drivers/infiniband/core/cm.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   4  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   5  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   6  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   7  * Copyright (c) 2019, Mellanox Technologies inc.  All rights reserved.
   8  */
   9
  10 #include <linux/completion.h>
  11 #include <linux/dma-mapping.h>
  12 #include <linux/device.h>
  13 #include <linux/module.h>
  14 #include <linux/err.h>
  15 #include <linux/idr.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/random.h>
  18 #include <linux/rbtree.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/slab.h>
  21 #include <linux/sysfs.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/kdev_t.h>
  24 #include <linux/etherdevice.h>
  25
  26 #include <rdma/ib_cache.h>
  27 #include <rdma/ib_cm.h>
  28 #include <rdma/ib_sysfs.h>
  29 #include "cm_msgs.h"
  30 #include "core_priv.h"
  31 #include "cm_trace.h"
  32
  33 MODULE_AUTHOR("Sean Hefty");
  34 MODULE_DESCRIPTION("InfiniBand CM");
  35 MODULE_LICENSE("Dual BSD/GPL");
  36
  37 static const char * const ibcm_rej_reason_strs[] = {
  38         [IB_CM_REJ_NO_QP]                       = "no QP",
  39         [IB_CM_REJ_NO_EEC]                      = "no EEC",
  40         [IB_CM_REJ_NO_RESOURCES]                = "no resources",
  41         [IB_CM_REJ_TIMEOUT]                     = "timeout",
  42         [IB_CM_REJ_UNSUPPORTED]                 = "unsupported",
  43         [IB_CM_REJ_INVALID_COMM_ID]             = "invalid comm ID",
  44         [IB_CM_REJ_INVALID_COMM_INSTANCE]       = "invalid comm instance",
  45         [IB_CM_REJ_INVALID_SERVICE_ID]          = "invalid service ID",
  46         [IB_CM_REJ_INVALID_TRANSPORT_TYPE]      = "invalid transport type",
  47         [IB_CM_REJ_STALE_CONN]                  = "stale conn",
  48         [IB_CM_REJ_RDC_NOT_EXIST]               = "RDC not exist",
  49         [IB_CM_REJ_INVALID_GID]                 = "invalid GID",
  50         [IB_CM_REJ_INVALID_LID]                 = "invalid LID",
  51         [IB_CM_REJ_INVALID_SL]                  = "invalid SL",
  52         [IB_CM_REJ_INVALID_TRAFFIC_CLASS]       = "invalid traffic class",
  53         [IB_CM_REJ_INVALID_HOP_LIMIT]           = "invalid hop limit",
  54         [IB_CM_REJ_INVALID_PACKET_RATE]         = "invalid packet rate",
  55         [IB_CM_REJ_INVALID_ALT_GID]             = "invalid alt GID",
  56         [IB_CM_REJ_INVALID_ALT_LID]             = "invalid alt LID",
  57         [IB_CM_REJ_INVALID_ALT_SL]              = "invalid alt SL",
  58         [IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS]   = "invalid alt traffic class",
  59         [IB_CM_REJ_INVALID_ALT_HOP_LIMIT]       = "invalid alt hop limit",
  60         [IB_CM_REJ_INVALID_ALT_PACKET_RATE]     = "invalid alt packet rate",
  61         [IB_CM_REJ_PORT_CM_REDIRECT]            = "port CM redirect",
  62         [IB_CM_REJ_PORT_REDIRECT]               = "port redirect",
  63         [IB_CM_REJ_INVALID_MTU]                 = "invalid MTU",
  64         [IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES] = "insufficient resp resources",
  65         [IB_CM_REJ_CONSUMER_DEFINED]            = "consumer defined",
  66         [IB_CM_REJ_INVALID_RNR_RETRY]           = "invalid RNR retry",
  67         [IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID]     = "duplicate local comm ID",
  68         [IB_CM_REJ_INVALID_CLASS_VERSION]       = "invalid class version",
  69         [IB_CM_REJ_INVALID_FLOW_LABEL]          = "invalid flow label",
  70         [IB_CM_REJ_INVALID_ALT_FLOW_LABEL]      = "invalid alt flow label",
  71         [IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED] =
  72                 "vendor option is not supported",
  73 };
  74
  75 const char *__attribute_const__ ibcm_reject_msg(int reason)
  76 {
  77         size_t index = reason;
  78
  79         if (index < ARRAY_SIZE(ibcm_rej_reason_strs) &&
  80             ibcm_rej_reason_strs[index])
  81                 return ibcm_rej_reason_strs[index];
  82         else
  83                 return "unrecognized reason";
  84 }
  85 EXPORT_SYMBOL(ibcm_reject_msg);
  86
  87 struct cm_id_private;
  88 struct cm_work;
  89 static int cm_add_one(struct ib_device *device);
  90 static void cm_remove_one(struct ib_device *device, void *client_data);
  91 static void cm_process_work(struct cm_id_private *cm_id_priv,
  92                             struct cm_work *work);
  93 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
  94                                    struct ib_cm_sidr_rep_param *param);
  95 static int cm_send_dreq_locked(struct cm_id_private *cm_id_priv,
  96                                const void *private_data, u8 private_data_len);
  97 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
  98                                void *private_data, u8 private_data_len);
  99 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
 100                               enum ib_cm_rej_reason reason, void *ari,
 101                               u8 ari_length, const void *private_data,
 102                               u8 private_data_len);
 103
 104 static struct ib_client cm_client = {
 105         .name   = "cm",
 106         .add    = cm_add_one,
 107         .remove = cm_remove_one
 108 };
 109
 110 static struct ib_cm {
 111         spinlock_t lock;
 112         struct list_head device_list;
 113         rwlock_t device_lock;
 114         struct rb_root listen_service_table;
 115         u64 listen_service_id;
 116         /* struct rb_root peer_service_table; todo: fix peer to peer */
 117         struct rb_root remote_qp_table;
 118         struct rb_root remote_id_table;
 119         struct rb_root remote_sidr_table;
 120         struct xarray local_id_table;
 121         u32 local_id_next;
 122         __be32 random_id_operand;
 123         struct list_head timewait_list;
 124         struct workqueue_struct *wq;
 125 } cm;
 126
 127 /* Counter indexes ordered by attribute ID */
 128 enum {
 129         CM_REQ_COUNTER,
 130         CM_MRA_COUNTER,
 131         CM_REJ_COUNTER,
 132         CM_REP_COUNTER,
 133         CM_RTU_COUNTER,
 134         CM_DREQ_COUNTER,
 135         CM_DREP_COUNTER,
 136         CM_SIDR_REQ_COUNTER,
 137         CM_SIDR_REP_COUNTER,
 138         CM_LAP_COUNTER,
 139         CM_APR_COUNTER,
 140         CM_ATTR_COUNT,
 141         CM_ATTR_ID_OFFSET = 0x0010,
 142 };
 143
 144 enum {
 145         CM_XMIT,
 146         CM_XMIT_RETRIES,
 147         CM_RECV,
 148         CM_RECV_DUPLICATES,
 149         CM_COUNTER_GROUPS
 150 };
 151
 152 struct cm_counter_attribute {
 153         struct ib_port_attribute attr;
 154         unsigned short group;
 155         unsigned short index;
 156 };
 157
 158 struct cm_port {
 159         struct cm_device *cm_dev;
 160         struct ib_mad_agent *mad_agent;
 161         u32 port_num;
 162         atomic_long_t counters[CM_COUNTER_GROUPS][CM_ATTR_COUNT];
 163 };
 164
 165 struct cm_device {
 166         struct kref kref;
 167         struct list_head list;
 168         spinlock_t mad_agent_lock;
 169         struct ib_device *ib_device;
 170         u8 ack_delay;
 171         int going_down;
 172         struct cm_port *port[];
 173 };
 174
 175 struct cm_av {
 176         struct cm_port *port;
 177         struct rdma_ah_attr ah_attr;
 178         u16 dlid_datapath;
 179         u16 pkey_index;
 180         u8 timeout;
 181 };
 182
 183 struct cm_work {
 184         struct delayed_work work;
 185         struct list_head list;
 186         struct cm_port *port;
 187         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 188         __be32 local_id;                        /* Established / timewait */
 189         __be32 remote_id;
 190         struct ib_cm_event cm_event;
 191         struct sa_path_rec path[];
 192 };
 193
 194 struct cm_timewait_info {
 195         struct cm_work work;
 196         struct list_head list;
 197         struct rb_node remote_qp_node;
 198         struct rb_node remote_id_node;
 199         __be64 remote_ca_guid;
 200         __be32 remote_qpn;
 201         u8 inserted_remote_qp;
 202         u8 inserted_remote_id;
 203 };
 204
 205 struct cm_id_private {
 206         struct ib_cm_id id;
 207
 208         struct rb_node service_node;
 209         struct rb_node sidr_id_node;
 210         u32 sidr_slid;
 211         spinlock_t lock;        /* Do not acquire inside cm.lock */
 212         struct completion comp;
 213         refcount_t refcount;
 214         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 215          * Protected by the cm.lock spinlock.
 216          */
 217         int listen_sharecount;
 218         struct rcu_head rcu;
 219
 220         struct ib_mad_send_buf *msg;
 221         struct cm_timewait_info *timewait_info;
 222         /* todo: use alternate port on send failure */
 223         struct cm_av av;
 224         struct cm_av alt_av;
 225
 226         void *private_data;
 227         __be64 tid;
 228         __be32 local_qpn;
 229         __be32 remote_qpn;
 230         enum ib_qp_type qp_type;
 231         __be32 sq_psn;
 232         __be32 rq_psn;
 233         int timeout_ms;
 234         enum ib_mtu path_mtu;
 235         __be16 pkey;
 236         u8 private_data_len;
 237         u8 max_cm_retries;
 238         u8 responder_resources;
 239         u8 initiator_depth;
 240         u8 retry_count;
 241         u8 rnr_retry_count;
 242         u8 service_timeout;
 243         u8 target_ack_delay;
 244
 245         struct list_head work_list;
 246         atomic_t work_count;
 247
 248         struct rdma_ucm_ece ece;
 249 };
 250
 251 static void cm_dev_release(struct kref *kref)
 252 {
 253         struct cm_device *cm_dev = container_of(kref, struct cm_device, kref);
 254         u32 i;
 255
 256         rdma_for_each_port(cm_dev->ib_device, i)
 257                 kfree(cm_dev->port[i - 1]);
 258
 259         kfree(cm_dev);
 260 }
 261
 262 static void cm_device_put(struct cm_device *cm_dev)
 263 {
 264         kref_put(&cm_dev->kref, cm_dev_release);
 265 }
 266
 267 static void cm_work_handler(struct work_struct *work);
 268
 269 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 270 {
 271         if (refcount_dec_and_test(&cm_id_priv->refcount))
 272                 complete(&cm_id_priv->comp);
 273 }
 274
 275 static struct ib_mad_send_buf *cm_alloc_msg(struct cm_id_private *cm_id_priv)
 276 {
 277         struct ib_mad_agent *mad_agent;
 278         struct ib_mad_send_buf *m;
 279         struct ib_ah *ah;
 280
 281         lockdep_assert_held(&cm_id_priv->lock);
 282
 283         if (!cm_id_priv->av.port)
 284                 return ERR_PTR(-EINVAL);
 285
 286         spin_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
 287         mad_agent = cm_id_priv->av.port->mad_agent;
 288         if (!mad_agent) {
 289                 m = ERR_PTR(-EINVAL);
 290                 goto out;
 291         }
 292
 293         ah = rdma_create_ah(mad_agent->qp->pd, &cm_id_priv->av.ah_attr, 0);
 294         if (IS_ERR(ah)) {
 295                 m = ERR_CAST(ah);
 296                 goto out;
 297         }
 298
 299         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 300                                cm_id_priv->av.pkey_index,
 301                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 302                                GFP_ATOMIC,
 303                                IB_MGMT_BASE_VERSION);
 304         if (IS_ERR(m)) {
 305                 rdma_destroy_ah(ah, 0);
 306                 goto out;
 307         }
 308
 309         /* Timeout set by caller if response is expected. */
 310         m->ah = ah;
 311         m->retries = cm_id_priv->max_cm_retries;
 312
 313         refcount_inc(&cm_id_priv->refcount);
 314         m->context[0] = cm_id_priv;
 315
 316 out:
 317         spin_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
 318         return m;
 319 }
 320
 321 static void cm_free_msg(struct ib_mad_send_buf *msg)
 322 {
 323         struct cm_id_private *cm_id_priv = msg->context[0];
 324
 325         if (msg->ah)
 326                 rdma_destroy_ah(msg->ah, 0);
 327         cm_deref_id(cm_id_priv);
 328         ib_free_send_mad(msg);
 329 }
 330
 331 static struct ib_mad_send_buf *
 332 cm_alloc_priv_msg(struct cm_id_private *cm_id_priv)
 333 {
 334         struct ib_mad_send_buf *msg;
 335
 336         lockdep_assert_held(&cm_id_priv->lock);
 337
 338         msg = cm_alloc_msg(cm_id_priv);
 339         if (IS_ERR(msg))
 340                 return msg;
 341         cm_id_priv->msg = msg;
 342         return msg;
 343 }
 344
 345 static void cm_free_priv_msg(struct ib_mad_send_buf *msg)
 346 {
 347         struct cm_id_private *cm_id_priv = msg->context[0];
 348
 349         lockdep_assert_held(&cm_id_priv->lock);
 350
 351         if (!WARN_ON(cm_id_priv->msg != msg))
 352                 cm_id_priv->msg = NULL;
 353
 354         if (msg->ah)
 355                 rdma_destroy_ah(msg->ah, 0);
 356         cm_deref_id(cm_id_priv);
 357         ib_free_send_mad(msg);
 358 }
 359
 360 static struct ib_mad_send_buf *cm_alloc_response_msg_no_ah(struct cm_port *port,
 361                                                            struct ib_mad_recv_wc *mad_recv_wc)
 362 {
 363         return ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 364                                   0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 365                                   GFP_ATOMIC,
 366                                   IB_MGMT_BASE_VERSION);
 367 }
 368
 369 static int cm_create_response_msg_ah(struct cm_port *port,
 370                                      struct ib_mad_recv_wc *mad_recv_wc,
 371                                      struct ib_mad_send_buf *msg)
 372 {
 373         struct ib_ah *ah;
 374
 375         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 376                                   mad_recv_wc->recv_buf.grh, port->port_num);
 377         if (IS_ERR(ah))
 378                 return PTR_ERR(ah);
 379
 380         msg->ah = ah;
 381         return 0;
 382 }
 383
 384 static int cm_alloc_response_msg(struct cm_port *port,
 385                                  struct ib_mad_recv_wc *mad_recv_wc,
 386                                  struct ib_mad_send_buf **msg)
 387 {
 388         struct ib_mad_send_buf *m;
 389         int ret;
 390
 391         m = cm_alloc_response_msg_no_ah(port, mad_recv_wc);
 392         if (IS_ERR(m))
 393                 return PTR_ERR(m);
 394
 395         ret = cm_create_response_msg_ah(port, mad_recv_wc, m);
 396         if (ret) {
 397                 ib_free_send_mad(m);
 398                 return ret;
 399         }
 400
 401         *msg = m;
 402         return 0;
 403 }
 404
 405 static void cm_free_response_msg(struct ib_mad_send_buf *msg)
 406 {
 407         if (msg->ah)
 408                 rdma_destroy_ah(msg->ah, 0);
 409         ib_free_send_mad(msg);
 410 }
 411
 412 static void *cm_copy_private_data(const void *private_data, u8 private_data_len)
 413 {
 414         void *data;
 415
 416         if (!private_data || !private_data_len)
 417                 return NULL;
 418
 419         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 420         if (!data)
 421                 return ERR_PTR(-ENOMEM);
 422
 423         return data;
 424 }
 425
 426 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 427                                  void *private_data, u8 private_data_len)
 428 {
 429         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 430                 kfree(cm_id_priv->private_data);
 431
 432         cm_id_priv->private_data = private_data;
 433         cm_id_priv->private_data_len = private_data_len;
 434 }
 435
 436 static void cm_set_av_port(struct cm_av *av, struct cm_port *port)
 437 {
 438         struct cm_port *old_port = av->port;
 439
 440         if (old_port == port)
 441                 return;
 442
 443         av->port = port;
 444         if (old_port)
 445                 cm_device_put(old_port->cm_dev);
 446         if (port)
 447                 kref_get(&port->cm_dev->kref);
 448 }
 449
 450 static void cm_init_av_for_lap(struct cm_port *port, struct ib_wc *wc,
 451                                struct rdma_ah_attr *ah_attr, struct cm_av *av)
 452 {
 453         cm_set_av_port(av, port);
 454         av->pkey_index = wc->pkey_index;
 455         rdma_move_ah_attr(&av->ah_attr, ah_attr);
 456 }
 457
 458 static int cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 459                                    struct ib_grh *grh, struct cm_av *av)
 460 {
 461         cm_set_av_port(av, port);
 462         av->pkey_index = wc->pkey_index;
 463         return ib_init_ah_attr_from_wc(port->cm_dev->ib_device,
 464                                        port->port_num, wc,
 465                                        grh, &av->ah_attr);
 466 }
 467
 468 static struct cm_port *
 469 get_cm_port_from_path(struct sa_path_rec *path, const struct ib_gid_attr *attr)
 470 {
 471         struct cm_device *cm_dev;
 472         struct cm_port *port = NULL;
 473         unsigned long flags;
 474
 475         if (attr) {
 476                 read_lock_irqsave(&cm.device_lock, flags);
 477                 list_for_each_entry(cm_dev, &cm.device_list, list) {
 478                         if (cm_dev->ib_device == attr->device) {
 479                                 port = cm_dev->port[attr->port_num - 1];
 480                                 break;
 481                         }
 482                 }
 483                 read_unlock_irqrestore(&cm.device_lock, flags);
 484         } else {
 485                 /* SGID attribute can be NULL in following
 486                  * conditions.
 487                  * (a) Alternative path
 488                  * (b) IB link layer without GRH
 489                  * (c) LAP send messages
 490                  */
 491                 read_lock_irqsave(&cm.device_lock, flags);
 492                 list_for_each_entry(cm_dev, &cm.device_list, list) {
 493                         attr = rdma_find_gid(cm_dev->ib_device,
 494                                              &path->sgid,
 495                                              sa_conv_pathrec_to_gid_type(path),
 496                                              NULL);
 497                         if (!IS_ERR(attr)) {
 498                                 port = cm_dev->port[attr->port_num - 1];
 499                                 break;
 500                         }
 501                 }
 502                 read_unlock_irqrestore(&cm.device_lock, flags);
 503                 if (port)
 504                         rdma_put_gid_attr(attr);
 505         }
 506         return port;
 507 }
 508
 509 static int cm_init_av_by_path(struct sa_path_rec *path,
 510                               const struct ib_gid_attr *sgid_attr,
 511                               struct cm_av *av)
 512 {
 513         struct rdma_ah_attr new_ah_attr;
 514         struct cm_device *cm_dev;
 515         struct cm_port *port;
 516         int ret;
 517
 518         port = get_cm_port_from_path(path, sgid_attr);
 519         if (!port)
 520                 return -EINVAL;
 521         cm_dev = port->cm_dev;
 522
 523         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 524                                   be16_to_cpu(path->pkey), &av->pkey_index);
 525         if (ret)
 526                 return ret;
 527
 528         cm_set_av_port(av, port);
 529
 530         /*
 531          * av->ah_attr might be initialized based on wc or during
 532          * request processing time which might have reference to sgid_attr.
