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         struct net_device *pdev = NULL;
 694
 695         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 696                 goto out;
 697
 698         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 699                 goto out;
 700
 701         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 702                 goto out;
 703
 704         /*
 705          * For drivers that do not associate more than one net device with
 706          * their gid tables, such as iWARP drivers, it is sufficient to
 707          * return the first table entry.
 708          *
 709          * Other driver classes might be included in the future.
 710          */
 711         if (rdma_protocol_iwarp(device, port)) {
 712                 sgid_attr = rdma_get_gid_attr(device, port, 0);
 713                 if (IS_ERR(sgid_attr))
 714                         goto out;
 715
 716                 rcu_read_lock();
 717                 ndev = rcu_dereference(sgid_attr->ndev);
 718                 if (ndev->ifindex != bound_if_index) {
 719                         pdev = dev_get_by_index_rcu(dev_addr->net, bound_if_index);
 720                         if (pdev) {
 721                                 if (is_vlan_dev(pdev)) {
 722                                         pdev = vlan_dev_real_dev(pdev);
 723                                         if (ndev->ifindex == pdev->ifindex)
 724                                                 bound_if_index = pdev->ifindex;
 725                                 }
 726                                 if (is_vlan_dev(ndev)) {
 727                                         pdev = vlan_dev_real_dev(ndev);
 728                                         if (bound_if_index == pdev->ifindex)
 729                                                 bound_if_index = ndev->ifindex;
 730                                 }
 731                         }
 732                 }
 733                 if (!net_eq(dev_net(ndev), dev_addr->net) ||
 734                     ndev->ifindex != bound_if_index) {
 735                         rdma_put_gid_attr(sgid_attr);
 736                         sgid_attr = ERR_PTR(-ENODEV);
 737                 }
 738                 rcu_read_unlock();
 739                 goto out;
 740         }
 741
 742         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 743                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 744                 if (!ndev)
 745                         goto out;
 746         } else {
 747                 gid_type = IB_GID_TYPE_IB;
 748         }
 749
 750         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 751         dev_put(ndev);
 752 out:
 753         return sgid_attr;
 754 }
 755
 756 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 757                                const struct ib_gid_attr *sgid_attr)
 758 {
 759         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 760         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 761 }
 762
 763 /**
 764  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 765  * based on source ip address.
 766  * @id_priv:    cm_id which should be bound to cma device
 767  *
 768  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 769  * based on source IP address. It returns 0 on success or error code otherwise.
 770  * It is applicable to active and passive side cm_id.
 771  */
 772 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 773 {
 774         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 775         const struct ib_gid_attr *sgid_attr;
 776         union ib_gid gid, iboe_gid, *gidp;
 777         struct cma_device *cma_dev;
 778         enum ib_gid_type gid_type;
 779         int ret = -ENODEV;
 780         u32 port;
 781
 782         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 783             id_priv->id.ps == RDMA_PS_IPOIB)
 784                 return -EINVAL;
 785
 786         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 787                     &iboe_gid);
 788
 789         memcpy(&gid, dev_addr->src_dev_addr +
 790                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 791
 792         mutex_lock(&lock);
 793         list_for_each_entry(cma_dev, &dev_list, list) {
 794                 rdma_for_each_port (cma_dev->device, port) {
 795                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 796                                &iboe_gid : &gid;
 797                         gid_type = cma_dev->default_gid_type[port - 1];
 798                         sgid_attr = cma_validate_port(cma_dev->device, port,
 799                                                       gid_type, gidp, id_priv);
 800                         if (!IS_ERR(sgid_attr)) {
 801                                 id_priv->id.port_num = port;
 802                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 803                                 cma_attach_to_dev(id_priv, cma_dev);
 804                                 ret = 0;
 805                                 goto out;
 806                         }
 807                 }
 808         }
 809 out:
 810         mutex_unlock(&lock);
 811         return ret;
 812 }
 813
 814 /**
 815  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 816  * @id_priv:            cm id to bind to cma device
 817  * @listen_id_priv:     listener cm id to match against
 818  * @req:                Pointer to req structure containaining incoming
 819  *                      request information
 820  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 821  * rdma device matches for listen_id and incoming request. It also verifies
 822  * that a GID table entry is present for the source address.
 823  * Returns 0 on success, or returns error code otherwise.
 824  */
 825 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 826                               const struct rdma_id_private *listen_id_priv,
 827                               struct cma_req_info *req)
 828 {
 829         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 830         const struct ib_gid_attr *sgid_attr;
 831         enum ib_gid_type gid_type;
 832         union ib_gid gid;
 833
 834         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 835             id_priv->id.ps == RDMA_PS_IPOIB)
 836                 return -EINVAL;
 837
 838         if (rdma_protocol_roce(req->device, req->port))
 839                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 840                             &gid);
 841         else
 842                 memcpy(&gid, dev_addr->src_dev_addr +
 843                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 844
 845         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 846         sgid_attr = cma_validate_port(req->device, req->port,
 847                                       gid_type, &gid, id_priv);
 848         if (IS_ERR(sgid_attr))
 849                 return PTR_ERR(sgid_attr);
 850
 851         id_priv->id.port_num = req->port;
 852         cma_bind_sgid_attr(id_priv, sgid_attr);
 853         /* Need to acquire lock to protect against reader
 854          * of cma_dev->id_list such as cma_netdev_callback() and
 855          * cma_process_remove().
 856          */
 857         mutex_lock(&lock);
 858         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 859         mutex_unlock(&lock);
 860         rdma_restrack_add(&id_priv->res);
 861         return 0;
 862 }
 863
 864 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 865                               const struct rdma_id_private *listen_id_priv)
 866 {
 867         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 868         const struct ib_gid_attr *sgid_attr;
 869         struct cma_device *cma_dev;
 870         enum ib_gid_type gid_type;
 871         int ret = -ENODEV;
 872         union ib_gid gid;
 873         u32 port;
 874
 875         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 876             id_priv->id.ps == RDMA_PS_IPOIB)
 877                 return -EINVAL;
 878
 879         memcpy(&gid, dev_addr->src_dev_addr +
 880                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 881
 882         mutex_lock(&lock);
 883
 884         cma_dev = listen_id_priv->cma_dev;
 885         port = listen_id_priv->id.port_num;
 886         gid_type = listen_id_priv->gid_type;
 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         list_for_each_entry(cma_dev, &dev_list, list) {
 897                 rdma_for_each_port (cma_dev->device, port) {
 898                         if (listen_id_priv->cma_dev == cma_dev &&
 899                             listen_id_priv->id.port_num == port)
 900                                 continue;
 901
 902                         gid_type = cma_dev->default_gid_type[port - 1];
 903                         sgid_attr = cma_validate_port(cma_dev->device, port,
 904                                                       gid_type, &gid, id_priv);
 905                         if (!IS_ERR(sgid_attr)) {
 906                                 id_priv->id.port_num = port;
 907                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 908                                 ret = 0;
 909                                 goto out;
 910                         }
 911                 }
 912         }
 913
 914 out:
 915         if (!ret) {
 916                 cma_attach_to_dev(id_priv, cma_dev);
 917                 rdma_restrack_add(&id_priv->res);
 918         }
 919
 920         mutex_unlock(&lock);
 921         return ret;
 922 }
 923
 924 /*
 925  * Select the source IB device and address to reach the destination IB address.
 926  */
 927 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 928 {
 929         struct cma_device *cma_dev, *cur_dev;
 930         struct sockaddr_ib *addr;
 931         union ib_gid gid, sgid, *dgid;
 932         unsigned int p;
 933         u16 pkey, index;
 934         enum ib_port_state port_state;
 935         int ret;
 936         int i;
 937
 938         cma_dev = NULL;
 939         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 940         dgid = (union ib_gid *) &addr->sib_addr;
 941         pkey = ntohs(addr->sib_pkey);
 942
 943         mutex_lock(&lock);
 944         list_for_each_entry(cur_dev, &dev_list, list) {
 945                 rdma_for_each_port (cur_dev->device, p) {
 946                         if (!rdma_cap_af_ib(cur_dev->device, p))
 947                                 continue;
 948
 949                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 950                                 continue;
 951
 952                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 953                                 continue;
 954
 955                         for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
 956                              ++i) {
 957                                 ret = rdma_query_gid(cur_dev->device, p, i,
 958                                                      &gid);
 959                                 if (ret)
 960                                         continue;
 961
 962                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 963                                         cma_dev = cur_dev;
 964                                         sgid = gid;
 965                                         id_priv->id.port_num = p;
 966                                         goto found;
 967                                 }
 968
 969                                 if (!cma_dev && (gid.global.subnet_prefix ==
 970                                     dgid->global.subnet_prefix) &&
 971                                     port_state == IB_PORT_ACTIVE) {
 972                                         cma_dev = cur_dev;
 973                                         sgid = gid;
 974                                         id_priv->id.port_num = p;
 975                                         goto found;
 976                                 }
 977                         }
 978                 }
 979         }
 980         mutex_unlock(&lock);
 981         return -ENODEV;
 982
 983 found:
 984         cma_attach_to_dev(id_priv, cma_dev);
 985         rdma_restrack_add(&id_priv->res);
 986         mutex_unlock(&lock);
 987         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 988         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 989         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 990         return 0;
 991 }
 992
 993 static void cma_id_get(struct rdma_id_private *id_priv)
 994 {
 995         refcount_inc(&id_priv->refcount);
 996 }
 997
 998 static void cma_id_put(struct rdma_id_private *id_priv)
 999 {
1000         if (refcount_dec_and_test(&id_priv->refcount))
1001                 complete(&id_priv->comp);
1002 }
1003
1004 static struct rdma_id_private *
1005 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
1006                  void *context, enum rdma_ucm_port_space ps,
1007                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
1008 {
1009         struct rdma_id_private *id_priv;
1010
1011         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
1012         if (!id_priv)
1013                 return ERR_PTR(-ENOMEM);
1014
1015         id_priv->state = RDMA_CM_IDLE;
1016         id_priv->id.context = context;
1017         id_priv->id.event_handler = event_handler;
1018         id_priv->id.ps = ps;
1019         id_priv->id.qp_type = qp_type;
1020         id_priv->tos_set = false;
1021         id_priv->timeout_set = false;
1022         id_priv->min_rnr_timer_set = false;
1023         id_priv->gid_type = IB_GID_TYPE_IB;
1024         spin_lock_init(&id_priv->lock);
1025         mutex_init(&id_priv->qp_mutex);
1026         init_completion(&id_priv->comp);
1027         refcount_set(&id_priv->refcount, 1);
1028         mutex_init(&id_priv->handler_mutex);
1029         INIT_LIST_HEAD(&id_priv->device_item);
1030         INIT_LIST_HEAD(&id_priv->id_list_entry);
1031         INIT_LIST_HEAD(&id_priv->listen_list);
1032         INIT_LIST_HEAD(&id_priv->mc_list);
1033         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
1034         id_priv->id.route.addr.dev_addr.net = get_net(net);
1035         id_priv->seq_num &= 0x00ffffff;
1036
1037         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
1038         if (parent)
1039                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
1040
1041         return id_priv;
1042 }
1043
1044 struct rdma_cm_id *
1045 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
1046                         void *context, enum rdma_ucm_port_space ps,
1047                         enum ib_qp_type qp_type, const char *caller)
1048 {
1049         struct rdma_id_private *ret;
1050
1051         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
1052         if (IS_ERR(ret))
1053                 return ERR_CAST(ret);
1054
1055         rdma_restrack_set_name(&ret->res, caller);
1056         return &ret->id;
1057 }
1058 EXPORT_SYMBOL(__rdma_create_kernel_id);
1059
1060 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
1061                                        void *context,
1062                                        enum rdma_ucm_port_space ps,
1063                                        enum ib_qp_type qp_type)
1064 {
1065         struct rdma_id_private *ret;
1066
1067         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
1068                                ps, qp_type, NULL);
1069         if (IS_ERR(ret))
1070                 return ERR_CAST(ret);
1071
1072         rdma_restrack_set_name(&ret->res, NULL);
1073         return &ret->id;
1074 }
1075 EXPORT_SYMBOL(rdma_create_user_id);
1076
1077 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1078 {
1079         struct ib_qp_attr qp_attr;
1080         int qp_attr_mask, ret;
1081
1082         qp_attr.qp_state = IB_QPS_INIT;
1083         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1084         if (ret)
1085                 return ret;
1086
1087         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1088         if (ret)
1089                 return ret;
1090
1091         qp_attr.qp_state = IB_QPS_RTR;
1092         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
1093         if (ret)
1094                 return ret;
1095
1096         qp_attr.qp_state = IB_QPS_RTS;
1097         qp_attr.sq_psn = 0;
1098         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
1099
1100         return ret;
1101 }
1102
1103 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1104 {
1105         struct ib_qp_attr qp_attr;
1106         int qp_attr_mask, ret;
1107
1108         qp_attr.qp_state = IB_QPS_INIT;
1109         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1110         if (ret)
1111                 return ret;
1112
1113         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1114 }
1115
1116 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
1117                    struct ib_qp_init_attr *qp_init_attr)
1118 {
1119         struct rdma_id_private *id_priv;
1120         struct ib_qp *qp;
1121         int ret;
1122
1123         id_priv = container_of(id, struct rdma_id_private, id);
1124         if (id->device != pd->device) {
1125                 ret = -EINVAL;
1126                 goto out_err;
1127         }
1128
1129         qp_init_attr->port_num = id->port_num;
1130         qp = ib_create_qp(pd, qp_init_attr);
1131         if (IS_ERR(qp)) {
1132                 ret = PTR_ERR(qp);
1133                 goto out_err;
1134         }
1135
1136         if (id->qp_type == IB_QPT_UD)
1137                 ret = cma_init_ud_qp(id_priv, qp);
1138         else
1139                 ret = cma_init_conn_qp(id_priv, qp);
1140         if (ret)
1141                 goto out_destroy;
1142
1143         id->qp = qp;
1144         id_priv->qp_num = qp->qp_num;
1145         id_priv->srq = (qp->srq != NULL);
1146         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
1147         return 0;
1148 out_destroy:
1149         ib_destroy_qp(qp);
1150 out_err:
1151         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
1152         return ret;
1153 }
1154 EXPORT_SYMBOL(rdma_create_qp);
1155
1156 void rdma_destroy_qp(struct rdma_cm_id *id)
1157 {
1158         struct rdma_id_private *id_priv;
1159
1160         id_priv = container_of(id, struct rdma_id_private, id);
1161         trace_cm_qp_destroy(id_priv);
1162         mutex_lock(&id_priv->qp_mutex);
1163         ib_destroy_qp(id_priv->id.qp);
1164         id_priv->id.qp = NULL;
1165         mutex_unlock(&id_priv->qp_mutex);
1166 }
1167 EXPORT_SYMBOL(rdma_destroy_qp);
1168
1169 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1170                              struct rdma_conn_param *conn_param)
1171 {
1172         struct ib_qp_attr qp_attr;
1173         int qp_attr_mask, ret;
1174
1175         mutex_lock(&id_priv->qp_mutex);
1176         if (!id_priv->id.qp) {
1177                 ret = 0;
1178                 goto out;
1179         }
1180
1181         /* Need to update QP attributes from default values. */
1182         qp_attr.qp_state = IB_QPS_INIT;
1183         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1184         if (ret)
1185                 goto out;
1186
1187         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1188         if (ret)
1189                 goto out;
1190
1191         qp_attr.qp_state = IB_QPS_RTR;
1192         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1193         if (ret)
1194                 goto out;
1195
1196         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1197
1198         if (conn_param)
1199                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1200         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1201 out:
1202         mutex_unlock(&id_priv->qp_mutex);
1203         return ret;
1204 }
1205
1206 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1207                              struct rdma_conn_param *conn_param)
1208 {
1209         struct ib_qp_attr qp_attr;
1210         int qp_attr_mask, ret;
1211
1212         mutex_lock(&id_priv->qp_mutex);
1213         if (!id_priv->id.qp) {
1214                 ret = 0;
1215                 goto out;
1216         }
1217
1218         qp_attr.qp_state = IB_QPS_RTS;
1219         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1220         if (ret)
1221                 goto out;
1222
1223         if (conn_param)
1224                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1225         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1226 out:
1227         mutex_unlock(&id_priv->qp_mutex);
1228         return ret;
1229 }
1230
1231 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1232 {
1233         struct ib_qp_attr qp_attr;
1234         int ret;
1235
1236         mutex_lock(&id_priv->qp_mutex);
1237         if (!id_priv->id.qp) {
1238                 ret = 0;
1239                 goto out;
1240         }
1241
1242         qp_attr.qp_state = IB_QPS_ERR;
1243         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1244 out:
1245         mutex_unlock(&id_priv->qp_mutex);
1246         return ret;
1247 }
1248
1249 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1250                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1251 {
1252         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1253         int ret;
1254         u16 pkey;
1255
1256         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1257                 pkey = 0xffff;
1258         else
1259                 pkey = ib_addr_get_pkey(dev_addr);
1260
1261         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1262                                   pkey, &qp_attr->pkey_index);
1263         if (ret)
1264                 return ret;
1265
1266         qp_attr->port_num = id_priv->id.port_num;
1267         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1268
1269         if (id_priv->id.qp_type == IB_QPT_UD) {
1270                 ret = cma_set_default_qkey(id_priv);
1271                 if (ret)
1272                         return ret;
1273
1274                 qp_attr->qkey = id_priv->qkey;
1275                 *qp_attr_mask |= IB_QP_QKEY;
1276         } else {
1277                 qp_attr->qp_access_flags = 0;
1278                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1279         }
1280         return 0;
1281 }
1282
1283 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1284                        int *qp_attr_mask)
1285 {
1286         struct rdma_id_private *id_priv;
1287         int ret = 0;
1288
1289         id_priv = container_of(id, struct rdma_id_private, id);
1290         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1291                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1292                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1293                 else
1294                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1295                                                  qp_attr_mask);
1296
1297                 if (qp_attr->qp_state == IB_QPS_RTR)
1298                         qp_attr->rq_psn = id_priv->seq_num;
1299         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1300                 if (!id_priv->cm_id.iw) {
1301                         qp_attr->qp_access_flags = 0;
1302                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1303                 } else
1304                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1305                                                  qp_attr_mask);
1306                 qp_attr->port_num = id_priv->id.port_num;
1307                 *qp_attr_mask |= IB_QP_PORT;
1308         } else {
1309                 ret = -ENOSYS;
1310         }
1311
1312         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1313                 qp_attr->timeout = id_priv->timeout;
1314
1315         if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1316                 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1317
1318         return ret;
1319 }
1320 EXPORT_SYMBOL(rdma_init_qp_attr);
1321
1322 static inline bool cma_zero_addr(const struct sockaddr *addr)
1323 {
1324         switch (addr->sa_family) {
1325         case AF_INET:
1326                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1327         case AF_INET6:
