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_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  47 #define CMA_MAX_CM_RETRIES 15
  48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  49 #define CMA_IBOE_PACKET_LIFETIME 18
  50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  51
  52 static const char * const cma_events[] = {
  53         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  54         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  55         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  56         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  57         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  58         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  59         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  60         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  61         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  62         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  63         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  64         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  65         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  66         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  67         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  68         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  69 };
  70
  71 static void cma_set_mgid(struct rdma_id_private *id_priv, struct sockaddr *addr,
  72                          union ib_gid *mgid);
  73
  74 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  75 {
  76         size_t index = event;
  77
  78         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  79                         cma_events[index] : "unrecognized event";
  80 }
  81 EXPORT_SYMBOL(rdma_event_msg);
  82
  83 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  84                                                 int reason)
  85 {
  86         if (rdma_ib_or_roce(id->device, id->port_num))
  87                 return ibcm_reject_msg(reason);
  88
  89         if (rdma_protocol_iwarp(id->device, id->port_num))
  90                 return iwcm_reject_msg(reason);
  91
  92         WARN_ON_ONCE(1);
  93         return "unrecognized transport";
  94 }
  95 EXPORT_SYMBOL(rdma_reject_msg);
  96
  97 /**
  98  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  99  *                           request.
 100  * @id: Communication identifier that received the REJECT event.
 101  * @reason: Value returned in the REJECT event status field.
 102  */
 103 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 104 {
 105         if (rdma_ib_or_roce(id->device, id->port_num))
 106                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 107
 108         if (rdma_protocol_iwarp(id->device, id->port_num))
 109                 return reason == -ECONNREFUSED;
 110
 111         WARN_ON_ONCE(1);
 112         return false;
 113 }
 114
 115 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 116                                       struct rdma_cm_event *ev, u8 *data_len)
 117 {
 118         const void *p;
 119
 120         if (rdma_is_consumer_reject(id, ev->status)) {
 121                 *data_len = ev->param.conn.private_data_len;
 122                 p = ev->param.conn.private_data;
 123         } else {
 124                 *data_len = 0;
 125                 p = NULL;
 126         }
 127         return p;
 128 }
 129 EXPORT_SYMBOL(rdma_consumer_reject_data);
 130
 131 /**
 132  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 133  * @id: Communication Identifier
 134  */
 135 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 136 {
 137         struct rdma_id_private *id_priv;
 138
 139         id_priv = container_of(id, struct rdma_id_private, id);
 140         if (id->device->node_type == RDMA_NODE_RNIC)
 141                 return id_priv->cm_id.iw;
 142         return NULL;
 143 }
 144 EXPORT_SYMBOL(rdma_iw_cm_id);
 145
 146 /**
 147  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 148  * @res: rdma resource tracking entry pointer
 149  */
 150 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 151 {
 152         struct rdma_id_private *id_priv =
 153                 container_of(res, struct rdma_id_private, res);
 154
 155         return &id_priv->id;
 156 }
 157 EXPORT_SYMBOL(rdma_res_to_id);
 158
 159 static int cma_add_one(struct ib_device *device);
 160 static void cma_remove_one(struct ib_device *device, void *client_data);
 161
 162 static struct ib_client cma_client = {
 163         .name   = "cma",
 164         .add    = cma_add_one,
 165         .remove = cma_remove_one
 166 };
 167
 168 static struct ib_sa_client sa_client;
 169 static LIST_HEAD(dev_list);
 170 static LIST_HEAD(listen_any_list);
 171 static DEFINE_MUTEX(lock);
 172 static struct workqueue_struct *cma_wq;
 173 static unsigned int cma_pernet_id;
 174
 175 struct cma_pernet {
 176         struct xarray tcp_ps;
 177         struct xarray udp_ps;
 178         struct xarray ipoib_ps;
 179         struct xarray ib_ps;
 180 };
 181
 182 static struct cma_pernet *cma_pernet(struct net *net)
 183 {
 184         return net_generic(net, cma_pernet_id);
 185 }
 186
 187 static
 188 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 189 {
 190         struct cma_pernet *pernet = cma_pernet(net);
 191
 192         switch (ps) {
 193         case RDMA_PS_TCP:
 194                 return &pernet->tcp_ps;
 195         case RDMA_PS_UDP:
 196                 return &pernet->udp_ps;
 197         case RDMA_PS_IPOIB:
 198                 return &pernet->ipoib_ps;
 199         case RDMA_PS_IB:
 200                 return &pernet->ib_ps;
 201         default:
 202                 return NULL;
 203         }
 204 }
 205
 206 struct cma_device {
 207         struct list_head        list;
 208         struct ib_device        *device;
 209         struct completion       comp;
 210         refcount_t refcount;
 211         struct list_head        id_list;
 212         enum ib_gid_type        *default_gid_type;
 213         u8                      *default_roce_tos;
 214 };
 215
 216 struct rdma_bind_list {
 217         enum rdma_ucm_port_space ps;
 218         struct hlist_head       owners;
 219         unsigned short          port;
 220 };
 221
 222 struct class_port_info_context {
 223         struct ib_class_port_info       *class_port_info;
 224         struct ib_device                *device;
 225         struct completion               done;
 226         struct ib_sa_query              *sa_query;
 227         u8                              port_num;
 228 };
 229
 230 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 231                         struct rdma_bind_list *bind_list, int snum)
 232 {
 233         struct xarray *xa = cma_pernet_xa(net, ps);
 234
 235         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 236 }
 237
 238 static struct rdma_bind_list *cma_ps_find(struct net *net,
 239                                           enum rdma_ucm_port_space ps, int snum)
 240 {
 241         struct xarray *xa = cma_pernet_xa(net, ps);
 242
 243         return xa_load(xa, snum);
 244 }
 245
 246 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 247                           int snum)
 248 {
 249         struct xarray *xa = cma_pernet_xa(net, ps);
 250
 251         xa_erase(xa, snum);
 252 }
 253
 254 enum {
 255         CMA_OPTION_AFONLY,
 256 };
 257
 258 void cma_dev_get(struct cma_device *cma_dev)
 259 {
 260         refcount_inc(&cma_dev->refcount);
 261 }
 262
 263 void cma_dev_put(struct cma_device *cma_dev)
 264 {
 265         if (refcount_dec_and_test(&cma_dev->refcount))
 266                 complete(&cma_dev->comp);
 267 }
 268
 269 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 270                                              void               *cookie)
 271 {
 272         struct cma_device *cma_dev;
 273         struct cma_device *found_cma_dev = NULL;
 274
 275         mutex_lock(&lock);
 276
 277         list_for_each_entry(cma_dev, &dev_list, list)
 278                 if (filter(cma_dev->device, cookie)) {
 279                         found_cma_dev = cma_dev;
 280                         break;
 281                 }
 282
 283         if (found_cma_dev)
 284                 cma_dev_get(found_cma_dev);
 285         mutex_unlock(&lock);
 286         return found_cma_dev;
 287 }
 288
 289 int cma_get_default_gid_type(struct cma_device *cma_dev,
 290                              unsigned int port)
 291 {
 292         if (!rdma_is_port_valid(cma_dev->device, port))
 293                 return -EINVAL;
 294
 295         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 296 }
 297
 298 int cma_set_default_gid_type(struct cma_device *cma_dev,
 299                              unsigned int port,
 300                              enum ib_gid_type default_gid_type)
 301 {
 302         unsigned long supported_gids;
 303
 304         if (!rdma_is_port_valid(cma_dev->device, port))
 305                 return -EINVAL;
 306
 307         if (default_gid_type == IB_GID_TYPE_IB &&
 308             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 309                 default_gid_type = IB_GID_TYPE_ROCE;
 310
 311         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 312
 313         if (!(supported_gids & 1 << default_gid_type))
 314                 return -EINVAL;
 315
 316         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 317                 default_gid_type;
 318
 319         return 0;
 320 }
 321
 322 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
 323 {
 324         if (!rdma_is_port_valid(cma_dev->device, port))
 325                 return -EINVAL;
 326
 327         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 328 }
 329
 330 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
 331                              u8 default_roce_tos)
 332 {
 333         if (!rdma_is_port_valid(cma_dev->device, port))
 334                 return -EINVAL;
 335
 336         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 337                  default_roce_tos;
 338
 339         return 0;
 340 }
 341 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 342 {
 343         return cma_dev->device;
 344 }
 345
 346 /*
 347  * Device removal can occur at anytime, so we need extra handling to
 348  * serialize notifying the user of device removal with other callbacks.
 349  * We do this by disabling removal notification while a callback is in process,
 350  * and reporting it after the callback completes.
 351  */
 352
 353 struct cma_multicast {
 354         struct rdma_id_private *id_priv;
 355         union {
 356                 struct ib_sa_multicast *sa_mc;
 357                 struct {
 358                         struct work_struct work;
 359                         struct rdma_cm_event event;
 360                 } iboe_join;
 361         };
 362         struct list_head        list;
 363         void                    *context;
 364         struct sockaddr_storage addr;
 365         u8                      join_state;
 366 };
 367
 368 struct cma_work {
 369         struct work_struct      work;
 370         struct rdma_id_private  *id;
 371         enum rdma_cm_state      old_state;
 372         enum rdma_cm_state      new_state;
 373         struct rdma_cm_event    event;
 374 };
 375
 376 union cma_ip_addr {
 377         struct in6_addr ip6;
 378         struct {
 379                 __be32 pad[3];
 380                 __be32 addr;
 381         } ip4;
 382 };
 383
 384 struct cma_hdr {
 385         u8 cma_version;
 386         u8 ip_version;  /* IP version: 7:4 */
 387         __be16 port;
 388         union cma_ip_addr src_addr;
 389         union cma_ip_addr dst_addr;
 390 };
 391
 392 #define CMA_VERSION 0x00
 393
 394 struct cma_req_info {
 395         struct sockaddr_storage listen_addr_storage;
 396         struct sockaddr_storage src_addr_storage;
 397         struct ib_device *device;
 398         union ib_gid local_gid;
 399         __be64 service_id;
 400         int port;
 401         bool has_gid;
 402         u16 pkey;
 403 };
 404
 405 static int cma_comp_exch(struct rdma_id_private *id_priv,
 406                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 407 {
 408         unsigned long flags;
 409         int ret;
 410
 411         /*
 412          * The FSM uses a funny double locking where state is protected by both
 413          * the handler_mutex and the spinlock. State is not allowed to change
 414          * to/from a handler_mutex protected value without also holding
 415          * handler_mutex.
 416          */
 417         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 418                 lockdep_assert_held(&id_priv->handler_mutex);
 419
 420         spin_lock_irqsave(&id_priv->lock, flags);
 421         if ((ret = (id_priv->state == comp)))
 422                 id_priv->state = exch;
 423         spin_unlock_irqrestore(&id_priv->lock, flags);
 424         return ret;
 425 }
 426
 427 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 428 {
 429         return hdr->ip_version >> 4;
 430 }
 431
 432 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 433 {
 434         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 435 }
 436
 437 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 438 {
 439         struct in_device *in_dev = NULL;
 440
 441         if (ndev) {
 442                 rtnl_lock();
 443                 in_dev = __in_dev_get_rtnl(ndev);
 444                 if (in_dev) {
 445                         if (join)
 446                                 ip_mc_inc_group(in_dev,
 447                                                 *(__be32 *)(mgid->raw + 12));
 448                         else
 449                                 ip_mc_dec_group(in_dev,
 450                                                 *(__be32 *)(mgid->raw + 12));
 451                 }
 452                 rtnl_unlock();
 453         }
 454         return (in_dev) ? 0 : -ENODEV;
 455 }
 456
 457 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 458                                struct cma_device *cma_dev)
 459 {
 460         cma_dev_get(cma_dev);
 461         id_priv->cma_dev = cma_dev;
 462         id_priv->id.device = cma_dev->device;
 463         id_priv->id.route.addr.dev_addr.transport =
 464                 rdma_node_get_transport(cma_dev->device->node_type);
 465         list_add_tail(&id_priv->list, &cma_dev->id_list);
 466         rdma_restrack_add(&id_priv->res);
 467
 468         trace_cm_id_attach(id_priv, cma_dev->device);
 469 }
 470
 471 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 472                               struct cma_device *cma_dev)
 473 {
 474         _cma_attach_to_dev(id_priv, cma_dev);
 475         id_priv->gid_type =
 476                 cma_dev->default_gid_type[id_priv->id.port_num -
 477                                           rdma_start_port(cma_dev->device)];
 478 }
 479
 480 static void cma_release_dev(struct rdma_id_private *id_priv)
 481 {
 482         mutex_lock(&lock);
 483         list_del(&id_priv->list);
 484         cma_dev_put(id_priv->cma_dev);
 485         id_priv->cma_dev = NULL;
 486         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 487                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 488                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 489         }
 490         mutex_unlock(&lock);
 491 }
 492
 493 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 494 {
 495         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 496 }
 497
 498 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 499 {
 500         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 501 }
 502
 503 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 504 {
 505         return id_priv->id.route.addr.src_addr.ss_family;
 506 }
 507
 508 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 509 {
 510         struct ib_sa_mcmember_rec rec;
 511         int ret = 0;
 512
 513         if (id_priv->qkey) {
 514                 if (qkey && id_priv->qkey != qkey)
 515                         return -EINVAL;
 516                 return 0;
 517         }
 518
 519         if (qkey) {
 520                 id_priv->qkey = qkey;
 521                 return 0;
 522         }
 523
 524         switch (id_priv->id.ps) {
 525         case RDMA_PS_UDP:
 526         case RDMA_PS_IB:
 527                 id_priv->qkey = RDMA_UDP_QKEY;
 528                 break;
 529         case RDMA_PS_IPOIB:
 530                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 531                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 532                                              id_priv->id.port_num, &rec.mgid,
 533                                              &rec);
 534                 if (!ret)
 535                         id_priv->qkey = be32_to_cpu(rec.qkey);
 536                 break;
 537         default:
 538                 break;
 539         }
 540         return ret;
 541 }
 542
 543 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 544 {
 545         dev_addr->dev_type = ARPHRD_INFINIBAND;
 546         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 547         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 548 }
 549
 550 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 551 {
 552         int ret;
 553
 554         if (addr->sa_family != AF_IB) {
 555                 ret = rdma_translate_ip(addr, dev_addr);
 556         } else {
 557                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 558                 ret = 0;
 559         }
 560
 561         return ret;
 562 }
 563
 564 static const struct ib_gid_attr *
 565 cma_validate_port(struct ib_device *device, u8 port,
 566                   enum ib_gid_type gid_type,
 567                   union ib_gid *gid,
 568                   struct rdma_id_private *id_priv)
 569 {
 570         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 571         int bound_if_index = dev_addr->bound_dev_if;
 572         const struct ib_gid_attr *sgid_attr;
 573         int dev_type = dev_addr->dev_type;
 574         struct net_device *ndev = NULL;
 575
 576         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 577                 return ERR_PTR(-ENODEV);
 578
 579         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 580                 return ERR_PTR(-ENODEV);
 581
 582         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 583                 return ERR_PTR(-ENODEV);
 584
 585         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 586                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 587                 if (!ndev)
 588                         return ERR_PTR(-ENODEV);
 589         } else {
 590                 gid_type = IB_GID_TYPE_IB;
 591         }
 592
 593         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 594         if (ndev)
 595                 dev_put(ndev);
 596         return sgid_attr;
 597 }
 598
 599 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 600                                const struct ib_gid_attr *sgid_attr)
 601 {
 602         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 603         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 604 }
 605
 606 /**
 607  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 608  * based on source ip address.
