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(&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(&cm.local_id_table, cm_local_id(cm_id_priv->id.local_id),
 882                  cm_id_priv, GFP_ATOMIC);
 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(&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                 cm_id_priv->timewait_info = NULL;
1526                 goto out;
1527         }
1528
1529         ret = cm_init_av_by_path(param->primary_path,
1530                                  param->ppath_sgid_attr, &cm_id_priv->av,
1531                                  cm_id_priv);
1532         if (ret)
1533                 goto out;
1534         if (param->alternate_path) {
1535                 ret = cm_init_av_by_path(param->alternate_path, NULL,
1536                                          &cm_id_priv->alt_av, cm_id_priv);
1537                 if (ret)
1538                         goto out;
1539         }
1540         cm_id->service_id = param->service_id;
1541         cm_id->service_mask = ~cpu_to_be64(0);
1542         cm_id_priv->timeout_ms = cm_convert_to_ms(
1543                                     param->primary_path->packet_life_time) * 2 +
1544                                  cm_convert_to_ms(
1545                                     param->remote_cm_response_timeout);
1546         cm_id_priv->max_cm_retries = param->max_cm_retries;
1547         cm_id_priv->initiator_depth = param->initiator_depth;
1548         cm_id_priv->responder_resources = param->responder_resources;
1549         cm_id_priv->retry_count = param->retry_count;
1550         cm_id_priv->path_mtu = param->primary_path->mtu;
1551         cm_id_priv->pkey = param->primary_path->pkey;
1552         cm_id_priv->qp_type = param->qp_type;
1553
1554         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1555         if (ret)
1556                 goto out;
1557
1558         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1559         cm_format_req(req_msg, cm_id_priv, param);
1560         cm_id_priv->tid = req_msg->hdr.tid;
1561         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1562         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1563
1564         cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
1565         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
1566
1567         trace_icm_send_req(&cm_id_priv->id);
1568         spin_lock_irqsave(&cm_id_priv->lock, flags);
1569         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1570         if (ret) {
1571                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1572                 goto error2;
1573         }
1574         BUG_ON(cm_id->state != IB_CM_IDLE);
1575         cm_id->state = IB_CM_REQ_SENT;
1576         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1577         return 0;
1578
1579 error2: cm_free_msg(cm_id_priv->msg);
1580 out:    return ret;
1581 }
1582 EXPORT_SYMBOL(ib_send_cm_req);
1583
1584 static int cm_issue_rej(struct cm_port *port,
1585                         struct ib_mad_recv_wc *mad_recv_wc,
1586                         enum ib_cm_rej_reason reason,
1587                         enum cm_msg_response msg_rejected,
1588                         void *ari, u8 ari_length)
1589 {
1590         struct ib_mad_send_buf *msg = NULL;
1591         struct cm_rej_msg *rej_msg, *rcv_msg;
1592         int ret;
1593
1594         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1595         if (ret)
1596                 return ret;
1597
1598         /* We just need common CM header information.  Cast to any message. */
1599         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1600         rej_msg = (struct cm_rej_msg *) msg->mad;
1601
1602         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1603         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1604                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg));
1605         IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1606                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1607         IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected);
1608         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1609
1610         if (ari && ari_length) {
1611                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1612                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1613         }
1614
1615         trace_icm_issue_rej(
1616                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg),
1617                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1618         ret = ib_post_send_mad(msg, NULL);
1619         if (ret)
1620                 cm_free_msg(msg);
1621
1622         return ret;
1623 }
1624
1625 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1626 {
1627         return ((cpu_to_be16(
1628                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) ||
1629                 (ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1630                                                req_msg))));
1631 }
1632
1633 static void cm_path_set_rec_type(struct ib_device *ib_device, u8 port_num,
1634                                  struct sa_path_rec *path, union ib_gid *gid)
1635 {
1636         if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1637                 path->rec_type = SA_PATH_REC_TYPE_OPA;
1638         else
1639                 path->rec_type = SA_PATH_REC_TYPE_IB;
1640 }
1641
1642 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1643                                         struct sa_path_rec *primary_path,
1644                                         struct sa_path_rec *alt_path)
1645 {
1646         u32 lid;
1647
1648         if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1649                 sa_path_set_dlid(primary_path,
1650                                  IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
1651                                          req_msg));
1652                 sa_path_set_slid(primary_path,
1653                                  IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
1654                                          req_msg));
1655         } else {
1656                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1657                         CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg));
1658                 sa_path_set_dlid(primary_path, lid);
1659
1660                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1661                         CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg));
1662                 sa_path_set_slid(primary_path, lid);
1663         }
1664
1665         if (!cm_req_has_alt_path(req_msg))
1666                 return;
1667
1668         if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1669                 sa_path_set_dlid(alt_path,
1670                                  IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
1671                                          req_msg));
1672                 sa_path_set_slid(alt_path,
1673                                  IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
1674                                          req_msg));
1675         } else {
1676                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1677                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg));
1678                 sa_path_set_dlid(alt_path, lid);
1679
1680                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1681                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg));
1682                 sa_path_set_slid(alt_path, lid);
1683         }
1684 }
1685
1686 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1687                                      struct sa_path_rec *primary_path,
1688                                      struct sa_path_rec *alt_path)
1689 {
1690         primary_path->dgid =
1691                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg);
1692         primary_path->sgid =
1693                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg);
1694         primary_path->flow_label =
1695                 cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg));
1696         primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg);
1697         primary_path->traffic_class =
1698                 IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg);
1699         primary_path->reversible = 1;
1700         primary_path->pkey =
1701                 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1702         primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg);
1703         primary_path->mtu_selector = IB_SA_EQ;
1704         primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1705         primary_path->rate_selector = IB_SA_EQ;
1706         primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg);
1707         primary_path->packet_life_time_selector = IB_SA_EQ;
1708         primary_path->packet_life_time =
1709                 IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg);
1710         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1711         primary_path->service_id =
1712                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1713         if (sa_path_is_roce(primary_path))
1714                 primary_path->roce.route_resolved = false;
1715
1716         if (cm_req_has_alt_path(req_msg)) {
1717                 alt_path->dgid = *IBA_GET_MEM_PTR(
1718                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg);
1719                 alt_path->sgid = *IBA_GET_MEM_PTR(
1720                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg);
1721                 alt_path->flow_label = cpu_to_be32(
1722                         IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg));
1723                 alt_path->hop_limit =
1724                         IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg);
1725                 alt_path->traffic_class =
1726                         IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg);
1727                 alt_path->reversible = 1;
1728                 alt_path->pkey =
1729                         cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1730                 alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg);
1731                 alt_path->mtu_selector = IB_SA_EQ;
1732                 alt_path->mtu =
1733                         IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1734                 alt_path->rate_selector = IB_SA_EQ;
1735                 alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg);
1736                 alt_path->packet_life_time_selector = IB_SA_EQ;
1737                 alt_path->packet_life_time =
1738                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg);
1739                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1740                 alt_path->service_id =
1741                         cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1742
1743                 if (sa_path_is_roce(alt_path))
1744                         alt_path->roce.route_resolved = false;
1745         }
1746         cm_format_path_lid_from_req(req_msg, primary_path, alt_path);
1747 }
1748
1749 static u16 cm_get_bth_pkey(struct cm_work *work)
1750 {
1751         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1752         u8 port_num = work->port->port_num;
1753         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1754         u16 pkey;
1755         int ret;
1756
1757         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1758         if (ret) {
1759                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1760                                      port_num, pkey_index, ret);
1761                 return 0;
1762         }
1763
1764         return pkey;
1765 }
1766
1767 /**
1768  * Convert OPA SGID to IB SGID
1769  * ULPs (such as IPoIB) do not understand OPA GIDs and will
1770  * reject them as the local_gid will not match the sgid. Therefore,
1771  * change the pathrec's SGID to an IB SGID.
1772  *
1773  * @work: Work completion
1774  * @path: Path record
1775  */
1776 static void cm_opa_to_ib_sgid(struct cm_work *work,
1777                               struct sa_path_rec *path)
1778 {
1779         struct ib_device *dev = work->port->cm_dev->ib_device;
1780         u8 port_num = work->port->port_num;
1781
1782         if (rdma_cap_opa_ah(dev, port_num) &&
1783             (ib_is_opa_gid(&path->sgid))) {
1784                 union ib_gid sgid;
1785
1786                 if (rdma_query_gid(dev, port_num, 0, &sgid)) {
1787                         dev_warn(&dev->dev,
1788                                  "Error updating sgid in CM request\n");
1789                         return;
1790                 }
1791
1792                 path->sgid = sgid;
1793         }
1794 }
1795
1796 static void cm_format_req_event(struct cm_work *work,
1797                                 struct cm_id_private *cm_id_priv,
1798                                 struct ib_cm_id *listen_id)
1799 {
1800         struct cm_req_msg *req_msg;
1801         struct ib_cm_req_event_param *param;
1802
1803         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1804         param = &work->cm_event.param.req_rcvd;
1805         param->listen_id = listen_id;
1806         param->bth_pkey = cm_get_bth_pkey(work);
1807         param->port = cm_id_priv->av.port->port_num;
1808         param->primary_path = &work->path[0];
1809         cm_opa_to_ib_sgid(work, param->primary_path);
1810         if (cm_req_has_alt_path(req_msg)) {
1811                 param->alternate_path = &work->path[1];
1812                 cm_opa_to_ib_sgid(work, param->alternate_path);
1813         } else {
1814                 param->alternate_path = NULL;
1815         }
1816         param->remote_ca_guid =
1817                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
1818         param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg);
1819         param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg);
1820         param->qp_type = cm_req_get_qp_type(req_msg);
1821         param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg);
1822         param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
1823         param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
1824         param->local_cm_response_timeout =
1825                 IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg);
1826         param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg);
1827         param->remote_cm_response_timeout =
1828                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg);
1829         param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
1830         param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
1831         param->srq = IBA_GET(CM_REQ_SRQ, req_msg);
1832         param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
1833         param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg);
1834         param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod);
1835
1836         work->cm_event.private_data =
1837                 IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg);
1838 }
1839
1840 static void cm_process_work(struct cm_id_private *cm_id_priv,
1841                             struct cm_work *work)
1842 {
1843         int ret;
1844
1845         /* We will typically only have the current event to report. */
1846         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1847         cm_free_work(work);
1848
1849         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1850                 spin_lock_irq(&cm_id_priv->lock);
1851                 work = cm_dequeue_work(cm_id_priv);
1852                 spin_unlock_irq(&cm_id_priv->lock);
1853                 if (!work)
1854                         return;
1855
1856                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1857                                                 &work->cm_event);
1858                 cm_free_work(work);
1859         }
1860         cm_deref_id(cm_id_priv);
1861         if (ret)
1862                 cm_destroy_id(&cm_id_priv->id, ret);
1863 }
1864
1865 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1866                           struct cm_id_private *cm_id_priv,
1867                           enum cm_msg_response msg_mraed, u8 service_timeout,
