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/rbtree.h>
  15 #include <linux/igmp.h>
  16 #include <linux/xarray.h>
  17 #include <linux/inetdevice.h>
  18 #include <linux/slab.h>
  19 #include <linux/module.h>
  20 #include <net/route.h>
  21
  22 #include <net/net_namespace.h>
  23 #include <net/netns/generic.h>
  24 #include <net/netevent.h>
  25 #include <net/tcp.h>
  26 #include <net/ipv6.h>
  27 #include <net/ip_fib.h>
  28 #include <net/ip6_route.h>
  29
  30 #include <rdma/rdma_cm.h>
  31 #include <rdma/rdma_cm_ib.h>
  32 #include <rdma/rdma_netlink.h>
  33 #include <rdma/ib.h>
  34 #include <rdma/ib_cache.h>
  35 #include <rdma/ib_cm.h>
  36 #include <rdma/ib_sa.h>
  37 #include <rdma/iw_cm.h>
  38
  39 #include "core_priv.h"
  40 #include "cma_priv.h"
  41 #include "cma_trace.h"
  42
  43 MODULE_AUTHOR("Sean Hefty");
  44 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  45 MODULE_LICENSE("Dual BSD/GPL");
  46
  47 #define CMA_CM_RESPONSE_TIMEOUT 20
  48 #define CMA_MAX_CM_RETRIES 15
  49 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  50 #define CMA_IBOE_PACKET_LIFETIME 16
  51 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  52
  53 static const char * const cma_events[] = {
  54         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  55         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  56         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  57         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  58         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  59         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  60         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  61         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  62         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  63         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  64         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  65         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  66         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  67         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  68         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  69         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  70 };
  71
  72 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
  73                               enum ib_gid_type gid_type);
  74
  75 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  76 {
  77         size_t index = event;
  78
  79         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  80                         cma_events[index] : "unrecognized event";
  81 }
  82 EXPORT_SYMBOL(rdma_event_msg);
  83
  84 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  85                                                 int reason)
  86 {
  87         if (rdma_ib_or_roce(id->device, id->port_num))
  88                 return ibcm_reject_msg(reason);
  89
  90         if (rdma_protocol_iwarp(id->device, id->port_num))
  91                 return iwcm_reject_msg(reason);
  92
  93         WARN_ON_ONCE(1);
  94         return "unrecognized transport";
  95 }
  96 EXPORT_SYMBOL(rdma_reject_msg);
  97
  98 /**
  99  * rdma_is_consumer_reject - return true if the consumer rejected the connect
 100  *                           request.
 101  * @id: Communication identifier that received the REJECT event.
 102  * @reason: Value returned in the REJECT event status field.
 103  */
 104 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 105 {
 106         if (rdma_ib_or_roce(id->device, id->port_num))
 107                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 108
 109         if (rdma_protocol_iwarp(id->device, id->port_num))
 110                 return reason == -ECONNREFUSED;
 111
 112         WARN_ON_ONCE(1);
 113         return false;
 114 }
 115
 116 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 117                                       struct rdma_cm_event *ev, u8 *data_len)
 118 {
 119         const void *p;
 120
 121         if (rdma_is_consumer_reject(id, ev->status)) {
 122                 *data_len = ev->param.conn.private_data_len;
 123                 p = ev->param.conn.private_data;
 124         } else {
 125                 *data_len = 0;
 126                 p = NULL;
 127         }
 128         return p;
 129 }
 130 EXPORT_SYMBOL(rdma_consumer_reject_data);
 131
 132 /**
 133  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 134  * @id: Communication Identifier
 135  */
 136 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 137 {
 138         struct rdma_id_private *id_priv;
 139
 140         id_priv = container_of(id, struct rdma_id_private, id);
 141         if (id->device->node_type == RDMA_NODE_RNIC)
 142                 return id_priv->cm_id.iw;
 143         return NULL;
 144 }
 145 EXPORT_SYMBOL(rdma_iw_cm_id);
 146
 147 /**
 148  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 149  * @res: rdma resource tracking entry pointer
 150  */
 151 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 152 {
 153         struct rdma_id_private *id_priv =
 154                 container_of(res, struct rdma_id_private, res);
 155
 156         return &id_priv->id;
 157 }
 158 EXPORT_SYMBOL(rdma_res_to_id);
 159
 160 static int cma_add_one(struct ib_device *device);
 161 static void cma_remove_one(struct ib_device *device, void *client_data);
 162
 163 static struct ib_client cma_client = {
 164         .name   = "cma",
 165         .add    = cma_add_one,
 166         .remove = cma_remove_one
 167 };
 168
 169 static struct ib_sa_client sa_client;
 170 static LIST_HEAD(dev_list);
 171 static LIST_HEAD(listen_any_list);
 172 static DEFINE_MUTEX(lock);
 173 static struct rb_root id_table = RB_ROOT;
 174 /* Serialize operations of id_table tree */
 175 static DEFINE_SPINLOCK(id_table_lock);
 176 static struct workqueue_struct *cma_wq;
 177 static unsigned int cma_pernet_id;
 178
 179 struct cma_pernet {
 180         struct xarray tcp_ps;
 181         struct xarray udp_ps;
 182         struct xarray ipoib_ps;
 183         struct xarray ib_ps;
 184 };
 185
 186 static struct cma_pernet *cma_pernet(struct net *net)
 187 {
 188         return net_generic(net, cma_pernet_id);
 189 }
 190
 191 static
 192 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 193 {
 194         struct cma_pernet *pernet = cma_pernet(net);
 195
 196         switch (ps) {
 197         case RDMA_PS_TCP:
 198                 return &pernet->tcp_ps;
 199         case RDMA_PS_UDP:
 200                 return &pernet->udp_ps;
 201         case RDMA_PS_IPOIB:
 202                 return &pernet->ipoib_ps;
 203         case RDMA_PS_IB:
 204                 return &pernet->ib_ps;
 205         default:
 206                 return NULL;
 207         }
 208 }
 209
 210 struct id_table_entry {
 211         struct list_head id_list;
 212         struct rb_node rb_node;
 213 };
 214
 215 struct cma_device {
 216         struct list_head        list;
 217         struct ib_device        *device;
 218         struct completion       comp;
 219         refcount_t refcount;
 220         struct list_head        id_list;
 221         enum ib_gid_type        *default_gid_type;
 222         u8                      *default_roce_tos;
 223 };
 224
 225 struct rdma_bind_list {
 226         enum rdma_ucm_port_space ps;
 227         struct hlist_head       owners;
 228         unsigned short          port;
 229 };
 230
 231 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 232                         struct rdma_bind_list *bind_list, int snum)
 233 {
 234         struct xarray *xa = cma_pernet_xa(net, ps);
 235
 236         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 237 }
 238
 239 static struct rdma_bind_list *cma_ps_find(struct net *net,
 240                                           enum rdma_ucm_port_space ps, int snum)
 241 {
 242         struct xarray *xa = cma_pernet_xa(net, ps);
 243
 244         return xa_load(xa, snum);
 245 }
 246
 247 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 248                           int snum)
 249 {
 250         struct xarray *xa = cma_pernet_xa(net, ps);
 251
 252         xa_erase(xa, snum);
 253 }
 254
 255 enum {
 256         CMA_OPTION_AFONLY,
 257 };
 258
 259 void cma_dev_get(struct cma_device *cma_dev)
 260 {
 261         refcount_inc(&cma_dev->refcount);
 262 }
 263
 264 void cma_dev_put(struct cma_device *cma_dev)
 265 {
 266         if (refcount_dec_and_test(&cma_dev->refcount))
 267                 complete(&cma_dev->comp);
 268 }
 269
 270 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 271                                              void               *cookie)
 272 {
 273         struct cma_device *cma_dev;
 274         struct cma_device *found_cma_dev = NULL;
 275
 276         mutex_lock(&lock);
 277
 278         list_for_each_entry(cma_dev, &dev_list, list)
 279                 if (filter(cma_dev->device, cookie)) {
 280                         found_cma_dev = cma_dev;
 281                         break;
 282                 }
 283
 284         if (found_cma_dev)
 285                 cma_dev_get(found_cma_dev);
 286         mutex_unlock(&lock);
 287         return found_cma_dev;
 288 }
 289
 290 int cma_get_default_gid_type(struct cma_device *cma_dev,
 291                              u32 port)
 292 {
 293         if (!rdma_is_port_valid(cma_dev->device, port))
 294                 return -EINVAL;
 295
 296         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 297 }
 298
 299 int cma_set_default_gid_type(struct cma_device *cma_dev,
 300                              u32 port,
 301                              enum ib_gid_type default_gid_type)
 302 {
 303         unsigned long supported_gids;
 304
 305         if (!rdma_is_port_valid(cma_dev->device, port))
 306                 return -EINVAL;
 307
 308         if (default_gid_type == IB_GID_TYPE_IB &&
 309             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 310                 default_gid_type = IB_GID_TYPE_ROCE;
 311
 312         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 313
 314         if (!(supported_gids & 1 << default_gid_type))
 315                 return -EINVAL;
 316
 317         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 318                 default_gid_type;
 319
 320         return 0;
 321 }
 322
 323 int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
 324 {
 325         if (!rdma_is_port_valid(cma_dev->device, port))
 326                 return -EINVAL;
 327
 328         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 329 }
 330
 331 int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
 332                              u8 default_roce_tos)
 333 {
 334         if (!rdma_is_port_valid(cma_dev->device, port))
 335                 return -EINVAL;
 336
 337         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 338                  default_roce_tos;
 339
 340         return 0;
 341 }
 342 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 343 {
 344         return cma_dev->device;
 345 }
 346
 347 /*
 348  * Device removal can occur at anytime, so we need extra handling to
 349  * serialize notifying the user of device removal with other callbacks.
 350  * We do this by disabling removal notification while a callback is in process,
 351  * and reporting it after the callback completes.
 352  */
 353
 354 struct cma_multicast {
 355         struct rdma_id_private *id_priv;
 356         union {
 357                 struct ib_sa_multicast *sa_mc;
 358                 struct {
 359                         struct work_struct work;
 360                         struct rdma_cm_event event;
 361                 } iboe_join;
 362         };
 363         struct list_head        list;
 364         void                    *context;
 365         struct sockaddr_storage addr;
 366         u8                      join_state;
 367 };
 368
 369 struct cma_work {
 370         struct work_struct      work;
 371         struct rdma_id_private  *id;
 372         enum rdma_cm_state      old_state;
 373         enum rdma_cm_state      new_state;
 374         struct rdma_cm_event    event;
 375 };
 376
 377 union cma_ip_addr {
 378         struct in6_addr ip6;
 379         struct {
 380                 __be32 pad[3];
 381                 __be32 addr;
 382         } ip4;
 383 };
 384
 385 struct cma_hdr {
 386         u8 cma_version;
 387         u8 ip_version;  /* IP version: 7:4 */
 388         __be16 port;
 389         union cma_ip_addr src_addr;
 390         union cma_ip_addr dst_addr;
 391 };
 392
 393 #define CMA_VERSION 0x00
 394
 395 struct cma_req_info {
 396         struct sockaddr_storage listen_addr_storage;
 397         struct sockaddr_storage src_addr_storage;
 398         struct ib_device *device;
 399         union ib_gid local_gid;
 400         __be64 service_id;
 401         int port;
 402         bool has_gid;
 403         u16 pkey;
 404 };
 405
 406 static int cma_comp_exch(struct rdma_id_private *id_priv,
 407                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 408 {
 409         unsigned long flags;
 410         int ret;
 411
 412         /*
 413          * The FSM uses a funny double locking where state is protected by both
 414          * the handler_mutex and the spinlock. State is not allowed to change
 415          * to/from a handler_mutex protected value without also holding
 416          * handler_mutex.
 417          */
 418         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 419                 lockdep_assert_held(&id_priv->handler_mutex);
 420
 421         spin_lock_irqsave(&id_priv->lock, flags);
 422         if ((ret = (id_priv->state == comp)))
 423                 id_priv->state = exch;
 424         spin_unlock_irqrestore(&id_priv->lock, flags);
 425         return ret;
 426 }
 427
 428 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 429 {
 430         return hdr->ip_version >> 4;
 431 }
 432
 433 static void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 434 {
 435         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 436 }
 437
 438 static struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 439 {
 440         return (struct sockaddr *)&id_priv->id.route.addr.src_addr;
 441 }
 442
 443 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 444 {
 445         return (struct sockaddr *)&id_priv->id.route.addr.dst_addr;
 446 }
 447
 448 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 449 {
 450         struct in_device *in_dev = NULL;
 451
 452         if (ndev) {
 453                 rtnl_lock();
 454                 in_dev = __in_dev_get_rtnl(ndev);
 455                 if (in_dev) {
 456                         if (join)
 457                                 ip_mc_inc_group(in_dev,
 458                                                 *(__be32 *)(mgid->raw + 12));
 459                         else
 460                                 ip_mc_dec_group(in_dev,
 461                                                 *(__be32 *)(mgid->raw + 12));
 462                 }
 463                 rtnl_unlock();
 464         }
 465         return (in_dev) ? 0 : -ENODEV;
 466 }
 467
 468 static int compare_netdev_and_ip(int ifindex_a, struct sockaddr *sa,
 469                                  struct id_table_entry *entry_b)
 470 {
 471         struct rdma_id_private *id_priv = list_first_entry(
 472                 &entry_b->id_list, struct rdma_id_private, id_list_entry);
 473         int ifindex_b = id_priv->id.route.addr.dev_addr.bound_dev_if;
 474         struct sockaddr *sb = cma_dst_addr(id_priv);
 475
 476         if (ifindex_a != ifindex_b)
 477                 return (ifindex_a > ifindex_b) ? 1 : -1;
 478
 479         if (sa->sa_family != sb->sa_family)
 480                 return sa->sa_family - sb->sa_family;
 481
 482         if (sa->sa_family == AF_INET &&
 483             __builtin_object_size(sa, 0) >= sizeof(struct sockaddr_in)) {
 484                 return memcmp(&((struct sockaddr_in *)sa)->sin_addr,
 485                               &((struct sockaddr_in *)sb)->sin_addr,
 486                               sizeof(((struct sockaddr_in *)sa)->sin_addr));
 487         }
 488
 489         if (sa->sa_family == AF_INET6 &&
 490             __builtin_object_size(sa, 0) >= sizeof(struct sockaddr_in6)) {
 491                 return ipv6_addr_cmp(&((struct sockaddr_in6 *)sa)->sin6_addr,
 492                                      &((struct sockaddr_in6 *)sb)->sin6_addr);
 493         }
 494
 495         return -1;
 496 }
 497
 498 static int cma_add_id_to_tree(struct rdma_id_private *node_id_priv)
 499 {
 500         struct rb_node **new, *parent = NULL;
 501         struct id_table_entry *this, *node;
 502         unsigned long flags;
 503         int result;
 504
 505         node = kzalloc(sizeof(*node), GFP_KERNEL);
 506         if (!node)
 507                 return -ENOMEM;
 508
 509         spin_lock_irqsave(&id_table_lock, flags);
 510         new = &id_table.rb_node;
 511         while (*new) {
 512                 this = container_of(*new, struct id_table_entry, rb_node);
 513                 result = compare_netdev_and_ip(
 514                         node_id_priv->id.route.addr.dev_addr.bound_dev_if,
 515                         cma_dst_addr(node_id_priv), this);
 516
 517                 parent = *new;
 518                 if (result < 0)
 519                         new = &((*new)->rb_left);
 520                 else if (result > 0)
 521                         new = &((*new)->rb_right);
 522                 else {
 523                         list_add_tail(&node_id_priv->id_list_entry,
 524                                       &this->id_list);
 525                         kfree(node);
 526                         goto unlock;
 527                 }
 528         }
 529
 530         INIT_LIST_HEAD(&node->id_list);
 531         list_add_tail(&node_id_priv->id_list_entry, &node->id_list);
 532
 533         rb_link_node(&node->rb_node, parent, new);
 534         rb_insert_color(&node->rb_node, &id_table);
 535
 536 unlock:
 537         spin_unlock_irqrestore(&id_table_lock, flags);
 538         return 0;
 539 }
 540
 541 static struct id_table_entry *
 542 node_from_ndev_ip(struct rb_root *root, int ifindex, struct sockaddr *sa)
 543 {
 544         struct rb_node *node = root->rb_node;
 545         struct id_table_entry *data;
 546         int result;
 547
 548         while (node) {
 549                 data = container_of(node, struct id_table_entry, rb_node);
 550                 result = compare_netdev_and_ip(ifindex, sa, data);
 551                 if (result < 0)
 552                         node = node->rb_left;
 553                 else if (result > 0)
 554                         node = node->rb_right;
 555                 else
 556                         return data;
 557         }
 558
 559         return NULL;
 560 }
 561
 562 static void cma_remove_id_from_tree(struct rdma_id_private *id_priv)
 563 {
 564         struct id_table_entry *data;
 565         unsigned long flags;
 566
 567         spin_lock_irqsave(&id_table_lock, flags);
 568         if (list_empty(&id_priv->id_list_entry))
 569                 goto out;
 570
 571         data = node_from_ndev_ip(&id_table,
 572                                  id_priv->id.route.addr.dev_addr.bound_dev_if,
 573                                  cma_dst_addr(id_priv));
 574         if (!data)
 575                 goto out;
 576
 577         list_del_init(&id_priv->id_list_entry);
 578         if (list_empty(&data->id_list)) {
 579                 rb_erase(&data->rb_node, &id_table);
 580                 kfree(data);
 581         }
 582 out:
 583         spin_unlock_irqrestore(&id_table_lock, flags);
 584 }
 585
 586 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 587                                struct cma_device *cma_dev)
 588 {
 589         cma_dev_get(cma_dev);
 590         id_priv->cma_dev = cma_dev;
 591         id_priv->id.device = cma_dev->device;
 592         id_priv->id.route.addr.dev_addr.transport =
 593                 rdma_node_get_transport(cma_dev->device->node_type);
 594         list_add_tail(&id_priv->device_item, &cma_dev->id_list);
 595
 596         trace_cm_id_attach(id_priv, cma_dev->device);
 597 }
 598
 599 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 600                               struct cma_device *cma_dev)
 601 {
 602         _cma_attach_to_dev(id_priv, cma_dev);
 603         id_priv->gid_type =
 604                 cma_dev->default_gid_type[id_priv->id.port_num -
 605                                           rdma_start_port(cma_dev->device)];
 606 }
 607
 608 static void cma_release_dev(struct rdma_id_private *id_priv)
 609 {
 610         mutex_lock(&lock);
 611         list_del_init(&id_priv->device_item);
 612         cma_dev_put(id_priv->cma_dev);
 613         id_priv->cma_dev = NULL;
 614         id_priv->id.device = NULL;
 615         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 616                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 617                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 618         }
 619         mutex_unlock(&lock);
 620 }
 621
 622 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 623 {
 624         return id_priv->id.route.addr.src_addr.ss_family;
 625 }
 626
 627 static int cma_set_default_qkey(struct rdma_id_private *id_priv)
 628 {
 629         struct ib_sa_mcmember_rec rec;
 630         int ret = 0;
 631
 632         switch (id_priv->id.ps) {
 633         case RDMA_PS_UDP:
 634         case RDMA_PS_IB:
 635                 id_priv->qkey = RDMA_UDP_QKEY;
 636                 break;
 637         case RDMA_PS_IPOIB:
 638                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 639                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 640                                              id_priv->id.port_num, &rec.mgid,
 641                                              &rec);
 642                 if (!ret)
 643                         id_priv->qkey = be32_to_cpu(rec.qkey);
 644                 break;
 645         default:
 646                 break;
 647         }
 648         return ret;
 649 }
 650
 651 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 652 {
 653         if (!qkey ||
 654             (id_priv->qkey && (id_priv->qkey != qkey)))
 655                 return -EINVAL;
 656
 657         id_priv->qkey = qkey;
 658         return 0;
 659 }
 660
 661 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 662 {
 663         dev_addr->dev_type = ARPHRD_INFINIBAND;
 664         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 665         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 666 }
 667
 668 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 669 {
 670         int ret;
 671
 672         if (addr->sa_family != AF_IB) {
 673                 ret = rdma_translate_ip(addr, dev_addr);
 674         } else {
 675                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 676                 ret = 0;
 677         }
 678
 679         return ret;
 680 }
 681
 682 static const struct ib_gid_attr *
 683 cma_validate_port(struct ib_device *device, u32 port,
 684                   enum ib_gid_type gid_type,
 685                   union ib_gid *gid,
 686                   struct rdma_id_private *id_priv)
 687 {
 688         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 689         const struct ib_gid_attr *sgid_attr = ERR_PTR(-ENODEV);
 690         int bound_if_index = dev_addr->bound_dev_if;
 691         int dev_type = dev_addr->dev_type;
 692         struct net_device *ndev = NULL;
 693
 694         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 695                 goto out;
 696
 697         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 698                 goto out;
 699
 700         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 701                 goto out;
 702
 703         /*
 704          * For drivers that do not associate more than one net device with
 705          * their gid tables, such as iWARP drivers, it is sufficient to
 706          * return the first table entry.