1328                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1329         case AF_IB:
1330                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1331         default:
1332                 return false;
1333         }
1334 }
1335
1336 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1337 {
1338         switch (addr->sa_family) {
1339         case AF_INET:
1340                 return ipv4_is_loopback(
1341                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1342         case AF_INET6:
1343                 return ipv6_addr_loopback(
1344                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1345         case AF_IB:
1346                 return ib_addr_loopback(
1347                         &((struct sockaddr_ib *)addr)->sib_addr);
1348         default:
1349                 return false;
1350         }
1351 }
1352
1353 static inline bool cma_any_addr(const struct sockaddr *addr)
1354 {
1355         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1356 }
1357
1358 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1359 {
1360         if (src->sa_family != dst->sa_family)
1361                 return -1;
1362
1363         switch (src->sa_family) {
1364         case AF_INET:
1365                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1366                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1367         case AF_INET6: {
1368                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1369                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1370                 bool link_local;
1371
1372                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1373                                           &dst_addr6->sin6_addr))
1374                         return 1;
1375                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1376                              IPV6_ADDR_LINKLOCAL;
1377                 /* Link local must match their scope_ids */
1378                 return link_local ? (src_addr6->sin6_scope_id !=
1379                                      dst_addr6->sin6_scope_id) :
1380                                     0;
1381         }
1382
1383         default:
1384                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1385                                    &((struct sockaddr_ib *) dst)->sib_addr);
1386         }
1387 }
1388
1389 static __be16 cma_port(const struct sockaddr *addr)
1390 {
1391         struct sockaddr_ib *sib;
1392
1393         switch (addr->sa_family) {
1394         case AF_INET:
1395                 return ((struct sockaddr_in *) addr)->sin_port;
1396         case AF_INET6:
1397                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1398         case AF_IB:
1399                 sib = (struct sockaddr_ib *) addr;
1400                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1401                                     be64_to_cpu(sib->sib_sid_mask)));
1402         default:
1403                 return 0;
1404         }
1405 }
1406
1407 static inline int cma_any_port(const struct sockaddr *addr)
1408 {
1409         return !cma_port(addr);
1410 }
1411
1412 static void cma_save_ib_info(struct sockaddr *src_addr,
1413                              struct sockaddr *dst_addr,
1414                              const struct rdma_cm_id *listen_id,
1415                              const struct sa_path_rec *path)
1416 {
1417         struct sockaddr_ib *listen_ib, *ib;
1418
1419         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1420         if (src_addr) {
1421                 ib = (struct sockaddr_ib *)src_addr;
1422                 ib->sib_family = AF_IB;
1423                 if (path) {
1424                         ib->sib_pkey = path->pkey;
1425                         ib->sib_flowinfo = path->flow_label;
1426                         memcpy(&ib->sib_addr, &path->sgid, 16);
1427                         ib->sib_sid = path->service_id;
1428                         ib->sib_scope_id = 0;
1429                 } else {
1430                         ib->sib_pkey = listen_ib->sib_pkey;
1431                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1432                         ib->sib_addr = listen_ib->sib_addr;
1433                         ib->sib_sid = listen_ib->sib_sid;
1434                         ib->sib_scope_id = listen_ib->sib_scope_id;
1435                 }
1436                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1437         }
1438         if (dst_addr) {
1439                 ib = (struct sockaddr_ib *)dst_addr;
1440                 ib->sib_family = AF_IB;
1441                 if (path) {
1442                         ib->sib_pkey = path->pkey;
1443                         ib->sib_flowinfo = path->flow_label;
1444                         memcpy(&ib->sib_addr, &path->dgid, 16);
1445                 }
1446         }
1447 }
1448
1449 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1450                               struct sockaddr_in *dst_addr,
1451                               struct cma_hdr *hdr,
1452                               __be16 local_port)
1453 {
1454         if (src_addr) {
1455                 *src_addr = (struct sockaddr_in) {
1456                         .sin_family = AF_INET,
1457                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1458                         .sin_port = local_port,
1459                 };
1460         }
1461
1462         if (dst_addr) {
1463                 *dst_addr = (struct sockaddr_in) {
1464                         .sin_family = AF_INET,
1465                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1466                         .sin_port = hdr->port,
1467                 };
1468         }
1469 }
1470
1471 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1472                               struct sockaddr_in6 *dst_addr,
1473                               struct cma_hdr *hdr,
1474                               __be16 local_port)
1475 {
1476         if (src_addr) {
1477                 *src_addr = (struct sockaddr_in6) {
1478                         .sin6_family = AF_INET6,
1479                         .sin6_addr = hdr->dst_addr.ip6,
1480                         .sin6_port = local_port,
1481                 };
1482         }
1483
1484         if (dst_addr) {
1485                 *dst_addr = (struct sockaddr_in6) {
1486                         .sin6_family = AF_INET6,
1487                         .sin6_addr = hdr->src_addr.ip6,
1488                         .sin6_port = hdr->port,
1489                 };
1490         }
1491 }
1492
1493 static u16 cma_port_from_service_id(__be64 service_id)
1494 {
1495         return (u16)be64_to_cpu(service_id);
1496 }
1497
1498 static int cma_save_ip_info(struct sockaddr *src_addr,
1499                             struct sockaddr *dst_addr,
1500                             const struct ib_cm_event *ib_event,
1501                             __be64 service_id)
1502 {
1503         struct cma_hdr *hdr;
1504         __be16 port;
1505
1506         hdr = ib_event->private_data;
1507         if (hdr->cma_version != CMA_VERSION)
1508                 return -EINVAL;
1509
1510         port = htons(cma_port_from_service_id(service_id));
1511
1512         switch (cma_get_ip_ver(hdr)) {
1513         case 4:
1514                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1515                                   (struct sockaddr_in *)dst_addr, hdr, port);
1516                 break;
1517         case 6:
1518                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1519                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1520                 break;
1521         default:
1522                 return -EAFNOSUPPORT;
1523         }
1524
1525         return 0;
1526 }
1527
1528 static int cma_save_net_info(struct sockaddr *src_addr,
1529                              struct sockaddr *dst_addr,
1530                              const struct rdma_cm_id *listen_id,
1531                              const struct ib_cm_event *ib_event,
1532                              sa_family_t sa_family, __be64 service_id)
1533 {
1534         if (sa_family == AF_IB) {
1535                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1536                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1537                                          ib_event->param.req_rcvd.primary_path);
1538                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1539                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1540                 return 0;
1541         }
1542
1543         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1544 }
1545
1546 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1547                              struct cma_req_info *req)
1548 {
1549         const struct ib_cm_req_event_param *req_param =
1550                 &ib_event->param.req_rcvd;
1551         const struct ib_cm_sidr_req_event_param *sidr_param =
1552                 &ib_event->param.sidr_req_rcvd;
1553
1554         switch (ib_event->event) {
1555         case IB_CM_REQ_RECEIVED:
1556                 req->device     = req_param->listen_id->device;
1557                 req->port       = req_param->port;
1558                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1559                        sizeof(req->local_gid));
1560                 req->has_gid    = true;
1561                 req->service_id = req_param->primary_path->service_id;
1562                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1563                 if (req->pkey != req_param->bth_pkey)
1564                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1565                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1566                                             req_param->bth_pkey, req->pkey);
1567                 break;
1568         case IB_CM_SIDR_REQ_RECEIVED:
1569                 req->device     = sidr_param->listen_id->device;
1570                 req->port       = sidr_param->port;
1571                 req->has_gid    = false;
1572                 req->service_id = sidr_param->service_id;
1573                 req->pkey       = sidr_param->pkey;
1574                 if (req->pkey != sidr_param->bth_pkey)
1575                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1576                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1577                                             sidr_param->bth_pkey, req->pkey);
1578                 break;
1579         default:
1580                 return -EINVAL;
1581         }
1582
1583         return 0;
1584 }
1585
1586 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1587                                   const struct sockaddr_in *dst_addr,
1588                                   const struct sockaddr_in *src_addr)
1589 {
1590         __be32 daddr = dst_addr->sin_addr.s_addr,
1591                saddr = src_addr->sin_addr.s_addr;
1592         struct fib_result res;
1593         struct flowi4 fl4;
1594         int err;
1595         bool ret;
1596
1597         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1598             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1599             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1600             ipv4_is_loopback(saddr))
1601                 return false;
1602
1603         memset(&fl4, 0, sizeof(fl4));
1604         fl4.flowi4_oif = net_dev->ifindex;
1605         fl4.daddr = daddr;
1606         fl4.saddr = saddr;
1607
1608         rcu_read_lock();
1609         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1610         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1611         rcu_read_unlock();
1612
1613         return ret;
1614 }
1615
1616 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1617                                   const struct sockaddr_in6 *dst_addr,
1618                                   const struct sockaddr_in6 *src_addr)
1619 {
1620 #if IS_ENABLED(CONFIG_IPV6)
1621         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1622                            IPV6_ADDR_LINKLOCAL;
1623         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1624                                          &src_addr->sin6_addr, net_dev->ifindex,
1625                                          NULL, strict);
1626         bool ret;
1627
1628         if (!rt)
1629                 return false;
1630
1631         ret = rt->rt6i_idev->dev == net_dev;
1632         ip6_rt_put(rt);
1633
1634         return ret;
1635 #else
1636         return false;
1637 #endif
1638 }
1639
1640 static bool validate_net_dev(struct net_device *net_dev,
1641                              const struct sockaddr *daddr,
1642                              const struct sockaddr *saddr)
1643 {
1644         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1645         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1646         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1647         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1648
1649         switch (daddr->sa_family) {
1650         case AF_INET:
1651                 return saddr->sa_family == AF_INET &&
1652                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1653
1654         case AF_INET6:
1655                 return saddr->sa_family == AF_INET6 &&
1656                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1657
1658         default:
1659                 return false;
1660         }
1661 }
1662
1663 static struct net_device *
1664 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1665 {
1666         const struct ib_gid_attr *sgid_attr = NULL;
1667         struct net_device *ndev;
1668
1669         if (ib_event->event == IB_CM_REQ_RECEIVED)
1670                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1671         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1672                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1673
1674         if (!sgid_attr)
1675                 return NULL;
1676
1677         rcu_read_lock();
1678         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1679         if (IS_ERR(ndev))
1680                 ndev = NULL;
1681         else
1682                 dev_hold(ndev);
1683         rcu_read_unlock();
1684         return ndev;
1685 }
1686
1687 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1688                                           struct cma_req_info *req)
1689 {
1690         struct sockaddr *listen_addr =
1691                         (struct sockaddr *)&req->listen_addr_storage;
1692         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1693         struct net_device *net_dev;
1694         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1695         int err;
1696
1697         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1698                                req->service_id);
1699         if (err)
1700                 return ERR_PTR(err);
1701
1702         if (rdma_protocol_roce(req->device, req->port))
1703                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1704         else
1705                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1706                                                    req->pkey,
1707                                                    gid, listen_addr);
1708         if (!net_dev)
1709                 return ERR_PTR(-ENODEV);
1710
1711         return net_dev;
1712 }
1713
1714 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1715 {
1716         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1717 }
1718
1719 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1720                                    const struct cma_hdr *hdr)
1721 {
1722         struct sockaddr *addr = cma_src_addr(id_priv);
1723         __be32 ip4_addr;
1724         struct in6_addr ip6_addr;
1725
1726         if (cma_any_addr(addr) && !id_priv->afonly)
1727                 return true;
1728
1729         switch (addr->sa_family) {
1730         case AF_INET:
1731                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1732                 if (cma_get_ip_ver(hdr) != 4)
1733                         return false;
1734                 if (!cma_any_addr(addr) &&
1735                     hdr->dst_addr.ip4.addr != ip4_addr)
1736                         return false;
1737                 break;
1738         case AF_INET6:
1739                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1740                 if (cma_get_ip_ver(hdr) != 6)
1741                         return false;
1742                 if (!cma_any_addr(addr) &&
1743                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1744                         return false;
1745                 break;
1746         case AF_IB:
1747                 return true;
1748         default:
1749                 return false;
1750         }
1751
1752         return true;
1753 }
1754
1755 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1756 {
1757         struct ib_device *device = id->device;
1758         const u32 port_num = id->port_num ?: rdma_start_port(device);
1759
1760         return rdma_protocol_roce(device, port_num);
1761 }
1762
1763 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1764 {
1765         const struct sockaddr *daddr =
1766                         (const struct sockaddr *)&req->listen_addr_storage;
1767         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1768
1769         /* Returns true if the req is for IPv6 link local */
1770         return (daddr->sa_family == AF_INET6 &&
1771                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1772 }
1773
1774 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1775                               const struct net_device *net_dev,
1776                               const struct cma_req_info *req)
1777 {
1778         const struct rdma_addr *addr = &id->route.addr;
1779
1780         if (!net_dev)
1781                 /* This request is an AF_IB request */
1782                 return (!id->port_num || id->port_num == req->port) &&
1783                        (addr->src_addr.ss_family == AF_IB);
1784
1785         /*
1786          * If the request is not for IPv6 link local, allow matching
1787          * request to any netdevice of the one or multiport rdma device.
1788          */
1789         if (!cma_is_req_ipv6_ll(req))
1790                 return true;
1791         /*
1792          * Net namespaces must match, and if the listner is listening
1793          * on a specific netdevice than netdevice must match as well.
1794          */
1795         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1796             (!!addr->dev_addr.bound_dev_if ==
1797              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1798                 return true;
1799         else
1800                 return false;
1801 }
1802
1803 static struct rdma_id_private *cma_find_listener(
1804                 const struct rdma_bind_list *bind_list,
1805                 const struct ib_cm_id *cm_id,
1806                 const struct ib_cm_event *ib_event,
1807                 const struct cma_req_info *req,
1808                 const struct net_device *net_dev)
1809 {
1810         struct rdma_id_private *id_priv, *id_priv_dev;
1811
1812         lockdep_assert_held(&lock);
1813
1814         if (!bind_list)
1815                 return ERR_PTR(-EINVAL);
1816
1817         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1818                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1819                         if (id_priv->id.device == cm_id->device &&
1820                             cma_match_net_dev(&id_priv->id, net_dev, req))
1821                                 return id_priv;
1822                         list_for_each_entry(id_priv_dev,
1823                                             &id_priv->listen_list,
1824                                             listen_item) {
1825                                 if (id_priv_dev->id.device == cm_id->device &&
1826                                     cma_match_net_dev(&id_priv_dev->id,
1827                                                       net_dev, req))
1828                                         return id_priv_dev;
1829                         }
1830                 }
1831         }
1832
1833         return ERR_PTR(-EINVAL);
1834 }
1835
1836 static struct rdma_id_private *
1837 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1838                      const struct ib_cm_event *ib_event,
1839                      struct cma_req_info *req,
1840                      struct net_device **net_dev)
1841 {
1842         struct rdma_bind_list *bind_list;
1843         struct rdma_id_private *id_priv;
1844         int err;
1845
1846         err = cma_save_req_info(ib_event, req);
1847         if (err)
1848                 return ERR_PTR(err);
1849
1850         *net_dev = cma_get_net_dev(ib_event, req);
1851         if (IS_ERR(*net_dev)) {
1852                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1853                         /* Assuming the protocol is AF_IB */
1854                         *net_dev = NULL;
1855                 } else {
1856                         return ERR_CAST(*net_dev);
1857                 }
1858         }
1859
1860         mutex_lock(&lock);
1861         /*
1862          * Net namespace might be getting deleted while route lookup,
1863          * cm_id lookup is in progress. Therefore, perform netdevice
1864          * validation, cm_id lookup under rcu lock.
1865          * RCU lock along with netdevice state check, synchronizes with
1866          * netdevice migrating to different net namespace and also avoids
1867          * case where net namespace doesn't get deleted while lookup is in
1868          * progress.
1869          * If the device state is not IFF_UP, its properties such as ifindex
1870          * and nd_net cannot be trusted to remain valid without rcu lock.
1871          * net/core/dev.c change_net_namespace() ensures to synchronize with
1872          * ongoing operations on net device after device is closed using
1873          * synchronize_net().
1874          */
1875         rcu_read_lock();
1876         if (*net_dev) {
1877                 /*
1878                  * If netdevice is down, it is likely that it is administratively
1879                  * down or it might be migrating to different namespace.
1880                  * In that case avoid further processing, as the net namespace
1881                  * or ifindex may change.
1882                  */
1883                 if (((*net_dev)->flags & IFF_UP) == 0) {
1884                         id_priv = ERR_PTR(-EHOSTUNREACH);
1885                         goto err;
1886                 }
1887
1888                 if (!validate_net_dev(*net_dev,
1889                                  (struct sockaddr *)&req->src_addr_storage,
1890                                  (struct sockaddr *)&req->listen_addr_storage)) {
1891                         id_priv = ERR_PTR(-EHOSTUNREACH);
1892                         goto err;
1893                 }
1894         }
1895
1896         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1897                                 rdma_ps_from_service_id(req->service_id),
1898                                 cma_port_from_service_id(req->service_id));
1899         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1900 err:
1901         rcu_read_unlock();
1902         mutex_unlock(&lock);
1903         if (IS_ERR(id_priv) && *net_dev) {
1904                 dev_put(*net_dev);
1905                 *net_dev = NULL;
1906         }
1907         return id_priv;
1908 }
1909
1910 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1911 {
1912         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1913 }
1914
1915 static void cma_cancel_route(struct rdma_id_private *id_priv)
1916 {
1917         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1918                 if (id_priv->query)
1919                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1920         }
1921 }
1922
1923 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1924 {
1925         struct rdma_id_private *dev_id_priv;
1926
1927         lockdep_assert_held(&lock);
1928
1929         /*
1930          * Remove from listen_any_list to prevent added devices from spawning
1931          * additional listen requests.
