drivers/infiniband/core/cma.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   4  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   5  * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
   6  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   7  */
   8
   9 #include <linux/completion.h>
  10 #include <linux/in.h>
  11 #include <linux/in6.h>
  12 #include <linux/mutex.h>
  13 #include <linux/random.h>
  14 #include <linux/igmp.h>
  15 #include <linux/xarray.h>
  16 #include <linux/inetdevice.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <net/route.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/netns/generic.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/ip_fib.h>
  26 #include <net/ip6_route.h>
  27
  28 #include <rdma/rdma_cm.h>
  29 #include <rdma/rdma_cm_ib.h>
  30 #include <rdma/rdma_netlink.h>
  31 #include <rdma/ib.h>
  32 #include <rdma/ib_cache.h>
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_sa.h>
  35 #include <rdma/iw_cm.h>
  36
  37 #include "core_priv.h"
  38 #include "cma_priv.h"
  39 #include "cma_trace.h"
  40
  41 MODULE_AUTHOR("Sean Hefty");
  42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  43 MODULE_LICENSE("Dual BSD/GPL");
  44
  45 #define CMA_CM_RESPONSE_TIMEOUT 20
  46 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  47 #define CMA_MAX_CM_RETRIES 15
  48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  49 #define CMA_IBOE_PACKET_LIFETIME 18
  50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  51
  52 static const char * const cma_events[] = {
  53         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  54         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  55         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  56         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  57         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  58         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  59         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  60         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  61         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  62         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  63         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  64         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  65         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  66         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  67         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  68         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  69 };
  70
  71 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
  72                               enum ib_gid_type gid_type);
  73
  74 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  75 {
  76         size_t index = event;
  77
  78         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  79                         cma_events[index] : "unrecognized event";
  80 }
  81 EXPORT_SYMBOL(rdma_event_msg);
  82
  83 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  84                                                 int reason)
  85 {
  86         if (rdma_ib_or_roce(id->device, id->port_num))
  87                 return ibcm_reject_msg(reason);
  88
  89         if (rdma_protocol_iwarp(id->device, id->port_num))
  90                 return iwcm_reject_msg(reason);
  91
  92         WARN_ON_ONCE(1);
  93         return "unrecognized transport";
  94 }
  95 EXPORT_SYMBOL(rdma_reject_msg);
  96
  97 /**
  98  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  99  *                           request.
 100  * @id: Communication identifier that received the REJECT event.
 101  * @reason: Value returned in the REJECT event status field.
 102  */
 103 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 104 {
 105         if (rdma_ib_or_roce(id->device, id->port_num))
 106                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 107
 108         if (rdma_protocol_iwarp(id->device, id->port_num))
 109                 return reason == -ECONNREFUSED;
 110
 111         WARN_ON_ONCE(1);
 112         return false;
 113 }
 114
 115 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 116                                       struct rdma_cm_event *ev, u8 *data_len)
 117 {
 118         const void *p;
 119
 120         if (rdma_is_consumer_reject(id, ev->status)) {
 121                 *data_len = ev->param.conn.private_data_len;
 122                 p = ev->param.conn.private_data;
 123         } else {
 124                 *data_len = 0;
 125                 p = NULL;
 126         }
 127         return p;
 128 }
 129 EXPORT_SYMBOL(rdma_consumer_reject_data);
 130
 131 /**
 132  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 133  * @id: Communication Identifier
 134  */
 135 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 136 {
 137         struct rdma_id_private *id_priv;
 138
 139         id_priv = container_of(id, struct rdma_id_private, id);
 140         if (id->device->node_type == RDMA_NODE_RNIC)
 141                 return id_priv->cm_id.iw;
 142         return NULL;
 143 }
 144 EXPORT_SYMBOL(rdma_iw_cm_id);
 145
 146 /**
 147  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 148  * @res: rdma resource tracking entry pointer
 149  */
 150 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 151 {
 152         struct rdma_id_private *id_priv =
 153                 container_of(res, struct rdma_id_private, res);
 154
 155         return &id_priv->id;
 156 }
 157 EXPORT_SYMBOL(rdma_res_to_id);
 158
 159 static int cma_add_one(struct ib_device *device);
 160 static void cma_remove_one(struct ib_device *device, void *client_data);
 161
 162 static struct ib_client cma_client = {
 163         .name   = "cma",
 164         .add    = cma_add_one,
 165         .remove = cma_remove_one
 166 };
 167
 168 static struct ib_sa_client sa_client;
 169 static LIST_HEAD(dev_list);
 170 static LIST_HEAD(listen_any_list);
 171 static DEFINE_MUTEX(lock);
 172 static struct workqueue_struct *cma_wq;
 173 static unsigned int cma_pernet_id;
 174
 175 struct cma_pernet {
 176         struct xarray tcp_ps;
 177         struct xarray udp_ps;
 178         struct xarray ipoib_ps;
 179         struct xarray ib_ps;
 180 };
 181
 182 static struct cma_pernet *cma_pernet(struct net *net)
 183 {
 184         return net_generic(net, cma_pernet_id);
 185 }
 186
 187 static
 188 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 189 {
 190         struct cma_pernet *pernet = cma_pernet(net);
 191
 192         switch (ps) {
 193         case RDMA_PS_TCP:
 194                 return &pernet->tcp_ps;
 195         case RDMA_PS_UDP:
 196                 return &pernet->udp_ps;
 197         case RDMA_PS_IPOIB:
 198                 return &pernet->ipoib_ps;
 199         case RDMA_PS_IB:
 200                 return &pernet->ib_ps;
 201         default:
 202                 return NULL;
 203         }
 204 }
 205
 206 struct cma_device {
 207         struct list_head        list;
 208         struct ib_device        *device;
 209         struct completion       comp;
 210         refcount_t refcount;
 211         struct list_head        id_list;
 212         enum ib_gid_type        *default_gid_type;
 213         u8                      *default_roce_tos;
 214 };
 215
 216 struct rdma_bind_list {
 217         enum rdma_ucm_port_space ps;
 218         struct hlist_head       owners;
 219         unsigned short          port;
 220 };
 221
 222 struct class_port_info_context {
 223         struct ib_class_port_info       *class_port_info;
 224         struct ib_device                *device;
 225         struct completion               done;
 226         struct ib_sa_query              *sa_query;
 227         u8                              port_num;
 228 };
 229
 230 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 231                         struct rdma_bind_list *bind_list, int snum)
 232 {
 233         struct xarray *xa = cma_pernet_xa(net, ps);
 234
 235         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 236 }
 237
 238 static struct rdma_bind_list *cma_ps_find(struct net *net,
 239                                           enum rdma_ucm_port_space ps, int snum)
 240 {
 241         struct xarray *xa = cma_pernet_xa(net, ps);
 242
 243         return xa_load(xa, snum);
 244 }
 245
 246 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 247                           int snum)
 248 {
 249         struct xarray *xa = cma_pernet_xa(net, ps);
 250
 251         xa_erase(xa, snum);
 252 }
 253
 254 enum {
 255         CMA_OPTION_AFONLY,
 256 };
 257
 258 void cma_dev_get(struct cma_device *cma_dev)
 259 {
 260         refcount_inc(&cma_dev->refcount);
 261 }
 262
 263 void cma_dev_put(struct cma_device *cma_dev)
 264 {
 265         if (refcount_dec_and_test(&cma_dev->refcount))
 266                 complete(&cma_dev->comp);
 267 }
 268
 269 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 270                                              void               *cookie)
 271 {
 272         struct cma_device *cma_dev;
 273         struct cma_device *found_cma_dev = NULL;
 274
 275         mutex_lock(&lock);
 276
 277         list_for_each_entry(cma_dev, &dev_list, list)
 278                 if (filter(cma_dev->device, cookie)) {
 279                         found_cma_dev = cma_dev;
 280                         break;
 281                 }
 282
 283         if (found_cma_dev)
 284                 cma_dev_get(found_cma_dev);
 285         mutex_unlock(&lock);
 286         return found_cma_dev;
 287 }
 288
 289 int cma_get_default_gid_type(struct cma_device *cma_dev,
 290                              unsigned int port)
 291 {
 292         if (!rdma_is_port_valid(cma_dev->device, port))
 293                 return -EINVAL;
 294
 295         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 296 }
 297
 298 int cma_set_default_gid_type(struct cma_device *cma_dev,
 299                              unsigned int port,
 300                              enum ib_gid_type default_gid_type)
 301 {
 302         unsigned long supported_gids;
 303
 304         if (!rdma_is_port_valid(cma_dev->device, port))
 305                 return -EINVAL;
 306
 307         if (default_gid_type == IB_GID_TYPE_IB &&
 308             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 309                 default_gid_type = IB_GID_TYPE_ROCE;
 310
 311         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 312
 313         if (!(supported_gids & 1 << default_gid_type))
 314                 return -EINVAL;
 315
 316         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 317                 default_gid_type;
 318
 319         return 0;
 320 }
 321
 322 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 323 {
 324         if (!rdma_is_port_valid(cma_dev->device, port))
 325                 return -EINVAL;
 326
 327         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 328 }
 329
 330 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 331                              u8 default_roce_tos)
 332 {
 333         if (!rdma_is_port_valid(cma_dev->device, port))
 334                 return -EINVAL;
 335
 336         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 337                  default_roce_tos;
 338
 339         return 0;
 340 }
 341 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 342 {
 343         return cma_dev->device;
 344 }
 345
 346 /*
 347  * Device removal can occur at anytime, so we need extra handling to
 348  * serialize notifying the user of device removal with other callbacks.
 349  * We do this by disabling removal notification while a callback is in process,
 350  * and reporting it after the callback completes.
 351  */
 352
 353 struct cma_multicast {
 354         struct rdma_id_private *id_priv;
 355         union {
 356                 struct ib_sa_multicast *sa_mc;
 357                 struct {
 358                         struct work_struct work;
 359                         struct rdma_cm_event event;
 360                 } iboe_join;
 361         };
 362         struct list_head        list;
 363         void                    *context;
 364         struct sockaddr_storage addr;
 365         u8                      join_state;
 366 };
 367
 368 struct cma_work {
 369         struct work_struct      work;
 370         struct rdma_id_private  *id;
 371         enum rdma_cm_state      old_state;
 372         enum rdma_cm_state      new_state;
 373         struct rdma_cm_event    event;
 374 };
 375
 376 union cma_ip_addr {
 377         struct in6_addr ip6;
 378         struct {
 379                 __be32 pad[3];
 380                 __be32 addr;
 381         } ip4;
 382 };
 383
 384 struct cma_hdr {
 385         u8 cma_version;
 386         u8 ip_version;  /* IP version: 7:4 */
 387         __be16 port;
 388         union cma_ip_addr src_addr;
 389         union cma_ip_addr dst_addr;
 390 };
 391
 392 #define CMA_VERSION 0x00
 393
 394 struct cma_req_info {
 395         struct sockaddr_storage listen_addr_storage;
 396         struct sockaddr_storage src_addr_storage;
 397         struct ib_device *device;
 398         union ib_gid local_gid;
 399         __be64 service_id;
 400         int port;
 401         bool has_gid;
 402         u16 pkey;
 403 };
 404
 405 static int cma_comp_exch(struct rdma_id_private *id_priv,
 406                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 407 {
 408         unsigned long flags;
 409         int ret;
 410
 411         /*
 412          * The FSM uses a funny double locking where state is protected by both
 413          * the handler_mutex and the spinlock. State is not allowed to change
 414          * to/from a handler_mutex protected value without also holding
 415          * handler_mutex.
 416          */
 417         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 418                 lockdep_assert_held(&id_priv->handler_mutex);
 419
 420         spin_lock_irqsave(&id_priv->lock, flags);
 421         if ((ret = (id_priv->state == comp)))
 422                 id_priv->state = exch;
 423         spin_unlock_irqrestore(&id_priv->lock, flags);
 424         return ret;
 425 }
 426
 427 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 428 {
 429         return hdr->ip_version >> 4;
 430 }
 431
 432 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 433 {
 434         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 435 }
 436
 437 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 438 {
 439         struct in_device *in_dev = NULL;
 440
 441         if (ndev) {
 442                 rtnl_lock();
 443                 in_dev = __in_dev_get_rtnl(ndev);
 444                 if (in_dev) {
 445                         if (join)
 446                                 ip_mc_inc_group(in_dev,
 447                                                 *(__be32 *)(mgid->raw + 12));
 448                         else
 449                                 ip_mc_dec_group(in_dev,
 450                                                 *(__be32 *)(mgid->raw + 12));
 451                 }
 452                 rtnl_unlock();
 453         }
 454         return (in_dev) ? 0 : -ENODEV;
 455 }
 456
 457 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 458                                struct cma_device *cma_dev)
 459 {
 460         cma_dev_get(cma_dev);
 461         id_priv->cma_dev = cma_dev;
 462         id_priv->id.device = cma_dev->device;
 463         id_priv->id.route.addr.dev_addr.transport =
 464                 rdma_node_get_transport(cma_dev->device->node_type);
 465         list_add_tail(&id_priv->list, &cma_dev->id_list);
 466
 467         trace_cm_id_attach(id_priv, cma_dev->device);
 468 }
 469
 470 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 471                               struct cma_device *cma_dev)
 472 {
 473         _cma_attach_to_dev(id_priv, cma_dev);
 474         id_priv->gid_type =
 475                 cma_dev->default_gid_type[id_priv->id.port_num -
 476                                           rdma_start_port(cma_dev->device)];
 477 }
 478
 479 static void cma_release_dev(struct rdma_id_private *id_priv)
 480 {
 481         mutex_lock(&lock);
 482         list_del(&id_priv->list);
 483         cma_dev_put(id_priv->cma_dev);
 484         id_priv->cma_dev = NULL;
 485         id_priv->id.device = NULL;
 486         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 487                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 488                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 489         }
 490         mutex_unlock(&lock);
 491 }
 492
 493 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 494 {
 495         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 496 }
 497
 498 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 499 {
 500         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 501 }
 502
 503 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 504 {
 505         return id_priv->id.route.addr.src_addr.ss_family;
 506 }
 507
 508 static int cma_set_default_qkey(struct rdma_id_private *id_priv)
 509 {
 510         struct ib_sa_mcmember_rec rec;
 511         int ret = 0;
 512
 513         switch (id_priv->id.ps) {
 514         case RDMA_PS_UDP:
 515         case RDMA_PS_IB:
 516                 id_priv->qkey = RDMA_UDP_QKEY;
 517                 break;
 518         case RDMA_PS_IPOIB:
 519                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 520                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 521                                              id_priv->id.port_num, &rec.mgid,
 522                                              &rec);
 523                 if (!ret)
 524                         id_priv->qkey = be32_to_cpu(rec.qkey);
 525                 break;
 526         default:
 527                 break;
 528         }
 529         return ret;
 530 }
 531
 532 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 533 {
 534         if (!qkey ||
 535             (id_priv->qkey && (id_priv->qkey != qkey)))
 536                 return -EINVAL;
 537
 538         id_priv->qkey = qkey;
 539         return 0;
 540 }
 541
 542 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 543 {
 544         dev_addr->dev_type = ARPHRD_INFINIBAND;
 545         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 546         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 547 }
 548
 549 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 550 {
 551         int ret;
 552
 553         if (addr->sa_family != AF_IB) {
 554                 ret = rdma_translate_ip(addr, dev_addr);
 555         } else {
 556                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 557                 ret = 0;
 558         }
 559
 560         return ret;
 561 }
 562
 563 static const struct ib_gid_attr *
 564 cma_validate_port(struct ib_device *device, u8 port,
 565                   enum ib_gid_type gid_type,
 566                   union ib_gid *gid,
 567                   struct rdma_id_private *id_priv)
 568 {
 569         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 570         int bound_if_index = dev_addr->bound_dev_if;
 571         const struct ib_gid_attr *sgid_attr;
 572         int dev_type = dev_addr->dev_type;
 573         struct net_device *ndev = NULL;
 574
 575         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 576                 return ERR_PTR(-ENODEV);
 577
 578         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 579                 return ERR_PTR(-ENODEV);
 580
 581         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 582                 return ERR_PTR(-ENODEV);
 583
 584         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 585                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 586                 if (!ndev)
 587                         return ERR_PTR(-ENODEV);
 588         } else {
 589                 gid_type = IB_GID_TYPE_IB;
 590         }
 591
 592         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 593         if (ndev)
 594                 dev_put(ndev);
 595         return sgid_attr;
 596 }
 597
 598 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 599                                const struct ib_gid_attr *sgid_attr)
 600 {
 601         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 602         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 603 }
 604
 605 /**
 606  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 607  * based on source ip address.
 608  * @id_priv:    cm_id which should be bound to cma device
 609  *
 610  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 611  * based on source IP address. It returns 0 on success or error code otherwise.
 612  * It is applicable to active and passive side cm_id.
 613  */
 614 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 615 {
 616         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 617         const struct ib_gid_attr *sgid_attr;
 618         union ib_gid gid, iboe_gid, *gidp;
 619         struct cma_device *cma_dev;
 620         enum ib_gid_type gid_type;
 621         int ret = -ENODEV;
 622         unsigned int port;
 623
 624         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 625             id_priv->id.ps == RDMA_PS_IPOIB)
 626                 return -EINVAL;
 627
 628         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 629                     &iboe_gid);
 630
 631         memcpy(&gid, dev_addr->src_dev_addr +
 632                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 633
 634         mutex_lock(&lock);
 635         list_for_each_entry(cma_dev, &dev_list, list) {
 636                 rdma_for_each_port (cma_dev->device, port) {
 637                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 638                                &iboe_gid : &gid;
 639                         gid_type = cma_dev->default_gid_type[port - 1];
 640                         sgid_attr = cma_validate_port(cma_dev->device, port,
 641                                                       gid_type, gidp, id_priv);
 642                         if (!IS_ERR(sgid_attr)) {
 643                                 id_priv->id.port_num = port;
 644                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 645                                 cma_attach_to_dev(id_priv, cma_dev);
 646                                 ret = 0;
 647                                 goto out;
 648                         }
 649                 }
 650         }
 651 out:
 652         mutex_unlock(&lock);
 653         return ret;
 654 }
 655
 656 /**
 657  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 658  * @id_priv:            cm id to bind to cma device
 659  * @listen_id_priv:     listener cm id to match against
 660  * @req:                Pointer to req structure containaining incoming
 661  *                      request information
 662  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 663  * rdma device matches for listen_id and incoming request. It also verifies
 664  * that a GID table entry is present for the source address.
 665  * Returns 0 on success, or returns error code otherwise.
 666  */
 667 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 668                               const struct rdma_id_private *listen_id_priv,
 669                               struct cma_req_info *req)
 670 {
 671         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 672         const struct ib_gid_attr *sgid_attr;
 673         enum ib_gid_type gid_type;
 674         union ib_gid gid;
 675
 676         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 677             id_priv->id.ps == RDMA_PS_IPOIB)
 678                 return -EINVAL;
 679
 680         if (rdma_protocol_roce(req->device, req->port))
 681                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 682                             &gid);
 683         else
 684                 memcpy(&gid, dev_addr->src_dev_addr +
 685                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 686
 687         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 688         sgid_attr = cma_validate_port(req->device, req->port,
 689                                       gid_type, &gid, id_priv);
 690         if (IS_ERR(sgid_attr))
 691                 return PTR_ERR(sgid_attr);
 692
 693         id_priv->id.port_num = req->port;
 694         cma_bind_sgid_attr(id_priv, sgid_attr);
 695         /* Need to acquire lock to protect against reader
 696          * of cma_dev->id_list such as cma_netdev_callback() and
 697          * cma_process_remove().
