
 213 };
 214
 215 struct rdma_bind_list {
 216         enum rdma_ucm_port_space ps;
 217         struct hlist_head       owners;
 218         unsigned short          port;
 219 };
 220
 221 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 222                         struct rdma_bind_list *bind_list, int snum)
 223 {
 224         struct xarray *xa = cma_pernet_xa(net, ps);
 225
 226         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 227 }
 228
 229 static struct rdma_bind_list *cma_ps_find(struct net *net,
 230                                           enum rdma_ucm_port_space ps, int snum)
 231 {
 232         struct xarray *xa = cma_pernet_xa(net, ps);
 233
 234         return xa_load(xa, snum);
 235 }
 236
 237 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 238                           int snum)
 239 {
 240         struct xarray *xa = cma_pernet_xa(net, ps);
 241
 242         xa_erase(xa, snum);
 243 }
 244
 245 enum {
 246         CMA_OPTION_AFONLY,
 247 };
 248
 249 void cma_dev_get(struct cma_device *cma_dev)
 250 {
 251         refcount_inc(&cma_dev->refcount);
 252 }
 253
 254 void cma_dev_put(struct cma_device *cma_dev)
 255 {
 256         if (refcount_dec_and_test(&cma_dev->refcount))
 257                 complete(&cma_dev->comp);
 258 }
 259
 260 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 261                                              void               *cookie)
 262 {
 263         struct cma_device *cma_dev;
 264         struct cma_device *found_cma_dev = NULL;
 265
 266         mutex_lock(&lock);
 267
 268         list_for_each_entry(cma_dev, &dev_list, list)
 269                 if (filter(cma_dev->device, cookie)) {
 270                         found_cma_dev = cma_dev;
 271                         break;
 272                 }
 273
 274         if (found_cma_dev)
 275                 cma_dev_get(found_cma_dev);
 276         mutex_unlock(&lock);
 277         return found_cma_dev;
 278 }
 279
 280 int cma_get_default_gid_type(struct cma_device *cma_dev,
 281                              u32 port)
 282 {
 283         if (!rdma_is_port_valid(cma_dev->device, port))
 284                 return -EINVAL;
 285
 286         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 287 }
 288
 289 int cma_set_default_gid_type(struct cma_device *cma_dev,
 290                              u32 port,
 291                              enum ib_gid_type default_gid_type)
 292 {
 293         unsigned long supported_gids;
 294
 295         if (!rdma_is_port_valid(cma_dev->device, port))
 296                 return -EINVAL;
 297
 298         if (default_gid_type == IB_GID_TYPE_IB &&
 299             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 300                 default_gid_type = IB_GID_TYPE_ROCE;
 301
 302         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 303
 304         if (!(supported_gids & 1 << default_gid_type))
 305                 return -EINVAL;
 306
 307         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 308                 default_gid_type;
 309
 310         return 0;
 311 }
 312
 313 int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
 314 {
 315         if (!rdma_is_port_valid(cma_dev->device, port))
 316                 return -EINVAL;
 317
 318         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 319 }
 320
 321 int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
 322                              u8 default_roce_tos)
 323 {
 324         if (!rdma_is_port_valid(cma_dev->device, port))
 325                 return -EINVAL;
 326
 327         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 328                  default_roce_tos;
 329
 330         return 0;
 331 }
 332 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 333 {
 334         return cma_dev->device;
 335 }
 336
 337 /*
 338  * Device removal can occur at anytime, so we need extra handling to
 339  * serialize notifying the user of device removal with other callbacks.
 340  * We do this by disabling removal notification while a callback is in process,
 341  * and reporting it after the callback completes.
 342  */
 343
 344 struct cma_multicast {
 345         struct rdma_id_private *id_priv;
 346         union {
 347                 struct ib_sa_multicast *sa_mc;
 348                 struct {
 349                         struct work_struct work;
 350                         struct rdma_cm_event event;
 351                 } iboe_join;
 352         };
 353         struct list_head        list;
 354         void                    *context;
 355         struct sockaddr_storage addr;
 356         u8                      join_state;
 357 };
 358
 359 struct cma_work {
 360         struct work_struct      work;
 361         struct rdma_id_private  *id;
 362         enum rdma_cm_state      old_state;
 363         enum rdma_cm_state      new_state;
 364         struct rdma_cm_event    event;
 365 };
 366
 367 union cma_ip_addr {
 368         struct in6_addr ip6;
 369         struct {
 370                 __be32 pad[3];
 371                 __be32 addr;
 372         } ip4;
 373 };
 374
 375 struct cma_hdr {
 376         u8 cma_version;
 377         u8 ip_version;  /* IP version: 7:4 */
 378         __be16 port;
 379         union cma_ip_addr src_addr;
 380         union cma_ip_addr dst_addr;
 381 };
 382
 383 #define CMA_VERSION 0x00
 384
 385 struct cma_req_info {
 386         struct sockaddr_storage listen_addr_storage;
 387         struct sockaddr_storage src_addr_storage;
 388         struct ib_device *device;
 389         union ib_gid local_gid;
 390         __be64 service_id;
 391         int port;
 392         bool has_gid;
 393         u16 pkey;
 394 };
 395
 396 static int cma_comp_exch(struct rdma_id_private *id_priv,
 397                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 398 {
 399         unsigned long flags;
 400         int ret;
 401
 402         /*
 403          * The FSM uses a funny double locking where state is protected by both
 404          * the handler_mutex and the spinlock. State is not allowed to change
 405          * to/from a handler_mutex protected value without also holding
 406          * handler_mutex.
 407          */
 408         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 409                 lockdep_assert_held(&id_priv->handler_mutex);
 410
 411         spin_lock_irqsave(&id_priv->lock, flags);
 412         if ((ret = (id_priv->state == comp)))
 413                 id_priv->state = exch;
 414         spin_unlock_irqrestore(&id_priv->lock, flags);
 415         return ret;
 416 }
 417
 418 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 419 {
 420         return hdr->ip_version >> 4;
 421 }
 422
 423 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 424 {
 425         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 426 }
 427
 428 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 429 {
 430         struct in_device *in_dev = NULL;
 431
 432         if (ndev) {
 433                 rtnl_lock();
 434                 in_dev = __in_dev_get_rtnl(ndev);
 435                 if (in_dev) {
 436                         if (join)
 437                                 ip_mc_inc_group(in_dev,
 438                                                 *(__be32 *)(mgid->raw + 12));
 439                         else
 440                                 ip_mc_dec_group(in_dev,
 441                                                 *(__be32 *)(mgid->raw + 12));
 442                 }
 443                 rtnl_unlock();
 444         }
 445         return (in_dev) ? 0 : -ENODEV;
 446 }
 447
 448 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 449                                struct cma_device *cma_dev)
 450 {
 451         cma_dev_get(cma_dev);
 452         id_priv->cma_dev = cma_dev;
 453         id_priv->id.device = cma_dev->device;
 454         id_priv->id.route.addr.dev_addr.transport =
 455                 rdma_node_get_transport(cma_dev->device->node_type);
 456         list_add_tail(&id_priv->device_item, &cma_dev->id_list);
 457
 458         trace_cm_id_attach(id_priv, cma_dev->device);
 459 }
 460
 461 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 462                               struct cma_device *cma_dev)
 463 {
 464         _cma_attach_to_dev(id_priv, cma_dev);
 465         id_priv->gid_type =
 466                 cma_dev->default_gid_type[id_priv->id.port_num -
 467                                           rdma_start_port(cma_dev->device)];
 468 }
 469
 470 static void cma_release_dev(struct rdma_id_private *id_priv)
 471 {
 472         mutex_lock(&lock);
 473         list_del_init(&id_priv->device_item);
 474         cma_dev_put(id_priv->cma_dev);
 475         id_priv->cma_dev = NULL;
 476         id_priv->id.device = NULL;
 477         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 478                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 479                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 480         }
 481         mutex_unlock(&lock);
 482 }
 483
 484 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 485 {
 486         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 487 }
 488
 489 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 490 {
 491         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 492 }
 493
 494 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 495 {
 496         return id_priv->id.route.addr.src_addr.ss_family;
 497 }
 498
 499 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 500 {
 501         struct ib_sa_mcmember_rec rec;
 502         int ret = 0;
 503
 504         if (id_priv->qkey) {
 505                 if (qkey && id_priv->qkey != qkey)
 506                         return -EINVAL;
 507                 return 0;
 508         }
 509
 510         if (qkey) {
 511                 id_priv->qkey = qkey;
 512                 return 0;
 513         }
 514
 515         switch (id_priv->id.ps) {
 516         case RDMA_PS_UDP:
 517         case RDMA_PS_IB:
 518                 id_priv->qkey = RDMA_UDP_QKEY;
 519                 break;
 520         case RDMA_PS_IPOIB:
 521                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 522                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 523                                              id_priv->id.port_num, &rec.mgid,
 524                                              &rec);
 525                 if (!ret)
 526                         id_priv->qkey = be32_to_cpu(rec.qkey);
 527                 break;
 528         default:
 529                 break;
 530         }
 531         return ret;
 532 }
 533
 534 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 535 {
 536         dev_addr->dev_type = ARPHRD_INFINIBAND;
 537         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 538         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 539 }
 540
 541 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 542 {
 543         int ret;
 544
 545         if (addr->sa_family != AF_IB) {
 546                 ret = rdma_translate_ip(addr, dev_addr);
 547         } else {
 548                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 549                 ret = 0;
 550         }
 551
 552         return ret;
 553 }
 554
 555 static const struct ib_gid_attr *
 556 cma_validate_port(struct ib_device *device, u32 port,
 557                   enum ib_gid_type gid_type,
 558                   union ib_gid *gid,
 559                   struct rdma_id_private *id_priv)
 560 {
 561         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 562         int bound_if_index = dev_addr->bound_dev_if;
 563         const struct ib_gid_attr *sgid_attr;
 564         int dev_type = dev_addr->dev_type;
 565         struct net_device *ndev = NULL;
 566
 567         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 568                 return ERR_PTR(-ENODEV);
 569
 570         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 571                 return ERR_PTR(-ENODEV);
 572
 573         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 574                 return ERR_PTR(-ENODEV);
 575
 576         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 577                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 578                 if (!ndev)
 579                         return ERR_PTR(-ENODEV);
 580         } else {
 581                 gid_type = IB_GID_TYPE_IB;
 582         }
 583
 584         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 585         if (ndev)
 586                 dev_put(ndev);
 587         return sgid_attr;
 588 }
 589
 590 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 591                                const struct ib_gid_attr *sgid_attr)
 592 {
 593         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 594         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 595 }
 596
 597 /**
 598  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 599  * based on source ip address.
 600  * @id_priv:    cm_id which should be bound to cma device
 601  *
 602  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 603  * based on source IP address. It returns 0 on success or error code otherwise.
 604  * It is applicable to active and passive side cm_id.
 605  */
 606 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 607 {
 608         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 609         const struct ib_gid_attr *sgid_attr;
 610         union ib_gid gid, iboe_gid, *gidp;
 611         struct cma_device *cma_dev;
 612         enum ib_gid_type gid_type;
 613         int ret = -ENODEV;
 614         u32 port;
 615
 616         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 617             id_priv->id.ps == RDMA_PS_IPOIB)
 618                 return -EINVAL;
 619
 620         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 621                     &iboe_gid);
 622
 623         memcpy(&gid, dev_addr->src_dev_addr +
 624                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 625
 626         mutex_lock(&lock);
 627         list_for_each_entry(cma_dev, &dev_list, list) {
 628                 rdma_for_each_port (cma_dev->device, port) {
 629                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 630                                &iboe_gid : &gid;
 631                         gid_type = cma_dev->default_gid_type[port - 1];
 632                         sgid_attr = cma_validate_port(cma_dev->device, port,
 633                                                       gid_type, gidp, id_priv);
 634                         if (!IS_ERR(sgid_attr)) {
 635                                 id_priv->id.port_num = port;
 636                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 637                                 cma_attach_to_dev(id_priv, cma_dev);
 638                                 ret = 0;
 639                                 goto out;
 640                         }
 641                 }
 642         }
 643 out:
 644         mutex_unlock(&lock);
 645         return ret;
 646 }
 647
 648 /**
 649  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 650  * @id_priv:            cm id to bind to cma device
 651  * @listen_id_priv:     listener cm id to match against
 652  * @req:                Pointer to req structure containaining incoming
 653  *                      request information
 654  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 655  * rdma device matches for listen_id and incoming request. It also verifies
 656  * that a GID table entry is present for the source address.
 657  * Returns 0 on success, or returns error code otherwise.
 658  */
 659 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 660                               const struct rdma_id_private *listen_id_priv,
 661                               struct cma_req_info *req)
 662 {
 663         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 664         const struct ib_gid_attr *sgid_attr;
 665         enum ib_gid_type gid_type;
 666         union ib_gid gid;
 667
 668         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 669             id_priv->id.ps == RDMA_PS_IPOIB)
 670                 return -EINVAL;
 671
 672         if (rdma_protocol_roce(req->device, req->port))
 673                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 674                             &gid);
 675         else
 676                 memcpy(&gid, dev_addr->src_dev_addr +
 677                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 678
 679         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 680         sgid_attr = cma_validate_port(req->device, req->port,
 681                                       gid_type, &gid, id_priv);
 682         if (IS_ERR(sgid_attr))
 683                 return PTR_ERR(sgid_attr);
 684
 685         id_priv->id.port_num = req->port;
 686         cma_bind_sgid_attr(id_priv, sgid_attr);
 687         /* Need to acquire lock to protect against reader
 688          * of cma_dev->id_list such as cma_netdev_callback() and
 689          * cma_process_remove().
 690          */
 691         mutex_lock(&lock);
 692         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 693         mutex_unlock(&lock);
 694         rdma_restrack_add(&id_priv->res);
 695         return 0;
 696 }
 697
 698 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 699                               const struct rdma_id_private *listen_id_priv)
 700 {
 701         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 702         const struct ib_gid_attr *sgid_attr;
 703         struct cma_device *cma_dev;
 704         enum ib_gid_type gid_type;
 705         int ret = -ENODEV;
 706         union ib_gid gid;
 707         u32 port;
 708
 709         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 710             id_priv->id.ps == RDMA_PS_IPOIB)
 711                 return -EINVAL;
 712
 713         memcpy(&gid, dev_addr->src_dev_addr +
 714                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 715
 716         mutex_lock(&lock);
 717
 718         cma_dev = listen_id_priv->cma_dev;
 719         port = listen_id_priv->id.port_num;
 720         gid_type = listen_id_priv->gid_type;
 721         sgid_attr = cma_validate_port(cma_dev->device, port,
 722                                       gid_type, &gid, id_priv);
 723         if (!IS_ERR(sgid_attr)) {
 724                 id_priv->id.port_num = port;
 725                 cma_bind_sgid_attr(id_priv, sgid_attr);
 726                 ret = 0;
 727                 goto out;
 728         }
 729
 730         list_for_each_entry(cma_dev, &dev_list, list) {
 731                 rdma_for_each_port (cma_dev->device, port) {
 732                         if (listen_id_priv->cma_dev == cma_dev &&
 733                             listen_id_priv->id.port_num == port)
 734                                 continue;
 735
 736                         gid_type = cma_dev->default_gid_type[port - 1];
 737                         sgid_attr = cma_validate_port(cma_dev->device, port,
 738                                                       gid_type, &gid, id_priv);
 739                         if (!IS_ERR(sgid_attr)) {
 740                                 id_priv->id.port_num = port;
 741                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 742                                 ret = 0;
 743                                 goto out;
 744                         }
 745                 }
 746         }
 747
 748 out:
 749         if (!ret) {
 750                 cma_attach_to_dev(id_priv, cma_dev);
 751                 rdma_restrack_add(&id_priv->res);
 752         }
 753
 754         mutex_unlock(&lock);
 755         return ret;
 756 }
 757
 758 /*
 759  * Select the source IB device and address to reach the destination IB address.
