drivers/infiniband/core/cm.c

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