drivers/infiniband/core/cma.c

   1 /*
   2  * Copyright (c) 2005 Voltaire Inc.  All rights reserved.
   3  * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
   4  * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
   5  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/in.h>
  38 #include <linux/in6.h>
  39 #include <linux/mutex.h>
  40 #include <linux/random.h>
  41 #include <linux/igmp.h>
  42 #include <linux/idr.h>
  43 #include <linux/inetdevice.h>
  44 #include <linux/slab.h>
  45 #include <linux/module.h>
  46 #include <net/route.h>
  47
  48 #include <net/net_namespace.h>
  49 #include <net/netns/generic.h>
  50 #include <net/tcp.h>
  51 #include <net/ipv6.h>
  52 #include <net/ip_fib.h>
  53 #include <net/ip6_route.h>
  54
  55 #include <rdma/rdma_cm.h>
  56 #include <rdma/rdma_cm_ib.h>
  57 #include <rdma/rdma_netlink.h>
  58 #include <rdma/ib.h>
  59 #include <rdma/ib_cache.h>
  60 #include <rdma/ib_cm.h>
  61 #include <rdma/ib_sa.h>
  62 #include <rdma/iw_cm.h>
  63
  64 #include "core_priv.h"
  65
  66 MODULE_AUTHOR("Sean Hefty");
  67 MODULE_DESCRIPTION("Generic RDMA CM Agent");
  68 MODULE_LICENSE("Dual BSD/GPL");
  69
  70 #define CMA_CM_RESPONSE_TIMEOUT 20
  71 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
  72 #define CMA_MAX_CM_RETRIES 15
  73 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
  74 #define CMA_IBOE_PACKET_LIFETIME 18
  75
  76 static const char * const cma_events[] = {
  77         [RDMA_CM_EVENT_ADDR_RESOLVED]    = "address resolved",
  78         [RDMA_CM_EVENT_ADDR_ERROR]       = "address error",
  79         [RDMA_CM_EVENT_ROUTE_RESOLVED]   = "route resolved ",
  80         [RDMA_CM_EVENT_ROUTE_ERROR]      = "route error",
  81         [RDMA_CM_EVENT_CONNECT_REQUEST]  = "connect request",
  82         [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
  83         [RDMA_CM_EVENT_CONNECT_ERROR]    = "connect error",
  84         [RDMA_CM_EVENT_UNREACHABLE]      = "unreachable",
  85         [RDMA_CM_EVENT_REJECTED]         = "rejected",
  86         [RDMA_CM_EVENT_ESTABLISHED]      = "established",
  87         [RDMA_CM_EVENT_DISCONNECTED]     = "disconnected",
  88         [RDMA_CM_EVENT_DEVICE_REMOVAL]   = "device removal",
  89         [RDMA_CM_EVENT_MULTICAST_JOIN]   = "multicast join",
  90         [RDMA_CM_EVENT_MULTICAST_ERROR]  = "multicast error",
  91         [RDMA_CM_EVENT_ADDR_CHANGE]      = "address change",
  92         [RDMA_CM_EVENT_TIMEWAIT_EXIT]    = "timewait exit",
  93 };
  94
  95 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
  96 {
  97         size_t index = event;
  98
  99         return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
 100                         cma_events[index] : "unrecognized event";
 101 }
 102 EXPORT_SYMBOL(rdma_event_msg);
 103
 104 static void cma_add_one(struct ib_device *device);
 105 static void cma_remove_one(struct ib_device *device, void *client_data);
 106
 107 static struct ib_client cma_client = {
 108         .name   = "cma",
 109         .add    = cma_add_one,
 110         .remove = cma_remove_one
 111 };
 112
 113 static struct ib_sa_client sa_client;
 114 static struct rdma_addr_client addr_client;
 115 static LIST_HEAD(dev_list);
 116 static LIST_HEAD(listen_any_list);
 117 static DEFINE_MUTEX(lock);
 118 static struct workqueue_struct *cma_wq;
 119 static int cma_pernet_id;
 120
 121 struct cma_pernet {
 122         struct idr tcp_ps;
 123         struct idr udp_ps;
 124         struct idr ipoib_ps;
 125         struct idr ib_ps;
 126 };
 127
 128 static struct cma_pernet *cma_pernet(struct net *net)
 129 {
 130         return net_generic(net, cma_pernet_id);
 131 }
 132
 133 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
 134 {
 135         struct cma_pernet *pernet = cma_pernet(net);
 136
 137         switch (ps) {
 138         case RDMA_PS_TCP:
 139                 return &pernet->tcp_ps;
 140         case RDMA_PS_UDP:
 141                 return &pernet->udp_ps;
 142         case RDMA_PS_IPOIB:
 143                 return &pernet->ipoib_ps;
 144         case RDMA_PS_IB:
 145                 return &pernet->ib_ps;
 146         default:
 147                 return NULL;
 148         }
 149 }
 150
 151 struct cma_device {
 152         struct list_head        list;
 153         struct ib_device        *device;
 154         struct completion       comp;
 155         atomic_t                refcount;
 156         struct list_head        id_list;
 157         enum ib_gid_type        *default_gid_type;
 158 };
 159
 160 struct rdma_bind_list {
 161         enum rdma_port_space    ps;
 162         struct hlist_head       owners;
 163         unsigned short          port;
 164 };
 165
 166 struct class_port_info_context {
 167         struct ib_class_port_info       *class_port_info;
 168         struct ib_device                *device;
 169         struct completion               done;
 170         struct ib_sa_query              *sa_query;
 171         u8                              port_num;
 172 };
 173
 174 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
 175                         struct rdma_bind_list *bind_list, int snum)
 176 {
 177         struct idr *idr = cma_pernet_idr(net, ps);
 178
 179         return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
 180 }
 181
 182 static struct rdma_bind_list *cma_ps_find(struct net *net,
 183                                           enum rdma_port_space ps, int snum)
 184 {
 185         struct idr *idr = cma_pernet_idr(net, ps);
 186
 187         return idr_find(idr, snum);
 188 }
 189
 190 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
 191 {
 192         struct idr *idr = cma_pernet_idr(net, ps);
 193
 194         idr_remove(idr, snum);
 195 }
 196
 197 enum {
 198         CMA_OPTION_AFONLY,
 199 };
 200
 201 void cma_ref_dev(struct cma_device *cma_dev)
 202 {
 203         atomic_inc(&cma_dev->refcount);
 204 }
 205
 206 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter  filter,
 207                                              void               *cookie)
 208 {
 209         struct cma_device *cma_dev;
 210         struct cma_device *found_cma_dev = NULL;
 211
 212         mutex_lock(&lock);
 213
 214         list_for_each_entry(cma_dev, &dev_list, list)
 215                 if (filter(cma_dev->device, cookie)) {
 216                         found_cma_dev = cma_dev;
 217                         break;
 218                 }
 219
 220         if (found_cma_dev)
 221                 cma_ref_dev(found_cma_dev);
 222         mutex_unlock(&lock);
 223         return found_cma_dev;
 224 }
 225
 226 int cma_get_default_gid_type(struct cma_device *cma_dev,
 227                              unsigned int port)
 228 {
 229         if (port < rdma_start_port(cma_dev->device) ||
 230             port > rdma_end_port(cma_dev->device))
 231                 return -EINVAL;
 232
 233         return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
 234 }
 235
 236 int cma_set_default_gid_type(struct cma_device *cma_dev,
 237                              unsigned int port,
 238                              enum ib_gid_type default_gid_type)
 239 {
 240         unsigned long supported_gids;
 241
 242         if (port < rdma_start_port(cma_dev->device) ||
 243             port > rdma_end_port(cma_dev->device))
 244                 return -EINVAL;
 245
 246         supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
 247
 248         if (!(supported_gids & 1 << default_gid_type))
 249                 return -EINVAL;
 250
 251         cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
 252                 default_gid_type;
 253
 254         return 0;
 255 }
 256
 257 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
 258 {
 259         return cma_dev->device;
 260 }
 261
 262 /*
 263  * Device removal can occur at anytime, so we need extra handling to
 264  * serialize notifying the user of device removal with other callbacks.
 265  * We do this by disabling removal notification while a callback is in process,
 266  * and reporting it after the callback completes.
 267  */
 268 struct rdma_id_private {
 269         struct rdma_cm_id       id;
 270
 271         struct rdma_bind_list   *bind_list;
 272         struct hlist_node       node;
 273         struct list_head        list; /* listen_any_list or cma_device.list */
 274         struct list_head        listen_list; /* per device listens */
 275         struct cma_device       *cma_dev;
 276         struct list_head        mc_list;
 277
 278         int                     internal_id;
 279         enum rdma_cm_state      state;
 280         spinlock_t              lock;
 281         struct mutex            qp_mutex;
 282
 283         struct completion       comp;
 284         atomic_t                refcount;
 285         struct mutex            handler_mutex;
 286
 287         int                     backlog;
 288         int                     timeout_ms;
 289         struct ib_sa_query      *query;
 290         int                     query_id;
 291         union {
 292                 struct ib_cm_id *ib;
 293                 struct iw_cm_id *iw;
 294         } cm_id;
 295
 296         u32                     seq_num;
 297         u32                     qkey;
 298         u32                     qp_num;
 299         pid_t                   owner;
 300         u32                     options;
 301         u8                      srq;
 302         u8                      tos;
 303         u8                      reuseaddr;
 304         u8                      afonly;
 305         enum ib_gid_type        gid_type;
 306 };
 307
 308 struct cma_multicast {
 309         struct rdma_id_private *id_priv;
 310         union {
 311                 struct ib_sa_multicast *ib;
 312         } multicast;
 313         struct list_head        list;
 314         void                    *context;
 315         struct sockaddr_storage addr;
 316         struct kref             mcref;
 317         bool                    igmp_joined;
 318         u8                      join_state;
 319 };
 320
 321 struct cma_work {
 322         struct work_struct      work;
 323         struct rdma_id_private  *id;
 324         enum rdma_cm_state      old_state;
 325         enum rdma_cm_state      new_state;
 326         struct rdma_cm_event    event;
 327 };
 328
 329 struct cma_ndev_work {
 330         struct work_struct      work;
 331         struct rdma_id_private  *id;
 332         struct rdma_cm_event    event;
 333 };
 334
 335 struct iboe_mcast_work {
 336         struct work_struct       work;
 337         struct rdma_id_private  *id;
 338         struct cma_multicast    *mc;
 339 };
 340
 341 union cma_ip_addr {
 342         struct in6_addr ip6;
 343         struct {
 344                 __be32 pad[3];
 345                 __be32 addr;
 346         } ip4;
 347 };
 348
 349 struct cma_hdr {
 350         u8 cma_version;
 351         u8 ip_version;  /* IP version: 7:4 */
 352         __be16 port;
 353         union cma_ip_addr src_addr;
 354         union cma_ip_addr dst_addr;
 355 };
 356
 357 #define CMA_VERSION 0x00
 358
 359 struct cma_req_info {
 360         struct ib_device *device;
 361         int port;
 362         union ib_gid local_gid;
 363         __be64 service_id;
 364         u16 pkey;
 365         bool has_gid:1;
 366 };
 367
 368 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
 369 {
 370         unsigned long flags;
 371         int ret;
 372
 373         spin_lock_irqsave(&id_priv->lock, flags);
 374         ret = (id_priv->state == comp);
 375         spin_unlock_irqrestore(&id_priv->lock, flags);
 376         return ret;
 377 }
 378
 379 static int cma_comp_exch(struct rdma_id_private *id_priv,
 380                          enum rdma_cm_state comp, enum rdma_cm_state exch)
 381 {
 382         unsigned long flags;
 383         int ret;
 384
 385         spin_lock_irqsave(&id_priv->lock, flags);
 386         if ((ret = (id_priv->state == comp)))
 387                 id_priv->state = exch;
 388         spin_unlock_irqrestore(&id_priv->lock, flags);
 389         return ret;
 390 }
 391
 392 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
 393                                    enum rdma_cm_state exch)
 394 {
 395         unsigned long flags;
 396         enum rdma_cm_state old;
 397
 398         spin_lock_irqsave(&id_priv->lock, flags);
 399         old = id_priv->state;
 400         id_priv->state = exch;
 401         spin_unlock_irqrestore(&id_priv->lock, flags);
 402         return old;
 403 }
 404
 405 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
 406 {
 407         return hdr->ip_version >> 4;
 408 }
 409
 410 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
 411 {
 412         hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
 413 }
 414
 415 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
 416 {
 417         struct in_device *in_dev = NULL;
 418
 419         if (ndev) {
 420                 rtnl_lock();
 421                 in_dev = __in_dev_get_rtnl(ndev);
 422                 if (in_dev) {
 423                         if (join)
 424                                 ip_mc_inc_group(in_dev,
 425                                                 *(__be32 *)(mgid->raw + 12));
 426                         else
 427                                 ip_mc_dec_group(in_dev,
 428                                                 *(__be32 *)(mgid->raw + 12));
 429                 }
 430                 rtnl_unlock();
 431         }
 432         return (in_dev) ? 0 : -ENODEV;
 433 }
 434
 435 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
 436                                struct cma_device *cma_dev)
 437 {
 438         cma_ref_dev(cma_dev);
 439         id_priv->cma_dev = cma_dev;
 440         id_priv->gid_type = 0;
 441         id_priv->id.device = cma_dev->device;
 442         id_priv->id.route.addr.dev_addr.transport =
 443                 rdma_node_get_transport(cma_dev->device->node_type);
 444         list_add_tail(&id_priv->list, &cma_dev->id_list);
 445 }
 446
 447 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
 448                               struct cma_device *cma_dev)
 449 {
 450         _cma_attach_to_dev(id_priv, cma_dev);
 451         id_priv->gid_type =
 452                 cma_dev->default_gid_type[id_priv->id.port_num -
 453                                           rdma_start_port(cma_dev->device)];
 454 }
 455
 456 void cma_deref_dev(struct cma_device *cma_dev)
 457 {
 458         if (atomic_dec_and_test(&cma_dev->refcount))
 459                 complete(&cma_dev->comp);
 460 }
 461
 462 static inline void release_mc(struct kref *kref)
 463 {
 464         struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
 465
 466         kfree(mc->multicast.ib);
 467         kfree(mc);
 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_deref_dev(id_priv->cma_dev);
 475         id_priv->cma_dev = NULL;
 476         mutex_unlock(&lock);
 477 }
 478
 479 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
 480 {
 481         return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
 482 }
 483
 484 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
 485 {
 486         return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
 487 }
 488
 489 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
 490 {
 491         return id_priv->id.route.addr.src_addr.ss_family;
 492 }
 493
 494 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
 495 {
 496         struct ib_sa_mcmember_rec rec;
 497         int ret = 0;
 498
 499         if (id_priv->qkey) {
 500                 if (qkey && id_priv->qkey != qkey)
 501                         return -EINVAL;
 502                 return 0;
 503         }
 504
 505         if (qkey) {
 506                 id_priv->qkey = qkey;
 507                 return 0;
 508         }
 509
 510         switch (id_priv->id.ps) {
 511         case RDMA_PS_UDP:
 512         case RDMA_PS_IB:
 513                 id_priv->qkey = RDMA_UDP_QKEY;
 514                 break;
 515         case RDMA_PS_IPOIB:
 516                 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
 517                 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
 518                                              id_priv->id.port_num, &rec.mgid,
 519                                              &rec);
 520                 if (!ret)
 521                         id_priv->qkey = be32_to_cpu(rec.qkey);
 522                 break;
 523         default:
 524                 break;
 525         }
 526         return ret;
 527 }
 528
 529 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
 530 {
 531         dev_addr->dev_type = ARPHRD_INFINIBAND;
 532         rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
 533         ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
 534 }
 535
 536 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
 537 {
 538         int ret;
 539
 540         if (addr->sa_family != AF_IB) {
 541                 ret = rdma_translate_ip(addr, dev_addr, NULL);
 542         } else {
 543                 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
 544                 ret = 0;
 545         }
 546
 547         return ret;
 548 }
 549
 550 static inline int cma_validate_port(struct ib_device *device, u8 port,
 551                                     enum ib_gid_type gid_type,
 552                                       union ib_gid *gid, int dev_type,
 553                                       int bound_if_index)
 554 {
 555         int ret = -ENODEV;
 556         struct net_device *ndev = NULL;
 557
 558         if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
 559                 return ret;
 560
 561         if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
 562                 return ret;
 563
 564         if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
 565                 ndev = dev_get_by_index(&init_net, bound_if_index);
 566                 if (ndev && ndev->flags & IFF_LOOPBACK) {
 567                         pr_info("detected loopback device\n");
 568                         dev_put(ndev);
 569
 570                         if (!device->get_netdev)
 571                                 return -EOPNOTSUPP;
 572
 573                         ndev = device->get_netdev(device, port);
 574                         if (!ndev)
 575                                 return -ENODEV;
 576                 }
 577         } else {
 578                 gid_type = IB_GID_TYPE_IB;
 579         }
 580
 581         ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
 582                                          ndev, NULL);
 583
 584         if (ndev)
 585                 dev_put(ndev);
 586
 587         return ret;
 588 }
 589
 590 static int cma_acquire_dev(struct rdma_id_private *id_priv,
 591                            struct rdma_id_private *listen_id_priv)
 592 {
 593         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 594         struct cma_device *cma_dev;
 595         union ib_gid gid, iboe_gid, *gidp;
 596         int ret = -ENODEV;
 597         u8 port;
 598
 599         if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
 600             id_priv->id.ps == RDMA_PS_IPOIB)
 601                 return -EINVAL;
 602
 603         mutex_lock(&lock);
 604         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
 605                     &iboe_gid);
 606
 607         memcpy(&gid, dev_addr->src_dev_addr +
 608                rdma_addr_gid_offset(dev_addr), sizeof gid);
 609
 610         if (listen_id_priv) {
 611                 cma_dev = listen_id_priv->cma_dev;
 612                 port = listen_id_priv->id.port_num;
 613                 gidp = rdma_protocol_roce(cma_dev->device, port) ?
