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  * ib_cm_insert_listen - Create a new listening ib_cm_id and listen on
1255  *                       the given service ID.
1256  *
1257  * If there's an existing ID listening on that same device and service ID,
1258  * return it.
1259  *
1260  * @device: Device associated with the cm_id.  All related communication will
1261  * be associated with the specified device.
1262  * @cm_handler: Callback invoked to notify the user of CM events.
1263  * @service_id: Service identifier matched against incoming connection
1264  *   and service ID resolution requests.  The service ID should be specified
1265  *   network-byte order.  If set to IB_CM_ASSIGN_SERVICE_ID, the CM will
1266  *   assign a service ID to the caller.
1267  *
1268  * Callers should call ib_destroy_cm_id when done with the listener ID.
1269  */
1270 struct ib_cm_id *ib_cm_insert_listen(struct ib_device *device,
1271                                      ib_cm_handler cm_handler,
1272                                      __be64 service_id)
1273 {
1274         struct cm_id_private *listen_id_priv;
1275         struct cm_id_private *cm_id_priv;
1276         int err = 0;
1277
1278         /* Create an ID in advance, since the creation may sleep */
1279         cm_id_priv = cm_alloc_id_priv(device, cm_handler, NULL);
1280         if (IS_ERR(cm_id_priv))
1281                 return ERR_CAST(cm_id_priv);
1282
1283         err = cm_init_listen(cm_id_priv, service_id, 0);
1284         if (err)
1285                 return ERR_PTR(err);
1286
1287         spin_lock_irq(&cm_id_priv->lock);
1288         listen_id_priv = cm_insert_listen(cm_id_priv, cm_handler);
1289         if (listen_id_priv != cm_id_priv) {
1290                 spin_unlock_irq(&cm_id_priv->lock);
1291                 ib_destroy_cm_id(&cm_id_priv->id);
1292                 if (!listen_id_priv)
1293                         return ERR_PTR(-EINVAL);
1294                 return &listen_id_priv->id;
1295         }
1296         cm_id_priv->id.state = IB_CM_LISTEN;
1297         spin_unlock_irq(&cm_id_priv->lock);
1298
1299         /*
1300          * A listen ID does not need to be in the xarray since it does not
1301          * receive mads, is not placed in the remote_id or remote_qpn rbtree,
1302          * and does not enter timewait.
1303          */
1304
1305         return &cm_id_priv->id;
1306 }
1307 EXPORT_SYMBOL(ib_cm_insert_listen);
1308
1309 static __be64 cm_form_tid(struct cm_id_private *cm_id_priv)
1310 {
1311         u64 hi_tid, low_tid;
1312
1313         hi_tid   = ((u64) cm_id_priv->av.port->mad_agent->hi_tid) << 32;
1314         low_tid  = (u64)cm_id_priv->id.local_id;
1315         return cpu_to_be64(hi_tid | low_tid);
1316 }
1317
1318 static void cm_format_mad_hdr(struct ib_mad_hdr *hdr,
1319                               __be16 attr_id, __be64 tid)
1320 {
1321         hdr->base_version  = IB_MGMT_BASE_VERSION;
1322         hdr->mgmt_class    = IB_MGMT_CLASS_CM;
1323         hdr->class_version = IB_CM_CLASS_VERSION;
1324         hdr->method        = IB_MGMT_METHOD_SEND;
1325         hdr->attr_id       = attr_id;
1326         hdr->tid           = tid;
1327 }
1328
1329 static void cm_format_mad_ece_hdr(struct ib_mad_hdr *hdr, __be16 attr_id,
1330                                   __be64 tid, u32 attr_mod)
1331 {
1332         cm_format_mad_hdr(hdr, attr_id, tid);
1333         hdr->attr_mod = cpu_to_be32(attr_mod);
1334 }
1335
1336 static void cm_format_req(struct cm_req_msg *req_msg,
1337                           struct cm_id_private *cm_id_priv,
1338                           struct ib_cm_req_param *param)
1339 {
1340         struct sa_path_rec *pri_path = param->primary_path;
1341         struct sa_path_rec *alt_path = param->alternate_path;
1342         bool pri_ext = false;
1343
1344         if (pri_path->rec_type == SA_PATH_REC_TYPE_OPA)
1345                 pri_ext = opa_is_extended_lid(pri_path->opa.dlid,
1346                                               pri_path->opa.slid);
1347
1348         cm_format_mad_ece_hdr(&req_msg->hdr, CM_REQ_ATTR_ID,
1349                               cm_form_tid(cm_id_priv), param->ece.attr_mod);
1350
1351         IBA_SET(CM_REQ_LOCAL_COMM_ID, req_msg,
1352                 be32_to_cpu(cm_id_priv->id.local_id));
1353         IBA_SET(CM_REQ_SERVICE_ID, req_msg, be64_to_cpu(param->service_id));
1354         IBA_SET(CM_REQ_LOCAL_CA_GUID, req_msg,
1355                 be64_to_cpu(cm_id_priv->id.device->node_guid));
1356         IBA_SET(CM_REQ_LOCAL_QPN, req_msg, param->qp_num);
1357         IBA_SET(CM_REQ_INITIATOR_DEPTH, req_msg, param->initiator_depth);
1358         IBA_SET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg,
1359                 param->remote_cm_response_timeout);
1360         cm_req_set_qp_type(req_msg, param->qp_type);
1361         IBA_SET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg, param->flow_control);
1362         IBA_SET(CM_REQ_STARTING_PSN, req_msg, param->starting_psn);
1363         IBA_SET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg,
1364                 param->local_cm_response_timeout);
1365         IBA_SET(CM_REQ_PARTITION_KEY, req_msg,
1366                 be16_to_cpu(param->primary_path->pkey));
1367         IBA_SET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg,
1368                 param->primary_path->mtu);
1369         IBA_SET(CM_REQ_MAX_CM_RETRIES, req_msg, param->max_cm_retries);
1370
1371         if (param->qp_type != IB_QPT_XRC_INI) {
1372                 IBA_SET(CM_REQ_RESPONDER_RESOURCES, req_msg,
1373                         param->responder_resources);
1374                 IBA_SET(CM_REQ_RETRY_COUNT, req_msg, param->retry_count);
1375                 IBA_SET(CM_REQ_RNR_RETRY_COUNT, req_msg,
1376                         param->rnr_retry_count);
1377                 IBA_SET(CM_REQ_SRQ, req_msg, param->srq);
1378         }
1379
1380         *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg) =
1381                 pri_path->sgid;
1382         *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg) =
1383                 pri_path->dgid;
1384         if (pri_ext) {
1385                 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg)
1386                         ->global.interface_id =
1387                         OPA_MAKE_ID(be32_to_cpu(pri_path->opa.slid));
1388                 IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg)
1389                         ->global.interface_id =
1390                         OPA_MAKE_ID(be32_to_cpu(pri_path->opa.dlid));
1391         }
1392         if (pri_path->hop_limit <= 1) {
1393                 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1394                         be16_to_cpu(pri_ext ? 0 :
1395                                               htons(ntohl(sa_path_get_slid(
1396                                                       pri_path)))));
1397                 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1398                         be16_to_cpu(pri_ext ? 0 :
1399                                               htons(ntohl(sa_path_get_dlid(
1400                                                       pri_path)))));
1401         } else {
1402                 /* Work-around until there's a way to obtain remote LID info */
1403                 IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
1404                         be16_to_cpu(IB_LID_PERMISSIVE));
1405                 IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
1406                         be16_to_cpu(IB_LID_PERMISSIVE));
1407         }
1408         IBA_SET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg,
1409                 be32_to_cpu(pri_path->flow_label));
1410         IBA_SET(CM_REQ_PRIMARY_PACKET_RATE, req_msg, pri_path->rate);
1411         IBA_SET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg, pri_path->traffic_class);
1412         IBA_SET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg, pri_path->hop_limit);
1413         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, pri_path->sl);
1414         IBA_SET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg,
1415                 (pri_path->hop_limit <= 1));
1416         IBA_SET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg,
1417                 cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1418                                pri_path->packet_life_time));
1419
1420         if (alt_path) {
1421                 bool alt_ext = false;
1422
1423                 if (alt_path->rec_type == SA_PATH_REC_TYPE_OPA)
1424                         alt_ext = opa_is_extended_lid(alt_path->opa.dlid,
1425                                                       alt_path->opa.slid);
1426
1427                 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg) =
1428                         alt_path->sgid;
1429                 *IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg) =
1430                         alt_path->dgid;
1431                 if (alt_ext) {
1432                         IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1433                                         req_msg)
1434                                 ->global.interface_id =
1435                                 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.slid));
1436                         IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_REMOTE_PORT_GID,
1437                                         req_msg)
1438                                 ->global.interface_id =
