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 18
  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                 return memcmp((char *)&((struct sockaddr_in *)sa)->sin_addr,
 484                               (char *)&((struct sockaddr_in *)sb)->sin_addr,
 485                               sizeof(((struct sockaddr_in *)sa)->sin_addr));
 486
 487         return ipv6_addr_cmp(&((struct sockaddr_in6 *)sa)->sin6_addr,
 488                              &((struct sockaddr_in6 *)sb)->sin6_addr);
 489 }
 490
 491 static int cma_add_id_to_tree(struct rdma_id_private *node_id_priv)
 492 {
 493         struct rb_node **new, *parent = NULL;
 494         struct id_table_entry *this, *node;
 495         unsigned long flags;
 496         int result;
 497
 498         node = kzalloc(sizeof(*node), GFP_KERNEL);
 499         if (!node)
 500                 return -ENOMEM;
 501
 502         spin_lock_irqsave(&id_table_lock, flags);
 503         new = &id_table.rb_node;
 504         while (*new) {
 505                 this = container_of(*new, struct id_table_entry, rb_node);
 506                 result = compare_netdev_and_ip(
 507                         node_id_priv->id.route.addr.dev_addr.bound_dev_if,
 508                         cma_dst_addr(node_id_priv), this);
 509
 510                 parent = *new;
 511                 if (result < 0)
 512                         new = &((*new)->rb_left);
 513                 else if (result > 0)
 514                         new = &((*new)->rb_right);
 515                 else {
 516                         list_add_tail(&node_id_priv->id_list_entry,
 517                                       &this->id_list);
 518                         kfree(node);
 519                         goto unlock;
 520                 }
 521         }
 522
 523         INIT_LIST_HEAD(&node->id_list);
 524         list_add_tail(&node_id_priv->id_list_entry, &node->id_list);
 525
 526         rb_link_node(&node->rb_node, parent, new);
 527         rb_insert_color(&node->rb_node, &id_table);
 528
 529 unlock:
 530         spin_unlock_irqrestore(&id_table_lock, flags);
 531         return 0;
 532 }
 533
 534 static struct id_table_entry *
 535 node_from_ndev_ip(struct rb_root *root, int ifindex, struct sockaddr *sa)
 536 {
 537         struct rb_node *node = root->rb_node;
 538         struct id_table_entry *data;
 539         int result;
 540
 541         while (node) {
 542                 data = container_of(node, struct id_table_entry, rb_node);
 543                 result = compare_netdev_and_ip(ifindex, sa, data);
 544                 if (result < 0)
 545                         node = node->rb_left;
 546                 else if (result > 0)
 547                         node = node->rb_right;
 548                 else
 549                         return data;
 550         }
 551
 552         return NULL;
 553 }
 554
 555 static void cma_remove_id_from_tree(struct rdma_id_private *id_priv)
 556 {
 557         struct id_table_entry *data;
 558         unsigned long flags;
 559
 560         spin_lock_irqsave(&id_table_lock, flags);
 561         if (list_empty(&id_priv->id_list_entry))
 562                 goto out;
 563
 564         data = node_from_ndev_ip(&id_table,
 565                                  id_priv->id.route.addr.dev_addr.bound_dev_if,
 566                                  cma_dst_addr(id_priv));
 567         if (!data)
 568                 goto out;
 569
 570         list_del_init(&id_priv->id_list_entry);
 571         if (list_empty(&data->id_list)) {
 572                 rb_erase(&data->rb_node, &id_table);
 573                 kfree(data);
 574         }
 575 out:
 576         spin_unlock_irqrestore(&id_table_lock, flags);
 577 }
 578
 579 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 580                                struct cma_device *cma_dev)
 581 {
 582         cma_dev_get(cma_dev);
 583         id_priv->cma_dev = cma_dev;
 584         id_priv->id.device = cma_dev->device;
 585         id_priv->id.route.addr.dev_addr.transport =
 586                 rdma_node_get_transport(cma_dev->device->node_type);
 587         list_add_tail(&id_priv->device_item, &cma_dev->id_list);
 588
 589         trace_cm_id_attach(id_priv, cma_dev->device);
 590 }
 591
 592 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 593                               struct cma_device *cma_dev)
 594 {
 595         _cma_attach_to_dev(id_priv, cma_dev);
 596         id_priv->gid_type =
 597                 cma_dev->default_gid_type[id_priv->id.port_num -
 598                                           rdma_start_port(cma_dev->device)];
 599 }
 600
 601 static void cma_release_dev(struct rdma_id_private *id_priv)
 602 {
 603         mutex_lock(&lock);
 604         list_del_init(&id_priv->device_item);
 605         cma_dev_put(id_priv->cma_dev);
 606         id_priv->cma_dev = NULL;
 607         id_priv->id.device = NULL;
 608         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 609                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 610                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 611         }
 612         mutex_unlock(&lock);
 613 }
 614
 615 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 616 {
 617         return id_priv->id.route.addr.src_addr.ss_family;
 618 }
 619
 620 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 621 {
 622         struct ib_sa_mcmember_rec rec;
 623         int ret = 0;
 624
 625         if (id_priv->qkey) {
 626                 if (qkey && id_priv->qkey != qkey)
 627                         return -EINVAL;
 628                 return 0;
 629         }
 630
 631         if (qkey) {
 632                 id_priv->qkey = qkey;
 633                 return 0;
 634         }
 635
 636         switch (id_priv->id.ps) {
 637         case RDMA_PS_UDP:
 638         case RDMA_PS_IB:
 639                 id_priv->qkey = RDMA_UDP_QKEY;
 640                 break;
 641         case RDMA_PS_IPOIB:
 642                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 643                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 644                                              id_priv->id.port_num, &rec.mgid,
 645                                              &rec);
 646                 if (!ret)
 647                         id_priv->qkey = be32_to_cpu(rec.qkey);
 648                 break;
 649         default:
 650                 break;
 651         }
 652         return ret;
 653 }
 654
 655 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 656 {
 657         dev_addr->dev_type = ARPHRD_INFINIBAND;
 658         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 659         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 660 }
 661
 662 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 663 {
 664         int ret;
 665
 666         if (addr->sa_family != AF_IB) {
 667                 ret = rdma_translate_ip(addr, dev_addr);
 668         } else {
 669                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 670                 ret = 0;
 671         }
 672
 673         return ret;
 674 }
 675
 676 static const struct ib_gid_attr *
 677 cma_validate_port(struct ib_device *device, u32 port,
 678                   enum ib_gid_type gid_type,
 679                   union ib_gid *gid,
 680                   struct rdma_id_private *id_priv)
 681 {
 682         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 683         int bound_if_index = dev_addr->bound_dev_if;
 684         const struct ib_gid_attr *sgid_attr;
 685         int dev_type = dev_addr->dev_type;
 686         struct net_device *ndev = NULL;
 687
 688         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 689                 return ERR_PTR(-ENODEV);
 690
 691         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 692                 return ERR_PTR(-ENODEV);
 693
 694         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 695                 return ERR_PTR(-ENODEV);
 696
 697         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 698                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 699                 if (!ndev)
 700                         return ERR_PTR(-ENODEV);
 701         } else {
 702                 gid_type = IB_GID_TYPE_IB;
 703         }
 704
 705         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 706         if (ndev)
 707                 dev_put(ndev);
 708         return sgid_attr;
 709 }
 710
 711 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 712                                const struct ib_gid_attr *sgid_attr)
 713 {
 714         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 715         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 716 }
 717
 718 /**
 719  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 720  * based on source ip address.
 721  * @id_priv:    cm_id which should be bound to cma device
 722  *
 723  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 724  * based on source IP address. It returns 0 on success or error code otherwise.
 725  * It is applicable to active and passive side cm_id.
 726  */
 727 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 728 {
 729         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 730         const struct ib_gid_attr *sgid_attr;
 731         union ib_gid gid, iboe_gid, *gidp;
 732         struct cma_device *cma_dev;
 733         enum ib_gid_type gid_type;
 734         int ret = -ENODEV;
 735         u32 port;
 736
 737         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 738             id_priv->id.ps == RDMA_PS_IPOIB)
 739                 return -EINVAL;
 740
 741         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 742                     &iboe_gid);
 743
 744         memcpy(&gid, dev_addr->src_dev_addr +
 745                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 746
 747         mutex_lock(&lock);
 748         list_for_each_entry(cma_dev, &dev_list, list) {
 749                 rdma_for_each_port (cma_dev->device, port) {
 750                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 751                                &iboe_gid : &gid;
 752                         gid_type = cma_dev->default_gid_type[port - 1];
 753                         sgid_attr = cma_validate_port(cma_dev->device, port,
 754                                                       gid_type, gidp, id_priv);
 755                         if (!IS_ERR(sgid_attr)) {
 756                                 id_priv->id.port_num = port;
 757                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 758                                 cma_attach_to_dev(id_priv, cma_dev);
 759                                 ret = 0;
 760                                 goto out;
 761                         }
 762                 }
 763         }
 764 out:
 765         mutex_unlock(&lock);
 766         return ret;
 767 }
 768
 769 /**
 770  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 771  * @id_priv:            cm id to bind to cma device
 772  * @listen_id_priv:     listener cm id to match against
 773  * @req:                Pointer to req structure containaining incoming
 774  *                      request information
 775  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 776  * rdma device matches for listen_id and incoming request. It also verifies
 777  * that a GID table entry is present for the source address.
 778  * Returns 0 on success, or returns error code otherwise.
 779  */
 780 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 781                               const struct rdma_id_private *listen_id_priv,
 782                               struct cma_req_info *req)
 783 {
 784         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 785         const struct ib_gid_attr *sgid_attr;
 786         enum ib_gid_type gid_type;
 787         union ib_gid gid;
 788
 789         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 790             id_priv->id.ps == RDMA_PS_IPOIB)
 791                 return -EINVAL;
 792
 793         if (rdma_protocol_roce(req->device, req->port))
 794                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 795                             &gid);
 796         else
 797                 memcpy(&gid, dev_addr->src_dev_addr +
 798                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 799
 800         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 801         sgid_attr = cma_validate_port(req->device, req->port,
 802                                       gid_type, &gid, id_priv);
 803         if (IS_ERR(sgid_attr))
 804                 return PTR_ERR(sgid_attr);
 805
 806         id_priv->id.port_num = req->port;
 807         cma_bind_sgid_attr(id_priv, sgid_attr);
 808         /* Need to acquire lock to protect against reader
 809          * of cma_dev->id_list such as cma_netdev_callback() and
 810          * cma_process_remove().
 811          */
 812         mutex_lock(&lock);
 813         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 814         mutex_unlock(&lock);
 815         rdma_restrack_add(&id_priv->res);
 816         return 0;
 817 }
 818
 819 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 820                               const struct rdma_id_private *listen_id_priv)
 821 {
 822         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 823         const struct ib_gid_attr *sgid_attr;
 824         struct cma_device *cma_dev;
 825         enum ib_gid_type gid_type;
 826         int ret = -ENODEV;
 827         union ib_gid gid;
 828         u32 port;
 829
 830         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 831             id_priv->id.ps == RDMA_PS_IPOIB)
 832                 return -EINVAL;
 833
 834         memcpy(&gid, dev_addr->src_dev_addr +
 835                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 836
 837         mutex_lock(&lock);
 838
 839         cma_dev = listen_id_priv->cma_dev;
 840         port = listen_id_priv->id.port_num;
 841         gid_type = listen_id_priv->gid_type;
 842         sgid_attr = cma_validate_port(cma_dev->device, port,
 843                                       gid_type, &gid, id_priv);
 844         if (!IS_ERR(sgid_attr)) {
 845                 id_priv->id.port_num = port;
 846                 cma_bind_sgid_attr(id_priv, sgid_attr);
 847                 ret = 0;
 848                 goto out;
 849         }
 850
 851         list_for_each_entry(cma_dev, &dev_list, list) {
 852                 rdma_for_each_port (cma_dev->device, port) {
 853                         if (listen_id_priv->cma_dev == cma_dev &&
 854                             listen_id_priv->id.port_num == port)
 855                                 continue;
 856
 857                         gid_type = cma_dev->default_gid_type[port - 1];
 858                         sgid_attr = cma_validate_port(cma_dev->device, port,
 859                                                       gid_type, &gid, id_priv);
 860                         if (!IS_ERR(sgid_attr)) {
 861                                 id_priv->id.port_num = port;
 862                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 863                                 ret = 0;
 864                                 goto out;
 865                         }
 866                 }
 867         }
 868
 869 out:
 870         if (!ret) {
 871                 cma_attach_to_dev(id_priv, cma_dev);
 872                 rdma_restrack_add(&id_priv->res);
 873         }
 874
 875         mutex_unlock(&lock);
 876         return ret;
 877 }
 878
 879 /*
 880  * Select the source IB device and address to reach the destination IB address.
 881  */
 882 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 883 {
 884         struct cma_device *cma_dev, *cur_dev;
 885         struct sockaddr_ib *addr;
 886         union ib_gid gid, sgid, *dgid;
 887         unsigned int p;
 888         u16 pkey, index;
 889         enum ib_port_state port_state;
 890         int ret;
 891         int i;
 892
 893         cma_dev = NULL;
 894         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 895         dgid = (union ib_gid *) &addr->sib_addr;
 896         pkey = ntohs(addr->sib_pkey);
 897
 898         mutex_lock(&lock);
 899         list_for_each_entry(cur_dev, &dev_list, list) {
 900                 rdma_for_each_port (cur_dev->device, p) {
 901                         if (!rdma_cap_af_ib(cur_dev->device, p))
 902                                 continue;
 903
 904                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 905                                 continue;
 906
 907                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 908                                 continue;
 909
 910                         for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
 911                              ++i) {
 912                                 ret = rdma_query_gid(cur_dev->device, p, i,
 913                                                      &gid);
 914                                 if (ret)
 915                                         continue;
 916
 917                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 918                                         cma_dev = cur_dev;
 919                                         sgid = gid;
 920                                         id_priv->id.port_num = p;
 921                                         goto found;
 922                                 }
 923
 924                                 if (!cma_dev && (gid.global.subnet_prefix ==
 925                                     dgid->global.subnet_prefix) &&
 926                                     port_state == IB_PORT_ACTIVE) {
 927                                         cma_dev = cur_dev;
 928                                         sgid = gid;
 929                                         id_priv->id.port_num = p;
 930                                         goto found;
 931                                 }
 932                         }
 933                 }
 934         }
 935         mutex_unlock(&lock);
 936         return -ENODEV;
 937
 938 found:
 939         cma_attach_to_dev(id_priv, cma_dev);
 940         rdma_restrack_add(&id_priv->res);
 941         mutex_unlock(&lock);
 942         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 943         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 944         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 945         return 0;
 946 }
 947
 948 static void cma_id_get(struct rdma_id_private *id_priv)
 949 {
 950         refcount_inc(&id_priv->refcount);
 951 }
 952
 953 static void cma_id_put(struct rdma_id_private *id_priv)
 954 {
 955         if (refcount_dec_and_test(&id_priv->refcount))
 956                 complete(&id_priv->comp);
 957 }
 958
 959 static struct rdma_id_private *
 960 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 961                  void *context, enum rdma_ucm_port_space ps,
 962                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 963 {
 964         struct rdma_id_private *id_priv;
 965
 966         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 967         if (!id_priv)
 968                 return ERR_PTR(-ENOMEM);
 969
 970         id_priv->state = RDMA_CM_IDLE;
 971         id_priv->id.context = context;
 972         id_priv->id.event_handler = event_handler;
 973         id_priv->id.ps = ps;
 974         id_priv->id.qp_type = qp_type;
 975         id_priv->tos_set = false;
 976         id_priv->timeout_set = false;
 977         id_priv->min_rnr_timer_set = false;
 978         id_priv->gid_type = IB_GID_TYPE_IB;
 979         spin_lock_init(&id_priv->lock);
 980         mutex_init(&id_priv->qp_mutex);
 981         init_completion(&id_priv->comp);
 982         refcount_set(&id_priv->refcount, 1);
 983         mutex_init(&id_priv->handler_mutex);
 984         INIT_LIST_HEAD(&id_priv->device_item);
 985         INIT_LIST_HEAD(&id_priv->id_list_entry);
 986         INIT_LIST_HEAD(&id_priv->listen_list);
 987         INIT_LIST_HEAD(&id_priv->mc_list);
 988         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 989         id_priv->id.route.addr.dev_addr.net = get_net(net);
 990         id_priv->seq_num &= 0x00ffffff;
 991
 992         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 993         if (parent)
 994                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 995
 996         return id_priv;
 997 }
 998
 999 struct rdma_cm_id *
1000 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
1001                         void *context, enum rdma_ucm_port_space ps,
1002                         enum ib_qp_type qp_type, const char *caller)
1003 {
1004         struct rdma_id_private *ret;
1005
1006         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
1007         if (IS_ERR(ret))
1008                 return ERR_CAST(ret);
1009
1010         rdma_restrack_set_name(&ret->res, caller);
1011         return &ret->id;
1012 }
1013 EXPORT_SYMBOL(__rdma_create_kernel_id);
1014
1015 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
1016                                        void *context,
1017                                        enum rdma_ucm_port_space ps,
1018                                        enum ib_qp_type qp_type)
1019 {
1020         struct rdma_id_private *ret;
1021
1022         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
1023                                ps, qp_type, NULL);
1024         if (IS_ERR(ret))
1025                 return ERR_CAST(ret);
1026
1027         rdma_restrack_set_name(&ret->res, NULL);
1028         return &ret->id;
1029 }
1030 EXPORT_SYMBOL(rdma_create_user_id);
1031
1032 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1033 {
1034         struct ib_qp_attr qp_attr;
1035         int qp_attr_mask, ret;
1036
1037         qp_attr.qp_state = IB_QPS_INIT;
1038         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1039         if (ret)
1040                 return ret;
1041
1042         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1043         if (ret)
1044                 return ret;
1045
1046         qp_attr.qp_state = IB_QPS_RTR;
1047         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
1048         if (ret)
1049                 return ret;
1050
1051         qp_attr.qp_state = IB_QPS_RTS;
1052         qp_attr.sq_psn = 0;
1053         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
1054
1055         return ret;
1056 }
1057
1058 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1059 {
1060         struct ib_qp_attr qp_attr;
1061         int qp_attr_mask, ret;
1062
1063         qp_attr.qp_state = IB_QPS_INIT;
1064         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1065         if (ret)
1066                 return ret;
1067
1068         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1069 }
1070
1071 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
1072                    struct ib_qp_init_attr *qp_init_attr)
1073 {
1074         struct rdma_id_private *id_priv;
1075         struct ib_qp *qp;
1076         int ret;
1077
1078         id_priv = container_of(id, struct rdma_id_private, id);
1079         if (id->device != pd->device) {
1080                 ret = -EINVAL;
1081                 goto out_err;
1082         }
1083
1084         qp_init_attr->port_num = id->port_num;
1085         qp = ib_create_qp(pd, qp_init_attr);
1086         if (IS_ERR(qp)) {
1087                 ret = PTR_ERR(qp);
1088                 goto out_err;
1089         }
1090
1091         if (id->qp_type == IB_QPT_UD)
1092                 ret = cma_init_ud_qp(id_priv, qp);
1093         else
1094                 ret = cma_init_conn_qp(id_priv, qp);
1095         if (ret)
1096                 goto out_destroy;
1097
1098         id->qp = qp;
1099         id_priv->qp_num = qp->qp_num;
1100         id_priv->srq = (qp->srq != NULL);
1101         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
1102         return 0;
1103 out_destroy:
1104         ib_destroy_qp(qp);
1105 out_err:
1106         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
1107         return ret;
1108 }
1109 EXPORT_SYMBOL(rdma_create_qp);
1110
1111 void rdma_destroy_qp(struct rdma_cm_id *id)
1112 {
1113         struct rdma_id_private *id_priv;
1114
