drivers/infiniband/core/cma.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   4  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   5  * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
   6  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   7  */
   8
   9 #include <linux/completion.h>
  10 #include <linux/in.h>
  11 #include <linux/in6.h>
  12 #include <linux/mutex.h>
  13 #include <linux/random.h>
  14 #include <linux/igmp.h>
  15 #include <linux/xarray.h>
  16 #include <linux/inetdevice.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <net/route.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/netns/generic.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/ip_fib.h>
  26 #include <net/ip6_route.h>
  27
  28 #include <rdma/rdma_cm.h>
  29 #include <rdma/rdma_cm_ib.h>
  30 #include <rdma/rdma_netlink.h>
  31 #include <rdma/ib.h>
  32 #include <rdma/ib_cache.h>
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_sa.h>
  35 #include <rdma/iw_cm.h>
  36
  37 #include "core_priv.h"
  38 #include "cma_priv.h"
  39 #include "cma_trace.h"
  40
  41 MODULE_AUTHOR("Sean Hefty");
  42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  43 MODULE_LICENSE("Dual BSD/GPL");
  44
  45 #define CMA_CM_RESPONSE_TIMEOUT 20
  46 #define CMA_MAX_CM_RETRIES 15
  47 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  48 #define CMA_IBOE_PACKET_LIFETIME 18
  49 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  50
  51 static const char * const cma_events[] = {
  52         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  53         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  54         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  55         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  56         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  57         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  58         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  59         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  60         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  61         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  62         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  63         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  64         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  65         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  66         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  67         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  68 };
  69
  70 static void cma_set_mgid(struct rdma_id_private *id_priv, struct sockaddr *addr,
  71                          union ib_gid *mgid);
  72
  73 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  74 {
  75         size_t index = event;
  76
  77         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  78                         cma_events[index] : "unrecognized event";
  79 }
  80 EXPORT_SYMBOL(rdma_event_msg);
  81
  82 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  83                                                 int reason)
  84 {
  85         if (rdma_ib_or_roce(id->device, id->port_num))
  86                 return ibcm_reject_msg(reason);
  87
  88         if (rdma_protocol_iwarp(id->device, id->port_num))
  89                 return iwcm_reject_msg(reason);
  90
  91         WARN_ON_ONCE(1);
  92         return "unrecognized transport";
  93 }
  94 EXPORT_SYMBOL(rdma_reject_msg);
  95
  96 /**
  97  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  98  *                           request.
  99  * @id: Communication identifier that received the REJECT event.
 100  * @reason: Value returned in the REJECT event status field.
 101  */
 102 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 103 {
 104         if (rdma_ib_or_roce(id->device, id->port_num))
 105                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 106
 107         if (rdma_protocol_iwarp(id->device, id->port_num))
 108                 return reason == -ECONNREFUSED;
 109
 110         WARN_ON_ONCE(1);
 111         return false;
 112 }
 113
 114 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 115                                       struct rdma_cm_event *ev, u8 *data_len)
 116 {
 117         const void *p;
 118
 119         if (rdma_is_consumer_reject(id, ev->status)) {
 120                 *data_len = ev->param.conn.private_data_len;
 121                 p = ev->param.conn.private_data;
 122         } else {
 123                 *data_len = 0;
 124                 p = NULL;
 125         }
 126         return p;
 127 }
 128 EXPORT_SYMBOL(rdma_consumer_reject_data);
 129
 130 /**
 131  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 132  * @id: Communication Identifier
 133  */
 134 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 135 {
 136         struct rdma_id_private *id_priv;
 137
 138         id_priv = container_of(id, struct rdma_id_private, id);
 139         if (id->device->node_type == RDMA_NODE_RNIC)
 140                 return id_priv->cm_id.iw;
 141         return NULL;
 142 }
 143 EXPORT_SYMBOL(rdma_iw_cm_id);
 144
 145 /**
 146  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 147  * @res: rdma resource tracking entry pointer
 148  */
 149 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 150 {
 151         struct rdma_id_private *id_priv =
 152                 container_of(res, struct rdma_id_private, res);
 153
 154         return &id_priv->id;
 155 }
 156 EXPORT_SYMBOL(rdma_res_to_id);
 157
 158 static int cma_add_one(struct ib_device *device);
 159 static void cma_remove_one(struct ib_device *device, void *client_data);
 160
 161 static struct ib_client cma_client = {
 162         .name   = "cma",
 163         .add    = cma_add_one,
 164         .remove = cma_remove_one
 165 };
 166
 167 static struct ib_sa_client sa_client;
 168 static LIST_HEAD(dev_list);
 169 static LIST_HEAD(listen_any_list);
 170 static DEFINE_MUTEX(lock);
 171 static struct workqueue_struct *cma_wq;
 172 static unsigned int cma_pernet_id;
 173
 174 struct cma_pernet {
 175         struct xarray tcp_ps;
 176         struct xarray udp_ps;
 177         struct xarray ipoib_ps;
 178         struct xarray ib_ps;
 179 };
 180
 181 static struct cma_pernet *cma_pernet(struct net *net)
 182 {
 183         return net_generic(net, cma_pernet_id);
 184 }
 185
 186 static
 187 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 188 {
 189         struct cma_pernet *pernet = cma_pernet(net);
 190
 191         switch (ps) {
 192         case RDMA_PS_TCP:
 193                 return &pernet->tcp_ps;
 194         case RDMA_PS_UDP:
 195                 return &pernet->udp_ps;
 196         case RDMA_PS_IPOIB:
 197                 return &pernet->ipoib_ps;
 198         case RDMA_PS_IB:
 199                 return &pernet->ib_ps;
 200         default:
 201                 return NULL;
 202         }
 203 }
 204
 205 struct cma_device {
 206         struct list_head        list;
 207         struct ib_device        *device;
 208         struct completion       comp;
 209         refcount_t refcount;
 210         struct list_head        id_list;
 211         enum ib_gid_type        *default_gid_type;
 212         u8                      *default_roce_tos;
 213 };
 214
 215 struct rdma_bind_list {
 216         enum rdma_ucm_port_space ps;
 217         struct hlist_head       owners;
 218         unsigned short          port;
 219 };
 220
 221 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 222                         struct rdma_bind_list *bind_list, int snum)
 223 {
 224         struct xarray *xa = cma_pernet_xa(net, ps);
 225
 226         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 227 }
 228
 229 static struct rdma_bind_list *cma_ps_find(struct net *net,
 230                                           enum rdma_ucm_port_space ps, int snum)
 231 {
 232         struct xarray *xa = cma_pernet_xa(net, ps);
 233
 234         return xa_load(xa, snum);
 235 }
 236
 237 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 238                           int snum)
 239 {
 240         struct xarray *xa = cma_pernet_xa(net, ps);
 241
 242         xa_erase(xa, snum);
 243 }
 244
 245 enum {
 246         CMA_OPTION_AFONLY,
 247 };
 248
 249 void cma_dev_get(struct cma_device *cma_dev)
 250 {
 251         refcount_inc(&cma_dev->refcount);
 252 }
 253
 254 void cma_dev_put(struct cma_device *cma_dev)
 255 {
 256         if (refcount_dec_and_test(&cma_dev->refcount))
 257                 complete(&cma_dev->comp);
 258 }
 259
 260 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 261                                              void               *cookie)
 262 {
 263         struct cma_device *cma_dev;
 264         struct cma_device *found_cma_dev = NULL;
 265
 266         mutex_lock(&lock);
 267
 268         list_for_each_entry(cma_dev, &dev_list, list)
 269                 if (filter(cma_dev->device, cookie)) {
 270                         found_cma_dev = cma_dev;
 271                         break;
 272                 }
 273
 274         if (found_cma_dev)
 275                 cma_dev_get(found_cma_dev);
 276         mutex_unlock(&lock);
 277         return found_cma_dev;
 278 }
 279
 280 int cma_get_default_gid_type(struct cma_device *cma_dev,
 281                              u32 port)
 282 {
 283         if (!rdma_is_port_valid(cma_dev->device, port))
 284                 return -EINVAL;
 285
 286         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 287 }
 288
 289 int cma_set_default_gid_type(struct cma_device *cma_dev,
 290                              u32 port,
 291                              enum ib_gid_type default_gid_type)
 292 {
 293         unsigned long supported_gids;
 294
 295         if (!rdma_is_port_valid(cma_dev->device, port))
 296                 return -EINVAL;
 297
 298         if (default_gid_type == IB_GID_TYPE_IB &&
 299             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 300                 default_gid_type = IB_GID_TYPE_ROCE;
 301
 302         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 303
 304         if (!(supported_gids & 1 << default_gid_type))
 305                 return -EINVAL;
 306
 307         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 308                 default_gid_type;
 309
 310         return 0;
 311 }
 312
 313 int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
 314 {
 315         if (!rdma_is_port_valid(cma_dev->device, port))
 316                 return -EINVAL;
 317
 318         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 319 }
 320
 321 int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
 322                              u8 default_roce_tos)
 323 {
 324         if (!rdma_is_port_valid(cma_dev->device, port))
 325                 return -EINVAL;
 326
 327         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 328                  default_roce_tos;
 329
 330         return 0;
 331 }
 332 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 333 {
 334         return cma_dev->device;
 335 }
 336
 337 /*
 338  * Device removal can occur at anytime, so we need extra handling to
 339  * serialize notifying the user of device removal with other callbacks.
 340  * We do this by disabling removal notification while a callback is in process,
 341  * and reporting it after the callback completes.
 342  */
 343
 344 struct cma_multicast {
 345         struct rdma_id_private *id_priv;
 346         union {
 347                 struct ib_sa_multicast *sa_mc;
 348                 struct {
 349                         struct work_struct work;
 350                         struct rdma_cm_event event;
 351                 } iboe_join;
 352         };
 353         struct list_head        list;
 354         void                    *context;
 355         struct sockaddr_storage addr;
 356         u8                      join_state;
 357 };
 358
 359 struct cma_work {
 360         struct work_struct      work;
 361         struct rdma_id_private  *id;
 362         enum rdma_cm_state      old_state;
 363         enum rdma_cm_state      new_state;
 364         struct rdma_cm_event    event;
 365 };
 366
 367 union cma_ip_addr {
 368         struct in6_addr ip6;
 369         struct {
 370                 __be32 pad[3];
 371                 __be32 addr;
 372         } ip4;
 373 };
 374
 375 struct cma_hdr {
 376         u8 cma_version;
 377         u8 ip_version;  /* IP version: 7:4 */
 378         __be16 port;
 379         union cma_ip_addr src_addr;
 380         union cma_ip_addr dst_addr;
 381 };
 382
 383 #define CMA_VERSION 0x00
 384
 385 struct cma_req_info {
 386         struct sockaddr_storage listen_addr_storage;
 387         struct sockaddr_storage src_addr_storage;
 388         struct ib_device *device;
 389         union ib_gid local_gid;
 390         __be64 service_id;
 391         int port;
 392         bool has_gid;
 393         u16 pkey;
 394 };
 395
 396 static int cma_comp_exch(struct rdma_id_private *id_priv,
 397                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 398 {
 399         unsigned long flags;
 400         int ret;
 401
 402         /*
 403          * The FSM uses a funny double locking where state is protected by both
 404          * the handler_mutex and the spinlock. State is not allowed to change
 405          * to/from a handler_mutex protected value without also holding
 406          * handler_mutex.
 407          */
 408         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 409                 lockdep_assert_held(&id_priv->handler_mutex);
 410
 411         spin_lock_irqsave(&id_priv->lock, flags);
 412         if ((ret = (id_priv->state == comp)))
 413                 id_priv->state = exch;
 414         spin_unlock_irqrestore(&id_priv->lock, flags);
 415         return ret;
 416 }
 417
 418 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 419 {
 420         return hdr->ip_version >> 4;
 421 }
 422
 423 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 424 {
 425         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 426 }
 427
 428 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 429 {
 430         struct in_device *in_dev = NULL;
 431
 432         if (ndev) {
 433                 rtnl_lock();
 434                 in_dev = __in_dev_get_rtnl(ndev);
 435                 if (in_dev) {
 436                         if (join)
 437                                 ip_mc_inc_group(in_dev,
 438                                                 *(__be32 *)(mgid->raw + 12));
 439                         else
 440                                 ip_mc_dec_group(in_dev,
 441                                                 *(__be32 *)(mgid->raw + 12));
 442                 }
 443                 rtnl_unlock();
 444         }
 445         return (in_dev) ? 0 : -ENODEV;
 446 }
 447
 448 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 449                                struct cma_device *cma_dev)
 450 {
 451         cma_dev_get(cma_dev);
 452         id_priv->cma_dev = cma_dev;
 453         id_priv->id.device = cma_dev->device;
 454         id_priv->id.route.addr.dev_addr.transport =
 455                 rdma_node_get_transport(cma_dev->device->node_type);
 456         list_add_tail(&id_priv->list, &cma_dev->id_list);
 457
 458         trace_cm_id_attach(id_priv, cma_dev->device);
 459 }
 460
 461 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 462                               struct cma_device *cma_dev)
 463 {
 464         _cma_attach_to_dev(id_priv, cma_dev);
 465         id_priv->gid_type =
 466                 cma_dev->default_gid_type[id_priv->id.port_num -
 467                                           rdma_start_port(cma_dev->device)];
 468 }
 469
 470 static void cma_release_dev(struct rdma_id_private *id_priv)
 471 {
 472         mutex_lock(&lock);
 473         list_del(&id_priv->list);
 474         cma_dev_put(id_priv->cma_dev);
 475         id_priv->cma_dev = NULL;
 476         id_priv->id.device = NULL;
 477         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 478                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 479                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 480         }
 481         mutex_unlock(&lock);
 482 }
 483
 484 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 485 {
 486         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 487 }
 488
 489 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 490 {
 491         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 492 }
 493
 494 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 495 {
 496         return id_priv->id.route.addr.src_addr.ss_family;
 497 }
 498
 499 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 500 {
 501         struct ib_sa_mcmember_rec rec;
 502         int ret = 0;
 503
 504         if (id_priv->qkey) {
 505                 if (qkey && id_priv->qkey != qkey)
 506                         return -EINVAL;
 507                 return 0;
 508         }
 509
 510         if (qkey) {
 511                 id_priv->qkey = qkey;
 512                 return 0;
 513         }
 514
 515         switch (id_priv->id.ps) {
 516         case RDMA_PS_UDP:
 517         case RDMA_PS_IB:
 518                 id_priv->qkey = RDMA_UDP_QKEY;
 519                 break;
 520         case RDMA_PS_IPOIB:
 521                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 522                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 523                                              id_priv->id.port_num, &rec.mgid,
 524                                              &rec);
 525                 if (!ret)
 526                         id_priv->qkey = be32_to_cpu(rec.qkey);
 527                 break;
 528         default:
 529                 break;
 530         }
 531         return ret;
 532 }
 533
 534 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 535 {
 536         dev_addr->dev_type = ARPHRD_INFINIBAND;
 537         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 538         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 539 }
 540
 541 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 542 {
 543         int ret;
 544
 545         if (addr->sa_family != AF_IB) {
 546                 ret = rdma_translate_ip(addr, dev_addr);
 547         } else {
 548                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 549                 ret = 0;
 550         }
 551
 552         return ret;
 553 }
 554
 555 static const struct ib_gid_attr *
 556 cma_validate_port(struct ib_device *device, u32 port,
 557                   enum ib_gid_type gid_type,
 558                   union ib_gid *gid,
 559                   struct rdma_id_private *id_priv)
 560 {
 561         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 562         int bound_if_index = dev_addr->bound_dev_if;
 563         const struct ib_gid_attr *sgid_attr;
 564         int dev_type = dev_addr->dev_type;
 565         struct net_device *ndev = NULL;
 566
 567         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 568                 return ERR_PTR(-ENODEV);
 569
 570         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 571                 return ERR_PTR(-ENODEV);
 572
 573         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 574                 return ERR_PTR(-ENODEV);
 575
 576         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 577                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 578                 if (!ndev)
 579                         return ERR_PTR(-ENODEV);
 580         } else {
 581                 gid_type = IB_GID_TYPE_IB;
 582         }
 583
 584         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 585         if (ndev)
 586                 dev_put(ndev);
 587         return sgid_attr;
 588 }
 589
 590 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 591                                const struct ib_gid_attr *sgid_attr)
 592 {
 593         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 594         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 595 }
 596
 597 /**
 598  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 599  * based on source ip address.
 600  * @id_priv:    cm_id which should be bound to cma device
 601  *
 602  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 603  * based on source IP address. It returns 0 on success or error code otherwise.
 604  * It is applicable to active and passive side cm_id.
 605  */
 606 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 607 {
 608         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 609         const struct ib_gid_attr *sgid_attr;
 610         union ib_gid gid, iboe_gid, *gidp;
 611         struct cma_device *cma_dev;
 612         enum ib_gid_type gid_type;
 613         int ret = -ENODEV;
 614         u32 port;
 615
 616         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 617             id_priv->id.ps == RDMA_PS_IPOIB)
 618                 return -EINVAL;
 619
 620         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 621                     &iboe_gid);
 622
 623         memcpy(&gid, dev_addr->src_dev_addr +
 624                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 625
 626         mutex_lock(&lock);
 627         list_for_each_entry(cma_dev, &dev_list, list) {
 628                 rdma_for_each_port (cma_dev->device, port) {
 629                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 630                                &iboe_gid : &gid;
 631                         gid_type = cma_dev->default_gid_type[port - 1];
 632                         sgid_attr = cma_validate_port(cma_dev->device, port,
 633                                                       gid_type, gidp, id_priv);
 634                         if (!IS_ERR(sgid_attr)) {
 635                                 id_priv->id.port_num = port;
 636                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 637                                 cma_attach_to_dev(id_priv, cma_dev);
 638                                 ret = 0;
 639                                 goto out;
 640                         }
 641                 }
 642         }
 643 out:
 644         mutex_unlock(&lock);
 645         return ret;
 646 }
 647
 648 /**
 649  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 650  * @id_priv:            cm id to bind to cma device
 651  * @listen_id_priv:     listener cm id to match against
 652  * @req:                Pointer to req structure containaining incoming
 653  *                      request information
 654  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 655  * rdma device matches for listen_id and incoming request. It also verifies
 656  * that a GID table entry is present for the source address.
 657  * Returns 0 on success, or returns error code otherwise.
 658  */
 659 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 660                               const struct rdma_id_private *listen_id_priv,
 661                               struct cma_req_info *req)
 662 {
 663         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 664         const struct ib_gid_attr *sgid_attr;
 665         enum ib_gid_type gid_type;
 666         union ib_gid gid;
 667
 668         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 669             id_priv->id.ps == RDMA_PS_IPOIB)
 670                 return -EINVAL;
 671
 672         if (rdma_protocol_roce(req->device, req->port))
 673                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 674                             &gid);
 675         else
 676                 memcpy(&gid, dev_addr->src_dev_addr +
 677                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 678
 679         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 680         sgid_attr = cma_validate_port(req->device, req->port,
 681                                       gid_type, &gid, id_priv);
 682         if (IS_ERR(sgid_attr))
 683                 return PTR_ERR(sgid_attr);
 684
 685         id_priv->id.port_num = req->port;
 686         cma_bind_sgid_attr(id_priv, sgid_attr);
 687         /* Need to acquire lock to protect against reader
 688          * of cma_dev->id_list such as cma_netdev_callback() and
 689          * cma_process_remove().