 533          * So initialize a new ah_attr on stack.
 534          * If initialization fails, old ah_attr is used for sending any
 535          * responses. If initialization is successful, than new ah_attr
 536          * is used by overwriting the old one. So that right ah_attr
 537          * can be used to return an error response.
 538          */
 539         ret = ib_init_ah_attr_from_path(cm_dev->ib_device, port->port_num, path,
 540                                         &new_ah_attr, sgid_attr);
 541         if (ret)
 542                 return ret;
 543
 544         av->timeout = path->packet_life_time + 1;
 545         rdma_move_ah_attr(&av->ah_attr, &new_ah_attr);
 546         return 0;
 547 }
 548
 549 /* Move av created by cm_init_av_by_path(), so av.dgid is not moved */
 550 static void cm_move_av_from_path(struct cm_av *dest, struct cm_av *src)
 551 {
 552         cm_set_av_port(dest, src->port);
 553         cm_set_av_port(src, NULL);
 554         dest->pkey_index = src->pkey_index;
 555         rdma_move_ah_attr(&dest->ah_attr, &src->ah_attr);
 556         dest->timeout = src->timeout;
 557 }
 558
 559 static void cm_destroy_av(struct cm_av *av)
 560 {
 561         rdma_destroy_ah_attr(&av->ah_attr);
 562         cm_set_av_port(av, NULL);
 563 }
 564
 565 static u32 cm_local_id(__be32 local_id)
 566 {
 567         return (__force u32) (local_id ^ cm.random_id_operand);
 568 }
 569
 570 static struct cm_id_private *cm_acquire_id(__be32 local_id, __be32 remote_id)
 571 {
 572         struct cm_id_private *cm_id_priv;
 573
 574         rcu_read_lock();
 575         cm_id_priv = xa_load(&cm.local_id_table, cm_local_id(local_id));
 576         if (!cm_id_priv || cm_id_priv->id.remote_id != remote_id ||
 577             !refcount_inc_not_zero(&cm_id_priv->refcount))
 578                 cm_id_priv = NULL;
 579         rcu_read_unlock();
 580
 581         return cm_id_priv;
 582 }
 583
 584 /*
 585  * Trivial helpers to strip endian annotation and compare; the
 586  * endianness doesn't actually matter since we just need a stable
 587  * order for the RB tree.
 588  */
 589 static int be32_lt(__be32 a, __be32 b)
 590 {
 591         return (__force u32) a < (__force u32) b;
 592 }
 593
 594 static int be32_gt(__be32 a, __be32 b)
 595 {
 596         return (__force u32) a > (__force u32) b;
 597 }
 598
 599 static int be64_lt(__be64 a, __be64 b)
 600 {
 601         return (__force u64) a < (__force u64) b;
 602 }
 603
 604 static int be64_gt(__be64 a, __be64 b)
 605 {
 606         return (__force u64) a > (__force u64) b;
 607 }
 608
 609 /*
 610  * Inserts a new cm_id_priv into the listen_service_table. Returns cm_id_priv
 611  * if the new ID was inserted, NULL if it could not be inserted due to a
 612  * collision, or the existing cm_id_priv ready for shared usage.
 613  */
 614 static struct cm_id_private *cm_insert_listen(struct cm_id_private *cm_id_priv,
 615                                               ib_cm_handler shared_handler)
 616 {
 617         struct rb_node **link = &cm.listen_service_table.rb_node;
 618         struct rb_node *parent = NULL;
 619         struct cm_id_private *cur_cm_id_priv;
 620         __be64 service_id = cm_id_priv->id.service_id;
 621         unsigned long flags;
 622
 623         spin_lock_irqsave(&cm.lock, flags);
 624         while (*link) {
 625                 parent = *link;
 626                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 627                                           service_node);
 628
 629                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 630                         link = &(*link)->rb_left;
 631                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 632                         link = &(*link)->rb_right;
 633                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 634                         link = &(*link)->rb_left;
 635                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 636                         link = &(*link)->rb_right;
 637                 else {
 638                         /*
 639                          * Sharing an ib_cm_id with different handlers is not
 640                          * supported
 641                          */
 642                         if (cur_cm_id_priv->id.cm_handler != shared_handler ||
 643                             cur_cm_id_priv->id.context ||
 644                             WARN_ON(!cur_cm_id_priv->id.cm_handler)) {
 645                                 spin_unlock_irqrestore(&cm.lock, flags);
 646                                 return NULL;
 647                         }
 648                         refcount_inc(&cur_cm_id_priv->refcount);
 649                         cur_cm_id_priv->listen_sharecount++;
 650                         spin_unlock_irqrestore(&cm.lock, flags);
 651                         return cur_cm_id_priv;
 652                 }
 653         }
 654         cm_id_priv->listen_sharecount++;
 655         rb_link_node(&cm_id_priv->service_node, parent, link);
 656         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 657         spin_unlock_irqrestore(&cm.lock, flags);
 658         return cm_id_priv;
 659 }
 660
 661 static struct cm_id_private *cm_find_listen(struct ib_device *device,
 662                                             __be64 service_id)
 663 {
 664         struct rb_node *node = cm.listen_service_table.rb_node;
 665         struct cm_id_private *cm_id_priv;
 666
 667         while (node) {
 668                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 669
 670                 if (device < cm_id_priv->id.device)
 671                         node = node->rb_left;
 672                 else if (device > cm_id_priv->id.device)
 673                         node = node->rb_right;
 674                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 675                         node = node->rb_left;
 676                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 677                         node = node->rb_right;
 678                 else {
 679                         refcount_inc(&cm_id_priv->refcount);
 680                         return cm_id_priv;
 681                 }
 682         }
 683         return NULL;
 684 }
 685
 686 static struct cm_timewait_info *
 687 cm_insert_remote_id(struct cm_timewait_info *timewait_info)
 688 {
 689         struct rb_node **link = &cm.remote_id_table.rb_node;
 690         struct rb_node *parent = NULL;
 691         struct cm_timewait_info *cur_timewait_info;
 692         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 693         __be32 remote_id = timewait_info->work.remote_id;
 694
 695         while (*link) {
 696                 parent = *link;
 697                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 698                                              remote_id_node);
 699                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 700                         link = &(*link)->rb_left;
 701                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 702                         link = &(*link)->rb_right;
 703                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 704                         link = &(*link)->rb_left;
 705                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 706                         link = &(*link)->rb_right;
 707                 else
 708                         return cur_timewait_info;
 709         }
 710         timewait_info->inserted_remote_id = 1;
 711         rb_link_node(&timewait_info->remote_id_node, parent, link);
 712         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 713         return NULL;
 714 }
 715
 716 static struct cm_id_private *cm_find_remote_id(__be64 remote_ca_guid,
 717                                                __be32 remote_id)
 718 {
 719         struct rb_node *node = cm.remote_id_table.rb_node;
 720         struct cm_timewait_info *timewait_info;
 721         struct cm_id_private *res = NULL;
 722
 723         spin_lock_irq(&cm.lock);
 724         while (node) {
 725                 timewait_info = rb_entry(node, struct cm_timewait_info,
 726                                          remote_id_node);
 727                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 728                         node = node->rb_left;
 729                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 730                         node = node->rb_right;
 731                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 732                         node = node->rb_left;
 733                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 734                         node = node->rb_right;
 735                 else {
 736                         res = cm_acquire_id(timewait_info->work.local_id,
 737                                              timewait_info->work.remote_id);
 738                         break;
 739                 }
 740         }
 741         spin_unlock_irq(&cm.lock);
 742         return res;
 743 }
 744
 745 static struct cm_timewait_info *
 746 cm_insert_remote_qpn(struct cm_timewait_info *timewait_info)
 747 {
 748         struct rb_node **link = &cm.remote_qp_table.rb_node;
 749         struct rb_node *parent = NULL;
 750         struct cm_timewait_info *cur_timewait_info;
 751         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 752         __be32 remote_qpn = timewait_info->remote_qpn;
 753
 754         while (*link) {
 755                 parent = *link;
 756                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 757                                              remote_qp_node);
 758                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 759                         link = &(*link)->rb_left;
 760                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 761                         link = &(*link)->rb_right;
 762                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 763                         link = &(*link)->rb_left;
 764                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 765                         link = &(*link)->rb_right;
 766                 else
 767                         return cur_timewait_info;
 768         }
 769         timewait_info->inserted_remote_qp = 1;
 770         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 771         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 772         return NULL;
 773 }
 774
 775 static struct cm_id_private *
 776 cm_insert_remote_sidr(struct cm_id_private *cm_id_priv)
 777 {
 778         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 779         struct rb_node *parent = NULL;
 780         struct cm_id_private *cur_cm_id_priv;
 781         __be32 remote_id = cm_id_priv->id.remote_id;
 782
 783         while (*link) {
 784                 parent = *link;
 785                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 786                                           sidr_id_node);
 787                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 788                         link = &(*link)->rb_left;
 789                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 790                         link = &(*link)->rb_right;
 791                 else {
 792                         if (cur_cm_id_priv->sidr_slid < cm_id_priv->sidr_slid)
 793                                 link = &(*link)->rb_left;
 794                         else if (cur_cm_id_priv->sidr_slid > cm_id_priv->sidr_slid)
 795                                 link = &(*link)->rb_right;
 796                         else
 797                                 return cur_cm_id_priv;
 798                 }
 799         }
 800         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 801         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 802         return NULL;
 803 }
 804
 805 static struct cm_id_private *cm_alloc_id_priv(struct ib_device *device,
 806                                               ib_cm_handler cm_handler,
 807                                               void *context)
 808 {
 809         struct cm_id_private *cm_id_priv;
 810         u32 id;
 811         int ret;
 812
 813         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 814         if (!cm_id_priv)
 815                 return ERR_PTR(-ENOMEM);
 816
 817         cm_id_priv->id.state = IB_CM_IDLE;
 818         cm_id_priv->id.device = device;
 819         cm_id_priv->id.cm_handler = cm_handler;
 820         cm_id_priv->id.context = context;
 821         cm_id_priv->id.remote_cm_qpn = 1;
 822
 823         RB_CLEAR_NODE(&cm_id_priv->service_node);
 824         RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
 825         spin_lock_init(&cm_id_priv->lock);
 826         init_completion(&cm_id_priv->comp);
 827         INIT_LIST_HEAD(&cm_id_priv->work_list);
 828         atomic_set(&cm_id_priv->work_count, -1);
 829         refcount_set(&cm_id_priv->refcount, 1);
 830
 831         ret = xa_alloc_cyclic(&cm.local_id_table, &id, NULL, xa_limit_32b,
 832                               &cm.local_id_next, GFP_KERNEL);
 833         if (ret < 0)
 834                 goto error;
 835         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 836
 837         return cm_id_priv;
 838
 839 error:
 840         kfree(cm_id_priv);
 841         return ERR_PTR(ret);
 842 }
 843
 844 /*
 845  * Make the ID visible to the MAD handlers and other threads that use the
 846  * xarray.
 847  */
 848 static void cm_finalize_id(struct cm_id_private *cm_id_priv)
 849 {
 850         xa_store(&cm.local_id_table, cm_local_id(cm_id_priv->id.local_id),
 851                  cm_id_priv, GFP_ATOMIC);
 852 }
 853
 854 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 855                                  ib_cm_handler cm_handler,
 856                                  void *context)
 857 {
 858         struct cm_id_private *cm_id_priv;
 859
 860         cm_id_priv = cm_alloc_id_priv(device, cm_handler, context);
 861         if (IS_ERR(cm_id_priv))
 862                 return ERR_CAST(cm_id_priv);
 863
 864         cm_finalize_id(cm_id_priv);
 865         return &cm_id_priv->id;
 866 }
 867 EXPORT_SYMBOL(ib_create_cm_id);
 868
 869 static struct cm_work *cm_dequeue_work(struct cm_id_private *cm_id_priv)
 870 {
 871         struct cm_work *work;
 872
 873         if (list_empty(&cm_id_priv->work_list))
 874                 return NULL;
 875
 876         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 877         list_del(&work->list);
 878         return work;
 879 }
 880
 881 static void cm_free_work(struct cm_work *work)
 882 {
 883         if (work->mad_recv_wc)
 884                 ib_free_recv_mad(work->mad_recv_wc);
 885         kfree(work);
 886 }
 887
 888 static void cm_queue_work_unlock(struct cm_id_private *cm_id_priv,
 889                                  struct cm_work *work)
 890         __releases(&cm_id_priv->lock)
 891 {
 892         bool immediate;
 893
 894         /*
 895          * To deliver the event to the user callback we have the drop the
 896          * spinlock, however, we need to ensure that the user callback is single
 897          * threaded and receives events in the temporal order. If there are
 898          * already events being processed then thread new events onto a list,
 899          * the thread currently processing will pick them up.
 900          */
 901         immediate = atomic_inc_and_test(&cm_id_priv->work_count);
 902         if (!immediate) {
 903                 list_add_tail(&work->list, &cm_id_priv->work_list);
 904                 /*
 905                  * This routine always consumes incoming reference. Once queued
 906                  * to the work_list then a reference is held by the thread
 907                  * currently running cm_process_work() and this reference is not
 908                  * needed.
 909                  */
 910                 cm_deref_id(cm_id_priv);
 911         }
 912         spin_unlock_irq(&cm_id_priv->lock);
 913
 914         if (immediate)
 915                 cm_process_work(cm_id_priv, work);
 916 }
 917
 918 static inline int cm_convert_to_ms(int iba_time)
 919 {
 920         /* approximate conversion to ms from 4.096us x 2^iba_time */
 921         return 1 << max(iba_time - 8, 0);
 922 }
 923
 924 /*
 925  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 926  * Because of how ack_timeout is stored, adding one doubles the timeout.
 927  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 928  * increment it (round up) only if the other is within 50%.
 929  */
 930 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 931 {
 932         int ack_timeout = packet_life_time + 1;
 933
 934         if (ack_timeout >= ca_ack_delay)
 935                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 936         else
 937                 ack_timeout = ca_ack_delay +
 938                               (ack_timeout >= (ca_ack_delay - 1));
 939
 940         return min(31, ack_timeout);
 941 }
 942
 943 static void cm_remove_remote(struct cm_id_private *cm_id_priv)
 944 {
 945         struct cm_timewait_info *timewait_info = cm_id_priv->timewait_info;
 946
 947         if (timewait_info->inserted_remote_id) {
 948                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 949                 timewait_info->inserted_remote_id = 0;
 950         }
 951
 952         if (timewait_info->inserted_remote_qp) {
 953                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 954                 timewait_info->inserted_remote_qp = 0;
 955         }
 956 }
 957
 958 static struct cm_timewait_info *cm_create_timewait_info(__be32 local_id)
 959 {
 960         struct cm_timewait_info *timewait_info;
 961
 962         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 963         if (!timewait_info)
 964                 return ERR_PTR(-ENOMEM);
 965
 966         timewait_info->work.local_id = local_id;
 967         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 968         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 969         return timewait_info;
 970 }
 971
 972 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 973 {
 974         int wait_time;
 975         unsigned long flags;
 976         struct cm_device *cm_dev;
 977
 978         lockdep_assert_held(&cm_id_priv->lock);
 979
 980         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 981         if (!cm_dev)
 982                 return;
 983
 984         spin_lock_irqsave(&cm.lock, flags);
 985         cm_remove_remote(cm_id_priv);
 986         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 987         spin_unlock_irqrestore(&cm.lock, flags);
 988
 989         /*
 990          * The cm_id could be destroyed by the user before we exit timewait.
 991          * To protect against this, we search for the cm_id after exiting
 992          * timewait before notifying the user that we've exited timewait.
 993          */
 994         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 995         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 996
 997         /* Check if the device started its remove_one */
 998         spin_lock_irqsave(&cm.lock, flags);
 999         if (!cm_dev->going_down)
1000                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
1001                                    msecs_to_jiffies(wait_time));
1002         spin_unlock_irqrestore(&cm.lock, flags);
1003
1004         /*
1005          * The timewait_info is converted into a work and gets freed during
1006          * cm_free_work() in cm_timewait_handler().
1007          */
1008         BUILD_BUG_ON(offsetof(struct cm_timewait_info, work) != 0);
1009         cm_id_priv->timewait_info = NULL;
1010 }
1011
1012 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
1013 {
1014         unsigned long flags;
1015
1016         lockdep_assert_held(&cm_id_priv->lock);
1017
1018         cm_id_priv->id.state = IB_CM_IDLE;
1019         if (cm_id_priv->timewait_info) {
1020                 spin_lock_irqsave(&cm.lock, flags);
1021                 cm_remove_remote(cm_id_priv);
1022                 spin_unlock_irqrestore(&cm.lock, flags);
1023                 kfree(cm_id_priv->timewait_info);
1024                 cm_id_priv->timewait_info = NULL;
1025         }
1026 }
1027
1028 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
1029 {
1030         struct cm_id_private *cm_id_priv;
1031         struct cm_work *work;
1032
1033         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1034         spin_lock_irq(&cm_id_priv->lock);
1035 retest:
1036         switch (cm_id->state) {
1037         case IB_CM_LISTEN:
1038                 spin_lock(&cm.lock);
1039                 if (--cm_id_priv->listen_sharecount > 0) {
1040                         /* The id is still shared. */
1041                         WARN_ON(refcount_read(&cm_id_priv->refcount) == 1);
1042                         spin_unlock(&cm.lock);
1043                         spin_unlock_irq(&cm_id_priv->lock);
1044                         cm_deref_id(cm_id_priv);
1045                         return;
1046                 }
1047                 cm_id->state = IB_CM_IDLE;
1048                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
1049                 RB_CLEAR_NODE(&cm_id_priv->service_node);
1050                 spin_unlock(&cm.lock);
1051                 break;
1052         case IB_CM_SIDR_REQ_SENT:
1053                 cm_id->state = IB_CM_IDLE;
1054                 ib_cancel_mad(cm_id_priv->msg);
1055                 break;
1056         case IB_CM_SIDR_REQ_RCVD:
1057                 cm_send_sidr_rep_locked(cm_id_priv,
1058                                         &(struct ib_cm_sidr_rep_param){
1059                                                 .status = IB_SIDR_REJECT });
1060                 /* cm_send_sidr_rep_locked will not move to IDLE if it fails */
1061                 cm_id->state = IB_CM_IDLE;
1062                 break;
1063         case IB_CM_REQ_SENT:
1064         case IB_CM_MRA_REQ_RCVD:
1065                 ib_cancel_mad(cm_id_priv->msg);
1066                 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_TIMEOUT,
1067                                    &cm_id_priv->id.device->node_guid,
1068                                    sizeof(cm_id_priv->id.device->node_guid),
1069                                    NULL, 0);
1070                 break;
1071         case IB_CM_REQ_RCVD:
1072                 if (err == -ENOMEM) {
1073                         /* Do not reject to allow future retries. */
1074                         cm_reset_to_idle(cm_id_priv);
1075                 } else {
1076                         cm_send_rej_locked(cm_id_priv,
1077                                            IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
1078                                            NULL, 0);
1079                 }
1080                 break;
1081         case IB_CM_REP_SENT:
1082         case IB_CM_MRA_REP_RCVD:
1083                 ib_cancel_mad(cm_id_priv->msg);
1084                 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL,
1085                                    0, NULL, 0);
1086                 goto retest;
1087         case IB_CM_MRA_REQ_SENT:
1088         case IB_CM_REP_RCVD:
1089         case IB_CM_MRA_REP_SENT:
1090                 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL,
1091                                    0, NULL, 0);
1092                 break;
1093         case IB_CM_ESTABLISHED:
1094                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
1095                         cm_id->state = IB_CM_IDLE;
1096                         break;
1097                 }
1098                 cm_send_dreq_locked(cm_id_priv, NULL, 0);
1099                 goto retest;
1100         case IB_CM_DREQ_SENT:
1101                 ib_cancel_mad(cm_id_priv->msg);
1102                 cm_enter_timewait(cm_id_priv);
1103                 goto retest;
1104         case IB_CM_DREQ_RCVD:
1105                 cm_send_drep_locked(cm_id_priv, NULL, 0);
1106                 WARN_ON(cm_id->state != IB_CM_TIMEWAIT);
1107                 goto retest;
1108         case IB_CM_TIMEWAIT:
1109                 /*
1110                  * The cm_acquire_id in cm_timewait_handler will stop working
1111                  * once we do xa_erase below, so just move to idle here for
1112                  * consistency.