 609  * @id_priv:    cm_id which should be bound to cma device
 610  *
 611  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 612  * based on source IP address. It returns 0 on success or error code otherwise.
 613  * It is applicable to active and passive side cm_id.
 614  */
 615 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 616 {
 617         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 618         const struct ib_gid_attr *sgid_attr;
 619         union ib_gid gid, iboe_gid, *gidp;
 620         struct cma_device *cma_dev;
 621         enum ib_gid_type gid_type;
 622         int ret = -ENODEV;
 623         unsigned int port;
 624
 625         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 626             id_priv->id.ps == RDMA_PS_IPOIB)
 627                 return -EINVAL;
 628
 629         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 630                     &iboe_gid);
 631
 632         memcpy(&gid, dev_addr->src_dev_addr +
 633                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 634
 635         mutex_lock(&lock);
 636         list_for_each_entry(cma_dev, &dev_list, list) {
 637                 rdma_for_each_port (cma_dev->device, port) {
 638                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 639                                &iboe_gid : &gid;
 640                         gid_type = cma_dev->default_gid_type[port - 1];
 641                         sgid_attr = cma_validate_port(cma_dev->device, port,
 642                                                       gid_type, gidp, id_priv);
 643                         if (!IS_ERR(sgid_attr)) {
 644                                 id_priv->id.port_num = port;
 645                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 646                                 cma_attach_to_dev(id_priv, cma_dev);
 647                                 ret = 0;
 648                                 goto out;
 649                         }
 650                 }
 651         }
 652 out:
 653         mutex_unlock(&lock);
 654         return ret;
 655 }
 656
 657 /**
 658  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 659  * @id_priv:            cm id to bind to cma device
 660  * @listen_id_priv:     listener cm id to match against
 661  * @req:                Pointer to req structure containaining incoming
 662  *                      request information
 663  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 664  * rdma device matches for listen_id and incoming request. It also verifies
 665  * that a GID table entry is present for the source address.
 666  * Returns 0 on success, or returns error code otherwise.
 667  */
 668 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 669                               const struct rdma_id_private *listen_id_priv,
 670                               struct cma_req_info *req)
 671 {
 672         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 673         const struct ib_gid_attr *sgid_attr;
 674         enum ib_gid_type gid_type;
 675         union ib_gid gid;
 676
 677         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 678             id_priv->id.ps == RDMA_PS_IPOIB)
 679                 return -EINVAL;
 680
 681         if (rdma_protocol_roce(req->device, req->port))
 682                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 683                             &gid);
 684         else
 685                 memcpy(&gid, dev_addr->src_dev_addr +
 686                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 687
 688         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 689         sgid_attr = cma_validate_port(req->device, req->port,
 690                                       gid_type, &gid, id_priv);
 691         if (IS_ERR(sgid_attr))
 692                 return PTR_ERR(sgid_attr);
 693
 694         id_priv->id.port_num = req->port;
 695         cma_bind_sgid_attr(id_priv, sgid_attr);
 696         /* Need to acquire lock to protect against reader
 697          * of cma_dev->id_list such as cma_netdev_callback() and
 698          * cma_process_remove().
 699          */
 700         mutex_lock(&lock);
 701         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 702         mutex_unlock(&lock);
 703         return 0;
 704 }
 705
 706 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 707                               const struct rdma_id_private *listen_id_priv)
 708 {
 709         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 710         const struct ib_gid_attr *sgid_attr;
 711         struct cma_device *cma_dev;
 712         enum ib_gid_type gid_type;
 713         int ret = -ENODEV;
 714         unsigned int port;
 715         union ib_gid gid;
 716
 717         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 718             id_priv->id.ps == RDMA_PS_IPOIB)
 719                 return -EINVAL;
 720
 721         memcpy(&gid, dev_addr->src_dev_addr +
 722                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 723
 724         mutex_lock(&lock);
 725
 726         cma_dev = listen_id_priv->cma_dev;
 727         port = listen_id_priv->id.port_num;
 728         gid_type = listen_id_priv->gid_type;
 729         sgid_attr = cma_validate_port(cma_dev->device, port,
 730                                       gid_type, &gid, id_priv);
 731         if (!IS_ERR(sgid_attr)) {
 732                 id_priv->id.port_num = port;
 733                 cma_bind_sgid_attr(id_priv, sgid_attr);
 734                 ret = 0;
 735                 goto out;
 736         }
 737
 738         list_for_each_entry(cma_dev, &dev_list, list) {
 739                 rdma_for_each_port (cma_dev->device, port) {
 740                         if (listen_id_priv->cma_dev == cma_dev &&
 741                             listen_id_priv->id.port_num == port)
 742                                 continue;
 743
 744                         gid_type = cma_dev->default_gid_type[port - 1];
 745                         sgid_attr = cma_validate_port(cma_dev->device, port,
 746                                                       gid_type, &gid, id_priv);
 747                         if (!IS_ERR(sgid_attr)) {
 748                                 id_priv->id.port_num = port;
 749                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 750                                 ret = 0;
 751                                 goto out;
 752                         }
 753                 }
 754         }
 755
 756 out:
 757         if (!ret)
 758                 cma_attach_to_dev(id_priv, cma_dev);
 759
 760         mutex_unlock(&lock);
 761         return ret;
 762 }
 763
 764 /*
 765  * Select the source IB device and address to reach the destination IB address.
 766  */
 767 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 768 {
 769         struct cma_device *cma_dev, *cur_dev;
 770         struct sockaddr_ib *addr;
 771         union ib_gid gid, sgid, *dgid;
 772         unsigned int p;
 773         u16 pkey, index;
 774         enum ib_port_state port_state;
 775         int i;
 776
 777         cma_dev = NULL;
 778         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 779         dgid = (union ib_gid *) &addr->sib_addr;
 780         pkey = ntohs(addr->sib_pkey);
 781
 782         mutex_lock(&lock);
 783         list_for_each_entry(cur_dev, &dev_list, list) {
 784                 rdma_for_each_port (cur_dev->device, p) {
 785                         if (!rdma_cap_af_ib(cur_dev->device, p))
 786                                 continue;
 787
 788                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 789                                 continue;
 790
 791                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 792                                 continue;
 793                         for (i = 0; !rdma_query_gid(cur_dev->device,
 794                                                     p, i, &gid);
 795                              i++) {
 796                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 797                                         cma_dev = cur_dev;
 798                                         sgid = gid;
 799                                         id_priv->id.port_num = p;
 800                                         goto found;
 801                                 }
 802
 803                                 if (!cma_dev && (gid.global.subnet_prefix ==
 804                                     dgid->global.subnet_prefix) &&
 805                                     port_state == IB_PORT_ACTIVE) {
 806                                         cma_dev = cur_dev;
 807                                         sgid = gid;
 808                                         id_priv->id.port_num = p;
 809                                         goto found;
 810                                 }
 811                         }
 812                 }
 813         }
 814         mutex_unlock(&lock);
 815         return -ENODEV;
 816
 817 found:
 818         cma_attach_to_dev(id_priv, cma_dev);
 819         mutex_unlock(&lock);
 820         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 821         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 822         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 823         return 0;
 824 }
 825
 826 static void cma_id_get(struct rdma_id_private *id_priv)
 827 {
 828         refcount_inc(&id_priv->refcount);
 829 }
 830
 831 static void cma_id_put(struct rdma_id_private *id_priv)
 832 {
 833         if (refcount_dec_and_test(&id_priv->refcount))
 834                 complete(&id_priv->comp);
 835 }
 836
 837 static struct rdma_id_private *
 838 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 839                  void *context, enum rdma_ucm_port_space ps,
 840                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 841 {
 842         struct rdma_id_private *id_priv;
 843
 844         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 845         if (!id_priv)
 846                 return ERR_PTR(-ENOMEM);
 847
 848         id_priv->state = RDMA_CM_IDLE;
 849         id_priv->id.context = context;
 850         id_priv->id.event_handler = event_handler;
 851         id_priv->id.ps = ps;
 852         id_priv->id.qp_type = qp_type;
 853         id_priv->tos_set = false;
 854         id_priv->timeout_set = false;
 855         id_priv->gid_type = IB_GID_TYPE_IB;
 856         spin_lock_init(&id_priv->lock);
 857         mutex_init(&id_priv->qp_mutex);
 858         init_completion(&id_priv->comp);
 859         refcount_set(&id_priv->refcount, 1);
 860         mutex_init(&id_priv->handler_mutex);
 861         INIT_LIST_HEAD(&id_priv->listen_list);
 862         INIT_LIST_HEAD(&id_priv->mc_list);
 863         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 864         id_priv->id.route.addr.dev_addr.net = get_net(net);
 865         id_priv->seq_num &= 0x00ffffff;
 866
 867         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 868         if (parent)
 869                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 870
 871         return id_priv;
 872 }
 873
 874 struct rdma_cm_id *
 875 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
 876                         void *context, enum rdma_ucm_port_space ps,
 877                         enum ib_qp_type qp_type, const char *caller)
 878 {
 879         struct rdma_id_private *ret;
 880
 881         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
 882         if (IS_ERR(ret))
 883                 return ERR_CAST(ret);
 884
 885         rdma_restrack_set_name(&ret->res, caller);
 886         return &ret->id;
 887 }
 888 EXPORT_SYMBOL(__rdma_create_kernel_id);
 889
 890 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
 891                                        void *context,
 892                                        enum rdma_ucm_port_space ps,
 893                                        enum ib_qp_type qp_type)
 894 {
 895         struct rdma_id_private *ret;
 896
 897         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
 898                                ps, qp_type, NULL);
 899         if (IS_ERR(ret))
 900                 return ERR_CAST(ret);
 901
 902         rdma_restrack_set_name(&ret->res, NULL);
 903         return &ret->id;
 904 }
 905 EXPORT_SYMBOL(rdma_create_user_id);
 906
 907 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 908 {
 909         struct ib_qp_attr qp_attr;
 910         int qp_attr_mask, ret;
 911
 912         qp_attr.qp_state = IB_QPS_INIT;
 913         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 914         if (ret)
 915                 return ret;
 916
 917         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 918         if (ret)
 919                 return ret;
 920
 921         qp_attr.qp_state = IB_QPS_RTR;
 922         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 923         if (ret)
 924                 return ret;
 925
 926         qp_attr.qp_state = IB_QPS_RTS;
 927         qp_attr.sq_psn = 0;
 928         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 929
 930         return ret;
 931 }
 932
 933 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 934 {
 935         struct ib_qp_attr qp_attr;
 936         int qp_attr_mask, ret;
 937
 938         qp_attr.qp_state = IB_QPS_INIT;
 939         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 940         if (ret)
 941                 return ret;
 942
 943         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 944 }
 945
 946 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 947                    struct ib_qp_init_attr *qp_init_attr)
 948 {
 949         struct rdma_id_private *id_priv;
 950         struct ib_qp *qp;
 951         int ret;
 952
 953         id_priv = container_of(id, struct rdma_id_private, id);
 954         if (id->device != pd->device) {
 955                 ret = -EINVAL;
 956                 goto out_err;
 957         }
 958
 959         qp_init_attr->port_num = id->port_num;
 960         qp = ib_create_qp(pd, qp_init_attr);
 961         if (IS_ERR(qp)) {
 962                 ret = PTR_ERR(qp);
 963                 goto out_err;
 964         }
 965
 966         if (id->qp_type == IB_QPT_UD)
 967                 ret = cma_init_ud_qp(id_priv, qp);
 968         else
 969                 ret = cma_init_conn_qp(id_priv, qp);
 970         if (ret)
 971                 goto out_destroy;
 972
 973         id->qp = qp;
 974         id_priv->qp_num = qp->qp_num;
 975         id_priv->srq = (qp->srq != NULL);
 976         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 977         return 0;
 978 out_destroy:
 979         ib_destroy_qp(qp);
 980 out_err:
 981         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 982         return ret;
 983 }
 984 EXPORT_SYMBOL(rdma_create_qp);
 985
 986 void rdma_destroy_qp(struct rdma_cm_id *id)
 987 {
 988         struct rdma_id_private *id_priv;
 989
 990         id_priv = container_of(id, struct rdma_id_private, id);
 991         trace_cm_qp_destroy(id_priv);
 992         mutex_lock(&id_priv->qp_mutex);
 993         ib_destroy_qp(id_priv->id.qp);
 994         id_priv->id.qp = NULL;
 995         mutex_unlock(&id_priv->qp_mutex);
 996 }
 997 EXPORT_SYMBOL(rdma_destroy_qp);
 998
 999 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1000                              struct rdma_conn_param *conn_param)
1001 {
1002         struct ib_qp_attr qp_attr;
1003         int qp_attr_mask, ret;
1004
1005         mutex_lock(&id_priv->qp_mutex);
1006         if (!id_priv->id.qp) {
1007                 ret = 0;
1008                 goto out;
1009         }
1010
1011         /* Need to update QP attributes from default values. */
1012         qp_attr.qp_state = IB_QPS_INIT;
1013         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1014         if (ret)
1015                 goto out;
1016
1017         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1018         if (ret)
1019                 goto out;
1020
1021         qp_attr.qp_state = IB_QPS_RTR;
1022         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1023         if (ret)
1024                 goto out;
1025
1026         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1027
1028         if (conn_param)
1029                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1030         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1031 out:
1032         mutex_unlock(&id_priv->qp_mutex);
1033         return ret;
1034 }
1035
1036 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1037                              struct rdma_conn_param *conn_param)
1038 {
1039         struct ib_qp_attr qp_attr;
1040         int qp_attr_mask, ret;
1041
1042         mutex_lock(&id_priv->qp_mutex);
1043         if (!id_priv->id.qp) {
1044                 ret = 0;
1045                 goto out;
1046         }
1047
1048         qp_attr.qp_state = IB_QPS_RTS;
1049         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1050         if (ret)
1051                 goto out;
1052
1053         if (conn_param)
1054                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1055         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1056 out:
1057         mutex_unlock(&id_priv->qp_mutex);
1058         return ret;
1059 }
1060
1061 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1062 {
1063         struct ib_qp_attr qp_attr;
1064         int ret;
1065
1066         mutex_lock(&id_priv->qp_mutex);