1932          */
1933         list_del_init(&id_priv->listen_any_item);
1934
1935         while (!list_empty(&id_priv->listen_list)) {
1936                 dev_id_priv =
1937                         list_first_entry(&id_priv->listen_list,
1938                                          struct rdma_id_private, listen_item);
1939                 /* sync with device removal to avoid duplicate destruction */
1940                 list_del_init(&dev_id_priv->device_item);
1941                 list_del_init(&dev_id_priv->listen_item);
1942                 mutex_unlock(&lock);
1943
1944                 rdma_destroy_id(&dev_id_priv->id);
1945                 mutex_lock(&lock);
1946         }
1947 }
1948
1949 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1950 {
1951         mutex_lock(&lock);
1952         _cma_cancel_listens(id_priv);
1953         mutex_unlock(&lock);
1954 }
1955
1956 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1957                                  enum rdma_cm_state state)
1958 {
1959         switch (state) {
1960         case RDMA_CM_ADDR_QUERY:
1961                 /*
1962                  * We can avoid doing the rdma_addr_cancel() based on state,
1963                  * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1964                  * Notice that the addr_handler work could still be exiting
1965                  * outside this state, however due to the interaction with the
1966                  * handler_mutex the work is guaranteed not to touch id_priv
1967                  * during exit.
1968                  */
1969                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1970                 break;
1971         case RDMA_CM_ROUTE_QUERY:
1972                 cma_cancel_route(id_priv);
1973                 break;
1974         case RDMA_CM_LISTEN:
1975                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1976                         cma_cancel_listens(id_priv);
1977                 break;
1978         default:
1979                 break;
1980         }
1981 }
1982
1983 static void cma_release_port(struct rdma_id_private *id_priv)
1984 {
1985         struct rdma_bind_list *bind_list = id_priv->bind_list;
1986         struct net *net = id_priv->id.route.addr.dev_addr.net;
1987
1988         if (!bind_list)
1989                 return;
1990
1991         mutex_lock(&lock);
1992         hlist_del(&id_priv->node);
1993         if (hlist_empty(&bind_list->owners)) {
1994                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1995                 kfree(bind_list);
1996         }
1997         mutex_unlock(&lock);
1998 }
1999
2000 static void destroy_mc(struct rdma_id_private *id_priv,
2001                        struct cma_multicast *mc)
2002 {
2003         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
2004
2005         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
2006                 ib_sa_free_multicast(mc->sa_mc);
2007
2008         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
2009                 struct rdma_dev_addr *dev_addr =
2010                         &id_priv->id.route.addr.dev_addr;
2011                 struct net_device *ndev = NULL;
2012
2013                 if (dev_addr->bound_dev_if)
2014                         ndev = dev_get_by_index(dev_addr->net,
2015                                                 dev_addr->bound_dev_if);
2016                 if (ndev && !send_only) {
2017                         enum ib_gid_type gid_type;
2018                         union ib_gid mgid;
2019
2020                         gid_type = id_priv->cma_dev->default_gid_type
2021                                            [id_priv->id.port_num -
2022                                             rdma_start_port(
2023                                                     id_priv->cma_dev->device)];
2024                         cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
2025                                           gid_type);
2026                         cma_igmp_send(ndev, &mgid, false);
2027                 }
2028                 dev_put(ndev);
2029
2030                 cancel_work_sync(&mc->iboe_join.work);
2031         }
2032         kfree(mc);
2033 }
2034
2035 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
2036 {
2037         struct cma_multicast *mc;
2038
2039         while (!list_empty(&id_priv->mc_list)) {
2040                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
2041                                       list);
2042                 list_del(&mc->list);
2043                 destroy_mc(id_priv, mc);
2044         }
2045 }
2046
2047 static void _destroy_id(struct rdma_id_private *id_priv,
2048                         enum rdma_cm_state state)
2049 {
2050         cma_cancel_operation(id_priv, state);
2051
2052         rdma_restrack_del(&id_priv->res);
2053         cma_remove_id_from_tree(id_priv);
2054         if (id_priv->cma_dev) {
2055                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
2056                         if (id_priv->cm_id.ib)
2057                                 ib_destroy_cm_id(id_priv->cm_id.ib);
2058                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
2059                         if (id_priv->cm_id.iw)
2060                                 iw_destroy_cm_id(id_priv->cm_id.iw);
2061                 }
2062                 cma_leave_mc_groups(id_priv);
2063                 cma_release_dev(id_priv);
2064         }
2065
2066         cma_release_port(id_priv);
2067         cma_id_put(id_priv);
2068         wait_for_completion(&id_priv->comp);
2069
2070         if (id_priv->internal_id)
2071                 cma_id_put(id_priv->id.context);
2072
2073         kfree(id_priv->id.route.path_rec);
2074         kfree(id_priv->id.route.path_rec_inbound);
2075         kfree(id_priv->id.route.path_rec_outbound);
2076
2077         put_net(id_priv->id.route.addr.dev_addr.net);
2078         kfree(id_priv);
2079 }
2080
2081 /*
2082  * destroy an ID from within the handler_mutex. This ensures that no other
2083  * handlers can start running concurrently.
2084  */
2085 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
2086         __releases(&idprv->handler_mutex)
2087 {
2088         enum rdma_cm_state state;
2089         unsigned long flags;
2090
2091         trace_cm_id_destroy(id_priv);
2092
2093         /*
2094          * Setting the state to destroyed under the handler mutex provides a
2095          * fence against calling handler callbacks. If this is invoked due to
2096          * the failure of a handler callback then it guarentees that no future
2097          * handlers will be called.
2098          */
2099         lockdep_assert_held(&id_priv->handler_mutex);
2100         spin_lock_irqsave(&id_priv->lock, flags);
2101         state = id_priv->state;
2102         id_priv->state = RDMA_CM_DESTROYING;
2103         spin_unlock_irqrestore(&id_priv->lock, flags);
2104         mutex_unlock(&id_priv->handler_mutex);
2105         _destroy_id(id_priv, state);
2106 }
2107
2108 void rdma_destroy_id(struct rdma_cm_id *id)
2109 {
2110         struct rdma_id_private *id_priv =
2111                 container_of(id, struct rdma_id_private, id);
2112
2113         mutex_lock(&id_priv->handler_mutex);
2114         destroy_id_handler_unlock(id_priv);
2115 }
2116 EXPORT_SYMBOL(rdma_destroy_id);
2117
2118 static int cma_rep_recv(struct rdma_id_private *id_priv)
2119 {
2120         int ret;
2121
2122         ret = cma_modify_qp_rtr(id_priv, NULL);
2123         if (ret)
2124                 goto reject;
2125
2126         ret = cma_modify_qp_rts(id_priv, NULL);
2127         if (ret)
2128                 goto reject;
2129
2130         trace_cm_send_rtu(id_priv);
2131         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
2132         if (ret)
2133                 goto reject;
2134
2135         return 0;
2136 reject:
2137         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
2138         cma_modify_qp_err(id_priv);
2139         trace_cm_send_rej(id_priv);
2140         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
2141                        NULL, 0, NULL, 0);
2142         return ret;
2143 }
2144
2145 static void cma_set_rep_event_data(struct rdma_cm_event *event,
2146                                    const struct ib_cm_rep_event_param *rep_data,
2147                                    void *private_data)
2148 {
2149         event->param.conn.private_data = private_data;
2150         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
2151         event->param.conn.responder_resources = rep_data->responder_resources;
2152         event->param.conn.initiator_depth = rep_data->initiator_depth;
2153         event->param.conn.flow_control = rep_data->flow_control;
2154         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
2155         event->param.conn.srq = rep_data->srq;
2156         event->param.conn.qp_num = rep_data->remote_qpn;
2157
2158         event->ece.vendor_id = rep_data->ece.vendor_id;
2159         event->ece.attr_mod = rep_data->ece.attr_mod;
2160 }
2161
2162 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
2163                                 struct rdma_cm_event *event)
2164 {
2165         int ret;
2166
2167         lockdep_assert_held(&id_priv->handler_mutex);
2168
2169         trace_cm_event_handler(id_priv, event);
2170         ret = id_priv->id.event_handler(&id_priv->id, event);
2171         trace_cm_event_done(id_priv, event, ret);
2172         return ret;
2173 }
2174
2175 static int cma_ib_handler(struct ib_cm_id *cm_id,
2176                           const struct ib_cm_event *ib_event)
2177 {
2178         struct rdma_id_private *id_priv = cm_id->context;
2179         struct rdma_cm_event event = {};
2180         enum rdma_cm_state state;
2181         int ret;
2182
2183         mutex_lock(&id_priv->handler_mutex);
2184         state = READ_ONCE(id_priv->state);
2185         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2186              state != RDMA_CM_CONNECT) ||
2187             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2188              state != RDMA_CM_DISCONNECT))
2189                 goto out;
2190
2191         switch (ib_event->event) {
2192         case IB_CM_REQ_ERROR:
2193         case IB_CM_REP_ERROR:
2194                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2195                 event.status = -ETIMEDOUT;
2196                 break;
2197         case IB_CM_REP_RECEIVED:
2198                 if (state == RDMA_CM_CONNECT &&
2199                     (id_priv->id.qp_type != IB_QPT_UD)) {
2200                         trace_cm_send_mra(id_priv);
2201                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2202                 }
2203                 if (id_priv->id.qp) {
2204                         event.status = cma_rep_recv(id_priv);
2205                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2206                                                      RDMA_CM_EVENT_ESTABLISHED;
2207                 } else {
2208                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2209                 }
2210                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2211                                        ib_event->private_data);
2212                 break;
2213         case IB_CM_RTU_RECEIVED:
2214         case IB_CM_USER_ESTABLISHED:
2215                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2216                 break;
2217         case IB_CM_DREQ_ERROR:
2218                 event.status = -ETIMEDOUT;
2219                 fallthrough;
2220         case IB_CM_DREQ_RECEIVED:
2221         case IB_CM_DREP_RECEIVED:
2222                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2223                                    RDMA_CM_DISCONNECT))
2224                         goto out;
2225                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2226                 break;
2227         case IB_CM_TIMEWAIT_EXIT:
2228                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2229                 break;
2230         case IB_CM_MRA_RECEIVED:
2231                 /* ignore event */
2232                 goto out;
2233         case IB_CM_REJ_RECEIVED:
2234                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2235                                                                                 ib_event->param.rej_rcvd.reason));
2236                 cma_modify_qp_err(id_priv);
2237                 event.status = ib_event->param.rej_rcvd.reason;
2238                 event.event = RDMA_CM_EVENT_REJECTED;
2239                 event.param.conn.private_data = ib_event->private_data;
2240                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2241                 break;
2242         default:
2243                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2244                        ib_event->event);
2245                 goto out;
2246         }
2247
2248         ret = cma_cm_event_handler(id_priv, &event);
2249         if (ret) {
2250                 /* Destroy the CM ID by returning a non-zero value. */
2251                 id_priv->cm_id.ib = NULL;
2252                 destroy_id_handler_unlock(id_priv);
2253                 return ret;
2254         }
2255 out:
2256         mutex_unlock(&id_priv->handler_mutex);
2257         return 0;
2258 }
2259
2260 static struct rdma_id_private *
2261 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2262                    const struct ib_cm_event *ib_event,
2263                    struct net_device *net_dev)
2264 {
2265         struct rdma_id_private *listen_id_priv;
2266         struct rdma_id_private *id_priv;
2267         struct rdma_cm_id *id;
2268         struct rdma_route *rt;
2269         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2270         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2271         const __be64 service_id =
2272                 ib_event->param.req_rcvd.primary_path->service_id;
2273         int ret;
2274
2275         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2276         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2277                                    listen_id->event_handler, listen_id->context,
2278                                    listen_id->ps,
2279                                    ib_event->param.req_rcvd.qp_type,
2280                                    listen_id_priv);
2281         if (IS_ERR(id_priv))
2282                 return NULL;
2283
2284         id = &id_priv->id;
2285         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2286                               (struct sockaddr *)&id->route.addr.dst_addr,
2287                               listen_id, ib_event, ss_family, service_id))
2288                 goto err;
2289
2290         rt = &id->route;
2291         rt->num_pri_alt_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2292         rt->path_rec = kmalloc_array(rt->num_pri_alt_paths,
2293                                      sizeof(*rt->path_rec), GFP_KERNEL);
2294         if (!rt->path_rec)
2295                 goto err;
2296
2297         rt->path_rec[0] = *path;
2298         if (rt->num_pri_alt_paths == 2)
2299                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2300
2301         if (net_dev) {
2302                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2303         } else {
2304                 if (!cma_protocol_roce(listen_id) &&
2305                     cma_any_addr(cma_src_addr(id_priv))) {
2306                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2307                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2308                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2309                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2310                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2311                         if (ret)
2312                                 goto err;
2313                 }
2314         }
2315         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2316
2317         id_priv->state = RDMA_CM_CONNECT;
2318         return id_priv;
2319
2320 err:
2321         rdma_destroy_id(id);
2322         return NULL;
2323 }
2324
2325 static struct rdma_id_private *
2326 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2327                   const struct ib_cm_event *ib_event,
2328                   struct net_device *net_dev)
2329 {
2330         const struct rdma_id_private *listen_id_priv;
2331         struct rdma_id_private *id_priv;
2332         struct rdma_cm_id *id;
2333         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2334         struct net *net = listen_id->route.addr.dev_addr.net;
2335         int ret;
2336
2337         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2338         id_priv = __rdma_create_id(net, listen_id->event_handler,
2339                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2340                                    listen_id_priv);
2341         if (IS_ERR(id_priv))
2342                 return NULL;
2343
2344         id = &id_priv->id;
2345         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2346                               (struct sockaddr *)&id->route.addr.dst_addr,
2347                               listen_id, ib_event, ss_family,
2348                               ib_event->param.sidr_req_rcvd.service_id))
2349                 goto err;
2350
2351         if (net_dev) {
2352                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2353         } else {
2354                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2355                         ret = cma_translate_addr(cma_src_addr(id_priv),
2356                                                  &id->route.addr.dev_addr);
2357                         if (ret)
2358                                 goto err;
2359                 }
2360         }
2361
2362         id_priv->state = RDMA_CM_CONNECT;
2363         return id_priv;
2364 err:
2365         rdma_destroy_id(id);
2366         return NULL;
2367 }
2368
2369 static void cma_set_req_event_data(struct rdma_cm_event *event,
2370                                    const struct ib_cm_req_event_param *req_data,
2371                                    void *private_data, int offset)
2372 {
2373         event->param.conn.private_data = private_data + offset;
2374         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2375         event->param.conn.responder_resources = req_data->responder_resources;
2376         event->param.conn.initiator_depth = req_data->initiator_depth;
2377         event->param.conn.flow_control = req_data->flow_control;
2378         event->param.conn.retry_count = req_data->retry_count;
2379         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2380         event->param.conn.srq = req_data->srq;
2381         event->param.conn.qp_num = req_data->remote_qpn;
2382
2383         event->ece.vendor_id = req_data->ece.vendor_id;
2384         event->ece.attr_mod = req_data->ece.attr_mod;
2385 }
2386
2387 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2388                                     const struct ib_cm_event *ib_event)
2389 {
2390         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2391                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2392                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2393                  (id->qp_type == IB_QPT_UD)) ||
2394                 (!id->qp_type));
2395 }
2396
2397 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2398                               const struct ib_cm_event *ib_event)
2399 {
2400         struct rdma_id_private *listen_id, *conn_id = NULL;
2401         struct rdma_cm_event event = {};
2402         struct cma_req_info req = {};
2403         struct net_device *net_dev;
2404         u8 offset;
2405         int ret;
2406
2407         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2408         if (IS_ERR(listen_id))
2409                 return PTR_ERR(listen_id);
2410
2411         trace_cm_req_handler(listen_id, ib_event->event);
2412         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2413                 ret = -EINVAL;
2414                 goto net_dev_put;
2415         }
2416
2417         mutex_lock(&listen_id->handler_mutex);
2418         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2419                 ret = -ECONNABORTED;
2420                 goto err_unlock;
2421         }
2422
2423         offset = cma_user_data_offset(listen_id);
2424         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2425         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2426                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2427                 event.param.ud.private_data = ib_event->private_data + offset;
2428                 event.param.ud.private_data_len =
2429                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2430         } else {
2431                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2432                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2433                                        ib_event->private_data, offset);
2434         }
2435         if (!conn_id) {
2436                 ret = -ENOMEM;
2437                 goto err_unlock;
2438         }
2439
2440         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2441         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2442         if (ret) {
2443                 destroy_id_handler_unlock(conn_id);
2444                 goto err_unlock;
2445         }
2446
2447         conn_id->cm_id.ib = cm_id;
2448         cm_id->context = conn_id;
2449         cm_id->cm_handler = cma_ib_handler;
2450
2451         ret = cma_cm_event_handler(conn_id, &event);
2452         if (ret) {
2453                 /* Destroy the CM ID by returning a non-zero value. */
2454                 conn_id->cm_id.ib = NULL;
2455                 mutex_unlock(&listen_id->handler_mutex);
2456                 destroy_id_handler_unlock(conn_id);
2457                 goto net_dev_put;
2458         }
2459
2460         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2461             conn_id->id.qp_type != IB_QPT_UD) {
2462                 trace_cm_send_mra(cm_id->context);
2463                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2464         }
2465         mutex_unlock(&conn_id->handler_mutex);
2466
2467 err_unlock:
2468         mutex_unlock(&listen_id->handler_mutex);
2469
2470 net_dev_put:
2471         dev_put(net_dev);
2472
2473         return ret;
2474 }
2475
2476 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2477 {
2478         if (addr->sa_family == AF_IB)
2479                 return ((struct sockaddr_ib *) addr)->sib_sid;
2480
2481         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2482 }
2483 EXPORT_SYMBOL(rdma_get_service_id);
2484
2485 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2486                     union ib_gid *dgid)
2487 {
2488         struct rdma_addr *addr = &cm_id->route.addr;
2489
2490         if (!cm_id->device) {
2491                 if (sgid)
2492                         memset(sgid, 0, sizeof(*sgid));
2493                 if (dgid)
2494                         memset(dgid, 0, sizeof(*dgid));
2495                 return;
2496         }
2497
2498         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2499                 if (sgid)
2500                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2501                 if (dgid)
2502                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2503         } else {
2504                 if (sgid)
2505                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2506                 if (dgid)
2507                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2508         }