 760  */
 761 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 762 {
 763         struct cma_device *cma_dev, *cur_dev;
 764         struct sockaddr_ib *addr;
 765         union ib_gid gid, sgid, *dgid;
 766         unsigned int p;
 767         u16 pkey, index;
 768         enum ib_port_state port_state;
 769         int ret;
 770         int i;
 771
 772         cma_dev = NULL;
 773         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 774         dgid = (union ib_gid *) &addr->sib_addr;
 775         pkey = ntohs(addr->sib_pkey);
 776
 777         mutex_lock(&lock);
 778         list_for_each_entry(cur_dev, &dev_list, list) {
 779                 rdma_for_each_port (cur_dev->device, p) {
 780                         if (!rdma_cap_af_ib(cur_dev->device, p))
 781                                 continue;
 782
 783                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 784                                 continue;
 785
 786                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 787                                 continue;
 788
 789                         for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
 790                              ++i) {
 791                                 ret = rdma_query_gid(cur_dev->device, p, i,
 792                                                      &gid);
 793                                 if (ret)
 794                                         continue;
 795
 796                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 797                                         cma_dev = cur_dev;
 798                                         sgid = gid;
 799                                         id_priv->id.port_num = p;
 800                                         goto found;
 801                                 }
 802
 803                                 if (!cma_dev && (gid.global.subnet_prefix ==
 804                                     dgid->global.subnet_prefix) &&
 805                                     port_state == IB_PORT_ACTIVE) {
 806                                         cma_dev = cur_dev;
 807                                         sgid = gid;
 808                                         id_priv->id.port_num = p;
 809                                         goto found;
 810                                 }
 811                         }
 812                 }
 813         }
 814         mutex_unlock(&lock);
 815         return -ENODEV;
 816
 817 found:
 818         cma_attach_to_dev(id_priv, cma_dev);
 819         rdma_restrack_add(&id_priv->res);
 820         mutex_unlock(&lock);
 821         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 822         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 823         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 824         return 0;
 825 }
 826
 827 static void cma_id_get(struct rdma_id_private *id_priv)
 828 {
 829         refcount_inc(&id_priv->refcount);
 830 }
 831
 832 static void cma_id_put(struct rdma_id_private *id_priv)
 833 {
 834         if (refcount_dec_and_test(&id_priv->refcount))
 835                 complete(&id_priv->comp);
 836 }
 837
 838 static struct rdma_id_private *
 839 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 840                  void *context, enum rdma_ucm_port_space ps,
 841                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 842 {
 843         struct rdma_id_private *id_priv;
 844
 845         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 846         if (!id_priv)
 847                 return ERR_PTR(-ENOMEM);
 848
 849         id_priv->state = RDMA_CM_IDLE;
 850         id_priv->id.context = context;
 851         id_priv->id.event_handler = event_handler;
 852         id_priv->id.ps = ps;
 853         id_priv->id.qp_type = qp_type;
 854         id_priv->tos_set = false;
 855         id_priv->timeout_set = false;
 856         id_priv->min_rnr_timer_set = false;
 857         id_priv->gid_type = IB_GID_TYPE_IB;
 858         spin_lock_init(&id_priv->lock);
 859         mutex_init(&id_priv->qp_mutex);
 860         init_completion(&id_priv->comp);
 861         refcount_set(&id_priv->refcount, 1);
 862         mutex_init(&id_priv->handler_mutex);
 863         INIT_LIST_HEAD(&id_priv->device_item);
 864         INIT_LIST_HEAD(&id_priv->listen_list);
 865         INIT_LIST_HEAD(&id_priv->mc_list);
 866         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 867         id_priv->id.route.addr.dev_addr.net = get_net(net);
 868         id_priv->seq_num &= 0x00ffffff;
 869
 870         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 871         if (parent)
 872                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 873
 874         return id_priv;
 875 }
 876
 877 struct rdma_cm_id *
 878 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
 879                         void *context, enum rdma_ucm_port_space ps,
 880                         enum ib_qp_type qp_type, const char *caller)
 881 {
 882         struct rdma_id_private *ret;
 883
 884         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
 885         if (IS_ERR(ret))
 886                 return ERR_CAST(ret);
 887
 888         rdma_restrack_set_name(&ret->res, caller);
 889         return &ret->id;
 890 }
 891 EXPORT_SYMBOL(__rdma_create_kernel_id);
 892
 893 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
 894                                        void *context,
 895                                        enum rdma_ucm_port_space ps,
 896                                        enum ib_qp_type qp_type)
 897 {
 898         struct rdma_id_private *ret;
 899
 900         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
 901                                ps, qp_type, NULL);
 902         if (IS_ERR(ret))
 903                 return ERR_CAST(ret);
 904
 905         rdma_restrack_set_name(&ret->res, NULL);
 906         return &ret->id;
 907 }
 908 EXPORT_SYMBOL(rdma_create_user_id);
 909
 910 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 911 {
 912         struct ib_qp_attr qp_attr;
 913         int qp_attr_mask, ret;
 914
 915         qp_attr.qp_state = IB_QPS_INIT;
 916         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 917         if (ret)
 918                 return ret;
 919
 920         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 921         if (ret)
 922                 return ret;
 923
 924         qp_attr.qp_state = IB_QPS_RTR;
 925         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 926         if (ret)
 927                 return ret;
 928
 929         qp_attr.qp_state = IB_QPS_RTS;
 930         qp_attr.sq_psn = 0;
 931         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 932
 933         return ret;
 934 }
 935
 936 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 937 {
 938         struct ib_qp_attr qp_attr;
 939         int qp_attr_mask, ret;
 940
 941         qp_attr.qp_state = IB_QPS_INIT;
 942         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 943         if (ret)
 944                 return ret;
 945
 946         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 947 }
 948
 949 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 950                    struct ib_qp_init_attr *qp_init_attr)
 951 {
 952         struct rdma_id_private *id_priv;
 953         struct ib_qp *qp;
 954         int ret;
 955
 956         id_priv = container_of(id, struct rdma_id_private, id);
 957         if (id->device != pd->device) {
 958                 ret = -EINVAL;
 959                 goto out_err;
 960         }
 961
 962         qp_init_attr->port_num = id->port_num;
 963         qp = ib_create_qp(pd, qp_init_attr);
 964         if (IS_ERR(qp)) {
 965                 ret = PTR_ERR(qp);
 966                 goto out_err;
 967         }
 968
 969         if (id->qp_type == IB_QPT_UD)
 970                 ret = cma_init_ud_qp(id_priv, qp);
 971         else
 972                 ret = cma_init_conn_qp(id_priv, qp);
 973         if (ret)
 974                 goto out_destroy;
 975
 976         id->qp = qp;
 977         id_priv->qp_num = qp->qp_num;
 978         id_priv->srq = (qp->srq != NULL);
 979         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 980         return 0;
 981 out_destroy:
 982         ib_destroy_qp(qp);
 983 out_err:
 984         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 985         return ret;
 986 }
 987 EXPORT_SYMBOL(rdma_create_qp);
 988
 989 void rdma_destroy_qp(struct rdma_cm_id *id)
 990 {
 991         struct rdma_id_private *id_priv;
 992
 993         id_priv = container_of(id, struct rdma_id_private, id);
 994         trace_cm_qp_destroy(id_priv);
 995         mutex_lock(&id_priv->qp_mutex);
 996         ib_destroy_qp(id_priv->id.qp);
 997         id_priv->id.qp = NULL;
 998         mutex_unlock(&id_priv->qp_mutex);
 999 }
1000 EXPORT_SYMBOL(rdma_destroy_qp);
1001
1002 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1003                              struct rdma_conn_param *conn_param)
1004 {
1005         struct ib_qp_attr qp_attr;
1006         int qp_attr_mask, ret;
1007
1008         mutex_lock(&id_priv->qp_mutex);
1009         if (!id_priv->id.qp) {
1010                 ret = 0;
1011                 goto out;
1012         }
1013
1014         /* Need to update QP attributes from default values. */
1015         qp_attr.qp_state = IB_QPS_INIT;
1016         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1017         if (ret)
1018                 goto out;
1019
1020         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1021         if (ret)
1022                 goto out;
1023
1024         qp_attr.qp_state = IB_QPS_RTR;
1025         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1026         if (ret)
1027                 goto out;
1028
1029         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1030
1031         if (conn_param)
1032                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1033         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1034 out:
1035         mutex_unlock(&id_priv->qp_mutex);
1036         return ret;
1037 }
1038
1039 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1040                              struct rdma_conn_param *conn_param)
1041 {
1042         struct ib_qp_attr qp_attr;
1043         int qp_attr_mask, ret;
1044
1045         mutex_lock(&id_priv->qp_mutex);
1046         if (!id_priv->id.qp) {
1047                 ret = 0;
1048                 goto out;
1049         }
1050
1051         qp_attr.qp_state = IB_QPS_RTS;
1052         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1053         if (ret)
1054                 goto out;
1055
1056         if (conn_param)
1057                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1058         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1059 out:
1060         mutex_unlock(&id_priv->qp_mutex);
1061         return ret;
1062 }
1063
1064 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1065 {
1066         struct ib_qp_attr qp_attr;
1067         int ret;
1068
1069         mutex_lock(&id_priv->qp_mutex);
1070         if (!id_priv->id.qp) {
1071                 ret = 0;
1072                 goto out;
1073         }
1074
1075         qp_attr.qp_state = IB_QPS_ERR;
1076         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1077 out:
1078         mutex_unlock(&id_priv->qp_mutex);
1079         return ret;
1080 }
1081
1082 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1083                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1084 {
1085         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1086         int ret;
1087         u16 pkey;
1088
1089         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1090                 pkey = 0xffff;
1091         else
1092                 pkey = ib_addr_get_pkey(dev_addr);
1093
1094         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1095                                   pkey, &qp_attr->pkey_index);
1096         if (ret)
1097                 return ret;
1098
1099         qp_attr->port_num = id_priv->id.port_num;
1100         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1101
1102         if (id_priv->id.qp_type == IB_QPT_UD) {
1103                 ret = cma_set_qkey(id_priv, 0);
1104                 if (ret)
1105                         return ret;
1106
1107                 qp_attr->qkey = id_priv->qkey;
1108                 *qp_attr_mask |= IB_QP_QKEY;
1109         } else {
1110                 qp_attr->qp_access_flags = 0;
1111                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1112         }
1113         return 0;
1114 }
1115
1116 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1117                        int *qp_attr_mask)
1118 {
1119         struct rdma_id_private *id_priv;
1120         int ret = 0;
1121
1122         id_priv = container_of(id, struct rdma_id_private, id);
1123         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1124                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1125                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1126                 else
1127                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1128                                                  qp_attr_mask);
1129
1130                 if (qp_attr->qp_state == IB_QPS_RTR)
1131                         qp_attr->rq_psn = id_priv->seq_num;
1132         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1133                 if (!id_priv->cm_id.iw) {
1134                         qp_attr->qp_access_flags = 0;
1135                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1136                 } else
1137                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1138                                                  qp_attr_mask);
1139                 qp_attr->port_num = id_priv->id.port_num;
1140                 *qp_attr_mask |= IB_QP_PORT;
1141         } else {
1142                 ret = -ENOSYS;
1143         }
1144
1145         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1146                 qp_attr->timeout = id_priv->timeout;
1147
1148         if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1149                 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1150
1151         return ret;
1152 }
1153 EXPORT_SYMBOL(rdma_init_qp_attr);
1154
1155 static inline bool cma_zero_addr(const struct sockaddr *addr)
1156 {
1157         switch (addr->sa_family) {
1158         case AF_INET:
1159                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1160         case AF_INET6:
1161                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1162         case AF_IB:
1163                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1164         default:
1165                 return false;
1166         }
1167 }
1168
1169 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1170 {
1171         switch (addr->sa_family) {
1172         case AF_INET:
1173                 return ipv4_is_loopback(
1174                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1175         case AF_INET6:
1176                 return ipv6_addr_loopback(
1177                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1178         case AF_IB:
1179                 return ib_addr_loopback(
1180                         &((struct sockaddr_ib *)addr)->sib_addr);
1181         default:
1182                 return false;
1183         }
1184 }
1185
1186 static inline bool cma_any_addr(const struct sockaddr *addr)
1187 {
1188         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1189 }
1190
1191 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1192 {
1193         if (src->sa_family != dst->sa_family)
1194                 return -1;
1195
1196         switch (src->sa_family) {
1197         case AF_INET:
1198                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1199                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1200         case AF_INET6: {
1201                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1202                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1203                 bool link_local;
1204
1205                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1206                                           &dst_addr6->sin6_addr))
1207                         return 1;
1208                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1209                              IPV6_ADDR_LINKLOCAL;
1210                 /* Link local must match their scope_ids */
1211                 return link_local ? (src_addr6->sin6_scope_id !=
1212                                      dst_addr6->sin6_scope_id) :
1213                                     0;
1214         }
1215
1216         default:
1217                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1218                                    &((struct sockaddr_ib *) dst)->sib_addr);
1219         }
1220 }
1221
1222 static __be16 cma_port(const struct sockaddr *addr)
1223 {
1224         struct sockaddr_ib *sib;
1225
1226         switch (addr->sa_family) {
1227         case AF_INET:
1228                 return ((struct sockaddr_in *) addr)->sin_port;
1229         case AF_INET6:
1230                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1231         case AF_IB:
1232                 sib = (struct sockaddr_ib *) addr;
1233                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1234                                     be64_to_cpu(sib->sib_sid_mask)));
1235         default:
1236                 return 0;
1237         }
1238 }
1239
1240 static inline int cma_any_port(const struct sockaddr *addr)
1241 {
1242         return !cma_port(addr);
1243 }
1244
1245 static void cma_save_ib_info(struct sockaddr *src_addr,
1246                              struct sockaddr *dst_addr,
1247                              const struct rdma_cm_id *listen_id,
1248                              const struct sa_path_rec *path)
1249 {
1250         struct sockaddr_ib *listen_ib, *ib;
1251
1252         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1253         if (src_addr) {
1254                 ib = (struct sockaddr_ib *)src_addr;
1255                 ib->sib_family = AF_IB;
1256                 if (path) {
1257                         ib->sib_pkey = path->pkey;
1258                         ib->sib_flowinfo = path->flow_label;
1259                         memcpy(&ib->sib_addr, &path->sgid, 16);
1260                         ib->sib_sid = path->service_id;
1261                         ib->sib_scope_id = 0;
1262                 } else {
1263                         ib->sib_pkey = listen_ib->sib_pkey;
1264                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1265                         ib->sib_addr = listen_ib->sib_addr;
1266                         ib->sib_sid = listen_ib->sib_sid;
1267                         ib->sib_scope_id = listen_ib->sib_scope_id;
1268                 }
1269                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1270         }
1271         if (dst_addr) {
1272                 ib = (struct sockaddr_ib *)dst_addr;
1273                 ib->sib_family = AF_IB;
1274                 if (path) {
1275                         ib->sib_pkey = path->pkey;
1276                         ib->sib_flowinfo = path->flow_label;
1277                         memcpy(&ib->sib_addr, &path->dgid, 16);
1278                 }
1279         }
1280 }
1281
1282 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1283                               struct sockaddr_in *dst_addr,
1284                               struct cma_hdr *hdr,
1285                               __be16 local_port)
1286 {
1287         if (src_addr) {
1288                 *src_addr = (struct sockaddr_in) {
1289                         .sin_family = AF_INET,
1290                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1291                         .sin_port = local_port,
1292                 };
1293         }
1294
1295         if (dst_addr) {
1296                 *dst_addr = (struct sockaddr_in) {
1297                         .sin_family = AF_INET,
1298                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1299                         .sin_port = hdr->port,
1300                 };
1301         }
1302 }
1303
1304 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1305                               struct sockaddr_in6 *dst_addr,
1306                               struct cma_hdr *hdr,
1307                               __be16 local_port)
1308 {
1309         if (src_addr) {
1310                 *src_addr = (struct sockaddr_in6) {
1311                         .sin6_family = AF_INET6,
1312                         .sin6_addr = hdr->dst_addr.ip6,
1313                         .sin6_port = local_port,
1314                 };
1315         }
1316
1317         if (dst_addr) {
1318                 *dst_addr = (struct sockaddr_in6) {
1319                         .sin6_family = AF_INET6,
1320                         .sin6_addr = hdr->src_addr.ip6,
1321                         .sin6_port = hdr->port,
1322                 };
1323         }
1324 }
1325
1326 static u16 cma_port_from_service_id(__be64 service_id)
1327 {
1328         return (u16)be64_to_cpu(service_id);
1329 }
1330
1331 static int cma_save_ip_info(struct sockaddr *src_addr,
1332                             struct sockaddr *dst_addr,
1333                             const struct ib_cm_event *ib_event,
1334                             __be64 service_id)
1335 {
1336         struct cma_hdr *hdr;
1337         __be16 port;
1338
1339         hdr = ib_event->private_data;
1340         if (hdr->cma_version != CMA_VERSION)
1341                 return -EINVAL;
1342
1343         port = htons(cma_port_from_service_id(service_id));
1344
1345         switch (cma_get_ip_ver(hdr)) {
1346         case 4:
1347                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1348                                   (struct sockaddr_in *)dst_addr, hdr, port);
1349                 break;
1350         case 6:
1351                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1352                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1353                 break;
1354         default:
1355                 return -EAFNOSUPPORT;
1356         }
1357
1358         return 0;
1359 }
1360
1361 static int cma_save_net_info(struct sockaddr *src_addr,
1362                              struct sockaddr *dst_addr,
1363                              const struct rdma_cm_id *listen_id,
1364                              const struct ib_cm_event *ib_event,
1365                              sa_family_t sa_family, __be64 service_id)
1366 {
1367         if (sa_family == AF_IB) {
1368                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1369                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1370                                          ib_event->param.req_rcvd.primary_path);
1371                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1372                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1373                 return 0;
1374         }
1375
1376         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1377 }
1378
1379 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1380                              struct cma_req_info *req)
1381 {
1382         const struct ib_cm_req_event_param *req_param =
1383                 &ib_event->param.req_rcvd;
1384         const struct ib_cm_sidr_req_event_param *sidr_param =
1385                 &ib_event->param.sidr_req_rcvd;
1386
1387         switch (ib_event->event) {
1388         case IB_CM_REQ_RECEIVED:
1389                 req->device     = req_param->listen_id->device;
1390                 req->port       = req_param->port;
1391                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1392                        sizeof(req->local_gid));
1393                 req->has_gid    = true;
1394                 req->service_id = req_param->primary_path->service_id;
1395                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1396                 if (req->pkey != req_param->bth_pkey)
1397                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1398                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1399                                             req_param->bth_pkey, req->pkey);
1400                 break;
1401         case IB_CM_SIDR_REQ_RECEIVED:
1402                 req->device     = sidr_param->listen_id->device;
1403                 req->port       = sidr_param->port;
1404                 req->has_gid    = false;
1405                 req->service_id = sidr_param->service_id;
1406                 req->pkey       = sidr_param->pkey;
1407                 if (req->pkey != sidr_param->bth_pkey)
1408                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1409                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1410                                             sidr_param->bth_pkey, req->pkey);
1411                 break;
1412         default:
1413                 return -EINVAL;
1414         }
1415
1416         return 0;
1417 }
1418
1419 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1420                                   const struct sockaddr_in *dst_addr,
1421                                   const struct sockaddr_in *src_addr)
1422 {
1423         __be32 daddr = dst_addr->sin_addr.s_addr,
1424                saddr = src_addr->sin_addr.s_addr;
1425         struct fib_result res;
1426         struct flowi4 fl4;
1427         int err;
1428         bool ret;
1429
1430         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1431             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1432             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1433             ipv4_is_loopback(saddr))
1434                 return false;
1435
1436         memset(&fl4, 0, sizeof(fl4));
1437         fl4.flowi4_iif = net_dev->ifindex;
1438         fl4.daddr = daddr;
1439         fl4.saddr = saddr;
1440
1441         rcu_read_lock();
1442         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1443         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1444         rcu_read_unlock();
1445
1446         return ret;
1447 }
1448
1449 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1450                                   const struct sockaddr_in6 *dst_addr,
1451                                   const struct sockaddr_in6 *src_addr)
1452 {
1453 #if IS_ENABLED(CONFIG_IPV6)
1454         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1455                            IPV6_ADDR_LINKLOCAL;
1456         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1457                                          &src_addr->sin6_addr, net_dev->ifindex,
1458                                          NULL, strict);
1459         bool ret;
1460
1461         if (!rt)
1462                 return false;
1463
1464         ret = rt->rt6i_idev->dev == net_dev;
1465         ip6_rt_put(rt);
1466
1467         return ret;
1468 #else
1469         return false;
1470 #endif
1471 }
1472
1473 static bool validate_net_dev(struct net_device *net_dev,
1474                              const struct sockaddr *daddr,
1475                              const struct sockaddr *saddr)
1476 {
1477         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1478         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1479         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1480         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1481
1482         switch (daddr->sa_family) {
1483         case AF_INET:
1484                 return saddr->sa_family == AF_INET &&
1485                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1486
1487         case AF_INET6:
1488                 return saddr->sa_family == AF_INET6 &&
1489                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1490
1491         default:
1492                 return false;
1493         }
1494 }
1495
1496 static struct net_device *
1497 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1498 {
1499         const struct ib_gid_attr *sgid_attr = NULL;
1500         struct net_device *ndev;
1501
1502         if (ib_event->event == IB_CM_REQ_RECEIVED)
1503                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1504         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1505                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1506
1507         if (!sgid_attr)
1508                 return NULL;
1509
1510         rcu_read_lock();
1511         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1512         if (IS_ERR(ndev))
1513                 ndev = NULL;
1514         else
1515                 dev_hold(ndev);
1516         rcu_read_unlock();
1517         return ndev;
1518 }
1519
1520 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1521                                           struct cma_req_info *req)
1522 {
1523         struct sockaddr *listen_addr =
1524                         (struct sockaddr *)&req->listen_addr_storage;
1525         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1526         struct net_device *net_dev;
1527         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1528         int err;
1529
1530         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1531                                req->service_id);
1532         if (err)
1533                 return ERR_PTR(err);
1534
1535         if (rdma_protocol_roce(req->device, req->port))
1536                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1537         else
1538                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1539                                                    req->pkey,
1540                                                    gid, listen_addr);
1541         if (!net_dev)
1542                 return ERR_PTR(-ENODEV);
1543
1544         return net_dev;
1545 }
1546
1547 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1548 {
1549         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1550 }
1551
1552 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1553                                    const struct cma_hdr *hdr)
1554 {
1555         struct sockaddr *addr = cma_src_addr(id_priv);
1556         __be32 ip4_addr;
1557         struct in6_addr ip6_addr;
1558
1559         if (cma_any_addr(addr) && !id_priv->afonly)
1560                 return true;
1561
1562         switch (addr->sa_family) {
1563         case AF_INET:
1564                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1565                 if (cma_get_ip_ver(hdr) != 4)
1566                         return false;
1567                 if (!cma_any_addr(addr) &&
1568                     hdr->dst_addr.ip4.addr != ip4_addr)
1569                         return false;
1570                 break;
1571         case AF_INET6:
1572                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1573                 if (cma_get_ip_ver(hdr) != 6)
1574                         return false;
1575                 if (!cma_any_addr(addr) &&
1576                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1577                         return false;
1578                 break;
1579         case AF_IB:
1580                 return true;
1581         default:
1582                 return false;
1583         }
1584
1585         return true;
1586 }
1587
1588 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1589 {
1590         struct ib_device *device = id->device;
1591         const u32 port_num = id->port_num ?: rdma_start_port(device);
1592
1593         return rdma_protocol_roce(device, port_num);
1594 }
1595
1596 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1597 {
1598         const struct sockaddr *daddr =
1599                         (const struct sockaddr *)&req->listen_addr_storage;
1600         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1601
1602         /* Returns true if the req is for IPv6 link local */
1603         return (daddr->sa_family == AF_INET6 &&
1604                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1605 }
1606
1607 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1608                               const struct net_device *net_dev,
1609                               const struct cma_req_info *req)
1610 {
1611         const struct rdma_addr *addr = &id->route.addr;
1612
1613         if (!net_dev)
1614                 /* This request is an AF_IB request */
1615                 return (!id->port_num || id->port_num == req->port) &&
1616                        (addr->src_addr.ss_family == AF_IB);
1617
1618         /*
1619          * If the request is not for IPv6 link local, allow matching
1620          * request to any netdevice of the one or multiport rdma device.