1113                  */
1114                 cm_id->state = IB_CM_IDLE;
1115                 break;
1116         case IB_CM_IDLE:
1117                 break;
1118         }
1119         WARN_ON(cm_id->state != IB_CM_IDLE);
1120
1121         spin_lock(&cm.lock);
1122         /* Required for cleanup paths related cm_req_handler() */
1123         if (cm_id_priv->timewait_info) {
1124                 cm_remove_remote(cm_id_priv);
1125                 kfree(cm_id_priv->timewait_info);
1126                 cm_id_priv->timewait_info = NULL;
1127         }
1128
1129         WARN_ON(cm_id_priv->listen_sharecount);
1130         WARN_ON(!RB_EMPTY_NODE(&cm_id_priv->service_node));
1131         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
1132                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
1133         spin_unlock(&cm.lock);
1134         spin_unlock_irq(&cm_id_priv->lock);
1135
1136         xa_erase(&cm.local_id_table, cm_local_id(cm_id->local_id));
1137         cm_deref_id(cm_id_priv);
1138         wait_for_completion(&cm_id_priv->comp);
1139         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1140                 cm_free_work(work);
1141
1142         cm_destroy_av(&cm_id_priv->av);
1143         cm_destroy_av(&cm_id_priv->alt_av);
1144         kfree(cm_id_priv->private_data);
1145         kfree_rcu(cm_id_priv, rcu);
1146 }
1147
1148 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1149 {
1150         cm_destroy_id(cm_id, 0);
1151 }
1152 EXPORT_SYMBOL(ib_destroy_cm_id);
1153
1154 static int cm_init_listen(struct cm_id_private *cm_id_priv, __be64 service_id)
1155 {
1156         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1157             (service_id != IB_CM_ASSIGN_SERVICE_ID))
1158                 return -EINVAL;
1159
1160         if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1161                 cm_id_priv->id.service_id = cpu_to_be64(cm.listen_service_id++);
1162         else
1163                 cm_id_priv->id.service_id = service_id;
1164
1165         return 0;
1166 }
1167
1168 /**
1169  * ib_cm_listen - Initiates listening on the specified service ID for
1170  *   connection and service ID resolution requests.
1171  * @cm_id: Connection identifier associated with the listen request.
1172  * @service_id: Service identifier matched against incoming connection
1173  *   and service ID resolution requests.  The service ID should be specified
1174  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1175  *   assign a service ID to the caller.
1176  */
1177 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id)
1178 {
1179         struct cm_id_private *cm_id_priv =
1180                 container_of(cm_id, struct cm_id_private, id);
1181         unsigned long flags;
1182         int ret;
1183
1184         spin_lock_irqsave(&cm_id_priv->lock, flags);
1185         if (cm_id_priv->id.state != IB_CM_IDLE) {
1186                 ret = -EINVAL;
1187                 goto out;
1188         }
1189
1190         ret = cm_init_listen(cm_id_priv, service_id);
1191         if (ret)
1192                 goto out;
1193
1194         if (!cm_insert_listen(cm_id_priv, NULL)) {
1195                 ret = -EBUSY;
1196                 goto out;
1197         }
1198
1199         cm_id_priv->id.state = IB_CM_LISTEN;
1200         ret = 0;
1201
1202 out:
1203         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1204         return ret;
1205 }
1206 EXPORT_SYMBOL(ib_cm_listen);
1207
1208 /**
1209  * ib_cm_insert_listen - Create a new listening ib_cm_id and listen on
1210  *                       the given service ID.
1211  *
1212  * If there's an existing ID listening on that same device and service ID,
1213  * return it.
1214  *
1215  * @device: Device associated with the cm_id.  All related communication will
1216  * be associated with the specified device.
1217  * @cm_handler: Callback invoked to notify the user of CM events.
1218  * @service_id: Service identifier matched against incoming connection
1219  *   and service ID resolution requests.  The service ID should be specified
1220  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1221  *   assign a service ID to the caller.
1222  *
1223  * Callers should call ib_destroy_cm_id when done with the listener ID.
1224  */
1225 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1226                                      ib_cm_handler cm_handler,
1227                                      __be64 service_id)
1228 {
1229         struct cm_id_private *listen_id_priv;
1230         struct cm_id_private *cm_id_priv;
1231         int err = 0;
1232
1233         /* Create an ID in advance, since the creation may sleep */
1234         cm_id_priv = cm_alloc_id_priv(device, cm_handler, NULL);
1235         if (IS_ERR(cm_id_priv))
1236                 return ERR_CAST(cm_id_priv);
1237
1238         err = cm_init_listen(cm_id_priv, service_id);
1239         if (err) {
1240                 ib_destroy_cm_id(&cm_id_priv->id);
1241                 return ERR_PTR(err);
1242         }
1243
1244         spin_lock_irq(&cm_id_priv->lock);
1245         listen_id_priv = cm_insert_listen(cm_id_priv, cm_handler);
1246         if (listen_id_priv != cm_id_priv) {
1247                 spin_unlock_irq(&cm_id_priv->lock);
1248                 ib_destroy_cm_id(&cm_id_priv->id);
1249                 if (!listen_id_priv)
1250                         return ERR_PTR(-EINVAL);
1251                 return &listen_id_priv->id;
1252         }
1253         cm_id_priv->id.state = IB_CM_LISTEN;
1254         spin_unlock_irq(&cm_id_priv->lock);
1255
1256         /*
1257          * A listen ID does not need to be in the xarray since it does not
1258          * receive mads, is not placed in the remote_id or remote_qpn rbtree,
1259          * and does not enter timewait.
1260          */
1261
1262         return &cm_id_priv->id;
1263 }
1264 EXPORT_SYMBOL(ib_cm_insert_listen);
1265
1266 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv)
1267 {
1268         u64 hi_tid = 0, low_tid;
1269
1270         lockdep_assert_held(&cm_id_priv->lock);
1271
1272         low_tid = (u64)cm_id_priv->id.local_id;
1273         if (!cm_id_priv->av.port)
1274                 return cpu_to_be64(low_tid);
1275
1276         spin_lock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
1277         if (cm_id_priv->av.port->mad_agent)
1278                 hi_tid = ((u64)cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1279         spin_unlock(&cm_id_priv->av.port->cm_dev->mad_agent_lock);
1280         return cpu_to_be64(hi_tid | low_tid);
1281 }
1282
1283 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1284                               __be16 attr_id, __be64 tid)
1285 {
1286         hdr->base_version  = IB_MGMT_BASE_VERSION;
1287         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1288         hdr->class_version = IB_CM_CLASS_VERSION;
1289         hdr->method        = IB_MGMT_METHOD_SEND;
1290         hdr->attr_id       = attr_id;
1291         hdr->tid           = tid;
1292 }
1293
1294 static void cm_format_mad_ece_hdr(struct ib_mad_hdr *hdr, __be16 attr_id,
1295                                   __be64 tid, u32 attr_mod)
1296 {
1297         cm_format_mad_hdr(hdr, attr_id, tid);
1298         hdr->attr_mod = cpu_to_be32(attr_mod);
1299 }
1300
1301 static void cm_format_req(struct cm_req_msg *req_msg,
1302                           struct cm_id_private *cm_id_priv,
1303                           struct ib_cm_req_param *param)
1304 {
1305         struct sa_path_rec *pri_path = param->primary_path;
1306         struct sa_path_rec *alt_path = param->alternate_path;
1307         bool pri_ext = false;
1308         __be16 lid;
1309
1310         if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
1311                 pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
1312                                               pri_path->opa.slid);
1313
1314         cm_format_mad_ece_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1315                               cm_form_tid(cm_id_priv), param->ece.attr_mod);
1316
1317         IBA_SET(CM_REQ_LOCAL_COMM_ID, req_msg,
1318                 be32_to_cpu(cm_id_priv->id.local_id));
1319         IBA_SET(CM_REQ_SERVICE_ID, req_msg, be64_to_cpu(param->service_id));
1320         IBA_SET(CM_REQ_LOCAL_CA_GUID, req_msg,
1321                 be64_to_cpu(cm_id_priv->id.device->node_guid));
1322         IBA_SET(CM_REQ_LOCAL_QPN, req_msg, param->qp_num);
1323         IBA_SET(CM_REQ_INITIATOR_DEPTH, req_msg, param->initiator_depth);
1324         IBA_SET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg,
1325                 param->remote_cm_response_timeout);
1326         cm_req_set_qp_type(req_msg, param->qp_type);
1327         IBA_SET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg, param->flow_control);
1328         IBA_SET(CM_REQ_STARTING_PSN, req_msg, param->starting_psn);
1329         IBA_SET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg,
1330                 param->local_cm_response_timeout);
1331         IBA_SET(CM_REQ_PARTITION_KEY, req_msg,
1332                 be16_to_cpu(param->primary_path->pkey));
1333         IBA_SET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg,
1334                 param->primary_path->mtu);
1335         IBA_SET(CM_REQ_MAX_CM_RETRIES, req_msg, param->max_cm_retries);
1336
1337         if (param->qp_type != IB_QPT_XRC_INI) {
1338                 IBA_SET(CM_REQ_RESPONDER_RESOURCES, req_msg,
1339                         param->responder_resources);
1340                 IBA_SET(CM_REQ_RETRY_COUNT, req_msg, param->retry_count);
1341                 IBA_SET(CM_REQ_RNR_RETRY_COUNT, req_msg,
1342                         param->rnr_retry_count);
1343                 IBA_SET(CM_REQ_SRQ, req_msg, param->srq);
1344         }
1345
1346         *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) =
1347                 pri_path->sgid;
1348         *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) =
1349                 pri_path->dgid;
1350         if (pri_ext) {
1351                 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg)
1352                         ->global.interface_id =
1353                         OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid));
1354                 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg)
1355                         ->global.interface_id =
1356                         OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid));
1357         }
1358         if (pri_path->hop_limit <= 1) {
1359                 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1360                         be16_to_cpu(pri_ext ? 0 :
1361                                               htons(ntohl(sa_path_get_slid(
1362                                                       pri_path)))));
1363                 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1364                         be16_to_cpu(pri_ext ? 0 :
1365                                               htons(ntohl(sa_path_get_dlid(
1366                                                       pri_path)))));
1367         } else {
1368
1369                 if (param->primary_path_inbound) {
1370                         lid = param->primary_path_inbound->ib.dlid;
1371                         IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1372                                 be16_to_cpu(lid));
1373                 } else
1374                         IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1375                                 be16_to_cpu(IB_LID_PERMISSIVE));
1376
1377                 /* Work-around until there's a way to obtain remote LID info */
1378                 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1379                         be16_to_cpu(IB_LID_PERMISSIVE));
1380         }
1381         IBA_SET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg,
1382                 be32_to_cpu(pri_path->flow_label));
1383         IBA_SET(CM_REQ_PRIMARY_PACKET_RATE, req_msg, pri_path->rate);
1384         IBA_SET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg, pri_path->traffic_class);
1385         IBA_SET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg, pri_path->hop_limit);
1386         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, pri_path->sl);
1387         IBA_SET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg,
1388                 (pri_path->hop_limit <= 1));
1389         IBA_SET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg,
1390                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1391                                pri_path->packet_life_time));
1392
1393         if (alt_path) {
1394                 bool alt_ext = false;
1395
1396                 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA)
1397                         alt_ext = opa_is_extended_lid(alt_path->opa.dlid,
1398                                                       alt_path->opa.slid);
1399
1400                 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg) =
1401                         alt_path->sgid;
1402                 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg) =
1403                         alt_path->dgid;
1404                 if (alt_ext) {
1405                         IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1406                                         req_msg)
1407                                 ->global.interface_id =
1408                                 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid));
1409                         IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID,
1410                                         req_msg)
1411                                 ->global.interface_id =
1412                                 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid));
1413                 }
1414                 if (alt_path->hop_limit <= 1) {
1415                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1416                                 be16_to_cpu(
1417                                         alt_ext ? 0 :
1418                                                   htons(ntohl(sa_path_get_slid(
1419                                                           alt_path)))));
1420                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1421                                 be16_to_cpu(
1422                                         alt_ext ? 0 :
1423                                                   htons(ntohl(sa_path_get_dlid(
1424                                                           alt_path)))));
1425                 } else {
1426                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1427                                 be16_to_cpu(IB_LID_PERMISSIVE));
1428                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1429                                 be16_to_cpu(IB_LID_PERMISSIVE));
1430                 }
1431                 IBA_SET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg,
1432                         be32_to_cpu(alt_path->flow_label));
1433                 IBA_SET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg, alt_path->rate);
1434                 IBA_SET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg,
1435                         alt_path->traffic_class);
1436                 IBA_SET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg,
1437                         alt_path->hop_limit);
1438                 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, alt_path->sl);
1439                 IBA_SET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg,
1440                         (alt_path->hop_limit <= 1));
1441                 IBA_SET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg,
1442                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1443                                        alt_path->packet_life_time));
1444         }
1445         IBA_SET(CM_REQ_VENDOR_ID, req_msg, param->ece.vendor_id);
1446
1447         if (param->private_data && param->private_data_len)
1448                 IBA_SET_MEM(CM_REQ_PRIVATE_DATA, req_msg, param->private_data,
1449                             param->private_data_len);
1450 }
1451
1452 static int cm_validate_req_param(struct ib_cm_req_param *param)
1453 {
1454         if (!param->primary_path)
1455                 return -EINVAL;
1456
1457         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1458             param->qp_type != IB_QPT_XRC_INI)
1459                 return -EINVAL;
1460
1461         if (param->private_data &&
1462             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1463                 return -EINVAL;
1464
1465         if (param->alternate_path &&
1466             (param->alternate_path->pkey != param->primary_path->pkey ||
1467              param->alternate_path->mtu != param->primary_path->mtu))
1468                 return -EINVAL;
1469
1470         return 0;
1471 }
1472
1473 int ib_send_cm_req(struct ib_cm_id *cm_id,
1474                    struct ib_cm_req_param *param)
1475 {
1476         struct cm_av av = {}, alt_av = {};
1477         struct cm_id_private *cm_id_priv;
1478         struct ib_mad_send_buf *msg;
1479         struct cm_req_msg *req_msg;
1480         unsigned long flags;
1481         int ret;
1482
1483         ret = cm_validate_req_param(param);
1484         if (ret)
1485                 return ret;
1486
1487         /* Verify that we're not in timewait. */
1488         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1489         spin_lock_irqsave(&cm_id_priv->lock, flags);
1490         if (cm_id->state != IB_CM_IDLE || WARN_ON(cm_id_priv->timewait_info)) {
1491                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1492                 return -EINVAL;
1493         }
1494         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1495
1496         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1497                                                             id.local_id);
1498         if (IS_ERR(cm_id_priv->timewait_info)) {
1499                 ret = PTR_ERR(cm_id_priv->timewait_info);
1500                 cm_id_priv->timewait_info = NULL;
1501                 return ret;
1502         }
1503
1504         ret = cm_init_av_by_path(param->primary_path,
1505                                  param->ppath_sgid_attr, &av);
1506         if (ret)
1507                 return ret;
1508         if (param->alternate_path) {
1509                 ret = cm_init_av_by_path(param->alternate_path, NULL,
1510                                          &alt_av);
1511                 if (ret) {
1512                         cm_destroy_av(&av);
1513                         return ret;
1514                 }
1515         }
1516         cm_id->service_id = param->service_id;
1517         cm_id_priv->timeout_ms = cm_convert_to_ms(
1518                                     param->primary_path->packet_life_time) * 2 +
1519                                  cm_convert_to_ms(
1520                                     param->remote_cm_response_timeout);
1521         cm_id_priv->max_cm_retries = param->max_cm_retries;
1522         cm_id_priv->initiator_depth = param->initiator_depth;
1523         cm_id_priv->responder_resources = param->responder_resources;
1524         cm_id_priv->retry_count = param->retry_count;
1525         cm_id_priv->path_mtu = param->primary_path->mtu;
1526         cm_id_priv->pkey = param->primary_path->pkey;
1527         cm_id_priv->qp_type = param->qp_type;
1528
1529         spin_lock_irqsave(&cm_id_priv->lock, flags);
1530
1531         cm_move_av_from_path(&cm_id_priv->av, &av);
1532         if (param->primary_path_outbound)
1533                 cm_id_priv->av.dlid_datapath =
1534                         be16_to_cpu(param->primary_path_outbound->ib.dlid);
1535
1536         if (param->alternate_path)
1537                 cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av);
1538
1539         msg = cm_alloc_priv_msg(cm_id_priv);
1540         if (IS_ERR(msg)) {
1541                 ret = PTR_ERR(msg);
1542                 goto out_unlock;
1543         }
1544
1545         req_msg = (struct cm_req_msg *)msg->mad;
1546         cm_format_req(req_msg, cm_id_priv, param);
1547         cm_id_priv->tid = req_msg->hdr.tid;
1548         msg->timeout_ms = cm_id_priv->timeout_ms;
1549         msg->context[1] = (void *)(unsigned long)IB_CM_REQ_SENT;
1550
1551         cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
1552         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
1553
1554         trace_icm_send_req(&cm_id_priv->id);
1555         ret = ib_post_send_mad(msg, NULL);
1556         if (ret)
1557                 goto out_free;
1558         BUG_ON(cm_id->state != IB_CM_IDLE);
1559         cm_id->state = IB_CM_REQ_SENT;
1560         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1561         return 0;
1562 out_free:
1563         cm_free_priv_msg(msg);
1564 out_unlock:
1565         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1566         return ret;
1567 }
1568 EXPORT_SYMBOL(ib_send_cm_req);
1569
1570 static int cm_issue_rej(struct cm_port *port,
1571                         struct ib_mad_recv_wc *mad_recv_wc,
1572                         enum ib_cm_rej_reason reason,
1573                         enum cm_msg_response msg_rejected,
1574                         void *ari, u8 ari_length)
1575 {
1576         struct ib_mad_send_buf *msg = NULL;
1577         struct cm_rej_msg *rej_msg, *rcv_msg;
1578         int ret;
1579
1580         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1581         if (ret)
1582                 return ret;
1583
1584         /* We just need common CM header information.  Cast to any message. */
1585         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1586         rej_msg = (struct cm_rej_msg *) msg->mad;
1587
1588         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1589         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1590                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg));
1591         IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1592                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1593         IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected);
1594         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1595
1596         if (ari && ari_length) {
1597                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1598                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1599         }
1600
1601         trace_icm_issue_rej(
1602                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg),
1603                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1604         ret = ib_post_send_mad(msg, NULL);
1605         if (ret)
1606                 cm_free_response_msg(msg);
1607
1608         return ret;
1609 }
1610
1611 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1612 {
1613         return ((cpu_to_be16(
1614                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) ||
1615                 (ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1616                                                req_msg))));
1617 }
1618
1619 static void cm_path_set_rec_type(struct ib_device *ib_device, u32 port_num,
1620                                  struct sa_path_rec *path, union ib_gid *gid)
1621 {
1622         if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1623                 path->rec_type = SA_PATH_REC_TYPE_OPA;
1624         else
1625                 path->rec_type = SA_PATH_REC_TYPE_IB;
1626 }
1627
1628 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1629                                         struct sa_path_rec *primary_path,
1630                                         struct sa_path_rec *alt_path,
1631                                         struct ib_wc *wc)
1632 {
1633         u32 lid;
1634
1635         if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1636                 sa_path_set_dlid(primary_path, wc->slid);
1637                 sa_path_set_slid(primary_path,
1638                                  IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
1639                                          req_msg));
1640         } else {
1641                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1642                         CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg));
1643                 sa_path_set_dlid(primary_path, lid);
1644
1645                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1646                         CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg));
1647                 sa_path_set_slid(primary_path, lid);
1648         }
1649
1650         if (!cm_req_has_alt_path(req_msg))
1651                 return;
1652
1653         if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1654                 sa_path_set_dlid(alt_path,
1655                                  IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
1656                                          req_msg));
1657                 sa_path_set_slid(alt_path,
1658                                  IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
1659                                          req_msg));
1660         } else {
1661                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1662                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg));
1663                 sa_path_set_dlid(alt_path, lid);
1664
1665                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1666                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg));
1667                 sa_path_set_slid(alt_path, lid);
1668         }
1669 }
1670
1671 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1672                                      struct sa_path_rec *primary_path,
1673                                      struct sa_path_rec *alt_path,
1674                                      struct ib_wc *wc)
1675 {
1676         primary_path->dgid =
1677                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg);
1678         primary_path->sgid =
1679                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg);
1680         primary_path->flow_label =
1681                 cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg));
1682         primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg);
1683         primary_path->traffic_class =
1684                 IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg);
1685         primary_path->reversible = 1;
1686         primary_path->pkey =
1687                 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1688         primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg);
1689         primary_path->mtu_selector = IB_SA_EQ;
1690         primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1691         primary_path->rate_selector = IB_SA_EQ;
1692         primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg);
1693         primary_path->packet_life_time_selector = IB_SA_EQ;
1694         primary_path->packet_life_time =
1695                 IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg);
1696         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1697         primary_path->service_id =
1698                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1699         if (sa_path_is_roce(primary_path))
1700                 primary_path->roce.route_resolved = false;
1701
1702         if (cm_req_has_alt_path(req_msg)) {
1703                 alt_path->dgid = *IBA_GET_MEM_PTR(
1704                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg);
1705                 alt_path->sgid = *IBA_GET_MEM_PTR(
1706                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg);
1707                 alt_path->flow_label = cpu_to_be32(
1708                         IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg));
1709                 alt_path->hop_limit =
1710                         IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg);
1711                 alt_path->traffic_class =
1712                         IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg);
1713                 alt_path->reversible = 1;
1714                 alt_path->pkey =
1715                         cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1716                 alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg);
1717                 alt_path->mtu_selector = IB_SA_EQ;
1718                 alt_path->mtu =
1719                         IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1720                 alt_path->rate_selector = IB_SA_EQ;
1721                 alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg);
1722                 alt_path->packet_life_time_selector = IB_SA_EQ;
1723                 alt_path->packet_life_time =
1724                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg);
1725                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1726                 alt_path->service_id =
1727                         cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1728
1729                 if (sa_path_is_roce(alt_path))
1730                         alt_path->roce.route_resolved = false;
1731         }
1732         cm_format_path_lid_from_req(req_msg, primary_path, alt_path, wc);
1733 }
1734
1735 static u16 cm_get_bth_pkey(struct cm_work *work)
1736 {
1737         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1738         u32 port_num = work->port->port_num;
1739         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1740         u16 pkey;
1741         int ret;
1742
1743         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1744         if (ret) {
1745                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %u, pkey index %u). %d\n",
1746                                      port_num, pkey_index, ret);
1747                 return 0;
1748         }
1749
1750         return pkey;
1751 }
1752
1753 /**
1754  * cm_opa_to_ib_sgid - Convert OPA SGID to IB SGID
1755  * ULPs (such as IPoIB) do not understand OPA GIDs and will
1756  * reject them as the local_gid will not match the sgid. Therefore,
1757  * change the pathrec's SGID to an IB SGID.