 690          */
 691         mutex_lock(&lock);
 692         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 693         mutex_unlock(&lock);
 694         rdma_restrack_add(&id_priv->res);
 695         return 0;
 696 }
 697
 698 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 699                               const struct rdma_id_private *listen_id_priv)
 700 {
 701         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 702         const struct ib_gid_attr *sgid_attr;
 703         struct cma_device *cma_dev;
 704         enum ib_gid_type gid_type;
 705         int ret = -ENODEV;
 706         union ib_gid gid;
 707         u32 port;
 708
 709         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 710             id_priv->id.ps == RDMA_PS_IPOIB)
 711                 return -EINVAL;
 712
 713         memcpy(&gid, dev_addr->src_dev_addr +
 714                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 715
 716         mutex_lock(&lock);
 717
 718         cma_dev = listen_id_priv->cma_dev;
 719         port = listen_id_priv->id.port_num;
 720         gid_type = listen_id_priv->gid_type;
 721         sgid_attr = cma_validate_port(cma_dev->device, port,
 722                                       gid_type, &gid, id_priv);
 723         if (!IS_ERR(sgid_attr)) {
 724                 id_priv->id.port_num = port;
 725                 cma_bind_sgid_attr(id_priv, sgid_attr);
 726                 ret = 0;
 727                 goto out;
 728         }
 729
 730         list_for_each_entry(cma_dev, &dev_list, list) {
 731                 rdma_for_each_port (cma_dev->device, port) {
 732                         if (listen_id_priv->cma_dev == cma_dev &&
 733                             listen_id_priv->id.port_num == port)
 734                                 continue;
 735
 736                         gid_type = cma_dev->default_gid_type[port - 1];
 737                         sgid_attr = cma_validate_port(cma_dev->device, port,
 738                                                       gid_type, &gid, id_priv);
 739                         if (!IS_ERR(sgid_attr)) {
 740                                 id_priv->id.port_num = port;
 741                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 742                                 ret = 0;
 743                                 goto out;
 744                         }
 745                 }
 746         }
 747
 748 out:
 749         if (!ret) {
 750                 cma_attach_to_dev(id_priv, cma_dev);
 751                 rdma_restrack_add(&id_priv->res);
 752         }
 753
 754         mutex_unlock(&lock);
 755         return ret;
 756 }
 757
 758 /*
 759  * Select the source IB device and address to reach the destination IB address.
 760  */
 761 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 762 {
 763         struct cma_device *cma_dev, *cur_dev;
 764         struct sockaddr_ib *addr;
 765         union ib_gid gid, sgid, *dgid;
 766         unsigned int p;
 767         u16 pkey, index;
 768         enum ib_port_state port_state;
 769         int i;
 770
 771         cma_dev = NULL;
 772         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 773         dgid = (union ib_gid *) &addr->sib_addr;
 774         pkey = ntohs(addr->sib_pkey);
 775
 776         mutex_lock(&lock);
 777         list_for_each_entry(cur_dev, &dev_list, list) {
 778                 rdma_for_each_port (cur_dev->device, p) {
 779                         if (!rdma_cap_af_ib(cur_dev->device, p))
 780                                 continue;
 781
 782                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 783                                 continue;
 784
 785                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 786                                 continue;
 787                         for (i = 0; !rdma_query_gid(cur_dev->device,
 788                                                     p, i, &gid);
 789                              i++) {
 790                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 791                                         cma_dev = cur_dev;
 792                                         sgid = gid;
 793                                         id_priv->id.port_num = p;
 794                                         goto found;
 795                                 }
 796
 797                                 if (!cma_dev && (gid.global.subnet_prefix ==
 798                                     dgid->global.subnet_prefix) &&
 799                                     port_state == IB_PORT_ACTIVE) {
 800                                         cma_dev = cur_dev;
 801                                         sgid = gid;
 802                                         id_priv->id.port_num = p;
 803                                         goto found;
 804                                 }
 805                         }
 806                 }
 807         }
 808         mutex_unlock(&lock);
 809         return -ENODEV;
 810
 811 found:
 812         cma_attach_to_dev(id_priv, cma_dev);
 813         rdma_restrack_add(&id_priv->res);
 814         mutex_unlock(&lock);
 815         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 816         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 817         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 818         return 0;
 819 }
 820
 821 static void cma_id_get(struct rdma_id_private *id_priv)
 822 {
 823         refcount_inc(&id_priv->refcount);
 824 }
 825
 826 static void cma_id_put(struct rdma_id_private *id_priv)
 827 {
 828         if (refcount_dec_and_test(&id_priv->refcount))
 829                 complete(&id_priv->comp);
 830 }
 831
 832 static struct rdma_id_private *
 833 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 834                  void *context, enum rdma_ucm_port_space ps,
 835                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 836 {
 837         struct rdma_id_private *id_priv;
 838
 839         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 840         if (!id_priv)
 841                 return ERR_PTR(-ENOMEM);
 842
 843         id_priv->state = RDMA_CM_IDLE;
 844         id_priv->id.context = context;
 845         id_priv->id.event_handler = event_handler;
 846         id_priv->id.ps = ps;
 847         id_priv->id.qp_type = qp_type;
 848         id_priv->tos_set = false;
 849         id_priv->timeout_set = false;
 850         id_priv->min_rnr_timer_set = false;
 851         id_priv->gid_type = IB_GID_TYPE_IB;
 852         spin_lock_init(&id_priv->lock);
 853         mutex_init(&id_priv->qp_mutex);
 854         init_completion(&id_priv->comp);
 855         refcount_set(&id_priv->refcount, 1);
 856         mutex_init(&id_priv->handler_mutex);
 857         INIT_LIST_HEAD(&id_priv->listen_list);
 858         INIT_LIST_HEAD(&id_priv->mc_list);
 859         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 860         id_priv->id.route.addr.dev_addr.net = get_net(net);
 861         id_priv->seq_num &= 0x00ffffff;
 862
 863         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 864         if (parent)
 865                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 866
 867         return id_priv;
 868 }
 869
 870 struct rdma_cm_id *
 871 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
 872                         void *context, enum rdma_ucm_port_space ps,
 873                         enum ib_qp_type qp_type, const char *caller)
 874 {
 875         struct rdma_id_private *ret;
 876
 877         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
 878         if (IS_ERR(ret))
 879                 return ERR_CAST(ret);
 880
 881         rdma_restrack_set_name(&ret->res, caller);
 882         return &ret->id;
 883 }
 884 EXPORT_SYMBOL(__rdma_create_kernel_id);
 885
 886 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
 887                                        void *context,
 888                                        enum rdma_ucm_port_space ps,
 889                                        enum ib_qp_type qp_type)
 890 {
 891         struct rdma_id_private *ret;
 892
 893         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
 894                                ps, qp_type, NULL);
 895         if (IS_ERR(ret))
 896                 return ERR_CAST(ret);
 897
 898         rdma_restrack_set_name(&ret->res, NULL);
 899         return &ret->id;
 900 }
 901 EXPORT_SYMBOL(rdma_create_user_id);
 902
 903 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 904 {
 905         struct ib_qp_attr qp_attr;
 906         int qp_attr_mask, ret;
 907
 908         qp_attr.qp_state = IB_QPS_INIT;
 909         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 910         if (ret)
 911                 return ret;
 912
 913         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 914         if (ret)
 915                 return ret;
 916
 917         qp_attr.qp_state = IB_QPS_RTR;
 918         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 919         if (ret)
 920                 return ret;
 921
 922         qp_attr.qp_state = IB_QPS_RTS;
 923         qp_attr.sq_psn = 0;
 924         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 925
 926         return ret;
 927 }
 928
 929 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 930                    struct ib_qp_init_attr *qp_init_attr)
 931 {
 932         struct rdma_id_private *id_priv;
 933         struct ib_qp *qp;
 934         int ret = 0;
 935
 936         id_priv = container_of(id, struct rdma_id_private, id);
 937         if (id->device != pd->device) {
 938                 ret = -EINVAL;
 939                 goto out_err;
 940         }
 941
 942         qp_init_attr->port_num = id->port_num;
 943         qp = ib_create_qp(pd, qp_init_attr);
 944         if (IS_ERR(qp)) {
 945                 ret = PTR_ERR(qp);
 946                 goto out_err;
 947         }
 948
 949         if (id->qp_type == IB_QPT_UD)
 950                 ret = cma_init_ud_qp(id_priv, qp);
 951         if (ret)
 952                 goto out_destroy;
 953
 954         id->qp = qp;
 955         id_priv->qp_num = qp->qp_num;
 956         id_priv->srq = (qp->srq != NULL);
 957         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 958         return 0;
 959 out_destroy:
 960         ib_destroy_qp(qp);
 961 out_err:
 962         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 963         return ret;
 964 }
 965 EXPORT_SYMBOL(rdma_create_qp);
 966
 967 void rdma_destroy_qp(struct rdma_cm_id *id)
 968 {
 969         struct rdma_id_private *id_priv;
 970
 971         id_priv = container_of(id, struct rdma_id_private, id);
 972         trace_cm_qp_destroy(id_priv);
 973         mutex_lock(&id_priv->qp_mutex);
 974         ib_destroy_qp(id_priv->id.qp);
 975         id_priv->id.qp = NULL;
 976         mutex_unlock(&id_priv->qp_mutex);
 977 }
 978 EXPORT_SYMBOL(rdma_destroy_qp);
 979
 980 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 981                              struct rdma_conn_param *conn_param)
 982 {
 983         struct ib_qp_attr qp_attr;
 984         int qp_attr_mask, ret;
 985
 986         mutex_lock(&id_priv->qp_mutex);
 987         if (!id_priv->id.qp) {
 988                 ret = 0;
 989                 goto out;
 990         }
 991
 992         /* Need to update QP attributes from default values. */
 993         qp_attr.qp_state = IB_QPS_INIT;
 994         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 995         if (ret)
 996                 goto out;
 997
 998         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 999         if (ret)
1000                 goto out;
1001
1002         qp_attr.qp_state = IB_QPS_RTR;
1003         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1004         if (ret)
1005                 goto out;
1006
1007         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1008
1009         if (conn_param)
1010                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1011         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1012 out:
1013         mutex_unlock(&id_priv->qp_mutex);
1014         return ret;
1015 }
1016
1017 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1018                              struct rdma_conn_param *conn_param)
1019 {
1020         struct ib_qp_attr qp_attr;
1021         int qp_attr_mask, ret;
1022
1023         mutex_lock(&id_priv->qp_mutex);
1024         if (!id_priv->id.qp) {
1025                 ret = 0;
1026                 goto out;
1027         }
1028
1029         qp_attr.qp_state = IB_QPS_RTS;
1030         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1031         if (ret)
1032                 goto out;
1033
1034         if (conn_param)
1035                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1036         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1037 out:
1038         mutex_unlock(&id_priv->qp_mutex);
1039         return ret;
1040 }
1041
1042 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1043 {
1044         struct ib_qp_attr qp_attr;
1045         int ret;
1046
1047         mutex_lock(&id_priv->qp_mutex);
1048         if (!id_priv->id.qp) {
1049                 ret = 0;
1050                 goto out;
1051         }
1052
1053         qp_attr.qp_state = IB_QPS_ERR;
1054         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1055 out:
1056         mutex_unlock(&id_priv->qp_mutex);
1057         return ret;
1058 }
1059
1060 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1061                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1062 {
1063         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1064         int ret;
1065         u16 pkey;
1066
1067         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1068                 pkey = 0xffff;
1069         else
1070                 pkey = ib_addr_get_pkey(dev_addr);
1071
1072         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1073                                   pkey, &qp_attr->pkey_index);
1074         if (ret)
1075                 return ret;
1076
1077         qp_attr->port_num = id_priv->id.port_num;
1078         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1079
1080         if (id_priv->id.qp_type == IB_QPT_UD) {
1081                 ret = cma_set_qkey(id_priv, 0);
1082                 if (ret)
1083                         return ret;
1084
1085                 qp_attr->qkey = id_priv->qkey;
1086                 *qp_attr_mask |= IB_QP_QKEY;
1087         } else {
1088                 qp_attr->qp_access_flags = 0;
1089                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1090         }
1091         return 0;
1092 }
1093
1094 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1095                        int *qp_attr_mask)
1096 {
1097         struct rdma_id_private *id_priv;
1098         int ret = 0;
1099
1100         id_priv = container_of(id, struct rdma_id_private, id);
1101         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1102                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1103                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1104                 else
1105                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1106                                                  qp_attr_mask);
1107
1108                 if (qp_attr->qp_state == IB_QPS_RTR)
1109                         qp_attr->rq_psn = id_priv->seq_num;
1110         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1111                 if (!id_priv->cm_id.iw) {
1112                         qp_attr->qp_access_flags = 0;
1113                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1114                 } else
1115                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1116                                                  qp_attr_mask);
1117                 qp_attr->port_num = id_priv->id.port_num;
1118                 *qp_attr_mask |= IB_QP_PORT;
1119         } else {
1120                 ret = -ENOSYS;
1121         }
1122
1123         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1124                 qp_attr->timeout = id_priv->timeout;
1125
1126         if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1127                 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1128
1129         return ret;
1130 }
1131 EXPORT_SYMBOL(rdma_init_qp_attr);
1132
1133 static inline bool cma_zero_addr(const struct sockaddr *addr)
1134 {
1135         switch (addr->sa_family) {
1136         case AF_INET:
1137                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1138         case AF_INET6:
1139                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1140         case AF_IB:
1141                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1142         default:
1143                 return false;
1144         }
1145 }
1146
1147 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1148 {
1149         switch (addr->sa_family) {
1150         case AF_INET:
1151                 return ipv4_is_loopback(
1152                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1153         case AF_INET6:
1154                 return ipv6_addr_loopback(
1155                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1156         case AF_IB:
1157                 return ib_addr_loopback(
1158                         &((struct sockaddr_ib *)addr)->sib_addr);
1159         default:
1160                 return false;
1161         }
1162 }
1163
1164 static inline bool cma_any_addr(const struct sockaddr *addr)
1165 {
1166         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1167 }
1168
1169 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1170 {
1171         if (src->sa_family != dst->sa_family)
1172                 return -1;
1173
1174         switch (src->sa_family) {
1175         case AF_INET:
1176                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1177                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1178         case AF_INET6: {
1179                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1180                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1181                 bool link_local;
1182
1183                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1184                                           &dst_addr6->sin6_addr))
1185                         return 1;
1186                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1187                              IPV6_ADDR_LINKLOCAL;
1188                 /* Link local must match their scope_ids */
1189                 return link_local ? (src_addr6->sin6_scope_id !=
1190                                      dst_addr6->sin6_scope_id) :
1191                                     0;
1192         }
1193
1194         default:
1195                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1196                                    &((struct sockaddr_ib *) dst)->sib_addr);
1197         }
1198 }
1199
1200 static __be16 cma_port(const struct sockaddr *addr)
1201 {
1202         struct sockaddr_ib *sib;
1203
1204         switch (addr->sa_family) {
1205         case AF_INET:
1206                 return ((struct sockaddr_in *) addr)->sin_port;
1207         case AF_INET6:
1208                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1209         case AF_IB:
1210                 sib = (struct sockaddr_ib *) addr;
1211                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1212                                     be64_to_cpu(sib->sib_sid_mask)));
1213         default:
1214                 return 0;
1215         }
1216 }
1217
1218 static inline int cma_any_port(const struct sockaddr *addr)
1219 {
1220         return !cma_port(addr);
1221 }
1222
1223 static void cma_save_ib_info(struct sockaddr *src_addr,
1224                              struct sockaddr *dst_addr,
1225                              const struct rdma_cm_id *listen_id,
1226                              const struct sa_path_rec *path)
1227 {
1228         struct sockaddr_ib *listen_ib, *ib;
1229
1230         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1231         if (src_addr) {
1232                 ib = (struct sockaddr_ib *)src_addr;
1233                 ib->sib_family = AF_IB;
1234                 if (path) {
1235                         ib->sib_pkey = path->pkey;
1236                         ib->sib_flowinfo = path->flow_label;
1237                         memcpy(&ib->sib_addr, &path->sgid, 16);
1238                         ib->sib_sid = path->service_id;
1239                         ib->sib_scope_id = 0;
1240                 } else {
1241                         ib->sib_pkey = listen_ib->sib_pkey;
1242                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1243                         ib->sib_addr = listen_ib->sib_addr;
1244                         ib->sib_sid = listen_ib->sib_sid;
1245                         ib->sib_scope_id = listen_ib->sib_scope_id;
1246                 }
1247                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1248         }
1249         if (dst_addr) {
1250                 ib = (struct sockaddr_ib *)dst_addr;
1251                 ib->sib_family = AF_IB;
1252                 if (path) {
1253                         ib->sib_pkey = path->pkey;
1254                         ib->sib_flowinfo = path->flow_label;
1255                         memcpy(&ib->sib_addr, &path->dgid, 16);
1256                 }
1257         }
1258 }
1259
1260 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1261                               struct sockaddr_in *dst_addr,
1262                               struct cma_hdr *hdr,
1263                               __be16 local_port)
1264 {
1265         if (src_addr) {
1266                 *src_addr = (struct sockaddr_in) {
1267                         .sin_family = AF_INET,
1268                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1269                         .sin_port = local_port,
1270                 };
1271         }
1272
1273         if (dst_addr) {
1274                 *dst_addr = (struct sockaddr_in) {
1275                         .sin_family = AF_INET,
1276                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1277                         .sin_port = hdr->port,
1278                 };
1279         }
1280 }
1281
1282 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1283                               struct sockaddr_in6 *dst_addr,
1284                               struct cma_hdr *hdr,
1285                               __be16 local_port)
1286 {
1287         if (src_addr) {
1288                 *src_addr = (struct sockaddr_in6) {
1289                         .sin6_family = AF_INET6,
1290                         .sin6_addr = hdr->dst_addr.ip6,
1291                         .sin6_port = local_port,
1292                 };
1293         }
1294
1295         if (dst_addr) {
1296                 *dst_addr = (struct sockaddr_in6) {
1297                         .sin6_family = AF_INET6,
1298                         .sin6_addr = hdr->src_addr.ip6,
1299                         .sin6_port = hdr->port,
