drivers/infiniband/core/cma.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   4  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   5  * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
   6  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   7  */
   8
   9 #include <linux/completion.h>
  10 #include <linux/in.h>
  11 #include <linux/in6.h>
  12 #include <linux/mutex.h>
  13 #include <linux/random.h>
  14 #include <linux/igmp.h>
  15 #include <linux/xarray.h>
  16 #include <linux/inetdevice.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <net/route.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/netns/generic.h>
  23 #include <net/tcp.h>
  24 #include <net/ipv6.h>
  25 #include <net/ip_fib.h>
  26 #include <net/ip6_route.h>
  27
  28 #include <rdma/rdma_cm.h>
  29 #include <rdma/rdma_cm_ib.h>
  30 #include <rdma/rdma_netlink.h>
  31 #include <rdma/ib.h>
  32 #include <rdma/ib_cache.h>
  33 #include <rdma/ib_cm.h>
  34 #include <rdma/ib_sa.h>
  35 #include <rdma/iw_cm.h>
  36
  37 #include "core_priv.h"
  38 #include "cma_priv.h"
  39 #include "cma_trace.h"
  40
  41 MODULE_AUTHOR("Sean Hefty");
  42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  43 MODULE_LICENSE("Dual BSD/GPL");
  44
  45 #define CMA_CM_RESPONSE_TIMEOUT 20
  46 #define CMA_MAX_CM_RETRIES 15
  47 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  48 #define CMA_IBOE_PACKET_LIFETIME 18
  49 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
  50
  51 static const char * const cma_events[] = {
  52         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  53         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  54         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  55         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  56         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  57         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  58         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  59         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  60         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  61         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  62         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  63         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  64         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  65         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  66         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  67         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  68 };
  69
  70 static void cma_set_mgid(struct rdma_id_private *id_priv, struct sockaddr *addr,
  71                          union ib_gid *mgid);
  72
  73 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  74 {
  75         size_t index = event;
  76
  77         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
  78                         cma_events[index] : "unrecognized event";
  79 }
  80 EXPORT_SYMBOL(rdma_event_msg);
  81
  82 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
  83                                                 int reason)
  84 {
  85         if (rdma_ib_or_roce(id->device, id->port_num))
  86                 return ibcm_reject_msg(reason);
  87
  88         if (rdma_protocol_iwarp(id->device, id->port_num))
  89                 return iwcm_reject_msg(reason);
  90
  91         WARN_ON_ONCE(1);
  92         return "unrecognized transport";
  93 }
  94 EXPORT_SYMBOL(rdma_reject_msg);
  95
  96 /**
  97  * rdma_is_consumer_reject - return true if the consumer rejected the connect
  98  *                           request.
  99  * @id: Communication identifier that received the REJECT event.
 100  * @reason: Value returned in the REJECT event status field.
 101  */
 102 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
 103 {
 104         if (rdma_ib_or_roce(id->device, id->port_num))
 105                 return reason == IB_CM_REJ_CONSUMER_DEFINED;
 106
 107         if (rdma_protocol_iwarp(id->device, id->port_num))
 108                 return reason == -ECONNREFUSED;
 109
 110         WARN_ON_ONCE(1);
 111         return false;
 112 }
 113
 114 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
 115                                       struct rdma_cm_event *ev, u8 *data_len)
 116 {
 117         const void *p;
 118
 119         if (rdma_is_consumer_reject(id, ev->status)) {
 120                 *data_len = ev->param.conn.private_data_len;
 121                 p = ev->param.conn.private_data;
 122         } else {
 123                 *data_len = 0;
 124                 p = NULL;
 125         }
 126         return p;
 127 }
 128 EXPORT_SYMBOL(rdma_consumer_reject_data);
 129
 130 /**
 131  * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
 132  * @id: Communication Identifier
 133  */
 134 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
 135 {
 136         struct rdma_id_private *id_priv;
 137
 138         id_priv = container_of(id, struct rdma_id_private, id);
 139         if (id->device->node_type == RDMA_NODE_RNIC)
 140                 return id_priv->cm_id.iw;
 141         return NULL;
 142 }
 143 EXPORT_SYMBOL(rdma_iw_cm_id);
 144
 145 /**
 146  * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
 147  * @res: rdma resource tracking entry pointer
 148  */
 149 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
 150 {
 151         struct rdma_id_private *id_priv =
 152                 container_of(res, struct rdma_id_private, res);
 153
 154         return &id_priv->id;
 155 }
 156 EXPORT_SYMBOL(rdma_res_to_id);
 157
 158 static int cma_add_one(struct ib_device *device);
 159 static void cma_remove_one(struct ib_device *device, void *client_data);
 160
 161 static struct ib_client cma_client = {
 162         .name   = "cma",
 163         .add    = cma_add_one,
 164         .remove = cma_remove_one
 165 };
 166
 167 static struct ib_sa_client sa_client;
 168 static LIST_HEAD(dev_list);
 169 static LIST_HEAD(listen_any_list);
 170 static DEFINE_MUTEX(lock);
 171 static struct workqueue_struct *cma_wq;
 172 static unsigned int cma_pernet_id;
 173
 174 struct cma_pernet {
 175         struct xarray tcp_ps;
 176         struct xarray udp_ps;
 177         struct xarray ipoib_ps;
 178         struct xarray ib_ps;
 179 };
 180
 181 static struct cma_pernet *cma_pernet(struct net *net)
 182 {
 183         return net_generic(net, cma_pernet_id);
 184 }
 185
 186 static
 187 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
 188 {
 189         struct cma_pernet *pernet = cma_pernet(net);
 190
 191         switch (ps) {
 192         case RDMA_PS_TCP:
 193                 return &pernet->tcp_ps;
 194         case RDMA_PS_UDP:
 195                 return &pernet->udp_ps;
 196         case RDMA_PS_IPOIB:
 197                 return &pernet->ipoib_ps;
 198         case RDMA_PS_IB:
 199                 return &pernet->ib_ps;
 200         default:
 201                 return NULL;
 202         }
 203 }
 204
 205 struct cma_device {
 206         struct list_head        list;
 207         struct ib_device        *device;
 208         struct completion       comp;
 209         refcount_t refcount;
 210         struct list_head        id_list;
 211         enum ib_gid_type        *default_gid_type;
 212         u8                      *default_roce_tos;
 213 };
 214
 215 struct rdma_bind_list {
 216         enum rdma_ucm_port_space ps;
 217         struct hlist_head       owners;
 218         unsigned short          port;
 219 };
 220
 221 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
 222                         struct rdma_bind_list *bind_list, int snum)
 223 {
 224         struct xarray *xa = cma_pernet_xa(net, ps);
 225
 226         return xa_insert(xa, snum, bind_list, GFP_KERNEL);
 227 }
 228
 229 static struct rdma_bind_list *cma_ps_find(struct net *net,
 230                                           enum rdma_ucm_port_space ps, int snum)
 231 {
 232         struct xarray *xa = cma_pernet_xa(net, ps);
 233
 234         return xa_load(xa, snum);
 235 }
 236
 237 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
 238                           int snum)
 239 {
 240         struct xarray *xa = cma_pernet_xa(net, ps);
 241
 242         xa_erase(xa, snum);
 243 }
 244
 245 enum {
 246         CMA_OPTION_AFONLY,
 247 };
 248
 249 void cma_dev_get(struct cma_device *cma_dev)
 250 {
 251         refcount_inc(&cma_dev->refcount);
 252 }
 253
 254 void cma_dev_put(struct cma_device *cma_dev)
 255 {
 256         if (refcount_dec_and_test(&cma_dev->refcount))
 257                 complete(&cma_dev->comp);
 258 }
 259
 260 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 261                                              void               *cookie)
 262 {
 263         struct cma_device *cma_dev;
 264         struct cma_device *found_cma_dev = NULL;
 265
 266         mutex_lock(&lock);
 267
 268         list_for_each_entry(cma_dev, &dev_list, list)
 269                 if (filter(cma_dev->device, cookie)) {
 270                         found_cma_dev = cma_dev;
 271                         break;
 272                 }
 273
 274         if (found_cma_dev)
 275                 cma_dev_get(found_cma_dev);
 276         mutex_unlock(&lock);
 277         return found_cma_dev;
 278 }
 279
 280 int cma_get_default_gid_type(struct cma_device *cma_dev,
 281                              u32 port)
 282 {
 283         if (!rdma_is_port_valid(cma_dev->device, port))
 284                 return -EINVAL;
 285
 286         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 287 }
 288
 289 int cma_set_default_gid_type(struct cma_device *cma_dev,
 290                              u32 port,
 291                              enum ib_gid_type default_gid_type)
 292 {
 293         unsigned long supported_gids;
 294
 295         if (!rdma_is_port_valid(cma_dev->device, port))
 296                 return -EINVAL;
 297
 298         if (default_gid_type == IB_GID_TYPE_IB &&
 299             rdma_protocol_roce_eth_encap(cma_dev->device, port))
 300                 default_gid_type = IB_GID_TYPE_ROCE;
 301
 302         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 303
 304         if (!(supported_gids & 1 << default_gid_type))
 305                 return -EINVAL;
 306
 307         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 308                 default_gid_type;
 309
 310         return 0;
 311 }
 312
 313 int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
 314 {
 315         if (!rdma_is_port_valid(cma_dev->device, port))
 316                 return -EINVAL;
 317
 318         return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
 319 }
 320
 321 int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
 322                              u8 default_roce_tos)
 323 {
 324         if (!rdma_is_port_valid(cma_dev->device, port))
 325                 return -EINVAL;
 326
 327         cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
 328                  default_roce_tos;
 329
 330         return 0;
 331 }
 332 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 333 {
 334         return cma_dev->device;
 335 }
 336
 337 /*
 338  * Device removal can occur at anytime, so we need extra handling to
 339  * serialize notifying the user of device removal with other callbacks.
 340  * We do this by disabling removal notification while a callback is in process,
 341  * and reporting it after the callback completes.
 342  */
 343
 344 struct cma_multicast {
 345         struct rdma_id_private *id_priv;
 346         union {
 347                 struct ib_sa_multicast *sa_mc;
 348                 struct {
 349                         struct work_struct work;
 350                         struct rdma_cm_event event;
 351                 } iboe_join;
 352         };
 353         struct list_head        list;
 354         void                    *context;
 355         struct sockaddr_storage addr;
 356         u8                      join_state;
 357 };
 358
 359 struct cma_work {
 360         struct work_struct      work;
 361         struct rdma_id_private  *id;
 362         enum rdma_cm_state      old_state;
 363         enum rdma_cm_state      new_state;
 364         struct rdma_cm_event    event;
 365 };
 366
 367 union cma_ip_addr {
 368         struct in6_addr ip6;
 369         struct {
 370                 __be32 pad[3];
 371                 __be32 addr;
 372         } ip4;
 373 };
 374
 375 struct cma_hdr {
 376         u8 cma_version;
 377         u8 ip_version;  /* IP version: 7:4 */
 378         __be16 port;
 379         union cma_ip_addr src_addr;
 380         union cma_ip_addr dst_addr;
 381 };
 382
 383 #define CMA_VERSION 0x00
 384
 385 struct cma_req_info {
 386         struct sockaddr_storage listen_addr_storage;
 387         struct sockaddr_storage src_addr_storage;
 388         struct ib_device *device;
 389         union ib_gid local_gid;
 390         __be64 service_id;
 391         int port;
 392         bool has_gid;
 393         u16 pkey;
 394 };
 395
 396 static int cma_comp_exch(struct rdma_id_private *id_priv,
 397                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 398 {
 399         unsigned long flags;
 400         int ret;
 401
 402         /*
 403          * The FSM uses a funny double locking where state is protected by both
 404          * the handler_mutex and the spinlock. State is not allowed to change
 405          * to/from a handler_mutex protected value without also holding
 406          * handler_mutex.
 407          */
 408         if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
 409                 lockdep_assert_held(&id_priv->handler_mutex);
 410
 411         spin_lock_irqsave(&id_priv->lock, flags);
 412         if ((ret = (id_priv->state == comp)))
 413                 id_priv->state = exch;
 414         spin_unlock_irqrestore(&id_priv->lock, flags);
 415         return ret;
 416 }
 417
 418 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 419 {
 420         return hdr->ip_version >> 4;
 421 }
 422
 423 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 424 {
 425         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 426 }
 427
 428 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 429 {
 430         struct in_device *in_dev = NULL;
 431
 432         if (ndev) {
 433                 rtnl_lock();
 434                 in_dev = __in_dev_get_rtnl(ndev);
 435                 if (in_dev) {
 436                         if (join)
 437                                 ip_mc_inc_group(in_dev,
 438                                                 *(__be32 *)(mgid->raw + 12));
 439                         else
 440                                 ip_mc_dec_group(in_dev,
 441                                                 *(__be32 *)(mgid->raw + 12));
 442                 }
 443                 rtnl_unlock();
 444         }
 445         return (in_dev) ? 0 : -ENODEV;
 446 }
 447
 448 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 449                                struct cma_device *cma_dev)
 450 {
 451         cma_dev_get(cma_dev);
 452         id_priv->cma_dev = cma_dev;
 453         id_priv->id.device = cma_dev->device;
 454         id_priv->id.route.addr.dev_addr.transport =
 455                 rdma_node_get_transport(cma_dev->device->node_type);
 456         list_add_tail(&id_priv->list, &cma_dev->id_list);
 457
 458         trace_cm_id_attach(id_priv, cma_dev->device);
 459 }
 460
 461 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 462                               struct cma_device *cma_dev)
 463 {
 464         _cma_attach_to_dev(id_priv, cma_dev);
 465         id_priv->gid_type =
 466                 cma_dev->default_gid_type[id_priv->id.port_num -
 467                                           rdma_start_port(cma_dev->device)];
 468 }
 469
 470 static void cma_release_dev(struct rdma_id_private *id_priv)
 471 {
 472         mutex_lock(&lock);
 473         list_del(&id_priv->list);
 474         cma_dev_put(id_priv->cma_dev);
 475         id_priv->cma_dev = NULL;
 476         id_priv->id.device = NULL;
 477         if (id_priv->id.route.addr.dev_addr.sgid_attr) {
 478                 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
 479                 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
 480         }
 481         mutex_unlock(&lock);
 482 }
 483
 484 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 485 {
 486         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 487 }
 488
 489 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 490 {
 491         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 492 }
 493
 494 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 495 {
 496         return id_priv->id.route.addr.src_addr.ss_family;
 497 }
 498
 499 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 500 {
 501         struct ib_sa_mcmember_rec rec;
 502         int ret = 0;
 503
 504         if (id_priv->qkey) {
 505                 if (qkey && id_priv->qkey != qkey)
 506                         return -EINVAL;
 507                 return 0;
 508         }
 509
 510         if (qkey) {
 511                 id_priv->qkey = qkey;
 512                 return 0;
 513         }
 514
 515         switch (id_priv->id.ps) {
 516         case RDMA_PS_UDP:
 517         case RDMA_PS_IB:
 518                 id_priv->qkey = RDMA_UDP_QKEY;
 519                 break;
 520         case RDMA_PS_IPOIB:
 521                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 522                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 523                                              id_priv->id.port_num, &rec.mgid,
 524                                              &rec);
 525                 if (!ret)
 526                         id_priv->qkey = be32_to_cpu(rec.qkey);
 527                 break;
 528         default:
 529                 break;
 530         }
 531         return ret;
 532 }
 533
 534 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 535 {
 536         dev_addr->dev_type = ARPHRD_INFINIBAND;
 537         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 538         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 539 }
 540
 541 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 542 {
 543         int ret;
 544
 545         if (addr->sa_family != AF_IB) {
 546                 ret = rdma_translate_ip(addr, dev_addr);
 547         } else {
 548                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 549                 ret = 0;
 550         }
 551
 552         return ret;
 553 }
 554
 555 static const struct ib_gid_attr *
 556 cma_validate_port(struct ib_device *device, u32 port,
 557                   enum ib_gid_type gid_type,
 558                   union ib_gid *gid,
 559                   struct rdma_id_private *id_priv)
 560 {
 561         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 562         int bound_if_index = dev_addr->bound_dev_if;
 563         const struct ib_gid_attr *sgid_attr;
 564         int dev_type = dev_addr->dev_type;
 565         struct net_device *ndev = NULL;
 566
 567         if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
 568                 return ERR_PTR(-ENODEV);
 569
 570         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 571                 return ERR_PTR(-ENODEV);
 572
 573         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 574                 return ERR_PTR(-ENODEV);
 575
 576         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 577                 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
 578                 if (!ndev)
 579                         return ERR_PTR(-ENODEV);
 580         } else {
 581                 gid_type = IB_GID_TYPE_IB;
 582         }
 583
 584         sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
 585         if (ndev)
 586                 dev_put(ndev);
 587         return sgid_attr;
 588 }
 589
 590 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
 591                                const struct ib_gid_attr *sgid_attr)
 592 {
 593         WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
 594         id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
 595 }
 596
 597 /**
 598  * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
 599  * based on source ip address.
 600  * @id_priv:    cm_id which should be bound to cma device
 601  *
 602  * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
 603  * based on source IP address. It returns 0 on success or error code otherwise.
 604  * It is applicable to active and passive side cm_id.
 605  */
 606 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
 607 {
 608         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 609         const struct ib_gid_attr *sgid_attr;
 610         union ib_gid gid, iboe_gid, *gidp;
 611         struct cma_device *cma_dev;
 612         enum ib_gid_type gid_type;
 613         int ret = -ENODEV;
 614         u32 port;
 615
 616         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 617             id_priv->id.ps == RDMA_PS_IPOIB)
 618                 return -EINVAL;
 619
 620         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 621                     &iboe_gid);
 622
 623         memcpy(&gid, dev_addr->src_dev_addr +
 624                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 625
 626         mutex_lock(&lock);
 627         list_for_each_entry(cma_dev, &dev_list, list) {
 628                 rdma_for_each_port (cma_dev->device, port) {
 629                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 630                                &iboe_gid : &gid;
 631                         gid_type = cma_dev->default_gid_type[port - 1];
 632                         sgid_attr = cma_validate_port(cma_dev->device, port,
 633                                                       gid_type, gidp, id_priv);
 634                         if (!IS_ERR(sgid_attr)) {
 635                                 id_priv->id.port_num = port;
 636                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 637                                 cma_attach_to_dev(id_priv, cma_dev);
 638                                 ret = 0;
 639                                 goto out;
 640                         }
 641                 }
 642         }
 643 out:
 644         mutex_unlock(&lock);
 645         return ret;
 646 }
 647
 648 /**
 649  * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
 650  * @id_priv:            cm id to bind to cma device
 651  * @listen_id_priv:     listener cm id to match against
 652  * @req:                Pointer to req structure containaining incoming
 653  *                      request information
 654  * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
 655  * rdma device matches for listen_id and incoming request. It also verifies
 656  * that a GID table entry is present for the source address.
 657  * Returns 0 on success, or returns error code otherwise.
 658  */
 659 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
 660                               const struct rdma_id_private *listen_id_priv,
 661                               struct cma_req_info *req)
 662 {
 663         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 664         const struct ib_gid_attr *sgid_attr;
 665         enum ib_gid_type gid_type;
 666         union ib_gid gid;
 667
 668         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 669             id_priv->id.ps == RDMA_PS_IPOIB)
 670                 return -EINVAL;
 671
 672         if (rdma_protocol_roce(req->device, req->port))
 673                 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 674                             &gid);
 675         else
 676                 memcpy(&gid, dev_addr->src_dev_addr +
 677                        rdma_addr_gid_offset(dev_addr), sizeof(gid));
 678
 679         gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
 680         sgid_attr = cma_validate_port(req->device, req->port,
 681                                       gid_type, &gid, id_priv);
 682         if (IS_ERR(sgid_attr))
 683                 return PTR_ERR(sgid_attr);
 684
 685         id_priv->id.port_num = req->port;
 686         cma_bind_sgid_attr(id_priv, sgid_attr);
 687         /* Need to acquire lock to protect against reader
 688          * of cma_dev->id_list such as cma_netdev_callback() and
 689          * cma_process_remove().