1868                           const void *private_data, u8 private_data_len)
1869 {
1870         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1871         IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed);
1872         IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg,
1873                 be32_to_cpu(cm_id_priv->id.local_id));
1874         IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg,
1875                 be32_to_cpu(cm_id_priv->id.remote_id));
1876         IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, service_timeout);
1877
1878         if (private_data && private_data_len)
1879                 IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data,
1880                             private_data_len);
1881 }
1882
1883 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1884                           struct cm_id_private *cm_id_priv,
1885                           enum ib_cm_rej_reason reason, void *ari,
1886                           u8 ari_length, const void *private_data,
1887                           u8 private_data_len, enum ib_cm_state state)
1888 {
1889         lockdep_assert_held(&cm_id_priv->lock);
1890
1891         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1892         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1893                 be32_to_cpu(cm_id_priv->id.remote_id));
1894
1895         switch (state) {
1896         case IB_CM_REQ_RCVD:
1897                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0));
1898                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1899                 break;
1900         case IB_CM_MRA_REQ_SENT:
1901                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1902                         be32_to_cpu(cm_id_priv->id.local_id));
1903                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1904                 break;
1905         case IB_CM_REP_RCVD:
1906         case IB_CM_MRA_REP_SENT:
1907                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1908                         be32_to_cpu(cm_id_priv->id.local_id));
1909                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP);
1910                 break;
1911         default:
1912                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1913                         be32_to_cpu(cm_id_priv->id.local_id));
1914                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg,
1915                         CM_MSG_RESPONSE_OTHER);
1916                 break;
1917         }
1918
1919         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1920         if (ari && ari_length) {
1921                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1922                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1923         }
1924
1925         if (private_data && private_data_len)
1926                 IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data,
1927                             private_data_len);
1928 }
1929
1930 static void cm_dup_req_handler(struct cm_work *work,
1931                                struct cm_id_private *cm_id_priv)
1932 {
1933         struct ib_mad_send_buf *msg = NULL;
1934         int ret;
1935
1936         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1937                         counter[CM_REQ_COUNTER]);
1938
1939         /* Quick state check to discard duplicate REQs. */
1940         spin_lock_irq(&cm_id_priv->lock);
1941         if (cm_id_priv->id.state == IB_CM_REQ_RCVD) {
1942                 spin_unlock_irq(&cm_id_priv->lock);
1943                 return;
1944         }
1945         spin_unlock_irq(&cm_id_priv->lock);
1946
1947         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1948         if (ret)
1949                 return;
1950
1951         spin_lock_irq(&cm_id_priv->lock);
1952         switch (cm_id_priv->id.state) {
1953         case IB_CM_MRA_REQ_SENT:
1954                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1955                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1956                               cm_id_priv->private_data,
1957                               cm_id_priv->private_data_len);
1958                 break;
1959         case IB_CM_TIMEWAIT:
1960                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv,
1961                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0,
1962                               IB_CM_TIMEWAIT);
1963                 break;
1964         default:
1965                 goto unlock;
1966         }
1967         spin_unlock_irq(&cm_id_priv->lock);
1968
1969         trace_icm_send_dup_req(&cm_id_priv->id);
1970         ret = ib_post_send_mad(msg, NULL);
1971         if (ret)
1972                 goto free;
1973         return;
1974
1975 unlock: spin_unlock_irq(&cm_id_priv->lock);
1976 free:   cm_free_msg(msg);
1977 }
1978
1979 static struct cm_id_private * cm_match_req(struct cm_work *work,
1980                                            struct cm_id_private *cm_id_priv)
1981 {
1982         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1983         struct cm_timewait_info *timewait_info;
1984         struct cm_req_msg *req_msg;
1985
1986         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1987
1988         /* Check for possible duplicate REQ. */
1989         spin_lock_irq(&cm.lock);
1990         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1991         if (timewait_info) {
1992                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
1993                                            timewait_info->work.remote_id);
1994                 spin_unlock_irq(&cm.lock);
1995                 if (cur_cm_id_priv) {
1996                         cm_dup_req_handler(work, cur_cm_id_priv);
1997                         cm_deref_id(cur_cm_id_priv);
1998                 }
1999                 return NULL;
2000         }
2001
2002         /* Check for stale connections. */
2003         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2004         if (timewait_info) {
2005                 cm_remove_remote(cm_id_priv);
2006                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2007                                            timewait_info->work.remote_id);
2008
2009                 spin_unlock_irq(&cm.lock);
2010                 cm_issue_rej(work->port, work->mad_recv_wc,
2011                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
2012                              NULL, 0);
2013                 if (cur_cm_id_priv) {
2014                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2015                         cm_deref_id(cur_cm_id_priv);
2016                 }
2017                 return NULL;
2018         }
2019
2020         /* Find matching listen request. */
2021         listen_cm_id_priv = cm_find_listen(
2022                 cm_id_priv->id.device,
2023                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)));
2024         if (!listen_cm_id_priv) {
2025                 cm_remove_remote(cm_id_priv);
2026                 spin_unlock_irq(&cm.lock);
2027                 cm_issue_rej(work->port, work->mad_recv_wc,
2028                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
2029                              NULL, 0);
2030                 return NULL;
2031         }
2032         spin_unlock_irq(&cm.lock);
2033         return listen_cm_id_priv;
2034 }
2035
2036 /*
2037  * Work-around for inter-subnet connections.  If the LIDs are permissive,
2038  * we need to override the LID/SL data in the REQ with the LID information
2039  * in the work completion.
2040  */
2041 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
2042 {
2043         if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) {
2044                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
2045                                         req_msg)) == IB_LID_PERMISSIVE) {
2046                         IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
2047                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2048                         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl);
2049                 }
2050
2051                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
2052                                         req_msg)) == IB_LID_PERMISSIVE)
2053                         IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
2054                                 wc->dlid_path_bits);
2055         }
2056
2057         if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) {
2058                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
2059                                         req_msg)) == IB_LID_PERMISSIVE) {
2060                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
2061                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2062                         IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl);
2063                 }
2064
2065                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
2066                                         req_msg)) == IB_LID_PERMISSIVE)
2067                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
2068                                 wc->dlid_path_bits);
2069         }
2070 }
2071
2072 static int cm_req_handler(struct cm_work *work)
2073 {
2074         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
2075         struct cm_req_msg *req_msg;
2076         const struct ib_global_route *grh;
2077         const struct ib_gid_attr *gid_attr;
2078         int ret;
2079
2080         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
2081
2082         cm_id_priv =
2083                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
2084         if (IS_ERR(cm_id_priv))
2085                 return PTR_ERR(cm_id_priv);
2086
2087         cm_id_priv->id.remote_id =
2088                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg));
2089         cm_id_priv->id.service_id =
2090                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
2091         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
2092         cm_id_priv->tid = req_msg->hdr.tid;
2093         cm_id_priv->timeout_ms = cm_convert_to_ms(
2094                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg));
2095         cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg);
2096         cm_id_priv->remote_qpn =
2097                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
2098         cm_id_priv->initiator_depth =
2099                 IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
2100         cm_id_priv->responder_resources =
2101                 IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
2102         cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
2103         cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
2104         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
2105         cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
2106         cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
2107         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
2108
2109         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2110                                       work->mad_recv_wc->recv_buf.grh,
2111                                       &cm_id_priv->av);
2112         if (ret)
2113                 goto destroy;
2114         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
2115                                                             id.local_id);
2116         if (IS_ERR(cm_id_priv->timewait_info)) {
2117                 ret = PTR_ERR(cm_id_priv->timewait_info);
2118                 cm_id_priv->timewait_info = NULL;
2119                 goto destroy;
2120         }
2121         cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id;
2122         cm_id_priv->timewait_info->remote_ca_guid =
2123                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
2124         cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn;
2125
2126         /*
2127          * Note that the ID pointer is not in the xarray at this point,
2128          * so this set is only visible to the local thread.
2129          */
2130         cm_id_priv->id.state = IB_CM_REQ_RCVD;
2131
2132         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
2133         if (!listen_cm_id_priv) {
2134                 trace_icm_no_listener_err(&cm_id_priv->id);
2135                 cm_id_priv->id.state = IB_CM_IDLE;
2136                 ret = -EINVAL;
2137                 goto destroy;
2138         }
2139
2140         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
2141
2142         memset(&work->path[0], 0, sizeof(work->path[0]));
2143         if (cm_req_has_alt_path(req_msg))
2144                 memset(&work->path[1], 0, sizeof(work->path[1]));
2145         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
2146         gid_attr = grh->sgid_attr;
2147
2148         if (gid_attr &&
2149             rdma_protocol_roce(work->port->cm_dev->ib_device,
2150                                work->port->port_num)) {
2151                 work->path[0].rec_type =
2152                         sa_conv_gid_to_pathrec_type(gid_attr->gid_type);
2153         } else {
2154                 cm_path_set_rec_type(
2155                         work->port->cm_dev->ib_device, work->port->port_num,
2156                         &work->path[0],
2157                         IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID,
2158                                         req_msg));
2159         }
2160         if (cm_req_has_alt_path(req_msg))
2161                 work->path[1].rec_type = work->path[0].rec_type;
2162         cm_format_paths_from_req(req_msg, &work->path[0],
2163                                  &work->path[1]);
2164         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
2165                 sa_path_set_dmac(&work->path[0],
2166                                  cm_id_priv->av.ah_attr.roce.dmac);
2167         work->path[0].hop_limit = grh->hop_limit;
2168         ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av,
2169                                  cm_id_priv);
2170         if (ret) {
2171                 int err;
2172
2173                 err = rdma_query_gid(work->port->cm_dev->ib_device,
2174                                      work->port->port_num, 0,
2175                                      &work->path[0].sgid);
2176                 if (err)
2177                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2178                                        NULL, 0, NULL, 0);
2179                 else
2180                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2181                                        &work->path[0].sgid,
2182                                        sizeof(work->path[0].sgid),
2183                                        NULL, 0);
2184                 goto rejected;
2185         }
2186         if (cm_req_has_alt_path(req_msg)) {
2187                 ret = cm_init_av_by_path(&work->path[1], NULL,
2188                                          &cm_id_priv->alt_av, cm_id_priv);
2189                 if (ret) {
2190                         ib_send_cm_rej(&cm_id_priv->id,
2191                                        IB_CM_REJ_INVALID_ALT_GID,
2192                                        &work->path[0].sgid,
2193                                        sizeof(work->path[0].sgid), NULL, 0);
2194                         goto rejected;
2195                 }
2196         }
2197
2198         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
2199         cm_id_priv->id.context = listen_cm_id_priv->id.context;
2200         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
2201
2202         /* Now MAD handlers can see the new ID */
2203         spin_lock_irq(&cm_id_priv->lock);
2204         cm_finalize_id(cm_id_priv);
2205
2206         /* Refcount belongs to the event, pairs with cm_process_work() */
2207         refcount_inc(&cm_id_priv->refcount);
2208         cm_queue_work_unlock(cm_id_priv, work);
2209         /*
2210          * Since this ID was just created and was not made visible to other MAD
2211          * handlers until the cm_finalize_id() above we know that the
2212          * cm_process_work() will deliver the event and the listen_cm_id
2213          * embedded in the event can be derefed here.