1067         if (!id_priv->id.qp) {
1068                 ret = 0;
1069                 goto out;
1070         }
1071
1072         qp_attr.qp_state = IB_QPS_ERR;
1073         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1074 out:
1075         mutex_unlock(&id_priv->qp_mutex);
1076         return ret;
1077 }
1078
1079 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1080                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1081 {
1082         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1083         int ret;
1084         u16 pkey;
1085
1086         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1087                 pkey = 0xffff;
1088         else
1089                 pkey = ib_addr_get_pkey(dev_addr);
1090
1091         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1092                                   pkey, &qp_attr->pkey_index);
1093         if (ret)
1094                 return ret;
1095
1096         qp_attr->port_num = id_priv->id.port_num;
1097         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1098
1099         if (id_priv->id.qp_type == IB_QPT_UD) {
1100                 ret = cma_set_qkey(id_priv, 0);
1101                 if (ret)
1102                         return ret;
1103
1104                 qp_attr->qkey = id_priv->qkey;
1105                 *qp_attr_mask |= IB_QP_QKEY;
1106         } else {
1107                 qp_attr->qp_access_flags = 0;
1108                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1109         }
1110         return 0;
1111 }
1112
1113 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1114                        int *qp_attr_mask)
1115 {
1116         struct rdma_id_private *id_priv;
1117         int ret = 0;
1118
1119         id_priv = container_of(id, struct rdma_id_private, id);
1120         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1121                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1122                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1123                 else
1124                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1125                                                  qp_attr_mask);
1126
1127                 if (qp_attr->qp_state == IB_QPS_RTR)
1128                         qp_attr->rq_psn = id_priv->seq_num;
1129         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1130                 if (!id_priv->cm_id.iw) {
1131                         qp_attr->qp_access_flags = 0;
1132                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1133                 } else
1134                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1135                                                  qp_attr_mask);
1136                 qp_attr->port_num = id_priv->id.port_num;
1137                 *qp_attr_mask |= IB_QP_PORT;
1138         } else
1139                 ret = -ENOSYS;
1140
1141         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1142                 qp_attr->timeout = id_priv->timeout;
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 int 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         if (id_priv->cma_dev) {
1856                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1857                         if (id_priv->cm_id.ib)
1858                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1859                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1860                         if (id_priv->cm_id.iw)
1861                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1862                 }
1863                 cma_leave_mc_groups(id_priv);
1864                 cma_release_dev(id_priv);
1865         }
1866
1867         cma_release_port(id_priv);
1868         cma_id_put(id_priv);
1869         wait_for_completion(&id_priv->comp);
1870
1871         if (id_priv->internal_id)
1872                 cma_id_put(id_priv->id.context);
1873
1874         kfree(id_priv->id.route.path_rec);
1875
1876         put_net(id_priv->id.route.addr.dev_addr.net);
1877         rdma_restrack_del(&id_priv->res);
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         id->tos = id_priv->tos;
2476         id->tos_set = id_priv->tos_set;
2477         id->afonly = id_priv->afonly;
2478         id_priv->cm_id.iw = id;
2479
2480         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2481                rdma_addr_size(cma_src_addr(id_priv)));
2482
2483         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2484
2485         if (ret) {
2486                 iw_destroy_cm_id(id_priv->cm_id.iw);
2487                 id_priv->cm_id.iw = NULL;
2488         }
2489
2490         return ret;
2491 }
2492
2493 static int cma_listen_handler(struct rdma_cm_id *id,
2494                               struct rdma_cm_event *event)
2495 {
2496         struct rdma_id_private *id_priv = id->context;
2497
2498         /* Listening IDs are always destroyed on removal */
2499         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2500                 return -1;
2501
2502         id->context = id_priv->id.context;
2503         id->event_handler = id_priv->id.event_handler;
2504         trace_cm_event_handler(id_priv, event);
2505         return id_priv->id.event_handler(id, event);
2506 }
2507
2508 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2509                              struct cma_device *cma_dev,
2510                              struct rdma_id_private **to_destroy)
2511 {
2512         struct rdma_id_private *dev_id_priv;
2513         struct net *net = id_priv->id.route.addr.dev_addr.net;
2514         int ret;
2515
2516         lockdep_assert_held(&lock);
2517
2518         *to_destroy = NULL;
2519         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2520                 return 0;
2521
2522         dev_id_priv =
2523                 __rdma_create_id(net, cma_listen_handler, id_priv,
2524                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2525         if (IS_ERR(dev_id_priv))
2526                 return PTR_ERR(dev_id_priv);
2527
2528         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2529         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2530                rdma_addr_size(cma_src_addr(id_priv)));
2531
2532         _cma_attach_to_dev(dev_id_priv, cma_dev);
2533         cma_id_get(id_priv);
2534         dev_id_priv->internal_id = 1;
2535         dev_id_priv->afonly = id_priv->afonly;
2536         dev_id_priv->tos_set = id_priv->tos_set;
2537         dev_id_priv->tos = id_priv->tos;
2538
2539         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2540         if (ret)
2541                 goto err_listen;
2542         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2543         return 0;
2544 err_listen:
2545         /* Caller must destroy this after releasing lock */
2546         *to_destroy = dev_id_priv;
2547         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2548         return ret;
2549 }
2550
2551 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2552 {
2553         struct rdma_id_private *to_destroy;
2554         struct cma_device *cma_dev;
2555         int ret;
2556
2557         mutex_lock(&lock);
2558         list_add_tail(&id_priv->list, &listen_any_list);
2559         list_for_each_entry(cma_dev, &dev_list, list) {
2560                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2561                 if (ret) {
2562                         /* Prevent racing with cma_process_remove() */
2563                         if (to_destroy)
2564                                 list_del_init(&to_destroy->list);
2565                         goto err_listen;
2566                 }
2567         }
2568         mutex_unlock(&lock);
2569         return 0;
2570
2571 err_listen:
2572         list_del(&id_priv->list);
2573         mutex_unlock(&lock);
2574         if (to_destroy)
2575                 rdma_destroy_id(&to_destroy->id);
2576         return ret;
2577 }
2578
2579 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2580 {
2581         struct rdma_id_private *id_priv;
2582
2583         id_priv = container_of(id, struct rdma_id_private, id);
2584         id_priv->tos = (u8) tos;
2585         id_priv->tos_set = true;
2586 }
2587 EXPORT_SYMBOL(rdma_set_service_type);
2588
2589 /**
2590  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2591  *                          with a connection identifier.
2592  * @id: Communication identifier to associated with service type.
2593  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2594  *
2595  * This function should be called before rdma_connect() on active side,
2596  * and on passive side before rdma_accept(). It is applicable to primary
2597  * path only. The timeout will affect the local side of the QP, it is not
2598  * negotiated with remote side and zero disables the timer. In case it is
2599  * set before rdma_resolve_route, the value will also be used to determine
2600  * PacketLifeTime for RoCE.
2601  *
2602  * Return: 0 for success
2603  */
2604 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2605 {
2606         struct rdma_id_private *id_priv;
2607
2608         if (id->qp_type != IB_QPT_RC)
2609                 return -EINVAL;
2610
2611         id_priv = container_of(id, struct rdma_id_private, id);
2612         id_priv->timeout = timeout;
2613         id_priv->timeout_set = true;
2614
2615         return 0;
2616 }
2617 EXPORT_SYMBOL(rdma_set_ack_timeout);
2618
2619 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2620                               void *context)
2621 {
2622         struct cma_work *work = context;
2623         struct rdma_route *route;
2624
2625         route = &work->id->id.route;
2626
2627         if (!status) {
2628                 route->num_paths = 1;
2629                 *route->path_rec = *path_rec;
2630         } else {
2631                 work->old_state = RDMA_CM_ROUTE_QUERY;
2632                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2633                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2634                 work->event.status = status;
2635                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2636                                      status);
2637         }
2638
2639         queue_work(cma_wq, &work->work);
2640 }
2641
2642 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2643                               unsigned long timeout_ms, struct cma_work *work)
2644 {
2645         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2646         struct sa_path_rec path_rec;
2647         ib_sa_comp_mask comp_mask;
2648         struct sockaddr_in6 *sin6;
2649         struct sockaddr_ib *sib;
2650
2651         memset(&path_rec, 0, sizeof path_rec);
2652
2653         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2654                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2655         else
2656                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2657         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2658         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2659         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2660         path_rec.numb_path = 1;
2661         path_rec.reversible = 1;
2662         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2663                                                   cma_dst_addr(id_priv));
2664
2665         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2666                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2667                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2668
2669         switch (cma_family(id_priv)) {
2670         case AF_INET:
2671                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2672                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2673                 break;
2674         case AF_INET6:
2675                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2676                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2677                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2678                 break;
2679         case AF_IB:
2680                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2681                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2682                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2683                 break;
2684         }
2685
2686         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2687                                                id_priv->id.port_num, &path_rec,
2688                                                comp_mask, timeout_ms,
2689                                                GFP_KERNEL, cma_query_handler,
2690                                                work, &id_priv->query);
2691
2692         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2693 }
2694
2695 static void cma_iboe_join_work_handler(struct work_struct *work)
2696 {
2697         struct cma_multicast *mc =
2698                 container_of(work, struct cma_multicast, iboe_join.work);
2699         struct rdma_cm_event *event = &mc->iboe_join.event;
2700         struct rdma_id_private *id_priv = mc->id_priv;
2701         int ret;
2702
2703         mutex_lock(&id_priv->handler_mutex);
2704         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2705             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2706                 goto out_unlock;
2707
2708         ret = cma_cm_event_handler(id_priv, event);
2709         WARN_ON(ret);
2710
2711 out_unlock:
2712         mutex_unlock(&id_priv->handler_mutex);
2713         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2714                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2715 }
2716
2717 static void cma_work_handler(struct work_struct *_work)
2718 {
2719         struct cma_work *work = container_of(_work, struct cma_work, work);
2720         struct rdma_id_private *id_priv = work->id;
2721
2722         mutex_lock(&id_priv->handler_mutex);
2723         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2724             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2725                 goto out_unlock;
2726         if (work->old_state != 0 || work->new_state != 0) {
2727                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2728                         goto out_unlock;
2729         }
2730
2731         if (cma_cm_event_handler(id_priv, &work->event)) {
2732                 cma_id_put(id_priv);
2733                 destroy_id_handler_unlock(id_priv);
2734                 goto out_free;
2735         }
2736
2737 out_unlock:
2738         mutex_unlock(&id_priv->handler_mutex);
2739         cma_id_put(id_priv);
2740 out_free:
2741         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2742                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2743         kfree(work);
2744 }
2745
2746 static void cma_init_resolve_route_work(struct cma_work *work,
2747                                         struct rdma_id_private *id_priv)
2748 {
2749         work->id = id_priv;
2750         INIT_WORK(&work->work, cma_work_handler);
2751         work->old_state = RDMA_CM_ROUTE_QUERY;
2752         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2753         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2754 }
2755
2756 static void enqueue_resolve_addr_work(struct cma_work *work,
2757                                       struct rdma_id_private *id_priv)
2758 {
2759         /* Balances with cma_id_put() in cma_work_handler */
2760         cma_id_get(id_priv);
2761
2762         work->id = id_priv;
2763         INIT_WORK(&work->work, cma_work_handler);
2764         work->old_state = RDMA_CM_ADDR_QUERY;
2765         work->new_state = RDMA_CM_ADDR_RESOLVED;
2766         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2767
2768         queue_work(cma_wq, &work->work);
2769 }
2770
2771 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2772                                 unsigned long timeout_ms)
2773 {
2774         struct rdma_route *route = &id_priv->id.route;
2775         struct cma_work *work;
2776         int ret;
2777
2778         work = kzalloc(sizeof *work, GFP_KERNEL);
2779         if (!work)
2780                 return -ENOMEM;
2781
2782         cma_init_resolve_route_work(work, id_priv);
2783
2784         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2785         if (!route->path_rec) {
2786                 ret = -ENOMEM;
2787                 goto err1;
2788         }
2789
2790         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2791         if (ret)
2792                 goto err2;
2793
2794         return 0;
2795 err2:
2796         kfree(route->path_rec);
2797         route->path_rec = NULL;
2798 err1:
2799         kfree(work);
2800         return ret;
2801 }
2802
2803 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2804                                            unsigned long supported_gids,
2805                                            enum ib_gid_type default_gid)
2806 {
2807         if ((network_type == RDMA_NETWORK_IPV4 ||
2808              network_type == RDMA_NETWORK_IPV6) &&
2809             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2810                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2811
2812         return default_gid;
2813 }
2814
2815 /*
2816  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2817  * path record type based on GID type.