 690          */
 691         mutex_lock(&lock);
 692         cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
 693         mutex_unlock(&lock);
 694         rdma_restrack_add(&id_priv->res);
 695         return 0;
 696 }
 697
 698 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
 699                               const struct rdma_id_private *listen_id_priv)
 700 {
 701         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 702         const struct ib_gid_attr *sgid_attr;
 703         struct cma_device *cma_dev;
 704         enum ib_gid_type gid_type;
 705         int ret = -ENODEV;
 706         union ib_gid gid;
 707         u32 port;
 708
 709         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 710             id_priv->id.ps == RDMA_PS_IPOIB)
 711                 return -EINVAL;
 712
 713         memcpy(&gid, dev_addr->src_dev_addr +
 714                rdma_addr_gid_offset(dev_addr), sizeof(gid));
 715
 716         mutex_lock(&lock);
 717
 718         cma_dev = listen_id_priv->cma_dev;
 719         port = listen_id_priv->id.port_num;
 720         gid_type = listen_id_priv->gid_type;
 721         sgid_attr = cma_validate_port(cma_dev->device, port,
 722                                       gid_type, &gid, id_priv);
 723         if (!IS_ERR(sgid_attr)) {
 724                 id_priv->id.port_num = port;
 725                 cma_bind_sgid_attr(id_priv, sgid_attr);
 726                 ret = 0;
 727                 goto out;
 728         }
 729
 730         list_for_each_entry(cma_dev, &dev_list, list) {
 731                 rdma_for_each_port (cma_dev->device, port) {
 732                         if (listen_id_priv->cma_dev == cma_dev &&
 733                             listen_id_priv->id.port_num == port)
 734                                 continue;
 735
 736                         gid_type = cma_dev->default_gid_type[port - 1];
 737                         sgid_attr = cma_validate_port(cma_dev->device, port,
 738                                                       gid_type, &gid, id_priv);
 739                         if (!IS_ERR(sgid_attr)) {
 740                                 id_priv->id.port_num = port;
 741                                 cma_bind_sgid_attr(id_priv, sgid_attr);
 742                                 ret = 0;
 743                                 goto out;
 744                         }
 745                 }
 746         }
 747
 748 out:
 749         if (!ret) {
 750                 cma_attach_to_dev(id_priv, cma_dev);
 751                 rdma_restrack_add(&id_priv->res);
 752         }
 753
 754         mutex_unlock(&lock);
 755         return ret;
 756 }
 757
 758 /*
 759  * Select the source IB device and address to reach the destination IB address.
 760  */
 761 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 762 {
 763         struct cma_device *cma_dev, *cur_dev;
 764         struct sockaddr_ib *addr;
 765         union ib_gid gid, sgid, *dgid;
 766         unsigned int p;
 767         u16 pkey, index;
 768         enum ib_port_state port_state;
 769         int i;
 770
 771         cma_dev = NULL;
 772         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 773         dgid = (union ib_gid *) &addr->sib_addr;
 774         pkey = ntohs(addr->sib_pkey);
 775
 776         mutex_lock(&lock);
 777         list_for_each_entry(cur_dev, &dev_list, list) {
 778                 rdma_for_each_port (cur_dev->device, p) {
 779                         if (!rdma_cap_af_ib(cur_dev->device, p))
 780                                 continue;
 781
 782                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 783                                 continue;
 784
 785                         if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
 786                                 continue;
 787                         for (i = 0; !rdma_query_gid(cur_dev->device,
 788                                                     p, i, &gid);
 789                              i++) {
 790                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 791                                         cma_dev = cur_dev;
 792                                         sgid = gid;
 793                                         id_priv->id.port_num = p;
 794                                         goto found;
 795                                 }
 796
 797                                 if (!cma_dev && (gid.global.subnet_prefix ==
 798                                     dgid->global.subnet_prefix) &&
 799                                     port_state == IB_PORT_ACTIVE) {
 800                                         cma_dev = cur_dev;
 801                                         sgid = gid;
 802                                         id_priv->id.port_num = p;
 803                                         goto found;
 804                                 }
 805                         }
 806                 }
 807         }
 808         mutex_unlock(&lock);
 809         return -ENODEV;
 810
 811 found:
 812         cma_attach_to_dev(id_priv, cma_dev);
 813         rdma_restrack_add(&id_priv->res);
 814         mutex_unlock(&lock);
 815         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 816         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 817         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 818         return 0;
 819 }
 820
 821 static void cma_id_get(struct rdma_id_private *id_priv)
 822 {
 823         refcount_inc(&id_priv->refcount);
 824 }
 825
 826 static void cma_id_put(struct rdma_id_private *id_priv)
 827 {
 828         if (refcount_dec_and_test(&id_priv->refcount))
 829                 complete(&id_priv->comp);
 830 }
 831
 832 static struct rdma_id_private *
 833 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
 834                  void *context, enum rdma_ucm_port_space ps,
 835                  enum ib_qp_type qp_type, const struct rdma_id_private *parent)
 836 {
 837         struct rdma_id_private *id_priv;
 838
 839         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 840         if (!id_priv)
 841                 return ERR_PTR(-ENOMEM);
 842
 843         id_priv->state = RDMA_CM_IDLE;
 844         id_priv->id.context = context;
 845         id_priv->id.event_handler = event_handler;
 846         id_priv->id.ps = ps;
 847         id_priv->id.qp_type = qp_type;
 848         id_priv->tos_set = false;
 849         id_priv->timeout_set = false;
 850         id_priv->min_rnr_timer_set = false;
 851         id_priv->gid_type = IB_GID_TYPE_IB;
 852         spin_lock_init(&id_priv->lock);
 853         mutex_init(&id_priv->qp_mutex);
 854         init_completion(&id_priv->comp);
 855         refcount_set(&id_priv->refcount, 1);
 856         mutex_init(&id_priv->handler_mutex);
 857         INIT_LIST_HEAD(&id_priv->listen_list);
 858         INIT_LIST_HEAD(&id_priv->mc_list);
 859         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 860         id_priv->id.route.addr.dev_addr.net = get_net(net);
 861         id_priv->seq_num &= 0x00ffffff;
 862
 863         rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
 864         if (parent)
 865                 rdma_restrack_parent_name(&id_priv->res, &parent->res);
 866
 867         return id_priv;
 868 }
 869
 870 struct rdma_cm_id *
 871 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
 872                         void *context, enum rdma_ucm_port_space ps,
 873                         enum ib_qp_type qp_type, const char *caller)
 874 {
 875         struct rdma_id_private *ret;
 876
 877         ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
 878         if (IS_ERR(ret))
 879                 return ERR_CAST(ret);
 880
 881         rdma_restrack_set_name(&ret->res, caller);
 882         return &ret->id;
 883 }
 884 EXPORT_SYMBOL(__rdma_create_kernel_id);
 885
 886 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
 887                                        void *context,
 888                                        enum rdma_ucm_port_space ps,
 889                                        enum ib_qp_type qp_type)
 890 {
 891         struct rdma_id_private *ret;
 892
 893         ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
 894                                ps, qp_type, NULL);
 895         if (IS_ERR(ret))
 896                 return ERR_CAST(ret);
 897
 898         rdma_restrack_set_name(&ret->res, NULL);
 899         return &ret->id;
 900 }
 901 EXPORT_SYMBOL(rdma_create_user_id);
 902
 903 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 904 {
 905         struct ib_qp_attr qp_attr;
 906         int qp_attr_mask, ret;
 907
 908         qp_attr.qp_state = IB_QPS_INIT;
 909         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 910         if (ret)
 911                 return ret;
 912
 913         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 914         if (ret)
 915                 return ret;
 916
 917         qp_attr.qp_state = IB_QPS_RTR;
 918         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 919         if (ret)
 920                 return ret;
 921
 922         qp_attr.qp_state = IB_QPS_RTS;
 923         qp_attr.sq_psn = 0;
 924         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 925
 926         return ret;
 927 }
 928
 929 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 930 {
 931         struct ib_qp_attr qp_attr;
 932         int qp_attr_mask, ret;
 933
 934         qp_attr.qp_state = IB_QPS_INIT;
 935         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 936         if (ret)
 937                 return ret;
 938
 939         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 940 }
 941
 942 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 943                    struct ib_qp_init_attr *qp_init_attr)
 944 {
 945         struct rdma_id_private *id_priv;
 946         struct ib_qp *qp;
 947         int ret;
 948
 949         id_priv = container_of(id, struct rdma_id_private, id);
 950         if (id->device != pd->device) {
 951                 ret = -EINVAL;
 952                 goto out_err;
 953         }
 954
 955         qp_init_attr->port_num = id->port_num;
 956         qp = ib_create_qp(pd, qp_init_attr);
 957         if (IS_ERR(qp)) {
 958                 ret = PTR_ERR(qp);
 959                 goto out_err;
 960         }
 961
 962         if (id->qp_type == IB_QPT_UD)
 963                 ret = cma_init_ud_qp(id_priv, qp);
 964         else
 965                 ret = cma_init_conn_qp(id_priv, qp);
 966         if (ret)
 967                 goto out_destroy;
 968
 969         id->qp = qp;
 970         id_priv->qp_num = qp->qp_num;
 971         id_priv->srq = (qp->srq != NULL);
 972         trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
 973         return 0;
 974 out_destroy:
 975         ib_destroy_qp(qp);
 976 out_err:
 977         trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
 978         return ret;
 979 }
 980 EXPORT_SYMBOL(rdma_create_qp);
 981
 982 void rdma_destroy_qp(struct rdma_cm_id *id)
 983 {
 984         struct rdma_id_private *id_priv;
 985
 986         id_priv = container_of(id, struct rdma_id_private, id);
 987         trace_cm_qp_destroy(id_priv);
 988         mutex_lock(&id_priv->qp_mutex);
 989         ib_destroy_qp(id_priv->id.qp);
 990         id_priv->id.qp = NULL;
 991         mutex_unlock(&id_priv->qp_mutex);
 992 }
 993 EXPORT_SYMBOL(rdma_destroy_qp);
 994
 995 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 996                              struct rdma_conn_param *conn_param)
 997 {
 998         struct ib_qp_attr qp_attr;
 999         int qp_attr_mask, ret;
1000
1001         mutex_lock(&id_priv->qp_mutex);
1002         if (!id_priv->id.qp) {
1003                 ret = 0;
1004                 goto out;
1005         }
1006
1007         /* Need to update QP attributes from default values. */
1008         qp_attr.qp_state = IB_QPS_INIT;
1009         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1010         if (ret)
1011                 goto out;
1012
1013         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1014         if (ret)
1015                 goto out;
1016
1017         qp_attr.qp_state = IB_QPS_RTR;
1018         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1019         if (ret)
1020                 goto out;
1021
1022         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1023
1024         if (conn_param)
1025                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1026         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1027 out:
1028         mutex_unlock(&id_priv->qp_mutex);
1029         return ret;
1030 }
1031
1032 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1033                              struct rdma_conn_param *conn_param)
1034 {
1035         struct ib_qp_attr qp_attr;
1036         int qp_attr_mask, ret;
1037
1038         mutex_lock(&id_priv->qp_mutex);
1039         if (!id_priv->id.qp) {
1040                 ret = 0;
1041                 goto out;
1042         }
1043
1044         qp_attr.qp_state = IB_QPS_RTS;
1045         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1046         if (ret)
1047                 goto out;
1048
1049         if (conn_param)
1050                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1051         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1052 out:
1053         mutex_unlock(&id_priv->qp_mutex);
1054         return ret;
1055 }
1056
1057 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1058 {
1059         struct ib_qp_attr qp_attr;
1060         int ret;
1061
1062         mutex_lock(&id_priv->qp_mutex);
1063         if (!id_priv->id.qp) {
1064                 ret = 0;
1065                 goto out;
1066         }
1067
1068         qp_attr.qp_state = IB_QPS_ERR;
1069         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1070 out:
1071         mutex_unlock(&id_priv->qp_mutex);
1072         return ret;
1073 }
1074
1075 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1076                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1077 {
1078         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1079         int ret;
1080         u16 pkey;
1081
1082         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1083                 pkey = 0xffff;
1084         else
1085                 pkey = ib_addr_get_pkey(dev_addr);
1086
1087         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1088                                   pkey, &qp_attr->pkey_index);
1089         if (ret)
1090                 return ret;
1091
1092         qp_attr->port_num = id_priv->id.port_num;
1093         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1094
1095         if (id_priv->id.qp_type == IB_QPT_UD) {
1096                 ret = cma_set_qkey(id_priv, 0);
1097                 if (ret)
1098                         return ret;
1099
1100                 qp_attr->qkey = id_priv->qkey;
1101                 *qp_attr_mask |= IB_QP_QKEY;
1102         } else {
1103                 qp_attr->qp_access_flags = 0;
1104                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1105         }
1106         return 0;
1107 }
1108
1109 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1110                        int *qp_attr_mask)
1111 {
1112         struct rdma_id_private *id_priv;
1113         int ret = 0;
1114
1115         id_priv = container_of(id, struct rdma_id_private, id);
1116         if (rdma_cap_ib_cm(id->device, id->port_num)) {
1117                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1118                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1119                 else
1120                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1121                                                  qp_attr_mask);
1122
1123                 if (qp_attr->qp_state == IB_QPS_RTR)
1124                         qp_attr->rq_psn = id_priv->seq_num;
1125         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1126                 if (!id_priv->cm_id.iw) {
1127                         qp_attr->qp_access_flags = 0;
1128                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1129                 } else
1130                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1131                                                  qp_attr_mask);
1132                 qp_attr->port_num = id_priv->id.port_num;
1133                 *qp_attr_mask |= IB_QP_PORT;
1134         } else {
1135                 ret = -ENOSYS;
1136         }
1137
1138         if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1139                 qp_attr->timeout = id_priv->timeout;
1140
1141         if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1142                 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1143
1144         return ret;
1145 }
1146 EXPORT_SYMBOL(rdma_init_qp_attr);
1147
1148 static inline bool cma_zero_addr(const struct sockaddr *addr)
1149 {
1150         switch (addr->sa_family) {
1151         case AF_INET:
1152                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1153         case AF_INET6:
1154                 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1155         case AF_IB:
1156                 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1157         default:
1158                 return false;
1159         }
1160 }
1161
1162 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1163 {
1164         switch (addr->sa_family) {
1165         case AF_INET:
1166                 return ipv4_is_loopback(
1167                         ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1168         case AF_INET6:
1169                 return ipv6_addr_loopback(
1170                         &((struct sockaddr_in6 *)addr)->sin6_addr);
1171         case AF_IB:
1172                 return ib_addr_loopback(
1173                         &((struct sockaddr_ib *)addr)->sib_addr);
1174         default:
1175                 return false;
1176         }
1177 }
1178
1179 static inline bool cma_any_addr(const struct sockaddr *addr)
1180 {
1181         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1182 }
1183
1184 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1185 {
1186         if (src->sa_family != dst->sa_family)
1187                 return -1;
1188
1189         switch (src->sa_family) {
1190         case AF_INET:
1191                 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1192                        ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1193         case AF_INET6: {
1194                 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1195                 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1196                 bool link_local;
1197
1198                 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1199                                           &dst_addr6->sin6_addr))
1200                         return 1;
1201                 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1202                              IPV6_ADDR_LINKLOCAL;
1203                 /* Link local must match their scope_ids */
1204                 return link_local ? (src_addr6->sin6_scope_id !=
1205                                      dst_addr6->sin6_scope_id) :
1206                                     0;
1207         }
1208
1209         default:
1210                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1211                                    &((struct sockaddr_ib *) dst)->sib_addr);
1212         }
1213 }
1214
1215 static __be16 cma_port(const struct sockaddr *addr)
1216 {
1217         struct sockaddr_ib *sib;
1218
1219         switch (addr->sa_family) {
1220         case AF_INET:
1221                 return ((struct sockaddr_in *) addr)->sin_port;
1222         case AF_INET6:
1223                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1224         case AF_IB:
1225                 sib = (struct sockaddr_ib *) addr;
1226                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1227                                     be64_to_cpu(sib->sib_sid_mask)));
1228         default:
1229                 return 0;
1230         }
1231 }
1232
1233 static inline int cma_any_port(const struct sockaddr *addr)
1234 {
1235         return !cma_port(addr);
1236 }
1237
1238 static void cma_save_ib_info(struct sockaddr *src_addr,
1239                              struct sockaddr *dst_addr,
1240                              const struct rdma_cm_id *listen_id,
1241                              const struct sa_path_rec *path)
1242 {
1243         struct sockaddr_ib *listen_ib, *ib;
1244
1245         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1246         if (src_addr) {
1247                 ib = (struct sockaddr_ib *)src_addr;
1248                 ib->sib_family = AF_IB;
1249                 if (path) {
1250                         ib->sib_pkey = path->pkey;
1251                         ib->sib_flowinfo = path->flow_label;
1252                         memcpy(&ib->sib_addr, &path->sgid, 16);
1253                         ib->sib_sid = path->service_id;
1254                         ib->sib_scope_id = 0;
1255                 } else {
1256                         ib->sib_pkey = listen_ib->sib_pkey;
1257                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1258                         ib->sib_addr = listen_ib->sib_addr;
1259                         ib->sib_sid = listen_ib->sib_sid;
1260                         ib->sib_scope_id = listen_ib->sib_scope_id;
1261                 }
1262                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1263         }
1264         if (dst_addr) {
1265                 ib = (struct sockaddr_ib *)dst_addr;
1266                 ib->sib_family = AF_IB;
1267                 if (path) {
1268                         ib->sib_pkey = path->pkey;
1269                         ib->sib_flowinfo = path->flow_label;
1270                         memcpy(&ib->sib_addr, &path->dgid, 16);
1271                 }
1272         }
1273 }
1274
1275 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1276                               struct sockaddr_in *dst_addr,
1277                               struct cma_hdr *hdr,
1278                               __be16 local_port)
1279 {
1280         if (src_addr) {
1281                 *src_addr = (struct sockaddr_in) {
1282                         .sin_family = AF_INET,
1283                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1284                         .sin_port = local_port,
1285                 };
1286         }
1287
1288         if (dst_addr) {
1289                 *dst_addr = (struct sockaddr_in) {
1290                         .sin_family = AF_INET,
1291                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1292                         .sin_port = hdr->port,
1293                 };
1294         }
1295 }
1296
1297 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1298                               struct sockaddr_in6 *dst_addr,
1299                               struct cma_hdr *hdr,
1300                               __be16 local_port)
1301 {
1302         if (src_addr) {
1303                 *src_addr = (struct sockaddr_in6) {
1304                         .sin6_family = AF_INET6,
1305                         .sin6_addr = hdr->dst_addr.ip6,
1306                         .sin6_port = local_port,
1307                 };
1308         }
1309
1310         if (dst_addr) {
1311                 *dst_addr = (struct sockaddr_in6) {
1312                         .sin6_family = AF_INET6,
1313                         .sin6_addr = hdr->src_addr.ip6,
1314                         .sin6_port = hdr->port,
1315                 };
1316         }
1317 }
1318
1319 static u16 cma_port_from_service_id(__be64 service_id)
1320 {
1321         return (u16)be64_to_cpu(service_id);