 614                        &iboe_gid : &gid;
 615
 616                 ret = cma_validate_port(cma_dev->device, port,
 617                                         rdma_protocol_ib(cma_dev->device, port) ?
 618                                         IB_GID_TYPE_IB :
 619                                         listen_id_priv->gid_type, gidp,
 620                                         dev_addr->dev_type,
 621                                         dev_addr->bound_dev_if);
 622                 if (!ret) {
 623                         id_priv->id.port_num = port;
 624                         goto out;
 625                 }
 626         }
 627
 628         list_for_each_entry(cma_dev, &dev_list, list) {
 629                 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
 630                         if (listen_id_priv &&
 631                             listen_id_priv->cma_dev == cma_dev &&
 632                             listen_id_priv->id.port_num == port)
 633                                 continue;
 634
 635                         gidp = rdma_protocol_roce(cma_dev->device, port) ?
 636                                &iboe_gid : &gid;
 637
 638                         ret = cma_validate_port(cma_dev->device, port,
 639                                                 rdma_protocol_ib(cma_dev->device, port) ?
 640                                                 IB_GID_TYPE_IB :
 641                                                 cma_dev->default_gid_type[port - 1],
 642                                                 gidp, dev_addr->dev_type,
 643                                                 dev_addr->bound_dev_if);
 644                         if (!ret) {
 645                                 id_priv->id.port_num = port;
 646                                 goto out;
 647                         }
 648                 }
 649         }
 650
 651 out:
 652         if (!ret)
 653                 cma_attach_to_dev(id_priv, cma_dev);
 654
 655         mutex_unlock(&lock);
 656         return ret;
 657 }
 658
 659 /*
 660  * Select the source IB device and address to reach the destination IB address.
 661  */
 662 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
 663 {
 664         struct cma_device *cma_dev, *cur_dev;
 665         struct sockaddr_ib *addr;
 666         union ib_gid gid, sgid, *dgid;
 667         u16 pkey, index;
 668         u8 p;
 669         int i;
 670
 671         cma_dev = NULL;
 672         addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
 673         dgid = (union ib_gid *) &addr->sib_addr;
 674         pkey = ntohs(addr->sib_pkey);
 675
 676         mutex_lock(&lock);
 677         list_for_each_entry(cur_dev, &dev_list, list) {
 678                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
 679                         if (!rdma_cap_af_ib(cur_dev->device, p))
 680                                 continue;
 681
 682                         if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
 683                                 continue;
 684
 685                         for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
 686                                                        &gid, NULL);
 687                              i++) {
 688                                 if (!memcmp(&gid, dgid, sizeof(gid))) {
 689                                         cma_dev = cur_dev;
 690                                         sgid = gid;
 691                                         id_priv->id.port_num = p;
 692                                         goto found;
 693                                 }
 694
 695                                 if (!cma_dev && (gid.global.subnet_prefix ==
 696                                                  dgid->global.subnet_prefix)) {
 697                                         cma_dev = cur_dev;
 698                                         sgid = gid;
 699                                         id_priv->id.port_num = p;
 700                                         goto found;
 701                                 }
 702                         }
 703                 }
 704         }
 705         mutex_unlock(&lock);
 706         return -ENODEV;
 707
 708 found:
 709         cma_attach_to_dev(id_priv, cma_dev);
 710         mutex_unlock(&lock);
 711         addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
 712         memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
 713         cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
 714         return 0;
 715 }
 716
 717 static void cma_deref_id(struct rdma_id_private *id_priv)
 718 {
 719         if (atomic_dec_and_test(&id_priv->refcount))
 720                 complete(&id_priv->comp);
 721 }
 722
 723 struct rdma_cm_id *rdma_create_id(struct net *net,
 724                                   rdma_cm_event_handler event_handler,
 725                                   void *context, enum rdma_port_space ps,
 726                                   enum ib_qp_type qp_type)
 727 {
 728         struct rdma_id_private *id_priv;
 729
 730         id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
 731         if (!id_priv)
 732                 return ERR_PTR(-ENOMEM);
 733
 734         id_priv->owner = task_pid_nr(current);
 735         id_priv->state = RDMA_CM_IDLE;
 736         id_priv->id.context = context;
 737         id_priv->id.event_handler = event_handler;
 738         id_priv->id.ps = ps;
 739         id_priv->id.qp_type = qp_type;
 740         spin_lock_init(&id_priv->lock);
 741         mutex_init(&id_priv->qp_mutex);
 742         init_completion(&id_priv->comp);
 743         atomic_set(&id_priv->refcount, 1);
 744         mutex_init(&id_priv->handler_mutex);
 745         INIT_LIST_HEAD(&id_priv->listen_list);
 746         INIT_LIST_HEAD(&id_priv->mc_list);
 747         get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
 748         id_priv->id.route.addr.dev_addr.net = get_net(net);
 749         id_priv->seq_num &= 0x00ffffff;
 750
 751         return &id_priv->id;
 752 }
 753 EXPORT_SYMBOL(rdma_create_id);
 754
 755 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 756 {
 757         struct ib_qp_attr qp_attr;
 758         int qp_attr_mask, ret;
 759
 760         qp_attr.qp_state = IB_QPS_INIT;
 761         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 762         if (ret)
 763                 return ret;
 764
 765         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 766         if (ret)
 767                 return ret;
 768
 769         qp_attr.qp_state = IB_QPS_RTR;
 770         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
 771         if (ret)
 772                 return ret;
 773
 774         qp_attr.qp_state = IB_QPS_RTS;
 775         qp_attr.sq_psn = 0;
 776         ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
 777
 778         return ret;
 779 }
 780
 781 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
 782 {
 783         struct ib_qp_attr qp_attr;
 784         int qp_attr_mask, ret;
 785
 786         qp_attr.qp_state = IB_QPS_INIT;
 787         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 788         if (ret)
 789                 return ret;
 790
 791         return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 792 }
 793
 794 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
 795                    struct ib_qp_init_attr *qp_init_attr)
 796 {
 797         struct rdma_id_private *id_priv;
 798         struct ib_qp *qp;
 799         int ret;
 800
 801         id_priv = container_of(id, struct rdma_id_private, id);
 802         if (id->device != pd->device)
 803                 return -EINVAL;
 804
 805         qp_init_attr->port_num = id->port_num;
 806         qp = ib_create_qp(pd, qp_init_attr);
 807         if (IS_ERR(qp))
 808                 return PTR_ERR(qp);
 809
 810         if (id->qp_type == IB_QPT_UD)
 811                 ret = cma_init_ud_qp(id_priv, qp);
 812         else
 813                 ret = cma_init_conn_qp(id_priv, qp);
 814         if (ret)
 815                 goto err;
 816
 817         id->qp = qp;
 818         id_priv->qp_num = qp->qp_num;
 819         id_priv->srq = (qp->srq != NULL);
 820         return 0;
 821 err:
 822         ib_destroy_qp(qp);
 823         return ret;
 824 }
 825 EXPORT_SYMBOL(rdma_create_qp);
 826
 827 void rdma_destroy_qp(struct rdma_cm_id *id)
 828 {
 829         struct rdma_id_private *id_priv;
 830
 831         id_priv = container_of(id, struct rdma_id_private, id);
 832         mutex_lock(&id_priv->qp_mutex);
 833         ib_destroy_qp(id_priv->id.qp);
 834         id_priv->id.qp = NULL;
 835         mutex_unlock(&id_priv->qp_mutex);
 836 }
 837 EXPORT_SYMBOL(rdma_destroy_qp);
 838
 839 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
 840                              struct rdma_conn_param *conn_param)
 841 {
 842         struct ib_qp_attr qp_attr;
 843         int qp_attr_mask, ret;
 844         union ib_gid sgid;
 845
 846         mutex_lock(&id_priv->qp_mutex);
 847         if (!id_priv->id.qp) {
 848                 ret = 0;
 849                 goto out;
 850         }
 851
 852         /* Need to update QP attributes from default values. */
 853         qp_attr.qp_state = IB_QPS_INIT;
 854         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 855         if (ret)
 856                 goto out;
 857
 858         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 859         if (ret)
 860                 goto out;
 861
 862         qp_attr.qp_state = IB_QPS_RTR;
 863         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 864         if (ret)
 865                 goto out;
 866
 867         ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
 868                            qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
 869         if (ret)
 870                 goto out;
 871
 872         BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
 873
 874         if (conn_param)
 875                 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
 876         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 877 out:
 878         mutex_unlock(&id_priv->qp_mutex);
 879         return ret;
 880 }
 881
 882 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
 883                              struct rdma_conn_param *conn_param)
 884 {
 885         struct ib_qp_attr qp_attr;
 886         int qp_attr_mask, ret;
 887
 888         mutex_lock(&id_priv->qp_mutex);
 889         if (!id_priv->id.qp) {
 890                 ret = 0;
 891                 goto out;
 892         }
 893
 894         qp_attr.qp_state = IB_QPS_RTS;
 895         ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
 896         if (ret)
 897                 goto out;
 898
 899         if (conn_param)
 900                 qp_attr.max_rd_atomic = conn_param->initiator_depth;
 901         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
 902 out:
 903         mutex_unlock(&id_priv->qp_mutex);
 904         return ret;
 905 }
 906
 907 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
 908 {
 909         struct ib_qp_attr qp_attr;
 910         int ret;
 911
 912         mutex_lock(&id_priv->qp_mutex);
 913         if (!id_priv->id.qp) {
 914                 ret = 0;
 915                 goto out;
 916         }
 917
 918         qp_attr.qp_state = IB_QPS_ERR;
 919         ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
 920 out:
 921         mutex_unlock(&id_priv->qp_mutex);
 922         return ret;
 923 }
 924
 925 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
 926                                struct ib_qp_attr *qp_attr, int *qp_attr_mask)
 927 {
 928         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
 929         int ret;
 930         u16 pkey;
 931
 932         if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
 933                 pkey = 0xffff;
 934         else
 935                 pkey = ib_addr_get_pkey(dev_addr);
 936
 937         ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
 938                                   pkey, &qp_attr->pkey_index);
 939         if (ret)
 940                 return ret;
 941
 942         qp_attr->port_num = id_priv->id.port_num;
 943         *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
 944
 945         if (id_priv->id.qp_type == IB_QPT_UD) {
 946                 ret = cma_set_qkey(id_priv, 0);
 947                 if (ret)
 948                         return ret;
 949
 950                 qp_attr->qkey = id_priv->qkey;
 951                 *qp_attr_mask |= IB_QP_QKEY;
 952         } else {
 953                 qp_attr->qp_access_flags = 0;
 954                 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
 955         }
 956         return 0;
 957 }
 958
 959 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
 960                        int *qp_attr_mask)
 961 {
 962         struct rdma_id_private *id_priv;
 963         int ret = 0;
 964
 965         id_priv = container_of(id, struct rdma_id_private, id);
 966         if (rdma_cap_ib_cm(id->device, id->port_num)) {
 967                 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
 968                         ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
 969                 else
 970                         ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
 971                                                  qp_attr_mask);
 972
 973                 if (qp_attr->qp_state == IB_QPS_RTR)
 974                         qp_attr->rq_psn = id_priv->seq_num;
 975         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
 976                 if (!id_priv->cm_id.iw) {
 977                         qp_attr->qp_access_flags = 0;
 978                         *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
 979                 } else
 980                         ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
 981                                                  qp_attr_mask);
 982                 qp_attr->port_num = id_priv->id.port_num;
 983                 *qp_attr_mask |= IB_QP_PORT;
 984         } else
 985                 ret = -ENOSYS;
 986
 987         return ret;
 988 }
 989 EXPORT_SYMBOL(rdma_init_qp_attr);
 990
 991 static inline int cma_zero_addr(struct sockaddr *addr)
 992 {
 993         switch (addr->sa_family) {
 994         case AF_INET:
 995                 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
 996         case AF_INET6:
 997                 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
 998         case AF_IB:
 999                 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1000         default:
1001                 return 0;
1002         }
1003 }
1004
1005 static inline int cma_loopback_addr(struct sockaddr *addr)
1006 {
1007         switch (addr->sa_family) {
1008         case AF_INET:
1009                 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1010         case AF_INET6:
1011                 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1012         case AF_IB:
1013                 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1014         default:
1015                 return 0;
1016         }
1017 }
1018
1019 static inline int cma_any_addr(struct sockaddr *addr)
1020 {
1021         return cma_zero_addr(addr) || cma_loopback_addr(addr);
1022 }
1023
1024 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1025 {
1026         if (src->sa_family != dst->sa_family)
1027                 return -1;
1028
1029         switch (src->sa_family) {
1030         case AF_INET:
1031                 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1032                        ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1033         case AF_INET6:
1034                 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1035                                      &((struct sockaddr_in6 *) dst)->sin6_addr);
1036         default:
1037                 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1038                                    &((struct sockaddr_ib *) dst)->sib_addr);
1039         }
1040 }
1041
1042 static __be16 cma_port(struct sockaddr *addr)
1043 {
1044         struct sockaddr_ib *sib;
1045
1046         switch (addr->sa_family) {
1047         case AF_INET:
1048                 return ((struct sockaddr_in *) addr)->sin_port;
1049         case AF_INET6:
1050                 return ((struct sockaddr_in6 *) addr)->sin6_port;
1051         case AF_IB:
1052                 sib = (struct sockaddr_ib *) addr;
1053                 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1054                                     be64_to_cpu(sib->sib_sid_mask)));
1055         default:
1056                 return 0;
1057         }
1058 }
1059
1060 static inline int cma_any_port(struct sockaddr *addr)
1061 {
1062         return !cma_port(addr);
1063 }
1064
1065 static void cma_save_ib_info(struct sockaddr *src_addr,
1066                              struct sockaddr *dst_addr,
1067                              struct rdma_cm_id *listen_id,
1068                              struct ib_sa_path_rec *path)
1069 {
1070         struct sockaddr_ib *listen_ib, *ib;
1071
1072         listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1073         if (src_addr) {
1074                 ib = (struct sockaddr_ib *)src_addr;
1075                 ib->sib_family = AF_IB;
1076                 if (path) {
1077                         ib->sib_pkey = path->pkey;
1078                         ib->sib_flowinfo = path->flow_label;
1079                         memcpy(&ib->sib_addr, &path->sgid, 16);
1080                         ib->sib_sid = path->service_id;
1081                         ib->sib_scope_id = 0;
1082                 } else {
1083                         ib->sib_pkey = listen_ib->sib_pkey;
1084                         ib->sib_flowinfo = listen_ib->sib_flowinfo;
1085                         ib->sib_addr = listen_ib->sib_addr;
1086                         ib->sib_sid = listen_ib->sib_sid;
1087                         ib->sib_scope_id = listen_ib->sib_scope_id;
1088                 }
1089                 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1090         }
1091         if (dst_addr) {
1092                 ib = (struct sockaddr_ib *)dst_addr;
1093                 ib->sib_family = AF_IB;
1094                 if (path) {
1095                         ib->sib_pkey = path->pkey;
1096                         ib->sib_flowinfo = path->flow_label;
1097                         memcpy(&ib->sib_addr, &path->dgid, 16);
1098                 }
1099         }
1100 }
1101
1102 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1103                               struct sockaddr_in *dst_addr,
1104                               struct cma_hdr *hdr,
1105                               __be16 local_port)
1106 {
1107         if (src_addr) {
1108                 *src_addr = (struct sockaddr_in) {
1109                         .sin_family = AF_INET,
1110                         .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1111                         .sin_port = local_port,
1112                 };
1113         }
1114
1115         if (dst_addr) {
1116                 *dst_addr = (struct sockaddr_in) {
1117                         .sin_family = AF_INET,
1118                         .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1119                         .sin_port = hdr->port,
1120                 };
1121         }
1122 }
1123
1124 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1125                               struct sockaddr_in6 *dst_addr,
1126                               struct cma_hdr *hdr,
1127                               __be16 local_port)
1128 {
1129         if (src_addr) {
