drivers/infiniband/core/cm.c

   1 /*
   2  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   3  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   4  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   5  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * OpenIB.org BSD license below:
  12  *
  13  *     Redistribution and use in source and binary forms, with or
  14  *     without modification, are permitted provided that the following
  15  *     conditions are met:
  16  *
  17  *      - Redistributions of source code must retain the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer.
  20  *
  21  *      - Redistributions in binary form must reproduce the above
  22  *        copyright notice, this list of conditions and the following
  23  *        disclaimer in the documentation and/or other materials
  24  *        provided with the distribution.
  25  *
  26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  33  * SOFTWARE.
  34  */
  35
  36 #include <linux/completion.h>
  37 #include <linux/dma-mapping.h>
  38 #include <linux/device.h>
  39 #include <linux/module.h>
  40 #include <linux/err.h>
  41 #include <linux/idr.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/random.h>
  44 #include <linux/rbtree.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/slab.h>
  47 #include <linux/sysfs.h>
  48 #include <linux/workqueue.h>
  49 #include <linux/kdev_t.h>
  50 #include <linux/etherdevice.h>
  51
  52 #include <rdma/ib_cache.h>
  53 #include <rdma/ib_cm.h>
  54 #include "cm_msgs.h"
  55
  56 MODULE_AUTHOR("Sean Hefty");
  57 MODULE_DESCRIPTION("InfiniBand CM");
  58 MODULE_LICENSE("Dual BSD/GPL");
  59
  60 static const char * const ibcm_rej_reason_strs[] = {
  61         [IB_CM_REJ_NO_QP]                       = "no QP",
  62         [IB_CM_REJ_NO_EEC]                      = "no EEC",
  63         [IB_CM_REJ_NO_RESOURCES]                = "no resources",
  64         [IB_CM_REJ_TIMEOUT]                     = "timeout",
  65         [IB_CM_REJ_UNSUPPORTED]                 = "unsupported",
  66         [IB_CM_REJ_INVALID_COMM_ID]             = "invalid comm ID",
  67         [IB_CM_REJ_INVALID_COMM_INSTANCE]       = "invalid comm instance",
  68         [IB_CM_REJ_INVALID_SERVICE_ID]          = "invalid service ID",
  69         [IB_CM_REJ_INVALID_TRANSPORT_TYPE]      = "invalid transport type",
  70         [IB_CM_REJ_STALE_CONN]                  = "stale conn",
  71         [IB_CM_REJ_RDC_NOT_EXIST]               = "RDC not exist",
  72         [IB_CM_REJ_INVALID_GID]                 = "invalid GID",
  73         [IB_CM_REJ_INVALID_LID]                 = "invalid LID",
  74         [IB_CM_REJ_INVALID_SL]                  = "invalid SL",
  75         [IB_CM_REJ_INVALID_TRAFFIC_CLASS]       = "invalid traffic class",
  76         [IB_CM_REJ_INVALID_HOP_LIMIT]           = "invalid hop limit",
  77         [IB_CM_REJ_INVALID_PACKET_RATE]         = "invalid packet rate",
  78         [IB_CM_REJ_INVALID_ALT_GID]             = "invalid alt GID",
  79         [IB_CM_REJ_INVALID_ALT_LID]             = "invalid alt LID",
  80         [IB_CM_REJ_INVALID_ALT_SL]              = "invalid alt SL",
  81         [IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS]   = "invalid alt traffic class",
  82         [IB_CM_REJ_INVALID_ALT_HOP_LIMIT]       = "invalid alt hop limit",
  83         [IB_CM_REJ_INVALID_ALT_PACKET_RATE]     = "invalid alt packet rate",
  84         [IB_CM_REJ_PORT_CM_REDIRECT]            = "port CM redirect",
  85         [IB_CM_REJ_PORT_REDIRECT]               = "port redirect",
  86         [IB_CM_REJ_INVALID_MTU]                 = "invalid MTU",
  87         [IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES] = "insufficient resp resources",
  88         [IB_CM_REJ_CONSUMER_DEFINED]            = "consumer defined",
  89         [IB_CM_REJ_INVALID_RNR_RETRY]           = "invalid RNR retry",
  90         [IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID]     = "duplicate local comm ID",
  91         [IB_CM_REJ_INVALID_CLASS_VERSION]       = "invalid class version",
  92         [IB_CM_REJ_INVALID_FLOW_LABEL]          = "invalid flow label",
  93         [IB_CM_REJ_INVALID_ALT_FLOW_LABEL]      = "invalid alt flow label",
  94 };
  95
  96 const char *__attribute_const__ ibcm_reject_msg(int reason)
  97 {
  98         size_t index = reason;
  99
 100         if (index < ARRAY_SIZE(ibcm_rej_reason_strs) &&
 101             ibcm_rej_reason_strs[index])
 102                 return ibcm_rej_reason_strs[index];
 103         else
 104                 return "unrecognized reason";
 105 }
 106 EXPORT_SYMBOL(ibcm_reject_msg);
 107
 108 static void cm_add_one(struct ib_device *device);
 109 static void cm_remove_one(struct ib_device *device, void *client_data);
 110
 111 static struct ib_client cm_client = {
 112         .name   = "cm",
 113         .add    = cm_add_one,
 114         .remove = cm_remove_one
 115 };
 116
 117 static struct ib_cm {
 118         spinlock_t lock;
 119         struct list_head device_list;
 120         rwlock_t device_lock;
 121         struct rb_root listen_service_table;
 122         u64 listen_service_id;
 123         /* struct rb_root peer_service_table; todo: fix peer to peer */
 124         struct rb_root remote_qp_table;
 125         struct rb_root remote_id_table;
 126         struct rb_root remote_sidr_table;
 127         struct idr local_id_table;
 128         __be32 random_id_operand;
 129         struct list_head timewait_list;
 130         struct workqueue_struct *wq;
 131         /* Sync on cm change port state */
 132         spinlock_t state_lock;
 133 } cm;
 134
 135 /* Counter indexes ordered by attribute ID */
 136 enum {
 137         CM_REQ_COUNTER,
 138         CM_MRA_COUNTER,
 139         CM_REJ_COUNTER,
 140         CM_REP_COUNTER,
 141         CM_RTU_COUNTER,
 142         CM_DREQ_COUNTER,
 143         CM_DREP_COUNTER,
 144         CM_SIDR_REQ_COUNTER,
 145         CM_SIDR_REP_COUNTER,
 146         CM_LAP_COUNTER,
 147         CM_APR_COUNTER,
 148         CM_ATTR_COUNT,
 149         CM_ATTR_ID_OFFSET = 0x0010,
 150 };
 151
 152 enum {
 153         CM_XMIT,
 154         CM_XMIT_RETRIES,
 155         CM_RECV,
 156         CM_RECV_DUPLICATES,
 157         CM_COUNTER_GROUPS
 158 };
 159
 160 static char const counter_group_names[CM_COUNTER_GROUPS]
 161                                      [sizeof("cm_rx_duplicates")] = {
 162         "cm_tx_msgs", "cm_tx_retries",
 163         "cm_rx_msgs", "cm_rx_duplicates"
 164 };
 165
 166 struct cm_counter_group {
 167         struct kobject obj;
 168         atomic_long_t counter[CM_ATTR_COUNT];
 169 };
 170
 171 struct cm_counter_attribute {
 172         struct attribute attr;
 173         int index;
 174 };
 175
 176 #define CM_COUNTER_ATTR(_name, _index) \
 177 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 178         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 179         .index = _index \
 180 }
 181
 182 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 183 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 184 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 185 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 186 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 187 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 188 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 189 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 190 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 191 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 192 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 193
 194 static struct attribute *cm_counter_default_attrs[] = {
 195         &cm_req_counter_attr.attr,
 196         &cm_mra_counter_attr.attr,
 197         &cm_rej_counter_attr.attr,
 198         &cm_rep_counter_attr.attr,
 199         &cm_rtu_counter_attr.attr,
 200         &cm_dreq_counter_attr.attr,
 201         &cm_drep_counter_attr.attr,
 202         &cm_sidr_req_counter_attr.attr,
 203         &cm_sidr_rep_counter_attr.attr,
 204         &cm_lap_counter_attr.attr,
 205         &cm_apr_counter_attr.attr,
 206         NULL
 207 };
 208
 209 struct cm_port {
 210         struct cm_device *cm_dev;
 211         struct ib_mad_agent *mad_agent;
 212         struct kobject port_obj;
 213         u8 port_num;
 214         struct list_head cm_priv_prim_list;
 215         struct list_head cm_priv_altr_list;
 216         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 217 };
 218
 219 struct cm_device {
 220         struct list_head list;
 221         struct ib_device *ib_device;
 222         struct device *device;
 223         u8 ack_delay;
 224         int going_down;
 225         struct cm_port *port[0];
 226 };
 227
 228 struct cm_av {
 229         struct cm_port *port;
 230         union ib_gid dgid;
 231         struct rdma_ah_attr ah_attr;
 232         u16 pkey_index;
 233         u8 timeout;
 234 };
 235
 236 struct cm_work {
 237         struct delayed_work work;
 238         struct list_head list;
 239         struct cm_port *port;
 240         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 241         __be32 local_id;                        /* Established / timewait */
 242         __be32 remote_id;
 243         struct ib_cm_event cm_event;
 244         struct sa_path_rec path[0];
 245 };
 246
 247 struct cm_timewait_info {
 248         struct cm_work work;                    /* Must be first. */
 249         struct list_head list;
 250         struct rb_node remote_qp_node;
 251         struct rb_node remote_id_node;
 252         __be64 remote_ca_guid;
 253         __be32 remote_qpn;
 254         u8 inserted_remote_qp;
 255         u8 inserted_remote_id;
 256 };
 257
 258 struct cm_id_private {
 259         struct ib_cm_id id;
 260
 261         struct rb_node service_node;
 262         struct rb_node sidr_id_node;
 263         spinlock_t lock;        /* Do not acquire inside cm.lock */
 264         struct completion comp;
 265         atomic_t refcount;
 266         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 267          * Protected by the cm.lock spinlock. */
 268         int listen_sharecount;
 269
 270         struct ib_mad_send_buf *msg;
 271         struct cm_timewait_info *timewait_info;
 272         /* todo: use alternate port on send failure */
 273         struct cm_av av;
 274         struct cm_av alt_av;
 275
 276         void *private_data;
 277         __be64 tid;
 278         __be32 local_qpn;
 279         __be32 remote_qpn;
 280         enum ib_qp_type qp_type;
 281         __be32 sq_psn;
 282         __be32 rq_psn;
 283         int timeout_ms;
 284         enum ib_mtu path_mtu;
 285         __be16 pkey;
 286         u8 private_data_len;
 287         u8 max_cm_retries;
 288         u8 peer_to_peer;
 289         u8 responder_resources;
 290         u8 initiator_depth;
 291         u8 retry_count;
 292         u8 rnr_retry_count;
 293         u8 service_timeout;
 294         u8 target_ack_delay;
 295
 296         struct list_head prim_list;
 297         struct list_head altr_list;
 298         /* Indicates that the send port mad is registered and av is set */
 299         int prim_send_port_not_ready;
 300         int altr_send_port_not_ready;
 301
 302         struct list_head work_list;
 303         atomic_t work_count;
 304 };
 305
 306 static void cm_work_handler(struct work_struct *work);
 307
 308 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 309 {
 310         if (atomic_dec_and_test(&cm_id_priv->refcount))
 311                 complete(&cm_id_priv->comp);
 312 }
 313
 314 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 315                         struct ib_mad_send_buf **msg)
 316 {
 317         struct ib_mad_agent *mad_agent;
 318         struct ib_mad_send_buf *m;
 319         struct ib_ah *ah;
 320         struct cm_av *av;
 321         unsigned long flags, flags2;
 322         int ret = 0;
 323
 324         /* don't let the port to be released till the agent is down */
 325         spin_lock_irqsave(&cm.state_lock, flags2);
 326         spin_lock_irqsave(&cm.lock, flags);
 327         if (!cm_id_priv->prim_send_port_not_ready)
 328                 av = &cm_id_priv->av;
 329         else if (!cm_id_priv->altr_send_port_not_ready &&
 330                  (cm_id_priv->alt_av.port))
 331                 av = &cm_id_priv->alt_av;
 332         else {
 333                 pr_info("%s: not valid CM id\n", __func__);
 334                 ret = -ENODEV;
 335                 spin_unlock_irqrestore(&cm.lock, flags);
 336                 goto out;
 337         }
 338         spin_unlock_irqrestore(&cm.lock, flags);
 339         /* Make sure the port haven't released the mad yet */
 340         mad_agent = cm_id_priv->av.port->mad_agent;
 341         if (!mad_agent) {
 342                 pr_info("%s: not a valid MAD agent\n", __func__);
 343                 ret = -ENODEV;
 344                 goto out;
 345         }
 346         ah = rdma_create_ah(mad_agent->qp->pd, &av->ah_attr);
 347         if (IS_ERR(ah)) {
 348                 ret = PTR_ERR(ah);
 349                 goto out;
 350         }
 351
 352         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 353                                av->pkey_index,
 354                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 355                                GFP_ATOMIC,
 356                                IB_MGMT_BASE_VERSION);
 357         if (IS_ERR(m)) {
 358                 rdma_destroy_ah(ah);
 359                 ret = PTR_ERR(m);
 360                 goto out;
 361         }
 362
 363         /* Timeout set by caller if response is expected. */
 364         m->ah = ah;
 365         m->retries = cm_id_priv->max_cm_retries;
 366
 367         atomic_inc(&cm_id_priv->refcount);
 368         m->context[0] = cm_id_priv;
 369         *msg = m;
 370
 371 out:
 372         spin_unlock_irqrestore(&cm.state_lock, flags2);
 373         return ret;
 374 }
 375
 376 static struct ib_mad_send_buf *cm_alloc_response_msg_no_ah(struct cm_port *port,
 377                                                            struct ib_mad_recv_wc *mad_recv_wc)
 378 {
 379         return ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 380                                   0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 381                                   GFP_ATOMIC,
 382                                   IB_MGMT_BASE_VERSION);
 383 }
 384
 385 static int cm_create_response_msg_ah(struct cm_port *port,
 386                                      struct ib_mad_recv_wc *mad_recv_wc,
 387                                      struct ib_mad_send_buf *msg)
 388 {
 389         struct ib_ah *ah;
 390
 391         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 392                                   mad_recv_wc->recv_buf.grh, port->port_num);
 393         if (IS_ERR(ah))
 394                 return PTR_ERR(ah);
 395
 396         msg->ah = ah;
 397         return 0;
 398 }
 399
 400 static void cm_free_msg(struct ib_mad_send_buf *msg)
 401 {
 402         if (msg->ah)
 403                 rdma_destroy_ah(msg->ah);
 404         if (msg->context[0])
 405                 cm_deref_id(msg->context[0]);
 406         ib_free_send_mad(msg);
 407 }
 408
 409 static int cm_alloc_response_msg(struct cm_port *port,
 410                                  struct ib_mad_recv_wc *mad_recv_wc,
 411                                  struct ib_mad_send_buf **msg)
 412 {
 413         struct ib_mad_send_buf *m;
 414         int ret;
 415
 416         m = cm_alloc_response_msg_no_ah(port, mad_recv_wc);
 417         if (IS_ERR(m))
 418                 return PTR_ERR(m);
 419
 420         ret = cm_create_response_msg_ah(port, mad_recv_wc, m);
 421         if (ret) {
 422                 cm_free_msg(m);
 423                 return ret;
 424         }
 425
 426         *msg = m;
 427         return 0;
 428 }
 429
 430 static void * cm_copy_private_data(const void *private_data,
 431                                    u8 private_data_len)
 432 {
 433         void *data;
 434
 435         if (!private_data || !private_data_len)
 436                 return NULL;
 437
 438         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 439         if (!data)
 440                 return ERR_PTR(-ENOMEM);
 441
 442         return data;
 443 }
 444
 445 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 446                                  void *private_data, u8 private_data_len)
 447 {
 448         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 449                 kfree(cm_id_priv->private_data);
 450
 451         cm_id_priv->private_data = private_data;
 452         cm_id_priv->private_data_len = private_data_len;
 453 }
 454
 455 static void cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 456                                     struct ib_grh *grh, struct cm_av *av)
 457 {
 458         av->port = port;
 459         av->pkey_index = wc->pkey_index;
 460         ib_init_ah_from_wc(port->cm_dev->ib_device, port->port_num, wc,
 461                            grh, &av->ah_attr);
 462 }
 463
 464 static int cm_init_av_by_path(struct sa_path_rec *path, struct cm_av *av,
 465                               struct cm_id_private *cm_id_priv)
 466 {
 467         struct cm_device *cm_dev;
 468         struct cm_port *port = NULL;
 469         unsigned long flags;
 470         int ret;
 471         u8 p;
 472         struct net_device *ndev = ib_get_ndev_from_path(path);
 473
 474         read_lock_irqsave(&cm.device_lock, flags);
 475         list_for_each_entry(cm_dev, &cm.device_list, list) {
 476                 if (!ib_find_cached_gid(cm_dev->ib_device, &path->sgid,
 477                                         sa_conv_pathrec_to_gid_type(path),
 478                                         ndev, &p, NULL)) {
 479                         port = cm_dev->port[p-1];
 480                         break;
 481                 }
 482         }
 483         read_unlock_irqrestore(&cm.device_lock, flags);
 484
 485         if (ndev)
 486                 dev_put(ndev);
 487
 488         if (!port)
 489                 return -EINVAL;
 490
 491         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 492                                   be16_to_cpu(path->pkey), &av->pkey_index);
 493         if (ret)
 494                 return ret;
 495
 496         av->port = port;
 497         ib_init_ah_from_path(cm_dev->ib_device, port->port_num, path,
 498                              &av->ah_attr);
 499         av->timeout = path->packet_life_time + 1;
 500
 501         spin_lock_irqsave(&cm.lock, flags);
 502         if (&cm_id_priv->av == av)
 503                 list_add_tail(&cm_id_priv->prim_list, &port->cm_priv_prim_list);
 504         else if (&cm_id_priv->alt_av == av)
 505                 list_add_tail(&cm_id_priv->altr_list, &port->cm_priv_altr_list);
 506         else
 507                 ret = -EINVAL;
 508
 509         spin_unlock_irqrestore(&cm.lock, flags);
 510
 511         return ret;
 512 }
 513
 514 static int cm_alloc_id(struct cm_id_private *cm_id_priv)
 515 {
 516         unsigned long flags;
 517         int id;
 518
 519         idr_preload(GFP_KERNEL);
 520         spin_lock_irqsave(&cm.lock, flags);
 521
 522         id = idr_alloc_cyclic(&cm.local_id_table, cm_id_priv, 0, 0, GFP_NOWAIT);
 523
 524         spin_unlock_irqrestore(&cm.lock, flags);
 525         idr_preload_end();
 526
 527         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 528         return id < 0 ? id : 0;
 529 }
 530
 531 static void cm_free_id(__be32 local_id)
 532 {
 533         spin_lock_irq(&cm.lock);
 534         idr_remove(&cm.local_id_table,
 535                    (__force int) (local_id ^ cm.random_id_operand));
 536         spin_unlock_irq(&cm.lock);
 537 }
 538
 539 static struct cm_id_private * cm_get_id(__be32 local_id, __be32 remote_id)
 540 {
 541         struct cm_id_private *cm_id_priv;
 542
 543         cm_id_priv = idr_find(&cm.local_id_table,
 544                               (__force int) (local_id ^ cm.random_id_operand));
 545         if (cm_id_priv) {
 546                 if (cm_id_priv->id.remote_id == remote_id)
 547                         atomic_inc(&cm_id_priv->refcount);
 548                 else
 549                         cm_id_priv = NULL;
 550         }
 551
 552         return cm_id_priv;
 553 }
 554
 555 static struct cm_id_private * cm_acquire_id(__be32 local_id, __be32 remote_id)
 556 {
 557         struct cm_id_private *cm_id_priv;
 558
 559         spin_lock_irq(&cm.lock);
 560         cm_id_priv = cm_get_id(local_id, remote_id);
 561         spin_unlock_irq(&cm.lock);
 562
 563         return cm_id_priv;
 564 }
 565
 566 /*
 567  * Trivial helpers to strip endian annotation and compare; the
 568  * endianness doesn't actually matter since we just need a stable
 569  * order for the RB tree.