1322 }
1323
1324 static int cma_save_ip_info(struct sockaddr *src_addr,
1325                             struct sockaddr *dst_addr,
1326                             const struct ib_cm_event *ib_event,
1327                             __be64 service_id)
1328 {
1329         struct cma_hdr *hdr;
1330         __be16 port;
1331
1332         hdr = ib_event->private_data;
1333         if (hdr->cma_version != CMA_VERSION)
1334                 return -EINVAL;
1335
1336         port = htons(cma_port_from_service_id(service_id));
1337
1338         switch (cma_get_ip_ver(hdr)) {
1339         case 4:
1340                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1341                                   (struct sockaddr_in *)dst_addr, hdr, port);
1342                 break;
1343         case 6:
1344                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1345                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1346                 break;
1347         default:
1348                 return -EAFNOSUPPORT;
1349         }
1350
1351         return 0;
1352 }
1353
1354 static int cma_save_net_info(struct sockaddr *src_addr,
1355                              struct sockaddr *dst_addr,
1356                              const struct rdma_cm_id *listen_id,
1357                              const struct ib_cm_event *ib_event,
1358                              sa_family_t sa_family, __be64 service_id)
1359 {
1360         if (sa_family == AF_IB) {
1361                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1362                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1363                                          ib_event->param.req_rcvd.primary_path);
1364                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1365                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1366                 return 0;
1367         }
1368
1369         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1370 }
1371
1372 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1373                              struct cma_req_info *req)
1374 {
1375         const struct ib_cm_req_event_param *req_param =
1376                 &ib_event->param.req_rcvd;
1377         const struct ib_cm_sidr_req_event_param *sidr_param =
1378                 &ib_event->param.sidr_req_rcvd;
1379
1380         switch (ib_event->event) {
1381         case IB_CM_REQ_RECEIVED:
1382                 req->device     = req_param->listen_id->device;
1383                 req->port       = req_param->port;
1384                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1385                        sizeof(req->local_gid));
1386                 req->has_gid    = true;
1387                 req->service_id = req_param->primary_path->service_id;
1388                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1389                 if (req->pkey != req_param->bth_pkey)
1390                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1391                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1392                                             req_param->bth_pkey, req->pkey);
1393                 break;
1394         case IB_CM_SIDR_REQ_RECEIVED:
1395                 req->device     = sidr_param->listen_id->device;
1396                 req->port       = sidr_param->port;
1397                 req->has_gid    = false;
1398                 req->service_id = sidr_param->service_id;
1399                 req->pkey       = sidr_param->pkey;
1400                 if (req->pkey != sidr_param->bth_pkey)
1401                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1402                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1403                                             sidr_param->bth_pkey, req->pkey);
1404                 break;
1405         default:
1406                 return -EINVAL;
1407         }
1408
1409         return 0;
1410 }
1411
1412 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1413                                   const struct sockaddr_in *dst_addr,
1414                                   const struct sockaddr_in *src_addr)
1415 {
1416         __be32 daddr = dst_addr->sin_addr.s_addr,
1417                saddr = src_addr->sin_addr.s_addr;
1418         struct fib_result res;
1419         struct flowi4 fl4;
1420         int err;
1421         bool ret;
1422
1423         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1424             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1425             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1426             ipv4_is_loopback(saddr))
1427                 return false;
1428
1429         memset(&fl4, 0, sizeof(fl4));
1430         fl4.flowi4_iif = net_dev->ifindex;
1431         fl4.daddr = daddr;
1432         fl4.saddr = saddr;
1433
1434         rcu_read_lock();
1435         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1436         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1437         rcu_read_unlock();
1438
1439         return ret;
1440 }
1441
1442 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1443                                   const struct sockaddr_in6 *dst_addr,
1444                                   const struct sockaddr_in6 *src_addr)
1445 {
1446 #if IS_ENABLED(CONFIG_IPV6)
1447         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1448                            IPV6_ADDR_LINKLOCAL;
1449         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1450                                          &src_addr->sin6_addr, net_dev->ifindex,
1451                                          NULL, strict);
1452         bool ret;
1453
1454         if (!rt)
1455                 return false;
1456
1457         ret = rt->rt6i_idev->dev == net_dev;
1458         ip6_rt_put(rt);
1459
1460         return ret;
1461 #else
1462         return false;
1463 #endif
1464 }
1465
1466 static bool validate_net_dev(struct net_device *net_dev,
1467                              const struct sockaddr *daddr,
1468                              const struct sockaddr *saddr)
1469 {
1470         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1471         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1472         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1473         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1474
1475         switch (daddr->sa_family) {
1476         case AF_INET:
1477                 return saddr->sa_family == AF_INET &&
1478                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1479
1480         case AF_INET6:
1481                 return saddr->sa_family == AF_INET6 &&
1482                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1483
1484         default:
1485                 return false;
1486         }
1487 }
1488
1489 static struct net_device *
1490 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1491 {
1492         const struct ib_gid_attr *sgid_attr = NULL;
1493         struct net_device *ndev;
1494
1495         if (ib_event->event == IB_CM_REQ_RECEIVED)
1496                 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1497         else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1498                 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1499
1500         if (!sgid_attr)
1501                 return NULL;
1502
1503         rcu_read_lock();
1504         ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1505         if (IS_ERR(ndev))
1506                 ndev = NULL;
1507         else
1508                 dev_hold(ndev);
1509         rcu_read_unlock();
1510         return ndev;
1511 }
1512
1513 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1514                                           struct cma_req_info *req)
1515 {
1516         struct sockaddr *listen_addr =
1517                         (struct sockaddr *)&req->listen_addr_storage;
1518         struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1519         struct net_device *net_dev;
1520         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1521         int err;
1522
1523         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1524                                req->service_id);
1525         if (err)
1526                 return ERR_PTR(err);
1527
1528         if (rdma_protocol_roce(req->device, req->port))
1529                 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1530         else
1531                 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1532                                                    req->pkey,
1533                                                    gid, listen_addr);
1534         if (!net_dev)
1535                 return ERR_PTR(-ENODEV);
1536
1537         return net_dev;
1538 }
1539
1540 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1541 {
1542         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1543 }
1544
1545 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1546                                    const struct cma_hdr *hdr)
1547 {
1548         struct sockaddr *addr = cma_src_addr(id_priv);
1549         __be32 ip4_addr;
1550         struct in6_addr ip6_addr;
1551
1552         if (cma_any_addr(addr) && !id_priv->afonly)
1553                 return true;
1554
1555         switch (addr->sa_family) {
1556         case AF_INET:
1557                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1558                 if (cma_get_ip_ver(hdr) != 4)
1559                         return false;
1560                 if (!cma_any_addr(addr) &&
1561                     hdr->dst_addr.ip4.addr != ip4_addr)
1562                         return false;
1563                 break;
1564         case AF_INET6:
1565                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1566                 if (cma_get_ip_ver(hdr) != 6)
1567                         return false;
1568                 if (!cma_any_addr(addr) &&
1569                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1570                         return false;
1571                 break;
1572         case AF_IB:
1573                 return true;
1574         default:
1575                 return false;
1576         }
1577
1578         return true;
1579 }
1580
1581 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1582 {
1583         struct ib_device *device = id->device;
1584         const u32 port_num = id->port_num ?: rdma_start_port(device);
1585
1586         return rdma_protocol_roce(device, port_num);
1587 }
1588
1589 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1590 {
1591         const struct sockaddr *daddr =
1592                         (const struct sockaddr *)&req->listen_addr_storage;
1593         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1594
1595         /* Returns true if the req is for IPv6 link local */
1596         return (daddr->sa_family == AF_INET6 &&
1597                 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1598 }
1599
1600 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1601                               const struct net_device *net_dev,
1602                               const struct cma_req_info *req)
1603 {
1604         const struct rdma_addr *addr = &id->route.addr;
1605
1606         if (!net_dev)
1607                 /* This request is an AF_IB request */
1608                 return (!id->port_num || id->port_num == req->port) &&
1609                        (addr->src_addr.ss_family == AF_IB);
1610
1611         /*
1612          * If the request is not for IPv6 link local, allow matching
1613          * request to any netdevice of the one or multiport rdma device.
1614          */
1615         if (!cma_is_req_ipv6_ll(req))
1616                 return true;
1617         /*
1618          * Net namespaces must match, and if the listner is listening
1619          * on a specific netdevice than netdevice must match as well.
1620          */
1621         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1622             (!!addr->dev_addr.bound_dev_if ==
1623              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1624                 return true;
1625         else
1626                 return false;
1627 }
1628
1629 static struct rdma_id_private *cma_find_listener(
1630                 const struct rdma_bind_list *bind_list,
1631                 const struct ib_cm_id *cm_id,
1632                 const struct ib_cm_event *ib_event,
1633                 const struct cma_req_info *req,
1634                 const struct net_device *net_dev)
1635 {
1636         struct rdma_id_private *id_priv, *id_priv_dev;
1637
1638         lockdep_assert_held(&lock);
1639
1640         if (!bind_list)
1641                 return ERR_PTR(-EINVAL);
1642
1643         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1644                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1645                         if (id_priv->id.device == cm_id->device &&
1646                             cma_match_net_dev(&id_priv->id, net_dev, req))
1647                                 return id_priv;
1648                         list_for_each_entry(id_priv_dev,
1649                                             &id_priv->listen_list,
1650                                             listen_list) {
1651                                 if (id_priv_dev->id.device == cm_id->device &&
1652                                     cma_match_net_dev(&id_priv_dev->id,
1653                                                       net_dev, req))
1654                                         return id_priv_dev;
1655                         }
1656                 }
1657         }
1658
1659         return ERR_PTR(-EINVAL);
1660 }
1661
1662 static struct rdma_id_private *
1663 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1664                      const struct ib_cm_event *ib_event,
1665                      struct cma_req_info *req,
1666                      struct net_device **net_dev)
1667 {
1668         struct rdma_bind_list *bind_list;
1669         struct rdma_id_private *id_priv;
1670         int err;
1671
1672         err = cma_save_req_info(ib_event, req);
1673         if (err)
1674                 return ERR_PTR(err);
1675
1676         *net_dev = cma_get_net_dev(ib_event, req);
1677         if (IS_ERR(*net_dev)) {
1678                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1679                         /* Assuming the protocol is AF_IB */
1680                         *net_dev = NULL;
1681                 } else {
1682                         return ERR_CAST(*net_dev);
1683                 }
1684         }
1685
1686         mutex_lock(&lock);
1687         /*
1688          * Net namespace might be getting deleted while route lookup,
1689          * cm_id lookup is in progress. Therefore, perform netdevice
1690          * validation, cm_id lookup under rcu lock.
1691          * RCU lock along with netdevice state check, synchronizes with
1692          * netdevice migrating to different net namespace and also avoids
1693          * case where net namespace doesn't get deleted while lookup is in
1694          * progress.
1695          * If the device state is not IFF_UP, its properties such as ifindex
1696          * and nd_net cannot be trusted to remain valid without rcu lock.
1697          * net/core/dev.c change_net_namespace() ensures to synchronize with
1698          * ongoing operations on net device after device is closed using
1699          * synchronize_net().
1700          */
1701         rcu_read_lock();
1702         if (*net_dev) {
1703                 /*
1704                  * If netdevice is down, it is likely that it is administratively
1705                  * down or it might be migrating to different namespace.
1706                  * In that case avoid further processing, as the net namespace
1707                  * or ifindex may change.
1708                  */
1709                 if (((*net_dev)->flags & IFF_UP) == 0) {
1710                         id_priv = ERR_PTR(-EHOSTUNREACH);
1711                         goto err;
1712                 }
1713
1714                 if (!validate_net_dev(*net_dev,
1715                                  (struct sockaddr *)&req->listen_addr_storage,
1716                                  (struct sockaddr *)&req->src_addr_storage)) {
1717                         id_priv = ERR_PTR(-EHOSTUNREACH);
1718                         goto err;
1719                 }
1720         }
1721
1722         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1723                                 rdma_ps_from_service_id(req->service_id),
1724                                 cma_port_from_service_id(req->service_id));
1725         id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1726 err:
1727         rcu_read_unlock();
1728         mutex_unlock(&lock);
1729         if (IS_ERR(id_priv) && *net_dev) {
1730                 dev_put(*net_dev);
1731                 *net_dev = NULL;
1732         }
1733         return id_priv;
1734 }
1735
1736 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1737 {
1738         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1739 }
1740
1741 static void cma_cancel_route(struct rdma_id_private *id_priv)
1742 {
1743         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1744                 if (id_priv->query)
1745                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1746         }
1747 }
1748
1749 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1750 {
1751         struct rdma_id_private *dev_id_priv;
1752
1753         lockdep_assert_held(&lock);
1754
1755         /*
1756          * Remove from listen_any_list to prevent added devices from spawning
1757          * additional listen requests.
1758          */
1759         list_del(&id_priv->list);
1760
1761         while (!list_empty(&id_priv->listen_list)) {
1762                 dev_id_priv = list_entry(id_priv->listen_list.next,
1763                                          struct rdma_id_private, listen_list);
1764                 /* sync with device removal to avoid duplicate destruction */
1765                 list_del_init(&dev_id_priv->list);
1766                 list_del(&dev_id_priv->listen_list);
1767                 mutex_unlock(&lock);
1768
1769                 rdma_destroy_id(&dev_id_priv->id);
1770                 mutex_lock(&lock);
1771         }
1772 }
1773
1774 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1775 {
1776         mutex_lock(&lock);
1777         _cma_cancel_listens(id_priv);
1778         mutex_unlock(&lock);
1779 }
1780
1781 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1782                                  enum rdma_cm_state state)
1783 {
1784         switch (state) {
1785         case RDMA_CM_ADDR_QUERY:
1786                 /*
1787                  * We can avoid doing the rdma_addr_cancel() based on state,
1788                  * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1789                  * Notice that the addr_handler work could still be exiting
1790                  * outside this state, however due to the interaction with the
1791                  * handler_mutex the work is guaranteed not to touch id_priv
1792                  * during exit.
1793                  */
1794                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1795                 break;
1796         case RDMA_CM_ROUTE_QUERY:
1797                 cma_cancel_route(id_priv);
1798                 break;
1799         case RDMA_CM_LISTEN:
1800                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1801                         cma_cancel_listens(id_priv);
1802                 break;
1803         default:
1804                 break;
1805         }
1806 }
1807
1808 static void cma_release_port(struct rdma_id_private *id_priv)
1809 {
1810         struct rdma_bind_list *bind_list = id_priv->bind_list;
1811         struct net *net = id_priv->id.route.addr.dev_addr.net;
1812
1813         if (!bind_list)
1814                 return;
1815
1816         mutex_lock(&lock);
1817         hlist_del(&id_priv->node);
1818         if (hlist_empty(&bind_list->owners)) {
1819                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1820                 kfree(bind_list);
1821         }
1822         mutex_unlock(&lock);
1823 }
1824
1825 static void destroy_mc(struct rdma_id_private *id_priv,
1826                        struct cma_multicast *mc)
1827 {
1828         bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1829
1830         if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1831                 ib_sa_free_multicast(mc->sa_mc);
1832
1833         if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1834                 struct rdma_dev_addr *dev_addr =
1835                         &id_priv->id.route.addr.dev_addr;
1836                 struct net_device *ndev = NULL;
1837
1838                 if (dev_addr->bound_dev_if)
1839                         ndev = dev_get_by_index(dev_addr->net,
1840                                                 dev_addr->bound_dev_if);
1841                 if (ndev) {
1842                         union ib_gid mgid;
1843
1844                         cma_set_mgid(id_priv, (struct sockaddr *)&mc->addr,
1845                                      &mgid);
1846
1847                         if (!send_only)
1848                                 cma_igmp_send(ndev, &mgid, false);
1849
1850                         dev_put(ndev);
1851                 }
1852
1853                 cancel_work_sync(&mc->iboe_join.work);
1854         }
1855         kfree(mc);
1856 }
1857
1858 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1859 {
1860         struct cma_multicast *mc;
1861
1862         while (!list_empty(&id_priv->mc_list)) {
1863                 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1864                                       list);
1865                 list_del(&mc->list);
1866                 destroy_mc(id_priv, mc);
1867         }
1868 }
1869
1870 static void _destroy_id(struct rdma_id_private *id_priv,
1871                         enum rdma_cm_state state)
1872 {
1873         cma_cancel_operation(id_priv, state);
1874
1875         rdma_restrack_del(&id_priv->res);
1876         if (id_priv->cma_dev) {
1877                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1878                         if (id_priv->cm_id.ib)
1879                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1880                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1881                         if (id_priv->cm_id.iw)
1882                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1883                 }
1884                 cma_leave_mc_groups(id_priv);
1885                 cma_release_dev(id_priv);
1886         }
1887
1888         cma_release_port(id_priv);
1889         cma_id_put(id_priv);
1890         wait_for_completion(&id_priv->comp);
1891
1892         if (id_priv->internal_id)
1893                 cma_id_put(id_priv->id.context);
1894
1895         kfree(id_priv->id.route.path_rec);
1896
1897         put_net(id_priv->id.route.addr.dev_addr.net);
1898         kfree(id_priv);
1899 }
1900
1901 /*
1902  * destroy an ID from within the handler_mutex. This ensures that no other
1903  * handlers can start running concurrently.
1904  */
1905 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1906         __releases(&idprv->handler_mutex)
1907 {
1908         enum rdma_cm_state state;
1909         unsigned long flags;
1910
1911         trace_cm_id_destroy(id_priv);
1912
1913         /*
1914          * Setting the state to destroyed under the handler mutex provides a
1915          * fence against calling handler callbacks. If this is invoked due to
1916          * the failure of a handler callback then it guarentees that no future
1917          * handlers will be called.