1130                 *src_addr = (struct sockaddr_in6) {
1131                         .sin6_family = AF_INET6,
1132                         .sin6_addr = hdr->dst_addr.ip6,
1133                         .sin6_port = local_port,
1134                 };
1135         }
1136
1137         if (dst_addr) {
1138                 *dst_addr = (struct sockaddr_in6) {
1139                         .sin6_family = AF_INET6,
1140                         .sin6_addr = hdr->src_addr.ip6,
1141                         .sin6_port = hdr->port,
1142                 };
1143         }
1144 }
1145
1146 static u16 cma_port_from_service_id(__be64 service_id)
1147 {
1148         return (u16)be64_to_cpu(service_id);
1149 }
1150
1151 static int cma_save_ip_info(struct sockaddr *src_addr,
1152                             struct sockaddr *dst_addr,
1153                             struct ib_cm_event *ib_event,
1154                             __be64 service_id)
1155 {
1156         struct cma_hdr *hdr;
1157         __be16 port;
1158
1159         hdr = ib_event->private_data;
1160         if (hdr->cma_version != CMA_VERSION)
1161                 return -EINVAL;
1162
1163         port = htons(cma_port_from_service_id(service_id));
1164
1165         switch (cma_get_ip_ver(hdr)) {
1166         case 4:
1167                 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1168                                   (struct sockaddr_in *)dst_addr, hdr, port);
1169                 break;
1170         case 6:
1171                 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1172                                   (struct sockaddr_in6 *)dst_addr, hdr, port);
1173                 break;
1174         default:
1175                 return -EAFNOSUPPORT;
1176         }
1177
1178         return 0;
1179 }
1180
1181 static int cma_save_net_info(struct sockaddr *src_addr,
1182                              struct sockaddr *dst_addr,
1183                              struct rdma_cm_id *listen_id,
1184                              struct ib_cm_event *ib_event,
1185                              sa_family_t sa_family, __be64 service_id)
1186 {
1187         if (sa_family == AF_IB) {
1188                 if (ib_event->event == IB_CM_REQ_RECEIVED)
1189                         cma_save_ib_info(src_addr, dst_addr, listen_id,
1190                                          ib_event->param.req_rcvd.primary_path);
1191                 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1192                         cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1193                 return 0;
1194         }
1195
1196         return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1197 }
1198
1199 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1200                              struct cma_req_info *req)
1201 {
1202         const struct ib_cm_req_event_param *req_param =
1203                 &ib_event->param.req_rcvd;
1204         const struct ib_cm_sidr_req_event_param *sidr_param =
1205                 &ib_event->param.sidr_req_rcvd;
1206
1207         switch (ib_event->event) {
1208         case IB_CM_REQ_RECEIVED:
1209                 req->device     = req_param->listen_id->device;
1210                 req->port       = req_param->port;
1211                 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1212                        sizeof(req->local_gid));
1213                 req->has_gid    = true;
1214                 req->service_id = req_param->primary_path->service_id;
1215                 req->pkey       = be16_to_cpu(req_param->primary_path->pkey);
1216                 if (req->pkey != req_param->bth_pkey)
1217                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1218                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1219                                             req_param->bth_pkey, req->pkey);
1220                 break;
1221         case IB_CM_SIDR_REQ_RECEIVED:
1222                 req->device     = sidr_param->listen_id->device;
1223                 req->port       = sidr_param->port;
1224                 req->has_gid    = false;
1225                 req->service_id = sidr_param->service_id;
1226                 req->pkey       = sidr_param->pkey;
1227                 if (req->pkey != sidr_param->bth_pkey)
1228                         pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1229                                             "RDMA CMA: in the future this may cause the request to be dropped\n",
1230                                             sidr_param->bth_pkey, req->pkey);
1231                 break;
1232         default:
1233                 return -EINVAL;
1234         }
1235
1236         return 0;
1237 }
1238
1239 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1240                                   const struct sockaddr_in *dst_addr,
1241                                   const struct sockaddr_in *src_addr)
1242 {
1243         __be32 daddr = dst_addr->sin_addr.s_addr,
1244                saddr = src_addr->sin_addr.s_addr;
1245         struct fib_result res;
1246         struct flowi4 fl4;
1247         int err;
1248         bool ret;
1249
1250         if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1251             ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1252             ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1253             ipv4_is_loopback(saddr))
1254                 return false;
1255
1256         memset(&fl4, 0, sizeof(fl4));
1257         fl4.flowi4_iif = net_dev->ifindex;
1258         fl4.daddr = daddr;
1259         fl4.saddr = saddr;
1260
1261         rcu_read_lock();
1262         err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1263         ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1264         rcu_read_unlock();
1265
1266         return ret;
1267 }
1268
1269 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1270                                   const struct sockaddr_in6 *dst_addr,
1271                                   const struct sockaddr_in6 *src_addr)
1272 {
1273 #if IS_ENABLED(CONFIG_IPV6)
1274         const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1275                            IPV6_ADDR_LINKLOCAL;
1276         struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1277                                          &src_addr->sin6_addr, net_dev->ifindex,
1278                                          strict);
1279         bool ret;
1280
1281         if (!rt)
1282                 return false;
1283
1284         ret = rt->rt6i_idev->dev == net_dev;
1285         ip6_rt_put(rt);
1286
1287         return ret;
1288 #else
1289         return false;
1290 #endif
1291 }
1292
1293 static bool validate_net_dev(struct net_device *net_dev,
1294                              const struct sockaddr *daddr,
1295                              const struct sockaddr *saddr)
1296 {
1297         const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1298         const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1299         const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1300         const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1301
1302         switch (daddr->sa_family) {
1303         case AF_INET:
1304                 return saddr->sa_family == AF_INET &&
1305                        validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1306
1307         case AF_INET6:
1308                 return saddr->sa_family == AF_INET6 &&
1309                        validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1310
1311         default:
1312                 return false;
1313         }
1314 }
1315
1316 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1317                                           const struct cma_req_info *req)
1318 {
1319         struct sockaddr_storage listen_addr_storage, src_addr_storage;
1320         struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1321                         *src_addr = (struct sockaddr *)&src_addr_storage;
1322         struct net_device *net_dev;
1323         const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1324         int err;
1325
1326         err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1327                                req->service_id);
1328         if (err)
1329                 return ERR_PTR(err);
1330
1331         net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1332                                            gid, listen_addr);
1333         if (!net_dev)
1334                 return ERR_PTR(-ENODEV);
1335
1336         if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1337                 dev_put(net_dev);
1338                 return ERR_PTR(-EHOSTUNREACH);
1339         }
1340
1341         return net_dev;
1342 }
1343
1344 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1345 {
1346         return (be64_to_cpu(service_id) >> 16) & 0xffff;
1347 }
1348
1349 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1350                                    const struct cma_hdr *hdr)
1351 {
1352         struct sockaddr *addr = cma_src_addr(id_priv);
1353         __be32 ip4_addr;
1354         struct in6_addr ip6_addr;
1355
1356         if (cma_any_addr(addr) && !id_priv->afonly)
1357                 return true;
1358
1359         switch (addr->sa_family) {
1360         case AF_INET:
1361                 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1362                 if (cma_get_ip_ver(hdr) != 4)
1363                         return false;
1364                 if (!cma_any_addr(addr) &&
1365                     hdr->dst_addr.ip4.addr != ip4_addr)
1366                         return false;
1367                 break;
1368         case AF_INET6:
1369                 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1370                 if (cma_get_ip_ver(hdr) != 6)
1371                         return false;
1372                 if (!cma_any_addr(addr) &&
1373                     memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1374                         return false;
1375                 break;
1376         case AF_IB:
1377                 return true;
1378         default:
1379                 return false;
1380         }
1381
1382         return true;
1383 }
1384
1385 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1386 {
1387         enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1388         enum rdma_transport_type transport =
1389                 rdma_node_get_transport(device->node_type);
1390
1391         return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1392 }
1393
1394 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1395 {
1396         struct ib_device *device = id->device;
1397         const int port_num = id->port_num ?: rdma_start_port(device);
1398
1399         return cma_protocol_roce_dev_port(device, port_num);
1400 }
1401
1402 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1403                               const struct net_device *net_dev,
1404                               u8 port_num)
1405 {
1406         const struct rdma_addr *addr = &id->route.addr;
1407
1408         if (!net_dev)
1409                 /* This request is an AF_IB request or a RoCE request */
1410                 return (!id->port_num || id->port_num == port_num) &&
1411                        (addr->src_addr.ss_family == AF_IB ||
1412                         cma_protocol_roce_dev_port(id->device, port_num));
1413
1414         /*
1415          * Net namespaces must match, and if the listner is listening
1416          * on a specific netdevice than netdevice must match as well.
1417          */
1418         if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1419             (!!addr->dev_addr.bound_dev_if ==
1420              (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1421                 return true;
1422         else
1423                 return false;
1424 }
1425
1426 static struct rdma_id_private *cma_find_listener(
1427                 const struct rdma_bind_list *bind_list,
1428                 const struct ib_cm_id *cm_id,
1429                 const struct ib_cm_event *ib_event,
1430                 const struct cma_req_info *req,
1431                 const struct net_device *net_dev)
1432 {
1433         struct rdma_id_private *id_priv, *id_priv_dev;
1434
1435         if (!bind_list)
1436                 return ERR_PTR(-EINVAL);
1437
1438         hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1439                 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1440                         if (id_priv->id.device == cm_id->device &&
1441                             cma_match_net_dev(&id_priv->id, net_dev, req->port))
1442                                 return id_priv;
1443                         list_for_each_entry(id_priv_dev,
1444                                             &id_priv->listen_list,
1445                                             listen_list) {
1446                                 if (id_priv_dev->id.device == cm_id->device &&
1447                                     cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1448                                         return id_priv_dev;
1449                         }
1450                 }
1451         }
1452
1453         return ERR_PTR(-EINVAL);
1454 }
1455
1456 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1457                                                  struct ib_cm_event *ib_event,
1458                                                  struct net_device **net_dev)
1459 {
1460         struct cma_req_info req;
1461         struct rdma_bind_list *bind_list;
1462         struct rdma_id_private *id_priv;
1463         int err;
1464
1465         err = cma_save_req_info(ib_event, &req);
1466         if (err)
1467                 return ERR_PTR(err);
1468
1469         *net_dev = cma_get_net_dev(ib_event, &req);
1470         if (IS_ERR(*net_dev)) {
1471                 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1472                         /* Assuming the protocol is AF_IB */
1473                         *net_dev = NULL;
1474                 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1475                         /* TODO find the net dev matching the request parameters
1476                          * through the RoCE GID table */
1477                         *net_dev = NULL;
1478                 } else {
1479                         return ERR_CAST(*net_dev);
1480                 }
1481         }
1482
1483         bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1484                                 rdma_ps_from_service_id(req.service_id),
1485                                 cma_port_from_service_id(req.service_id));
1486         id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1487         if (IS_ERR(id_priv) && *net_dev) {
1488                 dev_put(*net_dev);
1489                 *net_dev = NULL;
1490         }
1491
1492         return id_priv;
1493 }
1494
1495 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1496 {
1497         return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1498 }
1499
1500 static void cma_cancel_route(struct rdma_id_private *id_priv)
1501 {
1502         if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1503                 if (id_priv->query)
1504                         ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1505         }
1506 }
1507
1508 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1509 {
1510         struct rdma_id_private *dev_id_priv;
1511
1512         /*
1513          * Remove from listen_any_list to prevent added devices from spawning
1514          * additional listen requests.
1515          */
1516         mutex_lock(&lock);
1517         list_del(&id_priv->list);
1518
1519         while (!list_empty(&id_priv->listen_list)) {
1520                 dev_id_priv = list_entry(id_priv->listen_list.next,
1521                                          struct rdma_id_private, listen_list);
1522                 /* sync with device removal to avoid duplicate destruction */
1523                 list_del_init(&dev_id_priv->list);
1524                 list_del(&dev_id_priv->listen_list);
1525                 mutex_unlock(&lock);
1526
1527                 rdma_destroy_id(&dev_id_priv->id);
1528                 mutex_lock(&lock);
1529         }
1530         mutex_unlock(&lock);
1531 }
1532
1533 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1534                                  enum rdma_cm_state state)
1535 {
1536         switch (state) {
1537         case RDMA_CM_ADDR_QUERY:
1538                 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1539                 break;
1540         case RDMA_CM_ROUTE_QUERY:
1541                 cma_cancel_route(id_priv);
1542                 break;
1543         case RDMA_CM_LISTEN:
1544                 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1545                         cma_cancel_listens(id_priv);
1546                 break;
1547         default:
1548                 break;
1549         }
1550 }
1551
1552 static void cma_release_port(struct rdma_id_private *id_priv)
1553 {
1554         struct rdma_bind_list *bind_list = id_priv->bind_list;
1555         struct net *net = id_priv->id.route.addr.dev_addr.net;
1556
1557         if (!bind_list)
1558                 return;
1559
1560         mutex_lock(&lock);
1561         hlist_del(&id_priv->node);
1562         if (hlist_empty(&bind_list->owners)) {
1563                 cma_ps_remove(net, bind_list->ps, bind_list->port);
1564                 kfree(bind_list);
1565         }
1566         mutex_unlock(&lock);
1567 }
1568
1569 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1570 {
1571         struct cma_multicast *mc;
1572
1573         while (!list_empty(&id_priv->mc_list)) {
1574                 mc = container_of(id_priv->mc_list.next,
1575                                   struct cma_multicast, list);
1576                 list_del(&mc->list);
1577                 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1578                                       id_priv->id.port_num)) {
1579                         ib_sa_free_multicast(mc->multicast.ib);
1580                         kfree(mc);
1581                 } else {
1582                         if (mc->igmp_joined) {
1583                                 struct rdma_dev_addr *dev_addr =
1584                                         &id_priv->id.route.addr.dev_addr;
1585                                 struct net_device *ndev = NULL;
1586
1587                                 if (dev_addr->bound_dev_if)
1588                                         ndev = dev_get_by_index(&init_net,
1589                                                                 dev_addr->bound_dev_if);
1590                                 if (ndev) {
1591                                         cma_igmp_send(ndev,
1592                                                       &mc->multicast.ib->rec.mgid,
1593                                                       false);
1594                                         dev_put(ndev);
1595                                 }
1596                         }
1597                         kref_put(&mc->mcref, release_mc);
1598                 }
1599         }
1600 }
1601
1602 void rdma_destroy_id(struct rdma_cm_id *id)
1603 {
1604         struct rdma_id_private *id_priv;
1605         enum rdma_cm_state state;
1606
1607         id_priv = container_of(id, struct rdma_id_private, id);
1608         state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1609         cma_cancel_operation(id_priv, state);
1610
1611         /*
1612          * Wait for any active callback to finish.  New callbacks will find
1613          * the id_priv state set to destroying and abort.