2818  * It also sets up other L2 fields which includes destination mac address
2819  * netdev ifindex, of the path record.
2820  * It returns the netdev of the bound interface for this path record entry.
2821  */
2822 static struct net_device *
2823 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2824 {
2825         struct rdma_route *route = &id_priv->id.route;
2826         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2827         struct rdma_addr *addr = &route->addr;
2828         unsigned long supported_gids;
2829         struct net_device *ndev;
2830
2831         if (!addr->dev_addr.bound_dev_if)
2832                 return NULL;
2833
2834         ndev = dev_get_by_index(addr->dev_addr.net,
2835                                 addr->dev_addr.bound_dev_if);
2836         if (!ndev)
2837                 return NULL;
2838
2839         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2840                                                     id_priv->id.port_num);
2841         gid_type = cma_route_gid_type(addr->dev_addr.network,
2842                                       supported_gids,
2843                                       id_priv->gid_type);
2844         /* Use the hint from IP Stack to select GID Type */
2845         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2846                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2847         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2848
2849         route->path_rec->roce.route_resolved = true;
2850         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2851         return ndev;
2852 }
2853
2854 int rdma_set_ib_path(struct rdma_cm_id *id,
2855                      struct sa_path_rec *path_rec)
2856 {
2857         struct rdma_id_private *id_priv;
2858         struct net_device *ndev;
2859         int ret;
2860
2861         id_priv = container_of(id, struct rdma_id_private, id);
2862         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2863                            RDMA_CM_ROUTE_RESOLVED))
2864                 return -EINVAL;
2865
2866         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2867                                      GFP_KERNEL);
2868         if (!id->route.path_rec) {
2869                 ret = -ENOMEM;
2870                 goto err;
2871         }
2872
2873         if (rdma_protocol_roce(id->device, id->port_num)) {
2874                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2875                 if (!ndev) {
2876                         ret = -ENODEV;
2877                         goto err_free;
2878                 }
2879                 dev_put(ndev);
2880         }
2881
2882         id->route.num_paths = 1;
2883         return 0;
2884
2885 err_free:
2886         kfree(id->route.path_rec);
2887         id->route.path_rec = NULL;
2888 err:
2889         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2890         return ret;
2891 }
2892 EXPORT_SYMBOL(rdma_set_ib_path);
2893
2894 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2895 {
2896         struct cma_work *work;
2897
2898         work = kzalloc(sizeof *work, GFP_KERNEL);
2899         if (!work)
2900                 return -ENOMEM;
2901
2902         cma_init_resolve_route_work(work, id_priv);
2903         queue_work(cma_wq, &work->work);
2904         return 0;
2905 }
2906
2907 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2908 {
2909         struct net_device *dev;
2910
2911         dev = vlan_dev_real_dev(vlan_ndev);
2912         if (dev->num_tc)
2913                 return netdev_get_prio_tc_map(dev, prio);
2914
2915         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2916                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2917 }
2918
2919 struct iboe_prio_tc_map {
2920         int input_prio;
2921         int output_tc;
2922         bool found;
2923 };
2924
2925 static int get_lower_vlan_dev_tc(struct net_device *dev,
2926                                  struct netdev_nested_priv *priv)
2927 {
2928         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2929
2930         if (is_vlan_dev(dev))
2931                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2932         else if (dev->num_tc)
2933                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2934         else
2935                 map->output_tc = 0;
2936         /* We are interested only in first level VLAN device, so always
2937          * return 1 to stop iterating over next level devices.
2938          */
2939         map->found = true;
2940         return 1;
2941 }
2942
2943 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2944 {
2945         struct iboe_prio_tc_map prio_tc_map = {};
2946         int prio = rt_tos2priority(tos);
2947         struct netdev_nested_priv priv;
2948
2949         /* If VLAN device, get it directly from the VLAN netdev */
2950         if (is_vlan_dev(ndev))
2951                 return get_vlan_ndev_tc(ndev, prio);
2952
2953         prio_tc_map.input_prio = prio;
2954         priv.data = (void *)&prio_tc_map;
2955         rcu_read_lock();
2956         netdev_walk_all_lower_dev_rcu(ndev,
2957                                       get_lower_vlan_dev_tc,
2958                                       &priv);
2959         rcu_read_unlock();
2960         /* If map is found from lower device, use it; Otherwise
2961          * continue with the current netdevice to get priority to tc map.
2962          */
2963         if (prio_tc_map.found)
2964                 return prio_tc_map.output_tc;
2965         else if (ndev->num_tc)
2966                 return netdev_get_prio_tc_map(ndev, prio);
2967         else
2968                 return 0;
2969 }
2970
2971 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
2972 {
2973         struct sockaddr_in6 *addr6;
2974         u16 dport, sport;
2975         u32 hash, fl;
2976
2977         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
2978         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
2979         if ((cma_family(id_priv) != AF_INET6) || !fl) {
2980                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
2981                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
2982                 hash = (u32)sport * 31 + dport;
2983                 fl = hash & IB_GRH_FLOWLABEL_MASK;
2984         }
2985
2986         return cpu_to_be32(fl);
2987 }
2988
2989 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2990 {
2991         struct rdma_route *route = &id_priv->id.route;
2992         struct rdma_addr *addr = &route->addr;
2993         struct cma_work *work;
2994         int ret;
2995         struct net_device *ndev;
2996
2997         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2998                                         rdma_start_port(id_priv->cma_dev->device)];
2999         u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3000
3001
3002         work = kzalloc(sizeof *work, GFP_KERNEL);
3003         if (!work)
3004                 return -ENOMEM;
3005
3006         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3007         if (!route->path_rec) {
3008                 ret = -ENOMEM;
3009                 goto err1;
3010         }
3011
3012         route->num_paths = 1;
3013
3014         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3015         if (!ndev) {
3016                 ret = -ENODEV;
3017                 goto err2;
3018         }
3019
3020         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3021                     &route->path_rec->sgid);
3022         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3023                     &route->path_rec->dgid);
3024
3025         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3026                 /* TODO: get the hoplimit from the inet/inet6 device */
3027                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3028         else
3029                 route->path_rec->hop_limit = 1;
3030         route->path_rec->reversible = 1;
3031         route->path_rec->pkey = cpu_to_be16(0xffff);
3032         route->path_rec->mtu_selector = IB_SA_EQ;
3033         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3034         route->path_rec->traffic_class = tos;
3035         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3036         route->path_rec->rate_selector = IB_SA_EQ;
3037         route->path_rec->rate = iboe_get_rate(ndev);
3038         dev_put(ndev);
3039         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3040         /* In case ACK timeout is set, use this value to calculate
3041          * PacketLifeTime.  As per IBTA 12.7.34,
3042          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3043          * Assuming a negligible local ACK delay, we can use
3044          * PacketLifeTime = local ACK timeout/2
3045          * as a reasonable approximation for RoCE networks.
3046          */
3047         route->path_rec->packet_life_time = id_priv->timeout_set ?
3048                 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
3049
3050         if (!route->path_rec->mtu) {
3051                 ret = -EINVAL;
3052                 goto err2;
3053         }
3054
3055         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3056                                          id_priv->id.port_num))
3057                 route->path_rec->flow_label =
3058                         cma_get_roce_udp_flow_label(id_priv);
3059
3060         cma_init_resolve_route_work(work, id_priv);
3061         queue_work(cma_wq, &work->work);
3062
3063         return 0;
3064
3065 err2:
3066         kfree(route->path_rec);
3067         route->path_rec = NULL;
3068         route->num_paths = 0;
3069 err1:
3070         kfree(work);
3071         return ret;
3072 }
3073
3074 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3075 {
3076         struct rdma_id_private *id_priv;
3077         int ret;
3078
3079         id_priv = container_of(id, struct rdma_id_private, id);
3080         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3081                 return -EINVAL;
3082
3083         cma_id_get(id_priv);
3084         if (rdma_cap_ib_sa(id->device, id->port_num))
3085                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3086         else if (rdma_protocol_roce(id->device, id->port_num))
3087                 ret = cma_resolve_iboe_route(id_priv);
3088         else if (rdma_protocol_iwarp(id->device, id->port_num))
3089                 ret = cma_resolve_iw_route(id_priv);
3090         else
3091                 ret = -ENOSYS;
3092
3093         if (ret)
3094                 goto err;
3095
3096         return 0;
3097 err:
3098         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3099         cma_id_put(id_priv);
3100         return ret;
3101 }
3102 EXPORT_SYMBOL(rdma_resolve_route);
3103
3104 static void cma_set_loopback(struct sockaddr *addr)
3105 {
3106         switch (addr->sa_family) {
3107         case AF_INET:
3108                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3109                 break;
3110         case AF_INET6:
3111                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3112                               0, 0, 0, htonl(1));
3113                 break;
3114         default:
3115                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3116                             0, 0, 0, htonl(1));
3117                 break;
3118         }
3119 }
3120
3121 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3122 {
3123         struct cma_device *cma_dev, *cur_dev;
3124         union ib_gid gid;
3125         enum ib_port_state port_state;
3126         unsigned int p;
3127         u16 pkey;
3128         int ret;
3129
3130         cma_dev = NULL;
3131         mutex_lock(&lock);
3132         list_for_each_entry(cur_dev, &dev_list, list) {
3133                 if (cma_family(id_priv) == AF_IB &&
3134                     !rdma_cap_ib_cm(cur_dev->device, 1))
3135                         continue;
3136
3137                 if (!cma_dev)
3138                         cma_dev = cur_dev;
3139
3140                 rdma_for_each_port (cur_dev->device, p) {
3141                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3142                             port_state == IB_PORT_ACTIVE) {
3143                                 cma_dev = cur_dev;
3144                                 goto port_found;
3145                         }
3146                 }
3147         }
3148
3149         if (!cma_dev) {
3150                 ret = -ENODEV;
3151                 goto out;
3152         }
3153
3154         p = 1;
3155
3156 port_found:
3157         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3158         if (ret)
3159                 goto out;
3160
3161         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3162         if (ret)
3163                 goto out;
3164
3165         id_priv->id.route.addr.dev_addr.dev_type =
3166                 (rdma_protocol_ib(cma_dev->device, p)) ?
3167                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3168
3169         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3170         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3171         id_priv->id.port_num = p;
3172         cma_attach_to_dev(id_priv, cma_dev);
3173         cma_set_loopback(cma_src_addr(id_priv));
3174 out:
3175         mutex_unlock(&lock);
3176         return ret;
3177 }
3178
3179 static void addr_handler(int status, struct sockaddr *src_addr,
3180                          struct rdma_dev_addr *dev_addr, void *context)
3181 {
3182         struct rdma_id_private *id_priv = context;
3183         struct rdma_cm_event event = {};
3184         struct sockaddr *addr;
3185         struct sockaddr_storage old_addr;
3186
3187         mutex_lock(&id_priv->handler_mutex);
3188         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3189                            RDMA_CM_ADDR_RESOLVED))
3190                 goto out;
3191
3192         /*
3193          * Store the previous src address, so that if we fail to acquire
3194          * matching rdma device, old address can be restored back, which helps
3195          * to cancel the cma listen operation correctly.