2214          */
2215         cm_deref_id(listen_cm_id_priv);
2216         return 0;
2217
2218 rejected:
2219         cm_deref_id(listen_cm_id_priv);
2220 destroy:
2221         ib_destroy_cm_id(&cm_id_priv->id);
2222         return ret;
2223 }
2224
2225 static void cm_format_rep(struct cm_rep_msg *rep_msg,
2226                           struct cm_id_private *cm_id_priv,
2227                           struct ib_cm_rep_param *param)
2228 {
2229         cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid,
2230                               param->ece.attr_mod);
2231         IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg,
2232                 be32_to_cpu(cm_id_priv->id.local_id));
2233         IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg,
2234                 be32_to_cpu(cm_id_priv->id.remote_id));
2235         IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn);
2236         IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg,
2237                 param->responder_resources);
2238         IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg,
2239                 cm_id_priv->av.port->cm_dev->ack_delay);
2240         IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted);
2241         IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count);
2242         IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg,
2243                 be64_to_cpu(cm_id_priv->id.device->node_guid));
2244
2245         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
2246                 IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg,
2247                         param->initiator_depth);
2248                 IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg,
2249                         param->flow_control);
2250                 IBA_SET(CM_REP_SRQ, rep_msg, param->srq);
2251                 IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num);
2252         } else {
2253                 IBA_SET(CM_REP_SRQ, rep_msg, 1);
2254                 IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num);
2255         }
2256
2257         IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id);
2258         IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8);
2259         IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16);
2260
2261         if (param->private_data && param->private_data_len)
2262                 IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data,
2263                             param->private_data_len);
2264 }
2265
2266 int ib_send_cm_rep(struct ib_cm_id *cm_id,
2267                    struct ib_cm_rep_param *param)
2268 {
2269         struct cm_id_private *cm_id_priv;
2270         struct ib_mad_send_buf *msg;
2271         struct cm_rep_msg *rep_msg;
2272         unsigned long flags;
2273         int ret;
2274
2275         if (param->private_data &&
2276             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
2277                 return -EINVAL;
2278
2279         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2280         spin_lock_irqsave(&cm_id_priv->lock, flags);
2281         if (cm_id->state != IB_CM_REQ_RCVD &&
2282             cm_id->state != IB_CM_MRA_REQ_SENT) {
2283                 trace_icm_send_rep_err(cm_id_priv->id.local_id, cm_id->state);
2284                 ret = -EINVAL;
2285                 goto out;
2286         }
2287
2288         ret = cm_alloc_msg(cm_id_priv, &msg);
2289         if (ret)
2290                 goto out;
2291
2292         rep_msg = (struct cm_rep_msg *) msg->mad;
2293         cm_format_rep(rep_msg, cm_id_priv, param);
2294         msg->timeout_ms = cm_id_priv->timeout_ms;
2295         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
2296
2297         trace_icm_send_rep(cm_id);
2298         ret = ib_post_send_mad(msg, NULL);
2299         if (ret) {
2300                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2301                 cm_free_msg(msg);
2302                 return ret;
2303         }
2304
2305         cm_id->state = IB_CM_REP_SENT;
2306         cm_id_priv->msg = msg;
2307         cm_id_priv->initiator_depth = param->initiator_depth;
2308         cm_id_priv->responder_resources = param->responder_resources;
2309         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2310         WARN_ONCE(param->qp_num & 0xFF000000,
2311                   "IBTA declares QPN to be 24 bits, but it is 0x%X\n",
2312                   param->qp_num);
2313         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2314
2315 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2316         return ret;
2317 }
2318 EXPORT_SYMBOL(ib_send_cm_rep);
2319
2320 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2321                           struct cm_id_private *cm_id_priv,
2322                           const void *private_data,
2323                           u8 private_data_len)
2324 {
2325         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2326         IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg,
2327                 be32_to_cpu(cm_id_priv->id.local_id));
2328         IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg,
2329                 be32_to_cpu(cm_id_priv->id.remote_id));
2330
2331         if (private_data && private_data_len)
2332                 IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data,
2333                             private_data_len);
2334 }
2335
2336 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2337                    const void *private_data,
2338                    u8 private_data_len)
2339 {
2340         struct cm_id_private *cm_id_priv;
2341         struct ib_mad_send_buf *msg;
2342         unsigned long flags;
2343         void *data;
2344         int ret;
2345
2346         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2347                 return -EINVAL;
2348
2349         data = cm_copy_private_data(private_data, private_data_len);
2350         if (IS_ERR(data))
2351                 return PTR_ERR(data);
2352
2353         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2354         spin_lock_irqsave(&cm_id_priv->lock, flags);
2355         if (cm_id->state != IB_CM_REP_RCVD &&
2356             cm_id->state != IB_CM_MRA_REP_SENT) {
2357                 trace_icm_send_cm_rtu_err(cm_id);
2358                 ret = -EINVAL;
2359                 goto error;
2360         }
2361
2362         ret = cm_alloc_msg(cm_id_priv, &msg);
2363         if (ret)
2364                 goto error;
2365
2366         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2367                       private_data, private_data_len);
2368
2369         trace_icm_send_rtu(cm_id);
2370         ret = ib_post_send_mad(msg, NULL);
2371         if (ret) {
2372                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2373                 cm_free_msg(msg);
2374                 kfree(data);
2375                 return ret;
2376         }
2377
2378         cm_id->state = IB_CM_ESTABLISHED;
2379         cm_set_private_data(cm_id_priv, data, private_data_len);
2380         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2381         return 0;
2382
2383 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2384         kfree(data);
2385         return ret;
2386 }
2387 EXPORT_SYMBOL(ib_send_cm_rtu);
2388
2389 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2390 {
2391         struct cm_rep_msg *rep_msg;
2392         struct ib_cm_rep_event_param *param;
2393
2394         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2395         param = &work->cm_event.param.rep_rcvd;
2396         param->remote_ca_guid =
2397                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2398         param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg);
2399         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2400         param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg);
2401         param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2402         param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2403         param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2404         param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg);
2405         param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg);
2406         param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2407         param->srq = IBA_GET(CM_REP_SRQ, rep_msg);
2408         param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16;
2409         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8;
2410         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg);
2411         param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod);
2412
2413         work->cm_event.private_data =
2414                 IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg);
2415 }
2416
2417 static void cm_dup_rep_handler(struct cm_work *work)
2418 {
2419         struct cm_id_private *cm_id_priv;
2420         struct cm_rep_msg *rep_msg;
2421         struct ib_mad_send_buf *msg = NULL;
2422         int ret;
2423
2424         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2425         cm_id_priv = cm_acquire_id(
2426                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)),
2427                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)));
2428         if (!cm_id_priv)
2429                 return;
2430
2431         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2432                         counter[CM_REP_COUNTER]);
2433         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2434         if (ret)
2435                 goto deref;
2436
2437         spin_lock_irq(&cm_id_priv->lock);
2438         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2439                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2440                               cm_id_priv->private_data,
2441                               cm_id_priv->private_data_len);
2442         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2443                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2444                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2445                               cm_id_priv->private_data,
2446                               cm_id_priv->private_data_len);
2447         else
2448                 goto unlock;
2449         spin_unlock_irq(&cm_id_priv->lock);
2450
2451         trace_icm_send_dup_rep(&cm_id_priv->id);
2452         ret = ib_post_send_mad(msg, NULL);
2453         if (ret)
2454                 goto free;
2455         goto deref;
2456
2457 unlock: spin_unlock_irq(&cm_id_priv->lock);
2458 free:   cm_free_msg(msg);
2459 deref:  cm_deref_id(cm_id_priv);
2460 }
2461
2462 static int cm_rep_handler(struct cm_work *work)
2463 {
2464         struct cm_id_private *cm_id_priv;
2465         struct cm_rep_msg *rep_msg;
2466         int ret;
2467         struct cm_id_private *cur_cm_id_priv;
2468         struct cm_timewait_info *timewait_info;
2469
2470         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2471         cm_id_priv = cm_acquire_id(
2472                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0);
2473         if (!cm_id_priv) {
2474                 cm_dup_rep_handler(work);
2475                 trace_icm_remote_no_priv_err(
2476                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2477                 return -EINVAL;
2478         }
2479
2480         cm_format_rep_event(work, cm_id_priv->qp_type);
2481
2482         spin_lock_irq(&cm_id_priv->lock);
2483         switch (cm_id_priv->id.state) {
2484         case IB_CM_REQ_SENT:
2485         case IB_CM_MRA_REQ_RCVD:
2486                 break;
2487         default:
2488                 ret = -EINVAL;
2489                 trace_icm_rep_unknown_err(
2490                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2491                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg),
2492                         cm_id_priv->id.state);
2493                 spin_unlock_irq(&cm_id_priv->lock);
2494                 goto error;
2495         }
2496
2497         cm_id_priv->timewait_info->work.remote_id =
2498                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2499         cm_id_priv->timewait_info->remote_ca_guid =
2500                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2501         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2502
2503         spin_lock(&cm.lock);
2504         /* Check for duplicate REP. */
2505         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2506                 spin_unlock(&cm.lock);
2507                 spin_unlock_irq(&cm_id_priv->lock);
2508                 ret = -EINVAL;
2509                 trace_icm_insert_failed_err(
2510                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2511                 goto error;
2512         }
2513         /* Check for a stale connection. */
2514         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2515         if (timewait_info) {
2516                 cm_remove_remote(cm_id_priv);
2517                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2518                                            timewait_info->work.remote_id);
2519
2520                 spin_unlock(&cm.lock);
2521                 spin_unlock_irq(&cm_id_priv->lock);
2522                 cm_issue_rej(work->port, work->mad_recv_wc,
2523                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2524                              NULL, 0);
2525                 ret = -EINVAL;
2526                 trace_icm_staleconn_err(
2527                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2528                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2529
2530                 if (cur_cm_id_priv) {
2531                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2532                         cm_deref_id(cur_cm_id_priv);
2533                 }
2534
2535                 goto error;
2536         }
2537         spin_unlock(&cm.lock);
2538
2539         cm_id_priv->id.state = IB_CM_REP_RCVD;
2540         cm_id_priv->id.remote_id =
2541                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2542         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2543         cm_id_priv->initiator_depth =
2544                 IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2545         cm_id_priv->responder_resources =
2546                 IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2547         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2548         cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2549         cm_id_priv->target_ack_delay =
2550                 IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2551         cm_id_priv->av.timeout =
2552                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2553                                        cm_id_priv->av.timeout - 1);
2554         cm_id_priv->alt_av.timeout =
2555                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2556                                        cm_id_priv->alt_av.timeout - 1);
2557
2558         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2559         cm_queue_work_unlock(cm_id_priv, work);
2560         return 0;
2561
2562 error:
2563         cm_deref_id(cm_id_priv);
2564         return ret;
2565 }
2566
2567 static int cm_establish_handler(struct cm_work *work)
2568 {
2569         struct cm_id_private *cm_id_priv;
2570
2571         /* See comment in cm_establish about lookup. */
2572         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2573         if (!cm_id_priv)
2574                 return -EINVAL;
2575
2576         spin_lock_irq(&cm_id_priv->lock);
2577         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2578                 spin_unlock_irq(&cm_id_priv->lock);
2579                 goto out;
2580         }
2581
2582         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2583         cm_queue_work_unlock(cm_id_priv, work);
2584         return 0;
2585 out:
2586         cm_deref_id(cm_id_priv);
2587         return -EINVAL;
2588 }
2589
2590 static int cm_rtu_handler(struct cm_work *work)
2591 {
2592         struct cm_id_private *cm_id_priv;
2593         struct cm_rtu_msg *rtu_msg;
2594
2595         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2596         cm_id_priv = cm_acquire_id(
2597                 cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)),