 707          *
 708          * Other driver classes might be included in the future.
 709          */
 710         if (rdma_protocol_iwarp(device, port)) {
 711                 sgid_attr = rdma_get_gid_attr(device, port, 0);
 712                 if (IS_ERR(sgid_attr))
 713                         goto out;
 714
 715                 rcu_read_lock();
 716                 ndev = rcu_dereference(sgid_attr->ndev);
 717                 if (!net_eq(dev_net(ndev), dev_addr->net) ||
 718                     ndev->ifindex != bound_if_index) {
 719                         rdma_put_gid_attr(sgid_attr);
 720                         sgid_attr = ERR_PTR(-ENODEV);
 721                 }
 722                 rcu_read_unlock();
 723                 goto out;
 724         }
 725
 726         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 727                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 728                 if (!ndev)
 729                         goto out;
 730         } else {
 731                 gid_type = IB_GID_TYPE_IB;
 732         }
 733
 734         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 735         dev_put(ndev);
 736 out:
 737         return sgid_attr;
 738 }
 739
 740 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 741                                const struct ib_gid_attr *sgid_attr)
 742 {
 743         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 744         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 745 }
 746
 747 /**
 748  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 749  * based on source ip address.
 750  * @id_priv:    cm_id which should be bound to cma device
 751  *
 752  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 753  * based on source IP address. It returns 0 on success or error code otherwise.
 754  * It is applicable to active and passive side cm_id.
 755  */
 756 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 757 {
 758         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 759         const struct ib_gid_attr *sgid_attr;
 760         union ib_gid gid, iboe_gid, *gidp;
 761         struct cma_device *cma_dev;
 762         enum ib_gid_type gid_type;
 763         int ret = -ENODEV;
 764         u32 port;
 765
 766         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 767             id_priv->id.ps == RDMA_PS_IPOIB)
 768                 return -EINVAL;
 769
 770         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 771                     &iboe_gid);
 772
 773         memcpy(&gid, dev_addr->src_dev_addr +
 774                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 775
 776         mutex_lock(&lock);
 777         list_for_each_entry(cma_dev, &dev_list, list) {
 778                 rdma_for_each_port (cma_dev->device, port) {
 779                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 780                                &iboe_gid : &gid;
 781                         gid_type = cma_dev->default_gid_type[port - 1];
 782                         sgid_attr = cma_validate_port(cma_dev->device, port,
 783                                                       gid_type, gidp, id_priv);
 784                         if (!IS_ERR(sgid_attr)) {
 785                                 id_priv->id.port_num = port;
 786                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 787                                 cma_attach_to_dev(id_priv, cma_dev);
 788                                 ret = 0;
 789                                 goto out;
 790                         }
 791                 }
 792         }
 793 out:
 794         mutex_unlock(&lock);
 795         return ret;
 796 }
 797
 798 /**
 799  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 800  * @id_priv:            cm id to bind to cma device
 801  * @listen_id_priv:     listener cm id to match against
 802  * @req:                Pointer to req structure containaining incoming
 803  *                      request information
 804  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 805  * rdma device matches for listen_id and incoming request. It also verifies
 806  * that a GID table entry is present for the source address.
 807  * Returns 0 on success, or returns error code otherwise.
 808  */
 809 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 810                               const struct rdma_id_private *listen_id_priv,
 811                               struct cma_req_info *req)
 812 {
 813         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 814         const struct ib_gid_attr *sgid_attr;
 815         enum ib_gid_type gid_type;
 816         union ib_gid gid;
 817
 818         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 819             id_priv->id.ps == RDMA_PS_IPOIB)
 820                 return -EINVAL;
 821
 822         if (rdma_protocol_roce(req->device, req->port))
 823                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 824                             &gid);
 825         else
 826                 memcpy(&gid, dev_addr->src_dev_addr +
 827                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 828
 829         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 830         sgid_attr = cma_validate_port(req->device, req->port,
 831                                       gid_type, &gid, id_priv);
 832         if (IS_ERR(sgid_attr))
 833                 return PTR_ERR(sgid_attr);
 834
 835         id_priv->id.port_num = req->port;
 836         cma_bind_sgid_attr(id_priv, sgid_attr);
 837         /* Need to acquire lock to protect against reader
 838          * of cma_dev->id_list such as cma_netdev_callback() and
 839          * cma_process_remove().
 840          */
 841         mutex_lock(&lock);
 842         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 843         mutex_unlock(&lock);
 844         rdma_restrack_add(&id_priv->res);
 845         return 0;
 846 }
 847
 848 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 849                               const struct rdma_id_private *listen_id_priv)
 850 {
 851         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 852         const struct ib_gid_attr *sgid_attr;
 853         struct cma_device *cma_dev;
 854         enum ib_gid_type gid_type;
 855         int ret = -ENODEV;
 856         union ib_gid gid;
 857         u32 port;
 858
 859         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 860             id_priv->id.ps == RDMA_PS_IPOIB)
 861                 return -EINVAL;
 862
 863         memcpy(&gid, dev_addr->src_dev_addr +
 864                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 865
 866         mutex_lock(&lock);
 867
 868         cma_dev = listen_id_priv->cma_dev;
 869         port = listen_id_priv->id.port_num;
 870         gid_type = listen_id_priv->gid_type;
 871         sgid_attr = cma_validate_port(cma_dev->device, port,
 872                                       gid_type, &gid, id_priv);
 873         if (!IS_ERR(sgid_attr)) {
 874                 id_priv->id.port_num = port;
 875                 cma_bind_sgid_attr(id_priv, sgid_attr);
 876                 ret = 0;
 877                 goto out;
 878         }
 879
 880         list_for_each_entry(cma_dev, &dev_list, list) {
 881                 rdma_for_each_port (cma_dev->device, port) {
 882                         if (listen_id_priv->cma_dev == cma_dev &&
 883                             listen_id_priv->id.port_num == port)
 884                                 continue;
 885
 886                         gid_type = cma_dev->default_gid_type[port - 1];
 887                         sgid_attr = cma_validate_port(cma_dev->device, port,
 888                                                       gid_type, &gid, id_priv);
 889                         if (!IS_ERR(sgid_attr)) {
 890                                 id_priv->id.port_num = port;
 891                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 892                                 ret = 0;
 893                                 goto out;
 894                         }
 895                 }
 896         }
 897
 898 out:
 899         if (!ret) {
 900                 cma_attach_to_dev(id_priv, cma_dev);
 901                 rdma_restrack_add(&id_priv->res);
 902         }
 903
 904         mutex_unlock(&lock);
 905         return ret;
 906 }
 907
 908 /*
 909  * Select the source IB device and address to reach the destination IB address.
 910  */
 911 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 912 {
 913         struct cma_device *cma_dev, *cur_dev;
 914         struct sockaddr_ib *addr;
 915         union ib_gid gid, sgid, *dgid;
 916         unsigned int p;
 917         u16 pkey, index;
 918         enum ib_port_state port_state;
 919         int ret;
 920         int i;
 921
 922         cma_dev = NULL;
 923         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 924         dgid = (union ib_gid *) &addr->sib_addr;
 925         pkey = ntohs(addr->sib_pkey);
 926
 927         mutex_lock(&lock);
 928         list_for_each_entry(cur_dev, &dev_list, list) {
 929                 rdma_for_each_port (cur_dev->device, p) {
 930                         if (!rdma_cap_af_ib(cur_dev->device, p))
 931                                 continue;
 932
 933                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 934                                 continue;
 935
 936                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 937                                 continue;
 938
 939                         for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
 940                              ++i) {
 941                                 ret = rdma_query_gid(cur_dev->device, p, i,
 942                                                      &gid);
 943                                 if (ret)
 944                                         continue;
 945
 946                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 947                                         cma_dev = cur_dev;
 948                                         sgid = gid;
 949                                         id_priv->id.port_num = p;
 950                                         goto found;
 951                                 }
 952
 953                                 if (!cma_dev && (gid.global.subnet_prefix ==
 954                                     dgid->global.subnet_prefix) &&
 955                                     port_state == IB_PORT_ACTIVE) {
 956                                         cma_dev = cur_dev;
 957                                         sgid = gid;
 958                                         id_priv->id.port_num = p;
 959                                         goto found;
 960                                 }
 961                         }
 962                 }
 963         }
 964         mutex_unlock(&lock);
 965         return -ENODEV;
 966
 967 found:
 968         cma_attach_to_dev(id_priv, cma_dev);
 969         rdma_restrack_add(&id_priv->res);
 970         mutex_unlock(&lock);
 971         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 972         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 973         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 974         return 0;
 975 }
 976
 977 static void cma_id_get(struct rdma_id_private *id_priv)
 978 {
 979         refcount_inc(&id_priv->refcount);
 980 }
 981
 982 static void cma_id_put(struct rdma_id_private *id_priv)
 983 {
 984         if (refcount_dec_and_test(&id_priv->refcount))
 985                 complete(&id_priv->comp);
 986 }
 987
 988 static struct rdma_id_private *
 989 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 990                  void *context, enum rdma_ucm_port_space ps,
 991                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 992 {
 993         struct rdma_id_private *id_priv;
 994
 995         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 996         if (!id_priv)
 997                 return ERR_PTR(-ENOMEM);
 998
 999         id_priv->state = RDMA_CM_IDLE;
1000         id_priv->id.context = context;
1001         id_priv->id.event_handler = event_handler;
1002         id_priv->id.ps = ps;
1003         id_priv->id.qp_type = qp_type;
1004         id_priv->tos_set = false;
1005         id_priv->timeout_set = false;
1006         id_priv->min_rnr_timer_set = false;
1007         id_priv->gid_type = IB_GID_TYPE_IB;
1008         spin_lock_init(&id_priv->lock);
1009         mutex_init(&id_priv->qp_mutex);
1010         init_completion(&id_priv->comp);
1011         refcount_set(&id_priv->refcount, 1);
1012         mutex_init(&id_priv->handler_mutex);
1013         INIT_LIST_HEAD(&id_priv->device_item);
1014         INIT_LIST_HEAD(&id_priv->id_list_entry);
1015         INIT_LIST_HEAD(&id_priv->listen_list);
1016         INIT_LIST_HEAD(&id_priv->mc_list);
1017         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
1018         id_priv->id.route.addr.dev_addr.net = get_net(net);
1019         id_priv->seq_num &= 0x00ffffff;
1020
1021         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
1022         if (parent)
1023                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
1024
1025         return id_priv;
1026 }
1027
1028 struct rdma_cm_id *
1029 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
1030                         void *context, enum rdma_ucm_port_space ps,
1031                         enum ib_qp_type qp_type, const char *caller)
1032 {
1033         struct rdma_id_private *ret;
1034
1035         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
1036         if (IS_ERR(ret))
1037                 return ERR_CAST(ret);
1038
1039         rdma_restrack_set_name(&ret->res, caller);
1040         return &ret->id;
1041 }
1042 EXPORT_SYMBOL(__rdma_create_kernel_id);
1043
1044 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
1045                                        void *context,
1046                                        enum rdma_ucm_port_space ps,
1047                                        enum ib_qp_type qp_type)
1048 {
1049         struct rdma_id_private *ret;
1050
1051         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
1052                                ps, qp_type, NULL);
1053         if (IS_ERR(ret))
1054                 return ERR_CAST(ret);
1055
1056         rdma_restrack_set_name(&ret->res, NULL);
1057         return &ret->id;
1058 }
1059 EXPORT_SYMBOL(rdma_create_user_id);
1060
1061 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1062 {
1063         struct ib_qp_attr qp_attr;
1064         int qp_attr_mask, ret;
1065
1066         qp_attr.qp_state = IB_QPS_INIT;
1067         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1068         if (ret)
1069                 return ret;
1070
1071         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1072         if (ret)
1073                 return ret;
1074
1075         qp_attr.qp_state = IB_QPS_RTR;
1076         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
1077         if (ret)
1078                 return ret;
1079
1080         qp_attr.qp_state = IB_QPS_RTS;
1081         qp_attr.sq_psn = 0;
1082         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
1083
1084         return ret;
1085 }
1086
1087 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1088 {
1089         struct ib_qp_attr qp_attr;
1090         int qp_attr_mask, ret;
1091
1092         qp_attr.qp_state = IB_QPS_INIT;
1093         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1094         if (ret)
1095                 return ret;
1096
1097         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1098 }
1099
1100 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
1101                    struct ib_qp_init_attr *qp_init_attr)
1102 {
1103         struct rdma_id_private *id_priv;
1104         struct ib_qp *qp;
1105         int ret;
1106
1107         id_priv = container_of(id, struct rdma_id_private, id);
1108         if (id->device != pd->device) {
1109                 ret = -EINVAL;
1110                 goto out_err;
1111         }
1112
1113         qp_init_attr->port_num = id->port_num;
1114         qp = ib_create_qp(pd, qp_init_attr);
1115         if (IS_ERR(qp)) {
1116                 ret = PTR_ERR(qp);
1117                 goto out_err;
1118         }
1119
1120         if (id->qp_type == IB_QPT_UD)
1121                 ret = cma_init_ud_qp(id_priv, qp);
1122         else
1123                 ret = cma_init_conn_qp(id_priv, qp);
1124         if (ret)
1125                 goto out_destroy;
1126
1127         id->qp = qp;
1128         id_priv->qp_num = qp->qp_num;
1129         id_priv->srq = (qp->srq != NULL);
1130         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
1131         return 0;
1132 out_destroy:
1133         ib_destroy_qp(qp);
1134 out_err:
1135         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
1136         return ret;
1137 }
1138 EXPORT_SYMBOL(rdma_create_qp);
1139
1140 void rdma_destroy_qp(struct rdma_cm_id *id)
1141 {
1142         struct rdma_id_private *id_priv;
1143
1144         id_priv = container_of(id, struct rdma_id_private, id);
1145         trace_cm_qp_destroy(id_priv);
1146         mutex_lock(&id_priv->qp_mutex);
1147         ib_destroy_qp(id_priv->id.qp);
1148         id_priv->id.qp = NULL;
1149         mutex_unlock(&id_priv->qp_mutex);
1150 }
1151 EXPORT_SYMBOL(rdma_destroy_qp);
1152
1153 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1154                              struct rdma_conn_param *conn_param)
1155 {
1156         struct ib_qp_attr qp_attr;
1157         int qp_attr_mask, ret;
1158
1159         mutex_lock(&id_priv->qp_mutex);
1160         if (!id_priv->id.qp) {
1161                 ret = 0;
1162                 goto out;
1163         }
1164
1165         /* Need to update QP attributes from default values. */
1166         qp_attr.qp_state = IB_QPS_INIT;
1167         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1168         if (ret)
1169                 goto out;
1170
1171         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1172         if (ret)
1173                 goto out;
1174
1175         qp_attr.qp_state = IB_QPS_RTR;
1176         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1177         if (ret)
1178                 goto out;
1179
1180         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1181
1182         if (conn_param)
1183                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1184         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1185 out:
1186         mutex_unlock(&id_priv->qp_mutex);
1187         return ret;
1188 }
1189
1190 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1191                              struct rdma_conn_param *conn_param)
1192 {
1193         struct ib_qp_attr qp_attr;
1194         int qp_attr_mask, ret;
1195
1196         mutex_lock(&id_priv->qp_mutex);
1197         if (!id_priv->id.qp) {
1198                 ret = 0;
1199                 goto out;
1200         }
1201
1202         qp_attr.qp_state = IB_QPS_RTS;
1203         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1204         if (ret)
1205                 goto out;
1206
1207         if (conn_param)
1208                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1209         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1210 out:
1211         mutex_unlock(&id_priv->qp_mutex);
1212         return ret;
1213 }
1214
1215 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1216 {
1217         struct ib_qp_attr qp_attr;
1218         int ret;
1219
1220         mutex_lock(&id_priv->qp_mutex);
1221         if (!id_priv->id.qp) {
1222                 ret = 0;
1223                 goto out;
1224         }
1225
1226         qp_attr.qp_state = IB_QPS_ERR;
1227         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1228 out:
1229         mutex_unlock(&id_priv->qp_mutex);
1230         return ret;
1231 }
1232
1233 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1234                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1235 {
1236         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1237         int ret;
1238         u16 pkey;
1239
1240         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1241                 pkey = 0xffff;
1242         else
1243                 pkey = ib_addr_get_pkey(dev_addr);
1244
1245         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1246                                   pkey, &qp_attr->pkey_index);
1247         if (ret)
1248                 return ret;
1249
1250         qp_attr->port_num = id_priv->id.port_num;
1251         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1252
1253         if (id_priv->id.qp_type == IB_QPT_UD) {
1254                 ret = cma_set_default_qkey(id_priv);
1255                 if (ret)
1256                         return ret;
1257
1258                 qp_attr->qkey = id_priv->qkey;
1259                 *qp_attr_mask |= IB_QP_QKEY;
1260         } else {
1261                 qp_attr->qp_access_flags = 0;
1262                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1263         }
1264         return 0;
1265 }
1266
1267 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1268                        int *qp_attr_mask)
1269 {
1270         struct rdma_id_private *id_priv;
1271         int ret = 0;
1272
1273         id_priv = container_of(id, struct rdma_id_private, id);
1274         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1275                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1276                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1277                 else
1278                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1279                                                  qp_attr_mask);
1280
1281                 if (qp_attr->qp_state == IB_QPS_RTR)
1282                         qp_attr->rq_psn = id_priv->seq_num;
1283         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1284                 if (!id_priv->cm_id.iw) {
1285                         qp_attr->qp_access_flags = 0;
1286                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1287                 } else
1288                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1289                                                  qp_attr_mask);
1290                 qp_attr->port_num = id_priv->id.port_num;
1291                 *qp_attr_mask |= IB_QP_PORT;
1292         } else {
1293                 ret = -ENOSYS;
1294         }
1295
1296         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1297                 qp_attr->timeout = id_priv->timeout;
1298
1299         if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1300                 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1301
1302         return ret;
1303 }
1304 EXPORT_SYMBOL(rdma_init_qp_attr);
1305
1306 static inline bool cma_zero_addr(const struct sockaddr *addr)
1307 {
1308         switch (addr->sa_family) {
1309         case AF_INET:
1310                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1311         case AF_INET6:
1312                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1313         case AF_IB:
1314                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1315         default:
1316                 return false;
1317         }
1318 }
1319
1320 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1321 {
1322         switch (addr->sa_family) {
1323         case AF_INET:
1324                 return ipv4_is_loopback(
1325                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1326         case AF_INET6:
1327                 return ipv6_addr_loopback(
1328                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1329         case AF_IB:
1330                 return ib_addr_loopback(
1331                         &((struct sockaddr_ib *)addr)->sib_addr);
1332         default:
1333                 return false;
1334         }
1335 }
1336
1337 static inline bool cma_any_addr(const struct sockaddr *addr)
1338 {
1339         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1340 }
1341
1342 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1343 {
1344         if (src->sa_family != dst->sa_family)
1345                 return -1;
1346
1347         switch (src->sa_family) {
1348         case AF_INET:
1349                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1350                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1351         case AF_INET6: {
1352                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1353                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1354                 bool link_local;
1355
1356                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1357                                           &dst_addr6->sin6_addr))
1358                         return 1;
1359                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1360                              IPV6_ADDR_LINKLOCAL;
1361                 /* Link local must match their scope_ids */
1362                 return link_local ? (src_addr6->sin6_scope_id !=
1363                                      dst_addr6->sin6_scope_id) :
1364                                     0;
1365         }
1366
1367         default:
1368                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1369                                    &((struct sockaddr_ib *) dst)->sib_addr);
1370         }
1371 }
1372
1373 static __be16 cma_port(const struct sockaddr *addr)
1374 {
1375         struct sockaddr_ib *sib;
1376
1377         switch (addr->sa_family) {
1378         case AF_INET:
1379                 return ((struct sockaddr_in *) addr)->sin_port;
1380         case AF_INET6:
1381                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1382         case AF_IB:
1383                 sib = (struct sockaddr_ib *) addr;
1384                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1385                                     be64_to_cpu(sib->sib_sid_mask)));
1386         default:
1387                 return 0;
1388         }
1389 }
1390
1391 static inline int cma_any_port(const struct sockaddr *addr)
1392 {
1393         return !cma_port(addr);
1394 }
1395
1396 static void cma_save_ib_info(struct sockaddr *src_addr,
1397                              struct sockaddr *dst_addr,
1398                              const struct rdma_cm_id *listen_id,
1399                              const struct sa_path_rec *path)
1400 {
1401         struct sockaddr_ib *listen_ib, *ib;
1402
1403         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1404         if (src_addr) {
1405                 ib = (struct sockaddr_ib *)src_addr;
1406                 ib->sib_family = AF_IB;
1407                 if (path) {
1408                         ib->sib_pkey = path->pkey;
1409                         ib->sib_flowinfo = path->flow_label;
1410                         memcpy(&ib->sib_addr, &path->sgid, 16);
1411                         ib->sib_sid = path->service_id;