1918          */
1919         lockdep_assert_held(&id_priv->handler_mutex);
1920         spin_lock_irqsave(&id_priv->lock, flags);
1921         state = id_priv->state;
1922         id_priv->state = RDMA_CM_DESTROYING;
1923         spin_unlock_irqrestore(&id_priv->lock, flags);
1924         mutex_unlock(&id_priv->handler_mutex);
1925         _destroy_id(id_priv, state);
1926 }
1927
1928 void rdma_destroy_id(struct rdma_cm_id *id)
1929 {
1930         struct rdma_id_private *id_priv =
1931                 container_of(id, struct rdma_id_private, id);
1932
1933         mutex_lock(&id_priv->handler_mutex);
1934         destroy_id_handler_unlock(id_priv);
1935 }
1936 EXPORT_SYMBOL(rdma_destroy_id);
1937
1938 static int cma_rep_recv(struct rdma_id_private *id_priv)
1939 {
1940         int ret;
1941
1942         ret = cma_modify_qp_rtr(id_priv, NULL);
1943         if (ret)
1944                 goto reject;
1945
1946         ret = cma_modify_qp_rts(id_priv, NULL);
1947         if (ret)
1948                 goto reject;
1949
1950         trace_cm_send_rtu(id_priv);
1951         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1952         if (ret)
1953                 goto reject;
1954
1955         return 0;
1956 reject:
1957         pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1958         cma_modify_qp_err(id_priv);
1959         trace_cm_send_rej(id_priv);
1960         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1961                        NULL, 0, NULL, 0);
1962         return ret;
1963 }
1964
1965 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1966                                    const struct ib_cm_rep_event_param *rep_data,
1967                                    void *private_data)
1968 {
1969         event->param.conn.private_data = private_data;
1970         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1971         event->param.conn.responder_resources = rep_data->responder_resources;
1972         event->param.conn.initiator_depth = rep_data->initiator_depth;
1973         event->param.conn.flow_control = rep_data->flow_control;
1974         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1975         event->param.conn.srq = rep_data->srq;
1976         event->param.conn.qp_num = rep_data->remote_qpn;
1977
1978         event->ece.vendor_id = rep_data->ece.vendor_id;
1979         event->ece.attr_mod = rep_data->ece.attr_mod;
1980 }
1981
1982 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1983                                 struct rdma_cm_event *event)
1984 {
1985         int ret;
1986
1987         lockdep_assert_held(&id_priv->handler_mutex);
1988
1989         trace_cm_event_handler(id_priv, event);
1990         ret = id_priv->id.event_handler(&id_priv->id, event);
1991         trace_cm_event_done(id_priv, event, ret);
1992         return ret;
1993 }
1994
1995 static int cma_ib_handler(struct ib_cm_id *cm_id,
1996                           const struct ib_cm_event *ib_event)
1997 {
1998         struct rdma_id_private *id_priv = cm_id->context;
1999         struct rdma_cm_event event = {};
2000         enum rdma_cm_state state;
2001         int ret;
2002
2003         mutex_lock(&id_priv->handler_mutex);
2004         state = READ_ONCE(id_priv->state);
2005         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2006              state != RDMA_CM_CONNECT) ||
2007             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2008              state != RDMA_CM_DISCONNECT))
2009                 goto out;
2010
2011         switch (ib_event->event) {
2012         case IB_CM_REQ_ERROR:
2013         case IB_CM_REP_ERROR:
2014                 event.event = RDMA_CM_EVENT_UNREACHABLE;
2015                 event.status = -ETIMEDOUT;
2016                 break;
2017         case IB_CM_REP_RECEIVED:
2018                 if (state == RDMA_CM_CONNECT &&
2019                     (id_priv->id.qp_type != IB_QPT_UD)) {
2020                         trace_cm_send_mra(id_priv);
2021                         ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2022                 }
2023                 if (id_priv->id.qp) {
2024                         event.status = cma_rep_recv(id_priv);
2025                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2026                                                      RDMA_CM_EVENT_ESTABLISHED;
2027                 } else {
2028                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2029                 }
2030                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2031                                        ib_event->private_data);
2032                 break;
2033         case IB_CM_RTU_RECEIVED:
2034         case IB_CM_USER_ESTABLISHED:
2035                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2036                 break;
2037         case IB_CM_DREQ_ERROR:
2038                 event.status = -ETIMEDOUT;
2039                 fallthrough;
2040         case IB_CM_DREQ_RECEIVED:
2041         case IB_CM_DREP_RECEIVED:
2042                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2043                                    RDMA_CM_DISCONNECT))
2044                         goto out;
2045                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2046                 break;
2047         case IB_CM_TIMEWAIT_EXIT:
2048                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2049                 break;
2050         case IB_CM_MRA_RECEIVED:
2051                 /* ignore event */
2052                 goto out;
2053         case IB_CM_REJ_RECEIVED:
2054                 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2055                                                                                 ib_event->param.rej_rcvd.reason));
2056                 cma_modify_qp_err(id_priv);
2057                 event.status = ib_event->param.rej_rcvd.reason;
2058                 event.event = RDMA_CM_EVENT_REJECTED;
2059                 event.param.conn.private_data = ib_event->private_data;
2060                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2061                 break;
2062         default:
2063                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2064                        ib_event->event);
2065                 goto out;
2066         }
2067
2068         ret = cma_cm_event_handler(id_priv, &event);
2069         if (ret) {
2070                 /* Destroy the CM ID by returning a non-zero value. */
2071                 id_priv->cm_id.ib = NULL;
2072                 destroy_id_handler_unlock(id_priv);
2073                 return ret;
2074         }
2075 out:
2076         mutex_unlock(&id_priv->handler_mutex);
2077         return 0;
2078 }
2079
2080 static struct rdma_id_private *
2081 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2082                    const struct ib_cm_event *ib_event,
2083                    struct net_device *net_dev)
2084 {
2085         struct rdma_id_private *listen_id_priv;
2086         struct rdma_id_private *id_priv;
2087         struct rdma_cm_id *id;
2088         struct rdma_route *rt;
2089         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2090         struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2091         const __be64 service_id =
2092                 ib_event->param.req_rcvd.primary_path->service_id;
2093         int ret;
2094
2095         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2096         id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2097                                    listen_id->event_handler, listen_id->context,
2098                                    listen_id->ps,
2099                                    ib_event->param.req_rcvd.qp_type,
2100                                    listen_id_priv);
2101         if (IS_ERR(id_priv))
2102                 return NULL;
2103
2104         id = &id_priv->id;
2105         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2106                               (struct sockaddr *)&id->route.addr.dst_addr,
2107                               listen_id, ib_event, ss_family, service_id))
2108                 goto err;
2109
2110         rt = &id->route;
2111         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2112         rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2113                                      GFP_KERNEL);
2114         if (!rt->path_rec)
2115                 goto err;
2116
2117         rt->path_rec[0] = *path;
2118         if (rt->num_paths == 2)
2119                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2120
2121         if (net_dev) {
2122                 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2123         } else {
2124                 if (!cma_protocol_roce(listen_id) &&
2125                     cma_any_addr(cma_src_addr(id_priv))) {
2126                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2127                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2128                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2129                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2130                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2131                         if (ret)
2132                                 goto err;
2133                 }
2134         }
2135         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2136
2137         id_priv->state = RDMA_CM_CONNECT;
2138         return id_priv;
2139
2140 err:
2141         rdma_destroy_id(id);
2142         return NULL;
2143 }
2144
2145 static struct rdma_id_private *
2146 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2147                   const struct ib_cm_event *ib_event,
2148                   struct net_device *net_dev)
2149 {
2150         const struct rdma_id_private *listen_id_priv;
2151         struct rdma_id_private *id_priv;
2152         struct rdma_cm_id *id;
2153         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2154         struct net *net = listen_id->route.addr.dev_addr.net;
2155         int ret;
2156
2157         listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2158         id_priv = __rdma_create_id(net, listen_id->event_handler,
2159                                    listen_id->context, listen_id->ps, IB_QPT_UD,
2160                                    listen_id_priv);
2161         if (IS_ERR(id_priv))
2162                 return NULL;
2163
2164         id = &id_priv->id;
2165         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2166                               (struct sockaddr *)&id->route.addr.dst_addr,
2167                               listen_id, ib_event, ss_family,
2168                               ib_event->param.sidr_req_rcvd.service_id))
2169                 goto err;
2170
2171         if (net_dev) {
2172                 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2173         } else {
2174                 if (!cma_any_addr(cma_src_addr(id_priv))) {
2175                         ret = cma_translate_addr(cma_src_addr(id_priv),
2176                                                  &id->route.addr.dev_addr);
2177                         if (ret)
2178                                 goto err;
2179                 }
2180         }
2181
2182         id_priv->state = RDMA_CM_CONNECT;
2183         return id_priv;
2184 err:
2185         rdma_destroy_id(id);
2186         return NULL;
2187 }
2188
2189 static void cma_set_req_event_data(struct rdma_cm_event *event,
2190                                    const struct ib_cm_req_event_param *req_data,
2191                                    void *private_data, int offset)
2192 {
2193         event->param.conn.private_data = private_data + offset;
2194         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2195         event->param.conn.responder_resources = req_data->responder_resources;
2196         event->param.conn.initiator_depth = req_data->initiator_depth;
2197         event->param.conn.flow_control = req_data->flow_control;
2198         event->param.conn.retry_count = req_data->retry_count;
2199         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2200         event->param.conn.srq = req_data->srq;
2201         event->param.conn.qp_num = req_data->remote_qpn;
2202
2203         event->ece.vendor_id = req_data->ece.vendor_id;
2204         event->ece.attr_mod = req_data->ece.attr_mod;
2205 }
2206
2207 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2208                                     const struct ib_cm_event *ib_event)
2209 {
2210         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2211                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2212                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2213                  (id->qp_type == IB_QPT_UD)) ||
2214                 (!id->qp_type));
2215 }
2216
2217 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2218                               const struct ib_cm_event *ib_event)
2219 {
2220         struct rdma_id_private *listen_id, *conn_id = NULL;
2221         struct rdma_cm_event event = {};
2222         struct cma_req_info req = {};
2223         struct net_device *net_dev;
2224         u8 offset;
2225         int ret;
2226
2227         listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2228         if (IS_ERR(listen_id))
2229                 return PTR_ERR(listen_id);
2230
2231         trace_cm_req_handler(listen_id, ib_event->event);
2232         if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2233                 ret = -EINVAL;
2234                 goto net_dev_put;
2235         }
2236
2237         mutex_lock(&listen_id->handler_mutex);
2238         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2239                 ret = -ECONNABORTED;
2240                 goto err_unlock;
2241         }
2242
2243         offset = cma_user_data_offset(listen_id);
2244         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2245         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2246                 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2247                 event.param.ud.private_data = ib_event->private_data + offset;
2248                 event.param.ud.private_data_len =
2249                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2250         } else {
2251                 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2252                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2253                                        ib_event->private_data, offset);
2254         }
2255         if (!conn_id) {
2256                 ret = -ENOMEM;
2257                 goto err_unlock;
2258         }
2259
2260         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2261         ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2262         if (ret) {
2263                 destroy_id_handler_unlock(conn_id);
2264                 goto err_unlock;
2265         }
2266
2267         conn_id->cm_id.ib = cm_id;
2268         cm_id->context = conn_id;
2269         cm_id->cm_handler = cma_ib_handler;
2270
2271         ret = cma_cm_event_handler(conn_id, &event);
2272         if (ret) {
2273                 /* Destroy the CM ID by returning a non-zero value. */
2274                 conn_id->cm_id.ib = NULL;
2275                 mutex_unlock(&listen_id->handler_mutex);
2276                 destroy_id_handler_unlock(conn_id);
2277                 goto net_dev_put;
2278         }
2279
2280         if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2281             conn_id->id.qp_type != IB_QPT_UD) {
2282                 trace_cm_send_mra(cm_id->context);
2283                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2284         }
2285         mutex_unlock(&conn_id->handler_mutex);
2286
2287 err_unlock:
2288         mutex_unlock(&listen_id->handler_mutex);
2289
2290 net_dev_put:
2291         if (net_dev)
2292                 dev_put(net_dev);
2293
2294         return ret;
2295 }
2296
2297 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2298 {
2299         if (addr->sa_family == AF_IB)
2300                 return ((struct sockaddr_ib *) addr)->sib_sid;
2301
2302         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2303 }
2304 EXPORT_SYMBOL(rdma_get_service_id);
2305
2306 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2307                     union ib_gid *dgid)
2308 {
2309         struct rdma_addr *addr = &cm_id->route.addr;
2310
2311         if (!cm_id->device) {
2312                 if (sgid)
2313                         memset(sgid, 0, sizeof(*sgid));
2314                 if (dgid)
2315                         memset(dgid, 0, sizeof(*dgid));
2316                 return;
2317         }
2318
2319         if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2320                 if (sgid)
2321                         rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2322                 if (dgid)
2323                         rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2324         } else {
2325                 if (sgid)
2326                         rdma_addr_get_sgid(&addr->dev_addr, sgid);
2327                 if (dgid)
2328                         rdma_addr_get_dgid(&addr->dev_addr, dgid);
2329         }
2330 }
2331 EXPORT_SYMBOL(rdma_read_gids);
2332
2333 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2334 {
2335         struct rdma_id_private *id_priv = iw_id->context;
2336         struct rdma_cm_event event = {};
2337         int ret = 0;
2338         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2339         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2340
2341         mutex_lock(&id_priv->handler_mutex);
2342         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2343                 goto out;
2344
2345         switch (iw_event->event) {
2346         case IW_CM_EVENT_CLOSE:
2347                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2348                 break;
2349         case IW_CM_EVENT_CONNECT_REPLY:
2350                 memcpy(cma_src_addr(id_priv), laddr,
2351                        rdma_addr_size(laddr));
2352                 memcpy(cma_dst_addr(id_priv), raddr,
2353                        rdma_addr_size(raddr));
2354                 switch (iw_event->status) {
2355                 case 0:
2356                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2357                         event.param.conn.initiator_depth = iw_event->ird;
2358                         event.param.conn.responder_resources = iw_event->ord;
2359                         break;
2360                 case -ECONNRESET:
2361                 case -ECONNREFUSED:
2362                         event.event = RDMA_CM_EVENT_REJECTED;
2363                         break;
2364                 case -ETIMEDOUT:
2365                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2366                         break;
2367                 default:
2368                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2369                         break;
2370                 }
2371                 break;
2372         case IW_CM_EVENT_ESTABLISHED:
2373                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2374                 event.param.conn.initiator_depth = iw_event->ird;
2375                 event.param.conn.responder_resources = iw_event->ord;
2376                 break;
2377         default:
2378                 goto out;
2379         }
2380
2381         event.status = iw_event->status;
2382         event.param.conn.private_data = iw_event->private_data;
2383         event.param.conn.private_data_len = iw_event->private_data_len;
2384         ret = cma_cm_event_handler(id_priv, &event);
2385         if (ret) {
2386                 /* Destroy the CM ID by returning a non-zero value. */
2387                 id_priv->cm_id.iw = NULL;
2388                 destroy_id_handler_unlock(id_priv);
2389                 return ret;
2390         }
2391
2392 out:
2393         mutex_unlock(&id_priv->handler_mutex);
2394         return ret;
2395 }
2396
2397 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2398                                struct iw_cm_event *iw_event)
2399 {
2400         struct rdma_id_private *listen_id, *conn_id;
2401         struct rdma_cm_event event = {};
2402         int ret = -ECONNABORTED;
2403         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2404         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2405
2406         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2407         event.param.conn.private_data = iw_event->private_data;
2408         event.param.conn.private_data_len = iw_event->private_data_len;
2409         event.param.conn.initiator_depth = iw_event->ird;
2410         event.param.conn.responder_resources = iw_event->ord;
2411
2412         listen_id = cm_id->context;
2413
2414         mutex_lock(&listen_id->handler_mutex);
2415         if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2416                 goto out;
2417
2418         /* Create a new RDMA id for the new IW CM ID */
2419         conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2420                                    listen_id->id.event_handler,
2421                                    listen_id->id.context, RDMA_PS_TCP,
2422                                    IB_QPT_RC, listen_id);
2423         if (IS_ERR(conn_id)) {
2424                 ret = -ENOMEM;
2425                 goto out;
2426         }
2427         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2428         conn_id->state = RDMA_CM_CONNECT;
2429
2430         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2431         if (ret) {
2432                 mutex_unlock(&listen_id->handler_mutex);
2433                 destroy_id_handler_unlock(conn_id);
2434                 return ret;
2435         }
2436
2437         ret = cma_iw_acquire_dev(conn_id, listen_id);
2438         if (ret) {
2439                 mutex_unlock(&listen_id->handler_mutex);
2440                 destroy_id_handler_unlock(conn_id);
2441                 return ret;
2442         }
2443
2444         conn_id->cm_id.iw = cm_id;
2445         cm_id->context = conn_id;
2446         cm_id->cm_handler = cma_iw_handler;
2447
2448         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2449         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2450
2451         ret = cma_cm_event_handler(conn_id, &event);
2452         if (ret) {
2453                 /* User wants to destroy the CM ID */
2454                 conn_id->cm_id.iw = NULL;
2455                 mutex_unlock(&listen_id->handler_mutex);
2456                 destroy_id_handler_unlock(conn_id);
2457                 return ret;
2458         }
2459
2460         mutex_unlock(&conn_id->handler_mutex);
2461
2462 out:
2463         mutex_unlock(&listen_id->handler_mutex);
2464         return ret;
2465 }
2466
2467 static int cma_ib_listen(struct rdma_id_private *id_priv)
2468 {
2469         struct sockaddr *addr;
2470         struct ib_cm_id *id;
2471         __be64 svc_id;
2472
2473         addr = cma_src_addr(id_priv);
2474         svc_id = rdma_get_service_id(&id_priv->id, addr);
2475         id = ib_cm_insert_listen(id_priv->id.device,
2476                                  cma_ib_req_handler, svc_id);
2477         if (IS_ERR(id))
2478                 return PTR_ERR(id);
2479         id_priv->cm_id.ib = id;
2480
2481         return 0;
2482 }
2483
2484 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2485 {
2486         int ret;
2487         struct iw_cm_id *id;
2488
2489         id = iw_create_cm_id(id_priv->id.device,
2490                              iw_conn_req_handler,
2491                              id_priv);
2492         if (IS_ERR(id))
2493                 return PTR_ERR(id);
2494
2495         mutex_lock(&id_priv->qp_mutex);
2496         id->tos = id_priv->tos;
2497         id->tos_set = id_priv->tos_set;
2498         mutex_unlock(&id_priv->qp_mutex);
2499         id->afonly = id_priv->afonly;
2500         id_priv->cm_id.iw = id;
2501
2502         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2503                rdma_addr_size(cma_src_addr(id_priv)));
2504
2505         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2506
2507         if (ret) {
2508                 iw_destroy_cm_id(id_priv->cm_id.iw);
2509                 id_priv->cm_id.iw = NULL;
2510         }
2511
2512         return ret;
2513 }
2514
2515 static int cma_listen_handler(struct rdma_cm_id *id,
2516                               struct rdma_cm_event *event)
2517 {
2518         struct rdma_id_private *id_priv = id->context;
2519
2520         /* Listening IDs are always destroyed on removal */
2521         if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2522                 return -1;
2523
2524         id->context = id_priv->id.context;
2525         id->event_handler = id_priv->id.event_handler;
2526         trace_cm_event_handler(id_priv, event);
2527         return id_priv->id.event_handler(id, event);
2528 }
2529
2530 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2531                              struct cma_device *cma_dev,
2532                              struct rdma_id_private **to_destroy)
2533 {
2534         struct rdma_id_private *dev_id_priv;
2535         struct net *net = id_priv->id.route.addr.dev_addr.net;
2536         int ret;
2537
2538         lockdep_assert_held(&lock);
2539
2540         *to_destroy = NULL;
2541         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2542                 return 0;
2543
2544         dev_id_priv =
2545                 __rdma_create_id(net, cma_listen_handler, id_priv,
2546                                  id_priv->id.ps, id_priv->id.qp_type, id_priv);
2547         if (IS_ERR(dev_id_priv))
2548                 return PTR_ERR(dev_id_priv);
2549
2550         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2551         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2552                rdma_addr_size(cma_src_addr(id_priv)));
2553
2554         _cma_attach_to_dev(dev_id_priv, cma_dev);
2555         rdma_restrack_add(&dev_id_priv->res);
2556         cma_id_get(id_priv);
2557         dev_id_priv->internal_id = 1;
2558         dev_id_priv->afonly = id_priv->afonly;
2559         mutex_lock(&id_priv->qp_mutex);
2560         dev_id_priv->tos_set = id_priv->tos_set;
2561         dev_id_priv->tos = id_priv->tos;
2562         mutex_unlock(&id_priv->qp_mutex);
2563
2564         ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2565         if (ret)
2566                 goto err_listen;
2567         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2568         return 0;
2569 err_listen:
2570         /* Caller must destroy this after releasing lock */
2571         *to_destroy = dev_id_priv;
2572         dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2573         return ret;
2574 }
2575
2576 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2577 {
2578         struct rdma_id_private *to_destroy;
2579         struct cma_device *cma_dev;
2580         int ret;
2581
2582         mutex_lock(&lock);
2583         list_add_tail(&id_priv->list, &listen_any_list);
2584         list_for_each_entry(cma_dev, &dev_list, list) {
2585                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2586                 if (ret) {
2587                         /* Prevent racing with cma_process_remove() */
2588                         if (to_destroy)
2589                                 list_del_init(&to_destroy->list);
2590                         goto err_listen;
2591                 }
2592         }
2593         mutex_unlock(&lock);
2594         return 0;
2595
2596 err_listen:
2597         _cma_cancel_listens(id_priv);
2598         mutex_unlock(&lock);
2599         if (to_destroy)
2600                 rdma_destroy_id(&to_destroy->id);
2601         return ret;
2602 }
2603
2604 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2605 {
2606         struct rdma_id_private *id_priv;
2607
2608         id_priv = container_of(id, struct rdma_id_private, id);
2609         mutex_lock(&id_priv->qp_mutex);
2610         id_priv->tos = (u8) tos;
2611         id_priv->tos_set = true;
2612         mutex_unlock(&id_priv->qp_mutex);
2613 }
2614 EXPORT_SYMBOL(rdma_set_service_type);
2615
2616 /**
2617  * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2618  *                          with a connection identifier.