2598                 cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg)));
2599         if (!cm_id_priv)
2600                 return -EINVAL;
2601
2602         work->cm_event.private_data =
2603                 IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg);
2604
2605         spin_lock_irq(&cm_id_priv->lock);
2606         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2607             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2608                 spin_unlock_irq(&cm_id_priv->lock);
2609                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2610                                 counter[CM_RTU_COUNTER]);
2611                 goto out;
2612         }
2613         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2614
2615         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2616         cm_queue_work_unlock(cm_id_priv, work);
2617         return 0;
2618 out:
2619         cm_deref_id(cm_id_priv);
2620         return -EINVAL;
2621 }
2622
2623 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2624                           struct cm_id_private *cm_id_priv,
2625                           const void *private_data,
2626                           u8 private_data_len)
2627 {
2628         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2629                           cm_form_tid(cm_id_priv));
2630         IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg,
2631                 be32_to_cpu(cm_id_priv->id.local_id));
2632         IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg,
2633                 be32_to_cpu(cm_id_priv->id.remote_id));
2634         IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg,
2635                 be32_to_cpu(cm_id_priv->remote_qpn));
2636
2637         if (private_data && private_data_len)
2638                 IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data,
2639                             private_data_len);
2640 }
2641
2642 static int cm_send_dreq_locked(struct cm_id_private *cm_id_priv,
2643                                const void *private_data, u8 private_data_len)
2644 {
2645         struct ib_mad_send_buf *msg;
2646         int ret;
2647
2648         lockdep_assert_held(&cm_id_priv->lock);
2649
2650         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2651                 return -EINVAL;
2652
2653         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2654                 trace_icm_dreq_skipped(&cm_id_priv->id);
2655                 return -EINVAL;
2656         }
2657
2658         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2659             cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2660                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2661
2662         ret = cm_alloc_msg(cm_id_priv, &msg);
2663         if (ret) {
2664                 cm_enter_timewait(cm_id_priv);
2665                 return ret;
2666         }
2667
2668         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2669                        private_data, private_data_len);
2670         msg->timeout_ms = cm_id_priv->timeout_ms;
2671         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2672
2673         trace_icm_send_dreq(&cm_id_priv->id);
2674         ret = ib_post_send_mad(msg, NULL);
2675         if (ret) {
2676                 cm_enter_timewait(cm_id_priv);
2677                 cm_free_msg(msg);
2678                 return ret;
2679         }
2680
2681         cm_id_priv->id.state = IB_CM_DREQ_SENT;
2682         cm_id_priv->msg = msg;
2683         return 0;
2684 }
2685
2686 int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data,
2687                     u8 private_data_len)
2688 {
2689         struct cm_id_private *cm_id_priv =
2690                 container_of(cm_id, struct cm_id_private, id);
2691         unsigned long flags;
2692         int ret;
2693
2694         spin_lock_irqsave(&cm_id_priv->lock, flags);
2695         ret = cm_send_dreq_locked(cm_id_priv, private_data, private_data_len);
2696         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2697         return ret;
2698 }
2699 EXPORT_SYMBOL(ib_send_cm_dreq);
2700
2701 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2702                           struct cm_id_private *cm_id_priv,
2703                           const void *private_data,
2704                           u8 private_data_len)
2705 {
2706         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2707         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2708                 be32_to_cpu(cm_id_priv->id.local_id));
2709         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2710                 be32_to_cpu(cm_id_priv->id.remote_id));
2711
2712         if (private_data && private_data_len)
2713                 IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data,
2714                             private_data_len);
2715 }
2716
2717 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
2718                                void *private_data, u8 private_data_len)
2719 {
2720         struct ib_mad_send_buf *msg;
2721         int ret;
2722
2723         lockdep_assert_held(&cm_id_priv->lock);
2724
2725         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2726                 return -EINVAL;
2727
2728         if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2729                 trace_icm_send_drep_err(&cm_id_priv->id);
2730                 kfree(private_data);
2731                 return -EINVAL;
2732         }
2733
2734         cm_set_private_data(cm_id_priv, private_data, private_data_len);
2735         cm_enter_timewait(cm_id_priv);
2736
2737         ret = cm_alloc_msg(cm_id_priv, &msg);
2738         if (ret)
2739                 return ret;
2740
2741         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2742                        private_data, private_data_len);
2743
2744         trace_icm_send_drep(&cm_id_priv->id);
2745         ret = ib_post_send_mad(msg, NULL);
2746         if (ret) {
2747                 cm_free_msg(msg);
2748                 return ret;
2749         }
2750         return 0;
2751 }
2752
2753 int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data,
2754                     u8 private_data_len)
2755 {
2756         struct cm_id_private *cm_id_priv =
2757                 container_of(cm_id, struct cm_id_private, id);
2758         unsigned long flags;
2759         void *data;
2760         int ret;
2761
2762         data = cm_copy_private_data(private_data, private_data_len);
2763         if (IS_ERR(data))
2764                 return PTR_ERR(data);
2765
2766         spin_lock_irqsave(&cm_id_priv->lock, flags);
2767         ret = cm_send_drep_locked(cm_id_priv, data, private_data_len);
2768         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2769         return ret;
2770 }
2771 EXPORT_SYMBOL(ib_send_cm_drep);
2772
2773 static int cm_issue_drep(struct cm_port *port,
2774                          struct ib_mad_recv_wc *mad_recv_wc)
2775 {
2776         struct ib_mad_send_buf *msg = NULL;
2777         struct cm_dreq_msg *dreq_msg;
2778         struct cm_drep_msg *drep_msg;
2779         int ret;
2780
2781         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2782         if (ret)
2783                 return ret;
2784
2785         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2786         drep_msg = (struct cm_drep_msg *) msg->mad;
2787
2788         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2789         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2790                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg));
2791         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2792                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2793
2794         trace_icm_issue_drep(
2795                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2796                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2797         ret = ib_post_send_mad(msg, NULL);
2798         if (ret)
2799                 cm_free_msg(msg);
2800
2801         return ret;
2802 }
2803
2804 static int cm_dreq_handler(struct cm_work *work)
2805 {
2806         struct cm_id_private *cm_id_priv;
2807         struct cm_dreq_msg *dreq_msg;
2808         struct ib_mad_send_buf *msg = NULL;
2809
2810         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2811         cm_id_priv = cm_acquire_id(
2812                 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)),
2813                 cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)));
2814         if (!cm_id_priv) {
2815                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2816                                 counter[CM_DREQ_COUNTER]);
2817                 cm_issue_drep(work->port, work->mad_recv_wc);
2818                 trace_icm_no_priv_err(
2819                         IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2820                         IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2821                 return -EINVAL;
2822         }
2823
2824         work->cm_event.private_data =
2825                 IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg);
2826
2827         spin_lock_irq(&cm_id_priv->lock);
2828         if (cm_id_priv->local_qpn !=
2829             cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg)))
2830                 goto unlock;
2831
2832         switch (cm_id_priv->id.state) {
2833         case IB_CM_REP_SENT:
2834         case IB_CM_DREQ_SENT:
2835                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2836                 break;
2837         case IB_CM_ESTABLISHED:
2838                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2839                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2840                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2841                 break;
2842         case IB_CM_MRA_REP_RCVD:
2843                 break;
2844         case IB_CM_TIMEWAIT:
2845                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2846                                 counter[CM_DREQ_COUNTER]);
2847                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
2848                 if (IS_ERR(msg))
2849                         goto unlock;
2850
2851                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2852                                cm_id_priv->private_data,
2853                                cm_id_priv->private_data_len);
2854                 spin_unlock_irq(&cm_id_priv->lock);
2855
2856                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2857                     ib_post_send_mad(msg, NULL))
2858                         cm_free_msg(msg);
2859                 goto deref;
2860         case IB_CM_DREQ_RCVD:
2861                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2862                                 counter[CM_DREQ_COUNTER]);
2863                 goto unlock;
2864         default:
2865                 trace_icm_dreq_unknown_err(&cm_id_priv->id);
2866                 goto unlock;
2867         }
2868         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2869         cm_id_priv->tid = dreq_msg->hdr.tid;
2870         cm_queue_work_unlock(cm_id_priv, work);
2871         return 0;
2872
2873 unlock: spin_unlock_irq(&cm_id_priv->lock);
2874 deref:  cm_deref_id(cm_id_priv);
2875         return -EINVAL;
2876 }
2877
2878 static int cm_drep_handler(struct cm_work *work)
2879 {
2880         struct cm_id_private *cm_id_priv;
2881         struct cm_drep_msg *drep_msg;
2882
2883         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2884         cm_id_priv = cm_acquire_id(
2885                 cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)),
2886                 cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg)));
2887         if (!cm_id_priv)
2888                 return -EINVAL;
2889
2890         work->cm_event.private_data =
2891                 IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg);
2892
2893         spin_lock_irq(&cm_id_priv->lock);
2894         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2895             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2896                 spin_unlock_irq(&cm_id_priv->lock);
2897                 goto out;
2898         }
2899         cm_enter_timewait(cm_id_priv);
2900
2901         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2902         cm_queue_work_unlock(cm_id_priv, work);
2903         return 0;
2904 out:
2905         cm_deref_id(cm_id_priv);
2906         return -EINVAL;
2907 }
2908
2909 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
2910                               enum ib_cm_rej_reason reason, void *ari,
2911                               u8 ari_length, const void *private_data,
2912                               u8 private_data_len)
2913 {
2914         enum ib_cm_state state = cm_id_priv->id.state;
2915         struct ib_mad_send_buf *msg;
2916         int ret;
2917
2918         lockdep_assert_held(&cm_id_priv->lock);
2919
2920         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2921             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2922                 return -EINVAL;
2923
2924         switch (state) {
2925         case IB_CM_REQ_SENT:
2926         case IB_CM_MRA_REQ_RCVD:
2927         case IB_CM_REQ_RCVD:
2928         case IB_CM_MRA_REQ_SENT:
2929         case IB_CM_REP_RCVD:
2930         case IB_CM_MRA_REP_SENT:
2931                 cm_reset_to_idle(cm_id_priv);
2932                 ret = cm_alloc_msg(cm_id_priv, &msg);
2933                 if (ret)
2934                         return ret;
2935                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2936                               ari, ari_length, private_data, private_data_len,
2937                               state);
2938                 break;
2939         case IB_CM_REP_SENT:
2940         case IB_CM_MRA_REP_RCVD:
2941                 cm_enter_timewait(cm_id_priv);
2942                 ret = cm_alloc_msg(cm_id_priv, &msg);
2943                 if (ret)
2944                         return ret;
2945                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2946                               ari, ari_length, private_data, private_data_len,
2947                               state);
2948                 break;
2949         default:
2950                 trace_icm_send_unknown_rej_err(&cm_id_priv->id);
2951                 return -EINVAL;
2952         }
2953
2954         trace_icm_send_rej(&cm_id_priv->id, reason);
2955         ret = ib_post_send_mad(msg, NULL);
2956         if (ret) {
2957                 cm_free_msg(msg);
2958                 return ret;
2959         }
2960
2961         return 0;
2962 }
2963
2964 int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason,
2965                    void *ari, u8 ari_length, const void *private_data,
2966                    u8 private_data_len)
2967 {
2968         struct cm_id_private *cm_id_priv =
2969                 container_of(cm_id, struct cm_id_private, id);
2970         unsigned long flags;
2971         int ret;
2972
2973         spin_lock_irqsave(&cm_id_priv->lock, flags);
2974         ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length,
2975                                  private_data, private_data_len);
2976         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2977         return ret;
2978 }
2979 EXPORT_SYMBOL(ib_send_cm_rej);
2980
2981 static void cm_format_rej_event(struct cm_work *work)
2982 {
2983         struct cm_rej_msg *rej_msg;
2984         struct ib_cm_rej_event_param *param;
2985
2986         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2987         param = &work->cm_event.param.rej_rcvd;
2988         param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg);
2989         param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg);
2990         param->reason = IBA_GET(CM_REJ_REASON, rej_msg);
2991         work->cm_event.private_data =
2992                 IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg);
2993 }
2994
2995 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2996 {
2997         struct cm_id_private *cm_id_priv;
2998         __be32 remote_id;
2999
3000         remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg));
3001
3002         if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) {
3003                 cm_id_priv = cm_find_remote_id(
3004                         *((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)),