1412                         ib->sib_scope_id = 0;
1413                 } else {
1414                         ib->sib_pkey = listen_ib->sib_pkey;
1415                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1416                         ib->sib_addr = listen_ib->sib_addr;
1417                         ib->sib_sid = listen_ib->sib_sid;
1418                         ib->sib_scope_id = listen_ib->sib_scope_id;
1419                 }
1420                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1421         }
1422         if (dst_addr) {
1423                 ib = (struct sockaddr_ib *)dst_addr;
1424                 ib->sib_family = AF_IB;
1425                 if (path) {
1426                         ib->sib_pkey = path->pkey;
1427                         ib->sib_flowinfo = path->flow_label;
1428                         memcpy(&ib->sib_addr, &path->dgid, 16);
1429                 }
1430         }
1431 }
1432
1433 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1434                               struct sockaddr_in *dst_addr,
1435                               struct cma_hdr *hdr,
1436                               __be16 local_port)
1437 {
1438         if (src_addr) {
1439                 *src_addr = (struct sockaddr_in) {
1440                         .sin_family = AF_INET,
1441                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1442                         .sin_port = local_port,
1443                 };
1444         }
1445
1446         if (dst_addr) {
1447                 *dst_addr = (struct sockaddr_in) {
1448                         .sin_family = AF_INET,
1449                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1450                         .sin_port = hdr->port,
1451                 };
1452         }
1453 }
1454
1455 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1456                               struct sockaddr_in6 *dst_addr,
1457                               struct cma_hdr *hdr,
1458                               __be16 local_port)
1459 {
1460         if (src_addr) {
1461                 *src_addr = (struct sockaddr_in6) {
1462                         .sin6_family = AF_INET6,
1463                         .sin6_addr = hdr->dst_addr.ip6,
1464                         .sin6_port = local_port,
1465                 };
1466         }
1467
1468         if (dst_addr) {
1469                 *dst_addr = (struct sockaddr_in6) {
1470                         .sin6_family = AF_INET6,
1471                         .sin6_addr = hdr->src_addr.ip6,
1472                         .sin6_port = hdr->port,
1473                 };
1474         }
1475 }
1476
1477 static u16 cma_port_from_service_id(__be64 service_id)
1478 {
1479         return (u16)be64_to_cpu(service_id);
1480 }
1481
1482 static int cma_save_ip_info(struct sockaddr *src_addr,
1483                             struct sockaddr *dst_addr,
1484                             const struct ib_cm_event *ib_event,
1485                             __be64 service_id)
1486 {
1487         struct cma_hdr *hdr;
1488         __be16 port;
1489
1490         hdr = ib_event->private_data;
1491         if (hdr->cma_version != CMA_VERSION)
1492                 return -EINVAL;
1493
1494         port = htons(cma_port_from_service_id(service_id));
1495
1496         switch (cma_get_ip_ver(hdr)) {
1497         case 4:
1498                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1499                                   (struct sockaddr_in *)dst_addr, hdr, port);
1500                 break;
1501         case 6:
1502                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1503                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1504                 break;
1505         default:
1506                 return -EAFNOSUPPORT;
1507         }
1508
1509         return 0;
1510 }
1511
1512 static int cma_save_net_info(struct sockaddr *src_addr,
1513                              struct sockaddr *dst_addr,
1514                              const struct rdma_cm_id *listen_id,
1515                              const struct ib_cm_event *ib_event,
1516                              sa_family_t sa_family, __be64 service_id)
1517 {
1518         if (sa_family == AF_IB) {
1519                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1520                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1521                                          ib_event->param.req_rcvd.primary_path);
1522                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1523                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1524                 return 0;
1525         }
1526
1527         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1528 }
1529
1530 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1531                              struct cma_req_info *req)
1532 {
1533         const struct ib_cm_req_event_param *req_param =
1534                 &ib_event->param.req_rcvd;
1535         const struct ib_cm_sidr_req_event_param *sidr_param =
1536                 &ib_event->param.sidr_req_rcvd;
1537
1538         switch (ib_event->event) {
1539         case IB_CM_REQ_RECEIVED:
1540                 req->device     = req_param->listen_id->device;
1541                 req->port       = req_param->port;
1542                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1543                        sizeof(req->local_gid));
1544                 req->has_gid    = true;
1545                 req->service_id = req_param->primary_path->service_id;
1546                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1547                 if (req->pkey != req_param->bth_pkey)
1548                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1549                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1550                                             req_param->bth_pkey, req->pkey);
1551                 break;
1552         case IB_CM_SIDR_REQ_RECEIVED:
1553                 req->device     = sidr_param->listen_id->device;
1554                 req->port       = sidr_param->port;
1555                 req->has_gid    = false;
1556                 req->service_id = sidr_param->service_id;
1557                 req->pkey       = sidr_param->pkey;
1558                 if (req->pkey != sidr_param->bth_pkey)
1559                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1560                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1561                                             sidr_param->bth_pkey, req->pkey);
1562                 break;
1563         default:
1564                 return -EINVAL;
1565         }
1566
1567         return 0;
1568 }
1569
1570 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1571                                   const struct sockaddr_in *dst_addr,
1572                                   const struct sockaddr_in *src_addr)
1573 {
1574         __be32 daddr = dst_addr->sin_addr.s_addr,
1575                saddr = src_addr->sin_addr.s_addr;
1576         struct fib_result res;
1577         struct flowi4 fl4;
1578         int err;
1579         bool ret;
1580
1581         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1582             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1583             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1584             ipv4_is_loopback(saddr))
1585                 return false;
1586
1587         memset(&fl4, 0, sizeof(fl4));
1588         fl4.flowi4_oif = net_dev->ifindex;
1589         fl4.daddr = daddr;
1590         fl4.saddr = saddr;
1591
1592         rcu_read_lock();
1593         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1594         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1595         rcu_read_unlock();
1596
1597         return ret;
1598 }
1599
1600 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1601                                   const struct sockaddr_in6 *dst_addr,
1602                                   const struct sockaddr_in6 *src_addr)
1603 {
1604 #if IS_ENABLED(CONFIG_IPV6)
1605         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1606                            IPV6_ADDR_LINKLOCAL;
1607         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1608                                          &src_addr->sin6_addr, net_dev->ifindex,
1609                                          NULL, strict);
1610         bool ret;
1611
1612         if (!rt)
1613                 return false;
1614
1615         ret = rt->rt6i_idev->dev == net_dev;
1616         ip6_rt_put(rt);
1617
1618         return ret;
1619 #else
1620         return false;
1621 #endif
1622 }
1623
1624 static bool validate_net_dev(struct net_device *net_dev,
1625                              const struct sockaddr *daddr,
1626                              const struct sockaddr *saddr)
1627 {
1628         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1629         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1630         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1631         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1632
1633         switch (daddr->sa_family) {
1634         case AF_INET:
1635                 return saddr->sa_family == AF_INET &&
1636                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1637
1638         case AF_INET6:
1639                 return saddr->sa_family == AF_INET6 &&
1640                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1641
1642         default:
1643                 return false;
1644         }
1645 }
1646
1647 static struct net_device *
1648 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1649 {
1650         const struct ib_gid_attr *sgid_attr = NULL;
1651         struct net_device *ndev;
1652
1653         if (ib_event->event == IB_CM_REQ_RECEIVED)
1654                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1655         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1656                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1657
1658         if (!sgid_attr)
1659                 return NULL;
1660
1661         rcu_read_lock();
1662         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1663         if (IS_ERR(ndev))
1664                 ndev = NULL;
1665         else
1666                 dev_hold(ndev);
1667         rcu_read_unlock();
1668         return ndev;
1669 }
1670
1671 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1672                                           struct cma_req_info *req)
1673 {
1674         struct sockaddr *listen_addr =
1675                         (struct sockaddr *)&req->listen_addr_storage;
1676         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1677         struct net_device *net_dev;
1678         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1679         int err;
1680
1681         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1682                                req->service_id);
1683         if (err)
1684                 return ERR_PTR(err);
1685
1686         if (rdma_protocol_roce(req->device, req->port))
1687                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1688         else
1689                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1690                                                    req->pkey,
1691                                                    gid, listen_addr);
1692         if (!net_dev)
1693                 return ERR_PTR(-ENODEV);
1694
1695         return net_dev;
1696 }
1697
1698 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1699 {
1700         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1701 }
1702
1703 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1704                                    const struct cma_hdr *hdr)
1705 {
1706         struct sockaddr *addr = cma_src_addr(id_priv);
1707         __be32 ip4_addr;
1708         struct in6_addr ip6_addr;
1709
1710         if (cma_any_addr(addr) && !id_priv->afonly)
1711                 return true;
1712
1713         switch (addr->sa_family) {
1714         case AF_INET:
1715                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1716                 if (cma_get_ip_ver(hdr) != 4)
1717                         return false;
1718                 if (!cma_any_addr(addr) &&
1719                     hdr->dst_addr.ip4.addr != ip4_addr)
1720                         return false;
1721                 break;
1722         case AF_INET6:
1723                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1724                 if (cma_get_ip_ver(hdr) != 6)
1725                         return false;
1726                 if (!cma_any_addr(addr) &&
1727                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1728                         return false;
1729                 break;
1730         case AF_IB:
1731                 return true;
1732         default:
1733                 return false;
1734         }
1735
1736         return true;
1737 }
1738
1739 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1740 {
1741         struct ib_device *device = id->device;
1742         const u32 port_num = id->port_num ?: rdma_start_port(device);
1743
1744         return rdma_protocol_roce(device, port_num);
1745 }
1746
1747 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1748 {
1749         const struct sockaddr *daddr =
1750                         (const struct sockaddr *)&req->listen_addr_storage;
1751         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1752
1753         /* Returns true if the req is for IPv6 link local */
1754         return (daddr->sa_family == AF_INET6 &&
1755                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1756 }
1757
1758 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1759                               const struct net_device *net_dev,
1760                               const struct cma_req_info *req)
1761 {
1762         const struct rdma_addr *addr = &id->route.addr;
1763
1764         if (!net_dev)
1765                 /* This request is an AF_IB request */
1766                 return (!id->port_num || id->port_num == req->port) &&
1767                        (addr->src_addr.ss_family == AF_IB);
1768
1769         /*
1770          * If the request is not for IPv6 link local, allow matching
1771          * request to any netdevice of the one or multiport rdma device.
1772          */
1773         if (!cma_is_req_ipv6_ll(req))
1774                 return true;
1775         /*
1776          * Net namespaces must match, and if the listner is listening
1777          * on a specific netdevice than netdevice must match as well.
1778          */
1779         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1780             (!!addr->dev_addr.bound_dev_if ==
1781              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1782                 return true;
1783         else
1784                 return false;
1785 }
1786
1787 static struct rdma_id_private *cma_find_listener(
1788                 const struct rdma_bind_list *bind_list,
1789                 const struct ib_cm_id *cm_id,
1790                 const struct ib_cm_event *ib_event,
1791                 const struct cma_req_info *req,
1792                 const struct net_device *net_dev)
1793 {
1794         struct rdma_id_private *id_priv, *id_priv_dev;
1795
1796         lockdep_assert_held(&lock);
1797
1798         if (!bind_list)
1799                 return ERR_PTR(-EINVAL);
1800
1801         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1802                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1803                         if (id_priv->id.device == cm_id->device &&
1804                             cma_match_net_dev(&id_priv->id, net_dev, req))
1805                                 return id_priv;
1806                         list_for_each_entry(id_priv_dev,
1807                                             &id_priv->listen_list,
1808                                             listen_item) {
1809                                 if (id_priv_dev->id.device == cm_id->device &&
1810                                     cma_match_net_dev(&id_priv_dev->id,
1811                                                       net_dev, req))
1812                                         return id_priv_dev;
1813                         }
1814                 }
1815         }
1816
1817         return ERR_PTR(-EINVAL);
1818 }
1819
1820 static struct rdma_id_private *
1821 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1822                      const struct ib_cm_event *ib_event,
1823                      struct cma_req_info *req,
1824                      struct net_device **net_dev)
1825 {
1826         struct rdma_bind_list *bind_list;
1827         struct rdma_id_private *id_priv;
1828         int err;
1829
1830         err = cma_save_req_info(ib_event, req);
1831         if (err)
1832                 return ERR_PTR(err);
1833
1834         *net_dev = cma_get_net_dev(ib_event, req);
1835         if (IS_ERR(*net_dev)) {
1836                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1837                         /* Assuming the protocol is AF_IB */
1838                         *net_dev = NULL;
1839                 } else {
1840                         return ERR_CAST(*net_dev);
1841                 }
1842         }
1843
1844         mutex_lock(&lock);
1845         /*
1846          * Net namespace might be getting deleted while route lookup,
1847          * cm_id lookup is in progress. Therefore, perform netdevice
1848          * validation, cm_id lookup under rcu lock.
1849          * RCU lock along with netdevice state check, synchronizes with
1850          * netdevice migrating to different net namespace and also avoids
1851          * case where net namespace doesn't get deleted while lookup is in
1852          * progress.
1853          * If the device state is not IFF_UP, its properties such as ifindex
1854          * and nd_net cannot be trusted to remain valid without rcu lock.
1855          * net/core/dev.c change_net_namespace() ensures to synchronize with
1856          * ongoing operations on net device after device is closed using
1857          * synchronize_net().
1858          */
1859         rcu_read_lock();
1860         if (*net_dev) {
1861                 /*
1862                  * If netdevice is down, it is likely that it is administratively
1863                  * down or it might be migrating to different namespace.
1864                  * In that case avoid further processing, as the net namespace
1865                  * or ifindex may change.
1866                  */
1867                 if (((*net_dev)->flags & IFF_UP) == 0) {
1868                         id_priv = ERR_PTR(-EHOSTUNREACH);
1869                         goto err;
1870                 }
1871
1872                 if (!validate_net_dev(*net_dev,
1873                                  (struct sockaddr *)&req->src_addr_storage,
1874                                  (struct sockaddr *)&req->listen_addr_storage)) {
1875                         id_priv = ERR_PTR(-EHOSTUNREACH);
1876                         goto err;
1877                 }
1878         }
1879
1880         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1881                                 rdma_ps_from_service_id(req->service_id),
1882                                 cma_port_from_service_id(req->service_id));
1883         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1884 err:
1885         rcu_read_unlock();
1886         mutex_unlock(&lock);
1887         if (IS_ERR(id_priv) && *net_dev) {
1888                 dev_put(*net_dev);
1889                 *net_dev = NULL;
1890         }
1891         return id_priv;
1892 }
1893
1894 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1895 {
1896         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1897 }
1898
1899 static void cma_cancel_route(struct rdma_id_private *id_priv)
1900 {
1901         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1902                 if (id_priv->query)
1903                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1904         }
1905 }
1906
1907 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1908 {
1909         struct rdma_id_private *dev_id_priv;
1910
1911         lockdep_assert_held(&lock);
1912
1913         /*
1914          * Remove from listen_any_list to prevent added devices from spawning
1915          * additional listen requests.
1916          */
1917         list_del_init(&id_priv->listen_any_item);
1918
1919         while (!list_empty(&id_priv->listen_list)) {
1920                 dev_id_priv =
1921                         list_first_entry(&id_priv->listen_list,
1922                                          struct rdma_id_private, listen_item);
1923                 /* sync with device removal to avoid duplicate destruction */
1924                 list_del_init(&dev_id_priv->device_item);
1925                 list_del_init(&dev_id_priv->listen_item);
1926                 mutex_unlock(&lock);
1927
1928                 rdma_destroy_id(&dev_id_priv->id);
1929                 mutex_lock(&lock);
1930         }
1931 }
1932
1933 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1934 {
1935         mutex_lock(&lock);
1936         _cma_cancel_listens(id_priv);
1937         mutex_unlock(&lock);
1938 }
1939
1940 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1941                                  enum rdma_cm_state state)
1942 {
1943         switch (state) {
1944         case RDMA_CM_ADDR_QUERY:
1945                 /*
1946                  * We can avoid doing the rdma_addr_cancel() based on state,
1947                  * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1948                  * Notice that the addr_handler work could still be exiting
1949                  * outside this state, however due to the interaction with the
1950                  * handler_mutex the work is guaranteed not to touch id_priv
1951                  * during exit.
1952                  */
1953                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1954                 break;
1955         case RDMA_CM_ROUTE_QUERY:
1956                 cma_cancel_route(id_priv);
1957                 break;
1958         case RDMA_CM_LISTEN:
1959                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1960                         cma_cancel_listens(id_priv);
1961                 break;
1962         default:
1963                 break;
1964         }
1965 }
1966
1967 static void cma_release_port(struct rdma_id_private *id_priv)
1968 {
1969         struct rdma_bind_list *bind_list = id_priv->bind_list;
1970         struct net *net = id_priv->id.route.addr.dev_addr.net;
1971
1972         if (!bind_list)
1973                 return;
1974
1975         mutex_lock(&lock);
1976         hlist_del(&id_priv->node);
1977         if (hlist_empty(&bind_list->owners)) {
1978                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1979                 kfree(bind_list);
1980         }
1981         mutex_unlock(&lock);
1982 }
1983
1984 static void destroy_mc(struct rdma_id_private *id_priv,
1985                        struct cma_multicast *mc)
1986 {
1987         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1988
1989         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1990                 ib_sa_free_multicast(mc->sa_mc);
1991
1992         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1993                 struct rdma_dev_addr *dev_addr =
1994                         &id_priv->id.route.addr.dev_addr;
1995                 struct net_device *ndev = NULL;
1996
1997                 if (dev_addr->bound_dev_if)
1998                         ndev = dev_get_by_index(dev_addr->net,
1999                                                 dev_addr->bound_dev_if);
2000                 if (ndev && !send_only) {
2001                         enum ib_gid_type gid_type;
2002                         union ib_gid mgid;
2003
2004                         gid_type = id_priv->cma_dev->default_gid_type
2005                                            [id_priv->id.port_num -
2006                                             rdma_start_port(
2007                                                     id_priv->cma_dev->device)];
2008                         cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
2009                                           gid_type);
2010                         cma_igmp_send(ndev, &mgid, false);
2011                 }
2012                 dev_put(ndev);
2013
2014                 cancel_work_sync(&mc->iboe_join.work);
2015         }
2016         kfree(mc);
2017 }
2018
2019 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
2020 {
2021         struct cma_multicast *mc;
2022
2023         while (!list_empty(&id_priv->mc_list)) {
2024                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
2025                                       list);
2026                 list_del(&mc->list);
2027                 destroy_mc(id_priv, mc);
2028         }
2029 }
2030
2031 static void _destroy_id(struct rdma_id_private *id_priv,
2032                         enum rdma_cm_state state)
2033 {
2034         cma_cancel_operation(id_priv, state);
2035
2036         rdma_restrack_del(&id_priv->res);
2037         cma_remove_id_from_tree(id_priv);
2038         if (id_priv->cma_dev) {
2039                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
2040                         if (id_priv->cm_id.ib)
2041                                 ib_destroy_cm_id(id_priv->cm_id.ib);
2042                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
2043                         if (id_priv->cm_id.iw)
2044                                 iw_destroy_cm_id(id_priv->cm_id.iw);
2045                 }
2046                 cma_leave_mc_groups(id_priv);
2047                 cma_release_dev(id_priv);
2048         }
2049
2050         cma_release_port(id_priv);
2051         cma_id_put(id_priv);
2052         wait_for_completion(&id_priv->comp);
2053
2054         if (id_priv->internal_id)
2055                 cma_id_put(id_priv->id.context);
2056
2057         kfree(id_priv->id.route.path_rec);
2058         kfree(id_priv->id.route.path_rec_inbound);
2059         kfree(id_priv->id.route.path_rec_outbound);
2060
2061         put_net(id_priv->id.route.addr.dev_addr.net);
2062         kfree(id_priv);
2063 }
2064
2065 /*
2066  * destroy an ID from within the handler_mutex. This ensures that no other
2067  * handlers can start running concurrently.
2068  */
2069 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
2070         __releases(&idprv->handler_mutex)
2071 {
2072         enum rdma_cm_state state;
2073         unsigned long flags;
2074
2075         trace_cm_id_destroy(id_priv);
2076
2077         /*
2078          * Setting the state to destroyed under the handler mutex provides a
2079          * fence against calling handler callbacks. If this is invoked due to
2080          * the failure of a handler callback then it guarentees that no future
2081          * handlers will be called.