1300                 };
1301         }
1302 }
1303
1304 static u16 cma_port_from_service_id(__be64 service_id)
1305 {
1306         return (u16)be64_to_cpu(service_id);
1307 }
1308
1309 static int cma_save_ip_info(struct sockaddr *src_addr,
1310                             struct sockaddr *dst_addr,
1311                             const struct ib_cm_event *ib_event,
1312                             __be64 service_id)
1313 {
1314         struct cma_hdr *hdr;
1315         __be16 port;
1316
1317         hdr = ib_event->private_data;
1318         if (hdr->cma_version != CMA_VERSION)
1319                 return -EINVAL;
1320
1321         port = htons(cma_port_from_service_id(service_id));
1322
1323         switch (cma_get_ip_ver(hdr)) {
1324         case 4:
1325                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1326                                   (struct sockaddr_in *)dst_addr, hdr, port);
1327                 break;
1328         case 6:
1329                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1330                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1331                 break;
1332         default:
1333                 return -EAFNOSUPPORT;
1334         }
1335
1336         return 0;
1337 }
1338
1339 static int cma_save_net_info(struct sockaddr *src_addr,
1340                              struct sockaddr *dst_addr,
1341                              const struct rdma_cm_id *listen_id,
1342                              const struct ib_cm_event *ib_event,
1343                              sa_family_t sa_family, __be64 service_id)
1344 {
1345         if (sa_family == AF_IB) {
1346                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1347                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1348                                          ib_event->param.req_rcvd.primary_path);
1349                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1350                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1351                 return 0;
1352         }
1353
1354         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1355 }
1356
1357 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1358                              struct cma_req_info *req)
1359 {
1360         const struct ib_cm_req_event_param *req_param =
1361                 &ib_event->param.req_rcvd;
1362         const struct ib_cm_sidr_req_event_param *sidr_param =
1363                 &ib_event->param.sidr_req_rcvd;
1364
1365         switch (ib_event->event) {
1366         case IB_CM_REQ_RECEIVED:
1367                 req->device     = req_param->listen_id->device;
1368                 req->port       = req_param->port;
1369                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1370                        sizeof(req->local_gid));
1371                 req->has_gid    = true;
1372                 req->service_id = req_param->primary_path->service_id;
1373                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1374                 if (req->pkey != req_param->bth_pkey)
1375                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1376                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1377                                             req_param->bth_pkey, req->pkey);
1378                 break;
1379         case IB_CM_SIDR_REQ_RECEIVED:
1380                 req->device     = sidr_param->listen_id->device;
1381                 req->port       = sidr_param->port;
1382                 req->has_gid    = false;
1383                 req->service_id = sidr_param->service_id;
1384                 req->pkey       = sidr_param->pkey;
1385                 if (req->pkey != sidr_param->bth_pkey)
1386                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1387                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1388                                             sidr_param->bth_pkey, req->pkey);
1389                 break;
1390         default:
1391                 return -EINVAL;
1392         }
1393
1394         return 0;
1395 }
1396
1397 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1398                                   const struct sockaddr_in *dst_addr,
1399                                   const struct sockaddr_in *src_addr)
1400 {
1401         __be32 daddr = dst_addr->sin_addr.s_addr,
1402                saddr = src_addr->sin_addr.s_addr;
1403         struct fib_result res;
1404         struct flowi4 fl4;
1405         int err;
1406         bool ret;
1407
1408         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1409             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1410             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1411             ipv4_is_loopback(saddr))
1412                 return false;
1413
1414         memset(&fl4, 0, sizeof(fl4));
1415         fl4.flowi4_iif = net_dev->ifindex;
1416         fl4.daddr = daddr;
1417         fl4.saddr = saddr;
1418
1419         rcu_read_lock();
1420         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1421         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1422         rcu_read_unlock();
1423
1424         return ret;
1425 }
1426
1427 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1428                                   const struct sockaddr_in6 *dst_addr,
1429                                   const struct sockaddr_in6 *src_addr)
1430 {
1431 #if IS_ENABLED(CONFIG_IPV6)
1432         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1433                            IPV6_ADDR_LINKLOCAL;
1434         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1435                                          &src_addr->sin6_addr, net_dev->ifindex,
1436                                          NULL, strict);
1437         bool ret;
1438
1439         if (!rt)
1440                 return false;
1441
1442         ret = rt->rt6i_idev->dev == net_dev;
1443         ip6_rt_put(rt);
1444
1445         return ret;
1446 #else
1447         return false;
1448 #endif
1449 }
1450
1451 static bool validate_net_dev(struct net_device *net_dev,
1452                              const struct sockaddr *daddr,
1453                              const struct sockaddr *saddr)
1454 {
1455         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1456         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1457         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1458         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1459
1460         switch (daddr->sa_family) {
1461         case AF_INET:
1462                 return saddr->sa_family == AF_INET &&
1463                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1464
1465         case AF_INET6:
1466                 return saddr->sa_family == AF_INET6 &&
1467                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1468
1469         default:
1470                 return false;
1471         }
1472 }
1473
1474 static struct net_device *
1475 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1476 {
1477         const struct ib_gid_attr *sgid_attr = NULL;
1478         struct net_device *ndev;
1479
1480         if (ib_event->event == IB_CM_REQ_RECEIVED)
1481                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1482         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1483                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1484
1485         if (!sgid_attr)
1486                 return NULL;
1487
1488         rcu_read_lock();
1489         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1490         if (IS_ERR(ndev))
1491                 ndev = NULL;
1492         else
1493                 dev_hold(ndev);
1494         rcu_read_unlock();
1495         return ndev;
1496 }
1497
1498 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1499                                           struct cma_req_info *req)
1500 {
1501         struct sockaddr *listen_addr =
1502                         (struct sockaddr *)&req->listen_addr_storage;
1503         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1504         struct net_device *net_dev;
1505         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1506         int err;
1507
1508         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1509                                req->service_id);
1510         if (err)
1511                 return ERR_PTR(err);
1512
1513         if (rdma_protocol_roce(req->device, req->port))
1514                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1515         else
1516                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1517                                                    req->pkey,
1518                                                    gid, listen_addr);
1519         if (!net_dev)
1520                 return ERR_PTR(-ENODEV);
1521
1522         return net_dev;
1523 }
1524
1525 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1526 {
1527         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1528 }
1529
1530 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1531                                    const struct cma_hdr *hdr)
1532 {
1533         struct sockaddr *addr = cma_src_addr(id_priv);
1534         __be32 ip4_addr;
1535         struct in6_addr ip6_addr;
1536
1537         if (cma_any_addr(addr) && !id_priv->afonly)
1538                 return true;
1539
1540         switch (addr->sa_family) {
1541         case AF_INET:
1542                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1543                 if (cma_get_ip_ver(hdr) != 4)
1544                         return false;
1545                 if (!cma_any_addr(addr) &&
1546                     hdr->dst_addr.ip4.addr != ip4_addr)
1547                         return false;
1548                 break;
1549         case AF_INET6:
1550                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1551                 if (cma_get_ip_ver(hdr) != 6)
1552                         return false;
1553                 if (!cma_any_addr(addr) &&
1554                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1555                         return false;
1556                 break;
1557         case AF_IB:
1558                 return true;
1559         default:
1560                 return false;
1561         }
1562
1563         return true;
1564 }
1565
1566 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1567 {
1568         struct ib_device *device = id->device;
1569         const u32 port_num = id->port_num ?: rdma_start_port(device);
1570
1571         return rdma_protocol_roce(device, port_num);
1572 }
1573
1574 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1575 {
1576         const struct sockaddr *daddr =
1577                         (const struct sockaddr *)&req->listen_addr_storage;
1578         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1579
1580         /* Returns true if the req is for IPv6 link local */
1581         return (daddr->sa_family == AF_INET6 &&
1582                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1583 }
1584
1585 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1586                               const struct net_device *net_dev,
1587                               const struct cma_req_info *req)
1588 {
1589         const struct rdma_addr *addr = &id->route.addr;
1590
1591         if (!net_dev)
1592                 /* This request is an AF_IB request */
1593                 return (!id->port_num || id->port_num == req->port) &&
1594                        (addr->src_addr.ss_family == AF_IB);
1595
1596         /*
1597          * If the request is not for IPv6 link local, allow matching
1598          * request to any netdevice of the one or multiport rdma device.
1599          */
1600         if (!cma_is_req_ipv6_ll(req))
1601                 return true;
1602         /*
1603          * Net namespaces must match, and if the listner is listening
1604          * on a specific netdevice than netdevice must match as well.
1605          */
1606         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1607             (!!addr->dev_addr.bound_dev_if ==
1608              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1609                 return true;
1610         else
1611                 return false;
1612 }
1613
1614 static struct rdma_id_private *cma_find_listener(
1615                 const struct rdma_bind_list *bind_list,
1616                 const struct ib_cm_id *cm_id,
1617                 const struct ib_cm_event *ib_event,
1618                 const struct cma_req_info *req,
1619                 const struct net_device *net_dev)
1620 {
1621         struct rdma_id_private *id_priv, *id_priv_dev;
1622
1623         lockdep_assert_held(&lock);
1624
1625         if (!bind_list)
1626                 return ERR_PTR(-EINVAL);
1627
1628         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1629                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1630                         if (id_priv->id.device == cm_id->device &&
1631                             cma_match_net_dev(&id_priv->id, net_dev, req))
1632                                 return id_priv;
1633                         list_for_each_entry(id_priv_dev,
1634                                             &id_priv->listen_list,
1635                                             listen_list) {
1636                                 if (id_priv_dev->id.device == cm_id->device &&
1637                                     cma_match_net_dev(&id_priv_dev->id,
1638                                                       net_dev, req))
1639                                         return id_priv_dev;
1640                         }
1641                 }
1642         }
1643
1644         return ERR_PTR(-EINVAL);
1645 }
1646
1647 static struct rdma_id_private *
1648 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1649                      const struct ib_cm_event *ib_event,
1650                      struct cma_req_info *req,
1651                      struct net_device **net_dev)
1652 {
1653         struct rdma_bind_list *bind_list;
1654         struct rdma_id_private *id_priv;
1655         int err;
1656
1657         err = cma_save_req_info(ib_event, req);
1658         if (err)
1659                 return ERR_PTR(err);
1660
1661         *net_dev = cma_get_net_dev(ib_event, req);
1662         if (IS_ERR(*net_dev)) {
1663                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1664                         /* Assuming the protocol is AF_IB */
1665                         *net_dev = NULL;
1666                 } else {
1667                         return ERR_CAST(*net_dev);
1668                 }
1669         }
1670
1671         mutex_lock(&lock);
1672         /*
1673          * Net namespace might be getting deleted while route lookup,
1674          * cm_id lookup is in progress. Therefore, perform netdevice
1675          * validation, cm_id lookup under rcu lock.
1676          * RCU lock along with netdevice state check, synchronizes with
1677          * netdevice migrating to different net namespace and also avoids
1678          * case where net namespace doesn't get deleted while lookup is in
1679          * progress.
1680          * If the device state is not IFF_UP, its properties such as ifindex
1681          * and nd_net cannot be trusted to remain valid without rcu lock.
1682          * net/core/dev.c change_net_namespace() ensures to synchronize with
1683          * ongoing operations on net device after device is closed using
1684          * synchronize_net().
1685          */
1686         rcu_read_lock();
1687         if (*net_dev) {
1688                 /*
1689                  * If netdevice is down, it is likely that it is administratively
1690                  * down or it might be migrating to different namespace.
1691                  * In that case avoid further processing, as the net namespace
1692                  * or ifindex may change.
1693                  */
1694                 if (((*net_dev)->flags & IFF_UP) == 0) {
1695                         id_priv = ERR_PTR(-EHOSTUNREACH);
1696                         goto err;
1697                 }
1698
1699                 if (!validate_net_dev(*net_dev,
1700                                  (struct sockaddr *)&req->listen_addr_storage,
1701                                  (struct sockaddr *)&req->src_addr_storage)) {
1702                         id_priv = ERR_PTR(-EHOSTUNREACH);
1703                         goto err;
1704                 }
1705         }
1706
1707         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1708                                 rdma_ps_from_service_id(req->service_id),
1709                                 cma_port_from_service_id(req->service_id));
1710         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1711 err:
1712         rcu_read_unlock();
1713         mutex_unlock(&lock);
1714         if (IS_ERR(id_priv) && *net_dev) {
1715                 dev_put(*net_dev);
1716                 *net_dev = NULL;
1717         }
1718         return id_priv;
1719 }
1720
1721 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1722 {
1723         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1724 }
1725
1726 static void cma_cancel_route(struct rdma_id_private *id_priv)
1727 {
1728         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1729                 if (id_priv->query)
1730                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1731         }
1732 }
1733
1734 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1735 {
1736         struct rdma_id_private *dev_id_priv;
1737
1738         /*
1739          * Remove from listen_any_list to prevent added devices from spawning
1740          * additional listen requests.
1741          */
1742         mutex_lock(&lock);
1743         list_del(&id_priv->list);
1744
1745         while (!list_empty(&id_priv->listen_list)) {
1746                 dev_id_priv = list_entry(id_priv->listen_list.next,
1747                                          struct rdma_id_private, listen_list);
1748                 /* sync with device removal to avoid duplicate destruction */
1749                 list_del_init(&dev_id_priv->list);
1750                 list_del(&dev_id_priv->listen_list);
1751                 mutex_unlock(&lock);
1752
1753                 rdma_destroy_id(&dev_id_priv->id);
1754                 mutex_lock(&lock);
1755         }
1756         mutex_unlock(&lock);
1757 }
1758
1759 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1760                                  enum rdma_cm_state state)
1761 {
1762         switch (state) {
1763         case RDMA_CM_ADDR_QUERY:
1764                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1765                 break;
1766         case RDMA_CM_ROUTE_QUERY:
1767                 cma_cancel_route(id_priv);
1768                 break;
1769         case RDMA_CM_LISTEN:
1770                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1771                         cma_cancel_listens(id_priv);
1772                 break;
1773         default:
1774                 break;
1775         }
1776 }
1777
1778 static void cma_release_port(struct rdma_id_private *id_priv)
1779 {
1780         struct rdma_bind_list *bind_list = id_priv->bind_list;
1781         struct net *net = id_priv->id.route.addr.dev_addr.net;
1782
1783         if (!bind_list)
1784                 return;
1785
1786         mutex_lock(&lock);
1787         hlist_del(&id_priv->node);
1788         if (hlist_empty(&bind_list->owners)) {
1789                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1790                 kfree(bind_list);
1791         }
1792         mutex_unlock(&lock);
1793 }
1794
1795 static void destroy_mc(struct rdma_id_private *id_priv,
1796                        struct cma_multicast *mc)
1797 {
1798         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1799                 ib_sa_free_multicast(mc->sa_mc);
1800
1801         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1802                 struct rdma_dev_addr *dev_addr =
1803                         &id_priv->id.route.addr.dev_addr;
1804                 struct net_device *ndev = NULL;
1805
1806                 if (dev_addr->bound_dev_if)
1807                         ndev = dev_get_by_index(dev_addr->net,
1808                                                 dev_addr->bound_dev_if);
1809                 if (ndev) {
1810                         union ib_gid mgid;
1811
1812                         cma_set_mgid(id_priv, (struct sockaddr *)&mc->addr,
1813                                      &mgid);
1814                         cma_igmp_send(ndev, &mgid, false);
1815                         dev_put(ndev);
1816                 }
1817
1818                 cancel_work_sync(&mc->iboe_join.work);
1819         }
1820         kfree(mc);
1821 }
1822
1823 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1824 {
1825         struct cma_multicast *mc;
1826
1827         while (!list_empty(&id_priv->mc_list)) {
1828                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1829                                       list);
1830                 list_del(&mc->list);
1831                 destroy_mc(id_priv, mc);
1832         }
1833 }
1834
1835 static void _destroy_id(struct rdma_id_private *id_priv,
1836                         enum rdma_cm_state state)
1837 {
1838         cma_cancel_operation(id_priv, state);
1839
1840         rdma_restrack_del(&id_priv->res);
1841         if (id_priv->cma_dev) {
1842                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1843                         if (id_priv->cm_id.ib)
1844                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1845                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1846                         if (id_priv->cm_id.iw)
1847                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1848                 }
1849                 cma_leave_mc_groups(id_priv);
1850                 cma_release_dev(id_priv);
1851         }
1852
1853         cma_release_port(id_priv);
1854         cma_id_put(id_priv);
1855         wait_for_completion(&id_priv->comp);
1856
1857         if (id_priv->internal_id)
1858                 cma_id_put(id_priv->id.context);
1859
1860         kfree(id_priv->id.route.path_rec);
1861
1862         put_net(id_priv->id.route.addr.dev_addr.net);
1863         kfree(id_priv);
1864 }
1865
1866 /*
1867  * destroy an ID from within the handler_mutex. This ensures that no other
1868  * handlers can start running concurrently.
1869  */
1870 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1871         __releases(&idprv->handler_mutex)
1872 {
1873         enum rdma_cm_state state;
1874         unsigned long flags;
1875
1876         trace_cm_id_destroy(id_priv);
1877
1878         /*
1879          * Setting the state to destroyed under the handler mutex provides a
1880          * fence against calling handler callbacks. If this is invoked due to
1881          * the failure of a handler callback then it guarentees that no future
1882          * handlers will be called.