3005                         remote_id);
3006         } else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) ==
3007                    CM_MSG_RESPONSE_REQ)
3008                 cm_id_priv = cm_acquire_id(
3009                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3010                         0);
3011         else
3012                 cm_id_priv = cm_acquire_id(
3013                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3014                         remote_id);
3015
3016         return cm_id_priv;
3017 }
3018
3019 static int cm_rej_handler(struct cm_work *work)
3020 {
3021         struct cm_id_private *cm_id_priv;
3022         struct cm_rej_msg *rej_msg;
3023
3024         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
3025         cm_id_priv = cm_acquire_rejected_id(rej_msg);
3026         if (!cm_id_priv)
3027                 return -EINVAL;
3028
3029         cm_format_rej_event(work);
3030
3031         spin_lock_irq(&cm_id_priv->lock);
3032         switch (cm_id_priv->id.state) {
3033         case IB_CM_REQ_SENT:
3034         case IB_CM_MRA_REQ_RCVD:
3035         case IB_CM_REP_SENT:
3036         case IB_CM_MRA_REP_RCVD:
3037                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3038                 fallthrough;
3039         case IB_CM_REQ_RCVD:
3040         case IB_CM_MRA_REQ_SENT:
3041                 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN)
3042                         cm_enter_timewait(cm_id_priv);
3043                 else
3044                         cm_reset_to_idle(cm_id_priv);
3045                 break;
3046         case IB_CM_DREQ_SENT:
3047                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3048                 fallthrough;
3049         case IB_CM_REP_RCVD:
3050         case IB_CM_MRA_REP_SENT:
3051                 cm_enter_timewait(cm_id_priv);
3052                 break;
3053         case IB_CM_ESTABLISHED:
3054                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
3055                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
3056                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
3057                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
3058                                               cm_id_priv->msg);
3059                         cm_enter_timewait(cm_id_priv);
3060                         break;
3061                 }
3062                 fallthrough;
3063         default:
3064                 trace_icm_rej_unknown_err(&cm_id_priv->id);
3065                 spin_unlock_irq(&cm_id_priv->lock);
3066                 goto out;
3067         }
3068
3069         cm_queue_work_unlock(cm_id_priv, work);
3070         return 0;
3071 out:
3072         cm_deref_id(cm_id_priv);
3073         return -EINVAL;
3074 }
3075
3076 int ib_send_cm_mra(struct ib_cm_id *cm_id,
3077                    u8 service_timeout,
3078                    const void *private_data,
3079                    u8 private_data_len)
3080 {
3081         struct cm_id_private *cm_id_priv;
3082         struct ib_mad_send_buf *msg;
3083         enum ib_cm_state cm_state;
3084         enum ib_cm_lap_state lap_state;
3085         enum cm_msg_response msg_response;
3086         void *data;
3087         unsigned long flags;
3088         int ret;
3089
3090         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
3091                 return -EINVAL;
3092
3093         data = cm_copy_private_data(private_data, private_data_len);
3094         if (IS_ERR(data))
3095                 return PTR_ERR(data);
3096
3097         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3098
3099         spin_lock_irqsave(&cm_id_priv->lock, flags);
3100         switch(cm_id_priv->id.state) {
3101         case IB_CM_REQ_RCVD:
3102                 cm_state = IB_CM_MRA_REQ_SENT;
3103                 lap_state = cm_id->lap_state;
3104                 msg_response = CM_MSG_RESPONSE_REQ;
3105                 break;
3106         case IB_CM_REP_RCVD:
3107                 cm_state = IB_CM_MRA_REP_SENT;
3108                 lap_state = cm_id->lap_state;
3109                 msg_response = CM_MSG_RESPONSE_REP;
3110                 break;
3111         case IB_CM_ESTABLISHED:
3112                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
3113                         cm_state = cm_id->state;
3114                         lap_state = IB_CM_MRA_LAP_SENT;
3115                         msg_response = CM_MSG_RESPONSE_OTHER;
3116                         break;
3117                 }
3118                 fallthrough;
3119         default:
3120                 trace_icm_send_mra_unknown_err(&cm_id_priv->id);
3121                 ret = -EINVAL;
3122                 goto error1;
3123         }
3124
3125         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
3126                 ret = cm_alloc_msg(cm_id_priv, &msg);
3127                 if (ret)
3128                         goto error1;
3129
3130                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3131                               msg_response, service_timeout,
3132                               private_data, private_data_len);
3133                 trace_icm_send_mra(cm_id);
3134                 ret = ib_post_send_mad(msg, NULL);
3135                 if (ret)
3136                         goto error2;
3137         }
3138
3139         cm_id->state = cm_state;
3140         cm_id->lap_state = lap_state;
3141         cm_id_priv->service_timeout = service_timeout;
3142         cm_set_private_data(cm_id_priv, data, private_data_len);
3143         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3144         return 0;
3145
3146 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3147         kfree(data);
3148         return ret;
3149
3150 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3151         kfree(data);
3152         cm_free_msg(msg);
3153         return ret;
3154 }
3155 EXPORT_SYMBOL(ib_send_cm_mra);
3156
3157 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
3158 {
3159         switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) {
3160         case CM_MSG_RESPONSE_REQ:
3161                 return cm_acquire_id(
3162                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3163                         0);
3164         case CM_MSG_RESPONSE_REP:
3165         case CM_MSG_RESPONSE_OTHER:
3166                 return cm_acquire_id(
3167                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3168                         cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg)));
3169         default:
3170                 return NULL;
3171         }
3172 }
3173
3174 static int cm_mra_handler(struct cm_work *work)
3175 {
3176         struct cm_id_private *cm_id_priv;
3177         struct cm_mra_msg *mra_msg;
3178         int timeout;
3179
3180         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
3181         cm_id_priv = cm_acquire_mraed_id(mra_msg);
3182         if (!cm_id_priv)
3183                 return -EINVAL;
3184
3185         work->cm_event.private_data =
3186                 IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg);
3187         work->cm_event.param.mra_rcvd.service_timeout =
3188                 IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg);
3189         timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) +
3190                   cm_convert_to_ms(cm_id_priv->av.timeout);
3191
3192         spin_lock_irq(&cm_id_priv->lock);
3193         switch (cm_id_priv->id.state) {
3194         case IB_CM_REQ_SENT:
3195                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3196                             CM_MSG_RESPONSE_REQ ||
3197                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3198                                   cm_id_priv->msg, timeout))
3199                         goto out;
3200                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
3201                 break;
3202         case IB_CM_REP_SENT:
3203                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3204                             CM_MSG_RESPONSE_REP ||
3205                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3206                                   cm_id_priv->msg, timeout))
3207                         goto out;
3208                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
3209                 break;
3210         case IB_CM_ESTABLISHED:
3211                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3212                             CM_MSG_RESPONSE_OTHER ||
3213                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
3214                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3215                                   cm_id_priv->msg, timeout)) {
3216                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
3217                                 atomic_long_inc(&work->port->
3218                                                 counter_group[CM_RECV_DUPLICATES].
3219                                                 counter[CM_MRA_COUNTER]);
3220                         goto out;
3221                 }
3222                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
3223                 break;
3224         case IB_CM_MRA_REQ_RCVD:
3225         case IB_CM_MRA_REP_RCVD:
3226                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3227                                 counter[CM_MRA_COUNTER]);
3228                 fallthrough;
3229         default:
3230                 trace_icm_mra_unknown_err(&cm_id_priv->id);
3231                 goto out;
3232         }
3233
3234         cm_id_priv->msg->context[1] = (void *) (unsigned long)
3235                                       cm_id_priv->id.state;
3236         cm_queue_work_unlock(cm_id_priv, work);
3237         return 0;
3238 out:
3239         spin_unlock_irq(&cm_id_priv->lock);
3240         cm_deref_id(cm_id_priv);
3241         return -EINVAL;
3242 }
3243
3244 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3245                                         struct sa_path_rec *path)
3246 {
3247         u32 lid;
3248
3249         if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3250                 sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID,
3251                                                lap_msg));
3252                 sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID,
3253                                                lap_msg));
3254         } else {
3255                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3256                         CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg));
3257                 sa_path_set_dlid(path, lid);
3258
3259                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3260                         CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg));
3261                 sa_path_set_slid(path, lid);
3262         }
3263 }
3264
3265 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3266                                     struct sa_path_rec *path,
3267                                     struct cm_lap_msg *lap_msg)
3268 {
3269         path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg);
3270         path->sgid =
3271                 *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg);
3272         path->flow_label =
3273                 cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg));
3274         path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg);
3275         path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg);
3276         path->reversible = 1;
3277         path->pkey = cm_id_priv->pkey;
3278         path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg);
3279         path->mtu_selector = IB_SA_EQ;
3280         path->mtu = cm_id_priv->path_mtu;
3281         path->rate_selector = IB_SA_EQ;
3282         path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg);
3283         path->packet_life_time_selector = IB_SA_EQ;
3284         path->packet_life_time =
3285                 IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg);
3286         path->packet_life_time -= (path->packet_life_time > 0);
3287         cm_format_path_lid_from_lap(lap_msg, path);
3288 }
3289
3290 static int cm_lap_handler(struct cm_work *work)
3291 {
3292         struct cm_id_private *cm_id_priv;
3293         struct cm_lap_msg *lap_msg;
3294         struct ib_cm_lap_event_param *param;
3295         struct ib_mad_send_buf *msg = NULL;
3296         int ret;
3297
3298         /* Currently Alternate path messages are not supported for
3299          * RoCE link layer.
3300          */
3301         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3302                                work->port->port_num))
3303                 return -EINVAL;
3304
3305         /* todo: verify LAP request and send reject APR if invalid. */
3306         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3307         cm_id_priv = cm_acquire_id(
3308                 cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)),
3309                 cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg)));
3310         if (!cm_id_priv)
3311                 return -EINVAL;
3312
3313         param = &work->cm_event.param.lap_rcvd;
3314         memset(&work->path[0], 0, sizeof(work->path[1]));
3315         cm_path_set_rec_type(work->port->cm_dev->ib_device,
3316                              work->port->port_num, &work->path[0],
3317                              IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID,
3318                                              lap_msg));
3319         param->alternate_path = &work->path[0];
3320         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3321         work->cm_event.private_data =
3322                 IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg);
3323
3324         spin_lock_irq(&cm_id_priv->lock);
3325         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3326                 goto unlock;
3327
3328         switch (cm_id_priv->id.lap_state) {
3329         case IB_CM_LAP_UNINIT:
3330         case IB_CM_LAP_IDLE:
3331                 break;
3332         case IB_CM_MRA_LAP_SENT:
3333                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3334                                 counter[CM_LAP_COUNTER]);
3335                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
3336                 if (IS_ERR(msg))
3337                         goto unlock;
3338
3339                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3340                               CM_MSG_RESPONSE_OTHER,
3341                               cm_id_priv->service_timeout,
3342                               cm_id_priv->private_data,
3343                               cm_id_priv->private_data_len);
3344                 spin_unlock_irq(&cm_id_priv->lock);
3345
3346                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3347                     ib_post_send_mad(msg, NULL))
3348                         cm_free_msg(msg);
3349                 goto deref;
3350         case IB_CM_LAP_RCVD:
3351                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3352                                 counter[CM_LAP_COUNTER]);
3353                 goto unlock;
3354         default:
3355                 goto unlock;
3356         }
3357
3358         ret = cm_init_av_for_lap(work->port, work->mad_recv_wc->wc,
3359                                  work->mad_recv_wc->recv_buf.grh,
3360                                  &cm_id_priv->av);
3361         if (ret)
3362                 goto unlock;
3363
3364         ret = cm_init_av_by_path(param->alternate_path, NULL,
3365                                  &cm_id_priv->alt_av, cm_id_priv);
3366         if (ret)
3367                 goto unlock;
3368
3369         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3370         cm_id_priv->tid = lap_msg->hdr.tid;
3371         cm_queue_work_unlock(cm_id_priv, work);
3372         return 0;
3373
3374 unlock: spin_unlock_irq(&cm_id_priv->lock);
3375 deref:  cm_deref_id(cm_id_priv);
3376         return -EINVAL;
3377 }
3378
3379 static int cm_apr_handler(struct cm_work *work)
3380 {
3381         struct cm_id_private *cm_id_priv;
3382         struct cm_apr_msg *apr_msg;
3383
3384         /* Currently Alternate path messages are not supported for
3385          * RoCE link layer.
