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_set_mgid(struct rdma_id_private *id_priv, struct sockaddr *addr,
  71                          union ib_gid *mgid);
  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) {
1850                         union ib_gid mgid;
1851
1852                         cma_set_mgid(id_priv, (struct sockaddr *)&mc->addr,
1853                                      &mgid);
1854
1855                         if (!send_only)
1856                                 cma_igmp_send(ndev, &mgid, false);
1857
1858                         dev_put(ndev);
1859                 }
1860
1861                 cancel_work_sync(&mc->iboe_join.work);
1862         }
1863         kfree(mc);
1864 }
1865
1866 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1867 {
1868         struct cma_multicast *mc;
1869
1870         while (!list_empty(&id_priv->mc_list)) {
1871                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1872                                       list);
1873                 list_del(&mc->list);
1874                 destroy_mc(id_priv, mc);
1875         }
1876 }
1877
1878 static void _destroy_id(struct rdma_id_private *id_priv,
1879                         enum rdma_cm_state state)
1880 {
1881         cma_cancel_operation(id_priv, state);
1882
1883         rdma_restrack_del(&id_priv->res);
1884         if (id_priv->cma_dev) {
1885                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1886                         if (id_priv->cm_id.ib)
1887                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1888                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1889                         if (id_priv->cm_id.iw)
1890                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1891                 }
1892                 cma_leave_mc_groups(id_priv);
1893                 cma_release_dev(id_priv);
1894         }
1895
1896         cma_release_port(id_priv);
1897         cma_id_put(id_priv);
1898         wait_for_completion(&id_priv->comp);
1899
1900         if (id_priv->internal_id)
1901                 cma_id_put(id_priv->id.context);
1902
1903         kfree(id_priv->id.route.path_rec);
1904
1905         put_net(id_priv->id.route.addr.dev_addr.net);
1906         kfree(id_priv);
1907 }
1908
1909 /*
1910  * destroy an ID from within the handler_mutex. This ensures that no other
1911  * handlers can start running concurrently.
1912  */
1913 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1914         __releases(&idprv->handler_mutex)
1915 {
1916         enum rdma_cm_state state;
1917         unsigned long flags;
1918
1919         trace_cm_id_destroy(id_priv);
1920
1921         /*
1922          * Setting the state to destroyed under the handler mutex provides a
1923          * fence against calling handler callbacks. If this is invoked due to
1924          * the failure of a handler callback then it guarentees that no future
1925          * handlers will be called.
1926          */
1927         lockdep_assert_held(&id_priv->handler_mutex);
1928         spin_lock_irqsave(&id_priv->lock, flags);
1929         state = id_priv->state;
1930         id_priv->state = RDMA_CM_DESTROYING;
1931         spin_unlock_irqrestore(&id_priv->lock, flags);
1932         mutex_unlock(&id_priv->handler_mutex);
1933         _destroy_id(id_priv, state);
1934 }
1935
1936 void rdma_destroy_id(struct rdma_cm_id *id)
1937 {
1938         struct rdma_id_private *id_priv =
1939                 container_of(id, struct rdma_id_private, id);
1940
1941         mutex_lock(&id_priv->handler_mutex);
1942         destroy_id_handler_unlock(id_priv);
1943 }
1944 EXPORT_SYMBOL(rdma_destroy_id);
1945
1946 static int cma_rep_recv(struct rdma_id_private *id_priv)
1947 {
1948         int ret;
1949
1950         ret = cma_modify_qp_rtr(id_priv, NULL);
1951         if (ret)
1952                 goto reject;
1953
1954         ret = cma_modify_qp_rts(id_priv, NULL);
1955         if (ret)
1956                 goto reject;
1957
1958         trace_cm_send_rtu(id_priv);
1959         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1960         if (ret)
1961                 goto reject;
1962
1963         return 0;
1964 reject:
1965         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1966         cma_modify_qp_err(id_priv);
1967         trace_cm_send_rej(id_priv);
1968         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1969                        NULL, 0, NULL, 0);
1970         return ret;
1971 }
1972
1973 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1974                                    const struct ib_cm_rep_event_param *rep_data,
1975                                    void *private_data)
1976 {
1977         event->param.conn.private_data = private_data;
1978         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1979         event->param.conn.responder_resources = rep_data->responder_resources;
1980         event->param.conn.initiator_depth = rep_data->initiator_depth;
1981         event->param.conn.flow_control = rep_data->flow_control;
1982         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1983         event->param.conn.srq = rep_data->srq;
1984         event->param.conn.qp_num = rep_data->remote_qpn;
1985
1986         event->ece.vendor_id = rep_data->ece.vendor_id;
1987         event->ece.attr_mod = rep_data->ece.attr_mod;
1988 }
1989
1990 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1991                                 struct rdma_cm_event *event)
1992 {
1993         int ret;
1994
1995         lockdep_assert_held(&id_priv->handler_mutex);
1996
1997         trace_cm_event_handler(id_priv, event);
1998         ret = id_priv->id.event_handler(&id_priv->id, event);
1999         trace_cm_event_done(id_priv, event, ret);
2000         return ret;
2001 }
2002
2003 static int cma_ib_handler(struct ib_cm_id *cm_id,
2004                           const struct ib_cm_event *ib_event)
2005 {
2006         struct rdma_id_private *id_priv = cm_id->context;
2007         struct rdma_cm_event event = {};
2008         enum rdma_cm_state state;
2009         int ret;
2010
2011         mutex_lock(&id_priv->handler_mutex);
2012         state = READ_ONCE(id_priv->state);
2013         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2014              state != RDMA_CM_CONNECT) ||
2015             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2016              state != RDMA_CM_DISCONNECT))
2017                 goto out;
2018
2019         switch (ib_event->event) {
2020         case IB_CM_REQ_ERROR:
2021         case IB_CM_REP_ERROR:
2022                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2023                 event.status = -ETIMEDOUT;
2024                 break;
2025         case IB_CM_REP_RECEIVED:
2026                 if (state == RDMA_CM_CONNECT &&
2027                     (id_priv->id.qp_type != IB_QPT_UD)) {
2028                         trace_cm_send_mra(id_priv);
2029                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2030                 }
2031                 if (id_priv->id.qp) {
2032                         event.status = cma_rep_recv(id_priv);
2033                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2034                                                      RDMA_CM_EVENT_ESTABLISHED;
2035                 } else {
2036                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2037                 }
2038                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2039                                        ib_event->private_data);
2040                 break;
2041         case IB_CM_RTU_RECEIVED:
2042         case IB_CM_USER_ESTABLISHED:
2043                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2044                 break;
2045         case IB_CM_DREQ_ERROR:
2046                 event.status = -ETIMEDOUT;
2047                 fallthrough;
2048         case IB_CM_DREQ_RECEIVED:
2049         case IB_CM_DREP_RECEIVED:
2050                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2051                                    RDMA_CM_DISCONNECT))
2052                         goto out;
2053                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2054                 break;
2055         case IB_CM_TIMEWAIT_EXIT:
2056                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2057                 break;
2058         case IB_CM_MRA_RECEIVED:
2059                 /* ignore event */
2060                 goto out;
2061         case IB_CM_REJ_RECEIVED:
2062                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2063                                                                                 ib_event->param.rej_rcvd.reason));
2064                 cma_modify_qp_err(id_priv);
2065                 event.status = ib_event->param.rej_rcvd.reason;
2066                 event.event = RDMA_CM_EVENT_REJECTED;
2067                 event.param.conn.private_data = ib_event->private_data;
2068                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2069                 break;
2070         default:
2071                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2072                        ib_event->event);
2073                 goto out;
2074         }
2075
2076         ret = cma_cm_event_handler(id_priv, &event);
2077         if (ret) {
2078                 /* Destroy the CM ID by returning a non-zero value. */
2079                 id_priv->cm_id.ib = NULL;
2080                 destroy_id_handler_unlock(id_priv);
2081                 return ret;
2082         }
2083 out:
2084         mutex_unlock(&id_priv->handler_mutex);
2085         return 0;
2086 }
2087
2088 static struct rdma_id_private *
2089 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2090                    const struct ib_cm_event *ib_event,
2091                    struct net_device *net_dev)
2092 {
2093         struct rdma_id_private *listen_id_priv;
2094         struct rdma_id_private *id_priv;
2095         struct rdma_cm_id *id;
2096         struct rdma_route *rt;
2097         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2098         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2099         const __be64 service_id =
2100                 ib_event->param.req_rcvd.primary_path->service_id;
2101         int ret;
2102
2103         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2104         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2105                                    listen_id->event_handler, listen_id->context,
2106                                    listen_id->ps,
2107                                    ib_event->param.req_rcvd.qp_type,
2108                                    listen_id_priv);
2109         if (IS_ERR(id_priv))
2110                 return NULL;
2111
2112         id = &id_priv->id;
2113         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2114                               (struct sockaddr *)&id->route.addr.dst_addr,
2115                               listen_id, ib_event, ss_family, service_id))
2116                 goto err;
2117
2118         rt = &id->route;
2119         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2120         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2121                                      GFP_KERNEL);
2122         if (!rt->path_rec)
2123                 goto err;
2124
2125         rt->path_rec[0] = *path;
2126         if (rt->num_paths == 2)
2127                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2128
2129         if (net_dev) {
2130                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2131         } else {
2132                 if (!cma_protocol_roce(listen_id) &&
2133                     cma_any_addr(cma_src_addr(id_priv))) {
2134                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2135                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2136                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2137                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2138                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2139                         if (ret)
2140                                 goto err;
2141                 }
2142         }
2143         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2144
2145         id_priv->state = RDMA_CM_CONNECT;
2146         return id_priv;
2147
2148 err:
2149         rdma_destroy_id(id);
2150         return NULL;
2151 }
2152
2153 static struct rdma_id_private *
2154 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2155                   const struct ib_cm_event *ib_event,
2156                   struct net_device *net_dev)
2157 {
2158         const struct rdma_id_private *listen_id_priv;
2159         struct rdma_id_private *id_priv;
2160         struct rdma_cm_id *id;
2161         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2162         struct net *net = listen_id->route.addr.dev_addr.net;
2163         int ret;
2164
2165         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2166         id_priv = __rdma_create_id(net, listen_id->event_handler,
2167                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2168                                    listen_id_priv);
2169         if (IS_ERR(id_priv))
2170                 return NULL;
2171
2172         id = &id_priv->id;
2173         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2174                               (struct sockaddr *)&id->route.addr.dst_addr,
2175                               listen_id, ib_event, ss_family,
2176                               ib_event->param.sidr_req_rcvd.service_id))
2177                 goto err;
2178
2179         if (net_dev) {
2180                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2181         } else {
2182                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2183                         ret = cma_translate_addr(cma_src_addr(id_priv),
2184                                                  &id->route.addr.dev_addr);
2185                         if (ret)
2186                                 goto err;
2187                 }
2188         }
2189
2190         id_priv->state = RDMA_CM_CONNECT;
2191         return id_priv;
2192 err:
2193         rdma_destroy_id(id);
2194         return NULL;
2195 }
2196
2197 static void cma_set_req_event_data(struct rdma_cm_event *event,
2198                                    const struct ib_cm_req_event_param *req_data,
2199                                    void *private_data, int offset)
2200 {
2201         event->param.conn.private_data = private_data + offset;
2202         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2203         event->param.conn.responder_resources = req_data->responder_resources;
2204         event->param.conn.initiator_depth = req_data->initiator_depth;
2205         event->param.conn.flow_control = req_data->flow_control;
2206         event->param.conn.retry_count = req_data->retry_count;
2207         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2208         event->param.conn.srq = req_data->srq;
2209         event->param.conn.qp_num = req_data->remote_qpn;
2210
2211         event->ece.vendor_id = req_data->ece.vendor_id;
2212         event->ece.attr_mod = req_data->ece.attr_mod;
2213 }
2214
2215 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2216                                     const struct ib_cm_event *ib_event)
2217 {
2218         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2219                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2220                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2221                  (id->qp_type == IB_QPT_UD)) ||
2222                 (!id->qp_type));
2223 }
2224
2225 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2226                               const struct ib_cm_event *ib_event)
2227 {