1115         id_priv = container_of(id, struct rdma_id_private, id);
1116         trace_cm_qp_destroy(id_priv);
1117         mutex_lock(&id_priv->qp_mutex);
1118         ib_destroy_qp(id_priv->id.qp);
1119         id_priv->id.qp = NULL;
1120         mutex_unlock(&id_priv->qp_mutex);
1121 }
1122 EXPORT_SYMBOL(rdma_destroy_qp);
1123
1124 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1125                              struct rdma_conn_param *conn_param)
1126 {
1127         struct ib_qp_attr qp_attr;
1128         int qp_attr_mask, ret;
1129
1130         mutex_lock(&id_priv->qp_mutex);
1131         if (!id_priv->id.qp) {
1132                 ret = 0;
1133                 goto out;
1134         }
1135
1136         /* Need to update QP attributes from default values. */
1137         qp_attr.qp_state = IB_QPS_INIT;
1138         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1139         if (ret)
1140                 goto out;
1141
1142         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1143         if (ret)
1144                 goto out;
1145
1146         qp_attr.qp_state = IB_QPS_RTR;
1147         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1148         if (ret)
1149                 goto out;
1150
1151         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1152
1153         if (conn_param)
1154                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1155         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1156 out:
1157         mutex_unlock(&id_priv->qp_mutex);
1158         return ret;
1159 }
1160
1161 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1162                              struct rdma_conn_param *conn_param)
1163 {
1164         struct ib_qp_attr qp_attr;
1165         int qp_attr_mask, ret;
1166
1167         mutex_lock(&id_priv->qp_mutex);
1168         if (!id_priv->id.qp) {
1169                 ret = 0;
1170                 goto out;
1171         }
1172
1173         qp_attr.qp_state = IB_QPS_RTS;
1174         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1175         if (ret)
1176                 goto out;
1177
1178         if (conn_param)
1179                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1180         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1181 out:
1182         mutex_unlock(&id_priv->qp_mutex);
1183         return ret;
1184 }
1185
1186 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1187 {
1188         struct ib_qp_attr qp_attr;
1189         int ret;
1190
1191         mutex_lock(&id_priv->qp_mutex);
1192         if (!id_priv->id.qp) {
1193                 ret = 0;
1194                 goto out;
1195         }
1196
1197         qp_attr.qp_state = IB_QPS_ERR;
1198         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1199 out:
1200         mutex_unlock(&id_priv->qp_mutex);
1201         return ret;
1202 }
1203
1204 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1205                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1206 {
1207         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1208         int ret;
1209         u16 pkey;
1210
1211         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1212                 pkey = 0xffff;
1213         else
1214                 pkey = ib_addr_get_pkey(dev_addr);
1215
1216         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1217                                   pkey, &qp_attr->pkey_index);
1218         if (ret)
1219                 return ret;
1220
1221         qp_attr->port_num = id_priv->id.port_num;
1222         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1223
1224         if (id_priv->id.qp_type == IB_QPT_UD) {
1225                 ret = cma_set_qkey(id_priv, 0);
1226                 if (ret)
1227                         return ret;
1228
1229                 qp_attr->qkey = id_priv->qkey;
1230                 *qp_attr_mask |= IB_QP_QKEY;
1231         } else {
1232                 qp_attr->qp_access_flags = 0;
1233                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1234         }
1235         return 0;
1236 }
1237
1238 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1239                        int *qp_attr_mask)
1240 {
1241         struct rdma_id_private *id_priv;
1242         int ret = 0;
1243
1244         id_priv = container_of(id, struct rdma_id_private, id);
1245         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1246                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1247                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1248                 else
1249                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1250                                                  qp_attr_mask);
1251
1252                 if (qp_attr->qp_state == IB_QPS_RTR)
1253                         qp_attr->rq_psn = id_priv->seq_num;
1254         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1255                 if (!id_priv->cm_id.iw) {
1256                         qp_attr->qp_access_flags = 0;
1257                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1258                 } else
1259                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1260                                                  qp_attr_mask);
1261                 qp_attr->port_num = id_priv->id.port_num;
1262                 *qp_attr_mask |= IB_QP_PORT;
1263         } else {
1264                 ret = -ENOSYS;
1265         }
1266
1267         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1268                 qp_attr->timeout = id_priv->timeout;
1269
1270         if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1271                 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1272
1273         return ret;
1274 }
1275 EXPORT_SYMBOL(rdma_init_qp_attr);
1276
1277 static inline bool cma_zero_addr(const struct sockaddr *addr)
1278 {
1279         switch (addr->sa_family) {
1280         case AF_INET:
1281                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1282         case AF_INET6:
1283                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1284         case AF_IB:
1285                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1286         default:
1287                 return false;
1288         }
1289 }
1290
1291 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1292 {
1293         switch (addr->sa_family) {
1294         case AF_INET:
1295                 return ipv4_is_loopback(
1296                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1297         case AF_INET6:
1298                 return ipv6_addr_loopback(
1299                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1300         case AF_IB:
1301                 return ib_addr_loopback(
1302                         &((struct sockaddr_ib *)addr)->sib_addr);
1303         default:
1304                 return false;
1305         }
1306 }
1307
1308 static inline bool cma_any_addr(const struct sockaddr *addr)
1309 {
1310         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1311 }
1312
1313 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1314 {
1315         if (src->sa_family != dst->sa_family)
1316                 return -1;
1317
1318         switch (src->sa_family) {
1319         case AF_INET:
1320                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1321                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1322         case AF_INET6: {
1323                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1324                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1325                 bool link_local;
1326
1327                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1328                                           &dst_addr6->sin6_addr))
1329                         return 1;
1330                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1331                              IPV6_ADDR_LINKLOCAL;
1332                 /* Link local must match their scope_ids */
1333                 return link_local ? (src_addr6->sin6_scope_id !=
1334                                      dst_addr6->sin6_scope_id) :
1335                                     0;
1336         }
1337
1338         default:
1339                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1340                                    &((struct sockaddr_ib *) dst)->sib_addr);
1341         }
1342 }
1343
1344 static __be16 cma_port(const struct sockaddr *addr)
1345 {
1346         struct sockaddr_ib *sib;
1347
1348         switch (addr->sa_family) {
1349         case AF_INET:
1350                 return ((struct sockaddr_in *) addr)->sin_port;
1351         case AF_INET6:
1352                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1353         case AF_IB:
1354                 sib = (struct sockaddr_ib *) addr;
1355                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1356                                     be64_to_cpu(sib->sib_sid_mask)));
1357         default:
1358                 return 0;
1359         }
1360 }
1361
1362 static inline int cma_any_port(const struct sockaddr *addr)
1363 {
1364         return !cma_port(addr);
1365 }
1366
1367 static void cma_save_ib_info(struct sockaddr *src_addr,
1368                              struct sockaddr *dst_addr,
1369                              const struct rdma_cm_id *listen_id,
1370                              const struct sa_path_rec *path)
1371 {
1372         struct sockaddr_ib *listen_ib, *ib;
1373
1374         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1375         if (src_addr) {
1376                 ib = (struct sockaddr_ib *)src_addr;
1377                 ib->sib_family = AF_IB;
1378                 if (path) {
1379                         ib->sib_pkey = path->pkey;
1380                         ib->sib_flowinfo = path->flow_label;
1381                         memcpy(&ib->sib_addr, &path->sgid, 16);
1382                         ib->sib_sid = path->service_id;
1383                         ib->sib_scope_id = 0;
1384                 } else {
1385                         ib->sib_pkey = listen_ib->sib_pkey;
1386                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1387                         ib->sib_addr = listen_ib->sib_addr;
1388                         ib->sib_sid = listen_ib->sib_sid;
1389                         ib->sib_scope_id = listen_ib->sib_scope_id;
1390                 }
1391                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1392         }
1393         if (dst_addr) {
1394                 ib = (struct sockaddr_ib *)dst_addr;
1395                 ib->sib_family = AF_IB;
1396                 if (path) {
1397                         ib->sib_pkey = path->pkey;
1398                         ib->sib_flowinfo = path->flow_label;
1399                         memcpy(&ib->sib_addr, &path->dgid, 16);
1400                 }
1401         }
1402 }
1403
1404 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1405                               struct sockaddr_in *dst_addr,
1406                               struct cma_hdr *hdr,
1407                               __be16 local_port)
1408 {
1409         if (src_addr) {
1410                 *src_addr = (struct sockaddr_in) {
1411                         .sin_family = AF_INET,
1412                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1413                         .sin_port = local_port,
1414                 };
1415         }
1416
1417         if (dst_addr) {
1418                 *dst_addr = (struct sockaddr_in) {
1419                         .sin_family = AF_INET,
1420                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1421                         .sin_port = hdr->port,
1422                 };
1423         }
1424 }
1425
1426 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1427                               struct sockaddr_in6 *dst_addr,
1428                               struct cma_hdr *hdr,
1429                               __be16 local_port)
1430 {
1431         if (src_addr) {
1432                 *src_addr = (struct sockaddr_in6) {
1433                         .sin6_family = AF_INET6,
1434                         .sin6_addr = hdr->dst_addr.ip6,
1435                         .sin6_port = local_port,
1436                 };
1437         }
1438
1439         if (dst_addr) {
1440                 *dst_addr = (struct sockaddr_in6) {
1441                         .sin6_family = AF_INET6,
1442                         .sin6_addr = hdr->src_addr.ip6,
1443                         .sin6_port = hdr->port,
1444                 };
1445         }
1446 }
1447
1448 static u16 cma_port_from_service_id(__be64 service_id)
1449 {
1450         return (u16)be64_to_cpu(service_id);
1451 }
1452
1453 static int cma_save_ip_info(struct sockaddr *src_addr,
1454                             struct sockaddr *dst_addr,
1455                             const struct ib_cm_event *ib_event,
1456                             __be64 service_id)
1457 {
1458         struct cma_hdr *hdr;
1459         __be16 port;
1460
1461         hdr = ib_event->private_data;
1462         if (hdr->cma_version != CMA_VERSION)
1463                 return -EINVAL;
1464
1465         port = htons(cma_port_from_service_id(service_id));
1466
1467         switch (cma_get_ip_ver(hdr)) {
1468         case 4:
1469                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1470                                   (struct sockaddr_in *)dst_addr, hdr, port);
1471                 break;
1472         case 6:
1473                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1474                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1475                 break;
1476         default:
1477                 return -EAFNOSUPPORT;
1478         }
1479
1480         return 0;
1481 }
1482
1483 static int cma_save_net_info(struct sockaddr *src_addr,
1484                              struct sockaddr *dst_addr,
1485                              const struct rdma_cm_id *listen_id,
1486                              const struct ib_cm_event *ib_event,
1487                              sa_family_t sa_family, __be64 service_id)
1488 {
1489         if (sa_family == AF_IB) {
1490                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1491                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1492                                          ib_event->param.req_rcvd.primary_path);
1493                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1494                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1495                 return 0;
1496         }
1497
1498         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1499 }
1500
1501 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1502                              struct cma_req_info *req)
1503 {
1504         const struct ib_cm_req_event_param *req_param =
1505                 &ib_event->param.req_rcvd;
1506         const struct ib_cm_sidr_req_event_param *sidr_param =
1507                 &ib_event->param.sidr_req_rcvd;
1508
1509         switch (ib_event->event) {
1510         case IB_CM_REQ_RECEIVED:
1511                 req->device     = req_param->listen_id->device;
1512                 req->port       = req_param->port;
1513                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1514                        sizeof(req->local_gid));
1515                 req->has_gid    = true;
1516                 req->service_id = req_param->primary_path->service_id;
1517                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1518                 if (req->pkey != req_param->bth_pkey)
1519                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1520                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1521                                             req_param->bth_pkey, req->pkey);
1522                 break;
1523         case IB_CM_SIDR_REQ_RECEIVED:
1524                 req->device     = sidr_param->listen_id->device;
1525                 req->port       = sidr_param->port;
1526                 req->has_gid    = false;
1527                 req->service_id = sidr_param->service_id;
1528                 req->pkey       = sidr_param->pkey;
1529                 if (req->pkey != sidr_param->bth_pkey)
1530                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1531                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1532                                             sidr_param->bth_pkey, req->pkey);
1533                 break;
1534         default:
1535                 return -EINVAL;
1536         }
1537
1538         return 0;
1539 }
1540
1541 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1542                                   const struct sockaddr_in *dst_addr,
1543                                   const struct sockaddr_in *src_addr)
1544 {
1545         __be32 daddr = dst_addr->sin_addr.s_addr,
1546                saddr = src_addr->sin_addr.s_addr;
1547         struct fib_result res;
1548         struct flowi4 fl4;
1549         int err;
1550         bool ret;
1551
1552         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1553             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1554             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1555             ipv4_is_loopback(saddr))
1556                 return false;
1557
1558         memset(&fl4, 0, sizeof(fl4));
1559         fl4.flowi4_iif = net_dev->ifindex;
1560         fl4.daddr = daddr;
1561         fl4.saddr = saddr;
1562
1563         rcu_read_lock();
1564         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1565         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1566         rcu_read_unlock();
1567
1568         return ret;
1569 }
1570
1571 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1572                                   const struct sockaddr_in6 *dst_addr,
1573                                   const struct sockaddr_in6 *src_addr)
1574 {
1575 #if IS_ENABLED(CONFIG_IPV6)
1576         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1577                            IPV6_ADDR_LINKLOCAL;
1578         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1579                                          &src_addr->sin6_addr, net_dev->ifindex,
1580                                          NULL, strict);
1581         bool ret;
1582
1583         if (!rt)
1584                 return false;
1585
1586         ret = rt->rt6i_idev->dev == net_dev;
1587         ip6_rt_put(rt);
1588
1589         return ret;
1590 #else
1591         return false;
1592 #endif
1593 }
1594
1595 static bool validate_net_dev(struct net_device *net_dev,
1596                              const struct sockaddr *daddr,
1597                              const struct sockaddr *saddr)
1598 {
1599         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1600         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1601         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1602         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1603
1604         switch (daddr->sa_family) {
1605         case AF_INET:
1606                 return saddr->sa_family == AF_INET &&
1607                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1608
1609         case AF_INET6:
1610                 return saddr->sa_family == AF_INET6 &&
1611                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1612
1613         default:
1614                 return false;
1615         }
1616 }
1617
1618 static struct net_device *
1619 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1620 {
1621         const struct ib_gid_attr *sgid_attr = NULL;
1622         struct net_device *ndev;
1623
1624         if (ib_event->event == IB_CM_REQ_RECEIVED)
1625                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1626         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1627                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1628
1629         if (!sgid_attr)
1630                 return NULL;
1631
1632         rcu_read_lock();
1633         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1634         if (IS_ERR(ndev))
1635                 ndev = NULL;
1636         else
1637                 dev_hold(ndev);
1638         rcu_read_unlock();
1639         return ndev;
1640 }
1641
1642 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1643                                           struct cma_req_info *req)
1644 {
1645         struct sockaddr *listen_addr =
1646                         (struct sockaddr *)&req->listen_addr_storage;
1647         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1648         struct net_device *net_dev;
1649         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1650         int err;
1651
1652         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1653                                req->service_id);
1654         if (err)
1655                 return ERR_PTR(err);
1656
1657         if (rdma_protocol_roce(req->device, req->port))
1658                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1659         else
1660                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1661                                                    req->pkey,
1662                                                    gid, listen_addr);
1663         if (!net_dev)
1664                 return ERR_PTR(-ENODEV);
1665
1666         return net_dev;
1667 }
1668
1669 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1670 {
1671         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1672 }
1673
1674 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1675                                    const struct cma_hdr *hdr)
1676 {
1677         struct sockaddr *addr = cma_src_addr(id_priv);
1678         __be32 ip4_addr;
1679         struct in6_addr ip6_addr;
1680
1681         if (cma_any_addr(addr) && !id_priv->afonly)
1682                 return true;
1683
1684         switch (addr->sa_family) {
1685         case AF_INET:
1686                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1687                 if (cma_get_ip_ver(hdr) != 4)
1688                         return false;
1689                 if (!cma_any_addr(addr) &&
1690                     hdr->dst_addr.ip4.addr != ip4_addr)
1691                         return false;
1692                 break;
1693         case AF_INET6:
1694                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1695                 if (cma_get_ip_ver(hdr) != 6)
1696                         return false;
1697                 if (!cma_any_addr(addr) &&
1698                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1699                         return false;
1700                 break;
1701         case AF_IB:
1702                 return true;
1703         default:
1704                 return false;
1705         }
1706
1707         return true;
1708 }
1709
1710 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1711 {
1712         struct ib_device *device = id->device;
1713         const u32 port_num = id->port_num ?: rdma_start_port(device);
1714
1715         return rdma_protocol_roce(device, port_num);
1716 }
1717
1718 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1719 {
1720         const struct sockaddr *daddr =
1721                         (const struct sockaddr *)&req->listen_addr_storage;
1722         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1723
1724         /* Returns true if the req is for IPv6 link local */
1725         return (daddr->sa_family == AF_INET6 &&
1726                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1727 }
1728
1729 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1730                               const struct net_device *net_dev,
1731                               const struct cma_req_info *req)
1732 {
1733         const struct rdma_addr *addr = &id->route.addr;
1734
1735         if (!net_dev)
1736                 /* This request is an AF_IB request */
1737                 return (!id->port_num || id->port_num == req->port) &&
1738                        (addr->src_addr.ss_family == AF_IB);
1739
1740         /*
1741          * If the request is not for IPv6 link local, allow matching
1742          * request to any netdevice of the one or multiport rdma device.
