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