1883          */
1884         lockdep_assert_held(&id_priv->handler_mutex);
1885         spin_lock_irqsave(&id_priv->lock, flags);
1886         state = id_priv->state;
1887         id_priv->state = RDMA_CM_DESTROYING;
1888         spin_unlock_irqrestore(&id_priv->lock, flags);
1889         mutex_unlock(&id_priv->handler_mutex);
1890         _destroy_id(id_priv, state);
1891 }
1892
1893 void rdma_destroy_id(struct rdma_cm_id *id)
1894 {
1895         struct rdma_id_private *id_priv =
1896                 container_of(id, struct rdma_id_private, id);
1897
1898         mutex_lock(&id_priv->handler_mutex);
1899         destroy_id_handler_unlock(id_priv);
1900 }
1901 EXPORT_SYMBOL(rdma_destroy_id);
1902
1903 static int cma_rep_recv(struct rdma_id_private *id_priv)
1904 {
1905         int ret;
1906
1907         ret = cma_modify_qp_rtr(id_priv, NULL);
1908         if (ret)
1909                 goto reject;
1910
1911         ret = cma_modify_qp_rts(id_priv, NULL);
1912         if (ret)
1913                 goto reject;
1914
1915         trace_cm_send_rtu(id_priv);
1916         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1917         if (ret)
1918                 goto reject;
1919
1920         return 0;
1921 reject:
1922         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1923         cma_modify_qp_err(id_priv);
1924         trace_cm_send_rej(id_priv);
1925         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1926                        NULL, 0, NULL, 0);
1927         return ret;
1928 }
1929
1930 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1931                                    const struct ib_cm_rep_event_param *rep_data,
1932                                    void *private_data)
1933 {
1934         event->param.conn.private_data = private_data;
1935         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1936         event->param.conn.responder_resources = rep_data->responder_resources;
1937         event->param.conn.initiator_depth = rep_data->initiator_depth;
1938         event->param.conn.flow_control = rep_data->flow_control;
1939         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1940         event->param.conn.srq = rep_data->srq;
1941         event->param.conn.qp_num = rep_data->remote_qpn;
1942
1943         event->ece.vendor_id = rep_data->ece.vendor_id;
1944         event->ece.attr_mod = rep_data->ece.attr_mod;
1945 }
1946
1947 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1948                                 struct rdma_cm_event *event)
1949 {
1950         int ret;
1951
1952         lockdep_assert_held(&id_priv->handler_mutex);
1953
1954         trace_cm_event_handler(id_priv, event);
1955         ret = id_priv->id.event_handler(&id_priv->id, event);
1956         trace_cm_event_done(id_priv, event, ret);
1957         return ret;
1958 }
1959
1960 static int cma_ib_handler(struct ib_cm_id *cm_id,
1961                           const struct ib_cm_event *ib_event)
1962 {
1963         struct rdma_id_private *id_priv = cm_id->context;
1964         struct rdma_cm_event event = {};
1965         enum rdma_cm_state state;
1966         int ret;
1967
1968         mutex_lock(&id_priv->handler_mutex);
1969         state = READ_ONCE(id_priv->state);
1970         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1971              state != RDMA_CM_CONNECT) ||
1972             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1973              state != RDMA_CM_DISCONNECT))
1974                 goto out;
1975
1976         switch (ib_event->event) {
1977         case IB_CM_REQ_ERROR:
1978         case IB_CM_REP_ERROR:
1979                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1980                 event.status = -ETIMEDOUT;
1981                 break;
1982         case IB_CM_REP_RECEIVED:
1983                 if (state == RDMA_CM_CONNECT &&
1984                     (id_priv->id.qp_type != IB_QPT_UD)) {
1985                         trace_cm_send_mra(id_priv);
1986                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1987                 }
1988                 if (id_priv->id.qp) {
1989                         event.status = cma_rep_recv(id_priv);
1990                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1991                                                      RDMA_CM_EVENT_ESTABLISHED;
1992                 } else {
1993                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1994                 }
1995                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1996                                        ib_event->private_data);
1997                 break;
1998         case IB_CM_RTU_RECEIVED:
1999         case IB_CM_USER_ESTABLISHED:
2000                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2001                 break;
2002         case IB_CM_DREQ_ERROR:
2003                 event.status = -ETIMEDOUT;
2004                 fallthrough;
2005         case IB_CM_DREQ_RECEIVED:
2006         case IB_CM_DREP_RECEIVED:
2007                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2008                                    RDMA_CM_DISCONNECT))
2009                         goto out;
2010                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2011                 break;
2012         case IB_CM_TIMEWAIT_EXIT:
2013                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2014                 break;
2015         case IB_CM_MRA_RECEIVED:
2016                 /* ignore event */
2017                 goto out;
2018         case IB_CM_REJ_RECEIVED:
2019                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2020                                                                                 ib_event->param.rej_rcvd.reason));
2021                 cma_modify_qp_err(id_priv);
2022                 event.status = ib_event->param.rej_rcvd.reason;
2023                 event.event = RDMA_CM_EVENT_REJECTED;
2024                 event.param.conn.private_data = ib_event->private_data;
2025                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2026                 break;
2027         default:
2028                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2029                        ib_event->event);
2030                 goto out;
2031         }
2032
2033         ret = cma_cm_event_handler(id_priv, &event);
2034         if (ret) {
2035                 /* Destroy the CM ID by returning a non-zero value. */
2036                 id_priv->cm_id.ib = NULL;
2037                 destroy_id_handler_unlock(id_priv);
2038                 return ret;
2039         }
2040 out:
2041         mutex_unlock(&id_priv->handler_mutex);
2042         return 0;
2043 }
2044
2045 static struct rdma_id_private *
2046 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2047                    const struct ib_cm_event *ib_event,
2048                    struct net_device *net_dev)
2049 {
2050         struct rdma_id_private *listen_id_priv;
2051         struct rdma_id_private *id_priv;
2052         struct rdma_cm_id *id;
2053         struct rdma_route *rt;
2054         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2055         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2056         const __be64 service_id =
2057                 ib_event->param.req_rcvd.primary_path->service_id;
2058         int ret;
2059
2060         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2061         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2062                                    listen_id->event_handler, listen_id->context,
2063                                    listen_id->ps,
2064                                    ib_event->param.req_rcvd.qp_type,
2065                                    listen_id_priv);
2066         if (IS_ERR(id_priv))
2067                 return NULL;
2068
2069         id = &id_priv->id;
2070         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2071                               (struct sockaddr *)&id->route.addr.dst_addr,
2072                               listen_id, ib_event, ss_family, service_id))
2073                 goto err;
2074
2075         rt = &id->route;
2076         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2077         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2078                                      GFP_KERNEL);
2079         if (!rt->path_rec)
2080                 goto err;
2081
2082         rt->path_rec[0] = *path;
2083         if (rt->num_paths == 2)
2084                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2085
2086         if (net_dev) {
2087                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2088         } else {
2089                 if (!cma_protocol_roce(listen_id) &&
2090                     cma_any_addr(cma_src_addr(id_priv))) {
2091                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2092                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2093                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2094                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2095                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2096                         if (ret)
2097                                 goto err;
2098                 }
2099         }
2100         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2101
2102         id_priv->state = RDMA_CM_CONNECT;
2103         return id_priv;
2104
2105 err:
2106         rdma_destroy_id(id);
2107         return NULL;
2108 }
2109
2110 static struct rdma_id_private *
2111 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2112                   const struct ib_cm_event *ib_event,
2113                   struct net_device *net_dev)
2114 {
2115         const struct rdma_id_private *listen_id_priv;
2116         struct rdma_id_private *id_priv;
2117         struct rdma_cm_id *id;
2118         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2119         struct net *net = listen_id->route.addr.dev_addr.net;
2120         int ret;
2121
2122         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2123         id_priv = __rdma_create_id(net, listen_id->event_handler,
2124                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2125                                    listen_id_priv);
2126         if (IS_ERR(id_priv))
2127                 return NULL;
2128
2129         id = &id_priv->id;
2130         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2131                               (struct sockaddr *)&id->route.addr.dst_addr,
2132                               listen_id, ib_event, ss_family,
2133                               ib_event->param.sidr_req_rcvd.service_id))
2134                 goto err;
2135
2136         if (net_dev) {
2137                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2138         } else {
2139                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2140                         ret = cma_translate_addr(cma_src_addr(id_priv),
2141                                                  &id->route.addr.dev_addr);
2142                         if (ret)
2143                                 goto err;
2144                 }
2145         }
2146
2147         id_priv->state = RDMA_CM_CONNECT;
2148         return id_priv;
2149 err:
2150         rdma_destroy_id(id);
2151         return NULL;
2152 }
2153
2154 static void cma_set_req_event_data(struct rdma_cm_event *event,
2155                                    const struct ib_cm_req_event_param *req_data,
2156                                    void *private_data, int offset)
2157 {
2158         event->param.conn.private_data = private_data + offset;
2159         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2160         event->param.conn.responder_resources = req_data->responder_resources;
2161         event->param.conn.initiator_depth = req_data->initiator_depth;
2162         event->param.conn.flow_control = req_data->flow_control;
2163         event->param.conn.retry_count = req_data->retry_count;
2164         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2165         event->param.conn.srq = req_data->srq;
2166         event->param.conn.qp_num = req_data->remote_qpn;
2167
2168         event->ece.vendor_id = req_data->ece.vendor_id;
2169         event->ece.attr_mod = req_data->ece.attr_mod;
2170 }
2171
2172 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2173                                     const struct ib_cm_event *ib_event)
2174 {
2175         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2176                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2177                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2178                  (id->qp_type == IB_QPT_UD)) ||
2179                 (!id->qp_type));
2180 }
2181
2182 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2183                               const struct ib_cm_event *ib_event)
2184 {
2185         struct rdma_id_private *listen_id, *conn_id = NULL;
2186         struct rdma_cm_event event = {};
2187         struct cma_req_info req = {};
2188         struct net_device *net_dev;
2189         u8 offset;
2190         int ret;
2191
2192         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2193         if (IS_ERR(listen_id))
2194                 return PTR_ERR(listen_id);
2195
2196         trace_cm_req_handler(listen_id, ib_event->event);
2197         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2198                 ret = -EINVAL;
2199                 goto net_dev_put;
2200         }
2201
2202         mutex_lock(&listen_id->handler_mutex);
2203         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2204                 ret = -ECONNABORTED;
2205                 goto err_unlock;
2206         }
2207
2208         offset = cma_user_data_offset(listen_id);
2209         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2210         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2211                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2212                 event.param.ud.private_data = ib_event->private_data + offset;
2213                 event.param.ud.private_data_len =
2214                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2215         } else {
2216                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2217                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2218                                        ib_event->private_data, offset);
2219         }
2220         if (!conn_id) {
2221                 ret = -ENOMEM;
2222                 goto err_unlock;
2223         }
2224
2225         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2226         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2227         if (ret) {
2228                 destroy_id_handler_unlock(conn_id);
2229                 goto err_unlock;
2230         }
2231
2232         conn_id->cm_id.ib = cm_id;
2233         cm_id->context = conn_id;
2234         cm_id->cm_handler = cma_ib_handler;
2235
2236         ret = cma_cm_event_handler(conn_id, &event);
2237         if (ret) {
2238                 /* Destroy the CM ID by returning a non-zero value. */
2239                 conn_id->cm_id.ib = NULL;
2240                 mutex_unlock(&listen_id->handler_mutex);
2241                 destroy_id_handler_unlock(conn_id);
2242                 goto net_dev_put;
2243         }
2244
2245         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2246             conn_id->id.qp_type != IB_QPT_UD) {
2247                 trace_cm_send_mra(cm_id->context);
2248                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2249         }
2250         mutex_unlock(&conn_id->handler_mutex);
2251
2252 err_unlock:
2253         mutex_unlock(&listen_id->handler_mutex);
2254
2255 net_dev_put:
2256         if (net_dev)
2257                 dev_put(net_dev);
2258
2259         return ret;
2260 }
2261
2262 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2263 {
2264         if (addr->sa_family == AF_IB)
2265                 return ((struct sockaddr_ib *) addr)->sib_sid;
2266
2267         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2268 }
2269 EXPORT_SYMBOL(rdma_get_service_id);
2270
2271 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2272                     union ib_gid *dgid)
2273 {
2274         struct rdma_addr *addr = &cm_id->route.addr;
2275
2276         if (!cm_id->device) {
2277                 if (sgid)
2278                         memset(sgid, 0, sizeof(*sgid));
2279                 if (dgid)
2280                         memset(dgid, 0, sizeof(*dgid));
2281                 return;
2282         }
2283
2284         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2285                 if (sgid)
2286                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2287                 if (dgid)
2288                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2289         } else {
2290                 if (sgid)
2291                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2292                 if (dgid)
2293                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2294         }
2295 }
2296 EXPORT_SYMBOL(rdma_read_gids);
2297
2298 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2299 {
2300         struct rdma_id_private *id_priv = iw_id->context;
2301         struct rdma_cm_event event = {};
2302         int ret = 0;
2303         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2304         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2305
2306         mutex_lock(&id_priv->handler_mutex);
2307         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2308                 goto out;
2309
2310         switch (iw_event->event) {
2311         case IW_CM_EVENT_CLOSE:
2312                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2313                 break;
2314         case IW_CM_EVENT_CONNECT_REPLY:
2315                 memcpy(cma_src_addr(id_priv), laddr,
2316                        rdma_addr_size(laddr));
2317                 memcpy(cma_dst_addr(id_priv), raddr,
2318                        rdma_addr_size(raddr));
2319                 switch (iw_event->status) {
2320                 case 0:
2321                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2322                         event.param.conn.initiator_depth = iw_event->ird;
2323                         event.param.conn.responder_resources = iw_event->ord;
2324                         break;
2325                 case -ECONNRESET:
2326                 case -ECONNREFUSED:
2327                         event.event = RDMA_CM_EVENT_REJECTED;
2328                         break;
2329                 case -ETIMEDOUT:
2330                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2331                         break;
2332                 default:
2333                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2334                         break;
2335                 }
2336                 break;
2337         case IW_CM_EVENT_ESTABLISHED:
2338                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2339                 event.param.conn.initiator_depth = iw_event->ird;
2340                 event.param.conn.responder_resources = iw_event->ord;
2341                 break;
2342         default:
2343                 goto out;
2344         }
2345
2346         event.status = iw_event->status;
2347         event.param.conn.private_data = iw_event->private_data;
2348         event.param.conn.private_data_len = iw_event->private_data_len;
2349         ret = cma_cm_event_handler(id_priv, &event);
2350         if (ret) {
2351                 /* Destroy the CM ID by returning a non-zero value. */
2352                 id_priv->cm_id.iw = NULL;
2353                 destroy_id_handler_unlock(id_priv);
2354                 return ret;
2355         }
2356
2357 out:
2358         mutex_unlock(&id_priv->handler_mutex);
2359         return ret;
2360 }
2361
2362 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2363                                struct iw_cm_event *iw_event)
2364 {
2365         struct rdma_id_private *listen_id, *conn_id;
2366         struct rdma_cm_event event = {};
2367         int ret = -ECONNABORTED;
2368         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2369         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2370
2371         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2372         event.param.conn.private_data = iw_event->private_data;
2373         event.param.conn.private_data_len = iw_event->private_data_len;
2374         event.param.conn.initiator_depth = iw_event->ird;
2375         event.param.conn.responder_resources = iw_event->ord;
2376
2377         listen_id = cm_id->context;
2378
2379         mutex_lock(&listen_id->handler_mutex);
2380         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2381                 goto out;
2382
2383         /* Create a new RDMA id for the new IW CM ID */
2384         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2385                                    listen_id->id.event_handler,
2386                                    listen_id->id.context, RDMA_PS_TCP,
2387                                    IB_QPT_RC, listen_id);
2388         if (IS_ERR(conn_id)) {
2389                 ret = -ENOMEM;
2390                 goto out;
2391         }
2392         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2393         conn_id->state = RDMA_CM_CONNECT;
2394
2395         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2396         if (ret) {
2397                 mutex_unlock(&listen_id->handler_mutex);
2398                 destroy_id_handler_unlock(conn_id);
2399                 return ret;
2400         }
2401
2402         ret = cma_iw_acquire_dev(conn_id, listen_id);
2403         if (ret) {
2404                 mutex_unlock(&listen_id->handler_mutex);
2405                 destroy_id_handler_unlock(conn_id);
2406                 return ret;
2407         }
2408
2409         conn_id->cm_id.iw = cm_id;
2410         cm_id->context = conn_id;
2411         cm_id->cm_handler = cma_iw_handler;
2412
2413         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2414         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2415
2416         ret = cma_cm_event_handler(conn_id, &event);
2417         if (ret) {
2418                 /* User wants to destroy the CM ID */
2419                 conn_id->cm_id.iw = NULL;
2420                 mutex_unlock(&listen_id->handler_mutex);
2421                 destroy_id_handler_unlock(conn_id);
2422                 return ret;
2423         }
2424
2425         mutex_unlock(&conn_id->handler_mutex);
2426
2427 out:
2428         mutex_unlock(&listen_id->handler_mutex);
2429         return ret;
2430 }
2431
2432 static int cma_ib_listen(struct rdma_id_private *id_priv)
2433 {
2434         struct sockaddr *addr;
2435         struct ib_cm_id *id;
2436         __be64 svc_id;
2437
2438         addr = cma_src_addr(id_priv);
2439         svc_id = rdma_get_service_id(&id_priv->id, addr);
2440         id = ib_cm_insert_listen(id_priv->id.device,
2441                                  cma_ib_req_handler, svc_id);
2442         if (IS_ERR(id))
2443                 return PTR_ERR(id);
2444         id_priv->cm_id.ib = id;
2445
2446         return 0;
2447 }
2448
2449 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2450 {
2451         int ret;
2452         struct iw_cm_id *id;
2453
2454         id = iw_create_cm_id(id_priv->id.device,
2455                              iw_conn_req_handler,
2456                              id_priv);
2457         if (IS_ERR(id))
2458                 return PTR_ERR(id);
2459
2460         mutex_lock(&id_priv->qp_mutex);
2461         id->tos = id_priv->tos;
2462         id->tos_set = id_priv->tos_set;
2463         mutex_unlock(&id_priv->qp_mutex);
2464         id->afonly = id_priv->afonly;
2465         id_priv->cm_id.iw = id;
2466
2467         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2468                rdma_addr_size(cma_src_addr(id_priv)));
2469
2470         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2471
2472         if (ret) {
2473                 iw_destroy_cm_id(id_priv->cm_id.iw);
2474                 id_priv->cm_id.iw = NULL;
2475         }
2476
2477         return ret;
2478 }
2479
2480 static int cma_listen_handler(struct rdma_cm_id *id,
2481                               struct rdma_cm_event *event)
2482 {
2483         struct rdma_id_private *id_priv = id->context;
2484
2485         /* Listening IDs are always destroyed on removal */
2486         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2487                 return -1;
2488
2489         id->context = id_priv->id.context;
2490         id->event_handler = id_priv->id.event_handler;
2491         trace_cm_event_handler(id_priv, event);
2492         return id_priv->id.event_handler(id, event);
2493 }
2494
2495 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2496                              struct cma_device *cma_dev,
2497                              struct rdma_id_private **to_destroy)
2498 {
2499         struct rdma_id_private *dev_id_priv;
2500         struct net *net = id_priv->id.route.addr.dev_addr.net;
2501         int ret;
2502
2503         lockdep_assert_held(&lock);
2504
2505         *to_destroy = NULL;
2506         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2507                 return 0;
2508
2509         dev_id_priv =
2510                 __rdma_create_id(net, cma_listen_handler, id_priv,
2511                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2512         if (IS_ERR(dev_id_priv))
2513                 return PTR_ERR(dev_id_priv);
2514
2515         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2516         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2517                rdma_addr_size(cma_src_addr(id_priv)));
2518
2519         _cma_attach_to_dev(dev_id_priv, cma_dev);
2520         rdma_restrack_add(&dev_id_priv->res);
2521         cma_id_get(id_priv);
2522         dev_id_priv->internal_id = 1;
2523         dev_id_priv->afonly = id_priv->afonly;
2524         mutex_lock(&id_priv->qp_mutex);
2525         dev_id_priv->tos_set = id_priv->tos_set;
2526         dev_id_priv->tos = id_priv->tos;
2527         mutex_unlock(&id_priv->qp_mutex);
2528
2529         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2530         if (ret)
2531                 goto err_listen;
2532         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2533         return 0;
2534 err_listen:
2535         /* Caller must destroy this after releasing lock */
2536         *to_destroy = dev_id_priv;
2537         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2538         return ret;
2539 }
2540
2541 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2542 {
2543         struct rdma_id_private *to_destroy;
2544         struct cma_device *cma_dev;
2545         int ret;
2546
2547         mutex_lock(&lock);
2548         list_add_tail(&id_priv->list, &listen_any_list);
2549         list_for_each_entry(cma_dev, &dev_list, list) {
2550                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2551                 if (ret) {
2552                         /* Prevent racing with cma_process_remove() */
2553                         if (to_destroy)
2554                                 list_del_init(&to_destroy->list);
2555                         goto err_listen;
2556                 }
2557         }
2558         mutex_unlock(&lock);
2559         return 0;
2560
2561 err_listen:
2562         list_del(&id_priv->list);
2563         mutex_unlock(&lock);
2564         if (to_destroy)
2565                 rdma_destroy_id(&to_destroy->id);
2566         return ret;
2567 }
2568
2569 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2570 {
2571         struct rdma_id_private *id_priv;
2572
2573         id_priv = container_of(id, struct rdma_id_private, id);
2574         mutex_lock(&id_priv->qp_mutex);
2575         id_priv->tos = (u8) tos;
2576         id_priv->tos_set = true;
2577         mutex_unlock(&id_priv->qp_mutex);
2578 }
2579 EXPORT_SYMBOL(rdma_set_service_type);
2580
2581 /**
2582  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2583  *                          with a connection identifier.
