drivers/infiniband/core/cm.c

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