 698          */
 699         mutex_lock(&lock);
 700         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 701         mutex_unlock(&lock);
 702         rdma_restrack_add(&id_priv->res);
 703         return 0;
 704 }
 705
 706 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 707                               const struct rdma_id_private *listen_id_priv)
 708 {
 709         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 710         const struct ib_gid_attr *sgid_attr;
 711         struct cma_device *cma_dev;
 712         enum ib_gid_type gid_type;
 713         int ret = -ENODEV;
 714         unsigned int port;
 715         union ib_gid gid;
 716
 717         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 718             id_priv->id.ps == RDMA_PS_IPOIB)
 719                 return -EINVAL;
 720
 721         memcpy(&gid, dev_addr->src_dev_addr +
 722                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 723
 724         mutex_lock(&lock);
 725
 726         cma_dev = listen_id_priv->cma_dev;
 727         port = listen_id_priv->id.port_num;
 728         gid_type = listen_id_priv->gid_type;
 729         sgid_attr = cma_validate_port(cma_dev->device, port,
 730                                       gid_type, &gid, id_priv);
 731         if (!IS_ERR(sgid_attr)) {
 732                 id_priv->id.port_num = port;
 733                 cma_bind_sgid_attr(id_priv, sgid_attr);
 734                 ret = 0;
 735                 goto out;
 736         }
 737
 738         list_for_each_entry(cma_dev, &dev_list, list) {
 739                 rdma_for_each_port (cma_dev->device, port) {
 740                         if (listen_id_priv->cma_dev == cma_dev &&
 741                             listen_id_priv->id.port_num == port)
 742                                 continue;
 743
 744                         gid_type = cma_dev->default_gid_type[port - 1];
 745                         sgid_attr = cma_validate_port(cma_dev->device, port,
 746                                                       gid_type, &gid, id_priv);
 747                         if (!IS_ERR(sgid_attr)) {
 748                                 id_priv->id.port_num = port;
 749                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 750                                 ret = 0;
 751                                 goto out;
 752                         }
 753                 }
 754         }
 755
 756 out:
 757         if (!ret) {
 758                 cma_attach_to_dev(id_priv, cma_dev);
 759                 rdma_restrack_add(&id_priv->res);
 760         }
 761
 762         mutex_unlock(&lock);
 763         return ret;
 764 }
 765
 766 /*
 767  * Select the source IB device and address to reach the destination IB address.
 768  */
 769 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 770 {
 771         struct cma_device *cma_dev, *cur_dev;
 772         struct sockaddr_ib *addr;
 773         union ib_gid gid, sgid, *dgid;
 774         unsigned int p;
 775         u16 pkey, index;
 776         enum ib_port_state port_state;
 777         int ret;
 778         int i;
 779
 780         cma_dev = NULL;
 781         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 782         dgid = (union ib_gid *) &addr->sib_addr;
 783         pkey = ntohs(addr->sib_pkey);
 784
 785         mutex_lock(&lock);
 786         list_for_each_entry(cur_dev, &dev_list, list) {
 787                 rdma_for_each_port (cur_dev->device, p) {
 788                         if (!rdma_cap_af_ib(cur_dev->device, p))
 789                                 continue;
 790
 791                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 792                                 continue;
 793
 794                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 795                                 continue;
 796
 797                         for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
 798                              ++i) {
 799                                 ret = rdma_query_gid(cur_dev->device, p, i,
 800                                                      &gid);
 801                                 if (ret)
 802                                         continue;
 803
 804                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 805                                         cma_dev = cur_dev;
 806                                         sgid = gid;
 807                                         id_priv->id.port_num = p;
 808                                         goto found;
 809                                 }
 810
 811                                 if (!cma_dev && (gid.global.subnet_prefix ==
 812                                     dgid->global.subnet_prefix) &&
 813                                     port_state == IB_PORT_ACTIVE) {
 814                                         cma_dev = cur_dev;
 815                                         sgid = gid;
 816                                         id_priv->id.port_num = p;
 817                                         goto found;
 818                                 }
 819                         }
 820                 }
 821         }
 822         mutex_unlock(&lock);
 823         return -ENODEV;
 824
 825 found:
 826         cma_attach_to_dev(id_priv, cma_dev);
 827         rdma_restrack_add(&id_priv->res);
 828         mutex_unlock(&lock);
 829         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 830         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 831         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 832         return 0;
 833 }
 834
 835 static void cma_id_get(struct rdma_id_private *id_priv)
 836 {
 837         refcount_inc(&id_priv->refcount);
 838 }
 839
 840 static void cma_id_put(struct rdma_id_private *id_priv)
 841 {
 842         if (refcount_dec_and_test(&id_priv->refcount))
 843                 complete(&id_priv->comp);
 844 }
 845
 846 static struct rdma_id_private *
 847 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 848                  void *context, enum rdma_ucm_port_space ps,
 849                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 850 {
 851         struct rdma_id_private *id_priv;
 852
 853         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 854         if (!id_priv)
 855                 return ERR_PTR(-ENOMEM);
 856
 857         id_priv->state = RDMA_CM_IDLE;
 858         id_priv->id.context = context;
 859         id_priv->id.event_handler = event_handler;
 860         id_priv->id.ps = ps;
 861         id_priv->id.qp_type = qp_type;
 862         id_priv->tos_set = false;
 863         id_priv->timeout_set = false;
 864         id_priv->gid_type = IB_GID_TYPE_IB;
 865         spin_lock_init(&id_priv->lock);
 866         mutex_init(&id_priv->qp_mutex);
 867         init_completion(&id_priv->comp);
 868         refcount_set(&id_priv->refcount, 1);
 869         mutex_init(&id_priv->handler_mutex);
 870         INIT_LIST_HEAD(&id_priv->listen_list);
 871         INIT_LIST_HEAD(&id_priv->mc_list);
 872         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 873         id_priv->id.route.addr.dev_addr.net = get_net(net);
 874         id_priv->seq_num &= 0x00ffffff;
 875
 876         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 877         if (parent)
 878                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 879
 880         return id_priv;
 881 }
 882
 883 struct rdma_cm_id *
 884 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
 885                         void *context, enum rdma_ucm_port_space ps,
 886                         enum ib_qp_type qp_type, const char *caller)
 887 {
 888         struct rdma_id_private *ret;
 889
 890         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
 891         if (IS_ERR(ret))
 892                 return ERR_CAST(ret);
 893
 894         rdma_restrack_set_name(&ret->res, caller);
 895         return &ret->id;
 896 }
 897 EXPORT_SYMBOL(__rdma_create_kernel_id);
 898
 899 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
 900                                        void *context,
 901                                        enum rdma_ucm_port_space ps,
 902                                        enum ib_qp_type qp_type)
 903 {
 904         struct rdma_id_private *ret;
 905
 906         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
 907                                ps, qp_type, NULL);
 908         if (IS_ERR(ret))
 909                 return ERR_CAST(ret);
 910
 911         rdma_restrack_set_name(&ret->res, NULL);
 912         return &ret->id;
 913 }
 914 EXPORT_SYMBOL(rdma_create_user_id);
 915
 916 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 917 {
 918         struct ib_qp_attr qp_attr;
 919         int qp_attr_mask, ret;
 920
 921         qp_attr.qp_state = IB_QPS_INIT;
 922         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 923         if (ret)
 924                 return ret;
 925
 926         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 927         if (ret)
 928                 return ret;
 929
 930         qp_attr.qp_state = IB_QPS_RTR;
 931         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 932         if (ret)
 933                 return ret;
 934
 935         qp_attr.qp_state = IB_QPS_RTS;
 936         qp_attr.sq_psn = 0;
 937         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 938
 939         return ret;
 940 }
 941
 942 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 943 {
 944         struct ib_qp_attr qp_attr;
 945         int qp_attr_mask, ret;
 946
 947         qp_attr.qp_state = IB_QPS_INIT;
 948         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 949         if (ret)
 950                 return ret;
 951
 952         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 953 }
 954
 955 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 956                    struct ib_qp_init_attr *qp_init_attr)
 957 {
 958         struct rdma_id_private *id_priv;
 959         struct ib_qp *qp;
 960         int ret;
 961
 962         id_priv = container_of(id, struct rdma_id_private, id);
 963         if (id->device != pd->device) {
 964                 ret = -EINVAL;
 965                 goto out_err;
 966         }
 967
 968         qp_init_attr->port_num = id->port_num;
 969         qp = ib_create_qp(pd, qp_init_attr);
 970         if (IS_ERR(qp)) {
 971                 ret = PTR_ERR(qp);
 972                 goto out_err;
 973         }
 974
 975         if (id->qp_type == IB_QPT_UD)
 976                 ret = cma_init_ud_qp(id_priv, qp);
 977         else
 978                 ret = cma_init_conn_qp(id_priv, qp);
 979         if (ret)
 980                 goto out_destroy;
 981
 982         id->qp = qp;
 983         id_priv->qp_num = qp->qp_num;
 984         id_priv->srq = (qp->srq != NULL);
 985         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 986         return 0;
 987 out_destroy:
 988         ib_destroy_qp(qp);
 989 out_err:
 990         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 991         return ret;
 992 }
 993 EXPORT_SYMBOL(rdma_create_qp);
 994
 995 void rdma_destroy_qp(struct rdma_cm_id *id)
 996 {
 997         struct rdma_id_private *id_priv;
 998
 999         id_priv = container_of(id, struct rdma_id_private, id);
1000         trace_cm_qp_destroy(id_priv);
1001         mutex_lock(&id_priv->qp_mutex);
1002         ib_destroy_qp(id_priv->id.qp);
1003         id_priv->id.qp = NULL;
1004         mutex_unlock(&id_priv->qp_mutex);
1005 }
1006 EXPORT_SYMBOL(rdma_destroy_qp);
1007
1008 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1009                              struct rdma_conn_param *conn_param)
1010 {
1011         struct ib_qp_attr qp_attr;
1012         int qp_attr_mask, ret;
1013
1014         mutex_lock(&id_priv->qp_mutex);
1015         if (!id_priv->id.qp) {
1016                 ret = 0;
1017                 goto out;
1018         }
1019
1020         /* Need to update QP attributes from default values. */
1021         qp_attr.qp_state = IB_QPS_INIT;
1022         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1023         if (ret)
1024                 goto out;
1025
1026         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1027         if (ret)
1028                 goto out;
1029
1030         qp_attr.qp_state = IB_QPS_RTR;
1031         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1032         if (ret)
1033                 goto out;
1034
1035         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1036
1037         if (conn_param)
1038                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1039         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1040 out:
1041         mutex_unlock(&id_priv->qp_mutex);
1042         return ret;
1043 }
1044
1045 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1046                              struct rdma_conn_param *conn_param)
1047 {
1048         struct ib_qp_attr qp_attr;
1049         int qp_attr_mask, ret;
1050
1051         mutex_lock(&id_priv->qp_mutex);
1052         if (!id_priv->id.qp) {
1053                 ret = 0;
1054                 goto out;
1055         }
1056
1057         qp_attr.qp_state = IB_QPS_RTS;
1058         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1059         if (ret)
1060                 goto out;
1061
1062         if (conn_param)
1063                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1064         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1065 out:
1066         mutex_unlock(&id_priv->qp_mutex);
1067         return ret;
1068 }
1069
1070 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1071 {
1072         struct ib_qp_attr qp_attr;
1073         int ret;
1074
1075         mutex_lock(&id_priv->qp_mutex);
1076         if (!id_priv->id.qp) {
1077                 ret = 0;
1078                 goto out;
1079         }
1080
1081         qp_attr.qp_state = IB_QPS_ERR;
1082         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1083 out:
1084         mutex_unlock(&id_priv->qp_mutex);
1085         return ret;
1086 }
1087
1088 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1089                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1090 {
1091         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1092         int ret;
1093         u16 pkey;
1094
1095         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1096                 pkey = 0xffff;
1097         else
1098                 pkey = ib_addr_get_pkey(dev_addr);
1099
1100         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1101                                   pkey, &qp_attr->pkey_index);
1102         if (ret)
1103                 return ret;
1104
1105         qp_attr->port_num = id_priv->id.port_num;
1106         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1107
1108         if (id_priv->id.qp_type == IB_QPT_UD) {
1109                 ret = cma_set_default_qkey(id_priv);
1110                 if (ret)
1111                         return ret;
1112
1113                 qp_attr->qkey = id_priv->qkey;
1114                 *qp_attr_mask |= IB_QP_QKEY;
1115         } else {
1116                 qp_attr->qp_access_flags = 0;
1117                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1118         }
1119         return 0;
1120 }
1121
1122 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1123                        int *qp_attr_mask)
1124 {
1125         struct rdma_id_private *id_priv;
1126         int ret = 0;
1127
1128         id_priv = container_of(id, struct rdma_id_private, id);
1129         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1130                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1131                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1132                 else
1133                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1134                                                  qp_attr_mask);
1135
1136                 if (qp_attr->qp_state == IB_QPS_RTR)
1137                         qp_attr->rq_psn = id_priv->seq_num;
1138         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1139                 if (!id_priv->cm_id.iw) {
1140                         qp_attr->qp_access_flags = 0;
1141                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1142                 } else
1143                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1144                                                  qp_attr_mask);
1145                 qp_attr->port_num = id_priv->id.port_num;
1146                 *qp_attr_mask |= IB_QP_PORT;
1147         } else
1148                 ret = -ENOSYS;
1149
1150         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1151                 qp_attr->timeout = id_priv->timeout;
1152
1153         return ret;
1154 }
1155 EXPORT_SYMBOL(rdma_init_qp_attr);
1156
1157 static inline bool cma_zero_addr(const struct sockaddr *addr)
1158 {
1159         switch (addr->sa_family) {
1160         case AF_INET:
1161                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1162         case AF_INET6:
1163                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1164         case AF_IB:
1165                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1166         default:
1167                 return false;
1168         }
1169 }
1170
1171 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1172 {
1173         switch (addr->sa_family) {
1174         case AF_INET:
1175                 return ipv4_is_loopback(
1176                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1177         case AF_INET6:
1178                 return ipv6_addr_loopback(
1179                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1180         case AF_IB:
1181                 return ib_addr_loopback(
1182                         &((struct sockaddr_ib *)addr)->sib_addr);
1183         default:
1184                 return false;
1185         }
1186 }
1187
1188 static inline bool cma_any_addr(const struct sockaddr *addr)
1189 {
1190         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1191 }
1192
1193 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1194 {
1195         if (src->sa_family != dst->sa_family)
1196                 return -1;
1197
1198         switch (src->sa_family) {
1199         case AF_INET:
1200                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1201                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1202         case AF_INET6: {
1203                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1204                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1205                 bool link_local;
1206
1207                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1208                                           &dst_addr6->sin6_addr))
1209                         return 1;
1210                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1211                              IPV6_ADDR_LINKLOCAL;
1212                 /* Link local must match their scope_ids */
1213                 return link_local ? (src_addr6->sin6_scope_id !=
1214                                      dst_addr6->sin6_scope_id) :
1215                                     0;
1216         }
1217
1218         default:
1219                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1220                                    &((struct sockaddr_ib *) dst)->sib_addr);
1221         }
1222 }
1223
1224 static __be16 cma_port(const struct sockaddr *addr)
1225 {
1226         struct sockaddr_ib *sib;
1227
1228         switch (addr->sa_family) {
1229         case AF_INET:
1230                 return ((struct sockaddr_in *) addr)->sin_port;
1231         case AF_INET6:
1232                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1233         case AF_IB:
1234                 sib = (struct sockaddr_ib *) addr;
1235                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1236                                     be64_to_cpu(sib->sib_sid_mask)));
1237         default:
1238                 return 0;
1239         }
1240 }
1241
1242 static inline int cma_any_port(const struct sockaddr *addr)
1243 {
1244         return !cma_port(addr);
1245 }
1246
1247 static void cma_save_ib_info(struct sockaddr *src_addr,
1248                              struct sockaddr *dst_addr,
1249                              const struct rdma_cm_id *listen_id,
1250                              const struct sa_path_rec *path)
1251 {
1252         struct sockaddr_ib *listen_ib, *ib;
1253
1254         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1255         if (src_addr) {
1256                 ib = (struct sockaddr_ib *)src_addr;
1257                 ib->sib_family = AF_IB;
1258                 if (path) {
1259                         ib->sib_pkey = path->pkey;
1260                         ib->sib_flowinfo = path->flow_label;
1261                         memcpy(&ib->sib_addr, &path->sgid, 16);
1262                         ib->sib_sid = path->service_id;
1263                         ib->sib_scope_id = 0;
1264                 } else {
1265                         ib->sib_pkey = listen_ib->sib_pkey;
1266                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1267                         ib->sib_addr = listen_ib->sib_addr;
1268                         ib->sib_sid = listen_ib->sib_sid;
1269                         ib->sib_scope_id = listen_ib->sib_scope_id;
1270                 }
1271                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1272         }
1273         if (dst_addr) {
1274                 ib = (struct sockaddr_ib *)dst_addr;
1275                 ib->sib_family = AF_IB;
1276                 if (path) {
1277                         ib->sib_pkey = path->pkey;
1278                         ib->sib_flowinfo = path->flow_label;
1279                         memcpy(&ib->sib_addr, &path->dgid, 16);
1280                 }
1281         }
1282 }
1283
1284 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1285                               struct sockaddr_in *dst_addr,
1286                               struct cma_hdr *hdr,
1287                               __be16 local_port)
1288 {
1289         if (src_addr) {
1290                 *src_addr = (struct sockaddr_in) {
1291                         .sin_family = AF_INET,
1292                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1293                         .sin_port = local_port,
1294                 };
1295         }
1296
1297         if (dst_addr) {
1298                 *dst_addr = (struct sockaddr_in) {
1299                         .sin_family = AF_INET,
1300                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1301                         .sin_port = hdr->port,
1302                 };
1303         }
1304 }
1305
1306 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1307                               struct sockaddr_in6 *dst_addr,
1308                               struct cma_hdr *hdr,
1309                               __be16 local_port)
1310 {
1311         if (src_addr) {
1312                 *src_addr = (struct sockaddr_in6) {
1313                         .sin6_family = AF_INET6,
1314                         .sin6_addr = hdr->dst_addr.ip6,
1315                         .sin6_port = local_port,
1316                 };
1317         }
1318
1319         if (dst_addr) {
1320                 *dst_addr = (struct sockaddr_in6) {
1321                         .sin6_family = AF_INET6,
1322                         .sin6_addr = hdr->src_addr.ip6,
1323                         .sin6_port = hdr->port,
1324                 };
1325         }
1326 }
1327
1328 static u16 cma_port_from_service_id(__be64 service_id)
1329 {
1330         return (u16)be64_to_cpu(service_id);
1331 }
1332
1333 static int cma_save_ip_info(struct sockaddr *src_addr,
1334                             struct sockaddr *dst_addr,
