drivers/infiniband/core/cm.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2004-2007 Intel Corporation.  All rights reserved.
   4  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   5  * Copyright (c) 2004, 2005 Voltaire Corporation.  All rights reserved.
   6  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   7  * Copyright (c) 2019, Mellanox Technologies inc.  All rights reserved.
   8  */
   9
  10 #include <linux/completion.h>
  11 #include <linux/dma-mapping.h>
  12 #include <linux/device.h>
  13 #include <linux/module.h>
  14 #include <linux/err.h>
  15 #include <linux/idr.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/random.h>
  18 #include <linux/rbtree.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/slab.h>
  21 #include <linux/sysfs.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/kdev_t.h>
  24 #include <linux/etherdevice.h>
  25
  26 #include <rdma/ib_cache.h>
  27 #include <rdma/ib_cm.h>
  28 #include "cm_msgs.h"
  29 #include "core_priv.h"
  30
  31 MODULE_AUTHOR("Sean Hefty");
  32 MODULE_DESCRIPTION("InfiniBand CM");
  33 MODULE_LICENSE("Dual BSD/GPL");
  34
  35 static const char * const ibcm_rej_reason_strs[] = {
  36         [IB_CM_REJ_NO_QP]                       = "no QP",
  37         [IB_CM_REJ_NO_EEC]                      = "no EEC",
  38         [IB_CM_REJ_NO_RESOURCES]                = "no resources",
  39         [IB_CM_REJ_TIMEOUT]                     = "timeout",
  40         [IB_CM_REJ_UNSUPPORTED]                 = "unsupported",
  41         [IB_CM_REJ_INVALID_COMM_ID]             = "invalid comm ID",
  42         [IB_CM_REJ_INVALID_COMM_INSTANCE]       = "invalid comm instance",
  43         [IB_CM_REJ_INVALID_SERVICE_ID]          = "invalid service ID",
  44         [IB_CM_REJ_INVALID_TRANSPORT_TYPE]      = "invalid transport type",
  45         [IB_CM_REJ_STALE_CONN]                  = "stale conn",
  46         [IB_CM_REJ_RDC_NOT_EXIST]               = "RDC not exist",
  47         [IB_CM_REJ_INVALID_GID]                 = "invalid GID",
  48         [IB_CM_REJ_INVALID_LID]                 = "invalid LID",
  49         [IB_CM_REJ_INVALID_SL]                  = "invalid SL",
  50         [IB_CM_REJ_INVALID_TRAFFIC_CLASS]       = "invalid traffic class",
  51         [IB_CM_REJ_INVALID_HOP_LIMIT]           = "invalid hop limit",
  52         [IB_CM_REJ_INVALID_PACKET_RATE]         = "invalid packet rate",
  53         [IB_CM_REJ_INVALID_ALT_GID]             = "invalid alt GID",
  54         [IB_CM_REJ_INVALID_ALT_LID]             = "invalid alt LID",
  55         [IB_CM_REJ_INVALID_ALT_SL]              = "invalid alt SL",
  56         [IB_CM_REJ_INVALID_ALT_TRAFFIC_CLASS]   = "invalid alt traffic class",
  57         [IB_CM_REJ_INVALID_ALT_HOP_LIMIT]       = "invalid alt hop limit",
  58         [IB_CM_REJ_INVALID_ALT_PACKET_RATE]     = "invalid alt packet rate",
  59         [IB_CM_REJ_PORT_CM_REDIRECT]            = "port CM redirect",
  60         [IB_CM_REJ_PORT_REDIRECT]               = "port redirect",
  61         [IB_CM_REJ_INVALID_MTU]                 = "invalid MTU",
  62         [IB_CM_REJ_INSUFFICIENT_RESP_RESOURCES] = "insufficient resp resources",
  63         [IB_CM_REJ_CONSUMER_DEFINED]            = "consumer defined",
  64         [IB_CM_REJ_INVALID_RNR_RETRY]           = "invalid RNR retry",
  65         [IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID]     = "duplicate local comm ID",
  66         [IB_CM_REJ_INVALID_CLASS_VERSION]       = "invalid class version",
  67         [IB_CM_REJ_INVALID_FLOW_LABEL]          = "invalid flow label",
  68         [IB_CM_REJ_INVALID_ALT_FLOW_LABEL]      = "invalid alt flow label",
  69         [IB_CM_REJ_VENDOR_OPTION_NOT_SUPPORTED] =
  70                 "vendor option is not supported",
  71 };
  72
  73 const char *__attribute_const__ ibcm_reject_msg(int reason)
  74 {
  75         size_t index = reason;
  76
  77         if (index < ARRAY_SIZE(ibcm_rej_reason_strs) &&
  78             ibcm_rej_reason_strs[index])
  79                 return ibcm_rej_reason_strs[index];
  80         else
  81                 return "unrecognized reason";
  82 }
  83 EXPORT_SYMBOL(ibcm_reject_msg);
  84
  85 struct cm_id_private;
  86 struct cm_work;
  87 static int cm_add_one(struct ib_device *device);
  88 static void cm_remove_one(struct ib_device *device, void *client_data);
  89 static void cm_process_work(struct cm_id_private *cm_id_priv,
  90                             struct cm_work *work);
  91 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
  92                                    struct ib_cm_sidr_rep_param *param);
  93 static int cm_send_dreq_locked(struct cm_id_private *cm_id_priv,
  94                                const void *private_data, u8 private_data_len);
  95 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
  96                                void *private_data, u8 private_data_len);
  97 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
  98                               enum ib_cm_rej_reason reason, void *ari,
  99                               u8 ari_length, const void *private_data,
 100                               u8 private_data_len);
 101
 102 static struct ib_client cm_client = {
 103         .name   = "cm",
 104         .add    = cm_add_one,
 105         .remove = cm_remove_one
 106 };
 107
 108 static struct ib_cm {
 109         spinlock_t lock;
 110         struct list_head device_list;
 111         rwlock_t device_lock;
 112         struct rb_root listen_service_table;
 113         u64 listen_service_id;
 114         /* struct rb_root peer_service_table; todo: fix peer to peer */
 115         struct rb_root remote_qp_table;
 116         struct rb_root remote_id_table;
 117         struct rb_root remote_sidr_table;
 118         struct xarray local_id_table;
 119         u32 local_id_next;
 120         __be32 random_id_operand;
 121         struct list_head timewait_list;
 122         struct workqueue_struct *wq;
 123         /* Sync on cm change port state */
 124         spinlock_t state_lock;
 125 } cm;
 126
 127 /* Counter indexes ordered by attribute ID */
 128 enum {
 129         CM_REQ_COUNTER,
 130         CM_MRA_COUNTER,
 131         CM_REJ_COUNTER,
 132         CM_REP_COUNTER,
 133         CM_RTU_COUNTER,
 134         CM_DREQ_COUNTER,
 135         CM_DREP_COUNTER,
 136         CM_SIDR_REQ_COUNTER,
 137         CM_SIDR_REP_COUNTER,
 138         CM_LAP_COUNTER,
 139         CM_APR_COUNTER,
 140         CM_ATTR_COUNT,
 141         CM_ATTR_ID_OFFSET = 0x0010,
 142 };
 143
 144 enum {
 145         CM_XMIT,
 146         CM_XMIT_RETRIES,
 147         CM_RECV,
 148         CM_RECV_DUPLICATES,
 149         CM_COUNTER_GROUPS
 150 };
 151
 152 static char const counter_group_names[CM_COUNTER_GROUPS]
 153                                      [sizeof("cm_rx_duplicates")] = {
 154         "cm_tx_msgs", "cm_tx_retries",
 155         "cm_rx_msgs", "cm_rx_duplicates"
 156 };
 157
 158 struct cm_counter_group {
 159         struct kobject obj;
 160         atomic_long_t counter[CM_ATTR_COUNT];
 161 };
 162
 163 struct cm_counter_attribute {
 164         struct attribute attr;
 165         int index;
 166 };
 167
 168 #define CM_COUNTER_ATTR(_name, _index) \
 169 struct cm_counter_attribute cm_##_name##_counter_attr = { \
 170         .attr = { .name = __stringify(_name), .mode = 0444 }, \
 171         .index = _index \
 172 }
 173
 174 static CM_COUNTER_ATTR(req, CM_REQ_COUNTER);
 175 static CM_COUNTER_ATTR(mra, CM_MRA_COUNTER);
 176 static CM_COUNTER_ATTR(rej, CM_REJ_COUNTER);
 177 static CM_COUNTER_ATTR(rep, CM_REP_COUNTER);
 178 static CM_COUNTER_ATTR(rtu, CM_RTU_COUNTER);
 179 static CM_COUNTER_ATTR(dreq, CM_DREQ_COUNTER);
 180 static CM_COUNTER_ATTR(drep, CM_DREP_COUNTER);
 181 static CM_COUNTER_ATTR(sidr_req, CM_SIDR_REQ_COUNTER);
 182 static CM_COUNTER_ATTR(sidr_rep, CM_SIDR_REP_COUNTER);
 183 static CM_COUNTER_ATTR(lap, CM_LAP_COUNTER);
 184 static CM_COUNTER_ATTR(apr, CM_APR_COUNTER);
 185
 186 static struct attribute *cm_counter_default_attrs[] = {
 187         &cm_req_counter_attr.attr,
 188         &cm_mra_counter_attr.attr,
 189         &cm_rej_counter_attr.attr,
 190         &cm_rep_counter_attr.attr,
 191         &cm_rtu_counter_attr.attr,
 192         &cm_dreq_counter_attr.attr,
 193         &cm_drep_counter_attr.attr,
 194         &cm_sidr_req_counter_attr.attr,
 195         &cm_sidr_rep_counter_attr.attr,
 196         &cm_lap_counter_attr.attr,
 197         &cm_apr_counter_attr.attr,
 198         NULL
 199 };
 200
 201 struct cm_port {
 202         struct cm_device *cm_dev;
 203         struct ib_mad_agent *mad_agent;
 204         struct kobject port_obj;
 205         u8 port_num;
 206         struct list_head cm_priv_prim_list;
 207         struct list_head cm_priv_altr_list;
 208         struct cm_counter_group counter_group[CM_COUNTER_GROUPS];
 209 };
 210
 211 struct cm_device {
 212         struct list_head list;
 213         struct ib_device *ib_device;
 214         u8 ack_delay;
 215         int going_down;
 216         struct cm_port *port[];
 217 };
 218
 219 struct cm_av {
 220         struct cm_port *port;
 221         union ib_gid dgid;
 222         struct rdma_ah_attr ah_attr;
 223         u16 pkey_index;
 224         u8 timeout;
 225 };
 226
 227 struct cm_work {
 228         struct delayed_work work;
 229         struct list_head list;
 230         struct cm_port *port;
 231         struct ib_mad_recv_wc *mad_recv_wc;     /* Received MADs */
 232         __be32 local_id;                        /* Established / timewait */
 233         __be32 remote_id;
 234         struct ib_cm_event cm_event;
 235         struct sa_path_rec path[];
 236 };
 237
 238 struct cm_timewait_info {
 239         struct cm_work work;
 240         struct list_head list;
 241         struct rb_node remote_qp_node;
 242         struct rb_node remote_id_node;
 243         __be64 remote_ca_guid;
 244         __be32 remote_qpn;
 245         u8 inserted_remote_qp;
 246         u8 inserted_remote_id;
 247 };
 248
 249 struct cm_id_private {
 250         struct ib_cm_id id;
 251
 252         struct rb_node service_node;
 253         struct rb_node sidr_id_node;
 254         spinlock_t lock;        /* Do not acquire inside cm.lock */
 255         struct completion comp;
 256         refcount_t refcount;
 257         /* Number of clients sharing this ib_cm_id. Only valid for listeners.
 258          * Protected by the cm.lock spinlock. */
 259         int listen_sharecount;
 260         struct rcu_head rcu;
 261
 262         struct ib_mad_send_buf *msg;
 263         struct cm_timewait_info *timewait_info;
 264         /* todo: use alternate port on send failure */
 265         struct cm_av av;
 266         struct cm_av alt_av;
 267
 268         void *private_data;
 269         __be64 tid;
 270         __be32 local_qpn;
 271         __be32 remote_qpn;
 272         enum ib_qp_type qp_type;
 273         __be32 sq_psn;
 274         __be32 rq_psn;
 275         int timeout_ms;
 276         enum ib_mtu path_mtu;
 277         __be16 pkey;
 278         u8 private_data_len;
 279         u8 max_cm_retries;
 280         u8 responder_resources;
 281         u8 initiator_depth;
 282         u8 retry_count;
 283         u8 rnr_retry_count;
 284         u8 service_timeout;
 285         u8 target_ack_delay;
 286
 287         struct list_head prim_list;
 288         struct list_head altr_list;
 289         /* Indicates that the send port mad is registered and av is set */
 290         int prim_send_port_not_ready;
 291         int altr_send_port_not_ready;
 292
 293         struct list_head work_list;
 294         atomic_t work_count;
 295
 296         struct rdma_ucm_ece ece;
 297 };
 298
 299 static void cm_work_handler(struct work_struct *work);
 300
 301 static inline void cm_deref_id(struct cm_id_private *cm_id_priv)
 302 {
 303         if (refcount_dec_and_test(&cm_id_priv->refcount))
 304                 complete(&cm_id_priv->comp);
 305 }
 306
 307 static int cm_alloc_msg(struct cm_id_private *cm_id_priv,
 308                         struct ib_mad_send_buf **msg)
 309 {
 310         struct ib_mad_agent *mad_agent;
 311         struct ib_mad_send_buf *m;
 312         struct ib_ah *ah;
 313         struct cm_av *av;
 314         unsigned long flags, flags2;
 315         int ret = 0;
 316
 317         /* don't let the port to be released till the agent is down */
 318         spin_lock_irqsave(&cm.state_lock, flags2);
 319         spin_lock_irqsave(&cm.lock, flags);
 320         if (!cm_id_priv->prim_send_port_not_ready)
 321                 av = &cm_id_priv->av;
 322         else if (!cm_id_priv->altr_send_port_not_ready &&
 323                  (cm_id_priv->alt_av.port))
 324                 av = &cm_id_priv->alt_av;
 325         else {
 326                 pr_info("%s: not valid CM id\n", __func__);
 327                 ret = -ENODEV;
 328                 spin_unlock_irqrestore(&cm.lock, flags);
 329                 goto out;
 330         }
 331         spin_unlock_irqrestore(&cm.lock, flags);
 332         /* Make sure the port haven't released the mad yet */
 333         mad_agent = cm_id_priv->av.port->mad_agent;
 334         if (!mad_agent) {
 335                 pr_info("%s: not a valid MAD agent\n", __func__);
 336                 ret = -ENODEV;
 337                 goto out;
 338         }
 339         ah = rdma_create_ah(mad_agent->qp->pd, &av->ah_attr, 0);
 340         if (IS_ERR(ah)) {
 341                 ret = PTR_ERR(ah);
 342                 goto out;
 343         }
 344
 345         m = ib_create_send_mad(mad_agent, cm_id_priv->id.remote_cm_qpn,
 346                                av->pkey_index,
 347                                0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 348                                GFP_ATOMIC,
 349                                IB_MGMT_BASE_VERSION);
 350         if (IS_ERR(m)) {
 351                 rdma_destroy_ah(ah, 0);
 352                 ret = PTR_ERR(m);
 353                 goto out;
 354         }
 355
 356         /* Timeout set by caller if response is expected. */
 357         m->ah = ah;
 358         m->retries = cm_id_priv->max_cm_retries;
 359
 360         refcount_inc(&cm_id_priv->refcount);
 361         m->context[0] = cm_id_priv;
 362         *msg = m;
 363
 364 out:
 365         spin_unlock_irqrestore(&cm.state_lock, flags2);
 366         return ret;
 367 }
 368
 369 static struct ib_mad_send_buf *cm_alloc_response_msg_no_ah(struct cm_port *port,
 370                                                            struct ib_mad_recv_wc *mad_recv_wc)
 371 {
 372         return ib_create_send_mad(port->mad_agent, 1, mad_recv_wc->wc->pkey_index,
 373                                   0, IB_MGMT_MAD_HDR, IB_MGMT_MAD_DATA,
 374                                   GFP_ATOMIC,
 375                                   IB_MGMT_BASE_VERSION);
 376 }
 377
 378 static int cm_create_response_msg_ah(struct cm_port *port,
 379                                      struct ib_mad_recv_wc *mad_recv_wc,
 380                                      struct ib_mad_send_buf *msg)
 381 {
 382         struct ib_ah *ah;
 383
 384         ah = ib_create_ah_from_wc(port->mad_agent->qp->pd, mad_recv_wc->wc,
 385                                   mad_recv_wc->recv_buf.grh, port->port_num);
 386         if (IS_ERR(ah))
 387                 return PTR_ERR(ah);
 388
 389         msg->ah = ah;
 390         return 0;
 391 }
 392
 393 static void cm_free_msg(struct ib_mad_send_buf *msg)
 394 {
 395         if (msg->ah)
 396                 rdma_destroy_ah(msg->ah, 0);
 397         if (msg->context[0])
 398                 cm_deref_id(msg->context[0]);
 399         ib_free_send_mad(msg);
 400 }
 401
 402 static int cm_alloc_response_msg(struct cm_port *port,
 403                                  struct ib_mad_recv_wc *mad_recv_wc,
 404                                  struct ib_mad_send_buf **msg)
 405 {
 406         struct ib_mad_send_buf *m;
 407         int ret;
 408
 409         m = cm_alloc_response_msg_no_ah(port, mad_recv_wc);
 410         if (IS_ERR(m))
 411                 return PTR_ERR(m);
 412
 413         ret = cm_create_response_msg_ah(port, mad_recv_wc, m);
 414         if (ret) {
 415                 cm_free_msg(m);
 416                 return ret;
 417         }
 418
 419         *msg = m;
 420         return 0;
 421 }
 422
 423 static void * cm_copy_private_data(const void *private_data,
 424                                    u8 private_data_len)
 425 {
 426         void *data;
 427
 428         if (!private_data || !private_data_len)
 429                 return NULL;
 430
 431         data = kmemdup(private_data, private_data_len, GFP_KERNEL);
 432         if (!data)
 433                 return ERR_PTR(-ENOMEM);
 434
 435         return data;
 436 }
 437
 438 static void cm_set_private_data(struct cm_id_private *cm_id_priv,
 439                                  void *private_data, u8 private_data_len)
 440 {
 441         if (cm_id_priv->private_data && cm_id_priv->private_data_len)
 442                 kfree(cm_id_priv->private_data);
 443
 444         cm_id_priv->private_data = private_data;
 445         cm_id_priv->private_data_len = private_data_len;
 446 }
 447
 448 static int cm_init_av_for_lap(struct cm_port *port, struct ib_wc *wc,
 449                               struct ib_grh *grh, struct cm_av *av)
 450 {
 451         struct rdma_ah_attr new_ah_attr;
 452         int ret;
 453
 454         av->port = port;
 455         av->pkey_index = wc->pkey_index;
 456
 457         /*
 458          * av->ah_attr might be initialized based on past wc during incoming
 459          * connect request or while sending out connect request. So initialize
 460          * a new ah_attr on stack. If initialization fails, old ah_attr is
 461          * used for sending any responses. If initialization is successful,
 462          * than new ah_attr is used by overwriting old one.