1758  *
1759  * @work: Work completion
1760  * @path: Path record
1761  */
1762 static void cm_opa_to_ib_sgid(struct cm_work *work,
1763                               struct sa_path_rec *path)
1764 {
1765         struct ib_device *dev = work->port->cm_dev->ib_device;
1766         u32 port_num = work->port->port_num;
1767
1768         if (rdma_cap_opa_ah(dev, port_num) &&
1769             (ib_is_opa_gid(&path->sgid))) {
1770                 union ib_gid sgid;
1771
1772                 if (rdma_query_gid(dev, port_num, 0, &sgid)) {
1773                         dev_warn(&dev->dev,
1774                                  "Error updating sgid in CM request\n");
1775                         return;
1776                 }
1777
1778                 path->sgid = sgid;
1779         }
1780 }
1781
1782 static void cm_format_req_event(struct cm_work *work,
1783                                 struct cm_id_private *cm_id_priv,
1784                                 struct ib_cm_id *listen_id)
1785 {
1786         struct cm_req_msg *req_msg;
1787         struct ib_cm_req_event_param *param;
1788
1789         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1790         param = &work->cm_event.param.req_rcvd;
1791         param->listen_id = listen_id;
1792         param->bth_pkey = cm_get_bth_pkey(work);
1793         param->port = cm_id_priv->av.port->port_num;
1794         param->primary_path = &work->path[0];
1795         cm_opa_to_ib_sgid(work, param->primary_path);
1796         if (cm_req_has_alt_path(req_msg)) {
1797                 param->alternate_path = &work->path[1];
1798                 cm_opa_to_ib_sgid(work, param->alternate_path);
1799         } else {
1800                 param->alternate_path = NULL;
1801         }
1802         param->remote_ca_guid =
1803                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
1804         param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg);
1805         param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg);
1806         param->qp_type = cm_req_get_qp_type(req_msg);
1807         param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg);
1808         param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
1809         param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
1810         param->local_cm_response_timeout =
1811                 IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg);
1812         param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg);
1813         param->remote_cm_response_timeout =
1814                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg);
1815         param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
1816         param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
1817         param->srq = IBA_GET(CM_REQ_SRQ, req_msg);
1818         param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
1819         param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg);
1820         param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod);
1821
1822         work->cm_event.private_data =
1823                 IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg);
1824 }
1825
1826 static void cm_process_work(struct cm_id_private *cm_id_priv,
1827                             struct cm_work *work)
1828 {
1829         int ret;
1830
1831         /* We will typically only have the current event to report. */
1832         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1833         cm_free_work(work);
1834
1835         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1836                 spin_lock_irq(&cm_id_priv->lock);
1837                 work = cm_dequeue_work(cm_id_priv);
1838                 spin_unlock_irq(&cm_id_priv->lock);
1839                 if (!work)
1840                         return;
1841
1842                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1843                                                 &work->cm_event);
1844                 cm_free_work(work);
1845         }
1846         cm_deref_id(cm_id_priv);
1847         if (ret)
1848                 cm_destroy_id(&cm_id_priv->id, ret);
1849 }
1850
1851 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1852                           struct cm_id_private *cm_id_priv,
1853                           enum cm_msg_response msg_mraed, u8 service_timeout,
1854                           const void *private_data, u8 private_data_len)
1855 {
1856         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1857         IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed);
1858         IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg,
1859                 be32_to_cpu(cm_id_priv->id.local_id));
1860         IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg,
1861                 be32_to_cpu(cm_id_priv->id.remote_id));
1862         IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, service_timeout);
1863
1864         if (private_data && private_data_len)
1865                 IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data,
1866                             private_data_len);
1867 }
1868
1869 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1870                           struct cm_id_private *cm_id_priv,
1871                           enum ib_cm_rej_reason reason, void *ari,
1872                           u8 ari_length, const void *private_data,
1873                           u8 private_data_len, enum ib_cm_state state)
1874 {
1875         lockdep_assert_held(&cm_id_priv->lock);
1876
1877         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1878         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1879                 be32_to_cpu(cm_id_priv->id.remote_id));
1880
1881         switch (state) {
1882         case IB_CM_REQ_RCVD:
1883                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0));
1884                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1885                 break;
1886         case IB_CM_MRA_REQ_SENT:
1887                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1888                         be32_to_cpu(cm_id_priv->id.local_id));
1889                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1890                 break;
1891         case IB_CM_REP_RCVD:
1892         case IB_CM_MRA_REP_SENT:
1893                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1894                         be32_to_cpu(cm_id_priv->id.local_id));
1895                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP);
1896                 break;
1897         default:
1898                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1899                         be32_to_cpu(cm_id_priv->id.local_id));
1900                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg,
1901                         CM_MSG_RESPONSE_OTHER);
1902                 break;
1903         }
1904
1905         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1906         if (ari && ari_length) {
1907                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1908                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1909         }
1910
1911         if (private_data && private_data_len)
1912                 IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data,
1913                             private_data_len);
1914 }
1915
1916 static void cm_dup_req_handler(struct cm_work *work,
1917                                struct cm_id_private *cm_id_priv)
1918 {
1919         struct ib_mad_send_buf *msg = NULL;
1920         int ret;
1921
1922         atomic_long_inc(
1923                 &work->port->counters[CM_RECV_DUPLICATES][CM_REQ_COUNTER]);
1924
1925         /* Quick state check to discard duplicate REQs. */
1926         spin_lock_irq(&cm_id_priv->lock);
1927         if (cm_id_priv->id.state == IB_CM_REQ_RCVD) {
1928                 spin_unlock_irq(&cm_id_priv->lock);
1929                 return;
1930         }
1931         spin_unlock_irq(&cm_id_priv->lock);
1932
1933         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1934         if (ret)
1935                 return;
1936
1937         spin_lock_irq(&cm_id_priv->lock);
1938         switch (cm_id_priv->id.state) {
1939         case IB_CM_MRA_REQ_SENT:
1940                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1941                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1942                               cm_id_priv->private_data,
1943                               cm_id_priv->private_data_len);
1944                 break;
1945         case IB_CM_TIMEWAIT:
1946                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv,
1947                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0,
1948                               IB_CM_TIMEWAIT);
1949                 break;
1950         default:
1951                 goto unlock;
1952         }
1953         spin_unlock_irq(&cm_id_priv->lock);
1954
1955         trace_icm_send_dup_req(&cm_id_priv->id);
1956         ret = ib_post_send_mad(msg, NULL);
1957         if (ret)
1958                 goto free;
1959         return;
1960
1961 unlock: spin_unlock_irq(&cm_id_priv->lock);
1962 free:   cm_free_response_msg(msg);
1963 }
1964
1965 static struct cm_id_private *cm_match_req(struct cm_work *work,
1966                                           struct cm_id_private *cm_id_priv)
1967 {
1968         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1969         struct cm_timewait_info *timewait_info;
1970         struct cm_req_msg *req_msg;
1971
1972         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1973
1974         /* Check for possible duplicate REQ. */
1975         spin_lock_irq(&cm.lock);
1976         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1977         if (timewait_info) {
1978                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
1979                                            timewait_info->work.remote_id);
1980                 spin_unlock_irq(&cm.lock);
1981                 if (cur_cm_id_priv) {
1982                         cm_dup_req_handler(work, cur_cm_id_priv);
1983                         cm_deref_id(cur_cm_id_priv);
1984                 }
1985                 return NULL;
1986         }
1987
1988         /* Check for stale connections. */
1989         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1990         if (timewait_info) {
1991                 cm_remove_remote(cm_id_priv);
1992                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
1993                                            timewait_info->work.remote_id);
1994
1995                 spin_unlock_irq(&cm.lock);
1996                 cm_issue_rej(work->port, work->mad_recv_wc,
1997                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1998                              NULL, 0);
1999                 if (cur_cm_id_priv) {
2000                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2001                         cm_deref_id(cur_cm_id_priv);
2002                 }
2003                 return NULL;
2004         }
2005
2006         /* Find matching listen request. */
2007         listen_cm_id_priv = cm_find_listen(
2008                 cm_id_priv->id.device,
2009                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)));
2010         if (!listen_cm_id_priv) {
2011                 cm_remove_remote(cm_id_priv);
2012                 spin_unlock_irq(&cm.lock);
2013                 cm_issue_rej(work->port, work->mad_recv_wc,
2014                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
2015                              NULL, 0);
2016                 return NULL;
2017         }
2018         spin_unlock_irq(&cm.lock);
2019         return listen_cm_id_priv;
2020 }
2021
2022 /*
2023  * Work-around for inter-subnet connections.  If the LIDs are permissive,
2024  * we need to override the LID/SL data in the REQ with the LID information
2025  * in the work completion.
2026  */
2027 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
2028 {
2029         if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) {
2030                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
2031                                         req_msg)) == IB_LID_PERMISSIVE) {
2032                         IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
2033                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2034                         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl);
2035                 }
2036
2037                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
2038                                         req_msg)) == IB_LID_PERMISSIVE)
2039                         IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
2040                                 wc->dlid_path_bits);
2041         }
2042
2043         if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) {
2044                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
2045                                         req_msg)) == IB_LID_PERMISSIVE) {
2046                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
2047                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2048                         IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl);
2049                 }
2050
2051                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
2052                                         req_msg)) == IB_LID_PERMISSIVE)
2053                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
2054                                 wc->dlid_path_bits);
2055         }
2056 }
2057
2058 static int cm_req_handler(struct cm_work *work)
2059 {
2060         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
2061         struct cm_req_msg *req_msg;
2062         const struct ib_global_route *grh;
2063         const struct ib_gid_attr *gid_attr;
2064         int ret;
2065
2066         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
2067
2068         cm_id_priv =
2069                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
2070         if (IS_ERR(cm_id_priv))
2071                 return PTR_ERR(cm_id_priv);
2072
2073         cm_id_priv->id.remote_id =
2074                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg));
2075         cm_id_priv->id.service_id =
2076                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
2077         cm_id_priv->tid = req_msg->hdr.tid;
2078         cm_id_priv->timeout_ms = cm_convert_to_ms(
2079                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg));
2080         cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg);
2081         cm_id_priv->remote_qpn =
2082                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
2083         cm_id_priv->initiator_depth =
2084                 IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
2085         cm_id_priv->responder_resources =
2086                 IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
2087         cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
2088         cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
2089         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
2090         cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
2091         cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
2092         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
2093
2094         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2095                                       work->mad_recv_wc->recv_buf.grh,
2096                                       &cm_id_priv->av);
2097         if (ret)
2098                 goto destroy;
2099         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
2100                                                             id.local_id);
2101         if (IS_ERR(cm_id_priv->timewait_info)) {
2102                 ret = PTR_ERR(cm_id_priv->timewait_info);
2103                 cm_id_priv->timewait_info = NULL;
2104                 goto destroy;
2105         }
2106         cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id;
2107         cm_id_priv->timewait_info->remote_ca_guid =
2108                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
2109         cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn;
2110
2111         /*
2112          * Note that the ID pointer is not in the xarray at this point,
2113          * so this set is only visible to the local thread.
2114          */
2115         cm_id_priv->id.state = IB_CM_REQ_RCVD;
2116
2117         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
2118         if (!listen_cm_id_priv) {
2119                 trace_icm_no_listener_err(&cm_id_priv->id);
2120                 cm_id_priv->id.state = IB_CM_IDLE;
2121                 ret = -EINVAL;
2122                 goto destroy;
2123         }
2124
2125         memset(&work->path[0], 0, sizeof(work->path[0]));
2126         if (cm_req_has_alt_path(req_msg))
2127                 memset(&work->path[1], 0, sizeof(work->path[1]));
2128         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
2129         gid_attr = grh->sgid_attr;
2130
2131         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE) {
2132                 work->path[0].rec_type =
2133                         sa_conv_gid_to_pathrec_type(gid_attr->gid_type);
2134         } else {
2135                 cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
2136                 cm_path_set_rec_type(
2137                         work->port->cm_dev->ib_device, work->port->port_num,
2138                         &work->path[0],
2139                         IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID,
2140                                         req_msg));
2141         }
2142         if (cm_req_has_alt_path(req_msg))
2143                 work->path[1].rec_type = work->path[0].rec_type;
2144         cm_format_paths_from_req(req_msg, &work->path[0],
2145                                  &work->path[1], work->mad_recv_wc->wc);
2146         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
2147                 sa_path_set_dmac(&work->path[0],
2148                                  cm_id_priv->av.ah_attr.roce.dmac);
2149         work->path[0].hop_limit = grh->hop_limit;
2150
2151         /* This destroy call is needed to pair with cm_init_av_for_response */
2152         cm_destroy_av(&cm_id_priv->av);
2153         ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av);
2154         if (ret) {
2155                 int err;
2156
2157                 err = rdma_query_gid(work->port->cm_dev->ib_device,
2158                                      work->port->port_num, 0,
2159                                      &work->path[0].sgid);
2160                 if (err)
2161                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2162                                        NULL, 0, NULL, 0);
2163                 else
2164                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2165                                        &work->path[0].sgid,
2166                                        sizeof(work->path[0].sgid),
2167                                        NULL, 0);
2168                 goto rejected;
2169         }
2170         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_IB)
2171                 cm_id_priv->av.dlid_datapath =
2172                         IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg);
2173
2174         if (cm_req_has_alt_path(req_msg)) {
2175                 ret = cm_init_av_by_path(&work->path[1], NULL,
2176                                          &cm_id_priv->alt_av);
2177                 if (ret) {
2178                         ib_send_cm_rej(&cm_id_priv->id,
2179                                        IB_CM_REJ_INVALID_ALT_GID,
2180                                        &work->path[0].sgid,
2181                                        sizeof(work->path[0].sgid), NULL, 0);
2182                         goto rejected;
2183                 }
2184         }
2185
2186         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
2187         cm_id_priv->id.context = listen_cm_id_priv->id.context;
2188         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
2189
2190         /* Now MAD handlers can see the new ID */
2191         spin_lock_irq(&cm_id_priv->lock);
2192         cm_finalize_id(cm_id_priv);
2193
2194         /* Refcount belongs to the event, pairs with cm_process_work() */
2195         refcount_inc(&cm_id_priv->refcount);
2196         cm_queue_work_unlock(cm_id_priv, work);
2197         /*
2198          * Since this ID was just created and was not made visible to other MAD
2199          * handlers until the cm_finalize_id() above we know that the
2200          * cm_process_work() will deliver the event and the listen_cm_id
2201          * embedded in the event can be derefed here.