2584  * @id: Communication identifier to associated with service type.
2585  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2586  *
2587  * This function should be called before rdma_connect() on active side,
2588  * and on passive side before rdma_accept(). It is applicable to primary
2589  * path only. The timeout will affect the local side of the QP, it is not
2590  * negotiated with remote side and zero disables the timer. In case it is
2591  * set before rdma_resolve_route, the value will also be used to determine
2592  * PacketLifeTime for RoCE.
2593  *
2594  * Return: 0 for success
2595  */
2596 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2597 {
2598         struct rdma_id_private *id_priv;
2599
2600         if (id->qp_type != IB_QPT_RC)
2601                 return -EINVAL;
2602
2603         id_priv = container_of(id, struct rdma_id_private, id);
2604         mutex_lock(&id_priv->qp_mutex);
2605         id_priv->timeout = timeout;
2606         id_priv->timeout_set = true;
2607         mutex_unlock(&id_priv->qp_mutex);
2608
2609         return 0;
2610 }
2611 EXPORT_SYMBOL(rdma_set_ack_timeout);
2612
2613 /**
2614  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2615  *                            QP associated with a connection identifier.
2616  * @id: Communication identifier to associated with service type.
2617  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2618  *                 Timer Field" in the IBTA specification.
2619  *
2620  * This function should be called before rdma_connect() on active
2621  * side, and on passive side before rdma_accept(). The timer value
2622  * will be associated with the local QP. When it receives a send it is
2623  * not read to handle, typically if the receive queue is empty, an RNR
2624  * Retry NAK is returned to the requester with the min_rnr_timer
2625  * encoded. The requester will then wait at least the time specified
2626  * in the NAK before retrying. The default is zero, which translates
2627  * to a minimum RNR Timer value of 655 ms.
2628  *
2629  * Return: 0 for success
2630  */
2631 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2632 {
2633         struct rdma_id_private *id_priv;
2634
2635         /* It is a five-bit value */
2636         if (min_rnr_timer & 0xe0)
2637                 return -EINVAL;
2638
2639         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2640                 return -EINVAL;
2641
2642         id_priv = container_of(id, struct rdma_id_private, id);
2643         mutex_lock(&id_priv->qp_mutex);
2644         id_priv->min_rnr_timer = min_rnr_timer;
2645         id_priv->min_rnr_timer_set = true;
2646         mutex_unlock(&id_priv->qp_mutex);
2647
2648         return 0;
2649 }
2650 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2651
2652 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2653                               void *context)
2654 {
2655         struct cma_work *work = context;
2656         struct rdma_route *route;
2657
2658         route = &work->id->id.route;
2659
2660         if (!status) {
2661                 route->num_paths = 1;
2662                 *route->path_rec = *path_rec;
2663         } else {
2664                 work->old_state = RDMA_CM_ROUTE_QUERY;
2665                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2666                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2667                 work->event.status = status;
2668                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2669                                      status);
2670         }
2671
2672         queue_work(cma_wq, &work->work);
2673 }
2674
2675 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2676                               unsigned long timeout_ms, struct cma_work *work)
2677 {
2678         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2679         struct sa_path_rec path_rec;
2680         ib_sa_comp_mask comp_mask;
2681         struct sockaddr_in6 *sin6;
2682         struct sockaddr_ib *sib;
2683
2684         memset(&path_rec, 0, sizeof path_rec);
2685
2686         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2687                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2688         else
2689                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2690         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2691         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2692         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2693         path_rec.numb_path = 1;
2694         path_rec.reversible = 1;
2695         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2696                                                   cma_dst_addr(id_priv));
2697
2698         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2699                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2700                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2701
2702         switch (cma_family(id_priv)) {
2703         case AF_INET:
2704                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2705                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2706                 break;
2707         case AF_INET6:
2708                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2709                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2710                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2711                 break;
2712         case AF_IB:
2713                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2714                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2715                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2716                 break;
2717         }
2718
2719         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2720                                                id_priv->id.port_num, &path_rec,
2721                                                comp_mask, timeout_ms,
2722                                                GFP_KERNEL, cma_query_handler,
2723                                                work, &id_priv->query);
2724
2725         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2726 }
2727
2728 static void cma_iboe_join_work_handler(struct work_struct *work)
2729 {
2730         struct cma_multicast *mc =
2731                 container_of(work, struct cma_multicast, iboe_join.work);
2732         struct rdma_cm_event *event = &mc->iboe_join.event;
2733         struct rdma_id_private *id_priv = mc->id_priv;
2734         int ret;
2735
2736         mutex_lock(&id_priv->handler_mutex);
2737         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2738             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2739                 goto out_unlock;
2740
2741         ret = cma_cm_event_handler(id_priv, event);
2742         WARN_ON(ret);
2743
2744 out_unlock:
2745         mutex_unlock(&id_priv->handler_mutex);
2746         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2747                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2748 }
2749
2750 static void cma_work_handler(struct work_struct *_work)
2751 {
2752         struct cma_work *work = container_of(_work, struct cma_work, work);
2753         struct rdma_id_private *id_priv = work->id;
2754
2755         mutex_lock(&id_priv->handler_mutex);
2756         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2757             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2758                 goto out_unlock;
2759         if (work->old_state != 0 || work->new_state != 0) {
2760                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2761                         goto out_unlock;
2762         }
2763
2764         if (cma_cm_event_handler(id_priv, &work->event)) {
2765                 cma_id_put(id_priv);
2766                 destroy_id_handler_unlock(id_priv);
2767                 goto out_free;
2768         }
2769
2770 out_unlock:
2771         mutex_unlock(&id_priv->handler_mutex);
2772         cma_id_put(id_priv);
2773 out_free:
2774         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2775                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2776         kfree(work);
2777 }
2778
2779 static void cma_init_resolve_route_work(struct cma_work *work,
2780                                         struct rdma_id_private *id_priv)
2781 {
2782         work->id = id_priv;
2783         INIT_WORK(&work->work, cma_work_handler);
2784         work->old_state = RDMA_CM_ROUTE_QUERY;
2785         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2786         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2787 }
2788
2789 static void enqueue_resolve_addr_work(struct cma_work *work,
2790                                       struct rdma_id_private *id_priv)
2791 {
2792         /* Balances with cma_id_put() in cma_work_handler */
2793         cma_id_get(id_priv);
2794
2795         work->id = id_priv;
2796         INIT_WORK(&work->work, cma_work_handler);
2797         work->old_state = RDMA_CM_ADDR_QUERY;
2798         work->new_state = RDMA_CM_ADDR_RESOLVED;
2799         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2800
2801         queue_work(cma_wq, &work->work);
2802 }
2803
2804 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2805                                 unsigned long timeout_ms)
2806 {
2807         struct rdma_route *route = &id_priv->id.route;
2808         struct cma_work *work;
2809         int ret;
2810
2811         work = kzalloc(sizeof *work, GFP_KERNEL);
2812         if (!work)
2813                 return -ENOMEM;
2814
2815         cma_init_resolve_route_work(work, id_priv);
2816
2817         if (!route->path_rec)
2818                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2819         if (!route->path_rec) {
2820                 ret = -ENOMEM;
2821                 goto err1;
2822         }
2823
2824         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2825         if (ret)
2826                 goto err2;
2827
2828         return 0;
2829 err2:
2830         kfree(route->path_rec);
2831         route->path_rec = NULL;
2832 err1:
2833         kfree(work);
2834         return ret;
2835 }
2836
2837 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2838                                            unsigned long supported_gids,
2839                                            enum ib_gid_type default_gid)
2840 {
2841         if ((network_type == RDMA_NETWORK_IPV4 ||
2842              network_type == RDMA_NETWORK_IPV6) &&
2843             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2844                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2845
2846         return default_gid;
2847 }
2848
2849 /*
2850  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2851  * path record type based on GID type.
2852  * It also sets up other L2 fields which includes destination mac address
2853  * netdev ifindex, of the path record.
2854  * It returns the netdev of the bound interface for this path record entry.
2855  */
2856 static struct net_device *
2857 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2858 {
2859         struct rdma_route *route = &id_priv->id.route;
2860         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2861         struct rdma_addr *addr = &route->addr;
2862         unsigned long supported_gids;
2863         struct net_device *ndev;
2864
2865         if (!addr->dev_addr.bound_dev_if)
2866                 return NULL;
2867
2868         ndev = dev_get_by_index(addr->dev_addr.net,
2869                                 addr->dev_addr.bound_dev_if);
2870         if (!ndev)
2871                 return NULL;
2872
2873         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2874                                                     id_priv->id.port_num);
2875         gid_type = cma_route_gid_type(addr->dev_addr.network,
2876                                       supported_gids,
2877                                       id_priv->gid_type);
2878         /* Use the hint from IP Stack to select GID Type */
2879         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2880                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2881         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2882
2883         route->path_rec->roce.route_resolved = true;
2884         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2885         return ndev;
2886 }
2887
2888 int rdma_set_ib_path(struct rdma_cm_id *id,
2889                      struct sa_path_rec *path_rec)
2890 {
2891         struct rdma_id_private *id_priv;
2892         struct net_device *ndev;
2893         int ret;
2894
2895         id_priv = container_of(id, struct rdma_id_private, id);
2896         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2897                            RDMA_CM_ROUTE_RESOLVED))
2898                 return -EINVAL;
2899
2900         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2901                                      GFP_KERNEL);
2902         if (!id->route.path_rec) {
2903                 ret = -ENOMEM;
2904                 goto err;
2905         }
2906
2907         if (rdma_protocol_roce(id->device, id->port_num)) {
2908                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2909                 if (!ndev) {
2910                         ret = -ENODEV;
2911                         goto err_free;
2912                 }
2913                 dev_put(ndev);
2914         }
2915
2916         id->route.num_paths = 1;
2917         return 0;
2918
2919 err_free:
2920         kfree(id->route.path_rec);
2921         id->route.path_rec = NULL;
2922 err:
2923         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2924         return ret;
2925 }
2926 EXPORT_SYMBOL(rdma_set_ib_path);
2927
2928 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2929 {
2930         struct cma_work *work;
2931
2932         work = kzalloc(sizeof *work, GFP_KERNEL);
2933         if (!work)
2934                 return -ENOMEM;
2935
2936         cma_init_resolve_route_work(work, id_priv);
2937         queue_work(cma_wq, &work->work);
2938         return 0;
2939 }
2940
2941 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2942 {
2943         struct net_device *dev;
2944
2945         dev = vlan_dev_real_dev(vlan_ndev);
2946         if (dev->num_tc)
2947                 return netdev_get_prio_tc_map(dev, prio);
2948
2949         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2950                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2951 }
2952
2953 struct iboe_prio_tc_map {
2954         int input_prio;
2955         int output_tc;
2956         bool found;
2957 };
2958
2959 static int get_lower_vlan_dev_tc(struct net_device *dev,
2960                                  struct netdev_nested_priv *priv)
2961 {
2962         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2963
2964         if (is_vlan_dev(dev))
2965                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2966         else if (dev->num_tc)
2967                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2968         else
2969                 map->output_tc = 0;
2970         /* We are interested only in first level VLAN device, so always
2971          * return 1 to stop iterating over next level devices.
2972          */
2973         map->found = true;
2974         return 1;
2975 }
2976
2977 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2978 {
2979         struct iboe_prio_tc_map prio_tc_map = {};
2980         int prio = rt_tos2priority(tos);
2981         struct netdev_nested_priv priv;
2982
2983         /* If VLAN device, get it directly from the VLAN netdev */
2984         if (is_vlan_dev(ndev))
2985                 return get_vlan_ndev_tc(ndev, prio);
2986
2987         prio_tc_map.input_prio = prio;
2988         priv.data = (void *)&prio_tc_map;
2989         rcu_read_lock();
2990         netdev_walk_all_lower_dev_rcu(ndev,
2991                                       get_lower_vlan_dev_tc,
2992                                       &priv);
2993         rcu_read_unlock();
2994         /* If map is found from lower device, use it; Otherwise
2995          * continue with the current netdevice to get priority to tc map.
2996          */
2997         if (prio_tc_map.found)
2998                 return prio_tc_map.output_tc;
2999         else if (ndev->num_tc)
3000                 return netdev_get_prio_tc_map(ndev, prio);
3001         else
3002                 return 0;
3003 }
3004
3005 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3006 {
3007         struct sockaddr_in6 *addr6;
3008         u16 dport, sport;
3009         u32 hash, fl;
3010
3011         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3012         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3013         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3014                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3015                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3016                 hash = (u32)sport * 31 + dport;
3017                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3018         }
3019
3020         return cpu_to_be32(fl);
3021 }
3022
3023 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3024 {
3025         struct rdma_route *route = &id_priv->id.route;
3026         struct rdma_addr *addr = &route->addr;
3027         struct cma_work *work;
3028         int ret;
3029         struct net_device *ndev;
3030
3031         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3032                                         rdma_start_port(id_priv->cma_dev->device)];
3033         u8 tos;
3034
3035         mutex_lock(&id_priv->qp_mutex);
3036         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3037         mutex_unlock(&id_priv->qp_mutex);
3038
3039         work = kzalloc(sizeof *work, GFP_KERNEL);
3040         if (!work)
3041                 return -ENOMEM;
3042
3043         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3044         if (!route->path_rec) {
3045                 ret = -ENOMEM;
3046                 goto err1;
3047         }
3048
3049         route->num_paths = 1;
3050
3051         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3052         if (!ndev) {
3053                 ret = -ENODEV;
3054                 goto err2;
3055         }
3056
3057         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3058                     &route->path_rec->sgid);
3059         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3060                     &route->path_rec->dgid);
3061
3062         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3063                 /* TODO: get the hoplimit from the inet/inet6 device */
3064                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3065         else
3066                 route->path_rec->hop_limit = 1;
3067         route->path_rec->reversible = 1;
3068         route->path_rec->pkey = cpu_to_be16(0xffff);
3069         route->path_rec->mtu_selector = IB_SA_EQ;
3070         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3071         route->path_rec->traffic_class = tos;
3072         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3073         route->path_rec->rate_selector = IB_SA_EQ;
3074         route->path_rec->rate = iboe_get_rate(ndev);
3075         dev_put(ndev);
3076         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3077         /* In case ACK timeout is set, use this value to calculate
3078          * PacketLifeTime.  As per IBTA 12.7.34,
3079          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3080          * Assuming a negligible local ACK delay, we can use
3081          * PacketLifeTime = local ACK timeout/2
3082          * as a reasonable approximation for RoCE networks.
3083          */
3084         mutex_lock(&id_priv->qp_mutex);
3085         if (id_priv->timeout_set && id_priv->timeout)
3086                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3087         else
3088                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3089         mutex_unlock(&id_priv->qp_mutex);
3090
3091         if (!route->path_rec->mtu) {
3092                 ret = -EINVAL;
3093                 goto err2;
3094         }
3095
3096         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3097                                          id_priv->id.port_num))
3098                 route->path_rec->flow_label =
3099                         cma_get_roce_udp_flow_label(id_priv);
3100
3101         cma_init_resolve_route_work(work, id_priv);
3102         queue_work(cma_wq, &work->work);
3103
3104         return 0;
3105
3106 err2:
3107         kfree(route->path_rec);
3108         route->path_rec = NULL;
3109         route->num_paths = 0;
3110 err1:
3111         kfree(work);
3112         return ret;
3113 }
3114
3115 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3116 {
3117         struct rdma_id_private *id_priv;
3118         int ret;
3119
3120         id_priv = container_of(id, struct rdma_id_private, id);
3121         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3122                 return -EINVAL;
3123
3124         cma_id_get(id_priv);
3125         if (rdma_cap_ib_sa(id->device, id->port_num))
3126                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3127         else if (rdma_protocol_roce(id->device, id->port_num))
3128                 ret = cma_resolve_iboe_route(id_priv);
3129         else if (rdma_protocol_iwarp(id->device, id->port_num))
3130                 ret = cma_resolve_iw_route(id_priv);
3131         else
3132                 ret = -ENOSYS;
3133
3134         if (ret)
3135                 goto err;
3136
3137         return 0;
3138 err:
3139         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3140         cma_id_put(id_priv);
3141         return ret;
3142 }
3143 EXPORT_SYMBOL(rdma_resolve_route);
3144
3145 static void cma_set_loopback(struct sockaddr *addr)
3146 {
3147         switch (addr->sa_family) {
3148         case AF_INET:
3149                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3150                 break;
3151         case AF_INET6:
3152                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3153                               0, 0, 0, htonl(1));
3154                 break;
3155         default:
3156                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3157                             0, 0, 0, htonl(1));
3158                 break;
3159         }
3160 }
3161
3162 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3163 {
3164         struct cma_device *cma_dev, *cur_dev;
3165         union ib_gid gid;
3166         enum ib_port_state port_state;
3167         unsigned int p;
3168         u16 pkey;
3169         int ret;
3170
3171         cma_dev = NULL;
3172         mutex_lock(&lock);
3173         list_for_each_entry(cur_dev, &dev_list, list) {
3174                 if (cma_family(id_priv) == AF_IB &&
3175                     !rdma_cap_ib_cm(cur_dev->device, 1))
3176                         continue;
3177
3178                 if (!cma_dev)
3179                         cma_dev = cur_dev;
3180
3181                 rdma_for_each_port (cur_dev->device, p) {
3182                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3183                             port_state == IB_PORT_ACTIVE) {
3184                                 cma_dev = cur_dev;
3185                                 goto port_found;
3186                         }
3187                 }
3188         }
3189
3190         if (!cma_dev) {
3191                 ret = -ENODEV;
3192                 goto out;
3193         }
3194
3195         p = 1;
3196
3197 port_found:
3198         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3199         if (ret)
3200                 goto out;
3201
3202         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3203         if (ret)
3204                 goto out;
3205
3206         id_priv->id.route.addr.dev_addr.dev_type =
3207                 (rdma_protocol_ib(cma_dev->device, p)) ?