 463          */
 464         ret = ib_init_ah_attr_from_wc(port->cm_dev->ib_device,
 465                                       port->port_num, wc,
 466                                       grh, &new_ah_attr);
 467         if (ret)
 468                 return ret;
 469
 470         rdma_move_ah_attr(&av->ah_attr, &new_ah_attr);
 471         return 0;
 472 }
 473
 474 static int cm_init_av_for_response(struct cm_port *port, struct ib_wc *wc,
 475                                    struct ib_grh *grh, struct cm_av *av)
 476 {
 477         av->port = port;
 478         av->pkey_index = wc->pkey_index;
 479         return ib_init_ah_attr_from_wc(port->cm_dev->ib_device,
 480                                        port->port_num, wc,
 481                                        grh, &av->ah_attr);
 482 }
 483
 484 static void add_cm_id_to_port_list(struct cm_id_private *cm_id_priv,
 485                                    struct cm_av *av, struct cm_port *port)
 486 {
 487         unsigned long flags;
 488
 489         spin_lock_irqsave(&cm.lock, flags);
 490         if (&cm_id_priv->av == av)
 491                 list_add_tail(&cm_id_priv->prim_list, &port->cm_priv_prim_list);
 492         else if (&cm_id_priv->alt_av == av)
 493                 list_add_tail(&cm_id_priv->altr_list, &port->cm_priv_altr_list);
 494         else
 495                 WARN_ON(true);
 496         spin_unlock_irqrestore(&cm.lock, flags);
 497 }
 498
 499 static struct cm_port *
 500 get_cm_port_from_path(struct sa_path_rec *path, const struct ib_gid_attr *attr)
 501 {
 502         struct cm_device *cm_dev;
 503         struct cm_port *port = NULL;
 504         unsigned long flags;
 505
 506         if (attr) {
 507                 read_lock_irqsave(&cm.device_lock, flags);
 508                 list_for_each_entry(cm_dev, &cm.device_list, list) {
 509                         if (cm_dev->ib_device == attr->device) {
 510                                 port = cm_dev->port[attr->port_num - 1];
 511                                 break;
 512                         }
 513                 }
 514                 read_unlock_irqrestore(&cm.device_lock, flags);
 515         } else {
 516                 /* SGID attribute can be NULL in following
 517                  * conditions.
 518                  * (a) Alternative path
 519                  * (b) IB link layer without GRH
 520                  * (c) LAP send messages
 521                  */
 522                 read_lock_irqsave(&cm.device_lock, flags);
 523                 list_for_each_entry(cm_dev, &cm.device_list, list) {
 524                         attr = rdma_find_gid(cm_dev->ib_device,
 525                                              &path->sgid,
 526                                              sa_conv_pathrec_to_gid_type(path),
 527                                              NULL);
 528                         if (!IS_ERR(attr)) {
 529                                 port = cm_dev->port[attr->port_num - 1];
 530                                 break;
 531                         }
 532                 }
 533                 read_unlock_irqrestore(&cm.device_lock, flags);
 534                 if (port)
 535                         rdma_put_gid_attr(attr);
 536         }
 537         return port;
 538 }
 539
 540 static int cm_init_av_by_path(struct sa_path_rec *path,
 541                               const struct ib_gid_attr *sgid_attr,
 542                               struct cm_av *av,
 543                               struct cm_id_private *cm_id_priv)
 544 {
 545         struct rdma_ah_attr new_ah_attr;
 546         struct cm_device *cm_dev;
 547         struct cm_port *port;
 548         int ret;
 549
 550         port = get_cm_port_from_path(path, sgid_attr);
 551         if (!port)
 552                 return -EINVAL;
 553         cm_dev = port->cm_dev;
 554
 555         ret = ib_find_cached_pkey(cm_dev->ib_device, port->port_num,
 556                                   be16_to_cpu(path->pkey), &av->pkey_index);
 557         if (ret)
 558                 return ret;
 559
 560         av->port = port;
 561
 562         /*
 563          * av->ah_attr might be initialized based on wc or during
 564          * request processing time which might have reference to sgid_attr.
 565          * So initialize a new ah_attr on stack.
 566          * If initialization fails, old ah_attr is used for sending any
 567          * responses. If initialization is successful, than new ah_attr
 568          * is used by overwriting the old one. So that right ah_attr
 569          * can be used to return an error response.
 570          */
 571         ret = ib_init_ah_attr_from_path(cm_dev->ib_device, port->port_num, path,
 572                                         &new_ah_attr, sgid_attr);
 573         if (ret)
 574                 return ret;
 575
 576         av->timeout = path->packet_life_time + 1;
 577         add_cm_id_to_port_list(cm_id_priv, av, port);
 578         rdma_move_ah_attr(&av->ah_attr, &new_ah_attr);
 579         return 0;
 580 }
 581
 582 static u32 cm_local_id(__be32 local_id)
 583 {
 584         return (__force u32) (local_id ^ cm.random_id_operand);
 585 }
 586
 587 static struct cm_id_private *cm_acquire_id(__be32 local_id, __be32 remote_id)
 588 {
 589         struct cm_id_private *cm_id_priv;
 590
 591         rcu_read_lock();
 592         cm_id_priv = xa_load(&cm.local_id_table, cm_local_id(local_id));
 593         if (!cm_id_priv || cm_id_priv->id.remote_id != remote_id ||
 594             !refcount_inc_not_zero(&cm_id_priv->refcount))
 595                 cm_id_priv = NULL;
 596         rcu_read_unlock();
 597
 598         return cm_id_priv;
 599 }
 600
 601 /*
 602  * Trivial helpers to strip endian annotation and compare; the
 603  * endianness doesn't actually matter since we just need a stable
 604  * order for the RB tree.
 605  */
 606 static int be32_lt(__be32 a, __be32 b)
 607 {
 608         return (__force u32) a < (__force u32) b;
 609 }
 610
 611 static int be32_gt(__be32 a, __be32 b)
 612 {
 613         return (__force u32) a > (__force u32) b;
 614 }
 615
 616 static int be64_lt(__be64 a, __be64 b)
 617 {
 618         return (__force u64) a < (__force u64) b;
 619 }
 620
 621 static int be64_gt(__be64 a, __be64 b)
 622 {
 623         return (__force u64) a > (__force u64) b;
 624 }
 625
 626 /*
 627  * Inserts a new cm_id_priv into the listen_service_table. Returns cm_id_priv
 628  * if the new ID was inserted, NULL if it could not be inserted due to a
 629  * collision, or the existing cm_id_priv ready for shared usage.
 630  */
 631 static struct cm_id_private *cm_insert_listen(struct cm_id_private *cm_id_priv,
 632                                               ib_cm_handler shared_handler)
 633 {
 634         struct rb_node **link = &cm.listen_service_table.rb_node;
 635         struct rb_node *parent = NULL;
 636         struct cm_id_private *cur_cm_id_priv;
 637         __be64 service_id = cm_id_priv->id.service_id;
 638         __be64 service_mask = cm_id_priv->id.service_mask;
 639         unsigned long flags;
 640
 641         spin_lock_irqsave(&cm.lock, flags);
 642         while (*link) {
 643                 parent = *link;
 644                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 645                                           service_node);
 646                 if ((cur_cm_id_priv->id.service_mask & service_id) ==
 647                     (service_mask & cur_cm_id_priv->id.service_id) &&
 648                     (cm_id_priv->id.device == cur_cm_id_priv->id.device)) {
 649                         /*
 650                          * Sharing an ib_cm_id with different handlers is not
 651                          * supported
 652                          */
 653                         if (cur_cm_id_priv->id.cm_handler != shared_handler ||
 654                             cur_cm_id_priv->id.context ||
 655                             WARN_ON(!cur_cm_id_priv->id.cm_handler)) {
 656                                 spin_unlock_irqrestore(&cm.lock, flags);
 657                                 return NULL;
 658                         }
 659                         refcount_inc(&cur_cm_id_priv->refcount);
 660                         cur_cm_id_priv->listen_sharecount++;
 661                         spin_unlock_irqrestore(&cm.lock, flags);
 662                         return cur_cm_id_priv;
 663                 }
 664
 665                 if (cm_id_priv->id.device < cur_cm_id_priv->id.device)
 666                         link = &(*link)->rb_left;
 667                 else if (cm_id_priv->id.device > cur_cm_id_priv->id.device)
 668                         link = &(*link)->rb_right;
 669                 else if (be64_lt(service_id, cur_cm_id_priv->id.service_id))
 670                         link = &(*link)->rb_left;
 671                 else if (be64_gt(service_id, cur_cm_id_priv->id.service_id))
 672                         link = &(*link)->rb_right;
 673                 else
 674                         link = &(*link)->rb_right;
 675         }
 676         cm_id_priv->listen_sharecount++;
 677         rb_link_node(&cm_id_priv->service_node, parent, link);
 678         rb_insert_color(&cm_id_priv->service_node, &cm.listen_service_table);
 679         spin_unlock_irqrestore(&cm.lock, flags);
 680         return cm_id_priv;
 681 }
 682
 683 static struct cm_id_private * cm_find_listen(struct ib_device *device,
 684                                              __be64 service_id)
 685 {
 686         struct rb_node *node = cm.listen_service_table.rb_node;
 687         struct cm_id_private *cm_id_priv;
 688
 689         while (node) {
 690                 cm_id_priv = rb_entry(node, struct cm_id_private, service_node);
 691                 if ((cm_id_priv->id.service_mask & service_id) ==
 692                      cm_id_priv->id.service_id &&
 693                     (cm_id_priv->id.device == device)) {
 694                         refcount_inc(&cm_id_priv->refcount);
 695                         return cm_id_priv;
 696                 }
 697                 if (device < cm_id_priv->id.device)
 698                         node = node->rb_left;
 699                 else if (device > cm_id_priv->id.device)
 700                         node = node->rb_right;
 701                 else if (be64_lt(service_id, cm_id_priv->id.service_id))
 702                         node = node->rb_left;
 703                 else if (be64_gt(service_id, cm_id_priv->id.service_id))
 704                         node = node->rb_right;
 705                 else
 706                         node = node->rb_right;
 707         }
 708         return NULL;
 709 }
 710
 711 static struct cm_timewait_info * cm_insert_remote_id(struct cm_timewait_info
 712                                                      *timewait_info)
 713 {
 714         struct rb_node **link = &cm.remote_id_table.rb_node;
 715         struct rb_node *parent = NULL;
 716         struct cm_timewait_info *cur_timewait_info;
 717         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 718         __be32 remote_id = timewait_info->work.remote_id;
 719
 720         while (*link) {
 721                 parent = *link;
 722                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 723                                              remote_id_node);
 724                 if (be32_lt(remote_id, cur_timewait_info->work.remote_id))
 725                         link = &(*link)->rb_left;
 726                 else if (be32_gt(remote_id, cur_timewait_info->work.remote_id))
 727                         link = &(*link)->rb_right;
 728                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 729                         link = &(*link)->rb_left;
 730                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 731                         link = &(*link)->rb_right;
 732                 else
 733                         return cur_timewait_info;
 734         }
 735         timewait_info->inserted_remote_id = 1;
 736         rb_link_node(&timewait_info->remote_id_node, parent, link);
 737         rb_insert_color(&timewait_info->remote_id_node, &cm.remote_id_table);
 738         return NULL;
 739 }
 740
 741 static struct cm_id_private *cm_find_remote_id(__be64 remote_ca_guid,
 742                                                __be32 remote_id)
 743 {
 744         struct rb_node *node = cm.remote_id_table.rb_node;
 745         struct cm_timewait_info *timewait_info;
 746         struct cm_id_private *res = NULL;
 747
 748         spin_lock_irq(&cm.lock);
 749         while (node) {
 750                 timewait_info = rb_entry(node, struct cm_timewait_info,
 751                                          remote_id_node);
 752                 if (be32_lt(remote_id, timewait_info->work.remote_id))
 753                         node = node->rb_left;
 754                 else if (be32_gt(remote_id, timewait_info->work.remote_id))
 755                         node = node->rb_right;
 756                 else if (be64_lt(remote_ca_guid, timewait_info->remote_ca_guid))
 757                         node = node->rb_left;
 758                 else if (be64_gt(remote_ca_guid, timewait_info->remote_ca_guid))
 759                         node = node->rb_right;
 760                 else {
 761                         res = cm_acquire_id(timewait_info->work.local_id,
 762                                              timewait_info->work.remote_id);
 763                         break;
 764                 }
 765         }
 766         spin_unlock_irq(&cm.lock);
 767         return res;
 768 }
 769
 770 static struct cm_timewait_info * cm_insert_remote_qpn(struct cm_timewait_info
 771                                                       *timewait_info)
 772 {
 773         struct rb_node **link = &cm.remote_qp_table.rb_node;
 774         struct rb_node *parent = NULL;
 775         struct cm_timewait_info *cur_timewait_info;
 776         __be64 remote_ca_guid = timewait_info->remote_ca_guid;
 777         __be32 remote_qpn = timewait_info->remote_qpn;
 778
 779         while (*link) {
 780                 parent = *link;
 781                 cur_timewait_info = rb_entry(parent, struct cm_timewait_info,
 782                                              remote_qp_node);
 783                 if (be32_lt(remote_qpn, cur_timewait_info->remote_qpn))
 784                         link = &(*link)->rb_left;
 785                 else if (be32_gt(remote_qpn, cur_timewait_info->remote_qpn))
 786                         link = &(*link)->rb_right;
 787                 else if (be64_lt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 788                         link = &(*link)->rb_left;
 789                 else if (be64_gt(remote_ca_guid, cur_timewait_info->remote_ca_guid))
 790                         link = &(*link)->rb_right;
 791                 else
 792                         return cur_timewait_info;
 793         }
 794         timewait_info->inserted_remote_qp = 1;
 795         rb_link_node(&timewait_info->remote_qp_node, parent, link);
 796         rb_insert_color(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 797         return NULL;
 798 }
 799
 800 static struct cm_id_private * cm_insert_remote_sidr(struct cm_id_private
 801                                                     *cm_id_priv)
 802 {
 803         struct rb_node **link = &cm.remote_sidr_table.rb_node;
 804         struct rb_node *parent = NULL;
 805         struct cm_id_private *cur_cm_id_priv;
 806         union ib_gid *port_gid = &cm_id_priv->av.dgid;
 807         __be32 remote_id = cm_id_priv->id.remote_id;
 808
 809         while (*link) {
 810                 parent = *link;
 811                 cur_cm_id_priv = rb_entry(parent, struct cm_id_private,
 812                                           sidr_id_node);
 813                 if (be32_lt(remote_id, cur_cm_id_priv->id.remote_id))
 814                         link = &(*link)->rb_left;
 815                 else if (be32_gt(remote_id, cur_cm_id_priv->id.remote_id))
 816                         link = &(*link)->rb_right;
 817                 else {
 818                         int cmp;
 819                         cmp = memcmp(port_gid, &cur_cm_id_priv->av.dgid,
 820                                      sizeof *port_gid);
 821                         if (cmp < 0)
 822                                 link = &(*link)->rb_left;
 823                         else if (cmp > 0)
 824                                 link = &(*link)->rb_right;
 825                         else
 826                                 return cur_cm_id_priv;
 827                 }
 828         }
 829         rb_link_node(&cm_id_priv->sidr_id_node, parent, link);
 830         rb_insert_color(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
 831         return NULL;
 832 }
 833
 834 static struct cm_id_private *cm_alloc_id_priv(struct ib_device *device,
 835                                               ib_cm_handler cm_handler,
 836                                               void *context)
 837 {
 838         struct cm_id_private *cm_id_priv;
 839         u32 id;
 840         int ret;
 841
 842         cm_id_priv = kzalloc(sizeof *cm_id_priv, GFP_KERNEL);
 843         if (!cm_id_priv)
 844                 return ERR_PTR(-ENOMEM);
 845
 846         cm_id_priv->id.state = IB_CM_IDLE;
 847         cm_id_priv->id.device = device;
 848         cm_id_priv->id.cm_handler = cm_handler;
 849         cm_id_priv->id.context = context;
 850         cm_id_priv->id.remote_cm_qpn = 1;
 851
 852         RB_CLEAR_NODE(&cm_id_priv->service_node);
 853         RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
 854         spin_lock_init(&cm_id_priv->lock);
 855         init_completion(&cm_id_priv->comp);
 856         INIT_LIST_HEAD(&cm_id_priv->work_list);
 857         INIT_LIST_HEAD(&cm_id_priv->prim_list);
 858         INIT_LIST_HEAD(&cm_id_priv->altr_list);
 859         atomic_set(&cm_id_priv->work_count, -1);
 860         refcount_set(&cm_id_priv->refcount, 1);
 861
 862         ret = xa_alloc_cyclic_irq(&cm.local_id_table, &id, NULL, xa_limit_32b,
 863                                   &cm.local_id_next, GFP_KERNEL);
 864         if (ret < 0)
 865                 goto error;
 866         cm_id_priv->id.local_id = (__force __be32)id ^ cm.random_id_operand;
 867
 868         return cm_id_priv;
 869
 870 error:
 871         kfree(cm_id_priv);
 872         return ERR_PTR(ret);
 873 }
 874
 875 /*
 876  * Make the ID visible to the MAD handlers and other threads that use the
 877  * xarray.
 878  */
 879 static void cm_finalize_id(struct cm_id_private *cm_id_priv)
 880 {
 881         xa_store_irq(&cm.local_id_table, cm_local_id(cm_id_priv->id.local_id),
 882                      cm_id_priv, GFP_KERNEL);
 883 }
 884
 885 struct ib_cm_id *ib_create_cm_id(struct ib_device *device,
 886                                  ib_cm_handler cm_handler,
 887                                  void *context)
 888 {
 889         struct cm_id_private *cm_id_priv;
 890
 891         cm_id_priv = cm_alloc_id_priv(device, cm_handler, context);
 892         if (IS_ERR(cm_id_priv))
 893                 return ERR_CAST(cm_id_priv);
 894
 895         cm_finalize_id(cm_id_priv);
 896         return &cm_id_priv->id;
 897 }
 898 EXPORT_SYMBOL(ib_create_cm_id);
 899
 900 static struct cm_work * cm_dequeue_work(struct cm_id_private *cm_id_priv)
 901 {
 902         struct cm_work *work;
 903
 904         if (list_empty(&cm_id_priv->work_list))
 905                 return NULL;
 906
 907         work = list_entry(cm_id_priv->work_list.next, struct cm_work, list);
 908         list_del(&work->list);
 909         return work;
 910 }
 911
 912 static void cm_free_work(struct cm_work *work)
 913 {
 914         if (work->mad_recv_wc)
 915                 ib_free_recv_mad(work->mad_recv_wc);
 916         kfree(work);
 917 }
 918
 919 static void cm_queue_work_unlock(struct cm_id_private *cm_id_priv,
 920                                  struct cm_work *work)
 921 {
 922         bool immediate;
 923
 924         /*
 925          * To deliver the event to the user callback we have the drop the
 926          * spinlock, however, we need to ensure that the user callback is single
 927          * threaded and receives events in the temporal order. If there are
 928          * already events being processed then thread new events onto a list,
 929          * the thread currently processing will pick them up.
 930          */
 931         immediate = atomic_inc_and_test(&cm_id_priv->work_count);
 932         if (!immediate) {
 933                 list_add_tail(&work->list, &cm_id_priv->work_list);
 934                 /*
 935                  * This routine always consumes incoming reference. Once queued
 936                  * to the work_list then a reference is held by the thread
 937                  * currently running cm_process_work() and this reference is not
 938                  * needed.
 939                  */
 940                 cm_deref_id(cm_id_priv);
 941         }
 942         spin_unlock_irq(&cm_id_priv->lock);
 943
 944         if (immediate)
 945                 cm_process_work(cm_id_priv, work);
 946 }
 947
 948 static inline int cm_convert_to_ms(int iba_time)
 949 {
 950         /* approximate conversion to ms from 4.096us x 2^iba_time */
 951         return 1 << max(iba_time - 8, 0);
 952 }
 953
 954 /*
 955  * calculate: 4.096x2^ack_timeout = 4.096x2^ack_delay + 2x4.096x2^life_time
 956  * Because of how ack_timeout is stored, adding one doubles the timeout.
 957  * To avoid large timeouts, select the max(ack_delay, life_time + 1), and
 958  * increment it (round up) only if the other is within 50%.