3196          */
3197         addr = cma_src_addr(id_priv);
3198         memcpy(&old_addr, addr, rdma_addr_size(addr));
3199         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3200         if (!status && !id_priv->cma_dev) {
3201                 status = cma_acquire_dev_by_src_ip(id_priv);
3202                 if (status)
3203                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3204                                              status);
3205         } else if (status) {
3206                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3207         }
3208
3209         if (status) {
3210                 memcpy(addr, &old_addr,
3211                        rdma_addr_size((struct sockaddr *)&old_addr));
3212                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3213                                    RDMA_CM_ADDR_BOUND))
3214                         goto out;
3215                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3216                 event.status = status;
3217         } else
3218                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3219
3220         if (cma_cm_event_handler(id_priv, &event)) {
3221                 destroy_id_handler_unlock(id_priv);
3222                 return;
3223         }
3224 out:
3225         mutex_unlock(&id_priv->handler_mutex);
3226 }
3227
3228 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3229 {
3230         struct cma_work *work;
3231         union ib_gid gid;
3232         int ret;
3233
3234         work = kzalloc(sizeof *work, GFP_KERNEL);
3235         if (!work)
3236                 return -ENOMEM;
3237
3238         if (!id_priv->cma_dev) {
3239                 ret = cma_bind_loopback(id_priv);
3240                 if (ret)
3241                         goto err;
3242         }
3243
3244         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3245         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3246
3247         enqueue_resolve_addr_work(work, id_priv);
3248         return 0;
3249 err:
3250         kfree(work);
3251         return ret;
3252 }
3253
3254 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3255 {
3256         struct cma_work *work;
3257         int ret;
3258
3259         work = kzalloc(sizeof *work, GFP_KERNEL);
3260         if (!work)
3261                 return -ENOMEM;
3262
3263         if (!id_priv->cma_dev) {
3264                 ret = cma_resolve_ib_dev(id_priv);
3265                 if (ret)
3266                         goto err;
3267         }
3268
3269         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3270                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3271
3272         enqueue_resolve_addr_work(work, id_priv);
3273         return 0;
3274 err:
3275         kfree(work);
3276         return ret;
3277 }
3278
3279 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3280                          const struct sockaddr *dst_addr)
3281 {
3282         if (!src_addr || !src_addr->sa_family) {
3283                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3284                 src_addr->sa_family = dst_addr->sa_family;
3285                 if (IS_ENABLED(CONFIG_IPV6) &&
3286                     dst_addr->sa_family == AF_INET6) {
3287                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3288                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3289                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3290                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3291                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3292                 } else if (dst_addr->sa_family == AF_IB) {
3293                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3294                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3295                 }
3296         }
3297         return rdma_bind_addr(id, src_addr);
3298 }
3299
3300 /*
3301  * If required, resolve the source address for bind and leave the id_priv in
3302  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3303  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3304  * ignored.
3305  */
3306 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3307                                struct sockaddr *src_addr,
3308                                const struct sockaddr *dst_addr)
3309 {
3310         int ret;
3311
3312         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3313         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3314                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3315                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3316                 if (ret)
3317                         goto err_dst;
3318                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3319                                            RDMA_CM_ADDR_QUERY))) {
3320                         ret = -EINVAL;
3321                         goto err_dst;
3322                 }
3323         }
3324
3325         if (cma_family(id_priv) != dst_addr->sa_family) {
3326                 ret = -EINVAL;
3327                 goto err_state;
3328         }
3329         return 0;
3330
3331 err_state:
3332         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3333 err_dst:
3334         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3335         return ret;
3336 }
3337
3338 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3339                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3340 {
3341         struct rdma_id_private *id_priv =
3342                 container_of(id, struct rdma_id_private, id);
3343         int ret;
3344
3345         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3346         if (ret)
3347                 return ret;
3348
3349         if (cma_any_addr(dst_addr)) {
3350                 ret = cma_resolve_loopback(id_priv);
3351         } else {
3352                 if (dst_addr->sa_family == AF_IB) {
3353                         ret = cma_resolve_ib_addr(id_priv);
3354                 } else {
3355                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3356                                               &id->route.addr.dev_addr,
3357                                               timeout_ms, addr_handler,
3358                                               false, id_priv);
3359                 }
3360         }
3361         if (ret)
3362                 goto err;
3363
3364         return 0;
3365 err:
3366         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3367         return ret;
3368 }
3369 EXPORT_SYMBOL(rdma_resolve_addr);
3370
3371 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3372 {
3373         struct rdma_id_private *id_priv;
3374         unsigned long flags;
3375         int ret;
3376
3377         id_priv = container_of(id, struct rdma_id_private, id);
3378         spin_lock_irqsave(&id_priv->lock, flags);
3379         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3380             id_priv->state == RDMA_CM_IDLE) {
3381                 id_priv->reuseaddr = reuse;
3382                 ret = 0;
3383         } else {
3384                 ret = -EINVAL;
3385         }
3386         spin_unlock_irqrestore(&id_priv->lock, flags);
3387         return ret;
3388 }
3389 EXPORT_SYMBOL(rdma_set_reuseaddr);
3390
3391 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3392 {
3393         struct rdma_id_private *id_priv;
3394         unsigned long flags;
3395         int ret;
3396
3397         id_priv = container_of(id, struct rdma_id_private, id);
3398         spin_lock_irqsave(&id_priv->lock, flags);
3399         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3400                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3401                 id_priv->afonly = afonly;
3402                 ret = 0;
3403         } else {
3404                 ret = -EINVAL;
3405         }
3406         spin_unlock_irqrestore(&id_priv->lock, flags);
3407         return ret;
3408 }
3409 EXPORT_SYMBOL(rdma_set_afonly);
3410
3411 static void cma_bind_port(struct rdma_bind_list *bind_list,
3412                           struct rdma_id_private *id_priv)
3413 {
3414         struct sockaddr *addr;
3415         struct sockaddr_ib *sib;
3416         u64 sid, mask;
3417         __be16 port;
3418
3419         lockdep_assert_held(&lock);
3420
3421         addr = cma_src_addr(id_priv);
3422         port = htons(bind_list->port);
3423
3424         switch (addr->sa_family) {
3425         case AF_INET:
3426                 ((struct sockaddr_in *) addr)->sin_port = port;
3427                 break;
3428         case AF_INET6:
3429                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3430                 break;
3431         case AF_IB:
3432                 sib = (struct sockaddr_ib *) addr;
3433                 sid = be64_to_cpu(sib->sib_sid);
3434                 mask = be64_to_cpu(sib->sib_sid_mask);
3435                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3436                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3437                 break;
3438         }
3439         id_priv->bind_list = bind_list;
3440         hlist_add_head(&id_priv->node, &bind_list->owners);
3441 }
3442
3443 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3444                           struct rdma_id_private *id_priv, unsigned short snum)
3445 {
3446         struct rdma_bind_list *bind_list;
3447         int ret;
3448
3449         lockdep_assert_held(&lock);
3450
3451         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3452         if (!bind_list)
3453                 return -ENOMEM;
3454
3455         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3456                            snum);
3457         if (ret < 0)
3458                 goto err;
3459
3460         bind_list->ps = ps;
3461         bind_list->port = snum;
3462         cma_bind_port(bind_list, id_priv);
3463         return 0;
3464 err:
3465         kfree(bind_list);
3466         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3467 }
3468
3469 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3470                               struct rdma_id_private *id_priv)
3471 {
3472         struct rdma_id_private *cur_id;
3473         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3474         struct sockaddr  *saddr = cma_src_addr(id_priv);
3475         __be16 dport = cma_port(daddr);
3476
3477         lockdep_assert_held(&lock);
3478
3479         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3480                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3481                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3482                 __be16 cur_dport = cma_port(cur_daddr);
3483
3484                 if (id_priv == cur_id)
3485                         continue;
3486
3487                 /* different dest port -> unique */
3488                 if (!cma_any_port(daddr) &&
3489                     !cma_any_port(cur_daddr) &&
3490                     (dport != cur_dport))
3491                         continue;
3492
3493                 /* different src address -> unique */
3494                 if (!cma_any_addr(saddr) &&
3495                     !cma_any_addr(cur_saddr) &&
3496                     cma_addr_cmp(saddr, cur_saddr))
3497                         continue;
3498
3499                 /* different dst address -> unique */
3500                 if (!cma_any_addr(daddr) &&
3501                     !cma_any_addr(cur_daddr) &&
3502                     cma_addr_cmp(daddr, cur_daddr))
3503                         continue;
3504
3505                 return -EADDRNOTAVAIL;
3506         }
3507         return 0;
3508 }
3509
3510 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3511                               struct rdma_id_private *id_priv)
3512 {
3513         static unsigned int last_used_port;
3514         int low, high, remaining;
3515         unsigned int rover;
3516         struct net *net = id_priv->id.route.addr.dev_addr.net;
3517
3518         lockdep_assert_held(&lock);
3519
3520         inet_get_local_port_range(net, &low, &high);
3521         remaining = (high - low) + 1;
3522         rover = prandom_u32() % remaining + low;
3523 retry:
3524         if (last_used_port != rover) {
3525                 struct rdma_bind_list *bind_list;
3526                 int ret;
3527
3528                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3529
3530                 if (!bind_list) {
3531                         ret = cma_alloc_port(ps, id_priv, rover);
3532                 } else {
3533                         ret = cma_port_is_unique(bind_list, id_priv);
3534                         if (!ret)
3535                                 cma_bind_port(bind_list, id_priv);
3536                 }
3537                 /*
3538                  * Remember previously used port number in order to avoid
3539                  * re-using same port immediately after it is closed.
3540                  */
3541                 if (!ret)
3542                         last_used_port = rover;
3543                 if (ret != -EADDRNOTAVAIL)
3544                         return ret;
3545         }
3546         if (--remaining) {
3547                 rover++;
3548                 if ((rover < low) || (rover > high))
3549                         rover = low;
3550                 goto retry;
3551         }
3552         return -EADDRNOTAVAIL;
3553 }
3554
3555 /*
3556  * Check that the requested port is available.  This is called when trying to
3557  * bind to a specific port, or when trying to listen on a bound port.  In
3558  * the latter case, the provided id_priv may already be on the bind_list, but
3559  * we still need to check that it's okay to start listening.
3560  */
3561 static int cma_check_port(struct rdma_bind_list *bind_list,
3562                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3563 {
3564         struct rdma_id_private *cur_id;
3565         struct sockaddr *addr, *cur_addr;
3566
3567         lockdep_assert_held(&lock);
3568
3569         addr = cma_src_addr(id_priv);
3570         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3571                 if (id_priv == cur_id)
3572                         continue;
3573
3574                 if (reuseaddr && cur_id->reuseaddr)
3575                         continue;
3576
3577                 cur_addr = cma_src_addr(cur_id);
3578                 if (id_priv->afonly && cur_id->afonly &&
3579                     (addr->sa_family != cur_addr->sa_family))
3580                         continue;
3581
3582                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3583                         return -EADDRNOTAVAIL;
3584
3585                 if (!cma_addr_cmp(addr, cur_addr))
3586                         return -EADDRINUSE;
3587         }
3588         return 0;
3589 }
3590
3591 static int cma_use_port(enum rdma_ucm_port_space ps,
3592                         struct rdma_id_private *id_priv)
3593 {
3594         struct rdma_bind_list *bind_list;
3595         unsigned short snum;
3596         int ret;
3597
3598         lockdep_assert_held(&lock);
3599
3600         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3601         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3602                 return -EACCES;
3603
3604         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3605         if (!bind_list) {
3606                 ret = cma_alloc_port(ps, id_priv, snum);
3607         } else {
3608                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3609                 if (!ret)
3610                         cma_bind_port(bind_list, id_priv);
3611         }
3612         return ret;
3613 }
3614
3615 static enum rdma_ucm_port_space
3616 cma_select_inet_ps(struct rdma_id_private *id_priv)
3617 {
3618         switch (id_priv->id.ps) {
3619         case RDMA_PS_TCP:
3620         case RDMA_PS_UDP:
3621         case RDMA_PS_IPOIB:
3622         case RDMA_PS_IB:
3623                 return id_priv->id.ps;
3624         default:
3625
3626                 return 0;
3627         }
3628 }
3629
3630 static enum rdma_ucm_port_space
3631 cma_select_ib_ps(struct rdma_id_private *id_priv)
3632 {
3633         enum rdma_ucm_port_space ps = 0;
3634         struct sockaddr_ib *sib;
3635         u64 sid_ps, mask, sid;
3636
3637         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3638         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3639         sid = be64_to_cpu(sib->sib_sid) & mask;
3640
3641         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3642                 sid_ps = RDMA_IB_IP_PS_IB;
3643                 ps = RDMA_PS_IB;
3644         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3645                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3646                 sid_ps = RDMA_IB_IP_PS_TCP;
3647                 ps = RDMA_PS_TCP;
3648         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3649                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3650                 sid_ps = RDMA_IB_IP_PS_UDP;
3651                 ps = RDMA_PS_UDP;
3652         }
3653
3654         if (ps) {
3655                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3656                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3657                                                 be64_to_cpu(sib->sib_sid_mask));
3658         }
3659         return ps;
3660 }
3661
3662 static int cma_get_port(struct rdma_id_private *id_priv)
3663 {
3664         enum rdma_ucm_port_space ps;
3665         int ret;
3666
3667         if (cma_family(id_priv) != AF_IB)
3668                 ps = cma_select_inet_ps(id_priv);
3669         else
3670                 ps = cma_select_ib_ps(id_priv);
3671         if (!ps)
3672                 return -EPROTONOSUPPORT;
3673
3674         mutex_lock(&lock);
3675         if (cma_any_port(cma_src_addr(id_priv)))
3676                 ret = cma_alloc_any_port(ps, id_priv);
3677         else
3678                 ret = cma_use_port(ps, id_priv);
3679         mutex_unlock(&lock);
3680
3681         return ret;
3682 }
3683
3684 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3685                                struct sockaddr *addr)
3686 {
3687 #if IS_ENABLED(CONFIG_IPV6)
3688         struct sockaddr_in6 *sin6;
3689
3690         if (addr->sa_family != AF_INET6)
3691                 return 0;
3692
3693         sin6 = (struct sockaddr_in6 *) addr;
3694
3695         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3696                 return 0;
3697
3698         if (!sin6->sin6_scope_id)
3699                         return -EINVAL;
3700
3701         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3702 #endif
3703         return 0;
3704 }
3705
3706 int rdma_listen(struct rdma_cm_id *id, int backlog)
3707 {
3708         struct rdma_id_private *id_priv =
3709                 container_of(id, struct rdma_id_private, id);
3710         int ret;
3711
3712         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3713                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3714                 id->route.addr.src_addr.ss_family = AF_INET;
3715                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3716                 if (ret)
3717                         return ret;
3718                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3719                                            RDMA_CM_LISTEN)))
3720                         return -EINVAL;
3721         }
3722
3723         /*
3724          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3725          * any more, and has to be unique in the bind list.