2619  * @id: Communication identifier to associated with service type.
2620  * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2621  *
2622  * This function should be called before rdma_connect() on active side,
2623  * and on passive side before rdma_accept(). It is applicable to primary
2624  * path only. The timeout will affect the local side of the QP, it is not
2625  * negotiated with remote side and zero disables the timer. In case it is
2626  * set before rdma_resolve_route, the value will also be used to determine
2627  * PacketLifeTime for RoCE.
2628  *
2629  * Return: 0 for success
2630  */
2631 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2632 {
2633         struct rdma_id_private *id_priv;
2634
2635         if (id->qp_type != IB_QPT_RC)
2636                 return -EINVAL;
2637
2638         id_priv = container_of(id, struct rdma_id_private, id);
2639         mutex_lock(&id_priv->qp_mutex);
2640         id_priv->timeout = timeout;
2641         id_priv->timeout_set = true;
2642         mutex_unlock(&id_priv->qp_mutex);
2643
2644         return 0;
2645 }
2646 EXPORT_SYMBOL(rdma_set_ack_timeout);
2647
2648 /**
2649  * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2650  *                            QP associated with a connection identifier.
2651  * @id: Communication identifier to associated with service type.
2652  * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2653  *                 Timer Field" in the IBTA specification.
2654  *
2655  * This function should be called before rdma_connect() on active
2656  * side, and on passive side before rdma_accept(). The timer value
2657  * will be associated with the local QP. When it receives a send it is
2658  * not read to handle, typically if the receive queue is empty, an RNR
2659  * Retry NAK is returned to the requester with the min_rnr_timer
2660  * encoded. The requester will then wait at least the time specified
2661  * in the NAK before retrying. The default is zero, which translates
2662  * to a minimum RNR Timer value of 655 ms.
2663  *
2664  * Return: 0 for success
2665  */
2666 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2667 {
2668         struct rdma_id_private *id_priv;
2669
2670         /* It is a five-bit value */
2671         if (min_rnr_timer & 0xe0)
2672                 return -EINVAL;
2673
2674         if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2675                 return -EINVAL;
2676
2677         id_priv = container_of(id, struct rdma_id_private, id);
2678         mutex_lock(&id_priv->qp_mutex);
2679         id_priv->min_rnr_timer = min_rnr_timer;
2680         id_priv->min_rnr_timer_set = true;
2681         mutex_unlock(&id_priv->qp_mutex);
2682
2683         return 0;
2684 }
2685 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2686
2687 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2688                               void *context)
2689 {
2690         struct cma_work *work = context;
2691         struct rdma_route *route;
2692
2693         route = &work->id->id.route;
2694
2695         if (!status) {
2696                 route->num_paths = 1;
2697                 *route->path_rec = *path_rec;
2698         } else {
2699                 work->old_state = RDMA_CM_ROUTE_QUERY;
2700                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2701                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2702                 work->event.status = status;
2703                 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2704                                      status);
2705         }
2706
2707         queue_work(cma_wq, &work->work);
2708 }
2709
2710 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2711                               unsigned long timeout_ms, struct cma_work *work)
2712 {
2713         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2714         struct sa_path_rec path_rec;
2715         ib_sa_comp_mask comp_mask;
2716         struct sockaddr_in6 *sin6;
2717         struct sockaddr_ib *sib;
2718
2719         memset(&path_rec, 0, sizeof path_rec);
2720
2721         if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2722                 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2723         else
2724                 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2725         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2726         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2727         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2728         path_rec.numb_path = 1;
2729         path_rec.reversible = 1;
2730         path_rec.service_id = rdma_get_service_id(&id_priv->id,
2731                                                   cma_dst_addr(id_priv));
2732
2733         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2734                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2735                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2736
2737         switch (cma_family(id_priv)) {
2738         case AF_INET:
2739                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2740                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2741                 break;
2742         case AF_INET6:
2743                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2744                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2745                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2746                 break;
2747         case AF_IB:
2748                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2749                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2750                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2751                 break;
2752         }
2753
2754         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2755                                                id_priv->id.port_num, &path_rec,
2756                                                comp_mask, timeout_ms,
2757                                                GFP_KERNEL, cma_query_handler,
2758                                                work, &id_priv->query);
2759
2760         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2761 }
2762
2763 static void cma_iboe_join_work_handler(struct work_struct *work)
2764 {
2765         struct cma_multicast *mc =
2766                 container_of(work, struct cma_multicast, iboe_join.work);
2767         struct rdma_cm_event *event = &mc->iboe_join.event;
2768         struct rdma_id_private *id_priv = mc->id_priv;
2769         int ret;
2770
2771         mutex_lock(&id_priv->handler_mutex);
2772         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2773             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2774                 goto out_unlock;
2775
2776         ret = cma_cm_event_handler(id_priv, event);
2777         WARN_ON(ret);
2778
2779 out_unlock:
2780         mutex_unlock(&id_priv->handler_mutex);
2781         if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2782                 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2783 }
2784
2785 static void cma_work_handler(struct work_struct *_work)
2786 {
2787         struct cma_work *work = container_of(_work, struct cma_work, work);
2788         struct rdma_id_private *id_priv = work->id;
2789
2790         mutex_lock(&id_priv->handler_mutex);
2791         if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2792             READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2793                 goto out_unlock;
2794         if (work->old_state != 0 || work->new_state != 0) {
2795                 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2796                         goto out_unlock;
2797         }
2798
2799         if (cma_cm_event_handler(id_priv, &work->event)) {
2800                 cma_id_put(id_priv);
2801                 destroy_id_handler_unlock(id_priv);
2802                 goto out_free;
2803         }
2804
2805 out_unlock:
2806         mutex_unlock(&id_priv->handler_mutex);
2807         cma_id_put(id_priv);
2808 out_free:
2809         if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2810                 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2811         kfree(work);
2812 }
2813
2814 static void cma_init_resolve_route_work(struct cma_work *work,
2815                                         struct rdma_id_private *id_priv)
2816 {
2817         work->id = id_priv;
2818         INIT_WORK(&work->work, cma_work_handler);
2819         work->old_state = RDMA_CM_ROUTE_QUERY;
2820         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2821         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2822 }
2823
2824 static void enqueue_resolve_addr_work(struct cma_work *work,
2825                                       struct rdma_id_private *id_priv)
2826 {
2827         /* Balances with cma_id_put() in cma_work_handler */
2828         cma_id_get(id_priv);
2829
2830         work->id = id_priv;
2831         INIT_WORK(&work->work, cma_work_handler);
2832         work->old_state = RDMA_CM_ADDR_QUERY;
2833         work->new_state = RDMA_CM_ADDR_RESOLVED;
2834         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2835
2836         queue_work(cma_wq, &work->work);
2837 }
2838
2839 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2840                                 unsigned long timeout_ms)
2841 {
2842         struct rdma_route *route = &id_priv->id.route;
2843         struct cma_work *work;
2844         int ret;
2845
2846         work = kzalloc(sizeof *work, GFP_KERNEL);
2847         if (!work)
2848                 return -ENOMEM;
2849
2850         cma_init_resolve_route_work(work, id_priv);
2851
2852         if (!route->path_rec)
2853                 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2854         if (!route->path_rec) {
2855                 ret = -ENOMEM;
2856                 goto err1;
2857         }
2858
2859         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2860         if (ret)
2861                 goto err2;
2862
2863         return 0;
2864 err2:
2865         kfree(route->path_rec);
2866         route->path_rec = NULL;
2867 err1:
2868         kfree(work);
2869         return ret;
2870 }
2871
2872 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2873                                            unsigned long supported_gids,
2874                                            enum ib_gid_type default_gid)
2875 {
2876         if ((network_type == RDMA_NETWORK_IPV4 ||
2877              network_type == RDMA_NETWORK_IPV6) &&
2878             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2879                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2880
2881         return default_gid;
2882 }
2883
2884 /*
2885  * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2886  * path record type based on GID type.
2887  * It also sets up other L2 fields which includes destination mac address
2888  * netdev ifindex, of the path record.
2889  * It returns the netdev of the bound interface for this path record entry.
2890  */
2891 static struct net_device *
2892 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2893 {
2894         struct rdma_route *route = &id_priv->id.route;
2895         enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2896         struct rdma_addr *addr = &route->addr;
2897         unsigned long supported_gids;
2898         struct net_device *ndev;
2899
2900         if (!addr->dev_addr.bound_dev_if)
2901                 return NULL;
2902
2903         ndev = dev_get_by_index(addr->dev_addr.net,
2904                                 addr->dev_addr.bound_dev_if);
2905         if (!ndev)
2906                 return NULL;
2907
2908         supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2909                                                     id_priv->id.port_num);
2910         gid_type = cma_route_gid_type(addr->dev_addr.network,
2911                                       supported_gids,
2912                                       id_priv->gid_type);
2913         /* Use the hint from IP Stack to select GID Type */
2914         if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2915                 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2916         route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2917
2918         route->path_rec->roce.route_resolved = true;
2919         sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2920         return ndev;
2921 }
2922
2923 int rdma_set_ib_path(struct rdma_cm_id *id,
2924                      struct sa_path_rec *path_rec)
2925 {
2926         struct rdma_id_private *id_priv;
2927         struct net_device *ndev;
2928         int ret;
2929
2930         id_priv = container_of(id, struct rdma_id_private, id);
2931         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2932                            RDMA_CM_ROUTE_RESOLVED))
2933                 return -EINVAL;
2934
2935         id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2936                                      GFP_KERNEL);
2937         if (!id->route.path_rec) {
2938                 ret = -ENOMEM;
2939                 goto err;
2940         }
2941
2942         if (rdma_protocol_roce(id->device, id->port_num)) {
2943                 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2944                 if (!ndev) {
2945                         ret = -ENODEV;
2946                         goto err_free;
2947                 }
2948                 dev_put(ndev);
2949         }
2950
2951         id->route.num_paths = 1;
2952         return 0;
2953
2954 err_free:
2955         kfree(id->route.path_rec);
2956         id->route.path_rec = NULL;
2957 err:
2958         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2959         return ret;
2960 }
2961 EXPORT_SYMBOL(rdma_set_ib_path);
2962
2963 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2964 {
2965         struct cma_work *work;
2966
2967         work = kzalloc(sizeof *work, GFP_KERNEL);
2968         if (!work)
2969                 return -ENOMEM;
2970
2971         cma_init_resolve_route_work(work, id_priv);
2972         queue_work(cma_wq, &work->work);
2973         return 0;
2974 }
2975
2976 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2977 {
2978         struct net_device *dev;
2979
2980         dev = vlan_dev_real_dev(vlan_ndev);
2981         if (dev->num_tc)
2982                 return netdev_get_prio_tc_map(dev, prio);
2983
2984         return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2985                 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2986 }
2987
2988 struct iboe_prio_tc_map {
2989         int input_prio;
2990         int output_tc;
2991         bool found;
2992 };
2993
2994 static int get_lower_vlan_dev_tc(struct net_device *dev,
2995                                  struct netdev_nested_priv *priv)
2996 {
2997         struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2998
2999         if (is_vlan_dev(dev))
3000                 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3001         else if (dev->num_tc)
3002                 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3003         else
3004                 map->output_tc = 0;
3005         /* We are interested only in first level VLAN device, so always
3006          * return 1 to stop iterating over next level devices.
3007          */
3008         map->found = true;
3009         return 1;
3010 }
3011
3012 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3013 {
3014         struct iboe_prio_tc_map prio_tc_map = {};
3015         int prio = rt_tos2priority(tos);
3016         struct netdev_nested_priv priv;
3017
3018         /* If VLAN device, get it directly from the VLAN netdev */
3019         if (is_vlan_dev(ndev))
3020                 return get_vlan_ndev_tc(ndev, prio);
3021
3022         prio_tc_map.input_prio = prio;
3023         priv.data = (void *)&prio_tc_map;
3024         rcu_read_lock();
3025         netdev_walk_all_lower_dev_rcu(ndev,
3026                                       get_lower_vlan_dev_tc,
3027                                       &priv);
3028         rcu_read_unlock();
3029         /* If map is found from lower device, use it; Otherwise
3030          * continue with the current netdevice to get priority to tc map.
3031          */
3032         if (prio_tc_map.found)
3033                 return prio_tc_map.output_tc;
3034         else if (ndev->num_tc)
3035                 return netdev_get_prio_tc_map(ndev, prio);
3036         else
3037                 return 0;
3038 }
3039
3040 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3041 {
3042         struct sockaddr_in6 *addr6;
3043         u16 dport, sport;
3044         u32 hash, fl;
3045
3046         addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3047         fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3048         if ((cma_family(id_priv) != AF_INET6) || !fl) {
3049                 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3050                 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3051                 hash = (u32)sport * 31 + dport;
3052                 fl = hash & IB_GRH_FLOWLABEL_MASK;
3053         }
3054
3055         return cpu_to_be32(fl);
3056 }
3057
3058 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3059 {
3060         struct rdma_route *route = &id_priv->id.route;
3061         struct rdma_addr *addr = &route->addr;
3062         struct cma_work *work;
3063         int ret;
3064         struct net_device *ndev;
3065
3066         u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3067                                         rdma_start_port(id_priv->cma_dev->device)];
3068         u8 tos;
3069
3070         mutex_lock(&id_priv->qp_mutex);
3071         tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3072         mutex_unlock(&id_priv->qp_mutex);
3073
3074         work = kzalloc(sizeof *work, GFP_KERNEL);
3075         if (!work)
3076                 return -ENOMEM;
3077
3078         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3079         if (!route->path_rec) {
3080                 ret = -ENOMEM;
3081                 goto err1;
3082         }
3083
3084         route->num_paths = 1;
3085
3086         ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3087         if (!ndev) {
3088                 ret = -ENODEV;
3089                 goto err2;
3090         }
3091
3092         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3093                     &route->path_rec->sgid);
3094         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3095                     &route->path_rec->dgid);
3096
3097         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3098                 /* TODO: get the hoplimit from the inet/inet6 device */
3099                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3100         else
3101                 route->path_rec->hop_limit = 1;
3102         route->path_rec->reversible = 1;
3103         route->path_rec->pkey = cpu_to_be16(0xffff);
3104         route->path_rec->mtu_selector = IB_SA_EQ;
3105         route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3106         route->path_rec->traffic_class = tos;
3107         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3108         route->path_rec->rate_selector = IB_SA_EQ;
3109         route->path_rec->rate = iboe_get_rate(ndev);
3110         dev_put(ndev);
3111         route->path_rec->packet_life_time_selector = IB_SA_EQ;
3112         /* In case ACK timeout is set, use this value to calculate
3113          * PacketLifeTime.  As per IBTA 12.7.34,
3114          * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3115          * Assuming a negligible local ACK delay, we can use
3116          * PacketLifeTime = local ACK timeout/2
3117          * as a reasonable approximation for RoCE networks.
3118          */
3119         mutex_lock(&id_priv->qp_mutex);
3120         if (id_priv->timeout_set && id_priv->timeout)
3121                 route->path_rec->packet_life_time = id_priv->timeout - 1;
3122         else
3123                 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3124         mutex_unlock(&id_priv->qp_mutex);
3125
3126         if (!route->path_rec->mtu) {
3127                 ret = -EINVAL;
3128                 goto err2;
3129         }
3130
3131         if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3132                                          id_priv->id.port_num))
3133                 route->path_rec->flow_label =
3134                         cma_get_roce_udp_flow_label(id_priv);
3135
3136         cma_init_resolve_route_work(work, id_priv);
3137         queue_work(cma_wq, &work->work);
3138
3139         return 0;
3140
3141 err2:
3142         kfree(route->path_rec);
3143         route->path_rec = NULL;
3144         route->num_paths = 0;
3145 err1:
3146         kfree(work);
3147         return ret;
3148 }
3149
3150 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3151 {
3152         struct rdma_id_private *id_priv;
3153         int ret;
3154
3155         if (!timeout_ms)
3156                 return -EINVAL;
3157
3158         id_priv = container_of(id, struct rdma_id_private, id);
3159         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3160                 return -EINVAL;
3161
3162         cma_id_get(id_priv);
3163         if (rdma_cap_ib_sa(id->device, id->port_num))
3164                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3165         else if (rdma_protocol_roce(id->device, id->port_num))
3166                 ret = cma_resolve_iboe_route(id_priv);
3167         else if (rdma_protocol_iwarp(id->device, id->port_num))
3168                 ret = cma_resolve_iw_route(id_priv);
3169         else
3170                 ret = -ENOSYS;
3171
3172         if (ret)
3173                 goto err;
3174
3175         return 0;
3176 err:
3177         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3178         cma_id_put(id_priv);
3179         return ret;
3180 }
3181 EXPORT_SYMBOL(rdma_resolve_route);
3182
3183 static void cma_set_loopback(struct sockaddr *addr)
3184 {
3185         switch (addr->sa_family) {
3186         case AF_INET:
3187                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3188                 break;
3189         case AF_INET6:
3190                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3191                               0, 0, 0, htonl(1));
3192                 break;
3193         default:
3194                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3195                             0, 0, 0, htonl(1));
3196                 break;
3197         }
3198 }
3199
3200 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3201 {
3202         struct cma_device *cma_dev, *cur_dev;
3203         union ib_gid gid;
3204         enum ib_port_state port_state;
3205         unsigned int p;
3206         u16 pkey;
3207         int ret;
3208
3209         cma_dev = NULL;
3210         mutex_lock(&lock);
3211         list_for_each_entry(cur_dev, &dev_list, list) {
3212                 if (cma_family(id_priv) == AF_IB &&
3213                     !rdma_cap_ib_cm(cur_dev->device, 1))
3214                         continue;
3215
3216                 if (!cma_dev)
3217                         cma_dev = cur_dev;
3218
3219                 rdma_for_each_port (cur_dev->device, p) {
3220                         if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3221                             port_state == IB_PORT_ACTIVE) {
3222                                 cma_dev = cur_dev;
3223                                 goto port_found;
3224                         }
3225                 }
3226         }
3227
3228         if (!cma_dev) {
3229                 ret = -ENODEV;
3230                 goto out;
3231         }
3232
3233         p = 1;
3234
3235 port_found:
3236         ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3237         if (ret)
3238                 goto out;
3239
3240         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3241         if (ret)
3242                 goto out;
3243
3244         id_priv->id.route.addr.dev_addr.dev_type =
3245                 (rdma_protocol_ib(cma_dev->device, p)) ?