1743          */
1744         if (!cma_is_req_ipv6_ll(req))
1745                 return true;
1746         /*
1747          * Net namespaces must match, and if the listner is listening
1748          * on a specific netdevice than netdevice must match as well.
1749          */
1750         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1751             (!!addr->dev_addr.bound_dev_if ==
1752              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1753                 return true;
1754         else
1755                 return false;
1756 }
1757
1758 static struct rdma_id_private *cma_find_listener(
1759                 const struct rdma_bind_list *bind_list,
1760                 const struct ib_cm_id *cm_id,
1761                 const struct ib_cm_event *ib_event,
1762                 const struct cma_req_info *req,
1763                 const struct net_device *net_dev)
1764 {
1765         struct rdma_id_private *id_priv, *id_priv_dev;
1766
1767         lockdep_assert_held(&lock);
1768
1769         if (!bind_list)
1770                 return ERR_PTR(-EINVAL);
1771
1772         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1773                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1774                         if (id_priv->id.device == cm_id->device &&
1775                             cma_match_net_dev(&id_priv->id, net_dev, req))
1776                                 return id_priv;
1777                         list_for_each_entry(id_priv_dev,
1778                                             &id_priv->listen_list,
1779                                             listen_item) {
1780                                 if (id_priv_dev->id.device == cm_id->device &&
1781                                     cma_match_net_dev(&id_priv_dev->id,
1782                                                       net_dev, req))
1783                                         return id_priv_dev;
1784                         }
1785                 }
1786         }
1787
1788         return ERR_PTR(-EINVAL);
1789 }
1790
1791 static struct rdma_id_private *
1792 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1793                      const struct ib_cm_event *ib_event,
1794                      struct cma_req_info *req,
1795                      struct net_device **net_dev)
1796 {
1797         struct rdma_bind_list *bind_list;
1798         struct rdma_id_private *id_priv;
1799         int err;
1800
1801         err = cma_save_req_info(ib_event, req);
1802         if (err)
1803                 return ERR_PTR(err);
1804
1805         *net_dev = cma_get_net_dev(ib_event, req);
1806         if (IS_ERR(*net_dev)) {
1807                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1808                         /* Assuming the protocol is AF_IB */
1809                         *net_dev = NULL;
1810                 } else {
1811                         return ERR_CAST(*net_dev);
1812                 }
1813         }
1814
1815         mutex_lock(&lock);
1816         /*
1817          * Net namespace might be getting deleted while route lookup,
1818          * cm_id lookup is in progress. Therefore, perform netdevice
1819          * validation, cm_id lookup under rcu lock.
1820          * RCU lock along with netdevice state check, synchronizes with
1821          * netdevice migrating to different net namespace and also avoids
1822          * case where net namespace doesn't get deleted while lookup is in
1823          * progress.
1824          * If the device state is not IFF_UP, its properties such as ifindex
1825          * and nd_net cannot be trusted to remain valid without rcu lock.
1826          * net/core/dev.c change_net_namespace() ensures to synchronize with
1827          * ongoing operations on net device after device is closed using
1828          * synchronize_net().
1829          */
1830         rcu_read_lock();
1831         if (*net_dev) {
1832                 /*
1833                  * If netdevice is down, it is likely that it is administratively
1834                  * down or it might be migrating to different namespace.
1835                  * In that case avoid further processing, as the net namespace
1836                  * or ifindex may change.
1837                  */
1838                 if (((*net_dev)->flags & IFF_UP) == 0) {
1839                         id_priv = ERR_PTR(-EHOSTUNREACH);
1840                         goto err;
1841                 }
1842
1843                 if (!validate_net_dev(*net_dev,
1844                                  (struct sockaddr *)&req->listen_addr_storage,
1845                                  (struct sockaddr *)&req->src_addr_storage)) {
1846                         id_priv = ERR_PTR(-EHOSTUNREACH);
1847                         goto err;
1848                 }
1849         }
1850
1851         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1852                                 rdma_ps_from_service_id(req->service_id),
1853                                 cma_port_from_service_id(req->service_id));
1854         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1855 err:
1856         rcu_read_unlock();
1857         mutex_unlock(&lock);
1858         if (IS_ERR(id_priv) && *net_dev) {
1859                 dev_put(*net_dev);
1860                 *net_dev = NULL;
1861         }
1862         return id_priv;
1863 }
1864
1865 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1866 {
1867         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1868 }
1869
1870 static void cma_cancel_route(struct rdma_id_private *id_priv)
1871 {
1872         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1873                 if (id_priv->query)
1874                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1875         }
1876 }
1877
1878 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1879 {
1880         struct rdma_id_private *dev_id_priv;
1881
1882         lockdep_assert_held(&lock);
1883
1884         /*
1885          * Remove from listen_any_list to prevent added devices from spawning
1886          * additional listen requests.
1887          */
1888         list_del_init(&id_priv->listen_any_item);
1889
1890         while (!list_empty(&id_priv->listen_list)) {
1891                 dev_id_priv =
1892                         list_first_entry(&id_priv->listen_list,
1893                                          struct rdma_id_private, listen_item);
1894                 /* sync with device removal to avoid duplicate destruction */
1895                 list_del_init(&dev_id_priv->device_item);
1896                 list_del_init(&dev_id_priv->listen_item);
1897                 mutex_unlock(&lock);
1898
1899                 rdma_destroy_id(&dev_id_priv->id);
1900                 mutex_lock(&lock);
1901         }
1902 }
1903
1904 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1905 {
1906         mutex_lock(&lock);
1907         _cma_cancel_listens(id_priv);
1908         mutex_unlock(&lock);
1909 }
1910
1911 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1912                                  enum rdma_cm_state state)
1913 {
1914         switch (state) {
1915         case RDMA_CM_ADDR_QUERY:
1916                 /*
1917                  * We can avoid doing the rdma_addr_cancel() based on state,
1918                  * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1919                  * Notice that the addr_handler work could still be exiting
1920                  * outside this state, however due to the interaction with the
1921                  * handler_mutex the work is guaranteed not to touch id_priv
1922                  * during exit.
1923                  */
1924                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1925                 break;
1926         case RDMA_CM_ROUTE_QUERY:
1927                 cma_cancel_route(id_priv);
1928                 break;
1929         case RDMA_CM_LISTEN:
1930                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1931                         cma_cancel_listens(id_priv);
1932                 break;
1933         default:
1934                 break;
1935         }
1936 }
1937
1938 static void cma_release_port(struct rdma_id_private *id_priv)
1939 {
1940         struct rdma_bind_list *bind_list = id_priv->bind_list;
1941         struct net *net = id_priv->id.route.addr.dev_addr.net;
1942
1943         if (!bind_list)
1944                 return;
1945
1946         mutex_lock(&lock);
1947         hlist_del(&id_priv->node);
1948         if (hlist_empty(&bind_list->owners)) {
1949                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1950                 kfree(bind_list);
1951         }
1952         mutex_unlock(&lock);
1953 }
1954
1955 static void destroy_mc(struct rdma_id_private *id_priv,
1956                        struct cma_multicast *mc)
1957 {
1958         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1959
1960         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1961                 ib_sa_free_multicast(mc->sa_mc);
1962
1963         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1964                 struct rdma_dev_addr *dev_addr =
1965                         &id_priv->id.route.addr.dev_addr;
1966                 struct net_device *ndev = NULL;
1967
1968                 if (dev_addr->bound_dev_if)
1969                         ndev = dev_get_by_index(dev_addr->net,
1970                                                 dev_addr->bound_dev_if);
1971                 if (ndev && !send_only) {
1972                         enum ib_gid_type gid_type;
1973                         union ib_gid mgid;
1974
1975                         gid_type = id_priv->cma_dev->default_gid_type
1976                                            [id_priv->id.port_num -
1977                                             rdma_start_port(
1978                                                     id_priv->cma_dev->device)];
1979                         cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
1980                                           gid_type);
1981                         cma_igmp_send(ndev, &mgid, false);
1982                 }
1983                 dev_put(ndev);
1984
1985                 cancel_work_sync(&mc->iboe_join.work);
1986         }
1987         kfree(mc);
1988 }
1989
1990 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1991 {
1992         struct cma_multicast *mc;
1993
1994         while (!list_empty(&id_priv->mc_list)) {
1995                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1996                                       list);
1997                 list_del(&mc->list);
1998                 destroy_mc(id_priv, mc);
1999         }
2000 }
2001
2002 static void _destroy_id(struct rdma_id_private *id_priv,
2003                         enum rdma_cm_state state)
2004 {
2005         cma_cancel_operation(id_priv, state);
2006
2007         rdma_restrack_del(&id_priv->res);
2008         cma_remove_id_from_tree(id_priv);
2009         if (id_priv->cma_dev) {
2010                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
2011                         if (id_priv->cm_id.ib)
2012                                 ib_destroy_cm_id(id_priv->cm_id.ib);
2013                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
2014                         if (id_priv->cm_id.iw)
2015                                 iw_destroy_cm_id(id_priv->cm_id.iw);
2016                 }
2017                 cma_leave_mc_groups(id_priv);
2018                 cma_release_dev(id_priv);
2019         }
2020
2021         cma_release_port(id_priv);
2022         cma_id_put(id_priv);
2023         wait_for_completion(&id_priv->comp);
2024
2025         if (id_priv->internal_id)
2026                 cma_id_put(id_priv->id.context);
2027
2028         kfree(id_priv->id.route.path_rec);
2029
2030         put_net(id_priv->id.route.addr.dev_addr.net);
2031         kfree(id_priv);
2032 }
2033
2034 /*
2035  * destroy an ID from within the handler_mutex. This ensures that no other
2036  * handlers can start running concurrently.
2037  */
2038 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
2039         __releases(&idprv->handler_mutex)
2040 {
2041         enum rdma_cm_state state;
2042         unsigned long flags;
2043
2044         trace_cm_id_destroy(id_priv);
2045
2046         /*
2047          * Setting the state to destroyed under the handler mutex provides a
2048          * fence against calling handler callbacks. If this is invoked due to
2049          * the failure of a handler callback then it guarentees that no future
2050          * handlers will be called.
2051          */
2052         lockdep_assert_held(&id_priv->handler_mutex);
2053         spin_lock_irqsave(&id_priv->lock, flags);
2054         state = id_priv->state;
2055         id_priv->state = RDMA_CM_DESTROYING;
2056         spin_unlock_irqrestore(&id_priv->lock, flags);
2057         mutex_unlock(&id_priv->handler_mutex);
2058         _destroy_id(id_priv, state);
2059 }
2060
2061 void rdma_destroy_id(struct rdma_cm_id *id)
2062 {
2063         struct rdma_id_private *id_priv =
2064                 container_of(id, struct rdma_id_private, id);
2065
2066         mutex_lock(&id_priv->handler_mutex);
2067         destroy_id_handler_unlock(id_priv);
2068 }
2069 EXPORT_SYMBOL(rdma_destroy_id);
2070
2071 static int cma_rep_recv(struct rdma_id_private *id_priv)
2072 {
2073         int ret;
2074
2075         ret = cma_modify_qp_rtr(id_priv, NULL);
2076         if (ret)
2077                 goto reject;
2078
2079         ret = cma_modify_qp_rts(id_priv, NULL);
2080         if (ret)
2081                 goto reject;
2082
2083         trace_cm_send_rtu(id_priv);
2084         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
2085         if (ret)
2086                 goto reject;
2087
2088         return 0;
2089 reject:
2090         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
2091         cma_modify_qp_err(id_priv);
2092         trace_cm_send_rej(id_priv);
2093         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
2094                        NULL, 0, NULL, 0);
2095         return ret;
2096 }
2097
2098 static void cma_set_rep_event_data(struct rdma_cm_event *event,
2099                                    const struct ib_cm_rep_event_param *rep_data,
2100                                    void *private_data)
2101 {
2102         event->param.conn.private_data = private_data;
2103         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
2104         event->param.conn.responder_resources = rep_data->responder_resources;
2105         event->param.conn.initiator_depth = rep_data->initiator_depth;
2106         event->param.conn.flow_control = rep_data->flow_control;
2107         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
2108         event->param.conn.srq = rep_data->srq;
2109         event->param.conn.qp_num = rep_data->remote_qpn;
2110
2111         event->ece.vendor_id = rep_data->ece.vendor_id;
2112         event->ece.attr_mod = rep_data->ece.attr_mod;
2113 }
2114
2115 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
2116                                 struct rdma_cm_event *event)
2117 {
2118         int ret;
2119
2120         lockdep_assert_held(&id_priv->handler_mutex);
2121
2122         trace_cm_event_handler(id_priv, event);
2123         ret = id_priv->id.event_handler(&id_priv->id, event);
2124         trace_cm_event_done(id_priv, event, ret);
2125         return ret;
2126 }
2127
2128 static int cma_ib_handler(struct ib_cm_id *cm_id,
2129                           const struct ib_cm_event *ib_event)
2130 {
2131         struct rdma_id_private *id_priv = cm_id->context;
2132         struct rdma_cm_event event = {};
2133         enum rdma_cm_state state;
2134         int ret;
2135
2136         mutex_lock(&id_priv->handler_mutex);
2137         state = READ_ONCE(id_priv->state);
2138         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2139              state != RDMA_CM_CONNECT) ||
2140             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2141              state != RDMA_CM_DISCONNECT))
2142                 goto out;
2143
2144         switch (ib_event->event) {
2145         case IB_CM_REQ_ERROR:
2146         case IB_CM_REP_ERROR:
2147                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2148                 event.status = -ETIMEDOUT;
2149                 break;
2150         case IB_CM_REP_RECEIVED:
2151                 if (state == RDMA_CM_CONNECT &&
2152                     (id_priv->id.qp_type != IB_QPT_UD)) {
2153                         trace_cm_send_mra(id_priv);
2154                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2155                 }
2156                 if (id_priv->id.qp) {
2157                         event.status = cma_rep_recv(id_priv);
2158                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2159                                                      RDMA_CM_EVENT_ESTABLISHED;
2160                 } else {
2161                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2162                 }
2163                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2164                                        ib_event->private_data);
2165                 break;
2166         case IB_CM_RTU_RECEIVED:
2167         case IB_CM_USER_ESTABLISHED:
2168                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2169                 break;
2170         case IB_CM_DREQ_ERROR:
2171                 event.status = -ETIMEDOUT;
2172                 fallthrough;
2173         case IB_CM_DREQ_RECEIVED:
2174         case IB_CM_DREP_RECEIVED:
2175                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2176                                    RDMA_CM_DISCONNECT))
2177                         goto out;
2178                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2179                 break;
2180         case IB_CM_TIMEWAIT_EXIT:
2181                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2182                 break;
2183         case IB_CM_MRA_RECEIVED:
2184                 /* ignore event */
2185                 goto out;
2186         case IB_CM_REJ_RECEIVED:
2187                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2188                                                                                 ib_event->param.rej_rcvd.reason));
2189                 cma_modify_qp_err(id_priv);
2190                 event.status = ib_event->param.rej_rcvd.reason;
2191                 event.event = RDMA_CM_EVENT_REJECTED;
2192                 event.param.conn.private_data = ib_event->private_data;
2193                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2194                 break;
2195         default:
2196                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2197                        ib_event->event);
2198                 goto out;
2199         }
2200
2201         ret = cma_cm_event_handler(id_priv, &event);
2202         if (ret) {
2203                 /* Destroy the CM ID by returning a non-zero value. */
2204                 id_priv->cm_id.ib = NULL;
2205                 destroy_id_handler_unlock(id_priv);
2206                 return ret;
2207         }
2208 out:
2209         mutex_unlock(&id_priv->handler_mutex);
2210         return 0;
2211 }
2212
2213 static struct rdma_id_private *
2214 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2215                    const struct ib_cm_event *ib_event,
2216                    struct net_device *net_dev)
2217 {
2218         struct rdma_id_private *listen_id_priv;
2219         struct rdma_id_private *id_priv;
2220         struct rdma_cm_id *id;
2221         struct rdma_route *rt;
2222         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2223         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2224         const __be64 service_id =
2225                 ib_event->param.req_rcvd.primary_path->service_id;
2226         int ret;
2227
2228         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2229         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2230                                    listen_id->event_handler, listen_id->context,
2231                                    listen_id->ps,
2232                                    ib_event->param.req_rcvd.qp_type,
2233                                    listen_id_priv);
2234         if (IS_ERR(id_priv))
2235                 return NULL;
2236
2237         id = &id_priv->id;
2238         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2239                               (struct sockaddr *)&id->route.addr.dst_addr,
2240                               listen_id, ib_event, ss_family, service_id))
2241                 goto err;
2242
2243         rt = &id->route;
2244         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2245         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2246                                      GFP_KERNEL);
2247         if (!rt->path_rec)
2248                 goto err;
2249
2250         rt->path_rec[0] = *path;
2251         if (rt->num_paths == 2)
2252                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2253
2254         if (net_dev) {
2255                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2256         } else {
2257                 if (!cma_protocol_roce(listen_id) &&
2258                     cma_any_addr(cma_src_addr(id_priv))) {
2259                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2260                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2261                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2262                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2263                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2264                         if (ret)
2265                                 goto err;
2266                 }
2267         }
2268         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2269
2270         id_priv->state = RDMA_CM_CONNECT;
2271         return id_priv;
2272
2273 err:
2274         rdma_destroy_id(id);
2275         return NULL;
2276 }
2277
2278 static struct rdma_id_private *
2279 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2280                   const struct ib_cm_event *ib_event,
2281                   struct net_device *net_dev)
2282 {
2283         const struct rdma_id_private *listen_id_priv;
2284         struct rdma_id_private *id_priv;
2285         struct rdma_cm_id *id;
2286         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2287         struct net *net = listen_id->route.addr.dev_addr.net;
2288         int ret;
2289
2290         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2291         id_priv = __rdma_create_id(net, listen_id->event_handler,
2292                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2293                                    listen_id_priv);
2294         if (IS_ERR(id_priv))
2295                 return NULL;
2296
2297         id = &id_priv->id;
2298         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2299                               (struct sockaddr *)&id->route.addr.dst_addr,
2300                               listen_id, ib_event, ss_family,
2301                               ib_event->param.sidr_req_rcvd.service_id))
2302                 goto err;
2303
2304         if (net_dev) {
2305                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2306         } else {
2307                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2308                         ret = cma_translate_addr(cma_src_addr(id_priv),
2309                                                  &id->route.addr.dev_addr);
2310                         if (ret)
2311                                 goto err;
2312                 }
2313         }
2314
2315         id_priv->state = RDMA_CM_CONNECT;
2316         return id_priv;
2317 err:
2318         rdma_destroy_id(id);
2319         return NULL;
2320 }
2321
2322 static void cma_set_req_event_data(struct rdma_cm_event *event,
2323                                    const struct ib_cm_req_event_param *req_data,
2324                                    void *private_data, int offset)
2325 {
2326         event->param.conn.private_data = private_data + offset;
2327         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2328         event->param.conn.responder_resources = req_data->responder_resources;
2329         event->param.conn.initiator_depth = req_data->initiator_depth;
2330         event->param.conn.flow_control = req_data->flow_control;
2331         event->param.conn.retry_count = req_data->retry_count;
2332         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2333         event->param.conn.srq = req_data->srq;
2334         event->param.conn.qp_num = req_data->remote_qpn;
2335
2336         event->ece.vendor_id = req_data->ece.vendor_id;
2337         event->ece.attr_mod = req_data->ece.attr_mod;
2338 }
2339
2340 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2341                                     const struct ib_cm_event *ib_event)
2342 {
2343         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2344                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2345                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2346                  (id->qp_type == IB_QPT_UD)) ||
2347                 (!id->qp_type));
2348 }
2349
2350 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2351                               const struct ib_cm_event *ib_event)
2352 {
2353         struct rdma_id_private *listen_id, *conn_id = NULL;
2354         struct rdma_cm_event event = {};
2355         struct cma_req_info req = {};
2356         struct net_device *net_dev;
2357         u8 offset;
2358         int ret;
2359
2360         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2361         if (IS_ERR(listen_id))