2082          */
2083         lockdep_assert_held(&id_priv->handler_mutex);
2084         spin_lock_irqsave(&id_priv->lock, flags);
2085         state = id_priv->state;
2086         id_priv->state = RDMA_CM_DESTROYING;
2087         spin_unlock_irqrestore(&id_priv->lock, flags);
2088         mutex_unlock(&id_priv->handler_mutex);
2089         _destroy_id(id_priv, state);
2090 }
2091
2092 void rdma_destroy_id(struct rdma_cm_id *id)
2093 {
2094         struct rdma_id_private *id_priv =
2095                 container_of(id, struct rdma_id_private, id);
2096
2097         mutex_lock(&id_priv->handler_mutex);
2098         destroy_id_handler_unlock(id_priv);
2099 }
2100 EXPORT_SYMBOL(rdma_destroy_id);
2101
2102 static int cma_rep_recv(struct rdma_id_private *id_priv)
2103 {
2104         int ret;
2105
2106         ret = cma_modify_qp_rtr(id_priv, NULL);
2107         if (ret)
2108                 goto reject;
2109
2110         ret = cma_modify_qp_rts(id_priv, NULL);
2111         if (ret)
2112                 goto reject;
2113
2114         trace_cm_send_rtu(id_priv);
2115         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
2116         if (ret)
2117                 goto reject;
2118
2119         return 0;
2120 reject:
2121         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
2122         cma_modify_qp_err(id_priv);
2123         trace_cm_send_rej(id_priv);
2124         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
2125                        NULL, 0, NULL, 0);
2126         return ret;
2127 }
2128
2129 static void cma_set_rep_event_data(struct rdma_cm_event *event,
2130                                    const struct ib_cm_rep_event_param *rep_data,
2131                                    void *private_data)
2132 {
2133         event->param.conn.private_data = private_data;
2134         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
2135         event->param.conn.responder_resources = rep_data->responder_resources;
2136         event->param.conn.initiator_depth = rep_data->initiator_depth;
2137         event->param.conn.flow_control = rep_data->flow_control;
2138         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
2139         event->param.conn.srq = rep_data->srq;
2140         event->param.conn.qp_num = rep_data->remote_qpn;
2141
2142         event->ece.vendor_id = rep_data->ece.vendor_id;
2143         event->ece.attr_mod = rep_data->ece.attr_mod;
2144 }
2145
2146 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
2147                                 struct rdma_cm_event *event)
2148 {
2149         int ret;
2150
2151         lockdep_assert_held(&id_priv->handler_mutex);
2152
2153         trace_cm_event_handler(id_priv, event);
2154         ret = id_priv->id.event_handler(&id_priv->id, event);
2155         trace_cm_event_done(id_priv, event, ret);
2156         return ret;
2157 }
2158
2159 static int cma_ib_handler(struct ib_cm_id *cm_id,
2160                           const struct ib_cm_event *ib_event)
2161 {
2162         struct rdma_id_private *id_priv = cm_id->context;
2163         struct rdma_cm_event event = {};
2164         enum rdma_cm_state state;
2165         int ret;
2166
2167         mutex_lock(&id_priv->handler_mutex);
2168         state = READ_ONCE(id_priv->state);
2169         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2170              state != RDMA_CM_CONNECT) ||
2171             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2172              state != RDMA_CM_DISCONNECT))
2173                 goto out;
2174
2175         switch (ib_event->event) {
2176         case IB_CM_REQ_ERROR:
2177         case IB_CM_REP_ERROR:
2178                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2179                 event.status = -ETIMEDOUT;
2180                 break;
2181         case IB_CM_REP_RECEIVED:
2182                 if (state == RDMA_CM_CONNECT &&
2183                     (id_priv->id.qp_type != IB_QPT_UD)) {
2184                         trace_cm_send_mra(id_priv);
2185                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2186                 }
2187                 if (id_priv->id.qp) {
2188                         event.status = cma_rep_recv(id_priv);
2189                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2190                                                      RDMA_CM_EVENT_ESTABLISHED;
2191                 } else {
2192                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2193                 }
2194                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2195                                        ib_event->private_data);
2196                 break;
2197         case IB_CM_RTU_RECEIVED:
2198         case IB_CM_USER_ESTABLISHED:
2199                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2200                 break;
2201         case IB_CM_DREQ_ERROR:
2202                 event.status = -ETIMEDOUT;
2203                 fallthrough;
2204         case IB_CM_DREQ_RECEIVED:
2205         case IB_CM_DREP_RECEIVED:
2206                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2207                                    RDMA_CM_DISCONNECT))
2208                         goto out;
2209                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2210                 break;
2211         case IB_CM_TIMEWAIT_EXIT:
2212                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2213                 break;
2214         case IB_CM_MRA_RECEIVED:
2215                 /* ignore event */
2216                 goto out;
2217         case IB_CM_REJ_RECEIVED:
2218                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2219                                                                                 ib_event->param.rej_rcvd.reason));
2220                 cma_modify_qp_err(id_priv);
2221                 event.status = ib_event->param.rej_rcvd.reason;
2222                 event.event = RDMA_CM_EVENT_REJECTED;
2223                 event.param.conn.private_data = ib_event->private_data;
2224                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2225                 break;
2226         default:
2227                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2228                        ib_event->event);
2229                 goto out;
2230         }
2231
2232         ret = cma_cm_event_handler(id_priv, &event);
2233         if (ret) {
2234                 /* Destroy the CM ID by returning a non-zero value. */
2235                 id_priv->cm_id.ib = NULL;
2236                 destroy_id_handler_unlock(id_priv);
2237                 return ret;
2238         }
2239 out:
2240         mutex_unlock(&id_priv->handler_mutex);
2241         return 0;
2242 }
2243
2244 static struct rdma_id_private *
2245 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2246                    const struct ib_cm_event *ib_event,
2247                    struct net_device *net_dev)
2248 {
2249         struct rdma_id_private *listen_id_priv;
2250         struct rdma_id_private *id_priv;
2251         struct rdma_cm_id *id;
2252         struct rdma_route *rt;
2253         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2254         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2255         const __be64 service_id =
2256                 ib_event->param.req_rcvd.primary_path->service_id;
2257         int ret;
2258
2259         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2260         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2261                                    listen_id->event_handler, listen_id->context,
2262                                    listen_id->ps,
2263                                    ib_event->param.req_rcvd.qp_type,
2264                                    listen_id_priv);
2265         if (IS_ERR(id_priv))
2266                 return NULL;
2267
2268         id = &id_priv->id;
2269         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2270                               (struct sockaddr *)&id->route.addr.dst_addr,
2271                               listen_id, ib_event, ss_family, service_id))
2272                 goto err;
2273
2274         rt = &id->route;
2275         rt->num_pri_alt_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2276         rt->path_rec = kmalloc_array(rt->num_pri_alt_paths,
2277                                      sizeof(*rt->path_rec), GFP_KERNEL);
2278         if (!rt->path_rec)
2279                 goto err;
2280
2281         rt->path_rec[0] = *path;
2282         if (rt->num_pri_alt_paths == 2)
2283                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2284
2285         if (net_dev) {
2286                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2287         } else {
2288                 if (!cma_protocol_roce(listen_id) &&
2289                     cma_any_addr(cma_src_addr(id_priv))) {
2290                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2291                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2292                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2293                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2294                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2295                         if (ret)
2296                                 goto err;
2297                 }
2298         }
2299         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2300
2301         id_priv->state = RDMA_CM_CONNECT;
2302         return id_priv;
2303
2304 err:
2305         rdma_destroy_id(id);
2306         return NULL;
2307 }
2308
2309 static struct rdma_id_private *
2310 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2311                   const struct ib_cm_event *ib_event,
2312                   struct net_device *net_dev)
2313 {
2314         const struct rdma_id_private *listen_id_priv;
2315         struct rdma_id_private *id_priv;
2316         struct rdma_cm_id *id;
2317         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2318         struct net *net = listen_id->route.addr.dev_addr.net;
2319         int ret;
2320
2321         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2322         id_priv = __rdma_create_id(net, listen_id->event_handler,
2323                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2324                                    listen_id_priv);
2325         if (IS_ERR(id_priv))
2326                 return NULL;
2327
2328         id = &id_priv->id;
2329         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2330                               (struct sockaddr *)&id->route.addr.dst_addr,
2331                               listen_id, ib_event, ss_family,
2332                               ib_event->param.sidr_req_rcvd.service_id))
2333                 goto err;
2334
2335         if (net_dev) {
2336                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2337         } else {
2338                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2339                         ret = cma_translate_addr(cma_src_addr(id_priv),
2340                                                  &id->route.addr.dev_addr);
2341                         if (ret)
2342                                 goto err;
2343                 }
2344         }
2345
2346         id_priv->state = RDMA_CM_CONNECT;
2347         return id_priv;
2348 err:
2349         rdma_destroy_id(id);
2350         return NULL;
2351 }
2352
2353 static void cma_set_req_event_data(struct rdma_cm_event *event,
2354                                    const struct ib_cm_req_event_param *req_data,
2355                                    void *private_data, int offset)
2356 {
2357         event->param.conn.private_data = private_data + offset;
2358         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2359         event->param.conn.responder_resources = req_data->responder_resources;
2360         event->param.conn.initiator_depth = req_data->initiator_depth;
2361         event->param.conn.flow_control = req_data->flow_control;
2362         event->param.conn.retry_count = req_data->retry_count;
2363         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2364         event->param.conn.srq = req_data->srq;
2365         event->param.conn.qp_num = req_data->remote_qpn;
2366
2367         event->ece.vendor_id = req_data->ece.vendor_id;
2368         event->ece.attr_mod = req_data->ece.attr_mod;
2369 }
2370
2371 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2372                                     const struct ib_cm_event *ib_event)
2373 {
2374         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2375                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2376                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2377                  (id->qp_type == IB_QPT_UD)) ||
2378                 (!id->qp_type));
2379 }
2380
2381 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2382                               const struct ib_cm_event *ib_event)
2383 {
2384         struct rdma_id_private *listen_id, *conn_id = NULL;
2385         struct rdma_cm_event event = {};
2386         struct cma_req_info req = {};
2387         struct net_device *net_dev;
2388         u8 offset;
2389         int ret;
2390
2391         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2392         if (IS_ERR(listen_id))
2393                 return PTR_ERR(listen_id);
2394
2395         trace_cm_req_handler(listen_id, ib_event->event);
2396         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2397                 ret = -EINVAL;
2398                 goto net_dev_put;
2399         }
2400
2401         mutex_lock(&listen_id->handler_mutex);
2402         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2403                 ret = -ECONNABORTED;
2404                 goto err_unlock;
2405         }
2406
2407         offset = cma_user_data_offset(listen_id);
2408         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2409         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2410                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2411                 event.param.ud.private_data = ib_event->private_data + offset;
2412                 event.param.ud.private_data_len =
2413                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2414         } else {
2415                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2416                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2417                                        ib_event->private_data, offset);
2418         }
2419         if (!conn_id) {
2420                 ret = -ENOMEM;
2421                 goto err_unlock;
2422         }
2423
2424         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2425         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2426         if (ret) {
2427                 destroy_id_handler_unlock(conn_id);
2428                 goto err_unlock;
2429         }
2430
2431         conn_id->cm_id.ib = cm_id;
2432         cm_id->context = conn_id;
2433         cm_id->cm_handler = cma_ib_handler;
2434
2435         ret = cma_cm_event_handler(conn_id, &event);
2436         if (ret) {
2437                 /* Destroy the CM ID by returning a non-zero value. */
2438                 conn_id->cm_id.ib = NULL;
2439                 mutex_unlock(&listen_id->handler_mutex);
2440                 destroy_id_handler_unlock(conn_id);
2441                 goto net_dev_put;
2442         }
2443
2444         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2445             conn_id->id.qp_type != IB_QPT_UD) {
2446                 trace_cm_send_mra(cm_id->context);
2447                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2448         }
2449         mutex_unlock(&conn_id->handler_mutex);
2450
2451 err_unlock:
2452         mutex_unlock(&listen_id->handler_mutex);
2453
2454 net_dev_put:
2455         dev_put(net_dev);
2456
2457         return ret;
2458 }
2459
2460 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2461 {
2462         if (addr->sa_family == AF_IB)
2463                 return ((struct sockaddr_ib *) addr)->sib_sid;
2464
2465         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2466 }
2467 EXPORT_SYMBOL(rdma_get_service_id);
2468
2469 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2470                     union ib_gid *dgid)
2471 {
2472         struct rdma_addr *addr = &cm_id->route.addr;
2473
2474         if (!cm_id->device) {
2475                 if (sgid)
2476                         memset(sgid, 0, sizeof(*sgid));
2477                 if (dgid)
2478                         memset(dgid, 0, sizeof(*dgid));
2479                 return;
2480         }
2481
2482         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2483                 if (sgid)
2484                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2485                 if (dgid)
2486                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2487         } else {
2488                 if (sgid)
2489                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2490                 if (dgid)
2491                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2492         }
2493 }
2494 EXPORT_SYMBOL(rdma_read_gids);
2495
2496 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2497 {
2498         struct rdma_id_private *id_priv = iw_id->context;
2499         struct rdma_cm_event event = {};
2500         int ret = 0;
2501         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2502         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2503
2504         mutex_lock(&id_priv->handler_mutex);
2505         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2506                 goto out;
2507
2508         switch (iw_event->event) {
2509         case IW_CM_EVENT_CLOSE:
2510                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2511                 break;
2512         case IW_CM_EVENT_CONNECT_REPLY:
2513                 memcpy(cma_src_addr(id_priv), laddr,
2514                        rdma_addr_size(laddr));
2515                 memcpy(cma_dst_addr(id_priv), raddr,
2516                        rdma_addr_size(raddr));
2517                 switch (iw_event->status) {
2518                 case 0:
2519                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2520                         event.param.conn.initiator_depth = iw_event->ird;
2521                         event.param.conn.responder_resources = iw_event->ord;
2522                         break;
2523                 case -ECONNRESET:
2524                 case -ECONNREFUSED:
2525                         event.event = RDMA_CM_EVENT_REJECTED;
2526                         break;
2527                 case -ETIMEDOUT:
2528                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2529                         break;
2530                 default:
2531                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2532                         break;
2533                 }
2534                 break;
2535         case IW_CM_EVENT_ESTABLISHED:
2536                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2537                 event.param.conn.initiator_depth = iw_event->ird;
2538                 event.param.conn.responder_resources = iw_event->ord;
2539                 break;
2540         default:
2541                 goto out;
2542         }
2543
2544         event.status = iw_event->status;
2545         event.param.conn.private_data = iw_event->private_data;
2546         event.param.conn.private_data_len = iw_event->private_data_len;
2547         ret = cma_cm_event_handler(id_priv, &event);
2548         if (ret) {
2549                 /* Destroy the CM ID by returning a non-zero value. */
2550                 id_priv->cm_id.iw = NULL;
2551                 destroy_id_handler_unlock(id_priv);
2552                 return ret;
2553         }
2554
2555 out:
2556         mutex_unlock(&id_priv->handler_mutex);
2557         return ret;
2558 }
2559
2560 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2561                                struct iw_cm_event *iw_event)
2562 {
2563         struct rdma_id_private *listen_id, *conn_id;
2564         struct rdma_cm_event event = {};
2565         int ret = -ECONNABORTED;
2566         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2567         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2568
2569         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2570         event.param.conn.private_data = iw_event->private_data;
2571         event.param.conn.private_data_len = iw_event->private_data_len;
2572         event.param.conn.initiator_depth = iw_event->ird;
2573         event.param.conn.responder_resources = iw_event->ord;
2574
2575         listen_id = cm_id->context;
2576
2577         mutex_lock(&listen_id->handler_mutex);
2578         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2579                 goto out;
2580
2581         /* Create a new RDMA id for the new IW CM ID */
2582         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2583                                    listen_id->id.event_handler,
2584                                    listen_id->id.context, RDMA_PS_TCP,
2585                                    IB_QPT_RC, listen_id);
2586         if (IS_ERR(conn_id)) {
2587                 ret = -ENOMEM;
2588                 goto out;
2589         }
2590         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2591         conn_id->state = RDMA_CM_CONNECT;
2592
2593         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2594         if (ret) {
2595                 mutex_unlock(&listen_id->handler_mutex);
2596                 destroy_id_handler_unlock(conn_id);
2597                 return ret;
2598         }
2599
2600         ret = cma_iw_acquire_dev(conn_id, listen_id);
2601         if (ret) {
2602                 mutex_unlock(&listen_id->handler_mutex);
2603                 destroy_id_handler_unlock(conn_id);
2604                 return ret;
2605         }
2606
2607         conn_id->cm_id.iw = cm_id;
2608         cm_id->context = conn_id;
2609         cm_id->cm_handler = cma_iw_handler;
2610
2611         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2612         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2613
2614         ret = cma_cm_event_handler(conn_id, &event);
2615         if (ret) {
2616                 /* User wants to destroy the CM ID */
2617                 conn_id->cm_id.iw = NULL;
2618                 mutex_unlock(&listen_id->handler_mutex);
2619                 destroy_id_handler_unlock(conn_id);
2620                 return ret;
2621         }
2622
2623         mutex_unlock(&conn_id->handler_mutex);
2624
2625 out:
2626         mutex_unlock(&listen_id->handler_mutex);
2627         return ret;
2628 }
2629
2630 static int cma_ib_listen(struct rdma_id_private *id_priv)
2631 {
2632         struct sockaddr *addr;
2633         struct ib_cm_id *id;
2634         __be64 svc_id;
2635
2636         addr = cma_src_addr(id_priv);
2637         svc_id = rdma_get_service_id(&id_priv->id, addr);
2638         id = ib_cm_insert_listen(id_priv->id.device,
2639                                  cma_ib_req_handler, svc_id);
2640         if (IS_ERR(id))
2641                 return PTR_ERR(id);
2642         id_priv->cm_id.ib = id;
2643
2644         return 0;
2645 }
2646
2647 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2648 {
2649         int ret;
2650         struct iw_cm_id *id;
2651
2652         id = iw_create_cm_id(id_priv->id.device,
2653                              iw_conn_req_handler,
2654                              id_priv);
2655         if (IS_ERR(id))
2656                 return PTR_ERR(id);
2657
2658         mutex_lock(&id_priv->qp_mutex);
2659         id->tos = id_priv->tos;
2660         id->tos_set = id_priv->tos_set;
2661         mutex_unlock(&id_priv->qp_mutex);
2662         id->afonly = id_priv->afonly;
2663         id_priv->cm_id.iw = id;
2664
2665         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2666                rdma_addr_size(cma_src_addr(id_priv)));
2667
2668         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2669
2670         if (ret) {
2671                 iw_destroy_cm_id(id_priv->cm_id.iw);
2672                 id_priv->cm_id.iw = NULL;
2673         }
2674
2675         return ret;
2676 }
2677
2678 static int cma_listen_handler(struct rdma_cm_id *id,
2679                               struct rdma_cm_event *event)
2680 {
2681         struct rdma_id_private *id_priv = id->context;
2682
2683         /* Listening IDs are always destroyed on removal */
2684         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2685                 return -1;
2686
2687         id->context = id_priv->id.context;
2688         id->event_handler = id_priv->id.event_handler;
2689         trace_cm_event_handler(id_priv, event);
2690         return id_priv->id.event_handler(id, event);
2691 }
2692
2693 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2694                              struct cma_device *cma_dev,
2695                              struct rdma_id_private **to_destroy)
2696 {
2697         struct rdma_id_private *dev_id_priv;
2698         struct net *net = id_priv->id.route.addr.dev_addr.net;
2699         int ret;
2700
2701         lockdep_assert_held(&lock);
2702
2703         *to_destroy = NULL;
2704         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2705                 return 0;
2706
2707         dev_id_priv =
2708                 __rdma_create_id(net, cma_listen_handler, id_priv,
2709                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2710         if (IS_ERR(dev_id_priv))
2711                 return PTR_ERR(dev_id_priv);
2712
2713         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2714         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2715                rdma_addr_size(cma_src_addr(id_priv)));
2716
2717         _cma_attach_to_dev(dev_id_priv, cma_dev);
2718         rdma_restrack_add(&dev_id_priv->res);
2719         cma_id_get(id_priv);
2720         dev_id_priv->internal_id = 1;
2721         dev_id_priv->afonly = id_priv->afonly;
2722         mutex_lock(&id_priv->qp_mutex);
2723         dev_id_priv->tos_set = id_priv->tos_set;
2724         dev_id_priv->tos = id_priv->tos;
2725         mutex_unlock(&id_priv->qp_mutex);
2726
2727         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2728         if (ret)
2729                 goto err_listen;
2730         list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2731         return 0;
2732 err_listen:
2733         /* Caller must destroy this after releasing lock */
2734         *to_destroy = dev_id_priv;
2735         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2736         return ret;
2737 }
2738
2739 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2740 {
2741         struct rdma_id_private *to_destroy;
2742         struct cma_device *cma_dev;
2743         int ret;
2744
2745         mutex_lock(&lock);
2746         list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2747         list_for_each_entry(cma_dev, &dev_list, list) {
2748                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2749                 if (ret) {
2750                         /* Prevent racing with cma_process_remove() */
2751                         if (to_destroy)
2752                                 list_del_init(&to_destroy->device_item);
2753                         goto err_listen;
2754                 }
2755         }
2756         mutex_unlock(&lock);
2757         return 0;
2758
2759 err_listen:
2760         _cma_cancel_listens(id_priv);
2761         mutex_unlock(&lock);
2762         if (to_destroy)
2763                 rdma_destroy_id(&to_destroy->id);
2764         return ret;
2765 }
2766
2767 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2768 {
2769         struct rdma_id_private *id_priv;
2770
2771         id_priv = container_of(id, struct rdma_id_private, id);
2772         mutex_lock(&id_priv->qp_mutex);
2773         id_priv->tos = (u8) tos;
2774         id_priv->tos_set = true;
2775         mutex_unlock(&id_priv->qp_mutex);
2776 }
2777 EXPORT_SYMBOL(rdma_set_service_type);
2778
2779 /**
2780  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2781  *                          with a connection identifier.