1621          */
1622         if (!cma_is_req_ipv6_ll(req))
1623                 return true;
1624         /*
1625          * Net namespaces must match, and if the listner is listening
1626          * on a specific netdevice than netdevice must match as well.
1627          */
1628         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1629             (!!addr->dev_addr.bound_dev_if ==
1630              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1631                 return true;
1632         else
1633                 return false;
1634 }
1635
1636 static struct rdma_id_private *cma_find_listener(
1637                 const struct rdma_bind_list *bind_list,
1638                 const struct ib_cm_id *cm_id,
1639                 const struct ib_cm_event *ib_event,
1640                 const struct cma_req_info *req,
1641                 const struct net_device *net_dev)
1642 {
1643         struct rdma_id_private *id_priv, *id_priv_dev;
1644
1645         lockdep_assert_held(&lock);
1646
1647         if (!bind_list)
1648                 return ERR_PTR(-EINVAL);
1649
1650         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1651                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1652                         if (id_priv->id.device == cm_id->device &&
1653                             cma_match_net_dev(&id_priv->id, net_dev, req))
1654                                 return id_priv;
1655                         list_for_each_entry(id_priv_dev,
1656                                             &id_priv->listen_list,
1657                                             listen_item) {
1658                                 if (id_priv_dev->id.device == cm_id->device &&
1659                                     cma_match_net_dev(&id_priv_dev->id,
1660                                                       net_dev, req))
1661                                         return id_priv_dev;
1662                         }
1663                 }
1664         }
1665
1666         return ERR_PTR(-EINVAL);
1667 }
1668
1669 static struct rdma_id_private *
1670 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1671                      const struct ib_cm_event *ib_event,
1672                      struct cma_req_info *req,
1673                      struct net_device **net_dev)
1674 {
1675         struct rdma_bind_list *bind_list;
1676         struct rdma_id_private *id_priv;
1677         int err;
1678
1679         err = cma_save_req_info(ib_event, req);
1680         if (err)
1681                 return ERR_PTR(err);
1682
1683         *net_dev = cma_get_net_dev(ib_event, req);
1684         if (IS_ERR(*net_dev)) {
1685                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1686                         /* Assuming the protocol is AF_IB */
1687                         *net_dev = NULL;
1688                 } else {
1689                         return ERR_CAST(*net_dev);
1690                 }
1691         }
1692
1693         mutex_lock(&lock);
1694         /*
1695          * Net namespace might be getting deleted while route lookup,
1696          * cm_id lookup is in progress. Therefore, perform netdevice
1697          * validation, cm_id lookup under rcu lock.
1698          * RCU lock along with netdevice state check, synchronizes with
1699          * netdevice migrating to different net namespace and also avoids
1700          * case where net namespace doesn't get deleted while lookup is in
1701          * progress.
1702          * If the device state is not IFF_UP, its properties such as ifindex
1703          * and nd_net cannot be trusted to remain valid without rcu lock.
1704          * net/core/dev.c change_net_namespace() ensures to synchronize with
1705          * ongoing operations on net device after device is closed using
1706          * synchronize_net().
1707          */
1708         rcu_read_lock();
1709         if (*net_dev) {
1710                 /*
1711                  * If netdevice is down, it is likely that it is administratively
1712                  * down or it might be migrating to different namespace.
1713                  * In that case avoid further processing, as the net namespace
1714                  * or ifindex may change.
1715                  */
1716                 if (((*net_dev)->flags & IFF_UP) == 0) {
1717                         id_priv = ERR_PTR(-EHOSTUNREACH);
1718                         goto err;
1719                 }
1720
1721                 if (!validate_net_dev(*net_dev,
1722                                  (struct sockaddr *)&req->listen_addr_storage,
1723                                  (struct sockaddr *)&req->src_addr_storage)) {
1724                         id_priv = ERR_PTR(-EHOSTUNREACH);
1725                         goto err;
1726                 }
1727         }
1728
1729         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1730                                 rdma_ps_from_service_id(req->service_id),
1731                                 cma_port_from_service_id(req->service_id));
1732         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1733 err:
1734         rcu_read_unlock();
1735         mutex_unlock(&lock);
1736         if (IS_ERR(id_priv) && *net_dev) {
1737                 dev_put(*net_dev);
1738                 *net_dev = NULL;
1739         }
1740         return id_priv;
1741 }
1742
1743 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1744 {
1745         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1746 }
1747
1748 static void cma_cancel_route(struct rdma_id_private *id_priv)
1749 {
1750         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1751                 if (id_priv->query)
1752                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1753         }
1754 }
1755
1756 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1757 {
1758         struct rdma_id_private *dev_id_priv;
1759
1760         lockdep_assert_held(&lock);
1761
1762         /*
1763          * Remove from listen_any_list to prevent added devices from spawning
1764          * additional listen requests.
1765          */
1766         list_del_init(&id_priv->listen_any_item);
1767
1768         while (!list_empty(&id_priv->listen_list)) {
1769                 dev_id_priv =
1770                         list_first_entry(&id_priv->listen_list,
1771                                          struct rdma_id_private, listen_item);
1772                 /* sync with device removal to avoid duplicate destruction */
1773                 list_del_init(&dev_id_priv->device_item);
1774                 list_del_init(&dev_id_priv->listen_item);
1775                 mutex_unlock(&lock);
1776
1777                 rdma_destroy_id(&dev_id_priv->id);
1778                 mutex_lock(&lock);
1779         }
1780 }
1781
1782 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1783 {
1784         mutex_lock(&lock);
1785         _cma_cancel_listens(id_priv);
1786         mutex_unlock(&lock);
1787 }
1788
1789 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1790                                  enum rdma_cm_state state)
1791 {
1792         switch (state) {
1793         case RDMA_CM_ADDR_QUERY:
1794                 /*
1795                  * We can avoid doing the rdma_addr_cancel() based on state,
1796                  * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1797                  * Notice that the addr_handler work could still be exiting
1798                  * outside this state, however due to the interaction with the
1799                  * handler_mutex the work is guaranteed not to touch id_priv
1800                  * during exit.
1801                  */
1802                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1803                 break;
1804         case RDMA_CM_ROUTE_QUERY:
1805                 cma_cancel_route(id_priv);
1806                 break;
1807         case RDMA_CM_LISTEN:
1808                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1809                         cma_cancel_listens(id_priv);
1810                 break;
1811         default:
1812                 break;
1813         }
1814 }
1815
1816 static void cma_release_port(struct rdma_id_private *id_priv)
1817 {
1818         struct rdma_bind_list *bind_list = id_priv->bind_list;
1819         struct net *net = id_priv->id.route.addr.dev_addr.net;
1820
1821         if (!bind_list)
1822                 return;
1823
1824         mutex_lock(&lock);
1825         hlist_del(&id_priv->node);
1826         if (hlist_empty(&bind_list->owners)) {
1827                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1828                 kfree(bind_list);
1829         }
1830         mutex_unlock(&lock);
1831 }
1832
1833 static void destroy_mc(struct rdma_id_private *id_priv,
1834                        struct cma_multicast *mc)
1835 {
1836         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1837
1838         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1839                 ib_sa_free_multicast(mc->sa_mc);
1840
1841         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1842                 struct rdma_dev_addr *dev_addr =
1843                         &id_priv->id.route.addr.dev_addr;
1844                 struct net_device *ndev = NULL;
1845
1846                 if (dev_addr->bound_dev_if)
1847                         ndev = dev_get_by_index(dev_addr->net,
1848                                                 dev_addr->bound_dev_if);
1849                 if (ndev && !send_only) {
1850                         enum ib_gid_type gid_type;
1851                         union ib_gid mgid;
1852
1853                         gid_type = id_priv->cma_dev->default_gid_type
1854                                            [id_priv->id.port_num -
1855                                             rdma_start_port(
1856                                                     id_priv->cma_dev->device)];
1857                         cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
1858                                           gid_type);
1859                         cma_igmp_send(ndev, &mgid, false);
1860                 }
1861                 dev_put(ndev);
1862
1863                 cancel_work_sync(&mc->iboe_join.work);
1864         }
1865         kfree(mc);
1866 }
1867
1868 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1869 {
1870         struct cma_multicast *mc;
1871
1872         while (!list_empty(&id_priv->mc_list)) {
1873                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1874                                       list);
1875                 list_del(&mc->list);
1876                 destroy_mc(id_priv, mc);
1877         }
1878 }
1879
1880 static void _destroy_id(struct rdma_id_private *id_priv,
1881                         enum rdma_cm_state state)
1882 {
1883         cma_cancel_operation(id_priv, state);
1884
1885         rdma_restrack_del(&id_priv->res);
1886         if (id_priv->cma_dev) {
1887                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1888                         if (id_priv->cm_id.ib)
1889                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1890                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1891                         if (id_priv->cm_id.iw)
1892                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1893                 }
1894                 cma_leave_mc_groups(id_priv);
1895                 cma_release_dev(id_priv);
1896         }
1897
1898         cma_release_port(id_priv);
1899         cma_id_put(id_priv);
1900         wait_for_completion(&id_priv->comp);
1901
1902         if (id_priv->internal_id)
1903                 cma_id_put(id_priv->id.context);
1904
1905         kfree(id_priv->id.route.path_rec);
1906
1907         put_net(id_priv->id.route.addr.dev_addr.net);
1908         kfree(id_priv);
1909 }
1910
1911 /*
1912  * destroy an ID from within the handler_mutex. This ensures that no other
1913  * handlers can start running concurrently.
1914  */
1915 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1916         __releases(&idprv->handler_mutex)
1917 {
1918         enum rdma_cm_state state;
1919         unsigned long flags;
1920
1921         trace_cm_id_destroy(id_priv);
1922
1923         /*
1924          * Setting the state to destroyed under the handler mutex provides a
1925          * fence against calling handler callbacks. If this is invoked due to
1926          * the failure of a handler callback then it guarentees that no future
1927          * handlers will be called.