1335                             const struct ib_cm_event *ib_event,
1336                             __be64 service_id)
1337 {
1338         struct cma_hdr *hdr;
1339         __be16 port;
1340
1341         hdr = ib_event->private_data;
1342         if (hdr->cma_version != CMA_VERSION)
1343                 return -EINVAL;
1344
1345         port = htons(cma_port_from_service_id(service_id));
1346
1347         switch (cma_get_ip_ver(hdr)) {
1348         case 4:
1349                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1350                                   (struct sockaddr_in *)dst_addr, hdr, port);
1351                 break;
1352         case 6:
1353                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1354                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1355                 break;
1356         default:
1357                 return -EAFNOSUPPORT;
1358         }
1359
1360         return 0;
1361 }
1362
1363 static int cma_save_net_info(struct sockaddr *src_addr,
1364                              struct sockaddr *dst_addr,
1365                              const struct rdma_cm_id *listen_id,
1366                              const struct ib_cm_event *ib_event,
1367                              sa_family_t sa_family, __be64 service_id)
1368 {
1369         if (sa_family == AF_IB) {
1370                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1371                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1372                                          ib_event->param.req_rcvd.primary_path);
1373                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1374                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1375                 return 0;
1376         }
1377
1378         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1379 }
1380
1381 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1382                              struct cma_req_info *req)
1383 {
1384         const struct ib_cm_req_event_param *req_param =
1385                 &ib_event->param.req_rcvd;
1386         const struct ib_cm_sidr_req_event_param *sidr_param =
1387                 &ib_event->param.sidr_req_rcvd;
1388
1389         switch (ib_event->event) {
1390         case IB_CM_REQ_RECEIVED:
1391                 req->device     = req_param->listen_id->device;
1392                 req->port       = req_param->port;
1393                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1394                        sizeof(req->local_gid));
1395                 req->has_gid    = true;
1396                 req->service_id = req_param->primary_path->service_id;
1397                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1398                 if (req->pkey != req_param->bth_pkey)
1399                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1400                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1401                                             req_param->bth_pkey, req->pkey);
1402                 break;
1403         case IB_CM_SIDR_REQ_RECEIVED:
1404                 req->device     = sidr_param->listen_id->device;
1405                 req->port       = sidr_param->port;
1406                 req->has_gid    = false;
1407                 req->service_id = sidr_param->service_id;
1408                 req->pkey       = sidr_param->pkey;
1409                 if (req->pkey != sidr_param->bth_pkey)
1410                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1411                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1412                                             sidr_param->bth_pkey, req->pkey);
1413                 break;
1414         default:
1415                 return -EINVAL;
1416         }
1417
1418         return 0;
1419 }
1420
1421 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1422                                   const struct sockaddr_in *dst_addr,
1423                                   const struct sockaddr_in *src_addr)
1424 {
1425         __be32 daddr = dst_addr->sin_addr.s_addr,
1426                saddr = src_addr->sin_addr.s_addr;
1427         struct fib_result res;
1428         struct flowi4 fl4;
1429         int err;
1430         bool ret;
1431
1432         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1433             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1434             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1435             ipv4_is_loopback(saddr))
1436                 return false;
1437
1438         memset(&fl4, 0, sizeof(fl4));
1439         fl4.flowi4_oif = net_dev->ifindex;
1440         fl4.daddr = daddr;
1441         fl4.saddr = saddr;
1442
1443         rcu_read_lock();
1444         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1445         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1446         rcu_read_unlock();
1447
1448         return ret;
1449 }
1450
1451 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1452                                   const struct sockaddr_in6 *dst_addr,
1453                                   const struct sockaddr_in6 *src_addr)
1454 {
1455 #if IS_ENABLED(CONFIG_IPV6)
1456         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1457                            IPV6_ADDR_LINKLOCAL;
1458         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1459                                          &src_addr->sin6_addr, net_dev->ifindex,
1460                                          NULL, strict);
1461         bool ret;
1462
1463         if (!rt)
1464                 return false;
1465
1466         ret = rt->rt6i_idev->dev == net_dev;
1467         ip6_rt_put(rt);
1468
1469         return ret;
1470 #else
1471         return false;
1472 #endif
1473 }
1474
1475 static bool validate_net_dev(struct net_device *net_dev,
1476                              const struct sockaddr *daddr,
1477                              const struct sockaddr *saddr)
1478 {
1479         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1480         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1481         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1482         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1483
1484         switch (daddr->sa_family) {
1485         case AF_INET:
1486                 return saddr->sa_family == AF_INET &&
1487                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1488
1489         case AF_INET6:
1490                 return saddr->sa_family == AF_INET6 &&
1491                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1492
1493         default:
1494                 return false;
1495         }
1496 }
1497
1498 static struct net_device *
1499 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1500 {
1501         const struct ib_gid_attr *sgid_attr = NULL;
1502         struct net_device *ndev;
1503
1504         if (ib_event->event == IB_CM_REQ_RECEIVED)
1505                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1506         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1507                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1508
1509         if (!sgid_attr)
1510                 return NULL;
1511
1512         rcu_read_lock();
1513         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1514         if (IS_ERR(ndev))
1515                 ndev = NULL;
1516         else
1517                 dev_hold(ndev);
1518         rcu_read_unlock();
1519         return ndev;
1520 }
1521
1522 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1523                                           struct cma_req_info *req)
1524 {
1525         struct sockaddr *listen_addr =
1526                         (struct sockaddr *)&req->listen_addr_storage;
1527         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1528         struct net_device *net_dev;
1529         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1530         int err;
1531
1532         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1533                                req->service_id);
1534         if (err)
1535                 return ERR_PTR(err);
1536
1537         if (rdma_protocol_roce(req->device, req->port))
1538                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1539         else
1540                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1541                                                    req->pkey,
1542                                                    gid, listen_addr);
1543         if (!net_dev)
1544                 return ERR_PTR(-ENODEV);
1545
1546         return net_dev;
1547 }
1548
1549 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1550 {
1551         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1552 }
1553
1554 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1555                                    const struct cma_hdr *hdr)
1556 {
1557         struct sockaddr *addr = cma_src_addr(id_priv);
1558         __be32 ip4_addr;
1559         struct in6_addr ip6_addr;
1560
1561         if (cma_any_addr(addr) && !id_priv->afonly)
1562                 return true;
1563
1564         switch (addr->sa_family) {
1565         case AF_INET:
1566                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1567                 if (cma_get_ip_ver(hdr) != 4)
1568                         return false;
1569                 if (!cma_any_addr(addr) &&
1570                     hdr->dst_addr.ip4.addr != ip4_addr)
1571                         return false;
1572                 break;
1573         case AF_INET6:
1574                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1575                 if (cma_get_ip_ver(hdr) != 6)
1576                         return false;
1577                 if (!cma_any_addr(addr) &&
1578                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1579                         return false;
1580                 break;
1581         case AF_IB:
1582                 return true;
1583         default:
1584                 return false;
1585         }
1586
1587         return true;
1588 }
1589
1590 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1591 {
1592         struct ib_device *device = id->device;
1593         const int port_num = id->port_num ?: rdma_start_port(device);
1594
1595         return rdma_protocol_roce(device, port_num);
1596 }
1597
1598 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1599 {
1600         const struct sockaddr *daddr =
1601                         (const struct sockaddr *)&req->listen_addr_storage;
1602         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1603
1604         /* Returns true if the req is for IPv6 link local */
1605         return (daddr->sa_family == AF_INET6 &&
1606                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1607 }
1608
1609 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1610                               const struct net_device *net_dev,
1611                               const struct cma_req_info *req)
1612 {
1613         const struct rdma_addr *addr = &id->route.addr;
1614
1615         if (!net_dev)
1616                 /* This request is an AF_IB request */
1617                 return (!id->port_num || id->port_num == req->port) &&
1618                        (addr->src_addr.ss_family == AF_IB);
1619
1620         /*
1621          * If the request is not for IPv6 link local, allow matching
1622          * request to any netdevice of the one or multiport rdma device.
1623          */
1624         if (!cma_is_req_ipv6_ll(req))
1625                 return true;
1626         /*
1627          * Net namespaces must match, and if the listner is listening
1628          * on a specific netdevice than netdevice must match as well.
1629          */
1630         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1631             (!!addr->dev_addr.bound_dev_if ==
1632              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1633                 return true;
1634         else
1635                 return false;
1636 }
1637
1638 static struct rdma_id_private *cma_find_listener(
1639                 const struct rdma_bind_list *bind_list,
1640                 const struct ib_cm_id *cm_id,
1641                 const struct ib_cm_event *ib_event,
1642                 const struct cma_req_info *req,
1643                 const struct net_device *net_dev)
1644 {
1645         struct rdma_id_private *id_priv, *id_priv_dev;
1646
1647         lockdep_assert_held(&lock);
1648
1649         if (!bind_list)
1650                 return ERR_PTR(-EINVAL);
1651
1652         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1653                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1654                         if (id_priv->id.device == cm_id->device &&
1655                             cma_match_net_dev(&id_priv->id, net_dev, req))
1656                                 return id_priv;
1657                         list_for_each_entry(id_priv_dev,
1658                                             &id_priv->listen_list,
1659                                             listen_list) {
1660                                 if (id_priv_dev->id.device == cm_id->device &&
1661                                     cma_match_net_dev(&id_priv_dev->id,
1662                                                       net_dev, req))
1663                                         return id_priv_dev;
1664                         }
1665                 }
1666         }
1667
1668         return ERR_PTR(-EINVAL);
1669 }
1670
1671 static struct rdma_id_private *
1672 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1673                      const struct ib_cm_event *ib_event,
1674                      struct cma_req_info *req,
1675                      struct net_device **net_dev)
1676 {
1677         struct rdma_bind_list *bind_list;
1678         struct rdma_id_private *id_priv;
1679         int err;
1680
1681         err = cma_save_req_info(ib_event, req);
1682         if (err)
1683                 return ERR_PTR(err);
1684
1685         *net_dev = cma_get_net_dev(ib_event, req);
1686         if (IS_ERR(*net_dev)) {
1687                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1688                         /* Assuming the protocol is AF_IB */
1689                         *net_dev = NULL;
1690                 } else {
1691                         return ERR_CAST(*net_dev);
1692                 }
1693         }
1694
1695         mutex_lock(&lock);
1696         /*
1697          * Net namespace might be getting deleted while route lookup,
1698          * cm_id lookup is in progress. Therefore, perform netdevice
1699          * validation, cm_id lookup under rcu lock.
1700          * RCU lock along with netdevice state check, synchronizes with
1701          * netdevice migrating to different net namespace and also avoids
1702          * case where net namespace doesn't get deleted while lookup is in
1703          * progress.
1704          * If the device state is not IFF_UP, its properties such as ifindex
1705          * and nd_net cannot be trusted to remain valid without rcu lock.
1706          * net/core/dev.c change_net_namespace() ensures to synchronize with
1707          * ongoing operations on net device after device is closed using
1708          * synchronize_net().
1709          */
1710         rcu_read_lock();
1711         if (*net_dev) {
1712                 /*
1713                  * If netdevice is down, it is likely that it is administratively
1714                  * down or it might be migrating to different namespace.
1715                  * In that case avoid further processing, as the net namespace
1716                  * or ifindex may change.
1717                  */
1718                 if (((*net_dev)->flags & IFF_UP) == 0) {
1719                         id_priv = ERR_PTR(-EHOSTUNREACH);
1720                         goto err;
1721                 }
1722
1723                 if (!validate_net_dev(*net_dev,
1724                                  (struct sockaddr *)&req->src_addr_storage,
1725                                  (struct sockaddr *)&req->listen_addr_storage)) {
1726                         id_priv = ERR_PTR(-EHOSTUNREACH);
1727                         goto err;
1728                 }
1729         }
1730
1731         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1732                                 rdma_ps_from_service_id(req->service_id),
1733                                 cma_port_from_service_id(req->service_id));
1734         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1735 err:
1736         rcu_read_unlock();
1737         mutex_unlock(&lock);
1738         if (IS_ERR(id_priv) && *net_dev) {
1739                 dev_put(*net_dev);
1740                 *net_dev = NULL;
1741         }
1742         return id_priv;
1743 }
1744
1745 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1746 {
1747         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1748 }
1749
1750 static void cma_cancel_route(struct rdma_id_private *id_priv)
1751 {
1752         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1753                 if (id_priv->query)
1754                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1755         }
1756 }
1757
1758 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1759 {
1760         struct rdma_id_private *dev_id_priv;
1761
1762         lockdep_assert_held(&lock);
1763
1764         /*
1765          * Remove from listen_any_list to prevent added devices from spawning
1766          * additional listen requests.
1767          */
1768         list_del(&id_priv->list);
1769
1770         while (!list_empty(&id_priv->listen_list)) {
1771                 dev_id_priv = list_entry(id_priv->listen_list.next,
1772                                          struct rdma_id_private, listen_list);
1773                 /* sync with device removal to avoid duplicate destruction */
1774                 list_del_init(&dev_id_priv->list);
1775                 list_del(&dev_id_priv->listen_list);
1776                 mutex_unlock(&lock);
1777
1778                 rdma_destroy_id(&dev_id_priv->id);
1779                 mutex_lock(&lock);
1780         }
1781 }
1782
1783 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1784 {
1785         mutex_lock(&lock);
1786         _cma_cancel_listens(id_priv);
1787         mutex_unlock(&lock);
1788 }
1789
1790 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1791                                  enum rdma_cm_state state)
1792 {
1793         switch (state) {
1794         case RDMA_CM_ADDR_QUERY:
1795                 /*
1796                  * We can avoid doing the rdma_addr_cancel() based on state,
1797                  * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1798                  * Notice that the addr_handler work could still be exiting
1799                  * outside this state, however due to the interaction with the
1800                  * handler_mutex the work is guaranteed not to touch id_priv
1801                  * during exit.
1802                  */
1803                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1804                 break;
1805         case RDMA_CM_ROUTE_QUERY:
1806                 cma_cancel_route(id_priv);
1807                 break;
1808         case RDMA_CM_LISTEN:
1809                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1810                         cma_cancel_listens(id_priv);
1811                 break;
1812         default:
1813                 break;
1814         }
1815 }
1816
1817 static void cma_release_port(struct rdma_id_private *id_priv)
1818 {
1819         struct rdma_bind_list *bind_list = id_priv->bind_list;
1820         struct net *net = id_priv->id.route.addr.dev_addr.net;
1821
1822         if (!bind_list)
1823                 return;
1824
1825         mutex_lock(&lock);
1826         hlist_del(&id_priv->node);
1827         if (hlist_empty(&bind_list->owners)) {
1828                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1829                 kfree(bind_list);
1830         }
1831         mutex_unlock(&lock);
1832 }
1833
1834 static void destroy_mc(struct rdma_id_private *id_priv,
1835                        struct cma_multicast *mc)
1836 {
1837         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1838
1839         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1840                 ib_sa_free_multicast(mc->sa_mc);
1841
1842         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1843                 struct rdma_dev_addr *dev_addr =
1844                         &id_priv->id.route.addr.dev_addr;
1845                 struct net_device *ndev = NULL;
1846
1847                 if (dev_addr->bound_dev_if)
1848                         ndev = dev_get_by_index(dev_addr->net,
1849                                                 dev_addr->bound_dev_if);
1850                 if (ndev && !send_only) {
1851                         enum ib_gid_type gid_type;
1852                         union ib_gid mgid;
1853
1854                         gid_type = id_priv->cma_dev->default_gid_type
1855                                            [id_priv->id.port_num -
1856                                             rdma_start_port(
1857                                                     id_priv->cma_dev->device)];
1858                         cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
1859                                           gid_type);
1860                         cma_igmp_send(ndev, &mgid, false);
1861                 }
1862                 dev_put(ndev);
1863
1864                 cancel_work_sync(&mc->iboe_join.work);
1865         }
1866         kfree(mc);
1867 }
1868
1869 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1870 {
1871         struct cma_multicast *mc;
1872
1873         while (!list_empty(&id_priv->mc_list)) {
1874                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1875                                       list);
1876                 list_del(&mc->list);
1877                 destroy_mc(id_priv, mc);
1878         }
1879 }
1880
1881 static void _destroy_id(struct rdma_id_private *id_priv,
1882                         enum rdma_cm_state state)
1883 {
1884         cma_cancel_operation(id_priv, state);
1885
1886         rdma_restrack_del(&id_priv->res);
1887         if (id_priv->cma_dev) {
1888                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1889                         if (id_priv->cm_id.ib)
1890                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1891                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1892                         if (id_priv->cm_id.iw)
1893                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1894                 }
1895                 cma_leave_mc_groups(id_priv);
1896                 cma_release_dev(id_priv);
1897         }
1898
1899         cma_release_port(id_priv);
1900         cma_id_put(id_priv);
1901         wait_for_completion(&id_priv->comp);
1902
1903         if (id_priv->internal_id)
1904                 cma_id_put(id_priv->id.context);
1905
1906         kfree(id_priv->id.route.path_rec);
1907
1908         put_net(id_priv->id.route.addr.dev_addr.net);
1909         kfree(id_priv);
1910 }
1911
1912 /*
1913  * destroy an ID from within the handler_mutex. This ensures that no other
1914  * handlers can start running concurrently.
1915  */
1916 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1917         __releases(&idprv->handler_mutex)
1918 {
1919         enum rdma_cm_state state;
1920         unsigned long flags;
1921
1922         trace_cm_id_destroy(id_priv);
1923
1924         /*
1925          * Setting the state to destroyed under the handler mutex provides a
1926          * fence against calling handler callbacks. If this is invoked due to
1927          * the failure of a handler callback then it guarentees that no future
1928          * handlers will be called.