1439                                 OPA_MAKE_ID(be32_to_cpu(alt_path->opa.dlid));
1440                 }
1441                 if (alt_path->hop_limit <= 1) {
1442                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1443                                 be16_to_cpu(
1444                                         alt_ext ? 0 :
1445                                                   htons(ntohl(sa_path_get_slid(
1446                                                           alt_path)))));
1447                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1448                                 be16_to_cpu(
1449                                         alt_ext ? 0 :
1450                                                   htons(ntohl(sa_path_get_dlid(
1451                                                           alt_path)))));
1452                 } else {
1453                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
1454                                 be16_to_cpu(IB_LID_PERMISSIVE));
1455                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
1456                                 be16_to_cpu(IB_LID_PERMISSIVE));
1457                 }
1458                 IBA_SET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg,
1459                         be32_to_cpu(alt_path->flow_label));
1460                 IBA_SET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg, alt_path->rate);
1461                 IBA_SET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg,
1462                         alt_path->traffic_class);
1463                 IBA_SET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg,
1464                         alt_path->hop_limit);
1465                 IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, alt_path->sl);
1466                 IBA_SET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg,
1467                         (alt_path->hop_limit <= 1));
1468                 IBA_SET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg,
1469                         cm_ack_timeout(cm_id_priv->av.port->cm_dev->ack_delay,
1470                                        alt_path->packet_life_time));
1471         }
1472         IBA_SET(CM_REQ_VENDOR_ID, req_msg, param->ece.vendor_id);
1473
1474         if (param->private_data && param->private_data_len)
1475                 IBA_SET_MEM(CM_REQ_PRIVATE_DATA, req_msg, param->private_data,
1476                             param->private_data_len);
1477 }
1478
1479 static int cm_validate_req_param(struct ib_cm_req_param *param)
1480 {
1481         if (!param->primary_path)
1482                 return -EINVAL;
1483
1484         if (param->qp_type != IB_QPT_RC && param->qp_type != IB_QPT_UC &&
1485             param->qp_type != IB_QPT_XRC_INI)
1486                 return -EINVAL;
1487
1488         if (param->private_data &&
1489             param->private_data_len > IB_CM_REQ_PRIVATE_DATA_SIZE)
1490                 return -EINVAL;
1491
1492         if (param->alternate_path &&
1493             (param->alternate_path->pkey != param->primary_path->pkey ||
1494              param->alternate_path->mtu != param->primary_path->mtu))
1495                 return -EINVAL;
1496
1497         return 0;
1498 }
1499
1500 int ib_send_cm_req(struct ib_cm_id *cm_id,
1501                    struct ib_cm_req_param *param)
1502 {
1503         struct cm_id_private *cm_id_priv;
1504         struct cm_req_msg *req_msg;
1505         unsigned long flags;
1506         int ret;
1507
1508         ret = cm_validate_req_param(param);
1509         if (ret)
1510                 return ret;
1511
1512         /* Verify that we're not in timewait. */
1513         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
1514         spin_lock_irqsave(&cm_id_priv->lock, flags);
1515         if (cm_id->state != IB_CM_IDLE || WARN_ON(cm_id_priv->timewait_info)) {
1516                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1517                 ret = -EINVAL;
1518                 goto out;
1519         }
1520         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1521
1522         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
1523                                                             id.local_id);
1524         if (IS_ERR(cm_id_priv->timewait_info)) {
1525                 ret = PTR_ERR(cm_id_priv->timewait_info);
1526                 cm_id_priv->timewait_info = NULL;
1527                 goto out;
1528         }
1529
1530         ret = cm_init_av_by_path(param->primary_path,
1531                                  param->ppath_sgid_attr, &cm_id_priv->av,
1532                                  cm_id_priv);
1533         if (ret)
1534                 goto out;
1535         if (param->alternate_path) {
1536                 ret = cm_init_av_by_path(param->alternate_path, NULL,
1537                                          &cm_id_priv->alt_av, cm_id_priv);
1538                 if (ret)
1539                         goto out;
1540         }
1541         cm_id->service_id = param->service_id;
1542         cm_id->service_mask = ~cpu_to_be64(0);
1543         cm_id_priv->timeout_ms = cm_convert_to_ms(
1544                                     param->primary_path->packet_life_time) * 2 +
1545                                  cm_convert_to_ms(
1546                                     param->remote_cm_response_timeout);
1547         cm_id_priv->max_cm_retries = param->max_cm_retries;
1548         cm_id_priv->initiator_depth = param->initiator_depth;
1549         cm_id_priv->responder_resources = param->responder_resources;
1550         cm_id_priv->retry_count = param->retry_count;
1551         cm_id_priv->path_mtu = param->primary_path->mtu;
1552         cm_id_priv->pkey = param->primary_path->pkey;
1553         cm_id_priv->qp_type = param->qp_type;
1554
1555         ret = cm_alloc_msg(cm_id_priv, &cm_id_priv->msg);
1556         if (ret)
1557                 goto out;
1558
1559         req_msg = (struct cm_req_msg *) cm_id_priv->msg->mad;
1560         cm_format_req(req_msg, cm_id_priv, param);
1561         cm_id_priv->tid = req_msg->hdr.tid;
1562         cm_id_priv->msg->timeout_ms = cm_id_priv->timeout_ms;
1563         cm_id_priv->msg->context[1] = (void *) (unsigned long) IB_CM_REQ_SENT;
1564
1565         cm_id_priv->local_qpn = cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
1566         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
1567
1568         trace_icm_send_req(&cm_id_priv->id);
1569         spin_lock_irqsave(&cm_id_priv->lock, flags);
1570         ret = ib_post_send_mad(cm_id_priv->msg, NULL);
1571         if (ret) {
1572                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1573                 goto error2;
1574         }
1575         BUG_ON(cm_id->state != IB_CM_IDLE);
1576         cm_id->state = IB_CM_REQ_SENT;
1577         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1578         return 0;
1579
1580 error2: cm_free_msg(cm_id_priv->msg);
1581 out:    return ret;
1582 }
1583 EXPORT_SYMBOL(ib_send_cm_req);
1584
1585 static int cm_issue_rej(struct cm_port *port,
1586                         struct ib_mad_recv_wc *mad_recv_wc,
1587                         enum ib_cm_rej_reason reason,
1588                         enum cm_msg_response msg_rejected,
1589                         void *ari, u8 ari_length)
1590 {
1591         struct ib_mad_send_buf *msg = NULL;
1592         struct cm_rej_msg *rej_msg, *rcv_msg;
1593         int ret;
1594
1595         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
1596         if (ret)
1597                 return ret;
1598
1599         /* We just need common CM header information.  Cast to any message. */
1600         rcv_msg = (struct cm_rej_msg *) mad_recv_wc->recv_buf.mad;
1601         rej_msg = (struct cm_rej_msg *) msg->mad;
1602
1603         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, rcv_msg->hdr.tid);
1604         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1605                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg));
1606         IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1607                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1608         IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, msg_rejected);
1609         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1610
1611         if (ari && ari_length) {
1612                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1613                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1614         }
1615
1616         trace_icm_issue_rej(
1617                 IBA_GET(CM_REJ_LOCAL_COMM_ID, rcv_msg),
1618                 IBA_GET(CM_REJ_REMOTE_COMM_ID, rcv_msg));
1619         ret = ib_post_send_mad(msg, NULL);
1620         if (ret)
1621                 cm_free_msg(msg);
1622
1623         return ret;
1624 }
1625
1626 static bool cm_req_has_alt_path(struct cm_req_msg *req_msg)
1627 {
1628         return ((cpu_to_be16(
1629                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg))) ||
1630                 (ib_is_opa_gid(IBA_GET_MEM_PTR(CM_REQ_ALTERNATE_LOCAL_PORT_GID,
1631                                                req_msg))));
1632 }
1633
1634 static void cm_path_set_rec_type(struct ib_device *ib_device, u8 port_num,
1635                                  struct sa_path_rec *path, union ib_gid *gid)
1636 {
1637         if (ib_is_opa_gid(gid) && rdma_cap_opa_ah(ib_device, port_num))
1638                 path->rec_type = SA_PATH_REC_TYPE_OPA;