1928          */
1929         lockdep_assert_held(&id_priv->handler_mutex);
1930         spin_lock_irqsave(&id_priv->lock, flags);
1931         state = id_priv->state;
1932         id_priv->state = RDMA_CM_DESTROYING;
1933         spin_unlock_irqrestore(&id_priv->lock, flags);
1934         mutex_unlock(&id_priv->handler_mutex);
1935         _destroy_id(id_priv, state);
1936 }
1937
1938 void rdma_destroy_id(struct rdma_cm_id *id)
1939 {
1940         struct rdma_id_private *id_priv =
1941                 container_of(id, struct rdma_id_private, id);
1942
1943         mutex_lock(&id_priv->handler_mutex);
1944         destroy_id_handler_unlock(id_priv);
1945 }
1946 EXPORT_SYMBOL(rdma_destroy_id);
1947
1948 static int cma_rep_recv(struct rdma_id_private *id_priv)
1949 {
1950         int ret;
1951
1952         ret = cma_modify_qp_rtr(id_priv, NULL);
1953         if (ret)
1954                 goto reject;
1955
1956         ret = cma_modify_qp_rts(id_priv, NULL);
1957         if (ret)
1958                 goto reject;
1959
1960         trace_cm_send_rtu(id_priv);
1961         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1962         if (ret)
1963                 goto reject;
1964
1965         return 0;
1966 reject:
1967         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1968         cma_modify_qp_err(id_priv);
1969         trace_cm_send_rej(id_priv);
1970         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1971                        NULL, 0, NULL, 0);
1972         return ret;
1973 }
1974
1975 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1976                                    const struct ib_cm_rep_event_param *rep_data,
1977                                    void *private_data)
1978 {
1979         event->param.conn.private_data = private_data;
1980         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1981         event->param.conn.responder_resources = rep_data->responder_resources;
1982         event->param.conn.initiator_depth = rep_data->initiator_depth;
1983         event->param.conn.flow_control = rep_data->flow_control;
1984         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1985         event->param.conn.srq = rep_data->srq;
1986         event->param.conn.qp_num = rep_data->remote_qpn;
1987
1988         event->ece.vendor_id = rep_data->ece.vendor_id;
1989         event->ece.attr_mod = rep_data->ece.attr_mod;
1990 }
1991
1992 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1993                                 struct rdma_cm_event *event)
1994 {
1995         int ret;
1996
1997         lockdep_assert_held(&id_priv->handler_mutex);
1998
1999         trace_cm_event_handler(id_priv, event);
2000         ret = id_priv->id.event_handler(&id_priv->id, event);
2001         trace_cm_event_done(id_priv, event, ret);
2002         return ret;
2003 }
2004
2005 static int cma_ib_handler(struct ib_cm_id *cm_id,
2006                           const struct ib_cm_event *ib_event)
2007 {
2008         struct rdma_id_private *id_priv = cm_id->context;
2009         struct rdma_cm_event event = {};
2010         enum rdma_cm_state state;
2011         int ret;
2012
2013         mutex_lock(&id_priv->handler_mutex);
2014         state = READ_ONCE(id_priv->state);
2015         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2016              state != RDMA_CM_CONNECT) ||
2017             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2018              state != RDMA_CM_DISCONNECT))
2019                 goto out;
2020
2021         switch (ib_event->event) {
2022         case IB_CM_REQ_ERROR:
2023         case IB_CM_REP_ERROR:
2024                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2025                 event.status = -ETIMEDOUT;
2026                 break;
2027         case IB_CM_REP_RECEIVED:
2028                 if (state == RDMA_CM_CONNECT &&
2029                     (id_priv->id.qp_type != IB_QPT_UD)) {
2030                         trace_cm_send_mra(id_priv);
2031                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2032                 }
2033                 if (id_priv->id.qp) {
2034                         event.status = cma_rep_recv(id_priv);
2035                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2036                                                      RDMA_CM_EVENT_ESTABLISHED;
2037                 } else {
2038                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2039                 }
2040                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2041                                        ib_event->private_data);
2042                 break;
2043         case IB_CM_RTU_RECEIVED:
2044         case IB_CM_USER_ESTABLISHED:
2045                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2046                 break;
2047         case IB_CM_DREQ_ERROR:
2048                 event.status = -ETIMEDOUT;
2049                 fallthrough;
2050         case IB_CM_DREQ_RECEIVED:
2051         case IB_CM_DREP_RECEIVED:
2052                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2053                                    RDMA_CM_DISCONNECT))
2054                         goto out;
2055                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2056                 break;
2057         case IB_CM_TIMEWAIT_EXIT:
2058                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2059                 break;
2060         case IB_CM_MRA_RECEIVED:
2061                 /* ignore event */
2062                 goto out;
2063         case IB_CM_REJ_RECEIVED:
2064                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2065                                                                                 ib_event->param.rej_rcvd.reason));
2066                 cma_modify_qp_err(id_priv);
2067                 event.status = ib_event->param.rej_rcvd.reason;
2068                 event.event = RDMA_CM_EVENT_REJECTED;
2069                 event.param.conn.private_data = ib_event->private_data;
2070                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2071                 break;
2072         default:
2073                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2074                        ib_event->event);
2075                 goto out;
2076         }
2077
2078         ret = cma_cm_event_handler(id_priv, &event);
2079         if (ret) {
2080                 /* Destroy the CM ID by returning a non-zero value. */
2081                 id_priv->cm_id.ib = NULL;
2082                 destroy_id_handler_unlock(id_priv);
2083                 return ret;
2084         }
2085 out:
2086         mutex_unlock(&id_priv->handler_mutex);
2087         return 0;
2088 }
2089
2090 static struct rdma_id_private *
2091 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2092                    const struct ib_cm_event *ib_event,
2093                    struct net_device *net_dev)
2094 {
2095         struct rdma_id_private *listen_id_priv;
2096         struct rdma_id_private *id_priv;
2097         struct rdma_cm_id *id;
2098         struct rdma_route *rt;
2099         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2100         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2101         const __be64 service_id =
2102                 ib_event->param.req_rcvd.primary_path->service_id;
2103         int ret;
2104
2105         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2106         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2107                                    listen_id->event_handler, listen_id->context,
2108                                    listen_id->ps,
2109                                    ib_event->param.req_rcvd.qp_type,
2110                                    listen_id_priv);
2111         if (IS_ERR(id_priv))
2112                 return NULL;
2113
2114         id = &id_priv->id;
2115         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2116                               (struct sockaddr *)&id->route.addr.dst_addr,
2117                               listen_id, ib_event, ss_family, service_id))
2118                 goto err;
2119
2120         rt = &id->route;
2121         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2122         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2123                                      GFP_KERNEL);
2124         if (!rt->path_rec)
2125                 goto err;
2126
2127         rt->path_rec[0] = *path;
2128         if (rt->num_paths == 2)
2129                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2130
2131         if (net_dev) {
2132                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2133         } else {
2134                 if (!cma_protocol_roce(listen_id) &&
2135                     cma_any_addr(cma_src_addr(id_priv))) {
2136                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2137                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2138                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2139                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2140                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2141                         if (ret)
2142                                 goto err;
2143                 }
2144         }
2145         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2146
2147         id_priv->state = RDMA_CM_CONNECT;
2148         return id_priv;
2149
2150 err:
2151         rdma_destroy_id(id);
2152         return NULL;
2153 }
2154
2155 static struct rdma_id_private *
2156 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2157                   const struct ib_cm_event *ib_event,
2158                   struct net_device *net_dev)
2159 {
2160         const struct rdma_id_private *listen_id_priv;
2161         struct rdma_id_private *id_priv;
2162         struct rdma_cm_id *id;
2163         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2164         struct net *net = listen_id->route.addr.dev_addr.net;
2165         int ret;
2166
2167         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2168         id_priv = __rdma_create_id(net, listen_id->event_handler,
2169                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2170                                    listen_id_priv);
2171         if (IS_ERR(id_priv))
2172                 return NULL;
2173
2174         id = &id_priv->id;
2175         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2176                               (struct sockaddr *)&id->route.addr.dst_addr,
2177                               listen_id, ib_event, ss_family,
2178                               ib_event->param.sidr_req_rcvd.service_id))
2179                 goto err;
2180
2181         if (net_dev) {
2182                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2183         } else {
2184                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2185                         ret = cma_translate_addr(cma_src_addr(id_priv),
2186                                                  &id->route.addr.dev_addr);
2187                         if (ret)
2188                                 goto err;
2189                 }
2190         }
2191
2192         id_priv->state = RDMA_CM_CONNECT;
2193         return id_priv;
2194 err:
2195         rdma_destroy_id(id);
2196         return NULL;
2197 }
2198
2199 static void cma_set_req_event_data(struct rdma_cm_event *event,
2200                                    const struct ib_cm_req_event_param *req_data,
2201                                    void *private_data, int offset)
2202 {
2203         event->param.conn.private_data = private_data + offset;
2204         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2205         event->param.conn.responder_resources = req_data->responder_resources;
2206         event->param.conn.initiator_depth = req_data->initiator_depth;
2207         event->param.conn.flow_control = req_data->flow_control;
2208         event->param.conn.retry_count = req_data->retry_count;
2209         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2210         event->param.conn.srq = req_data->srq;
2211         event->param.conn.qp_num = req_data->remote_qpn;
2212
2213         event->ece.vendor_id = req_data->ece.vendor_id;
2214         event->ece.attr_mod = req_data->ece.attr_mod;
2215 }
2216
2217 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2218                                     const struct ib_cm_event *ib_event)
2219 {
2220         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2221                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2222                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2223                  (id->qp_type == IB_QPT_UD)) ||
2224                 (!id->qp_type));
2225 }
2226
2227 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2228                               const struct ib_cm_event *ib_event)
2229 {
2230         struct rdma_id_private *listen_id, *conn_id = NULL;
2231         struct rdma_cm_event event = {};
2232         struct cma_req_info req = {};
2233         struct net_device *net_dev;
2234         u8 offset;
2235         int ret;
2236
2237         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2238         if (IS_ERR(listen_id))
2239                 return PTR_ERR(listen_id);
2240
2241         trace_cm_req_handler(listen_id, ib_event->event);
2242         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2243                 ret = -EINVAL;
2244                 goto net_dev_put;
2245         }
2246
2247         mutex_lock(&listen_id->handler_mutex);
2248         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2249                 ret = -ECONNABORTED;
2250                 goto err_unlock;
2251         }
2252
2253         offset = cma_user_data_offset(listen_id);
2254         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2255         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2256                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2257                 event.param.ud.private_data = ib_event->private_data + offset;
2258                 event.param.ud.private_data_len =
2259                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2260         } else {
2261                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2262                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2263                                        ib_event->private_data, offset);
2264         }
2265         if (!conn_id) {
2266                 ret = -ENOMEM;
2267                 goto err_unlock;
2268         }
2269
2270         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2271         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2272         if (ret) {
2273                 destroy_id_handler_unlock(conn_id);
2274                 goto err_unlock;
2275         }
2276
2277         conn_id->cm_id.ib = cm_id;
2278         cm_id->context = conn_id;
2279         cm_id->cm_handler = cma_ib_handler;
2280
2281         ret = cma_cm_event_handler(conn_id, &event);
2282         if (ret) {
2283                 /* Destroy the CM ID by returning a non-zero value. */
2284                 conn_id->cm_id.ib = NULL;
2285                 mutex_unlock(&listen_id->handler_mutex);
2286                 destroy_id_handler_unlock(conn_id);
2287                 goto net_dev_put;
2288         }
2289
2290         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2291             conn_id->id.qp_type != IB_QPT_UD) {
2292                 trace_cm_send_mra(cm_id->context);
2293                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2294         }
2295         mutex_unlock(&conn_id->handler_mutex);
2296
2297 err_unlock:
2298         mutex_unlock(&listen_id->handler_mutex);
2299
2300 net_dev_put:
2301         if (net_dev)
2302                 dev_put(net_dev);
2303
2304         return ret;
2305 }
2306
2307 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2308 {
2309         if (addr->sa_family == AF_IB)
2310                 return ((struct sockaddr_ib *) addr)->sib_sid;
2311
2312         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2313 }
2314 EXPORT_SYMBOL(rdma_get_service_id);
2315
2316 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2317                     union ib_gid *dgid)
2318 {
2319         struct rdma_addr *addr = &cm_id->route.addr;
2320
2321         if (!cm_id->device) {
2322                 if (sgid)
2323                         memset(sgid, 0, sizeof(*sgid));
2324                 if (dgid)
2325                         memset(dgid, 0, sizeof(*dgid));
2326                 return;
2327         }
2328
2329         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2330                 if (sgid)
2331                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2332                 if (dgid)
2333                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2334         } else {
2335                 if (sgid)
2336                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2337                 if (dgid)
2338                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2339         }
2340 }
2341 EXPORT_SYMBOL(rdma_read_gids);
2342
2343 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2344 {
2345         struct rdma_id_private *id_priv = iw_id->context;
2346         struct rdma_cm_event event = {};
2347         int ret = 0;
2348         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2349         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2350
2351         mutex_lock(&id_priv->handler_mutex);
2352         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2353                 goto out;
2354
2355         switch (iw_event->event) {
2356         case IW_CM_EVENT_CLOSE:
2357                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2358                 break;
2359         case IW_CM_EVENT_CONNECT_REPLY:
2360                 memcpy(cma_src_addr(id_priv), laddr,
2361                        rdma_addr_size(laddr));
2362                 memcpy(cma_dst_addr(id_priv), raddr,
2363                        rdma_addr_size(raddr));
2364                 switch (iw_event->status) {
2365                 case 0:
2366                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2367                         event.param.conn.initiator_depth = iw_event->ird;
2368                         event.param.conn.responder_resources = iw_event->ord;
2369                         break;
2370                 case -ECONNRESET:
2371                 case -ECONNREFUSED:
2372                         event.event = RDMA_CM_EVENT_REJECTED;
2373                         break;
2374                 case -ETIMEDOUT:
2375                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2376                         break;
2377                 default:
2378                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2379                         break;
2380                 }
2381                 break;
2382         case IW_CM_EVENT_ESTABLISHED:
2383                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2384                 event.param.conn.initiator_depth = iw_event->ird;
2385                 event.param.conn.responder_resources = iw_event->ord;
2386                 break;
2387         default:
2388                 goto out;
2389         }
2390
2391         event.status = iw_event->status;
2392         event.param.conn.private_data = iw_event->private_data;
2393         event.param.conn.private_data_len = iw_event->private_data_len;
2394         ret = cma_cm_event_handler(id_priv, &event);
2395         if (ret) {
2396                 /* Destroy the CM ID by returning a non-zero value. */
2397                 id_priv->cm_id.iw = NULL;
2398                 destroy_id_handler_unlock(id_priv);
2399                 return ret;
2400         }
2401
2402 out:
2403         mutex_unlock(&id_priv->handler_mutex);
2404         return ret;
2405 }
2406
2407 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2408                                struct iw_cm_event *iw_event)
2409 {
2410         struct rdma_id_private *listen_id, *conn_id;
2411         struct rdma_cm_event event = {};
2412         int ret = -ECONNABORTED;
2413         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2414         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2415
2416         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2417         event.param.conn.private_data = iw_event->private_data;
2418         event.param.conn.private_data_len = iw_event->private_data_len;
2419         event.param.conn.initiator_depth = iw_event->ird;
2420         event.param.conn.responder_resources = iw_event->ord;
2421
2422         listen_id = cm_id->context;
2423
2424         mutex_lock(&listen_id->handler_mutex);
2425         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2426                 goto out;
2427
2428         /* Create a new RDMA id for the new IW CM ID */
2429         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2430                                    listen_id->id.event_handler,
2431                                    listen_id->id.context, RDMA_PS_TCP,
2432                                    IB_QPT_RC, listen_id);
2433         if (IS_ERR(conn_id)) {
2434                 ret = -ENOMEM;
2435                 goto out;
2436         }
2437         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2438         conn_id->state = RDMA_CM_CONNECT;
2439
2440         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2441         if (ret) {
2442                 mutex_unlock(&listen_id->handler_mutex);
2443                 destroy_id_handler_unlock(conn_id);
2444                 return ret;
2445         }
2446
2447         ret = cma_iw_acquire_dev(conn_id, listen_id);
2448         if (ret) {
2449                 mutex_unlock(&listen_id->handler_mutex);
2450                 destroy_id_handler_unlock(conn_id);
2451                 return ret;
2452         }
2453
2454         conn_id->cm_id.iw = cm_id;
2455         cm_id->context = conn_id;
2456         cm_id->cm_handler = cma_iw_handler;
2457
2458         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2459         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2460
2461         ret = cma_cm_event_handler(conn_id, &event);
2462         if (ret) {
2463                 /* User wants to destroy the CM ID */
2464                 conn_id->cm_id.iw = NULL;
2465                 mutex_unlock(&listen_id->handler_mutex);
2466                 destroy_id_handler_unlock(conn_id);
2467                 return ret;
2468         }
2469
2470         mutex_unlock(&conn_id->handler_mutex);
2471
2472 out:
2473         mutex_unlock(&listen_id->handler_mutex);
2474         return ret;
2475 }
2476
2477 static int cma_ib_listen(struct rdma_id_private *id_priv)
2478 {
2479         struct sockaddr *addr;
2480         struct ib_cm_id *id;
2481         __be64 svc_id;
2482
2483         addr = cma_src_addr(id_priv);
2484         svc_id = rdma_get_service_id(&id_priv->id, addr);
2485         id = ib_cm_insert_listen(id_priv->id.device,
2486                                  cma_ib_req_handler, svc_id);
2487         if (IS_ERR(id))
2488                 return PTR_ERR(id);
2489         id_priv->cm_id.ib = id;
2490
2491         return 0;
2492 }
2493
2494 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2495 {
2496         int ret;
2497         struct iw_cm_id *id;
2498
2499         id = iw_create_cm_id(id_priv->id.device,
2500                              iw_conn_req_handler,
2501                              id_priv);
2502         if (IS_ERR(id))
2503                 return PTR_ERR(id);
2504
2505         mutex_lock(&id_priv->qp_mutex);
2506         id->tos = id_priv->tos;
2507         id->tos_set = id_priv->tos_set;
2508         mutex_unlock(&id_priv->qp_mutex);
2509         id->afonly = id_priv->afonly;
2510         id_priv->cm_id.iw = id;
2511
2512         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2513                rdma_addr_size(cma_src_addr(id_priv)));
2514
2515         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2516
2517         if (ret) {
2518                 iw_destroy_cm_id(id_priv->cm_id.iw);
2519                 id_priv->cm_id.iw = NULL;
2520         }
2521
2522         return ret;
2523 }
2524
2525 static int cma_listen_handler(struct rdma_cm_id *id,
2526                               struct rdma_cm_event *event)
2527 {
2528         struct rdma_id_private *id_priv = id->context;
2529
2530         /* Listening IDs are always destroyed on removal */
2531         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2532                 return -1;
2533
2534         id->context = id_priv->id.context;
2535         id->event_handler = id_priv->id.event_handler;
2536         trace_cm_event_handler(id_priv, event);
2537         return id_priv->id.event_handler(id, event);
2538 }
2539
2540 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2541                              struct cma_device *cma_dev,
2542                              struct rdma_id_private **to_destroy)
2543 {
2544         struct rdma_id_private *dev_id_priv;
2545         struct net *net = id_priv->id.route.addr.dev_addr.net;
2546         int ret;
2547
2548         lockdep_assert_held(&lock);
2549
2550         *to_destroy = NULL;
2551         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2552                 return 0;
2553
2554         dev_id_priv =
2555                 __rdma_create_id(net, cma_listen_handler, id_priv,
2556                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2557         if (IS_ERR(dev_id_priv))
2558                 return PTR_ERR(dev_id_priv);
2559
2560         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2561         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2562                rdma_addr_size(cma_src_addr(id_priv)));
2563
2564         _cma_attach_to_dev(dev_id_priv, cma_dev);
2565         rdma_restrack_add(&dev_id_priv->res);
2566         cma_id_get(id_priv);
2567         dev_id_priv->internal_id = 1;
2568         dev_id_priv->afonly = id_priv->afonly;
2569         mutex_lock(&id_priv->qp_mutex);
2570         dev_id_priv->tos_set = id_priv->tos_set;
2571         dev_id_priv->tos = id_priv->tos;
2572         mutex_unlock(&id_priv->qp_mutex);
2573
2574         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2575         if (ret)
2576                 goto err_listen;
2577         list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2578         return 0;
2579 err_listen:
2580         /* Caller must destroy this after releasing lock */
2581         *to_destroy = dev_id_priv;
2582         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2583         return ret;
2584 }
2585
2586 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2587 {
2588         struct rdma_id_private *to_destroy;
2589         struct cma_device *cma_dev;
2590         int ret;
2591
2592         mutex_lock(&lock);
2593         list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2594         list_for_each_entry(cma_dev, &dev_list, list) {
2595                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2596                 if (ret) {
2597                         /* Prevent racing with cma_process_remove() */
2598                         if (to_destroy)
2599                                 list_del_init(&to_destroy->device_item);
2600                         goto err_listen;
2601                 }
2602         }
2603         mutex_unlock(&lock);
2604         return 0;
2605
2606 err_listen:
2607         _cma_cancel_listens(id_priv);
2608         mutex_unlock(&lock);
2609         if (to_destroy)
2610                 rdma_destroy_id(&to_destroy->id);
2611         return ret;
2612 }
2613
2614 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2615 {
2616         struct rdma_id_private *id_priv;
2617
2618         id_priv = container_of(id, struct rdma_id_private, id);
2619         mutex_lock(&id_priv->qp_mutex);
2620         id_priv->tos = (u8) tos;
2621         id_priv->tos_set = true;
2622         mutex_unlock(&id_priv->qp_mutex);
2623 }
2624 EXPORT_SYMBOL(rdma_set_service_type);
2625
2626 /**
2627  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2628  *                          with a connection identifier.
2629  * @id: Communication identifier to associated with service type.
2630  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2631  *
2632  * This function should be called before rdma_connect() on active side,
2633  * and on passive side before rdma_accept(). It is applicable to primary
2634  * path only. The timeout will affect the local side of the QP, it is not
2635  * negotiated with remote side and zero disables the timer. In case it is
2636  * set before rdma_resolve_route, the value will also be used to determine
2637  * PacketLifeTime for RoCE.
2638  *
2639  * Return: 0 for success
2640  */
2641 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2642 {
2643         struct rdma_id_private *id_priv;
2644
2645         if (id->qp_type != IB_QPT_RC)
2646                 return -EINVAL;
2647
2648         id_priv = container_of(id, struct rdma_id_private, id);
2649         mutex_lock(&id_priv->qp_mutex);
2650         id_priv->timeout = timeout;
2651         id_priv->timeout_set = true;
2652         mutex_unlock(&id_priv->qp_mutex);
2653
2654         return 0;
2655 }
2656 EXPORT_SYMBOL(rdma_set_ack_timeout);
2657
2658 /**
2659  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2660  *                            QP associated with a connection identifier.
2661  * @id: Communication identifier to associated with service type.
2662  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2663  *                 Timer Field" in the IBTA specification.
2664  *
2665  * This function should be called before rdma_connect() on active
2666  * side, and on passive side before rdma_accept(). The timer value
2667  * will be associated with the local QP. When it receives a send it is
2668  * not read to handle, typically if the receive queue is empty, an RNR
2669  * Retry NAK is returned to the requester with the min_rnr_timer
2670  * encoded. The requester will then wait at least the time specified
2671  * in the NAK before retrying. The default is zero, which translates
2672  * to a minimum RNR Timer value of 655 ms.