2509 }
2510 EXPORT_SYMBOL(rdma_read_gids);
2511
2512 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2513 {
2514         struct rdma_id_private *id_priv = iw_id->context;
2515         struct rdma_cm_event event = {};
2516         int ret = 0;
2517         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2518         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2519
2520         mutex_lock(&id_priv->handler_mutex);
2521         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2522                 goto out;
2523
2524         switch (iw_event->event) {
2525         case IW_CM_EVENT_CLOSE:
2526                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2527                 break;
2528         case IW_CM_EVENT_CONNECT_REPLY:
2529                 memcpy(cma_src_addr(id_priv), laddr,
2530                        rdma_addr_size(laddr));
2531                 memcpy(cma_dst_addr(id_priv), raddr,
2532                        rdma_addr_size(raddr));
2533                 switch (iw_event->status) {
2534                 case 0:
2535                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2536                         event.param.conn.initiator_depth = iw_event->ird;
2537                         event.param.conn.responder_resources = iw_event->ord;
2538                         break;
2539                 case -ECONNRESET:
2540                 case -ECONNREFUSED:
2541                         event.event = RDMA_CM_EVENT_REJECTED;
2542                         break;
2543                 case -ETIMEDOUT:
2544                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2545                         break;
2546                 default:
2547                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2548                         break;
2549                 }
2550                 break;
2551         case IW_CM_EVENT_ESTABLISHED:
2552                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2553                 event.param.conn.initiator_depth = iw_event->ird;
2554                 event.param.conn.responder_resources = iw_event->ord;
2555                 break;
2556         default:
2557                 goto out;
2558         }
2559
2560         event.status = iw_event->status;
2561         event.param.conn.private_data = iw_event->private_data;
2562         event.param.conn.private_data_len = iw_event->private_data_len;
2563         ret = cma_cm_event_handler(id_priv, &event);
2564         if (ret) {
2565                 /* Destroy the CM ID by returning a non-zero value. */
2566                 id_priv->cm_id.iw = NULL;
2567                 destroy_id_handler_unlock(id_priv);
2568                 return ret;
2569         }
2570
2571 out:
2572         mutex_unlock(&id_priv->handler_mutex);
2573         return ret;
2574 }
2575
2576 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2577                                struct iw_cm_event *iw_event)
2578 {
2579         struct rdma_id_private *listen_id, *conn_id;
2580         struct rdma_cm_event event = {};
2581         int ret = -ECONNABORTED;
2582         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2583         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2584
2585         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2586         event.param.conn.private_data = iw_event->private_data;
2587         event.param.conn.private_data_len = iw_event->private_data_len;
2588         event.param.conn.initiator_depth = iw_event->ird;
2589         event.param.conn.responder_resources = iw_event->ord;
2590
2591         listen_id = cm_id->context;
2592
2593         mutex_lock(&listen_id->handler_mutex);
2594         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2595                 goto out;
2596
2597         /* Create a new RDMA id for the new IW CM ID */
2598         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2599                                    listen_id->id.event_handler,
2600                                    listen_id->id.context, RDMA_PS_TCP,
2601                                    IB_QPT_RC, listen_id);
2602         if (IS_ERR(conn_id)) {
2603                 ret = -ENOMEM;
2604                 goto out;
2605         }
2606         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2607         conn_id->state = RDMA_CM_CONNECT;
2608
2609         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2610         if (ret) {
2611                 mutex_unlock(&listen_id->handler_mutex);
2612                 destroy_id_handler_unlock(conn_id);
2613                 return ret;
2614         }
2615
2616         ret = cma_iw_acquire_dev(conn_id, listen_id);
2617         if (ret) {
2618                 mutex_unlock(&listen_id->handler_mutex);
2619                 destroy_id_handler_unlock(conn_id);
2620                 return ret;
2621         }
2622
2623         conn_id->cm_id.iw = cm_id;
2624         cm_id->context = conn_id;
2625         cm_id->cm_handler = cma_iw_handler;
2626
2627         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2628         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2629
2630         ret = cma_cm_event_handler(conn_id, &event);
2631         if (ret) {
2632                 /* User wants to destroy the CM ID */
2633                 conn_id->cm_id.iw = NULL;
2634                 mutex_unlock(&listen_id->handler_mutex);
2635                 destroy_id_handler_unlock(conn_id);
2636                 return ret;
2637         }
2638
2639         mutex_unlock(&conn_id->handler_mutex);
2640
2641 out:
2642         mutex_unlock(&listen_id->handler_mutex);
2643         return ret;
2644 }
2645
2646 static int cma_ib_listen(struct rdma_id_private *id_priv)
2647 {
2648         struct sockaddr *addr;
2649         struct ib_cm_id *id;
2650         __be64 svc_id;
2651
2652         addr = cma_src_addr(id_priv);
2653         svc_id = rdma_get_service_id(&id_priv->id, addr);
2654         id = ib_cm_insert_listen(id_priv->id.device,
2655                                  cma_ib_req_handler, svc_id);
2656         if (IS_ERR(id))
2657                 return PTR_ERR(id);
2658         id_priv->cm_id.ib = id;
2659
2660         return 0;
2661 }
2662
2663 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2664 {
2665         int ret;
2666         struct iw_cm_id *id;
2667
2668         id = iw_create_cm_id(id_priv->id.device,
2669                              iw_conn_req_handler,
2670                              id_priv);
2671         if (IS_ERR(id))
2672                 return PTR_ERR(id);
2673
2674         mutex_lock(&id_priv->qp_mutex);
2675         id->tos = id_priv->tos;
2676         id->tos_set = id_priv->tos_set;
2677         mutex_unlock(&id_priv->qp_mutex);
2678         id->afonly = id_priv->afonly;
2679         id_priv->cm_id.iw = id;
2680
2681         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2682                rdma_addr_size(cma_src_addr(id_priv)));
2683
2684         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2685
2686         if (ret) {
2687                 iw_destroy_cm_id(id_priv->cm_id.iw);
2688                 id_priv->cm_id.iw = NULL;
2689         }
2690
2691         return ret;
2692 }
2693
2694 static int cma_listen_handler(struct rdma_cm_id *id,
2695                               struct rdma_cm_event *event)
2696 {
2697         struct rdma_id_private *id_priv = id->context;
2698
2699         /* Listening IDs are always destroyed on removal */
2700         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2701                 return -1;
2702
2703         id->context = id_priv->id.context;
2704         id->event_handler = id_priv->id.event_handler;
2705         trace_cm_event_handler(id_priv, event);
2706         return id_priv->id.event_handler(id, event);
2707 }
2708
2709 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2710                              struct cma_device *cma_dev,
2711                              struct rdma_id_private **to_destroy)
2712 {
2713         struct rdma_id_private *dev_id_priv;
2714         struct net *net = id_priv->id.route.addr.dev_addr.net;
2715         int ret;
2716
2717         lockdep_assert_held(&lock);
2718
2719         *to_destroy = NULL;
2720         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2721                 return 0;
2722
2723         dev_id_priv =
2724                 __rdma_create_id(net, cma_listen_handler, id_priv,
2725                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2726         if (IS_ERR(dev_id_priv))
2727                 return PTR_ERR(dev_id_priv);
2728
2729         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2730         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2731                rdma_addr_size(cma_src_addr(id_priv)));
2732
2733         _cma_attach_to_dev(dev_id_priv, cma_dev);
2734         rdma_restrack_add(&dev_id_priv->res);
2735         cma_id_get(id_priv);
2736         dev_id_priv->internal_id = 1;
2737         dev_id_priv->afonly = id_priv->afonly;
2738         mutex_lock(&id_priv->qp_mutex);
2739         dev_id_priv->tos_set = id_priv->tos_set;
2740         dev_id_priv->tos = id_priv->tos;
2741         mutex_unlock(&id_priv->qp_mutex);
2742
2743         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2744         if (ret)
2745                 goto err_listen;
2746         list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2747         return 0;
2748 err_listen:
2749         /* Caller must destroy this after releasing lock */
2750         *to_destroy = dev_id_priv;
2751         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2752         return ret;
2753 }
2754
2755 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2756 {
2757         struct rdma_id_private *to_destroy;
2758         struct cma_device *cma_dev;
2759         int ret;
2760
2761         mutex_lock(&lock);
2762         list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2763         list_for_each_entry(cma_dev, &dev_list, list) {
2764                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2765                 if (ret) {
2766                         /* Prevent racing with cma_process_remove() */
2767                         if (to_destroy)
2768                                 list_del_init(&to_destroy->device_item);
2769                         goto err_listen;
2770                 }
2771         }
2772         mutex_unlock(&lock);
2773         return 0;
2774
2775 err_listen:
2776         _cma_cancel_listens(id_priv);
2777         mutex_unlock(&lock);
2778         if (to_destroy)
2779                 rdma_destroy_id(&to_destroy->id);
2780         return ret;
2781 }
2782
2783 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2784 {
2785         struct rdma_id_private *id_priv;
2786
2787         id_priv = container_of(id, struct rdma_id_private, id);
2788         mutex_lock(&id_priv->qp_mutex);
2789         id_priv->tos = (u8) tos;
2790         id_priv->tos_set = true;
2791         mutex_unlock(&id_priv->qp_mutex);
2792 }
2793 EXPORT_SYMBOL(rdma_set_service_type);
2794
2795 /**
2796  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2797  *                          with a connection identifier.
2798  * @id: Communication identifier to associated with service type.
2799  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2800  *
2801  * This function should be called before rdma_connect() on active side,
2802  * and on passive side before rdma_accept(). It is applicable to primary
2803  * path only. The timeout will affect the local side of the QP, it is not
2804  * negotiated with remote side and zero disables the timer. In case it is
2805  * set before rdma_resolve_route, the value will also be used to determine
2806  * PacketLifeTime for RoCE.
2807  *
2808  * Return: 0 for success
2809  */
2810 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2811 {
2812         struct rdma_id_private *id_priv;
2813
2814         if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2815                 return -EINVAL;
2816
2817         id_priv = container_of(id, struct rdma_id_private, id);
2818         mutex_lock(&id_priv->qp_mutex);
2819         id_priv->timeout = timeout;
2820         id_priv->timeout_set = true;
2821         mutex_unlock(&id_priv->qp_mutex);
2822
2823         return 0;
2824 }
2825 EXPORT_SYMBOL(rdma_set_ack_timeout);
2826
2827 /**
2828  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2829  *                            QP associated with a connection identifier.
2830  * @id: Communication identifier to associated with service type.
2831  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2832  *                 Timer Field" in the IBTA specification.
2833  *
2834  * This function should be called before rdma_connect() on active
2835  * side, and on passive side before rdma_accept(). The timer value
2836  * will be associated with the local QP. When it receives a send it is
2837  * not read to handle, typically if the receive queue is empty, an RNR
2838  * Retry NAK is returned to the requester with the min_rnr_timer
2839  * encoded. The requester will then wait at least the time specified
2840  * in the NAK before retrying. The default is zero, which translates
2841  * to a minimum RNR Timer value of 655 ms.
2842  *
2843  * Return: 0 for success
2844  */
2845 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2846 {
2847         struct rdma_id_private *id_priv;
2848
2849         /* It is a five-bit value */
2850         if (min_rnr_timer & 0xe0)
2851                 return -EINVAL;
2852
2853         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2854                 return -EINVAL;
2855
2856         id_priv = container_of(id, struct rdma_id_private, id);
2857         mutex_lock(&id_priv->qp_mutex);
2858         id_priv->min_rnr_timer = min_rnr_timer;
2859         id_priv->min_rnr_timer_set = true;
2860         mutex_unlock(&id_priv->qp_mutex);
2861
2862         return 0;
2863 }
2864 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2865
2866 static int route_set_path_rec_inbound(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_inbound) {
2872                 route->path_rec_inbound =
2873                         kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
2874                 if (!route->path_rec_inbound)
2875                         return -ENOMEM;
2876         }
2877
2878         *route->path_rec_inbound = *path_rec;
2879         return 0;
2880 }
2881
2882 static int route_set_path_rec_outbound(struct cma_work *work,
2883                                        struct sa_path_rec *path_rec)
2884 {
2885         struct rdma_route *route = &work->id->id.route;
2886
2887         if (!route->path_rec_outbound) {
2888                 route->path_rec_outbound =
2889                         kzalloc(sizeof(*route->path_rec_outbound), GFP_KERNEL);
2890                 if (!route->path_rec_outbound)
2891                         return -ENOMEM;
2892         }
2893
2894         *route->path_rec_outbound = *path_rec;
2895         return 0;
2896 }
2897
2898 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2899                               unsigned int num_prs, void *context)
2900 {
2901         struct cma_work *work = context;
2902         struct rdma_route *route;
2903         int i;
2904
2905         route = &work->id->id.route;
2906
2907         if (status)
2908                 goto fail;
2909
2910         for (i = 0; i < num_prs; i++) {
2911                 if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
2912                         *route->path_rec = path_rec[i];
2913                 else if (path_rec[i].flags & IB_PATH_INBOUND)
2914                         status = route_set_path_rec_inbound(work, &path_rec[i]);
2915                 else if (path_rec[i].flags & IB_PATH_OUTBOUND)
2916                         status = route_set_path_rec_outbound(work,
2917                                                              &path_rec[i]);
2918                 else
2919                         status = -EINVAL;
2920
2921                 if (status)
2922                         goto fail;
2923         }
2924
2925         route->num_pri_alt_paths = 1;
2926         queue_work(cma_wq, &work->work);
2927         return;
2928
2929 fail:
2930         work->old_state = RDMA_CM_ROUTE_QUERY;
2931         work->new_state = RDMA_CM_ADDR_RESOLVED;
2932         work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2933         work->event.status = status;
2934         pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2935                              status);
2936         queue_work(cma_wq, &work->work);
2937 }
2938
2939 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2940                               unsigned long timeout_ms, struct cma_work *work)
2941 {
2942         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2943         struct sa_path_rec path_rec;
2944         ib_sa_comp_mask comp_mask;
2945         struct sockaddr_in6 *sin6;
2946         struct sockaddr_ib *sib;
2947
2948         memset(&path_rec, 0, sizeof path_rec);
2949
2950         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2951                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2952         else
2953                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2954         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2955         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2956         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2957         path_rec.numb_path = 1;
2958         path_rec.reversible = 1;
2959         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2960                                                   cma_dst_addr(id_priv));
2961
2962         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2963                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2964                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2965
2966         switch (cma_family(id_priv)) {
2967         case AF_INET:
2968                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2969                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2970                 break;
2971         case AF_INET6:
2972                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2973                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2974                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2975                 break;
2976         case AF_IB:
2977                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2978                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2979                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2980                 break;
2981         }
2982
2983         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2984                                                id_priv->id.port_num, &path_rec,
2985                                                comp_mask, timeout_ms,
2986                                                GFP_KERNEL, cma_query_handler,
2987                                                work, &id_priv->query);
2988
2989         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2990 }
2991
2992 static void cma_iboe_join_work_handler(struct work_struct *work)
2993 {
2994         struct cma_multicast *mc =
2995                 container_of(work, struct cma_multicast, iboe_join.work);
2996         struct rdma_cm_event *event = &mc->iboe_join.event;
2997         struct rdma_id_private *id_priv = mc->id_priv;
2998         int ret;
2999
3000         mutex_lock(&id_priv->handler_mutex);
3001         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
3002             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
3003                 goto out_unlock;
3004
3005         ret = cma_cm_event_handler(id_priv, event);
3006         WARN_ON(ret);
3007
3008 out_unlock:
3009         mutex_unlock(&id_priv->handler_mutex);
3010         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
3011                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
3012 }
3013
3014 static void cma_work_handler(struct work_struct *_work)
3015 {
3016         struct cma_work *work = container_of(_work, struct cma_work, work);
3017         struct rdma_id_private *id_priv = work->id;
3018
3019         mutex_lock(&id_priv->handler_mutex);
3020         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
3021             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
3022                 goto out_unlock;
3023         if (work->old_state != 0 || work->new_state != 0) {
3024                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
3025                         goto out_unlock;
3026         }
3027
3028         if (cma_cm_event_handler(id_priv, &work->event)) {
3029                 cma_id_put(id_priv);
3030                 destroy_id_handler_unlock(id_priv);
3031                 goto out_free;
3032         }
3033
3034 out_unlock:
3035         mutex_unlock(&id_priv->handler_mutex);
3036         cma_id_put(id_priv);
3037 out_free:
3038         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
3039                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
3040         kfree(work);
3041 }
3042
3043 static void cma_init_resolve_route_work(struct cma_work *work,
3044                                         struct rdma_id_private *id_priv)
3045 {
3046         work->id = id_priv;
3047         INIT_WORK(&work->work, cma_work_handler);
3048         work->old_state = RDMA_CM_ROUTE_QUERY;
3049         work->new_state = RDMA_CM_ROUTE_RESOLVED;
3050         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
3051 }
3052
3053 static void enqueue_resolve_addr_work(struct cma_work *work,
3054                                       struct rdma_id_private *id_priv)
3055 {
3056         /* Balances with cma_id_put() in cma_work_handler */
3057         cma_id_get(id_priv);
3058
3059         work->id = id_priv;
3060         INIT_WORK(&work->work, cma_work_handler);
3061         work->old_state = RDMA_CM_ADDR_QUERY;
3062         work->new_state = RDMA_CM_ADDR_RESOLVED;
3063         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3064
3065         queue_work(cma_wq, &work->work);
3066 }
3067
3068 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
3069                                 unsigned long timeout_ms)
3070 {
3071         struct rdma_route *route = &id_priv->id.route;
3072         struct cma_work *work;
3073         int ret;
3074
3075         work = kzalloc(sizeof *work, GFP_KERNEL);
3076         if (!work)
3077                 return -ENOMEM;
3078
3079         cma_init_resolve_route_work(work, id_priv);
3080
3081         if (!route->path_rec)
3082                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
3083         if (!route->path_rec) {
3084                 ret = -ENOMEM;
3085                 goto err1;
3086         }
3087
3088         ret = cma_query_ib_route(id_priv, timeout_ms, work);
3089         if (ret)
3090                 goto err2;
3091
3092         return 0;
3093 err2:
3094         kfree(route->path_rec);
3095         route->path_rec = NULL;
3096 err1:
3097         kfree(work);
3098         return ret;
3099 }
3100
3101 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
3102                                            unsigned long supported_gids,
3103                                            enum ib_gid_type default_gid)
3104 {
3105         if ((network_type == RDMA_NETWORK_IPV4 ||
3106              network_type == RDMA_NETWORK_IPV6) &&
3107             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
3108                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
3109
3110         return default_gid;
3111 }
3112
3113 /*
3114  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
3115  * path record type based on GID type.