1639         else
1640                 path->rec_type = SA_PATH_REC_TYPE_IB;
1641 }
1642
1643 static void cm_format_path_lid_from_req(struct cm_req_msg *req_msg,
1644                                         struct sa_path_rec *primary_path,
1645                                         struct sa_path_rec *alt_path)
1646 {
1647         u32 lid;
1648
1649         if (primary_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1650                 sa_path_set_dlid(primary_path,
1651                                  IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
1652                                          req_msg));
1653                 sa_path_set_slid(primary_path,
1654                                  IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
1655                                          req_msg));
1656         } else {
1657                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1658                         CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg));
1659                 sa_path_set_dlid(primary_path, lid);
1660
1661                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1662                         CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg));
1663                 sa_path_set_slid(primary_path, lid);
1664         }
1665
1666         if (!cm_req_has_alt_path(req_msg))
1667                 return;
1668
1669         if (alt_path->rec_type != SA_PATH_REC_TYPE_OPA) {
1670                 sa_path_set_dlid(alt_path,
1671                                  IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
1672                                          req_msg));
1673                 sa_path_set_slid(alt_path,
1674                                  IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
1675                                          req_msg));
1676         } else {
1677                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1678                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg));
1679                 sa_path_set_dlid(alt_path, lid);
1680
1681                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
1682                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg));
1683                 sa_path_set_slid(alt_path, lid);
1684         }
1685 }
1686
1687 static void cm_format_paths_from_req(struct cm_req_msg *req_msg,
1688                                      struct sa_path_rec *primary_path,
1689                                      struct sa_path_rec *alt_path)
1690 {
1691         primary_path->dgid =
1692                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID, req_msg);
1693         primary_path->sgid =
1694                 *IBA_GET_MEM_PTR(CM_REQ_PRIMARY_REMOTE_PORT_GID, req_msg);
1695         primary_path->flow_label =
1696                 cpu_to_be32(IBA_GET(CM_REQ_PRIMARY_FLOW_LABEL, req_msg));
1697         primary_path->hop_limit = IBA_GET(CM_REQ_PRIMARY_HOP_LIMIT, req_msg);
1698         primary_path->traffic_class =
1699                 IBA_GET(CM_REQ_PRIMARY_TRAFFIC_CLASS, req_msg);
1700         primary_path->reversible = 1;
1701         primary_path->pkey =
1702                 cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1703         primary_path->sl = IBA_GET(CM_REQ_PRIMARY_SL, req_msg);
1704         primary_path->mtu_selector = IB_SA_EQ;
1705         primary_path->mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1706         primary_path->rate_selector = IB_SA_EQ;
1707         primary_path->rate = IBA_GET(CM_REQ_PRIMARY_PACKET_RATE, req_msg);
1708         primary_path->packet_life_time_selector = IB_SA_EQ;
1709         primary_path->packet_life_time =
1710                 IBA_GET(CM_REQ_PRIMARY_LOCAL_ACK_TIMEOUT, req_msg);
1711         primary_path->packet_life_time -= (primary_path->packet_life_time > 0);
1712         primary_path->service_id =
1713                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1714         if (sa_path_is_roce(primary_path))
1715                 primary_path->roce.route_resolved = false;
1716
1717         if (cm_req_has_alt_path(req_msg)) {
1718                 alt_path->dgid = *IBA_GET_MEM_PTR(
1719                         CM_REQ_ALTERNATE_LOCAL_PORT_GID, req_msg);
1720                 alt_path->sgid = *IBA_GET_MEM_PTR(
1721                         CM_REQ_ALTERNATE_REMOTE_PORT_GID, req_msg);
1722                 alt_path->flow_label = cpu_to_be32(
1723                         IBA_GET(CM_REQ_ALTERNATE_FLOW_LABEL, req_msg));
1724                 alt_path->hop_limit =
1725                         IBA_GET(CM_REQ_ALTERNATE_HOP_LIMIT, req_msg);
1726                 alt_path->traffic_class =
1727                         IBA_GET(CM_REQ_ALTERNATE_TRAFFIC_CLASS, req_msg);
1728                 alt_path->reversible = 1;
1729                 alt_path->pkey =
1730                         cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
1731                 alt_path->sl = IBA_GET(CM_REQ_ALTERNATE_SL, req_msg);
1732                 alt_path->mtu_selector = IB_SA_EQ;
1733                 alt_path->mtu =
1734                         IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
1735                 alt_path->rate_selector = IB_SA_EQ;
1736                 alt_path->rate = IBA_GET(CM_REQ_ALTERNATE_PACKET_RATE, req_msg);
1737                 alt_path->packet_life_time_selector = IB_SA_EQ;
1738                 alt_path->packet_life_time =
1739                         IBA_GET(CM_REQ_ALTERNATE_LOCAL_ACK_TIMEOUT, req_msg);
1740                 alt_path->packet_life_time -= (alt_path->packet_life_time > 0);
1741                 alt_path->service_id =
1742                         cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
1743
1744                 if (sa_path_is_roce(alt_path))
1745                         alt_path->roce.route_resolved = false;
1746         }
1747         cm_format_path_lid_from_req(req_msg, primary_path, alt_path);
1748 }
1749
1750 static u16 cm_get_bth_pkey(struct cm_work *work)
1751 {
1752         struct ib_device *ib_dev = work->port->cm_dev->ib_device;
1753         u8 port_num = work->port->port_num;
1754         u16 pkey_index = work->mad_recv_wc->wc->pkey_index;
1755         u16 pkey;
1756         int ret;
1757
1758         ret = ib_get_cached_pkey(ib_dev, port_num, pkey_index, &pkey);
1759         if (ret) {
1760                 dev_warn_ratelimited(&ib_dev->dev, "ib_cm: Couldn't retrieve pkey for incoming request (port %d, pkey index %d). %d\n",
1761                                      port_num, pkey_index, ret);
1762                 return 0;
1763         }
1764
1765         return pkey;
1766 }
1767
1768 /**
1769  * cm_opa_to_ib_sgid - Convert OPA SGID to IB SGID
1770  * ULPs (such as IPoIB) do not understand OPA GIDs and will
1771  * reject them as the local_gid will not match the sgid. Therefore,
1772  * change the pathrec's SGID to an IB SGID.
1773  *
1774  * @work: Work completion
1775  * @path: Path record
1776  */
1777 static void cm_opa_to_ib_sgid(struct cm_work *work,
1778                               struct sa_path_rec *path)
1779 {
1780         struct ib_device *dev = work->port->cm_dev->ib_device;
1781         u8 port_num = work->port->port_num;
1782
1783         if (rdma_cap_opa_ah(dev, port_num) &&
1784             (ib_is_opa_gid(&path->sgid))) {
1785                 union ib_gid sgid;
1786
1787                 if (rdma_query_gid(dev, port_num, 0, &sgid)) {
1788                         dev_warn(&dev->dev,
1789                                  "Error updating sgid in CM request\n");
1790                         return;
1791                 }
1792
1793                 path->sgid = sgid;
1794         }
1795 }
1796
1797 static void cm_format_req_event(struct cm_work *work,
1798                                 struct cm_id_private *cm_id_priv,
1799                                 struct ib_cm_id *listen_id)
1800 {
1801         struct cm_req_msg *req_msg;
1802         struct ib_cm_req_event_param *param;
1803
1804         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1805         param = &work->cm_event.param.req_rcvd;
1806         param->listen_id = listen_id;
1807         param->bth_pkey = cm_get_bth_pkey(work);
1808         param->port = cm_id_priv->av.port->port_num;
1809         param->primary_path = &work->path[0];
1810         cm_opa_to_ib_sgid(work, param->primary_path);
1811         if (cm_req_has_alt_path(req_msg)) {
1812                 param->alternate_path = &work->path[1];
1813                 cm_opa_to_ib_sgid(work, param->alternate_path);
1814         } else {
1815                 param->alternate_path = NULL;
1816         }
1817         param->remote_ca_guid =
1818                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
1819         param->remote_qkey = IBA_GET(CM_REQ_LOCAL_Q_KEY, req_msg);
1820         param->remote_qpn = IBA_GET(CM_REQ_LOCAL_QPN, req_msg);
1821         param->qp_type = cm_req_get_qp_type(req_msg);
1822         param->starting_psn = IBA_GET(CM_REQ_STARTING_PSN, req_msg);
1823         param->responder_resources = IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
1824         param->initiator_depth = IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
1825         param->local_cm_response_timeout =
1826                 IBA_GET(CM_REQ_REMOTE_CM_RESPONSE_TIMEOUT, req_msg);
1827         param->flow_control = IBA_GET(CM_REQ_END_TO_END_FLOW_CONTROL, req_msg);
1828         param->remote_cm_response_timeout =