2782  * @id: Communication identifier to associated with service type.
2783  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2784  *
2785  * This function should be called before rdma_connect() on active side,
2786  * and on passive side before rdma_accept(). It is applicable to primary
2787  * path only. The timeout will affect the local side of the QP, it is not
2788  * negotiated with remote side and zero disables the timer. In case it is
2789  * set before rdma_resolve_route, the value will also be used to determine
2790  * PacketLifeTime for RoCE.
2791  *
2792  * Return: 0 for success
2793  */
2794 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2795 {
2796         struct rdma_id_private *id_priv;
2797
2798         if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2799                 return -EINVAL;
2800
2801         id_priv = container_of(id, struct rdma_id_private, id);
2802         mutex_lock(&id_priv->qp_mutex);
2803         id_priv->timeout = timeout;
2804         id_priv->timeout_set = true;
2805         mutex_unlock(&id_priv->qp_mutex);
2806
2807         return 0;
2808 }
2809 EXPORT_SYMBOL(rdma_set_ack_timeout);
2810
2811 /**
2812  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2813  *                            QP associated with a connection identifier.
2814  * @id: Communication identifier to associated with service type.
2815  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2816  *                 Timer Field" in the IBTA specification.
2817  *
2818  * This function should be called before rdma_connect() on active
2819  * side, and on passive side before rdma_accept(). The timer value
2820  * will be associated with the local QP. When it receives a send it is
2821  * not read to handle, typically if the receive queue is empty, an RNR
2822  * Retry NAK is returned to the requester with the min_rnr_timer
2823  * encoded. The requester will then wait at least the time specified
2824  * in the NAK before retrying. The default is zero, which translates
2825  * to a minimum RNR Timer value of 655 ms.
2826  *
2827  * Return: 0 for success
2828  */
2829 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2830 {
2831         struct rdma_id_private *id_priv;
2832
2833         /* It is a five-bit value */
2834         if (min_rnr_timer & 0xe0)
2835                 return -EINVAL;
2836
2837         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2838                 return -EINVAL;
2839
2840         id_priv = container_of(id, struct rdma_id_private, id);
2841         mutex_lock(&id_priv->qp_mutex);
2842         id_priv->min_rnr_timer = min_rnr_timer;
2843         id_priv->min_rnr_timer_set = true;
2844         mutex_unlock(&id_priv->qp_mutex);
2845
2846         return 0;
2847 }
2848 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2849
2850 static int route_set_path_rec_inbound(struct cma_work *work,
2851                                       struct sa_path_rec *path_rec)
2852 {
2853         struct rdma_route *route = &work->id->id.route;
2854
2855         if (!route->path_rec_inbound) {
2856                 route->path_rec_inbound =
2857                         kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
2858                 if (!route->path_rec_inbound)
2859                         return -ENOMEM;
2860         }
2861
2862         *route->path_rec_inbound = *path_rec;
2863         return 0;
2864 }
2865
2866 static int route_set_path_rec_outbound(struct cma_work *work,
2867                                        struct sa_path_rec *path_rec)
2868 {
2869         struct rdma_route *route = &work->id->id.route;
2870
2871         if (!route->path_rec_outbound) {
2872                 route->path_rec_outbound =
2873                         kzalloc(sizeof(*route->path_rec_outbound), GFP_KERNEL);
2874                 if (!route->path_rec_outbound)
2875                         return -ENOMEM;
2876         }
2877
2878         *route->path_rec_outbound = *path_rec;
2879         return 0;
2880 }
2881
2882 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2883                               unsigned int num_prs, void *context)
2884 {
2885         struct cma_work *work = context;
2886         struct rdma_route *route;
2887         int i;
2888
2889         route = &work->id->id.route;
2890
2891         if (status)
2892                 goto fail;
2893
2894         for (i = 0; i < num_prs; i++) {
2895                 if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
2896                         *route->path_rec = path_rec[i];
2897                 else if (path_rec[i].flags & IB_PATH_INBOUND)
2898                         status = route_set_path_rec_inbound(work, &path_rec[i]);
2899                 else if (path_rec[i].flags & IB_PATH_OUTBOUND)
2900                         status = route_set_path_rec_outbound(work,
2901                                                              &path_rec[i]);
2902                 else
2903                         status = -EINVAL;
2904
2905                 if (status)
2906                         goto fail;
2907         }
2908
2909         route->num_pri_alt_paths = 1;
2910         queue_work(cma_wq, &work->work);
2911         return;
2912
2913 fail:
2914         work->old_state = RDMA_CM_ROUTE_QUERY;
2915         work->new_state = RDMA_CM_ADDR_RESOLVED;
2916         work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2917         work->event.status = status;
2918         pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2919                              status);
2920         queue_work(cma_wq, &work->work);
2921 }
2922
2923 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2924                               unsigned long timeout_ms, struct cma_work *work)
2925 {
2926         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2927         struct sa_path_rec path_rec;
2928         ib_sa_comp_mask comp_mask;
2929         struct sockaddr_in6 *sin6;
2930         struct sockaddr_ib *sib;
2931
2932         memset(&path_rec, 0, sizeof path_rec);
2933
2934         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2935                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2936         else
2937                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2938         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2939         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2940         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2941         path_rec.numb_path = 1;
2942         path_rec.reversible = 1;
2943         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2944                                                   cma_dst_addr(id_priv));
2945
2946         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2947                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2948                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2949
2950         switch (cma_family(id_priv)) {
2951         case AF_INET:
2952                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2953                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2954                 break;
2955         case AF_INET6:
2956                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2957                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2958                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2959                 break;
2960         case AF_IB:
2961                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2962                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2963                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2964                 break;
2965         }
2966
2967         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2968                                                id_priv->id.port_num, &path_rec,
2969                                                comp_mask, timeout_ms,
2970                                                GFP_KERNEL, cma_query_handler,
2971                                                work, &id_priv->query);
2972
2973         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2974 }
2975
2976 static void cma_iboe_join_work_handler(struct work_struct *work)
2977 {
2978         struct cma_multicast *mc =
2979                 container_of(work, struct cma_multicast, iboe_join.work);
2980         struct rdma_cm_event *event = &mc->iboe_join.event;
2981         struct rdma_id_private *id_priv = mc->id_priv;
2982         int ret;
2983
2984         mutex_lock(&id_priv->handler_mutex);
2985         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2986             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2987                 goto out_unlock;
2988
2989         ret = cma_cm_event_handler(id_priv, event);
2990         WARN_ON(ret);
2991
2992 out_unlock:
2993         mutex_unlock(&id_priv->handler_mutex);
2994         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2995                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2996 }
2997
2998 static void cma_work_handler(struct work_struct *_work)
2999 {
3000         struct cma_work *work = container_of(_work, struct cma_work, work);
3001         struct rdma_id_private *id_priv = work->id;
3002
3003         mutex_lock(&id_priv->handler_mutex);
3004         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
3005             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
3006                 goto out_unlock;
3007         if (work->old_state != 0 || work->new_state != 0) {
3008                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
3009                         goto out_unlock;
3010         }
3011
3012         if (cma_cm_event_handler(id_priv, &work->event)) {
3013                 cma_id_put(id_priv);
3014                 destroy_id_handler_unlock(id_priv);
3015                 goto out_free;
3016         }
3017
3018 out_unlock:
3019         mutex_unlock(&id_priv->handler_mutex);
3020         cma_id_put(id_priv);
3021 out_free:
3022         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
3023                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
3024         kfree(work);
3025 }
3026
3027 static void cma_init_resolve_route_work(struct cma_work *work,
3028                                         struct rdma_id_private *id_priv)
3029 {
3030         work->id = id_priv;
3031         INIT_WORK(&work->work, cma_work_handler);
3032         work->old_state = RDMA_CM_ROUTE_QUERY;
3033         work->new_state = RDMA_CM_ROUTE_RESOLVED;
3034         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
3035 }
3036
3037 static void enqueue_resolve_addr_work(struct cma_work *work,
3038                                       struct rdma_id_private *id_priv)
3039 {
3040         /* Balances with cma_id_put() in cma_work_handler */
3041         cma_id_get(id_priv);
3042
3043         work->id = id_priv;
3044         INIT_WORK(&work->work, cma_work_handler);
3045         work->old_state = RDMA_CM_ADDR_QUERY;
3046         work->new_state = RDMA_CM_ADDR_RESOLVED;
3047         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3048
3049         queue_work(cma_wq, &work->work);
3050 }
3051
3052 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
3053                                 unsigned long timeout_ms)
3054 {
3055         struct rdma_route *route = &id_priv->id.route;
3056         struct cma_work *work;
3057         int ret;
3058
3059         work = kzalloc(sizeof *work, GFP_KERNEL);
3060         if (!work)
3061                 return -ENOMEM;
3062
3063         cma_init_resolve_route_work(work, id_priv);
3064
3065         if (!route->path_rec)
3066                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
3067         if (!route->path_rec) {
3068                 ret = -ENOMEM;
3069                 goto err1;
3070         }
3071
3072         ret = cma_query_ib_route(id_priv, timeout_ms, work);
3073         if (ret)
3074                 goto err2;
3075
3076         return 0;
3077 err2:
3078         kfree(route->path_rec);
3079         route->path_rec = NULL;
3080 err1:
3081         kfree(work);
3082         return ret;
3083 }
3084
3085 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
3086                                            unsigned long supported_gids,
3087                                            enum ib_gid_type default_gid)
3088 {
3089         if ((network_type == RDMA_NETWORK_IPV4 ||
3090              network_type == RDMA_NETWORK_IPV6) &&
3091             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
3092                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
3093
3094         return default_gid;
3095 }
3096
3097 /*
3098  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
3099  * path record type based on GID type.
3100  * It also sets up other L2 fields which includes destination mac address
3101  * netdev ifindex, of the path record.
3102  * It returns the netdev of the bound interface for this path record entry.
3103  */
3104 static struct net_device *
3105 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
3106 {
3107         struct rdma_route *route = &id_priv->id.route;
3108         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
3109         struct rdma_addr *addr = &route->addr;
3110         unsigned long supported_gids;
3111         struct net_device *ndev;
3112
3113         if (!addr->dev_addr.bound_dev_if)
3114                 return NULL;
3115
3116         ndev = dev_get_by_index(addr->dev_addr.net,
3117                                 addr->dev_addr.bound_dev_if);
3118         if (!ndev)
3119                 return NULL;
3120
3121         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
3122                                                     id_priv->id.port_num);
3123         gid_type = cma_route_gid_type(addr->dev_addr.network,
3124                                       supported_gids,
3125                                       id_priv->gid_type);
3126         /* Use the hint from IP Stack to select GID Type */
3127         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
3128                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
3129         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
3130
3131         route->path_rec->roce.route_resolved = true;
3132         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
3133         return ndev;
3134 }
3135
3136 int rdma_set_ib_path(struct rdma_cm_id *id,
3137                      struct sa_path_rec *path_rec)
3138 {
3139         struct rdma_id_private *id_priv;
3140         struct net_device *ndev;
3141         int ret;
3142
3143         id_priv = container_of(id, struct rdma_id_private, id);
3144         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3145                            RDMA_CM_ROUTE_RESOLVED))
3146                 return -EINVAL;
3147
3148         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
3149                                      GFP_KERNEL);
3150         if (!id->route.path_rec) {
3151                 ret = -ENOMEM;
3152                 goto err;
3153         }
3154
3155         if (rdma_protocol_roce(id->device, id->port_num)) {
3156                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3157                 if (!ndev) {
3158                         ret = -ENODEV;
3159                         goto err_free;
3160                 }
3161                 dev_put(ndev);
3162         }
3163
3164         id->route.num_pri_alt_paths = 1;
3165         return 0;
3166
3167 err_free:
3168         kfree(id->route.path_rec);
3169         id->route.path_rec = NULL;
3170 err:
3171         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
3172         return ret;
3173 }
3174 EXPORT_SYMBOL(rdma_set_ib_path);
3175
3176 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
3177 {
3178         struct cma_work *work;
3179
3180         work = kzalloc(sizeof *work, GFP_KERNEL);
3181         if (!work)
3182                 return -ENOMEM;
3183
3184         cma_init_resolve_route_work(work, id_priv);
3185         queue_work(cma_wq, &work->work);
3186         return 0;
3187 }
3188
3189 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
3190 {
3191         struct net_device *dev;
3192
3193         dev = vlan_dev_real_dev(vlan_ndev);
3194         if (dev->num_tc)
3195                 return netdev_get_prio_tc_map(dev, prio);
3196
3197         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
3198                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
3199 }
3200
3201 struct iboe_prio_tc_map {
3202         int input_prio;
3203         int output_tc;
3204         bool found;
3205 };
3206
3207 static int get_lower_vlan_dev_tc(struct net_device *dev,
3208                                  struct netdev_nested_priv *priv)
3209 {
3210         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3211
3212         if (is_vlan_dev(dev))
3213                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3214         else if (dev->num_tc)
3215                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3216         else
3217                 map->output_tc = 0;
3218         /* We are interested only in first level VLAN device, so always
3219          * return 1 to stop iterating over next level devices.
3220          */
3221         map->found = true;
3222         return 1;
3223 }
3224
3225 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3226 {
3227         struct iboe_prio_tc_map prio_tc_map = {};
3228         int prio = rt_tos2priority(tos);
3229         struct netdev_nested_priv priv;
3230
3231         /* If VLAN device, get it directly from the VLAN netdev */
3232         if (is_vlan_dev(ndev))
3233                 return get_vlan_ndev_tc(ndev, prio);
3234
3235         prio_tc_map.input_prio = prio;
3236         priv.data = (void *)&prio_tc_map;
3237         rcu_read_lock();
3238         netdev_walk_all_lower_dev_rcu(ndev,
3239                                       get_lower_vlan_dev_tc,
3240                                       &priv);
3241         rcu_read_unlock();
3242         /* If map is found from lower device, use it; Otherwise
3243          * continue with the current netdevice to get priority to tc map.
3244          */
3245         if (prio_tc_map.found)
3246                 return prio_tc_map.output_tc;
3247         else if (ndev->num_tc)
3248                 return netdev_get_prio_tc_map(ndev, prio);
3249         else
3250                 return 0;
3251 }
3252
3253 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3254 {
3255         struct sockaddr_in6 *addr6;
3256         u16 dport, sport;
3257         u32 hash, fl;
3258
3259         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3260         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3261         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3262                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3263                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3264                 hash = (u32)sport * 31 + dport;
3265                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3266         }
3267
3268         return cpu_to_be32(fl);
3269 }
3270
3271 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3272 {
3273         struct rdma_route *route = &id_priv->id.route;
3274         struct rdma_addr *addr = &route->addr;
3275         struct cma_work *work;
3276         int ret;
3277         struct net_device *ndev;
3278
3279         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3280                                         rdma_start_port(id_priv->cma_dev->device)];
3281         u8 tos;
3282
3283         mutex_lock(&id_priv->qp_mutex);
3284         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3285         mutex_unlock(&id_priv->qp_mutex);
3286
3287         work = kzalloc(sizeof *work, GFP_KERNEL);
3288         if (!work)
3289                 return -ENOMEM;
3290
3291         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3292         if (!route->path_rec) {
3293                 ret = -ENOMEM;
3294                 goto err1;
3295         }
3296
3297         route->num_pri_alt_paths = 1;
3298
3299         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3300         if (!ndev) {
3301                 ret = -ENODEV;
3302                 goto err2;
3303         }
3304
3305         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3306                     &route->path_rec->sgid);
3307         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3308                     &route->path_rec->dgid);
3309
3310         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3311                 /* TODO: get the hoplimit from the inet/inet6 device */
3312                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3313         else
3314                 route->path_rec->hop_limit = 1;
3315         route->path_rec->reversible = 1;
3316         route->path_rec->pkey = cpu_to_be16(0xffff);
3317         route->path_rec->mtu_selector = IB_SA_EQ;
3318         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3319         route->path_rec->traffic_class = tos;
3320         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3321         route->path_rec->rate_selector = IB_SA_EQ;
3322         route->path_rec->rate = IB_RATE_PORT_CURRENT;
3323         dev_put(ndev);
3324         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3325         /* In case ACK timeout is set, use this value to calculate
3326          * PacketLifeTime.  As per IBTA 12.7.34,
3327          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3328          * Assuming a negligible local ACK delay, we can use
3329          * PacketLifeTime = local ACK timeout/2
3330          * as a reasonable approximation for RoCE networks.
3331          */
3332         mutex_lock(&id_priv->qp_mutex);
3333         if (id_priv->timeout_set && id_priv->timeout)
3334                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3335         else
3336                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3337         mutex_unlock(&id_priv->qp_mutex);
3338
3339         if (!route->path_rec->mtu) {
3340                 ret = -EINVAL;
3341                 goto err2;
3342         }
3343
3344         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3345                                          id_priv->id.port_num))
3346                 route->path_rec->flow_label =
3347                         cma_get_roce_udp_flow_label(id_priv);
3348
3349         cma_init_resolve_route_work(work, id_priv);
3350         queue_work(cma_wq, &work->work);
3351
3352         return 0;
3353
3354 err2:
3355         kfree(route->path_rec);
3356         route->path_rec = NULL;
3357         route->num_pri_alt_paths = 0;
3358 err1:
3359         kfree(work);
3360         return ret;
3361 }
3362
3363 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3364 {
3365         struct rdma_id_private *id_priv;
3366         int ret;
3367
3368         if (!timeout_ms)
3369                 return -EINVAL;
3370
3371         id_priv = container_of(id, struct rdma_id_private, id);
3372         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3373                 return -EINVAL;
3374
3375         cma_id_get(id_priv);
3376         if (rdma_cap_ib_sa(id->device, id->port_num))
3377                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3378         else if (rdma_protocol_roce(id->device, id->port_num)) {
3379                 ret = cma_resolve_iboe_route(id_priv);
3380                 if (!ret)
3381                         cma_add_id_to_tree(id_priv);
3382         }
3383         else if (rdma_protocol_iwarp(id->device, id->port_num))
3384                 ret = cma_resolve_iw_route(id_priv);
3385         else
3386                 ret = -ENOSYS;
3387
3388         if (ret)
3389                 goto err;
3390
3391         return 0;
3392 err:
3393         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3394         cma_id_put(id_priv);
3395         return ret;
3396 }
3397 EXPORT_SYMBOL(rdma_resolve_route);
3398
3399 static void cma_set_loopback(struct sockaddr *addr)
3400 {
3401         switch (addr->sa_family) {
3402         case AF_INET:
3403                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3404                 break;
3405         case AF_INET6:
3406                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3407                               0, 0, 0, htonl(1));
3408                 break;
3409         default:
3410                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3411                             0, 0, 0, htonl(1));
3412                 break;
3413         }
3414 }
3415
3416 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3417 {
3418         struct cma_device *cma_dev, *cur_dev;
3419         union ib_gid gid;
3420         enum ib_port_state port_state;
3421         unsigned int p;
3422         u16 pkey;
3423         int ret;
3424
3425         cma_dev = NULL;
3426         mutex_lock(&lock);
3427         list_for_each_entry(cur_dev, &dev_list, list) {
3428                 if (cma_family(id_priv) == AF_IB &&
3429                     !rdma_cap_ib_cm(cur_dev->device, 1))
3430                         continue;
3431
3432                 if (!cma_dev)
3433                         cma_dev = cur_dev;
3434
3435                 rdma_for_each_port (cur_dev->device, p) {
3436                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3437                             port_state == IB_PORT_ACTIVE) {
3438                                 cma_dev = cur_dev;
3439                                 goto port_found;
3440                         }
3441                 }
3442         }
3443
3444         if (!cma_dev) {
3445                 ret = -ENODEV;
3446                 goto out;
3447         }
3448
3449         p = 1;
3450
3451 port_found:
3452         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3453         if (ret)
3454                 goto out;
3455
3456         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3457         if (ret)
3458                 goto out;
3459
3460         id_priv->id.route.addr.dev_addr.dev_type =
3461                 (rdma_protocol_ib(cma_dev->device, p)) ?
