drivers/infiniband/core/cma.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   4  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   5  * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
   6  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   7  */
   8
   9 #include <linux/completion.h>
  10 #include <linux/in.h>
  11 #include <linux/in6.h>
  12 #include <linux/mutex.h>
  13 #include <linux/random.h>
  14 #include <linux/igmp.h>
  15 #include <linux/xarray.h>
  16 #include <linux/inetdevice.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <net/route.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/netns/generic.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/ip_fib.h>
  26 #include <net/ip6_route.h>
  27
  28 #include <rdma/rdma_cm.h>
  29 #include <rdma/rdma_cm_ib.h>
  30 #include <rdma/rdma_netlink.h>
  31 #include <rdma/ib.h>
  32 #include <rdma/ib_cache.h>
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_sa.h>
  35 #include <rdma/iw_cm.h>
  36
  37 #include "core_priv.h"
  38 #include "cma_priv.h"
  39 #include "cma_trace.h"
  40
  41 MODULE_AUTHOR("Sean Hefty");
  42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  43 MODULE_LICENSE("Dual BSD/GPL");
  44
  45 #define CMA_CM_RESPONSE_TIMEOUT 20
  46 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  47 #define CMA_MAX_CM_RETRIES 15
  48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  49 #define CMA_IBOE_PACKET_LIFETIME 18
  50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  51
  52 static const char * const cma_events[] = {
  53         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  54         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  55         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  56         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  57         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  58         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  59         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  60         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  61         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  62         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  63         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  64         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  65         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  66         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  67         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  68         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  69 };
  70
  71 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  72 {
  73         size_t index = event;
  74
  75         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  76                         cma_events[index] : "unrecognized event";
  77 }
  78 EXPORT_SYMBOL(rdma_event_msg);
  79
  80 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  81                                                 int reason)
  82 {
  83         if (rdma_ib_or_roce(id->device, id->port_num))
  84                 return ibcm_reject_msg(reason);
  85
  86         if (rdma_protocol_iwarp(id->device, id->port_num))
  87                 return iwcm_reject_msg(reason);
  88
  89         WARN_ON_ONCE(1);
  90         return "unrecognized transport";
  91 }
  92 EXPORT_SYMBOL(rdma_reject_msg);
  93
  94 /**
  95  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  96  *                           request.
  97  * @id: Communication identifier that received the REJECT event.
  98  * @reason: Value returned in the REJECT event status field.
  99  */
 100 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 101 {
 102         if (rdma_ib_or_roce(id->device, id->port_num))
 103                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 104
 105         if (rdma_protocol_iwarp(id->device, id->port_num))
 106                 return reason == -ECONNREFUSED;
 107
 108         WARN_ON_ONCE(1);
 109         return false;
 110 }
 111
 112 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 113                                       struct rdma_cm_event *ev, u8 *data_len)
 114 {
 115         const void *p;
 116
 117         if (rdma_is_consumer_reject(id, ev->status)) {
 118                 *data_len = ev->param.conn.private_data_len;
 119                 p = ev->param.conn.private_data;
 120         } else {
 121                 *data_len = 0;
 122                 p = NULL;
 123         }
 124         return p;
 125 }
 126 EXPORT_SYMBOL(rdma_consumer_reject_data);
 127
 128 /**
 129  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 130  * @id: Communication Identifier
 131  */
 132 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 133 {
 134         struct rdma_id_private *id_priv;
 135
 136         id_priv = container_of(id, struct rdma_id_private, id);
 137         if (id->device->node_type == RDMA_NODE_RNIC)
 138                 return id_priv->cm_id.iw;
 139         return NULL;
 140 }
 141 EXPORT_SYMBOL(rdma_iw_cm_id);
 142
 143 /**
 144  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 145  * @res: rdma resource tracking entry pointer
 146  */
 147 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 148 {
 149         struct rdma_id_private *id_priv =
 150                 container_of(res, struct rdma_id_private, res);
 151
 152         return &id_priv->id;
 153 }
 154 EXPORT_SYMBOL(rdma_res_to_id);
 155
 156 static int cma_add_one(struct ib_device *device);
 157 static void cma_remove_one(struct ib_device *device, void *client_data);
 158
 159 static struct ib_client cma_client = {
 160         .name   = "cma",
 161         .add    = cma_add_one,
 162         .remove = cma_remove_one
 163 };
 164
 165 static struct ib_sa_client sa_client;
 166 static LIST_HEAD(dev_list);
 167 static LIST_HEAD(listen_any_list);
 168 static DEFINE_MUTEX(lock);
 169 static struct workqueue_struct *cma_wq;
 170 static unsigned int cma_pernet_id;
 171
 172 struct cma_pernet {
 173         struct xarray tcp_ps;
 174         struct xarray udp_ps;
 175         struct xarray ipoib_ps;
 176         struct xarray ib_ps;
 177 };
 178
 179 static struct cma_pernet *cma_pernet(struct net *net)
 180 {
 181         return net_generic(net, cma_pernet_id);
 182 }
 183
 184 static
 185 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 186 {
 187         struct cma_pernet *pernet = cma_pernet(net);
 188
 189         switch (ps) {
 190         case RDMA_PS_TCP:
 191                 return &pernet->tcp_ps;
 192         case RDMA_PS_UDP:
 193                 return &pernet->udp_ps;
 194         case RDMA_PS_IPOIB:
 195                 return &pernet->ipoib_ps;
 196         case RDMA_PS_IB:
 197                 return &pernet->ib_ps;
 198         default:
 199                 return NULL;
 200         }
 201 }
 202
 203 struct cma_device {
 204         struct list_head        list;
 205         struct ib_device        *device;
 206         struct completion       comp;
 207         refcount_t refcount;
 208         struct list_head        id_list;
 209         enum ib_gid_type        *default_gid_type;
 210         u8                      *default_roce_tos;
 211 };
 212
 213 struct rdma_bind_list {
 214         enum rdma_ucm_port_space ps;
 215         struct hlist_head       owners;
 216         unsigned short          port;
 217 };
 218
 219 struct class_port_info_context {
 220         struct ib_class_port_info       *class_port_info;
 221         struct ib_device                *device;
 222         struct completion               done;
 223         struct ib_sa_query              *sa_query;
 224         u8                              port_num;
 225 };
 226
 227 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 228                         struct rdma_bind_list *bind_list, int snum)
 229 {
 230         struct xarray *xa = cma_pernet_xa(net, ps);
 231
 232         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 233 }
 234
 235 static struct rdma_bind_list *cma_ps_find(struct net *net,
 236                                           enum rdma_ucm_port_space ps, int snum)
 237 {
 238         struct xarray *xa = cma_pernet_xa(net, ps);
 239
 240         return xa_load(xa, snum);
 241 }
 242
 243 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 244                           int snum)
 245 {
 246         struct xarray *xa = cma_pernet_xa(net, ps);
 247
 248         xa_erase(xa, snum);
 249 }
 250
 251 enum {
 252         CMA_OPTION_AFONLY,
 253 };
 254
 255 void cma_dev_get(struct cma_device *cma_dev)
 256 {
 257         refcount_inc(&cma_dev->refcount);
 258 }
 259
 260 void cma_dev_put(struct cma_device *cma_dev)
 261 {
 262         if (refcount_dec_and_test(&cma_dev->refcount))
 263                 complete(&cma_dev->comp);
 264 }
 265
 266 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 267                                              void               *cookie)
 268 {
 269         struct cma_device *cma_dev;
 270         struct cma_device *found_cma_dev = NULL;
 271
 272         mutex_lock(&lock);
 273
 274         list_for_each_entry(cma_dev, &dev_list, list)
 275                 if (filter(cma_dev->device, cookie)) {
 276                         found_cma_dev = cma_dev;
 277                         break;
 278                 }
 279
 280         if (found_cma_dev)
 281                 cma_dev_get(found_cma_dev);
 282         mutex_unlock(&lock);
 283         return found_cma_dev;
 284 }
 285
 286 int cma_get_default_gid_type(struct cma_device *cma_dev,
 287                              unsigned int port)
 288 {
 289         if (!rdma_is_port_valid(cma_dev->device, port))
 290                 return -EINVAL;
 291
 292         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 293 }
 294
 295 int cma_set_default_gid_type(struct cma_device *cma_dev,
 296                              unsigned int port,
 297                              enum ib_gid_type default_gid_type)
 298 {
 299         unsigned long supported_gids;
 300
 301         if (!rdma_is_port_valid(cma_dev->device, port))
 302                 return -EINVAL;
 303
 304         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 305
 306         if (!(supported_gids & 1 << default_gid_type))
 307                 return -EINVAL;
 308
 309         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 310                 default_gid_type;
 311
 312         return 0;
 313 }
 314
 315 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 316 {
 317         if (!rdma_is_port_valid(cma_dev->device, port))
 318                 return -EINVAL;
 319
 320         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 321 }
 322
 323 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 324                              u8 default_roce_tos)
 325 {
 326         if (!rdma_is_port_valid(cma_dev->device, port))
 327                 return -EINVAL;
 328
 329         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 330                  default_roce_tos;
 331
 332         return 0;
 333 }
 334 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 335 {
 336         return cma_dev->device;
 337 }
 338
 339 /*
 340  * Device removal can occur at anytime, so we need extra handling to
 341  * serialize notifying the user of device removal with other callbacks.
 342  * We do this by disabling removal notification while a callback is in process,
 343  * and reporting it after the callback completes.
 344  */
 345
 346 struct cma_multicast {
 347         struct rdma_id_private *id_priv;
 348         union {
 349                 struct ib_sa_multicast *ib;
 350         } multicast;
 351         struct list_head        list;
 352         void                    *context;
 353         struct sockaddr_storage addr;
 354         struct kref             mcref;
 355         u8                      join_state;
 356 };
 357
 358 struct cma_work {
 359         struct work_struct      work;
 360         struct rdma_id_private  *id;
 361         enum rdma_cm_state      old_state;
 362         enum rdma_cm_state      new_state;
 363         struct rdma_cm_event    event;
 364 };
 365
 366 struct cma_ndev_work {
 367         struct work_struct      work;
 368         struct rdma_id_private  *id;
 369         struct rdma_cm_event    event;
 370 };
 371
 372 struct iboe_mcast_work {
 373         struct work_struct       work;
 374         struct rdma_id_private  *id;
 375         struct cma_multicast    *mc;
 376 };
 377
 378 union cma_ip_addr {
 379         struct in6_addr ip6;
 380         struct {
 381                 __be32 pad[3];
 382                 __be32 addr;
 383         } ip4;
 384 };
 385
 386 struct cma_hdr {
 387         u8 cma_version;
 388         u8 ip_version;  /* IP version: 7:4 */
 389         __be16 port;
 390         union cma_ip_addr src_addr;
 391         union cma_ip_addr dst_addr;
 392 };
 393
 394 #define CMA_VERSION 0x00
 395
 396 struct cma_req_info {
 397         struct sockaddr_storage listen_addr_storage;
 398         struct sockaddr_storage src_addr_storage;
 399         struct ib_device *device;
 400         union ib_gid local_gid;
 401         __be64 service_id;
 402         int port;
 403         bool has_gid;
 404         u16 pkey;
 405 };
 406
 407 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 408 {
 409         unsigned long flags;
 410         int ret;
 411
 412         spin_lock_irqsave(&id_priv->lock, flags);
 413         ret = (id_priv->state == comp);
 414         spin_unlock_irqrestore(&id_priv->lock, flags);
 415         return ret;
 416 }
 417
 418 static int cma_comp_exch(struct rdma_id_private *id_priv,
 419                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 420 {
 421         unsigned long flags;
 422         int ret;
 423
 424         spin_lock_irqsave(&id_priv->lock, flags);
 425         if ((ret = (id_priv->state == comp)))
 426                 id_priv->state = exch;
 427         spin_unlock_irqrestore(&id_priv->lock, flags);
 428         return ret;
 429 }
 430
 431 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 432 {
 433         return hdr->ip_version >> 4;
 434 }
 435
 436 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 437 {
 438         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 439 }
 440
 441 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 442 {
 443         struct in_device *in_dev = NULL;
 444
 445         if (ndev) {
 446                 rtnl_lock();
 447                 in_dev = __in_dev_get_rtnl(ndev);
 448                 if (in_dev) {
 449                         if (join)
 450                                 ip_mc_inc_group(in_dev,
 451                                                 *(__be32 *)(mgid->raw + 12));
 452                         else
 453                                 ip_mc_dec_group(in_dev,
 454                                                 *(__be32 *)(mgid->raw + 12));
 455                 }
 456                 rtnl_unlock();
 457         }
 458         return (in_dev) ? 0 : -ENODEV;
 459 }
 460
 461 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 462                                struct cma_device *cma_dev)
 463 {
 464         cma_dev_get(cma_dev);
 465         id_priv->cma_dev = cma_dev;
 466         id_priv->id.device = cma_dev->device;
 467         id_priv->id.route.addr.dev_addr.transport =
 468                 rdma_node_get_transport(cma_dev->device->node_type);
 469         list_add_tail(&id_priv->list, &cma_dev->id_list);
 470         if (id_priv->res.kern_name)
 471                 rdma_restrack_kadd(&id_priv->res);
 472         else
 473                 rdma_restrack_uadd(&id_priv->res);
 474         trace_cm_id_attach(id_priv, cma_dev->device);
 475 }
 476
 477 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 478                               struct cma_device *cma_dev)
 479 {
 480         _cma_attach_to_dev(id_priv, cma_dev);
 481         id_priv->gid_type =
 482                 cma_dev->default_gid_type[id_priv->id.port_num -
 483                                           rdma_start_port(cma_dev->device)];
 484 }
 485
 486 static inline void release_mc(struct kref *kref)
 487 {
 488         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 489
 490         kfree(mc->multicast.ib);
 491         kfree(mc);
 492 }
 493
 494 static void cma_release_dev(struct rdma_id_private *id_priv)
 495 {
 496         mutex_lock(&lock);
 497         list_del(&id_priv->list);
 498         cma_dev_put(id_priv->cma_dev);
 499         id_priv->cma_dev = NULL;
 500         mutex_unlock(&lock);
 501 }
 502
 503 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 504 {
 505         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 506 }
 507
 508 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 509 {
 510         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 511 }
 512
 513 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 514 {
 515         return id_priv->id.route.addr.src_addr.ss_family;
 516 }
 517
 518 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 519 {
 520         struct ib_sa_mcmember_rec rec;
 521         int ret = 0;
 522
 523         if (id_priv->qkey) {
 524                 if (qkey && id_priv->qkey != qkey)
 525                         return -EINVAL;
 526                 return 0;
 527         }
 528
 529         if (qkey) {
 530                 id_priv->qkey = qkey;
 531                 return 0;
 532         }
 533
 534         switch (id_priv->id.ps) {
 535         case RDMA_PS_UDP:
 536         case RDMA_PS_IB:
 537                 id_priv->qkey = RDMA_UDP_QKEY;
 538                 break;
 539         case RDMA_PS_IPOIB:
 540                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 541                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 542                                              id_priv->id.port_num, &rec.mgid,
 543                                              &rec);
 544                 if (!ret)
 545                         id_priv->qkey = be32_to_cpu(rec.qkey);
 546                 break;
 547         default:
 548                 break;
 549         }
 550         return ret;
 551 }
 552
 553 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 554 {
 555         dev_addr->dev_type = ARPHRD_INFINIBAND;
 556         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 557         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 558 }
 559
 560 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 561 {
 562         int ret;
 563
 564         if (addr->sa_family != AF_IB) {
 565                 ret = rdma_translate_ip(addr, dev_addr);
 566         } else {
 567                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 568                 ret = 0;
 569         }
 570
 571         return ret;
 572 }
 573
 574 static const struct ib_gid_attr *
 575 cma_validate_port(struct ib_device *device, u8 port,
 576                   enum ib_gid_type gid_type,
 577                   union ib_gid *gid,
 578                   struct rdma_id_private *id_priv)
 579 {
 580         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 581         int bound_if_index = dev_addr->bound_dev_if;
 582         const struct ib_gid_attr *sgid_attr;
 583         int dev_type = dev_addr->dev_type;
 584         struct net_device *ndev = NULL;
 585
 586         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 587                 return ERR_PTR(-ENODEV);
 588
 589         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 590                 return ERR_PTR(-ENODEV);
 591
 592         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 593                 return ERR_PTR(-ENODEV);
 594
 595         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 596                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 597                 if (!ndev)
 598                         return ERR_PTR(-ENODEV);
 599         } else {
 600                 gid_type = IB_GID_TYPE_IB;
 601         }
 602
 603         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 604         if (ndev)
 605                 dev_put(ndev);
 606         return sgid_attr;
 607 }
 608
 609 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 610                                const struct ib_gid_attr *sgid_attr)
 611 {
 612         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 613         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 614 }
 615
 616 /**
 617  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 618  * based on source ip address.
 619  * @id_priv:    cm_id which should be bound to cma device
 620  *
 621  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 622  * based on source IP address. It returns 0 on success or error code otherwise.
 623  * It is applicable to active and passive side cm_id.
 624  */
 625 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 626 {
 627         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 628         const struct ib_gid_attr *sgid_attr;
 629         union ib_gid gid, iboe_gid, *gidp;
 630         struct cma_device *cma_dev;
 631         enum ib_gid_type gid_type;
 632         int ret = -ENODEV;
 633         unsigned int port;
 634
 635         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 636             id_priv->id.ps == RDMA_PS_IPOIB)
 637                 return -EINVAL;
 638
 639         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 640                     &iboe_gid);
 641
 642         memcpy(&gid, dev_addr->src_dev_addr +
 643                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 644
 645         mutex_lock(&lock);
 646         list_for_each_entry(cma_dev, &dev_list, list) {
 647                 rdma_for_each_port (cma_dev->device, port) {
 648                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 649                                &iboe_gid : &gid;
 650                         gid_type = cma_dev->default_gid_type[port - 1];
 651                         sgid_attr = cma_validate_port(cma_dev->device, port,
 652                                                       gid_type, gidp, id_priv);
 653                         if (!IS_ERR(sgid_attr)) {
 654                                 id_priv->id.port_num = port;
 655                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 656                                 cma_attach_to_dev(id_priv, cma_dev);
 657                                 ret = 0;
 658                                 goto out;
 659                         }
 660                 }
 661         }
 662 out:
 663         mutex_unlock(&lock);
 664         return ret;
 665 }
 666
 667 /**
 668  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 669  * @id_priv:            cm id to bind to cma device
 670  * @listen_id_priv:     listener cm id to match against
 671  * @req:                Pointer to req structure containaining incoming
 672  *                      request information
 673  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 674  * rdma device matches for listen_id and incoming request. It also verifies
 675  * that a GID table entry is present for the source address.