3246                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3247
3248         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3249         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3250         id_priv->id.port_num = p;
3251         cma_attach_to_dev(id_priv, cma_dev);
3252         rdma_restrack_add(&id_priv->res);
3253         cma_set_loopback(cma_src_addr(id_priv));
3254 out:
3255         mutex_unlock(&lock);
3256         return ret;
3257 }
3258
3259 static void addr_handler(int status, struct sockaddr *src_addr,
3260                          struct rdma_dev_addr *dev_addr, void *context)
3261 {
3262         struct rdma_id_private *id_priv = context;
3263         struct rdma_cm_event event = {};
3264         struct sockaddr *addr;
3265         struct sockaddr_storage old_addr;
3266
3267         mutex_lock(&id_priv->handler_mutex);
3268         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3269                            RDMA_CM_ADDR_RESOLVED))
3270                 goto out;
3271
3272         /*
3273          * Store the previous src address, so that if we fail to acquire
3274          * matching rdma device, old address can be restored back, which helps
3275          * to cancel the cma listen operation correctly.
3276          */
3277         addr = cma_src_addr(id_priv);
3278         memcpy(&old_addr, addr, rdma_addr_size(addr));
3279         memcpy(addr, src_addr, rdma_addr_size(src_addr));
3280         if (!status && !id_priv->cma_dev) {
3281                 status = cma_acquire_dev_by_src_ip(id_priv);
3282                 if (status)
3283                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3284                                              status);
3285                 rdma_restrack_add(&id_priv->res);
3286         } else if (status) {
3287                 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3288         }
3289
3290         if (status) {
3291                 memcpy(addr, &old_addr,
3292                        rdma_addr_size((struct sockaddr *)&old_addr));
3293                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3294                                    RDMA_CM_ADDR_BOUND))
3295                         goto out;
3296                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3297                 event.status = status;
3298         } else
3299                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3300
3301         if (cma_cm_event_handler(id_priv, &event)) {
3302                 destroy_id_handler_unlock(id_priv);
3303                 return;
3304         }
3305 out:
3306         mutex_unlock(&id_priv->handler_mutex);
3307 }
3308
3309 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3310 {
3311         struct cma_work *work;
3312         union ib_gid gid;
3313         int ret;
3314
3315         work = kzalloc(sizeof *work, GFP_KERNEL);
3316         if (!work)
3317                 return -ENOMEM;
3318
3319         if (!id_priv->cma_dev) {
3320                 ret = cma_bind_loopback(id_priv);
3321                 if (ret)
3322                         goto err;
3323         }
3324
3325         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3326         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3327
3328         enqueue_resolve_addr_work(work, id_priv);
3329         return 0;
3330 err:
3331         kfree(work);
3332         return ret;
3333 }
3334
3335 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3336 {
3337         struct cma_work *work;
3338         int ret;
3339
3340         work = kzalloc(sizeof *work, GFP_KERNEL);
3341         if (!work)
3342                 return -ENOMEM;
3343
3344         if (!id_priv->cma_dev) {
3345                 ret = cma_resolve_ib_dev(id_priv);
3346                 if (ret)
3347                         goto err;
3348         }
3349
3350         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3351                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3352
3353         enqueue_resolve_addr_work(work, id_priv);
3354         return 0;
3355 err:
3356         kfree(work);
3357         return ret;
3358 }
3359
3360 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3361                          const struct sockaddr *dst_addr)
3362 {
3363         if (!src_addr || !src_addr->sa_family) {
3364                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3365                 src_addr->sa_family = dst_addr->sa_family;
3366                 if (IS_ENABLED(CONFIG_IPV6) &&
3367                     dst_addr->sa_family == AF_INET6) {
3368                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3369                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3370                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3371                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3372                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3373                 } else if (dst_addr->sa_family == AF_IB) {
3374                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
3375                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3376                 }
3377         }
3378         return rdma_bind_addr(id, src_addr);
3379 }
3380
3381 /*
3382  * If required, resolve the source address for bind and leave the id_priv in
3383  * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3384  * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3385  * ignored.
3386  */
3387 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3388                                struct sockaddr *src_addr,
3389                                const struct sockaddr *dst_addr)
3390 {
3391         int ret;
3392
3393         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3394         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3395                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3396                 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3397                 if (ret)
3398                         goto err_dst;
3399                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3400                                            RDMA_CM_ADDR_QUERY))) {
3401                         ret = -EINVAL;
3402                         goto err_dst;
3403                 }
3404         }
3405
3406         if (cma_family(id_priv) != dst_addr->sa_family) {
3407                 ret = -EINVAL;
3408                 goto err_state;
3409         }
3410         return 0;
3411
3412 err_state:
3413         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3414 err_dst:
3415         memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3416         return ret;
3417 }
3418
3419 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3420                       const struct sockaddr *dst_addr, unsigned long timeout_ms)
3421 {
3422         struct rdma_id_private *id_priv =
3423                 container_of(id, struct rdma_id_private, id);
3424         int ret;
3425
3426         ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3427         if (ret)
3428                 return ret;
3429
3430         if (cma_any_addr(dst_addr)) {
3431                 ret = cma_resolve_loopback(id_priv);
3432         } else {
3433                 if (dst_addr->sa_family == AF_IB) {
3434                         ret = cma_resolve_ib_addr(id_priv);
3435                 } else {
3436                         /*
3437                          * The FSM can return back to RDMA_CM_ADDR_BOUND after
3438                          * rdma_resolve_ip() is called, eg through the error
3439                          * path in addr_handler(). If this happens the existing
3440                          * request must be canceled before issuing a new one.
3441                          * Since canceling a request is a bit slow and this
3442                          * oddball path is rare, keep track once a request has
3443                          * been issued. The track turns out to be a permanent
3444                          * state since this is the only cancel as it is
3445                          * immediately before rdma_resolve_ip().
3446                          */
3447                         if (id_priv->used_resolve_ip)
3448                                 rdma_addr_cancel(&id->route.addr.dev_addr);
3449                         else
3450                                 id_priv->used_resolve_ip = 1;
3451                         ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3452                                               &id->route.addr.dev_addr,
3453                                               timeout_ms, addr_handler,
3454                                               false, id_priv);
3455                 }
3456         }
3457         if (ret)
3458                 goto err;
3459
3460         return 0;
3461 err:
3462         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3463         return ret;
3464 }
3465 EXPORT_SYMBOL(rdma_resolve_addr);
3466
3467 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3468 {
3469         struct rdma_id_private *id_priv;
3470         unsigned long flags;
3471         int ret;
3472
3473         id_priv = container_of(id, struct rdma_id_private, id);
3474         spin_lock_irqsave(&id_priv->lock, flags);
3475         if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3476             id_priv->state == RDMA_CM_IDLE) {
3477                 id_priv->reuseaddr = reuse;
3478                 ret = 0;
3479         } else {
3480                 ret = -EINVAL;
3481         }
3482         spin_unlock_irqrestore(&id_priv->lock, flags);
3483         return ret;
3484 }
3485 EXPORT_SYMBOL(rdma_set_reuseaddr);
3486
3487 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3488 {
3489         struct rdma_id_private *id_priv;
3490         unsigned long flags;
3491         int ret;
3492
3493         id_priv = container_of(id, struct rdma_id_private, id);
3494         spin_lock_irqsave(&id_priv->lock, flags);
3495         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3496                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3497                 id_priv->afonly = afonly;
3498                 ret = 0;
3499         } else {
3500                 ret = -EINVAL;
3501         }
3502         spin_unlock_irqrestore(&id_priv->lock, flags);
3503         return ret;
3504 }
3505 EXPORT_SYMBOL(rdma_set_afonly);
3506
3507 static void cma_bind_port(struct rdma_bind_list *bind_list,
3508                           struct rdma_id_private *id_priv)
3509 {
3510         struct sockaddr *addr;
3511         struct sockaddr_ib *sib;
3512         u64 sid, mask;
3513         __be16 port;
3514
3515         lockdep_assert_held(&lock);
3516
3517         addr = cma_src_addr(id_priv);
3518         port = htons(bind_list->port);
3519
3520         switch (addr->sa_family) {
3521         case AF_INET:
3522                 ((struct sockaddr_in *) addr)->sin_port = port;
3523                 break;
3524         case AF_INET6:
3525                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3526                 break;
3527         case AF_IB:
3528                 sib = (struct sockaddr_ib *) addr;
3529                 sid = be64_to_cpu(sib->sib_sid);
3530                 mask = be64_to_cpu(sib->sib_sid_mask);
3531                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3532                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3533                 break;
3534         }
3535         id_priv->bind_list = bind_list;
3536         hlist_add_head(&id_priv->node, &bind_list->owners);
3537 }
3538
3539 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3540                           struct rdma_id_private *id_priv, unsigned short snum)
3541 {
3542         struct rdma_bind_list *bind_list;
3543         int ret;
3544
3545         lockdep_assert_held(&lock);
3546
3547         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3548         if (!bind_list)
3549                 return -ENOMEM;
3550
3551         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3552                            snum);
3553         if (ret < 0)
3554                 goto err;
3555
3556         bind_list->ps = ps;
3557         bind_list->port = snum;
3558         cma_bind_port(bind_list, id_priv);
3559         return 0;
3560 err:
3561         kfree(bind_list);
3562         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3563 }
3564
3565 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3566                               struct rdma_id_private *id_priv)
3567 {
3568         struct rdma_id_private *cur_id;
3569         struct sockaddr  *daddr = cma_dst_addr(id_priv);
3570         struct sockaddr  *saddr = cma_src_addr(id_priv);
3571         __be16 dport = cma_port(daddr);
3572
3573         lockdep_assert_held(&lock);
3574
3575         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3576                 struct sockaddr  *cur_daddr = cma_dst_addr(cur_id);
3577                 struct sockaddr  *cur_saddr = cma_src_addr(cur_id);
3578                 __be16 cur_dport = cma_port(cur_daddr);
3579
3580                 if (id_priv == cur_id)
3581                         continue;
3582
3583                 /* different dest port -> unique */
3584                 if (!cma_any_port(daddr) &&
3585                     !cma_any_port(cur_daddr) &&
3586                     (dport != cur_dport))
3587                         continue;
3588
3589                 /* different src address -> unique */
3590                 if (!cma_any_addr(saddr) &&
3591                     !cma_any_addr(cur_saddr) &&
3592                     cma_addr_cmp(saddr, cur_saddr))
3593                         continue;
3594
3595                 /* different dst address -> unique */
3596                 if (!cma_any_addr(daddr) &&
3597                     !cma_any_addr(cur_daddr) &&
3598                     cma_addr_cmp(daddr, cur_daddr))
3599                         continue;
3600
3601                 return -EADDRNOTAVAIL;
3602         }
3603         return 0;
3604 }
3605
3606 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3607                               struct rdma_id_private *id_priv)
3608 {
3609         static unsigned int last_used_port;
3610         int low, high, remaining;
3611         unsigned int rover;
3612         struct net *net = id_priv->id.route.addr.dev_addr.net;
3613
3614         lockdep_assert_held(&lock);
3615
3616         inet_get_local_port_range(net, &low, &high);
3617         remaining = (high - low) + 1;
3618         rover = prandom_u32() % remaining + low;
3619 retry:
3620         if (last_used_port != rover) {
3621                 struct rdma_bind_list *bind_list;
3622                 int ret;
3623
3624                 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3625
3626                 if (!bind_list) {
3627                         ret = cma_alloc_port(ps, id_priv, rover);
3628                 } else {
3629                         ret = cma_port_is_unique(bind_list, id_priv);
3630                         if (!ret)
3631                                 cma_bind_port(bind_list, id_priv);
3632                 }
3633                 /*
3634                  * Remember previously used port number in order to avoid
3635                  * re-using same port immediately after it is closed.
3636                  */
3637                 if (!ret)
3638                         last_used_port = rover;
3639                 if (ret != -EADDRNOTAVAIL)
3640                         return ret;
3641         }
3642         if (--remaining) {
3643                 rover++;
3644                 if ((rover < low) || (rover > high))
3645                         rover = low;
3646                 goto retry;
3647         }
3648         return -EADDRNOTAVAIL;
3649 }
3650
3651 /*
3652  * Check that the requested port is available.  This is called when trying to
3653  * bind to a specific port, or when trying to listen on a bound port.  In
3654  * the latter case, the provided id_priv may already be on the bind_list, but
3655  * we still need to check that it's okay to start listening.
3656  */
3657 static int cma_check_port(struct rdma_bind_list *bind_list,
3658                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
3659 {
3660         struct rdma_id_private *cur_id;
3661         struct sockaddr *addr, *cur_addr;
3662
3663         lockdep_assert_held(&lock);
3664
3665         addr = cma_src_addr(id_priv);
3666         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3667                 if (id_priv == cur_id)
3668                         continue;
3669
3670                 if (reuseaddr && cur_id->reuseaddr)
3671                         continue;
3672
3673                 cur_addr = cma_src_addr(cur_id);
3674                 if (id_priv->afonly && cur_id->afonly &&
3675                     (addr->sa_family != cur_addr->sa_family))
3676                         continue;
3677
3678                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3679                         return -EADDRNOTAVAIL;
3680
3681                 if (!cma_addr_cmp(addr, cur_addr))
3682                         return -EADDRINUSE;
3683         }
3684         return 0;
3685 }
3686
3687 static int cma_use_port(enum rdma_ucm_port_space ps,
3688                         struct rdma_id_private *id_priv)
3689 {
3690         struct rdma_bind_list *bind_list;
3691         unsigned short snum;
3692         int ret;
3693
3694         lockdep_assert_held(&lock);
3695
3696         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3697         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3698                 return -EACCES;
3699
3700         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3701         if (!bind_list) {
3702                 ret = cma_alloc_port(ps, id_priv, snum);
3703         } else {
3704                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3705                 if (!ret)
3706                         cma_bind_port(bind_list, id_priv);
3707         }
3708         return ret;
3709 }
3710
3711 static enum rdma_ucm_port_space
3712 cma_select_inet_ps(struct rdma_id_private *id_priv)
3713 {
3714         switch (id_priv->id.ps) {
3715         case RDMA_PS_TCP:
3716         case RDMA_PS_UDP:
3717         case RDMA_PS_IPOIB:
3718         case RDMA_PS_IB:
3719                 return id_priv->id.ps;
3720         default:
3721
3722                 return 0;
3723         }
3724 }
3725
3726 static enum rdma_ucm_port_space
3727 cma_select_ib_ps(struct rdma_id_private *id_priv)
3728 {
3729         enum rdma_ucm_port_space ps = 0;
3730         struct sockaddr_ib *sib;
3731         u64 sid_ps, mask, sid;
3732
3733         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3734         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3735         sid = be64_to_cpu(sib->sib_sid) & mask;
3736
3737         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3738                 sid_ps = RDMA_IB_IP_PS_IB;
3739                 ps = RDMA_PS_IB;
3740         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3741                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3742                 sid_ps = RDMA_IB_IP_PS_TCP;
3743                 ps = RDMA_PS_TCP;
3744         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3745                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3746                 sid_ps = RDMA_IB_IP_PS_UDP;
3747                 ps = RDMA_PS_UDP;
3748         }
3749
3750         if (ps) {
3751                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3752                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3753                                                 be64_to_cpu(sib->sib_sid_mask));
3754         }
3755         return ps;
3756 }
3757
3758 static int cma_get_port(struct rdma_id_private *id_priv)
3759 {
3760         enum rdma_ucm_port_space ps;
3761         int ret;
3762
3763         if (cma_family(id_priv) != AF_IB)
3764                 ps = cma_select_inet_ps(id_priv);
3765         else
3766                 ps = cma_select_ib_ps(id_priv);
3767         if (!ps)
3768                 return -EPROTONOSUPPORT;
3769
3770         mutex_lock(&lock);
3771         if (cma_any_port(cma_src_addr(id_priv)))
3772                 ret = cma_alloc_any_port(ps, id_priv);
3773         else
3774                 ret = cma_use_port(ps, id_priv);
3775         mutex_unlock(&lock);
3776
3777         return ret;
3778 }
3779
3780 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3781                                struct sockaddr *addr)
3782 {
3783 #if IS_ENABLED(CONFIG_IPV6)
3784         struct sockaddr_in6 *sin6;
3785
3786         if (addr->sa_family != AF_INET6)
3787                 return 0;
3788
3789         sin6 = (struct sockaddr_in6 *) addr;
3790
3791         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3792                 return 0;
3793
3794         if (!sin6->sin6_scope_id)
3795                         return -EINVAL;
3796
3797         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3798 #endif
3799         return 0;
3800 }
3801
3802 int rdma_listen(struct rdma_cm_id *id, int backlog)
3803 {
3804         struct rdma_id_private *id_priv =
3805                 container_of(id, struct rdma_id_private, id);
3806         int ret;
3807
3808         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3809                 struct sockaddr_in any_in = {
3810                         .sin_family = AF_INET,
3811                         .sin_addr.s_addr = htonl(INADDR_ANY),
3812                 };
3813
3814                 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3815                 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3816                 if (ret)
3817                         return ret;
3818                 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3819                                            RDMA_CM_LISTEN)))
3820                         return -EINVAL;
3821         }
3822
3823         /*
3824          * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3825          * any more, and has to be unique in the bind list.
3826          */
3827         if (id_priv->reuseaddr) {
3828                 mutex_lock(&lock);
3829                 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3830                 if (!ret)
3831                         id_priv->reuseaddr = 0;
3832                 mutex_unlock(&lock);
3833                 if (ret)
3834                         goto err;
3835         }
3836
3837         id_priv->backlog = backlog;
3838         if (id_priv->cma_dev) {
3839                 if (rdma_cap_ib_cm(id->device, 1)) {
3840                         ret = cma_ib_listen(id_priv);
3841                         if (ret)
3842                                 goto err;
3843                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3844                         ret = cma_iw_listen(id_priv, backlog);
3845                         if (ret)
3846                                 goto err;
3847                 } else {
3848                         ret = -ENOSYS;
3849                         goto err;
3850                 }
3851         } else {
3852                 ret = cma_listen_on_all(id_priv);
3853                 if (ret)
3854                         goto err;
3855         }
3856
3857         return 0;
3858 err:
3859         id_priv->backlog = 0;
3860         /*
3861          * All the failure paths that lead here will not allow the req_handler's
3862          * to have run.