2228         struct rdma_id_private *listen_id, *conn_id = NULL;
2229         struct rdma_cm_event event = {};
2230         struct cma_req_info req = {};
2231         struct net_device *net_dev;
2232         u8 offset;
2233         int ret;
2234
2235         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2236         if (IS_ERR(listen_id))
2237                 return PTR_ERR(listen_id);
2238
2239         trace_cm_req_handler(listen_id, ib_event->event);
2240         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2241                 ret = -EINVAL;
2242                 goto net_dev_put;
2243         }
2244
2245         mutex_lock(&listen_id->handler_mutex);
2246         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2247                 ret = -ECONNABORTED;
2248                 goto err_unlock;
2249         }
2250
2251         offset = cma_user_data_offset(listen_id);
2252         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2253         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2254                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2255                 event.param.ud.private_data = ib_event->private_data + offset;
2256                 event.param.ud.private_data_len =
2257                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2258         } else {
2259                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2260                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2261                                        ib_event->private_data, offset);
2262         }
2263         if (!conn_id) {
2264                 ret = -ENOMEM;
2265                 goto err_unlock;
2266         }
2267
2268         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2269         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2270         if (ret) {
2271                 destroy_id_handler_unlock(conn_id);
2272                 goto err_unlock;
2273         }
2274
2275         conn_id->cm_id.ib = cm_id;
2276         cm_id->context = conn_id;
2277         cm_id->cm_handler = cma_ib_handler;
2278
2279         ret = cma_cm_event_handler(conn_id, &event);
2280         if (ret) {
2281                 /* Destroy the CM ID by returning a non-zero value. */
2282                 conn_id->cm_id.ib = NULL;
2283                 mutex_unlock(&listen_id->handler_mutex);
2284                 destroy_id_handler_unlock(conn_id);
2285                 goto net_dev_put;
2286         }
2287
2288         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2289             conn_id->id.qp_type != IB_QPT_UD) {
2290                 trace_cm_send_mra(cm_id->context);
2291                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2292         }
2293         mutex_unlock(&conn_id->handler_mutex);
2294
2295 err_unlock:
2296         mutex_unlock(&listen_id->handler_mutex);
2297
2298 net_dev_put:
2299         if (net_dev)
2300                 dev_put(net_dev);
2301
2302         return ret;
2303 }
2304
2305 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2306 {
2307         if (addr->sa_family == AF_IB)
2308                 return ((struct sockaddr_ib *) addr)->sib_sid;
2309
2310         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2311 }
2312 EXPORT_SYMBOL(rdma_get_service_id);
2313
2314 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2315                     union ib_gid *dgid)
2316 {
2317         struct rdma_addr *addr = &cm_id->route.addr;
2318
2319         if (!cm_id->device) {
2320                 if (sgid)
2321                         memset(sgid, 0, sizeof(*sgid));
2322                 if (dgid)
2323                         memset(dgid, 0, sizeof(*dgid));
2324                 return;
2325         }
2326
2327         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2328                 if (sgid)
2329                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2330                 if (dgid)
2331                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2332         } else {
2333                 if (sgid)
2334                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2335                 if (dgid)
2336                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2337         }
2338 }
2339 EXPORT_SYMBOL(rdma_read_gids);
2340
2341 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2342 {
2343         struct rdma_id_private *id_priv = iw_id->context;
2344         struct rdma_cm_event event = {};
2345         int ret = 0;
2346         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2347         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2348
2349         mutex_lock(&id_priv->handler_mutex);
2350         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2351                 goto out;
2352
2353         switch (iw_event->event) {
2354         case IW_CM_EVENT_CLOSE:
2355                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2356                 break;
2357         case IW_CM_EVENT_CONNECT_REPLY:
2358                 memcpy(cma_src_addr(id_priv), laddr,
2359                        rdma_addr_size(laddr));
2360                 memcpy(cma_dst_addr(id_priv), raddr,
2361                        rdma_addr_size(raddr));
2362                 switch (iw_event->status) {
2363                 case 0:
2364                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2365                         event.param.conn.initiator_depth = iw_event->ird;
2366                         event.param.conn.responder_resources = iw_event->ord;
2367                         break;
2368                 case -ECONNRESET:
2369                 case -ECONNREFUSED:
2370                         event.event = RDMA_CM_EVENT_REJECTED;
2371                         break;
2372                 case -ETIMEDOUT:
2373                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2374                         break;
2375                 default:
2376                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2377                         break;
2378                 }
2379                 break;
2380         case IW_CM_EVENT_ESTABLISHED:
2381                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2382                 event.param.conn.initiator_depth = iw_event->ird;
2383                 event.param.conn.responder_resources = iw_event->ord;
2384                 break;
2385         default:
2386                 goto out;
2387         }
2388
2389         event.status = iw_event->status;
2390         event.param.conn.private_data = iw_event->private_data;
2391         event.param.conn.private_data_len = iw_event->private_data_len;
2392         ret = cma_cm_event_handler(id_priv, &event);
2393         if (ret) {
2394                 /* Destroy the CM ID by returning a non-zero value. */
2395                 id_priv->cm_id.iw = NULL;
2396                 destroy_id_handler_unlock(id_priv);
2397                 return ret;
2398         }
2399
2400 out:
2401         mutex_unlock(&id_priv->handler_mutex);
2402         return ret;
2403 }
2404
2405 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2406                                struct iw_cm_event *iw_event)
2407 {
2408         struct rdma_id_private *listen_id, *conn_id;
2409         struct rdma_cm_event event = {};
2410         int ret = -ECONNABORTED;
2411         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2412         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2413
2414         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2415         event.param.conn.private_data = iw_event->private_data;
2416         event.param.conn.private_data_len = iw_event->private_data_len;
2417         event.param.conn.initiator_depth = iw_event->ird;
2418         event.param.conn.responder_resources = iw_event->ord;
2419
2420         listen_id = cm_id->context;
2421
2422         mutex_lock(&listen_id->handler_mutex);
2423         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2424                 goto out;
2425
2426         /* Create a new RDMA id for the new IW CM ID */
2427         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2428                                    listen_id->id.event_handler,
2429                                    listen_id->id.context, RDMA_PS_TCP,
2430                                    IB_QPT_RC, listen_id);
2431         if (IS_ERR(conn_id)) {
2432                 ret = -ENOMEM;
2433                 goto out;
2434         }
2435         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2436         conn_id->state = RDMA_CM_CONNECT;
2437
2438         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2439         if (ret) {
2440                 mutex_unlock(&listen_id->handler_mutex);
2441                 destroy_id_handler_unlock(conn_id);
2442                 return ret;
2443         }
2444
2445         ret = cma_iw_acquire_dev(conn_id, listen_id);
2446         if (ret) {
2447                 mutex_unlock(&listen_id->handler_mutex);
2448                 destroy_id_handler_unlock(conn_id);
2449                 return ret;
2450         }
2451
2452         conn_id->cm_id.iw = cm_id;
2453         cm_id->context = conn_id;
2454         cm_id->cm_handler = cma_iw_handler;
2455
2456         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2457         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2458
2459         ret = cma_cm_event_handler(conn_id, &event);
2460         if (ret) {
2461                 /* User wants to destroy the CM ID */
2462                 conn_id->cm_id.iw = NULL;
2463                 mutex_unlock(&listen_id->handler_mutex);
2464                 destroy_id_handler_unlock(conn_id);
2465                 return ret;
2466         }
2467
2468         mutex_unlock(&conn_id->handler_mutex);
2469
2470 out:
2471         mutex_unlock(&listen_id->handler_mutex);
2472         return ret;
2473 }
2474
2475 static int cma_ib_listen(struct rdma_id_private *id_priv)
2476 {
2477         struct sockaddr *addr;
2478         struct ib_cm_id *id;
2479         __be64 svc_id;
2480
2481         addr = cma_src_addr(id_priv);
2482         svc_id = rdma_get_service_id(&id_priv->id, addr);
2483         id = ib_cm_insert_listen(id_priv->id.device,
2484                                  cma_ib_req_handler, svc_id);
2485         if (IS_ERR(id))
2486                 return PTR_ERR(id);
2487         id_priv->cm_id.ib = id;
2488
2489         return 0;
2490 }
2491
2492 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2493 {
2494         int ret;
2495         struct iw_cm_id *id;
2496
2497         id = iw_create_cm_id(id_priv->id.device,
2498                              iw_conn_req_handler,
2499                              id_priv);
2500         if (IS_ERR(id))
2501                 return PTR_ERR(id);
2502
2503         mutex_lock(&id_priv->qp_mutex);
2504         id->tos = id_priv->tos;
2505         id->tos_set = id_priv->tos_set;
2506         mutex_unlock(&id_priv->qp_mutex);
2507         id->afonly = id_priv->afonly;
2508         id_priv->cm_id.iw = id;
2509
2510         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2511                rdma_addr_size(cma_src_addr(id_priv)));
2512
2513         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2514
2515         if (ret) {
2516                 iw_destroy_cm_id(id_priv->cm_id.iw);
2517                 id_priv->cm_id.iw = NULL;
2518         }
2519
2520         return ret;
2521 }
2522
2523 static int cma_listen_handler(struct rdma_cm_id *id,
2524                               struct rdma_cm_event *event)
2525 {
2526         struct rdma_id_private *id_priv = id->context;
2527
2528         /* Listening IDs are always destroyed on removal */
2529         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2530                 return -1;
2531
2532         id->context = id_priv->id.context;
2533         id->event_handler = id_priv->id.event_handler;
2534         trace_cm_event_handler(id_priv, event);
2535         return id_priv->id.event_handler(id, event);
2536 }
2537
2538 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2539                              struct cma_device *cma_dev,
2540                              struct rdma_id_private **to_destroy)
2541 {
2542         struct rdma_id_private *dev_id_priv;
2543         struct net *net = id_priv->id.route.addr.dev_addr.net;
2544         int ret;
2545
2546         lockdep_assert_held(&lock);
2547
2548         *to_destroy = NULL;
2549         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2550                 return 0;
2551
2552         dev_id_priv =
2553                 __rdma_create_id(net, cma_listen_handler, id_priv,
2554                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2555         if (IS_ERR(dev_id_priv))
2556                 return PTR_ERR(dev_id_priv);
2557
2558         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2559         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2560                rdma_addr_size(cma_src_addr(id_priv)));
2561
2562         _cma_attach_to_dev(dev_id_priv, cma_dev);
2563         rdma_restrack_add(&dev_id_priv->res);
2564         cma_id_get(id_priv);
2565         dev_id_priv->internal_id = 1;
2566         dev_id_priv->afonly = id_priv->afonly;
2567         mutex_lock(&id_priv->qp_mutex);
2568         dev_id_priv->tos_set = id_priv->tos_set;
2569         dev_id_priv->tos = id_priv->tos;
2570         mutex_unlock(&id_priv->qp_mutex);
2571
2572         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2573         if (ret)
2574                 goto err_listen;
2575         list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2576         return 0;
2577 err_listen:
2578         /* Caller must destroy this after releasing lock */
2579         *to_destroy = dev_id_priv;
2580         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2581         return ret;
2582 }
2583
2584 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2585 {
2586         struct rdma_id_private *to_destroy;
2587         struct cma_device *cma_dev;
2588         int ret;
2589
2590         mutex_lock(&lock);
2591         list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2592         list_for_each_entry(cma_dev, &dev_list, list) {
2593                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2594                 if (ret) {
2595                         /* Prevent racing with cma_process_remove() */
2596                         if (to_destroy)
2597                                 list_del_init(&to_destroy->device_item);
2598                         goto err_listen;
2599                 }
2600         }
2601         mutex_unlock(&lock);
2602         return 0;
2603
2604 err_listen:
2605         _cma_cancel_listens(id_priv);
2606         mutex_unlock(&lock);
2607         if (to_destroy)
2608                 rdma_destroy_id(&to_destroy->id);
2609         return ret;
2610 }
2611
2612 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2613 {
2614         struct rdma_id_private *id_priv;
2615
2616         id_priv = container_of(id, struct rdma_id_private, id);
2617         mutex_lock(&id_priv->qp_mutex);
2618         id_priv->tos = (u8) tos;
2619         id_priv->tos_set = true;
2620         mutex_unlock(&id_priv->qp_mutex);
2621 }
2622 EXPORT_SYMBOL(rdma_set_service_type);
2623
2624 /**
2625  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2626  *                          with a connection identifier.
2627  * @id: Communication identifier to associated with service type.
2628  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2629  *
2630  * This function should be called before rdma_connect() on active side,
2631  * and on passive side before rdma_accept(). It is applicable to primary
2632  * path only. The timeout will affect the local side of the QP, it is not
2633  * negotiated with remote side and zero disables the timer. In case it is
2634  * set before rdma_resolve_route, the value will also be used to determine
2635  * PacketLifeTime for RoCE.