3116  * It also sets up other L2 fields which includes destination mac address
3117  * netdev ifindex, of the path record.
3118  * It returns the netdev of the bound interface for this path record entry.
3119  */
3120 static struct net_device *
3121 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
3122 {
3123         struct rdma_route *route = &id_priv->id.route;
3124         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
3125         struct rdma_addr *addr = &route->addr;
3126         unsigned long supported_gids;
3127         struct net_device *ndev;
3128
3129         if (!addr->dev_addr.bound_dev_if)
3130                 return NULL;
3131
3132         ndev = dev_get_by_index(addr->dev_addr.net,
3133                                 addr->dev_addr.bound_dev_if);
3134         if (!ndev)
3135                 return NULL;
3136
3137         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
3138                                                     id_priv->id.port_num);
3139         gid_type = cma_route_gid_type(addr->dev_addr.network,
3140                                       supported_gids,
3141                                       id_priv->gid_type);
3142         /* Use the hint from IP Stack to select GID Type */
3143         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
3144                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
3145         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
3146
3147         route->path_rec->roce.route_resolved = true;
3148         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
3149         return ndev;
3150 }
3151
3152 int rdma_set_ib_path(struct rdma_cm_id *id,
3153                      struct sa_path_rec *path_rec)
3154 {
3155         struct rdma_id_private *id_priv;
3156         struct net_device *ndev;
3157         int ret;
3158
3159         id_priv = container_of(id, struct rdma_id_private, id);
3160         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3161                            RDMA_CM_ROUTE_RESOLVED))
3162                 return -EINVAL;
3163
3164         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
3165                                      GFP_KERNEL);
3166         if (!id->route.path_rec) {
3167                 ret = -ENOMEM;
3168                 goto err;
3169         }
3170
3171         if (rdma_protocol_roce(id->device, id->port_num)) {
3172                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3173                 if (!ndev) {
3174                         ret = -ENODEV;
3175                         goto err_free;
3176                 }
3177                 dev_put(ndev);
3178         }
3179
3180         id->route.num_pri_alt_paths = 1;
3181         return 0;
3182
3183 err_free:
3184         kfree(id->route.path_rec);
3185         id->route.path_rec = NULL;
3186 err:
3187         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
3188         return ret;
3189 }
3190 EXPORT_SYMBOL(rdma_set_ib_path);
3191
3192 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
3193 {
3194         struct cma_work *work;
3195
3196         work = kzalloc(sizeof *work, GFP_KERNEL);
3197         if (!work)
3198                 return -ENOMEM;
3199
3200         cma_init_resolve_route_work(work, id_priv);
3201         queue_work(cma_wq, &work->work);
3202         return 0;
3203 }
3204
3205 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
3206 {
3207         struct net_device *dev;
3208
3209         dev = vlan_dev_real_dev(vlan_ndev);
3210         if (dev->num_tc)
3211                 return netdev_get_prio_tc_map(dev, prio);
3212
3213         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
3214                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
3215 }
3216
3217 struct iboe_prio_tc_map {
3218         int input_prio;
3219         int output_tc;
3220         bool found;
3221 };
3222
3223 static int get_lower_vlan_dev_tc(struct net_device *dev,
3224                                  struct netdev_nested_priv *priv)
3225 {
3226         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3227
3228         if (is_vlan_dev(dev))
3229                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3230         else if (dev->num_tc)
3231                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3232         else
3233                 map->output_tc = 0;
3234         /* We are interested only in first level VLAN device, so always
3235          * return 1 to stop iterating over next level devices.
3236          */
3237         map->found = true;
3238         return 1;
3239 }
3240
3241 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3242 {
3243         struct iboe_prio_tc_map prio_tc_map = {};
3244         int prio = rt_tos2priority(tos);
3245         struct netdev_nested_priv priv;
3246
3247         /* If VLAN device, get it directly from the VLAN netdev */
3248         if (is_vlan_dev(ndev))
3249                 return get_vlan_ndev_tc(ndev, prio);
3250
3251         prio_tc_map.input_prio = prio;
3252         priv.data = (void *)&prio_tc_map;
3253         rcu_read_lock();
3254         netdev_walk_all_lower_dev_rcu(ndev,
3255                                       get_lower_vlan_dev_tc,
3256                                       &priv);
3257         rcu_read_unlock();
3258         /* If map is found from lower device, use it; Otherwise
3259          * continue with the current netdevice to get priority to tc map.
3260          */
3261         if (prio_tc_map.found)
3262                 return prio_tc_map.output_tc;
3263         else if (ndev->num_tc)
3264                 return netdev_get_prio_tc_map(ndev, prio);
3265         else
3266                 return 0;
3267 }
3268
3269 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3270 {
3271         struct sockaddr_in6 *addr6;
3272         u16 dport, sport;
3273         u32 hash, fl;
3274
3275         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3276         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3277         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3278                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3279                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3280                 hash = (u32)sport * 31 + dport;
3281                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3282         }
3283
3284         return cpu_to_be32(fl);
3285 }
3286
3287 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3288 {
3289         struct rdma_route *route = &id_priv->id.route;
3290         struct rdma_addr *addr = &route->addr;
3291         struct cma_work *work;
3292         int ret;
3293         struct net_device *ndev;
3294
3295         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3296                                         rdma_start_port(id_priv->cma_dev->device)];
3297         u8 tos;
3298
3299         mutex_lock(&id_priv->qp_mutex);
3300         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3301         mutex_unlock(&id_priv->qp_mutex);
3302
3303         work = kzalloc(sizeof *work, GFP_KERNEL);
3304         if (!work)
3305                 return -ENOMEM;
3306
3307         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3308         if (!route->path_rec) {
3309                 ret = -ENOMEM;
3310                 goto err1;
3311         }
3312
3313         route->num_pri_alt_paths = 1;
3314
3315         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3316         if (!ndev) {
3317                 ret = -ENODEV;
3318                 goto err2;
3319         }
3320
3321         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3322                     &route->path_rec->sgid);
3323         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3324                     &route->path_rec->dgid);
3325
3326         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3327                 /* TODO: get the hoplimit from the inet/inet6 device */
3328                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3329         else
3330                 route->path_rec->hop_limit = 1;
3331         route->path_rec->reversible = 1;
3332         route->path_rec->pkey = cpu_to_be16(0xffff);
3333         route->path_rec->mtu_selector = IB_SA_EQ;
3334         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3335         route->path_rec->traffic_class = tos;
3336         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3337         route->path_rec->rate_selector = IB_SA_EQ;
3338         route->path_rec->rate = IB_RATE_PORT_CURRENT;
3339         dev_put(ndev);
3340         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3341         /* In case ACK timeout is set, use this value to calculate
3342          * PacketLifeTime.  As per IBTA 12.7.34,
3343          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3344          * Assuming a negligible local ACK delay, we can use
3345          * PacketLifeTime = local ACK timeout/2
3346          * as a reasonable approximation for RoCE networks.
3347          */
3348         mutex_lock(&id_priv->qp_mutex);
3349         if (id_priv->timeout_set && id_priv->timeout)
3350                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3351         else
3352                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3353         mutex_unlock(&id_priv->qp_mutex);
3354
3355         if (!route->path_rec->mtu) {
3356                 ret = -EINVAL;
3357                 goto err2;
3358         }
3359
3360         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3361                                          id_priv->id.port_num))
3362                 route->path_rec->flow_label =
3363                         cma_get_roce_udp_flow_label(id_priv);
3364
3365         cma_init_resolve_route_work(work, id_priv);
3366         queue_work(cma_wq, &work->work);
3367
3368         return 0;
3369
3370 err2:
3371         kfree(route->path_rec);
3372         route->path_rec = NULL;
3373         route->num_pri_alt_paths = 0;
3374 err1:
3375         kfree(work);
3376         return ret;
3377 }
3378
3379 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3380 {
3381         struct rdma_id_private *id_priv;
3382         int ret;
3383
3384         if (!timeout_ms)
3385                 return -EINVAL;
3386
3387         id_priv = container_of(id, struct rdma_id_private, id);
3388         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3389                 return -EINVAL;
3390
3391         cma_id_get(id_priv);
3392         if (rdma_cap_ib_sa(id->device, id->port_num))
3393                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3394         else if (rdma_protocol_roce(id->device, id->port_num)) {
3395                 ret = cma_resolve_iboe_route(id_priv);
3396                 if (!ret)
3397                         cma_add_id_to_tree(id_priv);
3398         }
3399         else if (rdma_protocol_iwarp(id->device, id->port_num))
3400                 ret = cma_resolve_iw_route(id_priv);
3401         else
3402                 ret = -ENOSYS;
3403
3404         if (ret)
3405                 goto err;
3406
3407         return 0;
3408 err:
3409         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3410         cma_id_put(id_priv);
3411         return ret;
3412 }
3413 EXPORT_SYMBOL(rdma_resolve_route);
3414
3415 static void cma_set_loopback(struct sockaddr *addr)
3416 {
3417         switch (addr->sa_family) {
3418         case AF_INET:
3419                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3420                 break;
3421         case AF_INET6:
3422                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3423                               0, 0, 0, htonl(1));
3424                 break;
3425         default:
3426                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3427                             0, 0, 0, htonl(1));
3428                 break;
3429         }
3430 }
3431
3432 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3433 {
3434         struct cma_device *cma_dev, *cur_dev;
3435         union ib_gid gid;
3436         enum ib_port_state port_state;
3437         unsigned int p;
3438         u16 pkey;
3439         int ret;
3440
3441         cma_dev = NULL;
3442         mutex_lock(&lock);
3443         list_for_each_entry(cur_dev, &dev_list, list) {
3444                 if (cma_family(id_priv) == AF_IB &&
3445                     !rdma_cap_ib_cm(cur_dev->device, 1))
3446                         continue;
3447
3448                 if (!cma_dev)
3449                         cma_dev = cur_dev;
3450
3451                 rdma_for_each_port (cur_dev->device, p) {
3452                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3453                             port_state == IB_PORT_ACTIVE) {
3454                                 cma_dev = cur_dev;
3455                                 goto port_found;
3456                         }
3457                 }
3458         }
3459
3460         if (!cma_dev) {
3461                 ret = -ENODEV;
3462                 goto out;
3463         }
3464
3465         p = 1;
3466
3467 port_found:
3468         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3469         if (ret)
3470                 goto out;
3471
3472         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3473         if (ret)
3474                 goto out;
3475
3476         id_priv->id.route.addr.dev_addr.dev_type =
3477                 (rdma_protocol_ib(cma_dev->device, p)) ?
3478                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3479
3480         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3481         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3482         id_priv->id.port_num = p;
3483         cma_attach_to_dev(id_priv, cma_dev);
3484         rdma_restrack_add(&id_priv->res);
3485         cma_set_loopback(cma_src_addr(id_priv));
3486 out:
3487         mutex_unlock(&lock);
3488         return ret;
3489 }
3490
3491 static void addr_handler(int status, struct sockaddr *src_addr,
3492                          struct rdma_dev_addr *dev_addr, void *context)
3493 {
3494         struct rdma_id_private *id_priv = context;
3495         struct rdma_cm_event event = {};
3496         struct sockaddr *addr;
3497         struct sockaddr_storage old_addr;
3498
3499         mutex_lock(&id_priv->handler_mutex);
3500         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3501                            RDMA_CM_ADDR_RESOLVED))
3502                 goto out;
3503
3504         /*
3505          * Store the previous src address, so that if we fail to acquire
3506          * matching rdma device, old address can be restored back, which helps
3507          * to cancel the cma listen operation correctly.
3508          */
3509         addr = cma_src_addr(id_priv);
3510         memcpy(&old_addr, addr, rdma_addr_size(addr));
3511         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3512         if (!status && !id_priv->cma_dev) {
3513                 status = cma_acquire_dev_by_src_ip(id_priv);
3514                 if (status)
3515                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3516                                              status);
3517                 rdma_restrack_add(&id_priv->res);
3518         } else if (status) {
3519                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3520         }
3521
3522         if (status) {
3523                 memcpy(addr, &old_addr,
3524                        rdma_addr_size((struct sockaddr *)&old_addr));
3525                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3526                                    RDMA_CM_ADDR_BOUND))
3527                         goto out;
3528                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3529                 event.status = status;
3530         } else
3531                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3532
3533         if (cma_cm_event_handler(id_priv, &event)) {
3534                 destroy_id_handler_unlock(id_priv);
3535                 return;
3536         }
3537 out:
3538         mutex_unlock(&id_priv->handler_mutex);
3539 }
3540
3541 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3542 {
3543         struct cma_work *work;
3544         union ib_gid gid;
3545         int ret;
3546
3547         work = kzalloc(sizeof *work, GFP_KERNEL);
3548         if (!work)
3549                 return -ENOMEM;
3550
3551         if (!id_priv->cma_dev) {
3552                 ret = cma_bind_loopback(id_priv);
3553                 if (ret)
3554                         goto err;
3555         }
3556
3557         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3558         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3559
3560         enqueue_resolve_addr_work(work, id_priv);
3561         return 0;
3562 err:
3563         kfree(work);
3564         return ret;
3565 }
3566
3567 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3568 {
3569         struct cma_work *work;
3570         int ret;
3571
3572         work = kzalloc(sizeof *work, GFP_KERNEL);
3573         if (!work)
3574                 return -ENOMEM;
3575
3576         if (!id_priv->cma_dev) {
3577                 ret = cma_resolve_ib_dev(id_priv);
3578                 if (ret)
3579                         goto err;
3580         }
3581
3582         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3583                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3584
3585         enqueue_resolve_addr_work(work, id_priv);
3586         return 0;
3587 err:
3588         kfree(work);
3589         return ret;
3590 }
3591
3592 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3593 {
3594         struct rdma_id_private *id_priv;
3595         unsigned long flags;
3596         int ret;
3597
3598         id_priv = container_of(id, struct rdma_id_private, id);
3599         spin_lock_irqsave(&id_priv->lock, flags);
3600         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3601             id_priv->state == RDMA_CM_IDLE) {
3602                 id_priv->reuseaddr = reuse;
3603                 ret = 0;
3604         } else {
3605                 ret = -EINVAL;
3606         }
3607         spin_unlock_irqrestore(&id_priv->lock, flags);
3608         return ret;
3609 }
3610 EXPORT_SYMBOL(rdma_set_reuseaddr);
3611
3612 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3613 {
3614         struct rdma_id_private *id_priv;
3615         unsigned long flags;
3616         int ret;
3617
3618         id_priv = container_of(id, struct rdma_id_private, id);
3619         spin_lock_irqsave(&id_priv->lock, flags);
3620         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3621                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3622                 id_priv->afonly = afonly;
3623                 ret = 0;
3624         } else {
3625                 ret = -EINVAL;
3626         }
3627         spin_unlock_irqrestore(&id_priv->lock, flags);
3628         return ret;
3629 }
3630 EXPORT_SYMBOL(rdma_set_afonly);
3631
3632 static void cma_bind_port(struct rdma_bind_list *bind_list,
3633                           struct rdma_id_private *id_priv)
3634 {
3635         struct sockaddr *addr;
3636         struct sockaddr_ib *sib;
3637         u64 sid, mask;
3638         __be16 port;
3639
3640         lockdep_assert_held(&lock);
3641
3642         addr = cma_src_addr(id_priv);
3643         port = htons(bind_list->port);
3644
3645         switch (addr->sa_family) {
3646         case AF_INET:
3647                 ((struct sockaddr_in *) addr)->sin_port = port;
3648                 break;
3649         case AF_INET6:
3650                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3651                 break;
3652         case AF_IB:
3653                 sib = (struct sockaddr_ib *) addr;
3654                 sid = be64_to_cpu(sib->sib_sid);
3655                 mask = be64_to_cpu(sib->sib_sid_mask);
3656                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3657                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3658                 break;
3659         }
3660         id_priv->bind_list = bind_list;
3661         hlist_add_head(&id_priv->node, &bind_list->owners);
3662 }
3663
3664 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3665                           struct rdma_id_private *id_priv, unsigned short snum)
3666 {
3667         struct rdma_bind_list *bind_list;
3668         int ret;
3669
3670         lockdep_assert_held(&lock);
3671
3672         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3673         if (!bind_list)
3674                 return -ENOMEM;
3675
3676         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3677                            snum);
3678         if (ret < 0)
3679                 goto err;
3680
3681         bind_list->ps = ps;
3682         bind_list->port = snum;
3683         cma_bind_port(bind_list, id_priv);
3684         return 0;
3685 err:
3686         kfree(bind_list);
3687         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3688 }
3689
3690 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3691                               struct rdma_id_private *id_priv)
3692 {
3693         struct rdma_id_private *cur_id;
3694         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3695         struct sockaddr  *saddr = cma_src_addr(id_priv);
3696         __be16 dport = cma_port(daddr);
3697
3698         lockdep_assert_held(&lock);
3699
3700         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3701                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3702                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3703                 __be16 cur_dport = cma_port(cur_daddr);
3704
3705                 if (id_priv == cur_id)
3706                         continue;
3707
3708                 /* different dest port -> unique */
3709                 if (!cma_any_port(daddr) &&
3710                     !cma_any_port(cur_daddr) &&
3711                     (dport != cur_dport))
3712                         continue;
3713
3714                 /* different src address -> unique */
3715                 if (!cma_any_addr(saddr) &&
3716                     !cma_any_addr(cur_saddr) &&
3717                     cma_addr_cmp(saddr, cur_saddr))
3718                         continue;
3719
3720                 /* different dst address -> unique */
3721                 if (!cma_any_addr(daddr) &&
3722                     !cma_any_addr(cur_daddr) &&
3723                     cma_addr_cmp(daddr, cur_daddr))
3724                         continue;
3725
3726                 return -EADDRNOTAVAIL;
3727         }
3728         return 0;
3729 }
3730
3731 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3732                               struct rdma_id_private *id_priv)
3733 {
3734         static unsigned int last_used_port;
3735         int low, high, remaining;
3736         unsigned int rover;
3737         struct net *net = id_priv->id.route.addr.dev_addr.net;
3738
3739         lockdep_assert_held(&lock);
3740
3741         inet_get_local_port_range(net, &low, &high);
3742         remaining = (high - low) + 1;
3743         rover = get_random_u32_inclusive(low, remaining + low - 1);
3744 retry:
3745         if (last_used_port != rover) {
3746                 struct rdma_bind_list *bind_list;
3747                 int ret;
3748
3749                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3750
3751                 if (!bind_list) {
3752                         ret = cma_alloc_port(ps, id_priv, rover);
3753                 } else {
3754                         ret = cma_port_is_unique(bind_list, id_priv);
3755                         if (!ret)
3756                                 cma_bind_port(bind_list, id_priv);
3757                 }
3758                 /*
3759                  * Remember previously used port number in order to avoid
3760                  * re-using same port immediately after it is closed.