3726          */
3727         if (id_priv->reuseaddr) {
3728                 mutex_lock(&lock);
3729                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3730                 if (!ret)
3731                         id_priv->reuseaddr = 0;
3732                 mutex_unlock(&lock);
3733                 if (ret)
3734                         goto err;
3735         }
3736
3737         id_priv->backlog = backlog;
3738         if (id->device) {
3739                 if (rdma_cap_ib_cm(id->device, 1)) {
3740                         ret = cma_ib_listen(id_priv);
3741                         if (ret)
3742                                 goto err;
3743                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3744                         ret = cma_iw_listen(id_priv, backlog);
3745                         if (ret)
3746                                 goto err;
3747                 } else {
3748                         ret = -ENOSYS;
3749                         goto err;
3750                 }
3751         } else {
3752                 ret = cma_listen_on_all(id_priv);
3753                 if (ret)
3754                         goto err;
3755         }
3756
3757         return 0;
3758 err:
3759         id_priv->backlog = 0;
3760         /*
3761          * All the failure paths that lead here will not allow the req_handler's
3762          * to have run.
3763          */
3764         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3765         return ret;
3766 }
3767 EXPORT_SYMBOL(rdma_listen);
3768
3769 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3770 {
3771         struct rdma_id_private *id_priv;
3772         int ret;
3773         struct sockaddr  *daddr;
3774
3775         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3776             addr->sa_family != AF_IB)
3777                 return -EAFNOSUPPORT;
3778
3779         id_priv = container_of(id, struct rdma_id_private, id);
3780         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3781                 return -EINVAL;
3782
3783         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3784         if (ret)
3785                 goto err1;
3786
3787         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3788         if (!cma_any_addr(addr)) {
3789                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3790                 if (ret)
3791                         goto err1;
3792
3793                 ret = cma_acquire_dev_by_src_ip(id_priv);
3794                 if (ret)
3795                         goto err1;
3796         }
3797
3798         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3799                 if (addr->sa_family == AF_INET)
3800                         id_priv->afonly = 1;
3801 #if IS_ENABLED(CONFIG_IPV6)
3802                 else if (addr->sa_family == AF_INET6) {
3803                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3804
3805                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3806                 }
3807 #endif
3808         }
3809         daddr = cma_dst_addr(id_priv);
3810         daddr->sa_family = addr->sa_family;
3811
3812         ret = cma_get_port(id_priv);
3813         if (ret)
3814                 goto err2;
3815
3816         return 0;
3817 err2:
3818         if (id_priv->cma_dev)
3819                 cma_release_dev(id_priv);
3820 err1:
3821         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3822         return ret;
3823 }
3824 EXPORT_SYMBOL(rdma_bind_addr);
3825
3826 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3827 {
3828         struct cma_hdr *cma_hdr;
3829
3830         cma_hdr = hdr;
3831         cma_hdr->cma_version = CMA_VERSION;
3832         if (cma_family(id_priv) == AF_INET) {
3833                 struct sockaddr_in *src4, *dst4;
3834
3835                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3836                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3837
3838                 cma_set_ip_ver(cma_hdr, 4);
3839                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3840                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3841                 cma_hdr->port = src4->sin_port;
3842         } else if (cma_family(id_priv) == AF_INET6) {
3843                 struct sockaddr_in6 *src6, *dst6;
3844
3845                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3846                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3847
3848                 cma_set_ip_ver(cma_hdr, 6);
3849                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3850                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3851                 cma_hdr->port = src6->sin6_port;
3852         }
3853         return 0;
3854 }
3855
3856 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3857                                 const struct ib_cm_event *ib_event)
3858 {
3859         struct rdma_id_private *id_priv = cm_id->context;
3860         struct rdma_cm_event event = {};
3861         const struct ib_cm_sidr_rep_event_param *rep =
3862                                 &ib_event->param.sidr_rep_rcvd;
3863         int ret;
3864
3865         mutex_lock(&id_priv->handler_mutex);
3866         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3867                 goto out;
3868
3869         switch (ib_event->event) {
3870         case IB_CM_SIDR_REQ_ERROR:
3871                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3872                 event.status = -ETIMEDOUT;
3873                 break;
3874         case IB_CM_SIDR_REP_RECEIVED:
3875                 event.param.ud.private_data = ib_event->private_data;
3876                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3877                 if (rep->status != IB_SIDR_SUCCESS) {
3878                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3879                         event.status = ib_event->param.sidr_rep_rcvd.status;
3880                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3881                                              event.status);
3882                         break;
3883                 }
3884                 ret = cma_set_qkey(id_priv, rep->qkey);
3885                 if (ret) {
3886                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3887                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3888                         event.status = ret;
3889                         break;
3890                 }
3891                 ib_init_ah_attr_from_path(id_priv->id.device,
3892                                           id_priv->id.port_num,
3893                                           id_priv->id.route.path_rec,
3894                                           &event.param.ud.ah_attr,
3895                                           rep->sgid_attr);
3896                 event.param.ud.qp_num = rep->qpn;
3897                 event.param.ud.qkey = rep->qkey;
3898                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3899                 event.status = 0;
3900                 break;
3901         default:
3902                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3903                        ib_event->event);
3904                 goto out;
3905         }
3906
3907         ret = cma_cm_event_handler(id_priv, &event);
3908
3909         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3910         if (ret) {
3911                 /* Destroy the CM ID by returning a non-zero value. */
3912                 id_priv->cm_id.ib = NULL;
3913                 destroy_id_handler_unlock(id_priv);
3914                 return ret;
3915         }
3916 out:
3917         mutex_unlock(&id_priv->handler_mutex);
3918         return 0;
3919 }
3920
3921 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3922                               struct rdma_conn_param *conn_param)
3923 {
3924         struct ib_cm_sidr_req_param req;
3925         struct ib_cm_id *id;
3926         void *private_data;
3927         u8 offset;
3928         int ret;
3929
3930         memset(&req, 0, sizeof req);
3931         offset = cma_user_data_offset(id_priv);
3932         req.private_data_len = offset + conn_param->private_data_len;
3933         if (req.private_data_len < conn_param->private_data_len)
3934                 return -EINVAL;
3935
3936         if (req.private_data_len) {
3937                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3938                 if (!private_data)
3939                         return -ENOMEM;
3940         } else {
3941                 private_data = NULL;
3942         }
3943
3944         if (conn_param->private_data && conn_param->private_data_len)
3945                 memcpy(private_data + offset, conn_param->private_data,
3946                        conn_param->private_data_len);
3947
3948         if (private_data) {
3949                 ret = cma_format_hdr(private_data, id_priv);
3950                 if (ret)
3951                         goto out;
3952                 req.private_data = private_data;
3953         }
3954
3955         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3956                              id_priv);
3957         if (IS_ERR(id)) {
3958                 ret = PTR_ERR(id);
3959                 goto out;
3960         }
3961         id_priv->cm_id.ib = id;
3962
3963         req.path = id_priv->id.route.path_rec;
3964         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3965         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3966         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3967         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3968
3969         trace_cm_send_sidr_req(id_priv);
3970         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3971         if (ret) {
3972                 ib_destroy_cm_id(id_priv->cm_id.ib);
3973                 id_priv->cm_id.ib = NULL;
3974         }
3975 out:
3976         kfree(private_data);
3977         return ret;
3978 }
3979
3980 static int cma_connect_ib(struct rdma_id_private *id_priv,
3981                           struct rdma_conn_param *conn_param)
3982 {
3983         struct ib_cm_req_param req;
3984         struct rdma_route *route;
3985         void *private_data;
3986         struct ib_cm_id *id;
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         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4009         if (IS_ERR(id)) {
4010                 ret = PTR_ERR(id);
4011                 goto out;
4012         }
4013         id_priv->cm_id.ib = id;
4014
4015         route = &id_priv->id.route;
4016         if (private_data) {
4017                 ret = cma_format_hdr(private_data, id_priv);
4018                 if (ret)
4019                         goto out;
4020                 req.private_data = private_data;
4021         }
4022
4023         req.primary_path = &route->path_rec[0];
4024         if (route->num_paths == 2)
4025                 req.alternate_path = &route->path_rec[1];
4026
4027         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4028         /* Alternate path SGID attribute currently unsupported */
4029         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4030         req.qp_num = id_priv->qp_num;
4031         req.qp_type = id_priv->id.qp_type;
4032         req.starting_psn = id_priv->seq_num;
4033         req.responder_resources = conn_param->responder_resources;
4034         req.initiator_depth = conn_param->initiator_depth;
4035         req.flow_control = conn_param->flow_control;
4036         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4037         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4038         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4039         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4040         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4041         req.srq = id_priv->srq ? 1 : 0;
4042         req.ece.vendor_id = id_priv->ece.vendor_id;
4043         req.ece.attr_mod = id_priv->ece.attr_mod;
4044
4045         trace_cm_send_req(id_priv);
4046         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4047 out:
4048         if (ret && !IS_ERR(id)) {
4049                 ib_destroy_cm_id(id);
4050                 id_priv->cm_id.ib = NULL;
4051         }
4052
4053         kfree(private_data);
4054         return ret;
4055 }
4056
4057 static int cma_connect_iw(struct rdma_id_private *id_priv,
4058                           struct rdma_conn_param *conn_param)
4059 {
4060         struct iw_cm_id *cm_id;
4061         int ret;
4062         struct iw_cm_conn_param iw_param;
4063
4064         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4065         if (IS_ERR(cm_id))
4066                 return PTR_ERR(cm_id);
4067
4068         cm_id->tos = id_priv->tos;
4069         cm_id->tos_set = id_priv->tos_set;
4070         id_priv->cm_id.iw = cm_id;
4071
4072         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4073                rdma_addr_size(cma_src_addr(id_priv)));
4074         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4075                rdma_addr_size(cma_dst_addr(id_priv)));
4076
4077         ret = cma_modify_qp_rtr(id_priv, conn_param);
4078         if (ret)
4079                 goto out;
4080
4081         if (conn_param) {
4082                 iw_param.ord = conn_param->initiator_depth;
4083                 iw_param.ird = conn_param->responder_resources;
4084                 iw_param.private_data = conn_param->private_data;
4085                 iw_param.private_data_len = conn_param->private_data_len;
4086                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4087         } else {
4088                 memset(&iw_param, 0, sizeof iw_param);
4089                 iw_param.qpn = id_priv->qp_num;
4090         }
4091         ret = iw_cm_connect(cm_id, &iw_param);
4092 out:
4093         if (ret) {
4094                 iw_destroy_cm_id(cm_id);
4095                 id_priv->cm_id.iw = NULL;
4096         }
4097         return ret;
4098 }
4099
4100 /**
4101  * rdma_connect_locked - Initiate an active connection request.
4102  * @id: Connection identifier to connect.
4103  * @conn_param: Connection information used for connected QPs.
4104  *
4105  * Same as rdma_connect() but can only be called from the
4106  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4107  */
4108 int rdma_connect_locked(struct rdma_cm_id *id,
4109                         struct rdma_conn_param *conn_param)
4110 {
4111         struct rdma_id_private *id_priv =
4112                 container_of(id, struct rdma_id_private, id);
4113         int ret;
4114
4115         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4116                 return -EINVAL;
4117
4118         if (!id->qp) {
4119                 id_priv->qp_num = conn_param->qp_num;
4120                 id_priv->srq = conn_param->srq;
4121         }
4122
4123         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4124                 if (id->qp_type == IB_QPT_UD)
4125                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4126                 else
4127                         ret = cma_connect_ib(id_priv, conn_param);
4128         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4129                 ret = cma_connect_iw(id_priv, conn_param);
4130         else
4131                 ret = -ENOSYS;
4132         if (ret)
4133                 goto err_state;
4134         return 0;
4135 err_state:
4136         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4137         return ret;
4138 }
4139 EXPORT_SYMBOL(rdma_connect_locked);
4140
4141 /**
4142  * rdma_connect - Initiate an active connection request.
4143  * @id: Connection identifier to connect.
4144  * @conn_param: Connection information used for connected QPs.
4145  *
4146  * Users must have resolved a route for the rdma_cm_id to connect with by having
4147  * called rdma_resolve_route before calling this routine.
4148  *
4149  * This call will either connect to a remote QP or obtain remote QP information
4150  * for unconnected rdma_cm_id's.  The actual operation is based on the
4151  * rdma_cm_id's port space.
4152  */
4153 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4154 {
4155         struct rdma_id_private *id_priv =
4156                 container_of(id, struct rdma_id_private, id);
4157         int ret;
4158
4159         mutex_lock(&id_priv->handler_mutex);
4160         ret = rdma_connect_locked(id, conn_param);
4161         mutex_unlock(&id_priv->handler_mutex);
4162         return ret;
4163 }
4164 EXPORT_SYMBOL(rdma_connect);
4165
4166 /**
4167  * rdma_connect_ece - Initiate an active connection request with ECE data.
4168  * @id: Connection identifier to connect.