2636  *
2637  * Return: 0 for success
2638  */
2639 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2640 {
2641         struct rdma_id_private *id_priv;
2642
2643         if (id->qp_type != IB_QPT_RC)
2644                 return -EINVAL;
2645
2646         id_priv = container_of(id, struct rdma_id_private, id);
2647         mutex_lock(&id_priv->qp_mutex);
2648         id_priv->timeout = timeout;
2649         id_priv->timeout_set = true;
2650         mutex_unlock(&id_priv->qp_mutex);
2651
2652         return 0;
2653 }
2654 EXPORT_SYMBOL(rdma_set_ack_timeout);
2655
2656 /**
2657  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2658  *                            QP associated with a connection identifier.
2659  * @id: Communication identifier to associated with service type.
2660  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2661  *                 Timer Field" in the IBTA specification.
2662  *
2663  * This function should be called before rdma_connect() on active
2664  * side, and on passive side before rdma_accept(). The timer value
2665  * will be associated with the local QP. When it receives a send it is
2666  * not read to handle, typically if the receive queue is empty, an RNR
2667  * Retry NAK is returned to the requester with the min_rnr_timer
2668  * encoded. The requester will then wait at least the time specified
2669  * in the NAK before retrying. The default is zero, which translates
2670  * to a minimum RNR Timer value of 655 ms.
2671  *
2672  * Return: 0 for success
2673  */
2674 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2675 {
2676         struct rdma_id_private *id_priv;
2677
2678         /* It is a five-bit value */
2679         if (min_rnr_timer & 0xe0)
2680                 return -EINVAL;
2681
2682         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2683                 return -EINVAL;
2684
2685         id_priv = container_of(id, struct rdma_id_private, id);
2686         mutex_lock(&id_priv->qp_mutex);
2687         id_priv->min_rnr_timer = min_rnr_timer;
2688         id_priv->min_rnr_timer_set = true;
2689         mutex_unlock(&id_priv->qp_mutex);
2690
2691         return 0;
2692 }
2693 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2694
2695 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2696                               void *context)
2697 {
2698         struct cma_work *work = context;
2699         struct rdma_route *route;
2700
2701         route = &work->id->id.route;
2702
2703         if (!status) {
2704                 route->num_paths = 1;
2705                 *route->path_rec = *path_rec;
2706         } else {
2707                 work->old_state = RDMA_CM_ROUTE_QUERY;
2708                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2709                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2710                 work->event.status = status;
2711                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2712                                      status);
2713         }
2714
2715         queue_work(cma_wq, &work->work);
2716 }
2717
2718 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2719                               unsigned long timeout_ms, struct cma_work *work)
2720 {
2721         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2722         struct sa_path_rec path_rec;
2723         ib_sa_comp_mask comp_mask;
2724         struct sockaddr_in6 *sin6;
2725         struct sockaddr_ib *sib;
2726
2727         memset(&path_rec, 0, sizeof path_rec);
2728
2729         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2730                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2731         else
2732                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2733         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2734         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2735         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2736         path_rec.numb_path = 1;
2737         path_rec.reversible = 1;
2738         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2739                                                   cma_dst_addr(id_priv));
2740
2741         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2742                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2743                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2744
2745         switch (cma_family(id_priv)) {
2746         case AF_INET:
2747                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2748                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2749                 break;
2750         case AF_INET6:
2751                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2752                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2753                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2754                 break;
2755         case AF_IB:
2756                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2757                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2758                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2759                 break;
2760         }
2761
2762         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2763                                                id_priv->id.port_num, &path_rec,
2764                                                comp_mask, timeout_ms,
2765                                                GFP_KERNEL, cma_query_handler,
2766                                                work, &id_priv->query);
2767
2768         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2769 }
2770
2771 static void cma_iboe_join_work_handler(struct work_struct *work)
2772 {
2773         struct cma_multicast *mc =
2774                 container_of(work, struct cma_multicast, iboe_join.work);
2775         struct rdma_cm_event *event = &mc->iboe_join.event;
2776         struct rdma_id_private *id_priv = mc->id_priv;
2777         int ret;
2778
2779         mutex_lock(&id_priv->handler_mutex);
2780         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2781             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2782                 goto out_unlock;
2783
2784         ret = cma_cm_event_handler(id_priv, event);
2785         WARN_ON(ret);
2786
2787 out_unlock:
2788         mutex_unlock(&id_priv->handler_mutex);
2789         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2790                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2791 }
2792
2793 static void cma_work_handler(struct work_struct *_work)
2794 {
2795         struct cma_work *work = container_of(_work, struct cma_work, work);
2796         struct rdma_id_private *id_priv = work->id;
2797
2798         mutex_lock(&id_priv->handler_mutex);
2799         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2800             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2801                 goto out_unlock;
2802         if (work->old_state != 0 || work->new_state != 0) {
2803                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2804                         goto out_unlock;
2805         }
2806
2807         if (cma_cm_event_handler(id_priv, &work->event)) {
2808                 cma_id_put(id_priv);
2809                 destroy_id_handler_unlock(id_priv);
2810                 goto out_free;
2811         }
2812
2813 out_unlock:
2814         mutex_unlock(&id_priv->handler_mutex);
2815         cma_id_put(id_priv);
2816 out_free:
2817         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2818                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2819         kfree(work);
2820 }
2821
2822 static void cma_init_resolve_route_work(struct cma_work *work,
2823                                         struct rdma_id_private *id_priv)
2824 {
2825         work->id = id_priv;
2826         INIT_WORK(&work->work, cma_work_handler);
2827         work->old_state = RDMA_CM_ROUTE_QUERY;
2828         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2829         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2830 }
2831
2832 static void enqueue_resolve_addr_work(struct cma_work *work,
2833                                       struct rdma_id_private *id_priv)
2834 {
2835         /* Balances with cma_id_put() in cma_work_handler */
2836         cma_id_get(id_priv);
2837
2838         work->id = id_priv;
2839         INIT_WORK(&work->work, cma_work_handler);
2840         work->old_state = RDMA_CM_ADDR_QUERY;
2841         work->new_state = RDMA_CM_ADDR_RESOLVED;
2842         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2843
2844         queue_work(cma_wq, &work->work);
2845 }
2846
2847 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2848                                 unsigned long timeout_ms)
2849 {
2850         struct rdma_route *route = &id_priv->id.route;
2851         struct cma_work *work;
2852         int ret;
2853
2854         work = kzalloc(sizeof *work, GFP_KERNEL);
2855         if (!work)
2856                 return -ENOMEM;
2857
2858         cma_init_resolve_route_work(work, id_priv);
2859
2860         if (!route->path_rec)
2861                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2862         if (!route->path_rec) {
2863                 ret = -ENOMEM;
2864                 goto err1;
2865         }
2866
2867         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2868         if (ret)
2869                 goto err2;
2870
2871         return 0;
2872 err2:
2873         kfree(route->path_rec);
2874         route->path_rec = NULL;
2875 err1:
2876         kfree(work);
2877         return ret;
2878 }
2879
2880 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2881                                            unsigned long supported_gids,
2882                                            enum ib_gid_type default_gid)
2883 {
2884         if ((network_type == RDMA_NETWORK_IPV4 ||
2885              network_type == RDMA_NETWORK_IPV6) &&
2886             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2887                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2888
2889         return default_gid;
2890 }
2891
2892 /*
2893  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2894  * path record type based on GID type.
2895  * It also sets up other L2 fields which includes destination mac address
2896  * netdev ifindex, of the path record.
2897  * It returns the netdev of the bound interface for this path record entry.
2898  */
2899 static struct net_device *
2900 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2901 {
2902         struct rdma_route *route = &id_priv->id.route;
2903         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2904         struct rdma_addr *addr = &route->addr;
2905         unsigned long supported_gids;
2906         struct net_device *ndev;
2907
2908         if (!addr->dev_addr.bound_dev_if)
2909                 return NULL;
2910
2911         ndev = dev_get_by_index(addr->dev_addr.net,
2912                                 addr->dev_addr.bound_dev_if);
2913         if (!ndev)
2914                 return NULL;
2915
2916         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2917                                                     id_priv->id.port_num);
2918         gid_type = cma_route_gid_type(addr->dev_addr.network,
2919                                       supported_gids,
2920                                       id_priv->gid_type);
2921         /* Use the hint from IP Stack to select GID Type */
2922         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2923                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2924         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2925
2926         route->path_rec->roce.route_resolved = true;
2927         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2928         return ndev;
2929 }
2930
2931 int rdma_set_ib_path(struct rdma_cm_id *id,
2932                      struct sa_path_rec *path_rec)
2933 {
2934         struct rdma_id_private *id_priv;
2935         struct net_device *ndev;
2936         int ret;
2937
2938         id_priv = container_of(id, struct rdma_id_private, id);
2939         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2940                            RDMA_CM_ROUTE_RESOLVED))
2941                 return -EINVAL;
2942
2943         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2944                                      GFP_KERNEL);
2945         if (!id->route.path_rec) {
2946                 ret = -ENOMEM;
2947                 goto err;
2948         }
2949
2950         if (rdma_protocol_roce(id->device, id->port_num)) {
2951                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2952                 if (!ndev) {
2953                         ret = -ENODEV;
2954                         goto err_free;
2955                 }
2956                 dev_put(ndev);
2957         }
2958
2959         id->route.num_paths = 1;
2960         return 0;
2961
2962 err_free:
2963         kfree(id->route.path_rec);
2964         id->route.path_rec = NULL;
2965 err:
2966         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2967         return ret;
2968 }
2969 EXPORT_SYMBOL(rdma_set_ib_path);
2970
2971 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2972 {
2973         struct cma_work *work;
2974
2975         work = kzalloc(sizeof *work, GFP_KERNEL);
2976         if (!work)
2977                 return -ENOMEM;
2978
2979         cma_init_resolve_route_work(work, id_priv);
2980         queue_work(cma_wq, &work->work);
2981         return 0;
2982 }
2983
2984 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2985 {
2986         struct net_device *dev;
2987
2988         dev = vlan_dev_real_dev(vlan_ndev);
2989         if (dev->num_tc)
2990                 return netdev_get_prio_tc_map(dev, prio);
2991
2992         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2993                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2994 }
2995
2996 struct iboe_prio_tc_map {
2997         int input_prio;
2998         int output_tc;
2999         bool found;
3000 };
3001
3002 static int get_lower_vlan_dev_tc(struct net_device *dev,
3003                                  struct netdev_nested_priv *priv)
3004 {
3005         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3006
3007         if (is_vlan_dev(dev))
3008                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3009         else if (dev->num_tc)
3010                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3011         else
3012                 map->output_tc = 0;
3013         /* We are interested only in first level VLAN device, so always
3014          * return 1 to stop iterating over next level devices.
3015          */
3016         map->found = true;
3017         return 1;
3018 }
3019
3020 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3021 {
3022         struct iboe_prio_tc_map prio_tc_map = {};
3023         int prio = rt_tos2priority(tos);
3024         struct netdev_nested_priv priv;
3025
3026         /* If VLAN device, get it directly from the VLAN netdev */
3027         if (is_vlan_dev(ndev))
3028                 return get_vlan_ndev_tc(ndev, prio);
3029
3030         prio_tc_map.input_prio = prio;
3031         priv.data = (void *)&prio_tc_map;
3032         rcu_read_lock();
3033         netdev_walk_all_lower_dev_rcu(ndev,
3034                                       get_lower_vlan_dev_tc,
3035                                       &priv);
3036         rcu_read_unlock();
3037         /* If map is found from lower device, use it; Otherwise
3038          * continue with the current netdevice to get priority to tc map.