 570  */
 571 static int be32_lt(__be32 a, __be32 b)
 572 {
 573         return (__force u32) a < (__force u32) b;
 574 }
 575
 576 static int be32_gt(__be32 a, __be32 b)
 577 {
 578         return (__force u32) a > (__force u32) b;
 579 }
 580
 581 static int be64_lt(__be64 a, __be64 b)
 582 {
 583         return (__force u64) a < (__force u64) b;
 584 }
 585
 586 static int be64_gt(__be64 a, __be64 b)
 587 {
 588         return (__force u64) a > (__force u64) b;
 589 }
 590
 591 static struct cm_id_private * cm_insert_listen(struct cm_id_private *cm_id_priv)
 592 {
 593         struct rb_node **link = &cm.listen_service_table.rb_node;
 594         struct rb_node *parent = NULL;
 595         struct cm_id_private *cur_cm_id_priv;
 596         __be64 service_id = cm_id_priv->id.service_id;
 597         __be64 service_mask = cm_id_priv->id.service_mask;
 598
 599         while (*link) {
 600                 parent = *link;
 601                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 602                                           service_node);
 603                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 604                     (service_mask & cur_cm_id_priv->id.service_id) &&
 605                     (cm_id_priv->id.device == cur_cm_id_priv->id.device))
 606                         return cur_cm_id_priv;
 607
 608                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 609                         link = &(*link)->rb_left;
 610                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 611                         link = &(*link)->rb_right;
 612                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 613                         link = &(*link)->rb_left;
 614                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 615                         link = &(*link)->rb_right;
 616                 else
 617                         link = &(*link)->rb_right;
 618         }
 619         rb_link_node(&cm_id_priv->service_node, parent, link);
 620         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 621         return NULL;
 622 }
 623
 624 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 625                                              __be64 service_id)
 626 {
 627         struct rb_node *node = cm.listen_service_table.rb_node;
 628         struct cm_id_private *cm_id_priv;
 629
 630         while (node) {
 631                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 632                 if ((cm_id_priv->id.service_mask & service_id) ==
 633                      cm_id_priv->id.service_id &&
 634                     (cm_id_priv->id.device == device))
 635                         return cm_id_priv;
 636
 637                 if (device < cm_id_priv->id.device)
 638                         node = node->rb_left;
 639                 else if (device > cm_id_priv->id.device)
 640                         node = node->rb_right;
 641                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 642                         node = node->rb_left;
 643                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 644                         node = node->rb_right;
 645                 else
 646                         node = node->rb_right;
 647         }
 648         return NULL;
 649 }
 650
 651 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 652                                                      *timewait_info)
 653 {
 654         struct rb_node **link = &cm.remote_id_table.rb_node;
 655         struct rb_node *parent = NULL;
 656         struct cm_timewait_info *cur_timewait_info;
 657         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 658         __be32 remote_id = timewait_info->work.remote_id;
 659
 660         while (*link) {
 661                 parent = *link;
 662                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 663                                              remote_id_node);
 664                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 665                         link = &(*link)->rb_left;
 666                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 667                         link = &(*link)->rb_right;
 668                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 669                         link = &(*link)->rb_left;
 670                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 671                         link = &(*link)->rb_right;
 672                 else
 673                         return cur_timewait_info;
 674         }
 675         timewait_info->inserted_remote_id = 1;
 676         rb_link_node(&timewait_info->remote_id_node, parent, link);
 677         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 678         return NULL;
 679 }
 680
 681 static struct cm_timewait_info * cm_find_remote_id(__be64 remote_ca_guid,
 682                                                    __be32 remote_id)
 683 {
 684         struct rb_node *node = cm.remote_id_table.rb_node;
 685         struct cm_timewait_info *timewait_info;
 686
 687         while (node) {
 688                 timewait_info = rb_entry(node, struct cm_timewait_info,
 689                                          remote_id_node);
 690                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 691                         node = node->rb_left;
 692                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 693                         node = node->rb_right;
 694                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 695                         node = node->rb_left;
 696                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 697                         node = node->rb_right;
 698                 else
 699                         return timewait_info;
 700         }
 701         return NULL;
 702 }
 703
 704 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 705                                                       *timewait_info)
 706 {
 707         struct rb_node **link = &cm.remote_qp_table.rb_node;
 708         struct rb_node *parent = NULL;
 709         struct cm_timewait_info *cur_timewait_info;
 710         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 711         __be32 remote_qpn = timewait_info->remote_qpn;
 712
 713         while (*link) {
 714                 parent = *link;
 715                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 716                                              remote_qp_node);
 717                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 718                         link = &(*link)->rb_left;
 719                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 720                         link = &(*link)->rb_right;
 721                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 722                         link = &(*link)->rb_left;
 723                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 724                         link = &(*link)->rb_right;
 725                 else
 726                         return cur_timewait_info;
 727         }
 728         timewait_info->inserted_remote_qp = 1;
 729         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 730         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 731         return NULL;
 732 }
 733
 734 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 735                                                     *cm_id_priv)
 736 {
 737         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 738         struct rb_node *parent = NULL;
 739         struct cm_id_private *cur_cm_id_priv;
 740         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 741         __be32 remote_id = cm_id_priv->id.remote_id;
 742
 743         while (*link) {
 744                 parent = *link;
 745                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 746                                           sidr_id_node);
 747                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 748                         link = &(*link)->rb_left;
 749                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 750                         link = &(*link)->rb_right;
 751                 else {
 752                         int cmp;
 753                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 754                                      sizeof *port_gid);
 755                         if (cmp < 0)
 756                                 link = &(*link)->rb_left;
 757                         else if (cmp > 0)
 758                                 link = &(*link)->rb_right;
 759                         else
 760                                 return cur_cm_id_priv;
 761                 }
 762         }
 763         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 764         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 765         return NULL;
 766 }
 767
 768 static void cm_reject_sidr_req(struct cm_id_private *cm_id_priv,
 769                                enum ib_cm_sidr_status status)
 770 {
 771         struct ib_cm_sidr_rep_param param;
 772
 773         memset(&param, 0, sizeof param);
 774         param.status = status;
 775         ib_send_cm_sidr_rep(&cm_id_priv->id, &param);
 776 }
 777
 778 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 779                                  ib_cm_handler cm_handler,
 780                                  void *context)
 781 {
 782         struct cm_id_private *cm_id_priv;
 783         int ret;
 784
 785         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 786         if (!cm_id_priv)
 787                 return ERR_PTR(-ENOMEM);
 788
 789         cm_id_priv->id.state = IB_CM_IDLE;
 790         cm_id_priv->id.device = device;
 791         cm_id_priv->id.cm_handler = cm_handler;
 792         cm_id_priv->id.context = context;
 793         cm_id_priv->id.remote_cm_qpn = 1;
 794         ret = cm_alloc_id(cm_id_priv);
 795         if (ret)
 796                 goto error;
 797
 798         spin_lock_init(&cm_id_priv->lock);
 799         init_completion(&cm_id_priv->comp);
 800         INIT_LIST_HEAD(&cm_id_priv->work_list);
 801         INIT_LIST_HEAD(&cm_id_priv->prim_list);
 802         INIT_LIST_HEAD(&cm_id_priv->altr_list);
 803         atomic_set(&cm_id_priv->work_count, -1);
 804         atomic_set(&cm_id_priv->refcount, 1);
 805         return &cm_id_priv->id;
 806
 807 error:
 808         kfree(cm_id_priv);
 809         return ERR_PTR(-ENOMEM);
 810 }
 811 EXPORT_SYMBOL(ib_create_cm_id);
 812
 813 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 814 {
 815         struct cm_work *work;
 816
 817         if (list_empty(&cm_id_priv->work_list))
 818                 return NULL;
 819
 820         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 821         list_del(&work->list);
 822         return work;
 823 }
 824
 825 static void cm_free_work(struct cm_work *work)
 826 {
 827         if (work->mad_recv_wc)
 828                 ib_free_recv_mad(work->mad_recv_wc);
 829         kfree(work);
 830 }
 831
 832 static inline int cm_convert_to_ms(int iba_time)
 833 {
 834         /* approximate conversion to ms from 4.096us x 2^iba_time */
 835         return 1 << max(iba_time - 8, 0);
 836 }
 837
 838 /*
 839  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 840  * Because of how ack_timeout is stored, adding one doubles the timeout.
 841  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 842  * increment it (round up) only if the other is within 50%.
 843  */
 844 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 845 {
 846         int ack_timeout = packet_life_time + 1;
 847
 848         if (ack_timeout >= ca_ack_delay)
 849                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 850         else
 851                 ack_timeout = ca_ack_delay +
 852                               (ack_timeout >= (ca_ack_delay - 1));
 853
 854         return min(31, ack_timeout);
 855 }
 856
 857 static void cm_cleanup_timewait(struct cm_timewait_info *timewait_info)
 858 {
 859         if (timewait_info->inserted_remote_id) {
 860                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 861                 timewait_info->inserted_remote_id = 0;
 862         }
 863
 864         if (timewait_info->inserted_remote_qp) {
 865                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 866                 timewait_info->inserted_remote_qp = 0;
 867         }
 868 }
 869
 870 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 871 {
 872         struct cm_timewait_info *timewait_info;
 873
 874         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 875         if (!timewait_info)
 876                 return ERR_PTR(-ENOMEM);
 877
 878         timewait_info->work.local_id = local_id;
 879         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 880         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 881         return timewait_info;
 882 }
 883
 884 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
 885 {
 886         int wait_time;
 887         unsigned long flags;
 888         struct cm_device *cm_dev;
 889
 890         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
 891         if (!cm_dev)
 892                 return;
 893
 894         spin_lock_irqsave(&cm.lock, flags);
 895         cm_cleanup_timewait(cm_id_priv->timewait_info);
 896         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
 897         spin_unlock_irqrestore(&cm.lock, flags);
 898
 899         /*
 900          * The cm_id could be destroyed by the user before we exit timewait.
 901          * To protect against this, we search for the cm_id after exiting
 902          * timewait before notifying the user that we've exited timewait.
 903          */
 904         cm_id_priv->id.state = IB_CM_TIMEWAIT;
 905         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
 906
 907         /* Check if the device started its remove_one */
 908         spin_lock_irqsave(&cm.lock, flags);
 909         if (!cm_dev->going_down)
 910                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
 911                                    msecs_to_jiffies(wait_time));
 912         spin_unlock_irqrestore(&cm.lock, flags);
 913
 914         cm_id_priv->timewait_info = NULL;
 915 }
 916
 917 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
 918 {
 919         unsigned long flags;
 920
 921         cm_id_priv->id.state = IB_CM_IDLE;
 922         if (cm_id_priv->timewait_info) {
 923                 spin_lock_irqsave(&cm.lock, flags);
 924                 cm_cleanup_timewait(cm_id_priv->timewait_info);
 925                 spin_unlock_irqrestore(&cm.lock, flags);
 926                 kfree(cm_id_priv->timewait_info);
 927                 cm_id_priv->timewait_info = NULL;
 928         }
 929 }
 930
 931 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
 932 {
 933         struct cm_id_private *cm_id_priv;
 934         struct cm_work *work;
 935
 936         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
 937 retest:
 938         spin_lock_irq(&cm_id_priv->lock);
 939         switch (cm_id->state) {
 940         case IB_CM_LISTEN:
 941                 spin_unlock_irq(&cm_id_priv->lock);
 942
 943                 spin_lock_irq(&cm.lock);
 944                 if (--cm_id_priv->listen_sharecount > 0) {
 945                         /* The id is still shared. */
 946                         cm_deref_id(cm_id_priv);
 947                         spin_unlock_irq(&cm.lock);
 948                         return;
 949                 }
 950                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
 951                 spin_unlock_irq(&cm.lock);
 952                 break;
 953         case IB_CM_SIDR_REQ_SENT:
 954                 cm_id->state = IB_CM_IDLE;
 955                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 956                 spin_unlock_irq(&cm_id_priv->lock);
 957                 break;
 958         case IB_CM_SIDR_REQ_RCVD:
 959                 spin_unlock_irq(&cm_id_priv->lock);
 960                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_REJECT);
 961                 spin_lock_irq(&cm.lock);
 962                 if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
 963                         rb_erase(&cm_id_priv->sidr_id_node,
 964                                  &cm.remote_sidr_table);
 965                 spin_unlock_irq(&cm.lock);
 966                 break;
 967         case IB_CM_REQ_SENT:
 968         case IB_CM_MRA_REQ_RCVD:
 969                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 970                 spin_unlock_irq(&cm_id_priv->lock);
 971                 ib_send_cm_rej(cm_id, IB_CM_REJ_TIMEOUT,
 972                                &cm_id_priv->id.device->node_guid,
 973                                sizeof cm_id_priv->id.device->node_guid,
 974                                NULL, 0);
 975                 break;
 976         case IB_CM_REQ_RCVD:
 977                 if (err == -ENOMEM) {
 978                         /* Do not reject to allow future retries. */
 979                         cm_reset_to_idle(cm_id_priv);
 980                         spin_unlock_irq(&cm_id_priv->lock);
 981                 } else {
 982                         spin_unlock_irq(&cm_id_priv->lock);
 983                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 984                                        NULL, 0, NULL, 0);
 985                 }
 986                 break;
 987         case IB_CM_REP_SENT:
 988         case IB_CM_MRA_REP_RCVD:
 989                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
 990                 /* Fall through */
 991         case IB_CM_MRA_REQ_SENT:
 992         case IB_CM_REP_RCVD:
 993         case IB_CM_MRA_REP_SENT:
 994                 spin_unlock_irq(&cm_id_priv->lock);
 995                 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
 996                                NULL, 0, NULL, 0);
 997                 break;
 998         case IB_CM_ESTABLISHED:
 999                 spin_unlock_irq(&cm_id_priv->lock);
1000                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT)
1001                         break;
1002                 ib_send_cm_dreq(cm_id, NULL, 0);
1003                 goto retest;
1004         case IB_CM_DREQ_SENT:
1005                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1006                 cm_enter_timewait(cm_id_priv);
1007                 spin_unlock_irq(&cm_id_priv->lock);
1008                 break;
1009         case IB_CM_DREQ_RCVD:
1010                 spin_unlock_irq(&cm_id_priv->lock);
1011                 ib_send_cm_drep(cm_id, NULL, 0);
1012                 break;
1013         default:
1014                 spin_unlock_irq(&cm_id_priv->lock);
1015                 break;
1016         }
1017
1018         spin_lock_irq(&cm.lock);
1019         if (!list_empty(&cm_id_priv->altr_list) &&
1020             (!cm_id_priv->altr_send_port_not_ready))
1021                 list_del(&cm_id_priv->altr_list);
1022         if (!list_empty(&cm_id_priv->prim_list) &&
1023             (!cm_id_priv->prim_send_port_not_ready))
1024                 list_del(&cm_id_priv->prim_list);
1025         spin_unlock_irq(&cm.lock);
1026
1027         cm_free_id(cm_id->local_id);
1028         cm_deref_id(cm_id_priv);
1029         wait_for_completion(&cm_id_priv->comp);
1030         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1031                 cm_free_work(work);
1032         kfree(cm_id_priv->private_data);
1033         kfree(cm_id_priv);
1034 }
1035
1036 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1037 {
1038         cm_destroy_id(cm_id, 0);
1039 }
1040 EXPORT_SYMBOL(ib_destroy_cm_id);
1041
1042 /**
1043  * __ib_cm_listen - Initiates listening on the specified service ID for
1044  *   connection and service ID resolution requests.
1045  * @cm_id: Connection identifier associated with the listen request.
1046  * @service_id: Service identifier matched against incoming connection
1047  *   and service ID resolution requests.  The service ID should be specified
1048  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1049  *   assign a service ID to the caller.
1050  * @service_mask: Mask applied to service ID used to listen across a
1051  *   range of service IDs.  If set to 0, the service ID is matched
1052  *   exactly.  This parameter is ignored if %service_id is set to
1053  *   IB_CM_ASSIGN_SERVICE_ID.
1054  */
1055 static int __ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id,
1056                           __be64 service_mask)
1057 {
1058         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
1059         int ret = 0;
1060
1061         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
1062         service_id &= service_mask;
1063         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1064             (service_id != IB_CM_ASSIGN_SERVICE_ID))
1065                 return -EINVAL;
1066
1067         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1068         if (cm_id->state != IB_CM_IDLE)
1069                 return -EINVAL;
1070
1071         cm_id->state = IB_CM_LISTEN;
1072         ++cm_id_priv->listen_sharecount;
1073
1074         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
1075                 cm_id->service_id = cpu_to_be64(cm.listen_service_id++);
1076                 cm_id->service_mask = ~cpu_to_be64(0);
1077         } else {
1078                 cm_id->service_id = service_id;
1079                 cm_id->service_mask = service_mask;
1080         }
1081         cur_cm_id_priv = cm_insert_listen(cm_id_priv);
1082
1083         if (cur_cm_id_priv) {
1084                 cm_id->state = IB_CM_IDLE;
1085                 --cm_id_priv->listen_sharecount;
1086                 ret = -EBUSY;
1087         }
1088         return ret;
1089 }
1090
1091 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask)
1092 {
1093         unsigned long flags;
1094         int ret;
1095
1096         spin_lock_irqsave(&cm.lock, flags);
1097         ret = __ib_cm_listen(cm_id, service_id, service_mask);
1098         spin_unlock_irqrestore(&cm.lock, flags);
1099
1100         return ret;
1101 }
1102 EXPORT_SYMBOL(ib_cm_listen);
1103
1104 /**
1105  * Create a new listening ib_cm_id and listen on the given service ID.
1106  *
1107  * If there's an existing ID listening on that same device and service ID,
1108  * return it.
1109  *
1110  * @device: Device associated with the cm_id.  All related communication will
1111  * be associated with the specified device.
1112  * @cm_handler: Callback invoked to notify the user of CM events.
1113  * @service_id: Service identifier matched against incoming connection
1114  *   and service ID resolution requests.  The service ID should be specified
1115  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1116  *   assign a service ID to the caller.
1117  *
1118  * Callers should call ib_destroy_cm_id when done with the listener ID.