 676  * Returns 0 on success, or returns error code otherwise.
 677  */
 678 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 679                               const struct rdma_id_private *listen_id_priv,
 680                               struct cma_req_info *req)
 681 {
 682         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 683         const struct ib_gid_attr *sgid_attr;
 684         enum ib_gid_type gid_type;
 685         union ib_gid gid;
 686
 687         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 688             id_priv->id.ps == RDMA_PS_IPOIB)
 689                 return -EINVAL;
 690
 691         if (rdma_protocol_roce(req->device, req->port))
 692                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 693                             &gid);
 694         else
 695                 memcpy(&gid, dev_addr->src_dev_addr +
 696                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 697
 698         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 699         sgid_attr = cma_validate_port(req->device, req->port,
 700                                       gid_type, &gid, id_priv);
 701         if (IS_ERR(sgid_attr))
 702                 return PTR_ERR(sgid_attr);
 703
 704         id_priv->id.port_num = req->port;
 705         cma_bind_sgid_attr(id_priv, sgid_attr);
 706         /* Need to acquire lock to protect against reader
 707          * of cma_dev->id_list such as cma_netdev_callback() and
 708          * cma_process_remove().
 709          */
 710         mutex_lock(&lock);
 711         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 712         mutex_unlock(&lock);
 713         return 0;
 714 }
 715
 716 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 717                               const struct rdma_id_private *listen_id_priv)
 718 {
 719         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 720         const struct ib_gid_attr *sgid_attr;
 721         struct cma_device *cma_dev;
 722         enum ib_gid_type gid_type;
 723         int ret = -ENODEV;
 724         unsigned int port;
 725         union ib_gid gid;
 726
 727         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 728             id_priv->id.ps == RDMA_PS_IPOIB)
 729                 return -EINVAL;
 730
 731         memcpy(&gid, dev_addr->src_dev_addr +
 732                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 733
 734         mutex_lock(&lock);
 735
 736         cma_dev = listen_id_priv->cma_dev;
 737         port = listen_id_priv->id.port_num;
 738         gid_type = listen_id_priv->gid_type;
 739         sgid_attr = cma_validate_port(cma_dev->device, port,
 740                                       gid_type, &gid, id_priv);
 741         if (!IS_ERR(sgid_attr)) {
 742                 id_priv->id.port_num = port;
 743                 cma_bind_sgid_attr(id_priv, sgid_attr);
 744                 ret = 0;
 745                 goto out;
 746         }
 747
 748         list_for_each_entry(cma_dev, &dev_list, list) {
 749                 rdma_for_each_port (cma_dev->device, port) {
 750                         if (listen_id_priv->cma_dev == cma_dev &&
 751                             listen_id_priv->id.port_num == port)
 752                                 continue;
 753
 754                         gid_type = cma_dev->default_gid_type[port - 1];
 755                         sgid_attr = cma_validate_port(cma_dev->device, port,
 756                                                       gid_type, &gid, id_priv);
 757                         if (!IS_ERR(sgid_attr)) {
 758                                 id_priv->id.port_num = port;
 759                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 760                                 ret = 0;
 761                                 goto out;
 762                         }
 763                 }
 764         }
 765
 766 out:
 767         if (!ret)
 768                 cma_attach_to_dev(id_priv, cma_dev);
 769
 770         mutex_unlock(&lock);
 771         return ret;
 772 }
 773
 774 /*
 775  * Select the source IB device and address to reach the destination IB address.
 776  */
 777 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 778 {
 779         struct cma_device *cma_dev, *cur_dev;
 780         struct sockaddr_ib *addr;
 781         union ib_gid gid, sgid, *dgid;
 782         unsigned int p;
 783         u16 pkey, index;
 784         enum ib_port_state port_state;
 785         int i;
 786
 787         cma_dev = NULL;
 788         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 789         dgid = (union ib_gid *) &addr->sib_addr;
 790         pkey = ntohs(addr->sib_pkey);
 791
 792         mutex_lock(&lock);
 793         list_for_each_entry(cur_dev, &dev_list, list) {
 794                 rdma_for_each_port (cur_dev->device, p) {
 795                         if (!rdma_cap_af_ib(cur_dev->device, p))
 796                                 continue;
 797
 798                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 799                                 continue;
 800
 801                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 802                                 continue;
 803                         for (i = 0; !rdma_query_gid(cur_dev->device,
 804                                                     p, i, &gid);
 805                              i++) {
 806                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 807                                         cma_dev = cur_dev;
 808                                         sgid = gid;
 809                                         id_priv->id.port_num = p;
 810                                         goto found;
 811                                 }
 812
 813                                 if (!cma_dev && (gid.global.subnet_prefix ==
 814                                     dgid->global.subnet_prefix) &&
 815                                     port_state == IB_PORT_ACTIVE) {
 816                                         cma_dev = cur_dev;
 817                                         sgid = gid;
 818                                         id_priv->id.port_num = p;
 819                                         goto found;
 820                                 }
 821                         }
 822                 }
 823         }
 824         mutex_unlock(&lock);
 825         return -ENODEV;
 826
 827 found:
 828         cma_attach_to_dev(id_priv, cma_dev);
 829         mutex_unlock(&lock);
 830         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 831         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 832         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 833         return 0;
 834 }
 835
 836 static void cma_id_get(struct rdma_id_private *id_priv)
 837 {
 838         refcount_inc(&id_priv->refcount);
 839 }
 840
 841 static void cma_id_put(struct rdma_id_private *id_priv)
 842 {
 843         if (refcount_dec_and_test(&id_priv->refcount))
 844                 complete(&id_priv->comp);
 845 }
 846
 847 struct rdma_cm_id *__rdma_create_id(struct net *net,
 848                                     rdma_cm_event_handler event_handler,
 849                                     void *context, enum rdma_ucm_port_space ps,
 850                                     enum ib_qp_type qp_type, const char *caller)
 851 {
 852         struct rdma_id_private *id_priv;
 853
 854         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 855         if (!id_priv)
 856                 return ERR_PTR(-ENOMEM);
 857
 858         rdma_restrack_set_task(&id_priv->res, caller);
 859         id_priv->res.type = RDMA_RESTRACK_CM_ID;
 860         id_priv->state = RDMA_CM_IDLE;
 861         id_priv->id.context = context;
 862         id_priv->id.event_handler = event_handler;
 863         id_priv->id.ps = ps;
 864         id_priv->id.qp_type = qp_type;
 865         id_priv->tos_set = false;
 866         id_priv->timeout_set = false;
 867         id_priv->gid_type = IB_GID_TYPE_IB;
 868         spin_lock_init(&id_priv->lock);
 869         mutex_init(&id_priv->qp_mutex);
 870         init_completion(&id_priv->comp);
 871         refcount_set(&id_priv->refcount, 1);
 872         mutex_init(&id_priv->handler_mutex);
 873         INIT_LIST_HEAD(&id_priv->listen_list);
 874         INIT_LIST_HEAD(&id_priv->mc_list);
 875         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 876         id_priv->id.route.addr.dev_addr.net = get_net(net);
 877         id_priv->seq_num &= 0x00ffffff;
 878
 879         return &id_priv->id;
 880 }
 881 EXPORT_SYMBOL(__rdma_create_id);
 882
 883 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 884 {
 885         struct ib_qp_attr qp_attr;
 886         int qp_attr_mask, ret;
 887
 888         qp_attr.qp_state = IB_QPS_INIT;
 889         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 890         if (ret)
 891                 return ret;
 892
 893         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 894         if (ret)
 895                 return ret;
 896
 897         qp_attr.qp_state = IB_QPS_RTR;
 898         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 899         if (ret)
 900                 return ret;
 901
 902         qp_attr.qp_state = IB_QPS_RTS;
 903         qp_attr.sq_psn = 0;
 904         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 905
 906         return ret;
 907 }
 908
 909 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 910 {
 911         struct ib_qp_attr qp_attr;
 912         int qp_attr_mask, ret;
 913
 914         qp_attr.qp_state = IB_QPS_INIT;
 915         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 916         if (ret)
 917                 return ret;
 918
 919         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 920 }
 921
 922 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 923                    struct ib_qp_init_attr *qp_init_attr)
 924 {
 925         struct rdma_id_private *id_priv;
 926         struct ib_qp *qp;
 927         int ret;
 928
 929         id_priv = container_of(id, struct rdma_id_private, id);
 930         if (id->device != pd->device) {
 931                 ret = -EINVAL;
 932                 goto out_err;
 933         }
 934
 935         qp_init_attr->port_num = id->port_num;
 936         qp = ib_create_qp(pd, qp_init_attr);
 937         if (IS_ERR(qp)) {
 938                 ret = PTR_ERR(qp);
 939                 goto out_err;
 940         }
 941
 942         if (id->qp_type == IB_QPT_UD)
 943                 ret = cma_init_ud_qp(id_priv, qp);
 944         else
 945                 ret = cma_init_conn_qp(id_priv, qp);
 946         if (ret)
 947                 goto out_destroy;
 948
 949         id->qp = qp;
 950         id_priv->qp_num = qp->qp_num;
 951         id_priv->srq = (qp->srq != NULL);
 952         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 953         return 0;
 954 out_destroy:
 955         ib_destroy_qp(qp);
 956 out_err:
 957         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 958         return ret;
 959 }
 960 EXPORT_SYMBOL(rdma_create_qp);
 961
 962 void rdma_destroy_qp(struct rdma_cm_id *id)
 963 {
 964         struct rdma_id_private *id_priv;
 965
 966         id_priv = container_of(id, struct rdma_id_private, id);
 967         trace_cm_qp_destroy(id_priv);
 968         mutex_lock(&id_priv->qp_mutex);
 969         ib_destroy_qp(id_priv->id.qp);
 970         id_priv->id.qp = NULL;
 971         mutex_unlock(&id_priv->qp_mutex);
 972 }
 973 EXPORT_SYMBOL(rdma_destroy_qp);
 974
 975 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 976                              struct rdma_conn_param *conn_param)
 977 {
 978         struct ib_qp_attr qp_attr;
 979         int qp_attr_mask, ret;
 980
 981         mutex_lock(&id_priv->qp_mutex);
 982         if (!id_priv->id.qp) {
 983                 ret = 0;
 984                 goto out;
 985         }
 986
 987         /* Need to update QP attributes from default values. */
 988         qp_attr.qp_state = IB_QPS_INIT;
 989         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 990         if (ret)
 991                 goto out;
 992
 993         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 994         if (ret)
 995                 goto out;
 996
 997         qp_attr.qp_state = IB_QPS_RTR;
 998         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 999         if (ret)
1000                 goto out;
1001
1002         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1003
1004         if (conn_param)
1005                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1006         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1007 out:
1008         mutex_unlock(&id_priv->qp_mutex);
1009         return ret;
1010 }
1011
1012 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1013                              struct rdma_conn_param *conn_param)
1014 {
1015         struct ib_qp_attr qp_attr;
1016         int qp_attr_mask, ret;
1017
1018         mutex_lock(&id_priv->qp_mutex);
1019         if (!id_priv->id.qp) {
1020                 ret = 0;
1021                 goto out;
1022         }
1023
1024         qp_attr.qp_state = IB_QPS_RTS;
1025         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1026         if (ret)
1027                 goto out;
1028
1029         if (conn_param)
1030                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1031         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1032 out:
1033         mutex_unlock(&id_priv->qp_mutex);
1034         return ret;
1035 }
1036
1037 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1038 {
1039         struct ib_qp_attr qp_attr;
1040         int ret;
1041
1042         mutex_lock(&id_priv->qp_mutex);
1043         if (!id_priv->id.qp) {
1044                 ret = 0;
1045                 goto out;
1046         }
1047
1048         qp_attr.qp_state = IB_QPS_ERR;
1049         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1050 out:
1051         mutex_unlock(&id_priv->qp_mutex);
1052         return ret;
1053 }
1054
1055 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1056                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1057 {
1058         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1059         int ret;
1060         u16 pkey;
1061
1062         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1063                 pkey = 0xffff;
1064         else
1065                 pkey = ib_addr_get_pkey(dev_addr);
1066
1067         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1068                                   pkey, &qp_attr->pkey_index);
1069         if (ret)
1070                 return ret;
1071
1072         qp_attr->port_num = id_priv->id.port_num;
1073         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1074
1075         if (id_priv->id.qp_type == IB_QPT_UD) {
1076                 ret = cma_set_qkey(id_priv, 0);
1077                 if (ret)
1078                         return ret;
1079
1080                 qp_attr->qkey = id_priv->qkey;
1081                 *qp_attr_mask |= IB_QP_QKEY;
1082         } else {
1083                 qp_attr->qp_access_flags = 0;
1084                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1085         }
1086         return 0;
1087 }
1088
1089 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1090                        int *qp_attr_mask)
1091 {
1092         struct rdma_id_private *id_priv;
1093         int ret = 0;
1094
1095         id_priv = container_of(id, struct rdma_id_private, id);
1096         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1097                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1098                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1099                 else
1100                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1101                                                  qp_attr_mask);
1102
1103                 if (qp_attr->qp_state == IB_QPS_RTR)
1104                         qp_attr->rq_psn = id_priv->seq_num;
1105         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1106                 if (!id_priv->cm_id.iw) {
1107                         qp_attr->qp_access_flags = 0;
1108                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1109                 } else
1110                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1111                                                  qp_attr_mask);
1112                 qp_attr->port_num = id_priv->id.port_num;
1113                 *qp_attr_mask |= IB_QP_PORT;
1114         } else
1115                 ret = -ENOSYS;
1116
1117         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1118                 qp_attr->timeout = id_priv->timeout;
1119
1120         return ret;
1121 }
1122 EXPORT_SYMBOL(rdma_init_qp_attr);
1123
1124 static inline bool cma_zero_addr(const struct sockaddr *addr)
1125 {
1126         switch (addr->sa_family) {
1127         case AF_INET:
1128                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1129         case AF_INET6:
1130                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1131         case AF_IB:
1132                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1133         default:
1134                 return false;
1135         }
1136 }
1137
1138 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1139 {
1140         switch (addr->sa_family) {
1141         case AF_INET:
1142                 return ipv4_is_loopback(
1143                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1144         case AF_INET6:
1145                 return ipv6_addr_loopback(
1146                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1147         case AF_IB:
1148                 return ib_addr_loopback(
1149                         &((struct sockaddr_ib *)addr)->sib_addr);
1150         default:
1151                 return false;
1152         }
1153 }
1154
1155 static inline bool cma_any_addr(const struct sockaddr *addr)
1156 {
1157         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1158 }
1159
1160 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1161 {
1162         if (src->sa_family != dst->sa_family)
1163                 return -1;
1164
1165         switch (src->sa_family) {
1166         case AF_INET:
1167                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1168                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1169         case AF_INET6: {
1170                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1171                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1172                 bool link_local;
1173
1174                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1175                                           &dst_addr6->sin6_addr))
1176                         return 1;
1177                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1178                              IPV6_ADDR_LINKLOCAL;
1179                 /* Link local must match their scope_ids */
1180                 return link_local ? (src_addr6->sin6_scope_id !=
1181                                      dst_addr6->sin6_scope_id) :
1182                                     0;
1183         }
1184
1185         default:
1186                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1187                                    &((struct sockaddr_ib *) dst)->sib_addr);
1188         }
1189 }
1190
1191 static __be16 cma_port(const struct sockaddr *addr)
1192 {
1193         struct sockaddr_ib *sib;
1194
1195         switch (addr->sa_family) {
1196         case AF_INET:
1197                 return ((struct sockaddr_in *) addr)->sin_port;
1198         case AF_INET6:
1199                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1200         case AF_IB:
1201                 sib = (struct sockaddr_ib *) addr;
1202                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1203                                     be64_to_cpu(sib->sib_sid_mask)));
1204         default:
1205                 return 0;
1206         }
1207 }
1208
1209 static inline int cma_any_port(const struct sockaddr *addr)
1210 {
1211         return !cma_port(addr);
1212 }
1213
1214 static void cma_save_ib_info(struct sockaddr *src_addr,
1215                              struct sockaddr *dst_addr,
1216                              const struct rdma_cm_id *listen_id,
1217                              const struct sa_path_rec *path)
1218 {
1219         struct sockaddr_ib *listen_ib, *ib;
1220
1221         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1222         if (src_addr) {
1223                 ib = (struct sockaddr_ib *)src_addr;
1224                 ib->sib_family = AF_IB;
1225                 if (path) {
1226                         ib->sib_pkey = path->pkey;
1227                         ib->sib_flowinfo = path->flow_label;
1228                         memcpy(&ib->sib_addr, &path->sgid, 16);
1229                         ib->sib_sid = path->service_id;
1230                         ib->sib_scope_id = 0;
1231                 } else {
1232                         ib->sib_pkey = listen_ib->sib_pkey;
1233                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1234                         ib->sib_addr = listen_ib->sib_addr;
1235                         ib->sib_sid = listen_ib->sib_sid;
1236                         ib->sib_scope_id = listen_ib->sib_scope_id;
1237                 }
1238                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1239         }
1240         if (dst_addr) {
1241                 ib = (struct sockaddr_ib *)dst_addr;
1242                 ib->sib_family = AF_IB;
1243                 if (path) {
1244                         ib->sib_pkey = path->pkey;
1245                         ib->sib_flowinfo = path->flow_label;
1246                         memcpy(&ib->sib_addr, &path->dgid, 16);
1247                 }
1248         }
1249 }
1250
1251 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1252                               struct sockaddr_in *dst_addr,
1253                               struct cma_hdr *hdr,
1254                               __be16 local_port)
1255 {
1256         if (src_addr) {
1257                 *src_addr = (struct sockaddr_in) {
1258                         .sin_family = AF_INET,
1259                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1260                         .sin_port = local_port,
1261                 };
1262         }
1263
1264         if (dst_addr) {
1265                 *dst_addr = (struct sockaddr_in) {
1266                         .sin_family = AF_INET,
1267                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1268                         .sin_port = hdr->port,
1269                 };
1270         }
1271 }
1272
1273 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1274                               struct sockaddr_in6 *dst_addr,
1275                               struct cma_hdr *hdr,
1276                               __be16 local_port)
1277 {
1278         if (src_addr) {
1279                 *src_addr = (struct sockaddr_in6) {
1280                         .sin6_family = AF_INET6,
1281                         .sin6_addr = hdr->dst_addr.ip6,
1282                         .sin6_port = local_port,
1283                 };