3462                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3463
3464         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3465         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3466         id_priv->id.port_num = p;
3467         cma_attach_to_dev(id_priv, cma_dev);
3468         rdma_restrack_add(&id_priv->res);
3469         cma_set_loopback(cma_src_addr(id_priv));
3470 out:
3471         mutex_unlock(&lock);
3472         return ret;
3473 }
3474
3475 static void addr_handler(int status, struct sockaddr *src_addr,
3476                          struct rdma_dev_addr *dev_addr, void *context)
3477 {
3478         struct rdma_id_private *id_priv = context;
3479         struct rdma_cm_event event = {};
3480         struct sockaddr *addr;
3481         struct sockaddr_storage old_addr;
3482
3483         mutex_lock(&id_priv->handler_mutex);
3484         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3485                            RDMA_CM_ADDR_RESOLVED))
3486                 goto out;
3487
3488         /*
3489          * Store the previous src address, so that if we fail to acquire
3490          * matching rdma device, old address can be restored back, which helps
3491          * to cancel the cma listen operation correctly.
3492          */
3493         addr = cma_src_addr(id_priv);
3494         memcpy(&old_addr, addr, rdma_addr_size(addr));
3495         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3496         if (!status && !id_priv->cma_dev) {
3497                 status = cma_acquire_dev_by_src_ip(id_priv);
3498                 if (status)
3499                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3500                                              status);
3501                 rdma_restrack_add(&id_priv->res);
3502         } else if (status) {
3503                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3504         }
3505
3506         if (status) {
3507                 memcpy(addr, &old_addr,
3508                        rdma_addr_size((struct sockaddr *)&old_addr));
3509                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3510                                    RDMA_CM_ADDR_BOUND))
3511                         goto out;
3512                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3513                 event.status = status;
3514         } else
3515                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3516
3517         if (cma_cm_event_handler(id_priv, &event)) {
3518                 destroy_id_handler_unlock(id_priv);
3519                 return;
3520         }
3521 out:
3522         mutex_unlock(&id_priv->handler_mutex);
3523 }
3524
3525 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3526 {
3527         struct cma_work *work;
3528         union ib_gid gid;
3529         int ret;
3530
3531         work = kzalloc(sizeof *work, GFP_KERNEL);
3532         if (!work)
3533                 return -ENOMEM;
3534
3535         if (!id_priv->cma_dev) {
3536                 ret = cma_bind_loopback(id_priv);
3537                 if (ret)
3538                         goto err;
3539         }
3540
3541         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3542         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3543
3544         enqueue_resolve_addr_work(work, id_priv);
3545         return 0;
3546 err:
3547         kfree(work);
3548         return ret;
3549 }
3550
3551 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3552 {
3553         struct cma_work *work;
3554         int ret;
3555
3556         work = kzalloc(sizeof *work, GFP_KERNEL);
3557         if (!work)
3558                 return -ENOMEM;
3559
3560         if (!id_priv->cma_dev) {
3561                 ret = cma_resolve_ib_dev(id_priv);
3562                 if (ret)
3563                         goto err;
3564         }
3565
3566         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3567                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3568
3569         enqueue_resolve_addr_work(work, id_priv);
3570         return 0;
3571 err:
3572         kfree(work);
3573         return ret;
3574 }
3575
3576 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3577 {
3578         struct rdma_id_private *id_priv;
3579         unsigned long flags;
3580         int ret;
3581
3582         id_priv = container_of(id, struct rdma_id_private, id);
3583         spin_lock_irqsave(&id_priv->lock, flags);
3584         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3585             id_priv->state == RDMA_CM_IDLE) {
3586                 id_priv->reuseaddr = reuse;
3587                 ret = 0;
3588         } else {
3589                 ret = -EINVAL;
3590         }
3591         spin_unlock_irqrestore(&id_priv->lock, flags);
3592         return ret;
3593 }
3594 EXPORT_SYMBOL(rdma_set_reuseaddr);
3595
3596 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3597 {
3598         struct rdma_id_private *id_priv;
3599         unsigned long flags;
3600         int ret;
3601
3602         id_priv = container_of(id, struct rdma_id_private, id);
3603         spin_lock_irqsave(&id_priv->lock, flags);
3604         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3605                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3606                 id_priv->afonly = afonly;
3607                 ret = 0;
3608         } else {
3609                 ret = -EINVAL;
3610         }
3611         spin_unlock_irqrestore(&id_priv->lock, flags);
3612         return ret;
3613 }
3614 EXPORT_SYMBOL(rdma_set_afonly);
3615
3616 static void cma_bind_port(struct rdma_bind_list *bind_list,
3617                           struct rdma_id_private *id_priv)
3618 {
3619         struct sockaddr *addr;
3620         struct sockaddr_ib *sib;
3621         u64 sid, mask;
3622         __be16 port;
3623
3624         lockdep_assert_held(&lock);
3625
3626         addr = cma_src_addr(id_priv);
3627         port = htons(bind_list->port);
3628
3629         switch (addr->sa_family) {
3630         case AF_INET:
3631                 ((struct sockaddr_in *) addr)->sin_port = port;
3632                 break;
3633         case AF_INET6:
3634                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3635                 break;
3636         case AF_IB:
3637                 sib = (struct sockaddr_ib *) addr;
3638                 sid = be64_to_cpu(sib->sib_sid);
3639                 mask = be64_to_cpu(sib->sib_sid_mask);
3640                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3641                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3642                 break;
3643         }
3644         id_priv->bind_list = bind_list;
3645         hlist_add_head(&id_priv->node, &bind_list->owners);
3646 }
3647
3648 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3649                           struct rdma_id_private *id_priv, unsigned short snum)
3650 {
3651         struct rdma_bind_list *bind_list;
3652         int ret;
3653
3654         lockdep_assert_held(&lock);
3655
3656         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3657         if (!bind_list)
3658                 return -ENOMEM;
3659
3660         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3661                            snum);
3662         if (ret < 0)
3663                 goto err;
3664
3665         bind_list->ps = ps;
3666         bind_list->port = snum;
3667         cma_bind_port(bind_list, id_priv);
3668         return 0;
3669 err:
3670         kfree(bind_list);
3671         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3672 }
3673
3674 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3675                               struct rdma_id_private *id_priv)
3676 {
3677         struct rdma_id_private *cur_id;
3678         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3679         struct sockaddr  *saddr = cma_src_addr(id_priv);
3680         __be16 dport = cma_port(daddr);
3681
3682         lockdep_assert_held(&lock);
3683
3684         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3685                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3686                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3687                 __be16 cur_dport = cma_port(cur_daddr);
3688
3689                 if (id_priv == cur_id)
3690                         continue;
3691
3692                 /* different dest port -> unique */
3693                 if (!cma_any_port(daddr) &&
3694                     !cma_any_port(cur_daddr) &&
3695                     (dport != cur_dport))
3696                         continue;
3697
3698                 /* different src address -> unique */
3699                 if (!cma_any_addr(saddr) &&
3700                     !cma_any_addr(cur_saddr) &&
3701                     cma_addr_cmp(saddr, cur_saddr))
3702                         continue;
3703
3704                 /* different dst address -> unique */
3705                 if (!cma_any_addr(daddr) &&
3706                     !cma_any_addr(cur_daddr) &&
3707                     cma_addr_cmp(daddr, cur_daddr))
3708                         continue;
3709
3710                 return -EADDRNOTAVAIL;
3711         }
3712         return 0;
3713 }
3714
3715 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3716                               struct rdma_id_private *id_priv)
3717 {
3718         static unsigned int last_used_port;
3719         int low, high, remaining;
3720         unsigned int rover;
3721         struct net *net = id_priv->id.route.addr.dev_addr.net;
3722
3723         lockdep_assert_held(&lock);
3724
3725         inet_get_local_port_range(net, &low, &high);
3726         remaining = (high - low) + 1;
3727         rover = get_random_u32_inclusive(low, remaining + low - 1);
3728 retry:
3729         if (last_used_port != rover) {
3730                 struct rdma_bind_list *bind_list;
3731                 int ret;
3732
3733                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3734
3735                 if (!bind_list) {
3736                         ret = cma_alloc_port(ps, id_priv, rover);
3737                 } else {
3738                         ret = cma_port_is_unique(bind_list, id_priv);
3739                         if (!ret)
3740                                 cma_bind_port(bind_list, id_priv);
3741                 }
3742                 /*
3743                  * Remember previously used port number in order to avoid
3744                  * re-using same port immediately after it is closed.
3745                  */
3746                 if (!ret)
3747                         last_used_port = rover;
3748                 if (ret != -EADDRNOTAVAIL)
3749                         return ret;
3750         }
3751         if (--remaining) {
3752                 rover++;
3753                 if ((rover < low) || (rover > high))
3754                         rover = low;
3755                 goto retry;
3756         }
3757         return -EADDRNOTAVAIL;
3758 }
3759
3760 /*
3761  * Check that the requested port is available.  This is called when trying to
3762  * bind to a specific port, or when trying to listen on a bound port.  In
3763  * the latter case, the provided id_priv may already be on the bind_list, but
3764  * we still need to check that it's okay to start listening.
3765  */
3766 static int cma_check_port(struct rdma_bind_list *bind_list,
3767                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3768 {
3769         struct rdma_id_private *cur_id;
3770         struct sockaddr *addr, *cur_addr;
3771
3772         lockdep_assert_held(&lock);
3773
3774         addr = cma_src_addr(id_priv);
3775         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3776                 if (id_priv == cur_id)
3777                         continue;
3778
3779                 if (reuseaddr && cur_id->reuseaddr)
3780                         continue;
3781
3782                 cur_addr = cma_src_addr(cur_id);
3783                 if (id_priv->afonly && cur_id->afonly &&
3784                     (addr->sa_family != cur_addr->sa_family))
3785                         continue;
3786
3787                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3788                         return -EADDRNOTAVAIL;
3789
3790                 if (!cma_addr_cmp(addr, cur_addr))
3791                         return -EADDRINUSE;
3792         }
3793         return 0;
3794 }
3795
3796 static int cma_use_port(enum rdma_ucm_port_space ps,
3797                         struct rdma_id_private *id_priv)
3798 {
3799         struct rdma_bind_list *bind_list;
3800         unsigned short snum;
3801         int ret;
3802
3803         lockdep_assert_held(&lock);
3804
3805         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3806         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3807                 return -EACCES;
3808
3809         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3810         if (!bind_list) {
3811                 ret = cma_alloc_port(ps, id_priv, snum);
3812         } else {
3813                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3814                 if (!ret)
3815                         cma_bind_port(bind_list, id_priv);
3816         }
3817         return ret;
3818 }
3819
3820 static enum rdma_ucm_port_space
3821 cma_select_inet_ps(struct rdma_id_private *id_priv)
3822 {
3823         switch (id_priv->id.ps) {
3824         case RDMA_PS_TCP:
3825         case RDMA_PS_UDP:
3826         case RDMA_PS_IPOIB:
3827         case RDMA_PS_IB:
3828                 return id_priv->id.ps;
3829         default:
3830
3831                 return 0;
3832         }
3833 }
3834
3835 static enum rdma_ucm_port_space
3836 cma_select_ib_ps(struct rdma_id_private *id_priv)
3837 {
3838         enum rdma_ucm_port_space ps = 0;
3839         struct sockaddr_ib *sib;
3840         u64 sid_ps, mask, sid;
3841
3842         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3843         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3844         sid = be64_to_cpu(sib->sib_sid) & mask;
3845
3846         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3847                 sid_ps = RDMA_IB_IP_PS_IB;
3848                 ps = RDMA_PS_IB;
3849         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3850                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3851                 sid_ps = RDMA_IB_IP_PS_TCP;
3852                 ps = RDMA_PS_TCP;
3853         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3854                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3855                 sid_ps = RDMA_IB_IP_PS_UDP;
3856                 ps = RDMA_PS_UDP;
3857         }
3858
3859         if (ps) {
3860                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3861                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3862                                                 be64_to_cpu(sib->sib_sid_mask));
3863         }
3864         return ps;
3865 }
3866
3867 static int cma_get_port(struct rdma_id_private *id_priv)
3868 {
3869         enum rdma_ucm_port_space ps;
3870         int ret;
3871
3872         if (cma_family(id_priv) != AF_IB)
3873                 ps = cma_select_inet_ps(id_priv);
3874         else
3875                 ps = cma_select_ib_ps(id_priv);
3876         if (!ps)
3877                 return -EPROTONOSUPPORT;
3878
3879         mutex_lock(&lock);
3880         if (cma_any_port(cma_src_addr(id_priv)))
3881                 ret = cma_alloc_any_port(ps, id_priv);
3882         else
3883                 ret = cma_use_port(ps, id_priv);
3884         mutex_unlock(&lock);
3885
3886         return ret;
3887 }
3888
3889 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3890                                struct sockaddr *addr)
3891 {
3892 #if IS_ENABLED(CONFIG_IPV6)
3893         struct sockaddr_in6 *sin6;
3894
3895         if (addr->sa_family != AF_INET6)
3896                 return 0;
3897
3898         sin6 = (struct sockaddr_in6 *) addr;
3899
3900         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3901                 return 0;
3902
3903         if (!sin6->sin6_scope_id)
3904                         return -EINVAL;
3905
3906         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3907 #endif
3908         return 0;
3909 }
3910
3911 int rdma_listen(struct rdma_cm_id *id, int backlog)
3912 {
3913         struct rdma_id_private *id_priv =
3914                 container_of(id, struct rdma_id_private, id);
3915         int ret;
3916
3917         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3918                 struct sockaddr_in any_in = {
3919                         .sin_family = AF_INET,
3920                         .sin_addr.s_addr = htonl(INADDR_ANY),
3921                 };
3922
3923                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3924                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3925                 if (ret)
3926                         return ret;
3927                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3928                                            RDMA_CM_LISTEN)))
3929                         return -EINVAL;
3930         }
3931
3932         /*
3933          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3934          * any more, and has to be unique in the bind list.
3935          */
3936         if (id_priv->reuseaddr) {
3937                 mutex_lock(&lock);
3938                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3939                 if (!ret)
3940                         id_priv->reuseaddr = 0;
3941                 mutex_unlock(&lock);
3942                 if (ret)
3943                         goto err;
3944         }
3945
3946         id_priv->backlog = backlog;
3947         if (id_priv->cma_dev) {
3948                 if (rdma_cap_ib_cm(id->device, 1)) {
3949                         ret = cma_ib_listen(id_priv);
3950                         if (ret)
3951                                 goto err;
3952                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3953                         ret = cma_iw_listen(id_priv, backlog);
3954                         if (ret)
3955                                 goto err;
3956                 } else {
3957                         ret = -ENOSYS;
3958                         goto err;
3959                 }
3960         } else {
3961                 ret = cma_listen_on_all(id_priv);
3962                 if (ret)
3963                         goto err;
3964         }
3965
3966         return 0;
3967 err:
3968         id_priv->backlog = 0;
3969         /*
3970          * All the failure paths that lead here will not allow the req_handler's
3971          * to have run.
3972          */
3973         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3974         return ret;
3975 }
3976 EXPORT_SYMBOL(rdma_listen);
3977
3978 static int rdma_bind_addr_dst(struct rdma_id_private *id_priv,
3979                               struct sockaddr *addr, const struct sockaddr *daddr)
3980 {
3981         struct sockaddr *id_daddr;
3982         int ret;
3983
3984         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3985             addr->sa_family != AF_IB)
3986                 return -EAFNOSUPPORT;
3987
3988         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3989                 return -EINVAL;
3990
3991         ret = cma_check_linklocal(&id_priv->id.route.addr.dev_addr, addr);
3992         if (ret)
3993                 goto err1;
3994
3995         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3996         if (!cma_any_addr(addr)) {
3997                 ret = cma_translate_addr(addr, &id_priv->id.route.addr.dev_addr);
3998                 if (ret)
3999                         goto err1;
4000
4001                 ret = cma_acquire_dev_by_src_ip(id_priv);
4002                 if (ret)
4003                         goto err1;
4004         }
4005
4006         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
4007                 if (addr->sa_family == AF_INET)
4008                         id_priv->afonly = 1;
4009 #if IS_ENABLED(CONFIG_IPV6)
4010                 else if (addr->sa_family == AF_INET6) {
4011                         struct net *net = id_priv->id.route.addr.dev_addr.net;
4012
4013                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
4014                 }
4015 #endif
4016         }
4017         id_daddr = cma_dst_addr(id_priv);
4018         if (daddr != id_daddr)
4019                 memcpy(id_daddr, daddr, rdma_addr_size(addr));
4020         id_daddr->sa_family = addr->sa_family;
4021
4022         ret = cma_get_port(id_priv);
4023         if (ret)
4024                 goto err2;
4025
4026         if (!cma_any_addr(addr))
4027                 rdma_restrack_add(&id_priv->res);
4028         return 0;
4029 err2:
4030         if (id_priv->cma_dev)
4031                 cma_release_dev(id_priv);
4032 err1:
4033         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
4034         return ret;
4035 }
4036
4037 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
4038                          const struct sockaddr *dst_addr)
4039 {
4040         struct rdma_id_private *id_priv =
4041                 container_of(id, struct rdma_id_private, id);
4042         struct sockaddr_storage zero_sock = {};
4043
4044         if (src_addr && src_addr->sa_family)
4045                 return rdma_bind_addr_dst(id_priv, src_addr, dst_addr);
4046
4047         /*
4048          * When the src_addr is not specified, automatically supply an any addr
4049          */
4050         zero_sock.ss_family = dst_addr->sa_family;
4051         if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
4052                 struct sockaddr_in6 *src_addr6 =
4053                         (struct sockaddr_in6 *)&zero_sock;
4054                 struct sockaddr_in6 *dst_addr6 =
4055                         (struct sockaddr_in6 *)dst_addr;
4056
4057                 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
4058                 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
4059                         id->route.addr.dev_addr.bound_dev_if =
4060                                 dst_addr6->sin6_scope_id;
4061         } else if (dst_addr->sa_family == AF_IB) {
4062                 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
4063                         ((struct sockaddr_ib *)dst_addr)->sib_pkey;
4064         }
4065         return rdma_bind_addr_dst(id_priv, (struct sockaddr *)&zero_sock, dst_addr);
4066 }
4067
4068 /*
4069  * If required, resolve the source address for bind and leave the id_priv in
4070  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
4071  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
4072  * ignored.
4073  */
4074 static int resolve_prepare_src(struct rdma_id_private *id_priv,
4075                                struct sockaddr *src_addr,
4076                                const struct sockaddr *dst_addr)
4077 {
4078         int ret;
4079
4080         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
4081                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
4082                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
4083                 if (ret)
4084                         return ret;
4085                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
4086                                            RDMA_CM_ADDR_QUERY)))
4087                         return -EINVAL;
4088
4089         } else {
4090                 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
4091         }
4092
4093         if (cma_family(id_priv) != dst_addr->sa_family) {
4094                 ret = -EINVAL;
4095                 goto err_state;
4096         }
4097         return 0;
4098
4099 err_state:
4100         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
4101         return ret;
4102 }
4103
4104 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
4105                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
4106 {
4107         struct rdma_id_private *id_priv =
4108                 container_of(id, struct rdma_id_private, id);
4109         int ret;
4110
4111         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
4112         if (ret)
4113                 return ret;
4114
4115         if (cma_any_addr(dst_addr)) {
4116                 ret = cma_resolve_loopback(id_priv);
4117         } else {
4118                 if (dst_addr->sa_family == AF_IB) {
4119                         ret = cma_resolve_ib_addr(id_priv);
4120                 } else {
4121                         /*
4122                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
4123                          * rdma_resolve_ip() is called, eg through the error
4124                          * path in addr_handler(). If this happens the existing
4125                          * request must be canceled before issuing a new one.
4126                          * Since canceling a request is a bit slow and this
4127                          * oddball path is rare, keep track once a request has
4128                          * been issued. The track turns out to be a permanent
4129                          * state since this is the only cancel as it is
4130                          * immediately before rdma_resolve_ip().