2362                 return PTR_ERR(listen_id);
2363
2364         trace_cm_req_handler(listen_id, ib_event->event);
2365         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2366                 ret = -EINVAL;
2367                 goto net_dev_put;
2368         }
2369
2370         mutex_lock(&listen_id->handler_mutex);
2371         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2372                 ret = -ECONNABORTED;
2373                 goto err_unlock;
2374         }
2375
2376         offset = cma_user_data_offset(listen_id);
2377         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2378         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2379                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2380                 event.param.ud.private_data = ib_event->private_data + offset;
2381                 event.param.ud.private_data_len =
2382                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2383         } else {
2384                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2385                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2386                                        ib_event->private_data, offset);
2387         }
2388         if (!conn_id) {
2389                 ret = -ENOMEM;
2390                 goto err_unlock;
2391         }
2392
2393         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2394         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2395         if (ret) {
2396                 destroy_id_handler_unlock(conn_id);
2397                 goto err_unlock;
2398         }
2399
2400         conn_id->cm_id.ib = cm_id;
2401         cm_id->context = conn_id;
2402         cm_id->cm_handler = cma_ib_handler;
2403
2404         ret = cma_cm_event_handler(conn_id, &event);
2405         if (ret) {
2406                 /* Destroy the CM ID by returning a non-zero value. */
2407                 conn_id->cm_id.ib = NULL;
2408                 mutex_unlock(&listen_id->handler_mutex);
2409                 destroy_id_handler_unlock(conn_id);
2410                 goto net_dev_put;
2411         }
2412
2413         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2414             conn_id->id.qp_type != IB_QPT_UD) {
2415                 trace_cm_send_mra(cm_id->context);
2416                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2417         }
2418         mutex_unlock(&conn_id->handler_mutex);
2419
2420 err_unlock:
2421         mutex_unlock(&listen_id->handler_mutex);
2422
2423 net_dev_put:
2424         if (net_dev)
2425                 dev_put(net_dev);
2426
2427         return ret;
2428 }
2429
2430 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2431 {
2432         if (addr->sa_family == AF_IB)
2433                 return ((struct sockaddr_ib *) addr)->sib_sid;
2434
2435         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2436 }
2437 EXPORT_SYMBOL(rdma_get_service_id);
2438
2439 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2440                     union ib_gid *dgid)
2441 {
2442         struct rdma_addr *addr = &cm_id->route.addr;
2443
2444         if (!cm_id->device) {
2445                 if (sgid)
2446                         memset(sgid, 0, sizeof(*sgid));
2447                 if (dgid)
2448                         memset(dgid, 0, sizeof(*dgid));
2449                 return;
2450         }
2451
2452         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2453                 if (sgid)
2454                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2455                 if (dgid)
2456                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2457         } else {
2458                 if (sgid)
2459                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2460                 if (dgid)
2461                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2462         }
2463 }
2464 EXPORT_SYMBOL(rdma_read_gids);
2465
2466 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2467 {
2468         struct rdma_id_private *id_priv = iw_id->context;
2469         struct rdma_cm_event event = {};
2470         int ret = 0;
2471         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2472         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2473
2474         mutex_lock(&id_priv->handler_mutex);
2475         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2476                 goto out;
2477
2478         switch (iw_event->event) {
2479         case IW_CM_EVENT_CLOSE:
2480                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2481                 break;
2482         case IW_CM_EVENT_CONNECT_REPLY:
2483                 memcpy(cma_src_addr(id_priv), laddr,
2484                        rdma_addr_size(laddr));
2485                 memcpy(cma_dst_addr(id_priv), raddr,
2486                        rdma_addr_size(raddr));
2487                 switch (iw_event->status) {
2488                 case 0:
2489                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2490                         event.param.conn.initiator_depth = iw_event->ird;
2491                         event.param.conn.responder_resources = iw_event->ord;
2492                         break;
2493                 case -ECONNRESET:
2494                 case -ECONNREFUSED:
2495                         event.event = RDMA_CM_EVENT_REJECTED;
2496                         break;
2497                 case -ETIMEDOUT:
2498                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2499                         break;
2500                 default:
2501                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2502                         break;
2503                 }
2504                 break;
2505         case IW_CM_EVENT_ESTABLISHED:
2506                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2507                 event.param.conn.initiator_depth = iw_event->ird;
2508                 event.param.conn.responder_resources = iw_event->ord;
2509                 break;
2510         default:
2511                 goto out;
2512         }
2513
2514         event.status = iw_event->status;
2515         event.param.conn.private_data = iw_event->private_data;
2516         event.param.conn.private_data_len = iw_event->private_data_len;
2517         ret = cma_cm_event_handler(id_priv, &event);
2518         if (ret) {
2519                 /* Destroy the CM ID by returning a non-zero value. */
2520                 id_priv->cm_id.iw = NULL;
2521                 destroy_id_handler_unlock(id_priv);
2522                 return ret;
2523         }
2524
2525 out:
2526         mutex_unlock(&id_priv->handler_mutex);
2527         return ret;
2528 }
2529
2530 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2531                                struct iw_cm_event *iw_event)
2532 {
2533         struct rdma_id_private *listen_id, *conn_id;
2534         struct rdma_cm_event event = {};
2535         int ret = -ECONNABORTED;
2536         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2537         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2538
2539         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2540         event.param.conn.private_data = iw_event->private_data;
2541         event.param.conn.private_data_len = iw_event->private_data_len;
2542         event.param.conn.initiator_depth = iw_event->ird;
2543         event.param.conn.responder_resources = iw_event->ord;
2544
2545         listen_id = cm_id->context;
2546
2547         mutex_lock(&listen_id->handler_mutex);
2548         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2549                 goto out;
2550
2551         /* Create a new RDMA id for the new IW CM ID */
2552         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2553                                    listen_id->id.event_handler,
2554                                    listen_id->id.context, RDMA_PS_TCP,
2555                                    IB_QPT_RC, listen_id);
2556         if (IS_ERR(conn_id)) {
2557                 ret = -ENOMEM;
2558                 goto out;
2559         }
2560         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2561         conn_id->state = RDMA_CM_CONNECT;
2562
2563         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2564         if (ret) {
2565                 mutex_unlock(&listen_id->handler_mutex);
2566                 destroy_id_handler_unlock(conn_id);
2567                 return ret;
2568         }
2569
2570         ret = cma_iw_acquire_dev(conn_id, listen_id);
2571         if (ret) {
2572                 mutex_unlock(&listen_id->handler_mutex);
2573                 destroy_id_handler_unlock(conn_id);
2574                 return ret;
2575         }
2576
2577         conn_id->cm_id.iw = cm_id;
2578         cm_id->context = conn_id;
2579         cm_id->cm_handler = cma_iw_handler;
2580
2581         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2582         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2583
2584         ret = cma_cm_event_handler(conn_id, &event);
2585         if (ret) {
2586                 /* User wants to destroy the CM ID */
2587                 conn_id->cm_id.iw = NULL;
2588                 mutex_unlock(&listen_id->handler_mutex);
2589                 destroy_id_handler_unlock(conn_id);
2590                 return ret;
2591         }
2592
2593         mutex_unlock(&conn_id->handler_mutex);
2594
2595 out:
2596         mutex_unlock(&listen_id->handler_mutex);
2597         return ret;
2598 }
2599
2600 static int cma_ib_listen(struct rdma_id_private *id_priv)
2601 {
2602         struct sockaddr *addr;
2603         struct ib_cm_id *id;
2604         __be64 svc_id;
2605
2606         addr = cma_src_addr(id_priv);
2607         svc_id = rdma_get_service_id(&id_priv->id, addr);
2608         id = ib_cm_insert_listen(id_priv->id.device,
2609                                  cma_ib_req_handler, svc_id);
2610         if (IS_ERR(id))
2611                 return PTR_ERR(id);
2612         id_priv->cm_id.ib = id;
2613
2614         return 0;
2615 }
2616
2617 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2618 {
2619         int ret;
2620         struct iw_cm_id *id;
2621
2622         id = iw_create_cm_id(id_priv->id.device,
2623                              iw_conn_req_handler,
2624                              id_priv);
2625         if (IS_ERR(id))
2626                 return PTR_ERR(id);
2627
2628         mutex_lock(&id_priv->qp_mutex);
2629         id->tos = id_priv->tos;
2630         id->tos_set = id_priv->tos_set;
2631         mutex_unlock(&id_priv->qp_mutex);
2632         id->afonly = id_priv->afonly;
2633         id_priv->cm_id.iw = id;
2634
2635         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2636                rdma_addr_size(cma_src_addr(id_priv)));
2637
2638         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2639
2640         if (ret) {
2641                 iw_destroy_cm_id(id_priv->cm_id.iw);
2642                 id_priv->cm_id.iw = NULL;
2643         }
2644
2645         return ret;
2646 }
2647
2648 static int cma_listen_handler(struct rdma_cm_id *id,
2649                               struct rdma_cm_event *event)
2650 {
2651         struct rdma_id_private *id_priv = id->context;
2652
2653         /* Listening IDs are always destroyed on removal */
2654         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2655                 return -1;
2656
2657         id->context = id_priv->id.context;
2658         id->event_handler = id_priv->id.event_handler;
2659         trace_cm_event_handler(id_priv, event);
2660         return id_priv->id.event_handler(id, event);
2661 }
2662
2663 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2664                              struct cma_device *cma_dev,
2665                              struct rdma_id_private **to_destroy)
2666 {
2667         struct rdma_id_private *dev_id_priv;
2668         struct net *net = id_priv->id.route.addr.dev_addr.net;
2669         int ret;
2670
2671         lockdep_assert_held(&lock);
2672
2673         *to_destroy = NULL;
2674         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2675                 return 0;
2676
2677         dev_id_priv =
2678                 __rdma_create_id(net, cma_listen_handler, id_priv,
2679                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2680         if (IS_ERR(dev_id_priv))
2681                 return PTR_ERR(dev_id_priv);
2682
2683         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2684         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2685                rdma_addr_size(cma_src_addr(id_priv)));
2686
2687         _cma_attach_to_dev(dev_id_priv, cma_dev);
2688         rdma_restrack_add(&dev_id_priv->res);
2689         cma_id_get(id_priv);
2690         dev_id_priv->internal_id = 1;
2691         dev_id_priv->afonly = id_priv->afonly;
2692         mutex_lock(&id_priv->qp_mutex);
2693         dev_id_priv->tos_set = id_priv->tos_set;
2694         dev_id_priv->tos = id_priv->tos;
2695         mutex_unlock(&id_priv->qp_mutex);
2696
2697         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2698         if (ret)
2699                 goto err_listen;
2700         list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2701         return 0;
2702 err_listen:
2703         /* Caller must destroy this after releasing lock */
2704         *to_destroy = dev_id_priv;
2705         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2706         return ret;
2707 }
2708
2709 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2710 {
2711         struct rdma_id_private *to_destroy;
2712         struct cma_device *cma_dev;
2713         int ret;
2714
2715         mutex_lock(&lock);
2716         list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2717         list_for_each_entry(cma_dev, &dev_list, list) {
2718                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2719                 if (ret) {
2720                         /* Prevent racing with cma_process_remove() */
2721                         if (to_destroy)
2722                                 list_del_init(&to_destroy->device_item);
2723                         goto err_listen;
2724                 }
2725         }
2726         mutex_unlock(&lock);
2727         return 0;
2728
2729 err_listen:
2730         _cma_cancel_listens(id_priv);
2731         mutex_unlock(&lock);
2732         if (to_destroy)
2733                 rdma_destroy_id(&to_destroy->id);
2734         return ret;
2735 }
2736
2737 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2738 {
2739         struct rdma_id_private *id_priv;
2740
2741         id_priv = container_of(id, struct rdma_id_private, id);
2742         mutex_lock(&id_priv->qp_mutex);
2743         id_priv->tos = (u8) tos;
2744         id_priv->tos_set = true;
2745         mutex_unlock(&id_priv->qp_mutex);
2746 }
2747 EXPORT_SYMBOL(rdma_set_service_type);
2748
2749 /**
2750  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2751  *                          with a connection identifier.
2752  * @id: Communication identifier to associated with service type.
2753  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2754  *
2755  * This function should be called before rdma_connect() on active side,
2756  * and on passive side before rdma_accept(). It is applicable to primary
2757  * path only. The timeout will affect the local side of the QP, it is not
2758  * negotiated with remote side and zero disables the timer. In case it is
2759  * set before rdma_resolve_route, the value will also be used to determine
2760  * PacketLifeTime for RoCE.
2761  *
2762  * Return: 0 for success
2763  */
2764 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2765 {
2766         struct rdma_id_private *id_priv;
2767
2768         if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2769                 return -EINVAL;
2770
2771         id_priv = container_of(id, struct rdma_id_private, id);
2772         mutex_lock(&id_priv->qp_mutex);
2773         id_priv->timeout = timeout;
2774         id_priv->timeout_set = true;
2775         mutex_unlock(&id_priv->qp_mutex);
2776
2777         return 0;
2778 }
2779 EXPORT_SYMBOL(rdma_set_ack_timeout);
2780
2781 /**
2782  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2783  *                            QP associated with a connection identifier.
2784  * @id: Communication identifier to associated with service type.
2785  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2786  *                 Timer Field" in the IBTA specification.
2787  *
2788  * This function should be called before rdma_connect() on active
2789  * side, and on passive side before rdma_accept(). The timer value
2790  * will be associated with the local QP. When it receives a send it is
2791  * not read to handle, typically if the receive queue is empty, an RNR
2792  * Retry NAK is returned to the requester with the min_rnr_timer
2793  * encoded. The requester will then wait at least the time specified
2794  * in the NAK before retrying. The default is zero, which translates
2795  * to a minimum RNR Timer value of 655 ms.
2796  *
2797  * Return: 0 for success
2798  */
2799 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2800 {
2801         struct rdma_id_private *id_priv;
2802
2803         /* It is a five-bit value */
2804         if (min_rnr_timer & 0xe0)
2805                 return -EINVAL;
2806
2807         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2808                 return -EINVAL;
2809
2810         id_priv = container_of(id, struct rdma_id_private, id);
2811         mutex_lock(&id_priv->qp_mutex);
2812         id_priv->min_rnr_timer = min_rnr_timer;
2813         id_priv->min_rnr_timer_set = true;
2814         mutex_unlock(&id_priv->qp_mutex);
2815
2816         return 0;
2817 }
2818 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2819
2820 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2821                               void *context)
2822 {
2823         struct cma_work *work = context;
2824         struct rdma_route *route;
2825
2826         route = &work->id->id.route;
2827
2828         if (!status) {
2829                 route->num_paths = 1;
2830                 *route->path_rec = *path_rec;
2831         } else {
2832                 work->old_state = RDMA_CM_ROUTE_QUERY;
2833                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2834                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2835                 work->event.status = status;
2836                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2837                                      status);
2838         }
2839
2840         queue_work(cma_wq, &work->work);
2841 }
2842
2843 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2844                               unsigned long timeout_ms, struct cma_work *work)
2845 {
2846         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2847         struct sa_path_rec path_rec;
2848         ib_sa_comp_mask comp_mask;
2849         struct sockaddr_in6 *sin6;
2850         struct sockaddr_ib *sib;
2851
2852         memset(&path_rec, 0, sizeof path_rec);
2853
2854         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2855                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2856         else
2857                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2858         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2859         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2860         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2861         path_rec.numb_path = 1;
2862         path_rec.reversible = 1;
2863         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2864                                                   cma_dst_addr(id_priv));
2865
2866         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2867                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2868                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2869
2870         switch (cma_family(id_priv)) {
2871         case AF_INET:
2872                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2873                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2874                 break;
2875         case AF_INET6:
2876                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2877                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2878                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2879                 break;
2880         case AF_IB:
2881                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2882                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2883                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2884                 break;
2885         }
2886
2887         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2888                                                id_priv->id.port_num, &path_rec,
2889                                                comp_mask, timeout_ms,
2890                                                GFP_KERNEL, cma_query_handler,
2891                                                work, &id_priv->query);
2892
2893         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2894 }
2895
2896 static void cma_iboe_join_work_handler(struct work_struct *work)
2897 {
2898         struct cma_multicast *mc =
2899                 container_of(work, struct cma_multicast, iboe_join.work);
2900         struct rdma_cm_event *event = &mc->iboe_join.event;
2901         struct rdma_id_private *id_priv = mc->id_priv;
2902         int ret;
2903
2904         mutex_lock(&id_priv->handler_mutex);
2905         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2906             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2907                 goto out_unlock;
2908
2909         ret = cma_cm_event_handler(id_priv, event);
2910         WARN_ON(ret);
2911
2912 out_unlock:
2913         mutex_unlock(&id_priv->handler_mutex);
2914         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2915                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2916 }
2917
2918 static void cma_work_handler(struct work_struct *_work)
2919 {
2920         struct cma_work *work = container_of(_work, struct cma_work, work);
2921         struct rdma_id_private *id_priv = work->id;
2922
2923         mutex_lock(&id_priv->handler_mutex);
2924         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2925             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2926                 goto out_unlock;
2927         if (work->old_state != 0 || work->new_state != 0) {
2928                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2929                         goto out_unlock;
2930         }
2931
2932         if (cma_cm_event_handler(id_priv, &work->event)) {
2933                 cma_id_put(id_priv);
2934                 destroy_id_handler_unlock(id_priv);
2935                 goto out_free;
2936         }
2937
2938 out_unlock:
2939         mutex_unlock(&id_priv->handler_mutex);
2940         cma_id_put(id_priv);
2941 out_free:
2942         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2943                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2944         kfree(work);
2945 }
2946
2947 static void cma_init_resolve_route_work(struct cma_work *work,
2948                                         struct rdma_id_private *id_priv)
2949 {
2950         work->id = id_priv;
2951         INIT_WORK(&work->work, cma_work_handler);
2952         work->old_state = RDMA_CM_ROUTE_QUERY;
2953         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2954         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2955 }
2956
2957 static void enqueue_resolve_addr_work(struct cma_work *work,
2958                                       struct rdma_id_private *id_priv)
2959 {
2960         /* Balances with cma_id_put() in cma_work_handler */
2961         cma_id_get(id_priv);
2962
2963         work->id = id_priv;
2964         INIT_WORK(&work->work, cma_work_handler);
2965         work->old_state = RDMA_CM_ADDR_QUERY;
2966         work->new_state = RDMA_CM_ADDR_RESOLVED;
2967         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2968
2969         queue_work(cma_wq, &work->work);
2970 }
2971
2972 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2973                                 unsigned long timeout_ms)
2974 {
2975         struct rdma_route *route = &id_priv->id.route;
2976         struct cma_work *work;
2977         int ret;
2978
2979         work = kzalloc(sizeof *work, GFP_KERNEL);
2980         if (!work)
2981                 return -ENOMEM;
2982
2983         cma_init_resolve_route_work(work, id_priv);
2984
2985         if (!route->path_rec)
2986                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2987         if (!route->path_rec) {
2988                 ret = -ENOMEM;
2989                 goto err1;
2990         }
2991
2992         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2993         if (ret)
2994                 goto err2;
2995
2996         return 0;
2997 err2:
2998         kfree(route->path_rec);
2999         route->path_rec = NULL;
3000 err1:
3001         kfree(work);
3002         return ret;
3003 }
3004
3005 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
3006                                            unsigned long supported_gids,
3007                                            enum ib_gid_type default_gid)
3008 {
3009         if ((network_type == RDMA_NETWORK_IPV4 ||
3010              network_type == RDMA_NETWORK_IPV6) &&
3011             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
3012                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
3013
3014         return default_gid;
3015 }
3016
3017 /*
3018  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
3019  * path record type based on GID type.