 959  */
 960 static u8 cm_ack_timeout(u8 ca_ack_delay, u8 packet_life_time)
 961 {
 962         int ack_timeout = packet_life_time + 1;
 963
 964         if (ack_timeout >= ca_ack_delay)
 965                 ack_timeout += (ca_ack_delay >= (ack_timeout - 1));
 966         else
 967                 ack_timeout = ca_ack_delay +
 968                               (ack_timeout >= (ca_ack_delay - 1));
 969
 970         return min(31, ack_timeout);
 971 }
 972
 973 static void cm_remove_remote(struct cm_id_private *cm_id_priv)
 974 {
 975         struct cm_timewait_info *timewait_info = cm_id_priv->timewait_info;
 976
 977         if (timewait_info->inserted_remote_id) {
 978                 rb_erase(&timewait_info->remote_id_node, &cm.remote_id_table);
 979                 timewait_info->inserted_remote_id = 0;
 980         }
 981
 982         if (timewait_info->inserted_remote_qp) {
 983                 rb_erase(&timewait_info->remote_qp_node, &cm.remote_qp_table);
 984                 timewait_info->inserted_remote_qp = 0;
 985         }
 986 }
 987
 988 static struct cm_timewait_info * cm_create_timewait_info(__be32 local_id)
 989 {
 990         struct cm_timewait_info *timewait_info;
 991
 992         timewait_info = kzalloc(sizeof *timewait_info, GFP_KERNEL);
 993         if (!timewait_info)
 994                 return ERR_PTR(-ENOMEM);
 995
 996         timewait_info->work.local_id = local_id;
 997         INIT_DELAYED_WORK(&timewait_info->work.work, cm_work_handler);
 998         timewait_info->work.cm_event.event = IB_CM_TIMEWAIT_EXIT;
 999         return timewait_info;
1000 }
1001
1002 static void cm_enter_timewait(struct cm_id_private *cm_id_priv)
1003 {
1004         int wait_time;
1005         unsigned long flags;
1006         struct cm_device *cm_dev;
1007
1008         lockdep_assert_held(&cm_id_priv->lock);
1009
1010         cm_dev = ib_get_client_data(cm_id_priv->id.device, &cm_client);
1011         if (!cm_dev)
1012                 return;
1013
1014         spin_lock_irqsave(&cm.lock, flags);
1015         cm_remove_remote(cm_id_priv);
1016         list_add_tail(&cm_id_priv->timewait_info->list, &cm.timewait_list);
1017         spin_unlock_irqrestore(&cm.lock, flags);
1018
1019         /*
1020          * The cm_id could be destroyed by the user before we exit timewait.
1021          * To protect against this, we search for the cm_id after exiting
1022          * timewait before notifying the user that we've exited timewait.
1023          */
1024         cm_id_priv->id.state = IB_CM_TIMEWAIT;
1025         wait_time = cm_convert_to_ms(cm_id_priv->av.timeout);
1026
1027         /* Check if the device started its remove_one */
1028         spin_lock_irqsave(&cm.lock, flags);
1029         if (!cm_dev->going_down)
1030                 queue_delayed_work(cm.wq, &cm_id_priv->timewait_info->work.work,
1031                                    msecs_to_jiffies(wait_time));
1032         spin_unlock_irqrestore(&cm.lock, flags);
1033
1034         /*
1035          * The timewait_info is converted into a work and gets freed during
1036          * cm_free_work() in cm_timewait_handler().
1037          */
1038         BUILD_BUG_ON(offsetof(struct cm_timewait_info, work) != 0);
1039         cm_id_priv->timewait_info = NULL;
1040 }
1041
1042 static void cm_reset_to_idle(struct cm_id_private *cm_id_priv)
1043 {
1044         unsigned long flags;
1045
1046         lockdep_assert_held(&cm_id_priv->lock);
1047
1048         cm_id_priv->id.state = IB_CM_IDLE;
1049         if (cm_id_priv->timewait_info) {
1050                 spin_lock_irqsave(&cm.lock, flags);
1051                 cm_remove_remote(cm_id_priv);
1052                 spin_unlock_irqrestore(&cm.lock, flags);
1053                 kfree(cm_id_priv->timewait_info);
1054                 cm_id_priv->timewait_info = NULL;
1055         }
1056 }
1057
1058 static void cm_destroy_id(struct ib_cm_id *cm_id, int err)
1059 {
1060         struct cm_id_private *cm_id_priv;
1061         struct cm_work *work;
1062
1063         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1064         spin_lock_irq(&cm_id_priv->lock);
1065 retest:
1066         switch (cm_id->state) {
1067         case IB_CM_LISTEN:
1068                 spin_lock(&cm.lock);
1069                 if (--cm_id_priv->listen_sharecount > 0) {
1070                         /* The id is still shared. */
1071                         WARN_ON(refcount_read(&cm_id_priv->refcount) == 1);
1072                         spin_unlock(&cm.lock);
1073                         spin_unlock_irq(&cm_id_priv->lock);
1074                         cm_deref_id(cm_id_priv);
1075                         return;
1076                 }
1077                 cm_id->state = IB_CM_IDLE;
1078                 rb_erase(&cm_id_priv->service_node, &cm.listen_service_table);
1079                 RB_CLEAR_NODE(&cm_id_priv->service_node);
1080                 spin_unlock(&cm.lock);
1081                 break;
1082         case IB_CM_SIDR_REQ_SENT:
1083                 cm_id->state = IB_CM_IDLE;
1084                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1085                 break;
1086         case IB_CM_SIDR_REQ_RCVD:
1087                 cm_send_sidr_rep_locked(cm_id_priv,
1088                                         &(struct ib_cm_sidr_rep_param){
1089                                                 .status = IB_SIDR_REJECT });
1090                 /* cm_send_sidr_rep_locked will not move to IDLE if it fails */
1091                 cm_id->state = IB_CM_IDLE;
1092                 break;
1093         case IB_CM_REQ_SENT:
1094         case IB_CM_MRA_REQ_RCVD:
1095                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1096                 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_TIMEOUT,
1097                                    &cm_id_priv->id.device->node_guid,
1098                                    sizeof(cm_id_priv->id.device->node_guid),
1099                                    NULL, 0);
1100                 break;
1101         case IB_CM_REQ_RCVD:
1102                 if (err == -ENOMEM) {
1103                         /* Do not reject to allow future retries. */
1104                         cm_reset_to_idle(cm_id_priv);
1105                 } else {
1106                         cm_send_rej_locked(cm_id_priv,
1107                                            IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
1108                                            NULL, 0);
1109                 }
1110                 break;
1111         case IB_CM_REP_SENT:
1112         case IB_CM_MRA_REP_RCVD:
1113                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1114                 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL,
1115                                    0, NULL, 0);
1116                 goto retest;
1117         case IB_CM_MRA_REQ_SENT:
1118         case IB_CM_REP_RCVD:
1119         case IB_CM_MRA_REP_SENT:
1120                 cm_send_rej_locked(cm_id_priv, IB_CM_REJ_CONSUMER_DEFINED, NULL,
1121                                    0, NULL, 0);
1122                 break;
1123         case IB_CM_ESTABLISHED:
1124                 if (cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
1125                         cm_id->state = IB_CM_IDLE;
1126                         break;
1127                 }
1128                 cm_send_dreq_locked(cm_id_priv, NULL, 0);
1129                 goto retest;
1130         case IB_CM_DREQ_SENT:
1131                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
1132                 cm_enter_timewait(cm_id_priv);
1133                 goto retest;
1134         case IB_CM_DREQ_RCVD:
1135                 cm_send_drep_locked(cm_id_priv, NULL, 0);
1136                 WARN_ON(cm_id->state != IB_CM_TIMEWAIT);
1137                 goto retest;
1138         case IB_CM_TIMEWAIT:
1139                 /*
1140                  * The cm_acquire_id in cm_timewait_handler will stop working
1141                  * once we do xa_erase below, so just move to idle here for
1142                  * consistency.
1143                  */
1144                 cm_id->state = IB_CM_IDLE;
1145                 break;
1146         case IB_CM_IDLE:
1147                 break;
1148         }
1149         WARN_ON(cm_id->state != IB_CM_IDLE);
1150
1151         spin_lock(&cm.lock);
1152         /* Required for cleanup paths related cm_req_handler() */
1153         if (cm_id_priv->timewait_info) {
1154                 cm_remove_remote(cm_id_priv);
1155                 kfree(cm_id_priv->timewait_info);
1156                 cm_id_priv->timewait_info = NULL;
1157         }
1158         if (!list_empty(&cm_id_priv->altr_list) &&
1159             (!cm_id_priv->altr_send_port_not_ready))
1160                 list_del(&cm_id_priv->altr_list);
1161         if (!list_empty(&cm_id_priv->prim_list) &&
1162             (!cm_id_priv->prim_send_port_not_ready))
1163                 list_del(&cm_id_priv->prim_list);
1164         WARN_ON(cm_id_priv->listen_sharecount);
1165         WARN_ON(!RB_EMPTY_NODE(&cm_id_priv->service_node));
1166         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node))
1167                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
1168         spin_unlock(&cm.lock);
1169         spin_unlock_irq(&cm_id_priv->lock);
1170
1171         xa_erase_irq(&cm.local_id_table, cm_local_id(cm_id->local_id));
1172         cm_deref_id(cm_id_priv);
1173         wait_for_completion(&cm_id_priv->comp);
1174         while ((work = cm_dequeue_work(cm_id_priv)) != NULL)
1175                 cm_free_work(work);
1176
1177         rdma_destroy_ah_attr(&cm_id_priv->av.ah_attr);
1178         rdma_destroy_ah_attr(&cm_id_priv->alt_av.ah_attr);
1179         kfree(cm_id_priv->private_data);
1180         kfree_rcu(cm_id_priv, rcu);
1181 }
1182
1183 void ib_destroy_cm_id(struct ib_cm_id *cm_id)
1184 {
1185         cm_destroy_id(cm_id, 0);
1186 }
1187 EXPORT_SYMBOL(ib_destroy_cm_id);
1188
1189 static int cm_init_listen(struct cm_id_private *cm_id_priv, __be64 service_id,
1190                           __be64 service_mask)
1191 {
1192         service_mask = service_mask ? service_mask : ~cpu_to_be64(0);
1193         service_id &= service_mask;
1194         if ((service_id & IB_SERVICE_ID_AGN_MASK) == IB_CM_ASSIGN_SERVICE_ID &&
1195             (service_id != IB_CM_ASSIGN_SERVICE_ID))
1196                 return -EINVAL;
1197
1198         if (service_id == IB_CM_ASSIGN_SERVICE_ID) {
1199                 cm_id_priv->id.service_id = cpu_to_be64(cm.listen_service_id++);
1200                 cm_id_priv->id.service_mask = ~cpu_to_be64(0);
1201         } else {
1202                 cm_id_priv->id.service_id = service_id;
1203                 cm_id_priv->id.service_mask = service_mask;
1204         }
1205         return 0;
1206 }
1207
1208 /**
1209  * ib_cm_listen - Initiates listening on the specified service ID for
1210  *   connection and service ID resolution requests.
1211  * @cm_id: Connection identifier associated with the listen request.
1212  * @service_id: Service identifier matched against incoming connection
1213  *   and service ID resolution requests.  The service ID should be specified
1214  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1215  *   assign a service ID to the caller.
1216  * @service_mask: Mask applied to service ID used to listen across a
1217  *   range of service IDs.  If set to 0, the service ID is matched
1218  *   exactly.  This parameter is ignored if %service_id is set to
1219  *   IB_CM_ASSIGN_SERVICE_ID.
1220  */
1221 int ib_cm_listen(struct ib_cm_id *cm_id, __be64 service_id, __be64 service_mask)
1222 {
1223         struct cm_id_private *cm_id_priv =
1224                 container_of(cm_id, struct cm_id_private, id);
1225         unsigned long flags;
1226         int ret;
1227
1228         spin_lock_irqsave(&cm_id_priv->lock, flags);
1229         if (cm_id_priv->id.state != IB_CM_IDLE) {
1230                 ret = -EINVAL;
1231                 goto out;
1232         }
1233
1234         ret = cm_init_listen(cm_id_priv, service_id, service_mask);
1235         if (ret)
1236                 goto out;
1237
1238         if (!cm_insert_listen(cm_id_priv, NULL)) {
1239                 ret = -EBUSY;
1240                 goto out;
1241         }
1242
1243         cm_id_priv->id.state = IB_CM_LISTEN;
1244         ret = 0;
1245
1246 out:
1247         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1248         return ret;
1249 }
1250 EXPORT_SYMBOL(ib_cm_listen);
1251
1252 /**
1253  * Create a new listening ib_cm_id and listen on the given service ID.
1254  *
1255  * If there's an existing ID listening on that same device and service ID,
1256  * return it.
1257  *
1258  * @device: Device associated with the cm_id.  All related communication will
1259  * be associated with the specified device.
1260  * @cm_handler: Callback invoked to notify the user of CM events.
1261  * @service_id: Service identifier matched against incoming connection
1262  *   and service ID resolution requests.  The service ID should be specified
1263  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1264  *   assign a service ID to the caller.
1265  *
1266  * Callers should call ib_destroy_cm_id when done with the listener ID.
1267  */
1268 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1269                                      ib_cm_handler cm_handler,
1270                                      __be64 service_id)
1271 {
1272         struct cm_id_private *listen_id_priv;
1273         struct cm_id_private *cm_id_priv;
1274         int err = 0;
1275
1276         /* Create an ID in advance, since the creation may sleep */
1277         cm_id_priv = cm_alloc_id_priv(device, cm_handler, NULL);
1278         if (IS_ERR(cm_id_priv))
1279                 return ERR_CAST(cm_id_priv);
1280
1281         err = cm_init_listen(cm_id_priv, service_id, 0);
1282         if (err)
1283                 return ERR_PTR(err);
1284
1285         spin_lock_irq(&cm_id_priv->lock);
1286         listen_id_priv = cm_insert_listen(cm_id_priv, cm_handler);
1287         if (listen_id_priv != cm_id_priv) {
1288                 spin_unlock_irq(&cm_id_priv->lock);
1289                 ib_destroy_cm_id(&cm_id_priv->id);
1290                 if (!listen_id_priv)
1291                         return ERR_PTR(-EINVAL);
1292                 return &listen_id_priv->id;
1293         }
1294         cm_id_priv->id.state = IB_CM_LISTEN;
1295         spin_unlock_irq(&cm_id_priv->lock);
1296
1297         /*
1298          * A listen ID does not need to be in the xarray since it does not
1299          * receive mads, is not placed in the remote_id or remote_qpn rbtree,
1300          * and does not enter timewait.