4169  * @conn_param: Connection information used for connected QPs.
4170  * @ece: ECE parameters
4171  *
4172  * See rdma_connect() explanation.
4173  */
4174 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4175                      struct rdma_ucm_ece *ece)
4176 {
4177         struct rdma_id_private *id_priv =
4178                 container_of(id, struct rdma_id_private, id);
4179
4180         id_priv->ece.vendor_id = ece->vendor_id;
4181         id_priv->ece.attr_mod = ece->attr_mod;
4182
4183         return rdma_connect(id, conn_param);
4184 }
4185 EXPORT_SYMBOL(rdma_connect_ece);
4186
4187 static int cma_accept_ib(struct rdma_id_private *id_priv,
4188                          struct rdma_conn_param *conn_param)
4189 {
4190         struct ib_cm_rep_param rep;
4191         int ret;
4192
4193         ret = cma_modify_qp_rtr(id_priv, conn_param);
4194         if (ret)
4195                 goto out;
4196
4197         ret = cma_modify_qp_rts(id_priv, conn_param);
4198         if (ret)
4199                 goto out;
4200
4201         memset(&rep, 0, sizeof rep);
4202         rep.qp_num = id_priv->qp_num;
4203         rep.starting_psn = id_priv->seq_num;
4204         rep.private_data = conn_param->private_data;
4205         rep.private_data_len = conn_param->private_data_len;
4206         rep.responder_resources = conn_param->responder_resources;
4207         rep.initiator_depth = conn_param->initiator_depth;
4208         rep.failover_accepted = 0;
4209         rep.flow_control = conn_param->flow_control;
4210         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4211         rep.srq = id_priv->srq ? 1 : 0;
4212         rep.ece.vendor_id = id_priv->ece.vendor_id;
4213         rep.ece.attr_mod = id_priv->ece.attr_mod;
4214
4215         trace_cm_send_rep(id_priv);
4216         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4217 out:
4218         return ret;
4219 }
4220
4221 static int cma_accept_iw(struct rdma_id_private *id_priv,
4222                   struct rdma_conn_param *conn_param)
4223 {
4224         struct iw_cm_conn_param iw_param;
4225         int ret;
4226
4227         if (!conn_param)
4228                 return -EINVAL;
4229
4230         ret = cma_modify_qp_rtr(id_priv, conn_param);
4231         if (ret)
4232                 return ret;
4233
4234         iw_param.ord = conn_param->initiator_depth;
4235         iw_param.ird = conn_param->responder_resources;
4236         iw_param.private_data = conn_param->private_data;
4237         iw_param.private_data_len = conn_param->private_data_len;
4238         if (id_priv->id.qp) {
4239                 iw_param.qpn = id_priv->qp_num;
4240         } else
4241                 iw_param.qpn = conn_param->qp_num;
4242
4243         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4244 }
4245
4246 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4247                              enum ib_cm_sidr_status status, u32 qkey,
4248                              const void *private_data, int private_data_len)
4249 {
4250         struct ib_cm_sidr_rep_param rep;
4251         int ret;
4252
4253         memset(&rep, 0, sizeof rep);
4254         rep.status = status;
4255         if (status == IB_SIDR_SUCCESS) {
4256                 ret = cma_set_qkey(id_priv, qkey);
4257                 if (ret)
4258                         return ret;
4259                 rep.qp_num = id_priv->qp_num;
4260                 rep.qkey = id_priv->qkey;
4261
4262                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4263                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4264         }
4265
4266         rep.private_data = private_data;
4267         rep.private_data_len = private_data_len;
4268
4269         trace_cm_send_sidr_rep(id_priv);
4270         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4271 }
4272
4273 /**
4274  * rdma_accept - Called to accept a connection request or response.
4275  * @id: Connection identifier associated with the request.
4276  * @conn_param: Information needed to establish the connection.  This must be
4277  *   provided if accepting a connection request.  If accepting a connection
4278  *   response, this parameter must be NULL.
4279  *
4280  * Typically, this routine is only called by the listener to accept a connection
4281  * request.  It must also be called on the active side of a connection if the
4282  * user is performing their own QP transitions.
4283  *
4284  * In the case of error, a reject message is sent to the remote side and the
4285  * state of the qp associated with the id is modified to error, such that any
4286  * previously posted receive buffers would be flushed.
4287  *
4288  * This function is for use by kernel ULPs and must be called from under the
4289  * handler callback.
4290  */
4291 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4292 {
4293         struct rdma_id_private *id_priv =
4294                 container_of(id, struct rdma_id_private, id);
4295         int ret;
4296
4297         lockdep_assert_held(&id_priv->handler_mutex);
4298
4299         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4300                 return -EINVAL;
4301
4302         if (!id->qp && conn_param) {
4303                 id_priv->qp_num = conn_param->qp_num;
4304                 id_priv->srq = conn_param->srq;
4305         }
4306
4307         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4308                 if (id->qp_type == IB_QPT_UD) {
4309                         if (conn_param)
4310                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4311                                                         conn_param->qkey,
4312                                                         conn_param->private_data,
4313                                                         conn_param->private_data_len);
4314                         else
4315                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4316                                                         0, NULL, 0);
4317                 } else {
4318                         if (conn_param)
4319                                 ret = cma_accept_ib(id_priv, conn_param);
4320                         else
4321                                 ret = cma_rep_recv(id_priv);
4322                 }
4323         } else if (rdma_cap_iw_cm(id->device, id->port_num))
4324                 ret = cma_accept_iw(id_priv, conn_param);
4325         else
4326                 ret = -ENOSYS;
4327
4328         if (ret)
4329                 goto reject;
4330
4331         return 0;
4332 reject:
4333         cma_modify_qp_err(id_priv);
4334         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4335         return ret;
4336 }
4337 EXPORT_SYMBOL(rdma_accept);
4338
4339 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4340                     struct rdma_ucm_ece *ece)
4341 {
4342         struct rdma_id_private *id_priv =
4343                 container_of(id, struct rdma_id_private, id);
4344
4345         id_priv->ece.vendor_id = ece->vendor_id;
4346         id_priv->ece.attr_mod = ece->attr_mod;
4347
4348         return rdma_accept(id, conn_param);
4349 }
4350 EXPORT_SYMBOL(rdma_accept_ece);
4351
4352 void rdma_lock_handler(struct rdma_cm_id *id)
4353 {
4354         struct rdma_id_private *id_priv =
4355                 container_of(id, struct rdma_id_private, id);
4356
4357         mutex_lock(&id_priv->handler_mutex);
4358 }
4359 EXPORT_SYMBOL(rdma_lock_handler);
4360
4361 void rdma_unlock_handler(struct rdma_cm_id *id)
4362 {
4363         struct rdma_id_private *id_priv =
4364                 container_of(id, struct rdma_id_private, id);
4365
4366         mutex_unlock(&id_priv->handler_mutex);
4367 }
4368 EXPORT_SYMBOL(rdma_unlock_handler);
4369
4370 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4371 {
4372         struct rdma_id_private *id_priv;
4373         int ret;
4374
4375         id_priv = container_of(id, struct rdma_id_private, id);
4376         if (!id_priv->cm_id.ib)
4377                 return -EINVAL;
4378
4379         switch (id->device->node_type) {
4380         case RDMA_NODE_IB_CA:
4381                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4382                 break;
4383         default:
4384                 ret = 0;
4385                 break;
4386         }
4387         return ret;
4388 }
4389 EXPORT_SYMBOL(rdma_notify);
4390
4391 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4392                 u8 private_data_len, u8 reason)
4393 {
4394         struct rdma_id_private *id_priv;
4395         int ret;
4396
4397         id_priv = container_of(id, struct rdma_id_private, id);
4398         if (!id_priv->cm_id.ib)
4399                 return -EINVAL;
4400
4401         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4402                 if (id->qp_type == IB_QPT_UD) {
4403                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4404                                                 private_data, private_data_len);
4405                 } else {
4406                         trace_cm_send_rej(id_priv);
4407                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4408                                              private_data, private_data_len);
4409                 }
4410         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4411                 ret = iw_cm_reject(id_priv->cm_id.iw,
4412                                    private_data, private_data_len);
4413         } else
4414                 ret = -ENOSYS;
4415
4416         return ret;
4417 }
4418 EXPORT_SYMBOL(rdma_reject);
4419
4420 int rdma_disconnect(struct rdma_cm_id *id)
4421 {
4422         struct rdma_id_private *id_priv;
4423         int ret;
4424
4425         id_priv = container_of(id, struct rdma_id_private, id);
4426         if (!id_priv->cm_id.ib)
4427                 return -EINVAL;
4428
4429         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4430                 ret = cma_modify_qp_err(id_priv);
4431                 if (ret)
4432                         goto out;
4433                 /* Initiate or respond to a disconnect. */
4434                 trace_cm_disconnect(id_priv);
4435                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4436                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4437                                 trace_cm_sent_drep(id_priv);
4438                 } else {
4439                         trace_cm_sent_dreq(id_priv);
4440                 }
4441         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4442                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4443         } else
4444                 ret = -EINVAL;
4445
4446 out:
4447         return ret;
4448 }
4449 EXPORT_SYMBOL(rdma_disconnect);
4450
4451 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4452                               struct ib_sa_multicast *multicast,
4453                               struct rdma_cm_event *event,
4454                               struct cma_multicast *mc)
4455 {
4456         struct rdma_dev_addr *dev_addr;
4457         enum ib_gid_type gid_type;
4458         struct net_device *ndev;
4459
4460         if (!status)
4461                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4462         else
4463                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4464                                      status);
4465
4466         event->status = status;
4467         event->param.ud.private_data = mc->context;
4468         if (status) {
4469                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4470                 return;
4471         }
4472
4473         dev_addr = &id_priv->id.route.addr.dev_addr;
4474         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4475         gid_type =
4476                 id_priv->cma_dev
4477                         ->default_gid_type[id_priv->id.port_num -
4478                                            rdma_start_port(
4479                                                    id_priv->cma_dev->device)];
4480
4481         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4482         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4483                                      &multicast->rec, ndev, gid_type,
4484                                      &event->param.ud.ah_attr)) {
4485                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4486                 goto out;
4487         }
4488
4489         event->param.ud.qp_num = 0xFFFFFF;
4490         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4491
4492 out:
4493         if (ndev)
4494                 dev_put(ndev);
4495 }
4496
4497 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4498 {
4499         struct cma_multicast *mc = multicast->context;
4500         struct rdma_id_private *id_priv = mc->id_priv;
4501         struct rdma_cm_event event = {};
4502         int ret = 0;
4503
4504         mutex_lock(&id_priv->handler_mutex);
4505         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4506             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4507                 goto out;
4508
4509         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4510         ret = cma_cm_event_handler(id_priv, &event);
4511         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4512         WARN_ON(ret);
4513
4514 out:
4515         mutex_unlock(&id_priv->handler_mutex);
4516         return 0;
4517 }
4518
4519 static void cma_set_mgid(struct rdma_id_private *id_priv,
4520                          struct sockaddr *addr, union ib_gid *mgid)
4521 {
4522         unsigned char mc_map[MAX_ADDR_LEN];
4523         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4524         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4525         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4526
4527         if (cma_any_addr(addr)) {
4528                 memset(mgid, 0, sizeof *mgid);
4529         } else if ((addr->sa_family == AF_INET6) &&
4530                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4531                                                                  0xFF10A01B)) {
4532                 /* IPv6 address is an SA assigned MGID. */
4533                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4534         } else if (addr->sa_family == AF_IB) {
4535                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4536         } else if (addr->sa_family == AF_INET6) {
4537                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4538                 if (id_priv->id.ps == RDMA_PS_UDP)
4539                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4540                 *mgid = *(union ib_gid *) (mc_map + 4);
4541         } else {
4542                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4543                 if (id_priv->id.ps == RDMA_PS_UDP)
4544                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4545                 *mgid = *(union ib_gid *) (mc_map + 4);
4546         }
4547 }
4548
4549 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4550                                  struct cma_multicast *mc)
4551 {
4552         struct ib_sa_mcmember_rec rec;
4553         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4554         ib_sa_comp_mask comp_mask;
4555         int ret;
4556
4557         ib_addr_get_mgid(dev_addr, &rec.mgid);
4558         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4559                                      &rec.mgid, &rec);
4560         if (ret)
4561                 return ret;
4562
4563         ret = cma_set_qkey(id_priv, 0);
4564         if (ret)
4565                 return ret;
4566
4567         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4568         rec.qkey = cpu_to_be32(id_priv->qkey);
4569         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4570         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4571         rec.join_state = mc->join_state;
4572
4573         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4574                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4575                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4576                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4577                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4578
4579         if (id_priv->id.ps == RDMA_PS_IPOIB)
4580                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4581                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4582                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4583                              IB_SA_MCMEMBER_REC_MTU |
4584                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4585
4586         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4587                                          id_priv->id.port_num, &rec, comp_mask,