3039          */
3040         if (prio_tc_map.found)
3041                 return prio_tc_map.output_tc;
3042         else if (ndev->num_tc)
3043                 return netdev_get_prio_tc_map(ndev, prio);
3044         else
3045                 return 0;
3046 }
3047
3048 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3049 {
3050         struct sockaddr_in6 *addr6;
3051         u16 dport, sport;
3052         u32 hash, fl;
3053
3054         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3055         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3056         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3057                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3058                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3059                 hash = (u32)sport * 31 + dport;
3060                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3061         }
3062
3063         return cpu_to_be32(fl);
3064 }
3065
3066 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3067 {
3068         struct rdma_route *route = &id_priv->id.route;
3069         struct rdma_addr *addr = &route->addr;
3070         struct cma_work *work;
3071         int ret;
3072         struct net_device *ndev;
3073
3074         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3075                                         rdma_start_port(id_priv->cma_dev->device)];
3076         u8 tos;
3077
3078         mutex_lock(&id_priv->qp_mutex);
3079         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3080         mutex_unlock(&id_priv->qp_mutex);
3081
3082         work = kzalloc(sizeof *work, GFP_KERNEL);
3083         if (!work)
3084                 return -ENOMEM;
3085
3086         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3087         if (!route->path_rec) {
3088                 ret = -ENOMEM;
3089                 goto err1;
3090         }
3091
3092         route->num_paths = 1;
3093
3094         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3095         if (!ndev) {
3096                 ret = -ENODEV;
3097                 goto err2;
3098         }
3099
3100         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3101                     &route->path_rec->sgid);
3102         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3103                     &route->path_rec->dgid);
3104
3105         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3106                 /* TODO: get the hoplimit from the inet/inet6 device */
3107                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3108         else
3109                 route->path_rec->hop_limit = 1;
3110         route->path_rec->reversible = 1;
3111         route->path_rec->pkey = cpu_to_be16(0xffff);
3112         route->path_rec->mtu_selector = IB_SA_EQ;
3113         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3114         route->path_rec->traffic_class = tos;
3115         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3116         route->path_rec->rate_selector = IB_SA_EQ;
3117         route->path_rec->rate = iboe_get_rate(ndev);
3118         dev_put(ndev);
3119         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3120         /* In case ACK timeout is set, use this value to calculate
3121          * PacketLifeTime.  As per IBTA 12.7.34,
3122          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3123          * Assuming a negligible local ACK delay, we can use
3124          * PacketLifeTime = local ACK timeout/2
3125          * as a reasonable approximation for RoCE networks.
3126          */
3127         mutex_lock(&id_priv->qp_mutex);
3128         if (id_priv->timeout_set && id_priv->timeout)
3129                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3130         else
3131                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3132         mutex_unlock(&id_priv->qp_mutex);
3133
3134         if (!route->path_rec->mtu) {
3135                 ret = -EINVAL;
3136                 goto err2;
3137         }
3138
3139         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3140                                          id_priv->id.port_num))
3141                 route->path_rec->flow_label =
3142                         cma_get_roce_udp_flow_label(id_priv);
3143
3144         cma_init_resolve_route_work(work, id_priv);
3145         queue_work(cma_wq, &work->work);
3146
3147         return 0;
3148
3149 err2:
3150         kfree(route->path_rec);
3151         route->path_rec = NULL;
3152         route->num_paths = 0;
3153 err1:
3154         kfree(work);
3155         return ret;
3156 }
3157
3158 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3159 {
3160         struct rdma_id_private *id_priv;
3161         int ret;
3162
3163         if (!timeout_ms)
3164                 return -EINVAL;
3165
3166         id_priv = container_of(id, struct rdma_id_private, id);
3167         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3168                 return -EINVAL;
3169
3170         cma_id_get(id_priv);
3171         if (rdma_cap_ib_sa(id->device, id->port_num))
3172                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3173         else if (rdma_protocol_roce(id->device, id->port_num))
3174                 ret = cma_resolve_iboe_route(id_priv);
3175         else if (rdma_protocol_iwarp(id->device, id->port_num))
3176                 ret = cma_resolve_iw_route(id_priv);
3177         else
3178                 ret = -ENOSYS;
3179
3180         if (ret)
3181                 goto err;
3182
3183         return 0;
3184 err:
3185         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3186         cma_id_put(id_priv);
3187         return ret;
3188 }
3189 EXPORT_SYMBOL(rdma_resolve_route);
3190
3191 static void cma_set_loopback(struct sockaddr *addr)
3192 {
3193         switch (addr->sa_family) {
3194         case AF_INET:
3195                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3196                 break;
3197         case AF_INET6:
3198                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3199                               0, 0, 0, htonl(1));
3200                 break;
3201         default:
3202                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3203                             0, 0, 0, htonl(1));
3204                 break;
3205         }
3206 }
3207
3208 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3209 {
3210         struct cma_device *cma_dev, *cur_dev;
3211         union ib_gid gid;
3212         enum ib_port_state port_state;
3213         unsigned int p;
3214         u16 pkey;
3215         int ret;
3216
3217         cma_dev = NULL;
3218         mutex_lock(&lock);
3219         list_for_each_entry(cur_dev, &dev_list, list) {
3220                 if (cma_family(id_priv) == AF_IB &&
3221                     !rdma_cap_ib_cm(cur_dev->device, 1))
3222                         continue;
3223
3224                 if (!cma_dev)
3225                         cma_dev = cur_dev;
3226
3227                 rdma_for_each_port (cur_dev->device, p) {
3228                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3229                             port_state == IB_PORT_ACTIVE) {
3230                                 cma_dev = cur_dev;
3231                                 goto port_found;
3232                         }
3233                 }
3234         }
3235
3236         if (!cma_dev) {
3237                 ret = -ENODEV;
3238                 goto out;
3239         }
3240
3241         p = 1;
3242
3243 port_found:
3244         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3245         if (ret)
3246                 goto out;
3247
3248         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3249         if (ret)
3250                 goto out;
3251
3252         id_priv->id.route.addr.dev_addr.dev_type =
3253                 (rdma_protocol_ib(cma_dev->device, p)) ?
3254                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3255
3256         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3257         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3258         id_priv->id.port_num = p;
3259         cma_attach_to_dev(id_priv, cma_dev);
3260         rdma_restrack_add(&id_priv->res);
3261         cma_set_loopback(cma_src_addr(id_priv));
3262 out:
3263         mutex_unlock(&lock);
3264         return ret;
3265 }
3266
3267 static void addr_handler(int status, struct sockaddr *src_addr,
3268                          struct rdma_dev_addr *dev_addr, void *context)
3269 {
3270         struct rdma_id_private *id_priv = context;
3271         struct rdma_cm_event event = {};
3272         struct sockaddr *addr;
3273         struct sockaddr_storage old_addr;
3274
3275         mutex_lock(&id_priv->handler_mutex);
3276         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3277                            RDMA_CM_ADDR_RESOLVED))
3278                 goto out;
3279
3280         /*
3281          * Store the previous src address, so that if we fail to acquire
3282          * matching rdma device, old address can be restored back, which helps
3283          * to cancel the cma listen operation correctly.
3284          */
3285         addr = cma_src_addr(id_priv);
3286         memcpy(&old_addr, addr, rdma_addr_size(addr));
3287         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3288         if (!status && !id_priv->cma_dev) {
3289                 status = cma_acquire_dev_by_src_ip(id_priv);
3290                 if (status)
3291                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3292                                              status);
3293                 rdma_restrack_add(&id_priv->res);
3294         } else if (status) {
3295                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3296         }
3297
3298         if (status) {
3299                 memcpy(addr, &old_addr,
3300                        rdma_addr_size((struct sockaddr *)&old_addr));
3301                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3302                                    RDMA_CM_ADDR_BOUND))
3303                         goto out;
3304                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3305                 event.status = status;
3306         } else
3307                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3308
3309         if (cma_cm_event_handler(id_priv, &event)) {
3310                 destroy_id_handler_unlock(id_priv);
3311                 return;
3312         }
3313 out:
3314         mutex_unlock(&id_priv->handler_mutex);
3315 }
3316
3317 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3318 {
3319         struct cma_work *work;
3320         union ib_gid gid;
3321         int ret;
3322
3323         work = kzalloc(sizeof *work, GFP_KERNEL);
3324         if (!work)
3325                 return -ENOMEM;
3326
3327         if (!id_priv->cma_dev) {
3328                 ret = cma_bind_loopback(id_priv);
3329                 if (ret)
3330                         goto err;
3331         }
3332
3333         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3334         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3335
3336         enqueue_resolve_addr_work(work, id_priv);
3337         return 0;
3338 err:
3339         kfree(work);
3340         return ret;
3341 }
3342
3343 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3344 {
3345         struct cma_work *work;
3346         int ret;
3347
3348         work = kzalloc(sizeof *work, GFP_KERNEL);
3349         if (!work)
3350                 return -ENOMEM;
3351
3352         if (!id_priv->cma_dev) {
3353                 ret = cma_resolve_ib_dev(id_priv);
3354                 if (ret)
3355                         goto err;
3356         }
3357
3358         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3359                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3360
3361         enqueue_resolve_addr_work(work, id_priv);
3362         return 0;
3363 err:
3364         kfree(work);
3365         return ret;
3366 }
3367
3368 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3369                          const struct sockaddr *dst_addr)
3370 {
3371         if (!src_addr || !src_addr->sa_family) {
3372                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3373                 src_addr->sa_family = dst_addr->sa_family;
3374                 if (IS_ENABLED(CONFIG_IPV6) &&
3375                     dst_addr->sa_family == AF_INET6) {
3376                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3377                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3378                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3379                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3380                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3381                 } else if (dst_addr->sa_family == AF_IB) {
3382                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3383                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3384                 }
3385         }
3386         return rdma_bind_addr(id, src_addr);
3387 }
3388
3389 /*
3390  * If required, resolve the source address for bind and leave the id_priv in
3391  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3392  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3393  * ignored.
3394  */
3395 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3396                                struct sockaddr *src_addr,
3397                                const struct sockaddr *dst_addr)
3398 {
3399         int ret;
3400
3401         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3402         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3403                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3404                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3405                 if (ret)
3406                         goto err_dst;
3407                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3408                                            RDMA_CM_ADDR_QUERY))) {
3409                         ret = -EINVAL;
3410                         goto err_dst;
3411                 }
3412         }
3413
3414         if (cma_family(id_priv) != dst_addr->sa_family) {
3415                 ret = -EINVAL;
3416                 goto err_state;
3417         }
3418         return 0;
3419
3420 err_state:
3421         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3422 err_dst:
3423         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3424         return ret;
3425 }
3426
3427 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3428                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3429 {
3430         struct rdma_id_private *id_priv =
3431                 container_of(id, struct rdma_id_private, id);
3432         int ret;
3433
3434         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3435         if (ret)
3436                 return ret;
3437
3438         if (cma_any_addr(dst_addr)) {
3439                 ret = cma_resolve_loopback(id_priv);
3440         } else {
3441                 if (dst_addr->sa_family == AF_IB) {
3442                         ret = cma_resolve_ib_addr(id_priv);
3443                 } else {
3444                         /*
3445                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
3446                          * rdma_resolve_ip() is called, eg through the error
3447                          * path in addr_handler(). If this happens the existing
3448                          * request must be canceled before issuing a new one.
3449                          * Since canceling a request is a bit slow and this
3450                          * oddball path is rare, keep track once a request has
3451                          * been issued. The track turns out to be a permanent
3452                          * state since this is the only cancel as it is
3453                          * immediately before rdma_resolve_ip().