1119  */
1120 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1121                                      ib_cm_handler cm_handler,
1122                                      __be64 service_id)
1123 {
1124         struct cm_id_private *cm_id_priv;
1125         struct ib_cm_id *cm_id;
1126         unsigned long flags;
1127         int err = 0;
1128
1129         /* Create an ID in advance, since the creation may sleep */
1130         cm_id = ib_create_cm_id(device, cm_handler, NULL);
1131         if (IS_ERR(cm_id))
1132                 return cm_id;
1133
1134         spin_lock_irqsave(&cm.lock, flags);
1135
1136         if (service_id == IB_CM_ASSIGN_SERVICE_ID)
1137                 goto new_id;
1138
1139         /* Find an existing ID */
1140         cm_id_priv = cm_find_listen(device, service_id);
1141         if (cm_id_priv) {
1142                 if (cm_id->cm_handler != cm_handler || cm_id->context) {
1143                         /* Sharing an ib_cm_id with different handlers is not
1144                          * supported */
1145                         spin_unlock_irqrestore(&cm.lock, flags);
1146                         return ERR_PTR(-EINVAL);
1147                 }
1148                 atomic_inc(&cm_id_priv->refcount);
1149                 ++cm_id_priv->listen_sharecount;
1150                 spin_unlock_irqrestore(&cm.lock, flags);
1151
1152                 ib_destroy_cm_id(cm_id);
1153                 cm_id = &cm_id_priv->id;
1154                 return cm_id;
1155         }
1156
1157 new_id:
1158         /* Use newly created ID */
1159         err = __ib_cm_listen(cm_id, service_id, 0);
1160
1161         spin_unlock_irqrestore(&cm.lock, flags);
1162
1163         if (err) {
1164                 ib_destroy_cm_id(cm_id);
1165                 return ERR_PTR(err);
1166         }
1167         return cm_id;
1168 }
1169 EXPORT_SYMBOL(ib_cm_insert_listen);
1170
1171 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv,
1172                           enum cm_msg_sequence msg_seq)
1173 {
1174         u64 hi_tid, low_tid;
1175
1176         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1177         low_tid  = (u64) ((__force u32)cm_id_priv->id.local_id |
1178                           (msg_seq << 30));
1179         return cpu_to_be64(hi_tid | low_tid);
1180 }
1181
1182 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1183                               __be16 attr_id, __be64 tid)
1184 {
1185         hdr->base_version  = IB_MGMT_BASE_VERSION;
1186         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1187         hdr->class_version = IB_CM_CLASS_VERSION;
1188         hdr->method        = IB_MGMT_METHOD_SEND;
1189         hdr->attr_id       = attr_id;
1190         hdr->tid           = tid;
1191 }
1192
1193 static void cm_format_req(struct cm_req_msg *req_msg,
1194                           struct cm_id_private *cm_id_priv,
1195                           struct ib_cm_req_param *param)
1196 {
1197         struct sa_path_rec *pri_path = param->primary_path;
1198         struct sa_path_rec *alt_path = param->alternate_path;
1199         bool pri_ext = false;
1200
1201         if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
1202                 pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
1203                                               pri_path->opa.slid);
1204
1205         cm_format_mad_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1206                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_REQ));
1207
1208         req_msg->local_comm_id = cm_id_priv->id.local_id;
1209         req_msg->service_id = param->service_id;
1210         req_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1211         cm_req_set_local_qpn(req_msg, cpu_to_be32(param->qp_num));
1212         cm_req_set_init_depth(req_msg, param->initiator_depth);
1213         cm_req_set_remote_resp_timeout(req_msg,
1214                                        param->remote_cm_response_timeout);
1215         cm_req_set_qp_type(req_msg, param->qp_type);
1216         cm_req_set_flow_ctrl(req_msg, param->flow_control);
1217         cm_req_set_starting_psn(req_msg, cpu_to_be32(param->starting_psn));
1218         cm_req_set_local_resp_timeout(req_msg,
1219                                       param->local_cm_response_timeout);
1220         req_msg->pkey = param->primary_path->pkey;
1221         cm_req_set_path_mtu(req_msg, param->primary_path->mtu);
1222         cm_req_set_max_cm_retries(req_msg, param->max_cm_retries);
1223
1224         if (param->qp_type != IB_QPT_XRC_INI) {
1225                 cm_req_set_resp_res(req_msg, param->responder_resources);
1226                 cm_req_set_retry_count(req_msg, param->retry_count);
1227                 cm_req_set_rnr_retry_count(req_msg, param->rnr_retry_count);
1228                 cm_req_set_srq(req_msg, param->srq);
1229         }
1230
1231         req_msg->primary_local_gid = pri_path->sgid;
1232         req_msg->primary_remote_gid = pri_path->dgid;
1233         if (pri_ext) {
1234                 req_msg->primary_local_gid.global.interface_id
1235                         = OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid));
1236                 req_msg->primary_remote_gid.global.interface_id
1237                         = OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid));
1238         }
1239         if (pri_path->hop_limit <= 1) {
1240                 req_msg->primary_local_lid = pri_ext ? 0 :
1241                         htons(ntohl(sa_path_get_slid(pri_path)));
1242                 req_msg->primary_remote_lid = pri_ext ? 0 :
1243                         htons(ntohl(sa_path_get_dlid(pri_path)));
1244         } else {
1245                 /* Work-around until there's a way to obtain remote LID info */
1246                 req_msg->primary_local_lid = IB_LID_PERMISSIVE;
1247                 req_msg->primary_remote_lid = IB_LID_PERMISSIVE;
1248         }
1249         cm_req_set_primary_flow_label(req_msg, pri_path->flow_label);
1250         cm_req_set_primary_packet_rate(req_msg, pri_path->rate);
1251         req_msg->primary_traffic_class = pri_path->traffic_class;
1252         req_msg->primary_hop_limit = pri_path->hop_limit;
1253         cm_req_set_primary_sl(req_msg, pri_path->sl);
1254         cm_req_set_primary_subnet_local(req_msg, (pri_path->hop_limit <= 1));
1255         cm_req_set_primary_local_ack_timeout(req_msg,
1256                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1257                                pri_path->packet_life_time));
1258
1259         if (alt_path) {
1260                 bool alt_ext = false;
1261
1262                 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA)
1263                         alt_ext = opa_is_extended_lid(alt_path->opa.dlid,
1264                                                       alt_path->opa.slid);
1265
1266                 req_msg->alt_local_gid = alt_path->sgid;
1267                 req_msg->alt_remote_gid = alt_path->dgid;
1268                 if (alt_ext) {
1269                         req_msg->alt_local_gid.global.interface_id
1270                                 = OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid));
1271                         req_msg->alt_remote_gid.global.interface_id
1272                                 = OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid));
1273                 }
1274                 if (alt_path->hop_limit <= 1) {
1275                         req_msg->alt_local_lid = alt_ext ? 0 :
1276                                 htons(ntohl(sa_path_get_slid(alt_path)));
1277                         req_msg->alt_remote_lid = alt_ext ? 0 :
1278                                 htons(ntohl(sa_path_get_dlid(alt_path)));
1279                 } else {
1280                         req_msg->alt_local_lid = IB_LID_PERMISSIVE;
1281                         req_msg->alt_remote_lid = IB_LID_PERMISSIVE;
1282                 }
1283                 cm_req_set_alt_flow_label(req_msg,
1284                                           alt_path->flow_label);
1285                 cm_req_set_alt_packet_rate(req_msg, alt_path->rate);
1286                 req_msg->alt_traffic_class = alt_path->traffic_class;
1287                 req_msg->alt_hop_limit = alt_path->hop_limit;
1288                 cm_req_set_alt_sl(req_msg, alt_path->sl);
1289                 cm_req_set_alt_subnet_local(req_msg, (alt_path->hop_limit <= 1));
1290                 cm_req_set_alt_local_ack_timeout(req_msg,
1291                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1292                                        alt_path->packet_life_time));
1293         }
1294
1295         if (param->private_data && param->private_data_len)
1296                 memcpy(req_msg->private_data, param->private_data,
1297                        param->private_data_len);
1298 }
1299
1300 static int cm_validate_req_param(struct ib_cm_req_param *param)
1301 {
1302         /* peer-to-peer not supported */
1303         if (param->peer_to_peer)
1304                 return -EINVAL;
1305
1306         if (!param->primary_path)
1307                 return -EINVAL;
1308
1309         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1310             param->qp_type != IB_QPT_XRC_INI)
1311                 return -EINVAL;
1312
1313         if (param->private_data &&
1314             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1315                 return -EINVAL;
1316
1317         if (param->alternate_path &&
1318             (param->alternate_path->pkey != param->primary_path->pkey ||
1319              param->alternate_path->mtu != param->primary_path->mtu))
1320                 return -EINVAL;
1321
1322         return 0;
1323 }
1324
1325 int ib_send_cm_req(struct ib_cm_id *cm_id,
1326                    struct ib_cm_req_param *param)
1327 {
1328         struct cm_id_private *cm_id_priv;
1329         struct cm_req_msg *req_msg;
1330         unsigned long flags;
1331         int ret;
1332
1333         ret = cm_validate_req_param(param);
1334         if (ret)
1335                 return ret;
1336
1337         /* Verify that we're not in timewait. */
1338         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1339         spin_lock_irqsave(&cm_id_priv->lock, flags);
1340         if (cm_id->state != IB_CM_IDLE) {
1341                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1342                 ret = -EINVAL;
1343                 goto out;
1344         }
1345         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1346
1347         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1348                                                             id.local_id);
1349         if (IS_ERR(cm_id_priv->timewait_info)) {
1350                 ret = PTR_ERR(cm_id_priv->timewait_info);
1351                 goto out;
1352         }
1353
1354         ret = cm_init_av_by_path(param->primary_path, &cm_id_priv->av,
1355                                  cm_id_priv);
1356         if (ret)
1357                 goto error1;
1358         if (param->alternate_path) {
1359                 ret = cm_init_av_by_path(param->alternate_path,
1360                                          &cm_id_priv->alt_av, cm_id_priv);
1361                 if (ret)
1362                         goto error1;
1363         }
1364         cm_id->service_id = param->service_id;
1365         cm_id->service_mask = ~cpu_to_be64(0);
1366         cm_id_priv->timeout_ms = cm_convert_to_ms(
1367                                     param->primary_path->packet_life_time) * 2 +
1368                                  cm_convert_to_ms(
1369                                     param->remote_cm_response_timeout);
1370         cm_id_priv->max_cm_retries = param->max_cm_retries;
1371         cm_id_priv->initiator_depth = param->initiator_depth;
1372         cm_id_priv->responder_resources = param->responder_resources;
1373         cm_id_priv->retry_count = param->retry_count;
1374         cm_id_priv->path_mtu = param->primary_path->mtu;
1375         cm_id_priv->pkey = param->primary_path->pkey;
1376         cm_id_priv->qp_type = param->qp_type;
1377
1378         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1379         if (ret)
1380                 goto error1;
1381
1382         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1383         cm_format_req(req_msg, cm_id_priv, param);
1384         cm_id_priv->tid = req_msg->hdr.tid;
1385         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1386         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1387
1388         cm_id_priv->local_qpn = cm_req_get_local_qpn(req_msg);
1389         cm_id_priv->rq_psn = cm_req_get_starting_psn(req_msg);
1390
1391         spin_lock_irqsave(&cm_id_priv->lock, flags);
1392         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1393         if (ret) {
1394                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1395                 goto error2;
1396         }
1397         BUG_ON(cm_id->state != IB_CM_IDLE);
1398         cm_id->state = IB_CM_REQ_SENT;
1399         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1400         return 0;
1401
1402 error2: cm_free_msg(cm_id_priv->msg);
1403 error1: kfree(cm_id_priv->timewait_info);
1404 out:    return ret;
1405 }
1406 EXPORT_SYMBOL(ib_send_cm_req);
1407
1408 static int cm_issue_rej(struct cm_port *port,
1409                         struct ib_mad_recv_wc *mad_recv_wc,
1410                         enum ib_cm_rej_reason reason,
1411                         enum cm_msg_response msg_rejected,
1412                         void *ari, u8 ari_length)
1413 {
1414         struct ib_mad_send_buf *msg = NULL;
1415         struct cm_rej_msg *rej_msg, *rcv_msg;
1416         int ret;
1417
1418         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1419         if (ret)
1420                 return ret;
1421
1422         /* We just need common CM header information.  Cast to any message. */
1423         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1424         rej_msg = (struct cm_rej_msg *) msg->mad;
1425
1426         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1427         rej_msg->remote_comm_id = rcv_msg->local_comm_id;
1428         rej_msg->local_comm_id = rcv_msg->remote_comm_id;
1429         cm_rej_set_msg_rejected(rej_msg, msg_rejected);
1430         rej_msg->reason = cpu_to_be16(reason);
1431
1432         if (ari && ari_length) {
1433                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1434                 memcpy(rej_msg->ari, ari, ari_length);
1435         }
1436
1437         ret = ib_post_send_mad(msg, NULL);
1438         if (ret)
1439                 cm_free_msg(msg);
1440
1441         return ret;
1442 }
1443
1444 static inline int cm_is_active_peer(__be64 local_ca_guid, __be64 remote_ca_guid,
1445                                     __be32 local_qpn, __be32 remote_qpn)
1446 {
1447         return (be64_to_cpu(local_ca_guid) > be64_to_cpu(remote_ca_guid) ||
1448                 ((local_ca_guid == remote_ca_guid) &&
1449                  (be32_to_cpu(local_qpn) > be32_to_cpu(remote_qpn))));
1450 }
1451
1452 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1453 {
1454         return ((req_msg->alt_local_lid) ||
1455                 (ib_is_opa_gid(&req_msg->alt_local_gid)));
1456 }
1457
1458 static void cm_path_set_rec_type(struct ib_device *ib_device, u8 port_num,
1459                                  struct sa_path_rec *path, union ib_gid *gid)
1460 {
1461         if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1462                 path->rec_type = SA_PATH_REC_TYPE_OPA;
1463         else
1464                 path->rec_type = SA_PATH_REC_TYPE_IB;
1465 }
1466
1467 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1468                                         struct sa_path_rec *primary_path,
1469                                         struct sa_path_rec *alt_path)
1470 {
1471         u32 lid;
1472
1473         if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1474                 sa_path_set_dlid(primary_path,
1475                                  htonl(ntohs(req_msg->primary_local_lid)));
1476                 sa_path_set_slid(primary_path,
1477                                  htonl(ntohs(req_msg->primary_remote_lid)));
1478         } else {
1479                 lid = opa_get_lid_from_gid(&req_msg->primary_local_gid);
1480                 sa_path_set_dlid(primary_path, cpu_to_be32(lid));
1481
1482                 lid = opa_get_lid_from_gid(&req_msg->primary_remote_gid);
1483                 sa_path_set_slid(primary_path, cpu_to_be32(lid));
1484         }
1485
1486         if (!cm_req_has_alt_path(req_msg))
1487                 return;
1488
1489         if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1490                 sa_path_set_dlid(alt_path,
1491                                  htonl(ntohs(req_msg->alt_local_lid)));
1492                 sa_path_set_slid(alt_path,
1493                                  htonl(ntohs(req_msg->alt_remote_lid)));
1494         } else {
1495                 lid = opa_get_lid_from_gid(&req_msg->alt_local_gid);
1496                 sa_path_set_dlid(alt_path, cpu_to_be32(lid));
1497
1498                 lid = opa_get_lid_from_gid(&req_msg->alt_remote_gid);
1499                 sa_path_set_slid(alt_path, cpu_to_be32(lid));
1500         }
1501 }
1502
1503 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1504                                      struct sa_path_rec *primary_path,
1505                                      struct sa_path_rec *alt_path)
1506 {
1507         primary_path->dgid = req_msg->primary_local_gid;
1508         primary_path->sgid = req_msg->primary_remote_gid;
1509         primary_path->flow_label = cm_req_get_primary_flow_label(req_msg);
1510         primary_path->hop_limit = req_msg->primary_hop_limit;
1511         primary_path->traffic_class = req_msg->primary_traffic_class;
1512         primary_path->reversible = 1;
1513         primary_path->pkey = req_msg->pkey;
1514         primary_path->sl = cm_req_get_primary_sl(req_msg);
1515         primary_path->mtu_selector = IB_SA_EQ;
1516         primary_path->mtu = cm_req_get_path_mtu(req_msg);
1517         primary_path->rate_selector = IB_SA_EQ;
1518         primary_path->rate = cm_req_get_primary_packet_rate(req_msg);
1519         primary_path->packet_life_time_selector = IB_SA_EQ;
1520         primary_path->packet_life_time =
1521                 cm_req_get_primary_local_ack_timeout(req_msg);
1522         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1523         primary_path->service_id = req_msg->service_id;
1524
1525         if (cm_req_has_alt_path(req_msg)) {
1526                 alt_path->dgid = req_msg->alt_local_gid;
1527                 alt_path->sgid = req_msg->alt_remote_gid;
1528                 alt_path->flow_label = cm_req_get_alt_flow_label(req_msg);
1529                 alt_path->hop_limit = req_msg->alt_hop_limit;
1530                 alt_path->traffic_class = req_msg->alt_traffic_class;
1531                 alt_path->reversible = 1;
1532                 alt_path->pkey = req_msg->pkey;
1533                 alt_path->sl = cm_req_get_alt_sl(req_msg);
1534                 alt_path->mtu_selector = IB_SA_EQ;
1535                 alt_path->mtu = cm_req_get_path_mtu(req_msg);
1536                 alt_path->rate_selector = IB_SA_EQ;
1537                 alt_path->rate = cm_req_get_alt_packet_rate(req_msg);
1538                 alt_path->packet_life_time_selector = IB_SA_EQ;
1539                 alt_path->packet_life_time =
1540                         cm_req_get_alt_local_ack_timeout(req_msg);
1541                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1542                 alt_path->service_id = req_msg->service_id;
1543         }
1544         cm_format_path_lid_from_req(req_msg, primary_path, alt_path);
1545 }
1546
1547 static u16 cm_get_bth_pkey(struct cm_work *work)
1548 {
1549         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1550         u8 port_num = work->port->port_num;
1551         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1552         u16 pkey;
1553         int ret;
1554
1555         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1556         if (ret) {
1557                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1558                                      port_num, pkey_index, ret);
1559                 return 0;
1560         }
1561
1562         return pkey;
1563 }
1564
1565 static void cm_format_req_event(struct cm_work *work,
1566                                 struct cm_id_private *cm_id_priv,
1567                                 struct ib_cm_id *listen_id)
1568 {
1569         struct cm_req_msg *req_msg;
1570         struct ib_cm_req_event_param *param;
1571
1572         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1573         param = &work->cm_event.param.req_rcvd;
1574         param->listen_id = listen_id;
1575         param->bth_pkey = cm_get_bth_pkey(work);
1576         param->port = cm_id_priv->av.port->port_num;
1577         param->primary_path = &work->path[0];
1578         if (req_msg->alt_local_lid)
1579                 param->alternate_path = &work->path[1];
1580         else
1581                 param->alternate_path = NULL;
1582         param->remote_ca_guid = req_msg->local_ca_guid;
1583         param->remote_qkey = be32_to_cpu(req_msg->local_qkey);
1584         param->remote_qpn = be32_to_cpu(cm_req_get_local_qpn(req_msg));
1585         param->qp_type = cm_req_get_qp_type(req_msg);
1586         param->starting_psn = be32_to_cpu(cm_req_get_starting_psn(req_msg));
1587         param->responder_resources = cm_req_get_init_depth(req_msg);
1588         param->initiator_depth = cm_req_get_resp_res(req_msg);
1589         param->local_cm_response_timeout =
1590                                         cm_req_get_remote_resp_timeout(req_msg);
1591         param->flow_control = cm_req_get_flow_ctrl(req_msg);
1592         param->remote_cm_response_timeout =
1593                                         cm_req_get_local_resp_timeout(req_msg);
1594         param->retry_count = cm_req_get_retry_count(req_msg);
1595         param->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1596         param->srq = cm_req_get_srq(req_msg);
1597         work->cm_event.private_data = &req_msg->private_data;
1598 }
1599
1600 static void cm_process_work(struct cm_id_private *cm_id_priv,
1601                             struct cm_work *work)