1614          */
1615         mutex_lock(&id_priv->handler_mutex);
1616         mutex_unlock(&id_priv->handler_mutex);
1617
1618         if (id_priv->cma_dev) {
1619                 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1620                         if (id_priv->cm_id.ib)
1621                                 ib_destroy_cm_id(id_priv->cm_id.ib);
1622                 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1623                         if (id_priv->cm_id.iw)
1624                                 iw_destroy_cm_id(id_priv->cm_id.iw);
1625                 }
1626                 cma_leave_mc_groups(id_priv);
1627                 cma_release_dev(id_priv);
1628         }
1629
1630         cma_release_port(id_priv);
1631         cma_deref_id(id_priv);
1632         wait_for_completion(&id_priv->comp);
1633
1634         if (id_priv->internal_id)
1635                 cma_deref_id(id_priv->id.context);
1636
1637         kfree(id_priv->id.route.path_rec);
1638         put_net(id_priv->id.route.addr.dev_addr.net);
1639         kfree(id_priv);
1640 }
1641 EXPORT_SYMBOL(rdma_destroy_id);
1642
1643 static int cma_rep_recv(struct rdma_id_private *id_priv)
1644 {
1645         int ret;
1646
1647         ret = cma_modify_qp_rtr(id_priv, NULL);
1648         if (ret)
1649                 goto reject;
1650
1651         ret = cma_modify_qp_rts(id_priv, NULL);
1652         if (ret)
1653                 goto reject;
1654
1655         ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1656         if (ret)
1657                 goto reject;
1658
1659         return 0;
1660 reject:
1661         cma_modify_qp_err(id_priv);
1662         ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1663                        NULL, 0, NULL, 0);
1664         return ret;
1665 }
1666
1667 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1668                                    struct ib_cm_rep_event_param *rep_data,
1669                                    void *private_data)
1670 {
1671         event->param.conn.private_data = private_data;
1672         event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1673         event->param.conn.responder_resources = rep_data->responder_resources;
1674         event->param.conn.initiator_depth = rep_data->initiator_depth;
1675         event->param.conn.flow_control = rep_data->flow_control;
1676         event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1677         event->param.conn.srq = rep_data->srq;
1678         event->param.conn.qp_num = rep_data->remote_qpn;
1679 }
1680
1681 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1682 {
1683         struct rdma_id_private *id_priv = cm_id->context;
1684         struct rdma_cm_event event;
1685         int ret = 0;
1686
1687         mutex_lock(&id_priv->handler_mutex);
1688         if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1689              id_priv->state != RDMA_CM_CONNECT) ||
1690             (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1691              id_priv->state != RDMA_CM_DISCONNECT))
1692                 goto out;
1693
1694         memset(&event, 0, sizeof event);
1695         switch (ib_event->event) {
1696         case IB_CM_REQ_ERROR:
1697         case IB_CM_REP_ERROR:
1698                 event.event = RDMA_CM_EVENT_UNREACHABLE;
1699                 event.status = -ETIMEDOUT;
1700                 break;
1701         case IB_CM_REP_RECEIVED:
1702                 if (id_priv->id.qp) {
1703                         event.status = cma_rep_recv(id_priv);
1704                         event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1705                                                      RDMA_CM_EVENT_ESTABLISHED;
1706                 } else {
1707                         event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1708                 }
1709                 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1710                                        ib_event->private_data);
1711                 break;
1712         case IB_CM_RTU_RECEIVED:
1713         case IB_CM_USER_ESTABLISHED:
1714                 event.event = RDMA_CM_EVENT_ESTABLISHED;
1715                 break;
1716         case IB_CM_DREQ_ERROR:
1717                 event.status = -ETIMEDOUT; /* fall through */
1718         case IB_CM_DREQ_RECEIVED:
1719         case IB_CM_DREP_RECEIVED:
1720                 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1721                                    RDMA_CM_DISCONNECT))
1722                         goto out;
1723                 event.event = RDMA_CM_EVENT_DISCONNECTED;
1724                 break;
1725         case IB_CM_TIMEWAIT_EXIT:
1726                 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1727                 break;
1728         case IB_CM_MRA_RECEIVED:
1729                 /* ignore event */
1730                 goto out;
1731         case IB_CM_REJ_RECEIVED:
1732                 cma_modify_qp_err(id_priv);
1733                 event.status = ib_event->param.rej_rcvd.reason;
1734                 event.event = RDMA_CM_EVENT_REJECTED;
1735                 event.param.conn.private_data = ib_event->private_data;
1736                 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1737                 break;
1738         default:
1739                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1740                        ib_event->event);
1741                 goto out;
1742         }
1743
1744         ret = id_priv->id.event_handler(&id_priv->id, &event);
1745         if (ret) {
1746                 /* Destroy the CM ID by returning a non-zero value. */
1747                 id_priv->cm_id.ib = NULL;
1748                 cma_exch(id_priv, RDMA_CM_DESTROYING);
1749                 mutex_unlock(&id_priv->handler_mutex);
1750                 rdma_destroy_id(&id_priv->id);
1751                 return ret;
1752         }
1753 out:
1754         mutex_unlock(&id_priv->handler_mutex);
1755         return ret;
1756 }
1757
1758 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1759                                                struct ib_cm_event *ib_event,
1760                                                struct net_device *net_dev)
1761 {
1762         struct rdma_id_private *id_priv;
1763         struct rdma_cm_id *id;
1764         struct rdma_route *rt;
1765         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1766         const __be64 service_id =
1767                       ib_event->param.req_rcvd.primary_path->service_id;
1768         int ret;
1769
1770         id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1771                             listen_id->event_handler, listen_id->context,
1772                             listen_id->ps, ib_event->param.req_rcvd.qp_type);
1773         if (IS_ERR(id))
1774                 return NULL;
1775
1776         id_priv = container_of(id, struct rdma_id_private, id);
1777         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1778                               (struct sockaddr *)&id->route.addr.dst_addr,
1779                               listen_id, ib_event, ss_family, service_id))
1780                 goto err;
1781
1782         rt = &id->route;
1783         rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1784         rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1785                                GFP_KERNEL);
1786         if (!rt->path_rec)
1787                 goto err;
1788
1789         rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1790         if (rt->num_paths == 2)
1791                 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1792
1793         if (net_dev) {
1794                 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1795                 if (ret)
1796                         goto err;
1797         } else {
1798                 if (!cma_protocol_roce(listen_id) &&
1799                     cma_any_addr(cma_src_addr(id_priv))) {
1800                         rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1801                         rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1802                         ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1803                 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1804                         ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1805                         if (ret)
1806                                 goto err;
1807                 }
1808         }
1809         rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1810
1811         id_priv->state = RDMA_CM_CONNECT;
1812         return id_priv;
1813
1814 err:
1815         rdma_destroy_id(id);
1816         return NULL;
1817 }
1818
1819 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1820                                               struct ib_cm_event *ib_event,
1821                                               struct net_device *net_dev)
1822 {
1823         struct rdma_id_private *id_priv;
1824         struct rdma_cm_id *id;
1825         const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1826         struct net *net = listen_id->route.addr.dev_addr.net;
1827         int ret;
1828
1829         id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1830                             listen_id->ps, IB_QPT_UD);
1831         if (IS_ERR(id))
1832                 return NULL;
1833
1834         id_priv = container_of(id, struct rdma_id_private, id);
1835         if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1836                               (struct sockaddr *)&id->route.addr.dst_addr,
1837                               listen_id, ib_event, ss_family,
1838                               ib_event->param.sidr_req_rcvd.service_id))
1839                 goto err;
1840
1841         if (net_dev) {
1842                 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1843                 if (ret)
1844                         goto err;
1845         } else {
1846                 if (!cma_any_addr(cma_src_addr(id_priv))) {
1847                         ret = cma_translate_addr(cma_src_addr(id_priv),
1848                                                  &id->route.addr.dev_addr);
1849                         if (ret)
1850                                 goto err;
1851                 }
1852         }
1853
1854         id_priv->state = RDMA_CM_CONNECT;
1855         return id_priv;
1856 err:
1857         rdma_destroy_id(id);
1858         return NULL;
1859 }
1860
1861 static void cma_set_req_event_data(struct rdma_cm_event *event,
1862                                    struct ib_cm_req_event_param *req_data,
1863                                    void *private_data, int offset)
1864 {
1865         event->param.conn.private_data = private_data + offset;
1866         event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1867         event->param.conn.responder_resources = req_data->responder_resources;
1868         event->param.conn.initiator_depth = req_data->initiator_depth;
1869         event->param.conn.flow_control = req_data->flow_control;
1870         event->param.conn.retry_count = req_data->retry_count;
1871         event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1872         event->param.conn.srq = req_data->srq;
1873         event->param.conn.qp_num = req_data->remote_qpn;
1874 }
1875
1876 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1877 {
1878         return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1879                  (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1880                 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1881                  (id->qp_type == IB_QPT_UD)) ||
1882                 (!id->qp_type));
1883 }
1884
1885 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1886 {
1887         struct rdma_id_private *listen_id, *conn_id = NULL;
1888         struct rdma_cm_event event;
1889         struct net_device *net_dev;
1890         u8 offset;
1891         int ret;
1892
1893         listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1894         if (IS_ERR(listen_id))
1895                 return PTR_ERR(listen_id);
1896
1897         if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1898                 ret = -EINVAL;
1899                 goto net_dev_put;
1900         }
1901
1902         mutex_lock(&listen_id->handler_mutex);
1903         if (listen_id->state != RDMA_CM_LISTEN) {
1904                 ret = -ECONNABORTED;
1905                 goto err1;
1906         }
1907
1908         memset(&event, 0, sizeof event);
1909         offset = cma_user_data_offset(listen_id);
1910         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1911         if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1912                 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1913                 event.param.ud.private_data = ib_event->private_data + offset;
1914                 event.param.ud.private_data_len =
1915                                 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1916         } else {
1917                 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1918                 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1919                                        ib_event->private_data, offset);
1920         }
1921         if (!conn_id) {
1922                 ret = -ENOMEM;
1923                 goto err1;
1924         }
1925
1926         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1927         ret = cma_acquire_dev(conn_id, listen_id);
1928         if (ret)
1929                 goto err2;
1930
1931         conn_id->cm_id.ib = cm_id;
1932         cm_id->context = conn_id;
1933         cm_id->cm_handler = cma_ib_handler;
1934
1935         /*
1936          * Protect against the user destroying conn_id from another thread
1937          * until we're done accessing it.
1938          */
1939         atomic_inc(&conn_id->refcount);
1940         ret = conn_id->id.event_handler(&conn_id->id, &event);
1941         if (ret)
1942                 goto err3;
1943         /*
1944          * Acquire mutex to prevent user executing rdma_destroy_id()
1945          * while we're accessing the cm_id.
1946          */
1947         mutex_lock(&lock);
1948         if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1949             (conn_id->id.qp_type != IB_QPT_UD))
1950                 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1951         mutex_unlock(&lock);
1952         mutex_unlock(&conn_id->handler_mutex);
1953         mutex_unlock(&listen_id->handler_mutex);
1954         cma_deref_id(conn_id);
1955         if (net_dev)
1956                 dev_put(net_dev);
1957         return 0;
1958
1959 err3:
1960         cma_deref_id(conn_id);
1961         /* Destroy the CM ID by returning a non-zero value. */
1962         conn_id->cm_id.ib = NULL;
1963 err2:
1964         cma_exch(conn_id, RDMA_CM_DESTROYING);
1965         mutex_unlock(&conn_id->handler_mutex);
1966 err1:
1967         mutex_unlock(&listen_id->handler_mutex);
1968         if (conn_id)
1969                 rdma_destroy_id(&conn_id->id);
1970
1971 net_dev_put:
1972         if (net_dev)
1973                 dev_put(net_dev);
1974
1975         return ret;
1976 }
1977
1978 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1979 {
1980         if (addr->sa_family == AF_IB)
1981                 return ((struct sockaddr_ib *) addr)->sib_sid;
1982
1983         return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1984 }
1985 EXPORT_SYMBOL(rdma_get_service_id);
1986
1987 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1988 {
1989         struct rdma_id_private *id_priv = iw_id->context;
1990         struct rdma_cm_event event;
1991         int ret = 0;
1992         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1993         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1994
1995         mutex_lock(&id_priv->handler_mutex);
1996         if (id_priv->state != RDMA_CM_CONNECT)
1997                 goto out;
1998
1999         memset(&event, 0, sizeof event);
2000         switch (iw_event->event) {
2001         case IW_CM_EVENT_CLOSE:
2002                 event.event = RDMA_CM_EVENT_DISCONNECTED;
2003                 break;
2004         case IW_CM_EVENT_CONNECT_REPLY:
2005                 memcpy(cma_src_addr(id_priv), laddr,
2006                        rdma_addr_size(laddr));
2007                 memcpy(cma_dst_addr(id_priv), raddr,
2008                        rdma_addr_size(raddr));
2009                 switch (iw_event->status) {
2010                 case 0:
2011                         event.event = RDMA_CM_EVENT_ESTABLISHED;
2012                         event.param.conn.initiator_depth = iw_event->ird;
2013                         event.param.conn.responder_resources = iw_event->ord;
2014                         break;
2015                 case -ECONNRESET:
2016                 case -ECONNREFUSED:
2017                         event.event = RDMA_CM_EVENT_REJECTED;
2018                         break;
2019                 case -ETIMEDOUT:
2020                         event.event = RDMA_CM_EVENT_UNREACHABLE;
2021                         break;
2022                 default:
2023                         event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2024                         break;
2025                 }
2026                 break;
2027         case IW_CM_EVENT_ESTABLISHED:
2028                 event.event = RDMA_CM_EVENT_ESTABLISHED;
2029                 event.param.conn.initiator_depth = iw_event->ird;
2030                 event.param.conn.responder_resources = iw_event->ord;
2031                 break;
2032         default:
2033                 BUG_ON(1);
2034         }
2035
2036         event.status = iw_event->status;
2037         event.param.conn.private_data = iw_event->private_data;
2038         event.param.conn.private_data_len = iw_event->private_data_len;
2039         ret = id_priv->id.event_handler(&id_priv->id, &event);
2040         if (ret) {
2041                 /* Destroy the CM ID by returning a non-zero value. */
2042                 id_priv->cm_id.iw = NULL;
2043                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2044                 mutex_unlock(&id_priv->handler_mutex);
2045                 rdma_destroy_id(&id_priv->id);
2046                 return ret;
2047         }
2048
2049 out:
2050         mutex_unlock(&id_priv->handler_mutex);
2051         return ret;
2052 }
2053
2054 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2055                                struct iw_cm_event *iw_event)
2056 {
2057         struct rdma_cm_id *new_cm_id;
2058         struct rdma_id_private *listen_id, *conn_id;
2059         struct rdma_cm_event event;
2060         int ret = -ECONNABORTED;
2061         struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2062         struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2063
2064         listen_id = cm_id->context;
2065
2066         mutex_lock(&listen_id->handler_mutex);
2067         if (listen_id->state != RDMA_CM_LISTEN)
2068                 goto out;
2069
2070         /* Create a new RDMA id for the new IW CM ID */
2071         new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2072                                    listen_id->id.event_handler,
2073                                    listen_id->id.context,
2074                                    RDMA_PS_TCP, IB_QPT_RC);
2075         if (IS_ERR(new_cm_id)) {
2076                 ret = -ENOMEM;
2077                 goto out;
2078         }
2079         conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2080         mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2081         conn_id->state = RDMA_CM_CONNECT;
2082
2083         ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2084         if (ret) {
2085                 mutex_unlock(&conn_id->handler_mutex);
2086                 rdma_destroy_id(new_cm_id);
2087                 goto out;
2088         }
2089
2090         ret = cma_acquire_dev(conn_id, listen_id);
2091         if (ret) {
2092                 mutex_unlock(&conn_id->handler_mutex);
2093                 rdma_destroy_id(new_cm_id);
2094                 goto out;
2095         }
2096
2097         conn_id->cm_id.iw = cm_id;
2098         cm_id->context = conn_id;
2099         cm_id->cm_handler = cma_iw_handler;
2100
2101         memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2102         memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2103
2104         memset(&event, 0, sizeof event);
2105         event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2106         event.param.conn.private_data = iw_event->private_data;
2107         event.param.conn.private_data_len = iw_event->private_data_len;
2108         event.param.conn.initiator_depth = iw_event->ird;
2109         event.param.conn.responder_resources = iw_event->ord;
2110
2111         /*
2112          * Protect against the user destroying conn_id from another thread
2113          * until we're done accessing it.