4131                          */
4132                         if (id_priv->used_resolve_ip)
4133                                 rdma_addr_cancel(&id->route.addr.dev_addr);
4134                         else
4135                                 id_priv->used_resolve_ip = 1;
4136                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
4137                                               &id->route.addr.dev_addr,
4138                                               timeout_ms, addr_handler,
4139                                               false, id_priv);
4140                 }
4141         }
4142         if (ret)
4143                 goto err;
4144
4145         return 0;
4146 err:
4147         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
4148         return ret;
4149 }
4150 EXPORT_SYMBOL(rdma_resolve_addr);
4151
4152 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
4153 {
4154         struct rdma_id_private *id_priv =
4155                 container_of(id, struct rdma_id_private, id);
4156
4157         return rdma_bind_addr_dst(id_priv, addr, cma_dst_addr(id_priv));
4158 }
4159 EXPORT_SYMBOL(rdma_bind_addr);
4160
4161 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
4162 {
4163         struct cma_hdr *cma_hdr;
4164
4165         cma_hdr = hdr;
4166         cma_hdr->cma_version = CMA_VERSION;
4167         if (cma_family(id_priv) == AF_INET) {
4168                 struct sockaddr_in *src4, *dst4;
4169
4170                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
4171                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
4172
4173                 cma_set_ip_ver(cma_hdr, 4);
4174                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
4175                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
4176                 cma_hdr->port = src4->sin_port;
4177         } else if (cma_family(id_priv) == AF_INET6) {
4178                 struct sockaddr_in6 *src6, *dst6;
4179
4180                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
4181                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
4182
4183                 cma_set_ip_ver(cma_hdr, 6);
4184                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
4185                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
4186                 cma_hdr->port = src6->sin6_port;
4187         }
4188         return 0;
4189 }
4190
4191 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
4192                                 const struct ib_cm_event *ib_event)
4193 {
4194         struct rdma_id_private *id_priv = cm_id->context;
4195         struct rdma_cm_event event = {};
4196         const struct ib_cm_sidr_rep_event_param *rep =
4197                                 &ib_event->param.sidr_rep_rcvd;
4198         int ret;
4199
4200         mutex_lock(&id_priv->handler_mutex);
4201         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4202                 goto out;
4203
4204         switch (ib_event->event) {
4205         case IB_CM_SIDR_REQ_ERROR:
4206                 event.event = RDMA_CM_EVENT_UNREACHABLE;
4207                 event.status = -ETIMEDOUT;
4208                 break;
4209         case IB_CM_SIDR_REP_RECEIVED:
4210                 event.param.ud.private_data = ib_event->private_data;
4211                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
4212                 if (rep->status != IB_SIDR_SUCCESS) {
4213                         event.event = RDMA_CM_EVENT_UNREACHABLE;
4214                         event.status = ib_event->param.sidr_rep_rcvd.status;
4215                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
4216                                              event.status);
4217                         break;
4218                 }
4219                 ret = cma_set_qkey(id_priv, rep->qkey);
4220                 if (ret) {
4221                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
4222                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
4223                         event.status = ret;
4224                         break;
4225                 }
4226                 ib_init_ah_attr_from_path(id_priv->id.device,
4227                                           id_priv->id.port_num,
4228                                           id_priv->id.route.path_rec,
4229                                           &event.param.ud.ah_attr,
4230                                           rep->sgid_attr);
4231                 event.param.ud.qp_num = rep->qpn;
4232                 event.param.ud.qkey = rep->qkey;
4233                 event.event = RDMA_CM_EVENT_ESTABLISHED;
4234                 event.status = 0;
4235                 break;
4236         default:
4237                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4238                        ib_event->event);
4239                 goto out;
4240         }
4241
4242         ret = cma_cm_event_handler(id_priv, &event);
4243
4244         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4245         if (ret) {
4246                 /* Destroy the CM ID by returning a non-zero value. */
4247                 id_priv->cm_id.ib = NULL;
4248                 destroy_id_handler_unlock(id_priv);
4249                 return ret;
4250         }
4251 out:
4252         mutex_unlock(&id_priv->handler_mutex);
4253         return 0;
4254 }
4255
4256 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4257                               struct rdma_conn_param *conn_param)
4258 {
4259         struct ib_cm_sidr_req_param req;
4260         struct ib_cm_id *id;
4261         void *private_data;
4262         u8 offset;
4263         int ret;
4264
4265         memset(&req, 0, sizeof req);
4266         offset = cma_user_data_offset(id_priv);
4267         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4268                 return -EINVAL;
4269
4270         if (req.private_data_len) {
4271                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4272                 if (!private_data)
4273                         return -ENOMEM;
4274         } else {
4275                 private_data = NULL;
4276         }
4277
4278         if (conn_param->private_data && conn_param->private_data_len)
4279                 memcpy(private_data + offset, conn_param->private_data,
4280                        conn_param->private_data_len);
4281
4282         if (private_data) {
4283                 ret = cma_format_hdr(private_data, id_priv);
4284                 if (ret)
4285                         goto out;
4286                 req.private_data = private_data;
4287         }
4288
4289         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4290                              id_priv);
4291         if (IS_ERR(id)) {
4292                 ret = PTR_ERR(id);
4293                 goto out;
4294         }
4295         id_priv->cm_id.ib = id;
4296
4297         req.path = id_priv->id.route.path_rec;
4298         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4299         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4300         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4301         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4302
4303         trace_cm_send_sidr_req(id_priv);
4304         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4305         if (ret) {
4306                 ib_destroy_cm_id(id_priv->cm_id.ib);
4307                 id_priv->cm_id.ib = NULL;
4308         }
4309 out:
4310         kfree(private_data);
4311         return ret;
4312 }
4313
4314 static int cma_connect_ib(struct rdma_id_private *id_priv,
4315                           struct rdma_conn_param *conn_param)
4316 {
4317         struct ib_cm_req_param req;
4318         struct rdma_route *route;
4319         void *private_data;
4320         struct ib_cm_id *id;
4321         u8 offset;
4322         int ret;
4323
4324         memset(&req, 0, sizeof req);
4325         offset = cma_user_data_offset(id_priv);
4326         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4327                 return -EINVAL;
4328
4329         if (req.private_data_len) {
4330                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4331                 if (!private_data)
4332                         return -ENOMEM;
4333         } else {
4334                 private_data = NULL;
4335         }
4336
4337         if (conn_param->private_data && conn_param->private_data_len)
4338                 memcpy(private_data + offset, conn_param->private_data,
4339                        conn_param->private_data_len);
4340
4341         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4342         if (IS_ERR(id)) {
4343                 ret = PTR_ERR(id);
4344                 goto out;
4345         }
4346         id_priv->cm_id.ib = id;
4347
4348         route = &id_priv->id.route;
4349         if (private_data) {
4350                 ret = cma_format_hdr(private_data, id_priv);
4351                 if (ret)
4352                         goto out;
4353                 req.private_data = private_data;
4354         }
4355
4356         req.primary_path = &route->path_rec[0];
4357         req.primary_path_inbound = route->path_rec_inbound;
4358         req.primary_path_outbound = route->path_rec_outbound;
4359         if (route->num_pri_alt_paths == 2)
4360                 req.alternate_path = &route->path_rec[1];
4361
4362         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4363         /* Alternate path SGID attribute currently unsupported */
4364         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4365         req.qp_num = id_priv->qp_num;
4366         req.qp_type = id_priv->id.qp_type;
4367         req.starting_psn = id_priv->seq_num;
4368         req.responder_resources = conn_param->responder_resources;
4369         req.initiator_depth = conn_param->initiator_depth;
4370         req.flow_control = conn_param->flow_control;
4371         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4372         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4373         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4374         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4375         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4376         req.srq = id_priv->srq ? 1 : 0;
4377         req.ece.vendor_id = id_priv->ece.vendor_id;
4378         req.ece.attr_mod = id_priv->ece.attr_mod;
4379
4380         trace_cm_send_req(id_priv);
4381         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4382 out:
4383         if (ret && !IS_ERR(id)) {
4384                 ib_destroy_cm_id(id);
4385                 id_priv->cm_id.ib = NULL;
4386         }
4387
4388         kfree(private_data);
4389         return ret;
4390 }
4391
4392 static int cma_connect_iw(struct rdma_id_private *id_priv,
4393                           struct rdma_conn_param *conn_param)
4394 {
4395         struct iw_cm_id *cm_id;
4396         int ret;
4397         struct iw_cm_conn_param iw_param;
4398
4399         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4400         if (IS_ERR(cm_id))
4401                 return PTR_ERR(cm_id);
4402
4403         mutex_lock(&id_priv->qp_mutex);
4404         cm_id->tos = id_priv->tos;
4405         cm_id->tos_set = id_priv->tos_set;
4406         mutex_unlock(&id_priv->qp_mutex);
4407
4408         id_priv->cm_id.iw = cm_id;
4409
4410         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4411                rdma_addr_size(cma_src_addr(id_priv)));
4412         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4413                rdma_addr_size(cma_dst_addr(id_priv)));
4414
4415         ret = cma_modify_qp_rtr(id_priv, conn_param);
4416         if (ret)
4417                 goto out;
4418
4419         if (conn_param) {
4420                 iw_param.ord = conn_param->initiator_depth;
4421                 iw_param.ird = conn_param->responder_resources;
4422                 iw_param.private_data = conn_param->private_data;
4423                 iw_param.private_data_len = conn_param->private_data_len;
4424                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4425         } else {
4426                 memset(&iw_param, 0, sizeof iw_param);
4427                 iw_param.qpn = id_priv->qp_num;
4428         }
4429         ret = iw_cm_connect(cm_id, &iw_param);
4430 out:
4431         if (ret) {
4432                 iw_destroy_cm_id(cm_id);
4433                 id_priv->cm_id.iw = NULL;
4434         }
4435         return ret;
4436 }
4437
4438 /**
4439  * rdma_connect_locked - Initiate an active connection request.
4440  * @id: Connection identifier to connect.
4441  * @conn_param: Connection information used for connected QPs.
4442  *
4443  * Same as rdma_connect() but can only be called from the
4444  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4445  */
4446 int rdma_connect_locked(struct rdma_cm_id *id,
4447                         struct rdma_conn_param *conn_param)
4448 {
4449         struct rdma_id_private *id_priv =
4450                 container_of(id, struct rdma_id_private, id);
4451         int ret;
4452
4453         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4454                 return -EINVAL;
4455
4456         if (!id->qp) {
4457                 id_priv->qp_num = conn_param->qp_num;
4458                 id_priv->srq = conn_param->srq;
4459         }
4460
4461         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4462                 if (id->qp_type == IB_QPT_UD)
4463                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4464                 else
4465                         ret = cma_connect_ib(id_priv, conn_param);
4466         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4467                 ret = cma_connect_iw(id_priv, conn_param);
4468         } else {
4469                 ret = -ENOSYS;
4470         }
4471         if (ret)
4472                 goto err_state;
4473         return 0;
4474 err_state:
4475         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4476         return ret;
4477 }
4478 EXPORT_SYMBOL(rdma_connect_locked);
4479
4480 /**
4481  * rdma_connect - Initiate an active connection request.
4482  * @id: Connection identifier to connect.
4483  * @conn_param: Connection information used for connected QPs.
4484  *
4485  * Users must have resolved a route for the rdma_cm_id to connect with by having
4486  * called rdma_resolve_route before calling this routine.
4487  *
4488  * This call will either connect to a remote QP or obtain remote QP information
4489  * for unconnected rdma_cm_id's.  The actual operation is based on the
4490  * rdma_cm_id's port space.
4491  */
4492 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4493 {
4494         struct rdma_id_private *id_priv =
4495                 container_of(id, struct rdma_id_private, id);
4496         int ret;
4497
4498         mutex_lock(&id_priv->handler_mutex);
4499         ret = rdma_connect_locked(id, conn_param);
4500         mutex_unlock(&id_priv->handler_mutex);
4501         return ret;
4502 }
4503 EXPORT_SYMBOL(rdma_connect);
4504
4505 /**
4506  * rdma_connect_ece - Initiate an active connection request with ECE data.
4507  * @id: Connection identifier to connect.
4508  * @conn_param: Connection information used for connected QPs.
4509  * @ece: ECE parameters
4510  *
4511  * See rdma_connect() explanation.
4512  */
4513 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4514                      struct rdma_ucm_ece *ece)
4515 {
4516         struct rdma_id_private *id_priv =
4517                 container_of(id, struct rdma_id_private, id);
4518
4519         id_priv->ece.vendor_id = ece->vendor_id;
4520         id_priv->ece.attr_mod = ece->attr_mod;
4521
4522         return rdma_connect(id, conn_param);
4523 }
4524 EXPORT_SYMBOL(rdma_connect_ece);
4525
4526 static int cma_accept_ib(struct rdma_id_private *id_priv,
4527                          struct rdma_conn_param *conn_param)
4528 {
4529         struct ib_cm_rep_param rep;
4530         int ret;
4531
4532         ret = cma_modify_qp_rtr(id_priv, conn_param);
4533         if (ret)
4534                 goto out;
4535
4536         ret = cma_modify_qp_rts(id_priv, conn_param);
4537         if (ret)
4538                 goto out;
4539
4540         memset(&rep, 0, sizeof rep);
4541         rep.qp_num = id_priv->qp_num;
4542         rep.starting_psn = id_priv->seq_num;
4543         rep.private_data = conn_param->private_data;
4544         rep.private_data_len = conn_param->private_data_len;
4545         rep.responder_resources = conn_param->responder_resources;
4546         rep.initiator_depth = conn_param->initiator_depth;
4547         rep.failover_accepted = 0;
4548         rep.flow_control = conn_param->flow_control;
4549         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4550         rep.srq = id_priv->srq ? 1 : 0;
4551         rep.ece.vendor_id = id_priv->ece.vendor_id;
4552         rep.ece.attr_mod = id_priv->ece.attr_mod;
4553
4554         trace_cm_send_rep(id_priv);
4555         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4556 out:
4557         return ret;
4558 }
4559
4560 static int cma_accept_iw(struct rdma_id_private *id_priv,
4561                   struct rdma_conn_param *conn_param)
4562 {
4563         struct iw_cm_conn_param iw_param;
4564         int ret;
4565
4566         if (!conn_param)
4567                 return -EINVAL;
4568
4569         ret = cma_modify_qp_rtr(id_priv, conn_param);
4570         if (ret)
4571                 return ret;
4572
4573         iw_param.ord = conn_param->initiator_depth;
4574         iw_param.ird = conn_param->responder_resources;
4575         iw_param.private_data = conn_param->private_data;
4576         iw_param.private_data_len = conn_param->private_data_len;
4577         if (id_priv->id.qp)
4578                 iw_param.qpn = id_priv->qp_num;
4579         else
4580                 iw_param.qpn = conn_param->qp_num;
4581
4582         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4583 }
4584
4585 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4586                              enum ib_cm_sidr_status status, u32 qkey,
4587                              const void *private_data, int private_data_len)
4588 {
4589         struct ib_cm_sidr_rep_param rep;
4590         int ret;
4591
4592         memset(&rep, 0, sizeof rep);
4593         rep.status = status;
4594         if (status == IB_SIDR_SUCCESS) {
4595                 if (qkey)
4596                         ret = cma_set_qkey(id_priv, qkey);
4597                 else
4598                         ret = cma_set_default_qkey(id_priv);
4599                 if (ret)
4600                         return ret;
4601                 rep.qp_num = id_priv->qp_num;
4602                 rep.qkey = id_priv->qkey;
4603
4604                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4605                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4606         }
4607
4608         rep.private_data = private_data;
4609         rep.private_data_len = private_data_len;
4610
4611         trace_cm_send_sidr_rep(id_priv);
4612         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4613 }
4614
4615 /**
4616  * rdma_accept - Called to accept a connection request or response.
4617  * @id: Connection identifier associated with the request.
4618  * @conn_param: Information needed to establish the connection.  This must be
4619  *   provided if accepting a connection request.  If accepting a connection
4620  *   response, this parameter must be NULL.
4621  *
4622  * Typically, this routine is only called by the listener to accept a connection
4623  * request.  It must also be called on the active side of a connection if the
4624  * user is performing their own QP transitions.
4625  *
4626  * In the case of error, a reject message is sent to the remote side and the
4627  * state of the qp associated with the id is modified to error, such that any
4628  * previously posted receive buffers would be flushed.
4629  *
4630  * This function is for use by kernel ULPs and must be called from under the
4631  * handler callback.