2673  *
2674  * Return: 0 for success
2675  */
2676 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2677 {
2678         struct rdma_id_private *id_priv;
2679
2680         /* It is a five-bit value */
2681         if (min_rnr_timer & 0xe0)
2682                 return -EINVAL;
2683
2684         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2685                 return -EINVAL;
2686
2687         id_priv = container_of(id, struct rdma_id_private, id);
2688         mutex_lock(&id_priv->qp_mutex);
2689         id_priv->min_rnr_timer = min_rnr_timer;
2690         id_priv->min_rnr_timer_set = true;
2691         mutex_unlock(&id_priv->qp_mutex);
2692
2693         return 0;
2694 }
2695 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2696
2697 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2698                               void *context)
2699 {
2700         struct cma_work *work = context;
2701         struct rdma_route *route;
2702
2703         route = &work->id->id.route;
2704
2705         if (!status) {
2706                 route->num_paths = 1;
2707                 *route->path_rec = *path_rec;
2708         } else {
2709                 work->old_state = RDMA_CM_ROUTE_QUERY;
2710                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2711                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2712                 work->event.status = status;
2713                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2714                                      status);
2715         }
2716
2717         queue_work(cma_wq, &work->work);
2718 }
2719
2720 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2721                               unsigned long timeout_ms, struct cma_work *work)
2722 {
2723         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2724         struct sa_path_rec path_rec;
2725         ib_sa_comp_mask comp_mask;
2726         struct sockaddr_in6 *sin6;
2727         struct sockaddr_ib *sib;
2728
2729         memset(&path_rec, 0, sizeof path_rec);
2730
2731         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2732                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2733         else
2734                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2735         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2736         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2737         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2738         path_rec.numb_path = 1;
2739         path_rec.reversible = 1;
2740         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2741                                                   cma_dst_addr(id_priv));
2742
2743         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2744                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2745                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2746
2747         switch (cma_family(id_priv)) {
2748         case AF_INET:
2749                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2750                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2751                 break;
2752         case AF_INET6:
2753                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2754                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2755                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2756                 break;
2757         case AF_IB:
2758                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2759                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2760                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2761                 break;
2762         }
2763
2764         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2765                                                id_priv->id.port_num, &path_rec,
2766                                                comp_mask, timeout_ms,
2767                                                GFP_KERNEL, cma_query_handler,
2768                                                work, &id_priv->query);
2769
2770         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2771 }
2772
2773 static void cma_iboe_join_work_handler(struct work_struct *work)
2774 {
2775         struct cma_multicast *mc =
2776                 container_of(work, struct cma_multicast, iboe_join.work);
2777         struct rdma_cm_event *event = &mc->iboe_join.event;
2778         struct rdma_id_private *id_priv = mc->id_priv;
2779         int ret;
2780
2781         mutex_lock(&id_priv->handler_mutex);
2782         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2783             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2784                 goto out_unlock;
2785
2786         ret = cma_cm_event_handler(id_priv, event);
2787         WARN_ON(ret);
2788
2789 out_unlock:
2790         mutex_unlock(&id_priv->handler_mutex);
2791         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2792                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2793 }
2794
2795 static void cma_work_handler(struct work_struct *_work)
2796 {
2797         struct cma_work *work = container_of(_work, struct cma_work, work);
2798         struct rdma_id_private *id_priv = work->id;
2799
2800         mutex_lock(&id_priv->handler_mutex);
2801         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2802             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2803                 goto out_unlock;
2804         if (work->old_state != 0 || work->new_state != 0) {
2805                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2806                         goto out_unlock;
2807         }
2808
2809         if (cma_cm_event_handler(id_priv, &work->event)) {
2810                 cma_id_put(id_priv);
2811                 destroy_id_handler_unlock(id_priv);
2812                 goto out_free;
2813         }
2814
2815 out_unlock:
2816         mutex_unlock(&id_priv->handler_mutex);
2817         cma_id_put(id_priv);
2818 out_free:
2819         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2820                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2821         kfree(work);
2822 }
2823
2824 static void cma_init_resolve_route_work(struct cma_work *work,
2825                                         struct rdma_id_private *id_priv)
2826 {
2827         work->id = id_priv;
2828         INIT_WORK(&work->work, cma_work_handler);
2829         work->old_state = RDMA_CM_ROUTE_QUERY;
2830         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2831         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2832 }
2833
2834 static void enqueue_resolve_addr_work(struct cma_work *work,
2835                                       struct rdma_id_private *id_priv)
2836 {
2837         /* Balances with cma_id_put() in cma_work_handler */
2838         cma_id_get(id_priv);
2839
2840         work->id = id_priv;
2841         INIT_WORK(&work->work, cma_work_handler);
2842         work->old_state = RDMA_CM_ADDR_QUERY;
2843         work->new_state = RDMA_CM_ADDR_RESOLVED;
2844         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2845
2846         queue_work(cma_wq, &work->work);
2847 }
2848
2849 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2850                                 unsigned long timeout_ms)
2851 {
2852         struct rdma_route *route = &id_priv->id.route;
2853         struct cma_work *work;
2854         int ret;
2855
2856         work = kzalloc(sizeof *work, GFP_KERNEL);
2857         if (!work)
2858                 return -ENOMEM;
2859
2860         cma_init_resolve_route_work(work, id_priv);
2861
2862         if (!route->path_rec)
2863                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2864         if (!route->path_rec) {
2865                 ret = -ENOMEM;
2866                 goto err1;
2867         }
2868
2869         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2870         if (ret)
2871                 goto err2;
2872
2873         return 0;
2874 err2:
2875         kfree(route->path_rec);
2876         route->path_rec = NULL;
2877 err1:
2878         kfree(work);
2879         return ret;
2880 }
2881
2882 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2883                                            unsigned long supported_gids,
2884                                            enum ib_gid_type default_gid)
2885 {
2886         if ((network_type == RDMA_NETWORK_IPV4 ||
2887              network_type == RDMA_NETWORK_IPV6) &&
2888             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2889                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2890
2891         return default_gid;
2892 }
2893
2894 /*
2895  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2896  * path record type based on GID type.
2897  * It also sets up other L2 fields which includes destination mac address
2898  * netdev ifindex, of the path record.
2899  * It returns the netdev of the bound interface for this path record entry.
2900  */
2901 static struct net_device *
2902 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2903 {
2904         struct rdma_route *route = &id_priv->id.route;
2905         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2906         struct rdma_addr *addr = &route->addr;
2907         unsigned long supported_gids;
2908         struct net_device *ndev;
2909
2910         if (!addr->dev_addr.bound_dev_if)
2911                 return NULL;
2912
2913         ndev = dev_get_by_index(addr->dev_addr.net,
2914                                 addr->dev_addr.bound_dev_if);
2915         if (!ndev)
2916                 return NULL;
2917
2918         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2919                                                     id_priv->id.port_num);
2920         gid_type = cma_route_gid_type(addr->dev_addr.network,
2921                                       supported_gids,
2922                                       id_priv->gid_type);
2923         /* Use the hint from IP Stack to select GID Type */
2924         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2925                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2926         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2927
2928         route->path_rec->roce.route_resolved = true;
2929         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2930         return ndev;
2931 }
2932
2933 int rdma_set_ib_path(struct rdma_cm_id *id,
2934                      struct sa_path_rec *path_rec)
2935 {
2936         struct rdma_id_private *id_priv;
2937         struct net_device *ndev;
2938         int ret;
2939
2940         id_priv = container_of(id, struct rdma_id_private, id);
2941         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2942                            RDMA_CM_ROUTE_RESOLVED))
2943                 return -EINVAL;
2944
2945         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2946                                      GFP_KERNEL);
2947         if (!id->route.path_rec) {
2948                 ret = -ENOMEM;
2949                 goto err;
2950         }
2951
2952         if (rdma_protocol_roce(id->device, id->port_num)) {
2953                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2954                 if (!ndev) {
2955                         ret = -ENODEV;
2956                         goto err_free;
2957                 }
2958                 dev_put(ndev);
2959         }
2960
2961         id->route.num_paths = 1;
2962         return 0;
2963
2964 err_free:
2965         kfree(id->route.path_rec);
2966         id->route.path_rec = NULL;
2967 err:
2968         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2969         return ret;
2970 }
2971 EXPORT_SYMBOL(rdma_set_ib_path);
2972
2973 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2974 {
2975         struct cma_work *work;
2976
2977         work = kzalloc(sizeof *work, GFP_KERNEL);
2978         if (!work)
2979                 return -ENOMEM;
2980
2981         cma_init_resolve_route_work(work, id_priv);
2982         queue_work(cma_wq, &work->work);
2983         return 0;
2984 }
2985
2986 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2987 {
2988         struct net_device *dev;
2989
2990         dev = vlan_dev_real_dev(vlan_ndev);
2991         if (dev->num_tc)
2992                 return netdev_get_prio_tc_map(dev, prio);
2993
2994         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2995                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2996 }
2997
2998 struct iboe_prio_tc_map {
2999         int input_prio;
3000         int output_tc;
3001         bool found;
3002 };
3003
3004 static int get_lower_vlan_dev_tc(struct net_device *dev,
3005                                  struct netdev_nested_priv *priv)
3006 {
3007         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3008
3009         if (is_vlan_dev(dev))
3010                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3011         else if (dev->num_tc)
3012                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3013         else
3014                 map->output_tc = 0;
3015         /* We are interested only in first level VLAN device, so always
3016          * return 1 to stop iterating over next level devices.
3017          */
3018         map->found = true;
3019         return 1;
3020 }
3021
3022 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3023 {
3024         struct iboe_prio_tc_map prio_tc_map = {};
3025         int prio = rt_tos2priority(tos);
3026         struct netdev_nested_priv priv;
3027
3028         /* If VLAN device, get it directly from the VLAN netdev */
3029         if (is_vlan_dev(ndev))
3030                 return get_vlan_ndev_tc(ndev, prio);
3031
3032         prio_tc_map.input_prio = prio;
3033         priv.data = (void *)&prio_tc_map;
3034         rcu_read_lock();
3035         netdev_walk_all_lower_dev_rcu(ndev,
3036                                       get_lower_vlan_dev_tc,
3037                                       &priv);
3038         rcu_read_unlock();
3039         /* If map is found from lower device, use it; Otherwise
3040          * continue with the current netdevice to get priority to tc map.
3041          */
3042         if (prio_tc_map.found)
3043                 return prio_tc_map.output_tc;
3044         else if (ndev->num_tc)
3045                 return netdev_get_prio_tc_map(ndev, prio);
3046         else
3047                 return 0;
3048 }
3049
3050 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3051 {
3052         struct sockaddr_in6 *addr6;
3053         u16 dport, sport;
3054         u32 hash, fl;
3055
3056         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3057         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3058         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3059                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3060                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3061                 hash = (u32)sport * 31 + dport;
3062                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3063         }
3064
3065         return cpu_to_be32(fl);
3066 }
3067
3068 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3069 {
3070         struct rdma_route *route = &id_priv->id.route;
3071         struct rdma_addr *addr = &route->addr;
3072         struct cma_work *work;
3073         int ret;
3074         struct net_device *ndev;
3075
3076         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3077                                         rdma_start_port(id_priv->cma_dev->device)];
3078         u8 tos;
3079
3080         mutex_lock(&id_priv->qp_mutex);
3081         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3082         mutex_unlock(&id_priv->qp_mutex);
3083
3084         work = kzalloc(sizeof *work, GFP_KERNEL);
3085         if (!work)
3086                 return -ENOMEM;
3087
3088         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3089         if (!route->path_rec) {
3090                 ret = -ENOMEM;
3091                 goto err1;
3092         }
3093
3094         route->num_paths = 1;
3095
3096         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3097         if (!ndev) {
3098                 ret = -ENODEV;
3099                 goto err2;
3100         }
3101
3102         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3103                     &route->path_rec->sgid);
3104         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3105                     &route->path_rec->dgid);
3106
3107         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3108                 /* TODO: get the hoplimit from the inet/inet6 device */
3109                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3110         else
3111                 route->path_rec->hop_limit = 1;
3112         route->path_rec->reversible = 1;
3113         route->path_rec->pkey = cpu_to_be16(0xffff);
3114         route->path_rec->mtu_selector = IB_SA_EQ;
3115         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3116         route->path_rec->traffic_class = tos;
3117         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3118         route->path_rec->rate_selector = IB_SA_EQ;
3119         route->path_rec->rate = iboe_get_rate(ndev);
3120         dev_put(ndev);
3121         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3122         /* In case ACK timeout is set, use this value to calculate
3123          * PacketLifeTime.  As per IBTA 12.7.34,
3124          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3125          * Assuming a negligible local ACK delay, we can use
3126          * PacketLifeTime = local ACK timeout/2
3127          * as a reasonable approximation for RoCE networks.
3128          */
3129         mutex_lock(&id_priv->qp_mutex);
3130         if (id_priv->timeout_set && id_priv->timeout)
3131                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3132         else
3133                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3134         mutex_unlock(&id_priv->qp_mutex);
3135
3136         if (!route->path_rec->mtu) {
3137                 ret = -EINVAL;
3138                 goto err2;
3139         }
3140
3141         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3142                                          id_priv->id.port_num))
3143                 route->path_rec->flow_label =
3144                         cma_get_roce_udp_flow_label(id_priv);
3145
3146         cma_init_resolve_route_work(work, id_priv);
3147         queue_work(cma_wq, &work->work);
3148
3149         return 0;
3150
3151 err2:
3152         kfree(route->path_rec);
3153         route->path_rec = NULL;
3154         route->num_paths = 0;
3155 err1:
3156         kfree(work);
3157         return ret;
3158 }
3159
3160 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3161 {
3162         struct rdma_id_private *id_priv;
3163         int ret;
3164
3165         if (!timeout_ms)
3166                 return -EINVAL;
3167
3168         id_priv = container_of(id, struct rdma_id_private, id);
3169         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3170                 return -EINVAL;
3171
3172         cma_id_get(id_priv);
3173         if (rdma_cap_ib_sa(id->device, id->port_num))
3174                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3175         else if (rdma_protocol_roce(id->device, id->port_num))
3176                 ret = cma_resolve_iboe_route(id_priv);
3177         else if (rdma_protocol_iwarp(id->device, id->port_num))
3178                 ret = cma_resolve_iw_route(id_priv);
3179         else
3180                 ret = -ENOSYS;
3181
3182         if (ret)
3183                 goto err;
3184
3185         return 0;
3186 err:
3187         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3188         cma_id_put(id_priv);
3189         return ret;
3190 }
3191 EXPORT_SYMBOL(rdma_resolve_route);
3192
3193 static void cma_set_loopback(struct sockaddr *addr)
3194 {
3195         switch (addr->sa_family) {
3196         case AF_INET:
3197                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3198                 break;
3199         case AF_INET6:
3200                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3201                               0, 0, 0, htonl(1));
3202                 break;
3203         default:
3204                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3205                             0, 0, 0, htonl(1));
3206                 break;
3207         }
3208 }
3209
3210 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3211 {
3212         struct cma_device *cma_dev, *cur_dev;
3213         union ib_gid gid;
3214         enum ib_port_state port_state;
3215         unsigned int p;
3216         u16 pkey;
3217         int ret;
3218
3219         cma_dev = NULL;
3220         mutex_lock(&lock);
3221         list_for_each_entry(cur_dev, &dev_list, list) {
3222                 if (cma_family(id_priv) == AF_IB &&
3223                     !rdma_cap_ib_cm(cur_dev->device, 1))
3224                         continue;
3225
3226                 if (!cma_dev)
3227                         cma_dev = cur_dev;
3228
3229                 rdma_for_each_port (cur_dev->device, p) {
3230                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3231                             port_state == IB_PORT_ACTIVE) {
3232                                 cma_dev = cur_dev;
3233                                 goto port_found;
3234                         }
3235                 }
3236         }
3237
3238         if (!cma_dev) {
3239                 ret = -ENODEV;
3240                 goto out;
3241         }
3242
3243         p = 1;
3244
3245 port_found:
3246         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3247         if (ret)
3248                 goto out;
3249
3250         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3251         if (ret)
3252                 goto out;
3253
3254         id_priv->id.route.addr.dev_addr.dev_type =
3255                 (rdma_protocol_ib(cma_dev->device, p)) ?
3256                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3257
3258         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3259         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3260         id_priv->id.port_num = p;
3261         cma_attach_to_dev(id_priv, cma_dev);
3262         rdma_restrack_add(&id_priv->res);
3263         cma_set_loopback(cma_src_addr(id_priv));
3264 out:
3265         mutex_unlock(&lock);
3266         return ret;
3267 }
3268
3269 static void addr_handler(int status, struct sockaddr *src_addr,
3270                          struct rdma_dev_addr *dev_addr, void *context)
3271 {
3272         struct rdma_id_private *id_priv = context;
3273         struct rdma_cm_event event = {};
3274         struct sockaddr *addr;
3275         struct sockaddr_storage old_addr;
3276
3277         mutex_lock(&id_priv->handler_mutex);
3278         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3279                            RDMA_CM_ADDR_RESOLVED))
3280                 goto out;
3281
3282         /*
3283          * Store the previous src address, so that if we fail to acquire
3284          * matching rdma device, old address can be restored back, which helps
3285          * to cancel the cma listen operation correctly.