1301          */
1302
1303         return &cm_id_priv->id;
1304 }
1305 EXPORT_SYMBOL(ib_cm_insert_listen);
1306
1307 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv)
1308 {
1309         u64 hi_tid, low_tid;
1310
1311         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1312         low_tid  = (u64)cm_id_priv->id.local_id;
1313         return cpu_to_be64(hi_tid | low_tid);
1314 }
1315
1316 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1317                               __be16 attr_id, __be64 tid)
1318 {
1319         hdr->base_version  = IB_MGMT_BASE_VERSION;
1320         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1321         hdr->class_version = IB_CM_CLASS_VERSION;
1322         hdr->method        = IB_MGMT_METHOD_SEND;
1323         hdr->attr_id       = attr_id;
1324         hdr->tid           = tid;
1325 }
1326
1327 static void cm_format_mad_ece_hdr(struct ib_mad_hdr *hdr, __be16 attr_id,
1328                                   __be64 tid, u32 attr_mod)
1329 {
1330         cm_format_mad_hdr(hdr, attr_id, tid);
1331         hdr->attr_mod = cpu_to_be32(attr_mod);
1332 }
1333
1334 static void cm_format_req(struct cm_req_msg *req_msg,
1335                           struct cm_id_private *cm_id_priv,
1336                           struct ib_cm_req_param *param)
1337 {
1338         struct sa_path_rec *pri_path = param->primary_path;
1339         struct sa_path_rec *alt_path = param->alternate_path;
1340         bool pri_ext = false;
1341
1342         if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
1343                 pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
1344                                               pri_path->opa.slid);
1345
1346         cm_format_mad_ece_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1347                               cm_form_tid(cm_id_priv), param->ece.attr_mod);
1348
1349         IBA_SET(CM_REQ_LOCAL_COMM_ID, req_msg,
1350                 be32_to_cpu(cm_id_priv->id.local_id));
1351         IBA_SET(CM_REQ_SERVICE_ID, req_msg, be64_to_cpu(param->service_id));
1352         IBA_SET(CM_REQ_LOCAL_CA_GUID, req_msg,
1353                 be64_to_cpu(cm_id_priv->id.device->node_guid));
1354         IBA_SET(CM_REQ_LOCAL_QPN, req_msg, param->qp_num);
1355         IBA_SET(CM_REQ_INITIATOR_DEPTH, req_msg, param->initiator_depth);
1356         IBA_SET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg,
1357                 param->remote_cm_response_timeout);
1358         cm_req_set_qp_type(req_msg, param->qp_type);
1359         IBA_SET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg, param->flow_control);
1360         IBA_SET(CM_REQ_STARTING_PSN, req_msg, param->starting_psn);
1361         IBA_SET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg,
1362                 param->local_cm_response_timeout);
1363         IBA_SET(CM_REQ_PARTITION_KEY, req_msg,
1364                 be16_to_cpu(param->primary_path->pkey));
1365         IBA_SET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg,
1366                 param->primary_path->mtu);
1367         IBA_SET(CM_REQ_MAX_CM_RETRIES, req_msg, param->max_cm_retries);
1368
1369         if (param->qp_type != IB_QPT_XRC_INI) {
1370                 IBA_SET(CM_REQ_RESPONDER_RESOURCES, req_msg,
1371                         param->responder_resources);
1372                 IBA_SET(CM_REQ_RETRY_COUNT, req_msg, param->retry_count);
1373                 IBA_SET(CM_REQ_RNR_RETRY_COUNT, req_msg,
1374                         param->rnr_retry_count);
1375                 IBA_SET(CM_REQ_SRQ, req_msg, param->srq);
1376         }
1377
1378         *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) =
1379                 pri_path->sgid;
1380         *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) =
1381                 pri_path->dgid;
1382         if (pri_ext) {
1383                 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg)
1384                         ->global.interface_id =
1385                         OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid));
1386                 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg)
1387                         ->global.interface_id =
1388                         OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid));
1389         }
1390         if (pri_path->hop_limit <= 1) {
1391                 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1392                         be16_to_cpu(pri_ext ? 0 :
1393                                               htons(ntohl(sa_path_get_slid(
1394                                                       pri_path)))));
1395                 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1396                         be16_to_cpu(pri_ext ? 0 :
1397                                               htons(ntohl(sa_path_get_dlid(
1398                                                       pri_path)))));
1399         } else {
1400                 /* Work-around until there's a way to obtain remote LID info */
1401                 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1402                         be16_to_cpu(IB_LID_PERMISSIVE));
1403                 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1404                         be16_to_cpu(IB_LID_PERMISSIVE));
1405         }
1406         IBA_SET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg,
1407                 be32_to_cpu(pri_path->flow_label));
1408         IBA_SET(CM_REQ_PRIMARY_PACKET_RATE, req_msg, pri_path->rate);
1409         IBA_SET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg, pri_path->traffic_class);
1410         IBA_SET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg, pri_path->hop_limit);
1411         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, pri_path->sl);
1412         IBA_SET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg,
1413                 (pri_path->hop_limit <= 1));
1414         IBA_SET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg,
1415                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1416                                pri_path->packet_life_time));
1417
1418         if (alt_path) {
1419                 bool alt_ext = false;
1420
1421                 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA)
1422                         alt_ext = opa_is_extended_lid(alt_path->opa.dlid,
1423                                                       alt_path->opa.slid);
1424
1425                 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg) =
1426                         alt_path->sgid;
1427                 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg) =
1428                         alt_path->dgid;
1429                 if (alt_ext) {
1430                         IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1431                                         req_msg)
1432                                 ->global.interface_id =
1433                                 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid));
1434                         IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID,
1435                                         req_msg)
1436                                 ->global.interface_id =
1437                                 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid));
1438                 }
1439                 if (alt_path->hop_limit <= 1) {
1440                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1441                                 be16_to_cpu(
1442                                         alt_ext ? 0 :
1443                                                   htons(ntohl(sa_path_get_slid(
1444                                                           alt_path)))));
1445                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1446                                 be16_to_cpu(
1447                                         alt_ext ? 0 :
1448                                                   htons(ntohl(sa_path_get_dlid(
1449                                                           alt_path)))));
1450                 } else {
1451                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1452                                 be16_to_cpu(IB_LID_PERMISSIVE));
1453                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1454                                 be16_to_cpu(IB_LID_PERMISSIVE));
1455                 }
1456                 IBA_SET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg,
1457                         be32_to_cpu(alt_path->flow_label));
1458                 IBA_SET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg, alt_path->rate);
1459                 IBA_SET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg,
1460                         alt_path->traffic_class);
1461                 IBA_SET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg,
1462                         alt_path->hop_limit);
1463                 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, alt_path->sl);
1464                 IBA_SET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg,
1465                         (alt_path->hop_limit <= 1));
1466                 IBA_SET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg,
1467                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1468                                        alt_path->packet_life_time));
1469         }
1470         IBA_SET(CM_REQ_VENDOR_ID, req_msg, param->ece.vendor_id);
1471
1472         if (param->private_data && param->private_data_len)
1473                 IBA_SET_MEM(CM_REQ_PRIVATE_DATA, req_msg, param->private_data,
1474                             param->private_data_len);
1475 }
1476
1477 static int cm_validate_req_param(struct ib_cm_req_param *param)
1478 {
1479         if (!param->primary_path)
1480                 return -EINVAL;
1481
1482         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1483             param->qp_type != IB_QPT_XRC_INI)
1484                 return -EINVAL;
1485
1486         if (param->private_data &&
1487             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1488                 return -EINVAL;
1489
1490         if (param->alternate_path &&
1491             (param->alternate_path->pkey != param->primary_path->pkey ||
1492              param->alternate_path->mtu != param->primary_path->mtu))
1493                 return -EINVAL;
1494
1495         return 0;
1496 }
1497
1498 int ib_send_cm_req(struct ib_cm_id *cm_id,
1499                    struct ib_cm_req_param *param)
1500 {
1501         struct cm_id_private *cm_id_priv;
1502         struct cm_req_msg *req_msg;
1503         unsigned long flags;
1504         int ret;
1505
1506         ret = cm_validate_req_param(param);
1507         if (ret)
1508                 return ret;
1509
1510         /* Verify that we're not in timewait. */
1511         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1512         spin_lock_irqsave(&cm_id_priv->lock, flags);
1513         if (cm_id->state != IB_CM_IDLE || WARN_ON(cm_id_priv->timewait_info)) {
1514                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1515                 ret = -EINVAL;
1516                 goto out;
1517         }
1518         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1519
1520         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1521                                                             id.local_id);
1522         if (IS_ERR(cm_id_priv->timewait_info)) {
1523                 ret = PTR_ERR(cm_id_priv->timewait_info);
1524                 goto out;
1525         }
1526
1527         ret = cm_init_av_by_path(param->primary_path,
1528                                  param->ppath_sgid_attr, &cm_id_priv->av,
1529                                  cm_id_priv);
1530         if (ret)
1531                 goto out;
1532         if (param->alternate_path) {
1533                 ret = cm_init_av_by_path(param->alternate_path, NULL,
1534                                          &cm_id_priv->alt_av, cm_id_priv);
1535                 if (ret)
1536                         goto out;
1537         }
1538         cm_id->service_id = param->service_id;
1539         cm_id->service_mask = ~cpu_to_be64(0);
1540         cm_id_priv->timeout_ms = cm_convert_to_ms(
1541                                     param->primary_path->packet_life_time) * 2 +
1542                                  cm_convert_to_ms(
1543                                     param->remote_cm_response_timeout);
1544         cm_id_priv->max_cm_retries = param->max_cm_retries;
1545         cm_id_priv->initiator_depth = param->initiator_depth;
1546         cm_id_priv->responder_resources = param->responder_resources;
1547         cm_id_priv->retry_count = param->retry_count;
1548         cm_id_priv->path_mtu = param->primary_path->mtu;
1549         cm_id_priv->pkey = param->primary_path->pkey;
1550         cm_id_priv->qp_type = param->qp_type;
1551
1552         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1553         if (ret)
1554                 goto out;
1555
1556         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1557         cm_format_req(req_msg, cm_id_priv, param);
1558         cm_id_priv->tid = req_msg->hdr.tid;
1559         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1560         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1561
1562         cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
1563         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
1564
1565         spin_lock_irqsave(&cm_id_priv->lock, flags);
1566         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1567         if (ret) {
1568                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1569                 goto error2;
1570         }
1571         BUG_ON(cm_id->state != IB_CM_IDLE);
1572         cm_id->state = IB_CM_REQ_SENT;
1573         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1574         return 0;
1575
1576 error2: cm_free_msg(cm_id_priv->msg);
1577 out:    return ret;
1578 }
1579 EXPORT_SYMBOL(ib_send_cm_req);
1580
1581 static int cm_issue_rej(struct cm_port *port,
1582                         struct ib_mad_recv_wc *mad_recv_wc,
1583                         enum ib_cm_rej_reason reason,
1584                         enum cm_msg_response msg_rejected,
1585                         void *ari, u8 ari_length)
1586 {
1587         struct ib_mad_send_buf *msg = NULL;
1588         struct cm_rej_msg *rej_msg, *rcv_msg;
1589         int ret;
1590
1591         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1592         if (ret)
1593                 return ret;
1594
1595         /* We just need common CM header information.  Cast to any message. */
1596         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1597         rej_msg = (struct cm_rej_msg *) msg->mad;
1598
1599         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1600         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1601                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg));
1602         IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1603                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1604         IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected);
1605         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1606
1607         if (ari && ari_length) {
1608                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1609                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1610         }
1611
1612         ret = ib_post_send_mad(msg, NULL);
1613         if (ret)
1614                 cm_free_msg(msg);
1615
1616         return ret;
1617 }
1618
1619 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1620 {
1621         return ((cpu_to_be16(
1622                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) ||
1623                 (ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1624                                                req_msg))));
1625 }
1626
1627 static void cm_path_set_rec_type(struct ib_device *ib_device, u8 port_num,
1628                                  struct sa_path_rec *path, union ib_gid *gid)
1629 {
1630         if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1631                 path->rec_type = SA_PATH_REC_TYPE_OPA;
1632         else
1633                 path->rec_type = SA_PATH_REC_TYPE_IB;
1634 }
1635
1636 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1637                                         struct sa_path_rec *primary_path,
1638                                         struct sa_path_rec *alt_path)
1639 {
1640         u32 lid;
1641
1642         if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1643                 sa_path_set_dlid(primary_path,
1644                                  IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
1645                                          req_msg));
1646                 sa_path_set_slid(primary_path,
1647                                  IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
1648                                          req_msg));
1649         } else {
1650                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1651                         CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg));
1652                 sa_path_set_dlid(primary_path, lid);
1653
1654                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1655                         CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg));
1656                 sa_path_set_slid(primary_path, lid);
1657         }
1658
1659         if (!cm_req_has_alt_path(req_msg))
1660                 return;
1661
1662         if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1663                 sa_path_set_dlid(alt_path,
1664                                  IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
1665                                          req_msg));
1666                 sa_path_set_slid(alt_path,
1667                                  IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
1668                                          req_msg));
1669         } else {
1670                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1671                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg));
1672                 sa_path_set_dlid(alt_path, lid);
1673
1674                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1675                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg));
1676                 sa_path_set_slid(alt_path, lid);
1677         }
1678 }
1679
1680 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1681                                      struct sa_path_rec *primary_path,
1682                                      struct sa_path_rec *alt_path)
1683 {
1684         primary_path->dgid =
1685                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg);
1686         primary_path->sgid =
1687                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg);
1688         primary_path->flow_label =
1689                 cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg));
1690         primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg);
1691         primary_path->traffic_class =
1692                 IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg);
1693         primary_path->reversible = 1;
1694         primary_path->pkey =
1695                 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1696         primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg);
1697         primary_path->mtu_selector = IB_SA_EQ;
1698         primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1699         primary_path->rate_selector = IB_SA_EQ;
1700         primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg);
1701         primary_path->packet_life_time_selector = IB_SA_EQ;
1702         primary_path->packet_life_time =
1703                 IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg);
1704         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1705         primary_path->service_id =
1706                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1707         if (sa_path_is_roce(primary_path))
1708                 primary_path->roce.route_resolved = false;
1709
1710         if (cm_req_has_alt_path(req_msg)) {
1711                 alt_path->dgid = *IBA_GET_MEM_PTR(
1712                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg);
1713                 alt_path->sgid = *IBA_GET_MEM_PTR(
1714                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg);
1715                 alt_path->flow_label = cpu_to_be32(
1716                         IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg));
1717                 alt_path->hop_limit =
1718                         IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg);
1719                 alt_path->traffic_class =
1720                         IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg);
1721                 alt_path->reversible = 1;
1722                 alt_path->pkey =
1723                         cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1724                 alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg);
1725                 alt_path->mtu_selector = IB_SA_EQ;
1726                 alt_path->mtu =
1727                         IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1728                 alt_path->rate_selector = IB_SA_EQ;
1729                 alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg);
1730                 alt_path->packet_life_time_selector = IB_SA_EQ;
1731                 alt_path->packet_life_time =
1732                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg);
1733                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1734                 alt_path->service_id =
1735                         cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1736
1737                 if (sa_path_is_roce(alt_path))
1738                         alt_path->roce.route_resolved = false;
1739         }
1740         cm_format_path_lid_from_req(req_msg, primary_path, alt_path);
1741 }
1742
1743 static u16 cm_get_bth_pkey(struct cm_work *work)
1744 {
1745         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1746         u8 port_num = work->port->port_num;
1747         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1748         u16 pkey;
1749         int ret;
1750
1751         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1752         if (ret) {
1753                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1754                                      port_num, pkey_index, ret);
1755                 return 0;
1756         }
1757
1758         return pkey;
1759 }
1760
1761 /**
1762  * Convert OPA SGID to IB SGID
1763  * ULPs (such as IPoIB) do not understand OPA GIDs and will
1764  * reject them as the local_gid will not match the sgid. Therefore,
1765  * change the pathrec's SGID to an IB SGID.
1766  *
1767  * @work: Work completion
1768  * @path: Path record
1769  */
1770 static void cm_opa_to_ib_sgid(struct cm_work *work,
1771                               struct sa_path_rec *path)
1772 {
1773         struct ib_device *dev = work->port->cm_dev->ib_device;
1774         u8 port_num = work->port->port_num;
1775
1776         if (rdma_cap_opa_ah(dev, port_num) &&
1777             (ib_is_opa_gid(&path->sgid))) {
1778                 union ib_gid sgid;
1779
1780                 if (rdma_query_gid(dev, port_num, 0, &sgid)) {
1781                         dev_warn(&dev->dev,
1782                                  "Error updating sgid in CM request\n");
1783                         return;
1784                 }
1785
1786                 path->sgid = sgid;
1787         }
1788 }
1789
1790 static void cm_format_req_event(struct cm_work *work,
1791                                 struct cm_id_private *cm_id_priv,
1792                                 struct ib_cm_id *listen_id)
1793 {
1794         struct cm_req_msg *req_msg;
1795         struct ib_cm_req_event_param *param;
1796
1797         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1798         param = &work->cm_event.param.req_rcvd;
1799         param->listen_id = listen_id;
1800         param->bth_pkey = cm_get_bth_pkey(work);
1801         param->port = cm_id_priv->av.port->port_num;
1802         param->primary_path = &work->path[0];
1803         cm_opa_to_ib_sgid(work, param->primary_path);
1804         if (cm_req_has_alt_path(req_msg)) {
1805                 param->alternate_path = &work->path[1];
1806                 cm_opa_to_ib_sgid(work, param->alternate_path);
1807         } else {
1808                 param->alternate_path = NULL;
1809         }
1810         param->remote_ca_guid =
1811                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
1812         param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg);
1813         param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg);
1814         param->qp_type = cm_req_get_qp_type(req_msg);
1815         param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg);
1816         param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
1817         param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
1818         param->local_cm_response_timeout =
1819                 IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg);
1820         param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg);
1821         param->remote_cm_response_timeout =
1822                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg);
1823         param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
1824         param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
1825         param->srq = IBA_GET(CM_REQ_SRQ, req_msg);
1826         param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
1827         param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg);
1828         param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod);
1829
1830         work->cm_event.private_data =
1831                 IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg);
1832 }
1833
1834 static void cm_process_work(struct cm_id_private *cm_id_priv,
1835                             struct cm_work *work)
1836 {
1837         int ret;
1838
1839         /* We will typically only have the current event to report. */
1840         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1841         cm_free_work(work);
1842
1843         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1844                 spin_lock_irq(&cm_id_priv->lock);
1845                 work = cm_dequeue_work(cm_id_priv);
1846                 spin_unlock_irq(&cm_id_priv->lock);
1847                 if (!work)
1848                         return;
1849
1850                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1851                                                 &work->cm_event);
1852                 cm_free_work(work);
1853         }
1854         cm_deref_id(cm_id_priv);
1855         if (ret)
1856                 cm_destroy_id(&cm_id_priv->id, ret);
1857 }
1858
1859 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1860                           struct cm_id_private *cm_id_priv,
1861                           enum cm_msg_response msg_mraed, u8 service_timeout,
1862                           const void *private_data, u8 private_data_len)
1863 {
1864         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1865         IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed);
1866         IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg,
1867                 be32_to_cpu(cm_id_priv->id.local_id));
1868         IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg,
1869                 be32_to_cpu(cm_id_priv->id.remote_id));
1870         IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, service_timeout);
1871
1872         if (private_data && private_data_len)
1873                 IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data,
1874                             private_data_len);
1875 }
1876
1877 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1878                           struct cm_id_private *cm_id_priv,
1879                           enum ib_cm_rej_reason reason, void *ari,
1880                           u8 ari_length, const void *private_data,
1881                           u8 private_data_len, enum ib_cm_state state)
1882 {
1883         lockdep_assert_held(&cm_id_priv->lock);
1884
1885         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1886         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1887                 be32_to_cpu(cm_id_priv->id.remote_id));
1888
1889         switch (state) {
1890         case IB_CM_REQ_RCVD:
1891                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0));
1892                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1893                 break;
1894         case IB_CM_MRA_REQ_SENT:
1895                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1896                         be32_to_cpu(cm_id_priv->id.local_id));
1897                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1898                 break;
1899         case IB_CM_REP_RCVD:
1900         case IB_CM_MRA_REP_SENT:
1901                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1902                         be32_to_cpu(cm_id_priv->id.local_id));
1903                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP);
1904                 break;
1905         default:
1906                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1907                         be32_to_cpu(cm_id_priv->id.local_id));
1908                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg,
1909                         CM_MSG_RESPONSE_OTHER);
1910                 break;
1911         }
1912
1913         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1914         if (ari && ari_length) {
1915                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1916                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1917         }
1918
1919         if (private_data && private_data_len)
1920                 IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data,
1921                             private_data_len);
1922 }
1923
1924 static void cm_dup_req_handler(struct cm_work *work,
1925                                struct cm_id_private *cm_id_priv)
1926 {
1927         struct ib_mad_send_buf *msg = NULL;
1928         int ret;
1929
1930         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1931                         counter[CM_REQ_COUNTER]);
1932
1933         /* Quick state check to discard duplicate REQs. */
1934         spin_lock_irq(&cm_id_priv->lock);
1935         if (cm_id_priv->id.state == IB_CM_REQ_RCVD) {
1936                 spin_unlock_irq(&cm_id_priv->lock);
1937                 return;
1938         }
1939         spin_unlock_irq(&cm_id_priv->lock);
1940
1941         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1942         if (ret)
1943                 return;
1944
1945         spin_lock_irq(&cm_id_priv->lock);
1946         switch (cm_id_priv->id.state) {
1947         case IB_CM_MRA_REQ_SENT:
1948                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1949                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1950                               cm_id_priv->private_data,
1951                               cm_id_priv->private_data_len);
1952                 break;
1953         case IB_CM_TIMEWAIT:
1954                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv,
1955                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0,
1956                               IB_CM_TIMEWAIT);
1957                 break;
1958         default:
1959                 goto unlock;
1960         }
1961         spin_unlock_irq(&cm_id_priv->lock);
1962
1963         ret = ib_post_send_mad(msg, NULL);
1964         if (ret)
1965                 goto free;
1966         return;
1967
1968 unlock: spin_unlock_irq(&cm_id_priv->lock);
1969 free:   cm_free_msg(msg);
1970 }
1971
1972 static struct cm_id_private * cm_match_req(struct cm_work *work,
1973                                            struct cm_id_private *cm_id_priv)
1974 {
1975         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1976         struct cm_timewait_info *timewait_info;
1977         struct cm_req_msg *req_msg;
1978
1979         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1980
1981         /* Check for possible duplicate REQ. */
1982         spin_lock_irq(&cm.lock);
1983         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1984         if (timewait_info) {
1985                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
1986                                            timewait_info->work.remote_id);
1987                 spin_unlock_irq(&cm.lock);
1988                 if (cur_cm_id_priv) {
1989                         cm_dup_req_handler(work, cur_cm_id_priv);
1990                         cm_deref_id(cur_cm_id_priv);
1991                 }
1992                 return NULL;
1993         }
1994
1995         /* Check for stale connections. */
1996         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
1997         if (timewait_info) {
1998                 cm_remove_remote(cm_id_priv);
1999                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2000                                            timewait_info->work.remote_id);
2001
2002                 spin_unlock_irq(&cm.lock);
2003                 cm_issue_rej(work->port, work->mad_recv_wc,
2004                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
2005                              NULL, 0);
2006                 if (cur_cm_id_priv) {
2007                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2008                         cm_deref_id(cur_cm_id_priv);
2009                 }
2010                 return NULL;
2011         }
2012
2013         /* Find matching listen request. */
2014         listen_cm_id_priv = cm_find_listen(
2015                 cm_id_priv->id.device,
2016                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)));
2017         if (!listen_cm_id_priv) {
2018                 cm_remove_remote(cm_id_priv);
2019                 spin_unlock_irq(&cm.lock);
2020                 cm_issue_rej(work->port, work->mad_recv_wc,
2021                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
2022                              NULL, 0);
2023                 return NULL;
2024         }
2025         spin_unlock_irq(&cm.lock);
2026         return listen_cm_id_priv;
2027 }
2028
2029 /*
2030  * Work-around for inter-subnet connections.  If the LIDs are permissive,
2031  * we need to override the LID/SL data in the REQ with the LID information
2032  * in the work completion.