3761                  */
3762                 if (!ret)
3763                         last_used_port = rover;
3764                 if (ret != -EADDRNOTAVAIL)
3765                         return ret;
3766         }
3767         if (--remaining) {
3768                 rover++;
3769                 if ((rover < low) || (rover > high))
3770                         rover = low;
3771                 goto retry;
3772         }
3773         return -EADDRNOTAVAIL;
3774 }
3775
3776 /*
3777  * Check that the requested port is available.  This is called when trying to
3778  * bind to a specific port, or when trying to listen on a bound port.  In
3779  * the latter case, the provided id_priv may already be on the bind_list, but
3780  * we still need to check that it's okay to start listening.
3781  */
3782 static int cma_check_port(struct rdma_bind_list *bind_list,
3783                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3784 {
3785         struct rdma_id_private *cur_id;
3786         struct sockaddr *addr, *cur_addr;
3787
3788         lockdep_assert_held(&lock);
3789
3790         addr = cma_src_addr(id_priv);
3791         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3792                 if (id_priv == cur_id)
3793                         continue;
3794
3795                 if (reuseaddr && cur_id->reuseaddr)
3796                         continue;
3797
3798                 cur_addr = cma_src_addr(cur_id);
3799                 if (id_priv->afonly && cur_id->afonly &&
3800                     (addr->sa_family != cur_addr->sa_family))
3801                         continue;
3802
3803                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3804                         return -EADDRNOTAVAIL;
3805
3806                 if (!cma_addr_cmp(addr, cur_addr))
3807                         return -EADDRINUSE;
3808         }
3809         return 0;
3810 }
3811
3812 static int cma_use_port(enum rdma_ucm_port_space ps,
3813                         struct rdma_id_private *id_priv)
3814 {
3815         struct rdma_bind_list *bind_list;
3816         unsigned short snum;
3817         int ret;
3818
3819         lockdep_assert_held(&lock);
3820
3821         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3822         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3823                 return -EACCES;
3824
3825         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3826         if (!bind_list) {
3827                 ret = cma_alloc_port(ps, id_priv, snum);
3828         } else {
3829                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3830                 if (!ret)
3831                         cma_bind_port(bind_list, id_priv);
3832         }
3833         return ret;
3834 }
3835
3836 static enum rdma_ucm_port_space
3837 cma_select_inet_ps(struct rdma_id_private *id_priv)
3838 {
3839         switch (id_priv->id.ps) {
3840         case RDMA_PS_TCP:
3841         case RDMA_PS_UDP:
3842         case RDMA_PS_IPOIB:
3843         case RDMA_PS_IB:
3844                 return id_priv->id.ps;
3845         default:
3846
3847                 return 0;
3848         }
3849 }
3850
3851 static enum rdma_ucm_port_space
3852 cma_select_ib_ps(struct rdma_id_private *id_priv)
3853 {
3854         enum rdma_ucm_port_space ps = 0;
3855         struct sockaddr_ib *sib;
3856         u64 sid_ps, mask, sid;
3857
3858         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3859         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3860         sid = be64_to_cpu(sib->sib_sid) & mask;
3861
3862         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3863                 sid_ps = RDMA_IB_IP_PS_IB;
3864                 ps = RDMA_PS_IB;
3865         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3866                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3867                 sid_ps = RDMA_IB_IP_PS_TCP;
3868                 ps = RDMA_PS_TCP;
3869         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3870                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3871                 sid_ps = RDMA_IB_IP_PS_UDP;
3872                 ps = RDMA_PS_UDP;
3873         }
3874
3875         if (ps) {
3876                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3877                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3878                                                 be64_to_cpu(sib->sib_sid_mask));
3879         }
3880         return ps;
3881 }
3882
3883 static int cma_get_port(struct rdma_id_private *id_priv)
3884 {
3885         enum rdma_ucm_port_space ps;
3886         int ret;
3887
3888         if (cma_family(id_priv) != AF_IB)
3889                 ps = cma_select_inet_ps(id_priv);
3890         else
3891                 ps = cma_select_ib_ps(id_priv);
3892         if (!ps)
3893                 return -EPROTONOSUPPORT;
3894
3895         mutex_lock(&lock);
3896         if (cma_any_port(cma_src_addr(id_priv)))
3897                 ret = cma_alloc_any_port(ps, id_priv);
3898         else
3899                 ret = cma_use_port(ps, id_priv);
3900         mutex_unlock(&lock);
3901
3902         return ret;
3903 }
3904
3905 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3906                                struct sockaddr *addr)
3907 {
3908 #if IS_ENABLED(CONFIG_IPV6)
3909         struct sockaddr_in6 *sin6;
3910
3911         if (addr->sa_family != AF_INET6)
3912                 return 0;
3913
3914         sin6 = (struct sockaddr_in6 *) addr;
3915
3916         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3917                 return 0;
3918
3919         if (!sin6->sin6_scope_id)
3920                         return -EINVAL;
3921
3922         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3923 #endif
3924         return 0;
3925 }
3926
3927 int rdma_listen(struct rdma_cm_id *id, int backlog)
3928 {
3929         struct rdma_id_private *id_priv =
3930                 container_of(id, struct rdma_id_private, id);
3931         int ret;
3932
3933         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3934                 struct sockaddr_in any_in = {
3935                         .sin_family = AF_INET,
3936                         .sin_addr.s_addr = htonl(INADDR_ANY),
3937                 };
3938
3939                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3940                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3941                 if (ret)
3942                         return ret;
3943                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3944                                            RDMA_CM_LISTEN)))
3945                         return -EINVAL;
3946         }
3947
3948         /*
3949          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3950          * any more, and has to be unique in the bind list.
3951          */
3952         if (id_priv->reuseaddr) {
3953                 mutex_lock(&lock);
3954                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3955                 if (!ret)
3956                         id_priv->reuseaddr = 0;
3957                 mutex_unlock(&lock);
3958                 if (ret)
3959                         goto err;
3960         }
3961
3962         id_priv->backlog = backlog;
3963         if (id_priv->cma_dev) {
3964                 if (rdma_cap_ib_cm(id->device, 1)) {
3965                         ret = cma_ib_listen(id_priv);
3966                         if (ret)
3967                                 goto err;
3968                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3969                         ret = cma_iw_listen(id_priv, backlog);
3970                         if (ret)
3971                                 goto err;
3972                 } else {
3973                         ret = -ENOSYS;
3974                         goto err;
3975                 }
3976         } else {
3977                 ret = cma_listen_on_all(id_priv);
3978                 if (ret)
3979                         goto err;
3980         }
3981
3982         return 0;
3983 err:
3984         id_priv->backlog = 0;
3985         /*
3986          * All the failure paths that lead here will not allow the req_handler's
3987          * to have run.
3988          */
3989         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3990         return ret;
3991 }
3992 EXPORT_SYMBOL(rdma_listen);
3993
3994 static int rdma_bind_addr_dst(struct rdma_id_private *id_priv,
3995                               struct sockaddr *addr, const struct sockaddr *daddr)
3996 {
3997         struct sockaddr *id_daddr;
3998         int ret;
3999
4000         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
4001             addr->sa_family != AF_IB)
4002                 return -EAFNOSUPPORT;
4003
4004         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
4005                 return -EINVAL;
4006
4007         ret = cma_check_linklocal(&id_priv->id.route.addr.dev_addr, addr);
4008         if (ret)
4009                 goto err1;
4010
4011         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
4012         if (!cma_any_addr(addr)) {
4013                 ret = cma_translate_addr(addr, &id_priv->id.route.addr.dev_addr);
4014                 if (ret)
4015                         goto err1;
4016
4017                 ret = cma_acquire_dev_by_src_ip(id_priv);
4018                 if (ret)
4019                         goto err1;
4020         }
4021
4022         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
4023                 if (addr->sa_family == AF_INET)
4024                         id_priv->afonly = 1;
4025 #if IS_ENABLED(CONFIG_IPV6)
4026                 else if (addr->sa_family == AF_INET6) {
4027                         struct net *net = id_priv->id.route.addr.dev_addr.net;
4028
4029                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
4030                 }
4031 #endif
4032         }
4033         id_daddr = cma_dst_addr(id_priv);
4034         if (daddr != id_daddr)
4035                 memcpy(id_daddr, daddr, rdma_addr_size(addr));
4036         id_daddr->sa_family = addr->sa_family;
4037
4038         ret = cma_get_port(id_priv);
4039         if (ret)
4040                 goto err2;
4041
4042         if (!cma_any_addr(addr))
4043                 rdma_restrack_add(&id_priv->res);
4044         return 0;
4045 err2:
4046         if (id_priv->cma_dev)
4047                 cma_release_dev(id_priv);
4048 err1:
4049         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
4050         return ret;
4051 }
4052
4053 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
4054                          const struct sockaddr *dst_addr)
4055 {
4056         struct rdma_id_private *id_priv =
4057                 container_of(id, struct rdma_id_private, id);
4058         struct sockaddr_storage zero_sock = {};
4059
4060         if (src_addr && src_addr->sa_family)
4061                 return rdma_bind_addr_dst(id_priv, src_addr, dst_addr);
4062
4063         /*
4064          * When the src_addr is not specified, automatically supply an any addr
4065          */
4066         zero_sock.ss_family = dst_addr->sa_family;
4067         if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
4068                 struct sockaddr_in6 *src_addr6 =
4069                         (struct sockaddr_in6 *)&zero_sock;
4070                 struct sockaddr_in6 *dst_addr6 =
4071                         (struct sockaddr_in6 *)dst_addr;
4072
4073                 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
4074                 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
4075                         id->route.addr.dev_addr.bound_dev_if =
4076                                 dst_addr6->sin6_scope_id;
4077         } else if (dst_addr->sa_family == AF_IB) {
4078                 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
4079                         ((struct sockaddr_ib *)dst_addr)->sib_pkey;
4080         }
4081         return rdma_bind_addr_dst(id_priv, (struct sockaddr *)&zero_sock, dst_addr);
4082 }
4083
4084 /*
4085  * If required, resolve the source address for bind and leave the id_priv in
4086  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
4087  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
4088  * ignored.
4089  */
4090 static int resolve_prepare_src(struct rdma_id_private *id_priv,
4091                                struct sockaddr *src_addr,
4092                                const struct sockaddr *dst_addr)
4093 {
4094         int ret;
4095
4096         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
4097                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
4098                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
4099                 if (ret)
4100                         return ret;
4101                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
4102                                            RDMA_CM_ADDR_QUERY)))
4103                         return -EINVAL;
4104
4105         } else {
4106                 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
4107         }
4108
4109         if (cma_family(id_priv) != dst_addr->sa_family) {
4110                 ret = -EINVAL;
4111                 goto err_state;
4112         }
4113         return 0;
4114
4115 err_state:
4116         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
4117         return ret;
4118 }
4119
4120 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
4121                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
4122 {
4123         struct rdma_id_private *id_priv =
4124                 container_of(id, struct rdma_id_private, id);
4125         int ret;
4126
4127         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
4128         if (ret)
4129                 return ret;
4130
4131         if (cma_any_addr(dst_addr)) {
4132                 ret = cma_resolve_loopback(id_priv);
4133         } else {
4134                 if (dst_addr->sa_family == AF_IB) {
4135                         ret = cma_resolve_ib_addr(id_priv);
4136                 } else {
4137                         /*
4138                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
4139                          * rdma_resolve_ip() is called, eg through the error
4140                          * path in addr_handler(). If this happens the existing
4141                          * request must be canceled before issuing a new one.
4142                          * Since canceling a request is a bit slow and this
4143                          * oddball path is rare, keep track once a request has
4144                          * been issued. The track turns out to be a permanent
4145                          * state since this is the only cancel as it is
4146                          * immediately before rdma_resolve_ip().
4147                          */
4148                         if (id_priv->used_resolve_ip)
4149                                 rdma_addr_cancel(&id->route.addr.dev_addr);
4150                         else
4151                                 id_priv->used_resolve_ip = 1;
4152                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
4153                                               &id->route.addr.dev_addr,
4154                                               timeout_ms, addr_handler,
4155                                               false, id_priv);
4156                 }
4157         }
4158         if (ret)
4159                 goto err;
4160
4161         return 0;
4162 err:
4163         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
4164         return ret;
4165 }
4166 EXPORT_SYMBOL(rdma_resolve_addr);
4167
4168 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
4169 {
4170         struct rdma_id_private *id_priv =
4171                 container_of(id, struct rdma_id_private, id);
4172
4173         return rdma_bind_addr_dst(id_priv, addr, cma_dst_addr(id_priv));
4174 }
4175 EXPORT_SYMBOL(rdma_bind_addr);
4176
4177 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
4178 {
4179         struct cma_hdr *cma_hdr;
4180
4181         cma_hdr = hdr;
4182         cma_hdr->cma_version = CMA_VERSION;
4183         if (cma_family(id_priv) == AF_INET) {
4184                 struct sockaddr_in *src4, *dst4;
4185
4186                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
4187                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
4188
4189                 cma_set_ip_ver(cma_hdr, 4);
4190                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
4191                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
4192                 cma_hdr->port = src4->sin_port;
4193         } else if (cma_family(id_priv) == AF_INET6) {
4194                 struct sockaddr_in6 *src6, *dst6;
4195
4196                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
4197                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
4198
4199                 cma_set_ip_ver(cma_hdr, 6);
4200                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
4201                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
4202                 cma_hdr->port = src6->sin6_port;
4203         }
4204         return 0;
4205 }
4206
4207 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
4208                                 const struct ib_cm_event *ib_event)
4209 {
4210         struct rdma_id_private *id_priv = cm_id->context;
4211         struct rdma_cm_event event = {};
4212         const struct ib_cm_sidr_rep_event_param *rep =
4213                                 &ib_event->param.sidr_rep_rcvd;
4214         int ret;
4215
4216         mutex_lock(&id_priv->handler_mutex);
4217         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4218                 goto out;
4219
4220         switch (ib_event->event) {
4221         case IB_CM_SIDR_REQ_ERROR:
4222                 event.event = RDMA_CM_EVENT_UNREACHABLE;
4223                 event.status = -ETIMEDOUT;
4224                 break;
4225         case IB_CM_SIDR_REP_RECEIVED:
4226                 event.param.ud.private_data = ib_event->private_data;
4227                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
4228                 if (rep->status != IB_SIDR_SUCCESS) {
4229                         event.event = RDMA_CM_EVENT_UNREACHABLE;
4230                         event.status = ib_event->param.sidr_rep_rcvd.status;
4231                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
4232                                              event.status);
4233                         break;
4234                 }
4235                 ret = cma_set_qkey(id_priv, rep->qkey);
4236                 if (ret) {
4237                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
4238                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
4239                         event.status = ret;
4240                         break;
4241                 }
4242                 ib_init_ah_attr_from_path(id_priv->id.device,
4243                                           id_priv->id.port_num,
4244                                           id_priv->id.route.path_rec,
4245                                           &event.param.ud.ah_attr,
4246                                           rep->sgid_attr);
4247                 event.param.ud.qp_num = rep->qpn;
4248                 event.param.ud.qkey = rep->qkey;
4249                 event.event = RDMA_CM_EVENT_ESTABLISHED;
4250                 event.status = 0;
4251                 break;
4252         default:
4253                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4254                        ib_event->event);
4255                 goto out;
4256         }
4257
4258         ret = cma_cm_event_handler(id_priv, &event);
4259
4260         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4261         if (ret) {
4262                 /* Destroy the CM ID by returning a non-zero value. */
4263                 id_priv->cm_id.ib = NULL;
4264                 destroy_id_handler_unlock(id_priv);
4265                 return ret;
4266         }
4267 out:
4268         mutex_unlock(&id_priv->handler_mutex);
4269         return 0;
4270 }
4271
4272 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4273                               struct rdma_conn_param *conn_param)
4274 {
4275         struct ib_cm_sidr_req_param req;
4276         struct ib_cm_id *id;
4277         void *private_data;
4278         u8 offset;
4279         int ret;
4280
4281         memset(&req, 0, sizeof req);
4282         offset = cma_user_data_offset(id_priv);
4283         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4284                 return -EINVAL;
4285
4286         if (req.private_data_len) {
4287                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4288                 if (!private_data)
4289                         return -ENOMEM;
4290         } else {
4291                 private_data = NULL;
4292         }
4293
4294         if (conn_param->private_data && conn_param->private_data_len)
4295                 memcpy(private_data + offset, conn_param->private_data,
4296                        conn_param->private_data_len);
4297
4298         if (private_data) {
4299                 ret = cma_format_hdr(private_data, id_priv);
4300                 if (ret)
4301                         goto out;
4302                 req.private_data = private_data;
4303         }
4304
4305         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4306                              id_priv);
4307         if (IS_ERR(id)) {
4308                 ret = PTR_ERR(id);
4309                 goto out;
4310         }
4311         id_priv->cm_id.ib = id;
4312
4313         req.path = id_priv->id.route.path_rec;
4314         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4315         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4316         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4317         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4318
4319         trace_cm_send_sidr_req(id_priv);
4320         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4321         if (ret) {
4322                 ib_destroy_cm_id(id_priv->cm_id.ib);
4323                 id_priv->cm_id.ib = NULL;
4324         }
4325 out:
4326         kfree(private_data);
4327         return ret;
4328 }
4329
4330 static int cma_connect_ib(struct rdma_id_private *id_priv,
4331                           struct rdma_conn_param *conn_param)
4332 {
4333         struct ib_cm_req_param req;
4334         struct rdma_route *route;
4335         void *private_data;
4336         struct ib_cm_id *id;
4337         u8 offset;
4338         int ret;
4339
4340         memset(&req, 0, sizeof req);
4341         offset = cma_user_data_offset(id_priv);
4342         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4343                 return -EINVAL;
4344
4345         if (req.private_data_len) {
4346                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4347                 if (!private_data)
4348                         return -ENOMEM;
4349         } else {
4350                 private_data = NULL;
4351         }
4352
4353         if (conn_param->private_data && conn_param->private_data_len)
4354                 memcpy(private_data + offset, conn_param->private_data,
4355                        conn_param->private_data_len);
4356
4357         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4358         if (IS_ERR(id)) {
4359                 ret = PTR_ERR(id);
4360                 goto out;
4361         }
4362         id_priv->cm_id.ib = id;
4363
4364         route = &id_priv->id.route;
4365         if (private_data) {
4366                 ret = cma_format_hdr(private_data, id_priv);
4367                 if (ret)
4368                         goto out;
4369                 req.private_data = private_data;
4370         }
4371
4372         req.primary_path = &route->path_rec[0];
4373         req.primary_path_inbound = route->path_rec_inbound;
4374         req.primary_path_outbound = route->path_rec_outbound;
4375         if (route->num_pri_alt_paths == 2)
4376                 req.alternate_path = &route->path_rec[1];
4377
4378         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4379         /* Alternate path SGID attribute currently unsupported */
4380         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4381         req.qp_num = id_priv->qp_num;
4382         req.qp_type = id_priv->id.qp_type;
4383         req.starting_psn = id_priv->seq_num;
4384         req.responder_resources = conn_param->responder_resources;
4385         req.initiator_depth = conn_param->initiator_depth;
4386         req.flow_control = conn_param->flow_control;
4387         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4388         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4389         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4390         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4391         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4392         req.srq = id_priv->srq ? 1 : 0;
4393         req.ece.vendor_id = id_priv->ece.vendor_id;
4394         req.ece.attr_mod = id_priv->ece.attr_mod;
4395
4396         trace_cm_send_req(id_priv);
4397         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4398 out:
4399         if (ret && !IS_ERR(id)) {
4400                 ib_destroy_cm_id(id);
4401                 id_priv->cm_id.ib = NULL;
4402         }
4403
4404         kfree(private_data);
4405         return ret;
4406 }
4407
4408 static int cma_connect_iw(struct rdma_id_private *id_priv,
4409                           struct rdma_conn_param *conn_param)
4410 {
4411         struct iw_cm_id *cm_id;
4412         int ret;
4413         struct iw_cm_conn_param iw_param;
4414
4415         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4416         if (IS_ERR(cm_id))
4417                 return PTR_ERR(cm_id);
4418
4419         mutex_lock(&id_priv->qp_mutex);
4420         cm_id->tos = id_priv->tos;
4421         cm_id->tos_set = id_priv->tos_set;
4422         mutex_unlock(&id_priv->qp_mutex);
4423
4424         id_priv->cm_id.iw = cm_id;
4425
4426         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4427                rdma_addr_size(cma_src_addr(id_priv)));
4428         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4429                rdma_addr_size(cma_dst_addr(id_priv)));
4430
4431         ret = cma_modify_qp_rtr(id_priv, conn_param);
4432         if (ret)
4433                 goto out;
4434
4435         if (conn_param) {
4436                 iw_param.ord = conn_param->initiator_depth;
4437                 iw_param.ird = conn_param->responder_resources;
4438                 iw_param.private_data = conn_param->private_data;
4439                 iw_param.private_data_len = conn_param->private_data_len;
4440                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4441         } else {
4442                 memset(&iw_param, 0, sizeof iw_param);
4443                 iw_param.qpn = id_priv->qp_num;
4444         }
4445         ret = iw_cm_connect(cm_id, &iw_param);
4446 out:
4447         if (ret) {
4448                 iw_destroy_cm_id(cm_id);
4449                 id_priv->cm_id.iw = NULL;
4450         }
4451         return ret;
4452 }
4453
4454 /**
4455  * rdma_connect_locked - Initiate an active connection request.