1602 {
1603         int ret;
1604
1605         /* We will typically only have the current event to report. */
1606         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1607         cm_free_work(work);
1608
1609         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1610                 spin_lock_irq(&cm_id_priv->lock);
1611                 work = cm_dequeue_work(cm_id_priv);
1612                 spin_unlock_irq(&cm_id_priv->lock);
1613                 BUG_ON(!work);
1614                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1615                                                 &work->cm_event);
1616                 cm_free_work(work);
1617         }
1618         cm_deref_id(cm_id_priv);
1619         if (ret)
1620                 cm_destroy_id(&cm_id_priv->id, ret);
1621 }
1622
1623 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1624                           struct cm_id_private *cm_id_priv,
1625                           enum cm_msg_response msg_mraed, u8 service_timeout,
1626                           const void *private_data, u8 private_data_len)
1627 {
1628         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1629         cm_mra_set_msg_mraed(mra_msg, msg_mraed);
1630         mra_msg->local_comm_id = cm_id_priv->id.local_id;
1631         mra_msg->remote_comm_id = cm_id_priv->id.remote_id;
1632         cm_mra_set_service_timeout(mra_msg, service_timeout);
1633
1634         if (private_data && private_data_len)
1635                 memcpy(mra_msg->private_data, private_data, private_data_len);
1636 }
1637
1638 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1639                           struct cm_id_private *cm_id_priv,
1640                           enum ib_cm_rej_reason reason,
1641                           void *ari,
1642                           u8 ari_length,
1643                           const void *private_data,
1644                           u8 private_data_len)
1645 {
1646         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1647         rej_msg->remote_comm_id = cm_id_priv->id.remote_id;
1648
1649         switch(cm_id_priv->id.state) {
1650         case IB_CM_REQ_RCVD:
1651                 rej_msg->local_comm_id = 0;
1652                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1653                 break;
1654         case IB_CM_MRA_REQ_SENT:
1655                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1656                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REQ);
1657                 break;
1658         case IB_CM_REP_RCVD:
1659         case IB_CM_MRA_REP_SENT:
1660                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1661                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_REP);
1662                 break;
1663         default:
1664                 rej_msg->local_comm_id = cm_id_priv->id.local_id;
1665                 cm_rej_set_msg_rejected(rej_msg, CM_MSG_RESPONSE_OTHER);
1666                 break;
1667         }
1668
1669         rej_msg->reason = cpu_to_be16(reason);
1670         if (ari && ari_length) {
1671                 cm_rej_set_reject_info_len(rej_msg, ari_length);
1672                 memcpy(rej_msg->ari, ari, ari_length);
1673         }
1674
1675         if (private_data && private_data_len)
1676                 memcpy(rej_msg->private_data, private_data, private_data_len);
1677 }
1678
1679 static void cm_dup_req_handler(struct cm_work *work,
1680                                struct cm_id_private *cm_id_priv)
1681 {
1682         struct ib_mad_send_buf *msg = NULL;
1683         int ret;
1684
1685         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1686                         counter[CM_REQ_COUNTER]);
1687
1688         /* Quick state check to discard duplicate REQs. */
1689         if (cm_id_priv->id.state == IB_CM_REQ_RCVD)
1690                 return;
1691
1692         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1693         if (ret)
1694                 return;
1695
1696         spin_lock_irq(&cm_id_priv->lock);
1697         switch (cm_id_priv->id.state) {
1698         case IB_CM_MRA_REQ_SENT:
1699                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1700                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1701                               cm_id_priv->private_data,
1702                               cm_id_priv->private_data_len);
1703                 break;
1704         case IB_CM_TIMEWAIT:
1705                 cm_format_rej((struct cm_rej_msg *) msg->mad, cm_id_priv,
1706                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0);
1707                 break;
1708         default:
1709                 goto unlock;
1710         }
1711         spin_unlock_irq(&cm_id_priv->lock);
1712
1713         ret = ib_post_send_mad(msg, NULL);
1714         if (ret)
1715                 goto free;
1716         return;
1717
1718 unlock: spin_unlock_irq(&cm_id_priv->lock);
1719 free:   cm_free_msg(msg);
1720 }
1721
1722 static struct cm_id_private * cm_match_req(struct cm_work *work,
1723                                            struct cm_id_private *cm_id_priv)
1724 {
1725         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1726         struct cm_timewait_info *timewait_info;
1727         struct cm_req_msg *req_msg;
1728         struct ib_cm_id *cm_id;
1729
1730         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1731
1732         /* Check for possible duplicate REQ. */
1733         spin_lock_irq(&cm.lock);
1734         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1735         if (timewait_info) {
1736                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1737                                            timewait_info->work.remote_id);
1738                 spin_unlock_irq(&cm.lock);
1739                 if (cur_cm_id_priv) {
1740                         cm_dup_req_handler(work, cur_cm_id_priv);
1741                         cm_deref_id(cur_cm_id_priv);
1742                 }
1743                 return NULL;
1744         }
1745
1746         /* Check for stale connections. */
1747         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1748         if (timewait_info) {
1749                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1750                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
1751                                            timewait_info->work.remote_id);
1752
1753                 spin_unlock_irq(&cm.lock);
1754                 cm_issue_rej(work->port, work->mad_recv_wc,
1755                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
1756                              NULL, 0);
1757                 if (cur_cm_id_priv) {
1758                         cm_id = &cur_cm_id_priv->id;
1759                         ib_send_cm_dreq(cm_id, NULL, 0);
1760                         cm_deref_id(cur_cm_id_priv);
1761                 }
1762                 return NULL;
1763         }
1764
1765         /* Find matching listen request. */
1766         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
1767                                            req_msg->service_id);
1768         if (!listen_cm_id_priv) {
1769                 cm_cleanup_timewait(cm_id_priv->timewait_info);
1770                 spin_unlock_irq(&cm.lock);
1771                 cm_issue_rej(work->port, work->mad_recv_wc,
1772                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
1773                              NULL, 0);
1774                 goto out;
1775         }
1776         atomic_inc(&listen_cm_id_priv->refcount);
1777         atomic_inc(&cm_id_priv->refcount);
1778         cm_id_priv->id.state = IB_CM_REQ_RCVD;
1779         atomic_inc(&cm_id_priv->work_count);
1780         spin_unlock_irq(&cm.lock);
1781 out:
1782         return listen_cm_id_priv;
1783 }
1784
1785 /*
1786  * Work-around for inter-subnet connections.  If the LIDs are permissive,
1787  * we need to override the LID/SL data in the REQ with the LID information
1788  * in the work completion.
1789  */
1790 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
1791 {
1792         if (!cm_req_get_primary_subnet_local(req_msg)) {
1793                 if (req_msg->primary_local_lid == IB_LID_PERMISSIVE) {
1794                         req_msg->primary_local_lid = ib_lid_be16(wc->slid);
1795                         cm_req_set_primary_sl(req_msg, wc->sl);
1796                 }
1797
1798                 if (req_msg->primary_remote_lid == IB_LID_PERMISSIVE)
1799                         req_msg->primary_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1800         }
1801
1802         if (!cm_req_get_alt_subnet_local(req_msg)) {
1803                 if (req_msg->alt_local_lid == IB_LID_PERMISSIVE) {
1804                         req_msg->alt_local_lid = ib_lid_be16(wc->slid);
1805                         cm_req_set_alt_sl(req_msg, wc->sl);
1806                 }
1807
1808                 if (req_msg->alt_remote_lid == IB_LID_PERMISSIVE)
1809                         req_msg->alt_remote_lid = cpu_to_be16(wc->dlid_path_bits);
1810         }
1811 }
1812
1813 static int cm_req_handler(struct cm_work *work)
1814 {
1815         struct ib_cm_id *cm_id;
1816         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
1817         struct cm_req_msg *req_msg;
1818         union ib_gid gid;
1819         struct ib_gid_attr gid_attr;
1820         const struct ib_global_route *grh;
1821         int ret;
1822
1823         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1824
1825         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
1826         if (IS_ERR(cm_id))
1827                 return PTR_ERR(cm_id);
1828
1829         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1830         cm_id_priv->id.remote_id = req_msg->local_comm_id;
1831         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
1832                                 work->mad_recv_wc->recv_buf.grh,
1833                                 &cm_id_priv->av);
1834         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1835                                                             id.local_id);
1836         if (IS_ERR(cm_id_priv->timewait_info)) {
1837                 ret = PTR_ERR(cm_id_priv->timewait_info);
1838                 goto destroy;
1839         }
1840         cm_id_priv->timewait_info->work.remote_id = req_msg->local_comm_id;
1841         cm_id_priv->timewait_info->remote_ca_guid = req_msg->local_ca_guid;
1842         cm_id_priv->timewait_info->remote_qpn = cm_req_get_local_qpn(req_msg);
1843
1844         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
1845         if (!listen_cm_id_priv) {
1846                 ret = -EINVAL;
1847                 kfree(cm_id_priv->timewait_info);
1848                 goto destroy;
1849         }
1850
1851         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
1852         cm_id_priv->id.context = listen_cm_id_priv->id.context;
1853         cm_id_priv->id.service_id = req_msg->service_id;
1854         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1855
1856         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
1857
1858         memset(&work->path[0], 0, sizeof(work->path[0]));
1859         memset(&work->path[1], 0, sizeof(work->path[1]));
1860         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
1861         ret = ib_get_cached_gid(work->port->cm_dev->ib_device,
1862                                 work->port->port_num,
1863                                 grh->sgid_index,
1864                                 &gid, &gid_attr);
1865         if (!ret) {
1866                 if (gid_attr.ndev) {
1867                         work->path[0].rec_type =
1868                                 sa_conv_gid_to_pathrec_type(gid_attr.gid_type);
1869                         sa_path_set_ifindex(&work->path[0],
1870                                             gid_attr.ndev->ifindex);
1871                         sa_path_set_ndev(&work->path[0],
1872                                          dev_net(gid_attr.ndev));
1873                         dev_put(gid_attr.ndev);
1874                 } else {
1875                         cm_path_set_rec_type(work->port->cm_dev->ib_device,
1876                                              work->port->port_num,
1877                                              &work->path[0],
1878                                              &req_msg->primary_local_gid);
1879                 }
1880                 if (cm_req_has_alt_path(req_msg))
1881                         work->path[1].rec_type = work->path[0].rec_type;
1882                 cm_format_paths_from_req(req_msg, &work->path[0],
1883                                          &work->path[1]);
1884                 if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
1885                         sa_path_set_dmac(&work->path[0],
1886                                          cm_id_priv->av.ah_attr.roce.dmac);
1887                 work->path[0].hop_limit = grh->hop_limit;
1888                 ret = cm_init_av_by_path(&work->path[0], &cm_id_priv->av,
1889                                          cm_id_priv);
1890         }
1891         if (ret) {
1892                 int err = ib_get_cached_gid(work->port->cm_dev->ib_device,
1893                                             work->port->port_num, 0,
1894                                             &work->path[0].sgid,
1895                                             &gid_attr);
1896                 if (!err && gid_attr.ndev) {
1897                         work->path[0].rec_type =
1898                                 sa_conv_gid_to_pathrec_type(gid_attr.gid_type);
1899                         sa_path_set_ifindex(&work->path[0],
1900                                             gid_attr.ndev->ifindex);
1901                         sa_path_set_ndev(&work->path[0],
1902                                          dev_net(gid_attr.ndev));
1903                         dev_put(gid_attr.ndev);
1904                 } else {
1905                         cm_path_set_rec_type(work->port->cm_dev->ib_device,
1906                                              work->port->port_num,
1907                                              &work->path[0],
1908                                              &req_msg->primary_local_gid);
1909                 }
1910                 if (cm_req_has_alt_path(req_msg))
1911                         work->path[1].rec_type = work->path[0].rec_type;
1912                 ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_GID,
1913                                &work->path[0].sgid, sizeof work->path[0].sgid,
1914                                NULL, 0);
1915                 goto rejected;
1916         }
1917         if (cm_req_has_alt_path(req_msg)) {
1918                 ret = cm_init_av_by_path(&work->path[1], &cm_id_priv->alt_av,
1919                                          cm_id_priv);
1920                 if (ret) {
1921                         ib_send_cm_rej(cm_id, IB_CM_REJ_INVALID_ALT_GID,
1922                                        &work->path[0].sgid,
1923                                        sizeof work->path[0].sgid, NULL, 0);
1924                         goto rejected;
1925                 }
1926         }
1927         cm_id_priv->tid = req_msg->hdr.tid;
1928         cm_id_priv->timeout_ms = cm_convert_to_ms(
1929                                         cm_req_get_local_resp_timeout(req_msg));
1930         cm_id_priv->max_cm_retries = cm_req_get_max_cm_retries(req_msg);
1931         cm_id_priv->remote_qpn = cm_req_get_local_qpn(req_msg);
1932         cm_id_priv->initiator_depth = cm_req_get_resp_res(req_msg);
1933         cm_id_priv->responder_resources = cm_req_get_init_depth(req_msg);
1934         cm_id_priv->path_mtu = cm_req_get_path_mtu(req_msg);
1935         cm_id_priv->pkey = req_msg->pkey;
1936         cm_id_priv->sq_psn = cm_req_get_starting_psn(req_msg);
1937         cm_id_priv->retry_count = cm_req_get_retry_count(req_msg);
1938         cm_id_priv->rnr_retry_count = cm_req_get_rnr_retry_count(req_msg);
1939         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
1940
1941         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
1942         cm_process_work(cm_id_priv, work);
1943         cm_deref_id(listen_cm_id_priv);
1944         return 0;
1945
1946 rejected:
1947         atomic_dec(&cm_id_priv->refcount);
1948         cm_deref_id(listen_cm_id_priv);
1949 destroy:
1950         ib_destroy_cm_id(cm_id);
1951         return ret;
1952 }
1953
1954 static void cm_format_rep(struct cm_rep_msg *rep_msg,
1955                           struct cm_id_private *cm_id_priv,
1956                           struct ib_cm_rep_param *param)
1957 {
1958         cm_format_mad_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid);
1959         rep_msg->local_comm_id = cm_id_priv->id.local_id;
1960         rep_msg->remote_comm_id = cm_id_priv->id.remote_id;
1961         cm_rep_set_starting_psn(rep_msg, cpu_to_be32(param->starting_psn));
1962         rep_msg->resp_resources = param->responder_resources;
1963         cm_rep_set_target_ack_delay(rep_msg,
1964                                     cm_id_priv->av.port->cm_dev->ack_delay);
1965         cm_rep_set_failover(rep_msg, param->failover_accepted);
1966         cm_rep_set_rnr_retry_count(rep_msg, param->rnr_retry_count);
1967         rep_msg->local_ca_guid = cm_id_priv->id.device->node_guid;
1968
1969         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
1970                 rep_msg->initiator_depth = param->initiator_depth;
1971                 cm_rep_set_flow_ctrl(rep_msg, param->flow_control);
1972                 cm_rep_set_srq(rep_msg, param->srq);
1973                 cm_rep_set_local_qpn(rep_msg, cpu_to_be32(param->qp_num));
1974         } else {
1975                 cm_rep_set_srq(rep_msg, 1);
1976                 cm_rep_set_local_eecn(rep_msg, cpu_to_be32(param->qp_num));
1977         }
1978
1979         if (param->private_data && param->private_data_len)
1980                 memcpy(rep_msg->private_data, param->private_data,
1981                        param->private_data_len);
1982 }
1983
1984 int ib_send_cm_rep(struct ib_cm_id *cm_id,
1985                    struct ib_cm_rep_param *param)
1986 {
1987         struct cm_id_private *cm_id_priv;
1988         struct ib_mad_send_buf *msg;
1989         struct cm_rep_msg *rep_msg;
1990         unsigned long flags;
1991         int ret;
1992
1993         if (param->private_data &&
1994             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
1995                 return -EINVAL;
1996
1997         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1998         spin_lock_irqsave(&cm_id_priv->lock, flags);
1999         if (cm_id->state != IB_CM_REQ_RCVD &&
2000             cm_id->state != IB_CM_MRA_REQ_SENT) {
2001                 ret = -EINVAL;
2002                 goto out;
2003         }
2004
2005         ret = cm_alloc_msg(cm_id_priv, &msg);
2006         if (ret)
2007                 goto out;
2008
2009         rep_msg = (struct cm_rep_msg *) msg->mad;
2010         cm_format_rep(rep_msg, cm_id_priv, param);
2011         msg->timeout_ms = cm_id_priv->timeout_ms;
2012         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
2013
2014         ret = ib_post_send_mad(msg, NULL);
2015         if (ret) {
2016                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2017                 cm_free_msg(msg);
2018                 return ret;
2019         }
2020
2021         cm_id->state = IB_CM_REP_SENT;
2022         cm_id_priv->msg = msg;
2023         cm_id_priv->initiator_depth = param->initiator_depth;
2024         cm_id_priv->responder_resources = param->responder_resources;
2025         cm_id_priv->rq_psn = cm_rep_get_starting_psn(rep_msg);
2026         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2027
2028 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2029         return ret;
2030 }
2031 EXPORT_SYMBOL(ib_send_cm_rep);
2032
2033 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2034                           struct cm_id_private *cm_id_priv,
2035                           const void *private_data,
2036                           u8 private_data_len)
2037 {
2038         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2039         rtu_msg->local_comm_id = cm_id_priv->id.local_id;
2040         rtu_msg->remote_comm_id = cm_id_priv->id.remote_id;
2041
2042         if (private_data && private_data_len)
2043                 memcpy(rtu_msg->private_data, private_data, private_data_len);
2044 }
2045
2046 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2047                    const void *private_data,
2048                    u8 private_data_len)
2049 {
2050         struct cm_id_private *cm_id_priv;
2051         struct ib_mad_send_buf *msg;
2052         unsigned long flags;
2053         void *data;
2054         int ret;
2055
2056         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2057                 return -EINVAL;
2058
2059         data = cm_copy_private_data(private_data, private_data_len);
2060         if (IS_ERR(data))
2061                 return PTR_ERR(data);
2062
2063         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2064         spin_lock_irqsave(&cm_id_priv->lock, flags);
2065         if (cm_id->state != IB_CM_REP_RCVD &&
2066             cm_id->state != IB_CM_MRA_REP_SENT) {
2067                 ret = -EINVAL;
2068                 goto error;
2069         }
2070
2071         ret = cm_alloc_msg(cm_id_priv, &msg);
2072         if (ret)
2073                 goto error;
2074
2075         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2076                       private_data, private_data_len);
2077
2078         ret = ib_post_send_mad(msg, NULL);
2079         if (ret) {
2080                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2081                 cm_free_msg(msg);
2082                 kfree(data);
2083                 return ret;
2084         }
2085
2086         cm_id->state = IB_CM_ESTABLISHED;
2087         cm_set_private_data(cm_id_priv, data, private_data_len);
2088         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2089         return 0;
2090
2091 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2092         kfree(data);
2093         return ret;
2094 }
2095 EXPORT_SYMBOL(ib_send_cm_rtu);
2096
2097 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2098 {
2099         struct cm_rep_msg *rep_msg;