2033  */
2034 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
2035 {
2036         if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) {
2037                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
2038                                         req_msg)) == IB_LID_PERMISSIVE) {
2039                         IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
2040                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2041                         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl);
2042                 }
2043
2044                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
2045                                         req_msg)) == IB_LID_PERMISSIVE)
2046                         IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
2047                                 wc->dlid_path_bits);
2048         }
2049
2050         if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) {
2051                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
2052                                         req_msg)) == IB_LID_PERMISSIVE) {
2053                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
2054                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2055                         IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl);
2056                 }
2057
2058                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
2059                                         req_msg)) == IB_LID_PERMISSIVE)
2060                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
2061                                 wc->dlid_path_bits);
2062         }
2063 }
2064
2065 static int cm_req_handler(struct cm_work *work)
2066 {
2067         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
2068         struct cm_req_msg *req_msg;
2069         const struct ib_global_route *grh;
2070         const struct ib_gid_attr *gid_attr;
2071         int ret;
2072
2073         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
2074
2075         cm_id_priv =
2076                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
2077         if (IS_ERR(cm_id_priv))
2078                 return PTR_ERR(cm_id_priv);
2079
2080         cm_id_priv->id.remote_id =
2081                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg));
2082         cm_id_priv->id.service_id =
2083                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
2084         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
2085         cm_id_priv->tid = req_msg->hdr.tid;
2086         cm_id_priv->timeout_ms = cm_convert_to_ms(
2087                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg));
2088         cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg);
2089         cm_id_priv->remote_qpn =
2090                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
2091         cm_id_priv->initiator_depth =
2092                 IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
2093         cm_id_priv->responder_resources =
2094                 IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
2095         cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
2096         cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
2097         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
2098         cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
2099         cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
2100         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
2101
2102         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2103                                       work->mad_recv_wc->recv_buf.grh,
2104                                       &cm_id_priv->av);
2105         if (ret)
2106                 goto destroy;
2107         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
2108                                                             id.local_id);
2109         if (IS_ERR(cm_id_priv->timewait_info)) {
2110                 ret = PTR_ERR(cm_id_priv->timewait_info);
2111                 goto destroy;
2112         }
2113         cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id;
2114         cm_id_priv->timewait_info->remote_ca_guid =
2115                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
2116         cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn;
2117
2118         /*
2119          * Note that the ID pointer is not in the xarray at this point,
2120          * so this set is only visible to the local thread.
2121          */
2122         cm_id_priv->id.state = IB_CM_REQ_RCVD;
2123
2124         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
2125         if (!listen_cm_id_priv) {
2126                 pr_debug("%s: local_id %d, no listen_cm_id_priv\n", __func__,
2127                          be32_to_cpu(cm_id_priv->id.local_id));
2128                 cm_id_priv->id.state = IB_CM_IDLE;
2129                 ret = -EINVAL;
2130                 goto destroy;
2131         }
2132
2133         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
2134
2135         memset(&work->path[0], 0, sizeof(work->path[0]));
2136         if (cm_req_has_alt_path(req_msg))
2137                 memset(&work->path[1], 0, sizeof(work->path[1]));
2138         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
2139         gid_attr = grh->sgid_attr;
2140
2141         if (gid_attr &&
2142             rdma_protocol_roce(work->port->cm_dev->ib_device,
2143                                work->port->port_num)) {
2144                 work->path[0].rec_type =
2145                         sa_conv_gid_to_pathrec_type(gid_attr->gid_type);
2146         } else {
2147                 cm_path_set_rec_type(
2148                         work->port->cm_dev->ib_device, work->port->port_num,
2149                         &work->path[0],
2150                         IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID,
2151                                         req_msg));
2152         }
2153         if (cm_req_has_alt_path(req_msg))
2154                 work->path[1].rec_type = work->path[0].rec_type;
2155         cm_format_paths_from_req(req_msg, &work->path[0],
2156                                  &work->path[1]);
2157         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
2158                 sa_path_set_dmac(&work->path[0],
2159                                  cm_id_priv->av.ah_attr.roce.dmac);
2160         work->path[0].hop_limit = grh->hop_limit;
2161         ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av,
2162                                  cm_id_priv);
2163         if (ret) {
2164                 int err;
2165
2166                 err = rdma_query_gid(work->port->cm_dev->ib_device,
2167                                      work->port->port_num, 0,
2168                                      &work->path[0].sgid);
2169                 if (err)
2170                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2171                                        NULL, 0, NULL, 0);
2172                 else
2173                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2174                                        &work->path[0].sgid,
2175                                        sizeof(work->path[0].sgid),
2176                                        NULL, 0);
2177                 goto rejected;
2178         }
2179         if (cm_req_has_alt_path(req_msg)) {
2180                 ret = cm_init_av_by_path(&work->path[1], NULL,
2181                                          &cm_id_priv->alt_av, cm_id_priv);
2182                 if (ret) {
2183                         ib_send_cm_rej(&cm_id_priv->id,
2184                                        IB_CM_REJ_INVALID_ALT_GID,
2185                                        &work->path[0].sgid,
2186                                        sizeof(work->path[0].sgid), NULL, 0);
2187                         goto rejected;
2188                 }
2189         }
2190
2191         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
2192         cm_id_priv->id.context = listen_cm_id_priv->id.context;
2193         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
2194
2195         /* Now MAD handlers can see the new ID */
2196         spin_lock_irq(&cm_id_priv->lock);
2197         cm_finalize_id(cm_id_priv);
2198
2199         /* Refcount belongs to the event, pairs with cm_process_work() */
2200         refcount_inc(&cm_id_priv->refcount);
2201         cm_queue_work_unlock(cm_id_priv, work);
2202         /*
2203          * Since this ID was just created and was not made visible to other MAD
2204          * handlers until the cm_finalize_id() above we know that the
2205          * cm_process_work() will deliver the event and the listen_cm_id
2206          * embedded in the event can be derefed here.
2207          */
2208         cm_deref_id(listen_cm_id_priv);
2209         return 0;
2210
2211 rejected:
2212         cm_deref_id(listen_cm_id_priv);
2213 destroy:
2214         ib_destroy_cm_id(&cm_id_priv->id);
2215         return ret;
2216 }
2217
2218 static void cm_format_rep(struct cm_rep_msg *rep_msg,
2219                           struct cm_id_private *cm_id_priv,
2220                           struct ib_cm_rep_param *param)
2221 {
2222         cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid,
2223                               param->ece.attr_mod);
2224         IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg,
2225                 be32_to_cpu(cm_id_priv->id.local_id));
2226         IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg,
2227                 be32_to_cpu(cm_id_priv->id.remote_id));
2228         IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn);
2229         IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg,
2230                 param->responder_resources);
2231         IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg,
2232                 cm_id_priv->av.port->cm_dev->ack_delay);
2233         IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted);
2234         IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count);
2235         IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg,
2236                 be64_to_cpu(cm_id_priv->id.device->node_guid));
2237
2238         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
2239                 IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg,
2240                         param->initiator_depth);
2241                 IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg,
2242                         param->flow_control);
2243                 IBA_SET(CM_REP_SRQ, rep_msg, param->srq);
2244                 IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num);
2245         } else {
2246                 IBA_SET(CM_REP_SRQ, rep_msg, 1);
2247                 IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num);
2248         }
2249
2250         IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id);
2251         IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8);
2252         IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16);
2253
2254         if (param->private_data && param->private_data_len)
2255                 IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data,
2256                             param->private_data_len);
2257 }
2258
2259 int ib_send_cm_rep(struct ib_cm_id *cm_id,
2260                    struct ib_cm_rep_param *param)
2261 {
2262         struct cm_id_private *cm_id_priv;
2263         struct ib_mad_send_buf *msg;
2264         struct cm_rep_msg *rep_msg;
2265         unsigned long flags;
2266         int ret;
2267
2268         if (param->private_data &&
2269             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
2270                 return -EINVAL;
2271
2272         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2273         spin_lock_irqsave(&cm_id_priv->lock, flags);
2274         if (cm_id->state != IB_CM_REQ_RCVD &&
2275             cm_id->state != IB_CM_MRA_REQ_SENT) {
2276                 pr_debug("%s: local_comm_id %d, cm_id->state: %d\n", __func__,
2277                          be32_to_cpu(cm_id_priv->id.local_id), cm_id->state);
2278                 ret = -EINVAL;
2279                 goto out;
2280         }
2281
2282         ret = cm_alloc_msg(cm_id_priv, &msg);
2283         if (ret)
2284                 goto out;
2285
2286         rep_msg = (struct cm_rep_msg *) msg->mad;
2287         cm_format_rep(rep_msg, cm_id_priv, param);
2288         msg->timeout_ms = cm_id_priv->timeout_ms;
2289         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
2290
2291         ret = ib_post_send_mad(msg, NULL);
2292         if (ret) {
2293                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2294                 cm_free_msg(msg);
2295                 return ret;
2296         }
2297
2298         cm_id->state = IB_CM_REP_SENT;
2299         cm_id_priv->msg = msg;
2300         cm_id_priv->initiator_depth = param->initiator_depth;
2301         cm_id_priv->responder_resources = param->responder_resources;
2302         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2303         WARN_ONCE(param->qp_num & 0xFF000000,
2304                   "IBTA declares QPN to be 24 bits, but it is 0x%X\n",
2305                   param->qp_num);
2306         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2307
2308 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2309         return ret;
2310 }
2311 EXPORT_SYMBOL(ib_send_cm_rep);
2312
2313 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2314                           struct cm_id_private *cm_id_priv,
2315                           const void *private_data,
2316                           u8 private_data_len)
2317 {
2318         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2319         IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg,
2320                 be32_to_cpu(cm_id_priv->id.local_id));
2321         IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg,
2322                 be32_to_cpu(cm_id_priv->id.remote_id));
2323
2324         if (private_data && private_data_len)
2325                 IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data,
2326                             private_data_len);
2327 }
2328
2329 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2330                    const void *private_data,
2331                    u8 private_data_len)
2332 {
2333         struct cm_id_private *cm_id_priv;
2334         struct ib_mad_send_buf *msg;
2335         unsigned long flags;
2336         void *data;
2337         int ret;
2338
2339         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2340                 return -EINVAL;
2341
2342         data = cm_copy_private_data(private_data, private_data_len);
2343         if (IS_ERR(data))
2344                 return PTR_ERR(data);
2345
2346         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2347         spin_lock_irqsave(&cm_id_priv->lock, flags);
2348         if (cm_id->state != IB_CM_REP_RCVD &&
2349             cm_id->state != IB_CM_MRA_REP_SENT) {
2350                 pr_debug("%s: local_id %d, cm_id->state %d\n", __func__,
2351                          be32_to_cpu(cm_id->local_id), cm_id->state);
2352                 ret = -EINVAL;
2353                 goto error;
2354         }
2355
2356         ret = cm_alloc_msg(cm_id_priv, &msg);
2357         if (ret)
2358                 goto error;
2359
2360         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2361                       private_data, private_data_len);
2362
2363         ret = ib_post_send_mad(msg, NULL);
2364         if (ret) {
2365                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2366                 cm_free_msg(msg);
2367                 kfree(data);
2368                 return ret;
2369         }
2370
2371         cm_id->state = IB_CM_ESTABLISHED;
2372         cm_set_private_data(cm_id_priv, data, private_data_len);
2373         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2374         return 0;
2375
2376 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2377         kfree(data);
2378         return ret;
2379 }
2380 EXPORT_SYMBOL(ib_send_cm_rtu);
2381
2382 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2383 {
2384         struct cm_rep_msg *rep_msg;
2385         struct ib_cm_rep_event_param *param;
2386
2387         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2388         param = &work->cm_event.param.rep_rcvd;
2389         param->remote_ca_guid =
2390                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2391         param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg);
2392         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2393         param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg);
2394         param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2395         param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2396         param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2397         param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg);
2398         param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg);
2399         param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2400         param->srq = IBA_GET(CM_REP_SRQ, rep_msg);
2401         param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16;
2402         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8;
2403         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg);
2404         param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod);
2405
2406         work->cm_event.private_data =
2407                 IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg);
2408 }
2409
2410 static void cm_dup_rep_handler(struct cm_work *work)
2411 {
2412         struct cm_id_private *cm_id_priv;
2413         struct cm_rep_msg *rep_msg;
2414         struct ib_mad_send_buf *msg = NULL;
2415         int ret;
2416
2417         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2418         cm_id_priv = cm_acquire_id(
2419                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)),
2420                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)));
2421         if (!cm_id_priv)
2422                 return;
2423
2424         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2425                         counter[CM_REP_COUNTER]);
2426         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2427         if (ret)
2428                 goto deref;
2429
2430         spin_lock_irq(&cm_id_priv->lock);
2431         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2432                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2433                               cm_id_priv->private_data,
2434                               cm_id_priv->private_data_len);
2435         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2436                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2437                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2438                               cm_id_priv->private_data,
2439                               cm_id_priv->private_data_len);
2440         else
2441                 goto unlock;
2442         spin_unlock_irq(&cm_id_priv->lock);
2443
2444         ret = ib_post_send_mad(msg, NULL);
2445         if (ret)
2446                 goto free;
2447         goto deref;
2448
2449 unlock: spin_unlock_irq(&cm_id_priv->lock);
2450 free:   cm_free_msg(msg);
2451 deref:  cm_deref_id(cm_id_priv);
2452 }
2453
2454 static int cm_rep_handler(struct cm_work *work)
2455 {
2456         struct cm_id_private *cm_id_priv;
2457         struct cm_rep_msg *rep_msg;
2458         int ret;
2459         struct cm_id_private *cur_cm_id_priv;
2460         struct cm_timewait_info *timewait_info;
2461
2462         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2463         cm_id_priv = cm_acquire_id(
2464                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0);
2465         if (!cm_id_priv) {
2466                 cm_dup_rep_handler(work);
2467                 pr_debug("%s: remote_comm_id %d, no cm_id_priv\n", __func__,
2468                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2469                 return -EINVAL;
2470         }
2471
2472         cm_format_rep_event(work, cm_id_priv->qp_type);
2473
2474         spin_lock_irq(&cm_id_priv->lock);
2475         switch (cm_id_priv->id.state) {
2476         case IB_CM_REQ_SENT:
2477         case IB_CM_MRA_REQ_RCVD:
2478                 break;
2479         default:
2480                 ret = -EINVAL;
2481                 pr_debug(
2482                         "%s: cm_id_priv->id.state: %d, local_comm_id %d, remote_comm_id %d\n",
2483                         __func__, cm_id_priv->id.state,
2484                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2485                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2486                 spin_unlock_irq(&cm_id_priv->lock);
2487                 goto error;
2488         }
2489
2490         cm_id_priv->timewait_info->work.remote_id =
2491                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2492         cm_id_priv->timewait_info->remote_ca_guid =
2493                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2494         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2495
2496         spin_lock(&cm.lock);
2497         /* Check for duplicate REP. */
2498         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2499                 spin_unlock(&cm.lock);
2500                 spin_unlock_irq(&cm_id_priv->lock);
2501                 ret = -EINVAL;
2502                 pr_debug("%s: Failed to insert remote id %d\n", __func__,
2503                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2504                 goto error;
2505         }
2506         /* Check for a stale connection. */
2507         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2508         if (timewait_info) {
2509                 cm_remove_remote(cm_id_priv);
2510                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2511                                            timewait_info->work.remote_id);
2512
2513                 spin_unlock(&cm.lock);
2514                 spin_unlock_irq(&cm_id_priv->lock);
2515                 cm_issue_rej(work->port, work->mad_recv_wc,
2516                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2517                              NULL, 0);
2518                 ret = -EINVAL;
2519                 pr_debug(
2520                         "%s: Stale connection. local_comm_id %d, remote_comm_id %d\n",
2521                         __func__, IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2522                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2523
2524                 if (cur_cm_id_priv) {
2525                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2526                         cm_deref_id(cur_cm_id_priv);
2527                 }
2528
2529                 goto error;
2530         }
2531         spin_unlock(&cm.lock);
2532
2533         cm_id_priv->id.state = IB_CM_REP_RCVD;
2534         cm_id_priv->id.remote_id =
2535                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2536         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2537         cm_id_priv->initiator_depth =
2538                 IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2539         cm_id_priv->responder_resources =
2540                 IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2541         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2542         cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2543         cm_id_priv->target_ack_delay =
2544                 IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2545         cm_id_priv->av.timeout =
2546                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2547                                        cm_id_priv->av.timeout - 1);
2548         cm_id_priv->alt_av.timeout =
2549                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2550                                        cm_id_priv->alt_av.timeout - 1);
2551
2552         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2553         cm_queue_work_unlock(cm_id_priv, work);
2554         return 0;
2555
2556 error:
2557         cm_deref_id(cm_id_priv);
2558         return ret;
2559 }
2560
2561 static int cm_establish_handler(struct cm_work *work)
2562 {
2563         struct cm_id_private *cm_id_priv;
2564
2565         /* See comment in cm_establish about lookup. */
2566         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2567         if (!cm_id_priv)
2568                 return -EINVAL;
2569
2570         spin_lock_irq(&cm_id_priv->lock);
2571         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2572                 spin_unlock_irq(&cm_id_priv->lock);
2573                 goto out;
2574         }
2575
2576         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2577         cm_queue_work_unlock(cm_id_priv, work);
2578         return 0;
2579 out:
2580         cm_deref_id(cm_id_priv);
2581         return -EINVAL;
2582 }
2583
2584 static int cm_rtu_handler(struct cm_work *work)
2585 {
2586         struct cm_id_private *cm_id_priv;
2587         struct cm_rtu_msg *rtu_msg;
2588
2589         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2590         cm_id_priv = cm_acquire_id(
2591                 cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)),
2592                 cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg)));
2593         if (!cm_id_priv)
2594                 return -EINVAL;
2595
2596         work->cm_event.private_data =
2597                 IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg);
2598
2599         spin_lock_irq(&cm_id_priv->lock);
2600         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2601             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2602                 spin_unlock_irq(&cm_id_priv->lock);
2603                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2604                                 counter[CM_RTU_COUNTER]);
2605                 goto out;
2606         }
2607         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2608
2609         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2610         cm_queue_work_unlock(cm_id_priv, work);
2611         return 0;
2612 out:
2613         cm_deref_id(cm_id_priv);
2614         return -EINVAL;
2615 }
2616
2617 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2618                           struct cm_id_private *cm_id_priv,
2619                           const void *private_data,
2620                           u8 private_data_len)
2621 {
2622         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2623                           cm_form_tid(cm_id_priv));
2624         IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg,
2625                 be32_to_cpu(cm_id_priv->id.local_id));
2626         IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg,
2627                 be32_to_cpu(cm_id_priv->id.remote_id));
2628         IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg,
2629                 be32_to_cpu(cm_id_priv->remote_qpn));
2630
2631         if (private_data && private_data_len)
2632                 IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data,
2633                             private_data_len);
2634 }
2635
2636 static int cm_send_dreq_locked(struct cm_id_private *cm_id_priv,
2637                                const void *private_data, u8 private_data_len)
2638 {
2639         struct ib_mad_send_buf *msg;
2640         int ret;