3286          */
3287         addr = cma_src_addr(id_priv);
3288         memcpy(&old_addr, addr, rdma_addr_size(addr));
3289         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3290         if (!status && !id_priv->cma_dev) {
3291                 status = cma_acquire_dev_by_src_ip(id_priv);
3292                 if (status)
3293                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3294                                              status);
3295                 rdma_restrack_add(&id_priv->res);
3296         } else if (status) {
3297                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3298         }
3299
3300         if (status) {
3301                 memcpy(addr, &old_addr,
3302                        rdma_addr_size((struct sockaddr *)&old_addr));
3303                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3304                                    RDMA_CM_ADDR_BOUND))
3305                         goto out;
3306                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3307                 event.status = status;
3308         } else
3309                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3310
3311         if (cma_cm_event_handler(id_priv, &event)) {
3312                 destroy_id_handler_unlock(id_priv);
3313                 return;
3314         }
3315 out:
3316         mutex_unlock(&id_priv->handler_mutex);
3317 }
3318
3319 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3320 {
3321         struct cma_work *work;
3322         union ib_gid gid;
3323         int ret;
3324
3325         work = kzalloc(sizeof *work, GFP_KERNEL);
3326         if (!work)
3327                 return -ENOMEM;
3328
3329         if (!id_priv->cma_dev) {
3330                 ret = cma_bind_loopback(id_priv);
3331                 if (ret)
3332                         goto err;
3333         }
3334
3335         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3336         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3337
3338         enqueue_resolve_addr_work(work, id_priv);
3339         return 0;
3340 err:
3341         kfree(work);
3342         return ret;
3343 }
3344
3345 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3346 {
3347         struct cma_work *work;
3348         int ret;
3349
3350         work = kzalloc(sizeof *work, GFP_KERNEL);
3351         if (!work)
3352                 return -ENOMEM;
3353
3354         if (!id_priv->cma_dev) {
3355                 ret = cma_resolve_ib_dev(id_priv);
3356                 if (ret)
3357                         goto err;
3358         }
3359
3360         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3361                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3362
3363         enqueue_resolve_addr_work(work, id_priv);
3364         return 0;
3365 err:
3366         kfree(work);
3367         return ret;
3368 }
3369
3370 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3371                          const struct sockaddr *dst_addr)
3372 {
3373         if (!src_addr || !src_addr->sa_family) {
3374                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3375                 src_addr->sa_family = dst_addr->sa_family;
3376                 if (IS_ENABLED(CONFIG_IPV6) &&
3377                     dst_addr->sa_family == AF_INET6) {
3378                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3379                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3380                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3381                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3382                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3383                 } else if (dst_addr->sa_family == AF_IB) {
3384                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3385                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3386                 }
3387         }
3388         return rdma_bind_addr(id, src_addr);
3389 }
3390
3391 /*
3392  * If required, resolve the source address for bind and leave the id_priv in
3393  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3394  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3395  * ignored.
3396  */
3397 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3398                                struct sockaddr *src_addr,
3399                                const struct sockaddr *dst_addr)
3400 {
3401         int ret;
3402
3403         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3404         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3405                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3406                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3407                 if (ret)
3408                         goto err_dst;
3409                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3410                                            RDMA_CM_ADDR_QUERY))) {
3411                         ret = -EINVAL;
3412                         goto err_dst;
3413                 }
3414         }
3415
3416         if (cma_family(id_priv) != dst_addr->sa_family) {
3417                 ret = -EINVAL;
3418                 goto err_state;
3419         }
3420         return 0;
3421
3422 err_state:
3423         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3424 err_dst:
3425         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3426         return ret;
3427 }
3428
3429 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3430                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3431 {
3432         struct rdma_id_private *id_priv =
3433                 container_of(id, struct rdma_id_private, id);
3434         int ret;
3435
3436         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3437         if (ret)
3438                 return ret;
3439
3440         if (cma_any_addr(dst_addr)) {
3441                 ret = cma_resolve_loopback(id_priv);
3442         } else {
3443                 if (dst_addr->sa_family == AF_IB) {
3444                         ret = cma_resolve_ib_addr(id_priv);
3445                 } else {
3446                         /*
3447                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
3448                          * rdma_resolve_ip() is called, eg through the error
3449                          * path in addr_handler(). If this happens the existing
3450                          * request must be canceled before issuing a new one.
3451                          * Since canceling a request is a bit slow and this
3452                          * oddball path is rare, keep track once a request has
3453                          * been issued. The track turns out to be a permanent
3454                          * state since this is the only cancel as it is
3455                          * immediately before rdma_resolve_ip().
3456                          */
3457                         if (id_priv->used_resolve_ip)
3458                                 rdma_addr_cancel(&id->route.addr.dev_addr);
3459                         else
3460                                 id_priv->used_resolve_ip = 1;
3461                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3462                                               &id->route.addr.dev_addr,
3463                                               timeout_ms, addr_handler,
3464                                               false, id_priv);
3465                 }
3466         }
3467         if (ret)
3468                 goto err;
3469
3470         return 0;
3471 err:
3472         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3473         return ret;
3474 }
3475 EXPORT_SYMBOL(rdma_resolve_addr);
3476
3477 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3478 {
3479         struct rdma_id_private *id_priv;
3480         unsigned long flags;
3481         int ret;
3482
3483         id_priv = container_of(id, struct rdma_id_private, id);
3484         spin_lock_irqsave(&id_priv->lock, flags);
3485         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3486             id_priv->state == RDMA_CM_IDLE) {
3487                 id_priv->reuseaddr = reuse;
3488                 ret = 0;
3489         } else {
3490                 ret = -EINVAL;
3491         }
3492         spin_unlock_irqrestore(&id_priv->lock, flags);
3493         return ret;
3494 }
3495 EXPORT_SYMBOL(rdma_set_reuseaddr);
3496
3497 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3498 {
3499         struct rdma_id_private *id_priv;
3500         unsigned long flags;
3501         int ret;
3502
3503         id_priv = container_of(id, struct rdma_id_private, id);
3504         spin_lock_irqsave(&id_priv->lock, flags);
3505         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3506                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3507                 id_priv->afonly = afonly;
3508                 ret = 0;
3509         } else {
3510                 ret = -EINVAL;
3511         }
3512         spin_unlock_irqrestore(&id_priv->lock, flags);
3513         return ret;
3514 }
3515 EXPORT_SYMBOL(rdma_set_afonly);
3516
3517 static void cma_bind_port(struct rdma_bind_list *bind_list,
3518                           struct rdma_id_private *id_priv)
3519 {
3520         struct sockaddr *addr;
3521         struct sockaddr_ib *sib;
3522         u64 sid, mask;
3523         __be16 port;
3524
3525         lockdep_assert_held(&lock);
3526
3527         addr = cma_src_addr(id_priv);
3528         port = htons(bind_list->port);
3529
3530         switch (addr->sa_family) {
3531         case AF_INET:
3532                 ((struct sockaddr_in *) addr)->sin_port = port;
3533                 break;
3534         case AF_INET6:
3535                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3536                 break;
3537         case AF_IB:
3538                 sib = (struct sockaddr_ib *) addr;
3539                 sid = be64_to_cpu(sib->sib_sid);
3540                 mask = be64_to_cpu(sib->sib_sid_mask);
3541                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3542                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3543                 break;
3544         }
3545         id_priv->bind_list = bind_list;
3546         hlist_add_head(&id_priv->node, &bind_list->owners);
3547 }
3548
3549 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3550                           struct rdma_id_private *id_priv, unsigned short snum)
3551 {
3552         struct rdma_bind_list *bind_list;
3553         int ret;
3554
3555         lockdep_assert_held(&lock);
3556
3557         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3558         if (!bind_list)
3559                 return -ENOMEM;
3560
3561         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3562                            snum);
3563         if (ret < 0)
3564                 goto err;
3565
3566         bind_list->ps = ps;
3567         bind_list->port = snum;
3568         cma_bind_port(bind_list, id_priv);
3569         return 0;
3570 err:
3571         kfree(bind_list);
3572         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3573 }
3574
3575 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3576                               struct rdma_id_private *id_priv)
3577 {
3578         struct rdma_id_private *cur_id;
3579         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3580         struct sockaddr  *saddr = cma_src_addr(id_priv);
3581         __be16 dport = cma_port(daddr);
3582
3583         lockdep_assert_held(&lock);
3584
3585         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3586                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3587                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3588                 __be16 cur_dport = cma_port(cur_daddr);
3589
3590                 if (id_priv == cur_id)
3591                         continue;
3592
3593                 /* different dest port -> unique */
3594                 if (!cma_any_port(daddr) &&
3595                     !cma_any_port(cur_daddr) &&
3596                     (dport != cur_dport))
3597                         continue;
3598
3599                 /* different src address -> unique */
3600                 if (!cma_any_addr(saddr) &&
3601                     !cma_any_addr(cur_saddr) &&
3602                     cma_addr_cmp(saddr, cur_saddr))
3603                         continue;
3604
3605                 /* different dst address -> unique */
3606                 if (!cma_any_addr(daddr) &&
3607                     !cma_any_addr(cur_daddr) &&
3608                     cma_addr_cmp(daddr, cur_daddr))
3609                         continue;
3610
3611                 return -EADDRNOTAVAIL;
3612         }
3613         return 0;
3614 }
3615
3616 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3617                               struct rdma_id_private *id_priv)
3618 {
3619         static unsigned int last_used_port;
3620         int low, high, remaining;
3621         unsigned int rover;
3622         struct net *net = id_priv->id.route.addr.dev_addr.net;
3623
3624         lockdep_assert_held(&lock);
3625
3626         inet_get_local_port_range(net, &low, &high);
3627         remaining = (high - low) + 1;
3628         rover = prandom_u32() % remaining + low;
3629 retry:
3630         if (last_used_port != rover) {
3631                 struct rdma_bind_list *bind_list;
3632                 int ret;
3633
3634                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3635
3636                 if (!bind_list) {
3637                         ret = cma_alloc_port(ps, id_priv, rover);
3638                 } else {
3639                         ret = cma_port_is_unique(bind_list, id_priv);
3640                         if (!ret)
3641                                 cma_bind_port(bind_list, id_priv);
3642                 }
3643                 /*
3644                  * Remember previously used port number in order to avoid
3645                  * re-using same port immediately after it is closed.
3646                  */
3647                 if (!ret)
3648                         last_used_port = rover;
3649                 if (ret != -EADDRNOTAVAIL)
3650                         return ret;
3651         }
3652         if (--remaining) {
3653                 rover++;
3654                 if ((rover < low) || (rover > high))
3655                         rover = low;
3656                 goto retry;
3657         }
3658         return -EADDRNOTAVAIL;
3659 }
3660
3661 /*
3662  * Check that the requested port is available.  This is called when trying to
3663  * bind to a specific port, or when trying to listen on a bound port.  In
3664  * the latter case, the provided id_priv may already be on the bind_list, but
3665  * we still need to check that it's okay to start listening.
3666  */
3667 static int cma_check_port(struct rdma_bind_list *bind_list,
3668                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3669 {
3670         struct rdma_id_private *cur_id;
3671         struct sockaddr *addr, *cur_addr;
3672
3673         lockdep_assert_held(&lock);
3674
3675         addr = cma_src_addr(id_priv);
3676         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3677                 if (id_priv == cur_id)
3678                         continue;
3679
3680                 if (reuseaddr && cur_id->reuseaddr)
3681                         continue;
3682
3683                 cur_addr = cma_src_addr(cur_id);
3684                 if (id_priv->afonly && cur_id->afonly &&
3685                     (addr->sa_family != cur_addr->sa_family))
3686                         continue;
3687
3688                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3689                         return -EADDRNOTAVAIL;
3690
3691                 if (!cma_addr_cmp(addr, cur_addr))
3692                         return -EADDRINUSE;
3693         }
3694         return 0;
3695 }
3696
3697 static int cma_use_port(enum rdma_ucm_port_space ps,
3698                         struct rdma_id_private *id_priv)
3699 {
3700         struct rdma_bind_list *bind_list;
3701         unsigned short snum;
3702         int ret;
3703
3704         lockdep_assert_held(&lock);
3705
3706         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3707         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3708                 return -EACCES;
3709
3710         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3711         if (!bind_list) {
3712                 ret = cma_alloc_port(ps, id_priv, snum);
3713         } else {
3714                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3715                 if (!ret)
3716                         cma_bind_port(bind_list, id_priv);
3717         }
3718         return ret;
3719 }
3720
3721 static enum rdma_ucm_port_space
3722 cma_select_inet_ps(struct rdma_id_private *id_priv)
3723 {
3724         switch (id_priv->id.ps) {
3725         case RDMA_PS_TCP:
3726         case RDMA_PS_UDP:
3727         case RDMA_PS_IPOIB:
3728         case RDMA_PS_IB:
3729                 return id_priv->id.ps;
3730         default:
3731
3732                 return 0;
3733         }
3734 }
3735
3736 static enum rdma_ucm_port_space
3737 cma_select_ib_ps(struct rdma_id_private *id_priv)
3738 {
3739         enum rdma_ucm_port_space ps = 0;
3740         struct sockaddr_ib *sib;
3741         u64 sid_ps, mask, sid;
3742
3743         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3744         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3745         sid = be64_to_cpu(sib->sib_sid) & mask;
3746
3747         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3748                 sid_ps = RDMA_IB_IP_PS_IB;
3749                 ps = RDMA_PS_IB;
3750         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3751                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3752                 sid_ps = RDMA_IB_IP_PS_TCP;
3753                 ps = RDMA_PS_TCP;
3754         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3755                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3756                 sid_ps = RDMA_IB_IP_PS_UDP;
3757                 ps = RDMA_PS_UDP;
3758         }
3759
3760         if (ps) {
3761                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3762                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3763                                                 be64_to_cpu(sib->sib_sid_mask));
3764         }
3765         return ps;
3766 }
3767
3768 static int cma_get_port(struct rdma_id_private *id_priv)
3769 {
3770         enum rdma_ucm_port_space ps;
3771         int ret;
3772
3773         if (cma_family(id_priv) != AF_IB)
3774                 ps = cma_select_inet_ps(id_priv);
3775         else
3776                 ps = cma_select_ib_ps(id_priv);
3777         if (!ps)
3778                 return -EPROTONOSUPPORT;
3779
3780         mutex_lock(&lock);
3781         if (cma_any_port(cma_src_addr(id_priv)))
3782                 ret = cma_alloc_any_port(ps, id_priv);
3783         else
3784                 ret = cma_use_port(ps, id_priv);
3785         mutex_unlock(&lock);
3786
3787         return ret;
3788 }
3789
3790 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3791                                struct sockaddr *addr)
3792 {
3793 #if IS_ENABLED(CONFIG_IPV6)
3794         struct sockaddr_in6 *sin6;
3795
3796         if (addr->sa_family != AF_INET6)
3797                 return 0;
3798
3799         sin6 = (struct sockaddr_in6 *) addr;
3800
3801         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3802                 return 0;
3803
3804         if (!sin6->sin6_scope_id)
3805                         return -EINVAL;
3806
3807         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3808 #endif
3809         return 0;
3810 }
3811
3812 int rdma_listen(struct rdma_cm_id *id, int backlog)
3813 {
3814         struct rdma_id_private *id_priv =
3815                 container_of(id, struct rdma_id_private, id);
3816         int ret;
3817
3818         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3819                 struct sockaddr_in any_in = {
3820                         .sin_family = AF_INET,
3821                         .sin_addr.s_addr = htonl(INADDR_ANY),
3822                 };
3823
3824                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3825                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3826                 if (ret)
3827                         return ret;
3828                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3829                                            RDMA_CM_LISTEN)))
3830                         return -EINVAL;
3831         }
3832
3833         /*
3834          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3835          * any more, and has to be unique in the bind list.
3836          */
3837         if (id_priv->reuseaddr) {
3838                 mutex_lock(&lock);
3839                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3840                 if (!ret)
3841                         id_priv->reuseaddr = 0;
3842                 mutex_unlock(&lock);
3843                 if (ret)
3844                         goto err;
3845         }
3846
3847         id_priv->backlog = backlog;
3848         if (id_priv->cma_dev) {
3849                 if (rdma_cap_ib_cm(id->device, 1)) {
3850                         ret = cma_ib_listen(id_priv);
3851                         if (ret)
3852                                 goto err;
3853                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3854                         ret = cma_iw_listen(id_priv, backlog);
3855                         if (ret)
3856                                 goto err;
3857                 } else {
3858                         ret = -ENOSYS;
3859                         goto err;
3860                 }
3861         } else {
3862                 ret = cma_listen_on_all(id_priv);
3863                 if (ret)
3864                         goto err;
3865         }
3866
3867         return 0;
3868 err:
3869         id_priv->backlog = 0;
3870         /*
3871          * All the failure paths that lead here will not allow the req_handler's
3872          * to have run.