1829                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg);
1830         param->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
1831         param->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
1832         param->srq = IBA_GET(CM_REQ_SRQ, req_msg);
1833         param->ppath_sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
1834         param->ece.vendor_id = IBA_GET(CM_REQ_VENDOR_ID, req_msg);
1835         param->ece.attr_mod = be32_to_cpu(req_msg->hdr.attr_mod);
1836
1837         work->cm_event.private_data =
1838                 IBA_GET_MEM_PTR(CM_REQ_PRIVATE_DATA, req_msg);
1839 }
1840
1841 static void cm_process_work(struct cm_id_private *cm_id_priv,
1842                             struct cm_work *work)
1843 {
1844         int ret;
1845
1846         /* We will typically only have the current event to report. */
1847         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
1848         cm_free_work(work);
1849
1850         while (!ret && !atomic_add_negative(-1, &cm_id_priv->work_count)) {
1851                 spin_lock_irq(&cm_id_priv->lock);
1852                 work = cm_dequeue_work(cm_id_priv);
1853                 spin_unlock_irq(&cm_id_priv->lock);
1854                 if (!work)
1855                         return;
1856
1857                 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id,
1858                                                 &work->cm_event);
1859                 cm_free_work(work);
1860         }
1861         cm_deref_id(cm_id_priv);
1862         if (ret)
1863                 cm_destroy_id(&cm_id_priv->id, ret);
1864 }
1865
1866 static void cm_format_mra(struct cm_mra_msg *mra_msg,
1867                           struct cm_id_private *cm_id_priv,
1868                           enum cm_msg_response msg_mraed, u8 service_timeout,
1869                           const void *private_data, u8 private_data_len)
1870 {
1871         cm_format_mad_hdr(&mra_msg->hdr, CM_MRA_ATTR_ID, cm_id_priv->tid);
1872         IBA_SET(CM_MRA_MESSAGE_MRAED, mra_msg, msg_mraed);
1873         IBA_SET(CM_MRA_LOCAL_COMM_ID, mra_msg,
1874                 be32_to_cpu(cm_id_priv->id.local_id));
1875         IBA_SET(CM_MRA_REMOTE_COMM_ID, mra_msg,
1876                 be32_to_cpu(cm_id_priv->id.remote_id));
1877         IBA_SET(CM_MRA_SERVICE_TIMEOUT, mra_msg, service_timeout);
1878
1879         if (private_data && private_data_len)
1880                 IBA_SET_MEM(CM_MRA_PRIVATE_DATA, mra_msg, private_data,
1881                             private_data_len);
1882 }
1883
1884 static void cm_format_rej(struct cm_rej_msg *rej_msg,
1885                           struct cm_id_private *cm_id_priv,
1886                           enum ib_cm_rej_reason reason, void *ari,
1887                           u8 ari_length, const void *private_data,
1888                           u8 private_data_len, enum ib_cm_state state)
1889 {
1890         lockdep_assert_held(&cm_id_priv->lock);
1891
1892         cm_format_mad_hdr(&rej_msg->hdr, CM_REJ_ATTR_ID, cm_id_priv->tid);
1893         IBA_SET(CM_REJ_REMOTE_COMM_ID, rej_msg,
1894                 be32_to_cpu(cm_id_priv->id.remote_id));
1895
1896         switch (state) {
1897         case IB_CM_REQ_RCVD:
1898                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg, be32_to_cpu(0));
1899                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1900                 break;
1901         case IB_CM_MRA_REQ_SENT:
1902                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1903                         be32_to_cpu(cm_id_priv->id.local_id));
1904                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REQ);
1905                 break;
1906         case IB_CM_REP_RCVD:
1907         case IB_CM_MRA_REP_SENT:
1908                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1909                         be32_to_cpu(cm_id_priv->id.local_id));
1910                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg, CM_MSG_RESPONSE_REP);
1911                 break;
1912         default:
1913                 IBA_SET(CM_REJ_LOCAL_COMM_ID, rej_msg,
1914                         be32_to_cpu(cm_id_priv->id.local_id));
1915                 IBA_SET(CM_REJ_MESSAGE_REJECTED, rej_msg,
1916                         CM_MSG_RESPONSE_OTHER);
1917                 break;
1918         }
1919
1920         IBA_SET(CM_REJ_REASON, rej_msg, reason);
1921         if (ari && ari_length) {
1922                 IBA_SET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg, ari_length);
1923                 IBA_SET_MEM(CM_REJ_ARI, rej_msg, ari, ari_length);
1924         }
1925
1926         if (private_data && private_data_len)
1927                 IBA_SET_MEM(CM_REJ_PRIVATE_DATA, rej_msg, private_data,
1928                             private_data_len);
1929 }
1930
1931 static void cm_dup_req_handler(struct cm_work *work,
1932                                struct cm_id_private *cm_id_priv)
1933 {
1934         struct ib_mad_send_buf *msg = NULL;
1935         int ret;
1936
1937         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
1938                         counter[CM_REQ_COUNTER]);
1939
1940         /* Quick state check to discard duplicate REQs. */
1941         spin_lock_irq(&cm_id_priv->lock);
1942         if (cm_id_priv->id.state == IB_CM_REQ_RCVD) {
1943                 spin_unlock_irq(&cm_id_priv->lock);
1944                 return;
1945         }
1946         spin_unlock_irq(&cm_id_priv->lock);
1947
1948         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
1949         if (ret)
1950                 return;
1951
1952         spin_lock_irq(&cm_id_priv->lock);
1953         switch (cm_id_priv->id.state) {
1954         case IB_CM_MRA_REQ_SENT:
1955                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
1956                               CM_MSG_RESPONSE_REQ, cm_id_priv->service_timeout,
1957                               cm_id_priv->private_data,
1958                               cm_id_priv->private_data_len);
1959                 break;
1960         case IB_CM_TIMEWAIT:
1961                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv,
1962                               IB_CM_REJ_STALE_CONN, NULL, 0, NULL, 0,
1963                               IB_CM_TIMEWAIT);
1964                 break;
1965         default:
1966                 goto unlock;
1967         }
1968         spin_unlock_irq(&cm_id_priv->lock);
1969
1970         trace_icm_send_dup_req(&cm_id_priv->id);
1971         ret = ib_post_send_mad(msg, NULL);
1972         if (ret)
1973                 goto free;
1974         return;
1975
1976 unlock: spin_unlock_irq(&cm_id_priv->lock);
1977 free:   cm_free_msg(msg);
1978 }
1979
1980 static struct cm_id_private * cm_match_req(struct cm_work *work,
1981                                            struct cm_id_private *cm_id_priv)
1982 {
1983         struct cm_id_private *listen_cm_id_priv, *cur_cm_id_priv;
1984         struct cm_timewait_info *timewait_info;
1985         struct cm_req_msg *req_msg;
1986
1987         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
1988
1989         /* Check for possible duplicate REQ. */
1990         spin_lock_irq(&cm.lock);
1991         timewait_info = cm_insert_remote_id(cm_id_priv->timewait_info);
1992         if (timewait_info) {
1993                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
1994                                            timewait_info->work.remote_id);
1995                 spin_unlock_irq(&cm.lock);
1996                 if (cur_cm_id_priv) {
1997                         cm_dup_req_handler(work, cur_cm_id_priv);
1998                         cm_deref_id(cur_cm_id_priv);
1999                 }
2000                 return NULL;
2001         }
2002
2003         /* Check for stale connections. */
2004         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2005         if (timewait_info) {
2006                 cm_remove_remote(cm_id_priv);
2007                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2008                                            timewait_info->work.remote_id);
2009
2010                 spin_unlock_irq(&cm.lock);
2011                 cm_issue_rej(work->port, work->mad_recv_wc,
2012                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REQ,
2013                              NULL, 0);
2014                 if (cur_cm_id_priv) {
2015                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2016                         cm_deref_id(cur_cm_id_priv);
2017                 }
2018                 return NULL;
2019         }
2020
2021         /* Find matching listen request. */
2022         listen_cm_id_priv = cm_find_listen(
2023                 cm_id_priv->id.device,
2024                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg)));
2025         if (!listen_cm_id_priv) {
2026                 cm_remove_remote(cm_id_priv);
2027                 spin_unlock_irq(&cm.lock);
2028                 cm_issue_rej(work->port, work->mad_recv_wc,
2029                              IB_CM_REJ_INVALID_SERVICE_ID, CM_MSG_RESPONSE_REQ,
2030                              NULL, 0);
2031                 return NULL;
2032         }
2033         spin_unlock_irq(&cm.lock);
2034         return listen_cm_id_priv;
2035 }
2036
2037 /*
2038  * Work-around for inter-subnet connections.  If the LIDs are permissive,
2039  * we need to override the LID/SL data in the REQ with the LID information
2040  * in the work completion.