3863          */
3864         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3865         return ret;
3866 }
3867 EXPORT_SYMBOL(rdma_listen);
3868
3869 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3870 {
3871         struct rdma_id_private *id_priv;
3872         int ret;
3873         struct sockaddr  *daddr;
3874
3875         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3876             addr->sa_family != AF_IB)
3877                 return -EAFNOSUPPORT;
3878
3879         id_priv = container_of(id, struct rdma_id_private, id);
3880         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3881                 return -EINVAL;
3882
3883         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3884         if (ret)
3885                 goto err1;
3886
3887         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3888         if (!cma_any_addr(addr)) {
3889                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3890                 if (ret)
3891                         goto err1;
3892
3893                 ret = cma_acquire_dev_by_src_ip(id_priv);
3894                 if (ret)
3895                         goto err1;
3896         }
3897
3898         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3899                 if (addr->sa_family == AF_INET)
3900                         id_priv->afonly = 1;
3901 #if IS_ENABLED(CONFIG_IPV6)
3902                 else if (addr->sa_family == AF_INET6) {
3903                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3904
3905                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3906                 }
3907 #endif
3908         }
3909         daddr = cma_dst_addr(id_priv);
3910         daddr->sa_family = addr->sa_family;
3911
3912         ret = cma_get_port(id_priv);
3913         if (ret)
3914                 goto err2;
3915
3916         if (!cma_any_addr(addr))
3917                 rdma_restrack_add(&id_priv->res);
3918         return 0;
3919 err2:
3920         if (id_priv->cma_dev)
3921                 cma_release_dev(id_priv);
3922 err1:
3923         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3924         return ret;
3925 }
3926 EXPORT_SYMBOL(rdma_bind_addr);
3927
3928 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3929 {
3930         struct cma_hdr *cma_hdr;
3931
3932         cma_hdr = hdr;
3933         cma_hdr->cma_version = CMA_VERSION;
3934         if (cma_family(id_priv) == AF_INET) {
3935                 struct sockaddr_in *src4, *dst4;
3936
3937                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3938                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3939
3940                 cma_set_ip_ver(cma_hdr, 4);
3941                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3942                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3943                 cma_hdr->port = src4->sin_port;
3944         } else if (cma_family(id_priv) == AF_INET6) {
3945                 struct sockaddr_in6 *src6, *dst6;
3946
3947                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3948                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3949
3950                 cma_set_ip_ver(cma_hdr, 6);
3951                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3952                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3953                 cma_hdr->port = src6->sin6_port;
3954         }
3955         return 0;
3956 }
3957
3958 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3959                                 const struct ib_cm_event *ib_event)
3960 {
3961         struct rdma_id_private *id_priv = cm_id->context;
3962         struct rdma_cm_event event = {};
3963         const struct ib_cm_sidr_rep_event_param *rep =
3964                                 &ib_event->param.sidr_rep_rcvd;
3965         int ret;
3966
3967         mutex_lock(&id_priv->handler_mutex);
3968         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3969                 goto out;
3970
3971         switch (ib_event->event) {
3972         case IB_CM_SIDR_REQ_ERROR:
3973                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3974                 event.status = -ETIMEDOUT;
3975                 break;
3976         case IB_CM_SIDR_REP_RECEIVED:
3977                 event.param.ud.private_data = ib_event->private_data;
3978                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3979                 if (rep->status != IB_SIDR_SUCCESS) {
3980                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3981                         event.status = ib_event->param.sidr_rep_rcvd.status;
3982                         pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3983                                              event.status);
3984                         break;
3985                 }
3986                 ret = cma_set_qkey(id_priv, rep->qkey);
3987                 if (ret) {
3988                         pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3989                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3990                         event.status = ret;
3991                         break;
3992                 }
3993                 ib_init_ah_attr_from_path(id_priv->id.device,
3994                                           id_priv->id.port_num,
3995                                           id_priv->id.route.path_rec,
3996                                           &event.param.ud.ah_attr,
3997                                           rep->sgid_attr);
3998                 event.param.ud.qp_num = rep->qpn;
3999                 event.param.ud.qkey = rep->qkey;
4000                 event.event = RDMA_CM_EVENT_ESTABLISHED;
4001                 event.status = 0;
4002                 break;
4003         default:
4004                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4005                        ib_event->event);
4006                 goto out;
4007         }
4008
4009         ret = cma_cm_event_handler(id_priv, &event);
4010
4011         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4012         if (ret) {
4013                 /* Destroy the CM ID by returning a non-zero value. */
4014                 id_priv->cm_id.ib = NULL;
4015                 destroy_id_handler_unlock(id_priv);
4016                 return ret;
4017         }
4018 out:
4019         mutex_unlock(&id_priv->handler_mutex);
4020         return 0;
4021 }
4022
4023 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4024                               struct rdma_conn_param *conn_param)
4025 {
4026         struct ib_cm_sidr_req_param req;
4027         struct ib_cm_id *id;
4028         void *private_data;
4029         u8 offset;
4030         int ret;
4031
4032         memset(&req, 0, sizeof req);
4033         offset = cma_user_data_offset(id_priv);
4034         req.private_data_len = offset + conn_param->private_data_len;
4035         if (req.private_data_len < conn_param->private_data_len)
4036                 return -EINVAL;
4037
4038         if (req.private_data_len) {
4039                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4040                 if (!private_data)
4041                         return -ENOMEM;
4042         } else {
4043                 private_data = NULL;
4044         }
4045
4046         if (conn_param->private_data && conn_param->private_data_len)
4047                 memcpy(private_data + offset, conn_param->private_data,
4048                        conn_param->private_data_len);
4049
4050         if (private_data) {
4051                 ret = cma_format_hdr(private_data, id_priv);
4052                 if (ret)
4053                         goto out;
4054                 req.private_data = private_data;
4055         }
4056
4057         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4058                              id_priv);
4059         if (IS_ERR(id)) {
4060                 ret = PTR_ERR(id);
4061                 goto out;
4062         }
4063         id_priv->cm_id.ib = id;
4064
4065         req.path = id_priv->id.route.path_rec;
4066         req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4067         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4068         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4069         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4070
4071         trace_cm_send_sidr_req(id_priv);
4072         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4073         if (ret) {
4074                 ib_destroy_cm_id(id_priv->cm_id.ib);
4075                 id_priv->cm_id.ib = NULL;
4076         }
4077 out:
4078         kfree(private_data);
4079         return ret;
4080 }
4081
4082 static int cma_connect_ib(struct rdma_id_private *id_priv,
4083                           struct rdma_conn_param *conn_param)
4084 {
4085         struct ib_cm_req_param req;
4086         struct rdma_route *route;
4087         void *private_data;
4088         struct ib_cm_id *id;
4089         u8 offset;
4090         int ret;
4091
4092         memset(&req, 0, sizeof req);
4093         offset = cma_user_data_offset(id_priv);
4094         req.private_data_len = offset + conn_param->private_data_len;
4095         if (req.private_data_len < conn_param->private_data_len)
4096                 return -EINVAL;
4097
4098         if (req.private_data_len) {
4099                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4100                 if (!private_data)
4101                         return -ENOMEM;
4102         } else {
4103                 private_data = NULL;
4104         }
4105
4106         if (conn_param->private_data && conn_param->private_data_len)
4107                 memcpy(private_data + offset, conn_param->private_data,
4108                        conn_param->private_data_len);
4109
4110         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4111         if (IS_ERR(id)) {
4112                 ret = PTR_ERR(id);
4113                 goto out;
4114         }
4115         id_priv->cm_id.ib = id;
4116
4117         route = &id_priv->id.route;
4118         if (private_data) {
4119                 ret = cma_format_hdr(private_data, id_priv);
4120                 if (ret)
4121                         goto out;
4122                 req.private_data = private_data;
4123         }
4124
4125         req.primary_path = &route->path_rec[0];
4126         if (route->num_paths == 2)
4127                 req.alternate_path = &route->path_rec[1];
4128
4129         req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4130         /* Alternate path SGID attribute currently unsupported */
4131         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4132         req.qp_num = id_priv->qp_num;
4133         req.qp_type = id_priv->id.qp_type;
4134         req.starting_psn = id_priv->seq_num;
4135         req.responder_resources = conn_param->responder_resources;
4136         req.initiator_depth = conn_param->initiator_depth;
4137         req.flow_control = conn_param->flow_control;
4138         req.retry_count = min_t(u8, 7, conn_param->retry_count);
4139         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4140         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4141         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4142         req.max_cm_retries = CMA_MAX_CM_RETRIES;
4143         req.srq = id_priv->srq ? 1 : 0;
4144         req.ece.vendor_id = id_priv->ece.vendor_id;
4145         req.ece.attr_mod = id_priv->ece.attr_mod;
4146
4147         trace_cm_send_req(id_priv);
4148         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4149 out:
4150         if (ret && !IS_ERR(id)) {
4151                 ib_destroy_cm_id(id);
4152                 id_priv->cm_id.ib = NULL;
4153         }
4154
4155         kfree(private_data);
4156         return ret;
4157 }
4158
4159 static int cma_connect_iw(struct rdma_id_private *id_priv,
4160                           struct rdma_conn_param *conn_param)
4161 {
4162         struct iw_cm_id *cm_id;
4163         int ret;
4164         struct iw_cm_conn_param iw_param;
4165
4166         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4167         if (IS_ERR(cm_id))
4168                 return PTR_ERR(cm_id);
4169
4170         mutex_lock(&id_priv->qp_mutex);
4171         cm_id->tos = id_priv->tos;
4172         cm_id->tos_set = id_priv->tos_set;
4173         mutex_unlock(&id_priv->qp_mutex);
4174
4175         id_priv->cm_id.iw = cm_id;
4176
4177         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4178                rdma_addr_size(cma_src_addr(id_priv)));
4179         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4180                rdma_addr_size(cma_dst_addr(id_priv)));
4181
4182         ret = cma_modify_qp_rtr(id_priv, conn_param);
4183         if (ret)
4184                 goto out;
4185
4186         if (conn_param) {
4187                 iw_param.ord = conn_param->initiator_depth;
4188                 iw_param.ird = conn_param->responder_resources;
4189                 iw_param.private_data = conn_param->private_data;
4190                 iw_param.private_data_len = conn_param->private_data_len;
4191                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4192         } else {
4193                 memset(&iw_param, 0, sizeof iw_param);
4194                 iw_param.qpn = id_priv->qp_num;
4195         }
4196         ret = iw_cm_connect(cm_id, &iw_param);
4197 out:
4198         if (ret) {
4199                 iw_destroy_cm_id(cm_id);
4200                 id_priv->cm_id.iw = NULL;
4201         }
4202         return ret;
4203 }
4204
4205 /**
4206  * rdma_connect_locked - Initiate an active connection request.
4207  * @id: Connection identifier to connect.
4208  * @conn_param: Connection information used for connected QPs.
4209  *
4210  * Same as rdma_connect() but can only be called from the
4211  * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4212  */
4213 int rdma_connect_locked(struct rdma_cm_id *id,
4214                         struct rdma_conn_param *conn_param)
4215 {
4216         struct rdma_id_private *id_priv =
4217                 container_of(id, struct rdma_id_private, id);
4218         int ret;
4219
4220         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4221                 return -EINVAL;
4222
4223         if (!id->qp) {
4224                 id_priv->qp_num = conn_param->qp_num;
4225                 id_priv->srq = conn_param->srq;
4226         }
4227
4228         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4229                 if (id->qp_type == IB_QPT_UD)
4230                         ret = cma_resolve_ib_udp(id_priv, conn_param);
4231                 else
4232                         ret = cma_connect_ib(id_priv, conn_param);
4233         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4234                 ret = cma_connect_iw(id_priv, conn_param);
4235         } else {
4236                 ret = -ENOSYS;
4237         }
4238         if (ret)
4239                 goto err_state;
4240         return 0;
4241 err_state:
4242         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4243         return ret;
4244 }
4245 EXPORT_SYMBOL(rdma_connect_locked);
4246
4247 /**
4248  * rdma_connect - Initiate an active connection request.
4249  * @id: Connection identifier to connect.
4250  * @conn_param: Connection information used for connected QPs.
4251  *
4252  * Users must have resolved a route for the rdma_cm_id to connect with by having
4253  * called rdma_resolve_route before calling this routine.
4254  *
4255  * This call will either connect to a remote QP or obtain remote QP information
4256  * for unconnected rdma_cm_id's.  The actual operation is based on the
4257  * rdma_cm_id's port space.
4258  */
4259 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4260 {
4261         struct rdma_id_private *id_priv =
4262                 container_of(id, struct rdma_id_private, id);
4263         int ret;
4264
4265         mutex_lock(&id_priv->handler_mutex);
4266         ret = rdma_connect_locked(id, conn_param);
4267         mutex_unlock(&id_priv->handler_mutex);
4268         return ret;
4269 }
4270 EXPORT_SYMBOL(rdma_connect);
4271
4272 /**
4273  * rdma_connect_ece - Initiate an active connection request with ECE data.
4274  * @id: Connection identifier to connect.
4275  * @conn_param: Connection information used for connected QPs.
4276  * @ece: ECE parameters
4277  *
4278  * See rdma_connect() explanation.
4279  */
4280 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4281                      struct rdma_ucm_ece *ece)
4282 {
4283         struct rdma_id_private *id_priv =
4284                 container_of(id, struct rdma_id_private, id);
4285
4286         id_priv->ece.vendor_id = ece->vendor_id;
4287         id_priv->ece.attr_mod = ece->attr_mod;
4288
4289         return rdma_connect(id, conn_param);
4290 }
4291 EXPORT_SYMBOL(rdma_connect_ece);
4292
4293 static int cma_accept_ib(struct rdma_id_private *id_priv,
4294                          struct rdma_conn_param *conn_param)
4295 {
4296         struct ib_cm_rep_param rep;
4297         int ret;
4298
4299         ret = cma_modify_qp_rtr(id_priv, conn_param);
4300         if (ret)
4301                 goto out;
4302
4303         ret = cma_modify_qp_rts(id_priv, conn_param);
4304         if (ret)
4305                 goto out;
4306
4307         memset(&rep, 0, sizeof rep);
4308         rep.qp_num = id_priv->qp_num;
4309         rep.starting_psn = id_priv->seq_num;
4310         rep.private_data = conn_param->private_data;
4311         rep.private_data_len = conn_param->private_data_len;
4312         rep.responder_resources = conn_param->responder_resources;
4313         rep.initiator_depth = conn_param->initiator_depth;
4314         rep.failover_accepted = 0;
4315         rep.flow_control = conn_param->flow_control;
4316         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4317         rep.srq = id_priv->srq ? 1 : 0;
4318         rep.ece.vendor_id = id_priv->ece.vendor_id;
4319         rep.ece.attr_mod = id_priv->ece.attr_mod;
4320
4321         trace_cm_send_rep(id_priv);
4322         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4323 out:
4324         return ret;
4325 }
4326
4327 static int cma_accept_iw(struct rdma_id_private *id_priv,
4328                   struct rdma_conn_param *conn_param)
4329 {
4330         struct iw_cm_conn_param iw_param;
4331         int ret;
4332
4333         if (!conn_param)
4334                 return -EINVAL;
4335
4336         ret = cma_modify_qp_rtr(id_priv, conn_param);
4337         if (ret)
4338                 return ret;
4339
4340         iw_param.ord = conn_param->initiator_depth;
4341         iw_param.ird = conn_param->responder_resources;
4342         iw_param.private_data = conn_param->private_data;
4343         iw_param.private_data_len = conn_param->private_data_len;
4344         if (id_priv->id.qp)
4345                 iw_param.qpn = id_priv->qp_num;
4346         else
4347                 iw_param.qpn = conn_param->qp_num;
4348
4349         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4350 }
4351
4352 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4353                              enum ib_cm_sidr_status status, u32 qkey,
4354                              const void *private_data, int private_data_len)
4355 {
4356         struct ib_cm_sidr_rep_param rep;
4357         int ret;
4358
4359         memset(&rep, 0, sizeof rep);
4360         rep.status = status;
4361         if (status == IB_SIDR_SUCCESS) {
4362                 ret = cma_set_qkey(id_priv, qkey);
4363                 if (ret)
4364                         return ret;
4365                 rep.qp_num = id_priv->qp_num;
4366                 rep.qkey = id_priv->qkey;
4367
4368                 rep.ece.vendor_id = id_priv->ece.vendor_id;
4369                 rep.ece.attr_mod = id_priv->ece.attr_mod;
4370         }
4371
4372         rep.private_data = private_data;
4373         rep.private_data_len = private_data_len;
4374
4375         trace_cm_send_sidr_rep(id_priv);
4376         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4377 }
4378
4379 /**
4380  * rdma_accept - Called to accept a connection request or response.
4381  * @id: Connection identifier associated with the request.
4382  * @conn_param: Information needed to establish the connection.  This must be
4383  *   provided if accepting a connection request.  If accepting a connection
4384  *   response, this parameter must be NULL.
4385  *
4386  * Typically, this routine is only called by the listener to accept a connection
4387  * request.  It must also be called on the active side of a connection if the
4388  * user is performing their own QP transitions.
4389  *
4390  * In the case of error, a reject message is sent to the remote side and the
4391  * state of the qp associated with the id is modified to error, such that any
4392  * previously posted receive buffers would be flushed.
4393  *
4394  * This function is for use by kernel ULPs and must be called from under the
4395  * handler callback.