2114          */
2115         atomic_inc(&conn_id->refcount);
2116         ret = conn_id->id.event_handler(&conn_id->id, &event);
2117         if (ret) {
2118                 /* User wants to destroy the CM ID */
2119                 conn_id->cm_id.iw = NULL;
2120                 cma_exch(conn_id, RDMA_CM_DESTROYING);
2121                 mutex_unlock(&conn_id->handler_mutex);
2122                 cma_deref_id(conn_id);
2123                 rdma_destroy_id(&conn_id->id);
2124                 goto out;
2125         }
2126
2127         mutex_unlock(&conn_id->handler_mutex);
2128         cma_deref_id(conn_id);
2129
2130 out:
2131         mutex_unlock(&listen_id->handler_mutex);
2132         return ret;
2133 }
2134
2135 static int cma_ib_listen(struct rdma_id_private *id_priv)
2136 {
2137         struct sockaddr *addr;
2138         struct ib_cm_id *id;
2139         __be64 svc_id;
2140
2141         addr = cma_src_addr(id_priv);
2142         svc_id = rdma_get_service_id(&id_priv->id, addr);
2143         id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2144         if (IS_ERR(id))
2145                 return PTR_ERR(id);
2146         id_priv->cm_id.ib = id;
2147
2148         return 0;
2149 }
2150
2151 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2152 {
2153         int ret;
2154         struct iw_cm_id *id;
2155
2156         id = iw_create_cm_id(id_priv->id.device,
2157                              iw_conn_req_handler,
2158                              id_priv);
2159         if (IS_ERR(id))
2160                 return PTR_ERR(id);
2161
2162         id->tos = id_priv->tos;
2163         id_priv->cm_id.iw = id;
2164
2165         memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2166                rdma_addr_size(cma_src_addr(id_priv)));
2167
2168         ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2169
2170         if (ret) {
2171                 iw_destroy_cm_id(id_priv->cm_id.iw);
2172                 id_priv->cm_id.iw = NULL;
2173         }
2174
2175         return ret;
2176 }
2177
2178 static int cma_listen_handler(struct rdma_cm_id *id,
2179                               struct rdma_cm_event *event)
2180 {
2181         struct rdma_id_private *id_priv = id->context;
2182
2183         id->context = id_priv->id.context;
2184         id->event_handler = id_priv->id.event_handler;
2185         return id_priv->id.event_handler(id, event);
2186 }
2187
2188 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2189                               struct cma_device *cma_dev)
2190 {
2191         struct rdma_id_private *dev_id_priv;
2192         struct rdma_cm_id *id;
2193         struct net *net = id_priv->id.route.addr.dev_addr.net;
2194         int ret;
2195
2196         if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2197                 return;
2198
2199         id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2200                             id_priv->id.qp_type);
2201         if (IS_ERR(id))
2202                 return;
2203
2204         dev_id_priv = container_of(id, struct rdma_id_private, id);
2205
2206         dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2207         memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2208                rdma_addr_size(cma_src_addr(id_priv)));
2209
2210         _cma_attach_to_dev(dev_id_priv, cma_dev);
2211         list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2212         atomic_inc(&id_priv->refcount);
2213         dev_id_priv->internal_id = 1;
2214         dev_id_priv->afonly = id_priv->afonly;
2215
2216         ret = rdma_listen(id, id_priv->backlog);
2217         if (ret)
2218                 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2219                         ret, cma_dev->device->name);
2220 }
2221
2222 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2223 {
2224         struct cma_device *cma_dev;
2225
2226         mutex_lock(&lock);
2227         list_add_tail(&id_priv->list, &listen_any_list);
2228         list_for_each_entry(cma_dev, &dev_list, list)
2229                 cma_listen_on_dev(id_priv, cma_dev);
2230         mutex_unlock(&lock);
2231 }
2232
2233 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2234 {
2235         struct rdma_id_private *id_priv;
2236
2237         id_priv = container_of(id, struct rdma_id_private, id);
2238         id_priv->tos = (u8) tos;
2239 }
2240 EXPORT_SYMBOL(rdma_set_service_type);
2241
2242 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2243                               void *context)
2244 {
2245         struct cma_work *work = context;
2246         struct rdma_route *route;
2247
2248         route = &work->id->id.route;
2249
2250         if (!status) {
2251                 route->num_paths = 1;
2252                 *route->path_rec = *path_rec;
2253         } else {
2254                 work->old_state = RDMA_CM_ROUTE_QUERY;
2255                 work->new_state = RDMA_CM_ADDR_RESOLVED;
2256                 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2257                 work->event.status = status;
2258         }
2259
2260         queue_work(cma_wq, &work->work);
2261 }
2262
2263 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2264                               struct cma_work *work)
2265 {
2266         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2267         struct ib_sa_path_rec path_rec;
2268         ib_sa_comp_mask comp_mask;
2269         struct sockaddr_in6 *sin6;
2270         struct sockaddr_ib *sib;
2271
2272         memset(&path_rec, 0, sizeof path_rec);
2273         rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2274         rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2275         path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2276         path_rec.numb_path = 1;
2277         path_rec.reversible = 1;
2278         path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2279
2280         comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2281                     IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2282                     IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2283
2284         switch (cma_family(id_priv)) {
2285         case AF_INET:
2286                 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2287                 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2288                 break;
2289         case AF_INET6:
2290                 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2291                 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2292                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2293                 break;
2294         case AF_IB:
2295                 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2296                 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2297                 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2298                 break;
2299         }
2300
2301         id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2302                                                id_priv->id.port_num, &path_rec,
2303                                                comp_mask, timeout_ms,
2304                                                GFP_KERNEL, cma_query_handler,
2305                                                work, &id_priv->query);
2306
2307         return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2308 }
2309
2310 static void cma_work_handler(struct work_struct *_work)
2311 {
2312         struct cma_work *work = container_of(_work, struct cma_work, work);
2313         struct rdma_id_private *id_priv = work->id;
2314         int destroy = 0;
2315
2316         mutex_lock(&id_priv->handler_mutex);
2317         if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2318                 goto out;
2319
2320         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2321                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2322                 destroy = 1;
2323         }
2324 out:
2325         mutex_unlock(&id_priv->handler_mutex);
2326         cma_deref_id(id_priv);
2327         if (destroy)
2328                 rdma_destroy_id(&id_priv->id);
2329         kfree(work);
2330 }
2331
2332 static void cma_ndev_work_handler(struct work_struct *_work)
2333 {
2334         struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2335         struct rdma_id_private *id_priv = work->id;
2336         int destroy = 0;
2337
2338         mutex_lock(&id_priv->handler_mutex);
2339         if (id_priv->state == RDMA_CM_DESTROYING ||
2340             id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2341                 goto out;
2342
2343         if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2344                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2345                 destroy = 1;
2346         }
2347
2348 out:
2349         mutex_unlock(&id_priv->handler_mutex);
2350         cma_deref_id(id_priv);
2351         if (destroy)
2352                 rdma_destroy_id(&id_priv->id);
2353         kfree(work);
2354 }
2355
2356 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2357 {
2358         struct rdma_route *route = &id_priv->id.route;
2359         struct cma_work *work;
2360         int ret;
2361
2362         work = kzalloc(sizeof *work, GFP_KERNEL);
2363         if (!work)
2364                 return -ENOMEM;
2365
2366         work->id = id_priv;
2367         INIT_WORK(&work->work, cma_work_handler);
2368         work->old_state = RDMA_CM_ROUTE_QUERY;
2369         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2370         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2371
2372         route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2373         if (!route->path_rec) {
2374                 ret = -ENOMEM;
2375                 goto err1;
2376         }
2377
2378         ret = cma_query_ib_route(id_priv, timeout_ms, work);
2379         if (ret)
2380                 goto err2;
2381
2382         return 0;
2383 err2:
2384         kfree(route->path_rec);
2385         route->path_rec = NULL;
2386 err1:
2387         kfree(work);
2388         return ret;
2389 }
2390
2391 int rdma_set_ib_paths(struct rdma_cm_id *id,
2392                       struct ib_sa_path_rec *path_rec, int num_paths)
2393 {
2394         struct rdma_id_private *id_priv;
2395         int ret;
2396
2397         id_priv = container_of(id, struct rdma_id_private, id);
2398         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2399                            RDMA_CM_ROUTE_RESOLVED))
2400                 return -EINVAL;
2401
2402         id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2403                                      GFP_KERNEL);
2404         if (!id->route.path_rec) {
2405                 ret = -ENOMEM;
2406                 goto err;
2407         }
2408
2409         id->route.num_paths = num_paths;
2410         return 0;
2411 err:
2412         cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2413         return ret;
2414 }
2415 EXPORT_SYMBOL(rdma_set_ib_paths);
2416
2417 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2418 {
2419         struct cma_work *work;
2420
2421         work = kzalloc(sizeof *work, GFP_KERNEL);
2422         if (!work)
2423                 return -ENOMEM;
2424
2425         work->id = id_priv;
2426         INIT_WORK(&work->work, cma_work_handler);
2427         work->old_state = RDMA_CM_ROUTE_QUERY;
2428         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2429         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2430         queue_work(cma_wq, &work->work);
2431         return 0;
2432 }
2433
2434 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2435 {
2436         int prio;
2437         struct net_device *dev;
2438
2439         prio = rt_tos2priority(tos);
2440         dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2441                 vlan_dev_real_dev(ndev) : ndev;
2442
2443         if (dev->num_tc)
2444                 return netdev_get_prio_tc_map(dev, prio);
2445
2446 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2447         if (ndev->priv_flags & IFF_802_1Q_VLAN)
2448                 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2449                         VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2450 #endif
2451         return 0;
2452 }
2453
2454 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2455                                            unsigned long supported_gids,
2456                                            enum ib_gid_type default_gid)
2457 {
2458         if ((network_type == RDMA_NETWORK_IPV4 ||
2459              network_type == RDMA_NETWORK_IPV6) &&
2460             test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2461                 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2462
2463         return default_gid;
2464 }
2465
2466 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2467 {
2468         struct rdma_route *route = &id_priv->id.route;
2469         struct rdma_addr *addr = &route->addr;
2470         struct cma_work *work;
2471         int ret;
2472         struct net_device *ndev = NULL;
2473
2474
2475         work = kzalloc(sizeof *work, GFP_KERNEL);
2476         if (!work)
2477                 return -ENOMEM;
2478
2479         work->id = id_priv;
2480         INIT_WORK(&work->work, cma_work_handler);
2481
2482         route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2483         if (!route->path_rec) {
2484                 ret = -ENOMEM;
2485                 goto err1;
2486         }
2487
2488         route->num_paths = 1;
2489
2490         if (addr->dev_addr.bound_dev_if) {
2491                 unsigned long supported_gids;
2492
2493                 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2494                 if (!ndev) {
2495                         ret = -ENODEV;
2496                         goto err2;
2497                 }
2498
2499                 if (ndev->flags & IFF_LOOPBACK) {
2500                         dev_put(ndev);
2501                         if (!id_priv->id.device->get_netdev) {
2502                                 ret = -EOPNOTSUPP;
2503                                 goto err2;
2504                         }
2505
2506                         ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2507                                                               id_priv->id.port_num);
2508                         if (!ndev) {
2509                                 ret = -ENODEV;
2510                                 goto err2;
2511                         }
2512                 }
2513
2514                 route->path_rec->net = &init_net;
2515                 route->path_rec->ifindex = ndev->ifindex;
2516                 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2517                                                             id_priv->id.port_num);
2518                 route->path_rec->gid_type =
2519                         cma_route_gid_type(addr->dev_addr.network,
2520                                            supported_gids,
2521                                            id_priv->gid_type);
2522         }
2523         if (!ndev) {
2524                 ret = -ENODEV;
2525                 goto err2;
2526         }
2527
2528         memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2529
2530         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2531                     &route->path_rec->sgid);
2532         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2533                     &route->path_rec->dgid);
2534
2535         /* Use the hint from IP Stack to select GID Type */
2536         if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2537                 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2538         if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2539                 /* TODO: get the hoplimit from the inet/inet6 device */
2540                 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2541         else
2542                 route->path_rec->hop_limit = 1;
2543         route->path_rec->reversible = 1;
2544         route->path_rec->pkey = cpu_to_be16(0xffff);
2545         route->path_rec->mtu_selector = IB_SA_EQ;
2546         route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2547         route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2548         route->path_rec->rate_selector = IB_SA_EQ;
2549         route->path_rec->rate = iboe_get_rate(ndev);
2550         dev_put(ndev);
2551         route->path_rec->packet_life_time_selector = IB_SA_EQ;
2552         route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2553         if (!route->path_rec->mtu) {
2554                 ret = -EINVAL;
2555                 goto err2;
2556         }
2557
2558         work->old_state = RDMA_CM_ROUTE_QUERY;
2559         work->new_state = RDMA_CM_ROUTE_RESOLVED;
2560         work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2561         work->event.status = 0;
2562
2563         queue_work(cma_wq, &work->work);
2564
2565         return 0;
2566
2567 err2:
2568         kfree(route->path_rec);
2569         route->path_rec = NULL;
2570 err1:
2571         kfree(work);
2572         return ret;
2573 }
2574
2575 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2576 {
2577         struct rdma_id_private *id_priv;
2578         int ret;
2579
2580         id_priv = container_of(id, struct rdma_id_private, id);
2581         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2582                 return -EINVAL;
2583
2584         atomic_inc(&id_priv->refcount);
2585         if (rdma_cap_ib_sa(id->device, id->port_num))
2586                 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2587         else if (rdma_protocol_roce(id->device, id->port_num))
2588                 ret = cma_resolve_iboe_route(id_priv);
2589         else if (rdma_protocol_iwarp(id->device, id->port_num))
2590                 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2591         else
2592                 ret = -ENOSYS;
2593
2594         if (ret)
2595                 goto err;
2596
2597         return 0;
2598 err:
2599         cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2600         cma_deref_id(id_priv);
2601         return ret;
2602 }
2603 EXPORT_SYMBOL(rdma_resolve_route);
2604
2605 static void cma_set_loopback(struct sockaddr *addr)
2606 {
2607         switch (addr->sa_family) {
2608         case AF_INET:
2609                 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2610                 break;
2611         case AF_INET6:
2612                 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2613                               0, 0, 0, htonl(1));
2614                 break;
2615         default:
2616                 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2617                             0, 0, 0, htonl(1));
2618                 break;
2619         }
2620 }
2621
2622 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2623 {
2624         struct cma_device *cma_dev, *cur_dev;
2625         struct ib_port_attr port_attr;
2626         union ib_gid gid;
2627         u16 pkey;
2628         int ret;
2629         u8 p;
2630
2631         cma_dev = NULL;
2632         mutex_lock(&lock);
2633         list_for_each_entry(cur_dev, &dev_list, list) {
2634                 if (cma_family(id_priv) == AF_IB &&
2635                     !rdma_cap_ib_cm(cur_dev->device, 1))
2636                         continue;
2637
2638                 if (!cma_dev)
2639                         cma_dev = cur_dev;
2640
2641                 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2642                         if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2643                             port_attr.state == IB_PORT_ACTIVE) {
2644                                 cma_dev = cur_dev;
2645                                 goto port_found;
2646                         }
2647                 }
2648         }
2649
2650         if (!cma_dev) {
2651                 ret = -ENODEV;
2652                 goto out;
2653         }
2654
2655         p = 1;
2656
2657 port_found:
2658         ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2659         if (ret)
2660                 goto out;
2661
2662         ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2663         if (ret)
2664                 goto out;
2665
2666         id_priv->id.route.addr.dev_addr.dev_type =
2667                 (rdma_protocol_ib(cma_dev->device, p)) ?
2668                 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2669
2670         rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2671         ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2672         id_priv->id.port_num = p;
2673         cma_attach_to_dev(id_priv, cma_dev);
2674         cma_set_loopback(cma_src_addr(id_priv));
2675 out:
2676         mutex_unlock(&lock);
2677         return ret;
2678 }
2679
2680 static void addr_handler(int status, struct sockaddr *src_addr,
2681                          struct rdma_dev_addr *dev_addr, void *context)
2682 {
2683         struct rdma_id_private *id_priv = context;
2684         struct rdma_cm_event event;
2685
2686         memset(&event, 0, sizeof event);
2687         mutex_lock(&id_priv->handler_mutex);
2688         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2689                            RDMA_CM_ADDR_RESOLVED))
2690                 goto out;
2691
2692         memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2693         if (!status && !id_priv->cma_dev)
2694                 status = cma_acquire_dev(id_priv, NULL);
2695
2696         if (status) {
2697                 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2698                                    RDMA_CM_ADDR_BOUND))
2699                         goto out;
2700                 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2701                 event.status = status;
2702         } else
2703                 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2704
2705         if (id_priv->id.event_handler(&id_priv->id, &event)) {
2706                 cma_exch(id_priv, RDMA_CM_DESTROYING);
2707                 mutex_unlock(&id_priv->handler_mutex);
2708                 cma_deref_id(id_priv);
2709                 rdma_destroy_id(&id_priv->id);
2710                 return;
2711         }
2712 out:
2713         mutex_unlock(&id_priv->handler_mutex);
2714         cma_deref_id(id_priv);
2715 }
2716
2717 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2718 {
2719         struct cma_work *work;
2720         union ib_gid gid;
2721         int ret;
2722
2723         work = kzalloc(sizeof *work, GFP_KERNEL);
2724         if (!work)
2725                 return -ENOMEM;
2726
2727         if (!id_priv->cma_dev) {
2728                 ret = cma_bind_loopback(id_priv);
2729                 if (ret)
2730                         goto err;
2731         }
2732
2733         rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2734         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2735
2736         work->id = id_priv;
2737         INIT_WORK(&work->work, cma_work_handler);
2738         work->old_state = RDMA_CM_ADDR_QUERY;
2739         work->new_state = RDMA_CM_ADDR_RESOLVED;
2740         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2741         queue_work(cma_wq, &work->work);
2742         return 0;
2743 err:
2744         kfree(work);
2745         return ret;
2746 }
2747
2748 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2749 {
2750         struct cma_work *work;
2751         int ret;
2752
2753         work = kzalloc(sizeof *work, GFP_KERNEL);
2754         if (!work)
2755                 return -ENOMEM;
2756
2757         if (!id_priv->cma_dev) {
2758                 ret = cma_resolve_ib_dev(id_priv);
2759                 if (ret)
2760                         goto err;
2761         }
2762
2763         rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2764                 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2765
2766         work->id = id_priv;
2767         INIT_WORK(&work->work, cma_work_handler);
2768         work->old_state = RDMA_CM_ADDR_QUERY;
2769         work->new_state = RDMA_CM_ADDR_RESOLVED;
2770         work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2771         queue_work(cma_wq, &work->work);
2772         return 0;
2773 err:
2774         kfree(work);
2775         return ret;
2776 }
2777
2778 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2779                          struct sockaddr *dst_addr)
2780 {
2781         if (!src_addr || !src_addr->sa_family) {
2782                 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2783                 src_addr->sa_family = dst_addr->sa_family;
2784                 if (IS_ENABLED(CONFIG_IPV6) &&
2785                     dst_addr->sa_family == AF_INET6) {
2786                         struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2787                         struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2788                         src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2789                         if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2790                                 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2791                 } else if (dst_addr->sa_family == AF_IB) {
2792                         ((struct sockaddr_ib *) src_addr)->sib_pkey =
2793                                 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2794                 }
2795         }
2796         return rdma_bind_addr(id, src_addr);
2797 }
2798
2799 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2800                       struct sockaddr *dst_addr, int timeout_ms)
2801 {
2802         struct rdma_id_private *id_priv;
2803         int ret;
2804
2805         id_priv = container_of(id, struct rdma_id_private, id);
2806         if (id_priv->state == RDMA_CM_IDLE) {
2807                 ret = cma_bind_addr(id, src_addr, dst_addr);
2808                 if (ret)
2809                         return ret;
2810         }
2811
2812         if (cma_family(id_priv) != dst_addr->sa_family)
2813                 return -EINVAL;
2814
2815         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2816                 return -EINVAL;
2817
2818         atomic_inc(&id_priv->refcount);
2819         memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2820         if (cma_any_addr(dst_addr)) {
2821                 ret = cma_resolve_loopback(id_priv);
2822         } else {
2823                 if (dst_addr->sa_family == AF_IB) {
2824                         ret = cma_resolve_ib_addr(id_priv);
2825                 } else {
2826                         ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2827                                               dst_addr, &id->route.addr.dev_addr,
2828                                               timeout_ms, addr_handler, id_priv);
2829                 }
2830         }
2831         if (ret)
2832                 goto err;
2833
2834         return 0;
2835 err:
2836         cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2837         cma_deref_id(id_priv);
2838         return ret;
2839 }
2840 EXPORT_SYMBOL(rdma_resolve_addr);
2841
2842 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2843 {
2844         struct rdma_id_private *id_priv;
2845         unsigned long flags;
2846         int ret;
2847
2848         id_priv = container_of(id, struct rdma_id_private, id);
2849         spin_lock_irqsave(&id_priv->lock, flags);
2850         if (reuse || id_priv->state == RDMA_CM_IDLE) {
2851                 id_priv->reuseaddr = reuse;
2852                 ret = 0;
2853         } else {
2854                 ret = -EINVAL;
2855         }
2856         spin_unlock_irqrestore(&id_priv->lock, flags);
2857         return ret;
2858 }
2859 EXPORT_SYMBOL(rdma_set_reuseaddr);
2860
2861 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2862 {
2863         struct rdma_id_private *id_priv;
2864         unsigned long flags;
2865         int ret;
2866
2867         id_priv = container_of(id, struct rdma_id_private, id);
2868         spin_lock_irqsave(&id_priv->lock, flags);
2869         if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2870                 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2871                 id_priv->afonly = afonly;
2872                 ret = 0;
2873         } else {
2874                 ret = -EINVAL;
2875         }
2876         spin_unlock_irqrestore(&id_priv->lock, flags);
2877         return ret;
2878 }
2879 EXPORT_SYMBOL(rdma_set_afonly);
2880
2881 static void cma_bind_port(struct rdma_bind_list *bind_list,
2882                           struct rdma_id_private *id_priv)
2883 {
2884         struct sockaddr *addr;
2885         struct sockaddr_ib *sib;
2886         u64 sid, mask;
2887         __be16 port;
2888
2889         addr = cma_src_addr(id_priv);
2890         port = htons(bind_list->port);
2891
2892         switch (addr->sa_family) {
2893         case AF_INET:
2894                 ((struct sockaddr_in *) addr)->sin_port = port;
2895                 break;
2896         case AF_INET6:
2897                 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2898                 break;
2899         case AF_IB:
2900                 sib = (struct sockaddr_ib *) addr;
2901                 sid = be64_to_cpu(sib->sib_sid);
2902                 mask = be64_to_cpu(sib->sib_sid_mask);
2903                 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2904                 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2905                 break;
2906         }
2907         id_priv->bind_list = bind_list;
2908         hlist_add_head(&id_priv->node, &bind_list->owners);
2909 }
2910
2911 static int cma_alloc_port(enum rdma_port_space ps,
2912                           struct rdma_id_private *id_priv, unsigned short snum)
2913 {
2914         struct rdma_bind_list *bind_list;
2915         int ret;
2916
2917         bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2918         if (!bind_list)
2919                 return -ENOMEM;
2920
2921         ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2922                            snum);
2923         if (ret < 0)
2924                 goto err;
2925
2926         bind_list->ps = ps;
2927         bind_list->port = (unsigned short)ret;
2928         cma_bind_port(bind_list, id_priv);
2929         return 0;
2930 err:
2931         kfree(bind_list);
2932         return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2933 }
2934
2935 static int cma_alloc_any_port(enum rdma_port_space ps,
2936                               struct rdma_id_private *id_priv)
2937 {
2938         static unsigned int last_used_port;
2939         int low, high, remaining;
2940         unsigned int rover;
2941         struct net *net = id_priv->id.route.addr.dev_addr.net;
2942
2943         inet_get_local_port_range(net, &low, &high);
2944         remaining = (high - low) + 1;
2945         rover = prandom_u32() % remaining + low;
2946 retry:
2947         if (last_used_port != rover &&
2948             !cma_ps_find(net, ps, (unsigned short)rover)) {
2949                 int ret = cma_alloc_port(ps, id_priv, rover);
2950                 /*
2951                  * Remember previously used port number in order to avoid
2952                  * re-using same port immediately after it is closed.