2202          */
2203         cm_deref_id(listen_cm_id_priv);
2204         return 0;
2205
2206 rejected:
2207         cm_deref_id(listen_cm_id_priv);
2208 destroy:
2209         ib_destroy_cm_id(&cm_id_priv->id);
2210         return ret;
2211 }
2212
2213 static void cm_format_rep(struct cm_rep_msg *rep_msg,
2214                           struct cm_id_private *cm_id_priv,
2215                           struct ib_cm_rep_param *param)
2216 {
2217         cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid,
2218                               param->ece.attr_mod);
2219         IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg,
2220                 be32_to_cpu(cm_id_priv->id.local_id));
2221         IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg,
2222                 be32_to_cpu(cm_id_priv->id.remote_id));
2223         IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn);
2224         IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg,
2225                 param->responder_resources);
2226         IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg,
2227                 cm_id_priv->av.port->cm_dev->ack_delay);
2228         IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted);
2229         IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count);
2230         IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg,
2231                 be64_to_cpu(cm_id_priv->id.device->node_guid));
2232
2233         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
2234                 IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg,
2235                         param->initiator_depth);
2236                 IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg,
2237                         param->flow_control);
2238                 IBA_SET(CM_REP_SRQ, rep_msg, param->srq);
2239                 IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num);
2240         } else {
2241                 IBA_SET(CM_REP_SRQ, rep_msg, 1);
2242                 IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num);
2243         }
2244
2245         IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id);
2246         IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8);
2247         IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16);
2248
2249         if (param->private_data && param->private_data_len)
2250                 IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data,
2251                             param->private_data_len);
2252 }
2253
2254 int ib_send_cm_rep(struct ib_cm_id *cm_id,
2255                    struct ib_cm_rep_param *param)
2256 {
2257         struct cm_id_private *cm_id_priv;
2258         struct ib_mad_send_buf *msg;
2259         struct cm_rep_msg *rep_msg;
2260         unsigned long flags;
2261         int ret;
2262
2263         if (param->private_data &&
2264             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
2265                 return -EINVAL;
2266
2267         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2268         spin_lock_irqsave(&cm_id_priv->lock, flags);
2269         if (cm_id->state != IB_CM_REQ_RCVD &&
2270             cm_id->state != IB_CM_MRA_REQ_SENT) {
2271                 trace_icm_send_rep_err(cm_id_priv->id.local_id, cm_id->state);
2272                 ret = -EINVAL;
2273                 goto out;
2274         }
2275
2276         msg = cm_alloc_priv_msg(cm_id_priv);
2277         if (IS_ERR(msg)) {
2278                 ret = PTR_ERR(msg);
2279                 goto out;
2280         }
2281
2282         rep_msg = (struct cm_rep_msg *) msg->mad;
2283         cm_format_rep(rep_msg, cm_id_priv, param);
2284         msg->timeout_ms = cm_id_priv->timeout_ms;
2285         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
2286
2287         trace_icm_send_rep(cm_id);
2288         ret = ib_post_send_mad(msg, NULL);
2289         if (ret)
2290                 goto out_free;
2291
2292         cm_id->state = IB_CM_REP_SENT;
2293         cm_id_priv->initiator_depth = param->initiator_depth;
2294         cm_id_priv->responder_resources = param->responder_resources;
2295         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2296         WARN_ONCE(param->qp_num & 0xFF000000,
2297                   "IBTA declares QPN to be 24 bits, but it is 0x%X\n",
2298                   param->qp_num);
2299         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2300         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2301         return 0;
2302
2303 out_free:
2304         cm_free_priv_msg(msg);
2305 out:
2306         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2307         return ret;
2308 }
2309 EXPORT_SYMBOL(ib_send_cm_rep);
2310
2311 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2312                           struct cm_id_private *cm_id_priv,
2313                           const void *private_data,
2314                           u8 private_data_len)
2315 {
2316         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2317         IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg,
2318                 be32_to_cpu(cm_id_priv->id.local_id));
2319         IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg,
2320                 be32_to_cpu(cm_id_priv->id.remote_id));
2321
2322         if (private_data && private_data_len)
2323                 IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data,
2324                             private_data_len);
2325 }
2326
2327 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2328                    const void *private_data,
2329                    u8 private_data_len)
2330 {
2331         struct cm_id_private *cm_id_priv;
2332         struct ib_mad_send_buf *msg;
2333         unsigned long flags;
2334         void *data;
2335         int ret;
2336
2337         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2338                 return -EINVAL;
2339
2340         data = cm_copy_private_data(private_data, private_data_len);
2341         if (IS_ERR(data))
2342                 return PTR_ERR(data);
2343
2344         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2345         spin_lock_irqsave(&cm_id_priv->lock, flags);
2346         if (cm_id->state != IB_CM_REP_RCVD &&
2347             cm_id->state != IB_CM_MRA_REP_SENT) {
2348                 trace_icm_send_cm_rtu_err(cm_id);
2349                 ret = -EINVAL;
2350                 goto error;
2351         }
2352
2353         msg = cm_alloc_msg(cm_id_priv);
2354         if (IS_ERR(msg)) {
2355                 ret = PTR_ERR(msg);
2356                 goto error;
2357         }
2358
2359         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2360                       private_data, private_data_len);
2361
2362         trace_icm_send_rtu(cm_id);
2363         ret = ib_post_send_mad(msg, NULL);
2364         if (ret) {
2365                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2366                 cm_free_msg(msg);
2367                 kfree(data);
2368                 return ret;
2369         }
2370
2371         cm_id->state = IB_CM_ESTABLISHED;
2372         cm_set_private_data(cm_id_priv, data, private_data_len);
2373         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2374         return 0;
2375
2376 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2377         kfree(data);
2378         return ret;
2379 }
2380 EXPORT_SYMBOL(ib_send_cm_rtu);
2381
2382 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2383 {
2384         struct cm_rep_msg *rep_msg;
2385         struct ib_cm_rep_event_param *param;
2386
2387         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2388         param = &work->cm_event.param.rep_rcvd;
2389         param->remote_ca_guid =
2390                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2391         param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg);
2392         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2393         param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg);
2394         param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2395         param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2396         param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2397         param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg);
2398         param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg);
2399         param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2400         param->srq = IBA_GET(CM_REP_SRQ, rep_msg);
2401         param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16;
2402         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8;
2403         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg);
2404         param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod);
2405
2406         work->cm_event.private_data =
2407                 IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg);
2408 }
2409
2410 static void cm_dup_rep_handler(struct cm_work *work)
2411 {
2412         struct cm_id_private *cm_id_priv;
2413         struct cm_rep_msg *rep_msg;
2414         struct ib_mad_send_buf *msg = NULL;
2415         int ret;
2416
2417         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2418         cm_id_priv = cm_acquire_id(
2419                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)),
2420                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)));
2421         if (!cm_id_priv)
2422                 return;
2423
2424         atomic_long_inc(
2425                 &work->port->counters[CM_RECV_DUPLICATES][CM_REP_COUNTER]);
2426         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2427         if (ret)
2428                 goto deref;
2429
2430         spin_lock_irq(&cm_id_priv->lock);
2431         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2432                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2433                               cm_id_priv->private_data,
2434                               cm_id_priv->private_data_len);
2435         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2436                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2437                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2438                               cm_id_priv->private_data,
2439                               cm_id_priv->private_data_len);
2440         else
2441                 goto unlock;
2442         spin_unlock_irq(&cm_id_priv->lock);
2443
2444         trace_icm_send_dup_rep(&cm_id_priv->id);
2445         ret = ib_post_send_mad(msg, NULL);
2446         if (ret)
2447                 goto free;
2448         goto deref;
2449
2450 unlock: spin_unlock_irq(&cm_id_priv->lock);
2451 free:   cm_free_response_msg(msg);
2452 deref:  cm_deref_id(cm_id_priv);
2453 }
2454
2455 static int cm_rep_handler(struct cm_work *work)
2456 {
2457         struct cm_id_private *cm_id_priv;
2458         struct cm_rep_msg *rep_msg;
2459         int ret;
2460         struct cm_id_private *cur_cm_id_priv;
2461         struct cm_timewait_info *timewait_info;
2462
2463         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2464         cm_id_priv = cm_acquire_id(
2465                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0);
2466         if (!cm_id_priv) {
2467                 cm_dup_rep_handler(work);
2468                 trace_icm_remote_no_priv_err(
2469                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2470                 return -EINVAL;
2471         }
2472
2473         cm_format_rep_event(work, cm_id_priv->qp_type);
2474
2475         spin_lock_irq(&cm_id_priv->lock);
2476         switch (cm_id_priv->id.state) {
2477         case IB_CM_REQ_SENT:
2478         case IB_CM_MRA_REQ_RCVD:
2479                 break;
2480         default:
2481                 ret = -EINVAL;
2482                 trace_icm_rep_unknown_err(
2483                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2484                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg),
2485                         cm_id_priv->id.state);
2486                 spin_unlock_irq(&cm_id_priv->lock);
2487                 goto error;
2488         }
2489
2490         cm_id_priv->timewait_info->work.remote_id =
2491                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2492         cm_id_priv->timewait_info->remote_ca_guid =
2493                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2494         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2495
2496         spin_lock(&cm.lock);
2497         /* Check for duplicate REP. */
2498         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2499                 spin_unlock(&cm.lock);
2500                 spin_unlock_irq(&cm_id_priv->lock);
2501                 ret = -EINVAL;
2502                 trace_icm_insert_failed_err(
2503                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2504                 goto error;
2505         }
2506         /* Check for a stale connection. */
2507         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2508         if (timewait_info) {
2509                 cm_remove_remote(cm_id_priv);
2510                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2511                                            timewait_info->work.remote_id);
2512
2513                 spin_unlock(&cm.lock);
2514                 spin_unlock_irq(&cm_id_priv->lock);
2515                 cm_issue_rej(work->port, work->mad_recv_wc,
2516                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2517                              NULL, 0);
2518                 ret = -EINVAL;
2519                 trace_icm_staleconn_err(
2520                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2521                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2522
2523                 if (cur_cm_id_priv) {
2524                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2525                         cm_deref_id(cur_cm_id_priv);
2526                 }
2527
2528                 goto error;
2529         }
2530         spin_unlock(&cm.lock);
2531
2532         cm_id_priv->id.state = IB_CM_REP_RCVD;
2533         cm_id_priv->id.remote_id =
2534                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2535         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2536         cm_id_priv->initiator_depth =
2537                 IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2538         cm_id_priv->responder_resources =
2539                 IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2540         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2541         cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2542         cm_id_priv->target_ack_delay =
2543                 IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2544         cm_id_priv->av.timeout =
2545                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2546                                        cm_id_priv->av.timeout - 1);
2547         cm_id_priv->alt_av.timeout =
2548                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2549                                        cm_id_priv->alt_av.timeout - 1);
2550
2551         ib_cancel_mad(cm_id_priv->msg);
2552         cm_queue_work_unlock(cm_id_priv, work);
2553         return 0;
2554
2555 error:
2556         cm_deref_id(cm_id_priv);
2557         return ret;
2558 }
2559
2560 static int cm_establish_handler(struct cm_work *work)
2561 {
2562         struct cm_id_private *cm_id_priv;
2563
2564         /* See comment in cm_establish about lookup. */
2565         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2566         if (!cm_id_priv)
2567                 return -EINVAL;
2568
2569         spin_lock_irq(&cm_id_priv->lock);
2570         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2571                 spin_unlock_irq(&cm_id_priv->lock);
2572                 goto out;
2573         }
2574
2575         ib_cancel_mad(cm_id_priv->msg);
2576         cm_queue_work_unlock(cm_id_priv, work);
2577         return 0;
2578 out:
2579         cm_deref_id(cm_id_priv);
2580         return -EINVAL;
2581 }
2582
2583 static int cm_rtu_handler(struct cm_work *work)
2584 {
2585         struct cm_id_private *cm_id_priv;
2586         struct cm_rtu_msg *rtu_msg;
2587
2588         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2589         cm_id_priv = cm_acquire_id(
2590                 cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)),
2591                 cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg)));
2592         if (!cm_id_priv)
2593                 return -EINVAL;
2594
2595         work->cm_event.private_data =
2596                 IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg);
2597
2598         spin_lock_irq(&cm_id_priv->lock);
2599         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2600             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2601                 spin_unlock_irq(&cm_id_priv->lock);
2602                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2603                                                      [CM_RTU_COUNTER]);
2604                 goto out;
2605         }
2606         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2607
2608         ib_cancel_mad(cm_id_priv->msg);
2609         cm_queue_work_unlock(cm_id_priv, work);
2610         return 0;
2611 out:
2612         cm_deref_id(cm_id_priv);
2613         return -EINVAL;
2614 }
2615
2616 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2617                           struct cm_id_private *cm_id_priv,
2618                           const void *private_data,
2619                           u8 private_data_len)
2620 {
2621         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2622                           cm_form_tid(cm_id_priv));
2623         IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg,
2624                 be32_to_cpu(cm_id_priv->id.local_id));
2625         IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg,
2626                 be32_to_cpu(cm_id_priv->id.remote_id));
2627         IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg,
2628                 be32_to_cpu(cm_id_priv->remote_qpn));
2629
2630         if (private_data && private_data_len)
2631                 IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data,
2632                             private_data_len);
2633 }
2634
2635 static int cm_send_dreq_locked(struct cm_id_private *cm_id_priv,
2636                                const void *private_data, u8 private_data_len)
2637 {
2638         struct ib_mad_send_buf *msg;
2639         int ret;
2640
2641         lockdep_assert_held(&cm_id_priv->lock);
2642
2643         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2644                 return -EINVAL;
2645
2646         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2647                 trace_icm_dreq_skipped(&cm_id_priv->id);
2648                 return -EINVAL;
2649         }
2650
2651         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2652             cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2653                 ib_cancel_mad(cm_id_priv->msg);
2654
2655         msg = cm_alloc_priv_msg(cm_id_priv);
2656         if (IS_ERR(msg)) {
2657                 cm_enter_timewait(cm_id_priv);
2658                 return PTR_ERR(msg);
2659         }
2660
2661         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2662                        private_data, private_data_len);
2663         msg->timeout_ms = cm_id_priv->timeout_ms;
2664         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2665
2666         trace_icm_send_dreq(&cm_id_priv->id);
2667         ret = ib_post_send_mad(msg, NULL);
2668         if (ret) {
2669                 cm_enter_timewait(cm_id_priv);
2670                 cm_free_priv_msg(msg);
2671                 return ret;
2672         }
2673
2674         cm_id_priv->id.state = IB_CM_DREQ_SENT;
2675         return 0;
2676 }
2677
2678 int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data,
2679                     u8 private_data_len)
2680 {
2681         struct cm_id_private *cm_id_priv =
2682                 container_of(cm_id, struct cm_id_private, id);
2683         unsigned long flags;
2684         int ret;
2685
2686         spin_lock_irqsave(&cm_id_priv->lock, flags);
2687         ret = cm_send_dreq_locked(cm_id_priv, private_data, private_data_len);
2688         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2689         return ret;
2690 }
2691 EXPORT_SYMBOL(ib_send_cm_dreq);
2692
2693 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2694                           struct cm_id_private *cm_id_priv,
2695                           const void *private_data,
2696                           u8 private_data_len)
2697 {
2698         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2699         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2700                 be32_to_cpu(cm_id_priv->id.local_id));
2701         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2702                 be32_to_cpu(cm_id_priv->id.remote_id));
2703
2704         if (private_data && private_data_len)
2705                 IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data,
2706                             private_data_len);
2707 }
2708
2709 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
2710                                void *private_data, u8 private_data_len)
2711 {
2712         struct ib_mad_send_buf *msg;
2713         int ret;
2714
2715         lockdep_assert_held(&cm_id_priv->lock);
2716
2717         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2718                 return -EINVAL;
2719
2720         if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2721                 trace_icm_send_drep_err(&cm_id_priv->id);
2722                 kfree(private_data);
2723                 return -EINVAL;
2724         }
2725
2726         cm_set_private_data(cm_id_priv, private_data, private_data_len);
2727         cm_enter_timewait(cm_id_priv);
2728
2729         msg = cm_alloc_msg(cm_id_priv);
2730         if (IS_ERR(msg))
2731                 return PTR_ERR(msg);
2732
2733         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2734                        private_data, private_data_len);
2735
2736         trace_icm_send_drep(&cm_id_priv->id);
2737         ret = ib_post_send_mad(msg, NULL);
2738         if (ret) {
2739                 cm_free_msg(msg);
2740                 return ret;
2741         }
2742         return 0;
2743 }
2744
2745 int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data,
2746                     u8 private_data_len)
2747 {
2748         struct cm_id_private *cm_id_priv =
2749                 container_of(cm_id, struct cm_id_private, id);
2750         unsigned long flags;
2751         void *data;
2752         int ret;
2753
2754         data = cm_copy_private_data(private_data, private_data_len);
2755         if (IS_ERR(data))
2756                 return PTR_ERR(data);
2757
2758         spin_lock_irqsave(&cm_id_priv->lock, flags);
2759         ret = cm_send_drep_locked(cm_id_priv, data, private_data_len);
2760         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2761         return ret;
2762 }
2763 EXPORT_SYMBOL(ib_send_cm_drep);
2764
2765 static int cm_issue_drep(struct cm_port *port,
2766                          struct ib_mad_recv_wc *mad_recv_wc)
2767 {
2768         struct ib_mad_send_buf *msg = NULL;
2769         struct cm_dreq_msg *dreq_msg;
2770         struct cm_drep_msg *drep_msg;
2771         int ret;
2772
2773         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2774         if (ret)
2775                 return ret;
2776
2777         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2778         drep_msg = (struct cm_drep_msg *) msg->mad;
2779
2780         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2781         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2782                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg));
2783         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2784                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2785
2786         trace_icm_issue_drep(
2787                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2788                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2789         ret = ib_post_send_mad(msg, NULL);
2790         if (ret)
2791                 cm_free_response_msg(msg);
2792
2793         return ret;
2794 }
2795
2796 static int cm_dreq_handler(struct cm_work *work)
2797 {
2798         struct cm_id_private *cm_id_priv;
2799         struct cm_dreq_msg *dreq_msg;
2800         struct ib_mad_send_buf *msg = NULL;
2801
2802         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2803         cm_id_priv = cm_acquire_id(