2041  */
2042 static void cm_process_routed_req(struct cm_req_msg *req_msg, struct ib_wc *wc)
2043 {
2044         if (!IBA_GET(CM_REQ_PRIMARY_SUBNET_LOCAL, req_msg)) {
2045                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_LOCAL_PORT_LID,
2046                                         req_msg)) == IB_LID_PERMISSIVE) {
2047                         IBA_SET(CM_REQ_PRIMARY_LOCAL_PORT_LID, req_msg,
2048                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2049                         IBA_SET(CM_REQ_PRIMARY_SL, req_msg, wc->sl);
2050                 }
2051
2052                 if (cpu_to_be16(IBA_GET(CM_REQ_PRIMARY_REMOTE_PORT_LID,
2053                                         req_msg)) == IB_LID_PERMISSIVE)
2054                         IBA_SET(CM_REQ_PRIMARY_REMOTE_PORT_LID, req_msg,
2055                                 wc->dlid_path_bits);
2056         }
2057
2058         if (!IBA_GET(CM_REQ_ALTERNATE_SUBNET_LOCAL, req_msg)) {
2059                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_LOCAL_PORT_LID,
2060                                         req_msg)) == IB_LID_PERMISSIVE) {
2061                         IBA_SET(CM_REQ_ALTERNATE_LOCAL_PORT_LID, req_msg,
2062                                 be16_to_cpu(ib_lid_be16(wc->slid)));
2063                         IBA_SET(CM_REQ_ALTERNATE_SL, req_msg, wc->sl);
2064                 }
2065
2066                 if (cpu_to_be16(IBA_GET(CM_REQ_ALTERNATE_REMOTE_PORT_LID,
2067                                         req_msg)) == IB_LID_PERMISSIVE)
2068                         IBA_SET(CM_REQ_ALTERNATE_REMOTE_PORT_LID, req_msg,
2069                                 wc->dlid_path_bits);
2070         }
2071 }
2072
2073 static int cm_req_handler(struct cm_work *work)
2074 {
2075         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
2076         struct cm_req_msg *req_msg;
2077         const struct ib_global_route *grh;
2078         const struct ib_gid_attr *gid_attr;
2079         int ret;
2080
2081         req_msg = (struct cm_req_msg *)work->mad_recv_wc->recv_buf.mad;
2082
2083         cm_id_priv =
2084                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
2085         if (IS_ERR(cm_id_priv))
2086                 return PTR_ERR(cm_id_priv);
2087
2088         cm_id_priv->id.remote_id =
2089                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_COMM_ID, req_msg));
2090         cm_id_priv->id.service_id =
2091                 cpu_to_be64(IBA_GET(CM_REQ_SERVICE_ID, req_msg));
2092         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
2093         cm_id_priv->tid = req_msg->hdr.tid;
2094         cm_id_priv->timeout_ms = cm_convert_to_ms(
2095                 IBA_GET(CM_REQ_LOCAL_CM_RESPONSE_TIMEOUT, req_msg));
2096         cm_id_priv->max_cm_retries = IBA_GET(CM_REQ_MAX_CM_RETRIES, req_msg);
2097         cm_id_priv->remote_qpn =
2098                 cpu_to_be32(IBA_GET(CM_REQ_LOCAL_QPN, req_msg));
2099         cm_id_priv->initiator_depth =
2100                 IBA_GET(CM_REQ_RESPONDER_RESOURCES, req_msg);
2101         cm_id_priv->responder_resources =
2102                 IBA_GET(CM_REQ_INITIATOR_DEPTH, req_msg);
2103         cm_id_priv->path_mtu = IBA_GET(CM_REQ_PATH_PACKET_PAYLOAD_MTU, req_msg);
2104         cm_id_priv->pkey = cpu_to_be16(IBA_GET(CM_REQ_PARTITION_KEY, req_msg));
2105         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REQ_STARTING_PSN, req_msg));
2106         cm_id_priv->retry_count = IBA_GET(CM_REQ_RETRY_COUNT, req_msg);
2107         cm_id_priv->rnr_retry_count = IBA_GET(CM_REQ_RNR_RETRY_COUNT, req_msg);
2108         cm_id_priv->qp_type = cm_req_get_qp_type(req_msg);
2109
2110         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
2111                                       work->mad_recv_wc->recv_buf.grh,
2112                                       &cm_id_priv->av);
2113         if (ret)
2114                 goto destroy;
2115         cm_id_priv->timewait_info = cm_create_timewait_info(cm_id_priv->
2116                                                             id.local_id);
2117         if (IS_ERR(cm_id_priv->timewait_info)) {
2118                 ret = PTR_ERR(cm_id_priv->timewait_info);
2119                 cm_id_priv->timewait_info = NULL;
2120                 goto destroy;
2121         }
2122         cm_id_priv->timewait_info->work.remote_id = cm_id_priv->id.remote_id;
2123         cm_id_priv->timewait_info->remote_ca_guid =
2124                 cpu_to_be64(IBA_GET(CM_REQ_LOCAL_CA_GUID, req_msg));
2125         cm_id_priv->timewait_info->remote_qpn = cm_id_priv->remote_qpn;
2126
2127         /*
2128          * Note that the ID pointer is not in the xarray at this point,
2129          * so this set is only visible to the local thread.
2130          */
2131         cm_id_priv->id.state = IB_CM_REQ_RCVD;
2132
2133         listen_cm_id_priv = cm_match_req(work, cm_id_priv);
2134         if (!listen_cm_id_priv) {
2135                 trace_icm_no_listener_err(&cm_id_priv->id);
2136                 cm_id_priv->id.state = IB_CM_IDLE;
2137                 ret = -EINVAL;
2138                 goto destroy;
2139         }
2140
2141         cm_process_routed_req(req_msg, work->mad_recv_wc->wc);
2142
2143         memset(&work->path[0], 0, sizeof(work->path[0]));
2144         if (cm_req_has_alt_path(req_msg))
2145                 memset(&work->path[1], 0, sizeof(work->path[1]));
2146         grh = rdma_ah_read_grh(&cm_id_priv->av.ah_attr);
2147         gid_attr = grh->sgid_attr;
2148
2149         if (gid_attr &&
2150             rdma_protocol_roce(work->port->cm_dev->ib_device,
2151                                work->port->port_num)) {
2152                 work->path[0].rec_type =
2153                         sa_conv_gid_to_pathrec_type(gid_attr->gid_type);
2154         } else {
2155                 cm_path_set_rec_type(
2156                         work->port->cm_dev->ib_device, work->port->port_num,
2157                         &work->path[0],
2158                         IBA_GET_MEM_PTR(CM_REQ_PRIMARY_LOCAL_PORT_GID,
2159                                         req_msg));
2160         }
2161         if (cm_req_has_alt_path(req_msg))
2162                 work->path[1].rec_type = work->path[0].rec_type;
2163         cm_format_paths_from_req(req_msg, &work->path[0],
2164                                  &work->path[1]);
2165         if (cm_id_priv->av.ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE)
2166                 sa_path_set_dmac(&work->path[0],
2167                                  cm_id_priv->av.ah_attr.roce.dmac);
2168         work->path[0].hop_limit = grh->hop_limit;
2169         ret = cm_init_av_by_path(&work->path[0], gid_attr, &cm_id_priv->av,
2170                                  cm_id_priv);
2171         if (ret) {
2172                 int err;
2173
2174                 err = rdma_query_gid(work->port->cm_dev->ib_device,
2175                                      work->port->port_num, 0,
2176                                      &work->path[0].sgid);
2177                 if (err)
2178                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2179                                        NULL, 0, NULL, 0);
2180                 else
2181                         ib_send_cm_rej(&cm_id_priv->id, IB_CM_REJ_INVALID_GID,
2182                                        &work->path[0].sgid,
2183                                        sizeof(work->path[0].sgid),
2184                                        NULL, 0);
2185                 goto rejected;
2186         }
2187         if (cm_req_has_alt_path(req_msg)) {
2188                 ret = cm_init_av_by_path(&work->path[1], NULL,
2189                                          &cm_id_priv->alt_av, cm_id_priv);
2190                 if (ret) {
2191                         ib_send_cm_rej(&cm_id_priv->id,
2192                                        IB_CM_REJ_INVALID_ALT_GID,
2193                                        &work->path[0].sgid,
2194                                        sizeof(work->path[0].sgid), NULL, 0);
2195                         goto rejected;
2196                 }
2197         }
2198
2199         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
2200         cm_id_priv->id.context = listen_cm_id_priv->id.context;
2201         cm_format_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
2202
2203         /* Now MAD handlers can see the new ID */
2204         spin_lock_irq(&cm_id_priv->lock);
2205         cm_finalize_id(cm_id_priv);
2206
2207         /* Refcount belongs to the event, pairs with cm_process_work() */
2208         refcount_inc(&cm_id_priv->refcount);
2209         cm_queue_work_unlock(cm_id_priv, work);
2210         /*
2211          * Since this ID was just created and was not made visible to other MAD
2212          * handlers until the cm_finalize_id() above we know that the
2213          * cm_process_work() will deliver the event and the listen_cm_id
2214          * embedded in the event can be derefed here.