3454                          */
3455                         if (id_priv->used_resolve_ip)
3456                                 rdma_addr_cancel(&id->route.addr.dev_addr);
3457                         else
3458                                 id_priv->used_resolve_ip = 1;
3459                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3460                                               &id->route.addr.dev_addr,
3461                                               timeout_ms, addr_handler,
3462                                               false, id_priv);
3463                 }
3464         }
3465         if (ret)
3466                 goto err;
3467
3468         return 0;
3469 err:
3470         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3471         return ret;
3472 }
3473 EXPORT_SYMBOL(rdma_resolve_addr);
3474
3475 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3476 {
3477         struct rdma_id_private *id_priv;
3478         unsigned long flags;
3479         int ret;
3480
3481         id_priv = container_of(id, struct rdma_id_private, id);
3482         spin_lock_irqsave(&id_priv->lock, flags);
3483         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3484             id_priv->state == RDMA_CM_IDLE) {
3485                 id_priv->reuseaddr = reuse;
3486                 ret = 0;
3487         } else {
3488                 ret = -EINVAL;
3489         }
3490         spin_unlock_irqrestore(&id_priv->lock, flags);
3491         return ret;
3492 }
3493 EXPORT_SYMBOL(rdma_set_reuseaddr);
3494
3495 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3496 {
3497         struct rdma_id_private *id_priv;
3498         unsigned long flags;
3499         int ret;
3500
3501         id_priv = container_of(id, struct rdma_id_private, id);
3502         spin_lock_irqsave(&id_priv->lock, flags);
3503         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3504                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3505                 id_priv->afonly = afonly;
3506                 ret = 0;
3507         } else {
3508                 ret = -EINVAL;
3509         }
3510         spin_unlock_irqrestore(&id_priv->lock, flags);
3511         return ret;
3512 }
3513 EXPORT_SYMBOL(rdma_set_afonly);
3514
3515 static void cma_bind_port(struct rdma_bind_list *bind_list,
3516                           struct rdma_id_private *id_priv)
3517 {
3518         struct sockaddr *addr;
3519         struct sockaddr_ib *sib;
3520         u64 sid, mask;
3521         __be16 port;
3522
3523         lockdep_assert_held(&lock);
3524
3525         addr = cma_src_addr(id_priv);
3526         port = htons(bind_list->port);
3527
3528         switch (addr->sa_family) {
3529         case AF_INET:
3530                 ((struct sockaddr_in *) addr)->sin_port = port;
3531                 break;
3532         case AF_INET6:
3533                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3534                 break;
3535         case AF_IB:
3536                 sib = (struct sockaddr_ib *) addr;
3537                 sid = be64_to_cpu(sib->sib_sid);
3538                 mask = be64_to_cpu(sib->sib_sid_mask);
3539                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3540                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3541                 break;
3542         }
3543         id_priv->bind_list = bind_list;
3544         hlist_add_head(&id_priv->node, &bind_list->owners);
3545 }
3546
3547 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3548                           struct rdma_id_private *id_priv, unsigned short snum)
3549 {
3550         struct rdma_bind_list *bind_list;
3551         int ret;
3552
3553         lockdep_assert_held(&lock);
3554
3555         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3556         if (!bind_list)
3557                 return -ENOMEM;
3558
3559         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3560                            snum);
3561         if (ret < 0)
3562                 goto err;
3563
3564         bind_list->ps = ps;
3565         bind_list->port = snum;
3566         cma_bind_port(bind_list, id_priv);
3567         return 0;
3568 err:
3569         kfree(bind_list);
3570         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3571 }
3572
3573 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3574                               struct rdma_id_private *id_priv)
3575 {
3576         struct rdma_id_private *cur_id;
3577         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3578         struct sockaddr  *saddr = cma_src_addr(id_priv);
3579         __be16 dport = cma_port(daddr);
3580
3581         lockdep_assert_held(&lock);
3582
3583         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3584                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3585                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3586                 __be16 cur_dport = cma_port(cur_daddr);
3587
3588                 if (id_priv == cur_id)
3589                         continue;
3590
3591                 /* different dest port -> unique */
3592                 if (!cma_any_port(daddr) &&
3593                     !cma_any_port(cur_daddr) &&
3594                     (dport != cur_dport))
3595                         continue;
3596
3597                 /* different src address -> unique */
3598                 if (!cma_any_addr(saddr) &&
3599                     !cma_any_addr(cur_saddr) &&
3600                     cma_addr_cmp(saddr, cur_saddr))
3601                         continue;
3602
3603                 /* different dst address -> unique */
3604                 if (!cma_any_addr(daddr) &&
3605                     !cma_any_addr(cur_daddr) &&
3606                     cma_addr_cmp(daddr, cur_daddr))
3607                         continue;
3608
3609                 return -EADDRNOTAVAIL;
3610         }
3611         return 0;
3612 }
3613
3614 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3615                               struct rdma_id_private *id_priv)
3616 {
3617         static unsigned int last_used_port;
3618         int low, high, remaining;
3619         unsigned int rover;
3620         struct net *net = id_priv->id.route.addr.dev_addr.net;
3621
3622         lockdep_assert_held(&lock);
3623
3624         inet_get_local_port_range(net, &low, &high);
3625         remaining = (high - low) + 1;
3626         rover = prandom_u32() % remaining + low;
3627 retry:
3628         if (last_used_port != rover) {
3629                 struct rdma_bind_list *bind_list;
3630                 int ret;
3631
3632                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3633
3634                 if (!bind_list) {
3635                         ret = cma_alloc_port(ps, id_priv, rover);
3636                 } else {
3637                         ret = cma_port_is_unique(bind_list, id_priv);
3638                         if (!ret)
3639                                 cma_bind_port(bind_list, id_priv);
3640                 }
3641                 /*
3642                  * Remember previously used port number in order to avoid
3643                  * re-using same port immediately after it is closed.
3644                  */
3645                 if (!ret)
3646                         last_used_port = rover;
3647                 if (ret != -EADDRNOTAVAIL)
3648                         return ret;
3649         }
3650         if (--remaining) {
3651                 rover++;
3652                 if ((rover < low) || (rover > high))
3653                         rover = low;
3654                 goto retry;
3655         }
3656         return -EADDRNOTAVAIL;
3657 }
3658
3659 /*
3660  * Check that the requested port is available.  This is called when trying to
3661  * bind to a specific port, or when trying to listen on a bound port.  In
3662  * the latter case, the provided id_priv may already be on the bind_list, but
3663  * we still need to check that it's okay to start listening.
3664  */
3665 static int cma_check_port(struct rdma_bind_list *bind_list,
3666                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3667 {
3668         struct rdma_id_private *cur_id;
3669         struct sockaddr *addr, *cur_addr;
3670
3671         lockdep_assert_held(&lock);
3672
3673         addr = cma_src_addr(id_priv);
3674         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3675                 if (id_priv == cur_id)
3676                         continue;
3677
3678                 if (reuseaddr && cur_id->reuseaddr)
3679                         continue;
3680
3681                 cur_addr = cma_src_addr(cur_id);
3682                 if (id_priv->afonly && cur_id->afonly &&
3683                     (addr->sa_family != cur_addr->sa_family))
3684                         continue;
3685
3686                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3687                         return -EADDRNOTAVAIL;
3688
3689                 if (!cma_addr_cmp(addr, cur_addr))
3690                         return -EADDRINUSE;
3691         }
3692         return 0;
3693 }
3694
3695 static int cma_use_port(enum rdma_ucm_port_space ps,
3696                         struct rdma_id_private *id_priv)
3697 {
3698         struct rdma_bind_list *bind_list;
3699         unsigned short snum;
3700         int ret;
3701
3702         lockdep_assert_held(&lock);
3703
3704         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3705         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3706                 return -EACCES;
3707
3708         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3709         if (!bind_list) {
3710                 ret = cma_alloc_port(ps, id_priv, snum);
3711         } else {
3712                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3713                 if (!ret)
3714                         cma_bind_port(bind_list, id_priv);
3715         }
3716         return ret;
3717 }
3718
3719 static enum rdma_ucm_port_space
3720 cma_select_inet_ps(struct rdma_id_private *id_priv)
3721 {
3722         switch (id_priv->id.ps) {
3723         case RDMA_PS_TCP:
3724         case RDMA_PS_UDP:
3725         case RDMA_PS_IPOIB:
3726         case RDMA_PS_IB:
3727                 return id_priv->id.ps;
3728         default:
3729
3730                 return 0;
3731         }
3732 }
3733
3734 static enum rdma_ucm_port_space
3735 cma_select_ib_ps(struct rdma_id_private *id_priv)
3736 {
3737         enum rdma_ucm_port_space ps = 0;
3738         struct sockaddr_ib *sib;
3739         u64 sid_ps, mask, sid;
3740
3741         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3742         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3743         sid = be64_to_cpu(sib->sib_sid) & mask;
3744
3745         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3746                 sid_ps = RDMA_IB_IP_PS_IB;
3747                 ps = RDMA_PS_IB;
3748         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3749                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3750                 sid_ps = RDMA_IB_IP_PS_TCP;
3751                 ps = RDMA_PS_TCP;
3752         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3753                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3754                 sid_ps = RDMA_IB_IP_PS_UDP;
3755                 ps = RDMA_PS_UDP;
3756         }
3757
3758         if (ps) {
3759                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3760                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3761                                                 be64_to_cpu(sib->sib_sid_mask));
3762         }
3763         return ps;
3764 }
3765
3766 static int cma_get_port(struct rdma_id_private *id_priv)
3767 {
3768         enum rdma_ucm_port_space ps;
3769         int ret;
3770
3771         if (cma_family(id_priv) != AF_IB)
3772                 ps = cma_select_inet_ps(id_priv);
3773         else
3774                 ps = cma_select_ib_ps(id_priv);
3775         if (!ps)
3776                 return -EPROTONOSUPPORT;
3777
3778         mutex_lock(&lock);
3779         if (cma_any_port(cma_src_addr(id_priv)))
3780                 ret = cma_alloc_any_port(ps, id_priv);
3781         else
3782                 ret = cma_use_port(ps, id_priv);
3783         mutex_unlock(&lock);
3784
3785         return ret;
3786 }
3787
3788 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3789                                struct sockaddr *addr)
3790 {
3791 #if IS_ENABLED(CONFIG_IPV6)
3792         struct sockaddr_in6 *sin6;
3793
3794         if (addr->sa_family != AF_INET6)
3795                 return 0;
3796
3797         sin6 = (struct sockaddr_in6 *) addr;
3798
3799         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3800                 return 0;
3801
3802         if (!sin6->sin6_scope_id)
3803                         return -EINVAL;
3804
3805         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3806 #endif
3807         return 0;
3808 }
3809
3810 int rdma_listen(struct rdma_cm_id *id, int backlog)
3811 {
3812         struct rdma_id_private *id_priv =
3813                 container_of(id, struct rdma_id_private, id);
3814         int ret;
3815
3816         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3817                 struct sockaddr_in any_in = {
3818                         .sin_family = AF_INET,
3819                         .sin_addr.s_addr = htonl(INADDR_ANY),
3820                 };
3821
3822                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3823                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3824                 if (ret)
3825                         return ret;
3826                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3827                                            RDMA_CM_LISTEN)))
3828                         return -EINVAL;
3829         }
3830
3831         /*
3832          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3833          * any more, and has to be unique in the bind list.
3834          */
3835         if (id_priv->reuseaddr) {
3836                 mutex_lock(&lock);
3837                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3838                 if (!ret)
3839                         id_priv->reuseaddr = 0;
3840                 mutex_unlock(&lock);
3841                 if (ret)
3842                         goto err;
3843         }
3844
3845         id_priv->backlog = backlog;
3846         if (id_priv->cma_dev) {
3847                 if (rdma_cap_ib_cm(id->device, 1)) {
3848                         ret = cma_ib_listen(id_priv);
3849                         if (ret)
3850                                 goto err;
3851                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3852                         ret = cma_iw_listen(id_priv, backlog);
3853                         if (ret)
3854                                 goto err;
3855                 } else {
3856                         ret = -ENOSYS;
3857                         goto err;
3858                 }
3859         } else {
3860                 ret = cma_listen_on_all(id_priv);
3861                 if (ret)
3862                         goto err;
3863         }
3864
3865         return 0;
3866 err:
3867         id_priv->backlog = 0;
3868         /*
3869          * All the failure paths that lead here will not allow the req_handler's
3870          * to have run.