2804                 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)),
2805                 cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)));
2806         if (!cm_id_priv) {
2807                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2808                                                      [CM_DREQ_COUNTER]);
2809                 cm_issue_drep(work->port, work->mad_recv_wc);
2810                 trace_icm_no_priv_err(
2811                         IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2812                         IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2813                 return -EINVAL;
2814         }
2815
2816         work->cm_event.private_data =
2817                 IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg);
2818
2819         spin_lock_irq(&cm_id_priv->lock);
2820         if (cm_id_priv->local_qpn !=
2821             cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg)))
2822                 goto unlock;
2823
2824         switch (cm_id_priv->id.state) {
2825         case IB_CM_REP_SENT:
2826         case IB_CM_DREQ_SENT:
2827         case IB_CM_MRA_REP_RCVD:
2828                 ib_cancel_mad(cm_id_priv->msg);
2829                 break;
2830         case IB_CM_ESTABLISHED:
2831                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2832                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2833                         ib_cancel_mad(cm_id_priv->msg);
2834                 break;
2835         case IB_CM_TIMEWAIT:
2836                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2837                                                      [CM_DREQ_COUNTER]);
2838                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
2839                 if (IS_ERR(msg))
2840                         goto unlock;
2841
2842                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2843                                cm_id_priv->private_data,
2844                                cm_id_priv->private_data_len);
2845                 spin_unlock_irq(&cm_id_priv->lock);
2846
2847                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2848                     ib_post_send_mad(msg, NULL))
2849                         cm_free_response_msg(msg);
2850                 goto deref;
2851         case IB_CM_DREQ_RCVD:
2852                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
2853                                                      [CM_DREQ_COUNTER]);
2854                 goto unlock;
2855         default:
2856                 trace_icm_dreq_unknown_err(&cm_id_priv->id);
2857                 goto unlock;
2858         }
2859         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2860         cm_id_priv->tid = dreq_msg->hdr.tid;
2861         cm_queue_work_unlock(cm_id_priv, work);
2862         return 0;
2863
2864 unlock: spin_unlock_irq(&cm_id_priv->lock);
2865 deref:  cm_deref_id(cm_id_priv);
2866         return -EINVAL;
2867 }
2868
2869 static int cm_drep_handler(struct cm_work *work)
2870 {
2871         struct cm_id_private *cm_id_priv;
2872         struct cm_drep_msg *drep_msg;
2873
2874         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2875         cm_id_priv = cm_acquire_id(
2876                 cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)),
2877                 cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg)));
2878         if (!cm_id_priv)
2879                 return -EINVAL;
2880
2881         work->cm_event.private_data =
2882                 IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg);
2883
2884         spin_lock_irq(&cm_id_priv->lock);
2885         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2886             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2887                 spin_unlock_irq(&cm_id_priv->lock);
2888                 goto out;
2889         }
2890         cm_enter_timewait(cm_id_priv);
2891
2892         ib_cancel_mad(cm_id_priv->msg);
2893         cm_queue_work_unlock(cm_id_priv, work);
2894         return 0;
2895 out:
2896         cm_deref_id(cm_id_priv);
2897         return -EINVAL;
2898 }
2899
2900 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
2901                               enum ib_cm_rej_reason reason, void *ari,
2902                               u8 ari_length, const void *private_data,
2903                               u8 private_data_len)
2904 {
2905         enum ib_cm_state state = cm_id_priv->id.state;
2906         struct ib_mad_send_buf *msg;
2907         int ret;
2908
2909         lockdep_assert_held(&cm_id_priv->lock);
2910
2911         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2912             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2913                 return -EINVAL;
2914
2915         trace_icm_send_rej(&cm_id_priv->id, reason);
2916
2917         switch (state) {
2918         case IB_CM_REQ_SENT:
2919         case IB_CM_MRA_REQ_RCVD:
2920         case IB_CM_REQ_RCVD:
2921         case IB_CM_MRA_REQ_SENT:
2922         case IB_CM_REP_RCVD:
2923         case IB_CM_MRA_REP_SENT:
2924                 cm_reset_to_idle(cm_id_priv);
2925                 msg = cm_alloc_msg(cm_id_priv);
2926                 if (IS_ERR(msg))
2927                         return PTR_ERR(msg);
2928                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2929                               ari, ari_length, private_data, private_data_len,
2930                               state);
2931                 break;
2932         case IB_CM_REP_SENT:
2933         case IB_CM_MRA_REP_RCVD:
2934                 cm_enter_timewait(cm_id_priv);
2935                 msg = cm_alloc_msg(cm_id_priv);
2936                 if (IS_ERR(msg))
2937                         return PTR_ERR(msg);
2938                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2939                               ari, ari_length, private_data, private_data_len,
2940                               state);
2941                 break;
2942         default:
2943                 trace_icm_send_unknown_rej_err(&cm_id_priv->id);
2944                 return -EINVAL;
2945         }
2946
2947         ret = ib_post_send_mad(msg, NULL);
2948         if (ret) {
2949                 cm_free_msg(msg);
2950                 return ret;
2951         }
2952
2953         return 0;
2954 }
2955
2956 int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason,
2957                    void *ari, u8 ari_length, const void *private_data,
2958                    u8 private_data_len)
2959 {
2960         struct cm_id_private *cm_id_priv =
2961                 container_of(cm_id, struct cm_id_private, id);
2962         unsigned long flags;
2963         int ret;
2964
2965         spin_lock_irqsave(&cm_id_priv->lock, flags);
2966         ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length,
2967                                  private_data, private_data_len);
2968         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2969         return ret;
2970 }
2971 EXPORT_SYMBOL(ib_send_cm_rej);
2972
2973 static void cm_format_rej_event(struct cm_work *work)
2974 {
2975         struct cm_rej_msg *rej_msg;
2976         struct ib_cm_rej_event_param *param;
2977
2978         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2979         param = &work->cm_event.param.rej_rcvd;
2980         param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg);
2981         param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg);
2982         param->reason = IBA_GET(CM_REJ_REASON, rej_msg);
2983         work->cm_event.private_data =
2984                 IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg);
2985 }
2986
2987 static struct cm_id_private *cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2988 {
2989         struct cm_id_private *cm_id_priv;
2990         __be32 remote_id;
2991
2992         remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg));
2993
2994         if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) {
2995                 cm_id_priv = cm_find_remote_id(
2996                         *((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)),
2997                         remote_id);
2998         } else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) ==
2999                    CM_MSG_RESPONSE_REQ)
3000                 cm_id_priv = cm_acquire_id(
3001                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3002                         0);
3003         else
3004                 cm_id_priv = cm_acquire_id(
3005                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3006                         remote_id);
3007
3008         return cm_id_priv;
3009 }
3010
3011 static int cm_rej_handler(struct cm_work *work)
3012 {
3013         struct cm_id_private *cm_id_priv;
3014         struct cm_rej_msg *rej_msg;
3015
3016         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
3017         cm_id_priv = cm_acquire_rejected_id(rej_msg);
3018         if (!cm_id_priv)
3019                 return -EINVAL;
3020
3021         cm_format_rej_event(work);
3022
3023         spin_lock_irq(&cm_id_priv->lock);
3024         switch (cm_id_priv->id.state) {
3025         case IB_CM_REQ_SENT:
3026         case IB_CM_MRA_REQ_RCVD:
3027         case IB_CM_REP_SENT:
3028         case IB_CM_MRA_REP_RCVD:
3029                 ib_cancel_mad(cm_id_priv->msg);
3030                 fallthrough;
3031         case IB_CM_REQ_RCVD:
3032         case IB_CM_MRA_REQ_SENT:
3033                 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN)
3034                         cm_enter_timewait(cm_id_priv);
3035                 else
3036                         cm_reset_to_idle(cm_id_priv);
3037                 break;
3038         case IB_CM_DREQ_SENT:
3039                 ib_cancel_mad(cm_id_priv->msg);
3040                 fallthrough;
3041         case IB_CM_REP_RCVD:
3042         case IB_CM_MRA_REP_SENT:
3043                 cm_enter_timewait(cm_id_priv);
3044                 break;
3045         case IB_CM_ESTABLISHED:
3046                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
3047                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
3048                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
3049                                 ib_cancel_mad(cm_id_priv->msg);
3050                         cm_enter_timewait(cm_id_priv);
3051                         break;
3052                 }
3053                 fallthrough;
3054         default:
3055                 trace_icm_rej_unknown_err(&cm_id_priv->id);
3056                 spin_unlock_irq(&cm_id_priv->lock);
3057                 goto out;
3058         }
3059
3060         cm_queue_work_unlock(cm_id_priv, work);
3061         return 0;
3062 out:
3063         cm_deref_id(cm_id_priv);
3064         return -EINVAL;
3065 }
3066
3067 int ib_send_cm_mra(struct ib_cm_id *cm_id,
3068                    u8 service_timeout,
3069                    const void *private_data,
3070                    u8 private_data_len)
3071 {
3072         struct cm_id_private *cm_id_priv;
3073         struct ib_mad_send_buf *msg;
3074         enum ib_cm_state cm_state;
3075         enum ib_cm_lap_state lap_state;
3076         enum cm_msg_response msg_response;
3077         void *data;
3078         unsigned long flags;
3079         int ret;
3080
3081         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
3082                 return -EINVAL;
3083
3084         data = cm_copy_private_data(private_data, private_data_len);
3085         if (IS_ERR(data))
3086                 return PTR_ERR(data);
3087
3088         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3089
3090         spin_lock_irqsave(&cm_id_priv->lock, flags);
3091         switch (cm_id_priv->id.state) {
3092         case IB_CM_REQ_RCVD:
3093                 cm_state = IB_CM_MRA_REQ_SENT;
3094                 lap_state = cm_id->lap_state;
3095                 msg_response = CM_MSG_RESPONSE_REQ;
3096                 break;
3097         case IB_CM_REP_RCVD:
3098                 cm_state = IB_CM_MRA_REP_SENT;
3099                 lap_state = cm_id->lap_state;
3100                 msg_response = CM_MSG_RESPONSE_REP;
3101                 break;
3102         case IB_CM_ESTABLISHED:
3103                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
3104                         cm_state = cm_id->state;
3105                         lap_state = IB_CM_MRA_LAP_SENT;
3106                         msg_response = CM_MSG_RESPONSE_OTHER;
3107                         break;
3108                 }
3109                 fallthrough;
3110         default:
3111                 trace_icm_send_mra_unknown_err(&cm_id_priv->id);
3112                 ret = -EINVAL;
3113                 goto error_unlock;
3114         }
3115
3116         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
3117                 msg = cm_alloc_msg(cm_id_priv);
3118                 if (IS_ERR(msg)) {
3119                         ret = PTR_ERR(msg);
3120                         goto error_unlock;
3121                 }
3122
3123                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3124                               msg_response, service_timeout,
3125                               private_data, private_data_len);
3126                 trace_icm_send_mra(cm_id);
3127                 ret = ib_post_send_mad(msg, NULL);
3128                 if (ret)
3129                         goto error_free_msg;
3130         }
3131
3132         cm_id->state = cm_state;
3133         cm_id->lap_state = lap_state;
3134         cm_id_priv->service_timeout = service_timeout;
3135         cm_set_private_data(cm_id_priv, data, private_data_len);
3136         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3137         return 0;
3138
3139 error_free_msg:
3140         cm_free_msg(msg);
3141 error_unlock:
3142         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3143         kfree(data);
3144         return ret;
3145 }
3146 EXPORT_SYMBOL(ib_send_cm_mra);
3147
3148 static struct cm_id_private *cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
3149 {
3150         switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) {
3151         case CM_MSG_RESPONSE_REQ:
3152                 return cm_acquire_id(
3153                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3154                         0);
3155         case CM_MSG_RESPONSE_REP:
3156         case CM_MSG_RESPONSE_OTHER:
3157                 return cm_acquire_id(
3158                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3159                         cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg)));
3160         default:
3161                 return NULL;
3162         }
3163 }
3164
3165 static int cm_mra_handler(struct cm_work *work)
3166 {
3167         struct cm_id_private *cm_id_priv;
3168         struct cm_mra_msg *mra_msg;
3169         int timeout;
3170
3171         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
3172         cm_id_priv = cm_acquire_mraed_id(mra_msg);
3173         if (!cm_id_priv)
3174                 return -EINVAL;
3175
3176         work->cm_event.private_data =
3177                 IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg);
3178         work->cm_event.param.mra_rcvd.service_timeout =
3179                 IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg);
3180         timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) +
3181                   cm_convert_to_ms(cm_id_priv->av.timeout);
3182
3183         spin_lock_irq(&cm_id_priv->lock);
3184         switch (cm_id_priv->id.state) {
3185         case IB_CM_REQ_SENT:
3186                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3187                             CM_MSG_RESPONSE_REQ ||
3188                     ib_modify_mad(cm_id_priv->msg, timeout))
3189                         goto out;
3190                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
3191                 break;
3192         case IB_CM_REP_SENT:
3193                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3194                             CM_MSG_RESPONSE_REP ||
3195                     ib_modify_mad(cm_id_priv->msg, timeout))
3196                         goto out;
3197                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
3198                 break;
3199         case IB_CM_ESTABLISHED:
3200                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3201                             CM_MSG_RESPONSE_OTHER ||
3202                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
3203                     ib_modify_mad(cm_id_priv->msg, timeout)) {
3204                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
3205                                 atomic_long_inc(
3206                                         &work->port->counters[CM_RECV_DUPLICATES]
3207                                                              [CM_MRA_COUNTER]);
3208                         goto out;
3209                 }
3210                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
3211                 break;
3212         case IB_CM_MRA_REQ_RCVD:
3213         case IB_CM_MRA_REP_RCVD:
3214                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3215                                                      [CM_MRA_COUNTER]);
3216                 fallthrough;
3217         default:
3218                 trace_icm_mra_unknown_err(&cm_id_priv->id);
3219                 goto out;
3220         }
3221
3222         cm_id_priv->msg->context[1] = (void *) (unsigned long)
3223                                       cm_id_priv->id.state;
3224         cm_queue_work_unlock(cm_id_priv, work);
3225         return 0;
3226 out:
3227         spin_unlock_irq(&cm_id_priv->lock);
3228         cm_deref_id(cm_id_priv);
3229         return -EINVAL;
3230 }
3231
3232 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3233                                         struct sa_path_rec *path)
3234 {
3235         u32 lid;
3236
3237         if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3238                 sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID,
3239                                                lap_msg));
3240                 sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID,
3241                                                lap_msg));
3242         } else {
3243                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3244                         CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg));
3245                 sa_path_set_dlid(path, lid);
3246
3247                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3248                         CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg));
3249                 sa_path_set_slid(path, lid);
3250         }
3251 }
3252
3253 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3254                                     struct sa_path_rec *path,
3255                                     struct cm_lap_msg *lap_msg)
3256 {
3257         path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg);
3258         path->sgid =
3259                 *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg);
3260         path->flow_label =
3261                 cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg));
3262         path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg);
3263         path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg);
3264         path->reversible = 1;
3265         path->pkey = cm_id_priv->pkey;
3266         path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg);
3267         path->mtu_selector = IB_SA_EQ;
3268         path->mtu = cm_id_priv->path_mtu;
3269         path->rate_selector = IB_SA_EQ;
3270         path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg);
3271         path->packet_life_time_selector = IB_SA_EQ;
3272         path->packet_life_time =
3273                 IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg);
3274         path->packet_life_time -= (path->packet_life_time > 0);
3275         cm_format_path_lid_from_lap(lap_msg, path);
3276 }
3277
3278 static int cm_lap_handler(struct cm_work *work)
3279 {
3280         struct cm_id_private *cm_id_priv;
3281         struct cm_lap_msg *lap_msg;
3282         struct ib_cm_lap_event_param *param;
3283         struct ib_mad_send_buf *msg = NULL;
3284         struct rdma_ah_attr ah_attr;
3285         struct cm_av alt_av = {};
3286         int ret;
3287
3288         /* Currently Alternate path messages are not supported for
3289          * RoCE link layer.
3290          */
3291         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3292                                work->port->port_num))
3293                 return -EINVAL;
3294
3295         /* todo: verify LAP request and send reject APR if invalid. */
3296         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3297         cm_id_priv = cm_acquire_id(
3298                 cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)),
3299                 cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg)));
3300         if (!cm_id_priv)
3301                 return -EINVAL;
3302
3303         param = &work->cm_event.param.lap_rcvd;
3304         memset(&work->path[0], 0, sizeof(work->path[1]));
3305         cm_path_set_rec_type(work->port->cm_dev->ib_device,
3306                              work->port->port_num, &work->path[0],
3307                              IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID,
3308                                              lap_msg));
3309         param->alternate_path = &work->path[0];
3310         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3311         work->cm_event.private_data =
3312                 IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg);
3313
3314         ret = ib_init_ah_attr_from_wc(work->port->cm_dev->ib_device,
3315                                       work->port->port_num,
3316                                       work->mad_recv_wc->wc,
3317                                       work->mad_recv_wc->recv_buf.grh,
3318                                       &ah_attr);
3319         if (ret)
3320                 goto deref;
3321
3322         ret = cm_init_av_by_path(param->alternate_path, NULL, &alt_av);
3323         if (ret) {
3324                 rdma_destroy_ah_attr(&ah_attr);
3325                 goto deref;
3326         }
3327
3328         spin_lock_irq(&cm_id_priv->lock);
3329         cm_init_av_for_lap(work->port, work->mad_recv_wc->wc,
3330                            &ah_attr, &cm_id_priv->av);
3331         cm_move_av_from_path(&cm_id_priv->alt_av, &alt_av);
3332
3333         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3334                 goto unlock;
3335
3336         switch (cm_id_priv->id.lap_state) {
3337         case IB_CM_LAP_UNINIT:
3338         case IB_CM_LAP_IDLE:
3339                 break;
3340         case IB_CM_MRA_LAP_SENT:
3341                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3342                                                      [CM_LAP_COUNTER]);
3343                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
3344                 if (IS_ERR(msg))
3345                         goto unlock;
3346
3347                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3348                               CM_MSG_RESPONSE_OTHER,
3349                               cm_id_priv->service_timeout,
3350                               cm_id_priv->private_data,
3351                               cm_id_priv->private_data_len);
3352                 spin_unlock_irq(&cm_id_priv->lock);
3353
3354                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3355                     ib_post_send_mad(msg, NULL))
3356                         cm_free_response_msg(msg);
3357                 goto deref;
3358         case IB_CM_LAP_RCVD:
3359                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3360                                                      [CM_LAP_COUNTER]);
3361                 goto unlock;
3362         default:
3363                 goto unlock;
3364         }
3365
3366         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3367         cm_id_priv->tid = lap_msg->hdr.tid;
3368         cm_queue_work_unlock(cm_id_priv, work);
3369         return 0;
3370
3371 unlock: spin_unlock_irq(&cm_id_priv->lock);
3372 deref:  cm_deref_id(cm_id_priv);
3373         return -EINVAL;
3374 }
3375
3376 static int cm_apr_handler(struct cm_work *work)
3377 {
3378         struct cm_id_private *cm_id_priv;
3379         struct cm_apr_msg *apr_msg;
3380
3381         /* Currently Alternate path messages are not supported for
3382          * RoCE link layer.