2215          */
2216         cm_deref_id(listen_cm_id_priv);
2217         return 0;
2218
2219 rejected:
2220         cm_deref_id(listen_cm_id_priv);
2221 destroy:
2222         ib_destroy_cm_id(&cm_id_priv->id);
2223         return ret;
2224 }
2225
2226 static void cm_format_rep(struct cm_rep_msg *rep_msg,
2227                           struct cm_id_private *cm_id_priv,
2228                           struct ib_cm_rep_param *param)
2229 {
2230         cm_format_mad_ece_hdr(&rep_msg->hdr, CM_REP_ATTR_ID, cm_id_priv->tid,
2231                               param->ece.attr_mod);
2232         IBA_SET(CM_REP_LOCAL_COMM_ID, rep_msg,
2233                 be32_to_cpu(cm_id_priv->id.local_id));
2234         IBA_SET(CM_REP_REMOTE_COMM_ID, rep_msg,
2235                 be32_to_cpu(cm_id_priv->id.remote_id));
2236         IBA_SET(CM_REP_STARTING_PSN, rep_msg, param->starting_psn);
2237         IBA_SET(CM_REP_RESPONDER_RESOURCES, rep_msg,
2238                 param->responder_resources);
2239         IBA_SET(CM_REP_TARGET_ACK_DELAY, rep_msg,
2240                 cm_id_priv->av.port->cm_dev->ack_delay);
2241         IBA_SET(CM_REP_FAILOVER_ACCEPTED, rep_msg, param->failover_accepted);
2242         IBA_SET(CM_REP_RNR_RETRY_COUNT, rep_msg, param->rnr_retry_count);
2243         IBA_SET(CM_REP_LOCAL_CA_GUID, rep_msg,
2244                 be64_to_cpu(cm_id_priv->id.device->node_guid));
2245
2246         if (cm_id_priv->qp_type != IB_QPT_XRC_TGT) {
2247                 IBA_SET(CM_REP_INITIATOR_DEPTH, rep_msg,
2248                         param->initiator_depth);
2249                 IBA_SET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg,
2250                         param->flow_control);
2251                 IBA_SET(CM_REP_SRQ, rep_msg, param->srq);
2252                 IBA_SET(CM_REP_LOCAL_QPN, rep_msg, param->qp_num);
2253         } else {
2254                 IBA_SET(CM_REP_SRQ, rep_msg, 1);
2255                 IBA_SET(CM_REP_LOCAL_EE_CONTEXT_NUMBER, rep_msg, param->qp_num);
2256         }
2257
2258         IBA_SET(CM_REP_VENDOR_ID_L, rep_msg, param->ece.vendor_id);
2259         IBA_SET(CM_REP_VENDOR_ID_M, rep_msg, param->ece.vendor_id >> 8);
2260         IBA_SET(CM_REP_VENDOR_ID_H, rep_msg, param->ece.vendor_id >> 16);
2261
2262         if (param->private_data && param->private_data_len)
2263                 IBA_SET_MEM(CM_REP_PRIVATE_DATA, rep_msg, param->private_data,
2264                             param->private_data_len);
2265 }
2266
2267 int ib_send_cm_rep(struct ib_cm_id *cm_id,
2268                    struct ib_cm_rep_param *param)
2269 {
2270         struct cm_id_private *cm_id_priv;
2271         struct ib_mad_send_buf *msg;
2272         struct cm_rep_msg *rep_msg;
2273         unsigned long flags;
2274         int ret;
2275
2276         if (param->private_data &&
2277             param->private_data_len > IB_CM_REP_PRIVATE_DATA_SIZE)
2278                 return -EINVAL;
2279
2280         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2281         spin_lock_irqsave(&cm_id_priv->lock, flags);
2282         if (cm_id->state != IB_CM_REQ_RCVD &&
2283             cm_id->state != IB_CM_MRA_REQ_SENT) {
2284                 trace_icm_send_rep_err(cm_id_priv->id.local_id, cm_id->state);
2285                 ret = -EINVAL;
2286                 goto out;
2287         }
2288
2289         ret = cm_alloc_msg(cm_id_priv, &msg);
2290         if (ret)
2291                 goto out;
2292
2293         rep_msg = (struct cm_rep_msg *) msg->mad;
2294         cm_format_rep(rep_msg, cm_id_priv, param);
2295         msg->timeout_ms = cm_id_priv->timeout_ms;
2296         msg->context[1] = (void *) (unsigned long) IB_CM_REP_SENT;
2297
2298         trace_icm_send_rep(cm_id);
2299         ret = ib_post_send_mad(msg, NULL);
2300         if (ret) {
2301                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2302                 cm_free_msg(msg);
2303                 return ret;
2304         }
2305
2306         cm_id->state = IB_CM_REP_SENT;
2307         cm_id_priv->msg = msg;
2308         cm_id_priv->initiator_depth = param->initiator_depth;
2309         cm_id_priv->responder_resources = param->responder_resources;
2310         cm_id_priv->rq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2311         WARN_ONCE(param->qp_num & 0xFF000000,
2312                   "IBTA declares QPN to be 24 bits, but it is 0x%X\n",
2313                   param->qp_num);
2314         cm_id_priv->local_qpn = cpu_to_be32(param->qp_num & 0xFFFFFF);
2315
2316 out:    spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2317         return ret;
2318 }
2319 EXPORT_SYMBOL(ib_send_cm_rep);
2320
2321 static void cm_format_rtu(struct cm_rtu_msg *rtu_msg,
2322                           struct cm_id_private *cm_id_priv,
2323                           const void *private_data,
2324                           u8 private_data_len)
2325 {
2326         cm_format_mad_hdr(&rtu_msg->hdr, CM_RTU_ATTR_ID, cm_id_priv->tid);
2327         IBA_SET(CM_RTU_LOCAL_COMM_ID, rtu_msg,
2328                 be32_to_cpu(cm_id_priv->id.local_id));
2329         IBA_SET(CM_RTU_REMOTE_COMM_ID, rtu_msg,
2330                 be32_to_cpu(cm_id_priv->id.remote_id));
2331
2332         if (private_data && private_data_len)
2333                 IBA_SET_MEM(CM_RTU_PRIVATE_DATA, rtu_msg, private_data,
2334                             private_data_len);
2335 }
2336
2337 int ib_send_cm_rtu(struct ib_cm_id *cm_id,
2338                    const void *private_data,
2339                    u8 private_data_len)
2340 {
2341         struct cm_id_private *cm_id_priv;
2342         struct ib_mad_send_buf *msg;
2343         unsigned long flags;
2344         void *data;
2345         int ret;
2346
2347         if (private_data && private_data_len > IB_CM_RTU_PRIVATE_DATA_SIZE)
2348                 return -EINVAL;
2349
2350         data = cm_copy_private_data(private_data, private_data_len);
2351         if (IS_ERR(data))
2352                 return PTR_ERR(data);
2353
2354         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
2355         spin_lock_irqsave(&cm_id_priv->lock, flags);
2356         if (cm_id->state != IB_CM_REP_RCVD &&
2357             cm_id->state != IB_CM_MRA_REP_SENT) {
2358                 trace_icm_send_cm_rtu_err(cm_id);
2359                 ret = -EINVAL;
2360                 goto error;
2361         }
2362
2363         ret = cm_alloc_msg(cm_id_priv, &msg);
2364         if (ret)
2365                 goto error;
2366
2367         cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2368                       private_data, private_data_len);
2369
2370         trace_icm_send_rtu(cm_id);
2371         ret = ib_post_send_mad(msg, NULL);
2372         if (ret) {
2373                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2374                 cm_free_msg(msg);
2375                 kfree(data);
2376                 return ret;
2377         }
2378
2379         cm_id->state = IB_CM_ESTABLISHED;
2380         cm_set_private_data(cm_id_priv, data, private_data_len);
2381         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2382         return 0;
2383
2384 error:  spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2385         kfree(data);
2386         return ret;
2387 }
2388 EXPORT_SYMBOL(ib_send_cm_rtu);
2389
2390 static void cm_format_rep_event(struct cm_work *work, enum ib_qp_type qp_type)
2391 {
2392         struct cm_rep_msg *rep_msg;
2393         struct ib_cm_rep_event_param *param;
2394
2395         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2396         param = &work->cm_event.param.rep_rcvd;
2397         param->remote_ca_guid =
2398                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2399         param->remote_qkey = IBA_GET(CM_REP_LOCAL_Q_KEY, rep_msg);
2400         param->remote_qpn = be32_to_cpu(cm_rep_get_qpn(rep_msg, qp_type));
2401         param->starting_psn = IBA_GET(CM_REP_STARTING_PSN, rep_msg);
2402         param->responder_resources = IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2403         param->initiator_depth = IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2404         param->target_ack_delay = IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2405         param->failover_accepted = IBA_GET(CM_REP_FAILOVER_ACCEPTED, rep_msg);