4456  * @id: Connection identifier to connect.
4457  * @conn_param: Connection information used for connected QPs.
4458  *
4459  * Same as rdma_connect() but can only be called from the
4460  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4461  */
4462 int rdma_connect_locked(struct rdma_cm_id *id,
4463                         struct rdma_conn_param *conn_param)
4464 {
4465         struct rdma_id_private *id_priv =
4466                 container_of(id, struct rdma_id_private, id);
4467         int ret;
4468
4469         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4470                 return -EINVAL;
4471
4472         if (!id->qp) {
4473                 id_priv->qp_num = conn_param->qp_num;
4474                 id_priv->srq = conn_param->srq;
4475         }
4476
4477         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4478                 if (id->qp_type == IB_QPT_UD)
4479                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4480                 else
4481                         ret = cma_connect_ib(id_priv, conn_param);
4482         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4483                 ret = cma_connect_iw(id_priv, conn_param);
4484         } else {
4485                 ret = -ENOSYS;
4486         }
4487         if (ret)
4488                 goto err_state;
4489         return 0;
4490 err_state:
4491         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4492         return ret;
4493 }
4494 EXPORT_SYMBOL(rdma_connect_locked);
4495
4496 /**
4497  * rdma_connect - Initiate an active connection request.
4498  * @id: Connection identifier to connect.
4499  * @conn_param: Connection information used for connected QPs.
4500  *
4501  * Users must have resolved a route for the rdma_cm_id to connect with by having
4502  * called rdma_resolve_route before calling this routine.
4503  *
4504  * This call will either connect to a remote QP or obtain remote QP information
4505  * for unconnected rdma_cm_id's.  The actual operation is based on the
4506  * rdma_cm_id's port space.
4507  */
4508 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4509 {
4510         struct rdma_id_private *id_priv =
4511                 container_of(id, struct rdma_id_private, id);
4512         int ret;
4513
4514         mutex_lock(&id_priv->handler_mutex);
4515         ret = rdma_connect_locked(id, conn_param);
4516         mutex_unlock(&id_priv->handler_mutex);
4517         return ret;
4518 }
4519 EXPORT_SYMBOL(rdma_connect);
4520
4521 /**
4522  * rdma_connect_ece - Initiate an active connection request with ECE data.
4523  * @id: Connection identifier to connect.
4524  * @conn_param: Connection information used for connected QPs.
4525  * @ece: ECE parameters
4526  *
4527  * See rdma_connect() explanation.
4528  */
4529 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4530                      struct rdma_ucm_ece *ece)
4531 {
4532         struct rdma_id_private *id_priv =
4533                 container_of(id, struct rdma_id_private, id);
4534
4535         id_priv->ece.vendor_id = ece->vendor_id;
4536         id_priv->ece.attr_mod = ece->attr_mod;
4537
4538         return rdma_connect(id, conn_param);
4539 }
4540 EXPORT_SYMBOL(rdma_connect_ece);
4541
4542 static int cma_accept_ib(struct rdma_id_private *id_priv,
4543                          struct rdma_conn_param *conn_param)
4544 {
4545         struct ib_cm_rep_param rep;
4546         int ret;
4547
4548         ret = cma_modify_qp_rtr(id_priv, conn_param);
4549         if (ret)
4550                 goto out;
4551
4552         ret = cma_modify_qp_rts(id_priv, conn_param);
4553         if (ret)
4554                 goto out;
4555
4556         memset(&rep, 0, sizeof rep);
4557         rep.qp_num = id_priv->qp_num;
4558         rep.starting_psn = id_priv->seq_num;
4559         rep.private_data = conn_param->private_data;
4560         rep.private_data_len = conn_param->private_data_len;
4561         rep.responder_resources = conn_param->responder_resources;
4562         rep.initiator_depth = conn_param->initiator_depth;
4563         rep.failover_accepted = 0;
4564         rep.flow_control = conn_param->flow_control;
4565         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4566         rep.srq = id_priv->srq ? 1 : 0;
4567         rep.ece.vendor_id = id_priv->ece.vendor_id;
4568         rep.ece.attr_mod = id_priv->ece.attr_mod;
4569
4570         trace_cm_send_rep(id_priv);
4571         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4572 out:
4573         return ret;
4574 }
4575
4576 static int cma_accept_iw(struct rdma_id_private *id_priv,
4577                   struct rdma_conn_param *conn_param)
4578 {
4579         struct iw_cm_conn_param iw_param;
4580         int ret;
4581
4582         if (!conn_param)
4583                 return -EINVAL;
4584
4585         ret = cma_modify_qp_rtr(id_priv, conn_param);
4586         if (ret)
4587                 return ret;
4588
4589         iw_param.ord = conn_param->initiator_depth;
4590         iw_param.ird = conn_param->responder_resources;
4591         iw_param.private_data = conn_param->private_data;
4592         iw_param.private_data_len = conn_param->private_data_len;
4593         if (id_priv->id.qp)
4594                 iw_param.qpn = id_priv->qp_num;
4595         else
4596                 iw_param.qpn = conn_param->qp_num;
4597
4598         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4599 }
4600
4601 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4602                              enum ib_cm_sidr_status status, u32 qkey,
4603                              const void *private_data, int private_data_len)
4604 {
4605         struct ib_cm_sidr_rep_param rep;
4606         int ret;
4607
4608         memset(&rep, 0, sizeof rep);
4609         rep.status = status;
4610         if (status == IB_SIDR_SUCCESS) {
4611                 if (qkey)
4612                         ret = cma_set_qkey(id_priv, qkey);
4613                 else
4614                         ret = cma_set_default_qkey(id_priv);
4615                 if (ret)
4616                         return ret;
4617                 rep.qp_num = id_priv->qp_num;
4618                 rep.qkey = id_priv->qkey;
4619
4620                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4621                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4622         }
4623
4624         rep.private_data = private_data;
4625         rep.private_data_len = private_data_len;
4626
4627         trace_cm_send_sidr_rep(id_priv);
4628         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4629 }
4630
4631 /**
4632  * rdma_accept - Called to accept a connection request or response.
4633  * @id: Connection identifier associated with the request.
4634  * @conn_param: Information needed to establish the connection.  This must be
4635  *   provided if accepting a connection request.  If accepting a connection
4636  *   response, this parameter must be NULL.
4637  *
4638  * Typically, this routine is only called by the listener to accept a connection
4639  * request.  It must also be called on the active side of a connection if the
4640  * user is performing their own QP transitions.
4641  *
4642  * In the case of error, a reject message is sent to the remote side and the
4643  * state of the qp associated with the id is modified to error, such that any
4644  * previously posted receive buffers would be flushed.
4645  *
4646  * This function is for use by kernel ULPs and must be called from under the
4647  * handler callback.
4648  */
4649 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4650 {
4651         struct rdma_id_private *id_priv =
4652                 container_of(id, struct rdma_id_private, id);
4653         int ret;
4654
4655         lockdep_assert_held(&id_priv->handler_mutex);
4656
4657         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4658                 return -EINVAL;
4659
4660         if (!id->qp && conn_param) {
4661                 id_priv->qp_num = conn_param->qp_num;
4662                 id_priv->srq = conn_param->srq;
4663         }
4664
4665         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4666                 if (id->qp_type == IB_QPT_UD) {
4667                         if (conn_param)
4668                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4669                                                         conn_param->qkey,
4670                                                         conn_param->private_data,
4671                                                         conn_param->private_data_len);
4672                         else
4673                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4674                                                         0, NULL, 0);
4675                 } else {
4676                         if (conn_param)
4677                                 ret = cma_accept_ib(id_priv, conn_param);
4678                         else
4679                                 ret = cma_rep_recv(id_priv);
4680                 }
4681         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4682                 ret = cma_accept_iw(id_priv, conn_param);
4683         } else {
4684                 ret = -ENOSYS;
4685         }
4686         if (ret)
4687                 goto reject;
4688
4689         return 0;
4690 reject:
4691         cma_modify_qp_err(id_priv);
4692         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4693         return ret;
4694 }
4695 EXPORT_SYMBOL(rdma_accept);
4696
4697 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4698                     struct rdma_ucm_ece *ece)
4699 {
4700         struct rdma_id_private *id_priv =
4701                 container_of(id, struct rdma_id_private, id);
4702
4703         id_priv->ece.vendor_id = ece->vendor_id;
4704         id_priv->ece.attr_mod = ece->attr_mod;
4705
4706         return rdma_accept(id, conn_param);
4707 }
4708 EXPORT_SYMBOL(rdma_accept_ece);
4709
4710 void rdma_lock_handler(struct rdma_cm_id *id)
4711 {
4712         struct rdma_id_private *id_priv =
4713                 container_of(id, struct rdma_id_private, id);
4714
4715         mutex_lock(&id_priv->handler_mutex);
4716 }
4717 EXPORT_SYMBOL(rdma_lock_handler);
4718
4719 void rdma_unlock_handler(struct rdma_cm_id *id)
4720 {
4721         struct rdma_id_private *id_priv =
4722                 container_of(id, struct rdma_id_private, id);
4723
4724         mutex_unlock(&id_priv->handler_mutex);
4725 }
4726 EXPORT_SYMBOL(rdma_unlock_handler);
4727
4728 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4729 {
4730         struct rdma_id_private *id_priv;
4731         int ret;
4732
4733         id_priv = container_of(id, struct rdma_id_private, id);
4734         if (!id_priv->cm_id.ib)
4735                 return -EINVAL;
4736
4737         switch (id->device->node_type) {
4738         case RDMA_NODE_IB_CA:
4739                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4740                 break;
4741         default:
4742                 ret = 0;
4743                 break;
4744         }
4745         return ret;
4746 }
4747 EXPORT_SYMBOL(rdma_notify);
4748
4749 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4750                 u8 private_data_len, u8 reason)
4751 {
4752         struct rdma_id_private *id_priv;
4753         int ret;
4754
4755         id_priv = container_of(id, struct rdma_id_private, id);
4756         if (!id_priv->cm_id.ib)
4757                 return -EINVAL;
4758
4759         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4760                 if (id->qp_type == IB_QPT_UD) {
4761                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4762                                                 private_data, private_data_len);
4763                 } else {
4764                         trace_cm_send_rej(id_priv);
4765                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4766                                              private_data, private_data_len);
4767                 }
4768         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4769                 ret = iw_cm_reject(id_priv->cm_id.iw,
4770                                    private_data, private_data_len);
4771         } else {
4772                 ret = -ENOSYS;
4773         }
4774
4775         return ret;
4776 }
4777 EXPORT_SYMBOL(rdma_reject);
4778
4779 int rdma_disconnect(struct rdma_cm_id *id)
4780 {
4781         struct rdma_id_private *id_priv;
4782         int ret;
4783
4784         id_priv = container_of(id, struct rdma_id_private, id);
4785         if (!id_priv->cm_id.ib)
4786                 return -EINVAL;
4787
4788         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4789                 ret = cma_modify_qp_err(id_priv);
4790                 if (ret)
4791                         goto out;
4792                 /* Initiate or respond to a disconnect. */
4793                 trace_cm_disconnect(id_priv);
4794                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4795                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4796                                 trace_cm_sent_drep(id_priv);
4797                 } else {
4798                         trace_cm_sent_dreq(id_priv);
4799                 }
4800         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4801                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4802         } else
4803                 ret = -EINVAL;
4804
4805 out:
4806         return ret;
4807 }
4808 EXPORT_SYMBOL(rdma_disconnect);
4809
4810 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4811                               struct ib_sa_multicast *multicast,
4812                               struct rdma_cm_event *event,
4813                               struct cma_multicast *mc)
4814 {
4815         struct rdma_dev_addr *dev_addr;
4816         enum ib_gid_type gid_type;
4817         struct net_device *ndev;
4818
4819         if (status)
4820                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4821                                      status);
4822
4823         event->status = status;
4824         event->param.ud.private_data = mc->context;
4825         if (status) {
4826                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4827                 return;
4828         }
4829
4830         dev_addr = &id_priv->id.route.addr.dev_addr;
4831         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4832         gid_type =
4833                 id_priv->cma_dev
4834                         ->default_gid_type[id_priv->id.port_num -
4835                                            rdma_start_port(
4836                                                    id_priv->cma_dev->device)];
4837
4838         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4839         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4840                                      &multicast->rec, ndev, gid_type,
4841                                      &event->param.ud.ah_attr)) {
4842                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4843                 goto out;
4844         }
4845
4846         event->param.ud.qp_num = 0xFFFFFF;
4847         event->param.ud.qkey = id_priv->qkey;
4848
4849 out:
4850         dev_put(ndev);
4851 }
4852
4853 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4854 {
4855         struct cma_multicast *mc = multicast->context;
4856         struct rdma_id_private *id_priv = mc->id_priv;
4857         struct rdma_cm_event event = {};
4858         int ret = 0;
4859
4860         mutex_lock(&id_priv->handler_mutex);
4861         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4862             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4863                 goto out;
4864
4865         ret = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4866         if (!ret) {
4867                 cma_make_mc_event(status, id_priv, multicast, &event, mc);
4868                 ret = cma_cm_event_handler(id_priv, &event);
4869         }
4870         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4871         WARN_ON(ret);
4872
4873 out:
4874         mutex_unlock(&id_priv->handler_mutex);
4875         return 0;
4876 }
4877
4878 static void cma_set_mgid(struct rdma_id_private *id_priv,
4879                          struct sockaddr *addr, union ib_gid *mgid)
4880 {
4881         unsigned char mc_map[MAX_ADDR_LEN];
4882         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4883         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4884         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4885
4886         if (cma_any_addr(addr)) {
4887                 memset(mgid, 0, sizeof *mgid);
4888         } else if ((addr->sa_family == AF_INET6) &&
4889                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4890                                                                  0xFF10A01B)) {
4891                 /* IPv6 address is an SA assigned MGID. */
4892                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4893         } else if (addr->sa_family == AF_IB) {
4894                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4895         } else if (addr->sa_family == AF_INET6) {
4896                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4897                 if (id_priv->id.ps == RDMA_PS_UDP)
4898                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4899                 *mgid = *(union ib_gid *) (mc_map + 4);
4900         } else {
4901                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4902                 if (id_priv->id.ps == RDMA_PS_UDP)
4903                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4904                 *mgid = *(union ib_gid *) (mc_map + 4);
4905         }
4906 }
4907
4908 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4909                                  struct cma_multicast *mc)
4910 {
4911         struct ib_sa_mcmember_rec rec;
4912         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4913         ib_sa_comp_mask comp_mask;
4914         int ret;
4915
4916         ib_addr_get_mgid(dev_addr, &rec.mgid);
4917         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4918                                      &rec.mgid, &rec);
4919         if (ret)
4920                 return ret;
4921
4922         if (!id_priv->qkey) {
4923                 ret = cma_set_default_qkey(id_priv);
4924                 if (ret)
4925                         return ret;
4926         }
4927
4928         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4929         rec.qkey = cpu_to_be32(id_priv->qkey);
4930         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4931         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4932         rec.join_state = mc->join_state;
4933
4934         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4935                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4936                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4937                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4938                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4939
4940         if (id_priv->id.ps == RDMA_PS_IPOIB)
4941                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4942                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4943                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4944                              IB_SA_MCMEMBER_REC_MTU |
4945                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4946
4947         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4948                                          id_priv->id.port_num, &rec, comp_mask,
4949                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4950         return PTR_ERR_OR_ZERO(mc->sa_mc);
4951 }
4952
4953 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4954                               enum ib_gid_type gid_type)
4955 {
4956         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4957         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4958
4959         if (cma_any_addr(addr)) {