2641
2642         lockdep_assert_held(&cm_id_priv->lock);
2643
2644         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2645                 return -EINVAL;
2646
2647         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2648                 pr_debug("%s: local_id %d, cm_id->state: %d\n", __func__,
2649                          be32_to_cpu(cm_id_priv->id.local_id),
2650                          cm_id_priv->id.state);
2651                 return -EINVAL;
2652         }
2653
2654         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2655             cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2656                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2657
2658         ret = cm_alloc_msg(cm_id_priv, &msg);
2659         if (ret) {
2660                 cm_enter_timewait(cm_id_priv);
2661                 return ret;
2662         }
2663
2664         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2665                        private_data, private_data_len);
2666         msg->timeout_ms = cm_id_priv->timeout_ms;
2667         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2668
2669         ret = ib_post_send_mad(msg, NULL);
2670         if (ret) {
2671                 cm_enter_timewait(cm_id_priv);
2672                 cm_free_msg(msg);
2673                 return ret;
2674         }
2675
2676         cm_id_priv->id.state = IB_CM_DREQ_SENT;
2677         cm_id_priv->msg = msg;
2678         return 0;
2679 }
2680
2681 int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data,
2682                     u8 private_data_len)
2683 {
2684         struct cm_id_private *cm_id_priv =
2685                 container_of(cm_id, struct cm_id_private, id);
2686         unsigned long flags;
2687         int ret;
2688
2689         spin_lock_irqsave(&cm_id_priv->lock, flags);
2690         ret = cm_send_dreq_locked(cm_id_priv, private_data, private_data_len);
2691         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2692         return ret;
2693 }
2694 EXPORT_SYMBOL(ib_send_cm_dreq);
2695
2696 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2697                           struct cm_id_private *cm_id_priv,
2698                           const void *private_data,
2699                           u8 private_data_len)
2700 {
2701         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2702         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2703                 be32_to_cpu(cm_id_priv->id.local_id));
2704         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2705                 be32_to_cpu(cm_id_priv->id.remote_id));
2706
2707         if (private_data && private_data_len)
2708                 IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data,
2709                             private_data_len);
2710 }
2711
2712 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
2713                                void *private_data, u8 private_data_len)
2714 {
2715         struct ib_mad_send_buf *msg;
2716         int ret;
2717
2718         lockdep_assert_held(&cm_id_priv->lock);
2719
2720         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2721                 return -EINVAL;
2722
2723         if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2724                 pr_debug(
2725                         "%s: local_id %d, cm_idcm_id->state(%d) != IB_CM_DREQ_RCVD\n",
2726                         __func__, be32_to_cpu(cm_id_priv->id.local_id),
2727                         cm_id_priv->id.state);
2728                 kfree(private_data);
2729                 return -EINVAL;
2730         }
2731
2732         cm_set_private_data(cm_id_priv, private_data, private_data_len);
2733         cm_enter_timewait(cm_id_priv);
2734
2735         ret = cm_alloc_msg(cm_id_priv, &msg);
2736         if (ret)
2737                 return ret;
2738
2739         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2740                        private_data, private_data_len);
2741
2742         ret = ib_post_send_mad(msg, NULL);
2743         if (ret) {
2744                 cm_free_msg(msg);
2745                 return ret;
2746         }
2747         return 0;
2748 }
2749
2750 int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data,
2751                     u8 private_data_len)
2752 {
2753         struct cm_id_private *cm_id_priv =
2754                 container_of(cm_id, struct cm_id_private, id);
2755         unsigned long flags;
2756         void *data;
2757         int ret;
2758
2759         data = cm_copy_private_data(private_data, private_data_len);
2760         if (IS_ERR(data))
2761                 return PTR_ERR(data);
2762
2763         spin_lock_irqsave(&cm_id_priv->lock, flags);
2764         ret = cm_send_drep_locked(cm_id_priv, data, private_data_len);
2765         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2766         return ret;
2767 }
2768 EXPORT_SYMBOL(ib_send_cm_drep);
2769
2770 static int cm_issue_drep(struct cm_port *port,
2771                          struct ib_mad_recv_wc *mad_recv_wc)
2772 {
2773         struct ib_mad_send_buf *msg = NULL;
2774         struct cm_dreq_msg *dreq_msg;
2775         struct cm_drep_msg *drep_msg;
2776         int ret;
2777
2778         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2779         if (ret)
2780                 return ret;
2781
2782         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2783         drep_msg = (struct cm_drep_msg *) msg->mad;
2784
2785         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2786         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2787                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg));
2788         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2789                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2790
2791         ret = ib_post_send_mad(msg, NULL);
2792         if (ret)
2793                 cm_free_msg(msg);
2794
2795         return ret;
2796 }
2797
2798 static int cm_dreq_handler(struct cm_work *work)
2799 {
2800         struct cm_id_private *cm_id_priv;
2801         struct cm_dreq_msg *dreq_msg;
2802         struct ib_mad_send_buf *msg = NULL;
2803
2804         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2805         cm_id_priv = cm_acquire_id(
2806                 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)),
2807                 cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)));
2808         if (!cm_id_priv) {
2809                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2810                                 counter[CM_DREQ_COUNTER]);
2811                 cm_issue_drep(work->port, work->mad_recv_wc);
2812                 pr_debug(
2813                         "%s: no cm_id_priv, local_comm_id %d, remote_comm_id %d\n",
2814                         __func__, IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2815                         IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2816                 return -EINVAL;
2817         }
2818
2819         work->cm_event.private_data =
2820                 IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg);
2821
2822         spin_lock_irq(&cm_id_priv->lock);
2823         if (cm_id_priv->local_qpn !=
2824             cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg)))
2825                 goto unlock;
2826
2827         switch (cm_id_priv->id.state) {
2828         case IB_CM_REP_SENT:
2829         case IB_CM_DREQ_SENT:
2830                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2831                 break;
2832         case IB_CM_ESTABLISHED:
2833                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2834                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2835                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2836                 break;
2837         case IB_CM_MRA_REP_RCVD:
2838                 break;
2839         case IB_CM_TIMEWAIT:
2840                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2841                                 counter[CM_DREQ_COUNTER]);
2842                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
2843                 if (IS_ERR(msg))
2844                         goto unlock;
2845
2846                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2847                                cm_id_priv->private_data,
2848                                cm_id_priv->private_data_len);
2849                 spin_unlock_irq(&cm_id_priv->lock);
2850
2851                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2852                     ib_post_send_mad(msg, NULL))
2853                         cm_free_msg(msg);
2854                 goto deref;
2855         case IB_CM_DREQ_RCVD:
2856                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2857                                 counter[CM_DREQ_COUNTER]);
2858                 goto unlock;
2859         default:
2860                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
2861                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
2862                          cm_id_priv->id.state);
2863                 goto unlock;
2864         }
2865         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2866         cm_id_priv->tid = dreq_msg->hdr.tid;
2867         cm_queue_work_unlock(cm_id_priv, work);
2868         return 0;
2869
2870 unlock: spin_unlock_irq(&cm_id_priv->lock);
2871 deref:  cm_deref_id(cm_id_priv);
2872         return -EINVAL;
2873 }
2874
2875 static int cm_drep_handler(struct cm_work *work)
2876 {
2877         struct cm_id_private *cm_id_priv;
2878         struct cm_drep_msg *drep_msg;
2879
2880         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2881         cm_id_priv = cm_acquire_id(
2882                 cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)),
2883                 cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg)));
2884         if (!cm_id_priv)
2885                 return -EINVAL;
2886
2887         work->cm_event.private_data =
2888                 IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg);
2889
2890         spin_lock_irq(&cm_id_priv->lock);
2891         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2892             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2893                 spin_unlock_irq(&cm_id_priv->lock);
2894                 goto out;
2895         }
2896         cm_enter_timewait(cm_id_priv);
2897
2898         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2899         cm_queue_work_unlock(cm_id_priv, work);
2900         return 0;
2901 out:
2902         cm_deref_id(cm_id_priv);
2903         return -EINVAL;
2904 }
2905
2906 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
2907                               enum ib_cm_rej_reason reason, void *ari,
2908                               u8 ari_length, const void *private_data,
2909                               u8 private_data_len)
2910 {
2911         enum ib_cm_state state = cm_id_priv->id.state;
2912         struct ib_mad_send_buf *msg;
2913         int ret;
2914
2915         lockdep_assert_held(&cm_id_priv->lock);
2916
2917         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2918             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2919                 return -EINVAL;
2920
2921         switch (state) {
2922         case IB_CM_REQ_SENT:
2923         case IB_CM_MRA_REQ_RCVD:
2924         case IB_CM_REQ_RCVD:
2925         case IB_CM_MRA_REQ_SENT:
2926         case IB_CM_REP_RCVD:
2927         case IB_CM_MRA_REP_SENT:
2928                 cm_reset_to_idle(cm_id_priv);
2929                 ret = cm_alloc_msg(cm_id_priv, &msg);
2930                 if (ret)
2931                         return ret;
2932                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2933                               ari, ari_length, private_data, private_data_len,
2934                               state);
2935                 break;
2936         case IB_CM_REP_SENT:
2937         case IB_CM_MRA_REP_RCVD:
2938                 cm_enter_timewait(cm_id_priv);
2939                 ret = cm_alloc_msg(cm_id_priv, &msg);
2940                 if (ret)
2941                         return ret;
2942                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2943                               ari, ari_length, private_data, private_data_len,
2944                               state);
2945                 break;
2946         default:
2947                 pr_debug("%s: local_id %d, cm_id->state: %d\n", __func__,
2948                          be32_to_cpu(cm_id_priv->id.local_id),
2949                          cm_id_priv->id.state);
2950                 return -EINVAL;
2951         }
2952
2953         ret = ib_post_send_mad(msg, NULL);
2954         if (ret) {
2955                 cm_free_msg(msg);
2956                 return ret;
2957         }
2958
2959         return 0;
2960 }
2961
2962 int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason,
2963                    void *ari, u8 ari_length, const void *private_data,
2964                    u8 private_data_len)
2965 {
2966         struct cm_id_private *cm_id_priv =
2967                 container_of(cm_id, struct cm_id_private, id);
2968         unsigned long flags;
2969         int ret;
2970
2971         spin_lock_irqsave(&cm_id_priv->lock, flags);
2972         ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length,
2973                                  private_data, private_data_len);
2974         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2975         return ret;
2976 }
2977 EXPORT_SYMBOL(ib_send_cm_rej);
2978
2979 static void cm_format_rej_event(struct cm_work *work)
2980 {
2981         struct cm_rej_msg *rej_msg;
2982         struct ib_cm_rej_event_param *param;
2983
2984         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2985         param = &work->cm_event.param.rej_rcvd;
2986         param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg);
2987         param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg);
2988         param->reason = IBA_GET(CM_REJ_REASON, rej_msg);
2989         work->cm_event.private_data =
2990                 IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg);
2991 }
2992
2993 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2994 {
2995         struct cm_id_private *cm_id_priv;
2996         __be32 remote_id;
2997
2998         remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg));
2999
3000         if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) {
3001                 cm_id_priv = cm_find_remote_id(
3002                         *((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)),
3003                         remote_id);
3004         } else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) ==
3005                    CM_MSG_RESPONSE_REQ)
3006                 cm_id_priv = cm_acquire_id(
3007                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3008                         0);
3009         else
3010                 cm_id_priv = cm_acquire_id(
3011                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3012                         remote_id);
3013
3014         return cm_id_priv;
3015 }
3016
3017 static int cm_rej_handler(struct cm_work *work)
3018 {
3019         struct cm_id_private *cm_id_priv;
3020         struct cm_rej_msg *rej_msg;
3021
3022         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
3023         cm_id_priv = cm_acquire_rejected_id(rej_msg);
3024         if (!cm_id_priv)
3025                 return -EINVAL;
3026
3027         cm_format_rej_event(work);
3028
3029         spin_lock_irq(&cm_id_priv->lock);
3030         switch (cm_id_priv->id.state) {
3031         case IB_CM_REQ_SENT:
3032         case IB_CM_MRA_REQ_RCVD:
3033         case IB_CM_REP_SENT:
3034         case IB_CM_MRA_REP_RCVD:
3035                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3036                 /* fall through */
3037         case IB_CM_REQ_RCVD:
3038         case IB_CM_MRA_REQ_SENT:
3039                 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN)
3040                         cm_enter_timewait(cm_id_priv);
3041                 else
3042                         cm_reset_to_idle(cm_id_priv);
3043                 break;
3044         case IB_CM_DREQ_SENT:
3045                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3046                 /* fall through */
3047         case IB_CM_REP_RCVD:
3048         case IB_CM_MRA_REP_SENT:
3049                 cm_enter_timewait(cm_id_priv);
3050                 break;
3051         case IB_CM_ESTABLISHED:
3052                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
3053                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
3054                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
3055                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
3056                                               cm_id_priv->msg);
3057                         cm_enter_timewait(cm_id_priv);
3058                         break;
3059                 }
3060                 /* fall through */
3061         default:
3062                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
3063                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
3064                          cm_id_priv->id.state);
3065                 spin_unlock_irq(&cm_id_priv->lock);
3066                 goto out;
3067         }
3068
3069         cm_queue_work_unlock(cm_id_priv, work);
3070         return 0;
3071 out:
3072         cm_deref_id(cm_id_priv);
3073         return -EINVAL;
3074 }
3075
3076 int ib_send_cm_mra(struct ib_cm_id *cm_id,
3077                    u8 service_timeout,
3078                    const void *private_data,
3079                    u8 private_data_len)
3080 {
3081         struct cm_id_private *cm_id_priv;
3082         struct ib_mad_send_buf *msg;
3083         enum ib_cm_state cm_state;
3084         enum ib_cm_lap_state lap_state;
3085         enum cm_msg_response msg_response;
3086         void *data;
3087         unsigned long flags;
3088         int ret;
3089
3090         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
3091                 return -EINVAL;
3092
3093         data = cm_copy_private_data(private_data, private_data_len);
3094         if (IS_ERR(data))
3095                 return PTR_ERR(data);
3096
3097         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3098
3099         spin_lock_irqsave(&cm_id_priv->lock, flags);
3100         switch(cm_id_priv->id.state) {
3101         case IB_CM_REQ_RCVD:
3102                 cm_state = IB_CM_MRA_REQ_SENT;
3103                 lap_state = cm_id->lap_state;
3104                 msg_response = CM_MSG_RESPONSE_REQ;
3105                 break;
3106         case IB_CM_REP_RCVD:
3107                 cm_state = IB_CM_MRA_REP_SENT;
3108                 lap_state = cm_id->lap_state;
3109                 msg_response = CM_MSG_RESPONSE_REP;
3110                 break;
3111         case IB_CM_ESTABLISHED:
3112                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
3113                         cm_state = cm_id->state;
3114                         lap_state = IB_CM_MRA_LAP_SENT;
3115                         msg_response = CM_MSG_RESPONSE_OTHER;
3116                         break;
3117                 }
3118                 /* fall through */
3119         default:
3120                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
3121                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
3122                          cm_id_priv->id.state);
3123                 ret = -EINVAL;
3124                 goto error1;
3125         }
3126
3127         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
3128                 ret = cm_alloc_msg(cm_id_priv, &msg);
3129                 if (ret)
3130                         goto error1;
3131
3132                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3133                               msg_response, service_timeout,
3134                               private_data, private_data_len);
3135                 ret = ib_post_send_mad(msg, NULL);
3136                 if (ret)
3137                         goto error2;
3138         }
3139
3140         cm_id->state = cm_state;
3141         cm_id->lap_state = lap_state;
3142         cm_id_priv->service_timeout = service_timeout;
3143         cm_set_private_data(cm_id_priv, data, private_data_len);
3144         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3145         return 0;
3146
3147 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3148         kfree(data);
3149         return ret;
3150
3151 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3152         kfree(data);
3153         cm_free_msg(msg);
3154         return ret;
3155 }
3156 EXPORT_SYMBOL(ib_send_cm_mra);
3157
3158 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
3159 {
3160         switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) {
3161         case CM_MSG_RESPONSE_REQ:
3162                 return cm_acquire_id(
3163                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3164                         0);
3165         case CM_MSG_RESPONSE_REP:
3166         case CM_MSG_RESPONSE_OTHER:
3167                 return cm_acquire_id(
3168                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3169                         cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg)));
3170         default:
3171                 return NULL;
3172         }
3173 }
3174
3175 static int cm_mra_handler(struct cm_work *work)
3176 {
3177         struct cm_id_private *cm_id_priv;
3178         struct cm_mra_msg *mra_msg;
3179         int timeout;
3180
3181         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
3182         cm_id_priv = cm_acquire_mraed_id(mra_msg);
3183         if (!cm_id_priv)
3184                 return -EINVAL;
3185
3186         work->cm_event.private_data =
3187                 IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg);
3188         work->cm_event.param.mra_rcvd.service_timeout =
3189                 IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg);
3190         timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) +
3191                   cm_convert_to_ms(cm_id_priv->av.timeout);
3192
3193         spin_lock_irq(&cm_id_priv->lock);
3194         switch (cm_id_priv->id.state) {
3195         case IB_CM_REQ_SENT:
3196                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3197                             CM_MSG_RESPONSE_REQ ||
3198                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3199                                   cm_id_priv->msg, timeout))
3200                         goto out;
3201                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
3202                 break;
3203         case IB_CM_REP_SENT:
3204                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3205                             CM_MSG_RESPONSE_REP ||
3206                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3207                                   cm_id_priv->msg, timeout))
3208                         goto out;
3209                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
3210                 break;
3211         case IB_CM_ESTABLISHED:
3212                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3213                             CM_MSG_RESPONSE_OTHER ||
3214                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
3215                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3216                                   cm_id_priv->msg, timeout)) {
3217                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
3218                                 atomic_long_inc(&work->port->
3219                                                 counter_group[CM_RECV_DUPLICATES].
3220                                                 counter[CM_MRA_COUNTER]);
3221                         goto out;
3222                 }
3223                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
3224                 break;
3225         case IB_CM_MRA_REQ_RCVD:
3226         case IB_CM_MRA_REP_RCVD:
3227                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3228                                 counter[CM_MRA_COUNTER]);
3229                 /* fall through */
3230         default:
3231                 pr_debug("%s local_id %d, cm_id_priv->id.state: %d\n",
3232                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
3233                          cm_id_priv->id.state);
3234                 goto out;
3235         }
3236
3237         cm_id_priv->msg->context[1] = (void *) (unsigned long)
3238                                       cm_id_priv->id.state;
3239         cm_queue_work_unlock(cm_id_priv, work);
3240         return 0;
3241 out:
3242         spin_unlock_irq(&cm_id_priv->lock);
3243         cm_deref_id(cm_id_priv);
3244         return -EINVAL;
3245 }
3246
3247 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3248                                         struct sa_path_rec *path)
3249 {
3250         u32 lid;
3251
3252         if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3253                 sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID,
3254                                                lap_msg));
3255                 sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID,
3256                                                lap_msg));
3257         } else {
3258                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3259                         CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg));
3260                 sa_path_set_dlid(path, lid);
3261
3262                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3263                         CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg));
3264                 sa_path_set_slid(path, lid);
3265         }
3266 }
3267
3268 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3269                                     struct sa_path_rec *path,
3270                                     struct cm_lap_msg *lap_msg)
3271 {
3272         path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg);
3273         path->sgid =
3274                 *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg);
3275         path->flow_label =
3276                 cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg));
3277         path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg);
3278         path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg);
3279         path->reversible = 1;
3280         path->pkey = cm_id_priv->pkey;
3281         path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg);
3282         path->mtu_selector = IB_SA_EQ;
3283         path->mtu = cm_id_priv->path_mtu;
3284         path->rate_selector = IB_SA_EQ;
3285         path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg);
3286         path->packet_life_time_selector = IB_SA_EQ;
3287         path->packet_life_time =
3288                 IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg);
3289         path->packet_life_time -= (path->packet_life_time > 0);
3290         cm_format_path_lid_from_lap(lap_msg, path);
3291 }
3292
3293 static int cm_lap_handler(struct cm_work *work)
3294 {
3295         struct cm_id_private *cm_id_priv;
3296         struct cm_lap_msg *lap_msg;
3297         struct ib_cm_lap_event_param *param;
3298         struct ib_mad_send_buf *msg = NULL;
3299         int ret;
3300
3301         /* Currently Alternate path messages are not supported for
3302          * RoCE link layer.