2100         struct ib_cm_rep_event_param *param;
2101
2102         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2103         param = &work->cm_event.param.rep_rcvd;
2104         param->remote_ca_guid = rep_msg->local_ca_guid;
2105         param->remote_qkey = be32_to_cpu(rep_msg->local_qkey);
2106         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2107         param->starting_psn = be32_to_cpu(cm_rep_get_starting_psn(rep_msg));
2108         param->responder_resources = rep_msg->initiator_depth;
2109         param->initiator_depth = rep_msg->resp_resources;
2110         param->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2111         param->failover_accepted = cm_rep_get_failover(rep_msg);
2112         param->flow_control = cm_rep_get_flow_ctrl(rep_msg);
2113         param->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2114         param->srq = cm_rep_get_srq(rep_msg);
2115         work->cm_event.private_data = &rep_msg->private_data;
2116 }
2117
2118 static void cm_dup_rep_handler(struct cm_work *work)
2119 {
2120         struct cm_id_private *cm_id_priv;
2121         struct cm_rep_msg *rep_msg;
2122         struct ib_mad_send_buf *msg = NULL;
2123         int ret;
2124
2125         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2126         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id,
2127                                    rep_msg->local_comm_id);
2128         if (!cm_id_priv)
2129                 return;
2130
2131         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2132                         counter[CM_REP_COUNTER]);
2133         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2134         if (ret)
2135                 goto deref;
2136
2137         spin_lock_irq(&cm_id_priv->lock);
2138         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2139                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2140                               cm_id_priv->private_data,
2141                               cm_id_priv->private_data_len);
2142         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2143                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2144                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2145                               cm_id_priv->private_data,
2146                               cm_id_priv->private_data_len);
2147         else
2148                 goto unlock;
2149         spin_unlock_irq(&cm_id_priv->lock);
2150
2151         ret = ib_post_send_mad(msg, NULL);
2152         if (ret)
2153                 goto free;
2154         goto deref;
2155
2156 unlock: spin_unlock_irq(&cm_id_priv->lock);
2157 free:   cm_free_msg(msg);
2158 deref:  cm_deref_id(cm_id_priv);
2159 }
2160
2161 static int cm_rep_handler(struct cm_work *work)
2162 {
2163         struct cm_id_private *cm_id_priv;
2164         struct cm_rep_msg *rep_msg;
2165         int ret;
2166         struct cm_id_private *cur_cm_id_priv;
2167         struct ib_cm_id *cm_id;
2168         struct cm_timewait_info *timewait_info;
2169
2170         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2171         cm_id_priv = cm_acquire_id(rep_msg->remote_comm_id, 0);
2172         if (!cm_id_priv) {
2173                 cm_dup_rep_handler(work);
2174                 return -EINVAL;
2175         }
2176
2177         cm_format_rep_event(work, cm_id_priv->qp_type);
2178
2179         spin_lock_irq(&cm_id_priv->lock);
2180         switch (cm_id_priv->id.state) {
2181         case IB_CM_REQ_SENT:
2182         case IB_CM_MRA_REQ_RCVD:
2183                 break;
2184         default:
2185                 spin_unlock_irq(&cm_id_priv->lock);
2186                 ret = -EINVAL;
2187                 goto error;
2188         }
2189
2190         cm_id_priv->timewait_info->work.remote_id = rep_msg->local_comm_id;
2191         cm_id_priv->timewait_info->remote_ca_guid = rep_msg->local_ca_guid;
2192         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2193
2194         spin_lock(&cm.lock);
2195         /* Check for duplicate REP. */
2196         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2197                 spin_unlock(&cm.lock);
2198                 spin_unlock_irq(&cm_id_priv->lock);
2199                 ret = -EINVAL;
2200                 goto error;
2201         }
2202         /* Check for a stale connection. */
2203         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2204         if (timewait_info) {
2205                 rb_erase(&cm_id_priv->timewait_info->remote_id_node,
2206                          &cm.remote_id_table);
2207                 cm_id_priv->timewait_info->inserted_remote_id = 0;
2208                 cur_cm_id_priv = cm_get_id(timewait_info->work.local_id,
2209                                            timewait_info->work.remote_id);
2210
2211                 spin_unlock(&cm.lock);
2212                 spin_unlock_irq(&cm_id_priv->lock);
2213                 cm_issue_rej(work->port, work->mad_recv_wc,
2214                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2215                              NULL, 0);
2216                 ret = -EINVAL;
2217                 if (cur_cm_id_priv) {
2218                         cm_id = &cur_cm_id_priv->id;
2219                         ib_send_cm_dreq(cm_id, NULL, 0);
2220                         cm_deref_id(cur_cm_id_priv);
2221                 }
2222
2223                 goto error;
2224         }
2225         spin_unlock(&cm.lock);
2226
2227         cm_id_priv->id.state = IB_CM_REP_RCVD;
2228         cm_id_priv->id.remote_id = rep_msg->local_comm_id;
2229         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2230         cm_id_priv->initiator_depth = rep_msg->resp_resources;
2231         cm_id_priv->responder_resources = rep_msg->initiator_depth;
2232         cm_id_priv->sq_psn = cm_rep_get_starting_psn(rep_msg);
2233         cm_id_priv->rnr_retry_count = cm_rep_get_rnr_retry_count(rep_msg);
2234         cm_id_priv->target_ack_delay = cm_rep_get_target_ack_delay(rep_msg);
2235         cm_id_priv->av.timeout =
2236                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2237                                        cm_id_priv->av.timeout - 1);
2238         cm_id_priv->alt_av.timeout =
2239                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2240                                        cm_id_priv->alt_av.timeout - 1);
2241
2242         /* todo: handle peer_to_peer */
2243
2244         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2245         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2246         if (!ret)
2247                 list_add_tail(&work->list, &cm_id_priv->work_list);
2248         spin_unlock_irq(&cm_id_priv->lock);
2249
2250         if (ret)
2251                 cm_process_work(cm_id_priv, work);
2252         else
2253                 cm_deref_id(cm_id_priv);
2254         return 0;
2255
2256 error:
2257         cm_deref_id(cm_id_priv);
2258         return ret;
2259 }
2260
2261 static int cm_establish_handler(struct cm_work *work)
2262 {
2263         struct cm_id_private *cm_id_priv;
2264         int ret;
2265
2266         /* See comment in cm_establish about lookup. */
2267         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2268         if (!cm_id_priv)
2269                 return -EINVAL;
2270
2271         spin_lock_irq(&cm_id_priv->lock);
2272         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2273                 spin_unlock_irq(&cm_id_priv->lock);
2274                 goto out;
2275         }
2276
2277         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2278         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2279         if (!ret)
2280                 list_add_tail(&work->list, &cm_id_priv->work_list);
2281         spin_unlock_irq(&cm_id_priv->lock);
2282
2283         if (ret)
2284                 cm_process_work(cm_id_priv, work);
2285         else
2286                 cm_deref_id(cm_id_priv);
2287         return 0;
2288 out:
2289         cm_deref_id(cm_id_priv);
2290         return -EINVAL;
2291 }
2292
2293 static int cm_rtu_handler(struct cm_work *work)
2294 {
2295         struct cm_id_private *cm_id_priv;
2296         struct cm_rtu_msg *rtu_msg;
2297         int ret;
2298
2299         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2300         cm_id_priv = cm_acquire_id(rtu_msg->remote_comm_id,
2301                                    rtu_msg->local_comm_id);
2302         if (!cm_id_priv)
2303                 return -EINVAL;
2304
2305         work->cm_event.private_data = &rtu_msg->private_data;
2306
2307         spin_lock_irq(&cm_id_priv->lock);
2308         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2309             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2310                 spin_unlock_irq(&cm_id_priv->lock);
2311                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2312                                 counter[CM_RTU_COUNTER]);
2313                 goto out;
2314         }
2315         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2316
2317         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2318         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2319         if (!ret)
2320                 list_add_tail(&work->list, &cm_id_priv->work_list);
2321         spin_unlock_irq(&cm_id_priv->lock);
2322
2323         if (ret)
2324                 cm_process_work(cm_id_priv, work);
2325         else
2326                 cm_deref_id(cm_id_priv);
2327         return 0;
2328 out:
2329         cm_deref_id(cm_id_priv);
2330         return -EINVAL;
2331 }
2332
2333 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2334                           struct cm_id_private *cm_id_priv,
2335                           const void *private_data,
2336                           u8 private_data_len)
2337 {
2338         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2339                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_DREQ));
2340         dreq_msg->local_comm_id = cm_id_priv->id.local_id;
2341         dreq_msg->remote_comm_id = cm_id_priv->id.remote_id;
2342         cm_dreq_set_remote_qpn(dreq_msg, cm_id_priv->remote_qpn);
2343
2344         if (private_data && private_data_len)
2345                 memcpy(dreq_msg->private_data, private_data, private_data_len);
2346 }
2347
2348 int ib_send_cm_dreq(struct ib_cm_id *cm_id,
2349                     const void *private_data,
2350                     u8 private_data_len)
2351 {
2352         struct cm_id_private *cm_id_priv;
2353         struct ib_mad_send_buf *msg;
2354         unsigned long flags;
2355         int ret;
2356
2357         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2358                 return -EINVAL;
2359
2360         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2361         spin_lock_irqsave(&cm_id_priv->lock, flags);
2362         if (cm_id->state != IB_CM_ESTABLISHED) {
2363                 ret = -EINVAL;
2364                 goto out;
2365         }
2366
2367         if (cm_id->lap_state == IB_CM_LAP_SENT ||
2368             cm_id->lap_state == IB_CM_MRA_LAP_RCVD)
2369                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2370
2371         ret = cm_alloc_msg(cm_id_priv, &msg);
2372         if (ret) {
2373                 cm_enter_timewait(cm_id_priv);
2374                 goto out;
2375         }
2376
2377         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2378                        private_data, private_data_len);
2379         msg->timeout_ms = cm_id_priv->timeout_ms;
2380         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2381
2382         ret = ib_post_send_mad(msg, NULL);
2383         if (ret) {
2384                 cm_enter_timewait(cm_id_priv);
2385                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2386                 cm_free_msg(msg);
2387                 return ret;
2388         }
2389
2390         cm_id->state = IB_CM_DREQ_SENT;
2391         cm_id_priv->msg = msg;
2392 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2393         return ret;
2394 }
2395 EXPORT_SYMBOL(ib_send_cm_dreq);
2396
2397 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2398                           struct cm_id_private *cm_id_priv,
2399                           const void *private_data,
2400                           u8 private_data_len)
2401 {
2402         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2403         drep_msg->local_comm_id = cm_id_priv->id.local_id;
2404         drep_msg->remote_comm_id = cm_id_priv->id.remote_id;
2405
2406         if (private_data && private_data_len)
2407                 memcpy(drep_msg->private_data, private_data, private_data_len);
2408 }
2409
2410 int ib_send_cm_drep(struct ib_cm_id *cm_id,
2411                     const void *private_data,
2412                     u8 private_data_len)
2413 {
2414         struct cm_id_private *cm_id_priv;
2415         struct ib_mad_send_buf *msg;
2416         unsigned long flags;
2417         void *data;
2418         int ret;
2419
2420         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2421                 return -EINVAL;
2422
2423         data = cm_copy_private_data(private_data, private_data_len);
2424         if (IS_ERR(data))
2425                 return PTR_ERR(data);
2426
2427         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2428         spin_lock_irqsave(&cm_id_priv->lock, flags);
2429         if (cm_id->state != IB_CM_DREQ_RCVD) {
2430                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2431                 kfree(data);
2432                 return -EINVAL;
2433         }
2434
2435         cm_set_private_data(cm_id_priv, data, private_data_len);
2436         cm_enter_timewait(cm_id_priv);
2437
2438         ret = cm_alloc_msg(cm_id_priv, &msg);
2439         if (ret)
2440                 goto out;
2441
2442         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2443                        private_data, private_data_len);
2444
2445         ret = ib_post_send_mad(msg, NULL);
2446         if (ret) {
2447                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2448                 cm_free_msg(msg);
2449                 return ret;
2450         }
2451
2452 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2453         return ret;
2454 }
2455 EXPORT_SYMBOL(ib_send_cm_drep);
2456
2457 static int cm_issue_drep(struct cm_port *port,
2458                          struct ib_mad_recv_wc *mad_recv_wc)
2459 {
2460         struct ib_mad_send_buf *msg = NULL;
2461         struct cm_dreq_msg *dreq_msg;
2462         struct cm_drep_msg *drep_msg;
2463         int ret;
2464
2465         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2466         if (ret)
2467                 return ret;
2468
2469         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2470         drep_msg = (struct cm_drep_msg *) msg->mad;
2471
2472         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2473         drep_msg->remote_comm_id = dreq_msg->local_comm_id;
2474         drep_msg->local_comm_id = dreq_msg->remote_comm_id;
2475
2476         ret = ib_post_send_mad(msg, NULL);
2477         if (ret)
2478                 cm_free_msg(msg);
2479
2480         return ret;
2481 }
2482
2483 static int cm_dreq_handler(struct cm_work *work)
2484 {
2485         struct cm_id_private *cm_id_priv;
2486         struct cm_dreq_msg *dreq_msg;
2487         struct ib_mad_send_buf *msg = NULL;
2488         int ret;
2489
2490         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2491         cm_id_priv = cm_acquire_id(dreq_msg->remote_comm_id,
2492                                    dreq_msg->local_comm_id);
2493         if (!cm_id_priv) {
2494                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2495                                 counter[CM_DREQ_COUNTER]);
2496                 cm_issue_drep(work->port, work->mad_recv_wc);
2497                 return -EINVAL;
2498         }
2499
2500         work->cm_event.private_data = &dreq_msg->private_data;
2501
2502         spin_lock_irq(&cm_id_priv->lock);
2503         if (cm_id_priv->local_qpn != cm_dreq_get_remote_qpn(dreq_msg))
2504                 goto unlock;
2505
2506         switch (cm_id_priv->id.state) {
2507         case IB_CM_REP_SENT:
2508         case IB_CM_DREQ_SENT:
2509                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2510                 break;
2511         case IB_CM_ESTABLISHED:
2512                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2513                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2514                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2515                 break;
2516         case IB_CM_MRA_REP_RCVD:
2517                 break;
2518         case IB_CM_TIMEWAIT:
2519                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2520                                 counter[CM_DREQ_COUNTER]);
2521                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
2522                 if (IS_ERR(msg))
2523                         goto unlock;
2524
2525                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2526                                cm_id_priv->private_data,
2527                                cm_id_priv->private_data_len);
2528                 spin_unlock_irq(&cm_id_priv->lock);
2529
2530                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2531                     ib_post_send_mad(msg, NULL))
2532                         cm_free_msg(msg);
2533                 goto deref;
2534         case IB_CM_DREQ_RCVD:
2535                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2536                                 counter[CM_DREQ_COUNTER]);
2537                 goto unlock;
2538         default:
2539                 goto unlock;
2540         }
2541         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2542         cm_id_priv->tid = dreq_msg->hdr.tid;
2543         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2544         if (!ret)
2545                 list_add_tail(&work->list, &cm_id_priv->work_list);
2546         spin_unlock_irq(&cm_id_priv->lock);
2547
2548         if (ret)
2549                 cm_process_work(cm_id_priv, work);
2550         else
2551                 cm_deref_id(cm_id_priv);
2552         return 0;
2553
2554 unlock: spin_unlock_irq(&cm_id_priv->lock);
2555 deref:  cm_deref_id(cm_id_priv);
2556         return -EINVAL;
2557 }
2558
2559 static int cm_drep_handler(struct cm_work *work)
2560 {
2561         struct cm_id_private *cm_id_priv;
2562         struct cm_drep_msg *drep_msg;
2563         int ret;
2564
2565         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2566         cm_id_priv = cm_acquire_id(drep_msg->remote_comm_id,
2567                                    drep_msg->local_comm_id);
2568         if (!cm_id_priv)
2569                 return -EINVAL;
2570
2571         work->cm_event.private_data = &drep_msg->private_data;
2572
2573         spin_lock_irq(&cm_id_priv->lock);
2574         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2575             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2576                 spin_unlock_irq(&cm_id_priv->lock);
2577                 goto out;
2578         }
2579         cm_enter_timewait(cm_id_priv);
2580
2581         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2582         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2583         if (!ret)
2584                 list_add_tail(&work->list, &cm_id_priv->work_list);
2585         spin_unlock_irq(&cm_id_priv->lock);
2586
2587         if (ret)
2588                 cm_process_work(cm_id_priv, work);
2589         else
2590                 cm_deref_id(cm_id_priv);
2591         return 0;
2592 out:
2593         cm_deref_id(cm_id_priv);
2594         return -EINVAL;
2595 }
2596
2597 int ib_send_cm_rej(struct ib_cm_id *cm_id,
2598                    enum ib_cm_rej_reason reason,
2599                    void *ari,
2600                    u8 ari_length,
2601                    const void *private_data,
2602                    u8 private_data_len)
2603 {
2604         struct cm_id_private *cm_id_priv;
2605         struct ib_mad_send_buf *msg;
2606         unsigned long flags;
2607         int ret;
2608
2609         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2610             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2611                 return -EINVAL;
2612
2613         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2614
2615         spin_lock_irqsave(&cm_id_priv->lock, flags);
2616         switch (cm_id->state) {
2617         case IB_CM_REQ_SENT:
2618         case IB_CM_MRA_REQ_RCVD:
2619         case IB_CM_REQ_RCVD:
2620         case IB_CM_MRA_REQ_SENT:
2621         case IB_CM_REP_RCVD:
2622         case IB_CM_MRA_REP_SENT:
2623                 ret = cm_alloc_msg(cm_id_priv, &msg);
2624                 if (!ret)
2625                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2626                                       cm_id_priv, reason, ari, ari_length,
2627                                       private_data, private_data_len);
2628
2629                 cm_reset_to_idle(cm_id_priv);
2630                 break;
2631         case IB_CM_REP_SENT:
2632         case IB_CM_MRA_REP_RCVD:
2633                 ret = cm_alloc_msg(cm_id_priv, &msg);
2634                 if (!ret)
2635                         cm_format_rej((struct cm_rej_msg *) msg->mad,
2636                                       cm_id_priv, reason, ari, ari_length,
2637                                       private_data, private_data_len);
2638
2639                 cm_enter_timewait(cm_id_priv);
2640                 break;
2641         default:
2642                 ret = -EINVAL;
2643                 goto out;
2644         }
2645
2646         if (ret)
2647                 goto out;
2648
2649         ret = ib_post_send_mad(msg, NULL);
2650         if (ret)
2651                 cm_free_msg(msg);
2652
2653 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2654         return ret;
2655 }
2656 EXPORT_SYMBOL(ib_send_cm_rej);
2657
2658 static void cm_format_rej_event(struct cm_work *work)
2659 {
2660         struct cm_rej_msg *rej_msg;
2661         struct ib_cm_rej_event_param *param;
2662
2663         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2664         param = &work->cm_event.param.rej_rcvd;
2665         param->ari = rej_msg->ari;