1284         }
1285
1286         if (dst_addr) {
1287                 *dst_addr = (struct sockaddr_in6) {
1288                         .sin6_family = AF_INET6,
1289                         .sin6_addr = hdr->src_addr.ip6,
1290                         .sin6_port = hdr->port,
1291                 };
1292         }
1293 }
1294
1295 static u16 cma_port_from_service_id(__be64 service_id)
1296 {
1297         return (u16)be64_to_cpu(service_id);
1298 }
1299
1300 static int cma_save_ip_info(struct sockaddr *src_addr,
1301                             struct sockaddr *dst_addr,
1302                             const struct ib_cm_event *ib_event,
1303                             __be64 service_id)
1304 {
1305         struct cma_hdr *hdr;
1306         __be16 port;
1307
1308         hdr = ib_event->private_data;
1309         if (hdr->cma_version != CMA_VERSION)
1310                 return -EINVAL;
1311
1312         port = htons(cma_port_from_service_id(service_id));
1313
1314         switch (cma_get_ip_ver(hdr)) {
1315         case 4:
1316                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1317                                   (struct sockaddr_in *)dst_addr, hdr, port);
1318                 break;
1319         case 6:
1320                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1321                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1322                 break;
1323         default:
1324                 return -EAFNOSUPPORT;
1325         }
1326
1327         return 0;
1328 }
1329
1330 static int cma_save_net_info(struct sockaddr *src_addr,
1331                              struct sockaddr *dst_addr,
1332                              const struct rdma_cm_id *listen_id,
1333                              const struct ib_cm_event *ib_event,
1334                              sa_family_t sa_family, __be64 service_id)
1335 {
1336         if (sa_family == AF_IB) {
1337                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1338                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1339                                          ib_event->param.req_rcvd.primary_path);
1340                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1341                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1342                 return 0;
1343         }
1344
1345         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1346 }
1347
1348 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1349                              struct cma_req_info *req)
1350 {
1351         const struct ib_cm_req_event_param *req_param =
1352                 &ib_event->param.req_rcvd;
1353         const struct ib_cm_sidr_req_event_param *sidr_param =
1354                 &ib_event->param.sidr_req_rcvd;
1355
1356         switch (ib_event->event) {
1357         case IB_CM_REQ_RECEIVED:
1358                 req->device     = req_param->listen_id->device;
1359                 req->port       = req_param->port;
1360                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1361                        sizeof(req->local_gid));
1362                 req->has_gid    = true;
1363                 req->service_id = req_param->primary_path->service_id;
1364                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1365                 if (req->pkey != req_param->bth_pkey)
1366                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1367                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1368                                             req_param->bth_pkey, req->pkey);
1369                 break;
1370         case IB_CM_SIDR_REQ_RECEIVED:
1371                 req->device     = sidr_param->listen_id->device;
1372                 req->port       = sidr_param->port;
1373                 req->has_gid    = false;
1374                 req->service_id = sidr_param->service_id;
1375                 req->pkey       = sidr_param->pkey;
1376                 if (req->pkey != sidr_param->bth_pkey)
1377                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1378                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1379                                             sidr_param->bth_pkey, req->pkey);
1380                 break;
1381         default:
1382                 return -EINVAL;
1383         }
1384
1385         return 0;
1386 }
1387
1388 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1389                                   const struct sockaddr_in *dst_addr,
1390                                   const struct sockaddr_in *src_addr)
1391 {
1392         __be32 daddr = dst_addr->sin_addr.s_addr,
1393                saddr = src_addr->sin_addr.s_addr;
1394         struct fib_result res;
1395         struct flowi4 fl4;
1396         int err;
1397         bool ret;
1398
1399         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1400             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1401             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1402             ipv4_is_loopback(saddr))
1403                 return false;
1404
1405         memset(&fl4, 0, sizeof(fl4));
1406         fl4.flowi4_iif = net_dev->ifindex;
1407         fl4.daddr = daddr;
1408         fl4.saddr = saddr;
1409
1410         rcu_read_lock();
1411         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1412         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1413         rcu_read_unlock();
1414
1415         return ret;
1416 }
1417
1418 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1419                                   const struct sockaddr_in6 *dst_addr,
1420                                   const struct sockaddr_in6 *src_addr)
1421 {
1422 #if IS_ENABLED(CONFIG_IPV6)
1423         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1424                            IPV6_ADDR_LINKLOCAL;
1425         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1426                                          &src_addr->sin6_addr, net_dev->ifindex,
1427                                          NULL, strict);
1428         bool ret;
1429
1430         if (!rt)
1431                 return false;
1432
1433         ret = rt->rt6i_idev->dev == net_dev;
1434         ip6_rt_put(rt);
1435
1436         return ret;
1437 #else
1438         return false;
1439 #endif
1440 }
1441
1442 static bool validate_net_dev(struct net_device *net_dev,
1443                              const struct sockaddr *daddr,
1444                              const struct sockaddr *saddr)
1445 {
1446         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1447         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1448         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1449         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1450
1451         switch (daddr->sa_family) {
1452         case AF_INET:
1453                 return saddr->sa_family == AF_INET &&
1454                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1455
1456         case AF_INET6:
1457                 return saddr->sa_family == AF_INET6 &&
1458                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1459
1460         default:
1461                 return false;
1462         }
1463 }
1464
1465 static struct net_device *
1466 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1467 {
1468         const struct ib_gid_attr *sgid_attr = NULL;
1469         struct net_device *ndev;
1470
1471         if (ib_event->event == IB_CM_REQ_RECEIVED)
1472                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1473         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1474                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1475
1476         if (!sgid_attr)
1477                 return NULL;
1478
1479         rcu_read_lock();
1480         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1481         if (IS_ERR(ndev))
1482                 ndev = NULL;
1483         else
1484                 dev_hold(ndev);
1485         rcu_read_unlock();
1486         return ndev;
1487 }
1488
1489 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1490                                           struct cma_req_info *req)
1491 {
1492         struct sockaddr *listen_addr =
1493                         (struct sockaddr *)&req->listen_addr_storage;
1494         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1495         struct net_device *net_dev;
1496         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1497         int err;
1498
1499         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1500                                req->service_id);
1501         if (err)
1502                 return ERR_PTR(err);
1503
1504         if (rdma_protocol_roce(req->device, req->port))
1505                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1506         else
1507                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1508                                                    req->pkey,
1509                                                    gid, listen_addr);
1510         if (!net_dev)
1511                 return ERR_PTR(-ENODEV);
1512
1513         return net_dev;
1514 }
1515
1516 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1517 {
1518         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1519 }
1520
1521 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1522                                    const struct cma_hdr *hdr)
1523 {
1524         struct sockaddr *addr = cma_src_addr(id_priv);
1525         __be32 ip4_addr;
1526         struct in6_addr ip6_addr;
1527
1528         if (cma_any_addr(addr) && !id_priv->afonly)
1529                 return true;
1530
1531         switch (addr->sa_family) {
1532         case AF_INET:
1533                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1534                 if (cma_get_ip_ver(hdr) != 4)
1535                         return false;
1536                 if (!cma_any_addr(addr) &&
1537                     hdr->dst_addr.ip4.addr != ip4_addr)
1538                         return false;
1539                 break;
1540         case AF_INET6:
1541                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1542                 if (cma_get_ip_ver(hdr) != 6)
1543                         return false;
1544                 if (!cma_any_addr(addr) &&
1545                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1546                         return false;
1547                 break;
1548         case AF_IB:
1549                 return true;
1550         default:
1551                 return false;
1552         }
1553
1554         return true;
1555 }
1556
1557 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1558 {
1559         struct ib_device *device = id->device;
1560         const int port_num = id->port_num ?: rdma_start_port(device);
1561
1562         return rdma_protocol_roce(device, port_num);
1563 }
1564
1565 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1566 {
1567         const struct sockaddr *daddr =
1568                         (const struct sockaddr *)&req->listen_addr_storage;
1569         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1570
1571         /* Returns true if the req is for IPv6 link local */
1572         return (daddr->sa_family == AF_INET6 &&
1573                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1574 }
1575
1576 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1577                               const struct net_device *net_dev,
1578                               const struct cma_req_info *req)
1579 {
1580         const struct rdma_addr *addr = &id->route.addr;
1581
1582         if (!net_dev)
1583                 /* This request is an AF_IB request */
1584                 return (!id->port_num || id->port_num == req->port) &&
1585                        (addr->src_addr.ss_family == AF_IB);
1586
1587         /*
1588          * If the request is not for IPv6 link local, allow matching
1589          * request to any netdevice of the one or multiport rdma device.
1590          */
1591         if (!cma_is_req_ipv6_ll(req))
1592                 return true;
1593         /*
1594          * Net namespaces must match, and if the listner is listening
1595          * on a specific netdevice than netdevice must match as well.
1596          */
1597         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1598             (!!addr->dev_addr.bound_dev_if ==
1599              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1600                 return true;
1601         else
1602                 return false;
1603 }
1604
1605 static struct rdma_id_private *cma_find_listener(
1606                 const struct rdma_bind_list *bind_list,
1607                 const struct ib_cm_id *cm_id,
1608                 const struct ib_cm_event *ib_event,
1609                 const struct cma_req_info *req,
1610                 const struct net_device *net_dev)
1611 {
1612         struct rdma_id_private *id_priv, *id_priv_dev;
1613
1614         lockdep_assert_held(&lock);
1615
1616         if (!bind_list)
1617                 return ERR_PTR(-EINVAL);
1618
1619         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1620                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1621                         if (id_priv->id.device == cm_id->device &&
1622                             cma_match_net_dev(&id_priv->id, net_dev, req))
1623                                 return id_priv;
1624                         list_for_each_entry(id_priv_dev,
1625                                             &id_priv->listen_list,
1626                                             listen_list) {
1627                                 if (id_priv_dev->id.device == cm_id->device &&
1628                                     cma_match_net_dev(&id_priv_dev->id,
1629                                                       net_dev, req))
1630                                         return id_priv_dev;
1631                         }
1632                 }
1633         }
1634
1635         return ERR_PTR(-EINVAL);
1636 }
1637
1638 static struct rdma_id_private *
1639 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1640                      const struct ib_cm_event *ib_event,
1641                      struct cma_req_info *req,
1642                      struct net_device **net_dev)
1643 {
1644         struct rdma_bind_list *bind_list;
1645         struct rdma_id_private *id_priv;
1646         int err;
1647
1648         err = cma_save_req_info(ib_event, req);
1649         if (err)
1650                 return ERR_PTR(err);
1651
1652         *net_dev = cma_get_net_dev(ib_event, req);
1653         if (IS_ERR(*net_dev)) {
1654                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1655                         /* Assuming the protocol is AF_IB */
1656                         *net_dev = NULL;
1657                 } else {
1658                         return ERR_CAST(*net_dev);
1659                 }
1660         }
1661
1662         mutex_lock(&lock);
1663         /*
1664          * Net namespace might be getting deleted while route lookup,
1665          * cm_id lookup is in progress. Therefore, perform netdevice
1666          * validation, cm_id lookup under rcu lock.
1667          * RCU lock along with netdevice state check, synchronizes with
1668          * netdevice migrating to different net namespace and also avoids
1669          * case where net namespace doesn't get deleted while lookup is in
1670          * progress.
1671          * If the device state is not IFF_UP, its properties such as ifindex
1672          * and nd_net cannot be trusted to remain valid without rcu lock.
1673          * net/core/dev.c change_net_namespace() ensures to synchronize with
1674          * ongoing operations on net device after device is closed using
1675          * synchronize_net().
1676          */
1677         rcu_read_lock();
1678         if (*net_dev) {
1679                 /*
1680                  * If netdevice is down, it is likely that it is administratively
1681                  * down or it might be migrating to different namespace.
1682                  * In that case avoid further processing, as the net namespace
1683                  * or ifindex may change.
1684                  */
1685                 if (((*net_dev)->flags & IFF_UP) == 0) {
1686                         id_priv = ERR_PTR(-EHOSTUNREACH);
1687                         goto err;
1688                 }
1689
1690                 if (!validate_net_dev(*net_dev,
1691                                  (struct sockaddr *)&req->listen_addr_storage,
1692                                  (struct sockaddr *)&req->src_addr_storage)) {
1693                         id_priv = ERR_PTR(-EHOSTUNREACH);
1694                         goto err;
1695                 }
1696         }
1697
1698         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1699                                 rdma_ps_from_service_id(req->service_id),
1700                                 cma_port_from_service_id(req->service_id));
1701         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1702 err:
1703         rcu_read_unlock();
1704         mutex_unlock(&lock);
1705         if (IS_ERR(id_priv) && *net_dev) {
1706                 dev_put(*net_dev);
1707                 *net_dev = NULL;
1708         }
1709         return id_priv;
1710 }
1711
1712 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1713 {
1714         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1715 }
1716
1717 static void cma_cancel_route(struct rdma_id_private *id_priv)
1718 {
1719         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1720                 if (id_priv->query)
1721                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1722         }
1723 }
1724
1725 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1726 {
1727         struct rdma_id_private *dev_id_priv;
1728
1729         /*
1730          * Remove from listen_any_list to prevent added devices from spawning
1731          * additional listen requests.
1732          */
1733         mutex_lock(&lock);
1734         list_del(&id_priv->list);
1735
1736         while (!list_empty(&id_priv->listen_list)) {
1737                 dev_id_priv = list_entry(id_priv->listen_list.next,
1738                                          struct rdma_id_private, listen_list);
1739                 /* sync with device removal to avoid duplicate destruction */
1740                 list_del_init(&dev_id_priv->list);
1741                 list_del(&dev_id_priv->listen_list);
1742                 mutex_unlock(&lock);
1743
1744                 rdma_destroy_id(&dev_id_priv->id);
1745                 mutex_lock(&lock);
1746         }
1747         mutex_unlock(&lock);
1748 }
1749
1750 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1751                                  enum rdma_cm_state state)
1752 {
1753         switch (state) {
1754         case RDMA_CM_ADDR_QUERY:
1755                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1756                 break;
1757         case RDMA_CM_ROUTE_QUERY:
1758                 cma_cancel_route(id_priv);
1759                 break;
1760         case RDMA_CM_LISTEN:
1761                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1762                         cma_cancel_listens(id_priv);
1763                 break;
1764         default:
1765                 break;
1766         }
1767 }
1768
1769 static void cma_release_port(struct rdma_id_private *id_priv)
1770 {
1771         struct rdma_bind_list *bind_list = id_priv->bind_list;
1772         struct net *net = id_priv->id.route.addr.dev_addr.net;
1773
1774         if (!bind_list)
1775                 return;
1776
1777         mutex_lock(&lock);
1778         hlist_del(&id_priv->node);
1779         if (hlist_empty(&bind_list->owners)) {
1780                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1781                 kfree(bind_list);
1782         }
1783         mutex_unlock(&lock);
1784 }
1785
1786 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv,
1787                                     struct cma_multicast *mc)
1788 {
1789         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1790         struct net_device *ndev = NULL;
1791
1792         if (dev_addr->bound_dev_if)
1793                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
1794         if (ndev) {
1795                 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1796                 dev_put(ndev);
1797         }
1798         kref_put(&mc->mcref, release_mc);
1799 }
1800
1801 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1802 {
1803         struct cma_multicast *mc;
1804
1805         while (!list_empty(&id_priv->mc_list)) {
1806                 mc = container_of(id_priv->mc_list.next,
1807                                   struct cma_multicast, list);
1808                 list_del(&mc->list);
1809                 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1810                                       id_priv->id.port_num)) {
1811                         ib_sa_free_multicast(mc->multicast.ib);
1812                         kfree(mc);
1813                 } else {
1814                         cma_leave_roce_mc_group(id_priv, mc);
1815                 }
1816         }
1817 }
1818
1819 static void _destroy_id(struct rdma_id_private *id_priv,
1820                         enum rdma_cm_state state)
1821 {
1822         cma_cancel_operation(id_priv, state);
1823
1824         rdma_restrack_del(&id_priv->res);
1825         if (id_priv->cma_dev) {
1826                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1827                         if (id_priv->cm_id.ib)
1828                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1829                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1830                         if (id_priv->cm_id.iw)
1831                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1832                 }
1833                 cma_leave_mc_groups(id_priv);
1834                 cma_release_dev(id_priv);
1835         }
1836
1837         cma_release_port(id_priv);
1838         cma_id_put(id_priv);
1839         wait_for_completion(&id_priv->comp);
1840
1841         if (id_priv->internal_id)
1842                 cma_id_put(id_priv->id.context);
1843
1844         kfree(id_priv->id.route.path_rec);
1845
1846         if (id_priv->id.route.addr.dev_addr.sgid_attr)
1847                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1848
1849         put_net(id_priv->id.route.addr.dev_addr.net);
1850         kfree(id_priv);
1851 }
1852
1853 /*
1854  * destroy an ID from within the handler_mutex. This ensures that no other
1855  * handlers can start running concurrently.
1856  */
1857 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1858         __releases(&idprv->handler_mutex)
1859 {
1860         enum rdma_cm_state state;
1861         unsigned long flags;
1862
1863         trace_cm_id_destroy(id_priv);
1864
1865         /*
1866          * Setting the state to destroyed under the handler mutex provides a
1867          * fence against calling handler callbacks. If this is invoked due to
1868          * the failure of a handler callback then it guarentees that no future
1869          * handlers will be called.