3383          */
3384         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3385                                work->port->port_num))
3386                 return -EINVAL;
3387
3388         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3389         cm_id_priv = cm_acquire_id(
3390                 cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)),
3391                 cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg)));
3392         if (!cm_id_priv)
3393                 return -EINVAL; /* Unmatched reply. */
3394
3395         work->cm_event.param.apr_rcvd.ap_status =
3396                 IBA_GET(CM_APR_AR_STATUS, apr_msg);
3397         work->cm_event.param.apr_rcvd.apr_info =
3398                 IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg);
3399         work->cm_event.param.apr_rcvd.info_len =
3400                 IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg);
3401         work->cm_event.private_data =
3402                 IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg);
3403
3404         spin_lock_irq(&cm_id_priv->lock);
3405         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3406             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3407              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3408                 spin_unlock_irq(&cm_id_priv->lock);
3409                 goto out;
3410         }
3411         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3412         ib_cancel_mad(cm_id_priv->msg);
3413         cm_queue_work_unlock(cm_id_priv, work);
3414         return 0;
3415 out:
3416         cm_deref_id(cm_id_priv);
3417         return -EINVAL;
3418 }
3419
3420 static int cm_timewait_handler(struct cm_work *work)
3421 {
3422         struct cm_timewait_info *timewait_info;
3423         struct cm_id_private *cm_id_priv;
3424
3425         timewait_info = container_of(work, struct cm_timewait_info, work);
3426         spin_lock_irq(&cm.lock);
3427         list_del(&timewait_info->list);
3428         spin_unlock_irq(&cm.lock);
3429
3430         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3431                                    timewait_info->work.remote_id);
3432         if (!cm_id_priv)
3433                 return -EINVAL;
3434
3435         spin_lock_irq(&cm_id_priv->lock);
3436         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3437             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3438                 spin_unlock_irq(&cm_id_priv->lock);
3439                 goto out;
3440         }
3441         cm_id_priv->id.state = IB_CM_IDLE;
3442         cm_queue_work_unlock(cm_id_priv, work);
3443         return 0;
3444 out:
3445         cm_deref_id(cm_id_priv);
3446         return -EINVAL;
3447 }
3448
3449 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3450                                struct cm_id_private *cm_id_priv,
3451                                struct ib_cm_sidr_req_param *param)
3452 {
3453         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3454                           cm_form_tid(cm_id_priv));
3455         IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg,
3456                 be32_to_cpu(cm_id_priv->id.local_id));
3457         IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg,
3458                 be16_to_cpu(param->path->pkey));
3459         IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg,
3460                 be64_to_cpu(param->service_id));
3461
3462         if (param->private_data && param->private_data_len)
3463                 IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg,
3464                             param->private_data, param->private_data_len);
3465 }
3466
3467 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3468                         struct ib_cm_sidr_req_param *param)
3469 {
3470         struct cm_id_private *cm_id_priv;
3471         struct ib_mad_send_buf *msg;
3472         struct cm_av av = {};
3473         unsigned long flags;
3474         int ret;
3475
3476         if (!param->path || (param->private_data &&
3477              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3478                 return -EINVAL;
3479
3480         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3481         ret = cm_init_av_by_path(param->path, param->sgid_attr, &av);
3482         if (ret)
3483                 return ret;
3484
3485         spin_lock_irqsave(&cm_id_priv->lock, flags);
3486         cm_move_av_from_path(&cm_id_priv->av, &av);
3487         cm_id->service_id = param->service_id;
3488         cm_id_priv->timeout_ms = param->timeout_ms;
3489         cm_id_priv->max_cm_retries = param->max_cm_retries;
3490         if (cm_id->state != IB_CM_IDLE) {
3491                 ret = -EINVAL;
3492                 goto out_unlock;
3493         }
3494
3495         msg = cm_alloc_priv_msg(cm_id_priv);
3496         if (IS_ERR(msg)) {
3497                 ret = PTR_ERR(msg);
3498                 goto out_unlock;
3499         }
3500
3501         cm_format_sidr_req((struct cm_sidr_req_msg *)msg->mad, cm_id_priv,
3502                            param);
3503         msg->timeout_ms = cm_id_priv->timeout_ms;
3504         msg->context[1] = (void *)(unsigned long)IB_CM_SIDR_REQ_SENT;
3505
3506         trace_icm_send_sidr_req(&cm_id_priv->id);
3507         ret = ib_post_send_mad(msg, NULL);
3508         if (ret)
3509                 goto out_free;
3510         cm_id->state = IB_CM_SIDR_REQ_SENT;
3511         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3512         return 0;
3513 out_free:
3514         cm_free_priv_msg(msg);
3515 out_unlock:
3516         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3517         return ret;
3518 }
3519 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3520
3521 static void cm_format_sidr_req_event(struct cm_work *work,
3522                                      const struct cm_id_private *rx_cm_id,
3523                                      struct ib_cm_id *listen_id)
3524 {
3525         struct cm_sidr_req_msg *sidr_req_msg;
3526         struct ib_cm_sidr_req_event_param *param;
3527
3528         sidr_req_msg = (struct cm_sidr_req_msg *)
3529                                 work->mad_recv_wc->recv_buf.mad;
3530         param = &work->cm_event.param.sidr_req_rcvd;
3531         param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg);
3532         param->listen_id = listen_id;
3533         param->service_id =
3534                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3535         param->bth_pkey = cm_get_bth_pkey(work);
3536         param->port = work->port->port_num;
3537         param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr;
3538         work->cm_event.private_data =
3539                 IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg);
3540 }
3541
3542 static int cm_sidr_req_handler(struct cm_work *work)
3543 {
3544         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
3545         struct cm_sidr_req_msg *sidr_req_msg;
3546         struct ib_wc *wc;
3547         int ret;
3548
3549         cm_id_priv =
3550                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
3551         if (IS_ERR(cm_id_priv))
3552                 return PTR_ERR(cm_id_priv);
3553
3554         /* Record SGID/SLID and request ID for lookup. */
3555         sidr_req_msg = (struct cm_sidr_req_msg *)
3556                                 work->mad_recv_wc->recv_buf.mad;
3557
3558         cm_id_priv->id.remote_id =
3559                 cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg));
3560         cm_id_priv->id.service_id =
3561                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3562         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3563
3564         wc = work->mad_recv_wc->wc;
3565         cm_id_priv->sidr_slid = wc->slid;
3566         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3567                                       work->mad_recv_wc->recv_buf.grh,
3568                                       &cm_id_priv->av);
3569         if (ret)
3570                 goto out;
3571
3572         spin_lock_irq(&cm.lock);
3573         listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3574         if (listen_cm_id_priv) {
3575                 spin_unlock_irq(&cm.lock);
3576                 atomic_long_inc(&work->port->counters[CM_RECV_DUPLICATES]
3577                                                      [CM_SIDR_REQ_COUNTER]);
3578                 goto out; /* Duplicate message. */
3579         }
3580         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3581         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
3582                                            cm_id_priv->id.service_id);
3583         if (!listen_cm_id_priv) {
3584                 spin_unlock_irq(&cm.lock);
3585                 ib_send_cm_sidr_rep(&cm_id_priv->id,
3586                                     &(struct ib_cm_sidr_rep_param){
3587                                             .status = IB_SIDR_UNSUPPORTED });
3588                 goto out; /* No match. */
3589         }
3590         spin_unlock_irq(&cm.lock);
3591
3592         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
3593         cm_id_priv->id.context = listen_cm_id_priv->id.context;
3594
3595         /*
3596          * A SIDR ID does not need to be in the xarray since it does not receive
3597          * mads, is not placed in the remote_id or remote_qpn rbtree, and does
3598          * not enter timewait.
3599          */
3600
3601         cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
3602         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
3603         cm_free_work(work);
3604         /*
3605          * A pointer to the listen_cm_id is held in the event, so this deref
3606          * must be after the event is delivered above.
3607          */
3608         cm_deref_id(listen_cm_id_priv);
3609         if (ret)
3610                 cm_destroy_id(&cm_id_priv->id, ret);
3611         return 0;
3612 out:
3613         ib_destroy_cm_id(&cm_id_priv->id);
3614         return -EINVAL;
3615 }
3616
3617 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3618                                struct cm_id_private *cm_id_priv,
3619                                struct ib_cm_sidr_rep_param *param)
3620 {
3621         cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3622                               cm_id_priv->tid, param->ece.attr_mod);
3623         IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg,
3624                 be32_to_cpu(cm_id_priv->id.remote_id));
3625         IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status);
3626         IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num);
3627         IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg,
3628                 be64_to_cpu(cm_id_priv->id.service_id));
3629         IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey);
3630         IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg,
3631                 param->ece.vendor_id & 0xFF);
3632         IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg,
3633                 (param->ece.vendor_id >> 8) & 0xFF);
3634
3635         if (param->info && param->info_length)
3636                 IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg,
3637                             param->info, param->info_length);
3638
3639         if (param->private_data && param->private_data_len)
3640                 IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg,
3641                             param->private_data, param->private_data_len);
3642 }
3643
3644 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
3645                                    struct ib_cm_sidr_rep_param *param)
3646 {
3647         struct ib_mad_send_buf *msg;
3648         unsigned long flags;
3649         int ret;
3650
3651         lockdep_assert_held(&cm_id_priv->lock);
3652
3653         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3654             (param->private_data &&
3655              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3656                 return -EINVAL;
3657
3658         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD)
3659                 return -EINVAL;
3660
3661         msg = cm_alloc_msg(cm_id_priv);
3662         if (IS_ERR(msg))
3663                 return PTR_ERR(msg);
3664
3665         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3666                            param);
3667         trace_icm_send_sidr_rep(&cm_id_priv->id);
3668         ret = ib_post_send_mad(msg, NULL);
3669         if (ret) {
3670                 cm_free_msg(msg);
3671                 return ret;
3672         }
3673         cm_id_priv->id.state = IB_CM_IDLE;
3674         spin_lock_irqsave(&cm.lock, flags);
3675         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3676                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3677                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3678         }
3679         spin_unlock_irqrestore(&cm.lock, flags);
3680         return 0;
3681 }
3682
3683 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3684                         struct ib_cm_sidr_rep_param *param)
3685 {
3686         struct cm_id_private *cm_id_priv =
3687                 container_of(cm_id, struct cm_id_private, id);
3688         unsigned long flags;
3689         int ret;
3690
3691         spin_lock_irqsave(&cm_id_priv->lock, flags);
3692         ret = cm_send_sidr_rep_locked(cm_id_priv, param);
3693         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3694         return ret;
3695 }
3696 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3697
3698 static void cm_format_sidr_rep_event(struct cm_work *work,
3699                                      const struct cm_id_private *cm_id_priv)
3700 {
3701         struct cm_sidr_rep_msg *sidr_rep_msg;
3702         struct ib_cm_sidr_rep_event_param *param;
3703
3704         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3705                                 work->mad_recv_wc->recv_buf.mad;
3706         param = &work->cm_event.param.sidr_rep_rcvd;
3707         param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg);
3708         param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg);
3709         param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg);
3710         param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION,
3711                                       sidr_rep_msg);
3712         param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH,
3713                                   sidr_rep_msg);
3714         param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
3715         work->cm_event.private_data =
3716                 IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg);
3717 }
3718
3719 static int cm_sidr_rep_handler(struct cm_work *work)
3720 {
3721         struct cm_sidr_rep_msg *sidr_rep_msg;
3722         struct cm_id_private *cm_id_priv;
3723
3724         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3725                                 work->mad_recv_wc->recv_buf.mad;
3726         cm_id_priv = cm_acquire_id(
3727                 cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0);
3728         if (!cm_id_priv)
3729                 return -EINVAL; /* Unmatched reply. */
3730
3731         spin_lock_irq(&cm_id_priv->lock);
3732         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3733                 spin_unlock_irq(&cm_id_priv->lock);
3734                 goto out;
3735         }
3736         cm_id_priv->id.state = IB_CM_IDLE;
3737         ib_cancel_mad(cm_id_priv->msg);
3738         spin_unlock_irq(&cm_id_priv->lock);
3739
3740         cm_format_sidr_rep_event(work, cm_id_priv);
3741         cm_process_work(cm_id_priv, work);
3742         return 0;
3743 out:
3744         cm_deref_id(cm_id_priv);
3745         return -EINVAL;
3746 }
3747
3748 static void cm_process_send_error(struct cm_id_private *cm_id_priv,
3749                                   struct ib_mad_send_buf *msg,
3750                                   enum ib_cm_state state,
3751                                   enum ib_wc_status wc_status)
3752 {
3753         struct ib_cm_event cm_event = {};
3754         int ret;
3755
3756         /* Discard old sends or ones without a response. */
3757         spin_lock_irq(&cm_id_priv->lock);
3758         if (msg != cm_id_priv->msg) {
3759                 spin_unlock_irq(&cm_id_priv->lock);
3760                 cm_free_msg(msg);
3761                 return;
3762         }
3763         cm_free_priv_msg(msg);
3764
3765         if (state != cm_id_priv->id.state || wc_status == IB_WC_SUCCESS ||
3766             wc_status == IB_WC_WR_FLUSH_ERR)
3767                 goto out_unlock;
3768
3769         trace_icm_mad_send_err(state, wc_status);
3770         switch (state) {
3771         case IB_CM_REQ_SENT:
3772         case IB_CM_MRA_REQ_RCVD:
3773                 cm_reset_to_idle(cm_id_priv);
3774                 cm_event.event = IB_CM_REQ_ERROR;
3775                 break;
3776         case IB_CM_REP_SENT:
3777         case IB_CM_MRA_REP_RCVD:
3778                 cm_reset_to_idle(cm_id_priv);
3779                 cm_event.event = IB_CM_REP_ERROR;
3780                 break;
3781         case IB_CM_DREQ_SENT:
3782                 cm_enter_timewait(cm_id_priv);
3783                 cm_event.event = IB_CM_DREQ_ERROR;
3784                 break;
3785         case IB_CM_SIDR_REQ_SENT:
3786                 cm_id_priv->id.state = IB_CM_IDLE;
3787                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3788                 break;
3789         default:
3790                 goto out_unlock;
3791         }
3792         spin_unlock_irq(&cm_id_priv->lock);
3793         cm_event.param.send_status = wc_status;
3794
3795         /* No other events can occur on the cm_id at this point. */
3796         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3797         if (ret)
3798                 ib_destroy_cm_id(&cm_id_priv->id);
3799         return;
3800 out_unlock:
3801         spin_unlock_irq(&cm_id_priv->lock);
3802 }
3803
3804 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3805                             struct ib_mad_send_wc *mad_send_wc)
3806 {
3807         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3808         struct cm_id_private *cm_id_priv = msg->context[0];
3809         enum ib_cm_state state =
3810                 (enum ib_cm_state)(unsigned long)msg->context[1];
3811         struct cm_port *port;
3812         u16 attr_index;
3813
3814         port = mad_agent->context;
3815         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3816                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3817
3818         /*
3819          * If the send was in response to a received message (context[0] is not
3820          * set to a cm_id), and is not a REJ, then it is a send that was
3821          * manually retried.
3822          */
3823         if (!cm_id_priv && (attr_index != CM_REJ_COUNTER))
3824                 msg->retries = 1;
3825
3826         atomic_long_add(1 + msg->retries, &port->counters[CM_XMIT][attr_index]);
3827         if (msg->retries)
3828                 atomic_long_add(msg->retries,
3829                                 &port->counters[CM_XMIT_RETRIES][attr_index]);
3830
3831         if (cm_id_priv)
3832                 cm_process_send_error(cm_id_priv, msg, state,
3833                                       mad_send_wc->status);
3834         else
3835                 cm_free_response_msg(msg);
3836 }
3837
3838 static void cm_work_handler(struct work_struct *_work)
3839 {
3840         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3841         int ret;
3842
3843         switch (work->cm_event.event) {
3844         case IB_CM_REQ_RECEIVED:
3845                 ret = cm_req_handler(work);
3846                 break;
3847         case IB_CM_MRA_RECEIVED:
3848                 ret = cm_mra_handler(work);
3849                 break;
3850         case IB_CM_REJ_RECEIVED:
3851                 ret = cm_rej_handler(work);
3852                 break;
3853         case IB_CM_REP_RECEIVED:
3854                 ret = cm_rep_handler(work);
3855                 break;
3856         case IB_CM_RTU_RECEIVED:
3857                 ret = cm_rtu_handler(work);
3858                 break;
3859         case IB_CM_USER_ESTABLISHED:
3860                 ret = cm_establish_handler(work);
3861                 break;
3862         case IB_CM_DREQ_RECEIVED:
3863                 ret = cm_dreq_handler(work);
3864                 break;
3865         case IB_CM_DREP_RECEIVED:
3866                 ret = cm_drep_handler(work);
3867                 break;
3868         case IB_CM_SIDR_REQ_RECEIVED:
3869                 ret = cm_sidr_req_handler(work);
3870                 break;
3871         case IB_CM_SIDR_REP_RECEIVED:
3872                 ret = cm_sidr_rep_handler(work);
3873                 break;
3874         case IB_CM_LAP_RECEIVED:
3875                 ret = cm_lap_handler(work);
3876                 break;
3877         case IB_CM_APR_RECEIVED:
3878                 ret = cm_apr_handler(work);
3879                 break;
3880         case IB_CM_TIMEWAIT_EXIT:
3881                 ret = cm_timewait_handler(work);
3882                 break;
3883         default:
3884                 trace_icm_handler_err(work->cm_event.event);
3885                 ret = -EINVAL;
3886                 break;
3887         }
3888         if (ret)
3889                 cm_free_work(work);
3890 }
3891
3892 static int cm_establish(struct ib_cm_id *cm_id)
3893 {
3894         struct cm_id_private *cm_id_priv;
3895         struct cm_work *work;
3896         unsigned long flags;
3897         int ret = 0;
3898         struct cm_device *cm_dev;
3899
3900         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3901         if (!cm_dev)
3902                 return -ENODEV;
3903
3904         work = kmalloc(sizeof *work, GFP_ATOMIC);
3905         if (!work)
3906                 return -ENOMEM;
3907
3908         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3909         spin_lock_irqsave(&cm_id_priv->lock, flags);
3910         switch (cm_id->state) {
3911         case IB_CM_REP_SENT:
3912         case IB_CM_MRA_REP_RCVD:
3913                 cm_id->state = IB_CM_ESTABLISHED;
3914                 break;
3915         case IB_CM_ESTABLISHED:
3916                 ret = -EISCONN;
3917                 break;
3918         default:
3919                 trace_icm_establish_err(cm_id);
3920                 ret = -EINVAL;
3921                 break;
3922         }
3923         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3924
3925         if (ret) {
3926                 kfree(work);
3927                 goto out;
3928         }
3929
3930         /*
3931          * The CM worker thread may try to destroy the cm_id before it
3932          * can execute this work item.  To prevent potential deadlock,
3933          * we need to find the cm_id once we're in the context of the
3934          * worker thread, rather than holding a reference on it.