2666         param->ari_length = cm_rej_get_reject_info_len(rej_msg);
2667         param->reason = __be16_to_cpu(rej_msg->reason);
2668         work->cm_event.private_data = &rej_msg->private_data;
2669 }
2670
2671 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2672 {
2673         struct cm_timewait_info *timewait_info;
2674         struct cm_id_private *cm_id_priv;
2675         __be32 remote_id;
2676
2677         remote_id = rej_msg->local_comm_id;
2678
2679         if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_TIMEOUT) {
2680                 spin_lock_irq(&cm.lock);
2681                 timewait_info = cm_find_remote_id( *((__be64 *) rej_msg->ari),
2682                                                   remote_id);
2683                 if (!timewait_info) {
2684                         spin_unlock_irq(&cm.lock);
2685                         return NULL;
2686                 }
2687                 cm_id_priv = idr_find(&cm.local_id_table, (__force int)
2688                                       (timewait_info->work.local_id ^
2689                                        cm.random_id_operand));
2690                 if (cm_id_priv) {
2691                         if (cm_id_priv->id.remote_id == remote_id)
2692                                 atomic_inc(&cm_id_priv->refcount);
2693                         else
2694                                 cm_id_priv = NULL;
2695                 }
2696                 spin_unlock_irq(&cm.lock);
2697         } else if (cm_rej_get_msg_rejected(rej_msg) == CM_MSG_RESPONSE_REQ)
2698                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, 0);
2699         else
2700                 cm_id_priv = cm_acquire_id(rej_msg->remote_comm_id, remote_id);
2701
2702         return cm_id_priv;
2703 }
2704
2705 static int cm_rej_handler(struct cm_work *work)
2706 {
2707         struct cm_id_private *cm_id_priv;
2708         struct cm_rej_msg *rej_msg;
2709         int ret;
2710
2711         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2712         cm_id_priv = cm_acquire_rejected_id(rej_msg);
2713         if (!cm_id_priv)
2714                 return -EINVAL;
2715
2716         cm_format_rej_event(work);
2717
2718         spin_lock_irq(&cm_id_priv->lock);
2719         switch (cm_id_priv->id.state) {
2720         case IB_CM_REQ_SENT:
2721         case IB_CM_MRA_REQ_RCVD:
2722         case IB_CM_REP_SENT:
2723         case IB_CM_MRA_REP_RCVD:
2724                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2725                 /* fall through */
2726         case IB_CM_REQ_RCVD:
2727         case IB_CM_MRA_REQ_SENT:
2728                 if (__be16_to_cpu(rej_msg->reason) == IB_CM_REJ_STALE_CONN)
2729                         cm_enter_timewait(cm_id_priv);
2730                 else
2731                         cm_reset_to_idle(cm_id_priv);
2732                 break;
2733         case IB_CM_DREQ_SENT:
2734                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2735                 /* fall through */
2736         case IB_CM_REP_RCVD:
2737         case IB_CM_MRA_REP_SENT:
2738                 cm_enter_timewait(cm_id_priv);
2739                 break;
2740         case IB_CM_ESTABLISHED:
2741                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
2742                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
2743                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
2744                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
2745                                               cm_id_priv->msg);
2746                         cm_enter_timewait(cm_id_priv);
2747                         break;
2748                 }
2749                 /* fall through */
2750         default:
2751                 spin_unlock_irq(&cm_id_priv->lock);
2752                 ret = -EINVAL;
2753                 goto out;
2754         }
2755
2756         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2757         if (!ret)
2758                 list_add_tail(&work->list, &cm_id_priv->work_list);
2759         spin_unlock_irq(&cm_id_priv->lock);
2760
2761         if (ret)
2762                 cm_process_work(cm_id_priv, work);
2763         else
2764                 cm_deref_id(cm_id_priv);
2765         return 0;
2766 out:
2767         cm_deref_id(cm_id_priv);
2768         return -EINVAL;
2769 }
2770
2771 int ib_send_cm_mra(struct ib_cm_id *cm_id,
2772                    u8 service_timeout,
2773                    const void *private_data,
2774                    u8 private_data_len)
2775 {
2776         struct cm_id_private *cm_id_priv;
2777         struct ib_mad_send_buf *msg;
2778         enum ib_cm_state cm_state;
2779         enum ib_cm_lap_state lap_state;
2780         enum cm_msg_response msg_response;
2781         void *data;
2782         unsigned long flags;
2783         int ret;
2784
2785         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
2786                 return -EINVAL;
2787
2788         data = cm_copy_private_data(private_data, private_data_len);
2789         if (IS_ERR(data))
2790                 return PTR_ERR(data);
2791
2792         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2793
2794         spin_lock_irqsave(&cm_id_priv->lock, flags);
2795         switch(cm_id_priv->id.state) {
2796         case IB_CM_REQ_RCVD:
2797                 cm_state = IB_CM_MRA_REQ_SENT;
2798                 lap_state = cm_id->lap_state;
2799                 msg_response = CM_MSG_RESPONSE_REQ;
2800                 break;
2801         case IB_CM_REP_RCVD:
2802                 cm_state = IB_CM_MRA_REP_SENT;
2803                 lap_state = cm_id->lap_state;
2804                 msg_response = CM_MSG_RESPONSE_REP;
2805                 break;
2806         case IB_CM_ESTABLISHED:
2807                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
2808                         cm_state = cm_id->state;
2809                         lap_state = IB_CM_MRA_LAP_SENT;
2810                         msg_response = CM_MSG_RESPONSE_OTHER;
2811                         break;
2812                 }
2813         default:
2814                 ret = -EINVAL;
2815                 goto error1;
2816         }
2817
2818         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
2819                 ret = cm_alloc_msg(cm_id_priv, &msg);
2820                 if (ret)
2821                         goto error1;
2822
2823                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2824                               msg_response, service_timeout,
2825                               private_data, private_data_len);
2826                 ret = ib_post_send_mad(msg, NULL);
2827                 if (ret)
2828                         goto error2;
2829         }
2830
2831         cm_id->state = cm_state;
2832         cm_id->lap_state = lap_state;
2833         cm_id_priv->service_timeout = service_timeout;
2834         cm_set_private_data(cm_id_priv, data, private_data_len);
2835         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2836         return 0;
2837
2838 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2839         kfree(data);
2840         return ret;
2841
2842 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2843         kfree(data);
2844         cm_free_msg(msg);
2845         return ret;
2846 }
2847 EXPORT_SYMBOL(ib_send_cm_mra);
2848
2849 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
2850 {
2851         switch (cm_mra_get_msg_mraed(mra_msg)) {
2852         case CM_MSG_RESPONSE_REQ:
2853                 return cm_acquire_id(mra_msg->remote_comm_id, 0);
2854         case CM_MSG_RESPONSE_REP:
2855         case CM_MSG_RESPONSE_OTHER:
2856                 return cm_acquire_id(mra_msg->remote_comm_id,
2857                                      mra_msg->local_comm_id);
2858         default:
2859                 return NULL;
2860         }
2861 }
2862
2863 static int cm_mra_handler(struct cm_work *work)
2864 {
2865         struct cm_id_private *cm_id_priv;
2866         struct cm_mra_msg *mra_msg;
2867         int timeout, ret;
2868
2869         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
2870         cm_id_priv = cm_acquire_mraed_id(mra_msg);
2871         if (!cm_id_priv)
2872                 return -EINVAL;
2873
2874         work->cm_event.private_data = &mra_msg->private_data;
2875         work->cm_event.param.mra_rcvd.service_timeout =
2876                                         cm_mra_get_service_timeout(mra_msg);
2877         timeout = cm_convert_to_ms(cm_mra_get_service_timeout(mra_msg)) +
2878                   cm_convert_to_ms(cm_id_priv->av.timeout);
2879
2880         spin_lock_irq(&cm_id_priv->lock);
2881         switch (cm_id_priv->id.state) {
2882         case IB_CM_REQ_SENT:
2883                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REQ ||
2884                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2885                                   cm_id_priv->msg, timeout))
2886                         goto out;
2887                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
2888                 break;
2889         case IB_CM_REP_SENT:
2890                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_REP ||
2891                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2892                                   cm_id_priv->msg, timeout))
2893                         goto out;
2894                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
2895                 break;
2896         case IB_CM_ESTABLISHED:
2897                 if (cm_mra_get_msg_mraed(mra_msg) != CM_MSG_RESPONSE_OTHER ||
2898                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
2899                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
2900                                   cm_id_priv->msg, timeout)) {
2901                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2902                                 atomic_long_inc(&work->port->
2903                                                 counter_group[CM_RECV_DUPLICATES].
2904                                                 counter[CM_MRA_COUNTER]);
2905                         goto out;
2906                 }
2907                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
2908                 break;
2909         case IB_CM_MRA_REQ_RCVD:
2910         case IB_CM_MRA_REP_RCVD:
2911                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2912                                 counter[CM_MRA_COUNTER]);
2913                 /* fall through */
2914         default:
2915                 goto out;
2916         }
2917
2918         cm_id_priv->msg->context[1] = (void *) (unsigned long)
2919                                       cm_id_priv->id.state;
2920         ret = atomic_inc_and_test(&cm_id_priv->work_count);
2921         if (!ret)
2922                 list_add_tail(&work->list, &cm_id_priv->work_list);
2923         spin_unlock_irq(&cm_id_priv->lock);
2924
2925         if (ret)
2926                 cm_process_work(cm_id_priv, work);
2927         else
2928                 cm_deref_id(cm_id_priv);
2929         return 0;
2930 out:
2931         spin_unlock_irq(&cm_id_priv->lock);
2932         cm_deref_id(cm_id_priv);
2933         return -EINVAL;
2934 }
2935
2936 static void cm_format_lap(struct cm_lap_msg *lap_msg,
2937                           struct cm_id_private *cm_id_priv,
2938                           struct sa_path_rec *alternate_path,
2939                           const void *private_data,
2940                           u8 private_data_len)
2941 {
2942         bool alt_ext = false;
2943
2944         if (alternate_path->rec_type == SA_PATH_REC_TYPE_OPA)
2945                 alt_ext = opa_is_extended_lid(alternate_path->opa.dlid,
2946                                               alternate_path->opa.slid);
2947         cm_format_mad_hdr(&lap_msg->hdr, CM_LAP_ATTR_ID,
2948                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_LAP));
2949         lap_msg->local_comm_id = cm_id_priv->id.local_id;
2950         lap_msg->remote_comm_id = cm_id_priv->id.remote_id;
2951         cm_lap_set_remote_qpn(lap_msg, cm_id_priv->remote_qpn);
2952         /* todo: need remote CM response timeout */
2953         cm_lap_set_remote_resp_timeout(lap_msg, 0x1F);
2954         lap_msg->alt_local_lid =
2955                 htons(ntohl(sa_path_get_slid(alternate_path)));
2956         lap_msg->alt_remote_lid =
2957                 htons(ntohl(sa_path_get_dlid(alternate_path)));
2958         lap_msg->alt_local_gid = alternate_path->sgid;
2959         lap_msg->alt_remote_gid = alternate_path->dgid;
2960         if (alt_ext) {
2961                 lap_msg->alt_local_gid.global.interface_id
2962                         = OPA_MAKE_ID(be32_to_cpu(alternate_path->opa.slid));
2963                 lap_msg->alt_remote_gid.global.interface_id
2964                         = OPA_MAKE_ID(be32_to_cpu(alternate_path->opa.dlid));
2965         }
2966         cm_lap_set_flow_label(lap_msg, alternate_path->flow_label);
2967         cm_lap_set_traffic_class(lap_msg, alternate_path->traffic_class);
2968         lap_msg->alt_hop_limit = alternate_path->hop_limit;
2969         cm_lap_set_packet_rate(lap_msg, alternate_path->rate);
2970         cm_lap_set_sl(lap_msg, alternate_path->sl);
2971         cm_lap_set_subnet_local(lap_msg, 1); /* local only... */
2972         cm_lap_set_local_ack_timeout(lap_msg,
2973                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
2974                                alternate_path->packet_life_time));
2975
2976         if (private_data && private_data_len)
2977                 memcpy(lap_msg->private_data, private_data, private_data_len);
2978 }
2979
2980 int ib_send_cm_lap(struct ib_cm_id *cm_id,
2981                    struct sa_path_rec *alternate_path,
2982                    const void *private_data,
2983                    u8 private_data_len)
2984 {
2985         struct cm_id_private *cm_id_priv;
2986         struct ib_mad_send_buf *msg;
2987         unsigned long flags;
2988         int ret;
2989
2990         if (private_data && private_data_len > IB_CM_LAP_PRIVATE_DATA_SIZE)
2991                 return -EINVAL;
2992
2993         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2994         spin_lock_irqsave(&cm_id_priv->lock, flags);
2995         if (cm_id->state != IB_CM_ESTABLISHED ||
2996             (cm_id->lap_state != IB_CM_LAP_UNINIT &&
2997              cm_id->lap_state != IB_CM_LAP_IDLE)) {
2998                 ret = -EINVAL;
2999                 goto out;
3000         }
3001
3002         ret = cm_init_av_by_path(alternate_path, &cm_id_priv->alt_av,
3003                                  cm_id_priv);
3004         if (ret)
3005                 goto out;
3006         cm_id_priv->alt_av.timeout =
3007                         cm_ack_timeout(cm_id_priv->target_ack_delay,
3008                                        cm_id_priv->alt_av.timeout - 1);
3009
3010         ret = cm_alloc_msg(cm_id_priv, &msg);
3011         if (ret)
3012                 goto out;
3013
3014         cm_format_lap((struct cm_lap_msg *) msg->mad, cm_id_priv,
3015                       alternate_path, private_data, private_data_len);
3016         msg->timeout_ms = cm_id_priv->timeout_ms;
3017         msg->context[1] = (void *) (unsigned long) IB_CM_ESTABLISHED;
3018
3019         ret = ib_post_send_mad(msg, NULL);
3020         if (ret) {
3021                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3022                 cm_free_msg(msg);
3023                 return ret;
3024         }
3025
3026         cm_id->lap_state = IB_CM_LAP_SENT;
3027         cm_id_priv->msg = msg;
3028
3029 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3030         return ret;
3031 }
3032 EXPORT_SYMBOL(ib_send_cm_lap);
3033
3034 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3035                                         struct sa_path_rec *path)
3036 {
3037         u32 lid;
3038
3039         if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3040                 sa_path_set_dlid(path, htonl(ntohs(lap_msg->alt_local_lid)));
3041                 sa_path_set_slid(path, htonl(ntohs(lap_msg->alt_remote_lid)));
3042         } else {
3043                 lid = opa_get_lid_from_gid(&lap_msg->alt_local_gid);
3044                 sa_path_set_dlid(path, cpu_to_be32(lid));
3045
3046                 lid = opa_get_lid_from_gid(&lap_msg->alt_remote_gid);
3047                 sa_path_set_slid(path, cpu_to_be32(lid));
3048         }
3049 }
3050
3051 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3052                                     struct sa_path_rec *path,
3053                                     struct cm_lap_msg *lap_msg)
3054 {
3055         path->dgid = lap_msg->alt_local_gid;
3056         path->sgid = lap_msg->alt_remote_gid;
3057         path->flow_label = cm_lap_get_flow_label(lap_msg);
3058         path->hop_limit = lap_msg->alt_hop_limit;
3059         path->traffic_class = cm_lap_get_traffic_class(lap_msg);
3060         path->reversible = 1;
3061         path->pkey = cm_id_priv->pkey;
3062         path->sl = cm_lap_get_sl(lap_msg);
3063         path->mtu_selector = IB_SA_EQ;
3064         path->mtu = cm_id_priv->path_mtu;
3065         path->rate_selector = IB_SA_EQ;
3066         path->rate = cm_lap_get_packet_rate(lap_msg);
3067         path->packet_life_time_selector = IB_SA_EQ;
3068         path->packet_life_time = cm_lap_get_local_ack_timeout(lap_msg);
3069         path->packet_life_time -= (path->packet_life_time > 0);
3070         cm_format_path_lid_from_lap(lap_msg, path);
3071 }
3072
3073 static int cm_lap_handler(struct cm_work *work)
3074 {
3075         struct cm_id_private *cm_id_priv;
3076         struct cm_lap_msg *lap_msg;
3077         struct ib_cm_lap_event_param *param;
3078         struct ib_mad_send_buf *msg = NULL;
3079         int ret;
3080
3081         /* todo: verify LAP request and send reject APR if invalid. */
3082         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3083         cm_id_priv = cm_acquire_id(lap_msg->remote_comm_id,
3084                                    lap_msg->local_comm_id);
3085         if (!cm_id_priv)
3086                 return -EINVAL;
3087
3088         param = &work->cm_event.param.lap_rcvd;
3089         memset(&work->path[0], 0, sizeof(work->path[1]));
3090         cm_path_set_rec_type(work->port->cm_dev->ib_device,
3091                              work->port->port_num,
3092                              &work->path[0],
3093                              &lap_msg->alt_local_gid);
3094         param->alternate_path = &work->path[0];
3095         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3096         work->cm_event.private_data = &lap_msg->private_data;
3097
3098         spin_lock_irq(&cm_id_priv->lock);
3099         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3100                 goto unlock;
3101
3102         switch (cm_id_priv->id.lap_state) {
3103         case IB_CM_LAP_UNINIT:
3104         case IB_CM_LAP_IDLE:
3105                 break;
3106         case IB_CM_MRA_LAP_SENT:
3107                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3108                                 counter[CM_LAP_COUNTER]);
3109                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
3110                 if (IS_ERR(msg))
3111                         goto unlock;
3112
3113                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3114                               CM_MSG_RESPONSE_OTHER,
3115                               cm_id_priv->service_timeout,
3116                               cm_id_priv->private_data,
3117                               cm_id_priv->private_data_len);
3118                 spin_unlock_irq(&cm_id_priv->lock);
3119
3120                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3121                     ib_post_send_mad(msg, NULL))
3122                         cm_free_msg(msg);
3123                 goto deref;
3124         case IB_CM_LAP_RCVD:
3125                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3126                                 counter[CM_LAP_COUNTER]);
3127                 goto unlock;
3128         default:
3129                 goto unlock;
3130         }
3131
3132         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3133         cm_id_priv->tid = lap_msg->hdr.tid;
3134         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3135                                 work->mad_recv_wc->recv_buf.grh,
3136                                 &cm_id_priv->av);
3137         cm_init_av_by_path(param->alternate_path, &cm_id_priv->alt_av,
3138                            cm_id_priv);
3139         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3140         if (!ret)
3141                 list_add_tail(&work->list, &cm_id_priv->work_list);
3142         spin_unlock_irq(&cm_id_priv->lock);
3143
3144         if (ret)
3145                 cm_process_work(cm_id_priv, work);
3146         else
3147                 cm_deref_id(cm_id_priv);
3148         return 0;
3149
3150 unlock: spin_unlock_irq(&cm_id_priv->lock);
3151 deref:  cm_deref_id(cm_id_priv);
3152         return -EINVAL;
3153 }
3154
3155 static void cm_format_apr(struct cm_apr_msg *apr_msg,
3156                           struct cm_id_private *cm_id_priv,
3157                           enum ib_cm_apr_status status,
3158                           void *info,
3159                           u8 info_length,
3160                           const void *private_data,
3161                           u8 private_data_len)
3162 {
3163         cm_format_mad_hdr(&apr_msg->hdr, CM_APR_ATTR_ID, cm_id_priv->tid);
3164         apr_msg->local_comm_id = cm_id_priv->id.local_id;
3165         apr_msg->remote_comm_id = cm_id_priv->id.remote_id;
3166         apr_msg->ap_status = (u8) status;
3167
3168         if (info && info_length) {
3169                 apr_msg->info_length = info_length;
3170                 memcpy(apr_msg->info, info, info_length);
3171         }
3172
3173         if (private_data && private_data_len)
3174                 memcpy(apr_msg->private_data, private_data, private_data_len);
3175 }
3176
3177 int ib_send_cm_apr(struct ib_cm_id *cm_id,
3178                    enum ib_cm_apr_status status,
3179                    void *info,
3180                    u8 info_length,
3181                    const void *private_data,
3182                    u8 private_data_len)
3183 {
3184         struct cm_id_private *cm_id_priv;