1929          */
1930         lockdep_assert_held(&id_priv->handler_mutex);
1931         spin_lock_irqsave(&id_priv->lock, flags);
1932         state = id_priv->state;
1933         id_priv->state = RDMA_CM_DESTROYING;
1934         spin_unlock_irqrestore(&id_priv->lock, flags);
1935         mutex_unlock(&id_priv->handler_mutex);
1936         _destroy_id(id_priv, state);
1937 }
1938
1939 void rdma_destroy_id(struct rdma_cm_id *id)
1940 {
1941         struct rdma_id_private *id_priv =
1942                 container_of(id, struct rdma_id_private, id);
1943
1944         mutex_lock(&id_priv->handler_mutex);
1945         destroy_id_handler_unlock(id_priv);
1946 }
1947 EXPORT_SYMBOL(rdma_destroy_id);
1948
1949 static int cma_rep_recv(struct rdma_id_private *id_priv)
1950 {
1951         int ret;
1952
1953         ret = cma_modify_qp_rtr(id_priv, NULL);
1954         if (ret)
1955                 goto reject;
1956
1957         ret = cma_modify_qp_rts(id_priv, NULL);
1958         if (ret)
1959                 goto reject;
1960
1961         trace_cm_send_rtu(id_priv);
1962         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1963         if (ret)
1964                 goto reject;
1965
1966         return 0;
1967 reject:
1968         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1969         cma_modify_qp_err(id_priv);
1970         trace_cm_send_rej(id_priv);
1971         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1972                        NULL, 0, NULL, 0);
1973         return ret;
1974 }
1975
1976 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1977                                    const struct ib_cm_rep_event_param *rep_data,
1978                                    void *private_data)
1979 {
1980         event->param.conn.private_data = private_data;
1981         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1982         event->param.conn.responder_resources = rep_data->responder_resources;
1983         event->param.conn.initiator_depth = rep_data->initiator_depth;
1984         event->param.conn.flow_control = rep_data->flow_control;
1985         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1986         event->param.conn.srq = rep_data->srq;
1987         event->param.conn.qp_num = rep_data->remote_qpn;
1988
1989         event->ece.vendor_id = rep_data->ece.vendor_id;
1990         event->ece.attr_mod = rep_data->ece.attr_mod;
1991 }
1992
1993 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1994                                 struct rdma_cm_event *event)
1995 {
1996         int ret;
1997
1998         lockdep_assert_held(&id_priv->handler_mutex);
1999
2000         trace_cm_event_handler(id_priv, event);
2001         ret = id_priv->id.event_handler(&id_priv->id, event);
2002         trace_cm_event_done(id_priv, event, ret);
2003         return ret;
2004 }
2005
2006 static int cma_ib_handler(struct ib_cm_id *cm_id,
2007                           const struct ib_cm_event *ib_event)
2008 {
2009         struct rdma_id_private *id_priv = cm_id->context;
2010         struct rdma_cm_event event = {};
2011         enum rdma_cm_state state;
2012         int ret;
2013
2014         mutex_lock(&id_priv->handler_mutex);
2015         state = READ_ONCE(id_priv->state);
2016         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2017              state != RDMA_CM_CONNECT) ||
2018             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2019              state != RDMA_CM_DISCONNECT))
2020                 goto out;
2021
2022         switch (ib_event->event) {
2023         case IB_CM_REQ_ERROR:
2024         case IB_CM_REP_ERROR:
2025                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2026                 event.status = -ETIMEDOUT;
2027                 break;
2028         case IB_CM_REP_RECEIVED:
2029                 if (state == RDMA_CM_CONNECT &&
2030                     (id_priv->id.qp_type != IB_QPT_UD)) {
2031                         trace_cm_send_mra(id_priv);
2032                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2033                 }
2034                 if (id_priv->id.qp) {
2035                         event.status = cma_rep_recv(id_priv);
2036                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2037                                                      RDMA_CM_EVENT_ESTABLISHED;
2038                 } else {
2039                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2040                 }
2041                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2042                                        ib_event->private_data);
2043                 break;
2044         case IB_CM_RTU_RECEIVED:
2045         case IB_CM_USER_ESTABLISHED:
2046                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2047                 break;
2048         case IB_CM_DREQ_ERROR:
2049                 event.status = -ETIMEDOUT;
2050                 fallthrough;
2051         case IB_CM_DREQ_RECEIVED:
2052         case IB_CM_DREP_RECEIVED:
2053                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2054                                    RDMA_CM_DISCONNECT))
2055                         goto out;
2056                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2057                 break;
2058         case IB_CM_TIMEWAIT_EXIT:
2059                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2060                 break;
2061         case IB_CM_MRA_RECEIVED:
2062                 /* ignore event */
2063                 goto out;
2064         case IB_CM_REJ_RECEIVED:
2065                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2066                                                                                 ib_event->param.rej_rcvd.reason));
2067                 cma_modify_qp_err(id_priv);
2068                 event.status = ib_event->param.rej_rcvd.reason;
2069                 event.event = RDMA_CM_EVENT_REJECTED;
2070                 event.param.conn.private_data = ib_event->private_data;
2071                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2072                 break;
2073         default:
2074                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2075                        ib_event->event);
2076                 goto out;
2077         }
2078
2079         ret = cma_cm_event_handler(id_priv, &event);
2080         if (ret) {
2081                 /* Destroy the CM ID by returning a non-zero value. */
2082                 id_priv->cm_id.ib = NULL;
2083                 destroy_id_handler_unlock(id_priv);
2084                 return ret;
2085         }
2086 out:
2087         mutex_unlock(&id_priv->handler_mutex);
2088         return 0;
2089 }
2090
2091 static struct rdma_id_private *
2092 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2093                    const struct ib_cm_event *ib_event,
2094                    struct net_device *net_dev)
2095 {
2096         struct rdma_id_private *listen_id_priv;
2097         struct rdma_id_private *id_priv;
2098         struct rdma_cm_id *id;
2099         struct rdma_route *rt;
2100         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2101         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2102         const __be64 service_id =
2103                 ib_event->param.req_rcvd.primary_path->service_id;
2104         int ret;
2105
2106         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2107         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2108                                    listen_id->event_handler, listen_id->context,
2109                                    listen_id->ps,
2110                                    ib_event->param.req_rcvd.qp_type,
2111                                    listen_id_priv);
2112         if (IS_ERR(id_priv))
2113                 return NULL;
2114
2115         id = &id_priv->id;
2116         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2117                               (struct sockaddr *)&id->route.addr.dst_addr,
2118                               listen_id, ib_event, ss_family, service_id))
2119                 goto err;
2120
2121         rt = &id->route;
2122         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2123         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2124                                      GFP_KERNEL);
2125         if (!rt->path_rec)
2126                 goto err;
2127
2128         rt->path_rec[0] = *path;
2129         if (rt->num_paths == 2)
2130                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2131
2132         if (net_dev) {
2133                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2134         } else {
2135                 if (!cma_protocol_roce(listen_id) &&
2136                     cma_any_addr(cma_src_addr(id_priv))) {
2137                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2138                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2139                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2140                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2141                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2142                         if (ret)
2143                                 goto err;
2144                 }
2145         }
2146         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2147
2148         id_priv->state = RDMA_CM_CONNECT;
2149         return id_priv;
2150
2151 err:
2152         rdma_destroy_id(id);
2153         return NULL;
2154 }
2155
2156 static struct rdma_id_private *
2157 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2158                   const struct ib_cm_event *ib_event,
2159                   struct net_device *net_dev)
2160 {
2161         const struct rdma_id_private *listen_id_priv;
2162         struct rdma_id_private *id_priv;
2163         struct rdma_cm_id *id;
2164         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2165         struct net *net = listen_id->route.addr.dev_addr.net;
2166         int ret;
2167
2168         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2169         id_priv = __rdma_create_id(net, listen_id->event_handler,
2170                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2171                                    listen_id_priv);
2172         if (IS_ERR(id_priv))
2173                 return NULL;
2174
2175         id = &id_priv->id;
2176         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2177                               (struct sockaddr *)&id->route.addr.dst_addr,
2178                               listen_id, ib_event, ss_family,
2179                               ib_event->param.sidr_req_rcvd.service_id))
2180                 goto err;
2181
2182         if (net_dev) {
2183                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2184         } else {
2185                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2186                         ret = cma_translate_addr(cma_src_addr(id_priv),
2187                                                  &id->route.addr.dev_addr);
2188                         if (ret)
2189                                 goto err;
2190                 }
2191         }
2192
2193         id_priv->state = RDMA_CM_CONNECT;
2194         return id_priv;
2195 err:
2196         rdma_destroy_id(id);
2197         return NULL;
2198 }
2199
2200 static void cma_set_req_event_data(struct rdma_cm_event *event,
2201                                    const struct ib_cm_req_event_param *req_data,
2202                                    void *private_data, int offset)
2203 {
2204         event->param.conn.private_data = private_data + offset;
2205         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2206         event->param.conn.responder_resources = req_data->responder_resources;
2207         event->param.conn.initiator_depth = req_data->initiator_depth;
2208         event->param.conn.flow_control = req_data->flow_control;
2209         event->param.conn.retry_count = req_data->retry_count;
2210         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2211         event->param.conn.srq = req_data->srq;
2212         event->param.conn.qp_num = req_data->remote_qpn;
2213
2214         event->ece.vendor_id = req_data->ece.vendor_id;
2215         event->ece.attr_mod = req_data->ece.attr_mod;
2216 }
2217
2218 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2219                                     const struct ib_cm_event *ib_event)
2220 {
2221         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2222                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2223                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2224                  (id->qp_type == IB_QPT_UD)) ||
2225                 (!id->qp_type));
2226 }
2227
2228 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2229                               const struct ib_cm_event *ib_event)
2230 {
2231         struct rdma_id_private *listen_id, *conn_id = NULL;
2232         struct rdma_cm_event event = {};
2233         struct cma_req_info req = {};
2234         struct net_device *net_dev;
2235         u8 offset;
2236         int ret;
2237
2238         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2239         if (IS_ERR(listen_id))
2240                 return PTR_ERR(listen_id);
2241
2242         trace_cm_req_handler(listen_id, ib_event->event);
2243         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2244                 ret = -EINVAL;
2245                 goto net_dev_put;
2246         }
2247
2248         mutex_lock(&listen_id->handler_mutex);
2249         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2250                 ret = -ECONNABORTED;
2251                 goto err_unlock;
2252         }
2253
2254         offset = cma_user_data_offset(listen_id);
2255         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2256         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2257                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2258                 event.param.ud.private_data = ib_event->private_data + offset;
2259                 event.param.ud.private_data_len =
2260                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2261         } else {
2262                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2263                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2264                                        ib_event->private_data, offset);
2265         }
2266         if (!conn_id) {
2267                 ret = -ENOMEM;
2268                 goto err_unlock;
2269         }
2270
2271         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2272         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2273         if (ret) {
2274                 destroy_id_handler_unlock(conn_id);
2275                 goto err_unlock;
2276         }
2277
2278         conn_id->cm_id.ib = cm_id;
2279         cm_id->context = conn_id;
2280         cm_id->cm_handler = cma_ib_handler;
2281
2282         ret = cma_cm_event_handler(conn_id, &event);
2283         if (ret) {
2284                 /* Destroy the CM ID by returning a non-zero value. */
2285                 conn_id->cm_id.ib = NULL;
2286                 mutex_unlock(&listen_id->handler_mutex);
2287                 destroy_id_handler_unlock(conn_id);
2288                 goto net_dev_put;
2289         }
2290
2291         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2292             conn_id->id.qp_type != IB_QPT_UD) {
2293                 trace_cm_send_mra(cm_id->context);
2294                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2295         }
2296         mutex_unlock(&conn_id->handler_mutex);
2297
2298 err_unlock:
2299         mutex_unlock(&listen_id->handler_mutex);
2300
2301 net_dev_put:
2302         if (net_dev)
2303                 dev_put(net_dev);
2304
2305         return ret;
2306 }
2307
2308 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2309 {
2310         if (addr->sa_family == AF_IB)
2311                 return ((struct sockaddr_ib *) addr)->sib_sid;
2312
2313         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2314 }
2315 EXPORT_SYMBOL(rdma_get_service_id);
2316
2317 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2318                     union ib_gid *dgid)
2319 {
2320         struct rdma_addr *addr = &cm_id->route.addr;
2321
2322         if (!cm_id->device) {
2323                 if (sgid)
2324                         memset(sgid, 0, sizeof(*sgid));
2325                 if (dgid)
2326                         memset(dgid, 0, sizeof(*dgid));
2327                 return;
2328         }
2329
2330         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2331                 if (sgid)
2332                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2333                 if (dgid)
2334                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2335         } else {
2336                 if (sgid)
2337                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2338                 if (dgid)
2339                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2340         }
2341 }
2342 EXPORT_SYMBOL(rdma_read_gids);
2343
2344 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2345 {
2346         struct rdma_id_private *id_priv = iw_id->context;
2347         struct rdma_cm_event event = {};
2348         int ret = 0;
2349         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2350         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2351
2352         mutex_lock(&id_priv->handler_mutex);
2353         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2354                 goto out;
2355
2356         switch (iw_event->event) {
2357         case IW_CM_EVENT_CLOSE:
2358                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2359                 break;
2360         case IW_CM_EVENT_CONNECT_REPLY:
2361                 memcpy(cma_src_addr(id_priv), laddr,
2362                        rdma_addr_size(laddr));
2363                 memcpy(cma_dst_addr(id_priv), raddr,
2364                        rdma_addr_size(raddr));
2365                 switch (iw_event->status) {
2366                 case 0:
2367                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2368                         event.param.conn.initiator_depth = iw_event->ird;
2369                         event.param.conn.responder_resources = iw_event->ord;
2370                         break;
2371                 case -ECONNRESET:
2372                 case -ECONNREFUSED:
2373                         event.event = RDMA_CM_EVENT_REJECTED;
2374                         break;
2375                 case -ETIMEDOUT:
2376                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2377                         break;
2378                 default:
2379                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2380                         break;
2381                 }
2382                 break;
2383         case IW_CM_EVENT_ESTABLISHED:
2384                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2385                 event.param.conn.initiator_depth = iw_event->ird;
2386                 event.param.conn.responder_resources = iw_event->ord;
2387                 break;
2388         default:
2389                 goto out;
2390         }
2391
2392         event.status = iw_event->status;
2393         event.param.conn.private_data = iw_event->private_data;
2394         event.param.conn.private_data_len = iw_event->private_data_len;
2395         ret = cma_cm_event_handler(id_priv, &event);
2396         if (ret) {
2397                 /* Destroy the CM ID by returning a non-zero value. */
2398                 id_priv->cm_id.iw = NULL;
2399                 destroy_id_handler_unlock(id_priv);
2400                 return ret;
2401         }
2402
2403 out:
2404         mutex_unlock(&id_priv->handler_mutex);
2405         return ret;
2406 }
2407
2408 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2409                                struct iw_cm_event *iw_event)
2410 {
2411         struct rdma_id_private *listen_id, *conn_id;
2412         struct rdma_cm_event event = {};
2413         int ret = -ECONNABORTED;
2414         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2415         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2416
2417         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2418         event.param.conn.private_data = iw_event->private_data;
2419         event.param.conn.private_data_len = iw_event->private_data_len;
2420         event.param.conn.initiator_depth = iw_event->ird;
2421         event.param.conn.responder_resources = iw_event->ord;
2422
2423         listen_id = cm_id->context;
2424
2425         mutex_lock(&listen_id->handler_mutex);
2426         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2427                 goto out;
2428
2429         /* Create a new RDMA id for the new IW CM ID */
2430         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2431                                    listen_id->id.event_handler,
2432                                    listen_id->id.context, RDMA_PS_TCP,
2433                                    IB_QPT_RC, listen_id);
2434         if (IS_ERR(conn_id)) {
2435                 ret = -ENOMEM;
2436                 goto out;
2437         }
2438         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2439         conn_id->state = RDMA_CM_CONNECT;
2440
2441         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2442         if (ret) {
2443                 mutex_unlock(&listen_id->handler_mutex);
2444                 destroy_id_handler_unlock(conn_id);
2445                 return ret;
2446         }
2447
2448         ret = cma_iw_acquire_dev(conn_id, listen_id);
2449         if (ret) {
2450                 mutex_unlock(&listen_id->handler_mutex);
2451                 destroy_id_handler_unlock(conn_id);
2452                 return ret;
2453         }
2454
2455         conn_id->cm_id.iw = cm_id;
2456         cm_id->context = conn_id;
2457         cm_id->cm_handler = cma_iw_handler;
2458
2459         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2460         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2461
2462         ret = cma_cm_event_handler(conn_id, &event);
2463         if (ret) {
2464                 /* User wants to destroy the CM ID */
2465                 conn_id->cm_id.iw = NULL;
2466                 mutex_unlock(&listen_id->handler_mutex);
2467                 destroy_id_handler_unlock(conn_id);
2468                 return ret;
2469         }
2470
2471         mutex_unlock(&conn_id->handler_mutex);
2472
2473 out:
2474         mutex_unlock(&listen_id->handler_mutex);
2475         return ret;
2476 }
2477
2478 static int cma_ib_listen(struct rdma_id_private *id_priv)
2479 {
2480         struct sockaddr *addr;
2481         struct ib_cm_id *id;
2482         __be64 svc_id;
2483
2484         addr = cma_src_addr(id_priv);
2485         svc_id = rdma_get_service_id(&id_priv->id, addr);
2486         id = ib_cm_insert_listen(id_priv->id.device,
2487                                  cma_ib_req_handler, svc_id);
2488         if (IS_ERR(id))
2489                 return PTR_ERR(id);
2490         id_priv->cm_id.ib = id;
2491
2492         return 0;
2493 }
2494
2495 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2496 {
2497         int ret;
2498         struct iw_cm_id *id;
2499
2500         id = iw_create_cm_id(id_priv->id.device,
2501                              iw_conn_req_handler,
2502                              id_priv);
2503         if (IS_ERR(id))
2504                 return PTR_ERR(id);
2505
2506         mutex_lock(&id_priv->qp_mutex);
2507         id->tos = id_priv->tos;
2508         id->tos_set = id_priv->tos_set;
2509         mutex_unlock(&id_priv->qp_mutex);
2510         id_priv->cm_id.iw = id;
2511
2512         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2513                rdma_addr_size(cma_src_addr(id_priv)));
2514
2515         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2516
2517         if (ret) {
2518                 iw_destroy_cm_id(id_priv->cm_id.iw);
2519                 id_priv->cm_id.iw = NULL;
2520         }
2521
2522         return ret;
2523 }
2524
2525 static int cma_listen_handler(struct rdma_cm_id *id,
2526                               struct rdma_cm_event *event)
2527 {
2528         struct rdma_id_private *id_priv = id->context;
2529
2530         /* Listening IDs are always destroyed on removal */
2531         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2532                 return -1;
2533
2534         id->context = id_priv->id.context;
2535         id->event_handler = id_priv->id.event_handler;
2536         trace_cm_event_handler(id_priv, event);
2537         return id_priv->id.event_handler(id, event);
2538 }
2539
2540 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2541                              struct cma_device *cma_dev,
2542                              struct rdma_id_private **to_destroy)
2543 {
2544         struct rdma_id_private *dev_id_priv;
2545         struct net *net = id_priv->id.route.addr.dev_addr.net;
2546         int ret;
2547
2548         lockdep_assert_held(&lock);
2549
2550         *to_destroy = NULL;
2551         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2552                 return 0;
2553
2554         dev_id_priv =
2555                 __rdma_create_id(net, cma_listen_handler, id_priv,
2556                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2557         if (IS_ERR(dev_id_priv))
2558                 return PTR_ERR(dev_id_priv);
2559
2560         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2561         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2562                rdma_addr_size(cma_src_addr(id_priv)));
2563
2564         _cma_attach_to_dev(dev_id_priv, cma_dev);
2565         rdma_restrack_add(&dev_id_priv->res);
2566         cma_id_get(id_priv);
2567         dev_id_priv->internal_id = 1;
2568         dev_id_priv->afonly = id_priv->afonly;
2569         mutex_lock(&id_priv->qp_mutex);
2570         dev_id_priv->tos_set = id_priv->tos_set;
2571         dev_id_priv->tos = id_priv->tos;
2572         mutex_unlock(&id_priv->qp_mutex);
2573
2574         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2575         if (ret)
2576                 goto err_listen;
2577         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2578         return 0;
2579 err_listen:
2580         /* Caller must destroy this after releasing lock */
2581         *to_destroy = dev_id_priv;
2582         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2583         return ret;
2584 }
2585
2586 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2587 {
2588         struct rdma_id_private *to_destroy;
2589         struct cma_device *cma_dev;
2590         int ret;
2591
2592         mutex_lock(&lock);
2593         list_add_tail(&id_priv->list, &listen_any_list);
2594         list_for_each_entry(cma_dev, &dev_list, list) {
2595                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2596                 if (ret) {
2597                         /* Prevent racing with cma_process_remove() */
2598                         if (to_destroy)
2599                                 list_del_init(&to_destroy->list);
2600                         goto err_listen;
2601                 }
2602         }
2603         mutex_unlock(&lock);
2604         return 0;
2605
2606 err_listen:
2607         _cma_cancel_listens(id_priv);
2608         mutex_unlock(&lock);
2609         if (to_destroy)
2610                 rdma_destroy_id(&to_destroy->id);
2611         return ret;
2612 }
2613
2614 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2615 {
2616         struct rdma_id_private *id_priv;
2617
2618         id_priv = container_of(id, struct rdma_id_private, id);
2619         mutex_lock(&id_priv->qp_mutex);
2620         id_priv->tos = (u8) tos;
2621         id_priv->tos_set = true;
2622         mutex_unlock(&id_priv->qp_mutex);
2623 }
2624 EXPORT_SYMBOL(rdma_set_service_type);
2625
2626 /**
2627  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2628  *                          with a connection identifier.