3208                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3209
3210         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3211         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3212         id_priv->id.port_num = p;
3213         cma_attach_to_dev(id_priv, cma_dev);
3214         rdma_restrack_add(&id_priv->res);
3215         cma_set_loopback(cma_src_addr(id_priv));
3216 out:
3217         mutex_unlock(&lock);
3218         return ret;
3219 }
3220
3221 static void addr_handler(int status, struct sockaddr *src_addr,
3222                          struct rdma_dev_addr *dev_addr, void *context)
3223 {
3224         struct rdma_id_private *id_priv = context;
3225         struct rdma_cm_event event = {};
3226         struct sockaddr *addr;
3227         struct sockaddr_storage old_addr;
3228
3229         mutex_lock(&id_priv->handler_mutex);
3230         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3231                            RDMA_CM_ADDR_RESOLVED))
3232                 goto out;
3233
3234         /*
3235          * Store the previous src address, so that if we fail to acquire
3236          * matching rdma device, old address can be restored back, which helps
3237          * to cancel the cma listen operation correctly.
3238          */
3239         addr = cma_src_addr(id_priv);
3240         memcpy(&old_addr, addr, rdma_addr_size(addr));
3241         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3242         if (!status && !id_priv->cma_dev) {
3243                 status = cma_acquire_dev_by_src_ip(id_priv);
3244                 if (status)
3245                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3246                                              status);
3247                 rdma_restrack_add(&id_priv->res);
3248         } else if (status) {
3249                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3250         }
3251
3252         if (status) {
3253                 memcpy(addr, &old_addr,
3254                        rdma_addr_size((struct sockaddr *)&old_addr));
3255                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3256                                    RDMA_CM_ADDR_BOUND))
3257                         goto out;
3258                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3259                 event.status = status;
3260         } else
3261                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3262
3263         if (cma_cm_event_handler(id_priv, &event)) {
3264                 destroy_id_handler_unlock(id_priv);
3265                 return;
3266         }
3267 out:
3268         mutex_unlock(&id_priv->handler_mutex);
3269 }
3270
3271 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3272 {
3273         struct cma_work *work;
3274         union ib_gid gid;
3275         int ret;
3276
3277         work = kzalloc(sizeof *work, GFP_KERNEL);
3278         if (!work)
3279                 return -ENOMEM;
3280
3281         if (!id_priv->cma_dev) {
3282                 ret = cma_bind_loopback(id_priv);
3283                 if (ret)
3284                         goto err;
3285         }
3286
3287         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3288         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3289
3290         enqueue_resolve_addr_work(work, id_priv);
3291         return 0;
3292 err:
3293         kfree(work);
3294         return ret;
3295 }
3296
3297 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3298 {
3299         struct cma_work *work;
3300         int ret;
3301
3302         work = kzalloc(sizeof *work, GFP_KERNEL);
3303         if (!work)
3304                 return -ENOMEM;
3305
3306         if (!id_priv->cma_dev) {
3307                 ret = cma_resolve_ib_dev(id_priv);
3308                 if (ret)
3309                         goto err;
3310         }
3311
3312         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3313                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3314
3315         enqueue_resolve_addr_work(work, id_priv);
3316         return 0;
3317 err:
3318         kfree(work);
3319         return ret;
3320 }
3321
3322 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3323                          const struct sockaddr *dst_addr)
3324 {
3325         if (!src_addr || !src_addr->sa_family) {
3326                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3327                 src_addr->sa_family = dst_addr->sa_family;
3328                 if (IS_ENABLED(CONFIG_IPV6) &&
3329                     dst_addr->sa_family == AF_INET6) {
3330                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3331                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3332                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3333                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3334                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3335                 } else if (dst_addr->sa_family == AF_IB) {
3336                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3337                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3338                 }
3339         }
3340         return rdma_bind_addr(id, src_addr);
3341 }
3342
3343 /*
3344  * If required, resolve the source address for bind and leave the id_priv in
3345  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3346  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3347  * ignored.
3348  */
3349 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3350                                struct sockaddr *src_addr,
3351                                const struct sockaddr *dst_addr)
3352 {
3353         int ret;
3354
3355         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3356         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3357                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3358                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3359                 if (ret)
3360                         goto err_dst;
3361                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3362                                            RDMA_CM_ADDR_QUERY))) {
3363                         ret = -EINVAL;
3364                         goto err_dst;
3365                 }
3366         }
3367
3368         if (cma_family(id_priv) != dst_addr->sa_family) {
3369                 ret = -EINVAL;
3370                 goto err_state;
3371         }
3372         return 0;
3373
3374 err_state:
3375         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3376 err_dst:
3377         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3378         return ret;
3379 }
3380
3381 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3382                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3383 {
3384         struct rdma_id_private *id_priv =
3385                 container_of(id, struct rdma_id_private, id);
3386         int ret;
3387
3388         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3389         if (ret)
3390                 return ret;
3391
3392         if (cma_any_addr(dst_addr)) {
3393                 ret = cma_resolve_loopback(id_priv);
3394         } else {
3395                 if (dst_addr->sa_family == AF_IB) {
3396                         ret = cma_resolve_ib_addr(id_priv);
3397                 } else {
3398                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3399                                               &id->route.addr.dev_addr,
3400                                               timeout_ms, addr_handler,
3401                                               false, id_priv);
3402                 }
3403         }
3404         if (ret)
3405                 goto err;
3406
3407         return 0;
3408 err:
3409         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3410         return ret;
3411 }
3412 EXPORT_SYMBOL(rdma_resolve_addr);
3413
3414 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3415 {
3416         struct rdma_id_private *id_priv;
3417         unsigned long flags;
3418         int ret;
3419
3420         id_priv = container_of(id, struct rdma_id_private, id);
3421         spin_lock_irqsave(&id_priv->lock, flags);
3422         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3423             id_priv->state == RDMA_CM_IDLE) {
3424                 id_priv->reuseaddr = reuse;
3425                 ret = 0;
3426         } else {
3427                 ret = -EINVAL;
3428         }
3429         spin_unlock_irqrestore(&id_priv->lock, flags);
3430         return ret;
3431 }
3432 EXPORT_SYMBOL(rdma_set_reuseaddr);
3433
3434 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3435 {
3436         struct rdma_id_private *id_priv;
3437         unsigned long flags;
3438         int ret;
3439
3440         id_priv = container_of(id, struct rdma_id_private, id);
3441         spin_lock_irqsave(&id_priv->lock, flags);
3442         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3443                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3444                 id_priv->afonly = afonly;
3445                 ret = 0;
3446         } else {
3447                 ret = -EINVAL;
3448         }
3449         spin_unlock_irqrestore(&id_priv->lock, flags);
3450         return ret;
3451 }
3452 EXPORT_SYMBOL(rdma_set_afonly);
3453
3454 static void cma_bind_port(struct rdma_bind_list *bind_list,
3455                           struct rdma_id_private *id_priv)
3456 {
3457         struct sockaddr *addr;
3458         struct sockaddr_ib *sib;
3459         u64 sid, mask;
3460         __be16 port;
3461
3462         lockdep_assert_held(&lock);
3463
3464         addr = cma_src_addr(id_priv);
3465         port = htons(bind_list->port);
3466
3467         switch (addr->sa_family) {
3468         case AF_INET:
3469                 ((struct sockaddr_in *) addr)->sin_port = port;
3470                 break;
3471         case AF_INET6:
3472                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3473                 break;
3474         case AF_IB:
3475                 sib = (struct sockaddr_ib *) addr;
3476                 sid = be64_to_cpu(sib->sib_sid);
3477                 mask = be64_to_cpu(sib->sib_sid_mask);
3478                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3479                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3480                 break;
3481         }
3482         id_priv->bind_list = bind_list;
3483         hlist_add_head(&id_priv->node, &bind_list->owners);
3484 }
3485
3486 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3487                           struct rdma_id_private *id_priv, unsigned short snum)
3488 {
3489         struct rdma_bind_list *bind_list;
3490         int ret;
3491
3492         lockdep_assert_held(&lock);
3493
3494         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3495         if (!bind_list)
3496                 return -ENOMEM;
3497
3498         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3499                            snum);
3500         if (ret < 0)
3501                 goto err;
3502
3503         bind_list->ps = ps;
3504         bind_list->port = snum;
3505         cma_bind_port(bind_list, id_priv);
3506         return 0;
3507 err:
3508         kfree(bind_list);
3509         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3510 }
3511
3512 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3513                               struct rdma_id_private *id_priv)
3514 {
3515         struct rdma_id_private *cur_id;
3516         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3517         struct sockaddr  *saddr = cma_src_addr(id_priv);
3518         __be16 dport = cma_port(daddr);
3519
3520         lockdep_assert_held(&lock);
3521
3522         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3523                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3524                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3525                 __be16 cur_dport = cma_port(cur_daddr);
3526
3527                 if (id_priv == cur_id)
3528                         continue;
3529
3530                 /* different dest port -> unique */
3531                 if (!cma_any_port(daddr) &&
3532                     !cma_any_port(cur_daddr) &&
3533                     (dport != cur_dport))
3534                         continue;
3535
3536                 /* different src address -> unique */
3537                 if (!cma_any_addr(saddr) &&
3538                     !cma_any_addr(cur_saddr) &&
3539                     cma_addr_cmp(saddr, cur_saddr))
3540                         continue;
3541
3542                 /* different dst address -> unique */
3543                 if (!cma_any_addr(daddr) &&
3544                     !cma_any_addr(cur_daddr) &&
3545                     cma_addr_cmp(daddr, cur_daddr))
3546                         continue;
3547
3548                 return -EADDRNOTAVAIL;
3549         }
3550         return 0;
3551 }
3552
3553 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3554                               struct rdma_id_private *id_priv)
3555 {
3556         static unsigned int last_used_port;
3557         int low, high, remaining;
3558         unsigned int rover;
3559         struct net *net = id_priv->id.route.addr.dev_addr.net;
3560
3561         lockdep_assert_held(&lock);
3562
3563         inet_get_local_port_range(net, &low, &high);
3564         remaining = (high - low) + 1;
3565         rover = prandom_u32() % remaining + low;
3566 retry:
3567         if (last_used_port != rover) {
3568                 struct rdma_bind_list *bind_list;
3569                 int ret;
3570
3571                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3572
3573                 if (!bind_list) {
3574                         ret = cma_alloc_port(ps, id_priv, rover);
3575                 } else {
3576                         ret = cma_port_is_unique(bind_list, id_priv);
3577                         if (!ret)
3578                                 cma_bind_port(bind_list, id_priv);
3579                 }
3580                 /*
3581                  * Remember previously used port number in order to avoid
3582                  * re-using same port immediately after it is closed.
3583                  */
3584                 if (!ret)
3585                         last_used_port = rover;
3586                 if (ret != -EADDRNOTAVAIL)
3587                         return ret;
3588         }
3589         if (--remaining) {
3590                 rover++;
3591                 if ((rover < low) || (rover > high))
3592                         rover = low;
3593                 goto retry;
3594         }
3595         return -EADDRNOTAVAIL;
3596 }
3597
3598 /*
3599  * Check that the requested port is available.  This is called when trying to
3600  * bind to a specific port, or when trying to listen on a bound port.  In
3601  * the latter case, the provided id_priv may already be on the bind_list, but
3602  * we still need to check that it's okay to start listening.
3603  */
3604 static int cma_check_port(struct rdma_bind_list *bind_list,
3605                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3606 {
3607         struct rdma_id_private *cur_id;
3608         struct sockaddr *addr, *cur_addr;
3609
3610         lockdep_assert_held(&lock);
3611
3612         addr = cma_src_addr(id_priv);
3613         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3614                 if (id_priv == cur_id)
3615                         continue;
3616
3617                 if (reuseaddr && cur_id->reuseaddr)
3618                         continue;
3619
3620                 cur_addr = cma_src_addr(cur_id);
3621                 if (id_priv->afonly && cur_id->afonly &&
3622                     (addr->sa_family != cur_addr->sa_family))
3623                         continue;
3624
3625                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3626                         return -EADDRNOTAVAIL;
3627
3628                 if (!cma_addr_cmp(addr, cur_addr))
3629                         return -EADDRINUSE;
3630         }
3631         return 0;
3632 }
3633
3634 static int cma_use_port(enum rdma_ucm_port_space ps,
3635                         struct rdma_id_private *id_priv)
3636 {
3637         struct rdma_bind_list *bind_list;
3638         unsigned short snum;
3639         int ret;
3640
3641         lockdep_assert_held(&lock);
3642
3643         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3644         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3645                 return -EACCES;
3646
3647         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3648         if (!bind_list) {
3649                 ret = cma_alloc_port(ps, id_priv, snum);
3650         } else {
3651                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3652                 if (!ret)
3653                         cma_bind_port(bind_list, id_priv);
3654         }
3655         return ret;
3656 }
3657
3658 static enum rdma_ucm_port_space
3659 cma_select_inet_ps(struct rdma_id_private *id_priv)
3660 {
3661         switch (id_priv->id.ps) {
3662         case RDMA_PS_TCP:
3663         case RDMA_PS_UDP:
3664         case RDMA_PS_IPOIB:
3665         case RDMA_PS_IB:
3666                 return id_priv->id.ps;
3667         default:
3668
3669                 return 0;
3670         }
3671 }
3672
3673 static enum rdma_ucm_port_space
3674 cma_select_ib_ps(struct rdma_id_private *id_priv)
3675 {
3676         enum rdma_ucm_port_space ps = 0;
3677         struct sockaddr_ib *sib;
3678         u64 sid_ps, mask, sid;
3679
3680         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3681         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3682         sid = be64_to_cpu(sib->sib_sid) & mask;
3683
3684         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3685                 sid_ps = RDMA_IB_IP_PS_IB;
3686                 ps = RDMA_PS_IB;
3687         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3688                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3689                 sid_ps = RDMA_IB_IP_PS_TCP;
3690                 ps = RDMA_PS_TCP;
3691         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3692                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3693                 sid_ps = RDMA_IB_IP_PS_UDP;
3694                 ps = RDMA_PS_UDP;
3695         }
3696
3697         if (ps) {
3698                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3699                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3700                                                 be64_to_cpu(sib->sib_sid_mask));
3701         }
3702         return ps;
3703 }
3704
3705 static int cma_get_port(struct rdma_id_private *id_priv)
3706 {
3707         enum rdma_ucm_port_space ps;
3708         int ret;
3709
3710         if (cma_family(id_priv) != AF_IB)
3711                 ps = cma_select_inet_ps(id_priv);
3712         else
3713                 ps = cma_select_ib_ps(id_priv);
3714         if (!ps)
3715                 return -EPROTONOSUPPORT;
3716
3717         mutex_lock(&lock);
3718         if (cma_any_port(cma_src_addr(id_priv)))
3719                 ret = cma_alloc_any_port(ps, id_priv);
3720         else
3721                 ret = cma_use_port(ps, id_priv);
3722         mutex_unlock(&lock);
3723
3724         return ret;
3725 }
3726
3727 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3728                                struct sockaddr *addr)
3729 {
3730 #if IS_ENABLED(CONFIG_IPV6)
3731         struct sockaddr_in6 *sin6;
3732
3733         if (addr->sa_family != AF_INET6)
3734                 return 0;
3735
3736         sin6 = (struct sockaddr_in6 *) addr;
3737
3738         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3739                 return 0;
3740
3741         if (!sin6->sin6_scope_id)
3742                         return -EINVAL;
3743
3744         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3745 #endif
3746         return 0;
3747 }
3748
3749 int rdma_listen(struct rdma_cm_id *id, int backlog)
3750 {
3751         struct rdma_id_private *id_priv =
3752                 container_of(id, struct rdma_id_private, id);
3753         int ret;
3754
3755         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3756                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3757                 id->route.addr.src_addr.ss_family = AF_INET;
3758                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3759                 if (ret)
3760                         return ret;
3761                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3762                                            RDMA_CM_LISTEN)))
3763                         return -EINVAL;
3764         }
3765
3766         /*
3767          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3768          * any more, and has to be unique in the bind list.
3769          */
3770         if (id_priv->reuseaddr) {
3771                 mutex_lock(&lock);
3772                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3773                 if (!ret)
3774                         id_priv->reuseaddr = 0;
3775                 mutex_unlock(&lock);
3776                 if (ret)
3777                         goto err;
3778         }
3779
3780         id_priv->backlog = backlog;
3781         if (id_priv->cma_dev) {
3782                 if (rdma_cap_ib_cm(id->device, 1)) {
3783                         ret = cma_ib_listen(id_priv);
3784                         if (ret)
3785                                 goto err;
3786                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3787                         ret = cma_iw_listen(id_priv, backlog);
3788                         if (ret)
3789                                 goto err;
3790                 } else {
3791                         ret = -ENOSYS;
3792                         goto err;
3793                 }
3794         } else {
3795                 ret = cma_listen_on_all(id_priv);
3796                 if (ret)
3797                         goto err;
3798         }
3799
3800         return 0;
3801 err:
3802         id_priv->backlog = 0;
3803         /*
3804          * All the failure paths that lead here will not allow the req_handler's
3805          * to have run.