1870          */
1871         lockdep_assert_held(&id_priv->handler_mutex);
1872         spin_lock_irqsave(&id_priv->lock, flags);
1873         state = id_priv->state;
1874         id_priv->state = RDMA_CM_DESTROYING;
1875         spin_unlock_irqrestore(&id_priv->lock, flags);
1876         mutex_unlock(&id_priv->handler_mutex);
1877         _destroy_id(id_priv, state);
1878 }
1879
1880 void rdma_destroy_id(struct rdma_cm_id *id)
1881 {
1882         struct rdma_id_private *id_priv =
1883                 container_of(id, struct rdma_id_private, id);
1884
1885         mutex_lock(&id_priv->handler_mutex);
1886         destroy_id_handler_unlock(id_priv);
1887 }
1888 EXPORT_SYMBOL(rdma_destroy_id);
1889
1890 static int cma_rep_recv(struct rdma_id_private *id_priv)
1891 {
1892         int ret;
1893
1894         ret = cma_modify_qp_rtr(id_priv, NULL);
1895         if (ret)
1896                 goto reject;
1897
1898         ret = cma_modify_qp_rts(id_priv, NULL);
1899         if (ret)
1900                 goto reject;
1901
1902         trace_cm_send_rtu(id_priv);
1903         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1904         if (ret)
1905                 goto reject;
1906
1907         return 0;
1908 reject:
1909         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1910         cma_modify_qp_err(id_priv);
1911         trace_cm_send_rej(id_priv);
1912         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1913                        NULL, 0, NULL, 0);
1914         return ret;
1915 }
1916
1917 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1918                                    const struct ib_cm_rep_event_param *rep_data,
1919                                    void *private_data)
1920 {
1921         event->param.conn.private_data = private_data;
1922         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1923         event->param.conn.responder_resources = rep_data->responder_resources;
1924         event->param.conn.initiator_depth = rep_data->initiator_depth;
1925         event->param.conn.flow_control = rep_data->flow_control;
1926         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1927         event->param.conn.srq = rep_data->srq;
1928         event->param.conn.qp_num = rep_data->remote_qpn;
1929
1930         event->ece.vendor_id = rep_data->ece.vendor_id;
1931         event->ece.attr_mod = rep_data->ece.attr_mod;
1932 }
1933
1934 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1935                                 struct rdma_cm_event *event)
1936 {
1937         int ret;
1938
1939         lockdep_assert_held(&id_priv->handler_mutex);
1940
1941         trace_cm_event_handler(id_priv, event);
1942         ret = id_priv->id.event_handler(&id_priv->id, event);
1943         trace_cm_event_done(id_priv, event, ret);
1944         return ret;
1945 }
1946
1947 static int cma_ib_handler(struct ib_cm_id *cm_id,
1948                           const struct ib_cm_event *ib_event)
1949 {
1950         struct rdma_id_private *id_priv = cm_id->context;
1951         struct rdma_cm_event event = {};
1952         int ret;
1953
1954         mutex_lock(&id_priv->handler_mutex);
1955         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1956              id_priv->state != RDMA_CM_CONNECT) ||
1957             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1958              id_priv->state != RDMA_CM_DISCONNECT))
1959                 goto out;
1960
1961         switch (ib_event->event) {
1962         case IB_CM_REQ_ERROR:
1963         case IB_CM_REP_ERROR:
1964                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1965                 event.status = -ETIMEDOUT;
1966                 break;
1967         case IB_CM_REP_RECEIVED:
1968                 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1969                     (id_priv->id.qp_type != IB_QPT_UD)) {
1970                         trace_cm_send_mra(id_priv);
1971                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1972                 }
1973                 if (id_priv->id.qp) {
1974                         event.status = cma_rep_recv(id_priv);
1975                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1976                                                      RDMA_CM_EVENT_ESTABLISHED;
1977                 } else {
1978                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1979                 }
1980                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1981                                        ib_event->private_data);
1982                 break;
1983         case IB_CM_RTU_RECEIVED:
1984         case IB_CM_USER_ESTABLISHED:
1985                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1986                 break;
1987         case IB_CM_DREQ_ERROR:
1988                 event.status = -ETIMEDOUT; /* fall through */
1989         case IB_CM_DREQ_RECEIVED:
1990         case IB_CM_DREP_RECEIVED:
1991                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1992                                    RDMA_CM_DISCONNECT))
1993                         goto out;
1994                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1995                 break;
1996         case IB_CM_TIMEWAIT_EXIT:
1997                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1998                 break;
1999         case IB_CM_MRA_RECEIVED:
2000                 /* ignore event */
2001                 goto out;
2002         case IB_CM_REJ_RECEIVED:
2003                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2004                                                                                 ib_event->param.rej_rcvd.reason));
2005                 cma_modify_qp_err(id_priv);
2006                 event.status = ib_event->param.rej_rcvd.reason;
2007                 event.event = RDMA_CM_EVENT_REJECTED;
2008                 event.param.conn.private_data = ib_event->private_data;
2009                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2010                 break;
2011         default:
2012                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2013                        ib_event->event);
2014                 goto out;
2015         }
2016
2017         ret = cma_cm_event_handler(id_priv, &event);
2018         if (ret) {
2019                 /* Destroy the CM ID by returning a non-zero value. */
2020                 id_priv->cm_id.ib = NULL;
2021                 destroy_id_handler_unlock(id_priv);
2022                 return ret;
2023         }
2024 out:
2025         mutex_unlock(&id_priv->handler_mutex);
2026         return 0;
2027 }
2028
2029 static struct rdma_id_private *
2030 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2031                    const struct ib_cm_event *ib_event,
2032                    struct net_device *net_dev)
2033 {
2034         struct rdma_id_private *listen_id_priv;
2035         struct rdma_id_private *id_priv;
2036         struct rdma_cm_id *id;
2037         struct rdma_route *rt;
2038         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2039         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2040         const __be64 service_id =
2041                 ib_event->param.req_rcvd.primary_path->service_id;
2042         int ret;
2043
2044         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2045         id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2046                             listen_id->event_handler, listen_id->context,
2047                             listen_id->ps, ib_event->param.req_rcvd.qp_type,
2048                             listen_id_priv->res.kern_name);
2049         if (IS_ERR(id))
2050                 return NULL;
2051
2052         id_priv = container_of(id, struct rdma_id_private, id);
2053         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2054                               (struct sockaddr *)&id->route.addr.dst_addr,
2055                               listen_id, ib_event, ss_family, service_id))
2056                 goto err;
2057
2058         rt = &id->route;
2059         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2060         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2061                                      GFP_KERNEL);
2062         if (!rt->path_rec)
2063                 goto err;
2064
2065         rt->path_rec[0] = *path;
2066         if (rt->num_paths == 2)
2067                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2068
2069         if (net_dev) {
2070                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2071         } else {
2072                 if (!cma_protocol_roce(listen_id) &&
2073                     cma_any_addr(cma_src_addr(id_priv))) {
2074                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2075                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2076                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2077                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2078                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2079                         if (ret)
2080                                 goto err;
2081                 }
2082         }
2083         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2084
2085         id_priv->state = RDMA_CM_CONNECT;
2086         return id_priv;
2087
2088 err:
2089         rdma_destroy_id(id);
2090         return NULL;
2091 }
2092
2093 static struct rdma_id_private *
2094 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2095                   const struct ib_cm_event *ib_event,
2096                   struct net_device *net_dev)
2097 {
2098         const struct rdma_id_private *listen_id_priv;
2099         struct rdma_id_private *id_priv;
2100         struct rdma_cm_id *id;
2101         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2102         struct net *net = listen_id->route.addr.dev_addr.net;
2103         int ret;
2104
2105         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2106         id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
2107                               listen_id->ps, IB_QPT_UD,
2108                               listen_id_priv->res.kern_name);
2109         if (IS_ERR(id))
2110                 return NULL;
2111
2112         id_priv = container_of(id, struct rdma_id_private, id);
2113         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2114                               (struct sockaddr *)&id->route.addr.dst_addr,
2115                               listen_id, ib_event, ss_family,
2116                               ib_event->param.sidr_req_rcvd.service_id))
2117                 goto err;
2118
2119         if (net_dev) {
2120                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2121         } else {
2122                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2123                         ret = cma_translate_addr(cma_src_addr(id_priv),
2124                                                  &id->route.addr.dev_addr);
2125                         if (ret)
2126                                 goto err;
2127                 }
2128         }
2129
2130         id_priv->state = RDMA_CM_CONNECT;
2131         return id_priv;
2132 err:
2133         rdma_destroy_id(id);
2134         return NULL;
2135 }
2136
2137 static void cma_set_req_event_data(struct rdma_cm_event *event,
2138                                    const struct ib_cm_req_event_param *req_data,
2139                                    void *private_data, int offset)
2140 {
2141         event->param.conn.private_data = private_data + offset;
2142         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2143         event->param.conn.responder_resources = req_data->responder_resources;
2144         event->param.conn.initiator_depth = req_data->initiator_depth;
2145         event->param.conn.flow_control = req_data->flow_control;
2146         event->param.conn.retry_count = req_data->retry_count;
2147         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2148         event->param.conn.srq = req_data->srq;
2149         event->param.conn.qp_num = req_data->remote_qpn;
2150
2151         event->ece.vendor_id = req_data->ece.vendor_id;
2152         event->ece.attr_mod = req_data->ece.attr_mod;
2153 }
2154
2155 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2156                                     const struct ib_cm_event *ib_event)
2157 {
2158         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2159                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2160                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2161                  (id->qp_type == IB_QPT_UD)) ||
2162                 (!id->qp_type));
2163 }
2164
2165 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2166                               const struct ib_cm_event *ib_event)
2167 {
2168         struct rdma_id_private *listen_id, *conn_id = NULL;
2169         struct rdma_cm_event event = {};
2170         struct cma_req_info req = {};
2171         struct net_device *net_dev;
2172         u8 offset;
2173         int ret;
2174
2175         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2176         if (IS_ERR(listen_id))
2177                 return PTR_ERR(listen_id);
2178
2179         trace_cm_req_handler(listen_id, ib_event->event);
2180         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2181                 ret = -EINVAL;
2182                 goto net_dev_put;
2183         }
2184
2185         mutex_lock(&listen_id->handler_mutex);
2186         if (listen_id->state != RDMA_CM_LISTEN) {
2187                 ret = -ECONNABORTED;
2188                 goto err_unlock;
2189         }
2190
2191         offset = cma_user_data_offset(listen_id);
2192         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2193         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2194                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2195                 event.param.ud.private_data = ib_event->private_data + offset;
2196                 event.param.ud.private_data_len =
2197                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2198         } else {
2199                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2200                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2201                                        ib_event->private_data, offset);
2202         }
2203         if (!conn_id) {
2204                 ret = -ENOMEM;
2205                 goto err_unlock;
2206         }
2207
2208         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2209         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2210         if (ret) {
2211                 destroy_id_handler_unlock(conn_id);
2212                 goto err_unlock;
2213         }
2214
2215         conn_id->cm_id.ib = cm_id;
2216         cm_id->context = conn_id;
2217         cm_id->cm_handler = cma_ib_handler;
2218
2219         ret = cma_cm_event_handler(conn_id, &event);
2220         if (ret) {
2221                 /* Destroy the CM ID by returning a non-zero value. */
2222                 conn_id->cm_id.ib = NULL;
2223                 mutex_unlock(&listen_id->handler_mutex);
2224                 destroy_id_handler_unlock(conn_id);
2225                 goto net_dev_put;
2226         }
2227
2228         if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2229             (conn_id->id.qp_type != IB_QPT_UD)) {
2230                 trace_cm_send_mra(cm_id->context);
2231                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2232         }
2233         mutex_unlock(&conn_id->handler_mutex);
2234
2235 err_unlock:
2236         mutex_unlock(&listen_id->handler_mutex);
2237
2238 net_dev_put:
2239         if (net_dev)
2240                 dev_put(net_dev);
2241
2242         return ret;
2243 }
2244
2245 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2246 {
2247         if (addr->sa_family == AF_IB)
2248                 return ((struct sockaddr_ib *) addr)->sib_sid;
2249
2250         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2251 }
2252 EXPORT_SYMBOL(rdma_get_service_id);
2253
2254 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2255                     union ib_gid *dgid)
2256 {
2257         struct rdma_addr *addr = &cm_id->route.addr;
2258
2259         if (!cm_id->device) {
2260                 if (sgid)
2261                         memset(sgid, 0, sizeof(*sgid));
2262                 if (dgid)
2263                         memset(dgid, 0, sizeof(*dgid));
2264                 return;
2265         }
2266
2267         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2268                 if (sgid)
2269                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2270                 if (dgid)
2271                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2272         } else {
2273                 if (sgid)
2274                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2275                 if (dgid)
2276                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2277         }
2278 }
2279 EXPORT_SYMBOL(rdma_read_gids);
2280
2281 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2282 {
2283         struct rdma_id_private *id_priv = iw_id->context;
2284         struct rdma_cm_event event = {};
2285         int ret = 0;
2286         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2287         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2288
2289         mutex_lock(&id_priv->handler_mutex);
2290         if (id_priv->state != RDMA_CM_CONNECT)
2291                 goto out;
2292
2293         switch (iw_event->event) {
2294         case IW_CM_EVENT_CLOSE:
2295                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2296                 break;
2297         case IW_CM_EVENT_CONNECT_REPLY:
2298                 memcpy(cma_src_addr(id_priv), laddr,
2299                        rdma_addr_size(laddr));
2300                 memcpy(cma_dst_addr(id_priv), raddr,
2301                        rdma_addr_size(raddr));
2302                 switch (iw_event->status) {
2303                 case 0:
2304                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2305                         event.param.conn.initiator_depth = iw_event->ird;
2306                         event.param.conn.responder_resources = iw_event->ord;
2307                         break;
2308                 case -ECONNRESET:
2309                 case -ECONNREFUSED:
2310                         event.event = RDMA_CM_EVENT_REJECTED;
2311                         break;
2312                 case -ETIMEDOUT:
2313                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2314                         break;
2315                 default:
2316                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2317                         break;
2318                 }
2319                 break;
2320         case IW_CM_EVENT_ESTABLISHED:
2321                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2322                 event.param.conn.initiator_depth = iw_event->ird;
2323                 event.param.conn.responder_resources = iw_event->ord;
2324                 break;
2325         default:
2326                 goto out;
2327         }
2328
2329         event.status = iw_event->status;
2330         event.param.conn.private_data = iw_event->private_data;
2331         event.param.conn.private_data_len = iw_event->private_data_len;
2332         ret = cma_cm_event_handler(id_priv, &event);
2333         if (ret) {
2334                 /* Destroy the CM ID by returning a non-zero value. */
2335                 id_priv->cm_id.iw = NULL;
2336                 destroy_id_handler_unlock(id_priv);
2337                 return ret;
2338         }
2339
2340 out:
2341         mutex_unlock(&id_priv->handler_mutex);
2342         return ret;
2343 }
2344
2345 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2346                                struct iw_cm_event *iw_event)
2347 {
2348         struct rdma_cm_id *new_cm_id;
2349         struct rdma_id_private *listen_id, *conn_id;
2350         struct rdma_cm_event event = {};
2351         int ret = -ECONNABORTED;
2352         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2353         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2354
2355         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2356         event.param.conn.private_data = iw_event->private_data;
2357         event.param.conn.private_data_len = iw_event->private_data_len;
2358         event.param.conn.initiator_depth = iw_event->ird;
2359         event.param.conn.responder_resources = iw_event->ord;
2360
2361         listen_id = cm_id->context;
2362
2363         mutex_lock(&listen_id->handler_mutex);
2364         if (listen_id->state != RDMA_CM_LISTEN)
2365                 goto out;
2366
2367         /* Create a new RDMA id for the new IW CM ID */
2368         new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2369                                      listen_id->id.event_handler,
2370                                      listen_id->id.context,
2371                                      RDMA_PS_TCP, IB_QPT_RC,
2372                                      listen_id->res.kern_name);
2373         if (IS_ERR(new_cm_id)) {
2374                 ret = -ENOMEM;
2375                 goto out;
2376         }
2377         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2378         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2379         conn_id->state = RDMA_CM_CONNECT;
2380
2381         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2382         if (ret) {
2383                 mutex_unlock(&listen_id->handler_mutex);
2384                 destroy_id_handler_unlock(conn_id);
2385                 return ret;
2386         }
2387
2388         ret = cma_iw_acquire_dev(conn_id, listen_id);
2389         if (ret) {
2390                 mutex_unlock(&listen_id->handler_mutex);
2391                 destroy_id_handler_unlock(conn_id);
2392                 return ret;
2393         }
2394
2395         conn_id->cm_id.iw = cm_id;
2396         cm_id->context = conn_id;
2397         cm_id->cm_handler = cma_iw_handler;
2398
2399         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2400         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2401
2402         ret = cma_cm_event_handler(conn_id, &event);
2403         if (ret) {
2404                 /* User wants to destroy the CM ID */
2405                 conn_id->cm_id.iw = NULL;
2406                 mutex_unlock(&listen_id->handler_mutex);
2407                 destroy_id_handler_unlock(conn_id);
2408                 return ret;
2409         }
2410
2411         mutex_unlock(&conn_id->handler_mutex);
2412
2413 out:
2414         mutex_unlock(&listen_id->handler_mutex);
2415         return ret;
2416 }
2417
2418 static int cma_ib_listen(struct rdma_id_private *id_priv)
2419 {
2420         struct sockaddr *addr;
2421         struct ib_cm_id *id;
2422         __be64 svc_id;
2423
2424         addr = cma_src_addr(id_priv);
2425         svc_id = rdma_get_service_id(&id_priv->id, addr);
2426         id = ib_cm_insert_listen(id_priv->id.device,
2427                                  cma_ib_req_handler, svc_id);
2428         if (IS_ERR(id))
2429                 return PTR_ERR(id);
2430         id_priv->cm_id.ib = id;
2431
2432         return 0;
2433 }
2434
2435 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2436 {
2437         int ret;
2438         struct iw_cm_id *id;
2439
2440         id = iw_create_cm_id(id_priv->id.device,
2441                              iw_conn_req_handler,
2442                              id_priv);
2443         if (IS_ERR(id))
2444                 return PTR_ERR(id);
2445
2446         id->tos = id_priv->tos;
2447         id->tos_set = id_priv->tos_set;
2448         id_priv->cm_id.iw = id;
2449
2450         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2451                rdma_addr_size(cma_src_addr(id_priv)));
2452
2453         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2454
2455         if (ret) {
2456                 iw_destroy_cm_id(id_priv->cm_id.iw);
2457                 id_priv->cm_id.iw = NULL;
2458         }
2459
2460         return ret;
2461 }
2462
2463 static int cma_listen_handler(struct rdma_cm_id *id,
2464                               struct rdma_cm_event *event)
2465 {
2466         struct rdma_id_private *id_priv = id->context;
2467
2468         /* Listening IDs are always destroyed on removal */
2469         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2470                 return -1;
2471
2472         id->context = id_priv->id.context;
2473         id->event_handler = id_priv->id.event_handler;
2474         trace_cm_event_handler(id_priv, event);
2475         return id_priv->id.event_handler(id, event);
2476 }
2477
2478 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2479                               struct cma_device *cma_dev)
2480 {
2481         struct rdma_id_private *dev_id_priv;
2482         struct rdma_cm_id *id;
2483         struct net *net = id_priv->id.route.addr.dev_addr.net;
2484         int ret;
2485
2486         lockdep_assert_held(&lock);
2487
2488         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2489                 return;
2490
2491         id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2492                               id_priv->id.qp_type, id_priv->res.kern_name);
2493         if (IS_ERR(id))
2494                 return;
2495
2496         dev_id_priv = container_of(id, struct rdma_id_private, id);
2497
2498         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2499         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2500                rdma_addr_size(cma_src_addr(id_priv)));
2501
2502         _cma_attach_to_dev(dev_id_priv, cma_dev);
2503         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2504         cma_id_get(id_priv);
2505         dev_id_priv->internal_id = 1;
2506         dev_id_priv->afonly = id_priv->afonly;
2507         dev_id_priv->tos_set = id_priv->tos_set;
2508         dev_id_priv->tos = id_priv->tos;
2509
2510         ret = rdma_listen(id, id_priv->backlog);
2511         if (ret)
2512                 dev_warn(&cma_dev->device->dev,
2513                          "RDMA CMA: cma_listen_on_dev, error %d\n", ret);
2514 }
2515
2516 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2517 {
2518         struct cma_device *cma_dev;
2519
2520         mutex_lock(&lock);
2521         list_add_tail(&id_priv->list, &listen_any_list);
2522         list_for_each_entry(cma_dev, &dev_list, list)
2523                 cma_listen_on_dev(id_priv, cma_dev);
2524         mutex_unlock(&lock);
2525 }
2526
2527 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2528 {
2529         struct rdma_id_private *id_priv;
2530
2531         id_priv = container_of(id, struct rdma_id_private, id);
2532         id_priv->tos = (u8) tos;
2533         id_priv->tos_set = true;
2534 }
2535 EXPORT_SYMBOL(rdma_set_service_type);
2536
2537 /**
2538  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2539  *                          with a connection identifier.