2953                  */
2954                 if (!ret)
2955                         last_used_port = rover;
2956                 if (ret != -EADDRNOTAVAIL)
2957                         return ret;
2958         }
2959         if (--remaining) {
2960                 rover++;
2961                 if ((rover < low) || (rover > high))
2962                         rover = low;
2963                 goto retry;
2964         }
2965         return -EADDRNOTAVAIL;
2966 }
2967
2968 /*
2969  * Check that the requested port is available.  This is called when trying to
2970  * bind to a specific port, or when trying to listen on a bound port.  In
2971  * the latter case, the provided id_priv may already be on the bind_list, but
2972  * we still need to check that it's okay to start listening.
2973  */
2974 static int cma_check_port(struct rdma_bind_list *bind_list,
2975                           struct rdma_id_private *id_priv, uint8_t reuseaddr)
2976 {
2977         struct rdma_id_private *cur_id;
2978         struct sockaddr *addr, *cur_addr;
2979
2980         addr = cma_src_addr(id_priv);
2981         hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2982                 if (id_priv == cur_id)
2983                         continue;
2984
2985                 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2986                     cur_id->reuseaddr)
2987                         continue;
2988
2989                 cur_addr = cma_src_addr(cur_id);
2990                 if (id_priv->afonly && cur_id->afonly &&
2991                     (addr->sa_family != cur_addr->sa_family))
2992                         continue;
2993
2994                 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2995                         return -EADDRNOTAVAIL;
2996
2997                 if (!cma_addr_cmp(addr, cur_addr))
2998                         return -EADDRINUSE;
2999         }
3000         return 0;
3001 }
3002
3003 static int cma_use_port(enum rdma_port_space ps,
3004                         struct rdma_id_private *id_priv)
3005 {
3006         struct rdma_bind_list *bind_list;
3007         unsigned short snum;
3008         int ret;
3009
3010         snum = ntohs(cma_port(cma_src_addr(id_priv)));
3011         if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3012                 return -EACCES;
3013
3014         bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3015         if (!bind_list) {
3016                 ret = cma_alloc_port(ps, id_priv, snum);
3017         } else {
3018                 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3019                 if (!ret)
3020                         cma_bind_port(bind_list, id_priv);
3021         }
3022         return ret;
3023 }
3024
3025 static int cma_bind_listen(struct rdma_id_private *id_priv)
3026 {
3027         struct rdma_bind_list *bind_list = id_priv->bind_list;
3028         int ret = 0;
3029
3030         mutex_lock(&lock);
3031         if (bind_list->owners.first->next)
3032                 ret = cma_check_port(bind_list, id_priv, 0);
3033         mutex_unlock(&lock);
3034         return ret;
3035 }
3036
3037 static enum rdma_port_space cma_select_inet_ps(
3038                 struct rdma_id_private *id_priv)
3039 {
3040         switch (id_priv->id.ps) {
3041         case RDMA_PS_TCP:
3042         case RDMA_PS_UDP:
3043         case RDMA_PS_IPOIB:
3044         case RDMA_PS_IB:
3045                 return id_priv->id.ps;
3046         default:
3047
3048                 return 0;
3049         }
3050 }
3051
3052 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3053 {
3054         enum rdma_port_space ps = 0;
3055         struct sockaddr_ib *sib;
3056         u64 sid_ps, mask, sid;
3057
3058         sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3059         mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3060         sid = be64_to_cpu(sib->sib_sid) & mask;
3061
3062         if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3063                 sid_ps = RDMA_IB_IP_PS_IB;
3064                 ps = RDMA_PS_IB;
3065         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3066                    (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3067                 sid_ps = RDMA_IB_IP_PS_TCP;
3068                 ps = RDMA_PS_TCP;
3069         } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3070                    (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3071                 sid_ps = RDMA_IB_IP_PS_UDP;
3072                 ps = RDMA_PS_UDP;
3073         }
3074
3075         if (ps) {
3076                 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3077                 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3078                                                 be64_to_cpu(sib->sib_sid_mask));
3079         }
3080         return ps;
3081 }
3082
3083 static int cma_get_port(struct rdma_id_private *id_priv)
3084 {
3085         enum rdma_port_space ps;
3086         int ret;
3087
3088         if (cma_family(id_priv) != AF_IB)
3089                 ps = cma_select_inet_ps(id_priv);
3090         else
3091                 ps = cma_select_ib_ps(id_priv);
3092         if (!ps)
3093                 return -EPROTONOSUPPORT;
3094
3095         mutex_lock(&lock);
3096         if (cma_any_port(cma_src_addr(id_priv)))
3097                 ret = cma_alloc_any_port(ps, id_priv);
3098         else
3099                 ret = cma_use_port(ps, id_priv);
3100         mutex_unlock(&lock);
3101
3102         return ret;
3103 }
3104
3105 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3106                                struct sockaddr *addr)
3107 {
3108 #if IS_ENABLED(CONFIG_IPV6)
3109         struct sockaddr_in6 *sin6;
3110
3111         if (addr->sa_family != AF_INET6)
3112                 return 0;
3113
3114         sin6 = (struct sockaddr_in6 *) addr;
3115
3116         if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3117                 return 0;
3118
3119         if (!sin6->sin6_scope_id)
3120                         return -EINVAL;
3121
3122         dev_addr->bound_dev_if = sin6->sin6_scope_id;
3123 #endif
3124         return 0;
3125 }
3126
3127 int rdma_listen(struct rdma_cm_id *id, int backlog)
3128 {
3129         struct rdma_id_private *id_priv;
3130         int ret;
3131
3132         id_priv = container_of(id, struct rdma_id_private, id);
3133         if (id_priv->state == RDMA_CM_IDLE) {
3134                 id->route.addr.src_addr.ss_family = AF_INET;
3135                 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3136                 if (ret)
3137                         return ret;
3138         }
3139
3140         if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3141                 return -EINVAL;
3142
3143         if (id_priv->reuseaddr) {
3144                 ret = cma_bind_listen(id_priv);
3145                 if (ret)
3146                         goto err;
3147         }
3148
3149         id_priv->backlog = backlog;
3150         if (id->device) {
3151                 if (rdma_cap_ib_cm(id->device, 1)) {
3152                         ret = cma_ib_listen(id_priv);
3153                         if (ret)
3154                                 goto err;
3155                 } else if (rdma_cap_iw_cm(id->device, 1)) {
3156                         ret = cma_iw_listen(id_priv, backlog);
3157                         if (ret)
3158                                 goto err;
3159                 } else {
3160                         ret = -ENOSYS;
3161                         goto err;
3162                 }
3163         } else
3164                 cma_listen_on_all(id_priv);
3165
3166         return 0;
3167 err:
3168         id_priv->backlog = 0;
3169         cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3170         return ret;
3171 }
3172 EXPORT_SYMBOL(rdma_listen);
3173
3174 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3175 {
3176         struct rdma_id_private *id_priv;
3177         int ret;
3178
3179         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3180             addr->sa_family != AF_IB)
3181                 return -EAFNOSUPPORT;
3182
3183         id_priv = container_of(id, struct rdma_id_private, id);
3184         if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3185                 return -EINVAL;
3186
3187         ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3188         if (ret)
3189                 goto err1;
3190
3191         memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3192         if (!cma_any_addr(addr)) {
3193                 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3194                 if (ret)
3195                         goto err1;
3196
3197                 ret = cma_acquire_dev(id_priv, NULL);
3198                 if (ret)
3199                         goto err1;
3200         }
3201
3202         if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3203                 if (addr->sa_family == AF_INET)
3204                         id_priv->afonly = 1;
3205 #if IS_ENABLED(CONFIG_IPV6)
3206                 else if (addr->sa_family == AF_INET6) {
3207                         struct net *net = id_priv->id.route.addr.dev_addr.net;
3208
3209                         id_priv->afonly = net->ipv6.sysctl.bindv6only;
3210                 }
3211 #endif
3212         }
3213         ret = cma_get_port(id_priv);
3214         if (ret)
3215                 goto err2;
3216
3217         return 0;
3218 err2:
3219         if (id_priv->cma_dev)
3220                 cma_release_dev(id_priv);
3221 err1:
3222         cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3223         return ret;
3224 }
3225 EXPORT_SYMBOL(rdma_bind_addr);
3226
3227 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3228 {
3229         struct cma_hdr *cma_hdr;
3230
3231         cma_hdr = hdr;
3232         cma_hdr->cma_version = CMA_VERSION;
3233         if (cma_family(id_priv) == AF_INET) {
3234                 struct sockaddr_in *src4, *dst4;
3235
3236                 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3237                 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3238
3239                 cma_set_ip_ver(cma_hdr, 4);
3240                 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3241                 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3242                 cma_hdr->port = src4->sin_port;
3243         } else if (cma_family(id_priv) == AF_INET6) {
3244                 struct sockaddr_in6 *src6, *dst6;
3245
3246                 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3247                 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3248
3249                 cma_set_ip_ver(cma_hdr, 6);
3250                 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3251                 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3252                 cma_hdr->port = src6->sin6_port;
3253         }
3254         return 0;
3255 }
3256
3257 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3258                                 struct ib_cm_event *ib_event)
3259 {
3260         struct rdma_id_private *id_priv = cm_id->context;
3261         struct rdma_cm_event event;
3262         struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3263         int ret = 0;
3264
3265         mutex_lock(&id_priv->handler_mutex);
3266         if (id_priv->state != RDMA_CM_CONNECT)
3267                 goto out;
3268
3269         memset(&event, 0, sizeof event);
3270         switch (ib_event->event) {
3271         case IB_CM_SIDR_REQ_ERROR:
3272                 event.event = RDMA_CM_EVENT_UNREACHABLE;
3273                 event.status = -ETIMEDOUT;
3274                 break;
3275         case IB_CM_SIDR_REP_RECEIVED:
3276                 event.param.ud.private_data = ib_event->private_data;
3277                 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3278                 if (rep->status != IB_SIDR_SUCCESS) {
3279                         event.event = RDMA_CM_EVENT_UNREACHABLE;
3280                         event.status = ib_event->param.sidr_rep_rcvd.status;
3281                         break;
3282                 }
3283                 ret = cma_set_qkey(id_priv, rep->qkey);
3284                 if (ret) {
3285                         event.event = RDMA_CM_EVENT_ADDR_ERROR;
3286                         event.status = ret;
3287                         break;
3288                 }
3289                 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3290                                      id_priv->id.route.path_rec,
3291                                      &event.param.ud.ah_attr);
3292                 event.param.ud.qp_num = rep->qpn;
3293                 event.param.ud.qkey = rep->qkey;
3294                 event.event = RDMA_CM_EVENT_ESTABLISHED;
3295                 event.status = 0;
3296                 break;
3297         default:
3298                 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3299                        ib_event->event);
3300                 goto out;
3301         }
3302
3303         ret = id_priv->id.event_handler(&id_priv->id, &event);
3304         if (ret) {
3305                 /* Destroy the CM ID by returning a non-zero value. */
3306                 id_priv->cm_id.ib = NULL;
3307                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3308                 mutex_unlock(&id_priv->handler_mutex);
3309                 rdma_destroy_id(&id_priv->id);
3310                 return ret;
3311         }
3312 out:
3313         mutex_unlock(&id_priv->handler_mutex);
3314         return ret;
3315 }
3316
3317 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3318                               struct rdma_conn_param *conn_param)
3319 {
3320         struct ib_cm_sidr_req_param req;
3321         struct ib_cm_id *id;
3322         void *private_data;
3323         u8 offset;
3324         int ret;
3325
3326         memset(&req, 0, sizeof req);
3327         offset = cma_user_data_offset(id_priv);
3328         req.private_data_len = offset + conn_param->private_data_len;
3329         if (req.private_data_len < conn_param->private_data_len)
3330                 return -EINVAL;
3331
3332         if (req.private_data_len) {
3333                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3334                 if (!private_data)
3335                         return -ENOMEM;
3336         } else {
3337                 private_data = NULL;
3338         }
3339
3340         if (conn_param->private_data && conn_param->private_data_len)
3341                 memcpy(private_data + offset, conn_param->private_data,
3342                        conn_param->private_data_len);
3343
3344         if (private_data) {
3345                 ret = cma_format_hdr(private_data, id_priv);
3346                 if (ret)
3347                         goto out;
3348                 req.private_data = private_data;
3349         }
3350
3351         id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3352                              id_priv);
3353         if (IS_ERR(id)) {
3354                 ret = PTR_ERR(id);
3355                 goto out;
3356         }
3357         id_priv->cm_id.ib = id;
3358
3359         req.path = id_priv->id.route.path_rec;
3360         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3361         req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3362         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3363
3364         ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3365         if (ret) {
3366                 ib_destroy_cm_id(id_priv->cm_id.ib);
3367                 id_priv->cm_id.ib = NULL;
3368         }
3369 out:
3370         kfree(private_data);
3371         return ret;
3372 }
3373
3374 static int cma_connect_ib(struct rdma_id_private *id_priv,
3375                           struct rdma_conn_param *conn_param)
3376 {
3377         struct ib_cm_req_param req;
3378         struct rdma_route *route;
3379         void *private_data;
3380         struct ib_cm_id *id;
3381         u8 offset;
3382         int ret;
3383
3384         memset(&req, 0, sizeof req);
3385         offset = cma_user_data_offset(id_priv);
3386         req.private_data_len = offset + conn_param->private_data_len;
3387         if (req.private_data_len < conn_param->private_data_len)
3388                 return -EINVAL;
3389
3390         if (req.private_data_len) {
3391                 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3392                 if (!private_data)
3393                         return -ENOMEM;
3394         } else {
3395                 private_data = NULL;
3396         }
3397
3398         if (conn_param->private_data && conn_param->private_data_len)
3399                 memcpy(private_data + offset, conn_param->private_data,
3400                        conn_param->private_data_len);
3401
3402         id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3403         if (IS_ERR(id)) {
3404                 ret = PTR_ERR(id);
3405                 goto out;
3406         }
3407         id_priv->cm_id.ib = id;
3408
3409         route = &id_priv->id.route;
3410         if (private_data) {
3411                 ret = cma_format_hdr(private_data, id_priv);
3412                 if (ret)
3413                         goto out;
3414                 req.private_data = private_data;
3415         }
3416
3417         req.primary_path = &route->path_rec[0];
3418         if (route->num_paths == 2)
3419                 req.alternate_path = &route->path_rec[1];
3420
3421         req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3422         req.qp_num = id_priv->qp_num;
3423         req.qp_type = id_priv->id.qp_type;
3424         req.starting_psn = id_priv->seq_num;
3425         req.responder_resources = conn_param->responder_resources;
3426         req.initiator_depth = conn_param->initiator_depth;
3427         req.flow_control = conn_param->flow_control;
3428         req.retry_count = min_t(u8, 7, conn_param->retry_count);
3429         req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3430         req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3431         req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3432         req.max_cm_retries = CMA_MAX_CM_RETRIES;
3433         req.srq = id_priv->srq ? 1 : 0;
3434
3435         ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3436 out:
3437         if (ret && !IS_ERR(id)) {
3438                 ib_destroy_cm_id(id);
3439                 id_priv->cm_id.ib = NULL;
3440         }
3441
3442         kfree(private_data);
3443         return ret;
3444 }
3445
3446 static int cma_connect_iw(struct rdma_id_private *id_priv,
3447                           struct rdma_conn_param *conn_param)
3448 {
3449         struct iw_cm_id *cm_id;
3450         int ret;
3451         struct iw_cm_conn_param iw_param;
3452
3453         cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3454         if (IS_ERR(cm_id))
3455                 return PTR_ERR(cm_id);
3456
3457         cm_id->tos = id_priv->tos;
3458         id_priv->cm_id.iw = cm_id;
3459
3460         memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3461                rdma_addr_size(cma_src_addr(id_priv)));
3462         memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3463                rdma_addr_size(cma_dst_addr(id_priv)));
3464
3465         ret = cma_modify_qp_rtr(id_priv, conn_param);
3466         if (ret)
3467                 goto out;
3468
3469         if (conn_param) {
3470                 iw_param.ord = conn_param->initiator_depth;
3471                 iw_param.ird = conn_param->responder_resources;
3472                 iw_param.private_data = conn_param->private_data;