3020  * It also sets up other L2 fields which includes destination mac address
3021  * netdev ifindex, of the path record.
3022  * It returns the netdev of the bound interface for this path record entry.
3023  */
3024 static struct net_device *
3025 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
3026 {
3027         struct rdma_route *route = &id_priv->id.route;
3028         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
3029         struct rdma_addr *addr = &route->addr;
3030         unsigned long supported_gids;
3031         struct net_device *ndev;
3032
3033         if (!addr->dev_addr.bound_dev_if)
3034                 return NULL;
3035
3036         ndev = dev_get_by_index(addr->dev_addr.net,
3037                                 addr->dev_addr.bound_dev_if);
3038         if (!ndev)
3039                 return NULL;
3040
3041         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
3042                                                     id_priv->id.port_num);
3043         gid_type = cma_route_gid_type(addr->dev_addr.network,
3044                                       supported_gids,
3045                                       id_priv->gid_type);
3046         /* Use the hint from IP Stack to select GID Type */
3047         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
3048                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
3049         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
3050
3051         route->path_rec->roce.route_resolved = true;
3052         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
3053         return ndev;
3054 }
3055
3056 int rdma_set_ib_path(struct rdma_cm_id *id,
3057                      struct sa_path_rec *path_rec)
3058 {
3059         struct rdma_id_private *id_priv;
3060         struct net_device *ndev;
3061         int ret;
3062
3063         id_priv = container_of(id, struct rdma_id_private, id);
3064         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3065                            RDMA_CM_ROUTE_RESOLVED))
3066                 return -EINVAL;
3067
3068         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
3069                                      GFP_KERNEL);
3070         if (!id->route.path_rec) {
3071                 ret = -ENOMEM;
3072                 goto err;
3073         }
3074
3075         if (rdma_protocol_roce(id->device, id->port_num)) {
3076                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3077                 if (!ndev) {
3078                         ret = -ENODEV;
3079                         goto err_free;
3080                 }
3081                 dev_put(ndev);
3082         }
3083
3084         id->route.num_paths = 1;
3085         return 0;
3086
3087 err_free:
3088         kfree(id->route.path_rec);
3089         id->route.path_rec = NULL;
3090 err:
3091         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
3092         return ret;
3093 }
3094 EXPORT_SYMBOL(rdma_set_ib_path);
3095
3096 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
3097 {
3098         struct cma_work *work;
3099
3100         work = kzalloc(sizeof *work, GFP_KERNEL);
3101         if (!work)
3102                 return -ENOMEM;
3103
3104         cma_init_resolve_route_work(work, id_priv);
3105         queue_work(cma_wq, &work->work);
3106         return 0;
3107 }
3108
3109 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
3110 {
3111         struct net_device *dev;
3112
3113         dev = vlan_dev_real_dev(vlan_ndev);
3114         if (dev->num_tc)
3115                 return netdev_get_prio_tc_map(dev, prio);
3116
3117         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
3118                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
3119 }
3120
3121 struct iboe_prio_tc_map {
3122         int input_prio;
3123         int output_tc;
3124         bool found;
3125 };
3126
3127 static int get_lower_vlan_dev_tc(struct net_device *dev,
3128                                  struct netdev_nested_priv *priv)
3129 {
3130         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3131
3132         if (is_vlan_dev(dev))
3133                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3134         else if (dev->num_tc)
3135                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3136         else
3137                 map->output_tc = 0;
3138         /* We are interested only in first level VLAN device, so always
3139          * return 1 to stop iterating over next level devices.
3140          */
3141         map->found = true;
3142         return 1;
3143 }
3144
3145 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3146 {
3147         struct iboe_prio_tc_map prio_tc_map = {};
3148         int prio = rt_tos2priority(tos);
3149         struct netdev_nested_priv priv;
3150
3151         /* If VLAN device, get it directly from the VLAN netdev */
3152         if (is_vlan_dev(ndev))
3153                 return get_vlan_ndev_tc(ndev, prio);
3154
3155         prio_tc_map.input_prio = prio;
3156         priv.data = (void *)&prio_tc_map;
3157         rcu_read_lock();
3158         netdev_walk_all_lower_dev_rcu(ndev,
3159                                       get_lower_vlan_dev_tc,
3160                                       &priv);
3161         rcu_read_unlock();
3162         /* If map is found from lower device, use it; Otherwise
3163          * continue with the current netdevice to get priority to tc map.
3164          */
3165         if (prio_tc_map.found)
3166                 return prio_tc_map.output_tc;
3167         else if (ndev->num_tc)
3168                 return netdev_get_prio_tc_map(ndev, prio);
3169         else
3170                 return 0;
3171 }
3172
3173 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3174 {
3175         struct sockaddr_in6 *addr6;
3176         u16 dport, sport;
3177         u32 hash, fl;
3178
3179         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3180         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3181         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3182                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3183                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3184                 hash = (u32)sport * 31 + dport;
3185                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3186         }
3187
3188         return cpu_to_be32(fl);
3189 }
3190
3191 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3192 {
3193         struct rdma_route *route = &id_priv->id.route;
3194         struct rdma_addr *addr = &route->addr;
3195         struct cma_work *work;
3196         int ret;
3197         struct net_device *ndev;
3198
3199         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3200                                         rdma_start_port(id_priv->cma_dev->device)];
3201         u8 tos;
3202
3203         mutex_lock(&id_priv->qp_mutex);
3204         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3205         mutex_unlock(&id_priv->qp_mutex);
3206
3207         work = kzalloc(sizeof *work, GFP_KERNEL);
3208         if (!work)
3209                 return -ENOMEM;
3210
3211         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3212         if (!route->path_rec) {
3213                 ret = -ENOMEM;
3214                 goto err1;
3215         }
3216
3217         route->num_paths = 1;
3218
3219         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3220         if (!ndev) {
3221                 ret = -ENODEV;
3222                 goto err2;
3223         }
3224
3225         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3226                     &route->path_rec->sgid);
3227         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3228                     &route->path_rec->dgid);
3229
3230         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3231                 /* TODO: get the hoplimit from the inet/inet6 device */
3232                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3233         else
3234                 route->path_rec->hop_limit = 1;
3235         route->path_rec->reversible = 1;
3236         route->path_rec->pkey = cpu_to_be16(0xffff);
3237         route->path_rec->mtu_selector = IB_SA_EQ;
3238         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3239         route->path_rec->traffic_class = tos;
3240         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3241         route->path_rec->rate_selector = IB_SA_EQ;
3242         route->path_rec->rate = iboe_get_rate(ndev);
3243         dev_put(ndev);
3244         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3245         /* In case ACK timeout is set, use this value to calculate
3246          * PacketLifeTime.  As per IBTA 12.7.34,
3247          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3248          * Assuming a negligible local ACK delay, we can use
3249          * PacketLifeTime = local ACK timeout/2
3250          * as a reasonable approximation for RoCE networks.
3251          */
3252         mutex_lock(&id_priv->qp_mutex);
3253         if (id_priv->timeout_set && id_priv->timeout)
3254                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3255         else
3256                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3257         mutex_unlock(&id_priv->qp_mutex);
3258
3259         if (!route->path_rec->mtu) {
3260                 ret = -EINVAL;
3261                 goto err2;
3262         }
3263
3264         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3265                                          id_priv->id.port_num))
3266                 route->path_rec->flow_label =
3267                         cma_get_roce_udp_flow_label(id_priv);
3268
3269         cma_init_resolve_route_work(work, id_priv);
3270         queue_work(cma_wq, &work->work);
3271
3272         return 0;
3273
3274 err2:
3275         kfree(route->path_rec);
3276         route->path_rec = NULL;
3277         route->num_paths = 0;
3278 err1:
3279         kfree(work);
3280         return ret;
3281 }
3282
3283 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3284 {
3285         struct rdma_id_private *id_priv;
3286         int ret;
3287
3288         if (!timeout_ms)
3289                 return -EINVAL;
3290
3291         id_priv = container_of(id, struct rdma_id_private, id);
3292         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3293                 return -EINVAL;
3294
3295         cma_id_get(id_priv);
3296         if (rdma_cap_ib_sa(id->device, id->port_num))
3297                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3298         else if (rdma_protocol_roce(id->device, id->port_num)) {
3299                 ret = cma_resolve_iboe_route(id_priv);
3300                 if (!ret)
3301                         cma_add_id_to_tree(id_priv);
3302         }
3303         else if (rdma_protocol_iwarp(id->device, id->port_num))
3304                 ret = cma_resolve_iw_route(id_priv);
3305         else
3306                 ret = -ENOSYS;
3307
3308         if (ret)
3309                 goto err;
3310
3311         return 0;
3312 err:
3313         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3314         cma_id_put(id_priv);
3315         return ret;
3316 }
3317 EXPORT_SYMBOL(rdma_resolve_route);
3318
3319 static void cma_set_loopback(struct sockaddr *addr)
3320 {
3321         switch (addr->sa_family) {
3322         case AF_INET:
3323                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3324                 break;
3325         case AF_INET6:
3326                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3327                               0, 0, 0, htonl(1));
3328                 break;
3329         default:
3330                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3331                             0, 0, 0, htonl(1));
3332                 break;
3333         }
3334 }
3335
3336 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3337 {
3338         struct cma_device *cma_dev, *cur_dev;
3339         union ib_gid gid;
3340         enum ib_port_state port_state;
3341         unsigned int p;
3342         u16 pkey;
3343         int ret;
3344
3345         cma_dev = NULL;
3346         mutex_lock(&lock);
3347         list_for_each_entry(cur_dev, &dev_list, list) {
3348                 if (cma_family(id_priv) == AF_IB &&
3349                     !rdma_cap_ib_cm(cur_dev->device, 1))
3350                         continue;
3351
3352                 if (!cma_dev)
3353                         cma_dev = cur_dev;
3354
3355                 rdma_for_each_port (cur_dev->device, p) {
3356                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3357                             port_state == IB_PORT_ACTIVE) {
3358                                 cma_dev = cur_dev;
3359                                 goto port_found;
3360                         }
3361                 }
3362         }
3363
3364         if (!cma_dev) {
3365                 ret = -ENODEV;
3366                 goto out;
3367         }
3368
3369         p = 1;
3370
3371 port_found:
3372         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3373         if (ret)
3374                 goto out;
3375
3376         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3377         if (ret)
3378                 goto out;
3379
3380         id_priv->id.route.addr.dev_addr.dev_type =
3381                 (rdma_protocol_ib(cma_dev->device, p)) ?
3382                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3383
3384         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3385         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3386         id_priv->id.port_num = p;
3387         cma_attach_to_dev(id_priv, cma_dev);
3388         rdma_restrack_add(&id_priv->res);
3389         cma_set_loopback(cma_src_addr(id_priv));
3390 out:
3391         mutex_unlock(&lock);
3392         return ret;
3393 }
3394
3395 static void addr_handler(int status, struct sockaddr *src_addr,
3396                          struct rdma_dev_addr *dev_addr, void *context)
3397 {
3398         struct rdma_id_private *id_priv = context;
3399         struct rdma_cm_event event = {};
3400         struct sockaddr *addr;
3401         struct sockaddr_storage old_addr;
3402
3403         mutex_lock(&id_priv->handler_mutex);
3404         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3405                            RDMA_CM_ADDR_RESOLVED))
3406                 goto out;
3407
3408         /*
3409          * Store the previous src address, so that if we fail to acquire
3410          * matching rdma device, old address can be restored back, which helps
3411          * to cancel the cma listen operation correctly.
3412          */
3413         addr = cma_src_addr(id_priv);
3414         memcpy(&old_addr, addr, rdma_addr_size(addr));
3415         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3416         if (!status && !id_priv->cma_dev) {
3417                 status = cma_acquire_dev_by_src_ip(id_priv);
3418                 if (status)
3419                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3420                                              status);
3421                 rdma_restrack_add(&id_priv->res);
3422         } else if (status) {
3423                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3424         }
3425
3426         if (status) {
3427                 memcpy(addr, &old_addr,
3428                        rdma_addr_size((struct sockaddr *)&old_addr));
3429                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3430                                    RDMA_CM_ADDR_BOUND))
3431                         goto out;
3432                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3433                 event.status = status;
3434         } else
3435                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3436
3437         if (cma_cm_event_handler(id_priv, &event)) {
3438                 destroy_id_handler_unlock(id_priv);
3439                 return;
3440         }
3441 out:
3442         mutex_unlock(&id_priv->handler_mutex);
3443 }
3444
3445 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3446 {
3447         struct cma_work *work;
3448         union ib_gid gid;
3449         int ret;
3450
3451         work = kzalloc(sizeof *work, GFP_KERNEL);
3452         if (!work)
3453                 return -ENOMEM;
3454
3455         if (!id_priv->cma_dev) {
3456                 ret = cma_bind_loopback(id_priv);
3457                 if (ret)
3458                         goto err;
3459         }
3460
3461         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3462         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3463
3464         enqueue_resolve_addr_work(work, id_priv);
3465         return 0;
3466 err:
3467         kfree(work);
3468         return ret;
3469 }
3470
3471 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3472 {
3473         struct cma_work *work;
3474         int ret;
3475
3476         work = kzalloc(sizeof *work, GFP_KERNEL);
3477         if (!work)
3478                 return -ENOMEM;
3479
3480         if (!id_priv->cma_dev) {
3481                 ret = cma_resolve_ib_dev(id_priv);
3482                 if (ret)
3483                         goto err;
3484         }
3485
3486         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3487                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3488
3489         enqueue_resolve_addr_work(work, id_priv);
3490         return 0;
3491 err:
3492         kfree(work);
3493         return ret;
3494 }
3495
3496 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3497                          const struct sockaddr *dst_addr)
3498 {
3499         struct sockaddr_storage zero_sock = {};
3500
3501         if (src_addr && src_addr->sa_family)
3502                 return rdma_bind_addr(id, src_addr);
3503
3504         /*
3505          * When the src_addr is not specified, automatically supply an any addr
3506          */
3507         zero_sock.ss_family = dst_addr->sa_family;
3508         if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
3509                 struct sockaddr_in6 *src_addr6 =
3510                         (struct sockaddr_in6 *)&zero_sock;
3511                 struct sockaddr_in6 *dst_addr6 =
3512                         (struct sockaddr_in6 *)dst_addr;
3513
3514                 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3515                 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3516                         id->route.addr.dev_addr.bound_dev_if =
3517                                 dst_addr6->sin6_scope_id;
3518         } else if (dst_addr->sa_family == AF_IB) {
3519                 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
3520                         ((struct sockaddr_ib *)dst_addr)->sib_pkey;
3521         }
3522         return rdma_bind_addr(id, (struct sockaddr *)&zero_sock);
3523 }
3524
3525 /*
3526  * If required, resolve the source address for bind and leave the id_priv in
3527  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3528  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3529  * ignored.
3530  */
3531 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3532                                struct sockaddr *src_addr,
3533                                const struct sockaddr *dst_addr)
3534 {
3535         int ret;
3536
3537         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3538         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3539                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3540                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3541                 if (ret)
3542                         goto err_dst;
3543                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3544                                            RDMA_CM_ADDR_QUERY))) {
3545                         ret = -EINVAL;
3546                         goto err_dst;
3547                 }
3548         }
3549
3550         if (cma_family(id_priv) != dst_addr->sa_family) {
3551                 ret = -EINVAL;
3552                 goto err_state;
3553         }
3554         return 0;
3555
3556 err_state:
3557         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3558 err_dst:
3559         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3560         return ret;
3561 }
3562
3563 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3564                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3565 {
3566         struct rdma_id_private *id_priv =
3567                 container_of(id, struct rdma_id_private, id);
3568         int ret;
3569
3570         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3571         if (ret)
3572                 return ret;
3573
3574         if (cma_any_addr(dst_addr)) {
3575                 ret = cma_resolve_loopback(id_priv);
3576         } else {
3577                 if (dst_addr->sa_family == AF_IB) {
3578                         ret = cma_resolve_ib_addr(id_priv);
3579                 } else {
3580                         /*
3581                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
3582                          * rdma_resolve_ip() is called, eg through the error
3583                          * path in addr_handler(). If this happens the existing
3584                          * request must be canceled before issuing a new one.