2629  * @id: Communication identifier to associated with service type.
2630  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2631  *
2632  * This function should be called before rdma_connect() on active side,
2633  * and on passive side before rdma_accept(). It is applicable to primary
2634  * path only. The timeout will affect the local side of the QP, it is not
2635  * negotiated with remote side and zero disables the timer. In case it is
2636  * set before rdma_resolve_route, the value will also be used to determine
2637  * PacketLifeTime for RoCE.
2638  *
2639  * Return: 0 for success
2640  */
2641 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2642 {
2643         struct rdma_id_private *id_priv;
2644
2645         if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2646                 return -EINVAL;
2647
2648         id_priv = container_of(id, struct rdma_id_private, id);
2649         mutex_lock(&id_priv->qp_mutex);
2650         id_priv->timeout = timeout;
2651         id_priv->timeout_set = true;
2652         mutex_unlock(&id_priv->qp_mutex);
2653
2654         return 0;
2655 }
2656 EXPORT_SYMBOL(rdma_set_ack_timeout);
2657
2658 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2659                               void *context)
2660 {
2661         struct cma_work *work = context;
2662         struct rdma_route *route;
2663
2664         route = &work->id->id.route;
2665
2666         if (!status) {
2667                 route->num_paths = 1;
2668                 *route->path_rec = *path_rec;
2669         } else {
2670                 work->old_state = RDMA_CM_ROUTE_QUERY;
2671                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2672                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2673                 work->event.status = status;
2674                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2675                                      status);
2676         }
2677
2678         queue_work(cma_wq, &work->work);
2679 }
2680
2681 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2682                               unsigned long timeout_ms, struct cma_work *work)
2683 {
2684         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2685         struct sa_path_rec path_rec;
2686         ib_sa_comp_mask comp_mask;
2687         struct sockaddr_in6 *sin6;
2688         struct sockaddr_ib *sib;
2689
2690         memset(&path_rec, 0, sizeof path_rec);
2691
2692         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2693                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2694         else
2695                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2696         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2697         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2698         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2699         path_rec.numb_path = 1;
2700         path_rec.reversible = 1;
2701         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2702                                                   cma_dst_addr(id_priv));
2703
2704         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2705                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2706                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2707
2708         switch (cma_family(id_priv)) {
2709         case AF_INET:
2710                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2711                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2712                 break;
2713         case AF_INET6:
2714                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2715                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2716                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2717                 break;
2718         case AF_IB:
2719                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2720                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2721                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2722                 break;
2723         }
2724
2725         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2726                                                id_priv->id.port_num, &path_rec,
2727                                                comp_mask, timeout_ms,
2728                                                GFP_KERNEL, cma_query_handler,
2729                                                work, &id_priv->query);
2730
2731         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2732 }
2733
2734 static void cma_iboe_join_work_handler(struct work_struct *work)
2735 {
2736         struct cma_multicast *mc =
2737                 container_of(work, struct cma_multicast, iboe_join.work);
2738         struct rdma_cm_event *event = &mc->iboe_join.event;
2739         struct rdma_id_private *id_priv = mc->id_priv;
2740         int ret;
2741
2742         mutex_lock(&id_priv->handler_mutex);
2743         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2744             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2745                 goto out_unlock;
2746
2747         ret = cma_cm_event_handler(id_priv, event);
2748         WARN_ON(ret);
2749
2750 out_unlock:
2751         mutex_unlock(&id_priv->handler_mutex);
2752         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2753                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2754 }
2755
2756 static void cma_work_handler(struct work_struct *_work)
2757 {
2758         struct cma_work *work = container_of(_work, struct cma_work, work);
2759         struct rdma_id_private *id_priv = work->id;
2760
2761         mutex_lock(&id_priv->handler_mutex);
2762         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2763             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2764                 goto out_unlock;
2765         if (work->old_state != 0 || work->new_state != 0) {
2766                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2767                         goto out_unlock;
2768         }
2769
2770         if (cma_cm_event_handler(id_priv, &work->event)) {
2771                 cma_id_put(id_priv);
2772                 destroy_id_handler_unlock(id_priv);
2773                 goto out_free;
2774         }
2775
2776 out_unlock:
2777         mutex_unlock(&id_priv->handler_mutex);
2778         cma_id_put(id_priv);
2779 out_free:
2780         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2781                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2782         kfree(work);
2783 }
2784
2785 static void cma_init_resolve_route_work(struct cma_work *work,
2786                                         struct rdma_id_private *id_priv)
2787 {
2788         work->id = id_priv;
2789         INIT_WORK(&work->work, cma_work_handler);
2790         work->old_state = RDMA_CM_ROUTE_QUERY;
2791         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2792         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2793 }
2794
2795 static void enqueue_resolve_addr_work(struct cma_work *work,
2796                                       struct rdma_id_private *id_priv)
2797 {
2798         /* Balances with cma_id_put() in cma_work_handler */
2799         cma_id_get(id_priv);
2800
2801         work->id = id_priv;
2802         INIT_WORK(&work->work, cma_work_handler);
2803         work->old_state = RDMA_CM_ADDR_QUERY;
2804         work->new_state = RDMA_CM_ADDR_RESOLVED;
2805         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2806
2807         queue_work(cma_wq, &work->work);
2808 }
2809
2810 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2811                                 unsigned long timeout_ms)
2812 {
2813         struct rdma_route *route = &id_priv->id.route;
2814         struct cma_work *work;
2815         int ret;
2816
2817         work = kzalloc(sizeof *work, GFP_KERNEL);
2818         if (!work)
2819                 return -ENOMEM;
2820
2821         cma_init_resolve_route_work(work, id_priv);
2822
2823         if (!route->path_rec)
2824                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2825         if (!route->path_rec) {
2826                 ret = -ENOMEM;
2827                 goto err1;
2828         }
2829
2830         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2831         if (ret)
2832                 goto err2;
2833
2834         return 0;
2835 err2:
2836         kfree(route->path_rec);
2837         route->path_rec = NULL;
2838 err1:
2839         kfree(work);
2840         return ret;
2841 }
2842
2843 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2844                                            unsigned long supported_gids,
2845                                            enum ib_gid_type default_gid)
2846 {
2847         if ((network_type == RDMA_NETWORK_IPV4 ||
2848              network_type == RDMA_NETWORK_IPV6) &&
2849             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2850                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2851
2852         return default_gid;
2853 }
2854
2855 /*
2856  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2857  * path record type based on GID type.
2858  * It also sets up other L2 fields which includes destination mac address
2859  * netdev ifindex, of the path record.
2860  * It returns the netdev of the bound interface for this path record entry.
2861  */
2862 static struct net_device *
2863 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2864 {
2865         struct rdma_route *route = &id_priv->id.route;
2866         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2867         struct rdma_addr *addr = &route->addr;
2868         unsigned long supported_gids;
2869         struct net_device *ndev;
2870
2871         if (!addr->dev_addr.bound_dev_if)
2872                 return NULL;
2873
2874         ndev = dev_get_by_index(addr->dev_addr.net,
2875                                 addr->dev_addr.bound_dev_if);
2876         if (!ndev)
2877                 return NULL;
2878
2879         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2880                                                     id_priv->id.port_num);
2881         gid_type = cma_route_gid_type(addr->dev_addr.network,
2882                                       supported_gids,
2883                                       id_priv->gid_type);
2884         /* Use the hint from IP Stack to select GID Type */
2885         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2886                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2887         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2888
2889         route->path_rec->roce.route_resolved = true;
2890         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2891         return ndev;
2892 }
2893
2894 int rdma_set_ib_path(struct rdma_cm_id *id,
2895                      struct sa_path_rec *path_rec)
2896 {
2897         struct rdma_id_private *id_priv;
2898         struct net_device *ndev;
2899         int ret;
2900
2901         id_priv = container_of(id, struct rdma_id_private, id);
2902         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2903                            RDMA_CM_ROUTE_RESOLVED))
2904                 return -EINVAL;
2905
2906         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2907                                      GFP_KERNEL);
2908         if (!id->route.path_rec) {
2909                 ret = -ENOMEM;
2910                 goto err;
2911         }
2912
2913         if (rdma_protocol_roce(id->device, id->port_num)) {
2914                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2915                 if (!ndev) {
2916                         ret = -ENODEV;
2917                         goto err_free;
2918                 }
2919                 dev_put(ndev);
2920         }
2921
2922         id->route.num_paths = 1;
2923         return 0;
2924
2925 err_free:
2926         kfree(id->route.path_rec);
2927         id->route.path_rec = NULL;
2928 err:
2929         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2930         return ret;
2931 }
2932 EXPORT_SYMBOL(rdma_set_ib_path);
2933
2934 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2935 {
2936         struct cma_work *work;
2937
2938         work = kzalloc(sizeof *work, GFP_KERNEL);
2939         if (!work)
2940                 return -ENOMEM;
2941
2942         cma_init_resolve_route_work(work, id_priv);
2943         queue_work(cma_wq, &work->work);
2944         return 0;
2945 }
2946
2947 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2948 {
2949         struct net_device *dev;
2950
2951         dev = vlan_dev_real_dev(vlan_ndev);
2952         if (dev->num_tc)
2953                 return netdev_get_prio_tc_map(dev, prio);
2954
2955         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2956                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2957 }
2958
2959 struct iboe_prio_tc_map {
2960         int input_prio;
2961         int output_tc;
2962         bool found;
2963 };
2964
2965 static int get_lower_vlan_dev_tc(struct net_device *dev,
2966                                  struct netdev_nested_priv *priv)
2967 {
2968         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2969
2970         if (is_vlan_dev(dev))
2971                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2972         else if (dev->num_tc)
2973                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2974         else
2975                 map->output_tc = 0;
2976         /* We are interested only in first level VLAN device, so always
2977          * return 1 to stop iterating over next level devices.
2978          */
2979         map->found = true;
2980         return 1;
2981 }
2982
2983 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2984 {
2985         struct iboe_prio_tc_map prio_tc_map = {};
2986         int prio = rt_tos2priority(tos);
2987         struct netdev_nested_priv priv;
2988
2989         /* If VLAN device, get it directly from the VLAN netdev */
2990         if (is_vlan_dev(ndev))
2991                 return get_vlan_ndev_tc(ndev, prio);
2992
2993         prio_tc_map.input_prio = prio;
2994         priv.data = (void *)&prio_tc_map;
2995         rcu_read_lock();
2996         netdev_walk_all_lower_dev_rcu(ndev,
2997                                       get_lower_vlan_dev_tc,
2998                                       &priv);
2999         rcu_read_unlock();
3000         /* If map is found from lower device, use it; Otherwise
3001          * continue with the current netdevice to get priority to tc map.
3002          */
3003         if (prio_tc_map.found)
3004                 return prio_tc_map.output_tc;
3005         else if (ndev->num_tc)
3006                 return netdev_get_prio_tc_map(ndev, prio);
3007         else
3008                 return 0;
3009 }
3010
3011 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3012 {
3013         struct sockaddr_in6 *addr6;
3014         u16 dport, sport;
3015         u32 hash, fl;
3016
3017         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3018         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3019         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3020                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3021                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3022                 hash = (u32)sport * 31 + dport;
3023                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3024         }
3025
3026         return cpu_to_be32(fl);
3027 }
3028
3029 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3030 {
3031         struct rdma_route *route = &id_priv->id.route;
3032         struct rdma_addr *addr = &route->addr;
3033         struct cma_work *work;
3034         int ret;
3035         struct net_device *ndev;
3036
3037         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3038                                         rdma_start_port(id_priv->cma_dev->device)];
3039         u8 tos;
3040
3041         mutex_lock(&id_priv->qp_mutex);
3042         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3043         mutex_unlock(&id_priv->qp_mutex);
3044
3045         work = kzalloc(sizeof *work, GFP_KERNEL);
3046         if (!work)
3047                 return -ENOMEM;
3048
3049         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3050         if (!route->path_rec) {
3051                 ret = -ENOMEM;
3052                 goto err1;
3053         }
3054
3055         route->num_paths = 1;
3056
3057         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3058         if (!ndev) {
3059                 ret = -ENODEV;
3060                 goto err2;
3061         }
3062
3063         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3064                     &route->path_rec->sgid);
3065         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3066                     &route->path_rec->dgid);
3067
3068         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3069                 /* TODO: get the hoplimit from the inet/inet6 device */
3070                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3071         else
3072                 route->path_rec->hop_limit = 1;
3073         route->path_rec->reversible = 1;
3074         route->path_rec->pkey = cpu_to_be16(0xffff);
3075         route->path_rec->mtu_selector = IB_SA_EQ;
3076         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3077         route->path_rec->traffic_class = tos;
3078         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3079         route->path_rec->rate_selector = IB_SA_EQ;
3080         route->path_rec->rate = IB_RATE_PORT_CURRENT;
3081         dev_put(ndev);
3082         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3083         /* In case ACK timeout is set, use this value to calculate
3084          * PacketLifeTime.  As per IBTA 12.7.34,
3085          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3086          * Assuming a negligible local ACK delay, we can use
3087          * PacketLifeTime = local ACK timeout/2
3088          * as a reasonable approximation for RoCE networks.
3089          */
3090         mutex_lock(&id_priv->qp_mutex);
3091         if (id_priv->timeout_set && id_priv->timeout)
3092                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3093         else
3094                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3095         mutex_unlock(&id_priv->qp_mutex);
3096
3097         if (!route->path_rec->mtu) {
3098                 ret = -EINVAL;
3099                 goto err2;
3100         }
3101
3102         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3103                                          id_priv->id.port_num))
3104                 route->path_rec->flow_label =
3105                         cma_get_roce_udp_flow_label(id_priv);
3106
3107         cma_init_resolve_route_work(work, id_priv);
3108         queue_work(cma_wq, &work->work);
3109
3110         return 0;
3111
3112 err2:
3113         kfree(route->path_rec);
3114         route->path_rec = NULL;
3115         route->num_paths = 0;
3116 err1:
3117         kfree(work);
3118         return ret;
3119 }
3120
3121 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3122 {
3123         struct rdma_id_private *id_priv;
3124         int ret;
3125
3126         id_priv = container_of(id, struct rdma_id_private, id);
3127         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3128                 return -EINVAL;
3129
3130         cma_id_get(id_priv);
3131         if (rdma_cap_ib_sa(id->device, id->port_num))
3132                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3133         else if (rdma_protocol_roce(id->device, id->port_num))
3134                 ret = cma_resolve_iboe_route(id_priv);
3135         else if (rdma_protocol_iwarp(id->device, id->port_num))
3136                 ret = cma_resolve_iw_route(id_priv);
3137         else
3138                 ret = -ENOSYS;
3139
3140         if (ret)
3141                 goto err;
3142
3143         return 0;
3144 err:
3145         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3146         cma_id_put(id_priv);
3147         return ret;
3148 }
3149 EXPORT_SYMBOL(rdma_resolve_route);
3150
3151 static void cma_set_loopback(struct sockaddr *addr)
3152 {
3153         switch (addr->sa_family) {
3154         case AF_INET:
3155                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3156                 break;
3157         case AF_INET6:
3158                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3159                               0, 0, 0, htonl(1));
3160                 break;
3161         default:
3162                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3163                             0, 0, 0, htonl(1));
3164                 break;
3165         }
3166 }
3167
3168 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3169 {
3170         struct cma_device *cma_dev, *cur_dev;
3171         union ib_gid gid;
3172         enum ib_port_state port_state;
3173         unsigned int p;
3174         u16 pkey;
3175         int ret;
3176
3177         cma_dev = NULL;
3178         mutex_lock(&lock);
3179         list_for_each_entry(cur_dev, &dev_list, list) {
3180                 if (cma_family(id_priv) == AF_IB &&
3181                     !rdma_cap_ib_cm(cur_dev->device, 1))
3182                         continue;
3183
3184                 if (!cma_dev)
3185                         cma_dev = cur_dev;
3186
3187                 rdma_for_each_port (cur_dev->device, p) {
3188                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3189                             port_state == IB_PORT_ACTIVE) {
3190                                 cma_dev = cur_dev;
3191                                 goto port_found;
3192                         }
3193                 }
3194         }
3195
3196         if (!cma_dev) {
3197                 ret = -ENODEV;
3198                 goto out;
3199         }
3200
3201         p = 1;
3202
3203 port_found:
3204         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3205         if (ret)
3206                 goto out;
3207
3208         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3209         if (ret)
3210                 goto out;
3211
3212         id_priv->id.route.addr.dev_addr.dev_type =
3213                 (rdma_protocol_ib(cma_dev->device, p)) ?
3214                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3215
3216         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3217         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3218         id_priv->id.port_num = p;
3219         cma_attach_to_dev(id_priv, cma_dev);
3220         rdma_restrack_add(&id_priv->res);
3221         cma_set_loopback(cma_src_addr(id_priv));
3222 out:
3223         mutex_unlock(&lock);
3224         return ret;
3225 }
3226
3227 static void addr_handler(int status, struct sockaddr *src_addr,
3228                          struct rdma_dev_addr *dev_addr, void *context)
3229 {
3230         struct rdma_id_private *id_priv = context;
3231         struct rdma_cm_event event = {};
3232         struct sockaddr *addr;
3233         struct sockaddr_storage old_addr;
3234
3235         mutex_lock(&id_priv->handler_mutex);
3236         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3237                            RDMA_CM_ADDR_RESOLVED))
3238                 goto out;
3239
3240         /*
3241          * Store the previous src address, so that if we fail to acquire
3242          * matching rdma device, old address can be restored back, which helps
3243          * to cancel the cma listen operation correctly.