3806          */
3807         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3808         return ret;
3809 }
3810 EXPORT_SYMBOL(rdma_listen);
3811
3812 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3813 {
3814         struct rdma_id_private *id_priv;
3815         int ret;
3816         struct sockaddr  *daddr;
3817
3818         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3819             addr->sa_family != AF_IB)
3820                 return -EAFNOSUPPORT;
3821
3822         id_priv = container_of(id, struct rdma_id_private, id);
3823         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3824                 return -EINVAL;
3825
3826         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3827         if (ret)
3828                 goto err1;
3829
3830         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3831         if (!cma_any_addr(addr)) {
3832                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3833                 if (ret)
3834                         goto err1;
3835
3836                 ret = cma_acquire_dev_by_src_ip(id_priv);
3837                 if (ret)
3838                         goto err1;
3839         }
3840
3841         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3842                 if (addr->sa_family == AF_INET)
3843                         id_priv->afonly = 1;
3844 #if IS_ENABLED(CONFIG_IPV6)
3845                 else if (addr->sa_family == AF_INET6) {
3846                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3847
3848                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3849                 }
3850 #endif
3851         }
3852         daddr = cma_dst_addr(id_priv);
3853         daddr->sa_family = addr->sa_family;
3854
3855         ret = cma_get_port(id_priv);
3856         if (ret)
3857                 goto err2;
3858
3859         if (!cma_any_addr(addr))
3860                 rdma_restrack_add(&id_priv->res);
3861         return 0;
3862 err2:
3863         if (id_priv->cma_dev)
3864                 cma_release_dev(id_priv);
3865 err1:
3866         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3867         return ret;
3868 }
3869 EXPORT_SYMBOL(rdma_bind_addr);
3870
3871 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3872 {
3873         struct cma_hdr *cma_hdr;
3874
3875         cma_hdr = hdr;
3876         cma_hdr->cma_version = CMA_VERSION;
3877         if (cma_family(id_priv) == AF_INET) {
3878                 struct sockaddr_in *src4, *dst4;
3879
3880                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3881                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3882
3883                 cma_set_ip_ver(cma_hdr, 4);
3884                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3885                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3886                 cma_hdr->port = src4->sin_port;
3887         } else if (cma_family(id_priv) == AF_INET6) {
3888                 struct sockaddr_in6 *src6, *dst6;
3889
3890                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3891                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3892
3893                 cma_set_ip_ver(cma_hdr, 6);
3894                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3895                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3896                 cma_hdr->port = src6->sin6_port;
3897         }
3898         return 0;
3899 }
3900
3901 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3902                                 const struct ib_cm_event *ib_event)
3903 {
3904         struct rdma_id_private *id_priv = cm_id->context;
3905         struct rdma_cm_event event = {};
3906         const struct ib_cm_sidr_rep_event_param *rep =
3907                                 &ib_event->param.sidr_rep_rcvd;
3908         int ret;
3909
3910         mutex_lock(&id_priv->handler_mutex);
3911         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3912                 goto out;
3913
3914         switch (ib_event->event) {
3915         case IB_CM_SIDR_REQ_ERROR:
3916                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3917                 event.status = -ETIMEDOUT;
3918                 break;
3919         case IB_CM_SIDR_REP_RECEIVED:
3920                 event.param.ud.private_data = ib_event->private_data;
3921                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3922                 if (rep->status != IB_SIDR_SUCCESS) {
3923                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3924                         event.status = ib_event->param.sidr_rep_rcvd.status;
3925                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3926                                              event.status);
3927                         break;
3928                 }
3929                 ret = cma_set_qkey(id_priv, rep->qkey);
3930                 if (ret) {
3931                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3932                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3933                         event.status = ret;
3934                         break;
3935                 }
3936                 ib_init_ah_attr_from_path(id_priv->id.device,
3937                                           id_priv->id.port_num,
3938                                           id_priv->id.route.path_rec,
3939                                           &event.param.ud.ah_attr,
3940                                           rep->sgid_attr);
3941                 event.param.ud.qp_num = rep->qpn;
3942                 event.param.ud.qkey = rep->qkey;
3943                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3944                 event.status = 0;
3945                 break;
3946         default:
3947                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3948                        ib_event->event);
3949                 goto out;
3950         }
3951
3952         ret = cma_cm_event_handler(id_priv, &event);
3953
3954         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3955         if (ret) {
3956                 /* Destroy the CM ID by returning a non-zero value. */
3957                 id_priv->cm_id.ib = NULL;
3958                 destroy_id_handler_unlock(id_priv);
3959                 return ret;
3960         }
3961 out:
3962         mutex_unlock(&id_priv->handler_mutex);
3963         return 0;
3964 }
3965
3966 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3967                               struct rdma_conn_param *conn_param)
3968 {
3969         struct ib_cm_sidr_req_param req;
3970         struct ib_cm_id *id;
3971         void *private_data;
3972         u8 offset;
3973         int ret;
3974
3975         memset(&req, 0, sizeof req);
3976         offset = cma_user_data_offset(id_priv);
3977         req.private_data_len = offset + conn_param->private_data_len;
3978         if (req.private_data_len < conn_param->private_data_len)
3979                 return -EINVAL;
3980
3981         if (req.private_data_len) {
3982                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3983                 if (!private_data)
3984                         return -ENOMEM;
3985         } else {
3986                 private_data = NULL;
3987         }
3988
3989         if (conn_param->private_data && conn_param->private_data_len)
3990                 memcpy(private_data + offset, conn_param->private_data,
3991                        conn_param->private_data_len);
3992
3993         if (private_data) {
3994                 ret = cma_format_hdr(private_data, id_priv);
3995                 if (ret)
3996                         goto out;
3997                 req.private_data = private_data;
3998         }
3999
4000         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4001                              id_priv);
4002         if (IS_ERR(id)) {
4003                 ret = PTR_ERR(id);
4004                 goto out;
4005         }
4006         id_priv->cm_id.ib = id;
4007
4008         req.path = id_priv->id.route.path_rec;
4009         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4010         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4011         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4012         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4013
4014         trace_cm_send_sidr_req(id_priv);
4015         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4016         if (ret) {
4017                 ib_destroy_cm_id(id_priv->cm_id.ib);
4018                 id_priv->cm_id.ib = NULL;
4019         }
4020 out:
4021         kfree(private_data);
4022         return ret;
4023 }
4024
4025 static int cma_connect_ib(struct rdma_id_private *id_priv,
4026                           struct rdma_conn_param *conn_param)
4027 {
4028         struct ib_cm_req_param req;
4029         struct rdma_route *route;
4030         void *private_data;
4031         struct ib_cm_id *id;
4032         u8 offset;
4033         int ret;
4034
4035         memset(&req, 0, sizeof req);
4036         offset = cma_user_data_offset(id_priv);
4037         req.private_data_len = offset + conn_param->private_data_len;
4038         if (req.private_data_len < conn_param->private_data_len)
4039                 return -EINVAL;
4040
4041         if (req.private_data_len) {
4042                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4043                 if (!private_data)
4044                         return -ENOMEM;
4045         } else {
4046                 private_data = NULL;
4047         }
4048
4049         if (conn_param->private_data && conn_param->private_data_len)
4050                 memcpy(private_data + offset, conn_param->private_data,
4051                        conn_param->private_data_len);
4052
4053         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4054         if (IS_ERR(id)) {
4055                 ret = PTR_ERR(id);
4056                 goto out;
4057         }
4058         id_priv->cm_id.ib = id;
4059
4060         route = &id_priv->id.route;
4061         if (private_data) {
4062                 ret = cma_format_hdr(private_data, id_priv);
4063                 if (ret)
4064                         goto out;
4065                 req.private_data = private_data;
4066         }
4067
4068         req.primary_path = &route->path_rec[0];
4069         if (route->num_paths == 2)
4070                 req.alternate_path = &route->path_rec[1];
4071
4072         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4073         /* Alternate path SGID attribute currently unsupported */
4074         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4075         req.qp_num = id_priv->qp_num;
4076         req.qp_type = id_priv->id.qp_type;
4077         req.starting_psn = id_priv->seq_num;
4078         req.responder_resources = conn_param->responder_resources;
4079         req.initiator_depth = conn_param->initiator_depth;
4080         req.flow_control = conn_param->flow_control;
4081         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4082         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4083         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4084         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4085         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4086         req.srq = id_priv->srq ? 1 : 0;
4087         req.ece.vendor_id = id_priv->ece.vendor_id;
4088         req.ece.attr_mod = id_priv->ece.attr_mod;
4089
4090         trace_cm_send_req(id_priv);
4091         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4092 out:
4093         if (ret && !IS_ERR(id)) {
4094                 ib_destroy_cm_id(id);
4095                 id_priv->cm_id.ib = NULL;
4096         }
4097
4098         kfree(private_data);
4099         return ret;
4100 }
4101
4102 static int cma_connect_iw(struct rdma_id_private *id_priv,
4103                           struct rdma_conn_param *conn_param)
4104 {
4105         struct iw_cm_id *cm_id;
4106         int ret;
4107         struct iw_cm_conn_param iw_param;
4108
4109         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4110         if (IS_ERR(cm_id))
4111                 return PTR_ERR(cm_id);
4112
4113         mutex_lock(&id_priv->qp_mutex);
4114         cm_id->tos = id_priv->tos;
4115         cm_id->tos_set = id_priv->tos_set;
4116         mutex_unlock(&id_priv->qp_mutex);
4117
4118         id_priv->cm_id.iw = cm_id;
4119
4120         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4121                rdma_addr_size(cma_src_addr(id_priv)));
4122         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4123                rdma_addr_size(cma_dst_addr(id_priv)));
4124
4125         ret = cma_modify_qp_rtr(id_priv, conn_param);
4126         if (ret)
4127                 goto out;
4128
4129         if (conn_param) {
4130                 iw_param.ord = conn_param->initiator_depth;
4131                 iw_param.ird = conn_param->responder_resources;
4132                 iw_param.private_data = conn_param->private_data;
4133                 iw_param.private_data_len = conn_param->private_data_len;
4134                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4135         } else {
4136                 memset(&iw_param, 0, sizeof iw_param);
4137                 iw_param.qpn = id_priv->qp_num;
4138         }
4139         ret = iw_cm_connect(cm_id, &iw_param);
4140 out:
4141         if (ret) {
4142                 iw_destroy_cm_id(cm_id);
4143                 id_priv->cm_id.iw = NULL;
4144         }
4145         return ret;
4146 }
4147
4148 /**
4149  * rdma_connect_locked - Initiate an active connection request.
4150  * @id: Connection identifier to connect.
4151  * @conn_param: Connection information used for connected QPs.
4152  *
4153  * Same as rdma_connect() but can only be called from the
4154  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4155  */
4156 int rdma_connect_locked(struct rdma_cm_id *id,
4157                         struct rdma_conn_param *conn_param)
4158 {
4159         struct rdma_id_private *id_priv =
4160                 container_of(id, struct rdma_id_private, id);
4161         int ret;
4162
4163         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4164                 return -EINVAL;
4165
4166         if (!id->qp) {
4167                 id_priv->qp_num = conn_param->qp_num;
4168                 id_priv->srq = conn_param->srq;
4169         }
4170
4171         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4172                 if (id->qp_type == IB_QPT_UD)
4173                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4174                 else
4175                         ret = cma_connect_ib(id_priv, conn_param);
4176         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4177                 ret = cma_connect_iw(id_priv, conn_param);
4178         } else {
4179                 ret = -ENOSYS;
4180         }
4181         if (ret)
4182                 goto err_state;
4183         return 0;
4184 err_state:
4185         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4186         return ret;
4187 }
4188 EXPORT_SYMBOL(rdma_connect_locked);
4189
4190 /**
4191  * rdma_connect - Initiate an active connection request.
4192  * @id: Connection identifier to connect.
4193  * @conn_param: Connection information used for connected QPs.
4194  *
4195  * Users must have resolved a route for the rdma_cm_id to connect with by having
4196  * called rdma_resolve_route before calling this routine.
4197  *
4198  * This call will either connect to a remote QP or obtain remote QP information
4199  * for unconnected rdma_cm_id's.  The actual operation is based on the
4200  * rdma_cm_id's port space.
4201  */
4202 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4203 {
4204         struct rdma_id_private *id_priv =
4205                 container_of(id, struct rdma_id_private, id);
4206         int ret;
4207
4208         mutex_lock(&id_priv->handler_mutex);
4209         ret = rdma_connect_locked(id, conn_param);
4210         mutex_unlock(&id_priv->handler_mutex);
4211         return ret;
4212 }
4213 EXPORT_SYMBOL(rdma_connect);
4214
4215 /**
4216  * rdma_connect_ece - Initiate an active connection request with ECE data.
4217  * @id: Connection identifier to connect.
4218  * @conn_param: Connection information used for connected QPs.
4219  * @ece: ECE parameters
4220  *
4221  * See rdma_connect() explanation.
4222  */
4223 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4224                      struct rdma_ucm_ece *ece)
4225 {
4226         struct rdma_id_private *id_priv =
4227                 container_of(id, struct rdma_id_private, id);
4228
4229         id_priv->ece.vendor_id = ece->vendor_id;
4230         id_priv->ece.attr_mod = ece->attr_mod;
4231
4232         return rdma_connect(id, conn_param);
4233 }
4234 EXPORT_SYMBOL(rdma_connect_ece);
4235
4236 static int cma_accept_ib(struct rdma_id_private *id_priv,
4237                          struct rdma_conn_param *conn_param)
4238 {
4239         struct ib_cm_rep_param rep;
4240         int ret;
4241
4242         ret = cma_modify_qp_rtr(id_priv, conn_param);
4243         if (ret)
4244                 goto out;
4245
4246         ret = cma_modify_qp_rts(id_priv, conn_param);
4247         if (ret)
4248                 goto out;
4249
4250         memset(&rep, 0, sizeof rep);
4251         rep.qp_num = id_priv->qp_num;
4252         rep.starting_psn = id_priv->seq_num;
4253         rep.private_data = conn_param->private_data;
4254         rep.private_data_len = conn_param->private_data_len;
4255         rep.responder_resources = conn_param->responder_resources;
4256         rep.initiator_depth = conn_param->initiator_depth;
4257         rep.failover_accepted = 0;
4258         rep.flow_control = conn_param->flow_control;
4259         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4260         rep.srq = id_priv->srq ? 1 : 0;
4261         rep.ece.vendor_id = id_priv->ece.vendor_id;
4262         rep.ece.attr_mod = id_priv->ece.attr_mod;
4263
4264         trace_cm_send_rep(id_priv);
4265         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4266 out:
4267         return ret;
4268 }
4269
4270 static int cma_accept_iw(struct rdma_id_private *id_priv,
4271                   struct rdma_conn_param *conn_param)
4272 {
4273         struct iw_cm_conn_param iw_param;
4274         int ret;
4275
4276         if (!conn_param)
4277                 return -EINVAL;
4278
4279         ret = cma_modify_qp_rtr(id_priv, conn_param);
4280         if (ret)
4281                 return ret;
4282
4283         iw_param.ord = conn_param->initiator_depth;
4284         iw_param.ird = conn_param->responder_resources;
4285         iw_param.private_data = conn_param->private_data;
4286         iw_param.private_data_len = conn_param->private_data_len;
4287         if (id_priv->id.qp)
4288                 iw_param.qpn = id_priv->qp_num;
4289         else
4290                 iw_param.qpn = conn_param->qp_num;
4291
4292         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4293 }
4294
4295 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4296                              enum ib_cm_sidr_status status, u32 qkey,
4297                              const void *private_data, int private_data_len)
4298 {
4299         struct ib_cm_sidr_rep_param rep;
4300         int ret;
4301
4302         memset(&rep, 0, sizeof rep);
4303         rep.status = status;
4304         if (status == IB_SIDR_SUCCESS) {
4305                 ret = cma_set_qkey(id_priv, qkey);
4306                 if (ret)
4307                         return ret;
4308                 rep.qp_num = id_priv->qp_num;
4309                 rep.qkey = id_priv->qkey;
4310
4311                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4312                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4313         }
4314
4315         rep.private_data = private_data;
4316         rep.private_data_len = private_data_len;
4317
4318         trace_cm_send_sidr_rep(id_priv);
4319         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4320 }
4321
4322 /**
4323  * rdma_accept - Called to accept a connection request or response.
4324  * @id: Connection identifier associated with the request.
4325  * @conn_param: Information needed to establish the connection.  This must be
4326  *   provided if accepting a connection request.  If accepting a connection
4327  *   response, this parameter must be NULL.
4328  *
4329  * Typically, this routine is only called by the listener to accept a connection
4330  * request.  It must also be called on the active side of a connection if the
4331  * user is performing their own QP transitions.
4332  *
4333  * In the case of error, a reject message is sent to the remote side and the
4334  * state of the qp associated with the id is modified to error, such that any
4335  * previously posted receive buffers would be flushed.
4336  *
4337  * This function is for use by kernel ULPs and must be called from under the
4338  * handler callback.