4960                 memset(mgid, 0, sizeof *mgid);
4961         } else if (addr->sa_family == AF_INET6) {
4962                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4963         } else {
4964                 mgid->raw[0] =
4965                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4966                 mgid->raw[1] =
4967                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4968                 mgid->raw[2] = 0;
4969                 mgid->raw[3] = 0;
4970                 mgid->raw[4] = 0;
4971                 mgid->raw[5] = 0;
4972                 mgid->raw[6] = 0;
4973                 mgid->raw[7] = 0;
4974                 mgid->raw[8] = 0;
4975                 mgid->raw[9] = 0;
4976                 mgid->raw[10] = 0xff;
4977                 mgid->raw[11] = 0xff;
4978                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4979         }
4980 }
4981
4982 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4983                                    struct cma_multicast *mc)
4984 {
4985         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4986         int err = 0;
4987         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4988         struct net_device *ndev = NULL;
4989         struct ib_sa_multicast ib = {};
4990         enum ib_gid_type gid_type;
4991         bool send_only;
4992
4993         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4994
4995         if (cma_zero_addr(addr))
4996                 return -EINVAL;
4997
4998         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4999                    rdma_start_port(id_priv->cma_dev->device)];
5000         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
5001
5002         ib.rec.pkey = cpu_to_be16(0xffff);
5003         if (dev_addr->bound_dev_if)
5004                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
5005         if (!ndev)
5006                 return -ENODEV;
5007
5008         ib.rec.rate = IB_RATE_PORT_CURRENT;
5009         ib.rec.hop_limit = 1;
5010         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
5011
5012         if (addr->sa_family == AF_INET) {
5013                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
5014                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
5015                         if (!send_only) {
5016                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
5017                                                     true);
5018                         }
5019                 }
5020         } else {
5021                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
5022                         err = -ENOTSUPP;
5023         }
5024         dev_put(ndev);
5025         if (err || !ib.rec.mtu)
5026                 return err ?: -EINVAL;
5027
5028         if (!id_priv->qkey)
5029                 cma_set_default_qkey(id_priv);
5030
5031         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
5032                     &ib.rec.port_gid);
5033         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
5034         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
5035         queue_work(cma_wq, &mc->iboe_join.work);
5036         return 0;
5037 }
5038
5039 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
5040                         u8 join_state, void *context)
5041 {
5042         struct rdma_id_private *id_priv =
5043                 container_of(id, struct rdma_id_private, id);
5044         struct cma_multicast *mc;
5045         int ret;
5046
5047         /* Not supported for kernel QPs */
5048         if (WARN_ON(id->qp))
5049                 return -EINVAL;
5050
5051         /* ULP is calling this wrong. */
5052         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
5053                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
5054                 return -EINVAL;
5055
5056         if (id_priv->id.qp_type != IB_QPT_UD)
5057                 return -EINVAL;
5058
5059         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
5060         if (!mc)
5061                 return -ENOMEM;
5062
5063         memcpy(&mc->addr, addr, rdma_addr_size(addr));
5064         mc->context = context;
5065         mc->id_priv = id_priv;
5066         mc->join_state = join_state;
5067
5068         if (rdma_protocol_roce(id->device, id->port_num)) {
5069                 ret = cma_iboe_join_multicast(id_priv, mc);
5070                 if (ret)
5071                         goto out_err;
5072         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
5073                 ret = cma_join_ib_multicast(id_priv, mc);
5074                 if (ret)
5075                         goto out_err;
5076         } else {
5077                 ret = -ENOSYS;
5078                 goto out_err;
5079         }
5080
5081         spin_lock(&id_priv->lock);
5082         list_add(&mc->list, &id_priv->mc_list);
5083         spin_unlock(&id_priv->lock);
5084
5085         return 0;
5086 out_err:
5087         kfree(mc);
5088         return ret;
5089 }
5090 EXPORT_SYMBOL(rdma_join_multicast);
5091
5092 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
5093 {
5094         struct rdma_id_private *id_priv;
5095         struct cma_multicast *mc;
5096
5097         id_priv = container_of(id, struct rdma_id_private, id);
5098         spin_lock_irq(&id_priv->lock);
5099         list_for_each_entry(mc, &id_priv->mc_list, list) {
5100                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
5101                         continue;
5102                 list_del(&mc->list);
5103                 spin_unlock_irq(&id_priv->lock);
5104
5105                 WARN_ON(id_priv->cma_dev->device != id->device);
5106                 destroy_mc(id_priv, mc);
5107                 return;
5108         }
5109         spin_unlock_irq(&id_priv->lock);
5110 }
5111 EXPORT_SYMBOL(rdma_leave_multicast);
5112
5113 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
5114 {
5115         struct rdma_dev_addr *dev_addr;
5116         struct cma_work *work;
5117
5118         dev_addr = &id_priv->id.route.addr.dev_addr;
5119
5120         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
5121             (net_eq(dev_net(ndev), dev_addr->net)) &&
5122             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
5123                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
5124                         ndev->name, &id_priv->id);
5125                 work = kzalloc(sizeof *work, GFP_KERNEL);
5126                 if (!work)
5127                         return -ENOMEM;
5128
5129                 INIT_WORK(&work->work, cma_work_handler);
5130                 work->id = id_priv;
5131                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
5132                 cma_id_get(id_priv);
5133                 queue_work(cma_wq, &work->work);
5134         }
5135
5136         return 0;
5137 }
5138
5139 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
5140                                void *ptr)
5141 {
5142         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
5143         struct cma_device *cma_dev;
5144         struct rdma_id_private *id_priv;
5145         int ret = NOTIFY_DONE;
5146
5147         if (event != NETDEV_BONDING_FAILOVER)
5148                 return NOTIFY_DONE;
5149
5150         if (!netif_is_bond_master(ndev))
5151                 return NOTIFY_DONE;
5152
5153         mutex_lock(&lock);
5154         list_for_each_entry(cma_dev, &dev_list, list)
5155                 list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
5156                         ret = cma_netdev_change(ndev, id_priv);
5157                         if (ret)
5158                                 goto out;
5159                 }
5160
5161 out:
5162         mutex_unlock(&lock);
5163         return ret;
5164 }
5165
5166 static void cma_netevent_work_handler(struct work_struct *_work)
5167 {
5168         struct rdma_id_private *id_priv =
5169                 container_of(_work, struct rdma_id_private, id.net_work);
5170         struct rdma_cm_event event = {};
5171
5172         mutex_lock(&id_priv->handler_mutex);
5173
5174         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
5175             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
5176                 goto out_unlock;
5177
5178         event.event = RDMA_CM_EVENT_UNREACHABLE;
5179         event.status = -ETIMEDOUT;
5180
5181         if (cma_cm_event_handler(id_priv, &event)) {
5182                 __acquire(&id_priv->handler_mutex);
5183                 id_priv->cm_id.ib = NULL;
5184                 cma_id_put(id_priv);
5185                 destroy_id_handler_unlock(id_priv);
5186                 return;
5187         }
5188
5189 out_unlock:
5190         mutex_unlock(&id_priv->handler_mutex);
5191         cma_id_put(id_priv);
5192 }
5193
5194 static int cma_netevent_callback(struct notifier_block *self,
5195                                  unsigned long event, void *ctx)
5196 {
5197         struct id_table_entry *ips_node = NULL;
5198         struct rdma_id_private *current_id;
5199         struct neighbour *neigh = ctx;
5200         unsigned long flags;
5201
5202         if (event != NETEVENT_NEIGH_UPDATE)
5203                 return NOTIFY_DONE;
5204
5205         spin_lock_irqsave(&id_table_lock, flags);
5206         if (neigh->tbl->family == AF_INET6) {
5207                 struct sockaddr_in6 neigh_sock_6;
5208
5209                 neigh_sock_6.sin6_family = AF_INET6;
5210                 neigh_sock_6.sin6_addr = *(struct in6_addr *)neigh->primary_key;
5211                 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5212                                              (struct sockaddr *)&neigh_sock_6);
5213         } else if (neigh->tbl->family == AF_INET) {
5214                 struct sockaddr_in neigh_sock_4;
5215
5216                 neigh_sock_4.sin_family = AF_INET;
5217                 neigh_sock_4.sin_addr.s_addr = *(__be32 *)(neigh->primary_key);
5218                 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5219                                              (struct sockaddr *)&neigh_sock_4);
5220         } else
5221                 goto out;
5222
5223         if (!ips_node)
5224                 goto out;
5225
5226         list_for_each_entry(current_id, &ips_node->id_list, id_list_entry) {
5227                 if (!memcmp(current_id->id.route.addr.dev_addr.dst_dev_addr,
5228                            neigh->ha, ETH_ALEN))
5229                         continue;
5230                 INIT_WORK(&current_id->id.net_work, cma_netevent_work_handler);
5231                 cma_id_get(current_id);
5232                 queue_work(cma_wq, &current_id->id.net_work);
5233         }
5234 out:
5235         spin_unlock_irqrestore(&id_table_lock, flags);
5236         return NOTIFY_DONE;
5237 }
5238
5239 static struct notifier_block cma_nb = {
5240         .notifier_call = cma_netdev_callback
5241 };
5242
5243 static struct notifier_block cma_netevent_cb = {
5244         .notifier_call = cma_netevent_callback
5245 };
5246
5247 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
5248 {
5249         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
5250         enum rdma_cm_state state;
5251         unsigned long flags;
5252
5253         mutex_lock(&id_priv->handler_mutex);
5254         /* Record that we want to remove the device */
5255         spin_lock_irqsave(&id_priv->lock, flags);
5256         state = id_priv->state;
5257         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
5258                 spin_unlock_irqrestore(&id_priv->lock, flags);
5259                 mutex_unlock(&id_priv->handler_mutex);
5260                 cma_id_put(id_priv);
5261                 return;
5262         }
5263         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
5264         spin_unlock_irqrestore(&id_priv->lock, flags);
5265
5266         if (cma_cm_event_handler(id_priv, &event)) {
5267                 /*
5268                  * At this point the ULP promises it won't call
5269                  * rdma_destroy_id() concurrently
5270                  */
5271                 cma_id_put(id_priv);
5272                 mutex_unlock(&id_priv->handler_mutex);
5273                 trace_cm_id_destroy(id_priv);
5274                 _destroy_id(id_priv, state);
5275                 return;
5276         }
5277         mutex_unlock(&id_priv->handler_mutex);
5278
5279         /*
5280          * If this races with destroy then the thread that first assigns state
5281          * to a destroying does the cancel.
5282          */
5283         cma_cancel_operation(id_priv, state);
5284         cma_id_put(id_priv);
5285 }
5286
5287 static void cma_process_remove(struct cma_device *cma_dev)
5288 {
5289         mutex_lock(&lock);
5290         while (!list_empty(&cma_dev->id_list)) {
5291                 struct rdma_id_private *id_priv = list_first_entry(
5292                         &cma_dev->id_list, struct rdma_id_private, device_item);
5293
5294                 list_del_init(&id_priv->listen_item);
5295                 list_del_init(&id_priv->device_item);
5296                 cma_id_get(id_priv);
5297                 mutex_unlock(&lock);
5298
5299                 cma_send_device_removal_put(id_priv);
5300
5301                 mutex_lock(&lock);
5302         }
5303         mutex_unlock(&lock);
5304
5305         cma_dev_put(cma_dev);
5306         wait_for_completion(&cma_dev->comp);
5307 }
5308
5309 static bool cma_supported(struct ib_device *device)
5310 {
5311         u32 i;
5312
5313         rdma_for_each_port(device, i) {
5314                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
5315                         return true;
5316         }
5317         return false;
5318 }
5319
5320 static int cma_add_one(struct ib_device *device)
5321 {
5322         struct rdma_id_private *to_destroy;
5323         struct cma_device *cma_dev;
5324         struct rdma_id_private *id_priv;
5325         unsigned long supported_gids = 0;
5326         int ret;
5327         u32 i;
5328
5329         if (!cma_supported(device))
5330                 return -EOPNOTSUPP;
5331
5332         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5333         if (!cma_dev)
5334                 return -ENOMEM;
5335
5336         cma_dev->device = device;
5337         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5338                                             sizeof(*cma_dev->default_gid_type),
5339                                             GFP_KERNEL);
5340         if (!cma_dev->default_gid_type) {
5341                 ret = -ENOMEM;
5342                 goto free_cma_dev;
5343         }
5344
5345         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5346                                             sizeof(*cma_dev->default_roce_tos),
5347                                             GFP_KERNEL);
5348         if (!cma_dev->default_roce_tos) {
5349                 ret = -ENOMEM;
5350                 goto free_gid_type;
5351         }
5352
5353         rdma_for_each_port (device, i) {
5354                 supported_gids = roce_gid_type_mask_support(device, i);
5355                 WARN_ON(!supported_gids);
5356                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5357                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5358                                 CMA_PREFERRED_ROCE_GID_TYPE;
5359                 else
5360                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5361                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5362                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5363         }
5364
5365         init_completion(&cma_dev->comp);
5366         refcount_set(&cma_dev->refcount, 1);
5367         INIT_LIST_HEAD(&cma_dev->id_list);
5368         ib_set_client_data(device, &cma_client, cma_dev);
5369
5370         mutex_lock(&lock);
5371         list_add_tail(&cma_dev->list, &dev_list);
5372         list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5373                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5374                 if (ret)
5375                         goto free_listen;
5376         }
5377         mutex_unlock(&lock);
5378
5379         trace_cm_add_one(device);
5380         return 0;
5381
5382 free_listen:
5383         list_del(&cma_dev->list);
5384         mutex_unlock(&lock);
5385
5386         /* cma_process_remove() will delete to_destroy */
5387         cma_process_remove(cma_dev);
5388         kfree(cma_dev->default_roce_tos);
5389 free_gid_type:
5390         kfree(cma_dev->default_gid_type);
5391
5392 free_cma_dev:
5393         kfree(cma_dev);
5394         return ret;
5395 }
5396
5397 static void cma_remove_one(struct ib_device *device, void *client_data)
5398 {
5399         struct cma_device *cma_dev = client_data;
5400
5401         trace_cm_remove_one(device);
5402
5403         mutex_lock(&lock);
5404         list_del(&cma_dev->list);
5405         mutex_unlock(&lock);
5406
5407         cma_process_remove(cma_dev);
5408         kfree(cma_dev->default_roce_tos);
5409         kfree(cma_dev->default_gid_type);
5410         kfree(cma_dev);
5411 }
5412
5413 static int cma_init_net(struct net *net)
5414 {
5415         struct cma_pernet *pernet = cma_pernet(net);
5416
5417         xa_init(&pernet->tcp_ps);
5418         xa_init(&pernet->udp_ps);
5419         xa_init(&pernet->ipoib_ps);
5420         xa_init(&pernet->ib_ps);
5421
5422         return 0;
5423 }
5424
5425 static void cma_exit_net(struct net *net)
5426 {
5427         struct cma_pernet *pernet = cma_pernet(net);
5428
5429         WARN_ON(!xa_empty(&pernet->tcp_ps));
5430         WARN_ON(!xa_empty(&pernet->udp_ps));
5431         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5432         WARN_ON(!xa_empty(&pernet->ib_ps));
5433 }
5434
5435 static struct pernet_operations cma_pernet_operations = {
5436         .init = cma_init_net,
5437         .exit = cma_exit_net,
5438         .id = &cma_pernet_id,
5439         .size = sizeof(struct cma_pernet),
5440 };
5441
5442 static int __init cma_init(void)
5443 {
5444         int ret;
5445
5446         /*
5447          * There is a rare lock ordering dependency in cma_netdev_callback()
5448          * that only happens when bonding is enabled. Teach lockdep that rtnl
5449          * must never be nested under lock so it can find these without having
5450          * to test with bonding.
5451          */
5452         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5453                 rtnl_lock();
5454                 mutex_lock(&lock);
5455                 mutex_unlock(&lock);
5456                 rtnl_unlock();
5457         }
5458
5459         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5460         if (!cma_wq)
5461                 return -ENOMEM;
5462
5463         ret = register_pernet_subsys(&cma_pernet_operations);
5464         if (ret)
5465                 goto err_wq;
5466
5467         ib_sa_register_client(&sa_client);
5468         register_netdevice_notifier(&cma_nb);
5469         register_netevent_notifier(&cma_netevent_cb);
5470
5471         ret = ib_register_client(&cma_client);
5472         if (ret)
5473                 goto err;
5474
5475         ret = cma_configfs_init();
5476         if (ret)
5477                 goto err_ib;
5478
5479         return 0;
5480
5481 err_ib:
5482         ib_unregister_client(&cma_client);
5483 err:
5484         unregister_netevent_notifier(&cma_netevent_cb);
5485         unregister_netdevice_notifier(&cma_nb);
5486         ib_sa_unregister_client(&sa_client);
5487         unregister_pernet_subsys(&cma_pernet_operations);
5488 err_wq:
5489         destroy_workqueue(cma_wq);
5490         return ret;
5491 }
5492
5493 static void __exit cma_cleanup(void)
5494 {
5495         cma_configfs_exit();
5496         ib_unregister_client(&cma_client);
5497         unregister_netevent_notifier(&cma_netevent_cb);
5498         unregister_netdevice_notifier(&cma_nb);
5499         ib_sa_unregister_client(&sa_client);
5500         unregister_pernet_subsys(&cma_pernet_operations);
5501         destroy_workqueue(cma_wq);
5502 }
5503
5504 module_init(cma_init);
5505 module_exit(cma_cleanup);