3303          */
3304         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3305                                work->port->port_num))
3306                 return -EINVAL;
3307
3308         /* todo: verify LAP request and send reject APR if invalid. */
3309         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3310         cm_id_priv = cm_acquire_id(
3311                 cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)),
3312                 cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg)));
3313         if (!cm_id_priv)
3314                 return -EINVAL;
3315
3316         param = &work->cm_event.param.lap_rcvd;
3317         memset(&work->path[0], 0, sizeof(work->path[1]));
3318         cm_path_set_rec_type(work->port->cm_dev->ib_device,
3319                              work->port->port_num, &work->path[0],
3320                              IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID,
3321                                              lap_msg));
3322         param->alternate_path = &work->path[0];
3323         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3324         work->cm_event.private_data =
3325                 IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg);
3326
3327         spin_lock_irq(&cm_id_priv->lock);
3328         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3329                 goto unlock;
3330
3331         switch (cm_id_priv->id.lap_state) {
3332         case IB_CM_LAP_UNINIT:
3333         case IB_CM_LAP_IDLE:
3334                 break;
3335         case IB_CM_MRA_LAP_SENT:
3336                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3337                                 counter[CM_LAP_COUNTER]);
3338                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
3339                 if (IS_ERR(msg))
3340                         goto unlock;
3341
3342                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3343                               CM_MSG_RESPONSE_OTHER,
3344                               cm_id_priv->service_timeout,
3345                               cm_id_priv->private_data,
3346                               cm_id_priv->private_data_len);
3347                 spin_unlock_irq(&cm_id_priv->lock);
3348
3349                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3350                     ib_post_send_mad(msg, NULL))
3351                         cm_free_msg(msg);
3352                 goto deref;
3353         case IB_CM_LAP_RCVD:
3354                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3355                                 counter[CM_LAP_COUNTER]);
3356                 goto unlock;
3357         default:
3358                 goto unlock;
3359         }
3360
3361         ret = cm_init_av_for_lap(work->port, work->mad_recv_wc->wc,
3362                                  work->mad_recv_wc->recv_buf.grh,
3363                                  &cm_id_priv->av);
3364         if (ret)
3365                 goto unlock;
3366
3367         ret = cm_init_av_by_path(param->alternate_path, NULL,
3368                                  &cm_id_priv->alt_av, cm_id_priv);
3369         if (ret)
3370                 goto unlock;
3371
3372         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3373         cm_id_priv->tid = lap_msg->hdr.tid;
3374         cm_queue_work_unlock(cm_id_priv, work);
3375         return 0;
3376
3377 unlock: spin_unlock_irq(&cm_id_priv->lock);
3378 deref:  cm_deref_id(cm_id_priv);
3379         return -EINVAL;
3380 }
3381
3382 static int cm_apr_handler(struct cm_work *work)
3383 {
3384         struct cm_id_private *cm_id_priv;
3385         struct cm_apr_msg *apr_msg;
3386
3387         /* Currently Alternate path messages are not supported for
3388          * RoCE link layer.
3389          */
3390         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3391                                work->port->port_num))
3392                 return -EINVAL;
3393
3394         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3395         cm_id_priv = cm_acquire_id(
3396                 cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)),
3397                 cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg)));
3398         if (!cm_id_priv)
3399                 return -EINVAL; /* Unmatched reply. */
3400
3401         work->cm_event.param.apr_rcvd.ap_status =
3402                 IBA_GET(CM_APR_AR_STATUS, apr_msg);
3403         work->cm_event.param.apr_rcvd.apr_info =
3404                 IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg);
3405         work->cm_event.param.apr_rcvd.info_len =
3406                 IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg);
3407         work->cm_event.private_data =
3408                 IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg);
3409
3410         spin_lock_irq(&cm_id_priv->lock);
3411         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3412             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3413              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3414                 spin_unlock_irq(&cm_id_priv->lock);
3415                 goto out;
3416         }
3417         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3418         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3419         cm_id_priv->msg = NULL;
3420         cm_queue_work_unlock(cm_id_priv, work);
3421         return 0;
3422 out:
3423         cm_deref_id(cm_id_priv);
3424         return -EINVAL;
3425 }
3426
3427 static int cm_timewait_handler(struct cm_work *work)
3428 {
3429         struct cm_timewait_info *timewait_info;
3430         struct cm_id_private *cm_id_priv;
3431
3432         timewait_info = container_of(work, struct cm_timewait_info, work);
3433         spin_lock_irq(&cm.lock);
3434         list_del(&timewait_info->list);
3435         spin_unlock_irq(&cm.lock);
3436
3437         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3438                                    timewait_info->work.remote_id);
3439         if (!cm_id_priv)
3440                 return -EINVAL;
3441
3442         spin_lock_irq(&cm_id_priv->lock);
3443         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3444             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3445                 spin_unlock_irq(&cm_id_priv->lock);
3446                 goto out;
3447         }
3448         cm_id_priv->id.state = IB_CM_IDLE;
3449         cm_queue_work_unlock(cm_id_priv, work);
3450         return 0;
3451 out:
3452         cm_deref_id(cm_id_priv);
3453         return -EINVAL;
3454 }
3455
3456 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3457                                struct cm_id_private *cm_id_priv,
3458                                struct ib_cm_sidr_req_param *param)
3459 {
3460         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3461                           cm_form_tid(cm_id_priv));
3462         IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg,
3463                 be32_to_cpu(cm_id_priv->id.local_id));
3464         IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg,
3465                 be16_to_cpu(param->path->pkey));
3466         IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg,
3467                 be64_to_cpu(param->service_id));
3468
3469         if (param->private_data && param->private_data_len)
3470                 IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg,
3471                             param->private_data, param->private_data_len);
3472 }
3473
3474 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3475                         struct ib_cm_sidr_req_param *param)
3476 {
3477         struct cm_id_private *cm_id_priv;
3478         struct ib_mad_send_buf *msg;
3479         unsigned long flags;
3480         int ret;
3481
3482         if (!param->path || (param->private_data &&
3483              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3484                 return -EINVAL;
3485
3486         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3487         ret = cm_init_av_by_path(param->path, param->sgid_attr,
3488                                  &cm_id_priv->av,
3489                                  cm_id_priv);
3490         if (ret)
3491                 goto out;
3492
3493         cm_id->service_id = param->service_id;
3494         cm_id->service_mask = ~cpu_to_be64(0);
3495         cm_id_priv->timeout_ms = param->timeout_ms;
3496         cm_id_priv->max_cm_retries = param->max_cm_retries;
3497         ret = cm_alloc_msg(cm_id_priv, &msg);
3498         if (ret)
3499                 goto out;
3500
3501         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3502                            param);
3503         msg->timeout_ms = cm_id_priv->timeout_ms;
3504         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3505
3506         spin_lock_irqsave(&cm_id_priv->lock, flags);
3507         if (cm_id->state == IB_CM_IDLE)
3508                 ret = ib_post_send_mad(msg, NULL);
3509         else
3510                 ret = -EINVAL;
3511
3512         if (ret) {
3513                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3514                 cm_free_msg(msg);
3515                 goto out;
3516         }
3517         cm_id->state = IB_CM_SIDR_REQ_SENT;
3518         cm_id_priv->msg = msg;
3519         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3520 out:
3521         return ret;
3522 }
3523 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3524
3525 static void cm_format_sidr_req_event(struct cm_work *work,
3526                                      const struct cm_id_private *rx_cm_id,
3527                                      struct ib_cm_id *listen_id)
3528 {
3529         struct cm_sidr_req_msg *sidr_req_msg;
3530         struct ib_cm_sidr_req_event_param *param;
3531
3532         sidr_req_msg = (struct cm_sidr_req_msg *)
3533                                 work->mad_recv_wc->recv_buf.mad;
3534         param = &work->cm_event.param.sidr_req_rcvd;
3535         param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg);
3536         param->listen_id = listen_id;
3537         param->service_id =
3538                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3539         param->bth_pkey = cm_get_bth_pkey(work);
3540         param->port = work->port->port_num;
3541         param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr;
3542         work->cm_event.private_data =
3543                 IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg);
3544 }
3545
3546 static int cm_sidr_req_handler(struct cm_work *work)
3547 {
3548         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
3549         struct cm_sidr_req_msg *sidr_req_msg;
3550         struct ib_wc *wc;
3551         int ret;
3552
3553         cm_id_priv =
3554                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
3555         if (IS_ERR(cm_id_priv))
3556                 return PTR_ERR(cm_id_priv);
3557
3558         /* Record SGID/SLID and request ID for lookup. */
3559         sidr_req_msg = (struct cm_sidr_req_msg *)
3560                                 work->mad_recv_wc->recv_buf.mad;
3561
3562         cm_id_priv->id.remote_id =
3563                 cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg));
3564         cm_id_priv->id.service_id =
3565                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3566         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3567         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3568
3569         wc = work->mad_recv_wc->wc;
3570         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3571         cm_id_priv->av.dgid.global.interface_id = 0;
3572         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3573                                       work->mad_recv_wc->recv_buf.grh,
3574                                       &cm_id_priv->av);
3575         if (ret)
3576                 goto out;
3577
3578         spin_lock_irq(&cm.lock);
3579         listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3580         if (listen_cm_id_priv) {
3581                 spin_unlock_irq(&cm.lock);
3582                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3583                                 counter[CM_SIDR_REQ_COUNTER]);
3584                 goto out; /* Duplicate message. */
3585         }
3586         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3587         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
3588                                            cm_id_priv->id.service_id);
3589         if (!listen_cm_id_priv) {
3590                 spin_unlock_irq(&cm.lock);
3591                 ib_send_cm_sidr_rep(&cm_id_priv->id,
3592                                     &(struct ib_cm_sidr_rep_param){
3593                                             .status = IB_SIDR_UNSUPPORTED });
3594                 goto out; /* No match. */
3595         }
3596         spin_unlock_irq(&cm.lock);
3597
3598         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
3599         cm_id_priv->id.context = listen_cm_id_priv->id.context;
3600
3601         /*
3602          * A SIDR ID does not need to be in the xarray since it does not receive
3603          * mads, is not placed in the remote_id or remote_qpn rbtree, and does
3604          * not enter timewait.
3605          */
3606
3607         cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
3608         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
3609         cm_free_work(work);
3610         /*
3611          * A pointer to the listen_cm_id is held in the event, so this deref
3612          * must be after the event is delivered above.
3613          */
3614         cm_deref_id(listen_cm_id_priv);
3615         if (ret)
3616                 cm_destroy_id(&cm_id_priv->id, ret);
3617         return 0;
3618 out:
3619         ib_destroy_cm_id(&cm_id_priv->id);
3620         return -EINVAL;
3621 }
3622
3623 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3624                                struct cm_id_private *cm_id_priv,
3625                                struct ib_cm_sidr_rep_param *param)
3626 {
3627         cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3628                               cm_id_priv->tid, param->ece.attr_mod);
3629         IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg,
3630                 be32_to_cpu(cm_id_priv->id.remote_id));
3631         IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status);
3632         IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num);
3633         IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg,
3634                 be64_to_cpu(cm_id_priv->id.service_id));
3635         IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey);
3636         IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg,
3637                 param->ece.vendor_id & 0xFF);
3638         IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg,
3639                 (param->ece.vendor_id >> 8) & 0xFF);
3640
3641         if (param->info && param->info_length)
3642                 IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg,
3643                             param->info, param->info_length);
3644
3645         if (param->private_data && param->private_data_len)
3646                 IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg,
3647                             param->private_data, param->private_data_len);
3648 }
3649
3650 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
3651                                    struct ib_cm_sidr_rep_param *param)
3652 {
3653         struct ib_mad_send_buf *msg;
3654         int ret;
3655
3656         lockdep_assert_held(&cm_id_priv->lock);
3657
3658         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3659             (param->private_data &&
3660              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3661                 return -EINVAL;
3662
3663         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD)
3664                 return -EINVAL;
3665
3666         ret = cm_alloc_msg(cm_id_priv, &msg);
3667         if (ret)
3668                 return ret;
3669
3670         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3671                            param);
3672         ret = ib_post_send_mad(msg, NULL);
3673         if (ret) {
3674                 cm_free_msg(msg);
3675                 return ret;
3676         }
3677         cm_id_priv->id.state = IB_CM_IDLE;
3678         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3679                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3680                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3681         }
3682         return 0;
3683 }
3684
3685 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3686                         struct ib_cm_sidr_rep_param *param)
3687 {
3688         struct cm_id_private *cm_id_priv =
3689                 container_of(cm_id, struct cm_id_private, id);
3690         unsigned long flags;
3691         int ret;
3692
3693         spin_lock_irqsave(&cm_id_priv->lock, flags);
3694         ret = cm_send_sidr_rep_locked(cm_id_priv, param);
3695         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3696         return ret;
3697 }
3698 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3699
3700 static void cm_format_sidr_rep_event(struct cm_work *work,
3701                                      const struct cm_id_private *cm_id_priv)
3702 {
3703         struct cm_sidr_rep_msg *sidr_rep_msg;
3704         struct ib_cm_sidr_rep_event_param *param;
3705
3706         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3707                                 work->mad_recv_wc->recv_buf.mad;
3708         param = &work->cm_event.param.sidr_rep_rcvd;
3709         param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg);
3710         param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg);
3711         param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg);
3712         param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION,
3713                                       sidr_rep_msg);
3714         param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH,
3715                                   sidr_rep_msg);
3716         param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
3717         work->cm_event.private_data =
3718                 IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg);
3719 }
3720
3721 static int cm_sidr_rep_handler(struct cm_work *work)
3722 {
3723         struct cm_sidr_rep_msg *sidr_rep_msg;
3724         struct cm_id_private *cm_id_priv;
3725
3726         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3727                                 work->mad_recv_wc->recv_buf.mad;
3728         cm_id_priv = cm_acquire_id(
3729                 cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0);
3730         if (!cm_id_priv)
3731                 return -EINVAL; /* Unmatched reply. */
3732
3733         spin_lock_irq(&cm_id_priv->lock);
3734         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3735                 spin_unlock_irq(&cm_id_priv->lock);
3736                 goto out;
3737         }
3738         cm_id_priv->id.state = IB_CM_IDLE;
3739         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3740         spin_unlock_irq(&cm_id_priv->lock);
3741
3742         cm_format_sidr_rep_event(work, cm_id_priv);
3743         cm_process_work(cm_id_priv, work);
3744         return 0;
3745 out:
3746         cm_deref_id(cm_id_priv);
3747         return -EINVAL;
3748 }
3749
3750 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3751                                   enum ib_wc_status wc_status)
3752 {
3753         struct cm_id_private *cm_id_priv;
3754         struct ib_cm_event cm_event;
3755         enum ib_cm_state state;
3756         int ret;
3757
3758         memset(&cm_event, 0, sizeof cm_event);
3759         cm_id_priv = msg->context[0];
3760
3761         /* Discard old sends or ones without a response. */
3762         spin_lock_irq(&cm_id_priv->lock);
3763         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3764         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3765                 goto discard;
3766
3767         pr_debug_ratelimited("CM: failed sending MAD in state %d. (%s)\n",
3768                              state, ib_wc_status_msg(wc_status));
3769         switch (state) {
3770         case IB_CM_REQ_SENT:
3771         case IB_CM_MRA_REQ_RCVD:
3772                 cm_reset_to_idle(cm_id_priv);
3773                 cm_event.event = IB_CM_REQ_ERROR;
3774                 break;
3775         case IB_CM_REP_SENT:
3776         case IB_CM_MRA_REP_RCVD:
3777                 cm_reset_to_idle(cm_id_priv);
3778                 cm_event.event = IB_CM_REP_ERROR;
3779                 break;
3780         case IB_CM_DREQ_SENT:
3781                 cm_enter_timewait(cm_id_priv);
3782                 cm_event.event = IB_CM_DREQ_ERROR;
3783                 break;
3784         case IB_CM_SIDR_REQ_SENT:
3785                 cm_id_priv->id.state = IB_CM_IDLE;
3786                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3787                 break;
3788         default:
3789                 goto discard;
3790         }
3791         spin_unlock_irq(&cm_id_priv->lock);
3792         cm_event.param.send_status = wc_status;
3793
3794         /* No other events can occur on the cm_id at this point. */
3795         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3796         cm_free_msg(msg);
3797         if (ret)
3798                 ib_destroy_cm_id(&cm_id_priv->id);
3799         return;
3800 discard:
3801         spin_unlock_irq(&cm_id_priv->lock);
3802         cm_free_msg(msg);
3803 }
3804
3805 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3806                             struct ib_mad_send_wc *mad_send_wc)
3807 {
3808         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3809         struct cm_port *port;
3810         u16 attr_index;
3811
3812         port = mad_agent->context;
3813         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3814                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3815
3816         /*
3817          * If the send was in response to a received message (context[0] is not
3818          * set to a cm_id), and is not a REJ, then it is a send that was
3819          * manually retried.
3820          */
3821         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3822                 msg->retries = 1;
3823
3824         atomic_long_add(1 + msg->retries,
3825                         &port->counter_group[CM_XMIT].counter[attr_index]);
3826         if (msg->retries)
3827                 atomic_long_add(msg->retries,
3828                                 &port->counter_group[CM_XMIT_RETRIES].
3829                                 counter[attr_index]);
3830
3831         switch (mad_send_wc->status) {
3832         case IB_WC_SUCCESS:
3833         case IB_WC_WR_FLUSH_ERR:
3834                 cm_free_msg(msg);
3835                 break;
3836         default:
3837                 if (msg->context[0] && msg->context[1])
3838                         cm_process_send_error(msg, mad_send_wc->status);
3839                 else
3840                         cm_free_msg(msg);
3841                 break;
3842         }
3843 }
3844
3845 static void cm_work_handler(struct work_struct *_work)
3846 {
3847         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3848         int ret;
3849
3850         switch (work->cm_event.event) {
3851         case IB_CM_REQ_RECEIVED:
3852                 ret = cm_req_handler(work);
3853                 break;
3854         case IB_CM_MRA_RECEIVED:
3855                 ret = cm_mra_handler(work);
3856                 break;
3857         case IB_CM_REJ_RECEIVED:
3858                 ret = cm_rej_handler(work);
3859                 break;
3860         case IB_CM_REP_RECEIVED:
3861                 ret = cm_rep_handler(work);
3862                 break;
3863         case IB_CM_RTU_RECEIVED:
3864                 ret = cm_rtu_handler(work);
3865                 break;
3866         case IB_CM_USER_ESTABLISHED:
3867                 ret = cm_establish_handler(work);
3868                 break;
3869         case IB_CM_DREQ_RECEIVED:
3870                 ret = cm_dreq_handler(work);
3871                 break;
3872         case IB_CM_DREP_RECEIVED:
3873                 ret = cm_drep_handler(work);
3874                 break;
3875         case IB_CM_SIDR_REQ_RECEIVED:
3876                 ret = cm_sidr_req_handler(work);
3877                 break;
3878         case IB_CM_SIDR_REP_RECEIVED:
3879                 ret = cm_sidr_rep_handler(work);
3880                 break;
3881         case IB_CM_LAP_RECEIVED:
3882                 ret = cm_lap_handler(work);
3883                 break;
3884         case IB_CM_APR_RECEIVED:
3885                 ret = cm_apr_handler(work);
3886                 break;
3887         case IB_CM_TIMEWAIT_EXIT:
3888                 ret = cm_timewait_handler(work);
3889                 break;
3890         default:
3891                 pr_debug("cm_event.event: 0x%x\n", work->cm_event.event);
3892                 ret = -EINVAL;
3893                 break;
3894         }
3895         if (ret)
3896                 cm_free_work(work);
3897 }
3898
3899 static int cm_establish(struct ib_cm_id *cm_id)
3900 {
3901         struct cm_id_private *cm_id_priv;
3902         struct cm_work *work;
3903         unsigned long flags;
3904         int ret = 0;
3905         struct cm_device *cm_dev;
3906
3907         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3908         if (!cm_dev)
3909                 return -ENODEV;
3910
3911         work = kmalloc(sizeof *work, GFP_ATOMIC);
3912         if (!work)
3913                 return -ENOMEM;
3914
3915         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3916         spin_lock_irqsave(&cm_id_priv->lock, flags);
3917         switch (cm_id->state)
3918         {
3919         case IB_CM_REP_SENT:
3920         case IB_CM_MRA_REP_RCVD:
3921                 cm_id->state = IB_CM_ESTABLISHED;
3922                 break;
3923         case IB_CM_ESTABLISHED:
3924                 ret = -EISCONN;
3925                 break;
3926         default:
3927                 pr_debug("%s: local_id %d, cm_id->state: %d\n", __func__,
3928                          be32_to_cpu(cm_id->local_id), cm_id->state);
3929                 ret = -EINVAL;
3930                 break;
3931         }
3932         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3933
3934         if (ret) {
3935                 kfree(work);
3936                 goto out;
3937         }
3938
3939         /*
3940          * The CM worker thread may try to destroy the cm_id before it
3941          * can execute this work item.  To prevent potential deadlock,
3942          * we need to find the cm_id once we're in the context of the
3943          * worker thread, rather than holding a reference on it.