2540  * @id: Communication identifier to associated with service type.
2541  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2542  *
2543  * This function should be called before rdma_connect() on active side,
2544  * and on passive side before rdma_accept(). It is applicable to primary
2545  * path only. The timeout will affect the local side of the QP, it is not
2546  * negotiated with remote side and zero disables the timer. In case it is
2547  * set before rdma_resolve_route, the value will also be used to determine
2548  * PacketLifeTime for RoCE.
2549  *
2550  * Return: 0 for success
2551  */
2552 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2553 {
2554         struct rdma_id_private *id_priv;
2555
2556         if (id->qp_type != IB_QPT_RC)
2557                 return -EINVAL;
2558
2559         id_priv = container_of(id, struct rdma_id_private, id);
2560         id_priv->timeout = timeout;
2561         id_priv->timeout_set = true;
2562
2563         return 0;
2564 }
2565 EXPORT_SYMBOL(rdma_set_ack_timeout);
2566
2567 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2568                               void *context)
2569 {
2570         struct cma_work *work = context;
2571         struct rdma_route *route;
2572
2573         route = &work->id->id.route;
2574
2575         if (!status) {
2576                 route->num_paths = 1;
2577                 *route->path_rec = *path_rec;
2578         } else {
2579                 work->old_state = RDMA_CM_ROUTE_QUERY;
2580                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2581                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2582                 work->event.status = status;
2583                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2584                                      status);
2585         }
2586
2587         queue_work(cma_wq, &work->work);
2588 }
2589
2590 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2591                               unsigned long timeout_ms, struct cma_work *work)
2592 {
2593         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2594         struct sa_path_rec path_rec;
2595         ib_sa_comp_mask comp_mask;
2596         struct sockaddr_in6 *sin6;
2597         struct sockaddr_ib *sib;
2598
2599         memset(&path_rec, 0, sizeof path_rec);
2600
2601         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2602                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2603         else
2604                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2605         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2606         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2607         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2608         path_rec.numb_path = 1;
2609         path_rec.reversible = 1;
2610         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2611                                                   cma_dst_addr(id_priv));
2612
2613         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2614                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2615                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2616
2617         switch (cma_family(id_priv)) {
2618         case AF_INET:
2619                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2620                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2621                 break;
2622         case AF_INET6:
2623                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2624                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2625                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2626                 break;
2627         case AF_IB:
2628                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2629                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2630                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2631                 break;
2632         }
2633
2634         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2635                                                id_priv->id.port_num, &path_rec,
2636                                                comp_mask, timeout_ms,
2637                                                GFP_KERNEL, cma_query_handler,
2638                                                work, &id_priv->query);
2639
2640         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2641 }
2642
2643 static void cma_work_handler(struct work_struct *_work)
2644 {
2645         struct cma_work *work = container_of(_work, struct cma_work, work);
2646         struct rdma_id_private *id_priv = work->id;
2647
2648         mutex_lock(&id_priv->handler_mutex);
2649         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2650                 goto out_unlock;
2651
2652         if (cma_cm_event_handler(id_priv, &work->event)) {
2653                 cma_id_put(id_priv);
2654                 destroy_id_handler_unlock(id_priv);
2655                 goto out_free;
2656         }
2657
2658 out_unlock:
2659         mutex_unlock(&id_priv->handler_mutex);
2660         cma_id_put(id_priv);
2661 out_free:
2662         kfree(work);
2663 }
2664
2665 static void cma_ndev_work_handler(struct work_struct *_work)
2666 {
2667         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2668         struct rdma_id_private *id_priv = work->id;
2669
2670         mutex_lock(&id_priv->handler_mutex);
2671         if (id_priv->state == RDMA_CM_DESTROYING ||
2672             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2673                 goto out_unlock;
2674
2675         if (cma_cm_event_handler(id_priv, &work->event)) {
2676                 cma_id_put(id_priv);
2677                 destroy_id_handler_unlock(id_priv);
2678                 goto out_free;
2679         }
2680
2681 out_unlock:
2682         mutex_unlock(&id_priv->handler_mutex);
2683         cma_id_put(id_priv);
2684 out_free:
2685         kfree(work);
2686 }
2687
2688 static void cma_init_resolve_route_work(struct cma_work *work,
2689                                         struct rdma_id_private *id_priv)
2690 {
2691         work->id = id_priv;
2692         INIT_WORK(&work->work, cma_work_handler);
2693         work->old_state = RDMA_CM_ROUTE_QUERY;
2694         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2695         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2696 }
2697
2698 static void enqueue_resolve_addr_work(struct cma_work *work,
2699                                       struct rdma_id_private *id_priv)
2700 {
2701         /* Balances with cma_id_put() in cma_work_handler */
2702         cma_id_get(id_priv);
2703
2704         work->id = id_priv;
2705         INIT_WORK(&work->work, cma_work_handler);
2706         work->old_state = RDMA_CM_ADDR_QUERY;
2707         work->new_state = RDMA_CM_ADDR_RESOLVED;
2708         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2709
2710         queue_work(cma_wq, &work->work);
2711 }
2712
2713 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2714                                 unsigned long timeout_ms)
2715 {
2716         struct rdma_route *route = &id_priv->id.route;
2717         struct cma_work *work;
2718         int ret;
2719
2720         work = kzalloc(sizeof *work, GFP_KERNEL);
2721         if (!work)
2722                 return -ENOMEM;
2723
2724         cma_init_resolve_route_work(work, id_priv);
2725
2726         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2727         if (!route->path_rec) {
2728                 ret = -ENOMEM;
2729                 goto err1;
2730         }
2731
2732         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2733         if (ret)
2734                 goto err2;
2735
2736         return 0;
2737 err2:
2738         kfree(route->path_rec);
2739         route->path_rec = NULL;
2740 err1:
2741         kfree(work);
2742         return ret;
2743 }
2744
2745 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2746                                            unsigned long supported_gids,
2747                                            enum ib_gid_type default_gid)
2748 {
2749         if ((network_type == RDMA_NETWORK_IPV4 ||
2750              network_type == RDMA_NETWORK_IPV6) &&
2751             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2752                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2753
2754         return default_gid;
2755 }
2756
2757 /*
2758  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2759  * path record type based on GID type.
2760  * It also sets up other L2 fields which includes destination mac address
2761  * netdev ifindex, of the path record.
2762  * It returns the netdev of the bound interface for this path record entry.
2763  */
2764 static struct net_device *
2765 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2766 {
2767         struct rdma_route *route = &id_priv->id.route;
2768         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2769         struct rdma_addr *addr = &route->addr;
2770         unsigned long supported_gids;
2771         struct net_device *ndev;
2772
2773         if (!addr->dev_addr.bound_dev_if)
2774                 return NULL;
2775
2776         ndev = dev_get_by_index(addr->dev_addr.net,
2777                                 addr->dev_addr.bound_dev_if);
2778         if (!ndev)
2779                 return NULL;
2780
2781         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2782                                                     id_priv->id.port_num);
2783         gid_type = cma_route_gid_type(addr->dev_addr.network,
2784                                       supported_gids,
2785                                       id_priv->gid_type);
2786         /* Use the hint from IP Stack to select GID Type */
2787         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2788                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2789         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2790
2791         route->path_rec->roce.route_resolved = true;
2792         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2793         return ndev;
2794 }
2795
2796 int rdma_set_ib_path(struct rdma_cm_id *id,
2797                      struct sa_path_rec *path_rec)
2798 {
2799         struct rdma_id_private *id_priv;
2800         struct net_device *ndev;
2801         int ret;
2802
2803         id_priv = container_of(id, struct rdma_id_private, id);
2804         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2805                            RDMA_CM_ROUTE_RESOLVED))
2806                 return -EINVAL;
2807
2808         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2809                                      GFP_KERNEL);
2810         if (!id->route.path_rec) {
2811                 ret = -ENOMEM;
2812                 goto err;
2813         }
2814
2815         if (rdma_protocol_roce(id->device, id->port_num)) {
2816                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2817                 if (!ndev) {
2818                         ret = -ENODEV;
2819                         goto err_free;
2820                 }
2821                 dev_put(ndev);
2822         }
2823
2824         id->route.num_paths = 1;
2825         return 0;
2826
2827 err_free:
2828         kfree(id->route.path_rec);
2829         id->route.path_rec = NULL;
2830 err:
2831         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2832         return ret;
2833 }
2834 EXPORT_SYMBOL(rdma_set_ib_path);
2835
2836 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2837 {
2838         struct cma_work *work;
2839
2840         work = kzalloc(sizeof *work, GFP_KERNEL);
2841         if (!work)
2842                 return -ENOMEM;
2843
2844         cma_init_resolve_route_work(work, id_priv);
2845         queue_work(cma_wq, &work->work);
2846         return 0;
2847 }
2848
2849 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2850 {
2851         struct net_device *dev;
2852
2853         dev = vlan_dev_real_dev(vlan_ndev);
2854         if (dev->num_tc)
2855                 return netdev_get_prio_tc_map(dev, prio);
2856
2857         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2858                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2859 }
2860
2861 struct iboe_prio_tc_map {
2862         int input_prio;
2863         int output_tc;
2864         bool found;
2865 };
2866
2867 static int get_lower_vlan_dev_tc(struct net_device *dev, void *data)
2868 {
2869         struct iboe_prio_tc_map *map = data;
2870
2871         if (is_vlan_dev(dev))
2872                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2873         else if (dev->num_tc)
2874                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2875         else
2876                 map->output_tc = 0;
2877         /* We are interested only in first level VLAN device, so always
2878          * return 1 to stop iterating over next level devices.
2879          */
2880         map->found = true;
2881         return 1;
2882 }
2883
2884 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2885 {
2886         struct iboe_prio_tc_map prio_tc_map = {};
2887         int prio = rt_tos2priority(tos);
2888
2889         /* If VLAN device, get it directly from the VLAN netdev */
2890         if (is_vlan_dev(ndev))
2891                 return get_vlan_ndev_tc(ndev, prio);
2892
2893         prio_tc_map.input_prio = prio;
2894         rcu_read_lock();
2895         netdev_walk_all_lower_dev_rcu(ndev,
2896                                       get_lower_vlan_dev_tc,
2897                                       &prio_tc_map);
2898         rcu_read_unlock();
2899         /* If map is found from lower device, use it; Otherwise
2900          * continue with the current netdevice to get priority to tc map.
2901          */
2902         if (prio_tc_map.found)
2903                 return prio_tc_map.output_tc;
2904         else if (ndev->num_tc)
2905                 return netdev_get_prio_tc_map(ndev, prio);
2906         else
2907                 return 0;
2908 }
2909
2910 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
2911 {
2912         struct sockaddr_in6 *addr6;
2913         u16 dport, sport;
2914         u32 hash, fl;
2915
2916         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
2917         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
2918         if ((cma_family(id_priv) != AF_INET6) || !fl) {
2919                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
2920                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
2921                 hash = (u32)sport * 31 + dport;
2922                 fl = hash & IB_GRH_FLOWLABEL_MASK;
2923         }
2924
2925         return cpu_to_be32(fl);
2926 }
2927
2928 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2929 {
2930         struct rdma_route *route = &id_priv->id.route;
2931         struct rdma_addr *addr = &route->addr;
2932         struct cma_work *work;
2933         int ret;
2934         struct net_device *ndev;
2935
2936         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2937                                         rdma_start_port(id_priv->cma_dev->device)];
2938         u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2939
2940
2941         work = kzalloc(sizeof *work, GFP_KERNEL);
2942         if (!work)
2943                 return -ENOMEM;
2944
2945         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2946         if (!route->path_rec) {
2947                 ret = -ENOMEM;
2948                 goto err1;
2949         }
2950
2951         route->num_paths = 1;
2952
2953         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2954         if (!ndev) {
2955                 ret = -ENODEV;
2956                 goto err2;
2957         }
2958
2959         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2960                     &route->path_rec->sgid);
2961         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2962                     &route->path_rec->dgid);
2963
2964         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2965                 /* TODO: get the hoplimit from the inet/inet6 device */
2966                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2967         else
2968                 route->path_rec->hop_limit = 1;
2969         route->path_rec->reversible = 1;
2970         route->path_rec->pkey = cpu_to_be16(0xffff);
2971         route->path_rec->mtu_selector = IB_SA_EQ;
2972         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2973         route->path_rec->traffic_class = tos;
2974         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2975         route->path_rec->rate_selector = IB_SA_EQ;
2976         route->path_rec->rate = iboe_get_rate(ndev);
2977         dev_put(ndev);
2978         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2979         /* In case ACK timeout is set, use this value to calculate
2980          * PacketLifeTime.  As per IBTA 12.7.34,
2981          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
2982          * Assuming a negligible local ACK delay, we can use
2983          * PacketLifeTime = local ACK timeout/2
2984          * as a reasonable approximation for RoCE networks.
2985          */
2986         route->path_rec->packet_life_time = id_priv->timeout_set ?
2987                 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
2988
2989         if (!route->path_rec->mtu) {
2990                 ret = -EINVAL;
2991                 goto err2;
2992         }
2993
2994         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
2995                                          id_priv->id.port_num))
2996                 route->path_rec->flow_label =
2997                         cma_get_roce_udp_flow_label(id_priv);
2998
2999         cma_init_resolve_route_work(work, id_priv);
3000         queue_work(cma_wq, &work->work);
3001
3002         return 0;
3003
3004 err2:
3005         kfree(route->path_rec);
3006         route->path_rec = NULL;
3007         route->num_paths = 0;
3008 err1:
3009         kfree(work);
3010         return ret;
3011 }
3012
3013 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3014 {
3015         struct rdma_id_private *id_priv;
3016         int ret;
3017
3018         id_priv = container_of(id, struct rdma_id_private, id);
3019         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3020                 return -EINVAL;
3021
3022         cma_id_get(id_priv);
3023         if (rdma_cap_ib_sa(id->device, id->port_num))
3024                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3025         else if (rdma_protocol_roce(id->device, id->port_num))
3026                 ret = cma_resolve_iboe_route(id_priv);
3027         else if (rdma_protocol_iwarp(id->device, id->port_num))
3028                 ret = cma_resolve_iw_route(id_priv);
3029         else
3030                 ret = -ENOSYS;
3031
3032         if (ret)
3033                 goto err;
3034
3035         return 0;
3036 err:
3037         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3038         cma_id_put(id_priv);
3039         return ret;
3040 }
3041 EXPORT_SYMBOL(rdma_resolve_route);
3042
3043 static void cma_set_loopback(struct sockaddr *addr)
3044 {
3045         switch (addr->sa_family) {
3046         case AF_INET:
3047                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3048                 break;
3049         case AF_INET6:
3050                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3051                               0, 0, 0, htonl(1));
3052                 break;
3053         default:
3054                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3055                             0, 0, 0, htonl(1));
3056                 break;
3057         }
3058 }
3059
3060 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3061 {
3062         struct cma_device *cma_dev, *cur_dev;
3063         union ib_gid gid;
3064         enum ib_port_state port_state;
3065         unsigned int p;
3066         u16 pkey;
3067         int ret;
3068
3069         cma_dev = NULL;
3070         mutex_lock(&lock);
3071         list_for_each_entry(cur_dev, &dev_list, list) {
3072                 if (cma_family(id_priv) == AF_IB &&
3073                     !rdma_cap_ib_cm(cur_dev->device, 1))
3074                         continue;
3075
3076                 if (!cma_dev)
3077                         cma_dev = cur_dev;
3078
3079                 rdma_for_each_port (cur_dev->device, p) {
3080                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3081                             port_state == IB_PORT_ACTIVE) {
3082                                 cma_dev = cur_dev;
3083                                 goto port_found;
3084                         }
3085                 }
3086         }
3087
3088         if (!cma_dev) {
3089                 ret = -ENODEV;
3090                 goto out;
3091         }
3092
3093         p = 1;
3094
3095 port_found:
3096         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3097         if (ret)
3098                 goto out;
3099
3100         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3101         if (ret)
3102                 goto out;
3103
3104         id_priv->id.route.addr.dev_addr.dev_type =
3105                 (rdma_protocol_ib(cma_dev->device, p)) ?