3386          */
3387         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3388                                work->port->port_num))
3389                 return -EINVAL;
3390
3391         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3392         cm_id_priv = cm_acquire_id(
3393                 cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)),
3394                 cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg)));
3395         if (!cm_id_priv)
3396                 return -EINVAL; /* Unmatched reply. */
3397
3398         work->cm_event.param.apr_rcvd.ap_status =
3399                 IBA_GET(CM_APR_AR_STATUS, apr_msg);
3400         work->cm_event.param.apr_rcvd.apr_info =
3401                 IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg);
3402         work->cm_event.param.apr_rcvd.info_len =
3403                 IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg);
3404         work->cm_event.private_data =
3405                 IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg);
3406
3407         spin_lock_irq(&cm_id_priv->lock);
3408         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3409             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3410              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3411                 spin_unlock_irq(&cm_id_priv->lock);
3412                 goto out;
3413         }
3414         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3415         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3416         cm_id_priv->msg = NULL;
3417         cm_queue_work_unlock(cm_id_priv, work);
3418         return 0;
3419 out:
3420         cm_deref_id(cm_id_priv);
3421         return -EINVAL;
3422 }
3423
3424 static int cm_timewait_handler(struct cm_work *work)
3425 {
3426         struct cm_timewait_info *timewait_info;
3427         struct cm_id_private *cm_id_priv;
3428
3429         timewait_info = container_of(work, struct cm_timewait_info, work);
3430         spin_lock_irq(&cm.lock);
3431         list_del(&timewait_info->list);
3432         spin_unlock_irq(&cm.lock);
3433
3434         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3435                                    timewait_info->work.remote_id);
3436         if (!cm_id_priv)
3437                 return -EINVAL;
3438
3439         spin_lock_irq(&cm_id_priv->lock);
3440         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3441             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3442                 spin_unlock_irq(&cm_id_priv->lock);
3443                 goto out;
3444         }
3445         cm_id_priv->id.state = IB_CM_IDLE;
3446         cm_queue_work_unlock(cm_id_priv, work);
3447         return 0;
3448 out:
3449         cm_deref_id(cm_id_priv);
3450         return -EINVAL;
3451 }
3452
3453 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3454                                struct cm_id_private *cm_id_priv,
3455                                struct ib_cm_sidr_req_param *param)
3456 {
3457         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3458                           cm_form_tid(cm_id_priv));
3459         IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg,
3460                 be32_to_cpu(cm_id_priv->id.local_id));
3461         IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg,
3462                 be16_to_cpu(param->path->pkey));
3463         IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg,
3464                 be64_to_cpu(param->service_id));
3465
3466         if (param->private_data && param->private_data_len)
3467                 IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg,
3468                             param->private_data, param->private_data_len);
3469 }
3470
3471 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3472                         struct ib_cm_sidr_req_param *param)
3473 {
3474         struct cm_id_private *cm_id_priv;
3475         struct ib_mad_send_buf *msg;
3476         unsigned long flags;
3477         int ret;
3478
3479         if (!param->path || (param->private_data &&
3480              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3481                 return -EINVAL;
3482
3483         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3484         ret = cm_init_av_by_path(param->path, param->sgid_attr,
3485                                  &cm_id_priv->av,
3486                                  cm_id_priv);
3487         if (ret)
3488                 goto out;
3489
3490         cm_id->service_id = param->service_id;
3491         cm_id->service_mask = ~cpu_to_be64(0);
3492         cm_id_priv->timeout_ms = param->timeout_ms;
3493         cm_id_priv->max_cm_retries = param->max_cm_retries;
3494         ret = cm_alloc_msg(cm_id_priv, &msg);
3495         if (ret)
3496                 goto out;
3497
3498         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3499                            param);
3500         msg->timeout_ms = cm_id_priv->timeout_ms;
3501         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3502
3503         spin_lock_irqsave(&cm_id_priv->lock, flags);
3504         if (cm_id->state == IB_CM_IDLE) {
3505                 trace_icm_send_sidr_req(&cm_id_priv->id);
3506                 ret = ib_post_send_mad(msg, NULL);
3507         } else {
3508                 ret = -EINVAL;
3509         }
3510
3511         if (ret) {
3512                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3513                 cm_free_msg(msg);
3514                 goto out;
3515         }
3516         cm_id->state = IB_CM_SIDR_REQ_SENT;
3517         cm_id_priv->msg = msg;
3518         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3519 out:
3520         return ret;
3521 }
3522 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3523
3524 static void cm_format_sidr_req_event(struct cm_work *work,
3525                                      const struct cm_id_private *rx_cm_id,
3526                                      struct ib_cm_id *listen_id)
3527 {
3528         struct cm_sidr_req_msg *sidr_req_msg;
3529         struct ib_cm_sidr_req_event_param *param;
3530
3531         sidr_req_msg = (struct cm_sidr_req_msg *)
3532                                 work->mad_recv_wc->recv_buf.mad;
3533         param = &work->cm_event.param.sidr_req_rcvd;
3534         param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg);
3535         param->listen_id = listen_id;
3536         param->service_id =
3537                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3538         param->bth_pkey = cm_get_bth_pkey(work);
3539         param->port = work->port->port_num;
3540         param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr;
3541         work->cm_event.private_data =
3542                 IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg);
3543 }
3544
3545 static int cm_sidr_req_handler(struct cm_work *work)
3546 {
3547         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
3548         struct cm_sidr_req_msg *sidr_req_msg;
3549         struct ib_wc *wc;
3550         int ret;
3551
3552         cm_id_priv =
3553                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
3554         if (IS_ERR(cm_id_priv))
3555                 return PTR_ERR(cm_id_priv);
3556
3557         /* Record SGID/SLID and request ID for lookup. */
3558         sidr_req_msg = (struct cm_sidr_req_msg *)
3559                                 work->mad_recv_wc->recv_buf.mad;
3560
3561         cm_id_priv->id.remote_id =
3562                 cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg));
3563         cm_id_priv->id.service_id =
3564                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3565         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3566         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3567
3568         wc = work->mad_recv_wc->wc;
3569         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3570         cm_id_priv->av.dgid.global.interface_id = 0;
3571         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3572                                       work->mad_recv_wc->recv_buf.grh,
3573                                       &cm_id_priv->av);
3574         if (ret)
3575                 goto out;
3576
3577         spin_lock_irq(&cm.lock);
3578         listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3579         if (listen_cm_id_priv) {
3580                 spin_unlock_irq(&cm.lock);
3581                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3582                                 counter[CM_SIDR_REQ_COUNTER]);
3583                 goto out; /* Duplicate message. */
3584         }
3585         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3586         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
3587                                            cm_id_priv->id.service_id);
3588         if (!listen_cm_id_priv) {
3589                 spin_unlock_irq(&cm.lock);
3590                 ib_send_cm_sidr_rep(&cm_id_priv->id,
3591                                     &(struct ib_cm_sidr_rep_param){
3592                                             .status = IB_SIDR_UNSUPPORTED });
3593                 goto out; /* No match. */
3594         }
3595         spin_unlock_irq(&cm.lock);
3596
3597         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
3598         cm_id_priv->id.context = listen_cm_id_priv->id.context;
3599
3600         /*
3601          * A SIDR ID does not need to be in the xarray since it does not receive
3602          * mads, is not placed in the remote_id or remote_qpn rbtree, and does
3603          * not enter timewait.
3604          */
3605
3606         cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
3607         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
3608         cm_free_work(work);
3609         /*
3610          * A pointer to the listen_cm_id is held in the event, so this deref
3611          * must be after the event is delivered above.
3612          */
3613         cm_deref_id(listen_cm_id_priv);
3614         if (ret)
3615                 cm_destroy_id(&cm_id_priv->id, ret);
3616         return 0;
3617 out:
3618         ib_destroy_cm_id(&cm_id_priv->id);
3619         return -EINVAL;
3620 }
3621
3622 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3623                                struct cm_id_private *cm_id_priv,
3624                                struct ib_cm_sidr_rep_param *param)
3625 {
3626         cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3627                               cm_id_priv->tid, param->ece.attr_mod);
3628         IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg,
3629                 be32_to_cpu(cm_id_priv->id.remote_id));
3630         IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status);
3631         IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num);
3632         IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg,
3633                 be64_to_cpu(cm_id_priv->id.service_id));
3634         IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey);
3635         IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg,
3636                 param->ece.vendor_id & 0xFF);
3637         IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg,
3638                 (param->ece.vendor_id >> 8) & 0xFF);
3639
3640         if (param->info && param->info_length)
3641                 IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg,
3642                             param->info, param->info_length);
3643
3644         if (param->private_data && param->private_data_len)
3645                 IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg,
3646                             param->private_data, param->private_data_len);
3647 }
3648
3649 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
3650                                    struct ib_cm_sidr_rep_param *param)
3651 {
3652         struct ib_mad_send_buf *msg;
3653         int ret;
3654
3655         lockdep_assert_held(&cm_id_priv->lock);
3656
3657         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3658             (param->private_data &&
3659              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3660                 return -EINVAL;
3661
3662         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD)
3663                 return -EINVAL;
3664
3665         ret = cm_alloc_msg(cm_id_priv, &msg);
3666         if (ret)
3667                 return ret;
3668
3669         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3670                            param);
3671         trace_icm_send_sidr_rep(&cm_id_priv->id);
3672         ret = ib_post_send_mad(msg, NULL);
3673         if (ret) {
3674                 cm_free_msg(msg);
3675                 return ret;
3676         }
3677         cm_id_priv->id.state = IB_CM_IDLE;
3678         spin_lock_irq(&cm.lock);
3679         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3680                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3681                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3682         }
3683         spin_unlock_irq(&cm.lock);
3684         return 0;
3685 }
3686
3687 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3688                         struct ib_cm_sidr_rep_param *param)
3689 {
3690         struct cm_id_private *cm_id_priv =
3691                 container_of(cm_id, struct cm_id_private, id);
3692         unsigned long flags;
3693         int ret;
3694
3695         spin_lock_irqsave(&cm_id_priv->lock, flags);
3696         ret = cm_send_sidr_rep_locked(cm_id_priv, param);
3697         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3698         return ret;
3699 }
3700 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3701
3702 static void cm_format_sidr_rep_event(struct cm_work *work,
3703                                      const struct cm_id_private *cm_id_priv)
3704 {
3705         struct cm_sidr_rep_msg *sidr_rep_msg;
3706         struct ib_cm_sidr_rep_event_param *param;
3707
3708         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3709                                 work->mad_recv_wc->recv_buf.mad;
3710         param = &work->cm_event.param.sidr_rep_rcvd;
3711         param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg);
3712         param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg);
3713         param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg);
3714         param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION,
3715                                       sidr_rep_msg);
3716         param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH,
3717                                   sidr_rep_msg);
3718         param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
3719         work->cm_event.private_data =
3720                 IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg);
3721 }
3722
3723 static int cm_sidr_rep_handler(struct cm_work *work)
3724 {
3725         struct cm_sidr_rep_msg *sidr_rep_msg;
3726         struct cm_id_private *cm_id_priv;
3727
3728         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3729                                 work->mad_recv_wc->recv_buf.mad;
3730         cm_id_priv = cm_acquire_id(
3731                 cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0);
3732         if (!cm_id_priv)
3733                 return -EINVAL; /* Unmatched reply. */
3734
3735         spin_lock_irq(&cm_id_priv->lock);
3736         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3737                 spin_unlock_irq(&cm_id_priv->lock);
3738                 goto out;
3739         }
3740         cm_id_priv->id.state = IB_CM_IDLE;
3741         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3742         spin_unlock_irq(&cm_id_priv->lock);
3743
3744         cm_format_sidr_rep_event(work, cm_id_priv);
3745         cm_process_work(cm_id_priv, work);
3746         return 0;
3747 out:
3748         cm_deref_id(cm_id_priv);
3749         return -EINVAL;
3750 }
3751
3752 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3753                                   enum ib_wc_status wc_status)
3754 {
3755         struct cm_id_private *cm_id_priv;
3756         struct ib_cm_event cm_event;
3757         enum ib_cm_state state;
3758         int ret;
3759
3760         memset(&cm_event, 0, sizeof cm_event);
3761         cm_id_priv = msg->context[0];
3762
3763         /* Discard old sends or ones without a response. */
3764         spin_lock_irq(&cm_id_priv->lock);
3765         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3766         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3767                 goto discard;
3768
3769         trace_icm_mad_send_err(state, wc_status);
3770         switch (state) {
3771         case IB_CM_REQ_SENT:
3772         case IB_CM_MRA_REQ_RCVD:
3773                 cm_reset_to_idle(cm_id_priv);
3774                 cm_event.event = IB_CM_REQ_ERROR;
3775                 break;
3776         case IB_CM_REP_SENT:
3777         case IB_CM_MRA_REP_RCVD:
3778                 cm_reset_to_idle(cm_id_priv);
3779                 cm_event.event = IB_CM_REP_ERROR;
3780                 break;
3781         case IB_CM_DREQ_SENT:
3782                 cm_enter_timewait(cm_id_priv);
3783                 cm_event.event = IB_CM_DREQ_ERROR;
3784                 break;
3785         case IB_CM_SIDR_REQ_SENT:
3786                 cm_id_priv->id.state = IB_CM_IDLE;
3787                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3788                 break;
3789         default:
3790                 goto discard;
3791         }
3792         spin_unlock_irq(&cm_id_priv->lock);
3793         cm_event.param.send_status = wc_status;
3794
3795         /* No other events can occur on the cm_id at this point. */
3796         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3797         cm_free_msg(msg);
3798         if (ret)
3799                 ib_destroy_cm_id(&cm_id_priv->id);
3800         return;
3801 discard:
3802         spin_unlock_irq(&cm_id_priv->lock);
3803         cm_free_msg(msg);
3804 }
3805
3806 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3807                             struct ib_mad_send_wc *mad_send_wc)
3808 {
3809         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3810         struct cm_port *port;
3811         u16 attr_index;
3812
3813         port = mad_agent->context;
3814         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3815                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3816
3817         /*
3818          * If the send was in response to a received message (context[0] is not
3819          * set to a cm_id), and is not a REJ, then it is a send that was
3820          * manually retried.