3873          */
3874         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3875         return ret;
3876 }
3877 EXPORT_SYMBOL(rdma_listen);
3878
3879 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3880 {
3881         struct rdma_id_private *id_priv;
3882         int ret;
3883         struct sockaddr  *daddr;
3884
3885         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3886             addr->sa_family != AF_IB)
3887                 return -EAFNOSUPPORT;
3888
3889         id_priv = container_of(id, struct rdma_id_private, id);
3890         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3891                 return -EINVAL;
3892
3893         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3894         if (ret)
3895                 goto err1;
3896
3897         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3898         if (!cma_any_addr(addr)) {
3899                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3900                 if (ret)
3901                         goto err1;
3902
3903                 ret = cma_acquire_dev_by_src_ip(id_priv);
3904                 if (ret)
3905                         goto err1;
3906         }
3907
3908         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3909                 if (addr->sa_family == AF_INET)
3910                         id_priv->afonly = 1;
3911 #if IS_ENABLED(CONFIG_IPV6)
3912                 else if (addr->sa_family == AF_INET6) {
3913                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3914
3915                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3916                 }
3917 #endif
3918         }
3919         daddr = cma_dst_addr(id_priv);
3920         daddr->sa_family = addr->sa_family;
3921
3922         ret = cma_get_port(id_priv);
3923         if (ret)
3924                 goto err2;
3925
3926         if (!cma_any_addr(addr))
3927                 rdma_restrack_add(&id_priv->res);
3928         return 0;
3929 err2:
3930         if (id_priv->cma_dev)
3931                 cma_release_dev(id_priv);
3932 err1:
3933         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3934         return ret;
3935 }
3936 EXPORT_SYMBOL(rdma_bind_addr);
3937
3938 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3939 {
3940         struct cma_hdr *cma_hdr;
3941
3942         cma_hdr = hdr;
3943         cma_hdr->cma_version = CMA_VERSION;
3944         if (cma_family(id_priv) == AF_INET) {
3945                 struct sockaddr_in *src4, *dst4;
3946
3947                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3948                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3949
3950                 cma_set_ip_ver(cma_hdr, 4);
3951                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3952                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3953                 cma_hdr->port = src4->sin_port;
3954         } else if (cma_family(id_priv) == AF_INET6) {
3955                 struct sockaddr_in6 *src6, *dst6;
3956
3957                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3958                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3959
3960                 cma_set_ip_ver(cma_hdr, 6);
3961                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3962                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3963                 cma_hdr->port = src6->sin6_port;
3964         }
3965         return 0;
3966 }
3967
3968 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3969                                 const struct ib_cm_event *ib_event)
3970 {
3971         struct rdma_id_private *id_priv = cm_id->context;
3972         struct rdma_cm_event event = {};
3973         const struct ib_cm_sidr_rep_event_param *rep =
3974                                 &ib_event->param.sidr_rep_rcvd;
3975         int ret;
3976
3977         mutex_lock(&id_priv->handler_mutex);
3978         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3979                 goto out;
3980
3981         switch (ib_event->event) {
3982         case IB_CM_SIDR_REQ_ERROR:
3983                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3984                 event.status = -ETIMEDOUT;
3985                 break;
3986         case IB_CM_SIDR_REP_RECEIVED:
3987                 event.param.ud.private_data = ib_event->private_data;
3988                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3989                 if (rep->status != IB_SIDR_SUCCESS) {
3990                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3991                         event.status = ib_event->param.sidr_rep_rcvd.status;
3992                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3993                                              event.status);
3994                         break;
3995                 }
3996                 ret = cma_set_qkey(id_priv, rep->qkey);
3997                 if (ret) {
3998                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3999                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
4000                         event.status = ret;
4001                         break;
4002                 }
4003                 ib_init_ah_attr_from_path(id_priv->id.device,
4004                                           id_priv->id.port_num,
4005                                           id_priv->id.route.path_rec,
4006                                           &event.param.ud.ah_attr,
4007                                           rep->sgid_attr);
4008                 event.param.ud.qp_num = rep->qpn;
4009                 event.param.ud.qkey = rep->qkey;
4010                 event.event = RDMA_CM_EVENT_ESTABLISHED;
4011                 event.status = 0;
4012                 break;
4013         default:
4014                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4015                        ib_event->event);
4016                 goto out;
4017         }
4018
4019         ret = cma_cm_event_handler(id_priv, &event);
4020
4021         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4022         if (ret) {
4023                 /* Destroy the CM ID by returning a non-zero value. */
4024                 id_priv->cm_id.ib = NULL;
4025                 destroy_id_handler_unlock(id_priv);
4026                 return ret;
4027         }
4028 out:
4029         mutex_unlock(&id_priv->handler_mutex);
4030         return 0;
4031 }
4032
4033 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4034                               struct rdma_conn_param *conn_param)
4035 {
4036         struct ib_cm_sidr_req_param req;
4037         struct ib_cm_id *id;
4038         void *private_data;
4039         u8 offset;
4040         int ret;
4041
4042         memset(&req, 0, sizeof req);
4043         offset = cma_user_data_offset(id_priv);
4044         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4045                 return -EINVAL;
4046
4047         if (req.private_data_len) {
4048                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4049                 if (!private_data)
4050                         return -ENOMEM;
4051         } else {
4052                 private_data = NULL;
4053         }
4054
4055         if (conn_param->private_data && conn_param->private_data_len)
4056                 memcpy(private_data + offset, conn_param->private_data,
4057                        conn_param->private_data_len);
4058
4059         if (private_data) {
4060                 ret = cma_format_hdr(private_data, id_priv);
4061                 if (ret)
4062                         goto out;
4063                 req.private_data = private_data;
4064         }
4065
4066         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4067                              id_priv);
4068         if (IS_ERR(id)) {
4069                 ret = PTR_ERR(id);
4070                 goto out;
4071         }
4072         id_priv->cm_id.ib = id;
4073
4074         req.path = id_priv->id.route.path_rec;
4075         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4076         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4077         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4078         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4079
4080         trace_cm_send_sidr_req(id_priv);
4081         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4082         if (ret) {
4083                 ib_destroy_cm_id(id_priv->cm_id.ib);
4084                 id_priv->cm_id.ib = NULL;
4085         }
4086 out:
4087         kfree(private_data);
4088         return ret;
4089 }
4090
4091 static int cma_connect_ib(struct rdma_id_private *id_priv,
4092                           struct rdma_conn_param *conn_param)
4093 {
4094         struct ib_cm_req_param req;
4095         struct rdma_route *route;
4096         void *private_data;
4097         struct ib_cm_id *id;
4098         u8 offset;
4099         int ret;
4100
4101         memset(&req, 0, sizeof req);
4102         offset = cma_user_data_offset(id_priv);
4103         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4104                 return -EINVAL;
4105
4106         if (req.private_data_len) {
4107                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4108                 if (!private_data)
4109                         return -ENOMEM;
4110         } else {
4111                 private_data = NULL;
4112         }
4113
4114         if (conn_param->private_data && conn_param->private_data_len)
4115                 memcpy(private_data + offset, conn_param->private_data,
4116                        conn_param->private_data_len);
4117
4118         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4119         if (IS_ERR(id)) {
4120                 ret = PTR_ERR(id);
4121                 goto out;
4122         }
4123         id_priv->cm_id.ib = id;
4124
4125         route = &id_priv->id.route;
4126         if (private_data) {
4127                 ret = cma_format_hdr(private_data, id_priv);
4128                 if (ret)
4129                         goto out;
4130                 req.private_data = private_data;
4131         }
4132
4133         req.primary_path = &route->path_rec[0];
4134         if (route->num_paths == 2)
4135                 req.alternate_path = &route->path_rec[1];
4136
4137         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4138         /* Alternate path SGID attribute currently unsupported */
4139         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4140         req.qp_num = id_priv->qp_num;
4141         req.qp_type = id_priv->id.qp_type;
4142         req.starting_psn = id_priv->seq_num;
4143         req.responder_resources = conn_param->responder_resources;
4144         req.initiator_depth = conn_param->initiator_depth;
4145         req.flow_control = conn_param->flow_control;
4146         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4147         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4148         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4149         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4150         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4151         req.srq = id_priv->srq ? 1 : 0;
4152         req.ece.vendor_id = id_priv->ece.vendor_id;
4153         req.ece.attr_mod = id_priv->ece.attr_mod;
4154
4155         trace_cm_send_req(id_priv);
4156         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4157 out:
4158         if (ret && !IS_ERR(id)) {
4159                 ib_destroy_cm_id(id);
4160                 id_priv->cm_id.ib = NULL;
4161         }
4162
4163         kfree(private_data);
4164         return ret;
4165 }
4166
4167 static int cma_connect_iw(struct rdma_id_private *id_priv,
4168                           struct rdma_conn_param *conn_param)
4169 {
4170         struct iw_cm_id *cm_id;
4171         int ret;
4172         struct iw_cm_conn_param iw_param;
4173
4174         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4175         if (IS_ERR(cm_id))
4176                 return PTR_ERR(cm_id);
4177
4178         mutex_lock(&id_priv->qp_mutex);
4179         cm_id->tos = id_priv->tos;
4180         cm_id->tos_set = id_priv->tos_set;
4181         mutex_unlock(&id_priv->qp_mutex);
4182
4183         id_priv->cm_id.iw = cm_id;
4184
4185         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4186                rdma_addr_size(cma_src_addr(id_priv)));
4187         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4188                rdma_addr_size(cma_dst_addr(id_priv)));
4189
4190         ret = cma_modify_qp_rtr(id_priv, conn_param);
4191         if (ret)
4192                 goto out;
4193
4194         if (conn_param) {
4195                 iw_param.ord = conn_param->initiator_depth;
4196                 iw_param.ird = conn_param->responder_resources;
4197                 iw_param.private_data = conn_param->private_data;
4198                 iw_param.private_data_len = conn_param->private_data_len;
4199                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4200         } else {
4201                 memset(&iw_param, 0, sizeof iw_param);
4202                 iw_param.qpn = id_priv->qp_num;
4203         }
4204         ret = iw_cm_connect(cm_id, &iw_param);
4205 out:
4206         if (ret) {
4207                 iw_destroy_cm_id(cm_id);
4208                 id_priv->cm_id.iw = NULL;
4209         }
4210         return ret;
4211 }
4212
4213 /**
4214  * rdma_connect_locked - Initiate an active connection request.
4215  * @id: Connection identifier to connect.
4216  * @conn_param: Connection information used for connected QPs.
4217  *
4218  * Same as rdma_connect() but can only be called from the
4219  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4220  */
4221 int rdma_connect_locked(struct rdma_cm_id *id,
4222                         struct rdma_conn_param *conn_param)
4223 {
4224         struct rdma_id_private *id_priv =
4225                 container_of(id, struct rdma_id_private, id);
4226         int ret;
4227
4228         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4229                 return -EINVAL;
4230
4231         if (!id->qp) {
4232                 id_priv->qp_num = conn_param->qp_num;
4233                 id_priv->srq = conn_param->srq;
4234         }
4235
4236         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4237                 if (id->qp_type == IB_QPT_UD)
4238                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4239                 else
4240                         ret = cma_connect_ib(id_priv, conn_param);
4241         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4242                 ret = cma_connect_iw(id_priv, conn_param);
4243         } else {
4244                 ret = -ENOSYS;
4245         }
4246         if (ret)
4247                 goto err_state;
4248         return 0;
4249 err_state:
4250         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4251         return ret;
4252 }
4253 EXPORT_SYMBOL(rdma_connect_locked);
4254
4255 /**
4256  * rdma_connect - Initiate an active connection request.
4257  * @id: Connection identifier to connect.
4258  * @conn_param: Connection information used for connected QPs.
4259  *
4260  * Users must have resolved a route for the rdma_cm_id to connect with by having
4261  * called rdma_resolve_route before calling this routine.
4262  *
4263  * This call will either connect to a remote QP or obtain remote QP information
4264  * for unconnected rdma_cm_id's.  The actual operation is based on the
4265  * rdma_cm_id's port space.
4266  */
4267 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4268 {
4269         struct rdma_id_private *id_priv =
4270                 container_of(id, struct rdma_id_private, id);
4271         int ret;
4272
4273         mutex_lock(&id_priv->handler_mutex);
4274         ret = rdma_connect_locked(id, conn_param);
4275         mutex_unlock(&id_priv->handler_mutex);
4276         return ret;
4277 }
4278 EXPORT_SYMBOL(rdma_connect);
4279
4280 /**
4281  * rdma_connect_ece - Initiate an active connection request with ECE data.
4282  * @id: Connection identifier to connect.
4283  * @conn_param: Connection information used for connected QPs.
4284  * @ece: ECE parameters
4285  *
4286  * See rdma_connect() explanation.
4287  */
4288 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4289                      struct rdma_ucm_ece *ece)
4290 {
4291         struct rdma_id_private *id_priv =
4292                 container_of(id, struct rdma_id_private, id);
4293
4294         id_priv->ece.vendor_id = ece->vendor_id;
4295         id_priv->ece.attr_mod = ece->attr_mod;
4296
4297         return rdma_connect(id, conn_param);
4298 }
4299 EXPORT_SYMBOL(rdma_connect_ece);
4300
4301 static int cma_accept_ib(struct rdma_id_private *id_priv,
4302                          struct rdma_conn_param *conn_param)
4303 {
4304         struct ib_cm_rep_param rep;
4305         int ret;
4306
4307         ret = cma_modify_qp_rtr(id_priv, conn_param);
4308         if (ret)
4309                 goto out;
4310
4311         ret = cma_modify_qp_rts(id_priv, conn_param);
4312         if (ret)
4313                 goto out;
4314
4315         memset(&rep, 0, sizeof rep);
4316         rep.qp_num = id_priv->qp_num;
4317         rep.starting_psn = id_priv->seq_num;
4318         rep.private_data = conn_param->private_data;
4319         rep.private_data_len = conn_param->private_data_len;
4320         rep.responder_resources = conn_param->responder_resources;
4321         rep.initiator_depth = conn_param->initiator_depth;
4322         rep.failover_accepted = 0;
4323         rep.flow_control = conn_param->flow_control;
4324         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4325         rep.srq = id_priv->srq ? 1 : 0;
4326         rep.ece.vendor_id = id_priv->ece.vendor_id;
4327         rep.ece.attr_mod = id_priv->ece.attr_mod;
4328
4329         trace_cm_send_rep(id_priv);
4330         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4331 out:
4332         return ret;
4333 }
4334
4335 static int cma_accept_iw(struct rdma_id_private *id_priv,
4336                   struct rdma_conn_param *conn_param)
4337 {
4338         struct iw_cm_conn_param iw_param;
4339         int ret;
4340
4341         if (!conn_param)
4342                 return -EINVAL;
4343
4344         ret = cma_modify_qp_rtr(id_priv, conn_param);
4345         if (ret)
4346                 return ret;
4347
4348         iw_param.ord = conn_param->initiator_depth;
4349         iw_param.ird = conn_param->responder_resources;
4350         iw_param.private_data = conn_param->private_data;
4351         iw_param.private_data_len = conn_param->private_data_len;
4352         if (id_priv->id.qp)
4353                 iw_param.qpn = id_priv->qp_num;
4354         else
4355                 iw_param.qpn = conn_param->qp_num;
4356
4357         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4358 }
4359
4360 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4361                              enum ib_cm_sidr_status status, u32 qkey,
4362                              const void *private_data, int private_data_len)
4363 {
4364         struct ib_cm_sidr_rep_param rep;
4365         int ret;
4366
4367         memset(&rep, 0, sizeof rep);
4368         rep.status = status;
4369         if (status == IB_SIDR_SUCCESS) {
4370                 ret = cma_set_qkey(id_priv, qkey);
4371                 if (ret)
4372                         return ret;
4373                 rep.qp_num = id_priv->qp_num;
4374                 rep.qkey = id_priv->qkey;
4375
4376                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4377                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4378         }
4379
4380         rep.private_data = private_data;
4381         rep.private_data_len = private_data_len;
4382
4383         trace_cm_send_sidr_rep(id_priv);
4384         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4385 }
4386
4387 /**
4388  * rdma_accept - Called to accept a connection request or response.
4389  * @id: Connection identifier associated with the request.
4390  * @conn_param: Information needed to establish the connection.  This must be
4391  *   provided if accepting a connection request.  If accepting a connection
4392  *   response, this parameter must be NULL.
4393  *
4394  * Typically, this routine is only called by the listener to accept a connection
4395  * request.  It must also be called on the active side of a connection if the
4396  * user is performing their own QP transitions.
4397  *
4398  * In the case of error, a reject message is sent to the remote side and the
4399  * state of the qp associated with the id is modified to error, such that any
4400  * previously posted receive buffers would be flushed.
4401  *
4402  * This function is for use by kernel ULPs and must be called from under the
4403  * handler callback.