3244          */
3245         addr = cma_src_addr(id_priv);
3246         memcpy(&old_addr, addr, rdma_addr_size(addr));
3247         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3248         if (!status && !id_priv->cma_dev) {
3249                 status = cma_acquire_dev_by_src_ip(id_priv);
3250                 if (status)
3251                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3252                                              status);
3253                 rdma_restrack_add(&id_priv->res);
3254         } else if (status) {
3255                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3256         }
3257
3258         if (status) {
3259                 memcpy(addr, &old_addr,
3260                        rdma_addr_size((struct sockaddr *)&old_addr));
3261                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3262                                    RDMA_CM_ADDR_BOUND))
3263                         goto out;
3264                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3265                 event.status = status;
3266         } else
3267                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3268
3269         if (cma_cm_event_handler(id_priv, &event)) {
3270                 destroy_id_handler_unlock(id_priv);
3271                 return;
3272         }
3273 out:
3274         mutex_unlock(&id_priv->handler_mutex);
3275 }
3276
3277 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3278 {
3279         struct cma_work *work;
3280         union ib_gid gid;
3281         int ret;
3282
3283         work = kzalloc(sizeof *work, GFP_KERNEL);
3284         if (!work)
3285                 return -ENOMEM;
3286
3287         if (!id_priv->cma_dev) {
3288                 ret = cma_bind_loopback(id_priv);
3289                 if (ret)
3290                         goto err;
3291         }
3292
3293         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3294         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3295
3296         enqueue_resolve_addr_work(work, id_priv);
3297         return 0;
3298 err:
3299         kfree(work);
3300         return ret;
3301 }
3302
3303 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3304 {
3305         struct cma_work *work;
3306         int ret;
3307
3308         work = kzalloc(sizeof *work, GFP_KERNEL);
3309         if (!work)
3310                 return -ENOMEM;
3311
3312         if (!id_priv->cma_dev) {
3313                 ret = cma_resolve_ib_dev(id_priv);
3314                 if (ret)
3315                         goto err;
3316         }
3317
3318         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3319                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3320
3321         enqueue_resolve_addr_work(work, id_priv);
3322         return 0;
3323 err:
3324         kfree(work);
3325         return ret;
3326 }
3327
3328 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3329                          const struct sockaddr *dst_addr)
3330 {
3331         struct sockaddr_storage zero_sock = {};
3332
3333         if (src_addr && src_addr->sa_family)
3334                 return rdma_bind_addr(id, src_addr);
3335
3336         /*
3337          * When the src_addr is not specified, automatically supply an any addr
3338          */
3339         zero_sock.ss_family = dst_addr->sa_family;
3340         if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
3341                 struct sockaddr_in6 *src_addr6 =
3342                         (struct sockaddr_in6 *)&zero_sock;
3343                 struct sockaddr_in6 *dst_addr6 =
3344                         (struct sockaddr_in6 *)dst_addr;
3345
3346                 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3347                 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3348                         id->route.addr.dev_addr.bound_dev_if =
3349                                 dst_addr6->sin6_scope_id;
3350         } else if (dst_addr->sa_family == AF_IB) {
3351                 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
3352                         ((struct sockaddr_ib *)dst_addr)->sib_pkey;
3353         }
3354         return rdma_bind_addr(id, (struct sockaddr *)&zero_sock);
3355 }
3356
3357 /*
3358  * If required, resolve the source address for bind and leave the id_priv in
3359  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3360  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3361  * ignored.
3362  */
3363 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3364                                struct sockaddr *src_addr,
3365                                const struct sockaddr *dst_addr)
3366 {
3367         int ret;
3368
3369         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3370         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3371                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3372                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3373                 if (ret)
3374                         goto err_dst;
3375                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3376                                            RDMA_CM_ADDR_QUERY))) {
3377                         ret = -EINVAL;
3378                         goto err_dst;
3379                 }
3380         }
3381
3382         if (cma_family(id_priv) != dst_addr->sa_family) {
3383                 ret = -EINVAL;
3384                 goto err_state;
3385         }
3386         return 0;
3387
3388 err_state:
3389         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3390 err_dst:
3391         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3392         return ret;
3393 }
3394
3395 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3396                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3397 {
3398         struct rdma_id_private *id_priv =
3399                 container_of(id, struct rdma_id_private, id);
3400         int ret;
3401
3402         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3403         if (ret)
3404                 return ret;
3405
3406         if (cma_any_addr(dst_addr)) {
3407                 ret = cma_resolve_loopback(id_priv);
3408         } else {
3409                 if (dst_addr->sa_family == AF_IB) {
3410                         ret = cma_resolve_ib_addr(id_priv);
3411                 } else {
3412                         /*
3413                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
3414                          * rdma_resolve_ip() is called, eg through the error
3415                          * path in addr_handler(). If this happens the existing
3416                          * request must be canceled before issuing a new one.
3417                          * Since canceling a request is a bit slow and this
3418                          * oddball path is rare, keep track once a request has
3419                          * been issued. The track turns out to be a permanent
3420                          * state since this is the only cancel as it is
3421                          * immediately before rdma_resolve_ip().
3422                          */
3423                         if (id_priv->used_resolve_ip)
3424                                 rdma_addr_cancel(&id->route.addr.dev_addr);
3425                         else
3426                                 id_priv->used_resolve_ip = 1;
3427                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3428                                               &id->route.addr.dev_addr,
3429                                               timeout_ms, addr_handler,
3430                                               false, id_priv);
3431                 }
3432         }
3433         if (ret)
3434                 goto err;
3435
3436         return 0;
3437 err:
3438         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3439         return ret;
3440 }
3441 EXPORT_SYMBOL(rdma_resolve_addr);
3442
3443 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3444 {
3445         struct rdma_id_private *id_priv;
3446         unsigned long flags;
3447         int ret;
3448
3449         id_priv = container_of(id, struct rdma_id_private, id);
3450         spin_lock_irqsave(&id_priv->lock, flags);
3451         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3452             id_priv->state == RDMA_CM_IDLE) {
3453                 id_priv->reuseaddr = reuse;
3454                 ret = 0;
3455         } else {
3456                 ret = -EINVAL;
3457         }
3458         spin_unlock_irqrestore(&id_priv->lock, flags);
3459         return ret;
3460 }
3461 EXPORT_SYMBOL(rdma_set_reuseaddr);
3462
3463 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3464 {
3465         struct rdma_id_private *id_priv;
3466         unsigned long flags;
3467         int ret;
3468
3469         id_priv = container_of(id, struct rdma_id_private, id);
3470         spin_lock_irqsave(&id_priv->lock, flags);
3471         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3472                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3473                 id_priv->afonly = afonly;
3474                 ret = 0;
3475         } else {
3476                 ret = -EINVAL;
3477         }
3478         spin_unlock_irqrestore(&id_priv->lock, flags);
3479         return ret;
3480 }
3481 EXPORT_SYMBOL(rdma_set_afonly);
3482
3483 static void cma_bind_port(struct rdma_bind_list *bind_list,
3484                           struct rdma_id_private *id_priv)
3485 {
3486         struct sockaddr *addr;
3487         struct sockaddr_ib *sib;
3488         u64 sid, mask;
3489         __be16 port;
3490
3491         lockdep_assert_held(&lock);
3492
3493         addr = cma_src_addr(id_priv);
3494         port = htons(bind_list->port);
3495
3496         switch (addr->sa_family) {
3497         case AF_INET:
3498                 ((struct sockaddr_in *) addr)->sin_port = port;
3499                 break;
3500         case AF_INET6:
3501                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3502                 break;
3503         case AF_IB:
3504                 sib = (struct sockaddr_ib *) addr;
3505                 sid = be64_to_cpu(sib->sib_sid);
3506                 mask = be64_to_cpu(sib->sib_sid_mask);
3507                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3508                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3509                 break;
3510         }
3511         id_priv->bind_list = bind_list;
3512         hlist_add_head(&id_priv->node, &bind_list->owners);
3513 }
3514
3515 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3516                           struct rdma_id_private *id_priv, unsigned short snum)
3517 {
3518         struct rdma_bind_list *bind_list;
3519         int ret;
3520
3521         lockdep_assert_held(&lock);
3522
3523         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3524         if (!bind_list)
3525                 return -ENOMEM;
3526
3527         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3528                            snum);
3529         if (ret < 0)
3530                 goto err;
3531
3532         bind_list->ps = ps;
3533         bind_list->port = snum;
3534         cma_bind_port(bind_list, id_priv);
3535         return 0;
3536 err:
3537         kfree(bind_list);
3538         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3539 }
3540
3541 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3542                               struct rdma_id_private *id_priv)
3543 {
3544         struct rdma_id_private *cur_id;
3545         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3546         struct sockaddr  *saddr = cma_src_addr(id_priv);
3547         __be16 dport = cma_port(daddr);
3548
3549         lockdep_assert_held(&lock);
3550
3551         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3552                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3553                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3554                 __be16 cur_dport = cma_port(cur_daddr);
3555
3556                 if (id_priv == cur_id)
3557                         continue;
3558
3559                 /* different dest port -> unique */
3560                 if (!cma_any_port(daddr) &&
3561                     !cma_any_port(cur_daddr) &&
3562                     (dport != cur_dport))
3563                         continue;
3564
3565                 /* different src address -> unique */
3566                 if (!cma_any_addr(saddr) &&
3567                     !cma_any_addr(cur_saddr) &&
3568                     cma_addr_cmp(saddr, cur_saddr))
3569                         continue;
3570
3571                 /* different dst address -> unique */
3572                 if (!cma_any_addr(daddr) &&
3573                     !cma_any_addr(cur_daddr) &&
3574                     cma_addr_cmp(daddr, cur_daddr))
3575                         continue;
3576
3577                 return -EADDRNOTAVAIL;
3578         }
3579         return 0;
3580 }
3581
3582 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3583                               struct rdma_id_private *id_priv)
3584 {
3585         static unsigned int last_used_port;
3586         int low, high, remaining;
3587         unsigned int rover;
3588         struct net *net = id_priv->id.route.addr.dev_addr.net;
3589
3590         lockdep_assert_held(&lock);
3591
3592         inet_get_local_port_range(net, &low, &high);
3593         remaining = (high - low) + 1;
3594         rover = prandom_u32() % remaining + low;
3595 retry:
3596         if (last_used_port != rover) {
3597                 struct rdma_bind_list *bind_list;
3598                 int ret;
3599
3600                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3601
3602                 if (!bind_list) {
3603                         ret = cma_alloc_port(ps, id_priv, rover);
3604                 } else {
3605                         ret = cma_port_is_unique(bind_list, id_priv);
3606                         if (!ret)
3607                                 cma_bind_port(bind_list, id_priv);
3608                 }
3609                 /*
3610                  * Remember previously used port number in order to avoid
3611                  * re-using same port immediately after it is closed.
3612                  */
3613                 if (!ret)
3614                         last_used_port = rover;
3615                 if (ret != -EADDRNOTAVAIL)
3616                         return ret;
3617         }
3618         if (--remaining) {
3619                 rover++;
3620                 if ((rover < low) || (rover > high))
3621                         rover = low;
3622                 goto retry;
3623         }
3624         return -EADDRNOTAVAIL;
3625 }
3626
3627 /*
3628  * Check that the requested port is available.  This is called when trying to
3629  * bind to a specific port, or when trying to listen on a bound port.  In
3630  * the latter case, the provided id_priv may already be on the bind_list, but
3631  * we still need to check that it's okay to start listening.
3632  */
3633 static int cma_check_port(struct rdma_bind_list *bind_list,
3634                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3635 {
3636         struct rdma_id_private *cur_id;
3637         struct sockaddr *addr, *cur_addr;
3638
3639         lockdep_assert_held(&lock);
3640
3641         addr = cma_src_addr(id_priv);
3642         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3643                 if (id_priv == cur_id)
3644                         continue;
3645
3646                 if (reuseaddr && cur_id->reuseaddr)
3647                         continue;
3648
3649                 cur_addr = cma_src_addr(cur_id);
3650                 if (id_priv->afonly && cur_id->afonly &&
3651                     (addr->sa_family != cur_addr->sa_family))
3652                         continue;
3653
3654                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3655                         return -EADDRNOTAVAIL;
3656
3657                 if (!cma_addr_cmp(addr, cur_addr))
3658                         return -EADDRINUSE;
3659         }
3660         return 0;
3661 }
3662
3663 static int cma_use_port(enum rdma_ucm_port_space ps,
3664                         struct rdma_id_private *id_priv)
3665 {
3666         struct rdma_bind_list *bind_list;
3667         unsigned short snum;
3668         int ret;
3669
3670         lockdep_assert_held(&lock);
3671
3672         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3673         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3674                 return -EACCES;
3675
3676         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3677         if (!bind_list) {
3678                 ret = cma_alloc_port(ps, id_priv, snum);
3679         } else {
3680                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3681                 if (!ret)
3682                         cma_bind_port(bind_list, id_priv);
3683         }
3684         return ret;
3685 }
3686
3687 static enum rdma_ucm_port_space
3688 cma_select_inet_ps(struct rdma_id_private *id_priv)
3689 {
3690         switch (id_priv->id.ps) {
3691         case RDMA_PS_TCP:
3692         case RDMA_PS_UDP:
3693         case RDMA_PS_IPOIB:
3694         case RDMA_PS_IB:
3695                 return id_priv->id.ps;
3696         default:
3697
3698                 return 0;
3699         }
3700 }
3701
3702 static enum rdma_ucm_port_space
3703 cma_select_ib_ps(struct rdma_id_private *id_priv)
3704 {
3705         enum rdma_ucm_port_space ps = 0;
3706         struct sockaddr_ib *sib;
3707         u64 sid_ps, mask, sid;
3708
3709         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3710         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3711         sid = be64_to_cpu(sib->sib_sid) & mask;
3712
3713         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3714                 sid_ps = RDMA_IB_IP_PS_IB;
3715                 ps = RDMA_PS_IB;
3716         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3717                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3718                 sid_ps = RDMA_IB_IP_PS_TCP;
3719                 ps = RDMA_PS_TCP;
3720         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3721                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3722                 sid_ps = RDMA_IB_IP_PS_UDP;
3723                 ps = RDMA_PS_UDP;
3724         }
3725
3726         if (ps) {
3727                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3728                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3729                                                 be64_to_cpu(sib->sib_sid_mask));
3730         }
3731         return ps;
3732 }
3733
3734 static int cma_get_port(struct rdma_id_private *id_priv)
3735 {
3736         enum rdma_ucm_port_space ps;
3737         int ret;
3738
3739         if (cma_family(id_priv) != AF_IB)
3740                 ps = cma_select_inet_ps(id_priv);
3741         else
3742                 ps = cma_select_ib_ps(id_priv);
3743         if (!ps)
3744                 return -EPROTONOSUPPORT;
3745
3746         mutex_lock(&lock);
3747         if (cma_any_port(cma_src_addr(id_priv)))
3748                 ret = cma_alloc_any_port(ps, id_priv);
3749         else
3750                 ret = cma_use_port(ps, id_priv);
3751         mutex_unlock(&lock);
3752
3753         return ret;
3754 }
3755
3756 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3757                                struct sockaddr *addr)
3758 {
3759 #if IS_ENABLED(CONFIG_IPV6)
3760         struct sockaddr_in6 *sin6;
3761
3762         if (addr->sa_family != AF_INET6)
3763                 return 0;
3764
3765         sin6 = (struct sockaddr_in6 *) addr;
3766
3767         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3768                 return 0;
3769
3770         if (!sin6->sin6_scope_id)
3771                         return -EINVAL;
3772
3773         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3774 #endif
3775         return 0;
3776 }
3777
3778 int rdma_listen(struct rdma_cm_id *id, int backlog)
3779 {
3780         struct rdma_id_private *id_priv =
3781                 container_of(id, struct rdma_id_private, id);
3782         int ret;
3783
3784         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3785                 struct sockaddr_in any_in = {
3786                         .sin_family = AF_INET,
3787                         .sin_addr.s_addr = htonl(INADDR_ANY),
3788                 };
3789
3790                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3791                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3792                 if (ret)
3793                         return ret;
3794                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3795                                            RDMA_CM_LISTEN)))
3796                         return -EINVAL;
3797         }
3798
3799         /*
3800          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3801          * any more, and has to be unique in the bind list.
3802          */
3803         if (id_priv->reuseaddr) {
3804                 mutex_lock(&lock);
3805                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3806                 if (!ret)
3807                         id_priv->reuseaddr = 0;
3808                 mutex_unlock(&lock);
3809                 if (ret)
3810                         goto err;
3811         }
3812
3813         id_priv->backlog = backlog;
3814         if (id_priv->cma_dev) {
3815                 if (rdma_cap_ib_cm(id->device, 1)) {
3816                         ret = cma_ib_listen(id_priv);
3817                         if (ret)
3818                                 goto err;
3819                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3820                         ret = cma_iw_listen(id_priv, backlog);
3821                         if (ret)
3822                                 goto err;
3823                 } else {
3824                         ret = -ENOSYS;
3825                         goto err;
3826                 }
3827         } else {
3828                 ret = cma_listen_on_all(id_priv);
3829                 if (ret)
3830                         goto err;
3831         }
3832
3833         return 0;
3834 err:
3835         id_priv->backlog = 0;
3836         /*
3837          * All the failure paths that lead here will not allow the req_handler's
3838          * to have run.