3821          */
3822         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3823                 msg->retries = 1;
3824
3825         atomic_long_add(1 + msg->retries,
3826                         &port->counter_group[CM_XMIT].counter[attr_index]);
3827         if (msg->retries)
3828                 atomic_long_add(msg->retries,
3829                                 &port->counter_group[CM_XMIT_RETRIES].
3830                                 counter[attr_index]);
3831
3832         switch (mad_send_wc->status) {
3833         case IB_WC_SUCCESS:
3834         case IB_WC_WR_FLUSH_ERR:
3835                 cm_free_msg(msg);
3836                 break;
3837         default:
3838                 if (msg->context[0] && msg->context[1])
3839                         cm_process_send_error(msg, mad_send_wc->status);
3840                 else
3841                         cm_free_msg(msg);
3842                 break;
3843         }
3844 }
3845
3846 static void cm_work_handler(struct work_struct *_work)
3847 {
3848         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3849         int ret;
3850
3851         switch (work->cm_event.event) {
3852         case IB_CM_REQ_RECEIVED:
3853                 ret = cm_req_handler(work);
3854                 break;
3855         case IB_CM_MRA_RECEIVED:
3856                 ret = cm_mra_handler(work);
3857                 break;
3858         case IB_CM_REJ_RECEIVED:
3859                 ret = cm_rej_handler(work);
3860                 break;
3861         case IB_CM_REP_RECEIVED:
3862                 ret = cm_rep_handler(work);
3863                 break;
3864         case IB_CM_RTU_RECEIVED:
3865                 ret = cm_rtu_handler(work);
3866                 break;
3867         case IB_CM_USER_ESTABLISHED:
3868                 ret = cm_establish_handler(work);
3869                 break;
3870         case IB_CM_DREQ_RECEIVED:
3871                 ret = cm_dreq_handler(work);
3872                 break;
3873         case IB_CM_DREP_RECEIVED:
3874                 ret = cm_drep_handler(work);
3875                 break;
3876         case IB_CM_SIDR_REQ_RECEIVED:
3877                 ret = cm_sidr_req_handler(work);
3878                 break;
3879         case IB_CM_SIDR_REP_RECEIVED:
3880                 ret = cm_sidr_rep_handler(work);
3881                 break;
3882         case IB_CM_LAP_RECEIVED:
3883                 ret = cm_lap_handler(work);
3884                 break;
3885         case IB_CM_APR_RECEIVED:
3886                 ret = cm_apr_handler(work);
3887                 break;
3888         case IB_CM_TIMEWAIT_EXIT:
3889                 ret = cm_timewait_handler(work);
3890                 break;
3891         default:
3892                 trace_icm_handler_err(work->cm_event.event);
3893                 ret = -EINVAL;
3894                 break;
3895         }
3896         if (ret)
3897                 cm_free_work(work);
3898 }
3899
3900 static int cm_establish(struct ib_cm_id *cm_id)
3901 {
3902         struct cm_id_private *cm_id_priv;
3903         struct cm_work *work;
3904         unsigned long flags;
3905         int ret = 0;
3906         struct cm_device *cm_dev;
3907
3908         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3909         if (!cm_dev)
3910                 return -ENODEV;
3911
3912         work = kmalloc(sizeof *work, GFP_ATOMIC);
3913         if (!work)
3914                 return -ENOMEM;
3915
3916         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3917         spin_lock_irqsave(&cm_id_priv->lock, flags);
3918         switch (cm_id->state)
3919         {
3920         case IB_CM_REP_SENT:
3921         case IB_CM_MRA_REP_RCVD:
3922                 cm_id->state = IB_CM_ESTABLISHED;
3923                 break;
3924         case IB_CM_ESTABLISHED:
3925                 ret = -EISCONN;
3926                 break;
3927         default:
3928                 trace_icm_establish_err(cm_id);
3929                 ret = -EINVAL;
3930                 break;
3931         }
3932         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3933
3934         if (ret) {
3935                 kfree(work);
3936                 goto out;
3937         }
3938
3939         /*
3940          * The CM worker thread may try to destroy the cm_id before it
3941          * can execute this work item.  To prevent potential deadlock,
3942          * we need to find the cm_id once we're in the context of the
3943          * worker thread, rather than holding a reference on it.
3944          */
3945         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3946         work->local_id = cm_id->local_id;
3947         work->remote_id = cm_id->remote_id;
3948         work->mad_recv_wc = NULL;
3949         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3950
3951         /* Check if the device started its remove_one */
3952         spin_lock_irqsave(&cm.lock, flags);
3953         if (!cm_dev->going_down) {
3954                 queue_delayed_work(cm.wq, &work->work, 0);
3955         } else {
3956                 kfree(work);
3957                 ret = -ENODEV;
3958         }
3959         spin_unlock_irqrestore(&cm.lock, flags);
3960
3961 out:
3962         return ret;
3963 }
3964
3965 static int cm_migrate(struct ib_cm_id *cm_id)
3966 {
3967         struct cm_id_private *cm_id_priv;
3968         struct cm_av tmp_av;
3969         unsigned long flags;
3970         int tmp_send_port_not_ready;
3971         int ret = 0;
3972
3973         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3974         spin_lock_irqsave(&cm_id_priv->lock, flags);
3975         if (cm_id->state == IB_CM_ESTABLISHED &&
3976             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3977              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3978                 cm_id->lap_state = IB_CM_LAP_IDLE;
3979                 /* Swap address vector */
3980                 tmp_av = cm_id_priv->av;
3981                 cm_id_priv->av = cm_id_priv->alt_av;
3982                 cm_id_priv->alt_av = tmp_av;
3983                 /* Swap port send ready state */
3984                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3985                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3986                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3987         } else
3988                 ret = -EINVAL;
3989         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3990
3991         return ret;
3992 }
3993
3994 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3995 {
3996         int ret;
3997
3998         switch (event) {
3999         case IB_EVENT_COMM_EST:
4000                 ret = cm_establish(cm_id);
4001                 break;
4002         case IB_EVENT_PATH_MIG:
4003                 ret = cm_migrate(cm_id);
4004                 break;
4005         default:
4006                 ret = -EINVAL;
4007         }
4008         return ret;
4009 }
4010 EXPORT_SYMBOL(ib_cm_notify);
4011
4012 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
4013                             struct ib_mad_send_buf *send_buf,
4014                             struct ib_mad_recv_wc *mad_recv_wc)
4015 {
4016         struct cm_port *port = mad_agent->context;
4017         struct cm_work *work;
4018         enum ib_cm_event_type event;
4019         bool alt_path = false;
4020         u16 attr_id;
4021         int paths = 0;
4022         int going_down = 0;
4023
4024         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
4025         case CM_REQ_ATTR_ID:
4026                 alt_path = cm_req_has_alt_path((struct cm_req_msg *)
4027                                                 mad_recv_wc->recv_buf.mad);
4028                 paths = 1 + (alt_path != 0);
4029                 event = IB_CM_REQ_RECEIVED;
4030                 break;
4031         case CM_MRA_ATTR_ID:
4032                 event = IB_CM_MRA_RECEIVED;
4033                 break;
4034         case CM_REJ_ATTR_ID:
4035                 event = IB_CM_REJ_RECEIVED;
4036                 break;
4037         case CM_REP_ATTR_ID:
4038                 event = IB_CM_REP_RECEIVED;
4039                 break;
4040         case CM_RTU_ATTR_ID:
4041                 event = IB_CM_RTU_RECEIVED;
4042                 break;
4043         case CM_DREQ_ATTR_ID:
4044                 event = IB_CM_DREQ_RECEIVED;
4045                 break;
4046         case CM_DREP_ATTR_ID:
4047                 event = IB_CM_DREP_RECEIVED;
4048                 break;
4049         case CM_SIDR_REQ_ATTR_ID:
4050                 event = IB_CM_SIDR_REQ_RECEIVED;
4051                 break;
4052         case CM_SIDR_REP_ATTR_ID:
4053                 event = IB_CM_SIDR_REP_RECEIVED;
4054                 break;
4055         case CM_LAP_ATTR_ID:
4056                 paths = 1;
4057                 event = IB_CM_LAP_RECEIVED;
4058                 break;
4059         case CM_APR_ATTR_ID:
4060                 event = IB_CM_APR_RECEIVED;
4061                 break;
4062         default:
4063                 ib_free_recv_mad(mad_recv_wc);
4064                 return;
4065         }
4066
4067         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
4068         atomic_long_inc(&port->counter_group[CM_RECV].