3871          */
3872         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3873         return ret;
3874 }
3875 EXPORT_SYMBOL(rdma_listen);
3876
3877 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3878 {
3879         struct rdma_id_private *id_priv;
3880         int ret;
3881         struct sockaddr  *daddr;
3882
3883         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3884             addr->sa_family != AF_IB)
3885                 return -EAFNOSUPPORT;
3886
3887         id_priv = container_of(id, struct rdma_id_private, id);
3888         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3889                 return -EINVAL;
3890
3891         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3892         if (ret)
3893                 goto err1;
3894
3895         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3896         if (!cma_any_addr(addr)) {
3897                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3898                 if (ret)
3899                         goto err1;
3900
3901                 ret = cma_acquire_dev_by_src_ip(id_priv);
3902                 if (ret)
3903                         goto err1;
3904         }
3905
3906         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3907                 if (addr->sa_family == AF_INET)
3908                         id_priv->afonly = 1;
3909 #if IS_ENABLED(CONFIG_IPV6)
3910                 else if (addr->sa_family == AF_INET6) {
3911                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3912
3913                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3914                 }
3915 #endif
3916         }
3917         daddr = cma_dst_addr(id_priv);
3918         daddr->sa_family = addr->sa_family;
3919
3920         ret = cma_get_port(id_priv);
3921         if (ret)
3922                 goto err2;
3923
3924         if (!cma_any_addr(addr))
3925                 rdma_restrack_add(&id_priv->res);
3926         return 0;
3927 err2:
3928         if (id_priv->cma_dev)
3929                 cma_release_dev(id_priv);
3930 err1:
3931         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3932         return ret;
3933 }
3934 EXPORT_SYMBOL(rdma_bind_addr);
3935
3936 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3937 {
3938         struct cma_hdr *cma_hdr;
3939
3940         cma_hdr = hdr;
3941         cma_hdr->cma_version = CMA_VERSION;
3942         if (cma_family(id_priv) == AF_INET) {
3943                 struct sockaddr_in *src4, *dst4;
3944
3945                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3946                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3947
3948                 cma_set_ip_ver(cma_hdr, 4);
3949                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3950                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3951                 cma_hdr->port = src4->sin_port;
3952         } else if (cma_family(id_priv) == AF_INET6) {
3953                 struct sockaddr_in6 *src6, *dst6;
3954
3955                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3956                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3957
3958                 cma_set_ip_ver(cma_hdr, 6);
3959                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3960                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3961                 cma_hdr->port = src6->sin6_port;
3962         }
3963         return 0;
3964 }
3965
3966 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3967                                 const struct ib_cm_event *ib_event)
3968 {
3969         struct rdma_id_private *id_priv = cm_id->context;
3970         struct rdma_cm_event event = {};
3971         const struct ib_cm_sidr_rep_event_param *rep =
3972                                 &ib_event->param.sidr_rep_rcvd;
3973         int ret;
3974
3975         mutex_lock(&id_priv->handler_mutex);
3976         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3977                 goto out;
3978
3979         switch (ib_event->event) {
3980         case IB_CM_SIDR_REQ_ERROR:
3981                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3982                 event.status = -ETIMEDOUT;
3983                 break;
3984         case IB_CM_SIDR_REP_RECEIVED:
3985                 event.param.ud.private_data = ib_event->private_data;
3986                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3987                 if (rep->status != IB_SIDR_SUCCESS) {
3988                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3989                         event.status = ib_event->param.sidr_rep_rcvd.status;
3990                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3991                                              event.status);
3992                         break;
3993                 }
3994                 ret = cma_set_qkey(id_priv, rep->qkey);
3995                 if (ret) {
3996                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3997                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3998                         event.status = ret;
3999                         break;
4000                 }
4001                 ib_init_ah_attr_from_path(id_priv->id.device,
4002                                           id_priv->id.port_num,
4003                                           id_priv->id.route.path_rec,
4004                                           &event.param.ud.ah_attr,
4005                                           rep->sgid_attr);
4006                 event.param.ud.qp_num = rep->qpn;
4007                 event.param.ud.qkey = rep->qkey;
4008                 event.event = RDMA_CM_EVENT_ESTABLISHED;
4009                 event.status = 0;
4010                 break;
4011         default:
4012                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4013                        ib_event->event);
4014                 goto out;
4015         }
4016
4017         ret = cma_cm_event_handler(id_priv, &event);
4018
4019         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4020         if (ret) {
4021                 /* Destroy the CM ID by returning a non-zero value. */
4022                 id_priv->cm_id.ib = NULL;
4023                 destroy_id_handler_unlock(id_priv);
4024                 return ret;
4025         }
4026 out:
4027         mutex_unlock(&id_priv->handler_mutex);
4028         return 0;
4029 }
4030
4031 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4032                               struct rdma_conn_param *conn_param)
4033 {
4034         struct ib_cm_sidr_req_param req;
4035         struct ib_cm_id *id;
4036         void *private_data;
4037         u8 offset;
4038         int ret;
4039
4040         memset(&req, 0, sizeof req);
4041         offset = cma_user_data_offset(id_priv);
4042         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4043                 return -EINVAL;
4044
4045         if (req.private_data_len) {
4046                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4047                 if (!private_data)
4048                         return -ENOMEM;
4049         } else {
4050                 private_data = NULL;
4051         }
4052
4053         if (conn_param->private_data && conn_param->private_data_len)
4054                 memcpy(private_data + offset, conn_param->private_data,
4055                        conn_param->private_data_len);
4056
4057         if (private_data) {
4058                 ret = cma_format_hdr(private_data, id_priv);
4059                 if (ret)
4060                         goto out;
4061                 req.private_data = private_data;
4062         }
4063
4064         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4065                              id_priv);
4066         if (IS_ERR(id)) {
4067                 ret = PTR_ERR(id);
4068                 goto out;
4069         }
4070         id_priv->cm_id.ib = id;
4071
4072         req.path = id_priv->id.route.path_rec;
4073         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4074         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4075         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4076         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4077
4078         trace_cm_send_sidr_req(id_priv);
4079         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4080         if (ret) {
4081                 ib_destroy_cm_id(id_priv->cm_id.ib);
4082                 id_priv->cm_id.ib = NULL;
4083         }
4084 out:
4085         kfree(private_data);
4086         return ret;
4087 }
4088
4089 static int cma_connect_ib(struct rdma_id_private *id_priv,
4090                           struct rdma_conn_param *conn_param)
4091 {
4092         struct ib_cm_req_param req;
4093         struct rdma_route *route;
4094         void *private_data;
4095         struct ib_cm_id *id;
4096         u8 offset;
4097         int ret;
4098
4099         memset(&req, 0, sizeof req);
4100         offset = cma_user_data_offset(id_priv);
4101         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4102                 return -EINVAL;
4103
4104         if (req.private_data_len) {
4105                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4106                 if (!private_data)
4107                         return -ENOMEM;
4108         } else {
4109                 private_data = NULL;
4110         }
4111
4112         if (conn_param->private_data && conn_param->private_data_len)
4113                 memcpy(private_data + offset, conn_param->private_data,
4114                        conn_param->private_data_len);
4115
4116         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4117         if (IS_ERR(id)) {
4118                 ret = PTR_ERR(id);
4119                 goto out;
4120         }
4121         id_priv->cm_id.ib = id;
4122
4123         route = &id_priv->id.route;
4124         if (private_data) {
4125                 ret = cma_format_hdr(private_data, id_priv);
4126                 if (ret)
4127                         goto out;
4128                 req.private_data = private_data;
4129         }
4130
4131         req.primary_path = &route->path_rec[0];
4132         if (route->num_paths == 2)
4133                 req.alternate_path = &route->path_rec[1];
4134
4135         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4136         /* Alternate path SGID attribute currently unsupported */
4137         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4138         req.qp_num = id_priv->qp_num;
4139         req.qp_type = id_priv->id.qp_type;
4140         req.starting_psn = id_priv->seq_num;
4141         req.responder_resources = conn_param->responder_resources;
4142         req.initiator_depth = conn_param->initiator_depth;
4143         req.flow_control = conn_param->flow_control;
4144         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4145         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4146         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4147         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4148         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4149         req.srq = id_priv->srq ? 1 : 0;
4150         req.ece.vendor_id = id_priv->ece.vendor_id;
4151         req.ece.attr_mod = id_priv->ece.attr_mod;
4152
4153         trace_cm_send_req(id_priv);
4154         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4155 out:
4156         if (ret && !IS_ERR(id)) {
4157                 ib_destroy_cm_id(id);
4158                 id_priv->cm_id.ib = NULL;
4159         }
4160
4161         kfree(private_data);
4162         return ret;
4163 }
4164
4165 static int cma_connect_iw(struct rdma_id_private *id_priv,
4166                           struct rdma_conn_param *conn_param)
4167 {
4168         struct iw_cm_id *cm_id;
4169         int ret;
4170         struct iw_cm_conn_param iw_param;
4171
4172         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4173         if (IS_ERR(cm_id))
4174                 return PTR_ERR(cm_id);
4175
4176         mutex_lock(&id_priv->qp_mutex);
4177         cm_id->tos = id_priv->tos;
4178         cm_id->tos_set = id_priv->tos_set;
4179         mutex_unlock(&id_priv->qp_mutex);
4180
4181         id_priv->cm_id.iw = cm_id;
4182
4183         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4184                rdma_addr_size(cma_src_addr(id_priv)));
4185         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4186                rdma_addr_size(cma_dst_addr(id_priv)));
4187
4188         ret = cma_modify_qp_rtr(id_priv, conn_param);
4189         if (ret)
4190                 goto out;
4191
4192         if (conn_param) {
4193                 iw_param.ord = conn_param->initiator_depth;
4194                 iw_param.ird = conn_param->responder_resources;
4195                 iw_param.private_data = conn_param->private_data;
4196                 iw_param.private_data_len = conn_param->private_data_len;
4197                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4198         } else {
4199                 memset(&iw_param, 0, sizeof iw_param);
4200                 iw_param.qpn = id_priv->qp_num;
4201         }
4202         ret = iw_cm_connect(cm_id, &iw_param);
4203 out:
4204         if (ret) {
4205                 iw_destroy_cm_id(cm_id);
4206                 id_priv->cm_id.iw = NULL;
4207         }
4208         return ret;
4209 }
4210
4211 /**
4212  * rdma_connect_locked - Initiate an active connection request.
4213  * @id: Connection identifier to connect.
4214  * @conn_param: Connection information used for connected QPs.
4215  *
4216  * Same as rdma_connect() but can only be called from the
4217  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4218  */
4219 int rdma_connect_locked(struct rdma_cm_id *id,
4220                         struct rdma_conn_param *conn_param)
4221 {
4222         struct rdma_id_private *id_priv =
4223                 container_of(id, struct rdma_id_private, id);
4224         int ret;
4225
4226         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4227                 return -EINVAL;
4228
4229         if (!id->qp) {
4230                 id_priv->qp_num = conn_param->qp_num;
4231                 id_priv->srq = conn_param->srq;
4232         }
4233
4234         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4235                 if (id->qp_type == IB_QPT_UD)
4236                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4237                 else
4238                         ret = cma_connect_ib(id_priv, conn_param);
4239         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4240                 ret = cma_connect_iw(id_priv, conn_param);
4241         } else {
4242                 ret = -ENOSYS;
4243         }
4244         if (ret)
4245                 goto err_state;
4246         return 0;
4247 err_state:
4248         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4249         return ret;
4250 }
4251 EXPORT_SYMBOL(rdma_connect_locked);
4252
4253 /**
4254  * rdma_connect - Initiate an active connection request.
4255  * @id: Connection identifier to connect.
4256  * @conn_param: Connection information used for connected QPs.
4257  *
4258  * Users must have resolved a route for the rdma_cm_id to connect with by having
4259  * called rdma_resolve_route before calling this routine.
4260  *
4261  * This call will either connect to a remote QP or obtain remote QP information
4262  * for unconnected rdma_cm_id's.  The actual operation is based on the
4263  * rdma_cm_id's port space.
4264  */
4265 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4266 {
4267         struct rdma_id_private *id_priv =
4268                 container_of(id, struct rdma_id_private, id);
4269         int ret;
4270
4271         mutex_lock(&id_priv->handler_mutex);
4272         ret = rdma_connect_locked(id, conn_param);
4273         mutex_unlock(&id_priv->handler_mutex);
4274         return ret;
4275 }
4276 EXPORT_SYMBOL(rdma_connect);
4277
4278 /**
4279  * rdma_connect_ece - Initiate an active connection request with ECE data.
4280  * @id: Connection identifier to connect.
4281  * @conn_param: Connection information used for connected QPs.
4282  * @ece: ECE parameters
4283  *
4284  * See rdma_connect() explanation.
4285  */
4286 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4287                      struct rdma_ucm_ece *ece)
4288 {
4289         struct rdma_id_private *id_priv =
4290                 container_of(id, struct rdma_id_private, id);
4291
4292         id_priv->ece.vendor_id = ece->vendor_id;
4293         id_priv->ece.attr_mod = ece->attr_mod;
4294
4295         return rdma_connect(id, conn_param);
4296 }
4297 EXPORT_SYMBOL(rdma_connect_ece);
4298
4299 static int cma_accept_ib(struct rdma_id_private *id_priv,
4300                          struct rdma_conn_param *conn_param)
4301 {
4302         struct ib_cm_rep_param rep;
4303         int ret;
4304
4305         ret = cma_modify_qp_rtr(id_priv, conn_param);
4306         if (ret)
4307                 goto out;
4308
4309         ret = cma_modify_qp_rts(id_priv, conn_param);
4310         if (ret)
4311                 goto out;
4312
4313         memset(&rep, 0, sizeof rep);
4314         rep.qp_num = id_priv->qp_num;
4315         rep.starting_psn = id_priv->seq_num;
4316         rep.private_data = conn_param->private_data;
4317         rep.private_data_len = conn_param->private_data_len;
4318         rep.responder_resources = conn_param->responder_resources;
4319         rep.initiator_depth = conn_param->initiator_depth;
4320         rep.failover_accepted = 0;
4321         rep.flow_control = conn_param->flow_control;
4322         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4323         rep.srq = id_priv->srq ? 1 : 0;
4324         rep.ece.vendor_id = id_priv->ece.vendor_id;
4325         rep.ece.attr_mod = id_priv->ece.attr_mod;
4326
4327         trace_cm_send_rep(id_priv);
4328         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4329 out:
4330         return ret;
4331 }
4332
4333 static int cma_accept_iw(struct rdma_id_private *id_priv,
4334                   struct rdma_conn_param *conn_param)
4335 {
4336         struct iw_cm_conn_param iw_param;
4337         int ret;
4338
4339         if (!conn_param)
4340                 return -EINVAL;
4341
4342         ret = cma_modify_qp_rtr(id_priv, conn_param);
4343         if (ret)
4344                 return ret;
4345
4346         iw_param.ord = conn_param->initiator_depth;
4347         iw_param.ird = conn_param->responder_resources;
4348         iw_param.private_data = conn_param->private_data;
4349         iw_param.private_data_len = conn_param->private_data_len;
4350         if (id_priv->id.qp)
4351                 iw_param.qpn = id_priv->qp_num;
4352         else
4353                 iw_param.qpn = conn_param->qp_num;
4354
4355         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4356 }
4357
4358 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4359                              enum ib_cm_sidr_status status, u32 qkey,
4360                              const void *private_data, int private_data_len)
4361 {
4362         struct ib_cm_sidr_rep_param rep;
4363         int ret;
4364
4365         memset(&rep, 0, sizeof rep);
4366         rep.status = status;
4367         if (status == IB_SIDR_SUCCESS) {
4368                 ret = cma_set_qkey(id_priv, qkey);
4369                 if (ret)
4370                         return ret;
4371                 rep.qp_num = id_priv->qp_num;
4372                 rep.qkey = id_priv->qkey;
4373
4374                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4375                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4376         }
4377
4378         rep.private_data = private_data;
4379         rep.private_data_len = private_data_len;
4380
4381         trace_cm_send_sidr_rep(id_priv);
4382         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4383 }
4384
4385 /**
4386  * rdma_accept - Called to accept a connection request or response.