3185         struct ib_mad_send_buf *msg;
3186         unsigned long flags;
3187         int ret;
3188
3189         if ((private_data && private_data_len > IB_CM_APR_PRIVATE_DATA_SIZE) ||
3190             (info && info_length > IB_CM_APR_INFO_LENGTH))
3191                 return -EINVAL;
3192
3193         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3194         spin_lock_irqsave(&cm_id_priv->lock, flags);
3195         if (cm_id->state != IB_CM_ESTABLISHED ||
3196             (cm_id->lap_state != IB_CM_LAP_RCVD &&
3197              cm_id->lap_state != IB_CM_MRA_LAP_SENT)) {
3198                 ret = -EINVAL;
3199                 goto out;
3200         }
3201
3202         ret = cm_alloc_msg(cm_id_priv, &msg);
3203         if (ret)
3204                 goto out;
3205
3206         cm_format_apr((struct cm_apr_msg *) msg->mad, cm_id_priv, status,
3207                       info, info_length, private_data, private_data_len);
3208         ret = ib_post_send_mad(msg, NULL);
3209         if (ret) {
3210                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3211                 cm_free_msg(msg);
3212                 return ret;
3213         }
3214
3215         cm_id->lap_state = IB_CM_LAP_IDLE;
3216 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3217         return ret;
3218 }
3219 EXPORT_SYMBOL(ib_send_cm_apr);
3220
3221 static int cm_apr_handler(struct cm_work *work)
3222 {
3223         struct cm_id_private *cm_id_priv;
3224         struct cm_apr_msg *apr_msg;
3225         int ret;
3226
3227         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3228         cm_id_priv = cm_acquire_id(apr_msg->remote_comm_id,
3229                                    apr_msg->local_comm_id);
3230         if (!cm_id_priv)
3231                 return -EINVAL; /* Unmatched reply. */
3232
3233         work->cm_event.param.apr_rcvd.ap_status = apr_msg->ap_status;
3234         work->cm_event.param.apr_rcvd.apr_info = &apr_msg->info;
3235         work->cm_event.param.apr_rcvd.info_len = apr_msg->info_length;
3236         work->cm_event.private_data = &apr_msg->private_data;
3237
3238         spin_lock_irq(&cm_id_priv->lock);
3239         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3240             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3241              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3242                 spin_unlock_irq(&cm_id_priv->lock);
3243                 goto out;
3244         }
3245         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3246         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3247         cm_id_priv->msg = NULL;
3248
3249         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3250         if (!ret)
3251                 list_add_tail(&work->list, &cm_id_priv->work_list);
3252         spin_unlock_irq(&cm_id_priv->lock);
3253
3254         if (ret)
3255                 cm_process_work(cm_id_priv, work);
3256         else
3257                 cm_deref_id(cm_id_priv);
3258         return 0;
3259 out:
3260         cm_deref_id(cm_id_priv);
3261         return -EINVAL;
3262 }
3263
3264 static int cm_timewait_handler(struct cm_work *work)
3265 {
3266         struct cm_timewait_info *timewait_info;
3267         struct cm_id_private *cm_id_priv;
3268         int ret;
3269
3270         timewait_info = (struct cm_timewait_info *)work;
3271         spin_lock_irq(&cm.lock);
3272         list_del(&timewait_info->list);
3273         spin_unlock_irq(&cm.lock);
3274
3275         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3276                                    timewait_info->work.remote_id);
3277         if (!cm_id_priv)
3278                 return -EINVAL;
3279
3280         spin_lock_irq(&cm_id_priv->lock);
3281         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3282             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3283                 spin_unlock_irq(&cm_id_priv->lock);
3284                 goto out;
3285         }
3286         cm_id_priv->id.state = IB_CM_IDLE;
3287         ret = atomic_inc_and_test(&cm_id_priv->work_count);
3288         if (!ret)
3289                 list_add_tail(&work->list, &cm_id_priv->work_list);
3290         spin_unlock_irq(&cm_id_priv->lock);
3291
3292         if (ret)
3293                 cm_process_work(cm_id_priv, work);
3294         else
3295                 cm_deref_id(cm_id_priv);
3296         return 0;
3297 out:
3298         cm_deref_id(cm_id_priv);
3299         return -EINVAL;
3300 }
3301
3302 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3303                                struct cm_id_private *cm_id_priv,
3304                                struct ib_cm_sidr_req_param *param)
3305 {
3306         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3307                           cm_form_tid(cm_id_priv, CM_MSG_SEQUENCE_SIDR));
3308         sidr_req_msg->request_id = cm_id_priv->id.local_id;
3309         sidr_req_msg->pkey = param->path->pkey;
3310         sidr_req_msg->service_id = param->service_id;
3311
3312         if (param->private_data && param->private_data_len)
3313                 memcpy(sidr_req_msg->private_data, param->private_data,
3314                        param->private_data_len);
3315 }
3316
3317 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3318                         struct ib_cm_sidr_req_param *param)
3319 {
3320         struct cm_id_private *cm_id_priv;
3321         struct ib_mad_send_buf *msg;
3322         unsigned long flags;
3323         int ret;
3324
3325         if (!param->path || (param->private_data &&
3326              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3327                 return -EINVAL;
3328
3329         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3330         ret = cm_init_av_by_path(param->path, &cm_id_priv->av, cm_id_priv);
3331         if (ret)
3332                 goto out;
3333
3334         cm_id->service_id = param->service_id;
3335         cm_id->service_mask = ~cpu_to_be64(0);
3336         cm_id_priv->timeout_ms = param->timeout_ms;
3337         cm_id_priv->max_cm_retries = param->max_cm_retries;
3338         ret = cm_alloc_msg(cm_id_priv, &msg);
3339         if (ret)
3340                 goto out;
3341
3342         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3343                            param);
3344         msg->timeout_ms = cm_id_priv->timeout_ms;
3345         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3346
3347         spin_lock_irqsave(&cm_id_priv->lock, flags);
3348         if (cm_id->state == IB_CM_IDLE)
3349                 ret = ib_post_send_mad(msg, NULL);
3350         else
3351                 ret = -EINVAL;
3352
3353         if (ret) {
3354                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3355                 cm_free_msg(msg);
3356                 goto out;
3357         }
3358         cm_id->state = IB_CM_SIDR_REQ_SENT;
3359         cm_id_priv->msg = msg;
3360         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3361 out:
3362         return ret;
3363 }
3364 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3365
3366 static void cm_format_sidr_req_event(struct cm_work *work,
3367                                      struct ib_cm_id *listen_id)
3368 {
3369         struct cm_sidr_req_msg *sidr_req_msg;
3370         struct ib_cm_sidr_req_event_param *param;
3371
3372         sidr_req_msg = (struct cm_sidr_req_msg *)
3373                                 work->mad_recv_wc->recv_buf.mad;
3374         param = &work->cm_event.param.sidr_req_rcvd;
3375         param->pkey = __be16_to_cpu(sidr_req_msg->pkey);
3376         param->listen_id = listen_id;
3377         param->service_id = sidr_req_msg->service_id;
3378         param->bth_pkey = cm_get_bth_pkey(work);
3379         param->port = work->port->port_num;
3380         work->cm_event.private_data = &sidr_req_msg->private_data;
3381 }
3382
3383 static int cm_sidr_req_handler(struct cm_work *work)
3384 {
3385         struct ib_cm_id *cm_id;
3386         struct cm_id_private *cm_id_priv, *cur_cm_id_priv;
3387         struct cm_sidr_req_msg *sidr_req_msg;
3388         struct ib_wc *wc;
3389
3390         cm_id = ib_create_cm_id(work->port->cm_dev->ib_device, NULL, NULL);
3391         if (IS_ERR(cm_id))
3392                 return PTR_ERR(cm_id);
3393         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3394
3395         /* Record SGID/SLID and request ID for lookup. */
3396         sidr_req_msg = (struct cm_sidr_req_msg *)
3397                                 work->mad_recv_wc->recv_buf.mad;
3398         wc = work->mad_recv_wc->wc;
3399         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3400         cm_id_priv->av.dgid.global.interface_id = 0;
3401         cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3402                                 work->mad_recv_wc->recv_buf.grh,
3403                                 &cm_id_priv->av);
3404         cm_id_priv->id.remote_id = sidr_req_msg->request_id;
3405         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3406         atomic_inc(&cm_id_priv->work_count);
3407
3408         spin_lock_irq(&cm.lock);
3409         cur_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3410         if (cur_cm_id_priv) {
3411                 spin_unlock_irq(&cm.lock);
3412                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3413                                 counter[CM_SIDR_REQ_COUNTER]);
3414                 goto out; /* Duplicate message. */
3415         }
3416         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3417         cur_cm_id_priv = cm_find_listen(cm_id->device,
3418                                         sidr_req_msg->service_id);
3419         if (!cur_cm_id_priv) {
3420                 spin_unlock_irq(&cm.lock);
3421                 cm_reject_sidr_req(cm_id_priv, IB_SIDR_UNSUPPORTED);
3422                 goto out; /* No match. */
3423         }
3424         atomic_inc(&cur_cm_id_priv->refcount);
3425         atomic_inc(&cm_id_priv->refcount);
3426         spin_unlock_irq(&cm.lock);
3427
3428         cm_id_priv->id.cm_handler = cur_cm_id_priv->id.cm_handler;
3429         cm_id_priv->id.context = cur_cm_id_priv->id.context;
3430         cm_id_priv->id.service_id = sidr_req_msg->service_id;
3431         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3432
3433         cm_format_sidr_req_event(work, &cur_cm_id_priv->id);
3434         cm_process_work(cm_id_priv, work);
3435         cm_deref_id(cur_cm_id_priv);
3436         return 0;
3437 out:
3438         ib_destroy_cm_id(&cm_id_priv->id);
3439         return -EINVAL;
3440 }
3441
3442 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3443                                struct cm_id_private *cm_id_priv,
3444                                struct ib_cm_sidr_rep_param *param)
3445 {
3446         cm_format_mad_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3447                           cm_id_priv->tid);
3448         sidr_rep_msg->request_id = cm_id_priv->id.remote_id;
3449         sidr_rep_msg->status = param->status;
3450         cm_sidr_rep_set_qpn(sidr_rep_msg, cpu_to_be32(param->qp_num));
3451         sidr_rep_msg->service_id = cm_id_priv->id.service_id;
3452         sidr_rep_msg->qkey = cpu_to_be32(param->qkey);
3453
3454         if (param->info && param->info_length)
3455                 memcpy(sidr_rep_msg->info, param->info, param->info_length);
3456
3457         if (param->private_data && param->private_data_len)
3458                 memcpy(sidr_rep_msg->private_data, param->private_data,
3459                        param->private_data_len);
3460 }
3461
3462 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3463                         struct ib_cm_sidr_rep_param *param)
3464 {
3465         struct cm_id_private *cm_id_priv;
3466         struct ib_mad_send_buf *msg;
3467         unsigned long flags;
3468         int ret;
3469
3470         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3471             (param->private_data &&
3472              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3473                 return -EINVAL;
3474
3475         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3476         spin_lock_irqsave(&cm_id_priv->lock, flags);
3477         if (cm_id->state != IB_CM_SIDR_REQ_RCVD) {
3478                 ret = -EINVAL;
3479                 goto error;
3480         }
3481
3482         ret = cm_alloc_msg(cm_id_priv, &msg);
3483         if (ret)
3484                 goto error;
3485
3486         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3487                            param);
3488         ret = ib_post_send_mad(msg, NULL);
3489         if (ret) {
3490                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3491                 cm_free_msg(msg);
3492                 return ret;
3493         }
3494         cm_id->state = IB_CM_IDLE;
3495         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3496
3497         spin_lock_irqsave(&cm.lock, flags);
3498         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3499                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3500                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3501         }
3502         spin_unlock_irqrestore(&cm.lock, flags);
3503         return 0;
3504
3505 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3506         return ret;
3507 }
3508 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3509
3510 static void cm_format_sidr_rep_event(struct cm_work *work)
3511 {
3512         struct cm_sidr_rep_msg *sidr_rep_msg;
3513         struct ib_cm_sidr_rep_event_param *param;
3514
3515         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3516                                 work->mad_recv_wc->recv_buf.mad;
3517         param = &work->cm_event.param.sidr_rep_rcvd;
3518         param->status = sidr_rep_msg->status;
3519         param->qkey = be32_to_cpu(sidr_rep_msg->qkey);
3520         param->qpn = be32_to_cpu(cm_sidr_rep_get_qpn(sidr_rep_msg));
3521         param->info = &sidr_rep_msg->info;
3522         param->info_len = sidr_rep_msg->info_length;
3523         work->cm_event.private_data = &sidr_rep_msg->private_data;
3524 }
3525
3526 static int cm_sidr_rep_handler(struct cm_work *work)
3527 {
3528         struct cm_sidr_rep_msg *sidr_rep_msg;
3529         struct cm_id_private *cm_id_priv;
3530
3531         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3532                                 work->mad_recv_wc->recv_buf.mad;
3533         cm_id_priv = cm_acquire_id(sidr_rep_msg->request_id, 0);
3534         if (!cm_id_priv)
3535                 return -EINVAL; /* Unmatched reply. */
3536
3537         spin_lock_irq(&cm_id_priv->lock);
3538         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3539                 spin_unlock_irq(&cm_id_priv->lock);
3540                 goto out;
3541         }
3542         cm_id_priv->id.state = IB_CM_IDLE;
3543         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3544         spin_unlock_irq(&cm_id_priv->lock);
3545
3546         cm_format_sidr_rep_event(work);
3547         cm_process_work(cm_id_priv, work);
3548         return 0;
3549 out:
3550         cm_deref_id(cm_id_priv);
3551         return -EINVAL;
3552 }
3553
3554 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3555                                   enum ib_wc_status wc_status)
3556 {
3557         struct cm_id_private *cm_id_priv;
3558         struct ib_cm_event cm_event;
3559         enum ib_cm_state state;
3560         int ret;
3561
3562         memset(&cm_event, 0, sizeof cm_event);
3563         cm_id_priv = msg->context[0];
3564
3565         /* Discard old sends or ones without a response. */
3566         spin_lock_irq(&cm_id_priv->lock);
3567         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3568         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3569                 goto discard;
3570
3571         pr_debug_ratelimited("CM: failed sending MAD in state %d. (%s)\n",
3572                              state, ib_wc_status_msg(wc_status));
3573         switch (state) {
3574         case IB_CM_REQ_SENT:
3575         case IB_CM_MRA_REQ_RCVD:
3576                 cm_reset_to_idle(cm_id_priv);
3577                 cm_event.event = IB_CM_REQ_ERROR;
3578                 break;
3579         case IB_CM_REP_SENT:
3580         case IB_CM_MRA_REP_RCVD:
3581                 cm_reset_to_idle(cm_id_priv);
3582                 cm_event.event = IB_CM_REP_ERROR;
3583                 break;
3584         case IB_CM_DREQ_SENT:
3585                 cm_enter_timewait(cm_id_priv);
3586                 cm_event.event = IB_CM_DREQ_ERROR;
3587                 break;
3588         case IB_CM_SIDR_REQ_SENT:
3589                 cm_id_priv->id.state = IB_CM_IDLE;
3590                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3591                 break;
3592         default:
3593                 goto discard;
3594         }
3595         spin_unlock_irq(&cm_id_priv->lock);
3596         cm_event.param.send_status = wc_status;
3597
3598         /* No other events can occur on the cm_id at this point. */
3599         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3600         cm_free_msg(msg);
3601         if (ret)
3602                 ib_destroy_cm_id(&cm_id_priv->id);
3603         return;
3604 discard:
3605         spin_unlock_irq(&cm_id_priv->lock);
3606         cm_free_msg(msg);
3607 }
3608
3609 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3610                             struct ib_mad_send_wc *mad_send_wc)
3611 {
3612         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3613         struct cm_port *port;
3614         u16 attr_index;
3615
3616         port = mad_agent->context;
3617         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3618                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3619
3620         /*
3621          * If the send was in response to a received message (context[0] is not
3622          * set to a cm_id), and is not a REJ, then it is a send that was
3623          * manually retried.
3624          */
3625         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3626                 msg->retries = 1;
3627
3628         atomic_long_add(1 + msg->retries,
3629                         &port->counter_group[CM_XMIT].counter[attr_index]);
3630         if (msg->retries)
3631                 atomic_long_add(msg->retries,
3632                                 &port->counter_group[CM_XMIT_RETRIES].
3633                                 counter[attr_index]);
3634
3635         switch (mad_send_wc->status) {
3636         case IB_WC_SUCCESS:
3637         case IB_WC_WR_FLUSH_ERR:
3638                 cm_free_msg(msg);
3639                 break;
3640         default:
3641                 if (msg->context[0] && msg->context[1])
3642                         cm_process_send_error(msg, mad_send_wc->status);
3643                 else
3644                         cm_free_msg(msg);
3645                 break;
3646         }
3647 }
3648
3649 static void cm_work_handler(struct work_struct *_work)
3650 {
3651         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3652         int ret;
3653
3654         switch (work->cm_event.event) {
3655         case IB_CM_REQ_RECEIVED:
3656                 ret = cm_req_handler(work);
3657                 break;
3658         case IB_CM_MRA_RECEIVED:
3659                 ret = cm_mra_handler(work);
3660                 break;
3661         case IB_CM_REJ_RECEIVED:
3662                 ret = cm_rej_handler(work);
3663                 break;
3664         case IB_CM_REP_RECEIVED:
3665                 ret = cm_rep_handler(work);
3666                 break;
3667         case IB_CM_RTU_RECEIVED:
3668                 ret = cm_rtu_handler(work);
3669                 break;
3670         case IB_CM_USER_ESTABLISHED:
3671                 ret = cm_establish_handler(work);
3672                 break;
3673         case IB_CM_DREQ_RECEIVED:
3674                 ret = cm_dreq_handler(work);
3675                 break;
3676         case IB_CM_DREP_RECEIVED:
3677                 ret = cm_drep_handler(work);
3678                 break;
3679         case IB_CM_SIDR_REQ_RECEIVED:
3680                 ret = cm_sidr_req_handler(work);
3681                 break;
3682         case IB_CM_SIDR_REP_RECEIVED:
3683                 ret = cm_sidr_rep_handler(work);
3684                 break;
3685         case IB_CM_LAP_RECEIVED:
3686                 ret = cm_lap_handler(work);
3687                 break;
3688         case IB_CM_APR_RECEIVED:
3689                 ret = cm_apr_handler(work);
3690                 break;
3691         case IB_CM_TIMEWAIT_EXIT:
3692                 ret = cm_timewait_handler(work);
3693                 break;
3694         default:
3695                 ret = -EINVAL;
3696                 break;
3697         }
3698         if (ret)
3699                 cm_free_work(work);
3700 }
3701
3702 static int cm_establish(struct ib_cm_id *cm_id)
3703 {
3704         struct cm_id_private *cm_id_priv;
3705         struct cm_work *work;
3706         unsigned long flags;
3707         int ret = 0;
3708         struct cm_device *cm_dev;
3709
3710         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3711         if (!cm_dev)
3712                 return -ENODEV;
3713
3714         work = kmalloc(sizeof *work, GFP_ATOMIC);
3715         if (!work)
3716                 return -ENOMEM;
3717
3718         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3719         spin_lock_irqsave(&cm_id_priv->lock, flags);
3720         switch (cm_id->state)
3721         {
3722         case IB_CM_REP_SENT:
3723         case IB_CM_MRA_REP_RCVD:
3724                 cm_id->state = IB_CM_ESTABLISHED;
3725                 break;
3726         case IB_CM_ESTABLISHED:
3727                 ret = -EISCONN;
3728                 break;
3729         default:
3730                 ret = -EINVAL;
3731                 break;
3732         }
3733         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3734
3735         if (ret) {
3736                 kfree(work);
3737                 goto out;
3738         }
3739
3740         /*
3741          * The CM worker thread may try to destroy the cm_id before it
3742          * can execute this work item.  To prevent potential deadlock,
3743          * we need to find the cm_id once we're in the context of the
3744          * worker thread, rather than holding a reference on it.