4069                         counter[attr_id - CM_ATTR_ID_OFFSET]);
4070
4071         work = kmalloc(struct_size(work, path, paths), GFP_KERNEL);
4072         if (!work) {
4073                 ib_free_recv_mad(mad_recv_wc);
4074                 return;
4075         }
4076
4077         INIT_DELAYED_WORK(&work->work, cm_work_handler);
4078         work->cm_event.event = event;
4079         work->mad_recv_wc = mad_recv_wc;
4080         work->port = port;
4081
4082         /* Check if the device started its remove_one */
4083         spin_lock_irq(&cm.lock);
4084         if (!port->cm_dev->going_down)
4085                 queue_delayed_work(cm.wq, &work->work, 0);
4086         else
4087                 going_down = 1;
4088         spin_unlock_irq(&cm.lock);
4089
4090         if (going_down) {
4091                 kfree(work);
4092                 ib_free_recv_mad(mad_recv_wc);
4093         }
4094 }
4095
4096 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
4097                                 struct ib_qp_attr *qp_attr,
4098                                 int *qp_attr_mask)
4099 {
4100         unsigned long flags;
4101         int ret;
4102
4103         spin_lock_irqsave(&cm_id_priv->lock, flags);
4104         switch (cm_id_priv->id.state) {
4105         case IB_CM_REQ_SENT:
4106         case IB_CM_MRA_REQ_RCVD:
4107         case IB_CM_REQ_RCVD:
4108         case IB_CM_MRA_REQ_SENT:
4109         case IB_CM_REP_RCVD:
4110         case IB_CM_MRA_REP_SENT:
4111         case IB_CM_REP_SENT:
4112         case IB_CM_MRA_REP_RCVD:
4113         case IB_CM_ESTABLISHED:
4114                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
4115                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
4116                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
4117                 if (cm_id_priv->responder_resources)
4118                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
4119                                                     IB_ACCESS_REMOTE_ATOMIC;
4120                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
4121                 qp_attr->port_num = cm_id_priv->av.port->port_num;
4122                 ret = 0;
4123                 break;
4124         default:
4125                 trace_icm_qp_init_err(&cm_id_priv->id);
4126                 ret = -EINVAL;
4127                 break;
4128         }
4129         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4130         return ret;
4131 }
4132
4133 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
4134                                struct ib_qp_attr *qp_attr,
4135                                int *qp_attr_mask)
4136 {
4137         unsigned long flags;
4138         int ret;
4139
4140         spin_lock_irqsave(&cm_id_priv->lock, flags);
4141         switch (cm_id_priv->id.state) {
4142         case IB_CM_REQ_RCVD:
4143         case IB_CM_MRA_REQ_SENT:
4144         case IB_CM_REP_RCVD:
4145         case IB_CM_MRA_REP_SENT:
4146         case IB_CM_REP_SENT:
4147         case IB_CM_MRA_REP_RCVD:
4148         case IB_CM_ESTABLISHED:
4149                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
4150                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
4151                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
4152                 qp_attr->path_mtu = cm_id_priv->path_mtu;
4153                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
4154                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
4155                 if (cm_id_priv->qp_type == IB_QPT_RC ||
4156                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
4157                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
4158                                          IB_QP_MIN_RNR_TIMER;
4159                         qp_attr->max_dest_rd_atomic =
4160                                         cm_id_priv->responder_resources;
4161                         qp_attr->min_rnr_timer = 0;
4162                 }
4163                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4164                         *qp_attr_mask |= IB_QP_ALT_PATH;
4165                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4166                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4167                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4168                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4169                 }
4170                 ret = 0;
4171                 break;
4172         default:
4173                 trace_icm_qp_rtr_err(&cm_id_priv->id);
4174                 ret = -EINVAL;
4175                 break;
4176         }
4177         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4178         return ret;
4179 }
4180
4181 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
4182                                struct ib_qp_attr *qp_attr,
4183                                int *qp_attr_mask)
4184 {
4185         unsigned long flags;
4186         int ret;
4187
4188         spin_lock_irqsave(&cm_id_priv->lock, flags);
4189         switch (cm_id_priv->id.state) {
4190         /* Allow transition to RTS before sending REP */
4191         case IB_CM_REQ_RCVD:
4192         case IB_CM_MRA_REQ_SENT:
4193
4194         case IB_CM_REP_RCVD:
4195         case IB_CM_MRA_REP_SENT:
4196         case IB_CM_REP_SENT:
4197         case IB_CM_MRA_REP_RCVD:
4198         case IB_CM_ESTABLISHED:
4199                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
4200                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
4201                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4202                         switch (cm_id_priv->qp_type) {
4203                         case IB_QPT_RC:
4204                         case IB_QPT_XRC_INI:
4205                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4206                                                  IB_QP_MAX_QP_RD_ATOMIC;
4207                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
4208                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4209                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4210                                 fallthrough;
4211                         case IB_QPT_XRC_TGT:
4212                                 *qp_attr_mask |= IB_QP_TIMEOUT;
4213                                 qp_attr->timeout = cm_id_priv->av.timeout;
4214                                 break;
4215                         default:
4216                                 break;
4217                         }
4218                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4219                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4220                                 qp_attr->path_mig_state = IB_MIG_REARM;
4221                         }
4222                 } else {
4223                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4224                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4225                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4226                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4227                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4228                         qp_attr->path_mig_state = IB_MIG_REARM;
4229                 }
4230                 ret = 0;
4231                 break;
4232         default:
4233                 trace_icm_qp_rts_err(&cm_id_priv->id);
4234                 ret = -EINVAL;
4235                 break;
4236         }
4237         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4238         return ret;
4239 }
4240
4241 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4242                        struct ib_qp_attr *qp_attr,
4243                        int *qp_attr_mask)
4244 {
4245         struct cm_id_private *cm_id_priv;
4246         int ret;
4247
4248         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4249         switch (qp_attr->qp_state) {
4250         case IB_QPS_INIT:
4251                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4252                 break;
4253         case IB_QPS_RTR:
4254                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4255                 break;
4256         case IB_QPS_RTS:
4257                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4258                 break;
4259         default:
4260                 ret = -EINVAL;
4261                 break;
4262         }
4263         return ret;
4264 }
4265 EXPORT_SYMBOL(ib_cm_init_qp_attr);
4266
4267 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
4268                                char *buf)
4269 {
4270         struct cm_counter_group *group;
4271         struct cm_counter_attribute *cm_attr;
4272
4273         group = container_of(obj, struct cm_counter_group, obj);
4274         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
4275
4276         return sprintf(buf, "%ld\n",
4277                        atomic_long_read(&group->counter[cm_attr->index]));
4278 }
4279
4280 static const struct sysfs_ops cm_counter_ops = {
4281         .show = cm_show_counter
4282 };
4283
4284 static struct kobj_type cm_counter_obj_type = {
4285         .sysfs_ops = &cm_counter_ops,
4286         .default_attrs = cm_counter_default_attrs
4287 };
4288
4289 static int cm_create_port_fs(struct cm_port *port)
4290 {
4291         int i, ret;
4292
4293         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
4294                 ret = ib_port_register_module_stat(port->cm_dev->ib_device,
4295                                                    port->port_num,
4296                                                    &port->counter_group[i].obj,
4297                                                    &cm_counter_obj_type,
4298                                                    counter_group_names[i]);
4299                 if (ret)
4300                         goto error;
4301         }
4302
4303         return 0;
4304
4305 error:
4306         while (i--)
4307                 ib_port_unregister_module_stat(&port->counter_group[i].obj);
4308         return ret;
4309
4310 }
4311
4312 static void cm_remove_port_fs(struct cm_port *port)
4313 {
4314         int i;
4315
4316         for (i = 0; i < CM_COUNTER_GROUPS; i++)
4317                 ib_port_unregister_module_stat(&port->counter_group[i].obj);
4318
4319 }
4320
4321 static int cm_add_one(struct ib_device *ib_device)
4322 {
4323         struct cm_device *cm_dev;
4324         struct cm_port *port;
4325         struct ib_mad_reg_req reg_req = {
4326                 .mgmt_class = IB_MGMT_CLASS_CM,
4327                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4328         };
4329         struct ib_port_modify port_modify = {
4330                 .set_port_cap_mask = IB_PORT_CM_SUP
4331         };
4332         unsigned long flags;
4333         int ret;
4334         int count = 0;
4335         u8 i;
4336
4337         cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt),
4338                          GFP_KERNEL);
4339         if (!cm_dev)
4340                 return -ENOMEM;
4341
4342         cm_dev->ib_device = ib_device;
4343         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4344         cm_dev->going_down = 0;
4345
4346         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4347         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4348                 if (!rdma_cap_ib_cm(ib_device, i))
4349                         continue;
4350
4351                 port = kzalloc(sizeof *port, GFP_KERNEL);
4352                 if (!port) {
4353                         ret = -ENOMEM;
4354                         goto error1;
4355                 }
4356
4357                 cm_dev->port[i-1] = port;
4358                 port->cm_dev = cm_dev;
4359                 port->port_num = i;
4360
4361                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4362                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4363
4364                 ret = cm_create_port_fs(port);
4365                 if (ret)
4366                         goto error1;
4367
4368                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4369                                                         IB_QPT_GSI,
4370                                                         &reg_req,
4371                                                         0,
4372                                                         cm_send_handler,
4373                                                         cm_recv_handler,
4374                                                         port,
4375                                                         0);
4376                 if (IS_ERR(port->mad_agent)) {
4377                         ret = PTR_ERR(port->mad_agent);
4378                         goto error2;
4379                 }
4380
4381                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4382                 if (ret)
4383                         goto error3;
4384
4385                 count++;
4386         }
4387
4388         if (!count) {
4389                 ret = -EOPNOTSUPP;
4390                 goto free;
4391         }
4392
4393         ib_set_client_data(ib_device, &cm_client, cm_dev);
4394
4395         write_lock_irqsave(&cm.device_lock, flags);
4396         list_add_tail(&cm_dev->list, &cm.device_list);
4397         write_unlock_irqrestore(&cm.device_lock, flags);
4398         return 0;
4399
4400 error3:
4401         ib_unregister_mad_agent(port->mad_agent);
4402 error2:
4403         cm_remove_port_fs(port);
4404 error1:
4405         port_modify.set_port_cap_mask = 0;
4406         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4407         kfree(port);
4408         while (--i) {
4409                 if (!rdma_cap_ib_cm(ib_device, i))
4410                         continue;
4411
4412                 port = cm_dev->port[i-1];
4413                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4414                 ib_unregister_mad_agent(port->mad_agent);
4415                 cm_remove_port_fs(port);
4416                 kfree(port);
4417         }
4418 free:
4419         kfree(cm_dev);
4420         return ret;
4421 }
4422
4423 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4424 {
4425         struct cm_device *cm_dev = client_data;
4426         struct cm_port *port;
4427         struct cm_id_private *cm_id_priv;
4428         struct ib_mad_agent *cur_mad_agent;
4429         struct ib_port_modify port_modify = {
4430                 .clr_port_cap_mask = IB_PORT_CM_SUP
4431         };
4432         unsigned long flags;
4433         int i;
4434
4435         write_lock_irqsave(&cm.device_lock, flags);
4436         list_del(&cm_dev->list);
4437         write_unlock_irqrestore(&cm.device_lock, flags);
4438
4439         spin_lock_irq(&cm.lock);
4440         cm_dev->going_down = 1;
4441         spin_unlock_irq(&cm.lock);
4442
4443         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4444                 if (!rdma_cap_ib_cm(ib_device, i))
4445                         continue;
4446
4447                 port = cm_dev->port[i-1];
4448                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4449                 /* Mark all the cm_id's as not valid */
4450                 spin_lock_irq(&cm.lock);
4451                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4452                         cm_id_priv->altr_send_port_not_ready = 1;
4453                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4454                         cm_id_priv->prim_send_port_not_ready = 1;
4455                 spin_unlock_irq(&cm.lock);
4456                 /*
4457                  * We flush the queue here after the going_down set, this
4458                  * verify that no new works will be queued in the recv handler,
4459                  * after that we can call the unregister_mad_agent
4460                  */
4461                 flush_workqueue(cm.wq);
4462                 spin_lock_irq(&cm.state_lock);
4463                 cur_mad_agent = port->mad_agent;
4464                 port->mad_agent = NULL;
4465                 spin_unlock_irq(&cm.state_lock);
4466                 ib_unregister_mad_agent(cur_mad_agent);
4467                 cm_remove_port_fs(port);
4468                 kfree(port);
4469         }
4470
4471         kfree(cm_dev);
4472 }
4473
4474 static int __init ib_cm_init(void)
4475 {
4476         int ret;
4477
4478         INIT_LIST_HEAD(&cm.device_list);
4479         rwlock_init(&cm.device_lock);
4480         spin_lock_init(&cm.lock);
4481         spin_lock_init(&cm.state_lock);
4482         cm.listen_service_table = RB_ROOT;
4483         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4484         cm.remote_id_table = RB_ROOT;
4485         cm.remote_qp_table = RB_ROOT;
4486         cm.remote_sidr_table = RB_ROOT;
4487         xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC);
4488         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4489         INIT_LIST_HEAD(&cm.timewait_list);
4490
4491         cm.wq = alloc_workqueue("ib_cm", 0, 1);
4492         if (!cm.wq) {
4493                 ret = -ENOMEM;
4494                 goto error2;
4495         }
4496
4497         ret = ib_register_client(&cm_client);
4498         if (ret)
4499                 goto error3;
4500
4501         return 0;
4502 error3:
4503         destroy_workqueue(cm.wq);
4504 error2:
4505         return ret;
4506 }
4507
4508 static void __exit ib_cm_cleanup(void)
4509 {
4510         struct cm_timewait_info *timewait_info, *tmp;
4511
4512         spin_lock_irq(&cm.lock);
4513         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4514                 cancel_delayed_work(&timewait_info->work.work);
4515         spin_unlock_irq(&cm.lock);
4516
4517         ib_unregister_client(&cm_client);
4518         destroy_workqueue(cm.wq);
4519
4520         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4521                 list_del(&timewait_info->list);
4522                 kfree(timewait_info);
4523         }
4524
4525         WARN_ON(!xa_empty(&cm.local_id_table));
4526 }
4527
4528 module_init(ib_cm_init);
4529 module_exit(ib_cm_cleanup);