3585                          * Since canceling a request is a bit slow and this
3586                          * oddball path is rare, keep track once a request has
3587                          * been issued. The track turns out to be a permanent
3588                          * state since this is the only cancel as it is
3589                          * immediately before rdma_resolve_ip().
3590                          */
3591                         if (id_priv->used_resolve_ip)
3592                                 rdma_addr_cancel(&id->route.addr.dev_addr);
3593                         else
3594                                 id_priv->used_resolve_ip = 1;
3595                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3596                                               &id->route.addr.dev_addr,
3597                                               timeout_ms, addr_handler,
3598                                               false, id_priv);
3599                 }
3600         }
3601         if (ret)
3602                 goto err;
3603
3604         return 0;
3605 err:
3606         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3607         return ret;
3608 }
3609 EXPORT_SYMBOL(rdma_resolve_addr);
3610
3611 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3612 {
3613         struct rdma_id_private *id_priv;
3614         unsigned long flags;
3615         int ret;
3616
3617         id_priv = container_of(id, struct rdma_id_private, id);
3618         spin_lock_irqsave(&id_priv->lock, flags);
3619         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3620             id_priv->state == RDMA_CM_IDLE) {
3621                 id_priv->reuseaddr = reuse;
3622                 ret = 0;
3623         } else {
3624                 ret = -EINVAL;
3625         }
3626         spin_unlock_irqrestore(&id_priv->lock, flags);
3627         return ret;
3628 }
3629 EXPORT_SYMBOL(rdma_set_reuseaddr);
3630
3631 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3632 {
3633         struct rdma_id_private *id_priv;
3634         unsigned long flags;
3635         int ret;
3636
3637         id_priv = container_of(id, struct rdma_id_private, id);
3638         spin_lock_irqsave(&id_priv->lock, flags);
3639         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3640                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3641                 id_priv->afonly = afonly;
3642                 ret = 0;
3643         } else {
3644                 ret = -EINVAL;
3645         }
3646         spin_unlock_irqrestore(&id_priv->lock, flags);
3647         return ret;
3648 }
3649 EXPORT_SYMBOL(rdma_set_afonly);
3650
3651 static void cma_bind_port(struct rdma_bind_list *bind_list,
3652                           struct rdma_id_private *id_priv)
3653 {
3654         struct sockaddr *addr;
3655         struct sockaddr_ib *sib;
3656         u64 sid, mask;
3657         __be16 port;
3658
3659         lockdep_assert_held(&lock);
3660
3661         addr = cma_src_addr(id_priv);
3662         port = htons(bind_list->port);
3663
3664         switch (addr->sa_family) {
3665         case AF_INET:
3666                 ((struct sockaddr_in *) addr)->sin_port = port;
3667                 break;
3668         case AF_INET6:
3669                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3670                 break;
3671         case AF_IB:
3672                 sib = (struct sockaddr_ib *) addr;
3673                 sid = be64_to_cpu(sib->sib_sid);
3674                 mask = be64_to_cpu(sib->sib_sid_mask);
3675                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3676                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3677                 break;
3678         }
3679         id_priv->bind_list = bind_list;
3680         hlist_add_head(&id_priv->node, &bind_list->owners);
3681 }
3682
3683 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3684                           struct rdma_id_private *id_priv, unsigned short snum)
3685 {
3686         struct rdma_bind_list *bind_list;
3687         int ret;
3688
3689         lockdep_assert_held(&lock);
3690
3691         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3692         if (!bind_list)
3693                 return -ENOMEM;
3694
3695         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3696                            snum);
3697         if (ret < 0)
3698                 goto err;
3699
3700         bind_list->ps = ps;
3701         bind_list->port = snum;
3702         cma_bind_port(bind_list, id_priv);
3703         return 0;
3704 err:
3705         kfree(bind_list);
3706         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3707 }
3708
3709 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3710                               struct rdma_id_private *id_priv)
3711 {
3712         struct rdma_id_private *cur_id;
3713         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3714         struct sockaddr  *saddr = cma_src_addr(id_priv);
3715         __be16 dport = cma_port(daddr);
3716
3717         lockdep_assert_held(&lock);
3718
3719         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3720                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3721                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3722                 __be16 cur_dport = cma_port(cur_daddr);
3723
3724                 if (id_priv == cur_id)
3725                         continue;
3726
3727                 /* different dest port -> unique */
3728                 if (!cma_any_port(daddr) &&
3729                     !cma_any_port(cur_daddr) &&
3730                     (dport != cur_dport))
3731                         continue;
3732
3733                 /* different src address -> unique */
3734                 if (!cma_any_addr(saddr) &&
3735                     !cma_any_addr(cur_saddr) &&
3736                     cma_addr_cmp(saddr, cur_saddr))
3737                         continue;
3738
3739                 /* different dst address -> unique */
3740                 if (!cma_any_addr(daddr) &&
3741                     !cma_any_addr(cur_daddr) &&
3742                     cma_addr_cmp(daddr, cur_daddr))
3743                         continue;
3744
3745                 return -EADDRNOTAVAIL;
3746         }
3747         return 0;
3748 }
3749
3750 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3751                               struct rdma_id_private *id_priv)
3752 {
3753         static unsigned int last_used_port;
3754         int low, high, remaining;
3755         unsigned int rover;
3756         struct net *net = id_priv->id.route.addr.dev_addr.net;
3757
3758         lockdep_assert_held(&lock);
3759
3760         inet_get_local_port_range(net, &low, &high);
3761         remaining = (high - low) + 1;
3762         rover = prandom_u32() % remaining + low;
3763 retry:
3764         if (last_used_port != rover) {
3765                 struct rdma_bind_list *bind_list;
3766                 int ret;
3767
3768                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3769
3770                 if (!bind_list) {
3771                         ret = cma_alloc_port(ps, id_priv, rover);
3772                 } else {
3773                         ret = cma_port_is_unique(bind_list, id_priv);
3774                         if (!ret)
3775                                 cma_bind_port(bind_list, id_priv);
3776                 }
3777                 /*
3778                  * Remember previously used port number in order to avoid
3779                  * re-using same port immediately after it is closed.
3780                  */
3781                 if (!ret)
3782                         last_used_port = rover;
3783                 if (ret != -EADDRNOTAVAIL)
3784                         return ret;
3785         }
3786         if (--remaining) {
3787                 rover++;
3788                 if ((rover < low) || (rover > high))
3789                         rover = low;
3790                 goto retry;
3791         }
3792         return -EADDRNOTAVAIL;
3793 }
3794
3795 /*
3796  * Check that the requested port is available.  This is called when trying to
3797  * bind to a specific port, or when trying to listen on a bound port.  In
3798  * the latter case, the provided id_priv may already be on the bind_list, but
3799  * we still need to check that it's okay to start listening.
3800  */
3801 static int cma_check_port(struct rdma_bind_list *bind_list,
3802                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3803 {
3804         struct rdma_id_private *cur_id;
3805         struct sockaddr *addr, *cur_addr;
3806
3807         lockdep_assert_held(&lock);
3808
3809         addr = cma_src_addr(id_priv);
3810         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3811                 if (id_priv == cur_id)
3812                         continue;
3813
3814                 if (reuseaddr && cur_id->reuseaddr)
3815                         continue;
3816
3817                 cur_addr = cma_src_addr(cur_id);
3818                 if (id_priv->afonly && cur_id->afonly &&
3819                     (addr->sa_family != cur_addr->sa_family))
3820                         continue;
3821
3822                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3823                         return -EADDRNOTAVAIL;
3824
3825                 if (!cma_addr_cmp(addr, cur_addr))
3826                         return -EADDRINUSE;
3827         }
3828         return 0;
3829 }
3830
3831 static int cma_use_port(enum rdma_ucm_port_space ps,
3832                         struct rdma_id_private *id_priv)
3833 {
3834         struct rdma_bind_list *bind_list;
3835         unsigned short snum;
3836         int ret;
3837
3838         lockdep_assert_held(&lock);
3839
3840         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3841         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3842                 return -EACCES;
3843
3844         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3845         if (!bind_list) {
3846                 ret = cma_alloc_port(ps, id_priv, snum);
3847         } else {
3848                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3849                 if (!ret)
3850                         cma_bind_port(bind_list, id_priv);
3851         }
3852         return ret;
3853 }
3854
3855 static enum rdma_ucm_port_space
3856 cma_select_inet_ps(struct rdma_id_private *id_priv)
3857 {
3858         switch (id_priv->id.ps) {
3859         case RDMA_PS_TCP:
3860         case RDMA_PS_UDP:
3861         case RDMA_PS_IPOIB:
3862         case RDMA_PS_IB:
3863                 return id_priv->id.ps;
3864         default:
3865
3866                 return 0;
3867         }
3868 }
3869
3870 static enum rdma_ucm_port_space
3871 cma_select_ib_ps(struct rdma_id_private *id_priv)
3872 {
3873         enum rdma_ucm_port_space ps = 0;
3874         struct sockaddr_ib *sib;
3875         u64 sid_ps, mask, sid;
3876
3877         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3878         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3879         sid = be64_to_cpu(sib->sib_sid) & mask;
3880
3881         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3882                 sid_ps = RDMA_IB_IP_PS_IB;
3883                 ps = RDMA_PS_IB;
3884         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3885                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3886                 sid_ps = RDMA_IB_IP_PS_TCP;
3887                 ps = RDMA_PS_TCP;
3888         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3889                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3890                 sid_ps = RDMA_IB_IP_PS_UDP;
3891                 ps = RDMA_PS_UDP;
3892         }
3893
3894         if (ps) {
3895                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3896                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3897                                                 be64_to_cpu(sib->sib_sid_mask));
3898         }
3899         return ps;
3900 }
3901
3902 static int cma_get_port(struct rdma_id_private *id_priv)
3903 {
3904         enum rdma_ucm_port_space ps;
3905         int ret;
3906
3907         if (cma_family(id_priv) != AF_IB)
3908                 ps = cma_select_inet_ps(id_priv);
3909         else
3910                 ps = cma_select_ib_ps(id_priv);
3911         if (!ps)
3912                 return -EPROTONOSUPPORT;
3913
3914         mutex_lock(&lock);
3915         if (cma_any_port(cma_src_addr(id_priv)))
3916                 ret = cma_alloc_any_port(ps, id_priv);
3917         else
3918                 ret = cma_use_port(ps, id_priv);
3919         mutex_unlock(&lock);
3920
3921         return ret;
3922 }
3923
3924 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3925                                struct sockaddr *addr)
3926 {
3927 #if IS_ENABLED(CONFIG_IPV6)
3928         struct sockaddr_in6 *sin6;
3929
3930         if (addr->sa_family != AF_INET6)
3931                 return 0;
3932
3933         sin6 = (struct sockaddr_in6 *) addr;
3934
3935         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3936                 return 0;
3937
3938         if (!sin6->sin6_scope_id)
3939                         return -EINVAL;
3940
3941         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3942 #endif
3943         return 0;
3944 }
3945
3946 int rdma_listen(struct rdma_cm_id *id, int backlog)
3947 {
3948         struct rdma_id_private *id_priv =
3949                 container_of(id, struct rdma_id_private, id);
3950         int ret;
3951
3952         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3953                 struct sockaddr_in any_in = {
3954                         .sin_family = AF_INET,
3955                         .sin_addr.s_addr = htonl(INADDR_ANY),
3956                 };
3957
3958                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3959                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3960                 if (ret)
3961                         return ret;
3962                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3963                                            RDMA_CM_LISTEN)))
3964                         return -EINVAL;
3965         }
3966
3967         /*
3968          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3969          * any more, and has to be unique in the bind list.
3970          */
3971         if (id_priv->reuseaddr) {
3972                 mutex_lock(&lock);
3973                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3974                 if (!ret)
3975                         id_priv->reuseaddr = 0;
3976                 mutex_unlock(&lock);
3977                 if (ret)
3978                         goto err;
3979         }
3980
3981         id_priv->backlog = backlog;
3982         if (id_priv->cma_dev) {
3983                 if (rdma_cap_ib_cm(id->device, 1)) {
3984                         ret = cma_ib_listen(id_priv);
3985                         if (ret)
3986                                 goto err;
3987                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3988                         ret = cma_iw_listen(id_priv, backlog);
3989                         if (ret)
3990                                 goto err;
3991                 } else {
3992                         ret = -ENOSYS;
3993                         goto err;
3994                 }
3995         } else {
3996                 ret = cma_listen_on_all(id_priv);
3997                 if (ret)
3998                         goto err;
3999         }
4000
4001         return 0;
4002 err:
4003         id_priv->backlog = 0;
4004         /*
4005          * All the failure paths that lead here will not allow the req_handler's
4006          * to have run.
4007          */
4008         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
4009         return ret;
4010 }
4011 EXPORT_SYMBOL(rdma_listen);
4012
4013 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
4014 {
4015         struct rdma_id_private *id_priv;
4016         int ret;
4017         struct sockaddr  *daddr;
4018
4019         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
4020             addr->sa_family != AF_IB)
4021                 return -EAFNOSUPPORT;
4022
4023         id_priv = container_of(id, struct rdma_id_private, id);
4024         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
4025                 return -EINVAL;
4026
4027         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
4028         if (ret)
4029                 goto err1;
4030
4031         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
4032         if (!cma_any_addr(addr)) {
4033                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
4034                 if (ret)
4035                         goto err1;
4036
4037                 ret = cma_acquire_dev_by_src_ip(id_priv);
4038                 if (ret)
4039                         goto err1;
4040         }
4041
4042         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
4043                 if (addr->sa_family == AF_INET)
4044                         id_priv->afonly = 1;
4045 #if IS_ENABLED(CONFIG_IPV6)
4046                 else if (addr->sa_family == AF_INET6) {
4047                         struct net *net = id_priv->id.route.addr.dev_addr.net;
4048
4049                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
4050                 }
4051 #endif
4052         }
4053         daddr = cma_dst_addr(id_priv);
4054         daddr->sa_family = addr->sa_family;
4055
4056         ret = cma_get_port(id_priv);
4057         if (ret)
4058                 goto err2;
4059
4060         if (!cma_any_addr(addr))
4061                 rdma_restrack_add(&id_priv->res);
4062         return 0;
4063 err2:
4064         if (id_priv->cma_dev)
4065                 cma_release_dev(id_priv);
4066 err1:
4067         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
4068         return ret;
4069 }
4070 EXPORT_SYMBOL(rdma_bind_addr);
4071
4072 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
4073 {
4074         struct cma_hdr *cma_hdr;
4075
4076         cma_hdr = hdr;
4077         cma_hdr->cma_version = CMA_VERSION;
4078         if (cma_family(id_priv) == AF_INET) {
4079                 struct sockaddr_in *src4, *dst4;
4080
4081                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
4082                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
4083
4084                 cma_set_ip_ver(cma_hdr, 4);
4085                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
4086                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
4087                 cma_hdr->port = src4->sin_port;
4088         } else if (cma_family(id_priv) == AF_INET6) {
4089                 struct sockaddr_in6 *src6, *dst6;
4090
4091                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
4092                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
4093
4094                 cma_set_ip_ver(cma_hdr, 6);
4095                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
4096                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
4097                 cma_hdr->port = src6->sin6_port;
4098         }
4099         return 0;
4100 }
4101
4102 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
4103                                 const struct ib_cm_event *ib_event)
4104 {
4105         struct rdma_id_private *id_priv = cm_id->context;
4106         struct rdma_cm_event event = {};
4107         const struct ib_cm_sidr_rep_event_param *rep =
4108                                 &ib_event->param.sidr_rep_rcvd;
4109         int ret;
4110
4111         mutex_lock(&id_priv->handler_mutex);
4112         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4113                 goto out;
4114
4115         switch (ib_event->event) {
4116         case IB_CM_SIDR_REQ_ERROR:
4117                 event.event = RDMA_CM_EVENT_UNREACHABLE;
4118                 event.status = -ETIMEDOUT;
4119                 break;
4120         case IB_CM_SIDR_REP_RECEIVED:
4121                 event.param.ud.private_data = ib_event->private_data;
4122                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
4123                 if (rep->status != IB_SIDR_SUCCESS) {
4124                         event.event = RDMA_CM_EVENT_UNREACHABLE;
4125                         event.status = ib_event->param.sidr_rep_rcvd.status;
4126                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
4127                                              event.status);
4128                         break;
4129                 }
4130                 ret = cma_set_qkey(id_priv, rep->qkey);
4131                 if (ret) {
4132                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
4133                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
4134                         event.status = ret;
4135                         break;
4136                 }
4137                 ib_init_ah_attr_from_path(id_priv->id.device,
4138                                           id_priv->id.port_num,
4139                                           id_priv->id.route.path_rec,
4140                                           &event.param.ud.ah_attr,
4141                                           rep->sgid_attr);
4142                 event.param.ud.qp_num = rep->qpn;
4143                 event.param.ud.qkey = rep->qkey;
4144                 event.event = RDMA_CM_EVENT_ESTABLISHED;
4145                 event.status = 0;
4146                 break;
4147         default:
4148                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4149                        ib_event->event);
4150                 goto out;
4151         }
4152
4153         ret = cma_cm_event_handler(id_priv, &event);
4154
4155         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4156         if (ret) {
4157                 /* Destroy the CM ID by returning a non-zero value. */
4158                 id_priv->cm_id.ib = NULL;
4159                 destroy_id_handler_unlock(id_priv);
4160                 return ret;
4161         }
4162 out:
4163         mutex_unlock(&id_priv->handler_mutex);
4164         return 0;
4165 }
4166
4167 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4168                               struct rdma_conn_param *conn_param)
4169 {
4170         struct ib_cm_sidr_req_param req;
4171         struct ib_cm_id *id;
4172         void *private_data;
4173         u8 offset;
4174         int ret;
4175
4176         memset(&req, 0, sizeof req);
4177         offset = cma_user_data_offset(id_priv);
4178         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4179                 return -EINVAL;
4180
4181         if (req.private_data_len) {
4182                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4183                 if (!private_data)
4184                         return -ENOMEM;
4185         } else {
4186                 private_data = NULL;
4187         }
4188
4189         if (conn_param->private_data && conn_param->private_data_len)
4190                 memcpy(private_data + offset, conn_param->private_data,
4191                        conn_param->private_data_len);
4192
4193         if (private_data) {
4194                 ret = cma_format_hdr(private_data, id_priv);
4195                 if (ret)
4196                         goto out;
4197                 req.private_data = private_data;
4198         }
4199
4200         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4201                              id_priv);
4202         if (IS_ERR(id)) {
4203                 ret = PTR_ERR(id);
4204                 goto out;
4205         }
4206         id_priv->cm_id.ib = id;
4207
4208         req.path = id_priv->id.route.path_rec;
4209         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4210         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4211         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4212         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4213
4214         trace_cm_send_sidr_req(id_priv);
4215         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4216         if (ret) {
4217                 ib_destroy_cm_id(id_priv->cm_id.ib);
4218                 id_priv->cm_id.ib = NULL;
4219         }
4220 out:
4221         kfree(private_data);
4222         return ret;
4223 }
4224
4225 static int cma_connect_ib(struct rdma_id_private *id_priv,
4226                           struct rdma_conn_param *conn_param)
4227 {
4228         struct ib_cm_req_param req;
4229         struct rdma_route *route;
4230         void *private_data;
4231         struct ib_cm_id *id;
4232         u8 offset;
4233         int ret;
4234
4235         memset(&req, 0, sizeof req);
4236         offset = cma_user_data_offset(id_priv);
4237         if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4238                 return -EINVAL;
4239
4240         if (req.private_data_len) {
4241                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4242                 if (!private_data)
4243                         return -ENOMEM;
4244         } else {
4245                 private_data = NULL;
4246         }
4247
4248         if (conn_param->private_data && conn_param->private_data_len)
4249                 memcpy(private_data + offset, conn_param->private_data,
4250                        conn_param->private_data_len);
4251
4252         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4253         if (IS_ERR(id)) {
4254                 ret = PTR_ERR(id);
4255                 goto out;
4256         }
4257         id_priv->cm_id.ib = id;
4258
4259         route = &id_priv->id.route;
4260         if (private_data) {
4261                 ret = cma_format_hdr(private_data, id_priv);
4262                 if (ret)
4263                         goto out;
4264                 req.private_data = private_data;
4265         }
4266
4267         req.primary_path = &route->path_rec[0];
4268         if (route->num_paths == 2)
4269                 req.alternate_path = &route->path_rec[1];
4270
4271         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4272         /* Alternate path SGID attribute currently unsupported */
4273         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4274         req.qp_num = id_priv->qp_num;
4275         req.qp_type = id_priv->id.qp_type;
4276         req.starting_psn = id_priv->seq_num;
4277         req.responder_resources = conn_param->responder_resources;
4278         req.initiator_depth = conn_param->initiator_depth;
4279         req.flow_control = conn_param->flow_control;
4280         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4281         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4282         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4283         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4284         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4285         req.srq = id_priv->srq ? 1 : 0;
4286         req.ece.vendor_id = id_priv->ece.vendor_id;
4287         req.ece.attr_mod = id_priv->ece.attr_mod;
4288
4289         trace_cm_send_req(id_priv);
4290         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4291 out:
4292         if (ret && !IS_ERR(id)) {
4293                 ib_destroy_cm_id(id);
4294                 id_priv->cm_id.ib = NULL;
4295         }
4296
4297         kfree(private_data);
4298         return ret;
4299 }
4300
4301 static int cma_connect_iw(struct rdma_id_private *id_priv,
4302                           struct rdma_conn_param *conn_param)
4303 {
4304         struct iw_cm_id *cm_id;
4305         int ret;
4306         struct iw_cm_conn_param iw_param;
4307
4308         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4309         if (IS_ERR(cm_id))
4310                 return PTR_ERR(cm_id);
4311
4312         mutex_lock(&id_priv->qp_mutex);
4313         cm_id->tos = id_priv->tos;
4314         cm_id->tos_set = id_priv->tos_set;
4315         mutex_unlock(&id_priv->qp_mutex);
4316
4317         id_priv->cm_id.iw = cm_id;
4318
4319         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4320                rdma_addr_size(cma_src_addr(id_priv)));
4321         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4322                rdma_addr_size(cma_dst_addr(id_priv)));
4323
4324         ret = cma_modify_qp_rtr(id_priv, conn_param);
4325         if (ret)
4326                 goto out;
4327
4328         if (conn_param) {
4329                 iw_param.ord = conn_param->initiator_depth;
4330                 iw_param.ird = conn_param->responder_resources;
4331                 iw_param.private_data = conn_param->private_data;
4332                 iw_param.private_data_len = conn_param->private_data_len;
4333                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4334         } else {
4335                 memset(&iw_param, 0, sizeof iw_param);
4336                 iw_param.qpn = id_priv->qp_num;
4337         }
4338         ret = iw_cm_connect(cm_id, &iw_param);
4339 out:
4340         if (ret) {
4341                 iw_destroy_cm_id(cm_id);
4342                 id_priv->cm_id.iw = NULL;
4343         }
4344         return ret;
4345 }
4346
4347 /**
4348  * rdma_connect_locked - Initiate an active connection request.