4632  */
4633 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4634 {
4635         struct rdma_id_private *id_priv =
4636                 container_of(id, struct rdma_id_private, id);
4637         int ret;
4638
4639         lockdep_assert_held(&id_priv->handler_mutex);
4640
4641         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4642                 return -EINVAL;
4643
4644         if (!id->qp && conn_param) {
4645                 id_priv->qp_num = conn_param->qp_num;
4646                 id_priv->srq = conn_param->srq;
4647         }
4648
4649         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4650                 if (id->qp_type == IB_QPT_UD) {
4651                         if (conn_param)
4652                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4653                                                         conn_param->qkey,
4654                                                         conn_param->private_data,
4655                                                         conn_param->private_data_len);
4656                         else
4657                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4658                                                         0, NULL, 0);
4659                 } else {
4660                         if (conn_param)
4661                                 ret = cma_accept_ib(id_priv, conn_param);
4662                         else
4663                                 ret = cma_rep_recv(id_priv);
4664                 }
4665         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4666                 ret = cma_accept_iw(id_priv, conn_param);
4667         } else {
4668                 ret = -ENOSYS;
4669         }
4670         if (ret)
4671                 goto reject;
4672
4673         return 0;
4674 reject:
4675         cma_modify_qp_err(id_priv);
4676         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4677         return ret;
4678 }
4679 EXPORT_SYMBOL(rdma_accept);
4680
4681 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4682                     struct rdma_ucm_ece *ece)
4683 {
4684         struct rdma_id_private *id_priv =
4685                 container_of(id, struct rdma_id_private, id);
4686
4687         id_priv->ece.vendor_id = ece->vendor_id;
4688         id_priv->ece.attr_mod = ece->attr_mod;
4689
4690         return rdma_accept(id, conn_param);
4691 }
4692 EXPORT_SYMBOL(rdma_accept_ece);
4693
4694 void rdma_lock_handler(struct rdma_cm_id *id)
4695 {
4696         struct rdma_id_private *id_priv =
4697                 container_of(id, struct rdma_id_private, id);
4698
4699         mutex_lock(&id_priv->handler_mutex);
4700 }
4701 EXPORT_SYMBOL(rdma_lock_handler);
4702
4703 void rdma_unlock_handler(struct rdma_cm_id *id)
4704 {
4705         struct rdma_id_private *id_priv =
4706                 container_of(id, struct rdma_id_private, id);
4707
4708         mutex_unlock(&id_priv->handler_mutex);
4709 }
4710 EXPORT_SYMBOL(rdma_unlock_handler);
4711
4712 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4713 {
4714         struct rdma_id_private *id_priv;
4715         int ret;
4716
4717         id_priv = container_of(id, struct rdma_id_private, id);
4718         if (!id_priv->cm_id.ib)
4719                 return -EINVAL;
4720
4721         switch (id->device->node_type) {
4722         case RDMA_NODE_IB_CA:
4723                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4724                 break;
4725         default:
4726                 ret = 0;
4727                 break;
4728         }
4729         return ret;
4730 }
4731 EXPORT_SYMBOL(rdma_notify);
4732
4733 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4734                 u8 private_data_len, u8 reason)
4735 {
4736         struct rdma_id_private *id_priv;
4737         int ret;
4738
4739         id_priv = container_of(id, struct rdma_id_private, id);
4740         if (!id_priv->cm_id.ib)
4741                 return -EINVAL;
4742
4743         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4744                 if (id->qp_type == IB_QPT_UD) {
4745                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4746                                                 private_data, private_data_len);
4747                 } else {
4748                         trace_cm_send_rej(id_priv);
4749                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4750                                              private_data, private_data_len);
4751                 }
4752         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4753                 ret = iw_cm_reject(id_priv->cm_id.iw,
4754                                    private_data, private_data_len);
4755         } else {
4756                 ret = -ENOSYS;
4757         }
4758
4759         return ret;
4760 }
4761 EXPORT_SYMBOL(rdma_reject);
4762
4763 int rdma_disconnect(struct rdma_cm_id *id)
4764 {
4765         struct rdma_id_private *id_priv;
4766         int ret;
4767
4768         id_priv = container_of(id, struct rdma_id_private, id);
4769         if (!id_priv->cm_id.ib)
4770                 return -EINVAL;
4771
4772         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4773                 ret = cma_modify_qp_err(id_priv);
4774                 if (ret)
4775                         goto out;
4776                 /* Initiate or respond to a disconnect. */
4777                 trace_cm_disconnect(id_priv);
4778                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4779                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4780                                 trace_cm_sent_drep(id_priv);
4781                 } else {
4782                         trace_cm_sent_dreq(id_priv);
4783                 }
4784         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4785                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4786         } else
4787                 ret = -EINVAL;
4788
4789 out:
4790         return ret;
4791 }
4792 EXPORT_SYMBOL(rdma_disconnect);
4793
4794 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4795                               struct ib_sa_multicast *multicast,
4796                               struct rdma_cm_event *event,
4797                               struct cma_multicast *mc)
4798 {
4799         struct rdma_dev_addr *dev_addr;
4800         enum ib_gid_type gid_type;
4801         struct net_device *ndev;
4802
4803         if (status)
4804                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4805                                      status);
4806
4807         event->status = status;
4808         event->param.ud.private_data = mc->context;
4809         if (status) {
4810                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4811                 return;
4812         }
4813
4814         dev_addr = &id_priv->id.route.addr.dev_addr;
4815         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4816         gid_type =
4817                 id_priv->cma_dev
4818                         ->default_gid_type[id_priv->id.port_num -
4819                                            rdma_start_port(
4820                                                    id_priv->cma_dev->device)];
4821
4822         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4823         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4824                                      &multicast->rec, ndev, gid_type,
4825                                      &event->param.ud.ah_attr)) {
4826                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4827                 goto out;
4828         }
4829
4830         event->param.ud.qp_num = 0xFFFFFF;
4831         event->param.ud.qkey = id_priv->qkey;
4832
4833 out:
4834         dev_put(ndev);
4835 }
4836
4837 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4838 {
4839         struct cma_multicast *mc = multicast->context;
4840         struct rdma_id_private *id_priv = mc->id_priv;
4841         struct rdma_cm_event event = {};
4842         int ret = 0;
4843
4844         mutex_lock(&id_priv->handler_mutex);
4845         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4846             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4847                 goto out;
4848
4849         ret = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4850         if (!ret) {
4851                 cma_make_mc_event(status, id_priv, multicast, &event, mc);
4852                 ret = cma_cm_event_handler(id_priv, &event);
4853         }
4854         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4855         WARN_ON(ret);
4856
4857 out:
4858         mutex_unlock(&id_priv->handler_mutex);
4859         return 0;
4860 }
4861
4862 static void cma_set_mgid(struct rdma_id_private *id_priv,
4863                          struct sockaddr *addr, union ib_gid *mgid)
4864 {
4865         unsigned char mc_map[MAX_ADDR_LEN];
4866         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4867         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4868         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4869
4870         if (cma_any_addr(addr)) {
4871                 memset(mgid, 0, sizeof *mgid);
4872         } else if ((addr->sa_family == AF_INET6) &&
4873                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4874                                                                  0xFF10A01B)) {
4875                 /* IPv6 address is an SA assigned MGID. */
4876                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4877         } else if (addr->sa_family == AF_IB) {
4878                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4879         } else if (addr->sa_family == AF_INET6) {
4880                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4881                 if (id_priv->id.ps == RDMA_PS_UDP)
4882                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4883                 *mgid = *(union ib_gid *) (mc_map + 4);
4884         } else {
4885                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4886                 if (id_priv->id.ps == RDMA_PS_UDP)
4887                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4888                 *mgid = *(union ib_gid *) (mc_map + 4);
4889         }
4890 }
4891
4892 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4893                                  struct cma_multicast *mc)
4894 {
4895         struct ib_sa_mcmember_rec rec;
4896         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4897         ib_sa_comp_mask comp_mask;
4898         int ret;
4899
4900         ib_addr_get_mgid(dev_addr, &rec.mgid);
4901         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4902                                      &rec.mgid, &rec);
4903         if (ret)
4904                 return ret;
4905
4906         if (!id_priv->qkey) {
4907                 ret = cma_set_default_qkey(id_priv);
4908                 if (ret)
4909                         return ret;
4910         }
4911
4912         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4913         rec.qkey = cpu_to_be32(id_priv->qkey);
4914         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4915         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4916         rec.join_state = mc->join_state;
4917
4918         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4919                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4920                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4921                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4922                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4923
4924         if (id_priv->id.ps == RDMA_PS_IPOIB)
4925                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4926                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4927                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4928                              IB_SA_MCMEMBER_REC_MTU |
4929                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4930
4931         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4932                                          id_priv->id.port_num, &rec, comp_mask,
4933                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4934         return PTR_ERR_OR_ZERO(mc->sa_mc);
4935 }
4936
4937 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4938                               enum ib_gid_type gid_type)
4939 {
4940         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4941         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4942
4943         if (cma_any_addr(addr)) {
4944                 memset(mgid, 0, sizeof *mgid);
4945         } else if (addr->sa_family == AF_INET6) {
4946                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4947         } else {
4948                 mgid->raw[0] =
4949                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4950                 mgid->raw[1] =
4951                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4952                 mgid->raw[2] = 0;
4953                 mgid->raw[3] = 0;
4954                 mgid->raw[4] = 0;
4955                 mgid->raw[5] = 0;
4956                 mgid->raw[6] = 0;
4957                 mgid->raw[7] = 0;
4958                 mgid->raw[8] = 0;
4959                 mgid->raw[9] = 0;
4960                 mgid->raw[10] = 0xff;
4961                 mgid->raw[11] = 0xff;
4962                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4963         }
4964 }
4965
4966 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4967                                    struct cma_multicast *mc)
4968 {
4969         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4970         int err = 0;
4971         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4972         struct net_device *ndev = NULL;
4973         struct ib_sa_multicast ib = {};
4974         enum ib_gid_type gid_type;
4975         bool send_only;
4976
4977         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4978
4979         if (cma_zero_addr(addr))
4980                 return -EINVAL;
4981
4982         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4983                    rdma_start_port(id_priv->cma_dev->device)];
4984         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4985
4986         ib.rec.pkey = cpu_to_be16(0xffff);
4987         if (dev_addr->bound_dev_if)
4988                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4989         if (!ndev)
4990                 return -ENODEV;
4991
4992         ib.rec.rate = IB_RATE_PORT_CURRENT;
4993         ib.rec.hop_limit = 1;
4994         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4995
4996         if (addr->sa_family == AF_INET) {
4997                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4998                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4999                         if (!send_only) {
5000                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
5001                                                     true);
5002                         }
5003                 }
5004         } else {
5005                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
5006                         err = -ENOTSUPP;
5007         }
5008         dev_put(ndev);
5009         if (err || !ib.rec.mtu)
5010                 return err ?: -EINVAL;
5011
5012         if (!id_priv->qkey)
5013                 cma_set_default_qkey(id_priv);
5014
5015         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
5016                     &ib.rec.port_gid);
5017         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
5018         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
5019         queue_work(cma_wq, &mc->iboe_join.work);
5020         return 0;
5021 }
5022
5023 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
5024                         u8 join_state, void *context)
5025 {
5026         struct rdma_id_private *id_priv =
5027                 container_of(id, struct rdma_id_private, id);
5028         struct cma_multicast *mc;
5029         int ret;
5030
5031         /* Not supported for kernel QPs */
5032         if (WARN_ON(id->qp))
5033                 return -EINVAL;
5034
5035         /* ULP is calling this wrong. */
5036         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
5037                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
5038                 return -EINVAL;
5039
5040         if (id_priv->id.qp_type != IB_QPT_UD)
5041                 return -EINVAL;
5042
5043         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
5044         if (!mc)
5045                 return -ENOMEM;
5046
5047         memcpy(&mc->addr, addr, rdma_addr_size(addr));
5048         mc->context = context;
5049         mc->id_priv = id_priv;
5050         mc->join_state = join_state;
5051
5052         if (rdma_protocol_roce(id->device, id->port_num)) {
5053                 ret = cma_iboe_join_multicast(id_priv, mc);
5054                 if (ret)
5055                         goto out_err;
5056         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
5057                 ret = cma_join_ib_multicast(id_priv, mc);
5058                 if (ret)
5059                         goto out_err;
5060         } else {
5061                 ret = -ENOSYS;
5062                 goto out_err;
5063         }
5064
5065         spin_lock(&id_priv->lock);
5066         list_add(&mc->list, &id_priv->mc_list);
5067         spin_unlock(&id_priv->lock);
5068
5069         return 0;
5070 out_err:
5071         kfree(mc);
5072         return ret;
5073 }
5074 EXPORT_SYMBOL(rdma_join_multicast);
5075
5076 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
5077 {
5078         struct rdma_id_private *id_priv;
5079         struct cma_multicast *mc;
5080
5081         id_priv = container_of(id, struct rdma_id_private, id);
5082         spin_lock_irq(&id_priv->lock);
5083         list_for_each_entry(mc, &id_priv->mc_list, list) {
5084                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
5085                         continue;
5086                 list_del(&mc->list);
5087                 spin_unlock_irq(&id_priv->lock);
5088
5089                 WARN_ON(id_priv->cma_dev->device != id->device);
5090                 destroy_mc(id_priv, mc);
5091                 return;
5092         }
5093         spin_unlock_irq(&id_priv->lock);
5094 }
5095 EXPORT_SYMBOL(rdma_leave_multicast);
5096
5097 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
5098 {
5099         struct rdma_dev_addr *dev_addr;
5100         struct cma_work *work;
5101
5102         dev_addr = &id_priv->id.route.addr.dev_addr;
5103
5104         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
5105             (net_eq(dev_net(ndev), dev_addr->net)) &&
5106             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
5107                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
5108                         ndev->name, &id_priv->id);
5109                 work = kzalloc(sizeof *work, GFP_KERNEL);
5110                 if (!work)
5111                         return -ENOMEM;
5112
5113                 INIT_WORK(&work->work, cma_work_handler);
5114                 work->id = id_priv;
5115                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
5116                 cma_id_get(id_priv);
5117                 queue_work(cma_wq, &work->work);
5118         }
5119
5120         return 0;
5121 }
5122
5123 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
5124                                void *ptr)
5125 {
5126         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
5127         struct cma_device *cma_dev;
5128         struct rdma_id_private *id_priv;
5129         int ret = NOTIFY_DONE;
5130
5131         if (event != NETDEV_BONDING_FAILOVER)
5132                 return NOTIFY_DONE;
5133
5134         if (!netif_is_bond_master(ndev))
5135                 return NOTIFY_DONE;
5136
5137         mutex_lock(&lock);
5138         list_for_each_entry(cma_dev, &dev_list, list)
5139                 list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
5140                         ret = cma_netdev_change(ndev, id_priv);
5141                         if (ret)
5142                                 goto out;
5143                 }
5144
5145 out:
5146         mutex_unlock(&lock);
5147         return ret;
5148 }
5149
5150 static void cma_netevent_work_handler(struct work_struct *_work)
5151 {
5152         struct rdma_id_private *id_priv =
5153                 container_of(_work, struct rdma_id_private, id.net_work);
5154         struct rdma_cm_event event = {};
5155
5156         mutex_lock(&id_priv->handler_mutex);
5157
5158         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
5159             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
5160                 goto out_unlock;
5161
5162         event.event = RDMA_CM_EVENT_UNREACHABLE;
5163         event.status = -ETIMEDOUT;
5164
5165         if (cma_cm_event_handler(id_priv, &event)) {
5166                 __acquire(&id_priv->handler_mutex);
5167                 id_priv->cm_id.ib = NULL;
5168                 cma_id_put(id_priv);
5169                 destroy_id_handler_unlock(id_priv);
5170                 return;
5171         }
5172
5173 out_unlock:
5174         mutex_unlock(&id_priv->handler_mutex);
5175         cma_id_put(id_priv);
5176 }
5177
5178 static int cma_netevent_callback(struct notifier_block *self,
5179                                  unsigned long event, void *ctx)
5180 {
5181         struct id_table_entry *ips_node = NULL;
5182         struct rdma_id_private *current_id;
5183         struct neighbour *neigh = ctx;
5184         unsigned long flags;
5185
5186         if (event != NETEVENT_NEIGH_UPDATE)
5187                 return NOTIFY_DONE;
5188
5189         spin_lock_irqsave(&id_table_lock, flags);
5190         if (neigh->tbl->family == AF_INET6) {
5191                 struct sockaddr_in6 neigh_sock_6;
5192
5193                 neigh_sock_6.sin6_family = AF_INET6;
5194                 neigh_sock_6.sin6_addr = *(struct in6_addr *)neigh->primary_key;
5195                 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5196                                              (struct sockaddr *)&neigh_sock_6);
5197         } else if (neigh->tbl->family == AF_INET) {
5198                 struct sockaddr_in neigh_sock_4;
5199
5200                 neigh_sock_4.sin_family = AF_INET;
5201                 neigh_sock_4.sin_addr.s_addr = *(__be32 *)(neigh->primary_key);
5202                 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5203                                              (struct sockaddr *)&neigh_sock_4);
5204         } else
5205                 goto out;
5206
5207         if (!ips_node)
5208                 goto out;
5209
5210         list_for_each_entry(current_id, &ips_node->id_list, id_list_entry) {
5211                 if (!memcmp(current_id->id.route.addr.dev_addr.dst_dev_addr,
5212                            neigh->ha, ETH_ALEN))
5213                         continue;
5214                 INIT_WORK(&current_id->id.net_work, cma_netevent_work_handler);
5215                 cma_id_get(current_id);
5216                 queue_work(cma_wq, &current_id->id.net_work);
5217         }
5218 out:
5219         spin_unlock_irqrestore(&id_table_lock, flags);
5220         return NOTIFY_DONE;
5221 }
5222
5223 static struct notifier_block cma_nb = {
5224         .notifier_call = cma_netdev_callback
5225 };
5226
5227 static struct notifier_block cma_netevent_cb = {
5228         .notifier_call = cma_netevent_callback
5229 };
5230
5231 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
5232 {
5233         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
5234         enum rdma_cm_state state;
5235         unsigned long flags;
5236
5237         mutex_lock(&id_priv->handler_mutex);
5238         /* Record that we want to remove the device */
5239         spin_lock_irqsave(&id_priv->lock, flags);
5240         state = id_priv->state;
5241         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
5242                 spin_unlock_irqrestore(&id_priv->lock, flags);
5243                 mutex_unlock(&id_priv->handler_mutex);
5244                 cma_id_put(id_priv);
5245                 return;
5246         }
5247         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
5248         spin_unlock_irqrestore(&id_priv->lock, flags);
5249
5250         if (cma_cm_event_handler(id_priv, &event)) {
5251                 /*
5252                  * At this point the ULP promises it won't call
5253                  * rdma_destroy_id() concurrently
5254                  */
5255                 cma_id_put(id_priv);
5256                 mutex_unlock(&id_priv->handler_mutex);
5257                 trace_cm_id_destroy(id_priv);
5258                 _destroy_id(id_priv, state);
5259                 return;
5260         }
5261         mutex_unlock(&id_priv->handler_mutex);
5262
5263         /*
5264          * If this races with destroy then the thread that first assigns state
5265          * to a destroying does the cancel.
5266          */
5267         cma_cancel_operation(id_priv, state);
5268         cma_id_put(id_priv);
5269 }
5270
5271 static void cma_process_remove(struct cma_device *cma_dev)
5272 {
5273         mutex_lock(&lock);
5274         while (!list_empty(&cma_dev->id_list)) {
5275                 struct rdma_id_private *id_priv = list_first_entry(
5276                         &cma_dev->id_list, struct rdma_id_private, device_item);
5277
5278                 list_del_init(&id_priv->listen_item);
5279                 list_del_init(&id_priv->device_item);
5280                 cma_id_get(id_priv);
5281                 mutex_unlock(&lock);
5282
5283                 cma_send_device_removal_put(id_priv);
5284
5285                 mutex_lock(&lock);
5286         }
5287         mutex_unlock(&lock);
5288
5289         cma_dev_put(cma_dev);
5290         wait_for_completion(&cma_dev->comp);
5291 }
5292
5293 static bool cma_supported(struct ib_device *device)
5294 {
5295         u32 i;
5296
5297         rdma_for_each_port(device, i) {
5298                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
5299                         return true;
5300         }
5301         return false;
5302 }
5303
5304 static int cma_add_one(struct ib_device *device)
5305 {
5306         struct rdma_id_private *to_destroy;
5307         struct cma_device *cma_dev;
5308         struct rdma_id_private *id_priv;
5309         unsigned long supported_gids = 0;
5310         int ret;
5311         u32 i;
5312
5313         if (!cma_supported(device))
5314                 return -EOPNOTSUPP;
5315
5316         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5317         if (!cma_dev)
5318                 return -ENOMEM;
5319
5320         cma_dev->device = device;
5321         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5322                                             sizeof(*cma_dev->default_gid_type),
5323                                             GFP_KERNEL);
5324         if (!cma_dev->default_gid_type) {
5325                 ret = -ENOMEM;
5326                 goto free_cma_dev;
5327         }
5328
5329         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5330                                             sizeof(*cma_dev->default_roce_tos),
5331                                             GFP_KERNEL);
5332         if (!cma_dev->default_roce_tos) {
5333                 ret = -ENOMEM;
5334                 goto free_gid_type;
5335         }
5336
5337         rdma_for_each_port (device, i) {
5338                 supported_gids = roce_gid_type_mask_support(device, i);
5339                 WARN_ON(!supported_gids);
5340                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5341                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5342                                 CMA_PREFERRED_ROCE_GID_TYPE;
5343                 else
5344                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5345                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5346                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5347         }
5348
5349         init_completion(&cma_dev->comp);
5350         refcount_set(&cma_dev->refcount, 1);
5351         INIT_LIST_HEAD(&cma_dev->id_list);
5352         ib_set_client_data(device, &cma_client, cma_dev);
5353
5354         mutex_lock(&lock);
5355         list_add_tail(&cma_dev->list, &dev_list);
5356         list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5357                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5358                 if (ret)
5359                         goto free_listen;
5360         }
5361         mutex_unlock(&lock);
5362
5363         trace_cm_add_one(device);
5364         return 0;
5365
5366 free_listen:
5367         list_del(&cma_dev->list);
5368         mutex_unlock(&lock);
5369
5370         /* cma_process_remove() will delete to_destroy */
5371         cma_process_remove(cma_dev);
5372         kfree(cma_dev->default_roce_tos);
5373 free_gid_type:
5374         kfree(cma_dev->default_gid_type);
5375
5376 free_cma_dev:
5377         kfree(cma_dev);
5378         return ret;
5379 }
5380
5381 static void cma_remove_one(struct ib_device *device, void *client_data)
5382 {
5383         struct cma_device *cma_dev = client_data;
5384
5385         trace_cm_remove_one(device);
5386
5387         mutex_lock(&lock);
5388         list_del(&cma_dev->list);
5389         mutex_unlock(&lock);
5390
5391         cma_process_remove(cma_dev);
5392         kfree(cma_dev->default_roce_tos);
5393         kfree(cma_dev->default_gid_type);
5394         kfree(cma_dev);
5395 }
5396
5397 static int cma_init_net(struct net *net)
5398 {
5399         struct cma_pernet *pernet = cma_pernet(net);
5400
5401         xa_init(&pernet->tcp_ps);
5402         xa_init(&pernet->udp_ps);
5403         xa_init(&pernet->ipoib_ps);
5404         xa_init(&pernet->ib_ps);
5405
5406         return 0;
5407 }
5408
5409 static void cma_exit_net(struct net *net)
5410 {
5411         struct cma_pernet *pernet = cma_pernet(net);
5412
5413         WARN_ON(!xa_empty(&pernet->tcp_ps));
5414         WARN_ON(!xa_empty(&pernet->udp_ps));
5415         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5416         WARN_ON(!xa_empty(&pernet->ib_ps));
5417 }
5418
5419 static struct pernet_operations cma_pernet_operations = {
5420         .init = cma_init_net,
5421         .exit = cma_exit_net,
5422         .id = &cma_pernet_id,
5423         .size = sizeof(struct cma_pernet),
5424 };
5425
5426 static int __init cma_init(void)
5427 {
5428         int ret;
5429
5430         /*
5431          * There is a rare lock ordering dependency in cma_netdev_callback()
5432          * that only happens when bonding is enabled. Teach lockdep that rtnl
5433          * must never be nested under lock so it can find these without having
5434          * to test with bonding.
5435          */
5436         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5437                 rtnl_lock();
5438                 mutex_lock(&lock);
5439                 mutex_unlock(&lock);
5440                 rtnl_unlock();
5441         }
5442
5443         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5444         if (!cma_wq)
5445                 return -ENOMEM;
5446
5447         ret = register_pernet_subsys(&cma_pernet_operations);
5448         if (ret)
5449                 goto err_wq;
5450
5451         ib_sa_register_client(&sa_client);
5452         register_netdevice_notifier(&cma_nb);
5453         register_netevent_notifier(&cma_netevent_cb);
5454
5455         ret = ib_register_client(&cma_client);
5456         if (ret)
5457                 goto err;
5458
5459         ret = cma_configfs_init();
5460         if (ret)
5461                 goto err_ib;
5462
5463         return 0;
5464
5465 err_ib:
5466         ib_unregister_client(&cma_client);
5467 err:
5468         unregister_netevent_notifier(&cma_netevent_cb);
5469         unregister_netdevice_notifier(&cma_nb);
5470         ib_sa_unregister_client(&sa_client);
5471         unregister_pernet_subsys(&cma_pernet_operations);
5472 err_wq:
5473         destroy_workqueue(cma_wq);
5474         return ret;
5475 }
5476
5477 static void __exit cma_cleanup(void)
5478 {
5479         cma_configfs_exit();
5480         ib_unregister_client(&cma_client);
5481         unregister_netevent_notifier(&cma_netevent_cb);
5482         unregister_netdevice_notifier(&cma_nb);
5483         ib_sa_unregister_client(&sa_client);
5484         unregister_pernet_subsys(&cma_pernet_operations);
5485         destroy_workqueue(cma_wq);
5486 }
5487
5488 module_init(cma_init);
5489 module_exit(cma_cleanup);