3473                 iw_param.private_data_len = conn_param->private_data_len;
3474                 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3475         } else {
3476                 memset(&iw_param, 0, sizeof iw_param);
3477                 iw_param.qpn = id_priv->qp_num;
3478         }
3479         ret = iw_cm_connect(cm_id, &iw_param);
3480 out:
3481         if (ret) {
3482                 iw_destroy_cm_id(cm_id);
3483                 id_priv->cm_id.iw = NULL;
3484         }
3485         return ret;
3486 }
3487
3488 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3489 {
3490         struct rdma_id_private *id_priv;
3491         int ret;
3492
3493         id_priv = container_of(id, struct rdma_id_private, id);
3494         if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3495                 return -EINVAL;
3496
3497         if (!id->qp) {
3498                 id_priv->qp_num = conn_param->qp_num;
3499                 id_priv->srq = conn_param->srq;
3500         }
3501
3502         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3503                 if (id->qp_type == IB_QPT_UD)
3504                         ret = cma_resolve_ib_udp(id_priv, conn_param);
3505                 else
3506                         ret = cma_connect_ib(id_priv, conn_param);
3507         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3508                 ret = cma_connect_iw(id_priv, conn_param);
3509         else
3510                 ret = -ENOSYS;
3511         if (ret)
3512                 goto err;
3513
3514         return 0;
3515 err:
3516         cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3517         return ret;
3518 }
3519 EXPORT_SYMBOL(rdma_connect);
3520
3521 static int cma_accept_ib(struct rdma_id_private *id_priv,
3522                          struct rdma_conn_param *conn_param)
3523 {
3524         struct ib_cm_rep_param rep;
3525         int ret;
3526
3527         ret = cma_modify_qp_rtr(id_priv, conn_param);
3528         if (ret)
3529                 goto out;
3530
3531         ret = cma_modify_qp_rts(id_priv, conn_param);
3532         if (ret)
3533                 goto out;
3534
3535         memset(&rep, 0, sizeof rep);
3536         rep.qp_num = id_priv->qp_num;
3537         rep.starting_psn = id_priv->seq_num;
3538         rep.private_data = conn_param->private_data;
3539         rep.private_data_len = conn_param->private_data_len;
3540         rep.responder_resources = conn_param->responder_resources;
3541         rep.initiator_depth = conn_param->initiator_depth;
3542         rep.failover_accepted = 0;
3543         rep.flow_control = conn_param->flow_control;
3544         rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3545         rep.srq = id_priv->srq ? 1 : 0;
3546
3547         ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3548 out:
3549         return ret;
3550 }
3551
3552 static int cma_accept_iw(struct rdma_id_private *id_priv,
3553                   struct rdma_conn_param *conn_param)
3554 {
3555         struct iw_cm_conn_param iw_param;
3556         int ret;
3557
3558         if (!conn_param)
3559                 return -EINVAL;
3560
3561         ret = cma_modify_qp_rtr(id_priv, conn_param);
3562         if (ret)
3563                 return ret;
3564
3565         iw_param.ord = conn_param->initiator_depth;
3566         iw_param.ird = conn_param->responder_resources;
3567         iw_param.private_data = conn_param->private_data;
3568         iw_param.private_data_len = conn_param->private_data_len;
3569         if (id_priv->id.qp) {
3570                 iw_param.qpn = id_priv->qp_num;
3571         } else
3572                 iw_param.qpn = conn_param->qp_num;
3573
3574         return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3575 }
3576
3577 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3578                              enum ib_cm_sidr_status status, u32 qkey,
3579                              const void *private_data, int private_data_len)
3580 {
3581         struct ib_cm_sidr_rep_param rep;
3582         int ret;
3583
3584         memset(&rep, 0, sizeof rep);
3585         rep.status = status;
3586         if (status == IB_SIDR_SUCCESS) {
3587                 ret = cma_set_qkey(id_priv, qkey);
3588                 if (ret)
3589                         return ret;
3590                 rep.qp_num = id_priv->qp_num;
3591                 rep.qkey = id_priv->qkey;
3592         }
3593         rep.private_data = private_data;
3594         rep.private_data_len = private_data_len;
3595
3596         return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3597 }
3598
3599 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3600 {
3601         struct rdma_id_private *id_priv;
3602         int ret;
3603
3604         id_priv = container_of(id, struct rdma_id_private, id);
3605
3606         id_priv->owner = task_pid_nr(current);
3607
3608         if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3609                 return -EINVAL;
3610
3611         if (!id->qp && conn_param) {
3612                 id_priv->qp_num = conn_param->qp_num;
3613                 id_priv->srq = conn_param->srq;
3614         }
3615
3616         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3617                 if (id->qp_type == IB_QPT_UD) {
3618                         if (conn_param)
3619                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3620                                                         conn_param->qkey,
3621                                                         conn_param->private_data,
3622                                                         conn_param->private_data_len);
3623                         else
3624                                 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3625                                                         0, NULL, 0);
3626                 } else {
3627                         if (conn_param)
3628                                 ret = cma_accept_ib(id_priv, conn_param);
3629                         else
3630                                 ret = cma_rep_recv(id_priv);
3631                 }
3632         } else if (rdma_cap_iw_cm(id->device, id->port_num))
3633                 ret = cma_accept_iw(id_priv, conn_param);
3634         else
3635                 ret = -ENOSYS;
3636
3637         if (ret)
3638                 goto reject;
3639
3640         return 0;
3641 reject:
3642         cma_modify_qp_err(id_priv);
3643         rdma_reject(id, NULL, 0);
3644         return ret;
3645 }
3646 EXPORT_SYMBOL(rdma_accept);
3647
3648 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3649 {
3650         struct rdma_id_private *id_priv;
3651         int ret;
3652
3653         id_priv = container_of(id, struct rdma_id_private, id);
3654         if (!id_priv->cm_id.ib)
3655                 return -EINVAL;
3656
3657         switch (id->device->node_type) {
3658         case RDMA_NODE_IB_CA:
3659                 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3660                 break;
3661         default:
3662                 ret = 0;
3663                 break;
3664         }
3665         return ret;
3666 }
3667 EXPORT_SYMBOL(rdma_notify);
3668
3669 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3670                 u8 private_data_len)
3671 {
3672         struct rdma_id_private *id_priv;
3673         int ret;
3674
3675         id_priv = container_of(id, struct rdma_id_private, id);
3676         if (!id_priv->cm_id.ib)
3677                 return -EINVAL;
3678
3679         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3680                 if (id->qp_type == IB_QPT_UD)
3681                         ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3682                                                 private_data, private_data_len);
3683                 else
3684                         ret = ib_send_cm_rej(id_priv->cm_id.ib,
3685                                              IB_CM_REJ_CONSUMER_DEFINED, NULL,
3686                                              0, private_data, private_data_len);
3687         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3688                 ret = iw_cm_reject(id_priv->cm_id.iw,
3689                                    private_data, private_data_len);
3690         } else
3691                 ret = -ENOSYS;
3692
3693         return ret;
3694 }
3695 EXPORT_SYMBOL(rdma_reject);
3696
3697 int rdma_disconnect(struct rdma_cm_id *id)
3698 {
3699         struct rdma_id_private *id_priv;
3700         int ret;
3701
3702         id_priv = container_of(id, struct rdma_id_private, id);
3703         if (!id_priv->cm_id.ib)
3704                 return -EINVAL;
3705
3706         if (rdma_cap_ib_cm(id->device, id->port_num)) {
3707                 ret = cma_modify_qp_err(id_priv);
3708                 if (ret)
3709                         goto out;
3710                 /* Initiate or respond to a disconnect. */
3711                 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3712                         ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3713         } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3714                 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3715         } else
3716                 ret = -EINVAL;
3717
3718 out:
3719         return ret;
3720 }
3721 EXPORT_SYMBOL(rdma_disconnect);
3722
3723 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3724 {
3725         struct rdma_id_private *id_priv;
3726         struct cma_multicast *mc = multicast->context;
3727         struct rdma_cm_event event;
3728         int ret = 0;
3729
3730         id_priv = mc->id_priv;
3731         mutex_lock(&id_priv->handler_mutex);
3732         if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3733             id_priv->state != RDMA_CM_ADDR_RESOLVED)
3734                 goto out;
3735
3736         if (!status)
3737                 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3738         mutex_lock(&id_priv->qp_mutex);
3739         if (!status && id_priv->id.qp)
3740                 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3741                                          be16_to_cpu(multicast->rec.mlid));
3742         mutex_unlock(&id_priv->qp_mutex);
3743
3744         memset(&event, 0, sizeof event);
3745         event.status = status;
3746         event.param.ud.private_data = mc->context;
3747         if (!status) {
3748                 struct rdma_dev_addr *dev_addr =
3749                         &id_priv->id.route.addr.dev_addr;
3750                 struct net_device *ndev =
3751                         dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3752                 enum ib_gid_type gid_type =
3753                         id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3754                         rdma_start_port(id_priv->cma_dev->device)];
3755
3756                 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3757                 ib_init_ah_from_mcmember(id_priv->id.device,
3758                                          id_priv->id.port_num, &multicast->rec,
3759                                          ndev, gid_type,
3760                                          &event.param.ud.ah_attr);
3761                 event.param.ud.qp_num = 0xFFFFFF;
3762                 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3763                 if (ndev)
3764                         dev_put(ndev);
3765         } else
3766                 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3767
3768         ret = id_priv->id.event_handler(&id_priv->id, &event);
3769         if (ret) {
3770                 cma_exch(id_priv, RDMA_CM_DESTROYING);
3771                 mutex_unlock(&id_priv->handler_mutex);
3772                 rdma_destroy_id(&id_priv->id);
3773                 return 0;
3774         }
3775
3776 out:
3777         mutex_unlock(&id_priv->handler_mutex);
3778         return 0;
3779 }
3780
3781 static void cma_set_mgid(struct rdma_id_private *id_priv,
3782                          struct sockaddr *addr, union ib_gid *mgid)
3783 {
3784         unsigned char mc_map[MAX_ADDR_LEN];
3785         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3786         struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3787         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3788
3789         if (cma_any_addr(addr)) {
3790                 memset(mgid, 0, sizeof *mgid);
3791         } else if ((addr->sa_family == AF_INET6) &&
3792                    ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3793                                                                  0xFF10A01B)) {
3794                 /* IPv6 address is an SA assigned MGID. */
3795                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3796         } else if (addr->sa_family == AF_IB) {
3797                 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3798         } else if ((addr->sa_family == AF_INET6)) {
3799                 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3800                 if (id_priv->id.ps == RDMA_PS_UDP)
3801                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3802                 *mgid = *(union ib_gid *) (mc_map + 4);
3803         } else {
3804                 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3805                 if (id_priv->id.ps == RDMA_PS_UDP)
3806                         mc_map[7] = 0x01;       /* Use RDMA CM signature */
3807                 *mgid = *(union ib_gid *) (mc_map + 4);
3808         }
3809 }
3810
3811 static void cma_query_sa_classport_info_cb(int status,
3812                                            struct ib_class_port_info *rec,
3813                                            void *context)
3814 {
3815         struct class_port_info_context *cb_ctx = context;
3816
3817         WARN_ON(!context);
3818
3819         if (status || !rec) {
3820                 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n",
3821                          cb_ctx->device->name, cb_ctx->port_num, status);
3822                 goto out;
3823         }
3824
3825         memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info));
3826
3827 out:
3828         complete(&cb_ctx->done);
3829 }
3830
3831 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num,
3832                                        struct ib_class_port_info *class_port_info)
3833 {
3834         struct class_port_info_context *cb_ctx;
3835         int ret;
3836
3837         cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL);
3838         if (!cb_ctx)
3839                 return -ENOMEM;
3840
3841         cb_ctx->device = device;
3842         cb_ctx->class_port_info = class_port_info;
3843         cb_ctx->port_num = port_num;
3844         init_completion(&cb_ctx->done);
3845
3846         ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num,
3847                                              CMA_QUERY_CLASSPORT_INFO_TIMEOUT,
3848                                              GFP_KERNEL, cma_query_sa_classport_info_cb,
3849                                              cb_ctx, &cb_ctx->sa_query);
3850         if (ret < 0) {
3851                 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n",
3852                        device->name, port_num, ret);
3853                 goto out;
3854         }
3855
3856         wait_for_completion(&cb_ctx->done);
3857
3858 out:
3859         kfree(cb_ctx);
3860         return ret;
3861 }
3862
3863 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3864                                  struct cma_multicast *mc)
3865 {
3866         struct ib_sa_mcmember_rec rec;
3867         struct ib_class_port_info class_port_info;
3868         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3869         ib_sa_comp_mask comp_mask;
3870         int ret;
3871
3872         ib_addr_get_mgid(dev_addr, &rec.mgid);
3873         ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3874                                      &rec.mgid, &rec);
3875         if (ret)
3876                 return ret;
3877
3878         ret = cma_set_qkey(id_priv, 0);
3879         if (ret)
3880                 return ret;
3881
3882         cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3883         rec.qkey = cpu_to_be32(id_priv->qkey);
3884         rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3885         rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3886         rec.join_state = mc->join_state;
3887
3888         if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) {
3889                 ret = cma_query_sa_classport_info(id_priv->id.device,
3890                                                   id_priv->id.port_num,
3891                                                   &class_port_info);
3892
3893                 if (ret)
3894                         return ret;
3895
3896                 if (!(ib_get_cpi_capmask2(&class_port_info) &
3897                       IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) {
3898                         pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3899                                 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3900                                 id_priv->id.device->name, id_priv->id.port_num);
3901                         return -EOPNOTSUPP;
3902                 }
3903         }
3904
3905         comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3906                     IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3907                     IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3908                     IB_SA_MCMEMBER_REC_FLOW_LABEL |
3909                     IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3910
3911         if (id_priv->id.ps == RDMA_PS_IPOIB)
3912                 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3913                              IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3914                              IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3915                              IB_SA_MCMEMBER_REC_MTU |
3916                              IB_SA_MCMEMBER_REC_HOP_LIMIT;
3917
3918         mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3919                                                 id_priv->id.port_num, &rec,
3920                                                 comp_mask, GFP_KERNEL,
3921                                                 cma_ib_mc_handler, mc);
3922         return PTR_ERR_OR_ZERO(mc->multicast.ib);
3923 }
3924
3925 static void iboe_mcast_work_handler(struct work_struct *work)
3926 {
3927         struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3928         struct cma_multicast *mc = mw->mc;
3929         struct ib_sa_multicast *m = mc->multicast.ib;
3930
3931         mc->multicast.ib->context = mc;
3932         cma_ib_mc_handler(0, m);
3933         kref_put(&mc->mcref, release_mc);
3934         kfree(mw);
3935 }
3936
3937 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3938 {
3939         struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3940         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3941
3942         if (cma_any_addr(addr)) {
3943                 memset(mgid, 0, sizeof *mgid);
3944         } else if (addr->sa_family == AF_INET6) {
3945                 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3946         } else {
3947                 mgid->raw[0] = 0xff;
3948                 mgid->raw[1] = 0x0e;
3949                 mgid->raw[2] = 0;
3950                 mgid->raw[3] = 0;
3951                 mgid->raw[4] = 0;
3952                 mgid->raw[5] = 0;