3839          */
3840         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3841         return ret;
3842 }
3843 EXPORT_SYMBOL(rdma_listen);
3844
3845 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3846 {
3847         struct rdma_id_private *id_priv;
3848         int ret;
3849         struct sockaddr  *daddr;
3850
3851         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3852             addr->sa_family != AF_IB)
3853                 return -EAFNOSUPPORT;
3854
3855         id_priv = container_of(id, struct rdma_id_private, id);
3856         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3857                 return -EINVAL;
3858
3859         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3860         if (ret)
3861                 goto err1;
3862
3863         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3864         if (!cma_any_addr(addr)) {
3865                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3866                 if (ret)
3867                         goto err1;
3868
3869                 ret = cma_acquire_dev_by_src_ip(id_priv);
3870                 if (ret)
3871                         goto err1;
3872         }
3873
3874         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3875                 if (addr->sa_family == AF_INET)
3876                         id_priv->afonly = 1;
3877 #if IS_ENABLED(CONFIG_IPV6)
3878                 else if (addr->sa_family == AF_INET6) {
3879                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3880
3881                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3882                 }
3883 #endif
3884         }
3885         daddr = cma_dst_addr(id_priv);
3886         daddr->sa_family = addr->sa_family;
3887
3888         ret = cma_get_port(id_priv);
3889         if (ret)
3890                 goto err2;
3891
3892         if (!cma_any_addr(addr))
3893                 rdma_restrack_add(&id_priv->res);
3894         return 0;
3895 err2:
3896         if (id_priv->cma_dev)
3897                 cma_release_dev(id_priv);
3898 err1:
3899         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3900         return ret;
3901 }
3902 EXPORT_SYMBOL(rdma_bind_addr);
3903
3904 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3905 {
3906         struct cma_hdr *cma_hdr;
3907
3908         cma_hdr = hdr;
3909         cma_hdr->cma_version = CMA_VERSION;
3910         if (cma_family(id_priv) == AF_INET) {
3911                 struct sockaddr_in *src4, *dst4;
3912
3913                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3914                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3915
3916                 cma_set_ip_ver(cma_hdr, 4);
3917                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3918                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3919                 cma_hdr->port = src4->sin_port;
3920         } else if (cma_family(id_priv) == AF_INET6) {
3921                 struct sockaddr_in6 *src6, *dst6;
3922
3923                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3924                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3925
3926                 cma_set_ip_ver(cma_hdr, 6);
3927                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3928                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3929                 cma_hdr->port = src6->sin6_port;
3930         }
3931         return 0;
3932 }
3933
3934 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3935                                 const struct ib_cm_event *ib_event)
3936 {
3937         struct rdma_id_private *id_priv = cm_id->context;
3938         struct rdma_cm_event event = {};
3939         const struct ib_cm_sidr_rep_event_param *rep =
3940                                 &ib_event->param.sidr_rep_rcvd;
3941         int ret;
3942
3943         mutex_lock(&id_priv->handler_mutex);
3944         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3945                 goto out;
3946
3947         switch (ib_event->event) {
3948         case IB_CM_SIDR_REQ_ERROR:
3949                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3950                 event.status = -ETIMEDOUT;
3951                 break;
3952         case IB_CM_SIDR_REP_RECEIVED:
3953                 event.param.ud.private_data = ib_event->private_data;
3954                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3955                 if (rep->status != IB_SIDR_SUCCESS) {
3956                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3957                         event.status = ib_event->param.sidr_rep_rcvd.status;
3958                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3959                                              event.status);
3960                         break;
3961                 }
3962                 ret = cma_set_qkey(id_priv, rep->qkey);
3963                 if (ret) {
3964                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3965                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3966                         event.status = ret;
3967                         break;
3968                 }
3969                 ib_init_ah_attr_from_path(id_priv->id.device,
3970                                           id_priv->id.port_num,
3971                                           id_priv->id.route.path_rec,
3972                                           &event.param.ud.ah_attr,
3973                                           rep->sgid_attr);
3974                 event.param.ud.qp_num = rep->qpn;
3975                 event.param.ud.qkey = rep->qkey;
3976                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3977                 event.status = 0;
3978                 break;
3979         default:
3980                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3981                        ib_event->event);
3982                 goto out;
3983         }
3984
3985         ret = cma_cm_event_handler(id_priv, &event);
3986
3987         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3988         if (ret) {
3989                 /* Destroy the CM ID by returning a non-zero value. */
3990                 id_priv->cm_id.ib = NULL;
3991                 destroy_id_handler_unlock(id_priv);
3992                 return ret;
3993         }
3994 out:
3995         mutex_unlock(&id_priv->handler_mutex);
3996         return 0;
3997 }
3998
3999 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4000                               struct rdma_conn_param *conn_param)
4001 {
4002         struct ib_cm_sidr_req_param req;
4003         struct ib_cm_id *id;
4004         void *private_data;
4005         u8 offset;
4006         int ret;
4007
4008         memset(&req, 0, sizeof req);
4009         offset = cma_user_data_offset(id_priv);
4010         req.private_data_len = offset + conn_param->private_data_len;
4011         if (req.private_data_len < conn_param->private_data_len)
4012                 return -EINVAL;
4013
4014         if (req.private_data_len) {
4015                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4016                 if (!private_data)
4017                         return -ENOMEM;
4018         } else {
4019                 private_data = NULL;
4020         }
4021
4022         if (conn_param->private_data && conn_param->private_data_len)
4023                 memcpy(private_data + offset, conn_param->private_data,
4024                        conn_param->private_data_len);
4025
4026         if (private_data) {
4027                 ret = cma_format_hdr(private_data, id_priv);
4028                 if (ret)
4029                         goto out;
4030                 req.private_data = private_data;
4031         }
4032
4033         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4034                              id_priv);
4035         if (IS_ERR(id)) {
4036                 ret = PTR_ERR(id);
4037                 goto out;
4038         }
4039         id_priv->cm_id.ib = id;
4040
4041         req.path = id_priv->id.route.path_rec;
4042         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4043         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4044         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4045         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4046
4047         trace_cm_send_sidr_req(id_priv);
4048         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4049         if (ret) {
4050                 ib_destroy_cm_id(id_priv->cm_id.ib);
4051                 id_priv->cm_id.ib = NULL;
4052         }
4053 out:
4054         kfree(private_data);
4055         return ret;
4056 }
4057
4058 static int cma_connect_ib(struct rdma_id_private *id_priv,
4059                           struct rdma_conn_param *conn_param)
4060 {
4061         struct ib_cm_req_param req;
4062         struct rdma_route *route;
4063         void *private_data;
4064         struct ib_cm_id *id;
4065         u8 offset;
4066         int ret;
4067
4068         memset(&req, 0, sizeof req);
4069         offset = cma_user_data_offset(id_priv);
4070         req.private_data_len = offset + conn_param->private_data_len;
4071         if (req.private_data_len < conn_param->private_data_len)
4072                 return -EINVAL;
4073
4074         if (req.private_data_len) {
4075                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4076                 if (!private_data)
4077                         return -ENOMEM;
4078         } else {
4079                 private_data = NULL;
4080         }
4081
4082         if (conn_param->private_data && conn_param->private_data_len)
4083                 memcpy(private_data + offset, conn_param->private_data,
4084                        conn_param->private_data_len);
4085
4086         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4087         if (IS_ERR(id)) {
4088                 ret = PTR_ERR(id);
4089                 goto out;
4090         }
4091         id_priv->cm_id.ib = id;
4092
4093         route = &id_priv->id.route;
4094         if (private_data) {
4095                 ret = cma_format_hdr(private_data, id_priv);
4096                 if (ret)
4097                         goto out;
4098                 req.private_data = private_data;
4099         }
4100
4101         req.primary_path = &route->path_rec[0];
4102         if (route->num_paths == 2)
4103                 req.alternate_path = &route->path_rec[1];
4104
4105         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4106         /* Alternate path SGID attribute currently unsupported */
4107         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4108         req.qp_num = id_priv->qp_num;
4109         req.qp_type = id_priv->id.qp_type;
4110         req.starting_psn = id_priv->seq_num;
4111         req.responder_resources = conn_param->responder_resources;
4112         req.initiator_depth = conn_param->initiator_depth;
4113         req.flow_control = conn_param->flow_control;
4114         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4115         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4116         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4117         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4118         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4119         req.srq = id_priv->srq ? 1 : 0;
4120         req.ece.vendor_id = id_priv->ece.vendor_id;
4121         req.ece.attr_mod = id_priv->ece.attr_mod;
4122
4123         trace_cm_send_req(id_priv);
4124         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4125 out:
4126         if (ret && !IS_ERR(id)) {
4127                 ib_destroy_cm_id(id);
4128                 id_priv->cm_id.ib = NULL;
4129         }
4130
4131         kfree(private_data);
4132         return ret;
4133 }
4134
4135 static int cma_connect_iw(struct rdma_id_private *id_priv,
4136                           struct rdma_conn_param *conn_param)
4137 {
4138         struct iw_cm_id *cm_id;
4139         int ret;
4140         struct iw_cm_conn_param iw_param;
4141
4142         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4143         if (IS_ERR(cm_id))
4144                 return PTR_ERR(cm_id);
4145
4146         mutex_lock(&id_priv->qp_mutex);
4147         cm_id->tos = id_priv->tos;
4148         cm_id->tos_set = id_priv->tos_set;
4149         mutex_unlock(&id_priv->qp_mutex);
4150
4151         id_priv->cm_id.iw = cm_id;
4152
4153         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4154                rdma_addr_size(cma_src_addr(id_priv)));
4155         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4156                rdma_addr_size(cma_dst_addr(id_priv)));
4157
4158         ret = cma_modify_qp_rtr(id_priv, conn_param);
4159         if (ret)
4160                 goto out;
4161
4162         if (conn_param) {
4163                 iw_param.ord = conn_param->initiator_depth;
4164                 iw_param.ird = conn_param->responder_resources;
4165                 iw_param.private_data = conn_param->private_data;
4166                 iw_param.private_data_len = conn_param->private_data_len;
4167                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4168         } else {
4169                 memset(&iw_param, 0, sizeof iw_param);
4170                 iw_param.qpn = id_priv->qp_num;
4171         }
4172         ret = iw_cm_connect(cm_id, &iw_param);
4173 out:
4174         if (ret) {
4175                 iw_destroy_cm_id(cm_id);
4176                 id_priv->cm_id.iw = NULL;
4177         }
4178         return ret;
4179 }
4180
4181 /**
4182  * rdma_connect_locked - Initiate an active connection request.
4183  * @id: Connection identifier to connect.
4184  * @conn_param: Connection information used for connected QPs.
4185  *
4186  * Same as rdma_connect() but can only be called from the
4187  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4188  */
4189 int rdma_connect_locked(struct rdma_cm_id *id,
4190                         struct rdma_conn_param *conn_param)
4191 {
4192         struct rdma_id_private *id_priv =
4193                 container_of(id, struct rdma_id_private, id);
4194         int ret;
4195
4196         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4197                 return -EINVAL;
4198
4199         if (!id->qp) {
4200                 id_priv->qp_num = conn_param->qp_num;
4201                 id_priv->srq = conn_param->srq;
4202         }
4203
4204         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4205                 if (id->qp_type == IB_QPT_UD)
4206                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4207                 else
4208                         ret = cma_connect_ib(id_priv, conn_param);
4209         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4210                 ret = cma_connect_iw(id_priv, conn_param);
4211         else
4212                 ret = -ENOSYS;
4213         if (ret)
4214                 goto err_state;
4215         return 0;
4216 err_state:
4217         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4218         return ret;
4219 }
4220 EXPORT_SYMBOL(rdma_connect_locked);
4221
4222 /**
4223  * rdma_connect - Initiate an active connection request.
4224  * @id: Connection identifier to connect.
4225  * @conn_param: Connection information used for connected QPs.
4226  *
4227  * Users must have resolved a route for the rdma_cm_id to connect with by having
4228  * called rdma_resolve_route before calling this routine.
4229  *
4230  * This call will either connect to a remote QP or obtain remote QP information
4231  * for unconnected rdma_cm_id's.  The actual operation is based on the
4232  * rdma_cm_id's port space.
4233  */
4234 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4235 {
4236         struct rdma_id_private *id_priv =
4237                 container_of(id, struct rdma_id_private, id);
4238         int ret;
4239
4240         mutex_lock(&id_priv->handler_mutex);
4241         ret = rdma_connect_locked(id, conn_param);
4242         mutex_unlock(&id_priv->handler_mutex);
4243         return ret;
4244 }
4245 EXPORT_SYMBOL(rdma_connect);
4246
4247 /**
4248  * rdma_connect_ece - Initiate an active connection request with ECE data.
4249  * @id: Connection identifier to connect.
4250  * @conn_param: Connection information used for connected QPs.
4251  * @ece: ECE parameters
4252  *
4253  * See rdma_connect() explanation.
4254  */
4255 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4256                      struct rdma_ucm_ece *ece)
4257 {
4258         struct rdma_id_private *id_priv =
4259                 container_of(id, struct rdma_id_private, id);
4260
4261         id_priv->ece.vendor_id = ece->vendor_id;
4262         id_priv->ece.attr_mod = ece->attr_mod;
4263
4264         return rdma_connect(id, conn_param);
4265 }
4266 EXPORT_SYMBOL(rdma_connect_ece);
4267
4268 static int cma_accept_ib(struct rdma_id_private *id_priv,
4269                          struct rdma_conn_param *conn_param)
4270 {
4271         struct ib_cm_rep_param rep;
4272         int ret;
4273
4274         ret = cma_modify_qp_rtr(id_priv, conn_param);
4275         if (ret)
4276                 goto out;
4277
4278         ret = cma_modify_qp_rts(id_priv, conn_param);
4279         if (ret)
4280                 goto out;
4281
4282         memset(&rep, 0, sizeof rep);
4283         rep.qp_num = id_priv->qp_num;
4284         rep.starting_psn = id_priv->seq_num;
4285         rep.private_data = conn_param->private_data;
4286         rep.private_data_len = conn_param->private_data_len;
4287         rep.responder_resources = conn_param->responder_resources;
4288         rep.initiator_depth = conn_param->initiator_depth;
4289         rep.failover_accepted = 0;
4290         rep.flow_control = conn_param->flow_control;
4291         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4292         rep.srq = id_priv->srq ? 1 : 0;
4293         rep.ece.vendor_id = id_priv->ece.vendor_id;
4294         rep.ece.attr_mod = id_priv->ece.attr_mod;
4295
4296         trace_cm_send_rep(id_priv);
4297         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4298 out:
4299         return ret;
4300 }
4301
4302 static int cma_accept_iw(struct rdma_id_private *id_priv,
4303                   struct rdma_conn_param *conn_param)
4304 {
4305         struct iw_cm_conn_param iw_param;
4306         int ret;
4307
4308         if (!conn_param)
4309                 return -EINVAL;
4310
4311         ret = cma_modify_qp_rtr(id_priv, conn_param);
4312         if (ret)
4313                 return ret;
4314
4315         iw_param.ord = conn_param->initiator_depth;
4316         iw_param.ird = conn_param->responder_resources;
4317         iw_param.private_data = conn_param->private_data;
4318         iw_param.private_data_len = conn_param->private_data_len;
4319         if (id_priv->id.qp) {
4320                 iw_param.qpn = id_priv->qp_num;
4321         } else
4322                 iw_param.qpn = conn_param->qp_num;
4323
4324         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4325 }
4326
4327 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4328                              enum ib_cm_sidr_status status, u32 qkey,
4329                              const void *private_data, int private_data_len)
4330 {
4331         struct ib_cm_sidr_rep_param rep;
4332         int ret;
4333
4334         memset(&rep, 0, sizeof rep);
4335         rep.status = status;
4336         if (status == IB_SIDR_SUCCESS) {
4337                 if (qkey)
4338                         ret = cma_set_qkey(id_priv, qkey);
4339                 else
4340                         ret = cma_set_default_qkey(id_priv);
4341                 if (ret)
4342                         return ret;
4343                 rep.qp_num = id_priv->qp_num;
4344                 rep.qkey = id_priv->qkey;
4345
4346                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4347                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4348         }
4349
4350         rep.private_data = private_data;
4351         rep.private_data_len = private_data_len;
4352
4353         trace_cm_send_sidr_rep(id_priv);
4354         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4355 }
4356
4357 /**
4358  * rdma_accept - Called to accept a connection request or response.
4359  * @id: Connection identifier associated with the request.
4360  * @conn_param: Information needed to establish the connection.  This must be
4361  *   provided if accepting a connection request.  If accepting a connection
4362  *   response, this parameter must be NULL.
4363  *
4364  * Typically, this routine is only called by the listener to accept a connection
4365  * request.  It must also be called on the active side of a connection if the
4366  * user is performing their own QP transitions.
4367  *
4368  * In the case of error, a reject message is sent to the remote side and the
4369  * state of the qp associated with the id is modified to error, such that any
4370  * previously posted receive buffers would be flushed.
4371  *
4372  * This function is for use by kernel ULPs and must be called from under the
4373  * handler callback.
