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