4387  * @id: Connection identifier associated with the request.
4388  * @conn_param: Information needed to establish the connection.  This must be
4389  *   provided if accepting a connection request.  If accepting a connection
4390  *   response, this parameter must be NULL.
4391  *
4392  * Typically, this routine is only called by the listener to accept a connection
4393  * request.  It must also be called on the active side of a connection if the
4394  * user is performing their own QP transitions.
4395  *
4396  * In the case of error, a reject message is sent to the remote side and the
4397  * state of the qp associated with the id is modified to error, such that any
4398  * previously posted receive buffers would be flushed.
4399  *
4400  * This function is for use by kernel ULPs and must be called from under the
4401  * handler callback.
4402  */
4403 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4404 {
4405         struct rdma_id_private *id_priv =
4406                 container_of(id, struct rdma_id_private, id);
4407         int ret;
4408
4409         lockdep_assert_held(&id_priv->handler_mutex);
4410
4411         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4412                 return -EINVAL;
4413
4414         if (!id->qp && conn_param) {
4415                 id_priv->qp_num = conn_param->qp_num;
4416                 id_priv->srq = conn_param->srq;
4417         }
4418
4419         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4420                 if (id->qp_type == IB_QPT_UD) {
4421                         if (conn_param)
4422                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4423                                                         conn_param->qkey,
4424                                                         conn_param->private_data,
4425                                                         conn_param->private_data_len);
4426                         else
4427                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4428                                                         0, NULL, 0);
4429                 } else {
4430                         if (conn_param)
4431                                 ret = cma_accept_ib(id_priv, conn_param);
4432                         else
4433                                 ret = cma_rep_recv(id_priv);
4434                 }
4435         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4436                 ret = cma_accept_iw(id_priv, conn_param);
4437         } else {
4438                 ret = -ENOSYS;
4439         }
4440         if (ret)
4441                 goto reject;
4442
4443         return 0;
4444 reject:
4445         cma_modify_qp_err(id_priv);
4446         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4447         return ret;
4448 }
4449 EXPORT_SYMBOL(rdma_accept);
4450
4451 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4452                     struct rdma_ucm_ece *ece)
4453 {
4454         struct rdma_id_private *id_priv =
4455                 container_of(id, struct rdma_id_private, id);
4456
4457         id_priv->ece.vendor_id = ece->vendor_id;
4458         id_priv->ece.attr_mod = ece->attr_mod;
4459
4460         return rdma_accept(id, conn_param);
4461 }
4462 EXPORT_SYMBOL(rdma_accept_ece);
4463
4464 void rdma_lock_handler(struct rdma_cm_id *id)
4465 {
4466         struct rdma_id_private *id_priv =
4467                 container_of(id, struct rdma_id_private, id);
4468
4469         mutex_lock(&id_priv->handler_mutex);
4470 }
4471 EXPORT_SYMBOL(rdma_lock_handler);
4472
4473 void rdma_unlock_handler(struct rdma_cm_id *id)
4474 {
4475         struct rdma_id_private *id_priv =
4476                 container_of(id, struct rdma_id_private, id);
4477
4478         mutex_unlock(&id_priv->handler_mutex);
4479 }
4480 EXPORT_SYMBOL(rdma_unlock_handler);
4481
4482 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4483 {
4484         struct rdma_id_private *id_priv;
4485         int ret;
4486
4487         id_priv = container_of(id, struct rdma_id_private, id);
4488         if (!id_priv->cm_id.ib)
4489                 return -EINVAL;
4490
4491         switch (id->device->node_type) {
4492         case RDMA_NODE_IB_CA:
4493                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4494                 break;
4495         default:
4496                 ret = 0;
4497                 break;
4498         }
4499         return ret;
4500 }
4501 EXPORT_SYMBOL(rdma_notify);
4502
4503 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4504                 u8 private_data_len, u8 reason)
4505 {
4506         struct rdma_id_private *id_priv;
4507         int ret;
4508
4509         id_priv = container_of(id, struct rdma_id_private, id);
4510         if (!id_priv->cm_id.ib)
4511                 return -EINVAL;
4512
4513         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4514                 if (id->qp_type == IB_QPT_UD) {
4515                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4516                                                 private_data, private_data_len);
4517                 } else {
4518                         trace_cm_send_rej(id_priv);
4519                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4520                                              private_data, private_data_len);
4521                 }
4522         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4523                 ret = iw_cm_reject(id_priv->cm_id.iw,
4524                                    private_data, private_data_len);
4525         } else {
4526                 ret = -ENOSYS;
4527         }
4528
4529         return ret;
4530 }
4531 EXPORT_SYMBOL(rdma_reject);
4532
4533 int rdma_disconnect(struct rdma_cm_id *id)
4534 {
4535         struct rdma_id_private *id_priv;
4536         int ret;
4537
4538         id_priv = container_of(id, struct rdma_id_private, id);
4539         if (!id_priv->cm_id.ib)
4540                 return -EINVAL;
4541
4542         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4543                 ret = cma_modify_qp_err(id_priv);
4544                 if (ret)
4545                         goto out;
4546                 /* Initiate or respond to a disconnect. */
4547                 trace_cm_disconnect(id_priv);
4548                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4549                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4550                                 trace_cm_sent_drep(id_priv);
4551                 } else {
4552                         trace_cm_sent_dreq(id_priv);
4553                 }
4554         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4555                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4556         } else
4557                 ret = -EINVAL;
4558
4559 out:
4560         return ret;
4561 }
4562 EXPORT_SYMBOL(rdma_disconnect);
4563
4564 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4565                               struct ib_sa_multicast *multicast,
4566                               struct rdma_cm_event *event,
4567                               struct cma_multicast *mc)
4568 {
4569         struct rdma_dev_addr *dev_addr;
4570         enum ib_gid_type gid_type;
4571         struct net_device *ndev;
4572
4573         if (!status)
4574                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4575         else
4576                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4577                                      status);
4578
4579         event->status = status;
4580         event->param.ud.private_data = mc->context;
4581         if (status) {
4582                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4583                 return;
4584         }
4585
4586         dev_addr = &id_priv->id.route.addr.dev_addr;
4587         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4588         gid_type =
4589                 id_priv->cma_dev
4590                         ->default_gid_type[id_priv->id.port_num -
4591                                            rdma_start_port(
4592                                                    id_priv->cma_dev->device)];
4593
4594         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4595         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4596                                      &multicast->rec, ndev, gid_type,
4597                                      &event->param.ud.ah_attr)) {
4598                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4599                 goto out;
4600         }
4601
4602         event->param.ud.qp_num = 0xFFFFFF;
4603         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4604
4605 out:
4606         if (ndev)
4607                 dev_put(ndev);
4608 }
4609
4610 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4611 {
4612         struct cma_multicast *mc = multicast->context;
4613         struct rdma_id_private *id_priv = mc->id_priv;
4614         struct rdma_cm_event event = {};
4615         int ret = 0;
4616
4617         mutex_lock(&id_priv->handler_mutex);
4618         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4619             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4620                 goto out;
4621
4622         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4623         ret = cma_cm_event_handler(id_priv, &event);
4624         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4625         WARN_ON(ret);
4626
4627 out:
4628         mutex_unlock(&id_priv->handler_mutex);
4629         return 0;
4630 }
4631
4632 static void cma_set_mgid(struct rdma_id_private *id_priv,
4633                          struct sockaddr *addr, union ib_gid *mgid)
4634 {
4635         unsigned char mc_map[MAX_ADDR_LEN];
4636         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4637         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4638         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4639
4640         if (cma_any_addr(addr)) {
4641                 memset(mgid, 0, sizeof *mgid);
4642         } else if ((addr->sa_family == AF_INET6) &&
4643                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4644                                                                  0xFF10A01B)) {
4645                 /* IPv6 address is an SA assigned MGID. */
4646                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4647         } else if (addr->sa_family == AF_IB) {
4648                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4649         } else if (addr->sa_family == AF_INET6) {
4650                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4651                 if (id_priv->id.ps == RDMA_PS_UDP)
4652                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4653                 *mgid = *(union ib_gid *) (mc_map + 4);
4654         } else {
4655                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4656                 if (id_priv->id.ps == RDMA_PS_UDP)
4657                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4658                 *mgid = *(union ib_gid *) (mc_map + 4);
4659         }
4660 }
4661
4662 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4663                                  struct cma_multicast *mc)
4664 {
4665         struct ib_sa_mcmember_rec rec;
4666         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4667         ib_sa_comp_mask comp_mask;
4668         int ret;
4669
4670         ib_addr_get_mgid(dev_addr, &rec.mgid);
4671         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4672                                      &rec.mgid, &rec);
4673         if (ret)
4674                 return ret;
4675
4676         ret = cma_set_qkey(id_priv, 0);
4677         if (ret)
4678                 return ret;
4679
4680         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4681         rec.qkey = cpu_to_be32(id_priv->qkey);
4682         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4683         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4684         rec.join_state = mc->join_state;
4685
4686         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4687                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4688                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4689                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4690                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4691
4692         if (id_priv->id.ps == RDMA_PS_IPOIB)
4693                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4694                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4695                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4696                              IB_SA_MCMEMBER_REC_MTU |
4697                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4698
4699         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4700                                          id_priv->id.port_num, &rec, comp_mask,
4701                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4702         return PTR_ERR_OR_ZERO(mc->sa_mc);
4703 }
4704
4705 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4706                               enum ib_gid_type gid_type)
4707 {
4708         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4709         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4710
4711         if (cma_any_addr(addr)) {
4712                 memset(mgid, 0, sizeof *mgid);
4713         } else if (addr->sa_family == AF_INET6) {
4714                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4715         } else {
4716                 mgid->raw[0] =
4717                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4718                 mgid->raw[1] =
4719                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4720                 mgid->raw[2] = 0;
4721                 mgid->raw[3] = 0;
4722                 mgid->raw[4] = 0;
4723                 mgid->raw[5] = 0;
4724                 mgid->raw[6] = 0;
4725                 mgid->raw[7] = 0;
4726                 mgid->raw[8] = 0;
4727                 mgid->raw[9] = 0;
4728                 mgid->raw[10] = 0xff;
4729                 mgid->raw[11] = 0xff;
4730                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4731         }
4732 }
4733
4734 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4735                                    struct cma_multicast *mc)
4736 {
4737         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4738         int err = 0;
4739         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4740         struct net_device *ndev = NULL;
4741         struct ib_sa_multicast ib;
4742         enum ib_gid_type gid_type;
4743         bool send_only;
4744
4745         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4746
4747         if (cma_zero_addr(addr))
4748                 return -EINVAL;
4749
4750         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4751                    rdma_start_port(id_priv->cma_dev->device)];
4752         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4753