4396  */
4397 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4398 {
4399         struct rdma_id_private *id_priv =
4400                 container_of(id, struct rdma_id_private, id);
4401         int ret;
4402
4403         lockdep_assert_held(&id_priv->handler_mutex);
4404
4405         if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4406                 return -EINVAL;
4407
4408         if (!id->qp && conn_param) {
4409                 id_priv->qp_num = conn_param->qp_num;
4410                 id_priv->srq = conn_param->srq;
4411         }
4412
4413         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4414                 if (id->qp_type == IB_QPT_UD) {
4415                         if (conn_param)
4416                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4417                                                         conn_param->qkey,
4418                                                         conn_param->private_data,
4419                                                         conn_param->private_data_len);
4420                         else
4421                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4422                                                         0, NULL, 0);
4423                 } else {
4424                         if (conn_param)
4425                                 ret = cma_accept_ib(id_priv, conn_param);
4426                         else
4427                                 ret = cma_rep_recv(id_priv);
4428                 }
4429         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4430                 ret = cma_accept_iw(id_priv, conn_param);
4431         } else {
4432                 ret = -ENOSYS;
4433         }
4434         if (ret)
4435                 goto reject;
4436
4437         return 0;
4438 reject:
4439         cma_modify_qp_err(id_priv);
4440         rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4441         return ret;
4442 }
4443 EXPORT_SYMBOL(rdma_accept);
4444
4445 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4446                     struct rdma_ucm_ece *ece)
4447 {
4448         struct rdma_id_private *id_priv =
4449                 container_of(id, struct rdma_id_private, id);
4450
4451         id_priv->ece.vendor_id = ece->vendor_id;
4452         id_priv->ece.attr_mod = ece->attr_mod;
4453
4454         return rdma_accept(id, conn_param);
4455 }
4456 EXPORT_SYMBOL(rdma_accept_ece);
4457
4458 void rdma_lock_handler(struct rdma_cm_id *id)
4459 {
4460         struct rdma_id_private *id_priv =
4461                 container_of(id, struct rdma_id_private, id);
4462
4463         mutex_lock(&id_priv->handler_mutex);
4464 }
4465 EXPORT_SYMBOL(rdma_lock_handler);
4466
4467 void rdma_unlock_handler(struct rdma_cm_id *id)
4468 {
4469         struct rdma_id_private *id_priv =
4470                 container_of(id, struct rdma_id_private, id);
4471
4472         mutex_unlock(&id_priv->handler_mutex);
4473 }
4474 EXPORT_SYMBOL(rdma_unlock_handler);
4475
4476 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4477 {
4478         struct rdma_id_private *id_priv;
4479         int ret;
4480
4481         id_priv = container_of(id, struct rdma_id_private, id);
4482         if (!id_priv->cm_id.ib)
4483                 return -EINVAL;
4484
4485         switch (id->device->node_type) {
4486         case RDMA_NODE_IB_CA:
4487                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4488                 break;
4489         default:
4490                 ret = 0;
4491                 break;
4492         }
4493         return ret;
4494 }
4495 EXPORT_SYMBOL(rdma_notify);
4496
4497 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4498                 u8 private_data_len, u8 reason)
4499 {
4500         struct rdma_id_private *id_priv;
4501         int ret;
4502
4503         id_priv = container_of(id, struct rdma_id_private, id);
4504         if (!id_priv->cm_id.ib)
4505                 return -EINVAL;
4506
4507         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4508                 if (id->qp_type == IB_QPT_UD) {
4509                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4510                                                 private_data, private_data_len);
4511                 } else {
4512                         trace_cm_send_rej(id_priv);
4513                         ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4514                                              private_data, private_data_len);
4515                 }
4516         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4517                 ret = iw_cm_reject(id_priv->cm_id.iw,
4518                                    private_data, private_data_len);
4519         } else {
4520                 ret = -ENOSYS;
4521         }
4522
4523         return ret;
4524 }
4525 EXPORT_SYMBOL(rdma_reject);
4526
4527 int rdma_disconnect(struct rdma_cm_id *id)
4528 {
4529         struct rdma_id_private *id_priv;
4530         int ret;
4531
4532         id_priv = container_of(id, struct rdma_id_private, id);
4533         if (!id_priv->cm_id.ib)
4534                 return -EINVAL;
4535
4536         if (rdma_cap_ib_cm(id->device, id->port_num)) {
4537                 ret = cma_modify_qp_err(id_priv);
4538                 if (ret)
4539                         goto out;
4540                 /* Initiate or respond to a disconnect. */
4541                 trace_cm_disconnect(id_priv);
4542                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4543                         if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4544                                 trace_cm_sent_drep(id_priv);
4545                 } else {
4546                         trace_cm_sent_dreq(id_priv);
4547                 }
4548         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4549                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4550         } else
4551                 ret = -EINVAL;
4552
4553 out:
4554         return ret;
4555 }
4556 EXPORT_SYMBOL(rdma_disconnect);
4557
4558 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4559                               struct ib_sa_multicast *multicast,
4560                               struct rdma_cm_event *event,
4561                               struct cma_multicast *mc)
4562 {
4563         struct rdma_dev_addr *dev_addr;
4564         enum ib_gid_type gid_type;
4565         struct net_device *ndev;
4566
4567         if (!status)
4568                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4569         else
4570                 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4571                                      status);
4572
4573         event->status = status;
4574         event->param.ud.private_data = mc->context;
4575         if (status) {
4576                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4577                 return;
4578         }
4579
4580         dev_addr = &id_priv->id.route.addr.dev_addr;
4581         ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4582         gid_type =
4583                 id_priv->cma_dev
4584                         ->default_gid_type[id_priv->id.port_num -
4585                                            rdma_start_port(
4586                                                    id_priv->cma_dev->device)];
4587
4588         event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4589         if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4590                                      &multicast->rec, ndev, gid_type,
4591                                      &event->param.ud.ah_attr)) {
4592                 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4593                 goto out;
4594         }
4595
4596         event->param.ud.qp_num = 0xFFFFFF;
4597         event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4598
4599 out:
4600         if (ndev)
4601                 dev_put(ndev);
4602 }
4603
4604 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4605 {
4606         struct cma_multicast *mc = multicast->context;
4607         struct rdma_id_private *id_priv = mc->id_priv;
4608         struct rdma_cm_event event = {};
4609         int ret = 0;
4610
4611         mutex_lock(&id_priv->handler_mutex);
4612         if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4613             READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4614                 goto out;
4615
4616         cma_make_mc_event(status, id_priv, multicast, &event, mc);
4617         ret = cma_cm_event_handler(id_priv, &event);
4618         rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4619         WARN_ON(ret);
4620
4621 out:
4622         mutex_unlock(&id_priv->handler_mutex);
4623         return 0;
4624 }
4625
4626 static void cma_set_mgid(struct rdma_id_private *id_priv,
4627                          struct sockaddr *addr, union ib_gid *mgid)
4628 {
4629         unsigned char mc_map[MAX_ADDR_LEN];
4630         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4631         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4632         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4633
4634         if (cma_any_addr(addr)) {
4635                 memset(mgid, 0, sizeof *mgid);
4636         } else if ((addr->sa_family == AF_INET6) &&
4637                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4638                                                                  0xFF10A01B)) {
4639                 /* IPv6 address is an SA assigned MGID. */
4640                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4641         } else if (addr->sa_family == AF_IB) {
4642                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4643         } else if (addr->sa_family == AF_INET6) {
4644                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4645                 if (id_priv->id.ps == RDMA_PS_UDP)
4646                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4647                 *mgid = *(union ib_gid *) (mc_map + 4);
4648         } else {
4649                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4650                 if (id_priv->id.ps == RDMA_PS_UDP)
4651                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
4652                 *mgid = *(union ib_gid *) (mc_map + 4);
4653         }
4654 }
4655
4656 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4657                                  struct cma_multicast *mc)
4658 {
4659         struct ib_sa_mcmember_rec rec;
4660         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4661         ib_sa_comp_mask comp_mask;
4662         int ret;
4663
4664         ib_addr_get_mgid(dev_addr, &rec.mgid);
4665         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4666                                      &rec.mgid, &rec);
4667         if (ret)
4668                 return ret;
4669
4670         ret = cma_set_qkey(id_priv, 0);
4671         if (ret)
4672                 return ret;
4673
4674         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4675         rec.qkey = cpu_to_be32(id_priv->qkey);
4676         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4677         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4678         rec.join_state = mc->join_state;
4679
4680         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4681                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4682                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4683                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
4684                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4685
4686         if (id_priv->id.ps == RDMA_PS_IPOIB)
4687                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4688                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4689                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4690                              IB_SA_MCMEMBER_REC_MTU |
4691                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
4692
4693         mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4694                                          id_priv->id.port_num, &rec, comp_mask,
4695                                          GFP_KERNEL, cma_ib_mc_handler, mc);
4696         return PTR_ERR_OR_ZERO(mc->sa_mc);
4697 }
4698
4699 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4700                               enum ib_gid_type gid_type)
4701 {
4702         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4703         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4704
4705         if (cma_any_addr(addr)) {
4706                 memset(mgid, 0, sizeof *mgid);
4707         } else if (addr->sa_family == AF_INET6) {
4708                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4709         } else {
4710                 mgid->raw[0] =
4711                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4712                 mgid->raw[1] =
4713                         (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4714                 mgid->raw[2] = 0;
4715                 mgid->raw[3] = 0;
4716                 mgid->raw[4] = 0;
4717                 mgid->raw[5] = 0;
4718                 mgid->raw[6] = 0;
4719                 mgid->raw[7] = 0;
4720                 mgid->raw[8] = 0;
4721                 mgid->raw[9] = 0;
4722                 mgid->raw[10] = 0xff;
4723                 mgid->raw[11] = 0xff;
4724                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4725         }
4726 }
4727
4728 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4729                                    struct cma_multicast *mc)
4730 {
4731         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4732         int err = 0;
4733         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4734         struct net_device *ndev = NULL;
4735         struct ib_sa_multicast ib;
4736         enum ib_gid_type gid_type;
4737         bool send_only;
4738
4739         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4740
4741         if (cma_zero_addr(addr))
4742                 return -EINVAL;
4743
4744         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4745                    rdma_start_port(id_priv->cma_dev->device)];
4746         cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4747
4748         ib.rec.pkey = cpu_to_be16(0xffff);
4749         if (id_priv->id.ps == RDMA_PS_UDP)
4750                 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4751
4752         if (dev_addr->bound_dev_if)
4753                 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4754         if (!ndev)
4755                 return -ENODEV;
4756
4757         ib.rec.rate = iboe_get_rate(ndev);
4758         ib.rec.hop_limit = 1;
4759         ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4760
4761         if (addr->sa_family == AF_INET) {
4762                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4763                         ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4764                         if (!send_only) {
4765                                 err = cma_igmp_send(ndev, &ib.rec.mgid,
4766                                                     true);
4767                         }
4768                 }
4769         } else {
4770                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4771                         err = -ENOTSUPP;
4772         }
4773         dev_put(ndev);
4774         if (err || !ib.rec.mtu)
4775                 return err ?: -EINVAL;
4776
4777         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4778                     &ib.rec.port_gid);
4779         INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4780         cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4781         queue_work(cma_wq, &mc->iboe_join.work);
4782         return 0;
4783 }
4784
4785 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4786                         u8 join_state, void *context)
4787 {
4788         struct rdma_id_private *id_priv =
4789                 container_of(id, struct rdma_id_private, id);
4790         struct cma_multicast *mc;
4791         int ret;
4792
4793         /* Not supported for kernel QPs */
4794         if (WARN_ON(id->qp))
4795                 return -EINVAL;
4796
4797         /* ULP is calling this wrong. */
4798         if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4799                             READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4800                 return -EINVAL;
4801
4802         mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4803         if (!mc)
4804                 return -ENOMEM;
4805
4806         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4807         mc->context = context;
4808         mc->id_priv = id_priv;
4809         mc->join_state = join_state;
4810
4811         if (rdma_protocol_roce(id->device, id->port_num)) {
4812                 ret = cma_iboe_join_multicast(id_priv, mc);
4813                 if (ret)
4814                         goto out_err;
4815         } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4816                 ret = cma_join_ib_multicast(id_priv, mc);
4817                 if (ret)
4818                         goto out_err;
4819         } else {
4820                 ret = -ENOSYS;
4821                 goto out_err;
4822         }
4823
4824         spin_lock(&id_priv->lock);
4825         list_add(&mc->list, &id_priv->mc_list);
4826         spin_unlock(&id_priv->lock);
4827
4828         return 0;
4829 out_err:
4830         kfree(mc);
4831         return ret;
4832 }
4833 EXPORT_SYMBOL(rdma_join_multicast);
4834
4835 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4836 {
4837         struct rdma_id_private *id_priv;
4838         struct cma_multicast *mc;
4839
4840         id_priv = container_of(id, struct rdma_id_private, id);
4841         spin_lock_irq(&id_priv->lock);
4842         list_for_each_entry(mc, &id_priv->mc_list, list) {
4843                 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4844                         continue;
4845                 list_del(&mc->list);
4846                 spin_unlock_irq(&id_priv->lock);
4847
4848                 WARN_ON(id_priv->cma_dev->device != id->device);
4849                 destroy_mc(id_priv, mc);
4850                 return;
4851         }
4852         spin_unlock_irq(&id_priv->lock);
4853 }
4854 EXPORT_SYMBOL(rdma_leave_multicast);
4855
4856 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4857 {
4858         struct rdma_dev_addr *dev_addr;
4859         struct cma_work *work;
4860
4861         dev_addr = &id_priv->id.route.addr.dev_addr;
4862
4863         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4864             (net_eq(dev_net(ndev), dev_addr->net)) &&
4865             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4866                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4867                         ndev->name, &id_priv->id);
4868                 work = kzalloc(sizeof *work, GFP_KERNEL);
4869                 if (!work)
4870                         return -ENOMEM;
4871
4872                 INIT_WORK(&work->work, cma_work_handler);
4873                 work->id = id_priv;
4874                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4875                 cma_id_get(id_priv);
4876                 queue_work(cma_wq, &work->work);
4877         }
4878
4879         return 0;
4880 }
4881
4882 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4883                                void *ptr)
4884 {
4885         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4886         struct cma_device *cma_dev;
4887         struct rdma_id_private *id_priv;
4888         int ret = NOTIFY_DONE;
4889
4890         if (event != NETDEV_BONDING_FAILOVER)
4891                 return NOTIFY_DONE;
4892
4893         if (!netif_is_bond_master(ndev))
4894                 return NOTIFY_DONE;
4895
4896         mutex_lock(&lock);
4897         list_for_each_entry(cma_dev, &dev_list, list)
4898                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4899                         ret = cma_netdev_change(ndev, id_priv);
4900                         if (ret)
4901                                 goto out;
4902                 }
4903
4904 out:
4905         mutex_unlock(&lock);
4906         return ret;
4907 }
4908
4909 static struct notifier_block cma_nb = {
4910         .notifier_call = cma_netdev_callback
4911 };
4912
4913 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4914 {
4915         struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4916         enum rdma_cm_state state;
4917         unsigned long flags;
4918
4919         mutex_lock(&id_priv->handler_mutex);
4920         /* Record that we want to remove the device */
4921         spin_lock_irqsave(&id_priv->lock, flags);
4922         state = id_priv->state;
4923         if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4924                 spin_unlock_irqrestore(&id_priv->lock, flags);
4925                 mutex_unlock(&id_priv->handler_mutex);
4926                 cma_id_put(id_priv);
4927                 return;
4928         }
4929         id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4930         spin_unlock_irqrestore(&id_priv->lock, flags);
4931
4932         if (cma_cm_event_handler(id_priv, &event)) {
4933                 /*
4934                  * At this point the ULP promises it won't call
4935                  * rdma_destroy_id() concurrently
4936                  */
4937                 cma_id_put(id_priv);
4938                 mutex_unlock(&id_priv->handler_mutex);
4939                 trace_cm_id_destroy(id_priv);
4940                 _destroy_id(id_priv, state);
4941                 return;
4942         }
4943         mutex_unlock(&id_priv->handler_mutex);
4944
4945         /*
4946          * If this races with destroy then the thread that first assigns state
4947          * to a destroying does the cancel.
4948          */
4949         cma_cancel_operation(id_priv, state);
4950         cma_id_put(id_priv);
4951 }
4952
4953 static void cma_process_remove(struct cma_device *cma_dev)
4954 {
4955         mutex_lock(&lock);
4956         while (!list_empty(&cma_dev->id_list)) {
4957                 struct rdma_id_private *id_priv = list_first_entry(
4958                         &cma_dev->id_list, struct rdma_id_private, list);
4959
4960                 list_del(&id_priv->listen_list);
4961                 list_del_init(&id_priv->list);
4962                 cma_id_get(id_priv);
4963                 mutex_unlock(&lock);
4964
4965                 cma_send_device_removal_put(id_priv);
4966
4967                 mutex_lock(&lock);
4968         }
4969         mutex_unlock(&lock);
4970
4971         cma_dev_put(cma_dev);
4972         wait_for_completion(&cma_dev->comp);
4973 }
4974
4975 static bool cma_supported(struct ib_device *device)
4976 {
4977         u32 i;
4978
4979         rdma_for_each_port(device, i) {
4980                 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
4981                         return true;
4982         }
4983         return false;
4984 }
4985
4986 static int cma_add_one(struct ib_device *device)
4987 {
4988         struct rdma_id_private *to_destroy;
4989         struct cma_device *cma_dev;
4990         struct rdma_id_private *id_priv;
4991         unsigned long supported_gids = 0;
4992         int ret;
4993         u32 i;
4994
4995         if (!cma_supported(device))
4996                 return -EOPNOTSUPP;
4997
4998         cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4999         if (!cma_dev)
5000                 return -ENOMEM;
5001
5002         cma_dev->device = device;
5003         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5004                                             sizeof(*cma_dev->default_gid_type),
5005                                             GFP_KERNEL);
5006         if (!cma_dev->default_gid_type) {
5007                 ret = -ENOMEM;
5008                 goto free_cma_dev;
5009         }
5010
5011         cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5012                                             sizeof(*cma_dev->default_roce_tos),
5013                                             GFP_KERNEL);
5014         if (!cma_dev->default_roce_tos) {
5015                 ret = -ENOMEM;
5016                 goto free_gid_type;
5017         }
5018
5019         rdma_for_each_port (device, i) {
5020                 supported_gids = roce_gid_type_mask_support(device, i);
5021                 WARN_ON(!supported_gids);
5022                 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5023                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5024                                 CMA_PREFERRED_ROCE_GID_TYPE;
5025                 else
5026                         cma_dev->default_gid_type[i - rdma_start_port(device)] =
5027                                 find_first_bit(&supported_gids, BITS_PER_LONG);
5028                 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5029         }
5030
5031         init_completion(&cma_dev->comp);
5032         refcount_set(&cma_dev->refcount, 1);
5033         INIT_LIST_HEAD(&cma_dev->id_list);
5034         ib_set_client_data(device, &cma_client, cma_dev);
5035
5036         mutex_lock(&lock);
5037         list_add_tail(&cma_dev->list, &dev_list);
5038         list_for_each_entry(id_priv, &listen_any_list, list) {
5039                 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5040                 if (ret)
5041                         goto free_listen;
5042         }
5043         mutex_unlock(&lock);
5044
5045         trace_cm_add_one(device);
5046         return 0;
5047
5048 free_listen:
5049         list_del(&cma_dev->list);
5050         mutex_unlock(&lock);
5051
5052         /* cma_process_remove() will delete to_destroy */
5053         cma_process_remove(cma_dev);
5054         kfree(cma_dev->default_roce_tos);
5055 free_gid_type:
5056         kfree(cma_dev->default_gid_type);
5057
5058 free_cma_dev:
5059         kfree(cma_dev);
5060         return ret;
5061 }
5062
5063 static void cma_remove_one(struct ib_device *device, void *client_data)
5064 {
5065         struct cma_device *cma_dev = client_data;
5066
5067         trace_cm_remove_one(device);
5068
5069         mutex_lock(&lock);
5070         list_del(&cma_dev->list);
5071         mutex_unlock(&lock);
5072
5073         cma_process_remove(cma_dev);
5074         kfree(cma_dev->default_roce_tos);
5075         kfree(cma_dev->default_gid_type);
5076         kfree(cma_dev);
5077 }
5078
5079 static int cma_init_net(struct net *net)
5080 {
5081         struct cma_pernet *pernet = cma_pernet(net);
5082
5083         xa_init(&pernet->tcp_ps);
5084         xa_init(&pernet->udp_ps);
5085         xa_init(&pernet->ipoib_ps);
5086         xa_init(&pernet->ib_ps);
5087
5088         return 0;
5089 }
5090
5091 static void cma_exit_net(struct net *net)
5092 {
5093         struct cma_pernet *pernet = cma_pernet(net);
5094
5095         WARN_ON(!xa_empty(&pernet->tcp_ps));
5096         WARN_ON(!xa_empty(&pernet->udp_ps));
5097         WARN_ON(!xa_empty(&pernet->ipoib_ps));
5098         WARN_ON(!xa_empty(&pernet->ib_ps));
5099 }
5100
5101 static struct pernet_operations cma_pernet_operations = {
5102         .init = cma_init_net,
5103         .exit = cma_exit_net,
5104         .id = &cma_pernet_id,
5105         .size = sizeof(struct cma_pernet),
5106 };
5107
5108 static int __init cma_init(void)
5109 {
5110         int ret;
5111
5112         /*
5113          * There is a rare lock ordering dependency in cma_netdev_callback()
5114          * that only happens when bonding is enabled. Teach lockdep that rtnl
5115          * must never be nested under lock so it can find these without having
5116          * to test with bonding.
5117          */
5118         if (IS_ENABLED(CONFIG_LOCKDEP)) {
5119                 rtnl_lock();
5120                 mutex_lock(&lock);
5121                 mutex_unlock(&lock);
5122                 rtnl_unlock();
5123         }
5124
5125         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5126         if (!cma_wq)
5127                 return -ENOMEM;
5128
5129         ret = register_pernet_subsys(&cma_pernet_operations);
5130         if (ret)
5131                 goto err_wq;
5132
5133         ib_sa_register_client(&sa_client);
5134         register_netdevice_notifier(&cma_nb);
5135
5136         ret = ib_register_client(&cma_client);
5137         if (ret)
5138                 goto err;
5139
5140         ret = cma_configfs_init();
5141         if (ret)
5142                 goto err_ib;
5143
5144         return 0;
5145
5146 err_ib:
5147         ib_unregister_client(&cma_client);
5148 err:
5149         unregister_netdevice_notifier(&cma_nb);
5150         ib_sa_unregister_client(&sa_client);
5151         unregister_pernet_subsys(&cma_pernet_operations);
5152 err_wq:
5153         destroy_workqueue(cma_wq);
5154         return ret;
5155 }
5156
5157 static void __exit cma_cleanup(void)
5158 {
5159         cma_configfs_exit();
5160         ib_unregister_client(&cma_client);
5161         unregister_netdevice_notifier(&cma_nb);
5162         ib_sa_unregister_client(&sa_client);
5163         unregister_pernet_subsys(&cma_pernet_operations);
5164         destroy_workqueue(cma_wq);
5165 }
5166
5167 module_init(cma_init);
5168 module_exit(cma_cleanup);