3106                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3107
3108         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3109         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3110         id_priv->id.port_num = p;
3111         cma_attach_to_dev(id_priv, cma_dev);
3112         cma_set_loopback(cma_src_addr(id_priv));
3113 out:
3114         mutex_unlock(&lock);
3115         return ret;
3116 }
3117
3118 static void addr_handler(int status, struct sockaddr *src_addr,
3119                          struct rdma_dev_addr *dev_addr, void *context)
3120 {
3121         struct rdma_id_private *id_priv = context;
3122         struct rdma_cm_event event = {};
3123         struct sockaddr *addr;
3124         struct sockaddr_storage old_addr;
3125
3126         mutex_lock(&id_priv->handler_mutex);
3127         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3128                            RDMA_CM_ADDR_RESOLVED))
3129                 goto out;
3130
3131         /*
3132          * Store the previous src address, so that if we fail to acquire
3133          * matching rdma device, old address can be restored back, which helps
3134          * to cancel the cma listen operation correctly.
3135          */
3136         addr = cma_src_addr(id_priv);
3137         memcpy(&old_addr, addr, rdma_addr_size(addr));
3138         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3139         if (!status && !id_priv->cma_dev) {
3140                 status = cma_acquire_dev_by_src_ip(id_priv);
3141                 if (status)
3142                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3143                                              status);
3144         } else if (status) {
3145                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3146         }
3147
3148         if (status) {
3149                 memcpy(addr, &old_addr,
3150                        rdma_addr_size((struct sockaddr *)&old_addr));
3151                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3152                                    RDMA_CM_ADDR_BOUND))
3153                         goto out;
3154                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3155                 event.status = status;
3156         } else
3157                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3158
3159         if (cma_cm_event_handler(id_priv, &event)) {
3160                 destroy_id_handler_unlock(id_priv);
3161                 return;
3162         }
3163 out:
3164         mutex_unlock(&id_priv->handler_mutex);
3165 }
3166
3167 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3168 {
3169         struct cma_work *work;
3170         union ib_gid gid;
3171         int ret;
3172
3173         work = kzalloc(sizeof *work, GFP_KERNEL);
3174         if (!work)
3175                 return -ENOMEM;
3176
3177         if (!id_priv->cma_dev) {
3178                 ret = cma_bind_loopback(id_priv);
3179                 if (ret)
3180                         goto err;
3181         }
3182
3183         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3184         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3185
3186         enqueue_resolve_addr_work(work, id_priv);
3187         return 0;
3188 err:
3189         kfree(work);
3190         return ret;
3191 }
3192
3193 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3194 {
3195         struct cma_work *work;
3196         int ret;
3197
3198         work = kzalloc(sizeof *work, GFP_KERNEL);
3199         if (!work)
3200                 return -ENOMEM;
3201
3202         if (!id_priv->cma_dev) {
3203                 ret = cma_resolve_ib_dev(id_priv);
3204                 if (ret)
3205                         goto err;
3206         }
3207
3208         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3209                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3210
3211         enqueue_resolve_addr_work(work, id_priv);
3212         return 0;
3213 err:
3214         kfree(work);
3215         return ret;
3216 }
3217
3218 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3219                          const struct sockaddr *dst_addr)
3220 {
3221         if (!src_addr || !src_addr->sa_family) {
3222                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3223                 src_addr->sa_family = dst_addr->sa_family;
3224                 if (IS_ENABLED(CONFIG_IPV6) &&
3225                     dst_addr->sa_family == AF_INET6) {
3226                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3227                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3228                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3229                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3230                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3231                 } else if (dst_addr->sa_family == AF_IB) {
3232                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3233                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3234                 }
3235         }
3236         return rdma_bind_addr(id, src_addr);
3237 }
3238
3239 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3240                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3241 {
3242         struct rdma_id_private *id_priv;
3243         int ret;
3244
3245         id_priv = container_of(id, struct rdma_id_private, id);
3246         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3247         if (id_priv->state == RDMA_CM_IDLE) {
3248                 ret = cma_bind_addr(id, src_addr, dst_addr);
3249                 if (ret) {
3250                         memset(cma_dst_addr(id_priv), 0,
3251                                rdma_addr_size(dst_addr));
3252                         return ret;
3253                 }
3254         }
3255
3256         if (cma_family(id_priv) != dst_addr->sa_family) {
3257                 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3258                 return -EINVAL;
3259         }
3260
3261         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3262                 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3263                 return -EINVAL;
3264         }
3265
3266         if (cma_any_addr(dst_addr)) {
3267                 ret = cma_resolve_loopback(id_priv);
3268         } else {
3269                 if (dst_addr->sa_family == AF_IB) {
3270                         ret = cma_resolve_ib_addr(id_priv);
3271                 } else {
3272                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3273                                               &id->route.addr.dev_addr,
3274                                               timeout_ms, addr_handler,
3275                                               false, id_priv);
3276                 }
3277         }
3278         if (ret)
3279                 goto err;
3280
3281         return 0;
3282 err:
3283         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3284         return ret;
3285 }
3286 EXPORT_SYMBOL(rdma_resolve_addr);
3287
3288 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3289 {
3290         struct rdma_id_private *id_priv;
3291         unsigned long flags;
3292         int ret;
3293
3294         id_priv = container_of(id, struct rdma_id_private, id);
3295         spin_lock_irqsave(&id_priv->lock, flags);
3296         if (reuse || id_priv->state == RDMA_CM_IDLE) {
3297                 id_priv->reuseaddr = reuse;
3298                 ret = 0;
3299         } else {
3300                 ret = -EINVAL;
3301         }
3302         spin_unlock_irqrestore(&id_priv->lock, flags);
3303         return ret;
3304 }
3305 EXPORT_SYMBOL(rdma_set_reuseaddr);
3306
3307 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3308 {
3309         struct rdma_id_private *id_priv;
3310         unsigned long flags;
3311         int ret;
3312
3313         id_priv = container_of(id, struct rdma_id_private, id);
3314         spin_lock_irqsave(&id_priv->lock, flags);
3315         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3316                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3317                 id_priv->afonly = afonly;
3318                 ret = 0;
3319         } else {
3320                 ret = -EINVAL;
3321         }
3322         spin_unlock_irqrestore(&id_priv->lock, flags);
3323         return ret;
3324 }
3325 EXPORT_SYMBOL(rdma_set_afonly);
3326
3327 static void cma_bind_port(struct rdma_bind_list *bind_list,
3328                           struct rdma_id_private *id_priv)
3329 {
3330         struct sockaddr *addr;
3331         struct sockaddr_ib *sib;
3332         u64 sid, mask;
3333         __be16 port;
3334
3335         lockdep_assert_held(&lock);
3336
3337         addr = cma_src_addr(id_priv);
3338         port = htons(bind_list->port);
3339
3340         switch (addr->sa_family) {
3341         case AF_INET:
3342                 ((struct sockaddr_in *) addr)->sin_port = port;
3343                 break;
3344         case AF_INET6:
3345                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3346                 break;
3347         case AF_IB:
3348                 sib = (struct sockaddr_ib *) addr;
3349                 sid = be64_to_cpu(sib->sib_sid);
3350                 mask = be64_to_cpu(sib->sib_sid_mask);
3351                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3352                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3353                 break;
3354         }
3355         id_priv->bind_list = bind_list;
3356         hlist_add_head(&id_priv->node, &bind_list->owners);
3357 }
3358
3359 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3360                           struct rdma_id_private *id_priv, unsigned short snum)
3361 {
3362         struct rdma_bind_list *bind_list;
3363         int ret;
3364
3365         lockdep_assert_held(&lock);
3366
3367         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3368         if (!bind_list)
3369                 return -ENOMEM;
3370
3371         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3372                            snum);
3373         if (ret < 0)
3374                 goto err;
3375
3376         bind_list->ps = ps;
3377         bind_list->port = snum;
3378         cma_bind_port(bind_list, id_priv);
3379         return 0;
3380 err:
3381         kfree(bind_list);
3382         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3383 }
3384
3385 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3386                               struct rdma_id_private *id_priv)
3387 {
3388         struct rdma_id_private *cur_id;
3389         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3390         struct sockaddr  *saddr = cma_src_addr(id_priv);
3391         __be16 dport = cma_port(daddr);
3392
3393         lockdep_assert_held(&lock);
3394
3395         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3396                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3397                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3398                 __be16 cur_dport = cma_port(cur_daddr);
3399
3400                 if (id_priv == cur_id)
3401                         continue;
3402
3403                 /* different dest port -> unique */
3404                 if (!cma_any_port(daddr) &&
3405                     !cma_any_port(cur_daddr) &&
3406                     (dport != cur_dport))
3407                         continue;
3408
3409                 /* different src address -> unique */
3410                 if (!cma_any_addr(saddr) &&
3411                     !cma_any_addr(cur_saddr) &&
3412                     cma_addr_cmp(saddr, cur_saddr))
3413                         continue;
3414
3415                 /* different dst address -> unique */
3416                 if (!cma_any_addr(daddr) &&
3417                     !cma_any_addr(cur_daddr) &&
3418                     cma_addr_cmp(daddr, cur_daddr))
3419                         continue;
3420
3421                 return -EADDRNOTAVAIL;
3422         }
3423         return 0;
3424 }
3425
3426 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3427                               struct rdma_id_private *id_priv)
3428 {
3429         static unsigned int last_used_port;
3430         int low, high, remaining;
3431         unsigned int rover;
3432         struct net *net = id_priv->id.route.addr.dev_addr.net;
3433
3434         lockdep_assert_held(&lock);
3435
3436         inet_get_local_port_range(net, &low, &high);
3437         remaining = (high - low) + 1;
3438         rover = prandom_u32() % remaining + low;
3439 retry:
3440         if (last_used_port != rover) {
3441                 struct rdma_bind_list *bind_list;
3442                 int ret;
3443
3444                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3445
3446                 if (!bind_list) {
3447                         ret = cma_alloc_port(ps, id_priv, rover);
3448                 } else {
3449                         ret = cma_port_is_unique(bind_list, id_priv);
3450                         if (!ret)
3451                                 cma_bind_port(bind_list, id_priv);
3452                 }
3453                 /*
3454                  * Remember previously used port number in order to avoid
3455                  * re-using same port immediately after it is closed.
3456                  */
3457                 if (!ret)
3458                         last_used_port = rover;
3459                 if (ret != -EADDRNOTAVAIL)
3460                         return ret;
3461         }
3462         if (--remaining) {
3463                 rover++;
3464                 if ((rover < low) || (rover > high))
3465                         rover = low;
3466                 goto retry;
3467         }
3468         return -EADDRNOTAVAIL;
3469 }
3470
3471 /*
3472  * Check that the requested port is available.  This is called when trying to
3473  * bind to a specific port, or when trying to listen on a bound port.  In
3474  * the latter case, the provided id_priv may already be on the bind_list, but
3475  * we still need to check that it's okay to start listening.
3476  */
3477 static int cma_check_port(struct rdma_bind_list *bind_list,
3478                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3479 {
3480         struct rdma_id_private *cur_id;
3481         struct sockaddr *addr, *cur_addr;
3482
3483         lockdep_assert_held(&lock);
3484
3485         addr = cma_src_addr(id_priv);
3486         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3487                 if (id_priv == cur_id)
3488                         continue;
3489
3490                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3491                     cur_id->reuseaddr)
3492                         continue;
3493
3494                 cur_addr = cma_src_addr(cur_id);
3495                 if (id_priv->afonly && cur_id->afonly &&
3496                     (addr->sa_family != cur_addr->sa_family))
3497                         continue;
3498
3499                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3500                         return -EADDRNOTAVAIL;
3501
3502                 if (!cma_addr_cmp(addr, cur_addr))
3503                         return -EADDRINUSE;
3504         }
3505         return 0;
3506 }
3507
3508 static int cma_use_port(enum rdma_ucm_port_space ps,
3509                         struct rdma_id_private *id_priv)
3510 {
3511         struct rdma_bind_list *bind_list;
3512         unsigned short snum;
3513         int ret;
3514
3515         lockdep_assert_held(&lock);
3516
3517         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3518         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3519                 return -EACCES;
3520
3521         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3522         if (!bind_list) {
3523                 ret = cma_alloc_port(ps, id_priv, snum);
3524         } else {
3525                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3526                 if (!ret)
3527                         cma_bind_port(bind_list, id_priv);
3528         }
3529         return ret;
3530 }
3531
3532 static int cma_bind_listen(struct rdma_id_private *id_priv)
3533 {
3534         struct rdma_bind_list *bind_list = id_priv->bind_list;
3535         int ret = 0;
3536
3537         mutex_lock(&lock);
3538         if (bind_list->owners.first->next)
3539                 ret = cma_check_port(bind_list, id_priv, 0);
3540         mutex_unlock(&lock);
3541         return ret;
3542 }
3543
3544 static enum rdma_ucm_port_space
3545 cma_select_inet_ps(struct rdma_id_private *id_priv)
3546 {
3547         switch (id_priv->id.ps) {
3548         case RDMA_PS_TCP:
3549         case RDMA_PS_UDP:
3550         case RDMA_PS_IPOIB:
3551         case RDMA_PS_IB:
3552                 return id_priv->id.ps;
3553         default:
3554
3555                 return 0;
3556         }
3557 }
3558
3559 static enum rdma_ucm_port_space
3560 cma_select_ib_ps(struct rdma_id_private *id_priv)
3561 {
3562         enum rdma_ucm_port_space ps = 0;
3563         struct sockaddr_ib *sib;
3564         u64 sid_ps, mask, sid;
3565
3566         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3567         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3568         sid = be64_to_cpu(sib->sib_sid) & mask;
3569
3570         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3571                 sid_ps = RDMA_IB_IP_PS_IB;
3572                 ps = RDMA_PS_IB;
3573         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3574                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3575                 sid_ps = RDMA_IB_IP_PS_TCP;
3576                 ps = RDMA_PS_TCP;
3577         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3578                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3579                 sid_ps = RDMA_IB_IP_PS_UDP;
3580                 ps = RDMA_PS_UDP;
3581         }
3582
3583         if (ps) {
3584                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3585                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3586                                                 be64_to_cpu(sib->sib_sid_mask));
3587         }
3588         return ps;
3589 }
3590
3591 static int cma_get_port(struct rdma_id_private *id_priv)
3592 {
3593         enum rdma_ucm_port_space ps;
3594         int ret;
3595
3596         if (cma_family(id_priv) != AF_IB)
3597                 ps = cma_select_inet_ps(id_priv);
3598         else
3599                 ps = cma_select_ib_ps(id_priv);
3600         if (!ps)
3601                 return -EPROTONOSUPPORT;
3602
3603         mutex_lock(&lock);
3604         if (cma_any_port(cma_src_addr(id_priv)))
3605                 ret = cma_alloc_any_port(ps, id_priv);
3606         else
3607                 ret = cma_use_port(ps, id_priv);
3608         mutex_unlock(&lock);
3609
3610         return ret;
3611 }
3612
3613 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3614                                struct sockaddr *addr)
3615 {
3616 #if IS_ENABLED(CONFIG_IPV6)
3617         struct sockaddr_in6 *sin6;
3618
3619         if (addr->sa_family != AF_INET6)
3620                 return 0;
3621
3622         sin6 = (struct sockaddr_in6 *) addr;
3623
3624         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3625                 return 0;
3626
3627         if (!sin6->sin6_scope_id)
3628                         return -EINVAL;
3629
3630         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3631 #endif
3632         return 0;
3633 }
3634
3635 int rdma_listen(struct rdma_cm_id *id, int backlog)
3636 {
3637         struct rdma_id_private *id_priv;
3638         int ret;
3639
3640         id_priv = container_of(id, struct rdma_id_private, id);
3641         if (id_priv->state == RDMA_CM_IDLE) {
3642                 id->route.addr.src_addr.ss_family = AF_INET;
3643                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3644                 if (ret)
3645                         return ret;
3646         }
3647
3648         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3649                 return -EINVAL;
3650
3651         if (id_priv->reuseaddr) {
3652                 ret = cma_bind_listen(id_priv);
3653                 if (ret)
3654                         goto err;
3655         }
3656
3657         id_priv->backlog = backlog;
3658         if (id->device) {
3659                 if (rdma_cap_ib_cm(id->device, 1)) {
3660                         ret = cma_ib_listen(id_priv);
3661                         if (ret)
3662                                 goto err;
3663                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3664                         ret = cma_iw_listen(id_priv, backlog);
3665                         if (ret)
3666                                 goto err;
3667                 } else {
3668                         ret = -ENOSYS;
3669                         goto err;
3670                 }
3671         } else
3672                 cma_listen_on_all(id_priv);
3673
3674         return 0;
3675 err:
3676         id_priv->backlog = 0;
3677         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3678         return ret;
3679 }
3680 EXPORT_SYMBOL(rdma_listen);
3681
3682 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3683 {
3684         struct rdma_id_private *id_priv;
3685         int ret;
3686         struct sockaddr  *daddr;
3687
3688         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3689             addr->sa_family != AF_IB)
3690                 return -EAFNOSUPPORT;
3691
3692         id_priv = container_of(id, struct rdma_id_private, id);
3693         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3694                 return -EINVAL;
3695
3696         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3697         if (ret)
3698                 goto err1;
3699
3700         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3701         if (!cma_any_addr(addr)) {
3702                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3703                 if (ret)
3704                         goto err1;
3705
3706                 ret = cma_acquire_dev_by_src_ip(id_priv);
3707                 if (ret)
3708                         goto err1;
3709         }
3710
3711         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3712                 if (addr->sa_family == AF_INET)
3713                         id_priv->afonly = 1;
3714 #if IS_ENABLED(CONFIG_IPV6)
3715                 else if (addr->sa_family == AF_INET6) {
3716                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3717
3718                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3719                 }
3720 #endif
3721         }
3722         daddr = cma_dst_addr(id_priv);
3723         daddr->sa_family = addr->sa_family;
3724
3725         ret = cma_get_port(id_priv);
3726         if (ret)
3727                 goto err2;
3728
3729         return 0;
3730 err2:
3731         rdma_restrack_del(&id_priv->res);
3732         if (id_priv->cma_dev)
3733                 cma_release_dev(id_priv);
3734 err1:
3735         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3736         return ret;
3737 }
3738 EXPORT_SYMBOL(rdma_bind_addr);
3739
3740 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3741 {
3742         struct cma_hdr *cma_hdr;
3743
3744         cma_hdr = hdr;
3745         cma_hdr->cma_version = CMA_VERSION;