4339  */
4340 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4341 {
4342         struct rdma_id_private *id_priv =
4343                 container_of(id, struct rdma_id_private, id);
4344         int ret;
4345
4346         lockdep_assert_held(&id_priv->handler_mutex);
4347
4348         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4349                 return -EINVAL;
4350
4351         if (!id->qp && conn_param) {
4352                 id_priv->qp_num = conn_param->qp_num;
4353                 id_priv->srq = conn_param->srq;
4354         }
4355
4356         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4357                 if (id->qp_type == IB_QPT_UD) {
4358                         if (conn_param)
4359                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4360                                                         conn_param->qkey,
4361                                                         conn_param->private_data,
4362                                                         conn_param->private_data_len);
4363                         else
4364                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4365                                                         0, NULL, 0);
4366                 } else {
4367                         if (conn_param)
4368                                 ret = cma_accept_ib(id_priv, conn_param);
4369                         else
4370                                 ret = cma_rep_recv(id_priv);
4371                 }
4372         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4373                 ret = cma_accept_iw(id_priv, conn_param);
4374         } else {
4375                 ret = -ENOSYS;
4376         }
4377         if (ret)
4378                 goto reject;
4379
4380         return 0;
4381 reject:
4382         cma_modify_qp_err(id_priv);
4383         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4384         return ret;
4385 }
4386 EXPORT_SYMBOL(rdma_accept);
4387
4388 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4389                     struct rdma_ucm_ece *ece)
4390 {
4391         struct rdma_id_private *id_priv =
4392                 container_of(id, struct rdma_id_private, id);
4393
4394         id_priv->ece.vendor_id = ece->vendor_id;
4395         id_priv->ece.attr_mod = ece->attr_mod;
4396
4397         return rdma_accept(id, conn_param);
4398 }
4399 EXPORT_SYMBOL(rdma_accept_ece);
4400
4401 void rdma_lock_handler(struct rdma_cm_id *id)
4402 {
4403         struct rdma_id_private *id_priv =
4404                 container_of(id, struct rdma_id_private, id);
4405
4406         mutex_lock(&id_priv->handler_mutex);
4407 }
4408 EXPORT_SYMBOL(rdma_lock_handler);
4409
4410 void rdma_unlock_handler(struct rdma_cm_id *id)
4411 {
4412         struct rdma_id_private *id_priv =
4413                 container_of(id, struct rdma_id_private, id);
4414
4415         mutex_unlock(&id_priv->handler_mutex);
4416 }
4417 EXPORT_SYMBOL(rdma_unlock_handler);
4418
4419 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4420 {
4421         struct rdma_id_private *id_priv;
4422         int ret;
4423
4424         id_priv = container_of(id, struct rdma_id_private, id);
4425         if (!id_priv->cm_id.ib)
4426                 return -EINVAL;
4427
4428         switch (id->device->node_type) {
4429         case RDMA_NODE_IB_CA:
4430                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4431                 break;
4432         default:
4433                 ret = 0;
4434                 break;
4435         }
4436         return ret;
4437 }
4438 EXPORT_SYMBOL(rdma_notify);
4439
4440 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4441                 u8 private_data_len, u8 reason)
4442 {
4443         struct rdma_id_private *id_priv;
4444         int ret;
4445
4446         id_priv = container_of(id, struct rdma_id_private, id);
4447         if (!id_priv->cm_id.ib)
4448                 return -EINVAL;
4449
4450         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4451                 if (id->qp_type == IB_QPT_UD) {
4452                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4453                                                 private_data, private_data_len);
4454                 } else {
4455                         trace_cm_send_rej(id_priv);
4456                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4457                                              private_data, private_data_len);
4458                 }
4459         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4460                 ret = iw_cm_reject(id_priv->cm_id.iw,
4461                                    private_data, private_data_len);
4462         } else {
4463                 ret = -ENOSYS;
4464         }
4465
4466         return ret;
4467 }
4468 EXPORT_SYMBOL(rdma_reject);
4469
4470 int rdma_disconnect(struct rdma_cm_id *id)
4471 {
4472         struct rdma_id_private *id_priv;
4473         int ret;
4474
4475         id_priv = container_of(id, struct rdma_id_private, id);
4476         if (!id_priv->cm_id.ib)
4477                 return -EINVAL;
4478
4479         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4480                 ret = cma_modify_qp_err(id_priv);
4481                 if (ret)
4482                         goto out;
4483                 /* Initiate or respond to a disconnect. */
4484                 trace_cm_disconnect(id_priv);
4485                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4486                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4487                                 trace_cm_sent_drep(id_priv);
4488                 } else {
4489                         trace_cm_sent_dreq(id_priv);
4490                 }
4491         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4492                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4493         } else
4494                 ret = -EINVAL;
4495
4496 out:
4497         return ret;
4498 }
4499 EXPORT_SYMBOL(rdma_disconnect);
4500
4501 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4502                               struct ib_sa_multicast *multicast,
4503                               struct rdma_cm_event *event,
4504                               struct cma_multicast *mc)
4505 {
4506         struct rdma_dev_addr *dev_addr;
4507         enum ib_gid_type gid_type;
4508         struct net_device *ndev;
4509
4510         if (!status)
4511                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4512         else
4513                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4514                                      status);
4515
4516         event->status = status;
4517         event->param.ud.private_data = mc->context;
4518         if (status) {
4519                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4520                 return;
4521         }
4522
4523         dev_addr = &id_priv->id.route.addr.dev_addr;
4524         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4525         gid_type =
4526                 id_priv->cma_dev
4527                         ->default_gid_type[id_priv->id.port_num -
4528                                            rdma_start_port(
4529                                                    id_priv->cma_dev->device)];
4530
4531         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4532         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4533                                      &multicast->rec, ndev, gid_type,
4534                                      &event->param.ud.ah_attr)) {
4535                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4536                 goto out;
4537         }
4538
4539         event->param.ud.qp_num = 0xFFFFFF;
4540         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4541
4542 out:
4543         if (ndev)
4544                 dev_put(ndev);
4545 }
4546
4547 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4548 {
4549         struct cma_multicast *mc = multicast->context;
4550         struct rdma_id_private *id_priv = mc->id_priv;
4551         struct rdma_cm_event event = {};
4552         int ret = 0;
4553
4554         mutex_lock(&id_priv->handler_mutex);
4555         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4556             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4557                 goto out;
4558
4559         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4560         ret = cma_cm_event_handler(id_priv, &event);
4561         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4562         WARN_ON(ret);
4563
4564 out:
4565         mutex_unlock(&id_priv->handler_mutex);
4566         return 0;
4567 }
4568
4569 static void cma_set_mgid(struct rdma_id_private *id_priv,
4570                          struct sockaddr *addr, union ib_gid *mgid)
4571 {
4572         unsigned char mc_map[MAX_ADDR_LEN];
4573         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4574         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4575         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4576
4577         if (cma_any_addr(addr)) {
4578                 memset(mgid, 0, sizeof *mgid);
4579         } else if ((addr->sa_family == AF_INET6) &&
4580                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4581                                                                  0xFF10A01B)) {
4582                 /* IPv6 address is an SA assigned MGID. */
4583                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4584         } else if (addr->sa_family == AF_IB) {
4585                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4586         } else if (addr->sa_family == AF_INET6) {
4587                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4588                 if (id_priv->id.ps == RDMA_PS_UDP)
4589                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4590                 *mgid = *(union ib_gid *) (mc_map + 4);
4591         } else {
4592                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4593                 if (id_priv->id.ps == RDMA_PS_UDP)
4594                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4595                 *mgid = *(union ib_gid *) (mc_map + 4);
4596         }
4597 }
4598
4599 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4600                                  struct cma_multicast *mc)
4601 {
4602         struct ib_sa_mcmember_rec rec;
4603         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4604         ib_sa_comp_mask comp_mask;
4605         int ret;
4606
4607         ib_addr_get_mgid(dev_addr, &rec.mgid);
4608         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4609                                      &rec.mgid, &rec);
4610         if (ret)
4611                 return ret;
4612
4613         ret = cma_set_qkey(id_priv, 0);
4614         if (ret)
4615                 return ret;
4616
4617         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4618         rec.qkey = cpu_to_be32(id_priv->qkey);
4619         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4620         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4621         rec.join_state = mc->join_state;
4622
4623         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4624                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4625                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4626                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4627                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4628
4629         if (id_priv->id.ps == RDMA_PS_IPOIB)
4630                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4631                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4632                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4633                              IB_SA_MCMEMBER_REC_MTU |
4634                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4635
4636         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4637                                          id_priv->id.port_num, &rec, comp_mask,
4638                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4639         return PTR_ERR_OR_ZERO(mc->sa_mc);
4640 }
4641
4642 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4643                               enum ib_gid_type gid_type)
4644 {
4645         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4646         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4647
4648         if (cma_any_addr(addr)) {
4649                 memset(mgid, 0, sizeof *mgid);
4650         } else if (addr->sa_family == AF_INET6) {
4651                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4652         } else {
4653                 mgid->raw[0] =
4654                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4655                 mgid->raw[1] =
4656                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4657                 mgid->raw[2] = 0;
4658                 mgid->raw[3] = 0;
4659                 mgid->raw[4] = 0;
4660                 mgid->raw[5] = 0;
4661                 mgid->raw[6] = 0;
4662                 mgid->raw[7] = 0;
4663                 mgid->raw[8] = 0;
4664                 mgid->raw[9] = 0;
4665                 mgid->raw[10] = 0xff;
4666                 mgid->raw[11] = 0xff;
4667                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4668         }
4669 }
4670
4671 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4672                                    struct cma_multicast *mc)
4673 {
4674         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4675         int err = 0;
4676         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4677         struct net_device *ndev = NULL;
4678         struct ib_sa_multicast ib;
4679         enum ib_gid_type gid_type;
4680         bool send_only;
4681
4682         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4683
4684         if (cma_zero_addr(addr))
4685                 return -EINVAL;
4686
4687         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4688                    rdma_start_port(id_priv->cma_dev->device)];
4689         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4690
4691         ib.rec.pkey = cpu_to_be16(0xffff);
4692         if (id_priv->id.ps == RDMA_PS_UDP)
4693                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4694
4695         if (dev_addr->bound_dev_if)
4696                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4697         if (!ndev)
4698                 return -ENODEV;
4699
4700         ib.rec.rate = iboe_get_rate(ndev);
4701         ib.rec.hop_limit = 1;
4702         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4703
4704         if (addr->sa_family == AF_INET) {
4705                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4706                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4707                         if (!send_only) {
4708                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4709                                                     true);
4710                         }
4711                 }
4712         } else {
4713                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4714                         err = -ENOTSUPP;
4715         }
4716         dev_put(ndev);
4717         if (err || !ib.rec.mtu)
4718                 return err ?: -EINVAL;
4719
4720         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4721                     &ib.rec.port_gid);
4722         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4723         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4724         queue_work(cma_wq, &mc->iboe_join.work);
4725         return 0;
4726 }
4727
4728 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4729                         u8 join_state, void *context)
4730 {
4731         struct rdma_id_private *id_priv =
4732                 container_of(id, struct rdma_id_private, id);
4733         struct cma_multicast *mc;
4734         int ret;
4735
4736         /* Not supported for kernel QPs */
4737         if (WARN_ON(id->qp))
4738                 return -EINVAL;
4739
4740         /* ULP is calling this wrong. */
4741         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4742                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4743                 return -EINVAL;
4744
4745         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4746         if (!mc)
4747                 return -ENOMEM;
4748
4749         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4750         mc->context = context;
4751         mc->id_priv = id_priv;
4752         mc->join_state = join_state;
4753
4754         if (rdma_protocol_roce(id->device, id->port_num)) {
4755                 ret = cma_iboe_join_multicast(id_priv, mc);
4756                 if (ret)
4757                         goto out_err;
4758         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4759                 ret = cma_join_ib_multicast(id_priv, mc);
4760                 if (ret)
4761                         goto out_err;
4762         } else {
4763                 ret = -ENOSYS;
4764                 goto out_err;
4765         }
4766
4767         spin_lock(&id_priv->lock);
4768         list_add(&mc->list, &id_priv->mc_list);
4769         spin_unlock(&id_priv->lock);
4770
4771         return 0;
4772 out_err:
4773         kfree(mc);
4774         return ret;
4775 }
4776 EXPORT_SYMBOL(rdma_join_multicast);
4777
4778 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4779 {
4780         struct rdma_id_private *id_priv;
4781         struct cma_multicast *mc;
4782
4783         id_priv = container_of(id, struct rdma_id_private, id);
4784         spin_lock_irq(&id_priv->lock);
4785         list_for_each_entry(mc, &id_priv->mc_list, list) {
4786                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4787                         continue;
4788                 list_del(&mc->list);
4789                 spin_unlock_irq(&id_priv->lock);
4790
4791                 WARN_ON(id_priv->cma_dev->device != id->device);
4792                 destroy_mc(id_priv, mc);
4793                 return;
4794         }
4795         spin_unlock_irq(&id_priv->lock);
4796 }
4797 EXPORT_SYMBOL(rdma_leave_multicast);
4798
4799 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4800 {
4801         struct rdma_dev_addr *dev_addr;
4802         struct cma_work *work;
4803
4804         dev_addr = &id_priv->id.route.addr.dev_addr;
4805
4806         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4807             (net_eq(dev_net(ndev), dev_addr->net)) &&
4808             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4809                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4810                         ndev->name, &id_priv->id);
4811                 work = kzalloc(sizeof *work, GFP_KERNEL);
4812                 if (!work)
4813                         return -ENOMEM;
4814
4815                 INIT_WORK(&work->work, cma_work_handler);
4816                 work->id = id_priv;
4817                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4818                 cma_id_get(id_priv);
4819                 queue_work(cma_wq, &work->work);
4820         }
4821
4822         return 0;
4823 }
4824
4825 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4826                                void *ptr)
4827 {
4828         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4829         struct cma_device *cma_dev;
4830         struct rdma_id_private *id_priv;
4831         int ret = NOTIFY_DONE;
4832
4833         if (event != NETDEV_BONDING_FAILOVER)
4834                 return NOTIFY_DONE;
4835
4836         if (!netif_is_bond_master(ndev))
4837                 return NOTIFY_DONE;
4838
4839         mutex_lock(&lock);
4840         list_for_each_entry(cma_dev, &dev_list, list)
4841                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4842                         ret = cma_netdev_change(ndev, id_priv);
4843                         if (ret)
4844                                 goto out;
4845                 }
4846
4847 out:
4848         mutex_unlock(&lock);
4849         return ret;
4850 }
4851
4852 static struct notifier_block cma_nb = {
4853         .notifier_call = cma_netdev_callback
4854 };
4855
4856 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4857 {
4858         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4859         enum rdma_cm_state state;
4860         unsigned long flags;
4861
4862         mutex_lock(&id_priv->handler_mutex);
4863         /* Record that we want to remove the device */
4864         spin_lock_irqsave(&id_priv->lock, flags);
4865         state = id_priv->state;
4866         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4867                 spin_unlock_irqrestore(&id_priv->lock, flags);
4868                 mutex_unlock(&id_priv->handler_mutex);
4869                 cma_id_put(id_priv);
4870                 return;
4871         }
4872         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4873         spin_unlock_irqrestore(&id_priv->lock, flags);
4874
4875         if (cma_cm_event_handler(id_priv, &event)) {
4876                 /*
4877                  * At this point the ULP promises it won't call
4878                  * rdma_destroy_id() concurrently
4879                  */
4880                 cma_id_put(id_priv);
4881                 mutex_unlock(&id_priv->handler_mutex);
4882                 trace_cm_id_destroy(id_priv);
4883                 _destroy_id(id_priv, state);
4884                 return;
4885         }
4886         mutex_unlock(&id_priv->handler_mutex);
4887
4888         /*
4889          * If this races with destroy then the thread that first assigns state
4890          * to a destroying does the cancel.
4891          */
4892         cma_cancel_operation(id_priv, state);
4893         cma_id_put(id_priv);
4894 }
4895
4896 static void cma_process_remove(struct cma_device *cma_dev)
4897 {
4898         mutex_lock(&lock);
4899         while (!list_empty(&cma_dev->id_list)) {
4900                 struct rdma_id_private *id_priv = list_first_entry(
4901                         &cma_dev->id_list, struct rdma_id_private, list);
4902
4903                 list_del(&id_priv->listen_list);
4904                 list_del_init(&id_priv->list);
4905                 cma_id_get(id_priv);
4906                 mutex_unlock(&lock);
4907
4908                 cma_send_device_removal_put(id_priv);
4909
4910                 mutex_lock(&lock);
4911         }
4912         mutex_unlock(&lock);
4913
4914         cma_dev_put(cma_dev);
4915         wait_for_completion(&cma_dev->comp);
4916 }
4917
4918 static bool cma_supported(struct ib_device *device)
4919 {
4920         u32 i;
4921
4922         rdma_for_each_port(device, i) {
4923                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
4924                         return true;
4925         }
4926         return false;
4927 }
4928
4929 static int cma_add_one(struct ib_device *device)
4930 {
4931         struct rdma_id_private *to_destroy;
4932         struct cma_device *cma_dev;
4933         struct rdma_id_private *id_priv;
4934         unsigned long supported_gids = 0;
4935         int ret;
4936         u32 i;
4937
4938         if (!cma_supported(device))
4939                 return -EOPNOTSUPP;
4940
4941         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4942         if (!cma_dev)
4943                 return -ENOMEM;
4944
4945         cma_dev->device = device;
4946         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4947                                             sizeof(*cma_dev->default_gid_type),
4948                                             GFP_KERNEL);
4949         if (!cma_dev->default_gid_type) {
4950                 ret = -ENOMEM;
4951                 goto free_cma_dev;
4952         }
4953
4954         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4955                                             sizeof(*cma_dev->default_roce_tos),
4956                                             GFP_KERNEL);
4957         if (!cma_dev->default_roce_tos) {
4958                 ret = -ENOMEM;
4959                 goto free_gid_type;
4960         }
4961
4962         rdma_for_each_port (device, i) {
4963                 supported_gids = roce_gid_type_mask_support(device, i);
4964                 WARN_ON(!supported_gids);
4965                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4966                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4967                                 CMA_PREFERRED_ROCE_GID_TYPE;
4968                 else
4969                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4970                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4971                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4972         }
4973
4974         init_completion(&cma_dev->comp);
4975         refcount_set(&cma_dev->refcount, 1);
4976         INIT_LIST_HEAD(&cma_dev->id_list);
4977         ib_set_client_data(device, &cma_client, cma_dev);
4978
4979         mutex_lock(&lock);
4980         list_add_tail(&cma_dev->list, &dev_list);
4981         list_for_each_entry(id_priv, &listen_any_list, list) {
4982                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
4983                 if (ret)
4984                         goto free_listen;
4985         }
4986         mutex_unlock(&lock);
4987
4988         trace_cm_add_one(device);
4989         return 0;
4990
4991 free_listen:
4992         list_del(&cma_dev->list);
4993         mutex_unlock(&lock);
4994
4995         /* cma_process_remove() will delete to_destroy */
4996         cma_process_remove(cma_dev);
4997         kfree(cma_dev->default_roce_tos);
4998 free_gid_type:
4999         kfree(cma_dev->default_gid_type);
5000
5001 free_cma_dev:
5002         kfree(cma_dev);
5003         return ret;
5004 }
5005
5006 static void cma_remove_one(struct ib_device *device, void *client_data)
5007 {
5008         struct cma_device *cma_dev = client_data;
5009
5010         trace_cm_remove_one(device);
5011
5012         mutex_lock(&lock);
5013         list_del(&cma_dev->list);
5014         mutex_unlock(&lock);
5015
5016         cma_process_remove(cma_dev);
5017         kfree(cma_dev->default_roce_tos);
5018         kfree(cma_dev->default_gid_type);
5019         kfree(cma_dev);
5020 }
5021
5022 static int cma_init_net(struct net *net)
5023 {
5024         struct cma_pernet *pernet = cma_pernet(net);
5025
5026         xa_init(&pernet->tcp_ps);
5027         xa_init(&pernet->udp_ps);
5028         xa_init(&pernet->ipoib_ps);
5029         xa_init(&pernet->ib_ps);
5030
5031         return 0;
5032 }
5033
5034 static void cma_exit_net(struct net *net)
5035 {
5036         struct cma_pernet *pernet = cma_pernet(net);
5037
5038         WARN_ON(!xa_empty(&pernet->tcp_ps));
5039         WARN_ON(!xa_empty(&pernet->udp_ps));
5040         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5041         WARN_ON(!xa_empty(&pernet->ib_ps));
5042 }
5043
5044 static struct pernet_operations cma_pernet_operations = {
5045         .init = cma_init_net,
5046         .exit = cma_exit_net,
5047         .id = &cma_pernet_id,
5048         .size = sizeof(struct cma_pernet),
5049 };
5050
5051 static int __init cma_init(void)
5052 {
5053         int ret;
5054
5055         /*
5056          * There is a rare lock ordering dependency in cma_netdev_callback()
5057          * that only happens when bonding is enabled. Teach lockdep that rtnl
5058          * must never be nested under lock so it can find these without having
5059          * to test with bonding.
5060          */
5061         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5062                 rtnl_lock();
5063                 mutex_lock(&lock);
5064                 mutex_unlock(&lock);
5065                 rtnl_unlock();
5066         }
5067
5068         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5069         if (!cma_wq)
5070                 return -ENOMEM;
5071
5072         ret = register_pernet_subsys(&cma_pernet_operations);
5073         if (ret)
5074                 goto err_wq;
5075
5076         ib_sa_register_client(&sa_client);
5077         register_netdevice_notifier(&cma_nb);
5078
5079         ret = ib_register_client(&cma_client);
5080         if (ret)
5081                 goto err;
5082
5083         ret = cma_configfs_init();
5084         if (ret)
5085                 goto err_ib;
5086
5087         return 0;
5088
5089 err_ib:
5090         ib_unregister_client(&cma_client);
5091 err:
5092         unregister_netdevice_notifier(&cma_nb);
5093         ib_sa_unregister_client(&sa_client);
5094         unregister_pernet_subsys(&cma_pernet_operations);
5095 err_wq:
5096         destroy_workqueue(cma_wq);
5097         return ret;
5098 }
5099
5100 static void __exit cma_cleanup(void)
5101 {
5102         cma_configfs_exit();
5103         ib_unregister_client(&cma_client);
5104         unregister_netdevice_notifier(&cma_nb);
5105         ib_sa_unregister_client(&sa_client);
5106         unregister_pernet_subsys(&cma_pernet_operations);
5107         destroy_workqueue(cma_wq);
5108 }
5109
5110 module_init(cma_init);
5111 module_exit(cma_cleanup);