3953                 mgid->raw[6] = 0;
3954                 mgid->raw[7] = 0;
3955                 mgid->raw[8] = 0;
3956                 mgid->raw[9] = 0;
3957                 mgid->raw[10] = 0xff;
3958                 mgid->raw[11] = 0xff;
3959                 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3960         }
3961 }
3962
3963 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3964                                    struct cma_multicast *mc)
3965 {
3966         struct iboe_mcast_work *work;
3967         struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3968         int err = 0;
3969         struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3970         struct net_device *ndev = NULL;
3971         enum ib_gid_type gid_type;
3972         bool send_only;
3973
3974         send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
3975
3976         if (cma_zero_addr((struct sockaddr *)&mc->addr))
3977                 return -EINVAL;
3978
3979         work = kzalloc(sizeof *work, GFP_KERNEL);
3980         if (!work)
3981                 return -ENOMEM;
3982
3983         mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3984         if (!mc->multicast.ib) {
3985                 err = -ENOMEM;
3986                 goto out1;
3987         }
3988
3989         cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3990
3991         mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3992         if (id_priv->id.ps == RDMA_PS_UDP)
3993                 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3994
3995         if (dev_addr->bound_dev_if)
3996                 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3997         if (!ndev) {
3998                 err = -ENODEV;
3999                 goto out2;
4000         }
4001         mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4002         mc->multicast.ib->rec.hop_limit = 1;
4003         mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4004
4005         gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4006                    rdma_start_port(id_priv->cma_dev->device)];
4007         if (addr->sa_family == AF_INET) {
4008                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4009                         mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4010                         if (!send_only) {
4011                                 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4012                                                     true);
4013                                 if (!err)
4014                                         mc->igmp_joined = true;
4015                         }
4016                 }
4017         } else {
4018                 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4019                         err = -ENOTSUPP;
4020         }
4021         dev_put(ndev);
4022         if (err || !mc->multicast.ib->rec.mtu) {
4023                 if (!err)
4024                         err = -EINVAL;
4025                 goto out2;
4026         }
4027         rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4028                     &mc->multicast.ib->rec.port_gid);
4029         work->id = id_priv;
4030         work->mc = mc;
4031         INIT_WORK(&work->work, iboe_mcast_work_handler);
4032         kref_get(&mc->mcref);
4033         queue_work(cma_wq, &work->work);
4034
4035         return 0;
4036
4037 out2:
4038         kfree(mc->multicast.ib);
4039 out1:
4040         kfree(work);
4041         return err;
4042 }
4043
4044 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4045                         u8 join_state, void *context)
4046 {
4047         struct rdma_id_private *id_priv;
4048         struct cma_multicast *mc;
4049         int ret;
4050
4051         if (!id->device)
4052                 return -EINVAL;
4053
4054         id_priv = container_of(id, struct rdma_id_private, id);
4055         if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4056             !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4057                 return -EINVAL;
4058
4059         mc = kmalloc(sizeof *mc, GFP_KERNEL);
4060         if (!mc)
4061                 return -ENOMEM;
4062
4063         memcpy(&mc->addr, addr, rdma_addr_size(addr));
4064         mc->context = context;
4065         mc->id_priv = id_priv;
4066         mc->igmp_joined = false;
4067         mc->join_state = join_state;
4068         spin_lock(&id_priv->lock);
4069         list_add(&mc->list, &id_priv->mc_list);
4070         spin_unlock(&id_priv->lock);
4071
4072         if (rdma_protocol_roce(id->device, id->port_num)) {
4073                 kref_init(&mc->mcref);
4074                 ret = cma_iboe_join_multicast(id_priv, mc);
4075         } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4076                 ret = cma_join_ib_multicast(id_priv, mc);
4077         else
4078                 ret = -ENOSYS;
4079
4080         if (ret) {
4081                 spin_lock_irq(&id_priv->lock);
4082                 list_del(&mc->list);
4083                 spin_unlock_irq(&id_priv->lock);
4084                 kfree(mc);
4085         }
4086         return ret;
4087 }
4088 EXPORT_SYMBOL(rdma_join_multicast);
4089
4090 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4091 {
4092         struct rdma_id_private *id_priv;
4093         struct cma_multicast *mc;
4094
4095         id_priv = container_of(id, struct rdma_id_private, id);
4096         spin_lock_irq(&id_priv->lock);
4097         list_for_each_entry(mc, &id_priv->mc_list, list) {
4098                 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4099                         list_del(&mc->list);
4100                         spin_unlock_irq(&id_priv->lock);
4101
4102                         if (id->qp)
4103                                 ib_detach_mcast(id->qp,
4104                                                 &mc->multicast.ib->rec.mgid,
4105                                                 be16_to_cpu(mc->multicast.ib->rec.mlid));
4106
4107                         BUG_ON(id_priv->cma_dev->device != id->device);
4108
4109                         if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4110                                 ib_sa_free_multicast(mc->multicast.ib);
4111                                 kfree(mc);
4112                         } else if (rdma_protocol_roce(id->device, id->port_num)) {
4113                                 if (mc->igmp_joined) {
4114                                         struct rdma_dev_addr *dev_addr =
4115                                                 &id->route.addr.dev_addr;
4116                                         struct net_device *ndev = NULL;
4117
4118                                         if (dev_addr->bound_dev_if)
4119                                                 ndev = dev_get_by_index(&init_net,
4120                                                                         dev_addr->bound_dev_if);
4121                                         if (ndev) {
4122                                                 cma_igmp_send(ndev,
4123                                                               &mc->multicast.ib->rec.mgid,
4124                                                               false);
4125                                                 dev_put(ndev);
4126                                         }
4127                                         mc->igmp_joined = false;
4128                                 }
4129                                 kref_put(&mc->mcref, release_mc);
4130                         }
4131                         return;
4132                 }
4133         }
4134         spin_unlock_irq(&id_priv->lock);
4135 }
4136 EXPORT_SYMBOL(rdma_leave_multicast);
4137
4138 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4139 {
4140         struct rdma_dev_addr *dev_addr;
4141         struct cma_ndev_work *work;
4142
4143         dev_addr = &id_priv->id.route.addr.dev_addr;
4144
4145         if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4146             (net_eq(dev_net(ndev), dev_addr->net)) &&
4147             memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4148                 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4149                         ndev->name, &id_priv->id);
4150                 work = kzalloc(sizeof *work, GFP_KERNEL);
4151                 if (!work)
4152                         return -ENOMEM;
4153
4154                 INIT_WORK(&work->work, cma_ndev_work_handler);
4155                 work->id = id_priv;
4156                 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4157                 atomic_inc(&id_priv->refcount);
4158                 queue_work(cma_wq, &work->work);
4159         }
4160
4161         return 0;
4162 }
4163
4164 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4165                                void *ptr)
4166 {
4167         struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4168         struct cma_device *cma_dev;
4169         struct rdma_id_private *id_priv;
4170         int ret = NOTIFY_DONE;
4171
4172         if (event != NETDEV_BONDING_FAILOVER)
4173                 return NOTIFY_DONE;
4174
4175         if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4176                 return NOTIFY_DONE;
4177
4178         mutex_lock(&lock);
4179         list_for_each_entry(cma_dev, &dev_list, list)
4180                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4181                         ret = cma_netdev_change(ndev, id_priv);
4182                         if (ret)
4183                                 goto out;
4184                 }
4185
4186 out:
4187         mutex_unlock(&lock);
4188         return ret;
4189 }
4190
4191 static struct notifier_block cma_nb = {
4192         .notifier_call = cma_netdev_callback
4193 };
4194
4195 static void cma_add_one(struct ib_device *device)
4196 {
4197         struct cma_device *cma_dev;
4198         struct rdma_id_private *id_priv;
4199         unsigned int i;
4200         unsigned long supported_gids = 0;
4201
4202         cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4203         if (!cma_dev)
4204                 return;
4205
4206         cma_dev->device = device;
4207         cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4208                                             sizeof(*cma_dev->default_gid_type),
4209                                             GFP_KERNEL);
4210         if (!cma_dev->default_gid_type) {
4211                 kfree(cma_dev);
4212                 return;
4213         }
4214         for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4215                 supported_gids = roce_gid_type_mask_support(device, i);
4216                 WARN_ON(!supported_gids);
4217                 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4218                         find_first_bit(&supported_gids, BITS_PER_LONG);
4219         }
4220
4221         init_completion(&cma_dev->comp);
4222         atomic_set(&cma_dev->refcount, 1);
4223         INIT_LIST_HEAD(&cma_dev->id_list);
4224         ib_set_client_data(device, &cma_client, cma_dev);
4225
4226         mutex_lock(&lock);
4227         list_add_tail(&cma_dev->list, &dev_list);
4228         list_for_each_entry(id_priv, &listen_any_list, list)
4229                 cma_listen_on_dev(id_priv, cma_dev);
4230         mutex_unlock(&lock);
4231 }
4232
4233 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4234 {
4235         struct rdma_cm_event event;
4236         enum rdma_cm_state state;
4237         int ret = 0;
4238
4239         /* Record that we want to remove the device */
4240         state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4241         if (state == RDMA_CM_DESTROYING)
4242                 return 0;
4243
4244         cma_cancel_operation(id_priv, state);
4245         mutex_lock(&id_priv->handler_mutex);
4246
4247         /* Check for destruction from another callback. */
4248         if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4249                 goto out;
4250
4251         memset(&event, 0, sizeof event);
4252         event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4253         ret = id_priv->id.event_handler(&id_priv->id, &event);
4254 out:
4255         mutex_unlock(&id_priv->handler_mutex);
4256         return ret;
4257 }
4258
4259 static void cma_process_remove(struct cma_device *cma_dev)
4260 {
4261         struct rdma_id_private *id_priv;
4262         int ret;
4263
4264         mutex_lock(&lock);
4265         while (!list_empty(&cma_dev->id_list)) {
4266                 id_priv = list_entry(cma_dev->id_list.next,
4267                                      struct rdma_id_private, list);
4268
4269                 list_del(&id_priv->listen_list);
4270                 list_del_init(&id_priv->list);
4271                 atomic_inc(&id_priv->refcount);
4272                 mutex_unlock(&lock);
4273
4274                 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4275                 cma_deref_id(id_priv);
4276                 if (ret)
4277                         rdma_destroy_id(&id_priv->id);
4278
4279                 mutex_lock(&lock);
4280         }
4281         mutex_unlock(&lock);
4282
4283         cma_deref_dev(cma_dev);
4284         wait_for_completion(&cma_dev->comp);
4285 }
4286
4287 static void cma_remove_one(struct ib_device *device, void *client_data)
4288 {
4289         struct cma_device *cma_dev = client_data;
4290
4291         if (!cma_dev)
4292                 return;
4293
4294         mutex_lock(&lock);
4295         list_del(&cma_dev->list);
4296         mutex_unlock(&lock);
4297
4298         cma_process_remove(cma_dev);
4299         kfree(cma_dev->default_gid_type);
4300         kfree(cma_dev);
4301 }
4302
4303 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4304 {
4305         struct nlmsghdr *nlh;
4306         struct rdma_cm_id_stats *id_stats;
4307         struct rdma_id_private *id_priv;
4308         struct rdma_cm_id *id = NULL;
4309         struct cma_device *cma_dev;
4310         int i_dev = 0, i_id = 0;
4311
4312         /*
4313          * We export all of the IDs as a sequence of messages.  Each
4314          * ID gets its own netlink message.
4315          */
4316         mutex_lock(&lock);
4317
4318         list_for_each_entry(cma_dev, &dev_list, list) {
4319                 if (i_dev < cb->args[0]) {
4320                         i_dev++;
4321                         continue;
4322                 }
4323
4324                 i_id = 0;
4325                 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4326                         if (i_id < cb->args[1]) {
4327                                 i_id++;
4328                                 continue;
4329                         }
4330
4331                         id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4332                                                 sizeof *id_stats, RDMA_NL_RDMA_CM,
4333                                                 RDMA_NL_RDMA_CM_ID_STATS,
4334                                                 NLM_F_MULTI);
4335                         if (!id_stats)
4336                                 goto out;
4337
4338                         memset(id_stats, 0, sizeof *id_stats);
4339                         id = &id_priv->id;
4340                         id_stats->node_type = id->route.addr.dev_addr.dev_type;
4341                         id_stats->port_num = id->port_num;
4342                         id_stats->bound_dev_if =
4343                                 id->route.addr.dev_addr.bound_dev_if;
4344
4345                         if (ibnl_put_attr(skb, nlh,
4346                                           rdma_addr_size(cma_src_addr(id_priv)),
4347                                           cma_src_addr(id_priv),
4348                                           RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4349                                 goto out;
4350                         if (ibnl_put_attr(skb, nlh,
4351                                           rdma_addr_size(cma_dst_addr(id_priv)),
4352                                           cma_dst_addr(id_priv),
4353                                           RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4354                                 goto out;
4355
4356                         id_stats->pid           = id_priv->owner;
4357                         id_stats->port_space    = id->ps;
4358                         id_stats->cm_state      = id_priv->state;
4359                         id_stats->qp_num        = id_priv->qp_num;
4360                         id_stats->qp_type       = id->qp_type;
4361
4362                         i_id++;
4363                 }
4364
4365                 cb->args[1] = 0;
4366                 i_dev++;
4367         }
4368
4369 out:
4370         mutex_unlock(&lock);
4371         cb->args[0] = i_dev;
4372         cb->args[1] = i_id;
4373
4374         return skb->len;
4375 }
4376
4377 static const struct ibnl_client_cbs cma_cb_table[] = {
4378         [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4379                                        .module = THIS_MODULE },
4380 };
4381
4382 static int cma_init_net(struct net *net)
4383 {
4384         struct cma_pernet *pernet = cma_pernet(net);
4385
4386         idr_init(&pernet->tcp_ps);
4387         idr_init(&pernet->udp_ps);
4388         idr_init(&pernet->ipoib_ps);
4389         idr_init(&pernet->ib_ps);
4390
4391         return 0;
4392 }
4393
4394 static void cma_exit_net(struct net *net)
4395 {
4396         struct cma_pernet *pernet = cma_pernet(net);
4397
4398         idr_destroy(&pernet->tcp_ps);
4399         idr_destroy(&pernet->udp_ps);
4400         idr_destroy(&pernet->ipoib_ps);
4401         idr_destroy(&pernet->ib_ps);
4402 }
4403
4404 static struct pernet_operations cma_pernet_operations = {
4405         .init = cma_init_net,
4406         .exit = cma_exit_net,
4407         .id = &cma_pernet_id,
4408         .size = sizeof(struct cma_pernet),
4409 };
4410
4411 static int __init cma_init(void)
4412 {
4413         int ret;
4414
4415         cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4416         if (!cma_wq)
4417                 return -ENOMEM;
4418
4419         ret = register_pernet_subsys(&cma_pernet_operations);
4420         if (ret)
4421                 goto err_wq;
4422
4423         ib_sa_register_client(&sa_client);
4424         rdma_addr_register_client(&addr_client);
4425         register_netdevice_notifier(&cma_nb);
4426
4427         ret = ib_register_client(&cma_client);
4428         if (ret)
4429                 goto err;
4430
4431         if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table),
4432                             cma_cb_table))
4433                 pr_warn("RDMA CMA: failed to add netlink callback\n");
4434         cma_configfs_init();
4435
4436         return 0;
4437
4438 err:
4439         unregister_netdevice_notifier(&cma_nb);
4440         rdma_addr_unregister_client(&addr_client);
4441         ib_sa_unregister_client(&sa_client);
4442 err_wq:
4443         destroy_workqueue(cma_wq);
4444         return ret;
4445 }
4446
4447 static void __exit cma_cleanup(void)
4448 {
4449         cma_configfs_exit();
4450         ibnl_remove_client(RDMA_NL_RDMA_CM);
4451         ib_unregister_client(&cma_client);
4452         unregister_netdevice_notifier(&cma_nb);
4453         rdma_addr_unregister_client(&addr_client);
4454         ib_sa_unregister_client(&sa_client);
4455         unregister_pernet_subsys(&cma_pernet_operations);
4456         destroy_workqueue(cma_wq);
4457 }
4458
4459 module_init(cma_init);
4460 module_exit(cma_cleanup);