4349  * @id: Connection identifier to connect.
4350  * @conn_param: Connection information used for connected QPs.
4351  *
4352  * Same as rdma_connect() but can only be called from the
4353  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4354  */
4355 int rdma_connect_locked(struct rdma_cm_id *id,
4356                         struct rdma_conn_param *conn_param)
4357 {
4358         struct rdma_id_private *id_priv =
4359                 container_of(id, struct rdma_id_private, id);
4360         int ret;
4361
4362         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4363                 return -EINVAL;
4364
4365         if (!id->qp) {
4366                 id_priv->qp_num = conn_param->qp_num;
4367                 id_priv->srq = conn_param->srq;
4368         }
4369
4370         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4371                 if (id->qp_type == IB_QPT_UD)
4372                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4373                 else
4374                         ret = cma_connect_ib(id_priv, conn_param);
4375         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4376                 ret = cma_connect_iw(id_priv, conn_param);
4377         } else {
4378                 ret = -ENOSYS;
4379         }
4380         if (ret)
4381                 goto err_state;
4382         return 0;
4383 err_state:
4384         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4385         return ret;
4386 }
4387 EXPORT_SYMBOL(rdma_connect_locked);
4388
4389 /**
4390  * rdma_connect - Initiate an active connection request.
4391  * @id: Connection identifier to connect.
4392  * @conn_param: Connection information used for connected QPs.
4393  *
4394  * Users must have resolved a route for the rdma_cm_id to connect with by having
4395  * called rdma_resolve_route before calling this routine.
4396  *
4397  * This call will either connect to a remote QP or obtain remote QP information
4398  * for unconnected rdma_cm_id's.  The actual operation is based on the
4399  * rdma_cm_id's port space.
4400  */
4401 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4402 {
4403         struct rdma_id_private *id_priv =
4404                 container_of(id, struct rdma_id_private, id);
4405         int ret;
4406
4407         mutex_lock(&id_priv->handler_mutex);
4408         ret = rdma_connect_locked(id, conn_param);
4409         mutex_unlock(&id_priv->handler_mutex);
4410         return ret;
4411 }
4412 EXPORT_SYMBOL(rdma_connect);
4413
4414 /**
4415  * rdma_connect_ece - Initiate an active connection request with ECE data.
4416  * @id: Connection identifier to connect.
4417  * @conn_param: Connection information used for connected QPs.
4418  * @ece: ECE parameters
4419  *
4420  * See rdma_connect() explanation.
4421  */
4422 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4423                      struct rdma_ucm_ece *ece)
4424 {
4425         struct rdma_id_private *id_priv =
4426                 container_of(id, struct rdma_id_private, id);
4427
4428         id_priv->ece.vendor_id = ece->vendor_id;
4429         id_priv->ece.attr_mod = ece->attr_mod;
4430
4431         return rdma_connect(id, conn_param);
4432 }
4433 EXPORT_SYMBOL(rdma_connect_ece);
4434
4435 static int cma_accept_ib(struct rdma_id_private *id_priv,
4436                          struct rdma_conn_param *conn_param)
4437 {
4438         struct ib_cm_rep_param rep;
4439         int ret;
4440
4441         ret = cma_modify_qp_rtr(id_priv, conn_param);
4442         if (ret)
4443                 goto out;
4444
4445         ret = cma_modify_qp_rts(id_priv, conn_param);
4446         if (ret)
4447                 goto out;
4448
4449         memset(&rep, 0, sizeof rep);
4450         rep.qp_num = id_priv->qp_num;
4451         rep.starting_psn = id_priv->seq_num;
4452         rep.private_data = conn_param->private_data;
4453         rep.private_data_len = conn_param->private_data_len;
4454         rep.responder_resources = conn_param->responder_resources;
4455         rep.initiator_depth = conn_param->initiator_depth;
4456         rep.failover_accepted = 0;
4457         rep.flow_control = conn_param->flow_control;
4458         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4459         rep.srq = id_priv->srq ? 1 : 0;
4460         rep.ece.vendor_id = id_priv->ece.vendor_id;
4461         rep.ece.attr_mod = id_priv->ece.attr_mod;
4462
4463         trace_cm_send_rep(id_priv);
4464         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4465 out:
4466         return ret;
4467 }
4468
4469 static int cma_accept_iw(struct rdma_id_private *id_priv,
4470                   struct rdma_conn_param *conn_param)
4471 {
4472         struct iw_cm_conn_param iw_param;
4473         int ret;
4474
4475         if (!conn_param)
4476                 return -EINVAL;
4477
4478         ret = cma_modify_qp_rtr(id_priv, conn_param);
4479         if (ret)
4480                 return ret;
4481
4482         iw_param.ord = conn_param->initiator_depth;
4483         iw_param.ird = conn_param->responder_resources;
4484         iw_param.private_data = conn_param->private_data;
4485         iw_param.private_data_len = conn_param->private_data_len;
4486         if (id_priv->id.qp)
4487                 iw_param.qpn = id_priv->qp_num;
4488         else
4489                 iw_param.qpn = conn_param->qp_num;
4490
4491         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4492 }
4493
4494 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4495                              enum ib_cm_sidr_status status, u32 qkey,
4496                              const void *private_data, int private_data_len)
4497 {
4498         struct ib_cm_sidr_rep_param rep;
4499         int ret;
4500
4501         memset(&rep, 0, sizeof rep);
4502         rep.status = status;
4503         if (status == IB_SIDR_SUCCESS) {
4504                 ret = cma_set_qkey(id_priv, qkey);
4505                 if (ret)
4506                         return ret;
4507                 rep.qp_num = id_priv->qp_num;
4508                 rep.qkey = id_priv->qkey;
4509
4510                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4511                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4512         }
4513
4514         rep.private_data = private_data;
4515         rep.private_data_len = private_data_len;
4516
4517         trace_cm_send_sidr_rep(id_priv);
4518         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4519 }
4520
4521 /**
4522  * rdma_accept - Called to accept a connection request or response.
4523  * @id: Connection identifier associated with the request.
4524  * @conn_param: Information needed to establish the connection.  This must be
4525  *   provided if accepting a connection request.  If accepting a connection
4526  *   response, this parameter must be NULL.
4527  *
4528  * Typically, this routine is only called by the listener to accept a connection
4529  * request.  It must also be called on the active side of a connection if the
4530  * user is performing their own QP transitions.
4531  *
4532  * In the case of error, a reject message is sent to the remote side and the
4533  * state of the qp associated with the id is modified to error, such that any
4534  * previously posted receive buffers would be flushed.
4535  *
4536  * This function is for use by kernel ULPs and must be called from under the
4537  * handler callback.
4538  */
4539 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4540 {
4541         struct rdma_id_private *id_priv =
4542                 container_of(id, struct rdma_id_private, id);
4543         int ret;
4544
4545         lockdep_assert_held(&id_priv->handler_mutex);
4546
4547         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4548                 return -EINVAL;
4549
4550         if (!id->qp && conn_param) {
4551                 id_priv->qp_num = conn_param->qp_num;
4552                 id_priv->srq = conn_param->srq;
4553         }
4554
4555         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4556                 if (id->qp_type == IB_QPT_UD) {
4557                         if (conn_param)
4558                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4559                                                         conn_param->qkey,
4560                                                         conn_param->private_data,
4561                                                         conn_param->private_data_len);
4562                         else
4563                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4564                                                         0, NULL, 0);
4565                 } else {
4566                         if (conn_param)
4567                                 ret = cma_accept_ib(id_priv, conn_param);
4568                         else
4569                                 ret = cma_rep_recv(id_priv);
4570                 }
4571         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4572                 ret = cma_accept_iw(id_priv, conn_param);
4573         } else {
4574                 ret = -ENOSYS;
4575         }
4576         if (ret)
4577                 goto reject;
4578
4579         return 0;
4580 reject:
4581         cma_modify_qp_err(id_priv);
4582         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4583         return ret;
4584 }
4585 EXPORT_SYMBOL(rdma_accept);
4586
4587 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4588                     struct rdma_ucm_ece *ece)
4589 {
4590         struct rdma_id_private *id_priv =
4591                 container_of(id, struct rdma_id_private, id);
4592
4593         id_priv->ece.vendor_id = ece->vendor_id;
4594         id_priv->ece.attr_mod = ece->attr_mod;
4595
4596         return rdma_accept(id, conn_param);
4597 }
4598 EXPORT_SYMBOL(rdma_accept_ece);
4599
4600 void rdma_lock_handler(struct rdma_cm_id *id)
4601 {
4602         struct rdma_id_private *id_priv =
4603                 container_of(id, struct rdma_id_private, id);
4604
4605         mutex_lock(&id_priv->handler_mutex);
4606 }
4607 EXPORT_SYMBOL(rdma_lock_handler);
4608
4609 void rdma_unlock_handler(struct rdma_cm_id *id)
4610 {
4611         struct rdma_id_private *id_priv =
4612                 container_of(id, struct rdma_id_private, id);
4613
4614         mutex_unlock(&id_priv->handler_mutex);
4615 }
4616 EXPORT_SYMBOL(rdma_unlock_handler);
4617
4618 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4619 {
4620         struct rdma_id_private *id_priv;
4621         int ret;
4622
4623         id_priv = container_of(id, struct rdma_id_private, id);
4624         if (!id_priv->cm_id.ib)
4625                 return -EINVAL;
4626
4627         switch (id->device->node_type) {
4628         case RDMA_NODE_IB_CA:
4629                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4630                 break;
4631         default:
4632                 ret = 0;
4633                 break;
4634         }
4635         return ret;
4636 }
4637 EXPORT_SYMBOL(rdma_notify);
4638
4639 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4640                 u8 private_data_len, u8 reason)
4641 {
4642         struct rdma_id_private *id_priv;
4643         int ret;
4644
4645         id_priv = container_of(id, struct rdma_id_private, id);
4646         if (!id_priv->cm_id.ib)
4647                 return -EINVAL;
4648
4649         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4650                 if (id->qp_type == IB_QPT_UD) {
4651                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4652                                                 private_data, private_data_len);
4653                 } else {
4654                         trace_cm_send_rej(id_priv);
4655                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4656                                              private_data, private_data_len);
4657                 }
4658         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4659                 ret = iw_cm_reject(id_priv->cm_id.iw,
4660                                    private_data, private_data_len);
4661         } else {
4662                 ret = -ENOSYS;
4663         }
4664
4665         return ret;
4666 }
4667 EXPORT_SYMBOL(rdma_reject);
4668
4669 int rdma_disconnect(struct rdma_cm_id *id)
4670 {
4671         struct rdma_id_private *id_priv;
4672         int ret;
4673
4674         id_priv = container_of(id, struct rdma_id_private, id);
4675         if (!id_priv->cm_id.ib)
4676                 return -EINVAL;
4677
4678         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4679                 ret = cma_modify_qp_err(id_priv);
4680                 if (ret)
4681                         goto out;
4682                 /* Initiate or respond to a disconnect. */
4683                 trace_cm_disconnect(id_priv);
4684                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4685                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4686                                 trace_cm_sent_drep(id_priv);
4687                 } else {
4688                         trace_cm_sent_dreq(id_priv);
4689                 }
4690         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4691                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4692         } else
4693                 ret = -EINVAL;
4694
4695 out:
4696         return ret;
4697 }
4698 EXPORT_SYMBOL(rdma_disconnect);
4699
4700 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4701                               struct ib_sa_multicast *multicast,
4702                               struct rdma_cm_event *event,
4703                               struct cma_multicast *mc)
4704 {
4705         struct rdma_dev_addr *dev_addr;
4706         enum ib_gid_type gid_type;
4707         struct net_device *ndev;
4708
4709         if (!status)
4710                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4711         else
4712                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4713                                      status);
4714
4715         event->status = status;
4716         event->param.ud.private_data = mc->context;
4717         if (status) {
4718                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4719                 return;
4720         }
4721
4722         dev_addr = &id_priv->id.route.addr.dev_addr;
4723         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4724         gid_type =
4725                 id_priv->cma_dev
4726                         ->default_gid_type[id_priv->id.port_num -
4727                                            rdma_start_port(
4728                                                    id_priv->cma_dev->device)];
4729
4730         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4731         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4732                                      &multicast->rec, ndev, gid_type,
4733                                      &event->param.ud.ah_attr)) {
4734                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4735                 goto out;
4736         }
4737
4738         event->param.ud.qp_num = 0xFFFFFF;
4739         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4740
4741 out:
4742         if (ndev)
4743                 dev_put(ndev);
4744 }
4745
4746 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4747 {
4748         struct cma_multicast *mc = multicast->context;
4749         struct rdma_id_private *id_priv = mc->id_priv;
4750         struct rdma_cm_event event = {};
4751         int ret = 0;
4752
4753         mutex_lock(&id_priv->handler_mutex);
4754         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4755             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4756                 goto out;
4757
4758         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4759         ret = cma_cm_event_handler(id_priv, &event);
4760         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4761         WARN_ON(ret);
4762
4763 out:
4764         mutex_unlock(&id_priv->handler_mutex);
4765         return 0;
4766 }
4767
4768 static void cma_set_mgid(struct rdma_id_private *id_priv,
4769                          struct sockaddr *addr, union ib_gid *mgid)
4770 {
4771         unsigned char mc_map[MAX_ADDR_LEN];
4772         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4773         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4774         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4775
4776         if (cma_any_addr(addr)) {
4777                 memset(mgid, 0, sizeof *mgid);
4778         } else if ((addr->sa_family == AF_INET6) &&
4779                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4780                                                                  0xFF10A01B)) {
4781                 /* IPv6 address is an SA assigned MGID. */
4782                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4783         } else if (addr->sa_family == AF_IB) {
4784                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4785         } else if (addr->sa_family == AF_INET6) {
4786                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4787                 if (id_priv->id.ps == RDMA_PS_UDP)
4788                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4789                 *mgid = *(union ib_gid *) (mc_map + 4);
4790         } else {
4791                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4792                 if (id_priv->id.ps == RDMA_PS_UDP)
4793                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4794                 *mgid = *(union ib_gid *) (mc_map + 4);
4795         }
4796 }
4797
4798 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4799                                  struct cma_multicast *mc)
4800 {
4801         struct ib_sa_mcmember_rec rec;
4802         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4803         ib_sa_comp_mask comp_mask;
4804         int ret;
4805
4806         ib_addr_get_mgid(dev_addr, &rec.mgid);
4807         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4808                                      &rec.mgid, &rec);
4809         if (ret)
4810                 return ret;
4811
4812         ret = cma_set_qkey(id_priv, 0);
4813         if (ret)
4814                 return ret;
4815
4816         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4817         rec.qkey = cpu_to_be32(id_priv->qkey);
4818         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4819         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4820         rec.join_state = mc->join_state;
4821
4822         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4823                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4824                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4825                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4826                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4827
4828         if (id_priv->id.ps == RDMA_PS_IPOIB)
4829                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4830                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4831                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4832                              IB_SA_MCMEMBER_REC_MTU |
4833                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4834
4835         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4836                                          id_priv->id.port_num, &rec, comp_mask,
4837                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4838         return PTR_ERR_OR_ZERO(mc->sa_mc);
4839 }
4840
4841 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4842                               enum ib_gid_type gid_type)
4843 {
4844         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4845         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4846
4847         if (cma_any_addr(addr)) {
4848                 memset(mgid, 0, sizeof *mgid);