4374  */
4375 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4376 {
4377         struct rdma_id_private *id_priv =
4378                 container_of(id, struct rdma_id_private, id);
4379         int ret;
4380
4381         lockdep_assert_held(&id_priv->handler_mutex);
4382
4383         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4384                 return -EINVAL;
4385
4386         if (!id->qp && conn_param) {
4387                 id_priv->qp_num = conn_param->qp_num;
4388                 id_priv->srq = conn_param->srq;
4389         }
4390
4391         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4392                 if (id->qp_type == IB_QPT_UD) {
4393                         if (conn_param)
4394                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4395                                                         conn_param->qkey,
4396                                                         conn_param->private_data,
4397                                                         conn_param->private_data_len);
4398                         else
4399                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4400                                                         0, NULL, 0);
4401                 } else {
4402                         if (conn_param)
4403                                 ret = cma_accept_ib(id_priv, conn_param);
4404                         else
4405                                 ret = cma_rep_recv(id_priv);
4406                 }
4407         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4408                 ret = cma_accept_iw(id_priv, conn_param);
4409         else
4410                 ret = -ENOSYS;
4411
4412         if (ret)
4413                 goto reject;
4414
4415         return 0;
4416 reject:
4417         cma_modify_qp_err(id_priv);
4418         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4419         return ret;
4420 }
4421 EXPORT_SYMBOL(rdma_accept);
4422
4423 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4424                     struct rdma_ucm_ece *ece)
4425 {
4426         struct rdma_id_private *id_priv =
4427                 container_of(id, struct rdma_id_private, id);
4428
4429         id_priv->ece.vendor_id = ece->vendor_id;
4430         id_priv->ece.attr_mod = ece->attr_mod;
4431
4432         return rdma_accept(id, conn_param);
4433 }
4434 EXPORT_SYMBOL(rdma_accept_ece);
4435
4436 void rdma_lock_handler(struct rdma_cm_id *id)
4437 {
4438         struct rdma_id_private *id_priv =
4439                 container_of(id, struct rdma_id_private, id);
4440
4441         mutex_lock(&id_priv->handler_mutex);
4442 }
4443 EXPORT_SYMBOL(rdma_lock_handler);
4444
4445 void rdma_unlock_handler(struct rdma_cm_id *id)
4446 {
4447         struct rdma_id_private *id_priv =
4448                 container_of(id, struct rdma_id_private, id);
4449
4450         mutex_unlock(&id_priv->handler_mutex);
4451 }
4452 EXPORT_SYMBOL(rdma_unlock_handler);
4453
4454 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4455 {
4456         struct rdma_id_private *id_priv;
4457         int ret;
4458
4459         id_priv = container_of(id, struct rdma_id_private, id);
4460         if (!id_priv->cm_id.ib)
4461                 return -EINVAL;
4462
4463         switch (id->device->node_type) {
4464         case RDMA_NODE_IB_CA:
4465                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4466                 break;
4467         default:
4468                 ret = 0;
4469                 break;
4470         }
4471         return ret;
4472 }
4473 EXPORT_SYMBOL(rdma_notify);
4474
4475 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4476                 u8 private_data_len, u8 reason)
4477 {
4478         struct rdma_id_private *id_priv;
4479         int ret;
4480
4481         id_priv = container_of(id, struct rdma_id_private, id);
4482         if (!id_priv->cm_id.ib)
4483                 return -EINVAL;
4484
4485         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4486                 if (id->qp_type == IB_QPT_UD) {
4487                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4488                                                 private_data, private_data_len);
4489                 } else {
4490                         trace_cm_send_rej(id_priv);
4491                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4492                                              private_data, private_data_len);
4493                 }
4494         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4495                 ret = iw_cm_reject(id_priv->cm_id.iw,
4496                                    private_data, private_data_len);
4497         } else
4498                 ret = -ENOSYS;
4499
4500         return ret;
4501 }
4502 EXPORT_SYMBOL(rdma_reject);
4503
4504 int rdma_disconnect(struct rdma_cm_id *id)
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                 ret = cma_modify_qp_err(id_priv);
4515                 if (ret)
4516                         goto out;
4517                 /* Initiate or respond to a disconnect. */
4518                 trace_cm_disconnect(id_priv);
4519                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4520                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4521                                 trace_cm_sent_drep(id_priv);
4522                 } else {
4523                         trace_cm_sent_dreq(id_priv);
4524                 }
4525         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4526                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4527         } else
4528                 ret = -EINVAL;
4529
4530 out:
4531         return ret;
4532 }
4533 EXPORT_SYMBOL(rdma_disconnect);
4534
4535 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4536                               struct ib_sa_multicast *multicast,
4537                               struct rdma_cm_event *event,
4538                               struct cma_multicast *mc)
4539 {
4540         struct rdma_dev_addr *dev_addr;
4541         enum ib_gid_type gid_type;
4542         struct net_device *ndev;
4543
4544         if (status)
4545                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4546                                      status);
4547
4548         event->status = status;
4549         event->param.ud.private_data = mc->context;
4550         if (status) {
4551                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4552                 return;
4553         }
4554
4555         dev_addr = &id_priv->id.route.addr.dev_addr;
4556         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4557         gid_type =
4558                 id_priv->cma_dev
4559                         ->default_gid_type[id_priv->id.port_num -
4560                                            rdma_start_port(
4561                                                    id_priv->cma_dev->device)];
4562
4563         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4564         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4565                                      &multicast->rec, ndev, gid_type,
4566                                      &event->param.ud.ah_attr)) {
4567                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4568                 goto out;
4569         }
4570
4571         event->param.ud.qp_num = 0xFFFFFF;
4572         event->param.ud.qkey = id_priv->qkey;
4573
4574 out:
4575         if (ndev)
4576                 dev_put(ndev);
4577 }
4578
4579 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4580 {
4581         struct cma_multicast *mc = multicast->context;
4582         struct rdma_id_private *id_priv = mc->id_priv;
4583         struct rdma_cm_event event = {};
4584         int ret = 0;
4585
4586         mutex_lock(&id_priv->handler_mutex);
4587         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4588             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4589                 goto out;
4590
4591         ret = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4592         if (!ret) {
4593                 cma_make_mc_event(status, id_priv, multicast, &event, mc);
4594                 ret = cma_cm_event_handler(id_priv, &event);
4595         }
4596         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4597         WARN_ON(ret);
4598
4599 out:
4600         mutex_unlock(&id_priv->handler_mutex);
4601         return 0;
4602 }
4603
4604 static void cma_set_mgid(struct rdma_id_private *id_priv,
4605                          struct sockaddr *addr, union ib_gid *mgid)
4606 {
4607         unsigned char mc_map[MAX_ADDR_LEN];
4608         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4609         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4610         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4611
4612         if (cma_any_addr(addr)) {
4613                 memset(mgid, 0, sizeof *mgid);
4614         } else if ((addr->sa_family == AF_INET6) &&
4615                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4616                                                                  0xFF10A01B)) {
4617                 /* IPv6 address is an SA assigned MGID. */
4618                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4619         } else if (addr->sa_family == AF_IB) {
4620                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4621         } else if (addr->sa_family == AF_INET6) {
4622                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4623                 if (id_priv->id.ps == RDMA_PS_UDP)
4624                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4625                 *mgid = *(union ib_gid *) (mc_map + 4);
4626         } else {
4627                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4628                 if (id_priv->id.ps == RDMA_PS_UDP)
4629                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4630                 *mgid = *(union ib_gid *) (mc_map + 4);
4631         }
4632 }
4633
4634 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4635                                  struct cma_multicast *mc)
4636 {
4637         struct ib_sa_mcmember_rec rec;
4638         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4639         ib_sa_comp_mask comp_mask;
4640         int ret;
4641
4642         ib_addr_get_mgid(dev_addr, &rec.mgid);
4643         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4644                                      &rec.mgid, &rec);
4645         if (ret)
4646                 return ret;
4647
4648         if (!id_priv->qkey) {
4649                 ret = cma_set_default_qkey(id_priv);
4650                 if (ret)
4651                         return ret;
4652         }
4653
4654         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4655         rec.qkey = cpu_to_be32(id_priv->qkey);
4656         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4657         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4658         rec.join_state = mc->join_state;
4659
4660         if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4661             (!ib_sa_sendonly_fullmem_support(&sa_client,
4662                                              id_priv->id.device,
4663                                              id_priv->id.port_num))) {
4664                 dev_warn(
4665                         &id_priv->id.device->dev,
4666                         "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4667                         id_priv->id.port_num);
4668                 return -EOPNOTSUPP;
4669         }
4670
4671         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4672                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4673                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4674                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4675                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4676
4677         if (id_priv->id.ps == RDMA_PS_IPOIB)
4678                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4679                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4680                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4681                              IB_SA_MCMEMBER_REC_MTU |
4682                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4683
4684         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4685                                          id_priv->id.port_num, &rec, comp_mask,
4686                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4687         return PTR_ERR_OR_ZERO(mc->sa_mc);
4688 }
4689
4690 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4691                               enum ib_gid_type gid_type)
4692 {
4693         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4694         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4695
4696         if (cma_any_addr(addr)) {
4697                 memset(mgid, 0, sizeof *mgid);
4698         } else if (addr->sa_family == AF_INET6) {
4699                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4700         } else {
4701                 mgid->raw[0] =
4702                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4703                 mgid->raw[1] =
4704                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4705                 mgid->raw[2] = 0;
4706                 mgid->raw[3] = 0;
4707                 mgid->raw[4] = 0;
4708                 mgid->raw[5] = 0;
4709                 mgid->raw[6] = 0;
4710                 mgid->raw[7] = 0;
4711                 mgid->raw[8] = 0;
4712                 mgid->raw[9] = 0;
4713                 mgid->raw[10] = 0xff;
4714                 mgid->raw[11] = 0xff;
4715                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4716         }
4717 }
4718
4719 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4720                                    struct cma_multicast *mc)
4721 {
4722         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4723         int err = 0;
4724         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4725         struct net_device *ndev = NULL;
4726         struct ib_sa_multicast ib = {};
4727         enum ib_gid_type gid_type;
4728         bool send_only;
4729
4730         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4731
4732         if (cma_zero_addr(addr))
4733                 return -EINVAL;
4734
4735         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4736                    rdma_start_port(id_priv->cma_dev->device)];
4737         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4738
4739         ib.rec.pkey = cpu_to_be16(0xffff);
4740         if (dev_addr->bound_dev_if)
4741                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4742         if (!ndev)
4743                 return -ENODEV;
4744
4745         ib.rec.rate = IB_RATE_PORT_CURRENT;
4746         ib.rec.hop_limit = 1;
4747         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4748
4749         if (addr->sa_family == AF_INET) {
4750                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4751                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4752                         if (!send_only) {
4753                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4754                                                     true);
4755                         }
4756                 }
4757         } else {
4758                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4759                         err = -ENOTSUPP;
4760         }
4761         dev_put(ndev);
4762         if (err || !ib.rec.mtu)
4763                 return err ?: -EINVAL;
4764
4765         if (!id_priv->qkey)
4766                 cma_set_default_qkey(id_priv);
4767
4768         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4769                     &ib.rec.port_gid);
4770         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4771         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4772         queue_work(cma_wq, &mc->iboe_join.work);
4773         return 0;
4774 }
4775
4776 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4777                         u8 join_state, void *context)
4778 {
4779         struct rdma_id_private *id_priv =
4780                 container_of(id, struct rdma_id_private, id);
4781         struct cma_multicast *mc;
4782         int ret;
4783
4784         /* Not supported for kernel QPs */
4785         if (WARN_ON(id->qp))
4786                 return -EINVAL;
4787
4788         /* ULP is calling this wrong. */
4789         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4790                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4791                 return -EINVAL;
4792
4793         if (id_priv->id.qp_type != IB_QPT_UD)
4794                 return -EINVAL;
4795
4796         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4797         if (!mc)
4798                 return -ENOMEM;
4799
4800         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4801         mc->context = context;
4802         mc->id_priv = id_priv;
4803         mc->join_state = join_state;
4804
4805         if (rdma_protocol_roce(id->device, id->port_num)) {
4806                 ret = cma_iboe_join_multicast(id_priv, mc);
4807                 if (ret)
4808                         goto out_err;
4809         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4810                 ret = cma_join_ib_multicast(id_priv, mc);
4811                 if (ret)
4812                         goto out_err;
4813         } else {
4814                 ret = -ENOSYS;
4815                 goto out_err;
4816         }
4817
4818         spin_lock(&id_priv->lock);
4819         list_add(&mc->list, &id_priv->mc_list);
4820         spin_unlock(&id_priv->lock);
4821
4822         return 0;
4823 out_err:
4824         kfree(mc);
4825         return ret;
4826 }
4827 EXPORT_SYMBOL(rdma_join_multicast);
4828
4829 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4830 {
4831         struct rdma_id_private *id_priv;
4832         struct cma_multicast *mc;
4833
4834         id_priv = container_of(id, struct rdma_id_private, id);
4835         spin_lock_irq(&id_priv->lock);
4836         list_for_each_entry(mc, &id_priv->mc_list, list) {
4837                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4838                         continue;
4839                 list_del(&mc->list);
4840                 spin_unlock_irq(&id_priv->lock);
4841
4842                 WARN_ON(id_priv->cma_dev->device != id->device);
4843                 destroy_mc(id_priv, mc);
4844                 return;
4845         }
4846         spin_unlock_irq(&id_priv->lock);
4847 }
4848 EXPORT_SYMBOL(rdma_leave_multicast);
4849
4850 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4851 {
4852         struct rdma_dev_addr *dev_addr;
4853         struct cma_work *work;
4854
4855         dev_addr = &id_priv->id.route.addr.dev_addr;
4856
4857         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4858             (net_eq(dev_net(ndev), dev_addr->net)) &&
4859             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4860                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4861                         ndev->name, &id_priv->id);
4862                 work = kzalloc(sizeof *work, GFP_KERNEL);
4863                 if (!work)
4864                         return -ENOMEM;
4865
4866                 INIT_WORK(&work->work, cma_work_handler);
4867                 work->id = id_priv;
4868                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4869                 cma_id_get(id_priv);
4870                 queue_work(cma_wq, &work->work);
4871         }
4872
4873         return 0;
4874 }
4875
4876 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4877                                void *ptr)
4878 {
4879         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4880         struct cma_device *cma_dev;
4881         struct rdma_id_private *id_priv;
4882         int ret = NOTIFY_DONE;
4883
4884         if (event != NETDEV_BONDING_FAILOVER)
4885                 return NOTIFY_DONE;
4886
4887         if (!netif_is_bond_master(ndev))
4888                 return NOTIFY_DONE;
4889
4890         mutex_lock(&lock);
4891         list_for_each_entry(cma_dev, &dev_list, list)
4892                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4893                         ret = cma_netdev_change(ndev, id_priv);
4894                         if (ret)
4895                                 goto out;
4896                 }
4897
4898 out:
4899         mutex_unlock(&lock);
4900         return ret;
4901 }
4902
4903 static struct notifier_block cma_nb = {
4904         .notifier_call = cma_netdev_callback
4905 };
4906
4907 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4908 {
4909         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4910         enum rdma_cm_state state;
4911         unsigned long flags;
4912
4913         mutex_lock(&id_priv->handler_mutex);
4914         /* Record that we want to remove the device */
4915         spin_lock_irqsave(&id_priv->lock, flags);
4916         state = id_priv->state;
4917         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4918                 spin_unlock_irqrestore(&id_priv->lock, flags);
4919                 mutex_unlock(&id_priv->handler_mutex);
4920                 cma_id_put(id_priv);
4921                 return;
4922         }
4923         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4924         spin_unlock_irqrestore(&id_priv->lock, flags);
4925
4926         if (cma_cm_event_handler(id_priv, &event)) {
4927                 /*
4928                  * At this point the ULP promises it won't call
4929                  * rdma_destroy_id() concurrently
4930                  */
4931                 cma_id_put(id_priv);
4932                 mutex_unlock(&id_priv->handler_mutex);
4933                 trace_cm_id_destroy(id_priv);
4934                 _destroy_id(id_priv, state);
4935                 return;
4936         }
4937         mutex_unlock(&id_priv->handler_mutex);
4938
4939         /*
4940          * If this races with destroy then the thread that first assigns state
4941          * to a destroying does the cancel.
4942          */
4943         cma_cancel_operation(id_priv, state);
4944         cma_id_put(id_priv);
4945 }
4946
4947 static void cma_process_remove(struct cma_device *cma_dev)
4948 {
4949         mutex_lock(&lock);
4950         while (!list_empty(&cma_dev->id_list)) {
4951                 struct rdma_id_private *id_priv = list_first_entry(
4952                         &cma_dev->id_list, struct rdma_id_private, list);
4953
4954                 list_del(&id_priv->listen_list);
4955                 list_del_init(&id_priv->list);
4956                 cma_id_get(id_priv);
4957                 mutex_unlock(&lock);
4958
4959                 cma_send_device_removal_put(id_priv);
4960
4961                 mutex_lock(&lock);
4962         }
4963         mutex_unlock(&lock);
4964
4965         cma_dev_put(cma_dev);
4966         wait_for_completion(&cma_dev->comp);
4967 }
4968
4969 static int cma_add_one(struct ib_device *device)
4970 {
4971         struct rdma_id_private *to_destroy;
4972         struct cma_device *cma_dev;
4973         struct rdma_id_private *id_priv;
4974         unsigned int i;
4975         unsigned long supported_gids = 0;
4976         int ret;
4977
4978         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4979         if (!cma_dev)
4980                 return -ENOMEM;
4981
4982         cma_dev->device = device;
4983         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4984                                             sizeof(*cma_dev->default_gid_type),
4985                                             GFP_KERNEL);
4986         if (!cma_dev->default_gid_type) {
4987                 ret = -ENOMEM;
4988                 goto free_cma_dev;
4989         }
4990
4991         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4992                                             sizeof(*cma_dev->default_roce_tos),
4993                                             GFP_KERNEL);
4994         if (!cma_dev->default_roce_tos) {
4995                 ret = -ENOMEM;
4996                 goto free_gid_type;
4997         }
4998
4999         rdma_for_each_port (device, i) {
5000                 supported_gids = roce_gid_type_mask_support(device, i);
5001                 WARN_ON(!supported_gids);
5002                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5003                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5004                                 CMA_PREFERRED_ROCE_GID_TYPE;
5005                 else
5006                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5007                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5008                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5009         }
5010
5011         init_completion(&cma_dev->comp);
5012         refcount_set(&cma_dev->refcount, 1);
5013         INIT_LIST_HEAD(&cma_dev->id_list);
5014         ib_set_client_data(device, &cma_client, cma_dev);
5015
5016         mutex_lock(&lock);
5017         list_add_tail(&cma_dev->list, &dev_list);
5018         list_for_each_entry(id_priv, &listen_any_list, list) {
5019                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5020                 if (ret)
5021                         goto free_listen;
5022         }
5023         mutex_unlock(&lock);
5024
5025         trace_cm_add_one(device);
5026         return 0;
5027
5028 free_listen:
5029         list_del(&cma_dev->list);
5030         mutex_unlock(&lock);
5031
5032         /* cma_process_remove() will delete to_destroy */
5033         cma_process_remove(cma_dev);
5034         kfree(cma_dev->default_roce_tos);
5035 free_gid_type:
5036         kfree(cma_dev->default_gid_type);
5037
5038 free_cma_dev:
5039         kfree(cma_dev);
5040         return ret;
5041 }
5042
5043 static void cma_remove_one(struct ib_device *device, void *client_data)
5044 {
5045         struct cma_device *cma_dev = client_data;
5046
5047         trace_cm_remove_one(device);
5048
5049         mutex_lock(&lock);
5050         list_del(&cma_dev->list);
5051         mutex_unlock(&lock);
5052
5053         cma_process_remove(cma_dev);
5054         kfree(cma_dev->default_roce_tos);
5055         kfree(cma_dev->default_gid_type);
5056         kfree(cma_dev);
5057 }
5058
5059 static int cma_init_net(struct net *net)
5060 {
5061         struct cma_pernet *pernet = cma_pernet(net);
5062
5063         xa_init(&pernet->tcp_ps);
5064         xa_init(&pernet->udp_ps);
5065         xa_init(&pernet->ipoib_ps);
5066         xa_init(&pernet->ib_ps);
5067
5068         return 0;
5069 }
5070
5071 static void cma_exit_net(struct net *net)
5072 {
5073         struct cma_pernet *pernet = cma_pernet(net);
5074
5075         WARN_ON(!xa_empty(&pernet->tcp_ps));
5076         WARN_ON(!xa_empty(&pernet->udp_ps));
5077         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5078         WARN_ON(!xa_empty(&pernet->ib_ps));
5079 }
5080
5081 static struct pernet_operations cma_pernet_operations = {
5082         .init = cma_init_net,
5083         .exit = cma_exit_net,
5084         .id = &cma_pernet_id,
5085         .size = sizeof(struct cma_pernet),
5086 };
5087
5088 static int __init cma_init(void)
5089 {
5090         int ret;
5091
5092         /*
5093          * There is a rare lock ordering dependency in cma_netdev_callback()
5094          * that only happens when bonding is enabled. Teach lockdep that rtnl
5095          * must never be nested under lock so it can find these without having
5096          * to test with bonding.
5097          */
5098         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5099                 rtnl_lock();
5100                 mutex_lock(&lock);
5101                 mutex_unlock(&lock);
5102                 rtnl_unlock();
5103         }
5104
5105         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5106         if (!cma_wq)
5107                 return -ENOMEM;
5108
5109         ret = register_pernet_subsys(&cma_pernet_operations);
5110         if (ret)
5111                 goto err_wq;
5112
5113         ib_sa_register_client(&sa_client);
5114         register_netdevice_notifier(&cma_nb);
5115
5116         ret = ib_register_client(&cma_client);
5117         if (ret)
5118                 goto err;
5119
5120         ret = cma_configfs_init();
5121         if (ret)
5122                 goto err_ib;
5123
5124         return 0;
5125
5126 err_ib:
5127         ib_unregister_client(&cma_client);
5128 err:
5129         unregister_netdevice_notifier(&cma_nb);
5130         ib_sa_unregister_client(&sa_client);
5131         unregister_pernet_subsys(&cma_pernet_operations);
5132 err_wq:
5133         destroy_workqueue(cma_wq);
5134         return ret;
5135 }
5136
5137 static void __exit cma_cleanup(void)
5138 {
5139         cma_configfs_exit();
5140         ib_unregister_client(&cma_client);
5141         unregister_netdevice_notifier(&cma_nb);
5142         ib_sa_unregister_client(&sa_client);
5143         unregister_pernet_subsys(&cma_pernet_operations);
5144         destroy_workqueue(cma_wq);
5145 }
5146
5147 module_init(cma_init);
5148 module_exit(cma_cleanup);