4754         ib.rec.pkey = cpu_to_be16(0xffff);
4755         if (id_priv->id.ps == RDMA_PS_UDP)
4756                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4757
4758         if (dev_addr->bound_dev_if)
4759                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4760         if (!ndev)
4761                 return -ENODEV;
4762
4763         ib.rec.rate = iboe_get_rate(ndev);
4764         ib.rec.hop_limit = 1;
4765         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4766
4767         if (addr->sa_family == AF_INET) {
4768                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4769                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4770                         if (!send_only) {
4771                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4772                                                     true);
4773                         }
4774                 }
4775         } else {
4776                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4777                         err = -ENOTSUPP;
4778         }
4779         dev_put(ndev);
4780         if (err || !ib.rec.mtu)
4781                 return err ?: -EINVAL;
4782
4783         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4784                     &ib.rec.port_gid);
4785         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4786         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4787         queue_work(cma_wq, &mc->iboe_join.work);
4788         return 0;
4789 }
4790
4791 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4792                         u8 join_state, void *context)
4793 {
4794         struct rdma_id_private *id_priv =
4795                 container_of(id, struct rdma_id_private, id);
4796         struct cma_multicast *mc;
4797         int ret;
4798
4799         /* Not supported for kernel QPs */
4800         if (WARN_ON(id->qp))
4801                 return -EINVAL;
4802
4803         /* ULP is calling this wrong. */
4804         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4805                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4806                 return -EINVAL;
4807
4808         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4809         if (!mc)
4810                 return -ENOMEM;
4811
4812         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4813         mc->context = context;
4814         mc->id_priv = id_priv;
4815         mc->join_state = join_state;
4816
4817         if (rdma_protocol_roce(id->device, id->port_num)) {
4818                 ret = cma_iboe_join_multicast(id_priv, mc);
4819                 if (ret)
4820                         goto out_err;
4821         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4822                 ret = cma_join_ib_multicast(id_priv, mc);
4823                 if (ret)
4824                         goto out_err;
4825         } else {
4826                 ret = -ENOSYS;
4827                 goto out_err;
4828         }
4829
4830         spin_lock(&id_priv->lock);
4831         list_add(&mc->list, &id_priv->mc_list);
4832         spin_unlock(&id_priv->lock);
4833
4834         return 0;
4835 out_err:
4836         kfree(mc);
4837         return ret;
4838 }
4839 EXPORT_SYMBOL(rdma_join_multicast);
4840
4841 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4842 {
4843         struct rdma_id_private *id_priv;
4844         struct cma_multicast *mc;
4845
4846         id_priv = container_of(id, struct rdma_id_private, id);
4847         spin_lock_irq(&id_priv->lock);
4848         list_for_each_entry(mc, &id_priv->mc_list, list) {
4849                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4850                         continue;
4851                 list_del(&mc->list);
4852                 spin_unlock_irq(&id_priv->lock);
4853
4854                 WARN_ON(id_priv->cma_dev->device != id->device);
4855                 destroy_mc(id_priv, mc);
4856                 return;
4857         }
4858         spin_unlock_irq(&id_priv->lock);
4859 }
4860 EXPORT_SYMBOL(rdma_leave_multicast);
4861
4862 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4863 {
4864         struct rdma_dev_addr *dev_addr;
4865         struct cma_work *work;
4866
4867         dev_addr = &id_priv->id.route.addr.dev_addr;
4868
4869         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4870             (net_eq(dev_net(ndev), dev_addr->net)) &&
4871             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4872                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4873                         ndev->name, &id_priv->id);
4874                 work = kzalloc(sizeof *work, GFP_KERNEL);
4875                 if (!work)
4876                         return -ENOMEM;
4877
4878                 INIT_WORK(&work->work, cma_work_handler);
4879                 work->id = id_priv;
4880                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4881                 cma_id_get(id_priv);
4882                 queue_work(cma_wq, &work->work);
4883         }
4884
4885         return 0;
4886 }
4887
4888 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4889                                void *ptr)
4890 {
4891         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4892         struct cma_device *cma_dev;
4893         struct rdma_id_private *id_priv;
4894         int ret = NOTIFY_DONE;
4895
4896         if (event != NETDEV_BONDING_FAILOVER)
4897                 return NOTIFY_DONE;
4898
4899         if (!netif_is_bond_master(ndev))
4900                 return NOTIFY_DONE;
4901
4902         mutex_lock(&lock);
4903         list_for_each_entry(cma_dev, &dev_list, list)
4904                 list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
4905                         ret = cma_netdev_change(ndev, id_priv);
4906                         if (ret)
4907                                 goto out;
4908                 }
4909
4910 out:
4911         mutex_unlock(&lock);
4912         return ret;
4913 }
4914
4915 static struct notifier_block cma_nb = {
4916         .notifier_call = cma_netdev_callback
4917 };
4918
4919 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4920 {
4921         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4922         enum rdma_cm_state state;
4923         unsigned long flags;
4924
4925         mutex_lock(&id_priv->handler_mutex);
4926         /* Record that we want to remove the device */
4927         spin_lock_irqsave(&id_priv->lock, flags);
4928         state = id_priv->state;
4929         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4930                 spin_unlock_irqrestore(&id_priv->lock, flags);
4931                 mutex_unlock(&id_priv->handler_mutex);
4932                 cma_id_put(id_priv);
4933                 return;
4934         }
4935         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4936         spin_unlock_irqrestore(&id_priv->lock, flags);
4937
4938         if (cma_cm_event_handler(id_priv, &event)) {
4939                 /*
4940                  * At this point the ULP promises it won't call
4941                  * rdma_destroy_id() concurrently
4942                  */
4943                 cma_id_put(id_priv);
4944                 mutex_unlock(&id_priv->handler_mutex);
4945                 trace_cm_id_destroy(id_priv);
4946                 _destroy_id(id_priv, state);
4947                 return;
4948         }
4949         mutex_unlock(&id_priv->handler_mutex);
4950
4951         /*
4952          * If this races with destroy then the thread that first assigns state
4953          * to a destroying does the cancel.
4954          */
4955         cma_cancel_operation(id_priv, state);
4956         cma_id_put(id_priv);
4957 }
4958
4959 static void cma_process_remove(struct cma_device *cma_dev)
4960 {
4961         mutex_lock(&lock);
4962         while (!list_empty(&cma_dev->id_list)) {
4963                 struct rdma_id_private *id_priv = list_first_entry(
4964                         &cma_dev->id_list, struct rdma_id_private, device_item);
4965
4966                 list_del_init(&id_priv->listen_item);
4967                 list_del_init(&id_priv->device_item);
4968                 cma_id_get(id_priv);
4969                 mutex_unlock(&lock);
4970
4971                 cma_send_device_removal_put(id_priv);
4972
4973                 mutex_lock(&lock);
4974         }
4975         mutex_unlock(&lock);
4976
4977         cma_dev_put(cma_dev);
4978         wait_for_completion(&cma_dev->comp);
4979 }
4980
4981 static bool cma_supported(struct ib_device *device)
4982 {
4983         u32 i;
4984
4985         rdma_for_each_port(device, i) {
4986                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
4987                         return true;
4988         }
4989         return false;
4990 }
4991
4992 static int cma_add_one(struct ib_device *device)
4993 {
4994         struct rdma_id_private *to_destroy;
4995         struct cma_device *cma_dev;
4996         struct rdma_id_private *id_priv;
4997         unsigned long supported_gids = 0;
4998         int ret;
4999         u32 i;
5000
5001         if (!cma_supported(device))
5002                 return -EOPNOTSUPP;
5003
5004         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5005         if (!cma_dev)
5006                 return -ENOMEM;
5007
5008         cma_dev->device = device;
5009         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5010                                             sizeof(*cma_dev->default_gid_type),
5011                                             GFP_KERNEL);
5012         if (!cma_dev->default_gid_type) {
5013                 ret = -ENOMEM;
5014                 goto free_cma_dev;
5015         }
5016
5017         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5018                                             sizeof(*cma_dev->default_roce_tos),
5019                                             GFP_KERNEL);
5020         if (!cma_dev->default_roce_tos) {
5021                 ret = -ENOMEM;
5022                 goto free_gid_type;
5023         }
5024
5025         rdma_for_each_port (device, i) {
5026                 supported_gids = roce_gid_type_mask_support(device, i);
5027                 WARN_ON(!supported_gids);
5028                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5029                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5030                                 CMA_PREFERRED_ROCE_GID_TYPE;
5031                 else
5032                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5033                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5034                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5035         }
5036
5037         init_completion(&cma_dev->comp);
5038         refcount_set(&cma_dev->refcount, 1);
5039         INIT_LIST_HEAD(&cma_dev->id_list);
5040         ib_set_client_data(device, &cma_client, cma_dev);
5041
5042         mutex_lock(&lock);
5043         list_add_tail(&cma_dev->list, &dev_list);
5044         list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5045                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5046                 if (ret)
5047                         goto free_listen;
5048         }
5049         mutex_unlock(&lock);
5050
5051         trace_cm_add_one(device);
5052         return 0;
5053
5054 free_listen:
5055         list_del(&cma_dev->list);
5056         mutex_unlock(&lock);
5057
5058         /* cma_process_remove() will delete to_destroy */
5059         cma_process_remove(cma_dev);
5060         kfree(cma_dev->default_roce_tos);
5061 free_gid_type:
5062         kfree(cma_dev->default_gid_type);
5063
5064 free_cma_dev:
5065         kfree(cma_dev);
5066         return ret;
5067 }
5068
5069 static void cma_remove_one(struct ib_device *device, void *client_data)
5070 {
5071         struct cma_device *cma_dev = client_data;
5072
5073         trace_cm_remove_one(device);
5074
5075         mutex_lock(&lock);
5076         list_del(&cma_dev->list);
5077         mutex_unlock(&lock);
5078
5079         cma_process_remove(cma_dev);
5080         kfree(cma_dev->default_roce_tos);
5081         kfree(cma_dev->default_gid_type);
5082         kfree(cma_dev);
5083 }
5084
5085 static int cma_init_net(struct net *net)
5086 {
5087         struct cma_pernet *pernet = cma_pernet(net);
5088
5089         xa_init(&pernet->tcp_ps);
5090         xa_init(&pernet->udp_ps);
5091         xa_init(&pernet->ipoib_ps);
5092         xa_init(&pernet->ib_ps);
5093
5094         return 0;
5095 }
5096
5097 static void cma_exit_net(struct net *net)
5098 {
5099         struct cma_pernet *pernet = cma_pernet(net);
5100
5101         WARN_ON(!xa_empty(&pernet->tcp_ps));
5102         WARN_ON(!xa_empty(&pernet->udp_ps));
5103         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5104         WARN_ON(!xa_empty(&pernet->ib_ps));
5105 }
5106
5107 static struct pernet_operations cma_pernet_operations = {
5108         .init = cma_init_net,
5109         .exit = cma_exit_net,
5110         .id = &cma_pernet_id,
5111         .size = sizeof(struct cma_pernet),
5112 };
5113
5114 static int __init cma_init(void)
5115 {
5116         int ret;
5117
5118         /*
5119          * There is a rare lock ordering dependency in cma_netdev_callback()
5120          * that only happens when bonding is enabled. Teach lockdep that rtnl
5121          * must never be nested under lock so it can find these without having
5122          * to test with bonding.
5123          */
5124         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5125                 rtnl_lock();
5126                 mutex_lock(&lock);
5127                 mutex_unlock(&lock);
5128                 rtnl_unlock();
5129         }
5130
5131         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5132         if (!cma_wq)
5133                 return -ENOMEM;
5134
5135         ret = register_pernet_subsys(&cma_pernet_operations);
5136         if (ret)
5137                 goto err_wq;
5138
5139         ib_sa_register_client(&sa_client);
5140         register_netdevice_notifier(&cma_nb);
5141
5142         ret = ib_register_client(&cma_client);
5143         if (ret)
5144                 goto err;
5145
5146         ret = cma_configfs_init();
5147         if (ret)
5148                 goto err_ib;
5149
5150         return 0;
5151
5152 err_ib:
5153         ib_unregister_client(&cma_client);
5154 err:
5155         unregister_netdevice_notifier(&cma_nb);
5156         ib_sa_unregister_client(&sa_client);
5157         unregister_pernet_subsys(&cma_pernet_operations);
5158 err_wq:
5159         destroy_workqueue(cma_wq);
5160         return ret;
5161 }
5162
5163 static void __exit cma_cleanup(void)
5164 {
5165         cma_configfs_exit();
5166         ib_unregister_client(&cma_client);
5167         unregister_netdevice_notifier(&cma_nb);
5168         ib_sa_unregister_client(&sa_client);
5169         unregister_pernet_subsys(&cma_pernet_operations);
5170         destroy_workqueue(cma_wq);
5171 }
5172
5173 module_init(cma_init);
5174 module_exit(cma_cleanup);