4849         } else if (addr->sa_family == AF_INET6) {
4850                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4851         } else {
4852                 mgid->raw[0] =
4853                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4854                 mgid->raw[1] =
4855                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4856                 mgid->raw[2] = 0;
4857                 mgid->raw[3] = 0;
4858                 mgid->raw[4] = 0;
4859                 mgid->raw[5] = 0;
4860                 mgid->raw[6] = 0;
4861                 mgid->raw[7] = 0;
4862                 mgid->raw[8] = 0;
4863                 mgid->raw[9] = 0;
4864                 mgid->raw[10] = 0xff;
4865                 mgid->raw[11] = 0xff;
4866                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4867         }
4868 }
4869
4870 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4871                                    struct cma_multicast *mc)
4872 {
4873         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4874         int err = 0;
4875         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4876         struct net_device *ndev = NULL;
4877         struct ib_sa_multicast ib;
4878         enum ib_gid_type gid_type;
4879         bool send_only;
4880
4881         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4882
4883         if (cma_zero_addr(addr))
4884                 return -EINVAL;
4885
4886         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4887                    rdma_start_port(id_priv->cma_dev->device)];
4888         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4889
4890         ib.rec.pkey = cpu_to_be16(0xffff);
4891         if (id_priv->id.ps == RDMA_PS_UDP)
4892                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4893
4894         if (dev_addr->bound_dev_if)
4895                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4896         if (!ndev)
4897                 return -ENODEV;
4898
4899         ib.rec.rate = iboe_get_rate(ndev);
4900         ib.rec.hop_limit = 1;
4901         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4902
4903         if (addr->sa_family == AF_INET) {
4904                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4905                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4906                         if (!send_only) {
4907                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4908                                                     true);
4909                         }
4910                 }
4911         } else {
4912                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4913                         err = -ENOTSUPP;
4914         }
4915         dev_put(ndev);
4916         if (err || !ib.rec.mtu)
4917                 return err ?: -EINVAL;
4918
4919         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4920                     &ib.rec.port_gid);
4921         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4922         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4923         queue_work(cma_wq, &mc->iboe_join.work);
4924         return 0;
4925 }
4926
4927 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4928                         u8 join_state, void *context)
4929 {
4930         struct rdma_id_private *id_priv =
4931                 container_of(id, struct rdma_id_private, id);
4932         struct cma_multicast *mc;
4933         int ret;
4934
4935         /* Not supported for kernel QPs */
4936         if (WARN_ON(id->qp))
4937                 return -EINVAL;
4938
4939         /* ULP is calling this wrong. */
4940         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4941                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4942                 return -EINVAL;
4943
4944         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4945         if (!mc)
4946                 return -ENOMEM;
4947
4948         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4949         mc->context = context;
4950         mc->id_priv = id_priv;
4951         mc->join_state = join_state;
4952
4953         if (rdma_protocol_roce(id->device, id->port_num)) {
4954                 ret = cma_iboe_join_multicast(id_priv, mc);
4955                 if (ret)
4956                         goto out_err;
4957         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4958                 ret = cma_join_ib_multicast(id_priv, mc);
4959                 if (ret)
4960                         goto out_err;
4961         } else {
4962                 ret = -ENOSYS;
4963                 goto out_err;
4964         }
4965
4966         spin_lock(&id_priv->lock);
4967         list_add(&mc->list, &id_priv->mc_list);
4968         spin_unlock(&id_priv->lock);
4969
4970         return 0;
4971 out_err:
4972         kfree(mc);
4973         return ret;
4974 }
4975 EXPORT_SYMBOL(rdma_join_multicast);
4976
4977 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4978 {
4979         struct rdma_id_private *id_priv;
4980         struct cma_multicast *mc;
4981
4982         id_priv = container_of(id, struct rdma_id_private, id);
4983         spin_lock_irq(&id_priv->lock);
4984         list_for_each_entry(mc, &id_priv->mc_list, list) {
4985                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4986                         continue;
4987                 list_del(&mc->list);
4988                 spin_unlock_irq(&id_priv->lock);
4989
4990                 WARN_ON(id_priv->cma_dev->device != id->device);
4991                 destroy_mc(id_priv, mc);
4992                 return;
4993         }
4994         spin_unlock_irq(&id_priv->lock);
4995 }
4996 EXPORT_SYMBOL(rdma_leave_multicast);
4997
4998 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4999 {
5000         struct rdma_dev_addr *dev_addr;
5001         struct cma_work *work;
5002
5003         dev_addr = &id_priv->id.route.addr.dev_addr;
5004
5005         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
5006             (net_eq(dev_net(ndev), dev_addr->net)) &&
5007             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
5008                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
5009                         ndev->name, &id_priv->id);
5010                 work = kzalloc(sizeof *work, GFP_KERNEL);
5011                 if (!work)
5012                         return -ENOMEM;
5013
5014                 INIT_WORK(&work->work, cma_work_handler);
5015                 work->id = id_priv;
5016                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
5017                 cma_id_get(id_priv);
5018                 queue_work(cma_wq, &work->work);
5019         }
5020
5021         return 0;
5022 }
5023
5024 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
5025                                void *ptr)
5026 {
5027         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
5028         struct cma_device *cma_dev;
5029         struct rdma_id_private *id_priv;
5030         int ret = NOTIFY_DONE;
5031
5032         if (event != NETDEV_BONDING_FAILOVER)
5033                 return NOTIFY_DONE;
5034
5035         if (!netif_is_bond_master(ndev))
5036                 return NOTIFY_DONE;
5037
5038         mutex_lock(&lock);
5039         list_for_each_entry(cma_dev, &dev_list, list)
5040                 list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
5041                         ret = cma_netdev_change(ndev, id_priv);
5042                         if (ret)
5043                                 goto out;
5044                 }
5045
5046 out:
5047         mutex_unlock(&lock);
5048         return ret;
5049 }
5050
5051 static void cma_netevent_work_handler(struct work_struct *_work)
5052 {
5053         struct rdma_id_private *id_priv =
5054                 container_of(_work, struct rdma_id_private, id.net_work);
5055         struct rdma_cm_event event = {};
5056
5057         mutex_lock(&id_priv->handler_mutex);
5058
5059         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
5060             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
5061                 goto out_unlock;
5062
5063         event.event = RDMA_CM_EVENT_UNREACHABLE;
5064         event.status = -ETIMEDOUT;
5065
5066         if (cma_cm_event_handler(id_priv, &event)) {
5067                 __acquire(&id_priv->handler_mutex);
5068                 id_priv->cm_id.ib = NULL;
5069                 cma_id_put(id_priv);
5070                 destroy_id_handler_unlock(id_priv);
5071                 return;
5072         }
5073
5074 out_unlock:
5075         mutex_unlock(&id_priv->handler_mutex);
5076         cma_id_put(id_priv);
5077 }
5078
5079 static int cma_netevent_callback(struct notifier_block *self,
5080                                  unsigned long event, void *ctx)
5081 {
5082         struct id_table_entry *ips_node = NULL;
5083         struct rdma_id_private *current_id;
5084         struct neighbour *neigh = ctx;
5085         unsigned long flags;
5086
5087         if (event != NETEVENT_NEIGH_UPDATE)
5088                 return NOTIFY_DONE;
5089
5090         spin_lock_irqsave(&id_table_lock, flags);
5091         if (neigh->tbl->family == AF_INET6) {
5092                 struct sockaddr_in6 neigh_sock_6;
5093
5094                 neigh_sock_6.sin6_family = AF_INET6;
5095                 neigh_sock_6.sin6_addr = *(struct in6_addr *)neigh->primary_key;
5096                 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5097                                              (struct sockaddr *)&neigh_sock_6);
5098         } else if (neigh->tbl->family == AF_INET) {
5099                 struct sockaddr_in neigh_sock_4;
5100
5101                 neigh_sock_4.sin_family = AF_INET;
5102                 neigh_sock_4.sin_addr.s_addr = *(__be32 *)(neigh->primary_key);
5103                 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5104                                              (struct sockaddr *)&neigh_sock_4);
5105         } else
5106                 goto out;
5107
5108         if (!ips_node)
5109                 goto out;
5110
5111         list_for_each_entry(current_id, &ips_node->id_list, id_list_entry) {
5112                 if (!memcmp(current_id->id.route.addr.dev_addr.dst_dev_addr,
5113                            neigh->ha, ETH_ALEN))
5114                         continue;
5115                 INIT_WORK(&current_id->id.net_work, cma_netevent_work_handler);
5116                 cma_id_get(current_id);
5117                 queue_work(cma_wq, &current_id->id.net_work);
5118         }
5119 out:
5120         spin_unlock_irqrestore(&id_table_lock, flags);
5121         return NOTIFY_DONE;
5122 }
5123
5124 static struct notifier_block cma_nb = {
5125         .notifier_call = cma_netdev_callback
5126 };
5127
5128 static struct notifier_block cma_netevent_cb = {
5129         .notifier_call = cma_netevent_callback
5130 };
5131
5132 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
5133 {
5134         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
5135         enum rdma_cm_state state;
5136         unsigned long flags;
5137
5138         mutex_lock(&id_priv->handler_mutex);
5139         /* Record that we want to remove the device */
5140         spin_lock_irqsave(&id_priv->lock, flags);
5141         state = id_priv->state;
5142         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
5143                 spin_unlock_irqrestore(&id_priv->lock, flags);
5144                 mutex_unlock(&id_priv->handler_mutex);
5145                 cma_id_put(id_priv);
5146                 return;
5147         }
5148         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
5149         spin_unlock_irqrestore(&id_priv->lock, flags);
5150
5151         if (cma_cm_event_handler(id_priv, &event)) {
5152                 /*
5153                  * At this point the ULP promises it won't call
5154                  * rdma_destroy_id() concurrently
5155                  */
5156                 cma_id_put(id_priv);
5157                 mutex_unlock(&id_priv->handler_mutex);
5158                 trace_cm_id_destroy(id_priv);
5159                 _destroy_id(id_priv, state);
5160                 return;
5161         }
5162         mutex_unlock(&id_priv->handler_mutex);
5163
5164         /*
5165          * If this races with destroy then the thread that first assigns state
5166          * to a destroying does the cancel.
5167          */
5168         cma_cancel_operation(id_priv, state);
5169         cma_id_put(id_priv);
5170 }
5171
5172 static void cma_process_remove(struct cma_device *cma_dev)
5173 {
5174         mutex_lock(&lock);
5175         while (!list_empty(&cma_dev->id_list)) {
5176                 struct rdma_id_private *id_priv = list_first_entry(
5177                         &cma_dev->id_list, struct rdma_id_private, device_item);
5178
5179                 list_del_init(&id_priv->listen_item);
5180                 list_del_init(&id_priv->device_item);
5181                 cma_id_get(id_priv);
5182                 mutex_unlock(&lock);
5183
5184                 cma_send_device_removal_put(id_priv);
5185
5186                 mutex_lock(&lock);
5187         }
5188         mutex_unlock(&lock);
5189
5190         cma_dev_put(cma_dev);
5191         wait_for_completion(&cma_dev->comp);
5192 }
5193
5194 static bool cma_supported(struct ib_device *device)
5195 {
5196         u32 i;
5197
5198         rdma_for_each_port(device, i) {
5199                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
5200                         return true;
5201         }
5202         return false;
5203 }
5204
5205 static int cma_add_one(struct ib_device *device)
5206 {
5207         struct rdma_id_private *to_destroy;
5208         struct cma_device *cma_dev;
5209         struct rdma_id_private *id_priv;
5210         unsigned long supported_gids = 0;
5211         int ret;
5212         u32 i;
5213
5214         if (!cma_supported(device))
5215                 return -EOPNOTSUPP;
5216
5217         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5218         if (!cma_dev)
5219                 return -ENOMEM;
5220
5221         cma_dev->device = device;
5222         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5223                                             sizeof(*cma_dev->default_gid_type),
5224                                             GFP_KERNEL);
5225         if (!cma_dev->default_gid_type) {
5226                 ret = -ENOMEM;
5227                 goto free_cma_dev;
5228         }
5229
5230         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5231                                             sizeof(*cma_dev->default_roce_tos),
5232                                             GFP_KERNEL);
5233         if (!cma_dev->default_roce_tos) {
5234                 ret = -ENOMEM;
5235                 goto free_gid_type;
5236         }
5237
5238         rdma_for_each_port (device, i) {
5239                 supported_gids = roce_gid_type_mask_support(device, i);
5240                 WARN_ON(!supported_gids);
5241                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5242                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5243                                 CMA_PREFERRED_ROCE_GID_TYPE;
5244                 else
5245                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5246                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5247                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5248         }
5249
5250         init_completion(&cma_dev->comp);
5251         refcount_set(&cma_dev->refcount, 1);
5252         INIT_LIST_HEAD(&cma_dev->id_list);
5253         ib_set_client_data(device, &cma_client, cma_dev);
5254
5255         mutex_lock(&lock);
5256         list_add_tail(&cma_dev->list, &dev_list);
5257         list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5258                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5259                 if (ret)
5260                         goto free_listen;
5261         }
5262         mutex_unlock(&lock);
5263
5264         trace_cm_add_one(device);
5265         return 0;
5266
5267 free_listen:
5268         list_del(&cma_dev->list);
5269         mutex_unlock(&lock);
5270
5271         /* cma_process_remove() will delete to_destroy */
5272         cma_process_remove(cma_dev);
5273         kfree(cma_dev->default_roce_tos);
5274 free_gid_type:
5275         kfree(cma_dev->default_gid_type);
5276
5277 free_cma_dev:
5278         kfree(cma_dev);
5279         return ret;
5280 }
5281
5282 static void cma_remove_one(struct ib_device *device, void *client_data)
5283 {
5284         struct cma_device *cma_dev = client_data;
5285
5286         trace_cm_remove_one(device);
5287
5288         mutex_lock(&lock);
5289         list_del(&cma_dev->list);
5290         mutex_unlock(&lock);
5291
5292         cma_process_remove(cma_dev);
5293         kfree(cma_dev->default_roce_tos);
5294         kfree(cma_dev->default_gid_type);
5295         kfree(cma_dev);
5296 }
5297
5298 static int cma_init_net(struct net *net)
5299 {
5300         struct cma_pernet *pernet = cma_pernet(net);
5301
5302         xa_init(&pernet->tcp_ps);
5303         xa_init(&pernet->udp_ps);
5304         xa_init(&pernet->ipoib_ps);
5305         xa_init(&pernet->ib_ps);
5306
5307         return 0;
5308 }
5309
5310 static void cma_exit_net(struct net *net)
5311 {
5312         struct cma_pernet *pernet = cma_pernet(net);
5313
5314         WARN_ON(!xa_empty(&pernet->tcp_ps));
5315         WARN_ON(!xa_empty(&pernet->udp_ps));
5316         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5317         WARN_ON(!xa_empty(&pernet->ib_ps));
5318 }
5319
5320 static struct pernet_operations cma_pernet_operations = {
5321         .init = cma_init_net,
5322         .exit = cma_exit_net,
5323         .id = &cma_pernet_id,
5324         .size = sizeof(struct cma_pernet),
5325 };
5326
5327 static int __init cma_init(void)
5328 {
5329         int ret;
5330
5331         /*
5332          * There is a rare lock ordering dependency in cma_netdev_callback()
5333          * that only happens when bonding is enabled. Teach lockdep that rtnl
5334          * must never be nested under lock so it can find these without having
5335          * to test with bonding.
5336          */
5337         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5338                 rtnl_lock();
5339                 mutex_lock(&lock);
5340                 mutex_unlock(&lock);
5341                 rtnl_unlock();
5342         }
5343
5344         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5345         if (!cma_wq)
5346                 return -ENOMEM;
5347
5348         ret = register_pernet_subsys(&cma_pernet_operations);
5349         if (ret)
5350                 goto err_wq;
5351
5352         ib_sa_register_client(&sa_client);
5353         register_netdevice_notifier(&cma_nb);
5354         register_netevent_notifier(&cma_netevent_cb);
5355
5356         ret = ib_register_client(&cma_client);
5357         if (ret)
5358                 goto err;
5359
5360         ret = cma_configfs_init();
5361         if (ret)
5362                 goto err_ib;
5363
5364         return 0;
5365
5366 err_ib:
5367         ib_unregister_client(&cma_client);
5368 err:
5369         unregister_netevent_notifier(&cma_netevent_cb);
5370         unregister_netdevice_notifier(&cma_nb);
5371         ib_sa_unregister_client(&sa_client);
5372         unregister_pernet_subsys(&cma_pernet_operations);
5373 err_wq:
5374         destroy_workqueue(cma_wq);
5375         return ret;
5376 }
5377
5378 static void __exit cma_cleanup(void)
5379 {
5380         cma_configfs_exit();
5381         ib_unregister_client(&cma_client);
5382         unregister_netevent_notifier(&cma_netevent_cb);
5383         unregister_netdevice_notifier(&cma_nb);
5384         ib_sa_unregister_client(&sa_client);
5385         unregister_pernet_subsys(&cma_pernet_operations);
5386         destroy_workqueue(cma_wq);
5387 }
5388
5389 module_init(cma_init);
5390 module_exit(cma_cleanup);