3745          */
3746         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3747         work->local_id = cm_id->local_id;
3748         work->remote_id = cm_id->remote_id;
3749         work->mad_recv_wc = NULL;
3750         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3751
3752         /* Check if the device started its remove_one */
3753         spin_lock_irqsave(&cm.lock, flags);
3754         if (!cm_dev->going_down) {
3755                 queue_delayed_work(cm.wq, &work->work, 0);
3756         } else {
3757                 kfree(work);
3758                 ret = -ENODEV;
3759         }
3760         spin_unlock_irqrestore(&cm.lock, flags);
3761
3762 out:
3763         return ret;
3764 }
3765
3766 static int cm_migrate(struct ib_cm_id *cm_id)
3767 {
3768         struct cm_id_private *cm_id_priv;
3769         struct cm_av tmp_av;
3770         unsigned long flags;
3771         int tmp_send_port_not_ready;
3772         int ret = 0;
3773
3774         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3775         spin_lock_irqsave(&cm_id_priv->lock, flags);
3776         if (cm_id->state == IB_CM_ESTABLISHED &&
3777             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3778              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3779                 cm_id->lap_state = IB_CM_LAP_IDLE;
3780                 /* Swap address vector */
3781                 tmp_av = cm_id_priv->av;
3782                 cm_id_priv->av = cm_id_priv->alt_av;
3783                 cm_id_priv->alt_av = tmp_av;
3784                 /* Swap port send ready state */
3785                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3786                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3787                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3788         } else
3789                 ret = -EINVAL;
3790         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3791
3792         return ret;
3793 }
3794
3795 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3796 {
3797         int ret;
3798
3799         switch (event) {
3800         case IB_EVENT_COMM_EST:
3801                 ret = cm_establish(cm_id);
3802                 break;
3803         case IB_EVENT_PATH_MIG:
3804                 ret = cm_migrate(cm_id);
3805                 break;
3806         default:
3807                 ret = -EINVAL;
3808         }
3809         return ret;
3810 }
3811 EXPORT_SYMBOL(ib_cm_notify);
3812
3813 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
3814                             struct ib_mad_send_buf *send_buf,
3815                             struct ib_mad_recv_wc *mad_recv_wc)
3816 {
3817         struct cm_port *port = mad_agent->context;
3818         struct cm_work *work;
3819         enum ib_cm_event_type event;
3820         u16 attr_id;
3821         int paths = 0;
3822         int going_down = 0;
3823
3824         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
3825         case CM_REQ_ATTR_ID:
3826                 paths = 1 + (((struct cm_req_msg *) mad_recv_wc->recv_buf.mad)->
3827                                                     alt_local_lid != 0);
3828                 event = IB_CM_REQ_RECEIVED;
3829                 break;
3830         case CM_MRA_ATTR_ID:
3831                 event = IB_CM_MRA_RECEIVED;
3832                 break;
3833         case CM_REJ_ATTR_ID:
3834                 event = IB_CM_REJ_RECEIVED;
3835                 break;
3836         case CM_REP_ATTR_ID:
3837                 event = IB_CM_REP_RECEIVED;
3838                 break;
3839         case CM_RTU_ATTR_ID:
3840                 event = IB_CM_RTU_RECEIVED;
3841                 break;
3842         case CM_DREQ_ATTR_ID:
3843                 event = IB_CM_DREQ_RECEIVED;
3844                 break;
3845         case CM_DREP_ATTR_ID:
3846                 event = IB_CM_DREP_RECEIVED;
3847                 break;
3848         case CM_SIDR_REQ_ATTR_ID:
3849                 event = IB_CM_SIDR_REQ_RECEIVED;
3850                 break;
3851         case CM_SIDR_REP_ATTR_ID:
3852                 event = IB_CM_SIDR_REP_RECEIVED;
3853                 break;
3854         case CM_LAP_ATTR_ID:
3855                 paths = 1;
3856                 event = IB_CM_LAP_RECEIVED;
3857                 break;
3858         case CM_APR_ATTR_ID:
3859                 event = IB_CM_APR_RECEIVED;
3860                 break;
3861         default:
3862                 ib_free_recv_mad(mad_recv_wc);
3863                 return;
3864         }
3865
3866         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
3867         atomic_long_inc(&port->counter_group[CM_RECV].
3868                         counter[attr_id - CM_ATTR_ID_OFFSET]);
3869
3870         work = kmalloc(sizeof(*work) + sizeof(struct sa_path_rec) * paths,
3871                        GFP_KERNEL);
3872         if (!work) {
3873                 ib_free_recv_mad(mad_recv_wc);
3874                 return;
3875         }
3876
3877         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3878         work->cm_event.event = event;
3879         work->mad_recv_wc = mad_recv_wc;
3880         work->port = port;
3881
3882         /* Check if the device started its remove_one */
3883         spin_lock_irq(&cm.lock);
3884         if (!port->cm_dev->going_down)
3885                 queue_delayed_work(cm.wq, &work->work, 0);
3886         else
3887                 going_down = 1;
3888         spin_unlock_irq(&cm.lock);
3889
3890         if (going_down) {
3891                 kfree(work);
3892                 ib_free_recv_mad(mad_recv_wc);
3893         }
3894 }
3895
3896 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
3897                                 struct ib_qp_attr *qp_attr,
3898                                 int *qp_attr_mask)
3899 {
3900         unsigned long flags;
3901         int ret;
3902
3903         spin_lock_irqsave(&cm_id_priv->lock, flags);
3904         switch (cm_id_priv->id.state) {
3905         case IB_CM_REQ_SENT:
3906         case IB_CM_MRA_REQ_RCVD:
3907         case IB_CM_REQ_RCVD:
3908         case IB_CM_MRA_REQ_SENT:
3909         case IB_CM_REP_RCVD:
3910         case IB_CM_MRA_REP_SENT:
3911         case IB_CM_REP_SENT:
3912         case IB_CM_MRA_REP_RCVD:
3913         case IB_CM_ESTABLISHED:
3914                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
3915                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
3916                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
3917                 if (cm_id_priv->responder_resources)
3918                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
3919                                                     IB_ACCESS_REMOTE_ATOMIC;
3920                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
3921                 qp_attr->port_num = cm_id_priv->av.port->port_num;
3922                 ret = 0;
3923                 break;
3924         default:
3925                 ret = -EINVAL;
3926                 break;
3927         }
3928         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3929         return ret;
3930 }
3931
3932 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
3933                                struct ib_qp_attr *qp_attr,
3934                                int *qp_attr_mask)
3935 {
3936         unsigned long flags;
3937         int ret;
3938
3939         spin_lock_irqsave(&cm_id_priv->lock, flags);
3940         switch (cm_id_priv->id.state) {
3941         case IB_CM_REQ_RCVD:
3942         case IB_CM_MRA_REQ_SENT:
3943         case IB_CM_REP_RCVD:
3944         case IB_CM_MRA_REP_SENT:
3945         case IB_CM_REP_SENT:
3946         case IB_CM_MRA_REP_RCVD:
3947         case IB_CM_ESTABLISHED:
3948                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
3949                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
3950                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
3951                 qp_attr->path_mtu = cm_id_priv->path_mtu;
3952                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
3953                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
3954                 if (cm_id_priv->qp_type == IB_QPT_RC ||
3955                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
3956                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
3957                                          IB_QP_MIN_RNR_TIMER;
3958                         qp_attr->max_dest_rd_atomic =
3959                                         cm_id_priv->responder_resources;
3960                         qp_attr->min_rnr_timer = 0;
3961                 }
3962                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
3963                         *qp_attr_mask |= IB_QP_ALT_PATH;
3964                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
3965                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
3966                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
3967                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
3968                 }
3969                 ret = 0;
3970                 break;
3971         default:
3972                 ret = -EINVAL;
3973                 break;
3974         }
3975         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3976         return ret;
3977 }
3978
3979 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
3980                                struct ib_qp_attr *qp_attr,
3981                                int *qp_attr_mask)
3982 {
3983         unsigned long flags;
3984         int ret;
3985
3986         spin_lock_irqsave(&cm_id_priv->lock, flags);
3987         switch (cm_id_priv->id.state) {
3988         /* Allow transition to RTS before sending REP */
3989         case IB_CM_REQ_RCVD:
3990         case IB_CM_MRA_REQ_SENT:
3991
3992         case IB_CM_REP_RCVD:
3993         case IB_CM_MRA_REP_SENT:
3994         case IB_CM_REP_SENT:
3995         case IB_CM_MRA_REP_RCVD:
3996         case IB_CM_ESTABLISHED:
3997                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
3998                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
3999                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4000                         switch (cm_id_priv->qp_type) {
4001                         case IB_QPT_RC:
4002                         case IB_QPT_XRC_INI:
4003                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4004                                                  IB_QP_MAX_QP_RD_ATOMIC;
4005                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
4006                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4007                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4008                                 /* fall through */
4009                         case IB_QPT_XRC_TGT:
4010                                 *qp_attr_mask |= IB_QP_TIMEOUT;
4011                                 qp_attr->timeout = cm_id_priv->av.timeout;
4012                                 break;
4013                         default:
4014                                 break;
4015                         }
4016                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4017                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4018                                 qp_attr->path_mig_state = IB_MIG_REARM;
4019                         }
4020                 } else {
4021                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4022                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4023                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4024                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4025                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4026                         qp_attr->path_mig_state = IB_MIG_REARM;
4027                 }
4028                 ret = 0;
4029                 break;
4030         default:
4031                 ret = -EINVAL;
4032                 break;
4033         }
4034         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4035         return ret;
4036 }
4037
4038 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4039                        struct ib_qp_attr *qp_attr,
4040                        int *qp_attr_mask)
4041 {
4042         struct cm_id_private *cm_id_priv;
4043         int ret;
4044
4045         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4046         switch (qp_attr->qp_state) {
4047         case IB_QPS_INIT:
4048                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4049                 break;
4050         case IB_QPS_RTR:
4051                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4052                 break;
4053         case IB_QPS_RTS:
4054                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4055                 break;
4056         default:
4057                 ret = -EINVAL;
4058                 break;
4059         }
4060         return ret;
4061 }
4062 EXPORT_SYMBOL(ib_cm_init_qp_attr);
4063
4064 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
4065                                char *buf)
4066 {
4067         struct cm_counter_group *group;
4068         struct cm_counter_attribute *cm_attr;
4069
4070         group = container_of(obj, struct cm_counter_group, obj);
4071         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
4072
4073         return sprintf(buf, "%ld\n",
4074                        atomic_long_read(&group->counter[cm_attr->index]));
4075 }
4076
4077 static const struct sysfs_ops cm_counter_ops = {
4078         .show = cm_show_counter
4079 };
4080
4081 static struct kobj_type cm_counter_obj_type = {
4082         .sysfs_ops = &cm_counter_ops,
4083         .default_attrs = cm_counter_default_attrs
4084 };
4085
4086 static void cm_release_port_obj(struct kobject *obj)
4087 {
4088         struct cm_port *cm_port;
4089
4090         cm_port = container_of(obj, struct cm_port, port_obj);
4091         kfree(cm_port);
4092 }
4093
4094 static struct kobj_type cm_port_obj_type = {
4095         .release = cm_release_port_obj
4096 };
4097
4098 static char *cm_devnode(struct device *dev, umode_t *mode)
4099 {
4100         if (mode)
4101                 *mode = 0666;
4102         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
4103 }
4104
4105 struct class cm_class = {
4106         .owner   = THIS_MODULE,
4107         .name    = "infiniband_cm",
4108         .devnode = cm_devnode,
4109 };
4110 EXPORT_SYMBOL(cm_class);
4111
4112 static int cm_create_port_fs(struct cm_port *port)
4113 {
4114         int i, ret;
4115
4116         ret = kobject_init_and_add(&port->port_obj, &cm_port_obj_type,
4117                                    &port->cm_dev->device->kobj,
4118                                    "%d", port->port_num);
4119         if (ret) {
4120                 kfree(port);
4121                 return ret;
4122         }
4123
4124         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
4125                 ret = kobject_init_and_add(&port->counter_group[i].obj,
4126                                            &cm_counter_obj_type,
4127                                            &port->port_obj,
4128                                            "%s", counter_group_names[i]);
4129                 if (ret)
4130                         goto error;
4131         }
4132
4133         return 0;
4134
4135 error:
4136         while (i--)
4137                 kobject_put(&port->counter_group[i].obj);
4138         kobject_put(&port->port_obj);
4139         return ret;
4140
4141 }
4142
4143 static void cm_remove_port_fs(struct cm_port *port)
4144 {
4145         int i;
4146
4147         for (i = 0; i < CM_COUNTER_GROUPS; i++)
4148                 kobject_put(&port->counter_group[i].obj);
4149
4150         kobject_put(&port->port_obj);
4151 }
4152
4153 static void cm_add_one(struct ib_device *ib_device)
4154 {
4155         struct cm_device *cm_dev;
4156         struct cm_port *port;
4157         struct ib_mad_reg_req reg_req = {
4158                 .mgmt_class = IB_MGMT_CLASS_CM,
4159                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4160         };
4161         struct ib_port_modify port_modify = {
4162                 .set_port_cap_mask = IB_PORT_CM_SUP
4163         };
4164         unsigned long flags;
4165         int ret;
4166         int count = 0;
4167         u8 i;
4168
4169         cm_dev = kzalloc(sizeof(*cm_dev) + sizeof(*port) *
4170                          ib_device->phys_port_cnt, GFP_KERNEL);
4171         if (!cm_dev)
4172                 return;
4173
4174         cm_dev->ib_device = ib_device;
4175         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4176         cm_dev->going_down = 0;
4177         cm_dev->device = device_create(&cm_class, &ib_device->dev,
4178                                        MKDEV(0, 0), NULL,
4179                                        "%s", ib_device->name);
4180         if (IS_ERR(cm_dev->device)) {
4181                 kfree(cm_dev);
4182                 return;
4183         }
4184
4185         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4186         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4187                 if (!rdma_cap_ib_cm(ib_device, i))
4188                         continue;
4189
4190                 port = kzalloc(sizeof *port, GFP_KERNEL);
4191                 if (!port)
4192                         goto error1;
4193
4194                 cm_dev->port[i-1] = port;
4195                 port->cm_dev = cm_dev;
4196                 port->port_num = i;
4197
4198                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4199                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4200
4201                 ret = cm_create_port_fs(port);
4202                 if (ret)
4203                         goto error1;
4204
4205                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4206                                                         IB_QPT_GSI,
4207                                                         &reg_req,
4208                                                         0,
4209                                                         cm_send_handler,
4210                                                         cm_recv_handler,
4211                                                         port,
4212                                                         0);
4213                 if (IS_ERR(port->mad_agent))
4214                         goto error2;
4215
4216                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4217                 if (ret)
4218                         goto error3;
4219
4220                 count++;
4221         }
4222
4223         if (!count)
4224                 goto free;
4225
4226         ib_set_client_data(ib_device, &cm_client, cm_dev);
4227
4228         write_lock_irqsave(&cm.device_lock, flags);
4229         list_add_tail(&cm_dev->list, &cm.device_list);
4230         write_unlock_irqrestore(&cm.device_lock, flags);
4231         return;
4232
4233 error3:
4234         ib_unregister_mad_agent(port->mad_agent);
4235 error2:
4236         cm_remove_port_fs(port);
4237 error1:
4238         port_modify.set_port_cap_mask = 0;
4239         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4240         while (--i) {
4241                 if (!rdma_cap_ib_cm(ib_device, i))
4242                         continue;
4243
4244                 port = cm_dev->port[i-1];
4245                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4246                 ib_unregister_mad_agent(port->mad_agent);
4247                 cm_remove_port_fs(port);
4248         }
4249 free:
4250         device_unregister(cm_dev->device);
4251         kfree(cm_dev);
4252 }
4253
4254 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4255 {
4256         struct cm_device *cm_dev = client_data;
4257         struct cm_port *port;
4258         struct cm_id_private *cm_id_priv;
4259         struct ib_mad_agent *cur_mad_agent;
4260         struct ib_port_modify port_modify = {
4261                 .clr_port_cap_mask = IB_PORT_CM_SUP
4262         };
4263         unsigned long flags;
4264         int i;
4265
4266         if (!cm_dev)
4267                 return;
4268
4269         write_lock_irqsave(&cm.device_lock, flags);
4270         list_del(&cm_dev->list);
4271         write_unlock_irqrestore(&cm.device_lock, flags);
4272
4273         spin_lock_irq(&cm.lock);
4274         cm_dev->going_down = 1;
4275         spin_unlock_irq(&cm.lock);
4276
4277         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4278                 if (!rdma_cap_ib_cm(ib_device, i))
4279                         continue;
4280
4281                 port = cm_dev->port[i-1];
4282                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4283                 /* Mark all the cm_id's as not valid */
4284                 spin_lock_irq(&cm.lock);
4285                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4286                         cm_id_priv->altr_send_port_not_ready = 1;
4287                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4288                         cm_id_priv->prim_send_port_not_ready = 1;
4289                 spin_unlock_irq(&cm.lock);
4290                 /*
4291                  * We flush the queue here after the going_down set, this
4292                  * verify that no new works will be queued in the recv handler,
4293                  * after that we can call the unregister_mad_agent
4294                  */
4295                 flush_workqueue(cm.wq);
4296                 spin_lock_irq(&cm.state_lock);
4297                 cur_mad_agent = port->mad_agent;
4298                 port->mad_agent = NULL;
4299                 spin_unlock_irq(&cm.state_lock);
4300                 ib_unregister_mad_agent(cur_mad_agent);
4301                 cm_remove_port_fs(port);
4302         }
4303
4304         device_unregister(cm_dev->device);
4305         kfree(cm_dev);
4306 }
4307
4308 static int __init ib_cm_init(void)
4309 {
4310         int ret;
4311
4312         memset(&cm, 0, sizeof cm);
4313         INIT_LIST_HEAD(&cm.device_list);
4314         rwlock_init(&cm.device_lock);
4315         spin_lock_init(&cm.lock);
4316         spin_lock_init(&cm.state_lock);
4317         cm.listen_service_table = RB_ROOT;
4318         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4319         cm.remote_id_table = RB_ROOT;
4320         cm.remote_qp_table = RB_ROOT;
4321         cm.remote_sidr_table = RB_ROOT;
4322         idr_init(&cm.local_id_table);
4323         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4324         INIT_LIST_HEAD(&cm.timewait_list);
4325
4326         ret = class_register(&cm_class);
4327         if (ret) {
4328                 ret = -ENOMEM;
4329                 goto error1;
4330         }
4331
4332         cm.wq = alloc_workqueue("ib_cm", 0, 1);
4333         if (!cm.wq) {
4334                 ret = -ENOMEM;
4335                 goto error2;
4336         }
4337
4338         ret = ib_register_client(&cm_client);
4339         if (ret)
4340                 goto error3;
4341
4342         return 0;
4343 error3:
4344         destroy_workqueue(cm.wq);
4345 error2:
4346         class_unregister(&cm_class);
4347 error1:
4348         idr_destroy(&cm.local_id_table);
4349         return ret;
4350 }
4351
4352 static void __exit ib_cm_cleanup(void)
4353 {
4354         struct cm_timewait_info *timewait_info, *tmp;
4355
4356         spin_lock_irq(&cm.lock);
4357         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4358                 cancel_delayed_work(&timewait_info->work.work);
4359         spin_unlock_irq(&cm.lock);
4360
4361         ib_unregister_client(&cm_client);
4362         destroy_workqueue(cm.wq);
4363
4364         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4365                 list_del(&timewait_info->list);
4366                 kfree(timewait_info);
4367         }
4368
4369         class_unregister(&cm_class);
4370         idr_destroy(&cm.local_id_table);
4371 }
4372
4373 module_init(ib_cm_init);
4374 module_exit(ib_cm_cleanup);
4375