2406         param->flow_control = IBA_GET(CM_REP_END_TO_END_FLOW_CONTROL, rep_msg);
2407         param->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2408         param->srq = IBA_GET(CM_REP_SRQ, rep_msg);
2409         param->ece.vendor_id = IBA_GET(CM_REP_VENDOR_ID_H, rep_msg) << 16;
2410         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_M, rep_msg) << 8;
2411         param->ece.vendor_id |= IBA_GET(CM_REP_VENDOR_ID_L, rep_msg);
2412         param->ece.attr_mod = be32_to_cpu(rep_msg->hdr.attr_mod);
2413
2414         work->cm_event.private_data =
2415                 IBA_GET_MEM_PTR(CM_REP_PRIVATE_DATA, rep_msg);
2416 }
2417
2418 static void cm_dup_rep_handler(struct cm_work *work)
2419 {
2420         struct cm_id_private *cm_id_priv;
2421         struct cm_rep_msg *rep_msg;
2422         struct ib_mad_send_buf *msg = NULL;
2423         int ret;
2424
2425         rep_msg = (struct cm_rep_msg *) work->mad_recv_wc->recv_buf.mad;
2426         cm_id_priv = cm_acquire_id(
2427                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)),
2428                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg)));
2429         if (!cm_id_priv)
2430                 return;
2431
2432         atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2433                         counter[CM_REP_COUNTER]);
2434         ret = cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg);
2435         if (ret)
2436                 goto deref;
2437
2438         spin_lock_irq(&cm_id_priv->lock);
2439         if (cm_id_priv->id.state == IB_CM_ESTABLISHED)
2440                 cm_format_rtu((struct cm_rtu_msg *) msg->mad, cm_id_priv,
2441                               cm_id_priv->private_data,
2442                               cm_id_priv->private_data_len);
2443         else if (cm_id_priv->id.state == IB_CM_MRA_REP_SENT)
2444                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
2445                               CM_MSG_RESPONSE_REP, cm_id_priv->service_timeout,
2446                               cm_id_priv->private_data,
2447                               cm_id_priv->private_data_len);
2448         else
2449                 goto unlock;
2450         spin_unlock_irq(&cm_id_priv->lock);
2451
2452         trace_icm_send_dup_rep(&cm_id_priv->id);
2453         ret = ib_post_send_mad(msg, NULL);
2454         if (ret)
2455                 goto free;
2456         goto deref;
2457
2458 unlock: spin_unlock_irq(&cm_id_priv->lock);
2459 free:   cm_free_msg(msg);
2460 deref:  cm_deref_id(cm_id_priv);
2461 }
2462
2463 static int cm_rep_handler(struct cm_work *work)
2464 {
2465         struct cm_id_private *cm_id_priv;
2466         struct cm_rep_msg *rep_msg;
2467         int ret;
2468         struct cm_id_private *cur_cm_id_priv;
2469         struct cm_timewait_info *timewait_info;
2470
2471         rep_msg = (struct cm_rep_msg *)work->mad_recv_wc->recv_buf.mad;
2472         cm_id_priv = cm_acquire_id(
2473                 cpu_to_be32(IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg)), 0);
2474         if (!cm_id_priv) {
2475                 cm_dup_rep_handler(work);
2476                 trace_icm_remote_no_priv_err(
2477                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2478                 return -EINVAL;
2479         }
2480
2481         cm_format_rep_event(work, cm_id_priv->qp_type);
2482
2483         spin_lock_irq(&cm_id_priv->lock);
2484         switch (cm_id_priv->id.state) {
2485         case IB_CM_REQ_SENT:
2486         case IB_CM_MRA_REQ_RCVD:
2487                 break;
2488         default:
2489                 ret = -EINVAL;
2490                 trace_icm_rep_unknown_err(
2491                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2492                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg),
2493                         cm_id_priv->id.state);
2494                 spin_unlock_irq(&cm_id_priv->lock);
2495                 goto error;
2496         }
2497
2498         cm_id_priv->timewait_info->work.remote_id =
2499                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2500         cm_id_priv->timewait_info->remote_ca_guid =
2501                 cpu_to_be64(IBA_GET(CM_REP_LOCAL_CA_GUID, rep_msg));
2502         cm_id_priv->timewait_info->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2503
2504         spin_lock(&cm.lock);
2505         /* Check for duplicate REP. */
2506         if (cm_insert_remote_id(cm_id_priv->timewait_info)) {
2507                 spin_unlock(&cm.lock);
2508                 spin_unlock_irq(&cm_id_priv->lock);
2509                 ret = -EINVAL;
2510                 trace_icm_insert_failed_err(
2511                          IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2512                 goto error;
2513         }
2514         /* Check for a stale connection. */
2515         timewait_info = cm_insert_remote_qpn(cm_id_priv->timewait_info);
2516         if (timewait_info) {
2517                 cm_remove_remote(cm_id_priv);
2518                 cur_cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
2519                                            timewait_info->work.remote_id);
2520
2521                 spin_unlock(&cm.lock);
2522                 spin_unlock_irq(&cm_id_priv->lock);
2523                 cm_issue_rej(work->port, work->mad_recv_wc,
2524                              IB_CM_REJ_STALE_CONN, CM_MSG_RESPONSE_REP,
2525                              NULL, 0);
2526                 ret = -EINVAL;
2527                 trace_icm_staleconn_err(
2528                         IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg),
2529                         IBA_GET(CM_REP_REMOTE_COMM_ID, rep_msg));
2530
2531                 if (cur_cm_id_priv) {
2532                         ib_send_cm_dreq(&cur_cm_id_priv->id, NULL, 0);
2533                         cm_deref_id(cur_cm_id_priv);
2534                 }
2535
2536                 goto error;
2537         }
2538         spin_unlock(&cm.lock);
2539
2540         cm_id_priv->id.state = IB_CM_REP_RCVD;
2541         cm_id_priv->id.remote_id =
2542                 cpu_to_be32(IBA_GET(CM_REP_LOCAL_COMM_ID, rep_msg));
2543         cm_id_priv->remote_qpn = cm_rep_get_qpn(rep_msg, cm_id_priv->qp_type);
2544         cm_id_priv->initiator_depth =
2545                 IBA_GET(CM_REP_RESPONDER_RESOURCES, rep_msg);
2546         cm_id_priv->responder_resources =
2547                 IBA_GET(CM_REP_INITIATOR_DEPTH, rep_msg);
2548         cm_id_priv->sq_psn = cpu_to_be32(IBA_GET(CM_REP_STARTING_PSN, rep_msg));
2549         cm_id_priv->rnr_retry_count = IBA_GET(CM_REP_RNR_RETRY_COUNT, rep_msg);
2550         cm_id_priv->target_ack_delay =
2551                 IBA_GET(CM_REP_TARGET_ACK_DELAY, rep_msg);
2552         cm_id_priv->av.timeout =
2553                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2554                                        cm_id_priv->av.timeout - 1);
2555         cm_id_priv->alt_av.timeout =
2556                         cm_ack_timeout(cm_id_priv->target_ack_delay,
2557                                        cm_id_priv->alt_av.timeout - 1);
2558
2559         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2560         cm_queue_work_unlock(cm_id_priv, work);
2561         return 0;
2562
2563 error:
2564         cm_deref_id(cm_id_priv);
2565         return ret;
2566 }
2567
2568 static int cm_establish_handler(struct cm_work *work)
2569 {
2570         struct cm_id_private *cm_id_priv;
2571
2572         /* See comment in cm_establish about lookup. */
2573         cm_id_priv = cm_acquire_id(work->local_id, work->remote_id);
2574         if (!cm_id_priv)
2575                 return -EINVAL;
2576
2577         spin_lock_irq(&cm_id_priv->lock);
2578         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2579                 spin_unlock_irq(&cm_id_priv->lock);
2580                 goto out;
2581         }
2582
2583         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2584         cm_queue_work_unlock(cm_id_priv, work);
2585         return 0;
2586 out:
2587         cm_deref_id(cm_id_priv);
2588         return -EINVAL;
2589 }
2590
2591 static int cm_rtu_handler(struct cm_work *work)
2592 {
2593         struct cm_id_private *cm_id_priv;
2594         struct cm_rtu_msg *rtu_msg;