3935          */
3936         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3937         work->local_id = cm_id->local_id;
3938         work->remote_id = cm_id->remote_id;
3939         work->mad_recv_wc = NULL;
3940         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3941
3942         /* Check if the device started its remove_one */
3943         spin_lock_irqsave(&cm.lock, flags);
3944         if (!cm_dev->going_down) {
3945                 queue_delayed_work(cm.wq, &work->work, 0);
3946         } else {
3947                 kfree(work);
3948                 ret = -ENODEV;
3949         }
3950         spin_unlock_irqrestore(&cm.lock, flags);
3951
3952 out:
3953         return ret;
3954 }
3955
3956 static int cm_migrate(struct ib_cm_id *cm_id)
3957 {
3958         struct cm_id_private *cm_id_priv;
3959         unsigned long flags;
3960         int ret = 0;
3961
3962         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3963         spin_lock_irqsave(&cm_id_priv->lock, flags);
3964         if (cm_id->state == IB_CM_ESTABLISHED &&
3965             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3966              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3967                 cm_id->lap_state = IB_CM_LAP_IDLE;
3968                 cm_id_priv->av = cm_id_priv->alt_av;
3969         } else
3970                 ret = -EINVAL;
3971         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3972
3973         return ret;
3974 }
3975
3976 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3977 {
3978         int ret;
3979
3980         switch (event) {
3981         case IB_EVENT_COMM_EST:
3982                 ret = cm_establish(cm_id);
3983                 break;
3984         case IB_EVENT_PATH_MIG:
3985                 ret = cm_migrate(cm_id);
3986                 break;
3987         default:
3988                 ret = -EINVAL;
3989         }
3990         return ret;
3991 }
3992 EXPORT_SYMBOL(ib_cm_notify);
3993
3994 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3995                             struct ib_mad_send_buf *send_buf,
3996                             struct ib_mad_recv_wc *mad_recv_wc)
3997 {
3998         struct cm_port *port = mad_agent->context;
3999         struct cm_work *work;
4000         enum ib_cm_event_type event;
4001         bool alt_path = false;
4002         u16 attr_id;
4003         int paths = 0;
4004         int going_down = 0;
4005
4006         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
4007         case CM_REQ_ATTR_ID:
4008                 alt_path = cm_req_has_alt_path((struct cm_req_msg *)
4009                                                 mad_recv_wc->recv_buf.mad);
4010                 paths = 1 + (alt_path != 0);
4011                 event = IB_CM_REQ_RECEIVED;
4012                 break;
4013         case CM_MRA_ATTR_ID:
4014                 event = IB_CM_MRA_RECEIVED;
4015                 break;
4016         case CM_REJ_ATTR_ID:
4017                 event = IB_CM_REJ_RECEIVED;
4018                 break;
4019         case CM_REP_ATTR_ID:
4020                 event = IB_CM_REP_RECEIVED;
4021                 break;
4022         case CM_RTU_ATTR_ID:
4023                 event = IB_CM_RTU_RECEIVED;
4024                 break;
4025         case CM_DREQ_ATTR_ID:
4026                 event = IB_CM_DREQ_RECEIVED;
4027                 break;
4028         case CM_DREP_ATTR_ID:
4029                 event = IB_CM_DREP_RECEIVED;
4030                 break;
4031         case CM_SIDR_REQ_ATTR_ID:
4032                 event = IB_CM_SIDR_REQ_RECEIVED;
4033                 break;
4034         case CM_SIDR_REP_ATTR_ID:
4035                 event = IB_CM_SIDR_REP_RECEIVED;
4036                 break;
4037         case CM_LAP_ATTR_ID:
4038                 paths = 1;
4039                 event = IB_CM_LAP_RECEIVED;
4040                 break;
4041         case CM_APR_ATTR_ID:
4042                 event = IB_CM_APR_RECEIVED;
4043                 break;
4044         default:
4045                 ib_free_recv_mad(mad_recv_wc);
4046                 return;
4047         }
4048
4049         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
4050         atomic_long_inc(&port->counters[CM_RECV][attr_id - CM_ATTR_ID_OFFSET]);
4051
4052         work = kmalloc(struct_size(work, path, paths), GFP_KERNEL);
4053         if (!work) {
4054                 ib_free_recv_mad(mad_recv_wc);
4055                 return;
4056         }
4057
4058         INIT_DELAYED_WORK(&work->work, cm_work_handler);
4059         work->cm_event.event = event;
4060         work->mad_recv_wc = mad_recv_wc;
4061         work->port = port;
4062
4063         /* Check if the device started its remove_one */
4064         spin_lock_irq(&cm.lock);
4065         if (!port->cm_dev->going_down)
4066                 queue_delayed_work(cm.wq, &work->work, 0);
4067         else
4068                 going_down = 1;
4069         spin_unlock_irq(&cm.lock);
4070
4071         if (going_down) {
4072                 kfree(work);
4073                 ib_free_recv_mad(mad_recv_wc);
4074         }
4075 }
4076
4077 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
4078                                 struct ib_qp_attr *qp_attr,
4079                                 int *qp_attr_mask)
4080 {
4081         unsigned long flags;
4082         int ret;
4083
4084         spin_lock_irqsave(&cm_id_priv->lock, flags);
4085         switch (cm_id_priv->id.state) {
4086         case IB_CM_REQ_SENT:
4087         case IB_CM_MRA_REQ_RCVD:
4088         case IB_CM_REQ_RCVD:
4089         case IB_CM_MRA_REQ_SENT:
4090         case IB_CM_REP_RCVD:
4091         case IB_CM_MRA_REP_SENT:
4092         case IB_CM_REP_SENT:
4093         case IB_CM_MRA_REP_RCVD:
4094         case IB_CM_ESTABLISHED:
4095                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
4096                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
4097                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
4098                 if (cm_id_priv->responder_resources) {
4099                         struct ib_device *ib_dev = cm_id_priv->id.device;
4100                         u64 support_flush = ib_dev->attrs.device_cap_flags &
4101                           (IB_DEVICE_FLUSH_GLOBAL | IB_DEVICE_FLUSH_PERSISTENT);
4102                         u32 flushable = support_flush ?
4103                                         (IB_ACCESS_FLUSH_GLOBAL |
4104                                          IB_ACCESS_FLUSH_PERSISTENT) : 0;
4105
4106                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
4107                                                     IB_ACCESS_REMOTE_ATOMIC |
4108                                                     flushable;
4109                 }
4110                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
4111                 if (cm_id_priv->av.port)
4112                         qp_attr->port_num = cm_id_priv->av.port->port_num;
4113                 ret = 0;
4114                 break;
4115         default:
4116                 trace_icm_qp_init_err(&cm_id_priv->id);
4117                 ret = -EINVAL;
4118                 break;
4119         }
4120         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4121         return ret;
4122 }
4123
4124 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
4125                                struct ib_qp_attr *qp_attr,
4126                                int *qp_attr_mask)
4127 {
4128         unsigned long flags;
4129         int ret;
4130
4131         spin_lock_irqsave(&cm_id_priv->lock, flags);
4132         switch (cm_id_priv->id.state) {
4133         case IB_CM_REQ_RCVD:
4134         case IB_CM_MRA_REQ_SENT:
4135         case IB_CM_REP_RCVD:
4136         case IB_CM_MRA_REP_SENT:
4137         case IB_CM_REP_SENT:
4138         case IB_CM_MRA_REP_RCVD:
4139         case IB_CM_ESTABLISHED:
4140                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
4141                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
4142                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
4143                 if ((qp_attr->ah_attr.type == RDMA_AH_ATTR_TYPE_IB) &&
4144                     cm_id_priv->av.dlid_datapath &&
4145                     (cm_id_priv->av.dlid_datapath != 0xffff))
4146                         qp_attr->ah_attr.ib.dlid = cm_id_priv->av.dlid_datapath;
4147                 qp_attr->path_mtu = cm_id_priv->path_mtu;
4148                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
4149                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
4150                 if (cm_id_priv->qp_type == IB_QPT_RC ||
4151                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
4152                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
4153                                          IB_QP_MIN_RNR_TIMER;
4154                         qp_attr->max_dest_rd_atomic =
4155                                         cm_id_priv->responder_resources;
4156                         qp_attr->min_rnr_timer = 0;
4157                 }
4158                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr) &&
4159                     cm_id_priv->alt_av.port) {
4160                         *qp_attr_mask |= IB_QP_ALT_PATH;
4161                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4162                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4163                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4164                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4165                 }
4166                 ret = 0;
4167                 break;
4168         default:
4169                 trace_icm_qp_rtr_err(&cm_id_priv->id);
4170                 ret = -EINVAL;
4171                 break;
4172         }
4173         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4174         return ret;
4175 }
4176
4177 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
4178                                struct ib_qp_attr *qp_attr,
4179                                int *qp_attr_mask)
4180 {
4181         unsigned long flags;
4182         int ret;
4183
4184         spin_lock_irqsave(&cm_id_priv->lock, flags);
4185         switch (cm_id_priv->id.state) {
4186         /* Allow transition to RTS before sending REP */
4187         case IB_CM_REQ_RCVD:
4188         case IB_CM_MRA_REQ_SENT:
4189
4190         case IB_CM_REP_RCVD:
4191         case IB_CM_MRA_REP_SENT:
4192         case IB_CM_REP_SENT:
4193         case IB_CM_MRA_REP_RCVD:
4194         case IB_CM_ESTABLISHED:
4195                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
4196                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
4197                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4198                         switch (cm_id_priv->qp_type) {
4199                         case IB_QPT_RC:
4200                         case IB_QPT_XRC_INI:
4201                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4202                                                  IB_QP_MAX_QP_RD_ATOMIC;
4203                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
4204                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4205                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4206                                 fallthrough;
4207                         case IB_QPT_XRC_TGT:
4208                                 *qp_attr_mask |= IB_QP_TIMEOUT;
4209                                 qp_attr->timeout = cm_id_priv->av.timeout;
4210                                 break;
4211                         default:
4212                                 break;
4213                         }
4214                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4215                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4216                                 qp_attr->path_mig_state = IB_MIG_REARM;
4217                         }
4218                 } else {
4219                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4220                         if (cm_id_priv->alt_av.port)
4221                                 qp_attr->alt_port_num =
4222                                         cm_id_priv->alt_av.port->port_num;
4223                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4224                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4225                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4226                         qp_attr->path_mig_state = IB_MIG_REARM;
4227                 }
4228                 ret = 0;
4229                 break;
4230         default:
4231                 trace_icm_qp_rts_err(&cm_id_priv->id);
4232                 ret = -EINVAL;
4233                 break;
4234         }
4235         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4236         return ret;
4237 }
4238
4239 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4240                        struct ib_qp_attr *qp_attr,
4241                        int *qp_attr_mask)
4242 {
4243         struct cm_id_private *cm_id_priv;
4244         int ret;
4245
4246         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4247         switch (qp_attr->qp_state) {
4248         case IB_QPS_INIT:
4249                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4250                 break;
4251         case IB_QPS_RTR:
4252                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4253                 break;
4254         case IB_QPS_RTS:
4255                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4256                 break;
4257         default:
4258                 ret = -EINVAL;
4259                 break;
4260         }
4261         return ret;
4262 }
4263 EXPORT_SYMBOL(ib_cm_init_qp_attr);
4264
4265 static ssize_t cm_show_counter(struct ib_device *ibdev, u32 port_num,
4266                                struct ib_port_attribute *attr, char *buf)
4267 {
4268         struct cm_counter_attribute *cm_attr =
4269                 container_of(attr, struct cm_counter_attribute, attr);
4270         struct cm_device *cm_dev = ib_get_client_data(ibdev, &cm_client);
4271
4272         if (WARN_ON(!cm_dev))
4273                 return -EINVAL;
4274
4275         return sysfs_emit(
4276                 buf, "%ld\n",
4277                 atomic_long_read(
4278                         &cm_dev->port[port_num - 1]
4279                                  ->counters[cm_attr->group][cm_attr->index]));
4280 }
4281
4282 #define CM_COUNTER_ATTR(_name, _group, _index)                                 \
4283         {                                                                      \
4284                 .attr = __ATTR(_name, 0444, cm_show_counter, NULL),            \
4285                 .group = _group, .index = _index                               \
4286         }
4287
4288 #define CM_COUNTER_GROUP(_group, _name)                                        \
4289         static struct cm_counter_attribute cm_counter_attr_##_group[] = {      \
4290                 CM_COUNTER_ATTR(req, _group, CM_REQ_COUNTER),                  \
4291                 CM_COUNTER_ATTR(mra, _group, CM_MRA_COUNTER),                  \
4292                 CM_COUNTER_ATTR(rej, _group, CM_REJ_COUNTER),                  \
4293                 CM_COUNTER_ATTR(rep, _group, CM_REP_COUNTER),                  \
4294                 CM_COUNTER_ATTR(rtu, _group, CM_RTU_COUNTER),                  \
4295                 CM_COUNTER_ATTR(dreq, _group, CM_DREQ_COUNTER),                \
4296                 CM_COUNTER_ATTR(drep, _group, CM_DREP_COUNTER),                \
4297                 CM_COUNTER_ATTR(sidr_req, _group, CM_SIDR_REQ_COUNTER),        \
4298                 CM_COUNTER_ATTR(sidr_rep, _group, CM_SIDR_REP_COUNTER),        \
4299                 CM_COUNTER_ATTR(lap, _group, CM_LAP_COUNTER),                  \
4300                 CM_COUNTER_ATTR(apr, _group, CM_APR_COUNTER),                  \
4301         };                                                                     \
4302         static struct attribute *cm_counter_attrs_##_group[] = {               \
4303                 &cm_counter_attr_##_group[0].attr.attr,                        \
4304                 &cm_counter_attr_##_group[1].attr.attr,                        \
4305                 &cm_counter_attr_##_group[2].attr.attr,                        \
4306                 &cm_counter_attr_##_group[3].attr.attr,                        \
4307                 &cm_counter_attr_##_group[4].attr.attr,                        \
4308                 &cm_counter_attr_##_group[5].attr.attr,                        \
4309                 &cm_counter_attr_##_group[6].attr.attr,                        \
4310                 &cm_counter_attr_##_group[7].attr.attr,                        \
4311                 &cm_counter_attr_##_group[8].attr.attr,                        \
4312                 &cm_counter_attr_##_group[9].attr.attr,                        \
4313                 &cm_counter_attr_##_group[10].attr.attr,                       \
4314                 NULL,                                                          \
4315         };                                                                     \
4316         static const struct attribute_group cm_counter_group_##_group = {      \
4317                 .name = _name,                                                 \
4318                 .attrs = cm_counter_attrs_##_group,                            \
4319         };
4320
4321 CM_COUNTER_GROUP(CM_XMIT, "cm_tx_msgs")
4322 CM_COUNTER_GROUP(CM_XMIT_RETRIES, "cm_tx_retries")
4323 CM_COUNTER_GROUP(CM_RECV, "cm_rx_msgs")
4324 CM_COUNTER_GROUP(CM_RECV_DUPLICATES, "cm_rx_duplicates")
4325
4326 static const struct attribute_group *cm_counter_groups[] = {
4327         &cm_counter_group_CM_XMIT,
4328         &cm_counter_group_CM_XMIT_RETRIES,
4329         &cm_counter_group_CM_RECV,
4330         &cm_counter_group_CM_RECV_DUPLICATES,
4331         NULL,
4332 };
4333
4334 static int cm_add_one(struct ib_device *ib_device)
4335 {
4336         struct cm_device *cm_dev;
4337         struct cm_port *port;
4338         struct ib_mad_reg_req reg_req = {
4339                 .mgmt_class = IB_MGMT_CLASS_CM,
4340                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4341         };
4342         struct ib_port_modify port_modify = {
4343                 .set_port_cap_mask = IB_PORT_CM_SUP
4344         };
4345         unsigned long flags;
4346         int ret;
4347         int count = 0;
4348         u32 i;
4349
4350         cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt),
4351                          GFP_KERNEL);
4352         if (!cm_dev)
4353                 return -ENOMEM;
4354
4355         kref_init(&cm_dev->kref);
4356         spin_lock_init(&cm_dev->mad_agent_lock);
4357         cm_dev->ib_device = ib_device;
4358         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4359         cm_dev->going_down = 0;
4360
4361         ib_set_client_data(ib_device, &cm_client, cm_dev);
4362
4363         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4364         rdma_for_each_port (ib_device, i) {
4365                 if (!rdma_cap_ib_cm(ib_device, i))
4366                         continue;
4367
4368                 port = kzalloc(sizeof *port, GFP_KERNEL);
4369                 if (!port) {
4370                         ret = -ENOMEM;
4371                         goto error1;
4372                 }
4373
4374                 cm_dev->port[i-1] = port;
4375                 port->cm_dev = cm_dev;
4376                 port->port_num = i;
4377
4378                 ret = ib_port_register_client_groups(ib_device, i,
4379                                                      cm_counter_groups);
4380                 if (ret)
4381                         goto error1;
4382
4383                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4384                                                         IB_QPT_GSI,
4385                                                         &reg_req,
4386                                                         0,
4387                                                         cm_send_handler,
4388                                                         cm_recv_handler,
4389                                                         port,
4390                                                         0);
4391                 if (IS_ERR(port->mad_agent)) {
4392                         ret = PTR_ERR(port->mad_agent);
4393                         goto error2;
4394                 }
4395
4396                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4397                 if (ret)
4398                         goto error3;
4399
4400                 count++;
4401         }
4402
4403         if (!count) {
4404                 ret = -EOPNOTSUPP;
4405                 goto free;
4406         }
4407
4408         write_lock_irqsave(&cm.device_lock, flags);
4409         list_add_tail(&cm_dev->list, &cm.device_list);
4410         write_unlock_irqrestore(&cm.device_lock, flags);
4411         return 0;
4412
4413 error3:
4414         ib_unregister_mad_agent(port->mad_agent);
4415 error2:
4416         ib_port_unregister_client_groups(ib_device, i, cm_counter_groups);
4417 error1:
4418         port_modify.set_port_cap_mask = 0;
4419         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4420         while (--i) {
4421                 if (!rdma_cap_ib_cm(ib_device, i))
4422                         continue;
4423
4424                 port = cm_dev->port[i-1];
4425                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4426                 ib_unregister_mad_agent(port->mad_agent);
4427                 ib_port_unregister_client_groups(ib_device, i,
4428                                                  cm_counter_groups);
4429         }
4430 free:
4431         cm_device_put(cm_dev);
4432         return ret;
4433 }
4434
4435 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4436 {
4437         struct cm_device *cm_dev = client_data;
4438         struct cm_port *port;
4439         struct ib_port_modify port_modify = {
4440                 .clr_port_cap_mask = IB_PORT_CM_SUP
4441         };
4442         unsigned long flags;
4443         u32 i;
4444
4445         write_lock_irqsave(&cm.device_lock, flags);
4446         list_del(&cm_dev->list);
4447         write_unlock_irqrestore(&cm.device_lock, flags);
4448
4449         spin_lock_irq(&cm.lock);
4450         cm_dev->going_down = 1;
4451         spin_unlock_irq(&cm.lock);
4452
4453         rdma_for_each_port (ib_device, i) {
4454                 struct ib_mad_agent *mad_agent;
4455
4456                 if (!rdma_cap_ib_cm(ib_device, i))
4457                         continue;
4458
4459                 port = cm_dev->port[i-1];
4460                 mad_agent = port->mad_agent;
4461                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4462                 /*
4463                  * We flush the queue here after the going_down set, this
4464                  * verify that no new works will be queued in the recv handler,
4465                  * after that we can call the unregister_mad_agent
4466                  */
4467                 flush_workqueue(cm.wq);
4468                 /*
4469                  * The above ensures no call paths from the work are running,
4470                  * the remaining paths all take the mad_agent_lock.
4471                  */
4472                 spin_lock(&cm_dev->mad_agent_lock);
4473                 port->mad_agent = NULL;
4474                 spin_unlock(&cm_dev->mad_agent_lock);
4475                 ib_unregister_mad_agent(mad_agent);
4476                 ib_port_unregister_client_groups(ib_device, i,
4477                                                  cm_counter_groups);
4478         }
4479
4480         cm_device_put(cm_dev);
4481 }
4482
4483 static int __init ib_cm_init(void)
4484 {
4485         int ret;
4486
4487         INIT_LIST_HEAD(&cm.device_list);
4488         rwlock_init(&cm.device_lock);
4489         spin_lock_init(&cm.lock);
4490         cm.listen_service_table = RB_ROOT;
4491         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4492         cm.remote_id_table = RB_ROOT;
4493         cm.remote_qp_table = RB_ROOT;
4494         cm.remote_sidr_table = RB_ROOT;
4495         xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC);
4496         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4497         INIT_LIST_HEAD(&cm.timewait_list);
4498
4499         cm.wq = alloc_workqueue("ib_cm", 0, 1);
4500         if (!cm.wq) {
4501                 ret = -ENOMEM;
4502                 goto error2;
4503         }
4504
4505         ret = ib_register_client(&cm_client);
4506         if (ret)
4507                 goto error3;
4508
4509         return 0;
4510 error3:
4511         destroy_workqueue(cm.wq);
4512 error2:
4513         return ret;
4514 }
4515
4516 static void __exit ib_cm_cleanup(void)
4517 {
4518         struct cm_timewait_info *timewait_info, *tmp;
4519
4520         spin_lock_irq(&cm.lock);
4521         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4522                 cancel_delayed_work(&timewait_info->work.work);
4523         spin_unlock_irq(&cm.lock);
4524
4525         ib_unregister_client(&cm_client);
4526         destroy_workqueue(cm.wq);
4527
4528         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4529                 list_del(&timewait_info->list);
4530                 kfree(timewait_info);
4531         }
4532
4533         WARN_ON(!xa_empty(&cm.local_id_table));
4534 }
4535
4536 module_init(ib_cm_init);
4537 module_exit(ib_cm_cleanup);