3944          */
3945         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3946         work->local_id = cm_id->local_id;
3947         work->remote_id = cm_id->remote_id;
3948         work->mad_recv_wc = NULL;
3949         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3950
3951         /* Check if the device started its remove_one */
3952         spin_lock_irqsave(&cm.lock, flags);
3953         if (!cm_dev->going_down) {
3954                 queue_delayed_work(cm.wq, &work->work, 0);
3955         } else {
3956                 kfree(work);
3957                 ret = -ENODEV;
3958         }
3959         spin_unlock_irqrestore(&cm.lock, flags);
3960
3961 out:
3962         return ret;
3963 }
3964
3965 static int cm_migrate(struct ib_cm_id *cm_id)
3966 {
3967         struct cm_id_private *cm_id_priv;
3968         struct cm_av tmp_av;
3969         unsigned long flags;
3970         int tmp_send_port_not_ready;
3971         int ret = 0;
3972
3973         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3974         spin_lock_irqsave(&cm_id_priv->lock, flags);
3975         if (cm_id->state == IB_CM_ESTABLISHED &&
3976             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3977              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3978                 cm_id->lap_state = IB_CM_LAP_IDLE;
3979                 /* Swap address vector */
3980                 tmp_av = cm_id_priv->av;
3981                 cm_id_priv->av = cm_id_priv->alt_av;
3982                 cm_id_priv->alt_av = tmp_av;
3983                 /* Swap port send ready state */
3984                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3985                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3986                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3987         } else
3988                 ret = -EINVAL;
3989         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3990
3991         return ret;
3992 }
3993
3994 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3995 {
3996         int ret;
3997
3998         switch (event) {
3999         case IB_EVENT_COMM_EST:
4000                 ret = cm_establish(cm_id);
4001                 break;
4002         case IB_EVENT_PATH_MIG:
4003                 ret = cm_migrate(cm_id);
4004                 break;
4005         default:
4006                 ret = -EINVAL;
4007         }
4008         return ret;
4009 }
4010 EXPORT_SYMBOL(ib_cm_notify);
4011
4012 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
4013                             struct ib_mad_send_buf *send_buf,
4014                             struct ib_mad_recv_wc *mad_recv_wc)
4015 {
4016         struct cm_port *port = mad_agent->context;
4017         struct cm_work *work;
4018         enum ib_cm_event_type event;
4019         bool alt_path = false;
4020         u16 attr_id;
4021         int paths = 0;
4022         int going_down = 0;
4023
4024         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
4025         case CM_REQ_ATTR_ID:
4026                 alt_path = cm_req_has_alt_path((struct cm_req_msg *)
4027                                                 mad_recv_wc->recv_buf.mad);
4028                 paths = 1 + (alt_path != 0);
4029                 event = IB_CM_REQ_RECEIVED;
4030                 break;
4031         case CM_MRA_ATTR_ID:
4032                 event = IB_CM_MRA_RECEIVED;
4033                 break;
4034         case CM_REJ_ATTR_ID:
4035                 event = IB_CM_REJ_RECEIVED;
4036                 break;
4037         case CM_REP_ATTR_ID:
4038                 event = IB_CM_REP_RECEIVED;
4039                 break;
4040         case CM_RTU_ATTR_ID:
4041                 event = IB_CM_RTU_RECEIVED;
4042                 break;
4043         case CM_DREQ_ATTR_ID:
4044                 event = IB_CM_DREQ_RECEIVED;
4045                 break;
4046         case CM_DREP_ATTR_ID:
4047                 event = IB_CM_DREP_RECEIVED;
4048                 break;
4049         case CM_SIDR_REQ_ATTR_ID:
4050                 event = IB_CM_SIDR_REQ_RECEIVED;
4051                 break;
4052         case CM_SIDR_REP_ATTR_ID:
4053                 event = IB_CM_SIDR_REP_RECEIVED;
4054                 break;
4055         case CM_LAP_ATTR_ID:
4056                 paths = 1;
4057                 event = IB_CM_LAP_RECEIVED;
4058                 break;
4059         case CM_APR_ATTR_ID:
4060                 event = IB_CM_APR_RECEIVED;
4061                 break;
4062         default:
4063                 ib_free_recv_mad(mad_recv_wc);
4064                 return;
4065         }
4066
4067         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
4068         atomic_long_inc(&port->counter_group[CM_RECV].
4069                         counter[attr_id - CM_ATTR_ID_OFFSET]);
4070
4071         work = kmalloc(struct_size(work, path, paths), GFP_KERNEL);
4072         if (!work) {
4073                 ib_free_recv_mad(mad_recv_wc);
4074                 return;
4075         }
4076
4077         INIT_DELAYED_WORK(&work->work, cm_work_handler);
4078         work->cm_event.event = event;
4079         work->mad_recv_wc = mad_recv_wc;
4080         work->port = port;
4081
4082         /* Check if the device started its remove_one */
4083         spin_lock_irq(&cm.lock);
4084         if (!port->cm_dev->going_down)
4085                 queue_delayed_work(cm.wq, &work->work, 0);
4086         else
4087                 going_down = 1;
4088         spin_unlock_irq(&cm.lock);
4089
4090         if (going_down) {
4091                 kfree(work);
4092                 ib_free_recv_mad(mad_recv_wc);
4093         }
4094 }
4095
4096 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
4097                                 struct ib_qp_attr *qp_attr,
4098                                 int *qp_attr_mask)
4099 {
4100         unsigned long flags;
4101         int ret;
4102
4103         spin_lock_irqsave(&cm_id_priv->lock, flags);
4104         switch (cm_id_priv->id.state) {
4105         case IB_CM_REQ_SENT:
4106         case IB_CM_MRA_REQ_RCVD:
4107         case IB_CM_REQ_RCVD:
4108         case IB_CM_MRA_REQ_SENT:
4109         case IB_CM_REP_RCVD:
4110         case IB_CM_MRA_REP_SENT:
4111         case IB_CM_REP_SENT:
4112         case IB_CM_MRA_REP_RCVD:
4113         case IB_CM_ESTABLISHED:
4114                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
4115                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
4116                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
4117                 if (cm_id_priv->responder_resources)
4118                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
4119                                                     IB_ACCESS_REMOTE_ATOMIC;
4120                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
4121                 qp_attr->port_num = cm_id_priv->av.port->port_num;
4122                 ret = 0;
4123                 break;
4124         default:
4125                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
4126                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
4127                          cm_id_priv->id.state);
4128                 ret = -EINVAL;
4129                 break;
4130         }
4131         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4132         return ret;
4133 }
4134
4135 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
4136                                struct ib_qp_attr *qp_attr,
4137                                int *qp_attr_mask)
4138 {
4139         unsigned long flags;
4140         int ret;
4141
4142         spin_lock_irqsave(&cm_id_priv->lock, flags);
4143         switch (cm_id_priv->id.state) {
4144         case IB_CM_REQ_RCVD:
4145         case IB_CM_MRA_REQ_SENT:
4146         case IB_CM_REP_RCVD:
4147         case IB_CM_MRA_REP_SENT:
4148         case IB_CM_REP_SENT:
4149         case IB_CM_MRA_REP_RCVD:
4150         case IB_CM_ESTABLISHED:
4151                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
4152                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
4153                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
4154                 qp_attr->path_mtu = cm_id_priv->path_mtu;
4155                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
4156                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
4157                 if (cm_id_priv->qp_type == IB_QPT_RC ||
4158                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
4159                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
4160                                          IB_QP_MIN_RNR_TIMER;
4161                         qp_attr->max_dest_rd_atomic =
4162                                         cm_id_priv->responder_resources;
4163                         qp_attr->min_rnr_timer = 0;
4164                 }
4165                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4166                         *qp_attr_mask |= IB_QP_ALT_PATH;
4167                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4168                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4169                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4170                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4171                 }
4172                 ret = 0;
4173                 break;
4174         default:
4175                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
4176                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
4177                          cm_id_priv->id.state);
4178                 ret = -EINVAL;
4179                 break;
4180         }
4181         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4182         return ret;
4183 }
4184
4185 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
4186                                struct ib_qp_attr *qp_attr,
4187                                int *qp_attr_mask)
4188 {
4189         unsigned long flags;
4190         int ret;
4191
4192         spin_lock_irqsave(&cm_id_priv->lock, flags);
4193         switch (cm_id_priv->id.state) {
4194         /* Allow transition to RTS before sending REP */
4195         case IB_CM_REQ_RCVD:
4196         case IB_CM_MRA_REQ_SENT:
4197
4198         case IB_CM_REP_RCVD:
4199         case IB_CM_MRA_REP_SENT:
4200         case IB_CM_REP_SENT:
4201         case IB_CM_MRA_REP_RCVD:
4202         case IB_CM_ESTABLISHED:
4203                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
4204                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
4205                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4206                         switch (cm_id_priv->qp_type) {
4207                         case IB_QPT_RC:
4208                         case IB_QPT_XRC_INI:
4209                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4210                                                  IB_QP_MAX_QP_RD_ATOMIC;
4211                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
4212                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4213                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4214                                 /* fall through */
4215                         case IB_QPT_XRC_TGT:
4216                                 *qp_attr_mask |= IB_QP_TIMEOUT;
4217                                 qp_attr->timeout = cm_id_priv->av.timeout;
4218                                 break;
4219                         default:
4220                                 break;
4221                         }
4222                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4223                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4224                                 qp_attr->path_mig_state = IB_MIG_REARM;
4225                         }
4226                 } else {
4227                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4228                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4229                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4230                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4231                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4232                         qp_attr->path_mig_state = IB_MIG_REARM;
4233                 }
4234                 ret = 0;
4235                 break;
4236         default:
4237                 pr_debug("%s: local_id %d, cm_id_priv->id.state: %d\n",
4238                          __func__, be32_to_cpu(cm_id_priv->id.local_id),
4239                          cm_id_priv->id.state);
4240                 ret = -EINVAL;
4241                 break;
4242         }
4243         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4244         return ret;
4245 }
4246
4247 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4248                        struct ib_qp_attr *qp_attr,
4249                        int *qp_attr_mask)
4250 {
4251         struct cm_id_private *cm_id_priv;
4252         int ret;
4253
4254         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4255         switch (qp_attr->qp_state) {
4256         case IB_QPS_INIT:
4257                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4258                 break;
4259         case IB_QPS_RTR:
4260                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4261                 break;
4262         case IB_QPS_RTS:
4263                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4264                 break;
4265         default:
4266                 ret = -EINVAL;
4267                 break;
4268         }
4269         return ret;
4270 }
4271 EXPORT_SYMBOL(ib_cm_init_qp_attr);
4272
4273 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
4274                                char *buf)
4275 {
4276         struct cm_counter_group *group;
4277         struct cm_counter_attribute *cm_attr;
4278
4279         group = container_of(obj, struct cm_counter_group, obj);
4280         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
4281
4282         return sprintf(buf, "%ld\n",
4283                        atomic_long_read(&group->counter[cm_attr->index]));
4284 }
4285
4286 static const struct sysfs_ops cm_counter_ops = {
4287         .show = cm_show_counter
4288 };
4289
4290 static struct kobj_type cm_counter_obj_type = {
4291         .sysfs_ops = &cm_counter_ops,
4292         .default_attrs = cm_counter_default_attrs
4293 };
4294
4295 static char *cm_devnode(struct device *dev, umode_t *mode)
4296 {
4297         if (mode)
4298                 *mode = 0666;
4299         return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
4300 }
4301
4302 struct class cm_class = {
4303         .owner   = THIS_MODULE,
4304         .name    = "infiniband_cm",
4305         .devnode = cm_devnode,
4306 };
4307 EXPORT_SYMBOL(cm_class);
4308
4309 static int cm_create_port_fs(struct cm_port *port)
4310 {
4311         int i, ret;
4312
4313         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
4314                 ret = ib_port_register_module_stat(port->cm_dev->ib_device,
4315                                                    port->port_num,
4316                                                    &port->counter_group[i].obj,
4317                                                    &cm_counter_obj_type,
4318                                                    counter_group_names[i]);
4319                 if (ret)
4320                         goto error;
4321         }
4322
4323         return 0;
4324
4325 error:
4326         while (i--)
4327                 ib_port_unregister_module_stat(&port->counter_group[i].obj);
4328         return ret;
4329
4330 }
4331
4332 static void cm_remove_port_fs(struct cm_port *port)
4333 {
4334         int i;
4335
4336         for (i = 0; i < CM_COUNTER_GROUPS; i++)
4337                 ib_port_unregister_module_stat(&port->counter_group[i].obj);
4338
4339 }
4340
4341 static int cm_add_one(struct ib_device *ib_device)
4342 {
4343         struct cm_device *cm_dev;
4344         struct cm_port *port;
4345         struct ib_mad_reg_req reg_req = {
4346                 .mgmt_class = IB_MGMT_CLASS_CM,
4347                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4348         };
4349         struct ib_port_modify port_modify = {
4350                 .set_port_cap_mask = IB_PORT_CM_SUP
4351         };
4352         unsigned long flags;
4353         int ret;
4354         int count = 0;
4355         u8 i;
4356
4357         cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt),
4358                          GFP_KERNEL);
4359         if (!cm_dev)
4360                 return -ENOMEM;
4361
4362         cm_dev->ib_device = ib_device;
4363         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4364         cm_dev->going_down = 0;
4365
4366         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4367         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4368                 if (!rdma_cap_ib_cm(ib_device, i))
4369                         continue;
4370
4371                 port = kzalloc(sizeof *port, GFP_KERNEL);
4372                 if (!port) {
4373                         ret = -ENOMEM;
4374                         goto error1;
4375                 }
4376
4377                 cm_dev->port[i-1] = port;
4378                 port->cm_dev = cm_dev;
4379                 port->port_num = i;
4380
4381                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4382                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4383
4384                 ret = cm_create_port_fs(port);
4385                 if (ret)
4386                         goto error1;
4387
4388                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4389                                                         IB_QPT_GSI,
4390                                                         &reg_req,
4391                                                         0,
4392                                                         cm_send_handler,
4393                                                         cm_recv_handler,
4394                                                         port,
4395                                                         0);
4396                 if (IS_ERR(port->mad_agent)) {
4397                         ret = PTR_ERR(port->mad_agent);
4398                         goto error2;
4399                 }
4400
4401                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4402                 if (ret)
4403                         goto error3;
4404
4405                 count++;
4406         }
4407
4408         if (!count) {
4409                 ret = -EOPNOTSUPP;
4410                 goto free;
4411         }
4412
4413         ib_set_client_data(ib_device, &cm_client, cm_dev);
4414
4415         write_lock_irqsave(&cm.device_lock, flags);
4416         list_add_tail(&cm_dev->list, &cm.device_list);
4417         write_unlock_irqrestore(&cm.device_lock, flags);
4418         return 0;
4419
4420 error3:
4421         ib_unregister_mad_agent(port->mad_agent);
4422 error2:
4423         cm_remove_port_fs(port);
4424 error1:
4425         port_modify.set_port_cap_mask = 0;
4426         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4427         kfree(port);
4428         while (--i) {
4429                 if (!rdma_cap_ib_cm(ib_device, i))
4430                         continue;
4431
4432                 port = cm_dev->port[i-1];
4433                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4434                 ib_unregister_mad_agent(port->mad_agent);
4435                 cm_remove_port_fs(port);
4436                 kfree(port);
4437         }
4438 free:
4439         kfree(cm_dev);
4440         return ret;
4441 }
4442
4443 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4444 {
4445         struct cm_device *cm_dev = client_data;
4446         struct cm_port *port;
4447         struct cm_id_private *cm_id_priv;
4448         struct ib_mad_agent *cur_mad_agent;
4449         struct ib_port_modify port_modify = {
4450                 .clr_port_cap_mask = IB_PORT_CM_SUP
4451         };
4452         unsigned long flags;
4453         int i;
4454
4455         write_lock_irqsave(&cm.device_lock, flags);
4456         list_del(&cm_dev->list);
4457         write_unlock_irqrestore(&cm.device_lock, flags);
4458
4459         spin_lock_irq(&cm.lock);
4460         cm_dev->going_down = 1;
4461         spin_unlock_irq(&cm.lock);
4462
4463         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4464                 if (!rdma_cap_ib_cm(ib_device, i))
4465                         continue;
4466
4467                 port = cm_dev->port[i-1];
4468                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4469                 /* Mark all the cm_id's as not valid */
4470                 spin_lock_irq(&cm.lock);
4471                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4472                         cm_id_priv->altr_send_port_not_ready = 1;
4473                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4474                         cm_id_priv->prim_send_port_not_ready = 1;
4475                 spin_unlock_irq(&cm.lock);
4476                 /*
4477                  * We flush the queue here after the going_down set, this
4478                  * verify that no new works will be queued in the recv handler,
4479                  * after that we can call the unregister_mad_agent
4480                  */
4481                 flush_workqueue(cm.wq);
4482                 spin_lock_irq(&cm.state_lock);
4483                 cur_mad_agent = port->mad_agent;
4484                 port->mad_agent = NULL;
4485                 spin_unlock_irq(&cm.state_lock);
4486                 ib_unregister_mad_agent(cur_mad_agent);
4487                 cm_remove_port_fs(port);
4488                 kfree(port);
4489         }
4490
4491         kfree(cm_dev);
4492 }
4493
4494 static int __init ib_cm_init(void)
4495 {
4496         int ret;
4497
4498         INIT_LIST_HEAD(&cm.device_list);
4499         rwlock_init(&cm.device_lock);
4500         spin_lock_init(&cm.lock);
4501         spin_lock_init(&cm.state_lock);
4502         cm.listen_service_table = RB_ROOT;
4503         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4504         cm.remote_id_table = RB_ROOT;
4505         cm.remote_qp_table = RB_ROOT;
4506         cm.remote_sidr_table = RB_ROOT;
4507         xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
4508         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4509         INIT_LIST_HEAD(&cm.timewait_list);
4510
4511         ret = class_register(&cm_class);
4512         if (ret) {
4513                 ret = -ENOMEM;
4514                 goto error1;
4515         }
4516
4517         cm.wq = alloc_workqueue("ib_cm", 0, 1);
4518         if (!cm.wq) {
4519                 ret = -ENOMEM;
4520                 goto error2;
4521         }
4522
4523         ret = ib_register_client(&cm_client);
4524         if (ret)
4525                 goto error3;
4526
4527         return 0;
4528 error3:
4529         destroy_workqueue(cm.wq);
4530 error2:
4531         class_unregister(&cm_class);
4532 error1:
4533         return ret;
4534 }
4535
4536 static void __exit ib_cm_cleanup(void)
4537 {
4538         struct cm_timewait_info *timewait_info, *tmp;
4539
4540         spin_lock_irq(&cm.lock);
4541         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4542                 cancel_delayed_work(&timewait_info->work.work);
4543         spin_unlock_irq(&cm.lock);
4544
4545         ib_unregister_client(&cm_client);
4546         destroy_workqueue(cm.wq);
4547
4548         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4549                 list_del(&timewait_info->list);
4550                 kfree(timewait_info);
4551         }
4552
4553         class_unregister(&cm_class);
4554         WARN_ON(!xa_empty(&cm.local_id_table));
4555 }
4556
4557 module_init(ib_cm_init);
4558 module_exit(ib_cm_cleanup);