4404  */
4405 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4406 {
4407         struct rdma_id_private *id_priv =
4408                 container_of(id, struct rdma_id_private, id);
4409         int ret;
4410
4411         lockdep_assert_held(&id_priv->handler_mutex);
4412
4413         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4414                 return -EINVAL;
4415
4416         if (!id->qp && conn_param) {
4417                 id_priv->qp_num = conn_param->qp_num;
4418                 id_priv->srq = conn_param->srq;
4419         }
4420
4421         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4422                 if (id->qp_type == IB_QPT_UD) {
4423                         if (conn_param)
4424                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4425                                                         conn_param->qkey,
4426                                                         conn_param->private_data,
4427                                                         conn_param->private_data_len);
4428                         else
4429                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4430                                                         0, NULL, 0);
4431                 } else {
4432                         if (conn_param)
4433                                 ret = cma_accept_ib(id_priv, conn_param);
4434                         else
4435                                 ret = cma_rep_recv(id_priv);
4436                 }
4437         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4438                 ret = cma_accept_iw(id_priv, conn_param);
4439         } else {
4440                 ret = -ENOSYS;
4441         }
4442         if (ret)
4443                 goto reject;
4444
4445         return 0;
4446 reject:
4447         cma_modify_qp_err(id_priv);
4448         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4449         return ret;
4450 }
4451 EXPORT_SYMBOL(rdma_accept);
4452
4453 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4454                     struct rdma_ucm_ece *ece)
4455 {
4456         struct rdma_id_private *id_priv =
4457                 container_of(id, struct rdma_id_private, id);
4458
4459         id_priv->ece.vendor_id = ece->vendor_id;
4460         id_priv->ece.attr_mod = ece->attr_mod;
4461
4462         return rdma_accept(id, conn_param);
4463 }
4464 EXPORT_SYMBOL(rdma_accept_ece);
4465
4466 void rdma_lock_handler(struct rdma_cm_id *id)
4467 {
4468         struct rdma_id_private *id_priv =
4469                 container_of(id, struct rdma_id_private, id);
4470
4471         mutex_lock(&id_priv->handler_mutex);
4472 }
4473 EXPORT_SYMBOL(rdma_lock_handler);
4474
4475 void rdma_unlock_handler(struct rdma_cm_id *id)
4476 {
4477         struct rdma_id_private *id_priv =
4478                 container_of(id, struct rdma_id_private, id);
4479
4480         mutex_unlock(&id_priv->handler_mutex);
4481 }
4482 EXPORT_SYMBOL(rdma_unlock_handler);
4483
4484 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4485 {
4486         struct rdma_id_private *id_priv;
4487         int ret;
4488
4489         id_priv = container_of(id, struct rdma_id_private, id);
4490         if (!id_priv->cm_id.ib)
4491                 return -EINVAL;
4492
4493         switch (id->device->node_type) {
4494         case RDMA_NODE_IB_CA:
4495                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4496                 break;
4497         default:
4498                 ret = 0;
4499                 break;
4500         }
4501         return ret;
4502 }
4503 EXPORT_SYMBOL(rdma_notify);
4504
4505 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4506                 u8 private_data_len, u8 reason)
4507 {
4508         struct rdma_id_private *id_priv;
4509         int ret;
4510
4511         id_priv = container_of(id, struct rdma_id_private, id);
4512         if (!id_priv->cm_id.ib)
4513                 return -EINVAL;
4514
4515         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4516                 if (id->qp_type == IB_QPT_UD) {
4517                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4518                                                 private_data, private_data_len);
4519                 } else {
4520                         trace_cm_send_rej(id_priv);
4521                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4522                                              private_data, private_data_len);
4523                 }
4524         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4525                 ret = iw_cm_reject(id_priv->cm_id.iw,
4526                                    private_data, private_data_len);
4527         } else {
4528                 ret = -ENOSYS;
4529         }
4530
4531         return ret;
4532 }
4533 EXPORT_SYMBOL(rdma_reject);
4534
4535 int rdma_disconnect(struct rdma_cm_id *id)
4536 {
4537         struct rdma_id_private *id_priv;
4538         int ret;
4539
4540         id_priv = container_of(id, struct rdma_id_private, id);
4541         if (!id_priv->cm_id.ib)
4542                 return -EINVAL;
4543
4544         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4545                 ret = cma_modify_qp_err(id_priv);
4546                 if (ret)
4547                         goto out;
4548                 /* Initiate or respond to a disconnect. */
4549                 trace_cm_disconnect(id_priv);
4550                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4551                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4552                                 trace_cm_sent_drep(id_priv);
4553                 } else {
4554                         trace_cm_sent_dreq(id_priv);
4555                 }
4556         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4557                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4558         } else
4559                 ret = -EINVAL;
4560
4561 out:
4562         return ret;
4563 }
4564 EXPORT_SYMBOL(rdma_disconnect);
4565
4566 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4567                               struct ib_sa_multicast *multicast,
4568                               struct rdma_cm_event *event,
4569                               struct cma_multicast *mc)
4570 {
4571         struct rdma_dev_addr *dev_addr;
4572         enum ib_gid_type gid_type;
4573         struct net_device *ndev;
4574
4575         if (!status)
4576                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4577         else
4578                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4579                                      status);
4580
4581         event->status = status;
4582         event->param.ud.private_data = mc->context;
4583         if (status) {
4584                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4585                 return;
4586         }
4587
4588         dev_addr = &id_priv->id.route.addr.dev_addr;
4589         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4590         gid_type =
4591                 id_priv->cma_dev
4592                         ->default_gid_type[id_priv->id.port_num -
4593                                            rdma_start_port(
4594                                                    id_priv->cma_dev->device)];
4595
4596         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4597         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4598                                      &multicast->rec, ndev, gid_type,
4599                                      &event->param.ud.ah_attr)) {
4600                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4601                 goto out;
4602         }
4603
4604         event->param.ud.qp_num = 0xFFFFFF;
4605         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4606
4607 out:
4608         if (ndev)
4609                 dev_put(ndev);
4610 }
4611
4612 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4613 {
4614         struct cma_multicast *mc = multicast->context;
4615         struct rdma_id_private *id_priv = mc->id_priv;
4616         struct rdma_cm_event event = {};
4617         int ret = 0;
4618
4619         mutex_lock(&id_priv->handler_mutex);
4620         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4621             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4622                 goto out;
4623
4624         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4625         ret = cma_cm_event_handler(id_priv, &event);
4626         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4627         WARN_ON(ret);
4628
4629 out:
4630         mutex_unlock(&id_priv->handler_mutex);
4631         return 0;
4632 }
4633
4634 static void cma_set_mgid(struct rdma_id_private *id_priv,
4635                          struct sockaddr *addr, union ib_gid *mgid)
4636 {
4637         unsigned char mc_map[MAX_ADDR_LEN];
4638         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4639         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4640         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4641
4642         if (cma_any_addr(addr)) {
4643                 memset(mgid, 0, sizeof *mgid);
4644         } else if ((addr->sa_family == AF_INET6) &&
4645                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4646                                                                  0xFF10A01B)) {
4647                 /* IPv6 address is an SA assigned MGID. */
4648                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4649         } else if (addr->sa_family == AF_IB) {
4650                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4651         } else if (addr->sa_family == AF_INET6) {
4652                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4653                 if (id_priv->id.ps == RDMA_PS_UDP)
4654                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4655                 *mgid = *(union ib_gid *) (mc_map + 4);
4656         } else {
4657                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4658                 if (id_priv->id.ps == RDMA_PS_UDP)
4659                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4660                 *mgid = *(union ib_gid *) (mc_map + 4);
4661         }
4662 }
4663
4664 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4665                                  struct cma_multicast *mc)
4666 {
4667         struct ib_sa_mcmember_rec rec;
4668         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4669         ib_sa_comp_mask comp_mask;
4670         int ret;
4671
4672         ib_addr_get_mgid(dev_addr, &rec.mgid);
4673         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4674                                      &rec.mgid, &rec);
4675         if (ret)
4676                 return ret;
4677
4678         ret = cma_set_qkey(id_priv, 0);
4679         if (ret)
4680                 return ret;
4681
4682         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4683         rec.qkey = cpu_to_be32(id_priv->qkey);
4684         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4685         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4686         rec.join_state = mc->join_state;
4687
4688         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4689                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4690                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4691                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4692                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4693
4694         if (id_priv->id.ps == RDMA_PS_IPOIB)
4695                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4696                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4697                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4698                              IB_SA_MCMEMBER_REC_MTU |
4699                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4700
4701         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4702                                          id_priv->id.port_num, &rec, comp_mask,
4703                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4704         return PTR_ERR_OR_ZERO(mc->sa_mc);
4705 }
4706
4707 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4708                               enum ib_gid_type gid_type)
4709 {
4710         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4711         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4712
4713         if (cma_any_addr(addr)) {
4714                 memset(mgid, 0, sizeof *mgid);
4715         } else if (addr->sa_family == AF_INET6) {
4716                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4717         } else {
4718                 mgid->raw[0] =
4719                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4720                 mgid->raw[1] =
4721                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4722                 mgid->raw[2] = 0;
4723                 mgid->raw[3] = 0;
4724                 mgid->raw[4] = 0;
4725                 mgid->raw[5] = 0;
4726                 mgid->raw[6] = 0;
4727                 mgid->raw[7] = 0;
4728                 mgid->raw[8] = 0;
4729                 mgid->raw[9] = 0;
4730                 mgid->raw[10] = 0xff;
4731                 mgid->raw[11] = 0xff;
4732                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4733         }
4734 }
4735
4736 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4737                                    struct cma_multicast *mc)
4738 {
4739         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4740         int err = 0;
4741         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4742         struct net_device *ndev = NULL;
4743         struct ib_sa_multicast ib;
4744         enum ib_gid_type gid_type;
4745         bool send_only;
4746
4747         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4748
4749         if (cma_zero_addr(addr))
4750                 return -EINVAL;
4751
4752         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4753                    rdma_start_port(id_priv->cma_dev->device)];
4754         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4755
4756         ib.rec.pkey = cpu_to_be16(0xffff);
4757         if (id_priv->id.ps == RDMA_PS_UDP)
4758                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4759
4760         if (dev_addr->bound_dev_if)
4761                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4762         if (!ndev)
4763                 return -ENODEV;
4764
4765         ib.rec.rate = iboe_get_rate(ndev);
4766         ib.rec.hop_limit = 1;
4767         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4768
4769         if (addr->sa_family == AF_INET) {
4770                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4771                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4772                         if (!send_only) {
4773                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4774                                                     true);
4775                         }
4776                 }
4777         } else {
4778                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4779                         err = -ENOTSUPP;
4780         }
4781         dev_put(ndev);
4782         if (err || !ib.rec.mtu)
4783                 return err ?: -EINVAL;
4784
4785         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4786                     &ib.rec.port_gid);
4787         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4788         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4789         queue_work(cma_wq, &mc->iboe_join.work);
4790         return 0;
4791 }
4792
4793 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4794                         u8 join_state, void *context)
4795 {
4796         struct rdma_id_private *id_priv =
4797                 container_of(id, struct rdma_id_private, id);
4798         struct cma_multicast *mc;
4799         int ret;
4800
4801         /* Not supported for kernel QPs */
4802         if (WARN_ON(id->qp))
4803                 return -EINVAL;
4804
4805         /* ULP is calling this wrong. */
4806         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4807                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4808                 return -EINVAL;
4809
4810         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4811         if (!mc)
4812                 return -ENOMEM;
4813
4814         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4815         mc->context = context;
4816         mc->id_priv = id_priv;
4817         mc->join_state = join_state;
4818
4819         if (rdma_protocol_roce(id->device, id->port_num)) {
4820                 ret = cma_iboe_join_multicast(id_priv, mc);
4821                 if (ret)
4822                         goto out_err;
4823         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4824                 ret = cma_join_ib_multicast(id_priv, mc);
4825                 if (ret)
4826                         goto out_err;
4827         } else {
4828                 ret = -ENOSYS;
4829                 goto out_err;
4830         }
4831
4832         spin_lock(&id_priv->lock);
4833         list_add(&mc->list, &id_priv->mc_list);
4834         spin_unlock(&id_priv->lock);
4835
4836         return 0;
4837 out_err:
4838         kfree(mc);
4839         return ret;
4840 }
4841 EXPORT_SYMBOL(rdma_join_multicast);
4842
4843 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4844 {
4845         struct rdma_id_private *id_priv;
4846         struct cma_multicast *mc;
4847
4848         id_priv = container_of(id, struct rdma_id_private, id);
4849         spin_lock_irq(&id_priv->lock);
4850         list_for_each_entry(mc, &id_priv->mc_list, list) {
4851                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4852                         continue;
4853                 list_del(&mc->list);
4854                 spin_unlock_irq(&id_priv->lock);
4855
4856                 WARN_ON(id_priv->cma_dev->device != id->device);
4857                 destroy_mc(id_priv, mc);
4858                 return;
4859         }
4860         spin_unlock_irq(&id_priv->lock);
4861 }
4862 EXPORT_SYMBOL(rdma_leave_multicast);
4863
4864 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4865 {
4866         struct rdma_dev_addr *dev_addr;
4867         struct cma_work *work;
4868
4869         dev_addr = &id_priv->id.route.addr.dev_addr;
4870
4871         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4872             (net_eq(dev_net(ndev), dev_addr->net)) &&
4873             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4874                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4875                         ndev->name, &id_priv->id);
4876                 work = kzalloc(sizeof *work, GFP_KERNEL);
4877                 if (!work)
4878                         return -ENOMEM;
4879
4880                 INIT_WORK(&work->work, cma_work_handler);
4881                 work->id = id_priv;
4882                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4883                 cma_id_get(id_priv);
4884                 queue_work(cma_wq, &work->work);
4885         }
4886
4887         return 0;
4888 }
4889
4890 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4891                                void *ptr)
4892 {
4893         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4894         struct cma_device *cma_dev;
4895         struct rdma_id_private *id_priv;
4896         int ret = NOTIFY_DONE;
4897
4898         if (event != NETDEV_BONDING_FAILOVER)
4899                 return NOTIFY_DONE;
4900
4901         if (!netif_is_bond_master(ndev))
4902                 return NOTIFY_DONE;
4903
4904         mutex_lock(&lock);
4905         list_for_each_entry(cma_dev, &dev_list, list)
4906                 list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
4907                         ret = cma_netdev_change(ndev, id_priv);
4908                         if (ret)
4909                                 goto out;
4910                 }
4911
4912 out:
4913         mutex_unlock(&lock);
4914         return ret;
4915 }
4916
4917 static struct notifier_block cma_nb = {
4918         .notifier_call = cma_netdev_callback
4919 };
4920
4921 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4922 {
4923         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4924         enum rdma_cm_state state;
4925         unsigned long flags;
4926
4927         mutex_lock(&id_priv->handler_mutex);
4928         /* Record that we want to remove the device */
4929         spin_lock_irqsave(&id_priv->lock, flags);
4930         state = id_priv->state;
4931         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4932                 spin_unlock_irqrestore(&id_priv->lock, flags);
4933                 mutex_unlock(&id_priv->handler_mutex);
4934                 cma_id_put(id_priv);
4935                 return;
4936         }
4937         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4938         spin_unlock_irqrestore(&id_priv->lock, flags);
4939
4940         if (cma_cm_event_handler(id_priv, &event)) {
4941                 /*
4942                  * At this point the ULP promises it won't call
4943                  * rdma_destroy_id() concurrently
4944                  */
4945                 cma_id_put(id_priv);
4946                 mutex_unlock(&id_priv->handler_mutex);
4947                 trace_cm_id_destroy(id_priv);
4948                 _destroy_id(id_priv, state);
4949                 return;
4950         }
4951         mutex_unlock(&id_priv->handler_mutex);
4952
4953         /*
4954          * If this races with destroy then the thread that first assigns state
4955          * to a destroying does the cancel.
4956          */
4957         cma_cancel_operation(id_priv, state);
4958         cma_id_put(id_priv);
4959 }
4960
4961 static void cma_process_remove(struct cma_device *cma_dev)
4962 {
4963         mutex_lock(&lock);
4964         while (!list_empty(&cma_dev->id_list)) {
4965                 struct rdma_id_private *id_priv = list_first_entry(
4966                         &cma_dev->id_list, struct rdma_id_private, device_item);
4967
4968                 list_del_init(&id_priv->listen_item);
4969                 list_del_init(&id_priv->device_item);
4970                 cma_id_get(id_priv);
4971                 mutex_unlock(&lock);
4972
4973                 cma_send_device_removal_put(id_priv);
4974
4975                 mutex_lock(&lock);
4976         }
4977         mutex_unlock(&lock);
4978
4979         cma_dev_put(cma_dev);
4980         wait_for_completion(&cma_dev->comp);
4981 }
4982
4983 static bool cma_supported(struct ib_device *device)
4984 {
4985         u32 i;
4986
4987         rdma_for_each_port(device, i) {
4988                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
4989                         return true;
4990         }
4991         return false;
4992 }
4993
4994 static int cma_add_one(struct ib_device *device)
4995 {
4996         struct rdma_id_private *to_destroy;
4997         struct cma_device *cma_dev;
4998         struct rdma_id_private *id_priv;
4999         unsigned long supported_gids = 0;
5000         int ret;
5001         u32 i;
5002
5003         if (!cma_supported(device))
5004                 return -EOPNOTSUPP;
5005
5006         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5007         if (!cma_dev)
5008                 return -ENOMEM;
5009
5010         cma_dev->device = device;
5011         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5012                                             sizeof(*cma_dev->default_gid_type),
5013                                             GFP_KERNEL);
5014         if (!cma_dev->default_gid_type) {
5015                 ret = -ENOMEM;
5016                 goto free_cma_dev;
5017         }
5018
5019         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5020                                             sizeof(*cma_dev->default_roce_tos),
5021                                             GFP_KERNEL);
5022         if (!cma_dev->default_roce_tos) {
5023                 ret = -ENOMEM;
5024                 goto free_gid_type;
5025         }
5026
5027         rdma_for_each_port (device, i) {
5028                 supported_gids = roce_gid_type_mask_support(device, i);
5029                 WARN_ON(!supported_gids);
5030                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5031                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5032                                 CMA_PREFERRED_ROCE_GID_TYPE;
5033                 else
5034                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5035                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5036                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5037         }
5038
5039         init_completion(&cma_dev->comp);
5040         refcount_set(&cma_dev->refcount, 1);
5041         INIT_LIST_HEAD(&cma_dev->id_list);
5042         ib_set_client_data(device, &cma_client, cma_dev);
5043
5044         mutex_lock(&lock);
5045         list_add_tail(&cma_dev->list, &dev_list);
5046         list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5047                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5048                 if (ret)
5049                         goto free_listen;
5050         }
5051         mutex_unlock(&lock);
5052
5053         trace_cm_add_one(device);
5054         return 0;
5055
5056 free_listen:
5057         list_del(&cma_dev->list);
5058         mutex_unlock(&lock);
5059
5060         /* cma_process_remove() will delete to_destroy */
5061         cma_process_remove(cma_dev);
5062         kfree(cma_dev->default_roce_tos);
5063 free_gid_type:
5064         kfree(cma_dev->default_gid_type);
5065
5066 free_cma_dev:
5067         kfree(cma_dev);
5068         return ret;
5069 }
5070
5071 static void cma_remove_one(struct ib_device *device, void *client_data)
5072 {
5073         struct cma_device *cma_dev = client_data;
5074
5075         trace_cm_remove_one(device);
5076
5077         mutex_lock(&lock);
5078         list_del(&cma_dev->list);
5079         mutex_unlock(&lock);
5080
5081         cma_process_remove(cma_dev);
5082         kfree(cma_dev->default_roce_tos);
5083         kfree(cma_dev->default_gid_type);
5084         kfree(cma_dev);
5085 }
5086
5087 static int cma_init_net(struct net *net)
5088 {
5089         struct cma_pernet *pernet = cma_pernet(net);
5090
5091         xa_init(&pernet->tcp_ps);
5092         xa_init(&pernet->udp_ps);
5093         xa_init(&pernet->ipoib_ps);
5094         xa_init(&pernet->ib_ps);
5095
5096         return 0;
5097 }
5098
5099 static void cma_exit_net(struct net *net)
5100 {
5101         struct cma_pernet *pernet = cma_pernet(net);
5102
5103         WARN_ON(!xa_empty(&pernet->tcp_ps));
5104         WARN_ON(!xa_empty(&pernet->udp_ps));
5105         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5106         WARN_ON(!xa_empty(&pernet->ib_ps));
5107 }
5108
5109 static struct pernet_operations cma_pernet_operations = {
5110         .init = cma_init_net,
5111         .exit = cma_exit_net,
5112         .id = &cma_pernet_id,
5113         .size = sizeof(struct cma_pernet),
5114 };
5115
5116 static int __init cma_init(void)
5117 {
5118         int ret;
5119
5120         /*
5121          * There is a rare lock ordering dependency in cma_netdev_callback()
5122          * that only happens when bonding is enabled. Teach lockdep that rtnl
5123          * must never be nested under lock so it can find these without having
5124          * to test with bonding.
5125          */
5126         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5127                 rtnl_lock();
5128                 mutex_lock(&lock);
5129                 mutex_unlock(&lock);
5130                 rtnl_unlock();
5131         }
5132
5133         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5134         if (!cma_wq)
5135                 return -ENOMEM;
5136
5137         ret = register_pernet_subsys(&cma_pernet_operations);
5138         if (ret)
5139                 goto err_wq;
5140
5141         ib_sa_register_client(&sa_client);
5142         register_netdevice_notifier(&cma_nb);
5143
5144         ret = ib_register_client(&cma_client);
5145         if (ret)
5146                 goto err;
5147
5148         ret = cma_configfs_init();
5149         if (ret)
5150                 goto err_ib;
5151
5152         return 0;
5153
5154 err_ib:
5155         ib_unregister_client(&cma_client);
5156 err:
5157         unregister_netdevice_notifier(&cma_nb);
5158         ib_sa_unregister_client(&sa_client);
5159         unregister_pernet_subsys(&cma_pernet_operations);
5160 err_wq:
5161         destroy_workqueue(cma_wq);
5162         return ret;
5163 }
5164
5165 static void __exit cma_cleanup(void)
5166 {
5167         cma_configfs_exit();
5168         ib_unregister_client(&cma_client);
5169         unregister_netdevice_notifier(&cma_nb);
5170         ib_sa_unregister_client(&sa_client);
5171         unregister_pernet_subsys(&cma_pernet_operations);
5172         destroy_workqueue(cma_wq);
5173 }
5174
5175 module_init(cma_init);
5176 module_exit(cma_cleanup);