3746         if (cma_family(id_priv) == AF_INET) {
3747                 struct sockaddr_in *src4, *dst4;
3748
3749                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3750                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3751
3752                 cma_set_ip_ver(cma_hdr, 4);
3753                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3754                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3755                 cma_hdr->port = src4->sin_port;
3756         } else if (cma_family(id_priv) == AF_INET6) {
3757                 struct sockaddr_in6 *src6, *dst6;
3758
3759                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3760                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3761
3762                 cma_set_ip_ver(cma_hdr, 6);
3763                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3764                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3765                 cma_hdr->port = src6->sin6_port;
3766         }
3767         return 0;
3768 }
3769
3770 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3771                                 const struct ib_cm_event *ib_event)
3772 {
3773         struct rdma_id_private *id_priv = cm_id->context;
3774         struct rdma_cm_event event = {};
3775         const struct ib_cm_sidr_rep_event_param *rep =
3776                                 &ib_event->param.sidr_rep_rcvd;
3777         int ret;
3778
3779         mutex_lock(&id_priv->handler_mutex);
3780         if (id_priv->state != RDMA_CM_CONNECT)
3781                 goto out;
3782
3783         switch (ib_event->event) {
3784         case IB_CM_SIDR_REQ_ERROR:
3785                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3786                 event.status = -ETIMEDOUT;
3787                 break;
3788         case IB_CM_SIDR_REP_RECEIVED:
3789                 event.param.ud.private_data = ib_event->private_data;
3790                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3791                 if (rep->status != IB_SIDR_SUCCESS) {
3792                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3793                         event.status = ib_event->param.sidr_rep_rcvd.status;
3794                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3795                                              event.status);
3796                         break;
3797                 }
3798                 ret = cma_set_qkey(id_priv, rep->qkey);
3799                 if (ret) {
3800                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3801                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3802                         event.status = ret;
3803                         break;
3804                 }
3805                 ib_init_ah_attr_from_path(id_priv->id.device,
3806                                           id_priv->id.port_num,
3807                                           id_priv->id.route.path_rec,
3808                                           &event.param.ud.ah_attr,
3809                                           rep->sgid_attr);
3810                 event.param.ud.qp_num = rep->qpn;
3811                 event.param.ud.qkey = rep->qkey;
3812                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3813                 event.status = 0;
3814                 break;
3815         default:
3816                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3817                        ib_event->event);
3818                 goto out;
3819         }
3820
3821         ret = cma_cm_event_handler(id_priv, &event);
3822
3823         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3824         if (ret) {
3825                 /* Destroy the CM ID by returning a non-zero value. */
3826                 id_priv->cm_id.ib = NULL;
3827                 destroy_id_handler_unlock(id_priv);
3828                 return ret;
3829         }
3830 out:
3831         mutex_unlock(&id_priv->handler_mutex);
3832         return 0;
3833 }
3834
3835 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3836                               struct rdma_conn_param *conn_param)
3837 {
3838         struct ib_cm_sidr_req_param req;
3839         struct ib_cm_id *id;
3840         void *private_data;
3841         u8 offset;
3842         int ret;
3843
3844         memset(&req, 0, sizeof req);
3845         offset = cma_user_data_offset(id_priv);
3846         req.private_data_len = offset + conn_param->private_data_len;
3847         if (req.private_data_len < conn_param->private_data_len)
3848                 return -EINVAL;
3849
3850         if (req.private_data_len) {
3851                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3852                 if (!private_data)
3853                         return -ENOMEM;
3854         } else {
3855                 private_data = NULL;
3856         }
3857
3858         if (conn_param->private_data && conn_param->private_data_len)
3859                 memcpy(private_data + offset, conn_param->private_data,
3860                        conn_param->private_data_len);
3861
3862         if (private_data) {
3863                 ret = cma_format_hdr(private_data, id_priv);
3864                 if (ret)
3865                         goto out;
3866                 req.private_data = private_data;
3867         }
3868
3869         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3870                              id_priv);
3871         if (IS_ERR(id)) {
3872                 ret = PTR_ERR(id);
3873                 goto out;
3874         }
3875         id_priv->cm_id.ib = id;
3876
3877         req.path = id_priv->id.route.path_rec;
3878         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3879         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3880         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3881         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3882
3883         trace_cm_send_sidr_req(id_priv);
3884         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3885         if (ret) {
3886                 ib_destroy_cm_id(id_priv->cm_id.ib);
3887                 id_priv->cm_id.ib = NULL;
3888         }
3889 out:
3890         kfree(private_data);
3891         return ret;
3892 }
3893
3894 static int cma_connect_ib(struct rdma_id_private *id_priv,
3895                           struct rdma_conn_param *conn_param)
3896 {
3897         struct ib_cm_req_param req;
3898         struct rdma_route *route;
3899         void *private_data;
3900         struct ib_cm_id *id;
3901         u8 offset;
3902         int ret;
3903
3904         memset(&req, 0, sizeof req);
3905         offset = cma_user_data_offset(id_priv);
3906         req.private_data_len = offset + conn_param->private_data_len;
3907         if (req.private_data_len < conn_param->private_data_len)
3908                 return -EINVAL;
3909
3910         if (req.private_data_len) {
3911                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3912                 if (!private_data)
3913                         return -ENOMEM;
3914         } else {
3915                 private_data = NULL;
3916         }
3917
3918         if (conn_param->private_data && conn_param->private_data_len)
3919                 memcpy(private_data + offset, conn_param->private_data,
3920                        conn_param->private_data_len);
3921
3922         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3923         if (IS_ERR(id)) {
3924                 ret = PTR_ERR(id);
3925                 goto out;
3926         }
3927         id_priv->cm_id.ib = id;
3928
3929         route = &id_priv->id.route;
3930         if (private_data) {
3931                 ret = cma_format_hdr(private_data, id_priv);
3932                 if (ret)
3933                         goto out;
3934                 req.private_data = private_data;
3935         }
3936
3937         req.primary_path = &route->path_rec[0];
3938         if (route->num_paths == 2)
3939                 req.alternate_path = &route->path_rec[1];
3940
3941         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3942         /* Alternate path SGID attribute currently unsupported */
3943         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3944         req.qp_num = id_priv->qp_num;
3945         req.qp_type = id_priv->id.qp_type;
3946         req.starting_psn = id_priv->seq_num;
3947         req.responder_resources = conn_param->responder_resources;
3948         req.initiator_depth = conn_param->initiator_depth;
3949         req.flow_control = conn_param->flow_control;
3950         req.retry_count = min_t(u8, 7, conn_param->retry_count);
3951         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3952         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3953         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3954         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3955         req.srq = id_priv->srq ? 1 : 0;
3956         req.ece.vendor_id = id_priv->ece.vendor_id;
3957         req.ece.attr_mod = id_priv->ece.attr_mod;
3958
3959         trace_cm_send_req(id_priv);
3960         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3961 out:
3962         if (ret && !IS_ERR(id)) {
3963                 ib_destroy_cm_id(id);
3964                 id_priv->cm_id.ib = NULL;
3965         }
3966
3967         kfree(private_data);
3968         return ret;
3969 }
3970
3971 static int cma_connect_iw(struct rdma_id_private *id_priv,
3972                           struct rdma_conn_param *conn_param)
3973 {
3974         struct iw_cm_id *cm_id;
3975         int ret;
3976         struct iw_cm_conn_param iw_param;
3977
3978         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3979         if (IS_ERR(cm_id))
3980                 return PTR_ERR(cm_id);
3981
3982         cm_id->tos = id_priv->tos;
3983         cm_id->tos_set = id_priv->tos_set;
3984         id_priv->cm_id.iw = cm_id;
3985
3986         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3987                rdma_addr_size(cma_src_addr(id_priv)));
3988         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3989                rdma_addr_size(cma_dst_addr(id_priv)));
3990
3991         ret = cma_modify_qp_rtr(id_priv, conn_param);
3992         if (ret)
3993                 goto out;
3994
3995         if (conn_param) {
3996                 iw_param.ord = conn_param->initiator_depth;
3997                 iw_param.ird = conn_param->responder_resources;
3998                 iw_param.private_data = conn_param->private_data;
3999                 iw_param.private_data_len = conn_param->private_data_len;
4000                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4001         } else {
4002                 memset(&iw_param, 0, sizeof iw_param);
4003                 iw_param.qpn = id_priv->qp_num;
4004         }
4005         ret = iw_cm_connect(cm_id, &iw_param);
4006 out:
4007         if (ret) {
4008                 iw_destroy_cm_id(cm_id);
4009                 id_priv->cm_id.iw = NULL;
4010         }
4011         return ret;
4012 }
4013
4014 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4015 {
4016         struct rdma_id_private *id_priv;
4017         int ret;
4018
4019         id_priv = container_of(id, struct rdma_id_private, id);
4020         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4021                 return -EINVAL;
4022
4023         if (!id->qp) {
4024                 id_priv->qp_num = conn_param->qp_num;
4025                 id_priv->srq = conn_param->srq;
4026         }
4027
4028         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4029                 if (id->qp_type == IB_QPT_UD)
4030                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4031                 else
4032                         ret = cma_connect_ib(id_priv, conn_param);
4033         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4034                 ret = cma_connect_iw(id_priv, conn_param);
4035         else
4036                 ret = -ENOSYS;
4037         if (ret)
4038                 goto err;
4039
4040         return 0;
4041 err:
4042         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4043         return ret;
4044 }
4045 EXPORT_SYMBOL(rdma_connect);
4046
4047 /**
4048  * rdma_connect_ece - Initiate an active connection request with ECE data.
4049  * @id: Connection identifier to connect.
4050  * @conn_param: Connection information used for connected QPs.
4051  * @ece: ECE parameters
4052  *
4053  * See rdma_connect() explanation.
4054  */
4055 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4056                      struct rdma_ucm_ece *ece)
4057 {
4058         struct rdma_id_private *id_priv =
4059                 container_of(id, struct rdma_id_private, id);
4060
4061         id_priv->ece.vendor_id = ece->vendor_id;
4062         id_priv->ece.attr_mod = ece->attr_mod;
4063
4064         return rdma_connect(id, conn_param);
4065 }
4066 EXPORT_SYMBOL(rdma_connect_ece);
4067
4068 static int cma_accept_ib(struct rdma_id_private *id_priv,
4069                          struct rdma_conn_param *conn_param)
4070 {
4071         struct ib_cm_rep_param rep;
4072         int ret;
4073
4074         ret = cma_modify_qp_rtr(id_priv, conn_param);
4075         if (ret)
4076                 goto out;
4077
4078         ret = cma_modify_qp_rts(id_priv, conn_param);
4079         if (ret)
4080                 goto out;
4081
4082         memset(&rep, 0, sizeof rep);
4083         rep.qp_num = id_priv->qp_num;
4084         rep.starting_psn = id_priv->seq_num;
4085         rep.private_data = conn_param->private_data;
4086         rep.private_data_len = conn_param->private_data_len;
4087         rep.responder_resources = conn_param->responder_resources;
4088         rep.initiator_depth = conn_param->initiator_depth;
4089         rep.failover_accepted = 0;
4090         rep.flow_control = conn_param->flow_control;
4091         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4092         rep.srq = id_priv->srq ? 1 : 0;
4093         rep.ece.vendor_id = id_priv->ece.vendor_id;
4094         rep.ece.attr_mod = id_priv->ece.attr_mod;
4095
4096         trace_cm_send_rep(id_priv);
4097         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4098 out:
4099         return ret;
4100 }
4101
4102 static int cma_accept_iw(struct rdma_id_private *id_priv,
4103                   struct rdma_conn_param *conn_param)
4104 {
4105         struct iw_cm_conn_param iw_param;
4106         int ret;
4107
4108         if (!conn_param)
4109                 return -EINVAL;
4110
4111         ret = cma_modify_qp_rtr(id_priv, conn_param);
4112         if (ret)
4113                 return ret;
4114
4115         iw_param.ord = conn_param->initiator_depth;
4116         iw_param.ird = conn_param->responder_resources;
4117         iw_param.private_data = conn_param->private_data;
4118         iw_param.private_data_len = conn_param->private_data_len;
4119         if (id_priv->id.qp) {
4120                 iw_param.qpn = id_priv->qp_num;
4121         } else
4122                 iw_param.qpn = conn_param->qp_num;
4123
4124         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4125 }
4126
4127 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4128                              enum ib_cm_sidr_status status, u32 qkey,
4129                              const void *private_data, int private_data_len)
4130 {
4131         struct ib_cm_sidr_rep_param rep;
4132         int ret;
4133
4134         memset(&rep, 0, sizeof rep);
4135         rep.status = status;
4136         if (status == IB_SIDR_SUCCESS) {
4137                 ret = cma_set_qkey(id_priv, qkey);
4138                 if (ret)
4139                         return ret;
4140                 rep.qp_num = id_priv->qp_num;
4141                 rep.qkey = id_priv->qkey;
4142
4143                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4144                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4145         }
4146
4147         rep.private_data = private_data;
4148         rep.private_data_len = private_data_len;
4149
4150         trace_cm_send_sidr_rep(id_priv);
4151         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4152 }
4153
4154 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4155                   const char *caller)
4156 {
4157         struct rdma_id_private *id_priv;
4158         int ret;
4159
4160         id_priv = container_of(id, struct rdma_id_private, id);
4161
4162         rdma_restrack_set_task(&id_priv->res, caller);
4163
4164         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
4165                 return -EINVAL;
4166
4167         if (!id->qp && conn_param) {
4168                 id_priv->qp_num = conn_param->qp_num;
4169                 id_priv->srq = conn_param->srq;
4170         }
4171
4172         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4173                 if (id->qp_type == IB_QPT_UD) {
4174                         if (conn_param)
4175                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4176                                                         conn_param->qkey,
4177                                                         conn_param->private_data,
4178                                                         conn_param->private_data_len);
4179                         else
4180                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4181                                                         0, NULL, 0);
4182                 } else {
4183                         if (conn_param)
4184                                 ret = cma_accept_ib(id_priv, conn_param);
4185                         else
4186                                 ret = cma_rep_recv(id_priv);
4187                 }
4188         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4189                 ret = cma_accept_iw(id_priv, conn_param);
4190         else
4191                 ret = -ENOSYS;
4192
4193         if (ret)
4194                 goto reject;
4195
4196         return 0;
4197 reject:
4198         cma_modify_qp_err(id_priv);
4199         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4200         return ret;
4201 }
4202 EXPORT_SYMBOL(__rdma_accept);
4203
4204 int __rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4205                       const char *caller, struct rdma_ucm_ece *ece)
4206 {
4207         struct rdma_id_private *id_priv =
4208                 container_of(id, struct rdma_id_private, id);
4209
4210         id_priv->ece.vendor_id = ece->vendor_id;
4211         id_priv->ece.attr_mod = ece->attr_mod;
4212
4213         return __rdma_accept(id, conn_param, caller);
4214 }
4215 EXPORT_SYMBOL(__rdma_accept_ece);
4216
4217 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4218 {
4219         struct rdma_id_private *id_priv;
4220         int ret;
4221
4222         id_priv = container_of(id, struct rdma_id_private, id);
4223         if (!id_priv->cm_id.ib)
4224                 return -EINVAL;
4225
4226         switch (id->device->node_type) {
4227         case RDMA_NODE_IB_CA:
4228                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4229                 break;
4230         default:
4231                 ret = 0;
4232                 break;
4233         }
4234         return ret;
4235 }
4236 EXPORT_SYMBOL(rdma_notify);
4237
4238 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4239                 u8 private_data_len, u8 reason)
4240 {
4241         struct rdma_id_private *id_priv;
4242         int ret;
4243
4244         id_priv = container_of(id, struct rdma_id_private, id);
4245         if (!id_priv->cm_id.ib)
4246                 return -EINVAL;
4247
4248         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4249                 if (id->qp_type == IB_QPT_UD) {
4250                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4251                                                 private_data, private_data_len);
4252                 } else {
4253                         trace_cm_send_rej(id_priv);
4254                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4255                                              private_data, private_data_len);
4256                 }
4257         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4258                 ret = iw_cm_reject(id_priv->cm_id.iw,
4259                                    private_data, private_data_len);
4260         } else
4261                 ret = -ENOSYS;
4262
4263         return ret;
4264 }
4265 EXPORT_SYMBOL(rdma_reject);
4266
4267 int rdma_disconnect(struct rdma_cm_id *id)
4268 {
4269         struct rdma_id_private *id_priv;
4270         int ret;
4271
4272         id_priv = container_of(id, struct rdma_id_private, id);
4273         if (!id_priv->cm_id.ib)
4274                 return -EINVAL;
4275
4276         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4277                 ret = cma_modify_qp_err(id_priv);
4278                 if (ret)
4279                         goto out;
4280                 /* Initiate or respond to a disconnect. */
4281                 trace_cm_disconnect(id_priv);
4282                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4283                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4284                                 trace_cm_sent_drep(id_priv);
4285                 } else {
4286                         trace_cm_sent_dreq(id_priv);
4287                 }
4288         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4289                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4290         } else
4291                 ret = -EINVAL;
4292
4293 out:
4294         return ret;
4295 }
4296 EXPORT_SYMBOL(rdma_disconnect);
4297
4298 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4299 {
4300         struct rdma_id_private *id_priv;
4301         struct cma_multicast *mc = multicast->context;
4302         struct rdma_cm_event event = {};
4303         int ret = 0;
4304
4305         id_priv = mc->id_priv;
4306         mutex_lock(&id_priv->handler_mutex);
4307         if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4308             id_priv->state != RDMA_CM_ADDR_RESOLVED)
4309                 goto out;
4310
4311         if (!status)
4312                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4313         else
4314                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4315                                      status);
4316         mutex_lock(&id_priv->qp_mutex);
4317         if (!status && id_priv->id.qp) {
4318                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
4319                                          be16_to_cpu(multicast->rec.mlid));
4320                 if (status)
4321                         pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
4322                                              status);
4323         }
4324         mutex_unlock(&id_priv->qp_mutex);
4325
4326         event.status = status;
4327         event.param.ud.private_data = mc->context;
4328         if (!status) {
4329                 struct rdma_dev_addr *dev_addr =
4330                         &id_priv->id.route.addr.dev_addr;
4331                 struct net_device *ndev =
4332                         dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4333                 enum ib_gid_type gid_type =
4334                         id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4335                         rdma_start_port(id_priv->cma_dev->device)];