4588                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4589         return PTR_ERR_OR_ZERO(mc->sa_mc);
4590 }
4591
4592 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4593                               enum ib_gid_type gid_type)
4594 {
4595         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4596         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4597
4598         if (cma_any_addr(addr)) {
4599                 memset(mgid, 0, sizeof *mgid);
4600         } else if (addr->sa_family == AF_INET6) {
4601                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4602         } else {
4603                 mgid->raw[0] =
4604                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4605                 mgid->raw[1] =
4606                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4607                 mgid->raw[2] = 0;
4608                 mgid->raw[3] = 0;
4609                 mgid->raw[4] = 0;
4610                 mgid->raw[5] = 0;
4611                 mgid->raw[6] = 0;
4612                 mgid->raw[7] = 0;
4613                 mgid->raw[8] = 0;
4614                 mgid->raw[9] = 0;
4615                 mgid->raw[10] = 0xff;
4616                 mgid->raw[11] = 0xff;
4617                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4618         }
4619 }
4620
4621 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4622                                    struct cma_multicast *mc)
4623 {
4624         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4625         int err = 0;
4626         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4627         struct net_device *ndev = NULL;
4628         struct ib_sa_multicast ib;
4629         enum ib_gid_type gid_type;
4630         bool send_only;
4631
4632         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4633
4634         if (cma_zero_addr(addr))
4635                 return -EINVAL;
4636
4637         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4638                    rdma_start_port(id_priv->cma_dev->device)];
4639         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4640
4641         ib.rec.pkey = cpu_to_be16(0xffff);
4642         if (id_priv->id.ps == RDMA_PS_UDP)
4643                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4644
4645         if (dev_addr->bound_dev_if)
4646                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4647         if (!ndev)
4648                 return -ENODEV;
4649
4650         ib.rec.rate = iboe_get_rate(ndev);
4651         ib.rec.hop_limit = 1;
4652         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4653
4654         if (addr->sa_family == AF_INET) {
4655                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4656                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4657                         if (!send_only) {
4658                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4659                                                     true);
4660                         }
4661                 }
4662         } else {
4663                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4664                         err = -ENOTSUPP;
4665         }
4666         dev_put(ndev);
4667         if (err || !ib.rec.mtu)
4668                 return err ?: -EINVAL;
4669
4670         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4671                     &ib.rec.port_gid);
4672         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4673         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4674         queue_work(cma_wq, &mc->iboe_join.work);
4675         return 0;
4676 }
4677
4678 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4679                         u8 join_state, void *context)
4680 {
4681         struct rdma_id_private *id_priv =
4682                 container_of(id, struct rdma_id_private, id);
4683         struct cma_multicast *mc;
4684         int ret;
4685
4686         /* Not supported for kernel QPs */
4687         if (WARN_ON(id->qp))
4688                 return -EINVAL;
4689
4690         /* ULP is calling this wrong. */
4691         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4692                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4693                 return -EINVAL;
4694
4695         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4696         if (!mc)
4697                 return -ENOMEM;
4698
4699         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4700         mc->context = context;
4701         mc->id_priv = id_priv;
4702         mc->join_state = join_state;
4703
4704         if (rdma_protocol_roce(id->device, id->port_num)) {
4705                 ret = cma_iboe_join_multicast(id_priv, mc);
4706                 if (ret)
4707                         goto out_err;
4708         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4709                 ret = cma_join_ib_multicast(id_priv, mc);
4710                 if (ret)
4711                         goto out_err;
4712         } else {
4713                 ret = -ENOSYS;
4714                 goto out_err;
4715         }
4716
4717         spin_lock(&id_priv->lock);
4718         list_add(&mc->list, &id_priv->mc_list);
4719         spin_unlock(&id_priv->lock);
4720
4721         return 0;
4722 out_err:
4723         kfree(mc);
4724         return ret;
4725 }
4726 EXPORT_SYMBOL(rdma_join_multicast);
4727
4728 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4729 {
4730         struct rdma_id_private *id_priv;
4731         struct cma_multicast *mc;
4732
4733         id_priv = container_of(id, struct rdma_id_private, id);
4734         spin_lock_irq(&id_priv->lock);
4735         list_for_each_entry(mc, &id_priv->mc_list, list) {
4736                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4737                         continue;
4738                 list_del(&mc->list);
4739                 spin_unlock_irq(&id_priv->lock);
4740
4741                 WARN_ON(id_priv->cma_dev->device != id->device);
4742                 destroy_mc(id_priv, mc);
4743                 return;
4744         }
4745         spin_unlock_irq(&id_priv->lock);
4746 }
4747 EXPORT_SYMBOL(rdma_leave_multicast);
4748
4749 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4750 {
4751         struct rdma_dev_addr *dev_addr;
4752         struct cma_work *work;
4753
4754         dev_addr = &id_priv->id.route.addr.dev_addr;
4755
4756         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4757             (net_eq(dev_net(ndev), dev_addr->net)) &&
4758             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4759                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4760                         ndev->name, &id_priv->id);
4761                 work = kzalloc(sizeof *work, GFP_KERNEL);
4762                 if (!work)
4763                         return -ENOMEM;
4764
4765                 INIT_WORK(&work->work, cma_work_handler);
4766                 work->id = id_priv;
4767                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4768                 cma_id_get(id_priv);
4769                 queue_work(cma_wq, &work->work);
4770         }
4771
4772         return 0;
4773 }
4774
4775 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4776                                void *ptr)
4777 {
4778         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4779         struct cma_device *cma_dev;
4780         struct rdma_id_private *id_priv;
4781         int ret = NOTIFY_DONE;
4782
4783         if (event != NETDEV_BONDING_FAILOVER)
4784                 return NOTIFY_DONE;
4785
4786         if (!netif_is_bond_master(ndev))
4787                 return NOTIFY_DONE;
4788
4789         mutex_lock(&lock);
4790         list_for_each_entry(cma_dev, &dev_list, list)
4791                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4792                         ret = cma_netdev_change(ndev, id_priv);
4793                         if (ret)
4794                                 goto out;
4795                 }
4796
4797 out:
4798         mutex_unlock(&lock);
4799         return ret;
4800 }
4801
4802 static struct notifier_block cma_nb = {
4803         .notifier_call = cma_netdev_callback
4804 };
4805
4806 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4807 {
4808         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4809         enum rdma_cm_state state;
4810         unsigned long flags;
4811
4812         mutex_lock(&id_priv->handler_mutex);
4813         /* Record that we want to remove the device */
4814         spin_lock_irqsave(&id_priv->lock, flags);
4815         state = id_priv->state;
4816         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4817                 spin_unlock_irqrestore(&id_priv->lock, flags);
4818                 mutex_unlock(&id_priv->handler_mutex);
4819                 cma_id_put(id_priv);
4820                 return;
4821         }
4822         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4823         spin_unlock_irqrestore(&id_priv->lock, flags);
4824
4825         if (cma_cm_event_handler(id_priv, &event)) {
4826                 /*
4827                  * At this point the ULP promises it won't call
4828                  * rdma_destroy_id() concurrently
4829                  */
4830                 cma_id_put(id_priv);
4831                 mutex_unlock(&id_priv->handler_mutex);
4832                 trace_cm_id_destroy(id_priv);
4833                 _destroy_id(id_priv, state);
4834                 return;
4835         }
4836         mutex_unlock(&id_priv->handler_mutex);
4837
4838         /*
4839          * If this races with destroy then the thread that first assigns state
4840          * to a destroying does the cancel.
4841          */
4842         cma_cancel_operation(id_priv, state);
4843         cma_id_put(id_priv);
4844 }
4845
4846 static void cma_process_remove(struct cma_device *cma_dev)
4847 {
4848         mutex_lock(&lock);
4849         while (!list_empty(&cma_dev->id_list)) {
4850                 struct rdma_id_private *id_priv = list_first_entry(
4851                         &cma_dev->id_list, struct rdma_id_private, list);
4852
4853                 list_del(&id_priv->listen_list);
4854                 list_del_init(&id_priv->list);
4855                 cma_id_get(id_priv);
4856                 mutex_unlock(&lock);
4857
4858                 cma_send_device_removal_put(id_priv);
4859
4860                 mutex_lock(&lock);
4861         }
4862         mutex_unlock(&lock);
4863
4864         cma_dev_put(cma_dev);
4865         wait_for_completion(&cma_dev->comp);
4866 }
4867
4868 static int cma_add_one(struct ib_device *device)
4869 {
4870         struct rdma_id_private *to_destroy;
4871         struct cma_device *cma_dev;
4872         struct rdma_id_private *id_priv;
4873         unsigned int i;
4874         unsigned long supported_gids = 0;
4875         int ret;
4876
4877         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4878         if (!cma_dev)
4879                 return -ENOMEM;
4880
4881         cma_dev->device = device;
4882         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4883                                             sizeof(*cma_dev->default_gid_type),
4884                                             GFP_KERNEL);
4885         if (!cma_dev->default_gid_type) {
4886                 ret = -ENOMEM;
4887                 goto free_cma_dev;
4888         }
4889
4890         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4891                                             sizeof(*cma_dev->default_roce_tos),
4892                                             GFP_KERNEL);
4893         if (!cma_dev->default_roce_tos) {
4894                 ret = -ENOMEM;
4895                 goto free_gid_type;
4896         }
4897
4898         rdma_for_each_port (device, i) {
4899                 supported_gids = roce_gid_type_mask_support(device, i);
4900                 WARN_ON(!supported_gids);
4901                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4902                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4903                                 CMA_PREFERRED_ROCE_GID_TYPE;
4904                 else
4905                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
4906                                 find_first_bit(&supported_gids, BITS_PER_LONG);
4907                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4908         }
4909
4910         init_completion(&cma_dev->comp);
4911         refcount_set(&cma_dev->refcount, 1);
4912         INIT_LIST_HEAD(&cma_dev->id_list);
4913         ib_set_client_data(device, &cma_client, cma_dev);
4914
4915         mutex_lock(&lock);
4916         list_add_tail(&cma_dev->list, &dev_list);
4917         list_for_each_entry(id_priv, &listen_any_list, list) {
4918                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
4919                 if (ret)
4920                         goto free_listen;
4921         }
4922         mutex_unlock(&lock);
4923
4924         trace_cm_add_one(device);
4925         return 0;
4926
4927 free_listen:
4928         list_del(&cma_dev->list);
4929         mutex_unlock(&lock);
4930
4931         /* cma_process_remove() will delete to_destroy */
4932         cma_process_remove(cma_dev);
4933         kfree(cma_dev->default_roce_tos);
4934 free_gid_type:
4935         kfree(cma_dev->default_gid_type);
4936
4937 free_cma_dev:
4938         kfree(cma_dev);
4939         return ret;
4940 }
4941
4942 static void cma_remove_one(struct ib_device *device, void *client_data)
4943 {
4944         struct cma_device *cma_dev = client_data;
4945
4946         trace_cm_remove_one(device);
4947
4948         mutex_lock(&lock);
4949         list_del(&cma_dev->list);
4950         mutex_unlock(&lock);
4951
4952         cma_process_remove(cma_dev);
4953         kfree(cma_dev->default_roce_tos);
4954         kfree(cma_dev->default_gid_type);
4955         kfree(cma_dev);
4956 }
4957
4958 static int cma_init_net(struct net *net)
4959 {
4960         struct cma_pernet *pernet = cma_pernet(net);
4961
4962         xa_init(&pernet->tcp_ps);
4963         xa_init(&pernet->udp_ps);
4964         xa_init(&pernet->ipoib_ps);
4965         xa_init(&pernet->ib_ps);
4966
4967         return 0;
4968 }
4969
4970 static void cma_exit_net(struct net *net)
4971 {
4972         struct cma_pernet *pernet = cma_pernet(net);
4973
4974         WARN_ON(!xa_empty(&pernet->tcp_ps));
4975         WARN_ON(!xa_empty(&pernet->udp_ps));
4976         WARN_ON(!xa_empty(&pernet->ipoib_ps));
4977         WARN_ON(!xa_empty(&pernet->ib_ps));
4978 }
4979
4980 static struct pernet_operations cma_pernet_operations = {
4981         .init = cma_init_net,
4982         .exit = cma_exit_net,
4983         .id = &cma_pernet_id,
4984         .size = sizeof(struct cma_pernet),
4985 };
4986
4987 static int __init cma_init(void)
4988 {
4989         int ret;
4990
4991         /*
4992          * There is a rare lock ordering dependency in cma_netdev_callback()
4993          * that only happens when bonding is enabled. Teach lockdep that rtnl
4994          * must never be nested under lock so it can find these without having
4995          * to test with bonding.
4996          */
4997         if (IS_ENABLED(CONFIG_LOCKDEP)) {
4998                 rtnl_lock();
4999                 mutex_lock(&lock);
5000                 mutex_unlock(&lock);
5001                 rtnl_unlock();
5002         }
5003
5004         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5005         if (!cma_wq)
5006                 return -ENOMEM;
5007
5008         ret = register_pernet_subsys(&cma_pernet_operations);
5009         if (ret)
5010                 goto err_wq;
5011
5012         ib_sa_register_client(&sa_client);
5013         register_netdevice_notifier(&cma_nb);
5014
5015         ret = ib_register_client(&cma_client);
5016         if (ret)
5017                 goto err;
5018
5019         ret = cma_configfs_init();
5020         if (ret)
5021                 goto err_ib;
5022
5023         return 0;
5024
5025 err_ib:
5026         ib_unregister_client(&cma_client);
5027 err:
5028         unregister_netdevice_notifier(&cma_nb);
5029         ib_sa_unregister_client(&sa_client);
5030         unregister_pernet_subsys(&cma_pernet_operations);
5031 err_wq:
5032         destroy_workqueue(cma_wq);
5033         return ret;
5034 }
5035
5036 static void __exit cma_cleanup(void)
5037 {
5038         cma_configfs_exit();
5039         ib_unregister_client(&cma_client);
5040         unregister_netdevice_notifier(&cma_nb);
5041         ib_sa_unregister_client(&sa_client);
5042         unregister_pernet_subsys(&cma_pernet_operations);
5043         destroy_workqueue(cma_wq);
5044 }
5045
5046 module_init(cma_init);
5047 module_exit(cma_cleanup);