4336
4337                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4338                 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4339                                                id_priv->id.port_num,
4340                                                &multicast->rec,
4341                                                ndev, gid_type,
4342                                                &event.param.ud.ah_attr);
4343                 if (ret)
4344                         event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4345
4346                 event.param.ud.qp_num = 0xFFFFFF;
4347                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4348                 if (ndev)
4349                         dev_put(ndev);
4350         } else
4351                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4352
4353         ret = cma_cm_event_handler(id_priv, &event);
4354
4355         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4356         if (ret) {
4357                 destroy_id_handler_unlock(id_priv);
4358                 return 0;
4359         }
4360
4361 out:
4362         mutex_unlock(&id_priv->handler_mutex);
4363         return 0;
4364 }
4365
4366 static void cma_set_mgid(struct rdma_id_private *id_priv,
4367                          struct sockaddr *addr, union ib_gid *mgid)
4368 {
4369         unsigned char mc_map[MAX_ADDR_LEN];
4370         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4371         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4372         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4373
4374         if (cma_any_addr(addr)) {
4375                 memset(mgid, 0, sizeof *mgid);
4376         } else if ((addr->sa_family == AF_INET6) &&
4377                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4378                                                                  0xFF10A01B)) {
4379                 /* IPv6 address is an SA assigned MGID. */
4380                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4381         } else if (addr->sa_family == AF_IB) {
4382                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4383         } else if (addr->sa_family == AF_INET6) {
4384                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4385                 if (id_priv->id.ps == RDMA_PS_UDP)
4386                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4387                 *mgid = *(union ib_gid *) (mc_map + 4);
4388         } else {
4389                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4390                 if (id_priv->id.ps == RDMA_PS_UDP)
4391                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4392                 *mgid = *(union ib_gid *) (mc_map + 4);
4393         }
4394 }
4395
4396 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4397                                  struct cma_multicast *mc)
4398 {
4399         struct ib_sa_mcmember_rec rec;
4400         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4401         ib_sa_comp_mask comp_mask;
4402         int ret;
4403
4404         ib_addr_get_mgid(dev_addr, &rec.mgid);
4405         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4406                                      &rec.mgid, &rec);
4407         if (ret)
4408                 return ret;
4409
4410         ret = cma_set_qkey(id_priv, 0);
4411         if (ret)
4412                 return ret;
4413
4414         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4415         rec.qkey = cpu_to_be32(id_priv->qkey);
4416         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4417         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4418         rec.join_state = mc->join_state;
4419
4420         if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4421             (!ib_sa_sendonly_fullmem_support(&sa_client,
4422                                              id_priv->id.device,
4423                                              id_priv->id.port_num))) {
4424                 dev_warn(
4425                         &id_priv->id.device->dev,
4426                         "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4427                         id_priv->id.port_num);
4428                 return -EOPNOTSUPP;
4429         }
4430
4431         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4432                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4433                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4434                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4435                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4436
4437         if (id_priv->id.ps == RDMA_PS_IPOIB)
4438                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4439                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4440                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4441                              IB_SA_MCMEMBER_REC_MTU |
4442                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4443
4444         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4445                                                 id_priv->id.port_num, &rec,
4446                                                 comp_mask, GFP_KERNEL,
4447                                                 cma_ib_mc_handler, mc);
4448         return PTR_ERR_OR_ZERO(mc->multicast.ib);
4449 }
4450
4451 static void iboe_mcast_work_handler(struct work_struct *work)
4452 {
4453         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4454         struct cma_multicast *mc = mw->mc;
4455         struct ib_sa_multicast *m = mc->multicast.ib;
4456
4457         mc->multicast.ib->context = mc;
4458         cma_ib_mc_handler(0, m);
4459         kref_put(&mc->mcref, release_mc);
4460         kfree(mw);
4461 }
4462
4463 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4464                               enum ib_gid_type gid_type)
4465 {
4466         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4467         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4468
4469         if (cma_any_addr(addr)) {
4470                 memset(mgid, 0, sizeof *mgid);
4471         } else if (addr->sa_family == AF_INET6) {
4472                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4473         } else {
4474                 mgid->raw[0] =
4475                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4476                 mgid->raw[1] =
4477                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4478                 mgid->raw[2] = 0;
4479                 mgid->raw[3] = 0;
4480                 mgid->raw[4] = 0;
4481                 mgid->raw[5] = 0;
4482                 mgid->raw[6] = 0;
4483                 mgid->raw[7] = 0;
4484                 mgid->raw[8] = 0;
4485                 mgid->raw[9] = 0;
4486                 mgid->raw[10] = 0xff;
4487                 mgid->raw[11] = 0xff;
4488                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4489         }
4490 }
4491
4492 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4493                                    struct cma_multicast *mc)
4494 {
4495         struct iboe_mcast_work *work;
4496         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4497         int err = 0;
4498         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4499         struct net_device *ndev = NULL;
4500         enum ib_gid_type gid_type;
4501         bool send_only;
4502
4503         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4504
4505         if (cma_zero_addr((struct sockaddr *)&mc->addr))
4506                 return -EINVAL;
4507
4508         work = kzalloc(sizeof *work, GFP_KERNEL);
4509         if (!work)
4510                 return -ENOMEM;
4511
4512         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4513         if (!mc->multicast.ib) {
4514                 err = -ENOMEM;
4515                 goto out1;
4516         }
4517
4518         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4519                    rdma_start_port(id_priv->cma_dev->device)];
4520         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4521
4522         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4523         if (id_priv->id.ps == RDMA_PS_UDP)
4524                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4525
4526         if (dev_addr->bound_dev_if)
4527                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4528         if (!ndev) {
4529                 err = -ENODEV;
4530                 goto out2;
4531         }
4532         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4533         mc->multicast.ib->rec.hop_limit = 1;
4534         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4535
4536         if (addr->sa_family == AF_INET) {
4537                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4538                         mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4539                         if (!send_only) {
4540                                 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4541                                                     true);
4542                         }
4543                 }
4544         } else {
4545                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4546                         err = -ENOTSUPP;
4547         }
4548         dev_put(ndev);
4549         if (err || !mc->multicast.ib->rec.mtu) {
4550                 if (!err)
4551                         err = -EINVAL;
4552                 goto out2;
4553         }
4554         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4555                     &mc->multicast.ib->rec.port_gid);
4556         work->id = id_priv;
4557         work->mc = mc;
4558         INIT_WORK(&work->work, iboe_mcast_work_handler);
4559         kref_get(&mc->mcref);
4560         queue_work(cma_wq, &work->work);
4561
4562         return 0;
4563
4564 out2:
4565         kfree(mc->multicast.ib);
4566 out1:
4567         kfree(work);
4568         return err;
4569 }
4570
4571 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4572                         u8 join_state, void *context)
4573 {
4574         struct rdma_id_private *id_priv;
4575         struct cma_multicast *mc;
4576         int ret;
4577
4578         if (!id->device)
4579                 return -EINVAL;
4580
4581         id_priv = container_of(id, struct rdma_id_private, id);
4582         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4583             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4584                 return -EINVAL;
4585
4586         mc = kmalloc(sizeof *mc, GFP_KERNEL);
4587         if (!mc)
4588                 return -ENOMEM;
4589
4590         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4591         mc->context = context;
4592         mc->id_priv = id_priv;
4593         mc->join_state = join_state;
4594
4595         if (rdma_protocol_roce(id->device, id->port_num)) {
4596                 kref_init(&mc->mcref);
4597                 ret = cma_iboe_join_multicast(id_priv, mc);
4598                 if (ret)
4599                         goto out_err;
4600         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4601                 ret = cma_join_ib_multicast(id_priv, mc);
4602                 if (ret)
4603                         goto out_err;
4604         } else {
4605                 ret = -ENOSYS;
4606                 goto out_err;
4607         }
4608
4609         spin_lock(&id_priv->lock);
4610         list_add(&mc->list, &id_priv->mc_list);
4611         spin_unlock(&id_priv->lock);
4612
4613         return 0;
4614 out_err:
4615         kfree(mc);
4616         return ret;
4617 }
4618 EXPORT_SYMBOL(rdma_join_multicast);
4619
4620 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4621 {
4622         struct rdma_id_private *id_priv;
4623         struct cma_multicast *mc;
4624
4625         id_priv = container_of(id, struct rdma_id_private, id);
4626         spin_lock_irq(&id_priv->lock);
4627         list_for_each_entry(mc, &id_priv->mc_list, list) {
4628                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4629                         list_del(&mc->list);
4630                         spin_unlock_irq(&id_priv->lock);
4631
4632                         if (id->qp)
4633                                 ib_detach_mcast(id->qp,
4634                                                 &mc->multicast.ib->rec.mgid,
4635                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
4636
4637                         BUG_ON(id_priv->cma_dev->device != id->device);
4638
4639                         if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4640                                 ib_sa_free_multicast(mc->multicast.ib);
4641                                 kfree(mc);
4642                         } else if (rdma_protocol_roce(id->device, id->port_num)) {
4643                                 cma_leave_roce_mc_group(id_priv, mc);
4644                         }
4645                         return;
4646                 }
4647         }
4648         spin_unlock_irq(&id_priv->lock);
4649 }
4650 EXPORT_SYMBOL(rdma_leave_multicast);
4651
4652 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4653 {
4654         struct rdma_dev_addr *dev_addr;
4655         struct cma_ndev_work *work;
4656
4657         dev_addr = &id_priv->id.route.addr.dev_addr;
4658
4659         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4660             (net_eq(dev_net(ndev), dev_addr->net)) &&
4661             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4662                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4663                         ndev->name, &id_priv->id);
4664                 work = kzalloc(sizeof *work, GFP_KERNEL);
4665                 if (!work)
4666                         return -ENOMEM;
4667
4668                 INIT_WORK(&work->work, cma_ndev_work_handler);
4669                 work->id = id_priv;
4670                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4671                 cma_id_get(id_priv);
4672                 queue_work(cma_wq, &work->work);
4673         }
4674
4675         return 0;
4676 }
4677
4678 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4679                                void *ptr)
4680 {
4681         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4682         struct cma_device *cma_dev;
4683         struct rdma_id_private *id_priv;
4684         int ret = NOTIFY_DONE;
4685
4686         if (event != NETDEV_BONDING_FAILOVER)
4687                 return NOTIFY_DONE;
4688
4689         if (!netif_is_bond_master(ndev))
4690                 return NOTIFY_DONE;
4691
4692         mutex_lock(&lock);
4693         list_for_each_entry(cma_dev, &dev_list, list)
4694                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4695                         ret = cma_netdev_change(ndev, id_priv);
4696                         if (ret)
4697                                 goto out;
4698                 }
4699
4700 out:
4701         mutex_unlock(&lock);
4702         return ret;
4703 }
4704
4705 static struct notifier_block cma_nb = {
4706         .notifier_call = cma_netdev_callback
4707 };
4708
4709 static int cma_add_one(struct ib_device *device)
4710 {
4711         struct cma_device *cma_dev;
4712         struct rdma_id_private *id_priv;
4713         unsigned int i;
4714         unsigned long supported_gids = 0;
4715         int ret;
4716
4717         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4718         if (!cma_dev)
4719                 return -ENOMEM;
4720
4721         cma_dev->device = device;
4722         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4723                                             sizeof(*cma_dev->default_gid_type),
4724                                             GFP_KERNEL);
4725         if (!cma_dev->default_gid_type) {
4726                 ret = -ENOMEM;
4727                 goto free_cma_dev;
4728         }
4729
4730         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4731                                             sizeof(*cma_dev->default_roce_tos),
4732                                             GFP_KERNEL);
4733         if (!cma_dev->default_roce_tos) {
4734                 ret = -ENOMEM;
4735                 goto free_gid_type;
4736         }
4737
4738         rdma_for_each_port (device, i) {
4739                 supported_gids = roce_gid_type_mask_support(device, i);
4740                 WARN_ON(!supported_gids);
4741                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4742                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4743                                 CMA_PREFERRED_ROCE_GID_TYPE;
4744                 else
4745                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4746                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4747                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4748         }
4749
4750         init_completion(&cma_dev->comp);
4751         refcount_set(&cma_dev->refcount, 1);
4752         INIT_LIST_HEAD(&cma_dev->id_list);
4753         ib_set_client_data(device, &cma_client, cma_dev);
4754
4755         mutex_lock(&lock);
4756         list_add_tail(&cma_dev->list, &dev_list);
4757         list_for_each_entry(id_priv, &listen_any_list, list)
4758                 cma_listen_on_dev(id_priv, cma_dev);
4759         mutex_unlock(&lock);
4760
4761         trace_cm_add_one(device);
4762         return 0;
4763
4764 free_gid_type:
4765         kfree(cma_dev->default_gid_type);
4766
4767 free_cma_dev:
4768         kfree(cma_dev);
4769         return ret;
4770 }
4771
4772 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4773 {
4774         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4775         enum rdma_cm_state state;
4776         unsigned long flags;
4777
4778         mutex_lock(&id_priv->handler_mutex);
4779         /* Record that we want to remove the device */
4780         spin_lock_irqsave(&id_priv->lock, flags);
4781         state = id_priv->state;
4782         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4783                 spin_unlock_irqrestore(&id_priv->lock, flags);
4784                 mutex_unlock(&id_priv->handler_mutex);
4785                 cma_id_put(id_priv);
4786                 return;
4787         }
4788         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4789         spin_unlock_irqrestore(&id_priv->lock, flags);
4790
4791         if (cma_cm_event_handler(id_priv, &event)) {
4792                 /*
4793                  * At this point the ULP promises it won't call
4794                  * rdma_destroy_id() concurrently
4795                  */
4796                 cma_id_put(id_priv);
4797                 mutex_unlock(&id_priv->handler_mutex);
4798                 trace_cm_id_destroy(id_priv);
4799                 _destroy_id(id_priv, state);
4800                 return;
4801         }
4802         mutex_unlock(&id_priv->handler_mutex);
4803
4804         /*
4805          * If this races with destroy then the thread that first assigns state
4806          * to a destroying does the cancel.
4807          */
4808         cma_cancel_operation(id_priv, state);
4809         cma_id_put(id_priv);
4810 }
4811
4812 static void cma_process_remove(struct cma_device *cma_dev)
4813 {
4814         mutex_lock(&lock);
4815         while (!list_empty(&cma_dev->id_list)) {
4816                 struct rdma_id_private *id_priv = list_first_entry(
4817                         &cma_dev->id_list, struct rdma_id_private, list);
4818
4819                 list_del(&id_priv->listen_list);
4820                 list_del_init(&id_priv->list);
4821                 cma_id_get(id_priv);
4822                 mutex_unlock(&lock);
4823
4824                 cma_send_device_removal_put(id_priv);
4825
4826                 mutex_lock(&lock);
4827         }
4828         mutex_unlock(&lock);
4829
4830         cma_dev_put(cma_dev);
4831         wait_for_completion(&cma_dev->comp);
4832 }
4833
4834 static void cma_remove_one(struct ib_device *device, void *client_data)
4835 {
4836         struct cma_device *cma_dev = client_data;
4837
4838         trace_cm_remove_one(device);
4839
4840         mutex_lock(&lock);
4841         list_del(&cma_dev->list);
4842         mutex_unlock(&lock);
4843
4844         cma_process_remove(cma_dev);
4845         kfree(cma_dev->default_roce_tos);
4846         kfree(cma_dev->default_gid_type);
4847         kfree(cma_dev);
4848 }
4849
4850 static int cma_init_net(struct net *net)
4851 {
4852         struct cma_pernet *pernet = cma_pernet(net);
4853
4854         xa_init(&pernet->tcp_ps);
4855         xa_init(&pernet->udp_ps);
4856         xa_init(&pernet->ipoib_ps);
4857         xa_init(&pernet->ib_ps);
4858
4859         return 0;
4860 }
4861
4862 static void cma_exit_net(struct net *net)
4863 {
4864         struct cma_pernet *pernet = cma_pernet(net);
4865
4866         WARN_ON(!xa_empty(&pernet->tcp_ps));
4867         WARN_ON(!xa_empty(&pernet->udp_ps));
4868         WARN_ON(!xa_empty(&pernet->ipoib_ps));
4869         WARN_ON(!xa_empty(&pernet->ib_ps));
4870 }
4871
4872 static struct pernet_operations cma_pernet_operations = {
4873         .init = cma_init_net,
4874         .exit = cma_exit_net,
4875         .id = &cma_pernet_id,
4876         .size = sizeof(struct cma_pernet),
4877 };
4878
4879 static int __init cma_init(void)
4880 {
4881         int ret;
4882
4883         /*
4884          * There is a rare lock ordering dependency in cma_netdev_callback()
4885          * that only happens when bonding is enabled. Teach lockdep that rtnl
4886          * must never be nested under lock so it can find these without having
4887          * to test with bonding.
4888          */
4889         if (IS_ENABLED(CONFIG_LOCKDEP)) {
4890                 rtnl_lock();
4891                 mutex_lock(&lock);
4892                 mutex_unlock(&lock);
4893                 rtnl_unlock();
4894         }
4895
4896         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4897         if (!cma_wq)
4898                 return -ENOMEM;
4899
4900         ret = register_pernet_subsys(&cma_pernet_operations);
4901         if (ret)
4902                 goto err_wq;
4903
4904         ib_sa_register_client(&sa_client);
4905         register_netdevice_notifier(&cma_nb);
4906
4907         ret = ib_register_client(&cma_client);
4908         if (ret)
4909                 goto err;
4910
4911         ret = cma_configfs_init();
4912         if (ret)
4913                 goto err_ib;
4914
4915         return 0;
4916
4917 err_ib:
4918         ib_unregister_client(&cma_client);
4919 err:
4920         unregister_netdevice_notifier(&cma_nb);
4921         ib_sa_unregister_client(&sa_client);
4922         unregister_pernet_subsys(&cma_pernet_operations);
4923 err_wq:
4924         destroy_workqueue(cma_wq);
4925         return ret;
4926 }
4927
4928 static void __exit cma_cleanup(void)
4929 {
4930         cma_configfs_exit();
4931         ib_unregister_client(&cma_client);
4932         unregister_netdevice_notifier(&cma_nb);
4933         ib_sa_unregister_client(&sa_client);
4934         unregister_pernet_subsys(&cma_pernet_operations);
4935         destroy_workqueue(cma_wq);
4936 }
4937
4938 module_init(cma_init);
4939 module_exit(cma_cleanup);