2595
2596         rtu_msg = (struct cm_rtu_msg *)work->mad_recv_wc->recv_buf.mad;
2597         cm_id_priv = cm_acquire_id(
2598                 cpu_to_be32(IBA_GET(CM_RTU_REMOTE_COMM_ID, rtu_msg)),
2599                 cpu_to_be32(IBA_GET(CM_RTU_LOCAL_COMM_ID, rtu_msg)));
2600         if (!cm_id_priv)
2601                 return -EINVAL;
2602
2603         work->cm_event.private_data =
2604                 IBA_GET_MEM_PTR(CM_RTU_PRIVATE_DATA, rtu_msg);
2605
2606         spin_lock_irq(&cm_id_priv->lock);
2607         if (cm_id_priv->id.state != IB_CM_REP_SENT &&
2608             cm_id_priv->id.state != IB_CM_MRA_REP_RCVD) {
2609                 spin_unlock_irq(&cm_id_priv->lock);
2610                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2611                                 counter[CM_RTU_COUNTER]);
2612                 goto out;
2613         }
2614         cm_id_priv->id.state = IB_CM_ESTABLISHED;
2615
2616         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2617         cm_queue_work_unlock(cm_id_priv, work);
2618         return 0;
2619 out:
2620         cm_deref_id(cm_id_priv);
2621         return -EINVAL;
2622 }
2623
2624 static void cm_format_dreq(struct cm_dreq_msg *dreq_msg,
2625                           struct cm_id_private *cm_id_priv,
2626                           const void *private_data,
2627                           u8 private_data_len)
2628 {
2629         cm_format_mad_hdr(&dreq_msg->hdr, CM_DREQ_ATTR_ID,
2630                           cm_form_tid(cm_id_priv));
2631         IBA_SET(CM_DREQ_LOCAL_COMM_ID, dreq_msg,
2632                 be32_to_cpu(cm_id_priv->id.local_id));
2633         IBA_SET(CM_DREQ_REMOTE_COMM_ID, dreq_msg,
2634                 be32_to_cpu(cm_id_priv->id.remote_id));
2635         IBA_SET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg,
2636                 be32_to_cpu(cm_id_priv->remote_qpn));
2637
2638         if (private_data && private_data_len)
2639                 IBA_SET_MEM(CM_DREQ_PRIVATE_DATA, dreq_msg, private_data,
2640                             private_data_len);
2641 }
2642
2643 static int cm_send_dreq_locked(struct cm_id_private *cm_id_priv,
2644                                const void *private_data, u8 private_data_len)
2645 {
2646         struct ib_mad_send_buf *msg;
2647         int ret;
2648
2649         lockdep_assert_held(&cm_id_priv->lock);
2650
2651         if (private_data && private_data_len > IB_CM_DREQ_PRIVATE_DATA_SIZE)
2652                 return -EINVAL;
2653
2654         if (cm_id_priv->id.state != IB_CM_ESTABLISHED) {
2655                 trace_icm_dreq_skipped(&cm_id_priv->id);
2656                 return -EINVAL;
2657         }
2658
2659         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2660             cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2661                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2662
2663         ret = cm_alloc_msg(cm_id_priv, &msg);
2664         if (ret) {
2665                 cm_enter_timewait(cm_id_priv);
2666                 return ret;
2667         }
2668
2669         cm_format_dreq((struct cm_dreq_msg *) msg->mad, cm_id_priv,
2670                        private_data, private_data_len);
2671         msg->timeout_ms = cm_id_priv->timeout_ms;
2672         msg->context[1] = (void *) (unsigned long) IB_CM_DREQ_SENT;
2673
2674         trace_icm_send_dreq(&cm_id_priv->id);
2675         ret = ib_post_send_mad(msg, NULL);
2676         if (ret) {
2677                 cm_enter_timewait(cm_id_priv);
2678                 cm_free_msg(msg);
2679                 return ret;
2680         }
2681
2682         cm_id_priv->id.state = IB_CM_DREQ_SENT;
2683         cm_id_priv->msg = msg;
2684         return 0;
2685 }
2686
2687 int ib_send_cm_dreq(struct ib_cm_id *cm_id, const void *private_data,
2688                     u8 private_data_len)
2689 {
2690         struct cm_id_private *cm_id_priv =
2691                 container_of(cm_id, struct cm_id_private, id);
2692         unsigned long flags;
2693         int ret;
2694
2695         spin_lock_irqsave(&cm_id_priv->lock, flags);
2696         ret = cm_send_dreq_locked(cm_id_priv, private_data, private_data_len);
2697         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2698         return ret;
2699 }
2700 EXPORT_SYMBOL(ib_send_cm_dreq);
2701
2702 static void cm_format_drep(struct cm_drep_msg *drep_msg,
2703                           struct cm_id_private *cm_id_priv,
2704                           const void *private_data,
2705                           u8 private_data_len)
2706 {
2707         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, cm_id_priv->tid);
2708         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2709                 be32_to_cpu(cm_id_priv->id.local_id));
2710         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2711                 be32_to_cpu(cm_id_priv->id.remote_id));
2712
2713         if (private_data && private_data_len)
2714                 IBA_SET_MEM(CM_DREP_PRIVATE_DATA, drep_msg, private_data,
2715                             private_data_len);
2716 }
2717
2718 static int cm_send_drep_locked(struct cm_id_private *cm_id_priv,
2719                                void *private_data, u8 private_data_len)
2720 {
2721         struct ib_mad_send_buf *msg;
2722         int ret;
2723
2724         lockdep_assert_held(&cm_id_priv->lock);
2725
2726         if (private_data && private_data_len > IB_CM_DREP_PRIVATE_DATA_SIZE)
2727                 return -EINVAL;
2728
2729         if (cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2730                 trace_icm_send_drep_err(&cm_id_priv->id);
2731                 kfree(private_data);
2732                 return -EINVAL;
2733         }
2734
2735         cm_set_private_data(cm_id_priv, private_data, private_data_len);
2736         cm_enter_timewait(cm_id_priv);
2737
2738         ret = cm_alloc_msg(cm_id_priv, &msg);
2739         if (ret)
2740                 return ret;
2741
2742         cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2743                        private_data, private_data_len);
2744
2745         trace_icm_send_drep(&cm_id_priv->id);
2746         ret = ib_post_send_mad(msg, NULL);
2747         if (ret) {
2748                 cm_free_msg(msg);
2749                 return ret;
2750         }
2751         return 0;
2752 }
2753
2754 int ib_send_cm_drep(struct ib_cm_id *cm_id, const void *private_data,
2755                     u8 private_data_len)
2756 {
2757         struct cm_id_private *cm_id_priv =
2758                 container_of(cm_id, struct cm_id_private, id);
2759         unsigned long flags;
2760         void *data;
2761         int ret;
2762
2763         data = cm_copy_private_data(private_data, private_data_len);
2764         if (IS_ERR(data))
2765                 return PTR_ERR(data);
2766
2767         spin_lock_irqsave(&cm_id_priv->lock, flags);
2768         ret = cm_send_drep_locked(cm_id_priv, data, private_data_len);
2769         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2770         return ret;
2771 }
2772 EXPORT_SYMBOL(ib_send_cm_drep);
2773
2774 static int cm_issue_drep(struct cm_port *port,
2775                          struct ib_mad_recv_wc *mad_recv_wc)
2776 {
2777         struct ib_mad_send_buf *msg = NULL;
2778         struct cm_dreq_msg *dreq_msg;
2779         struct cm_drep_msg *drep_msg;
2780         int ret;
2781
2782         ret = cm_alloc_response_msg(port, mad_recv_wc, &msg);
2783         if (ret)
2784                 return ret;
2785
2786         dreq_msg = (struct cm_dreq_msg *) mad_recv_wc->recv_buf.mad;
2787         drep_msg = (struct cm_drep_msg *) msg->mad;
2788
2789         cm_format_mad_hdr(&drep_msg->hdr, CM_DREP_ATTR_ID, dreq_msg->hdr.tid);
2790         IBA_SET(CM_DREP_REMOTE_COMM_ID, drep_msg,
2791                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg));
2792         IBA_SET(CM_DREP_LOCAL_COMM_ID, drep_msg,
2793                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2794
2795         trace_icm_issue_drep(
2796                 IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2797                 IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2798         ret = ib_post_send_mad(msg, NULL);
2799         if (ret)
2800                 cm_free_msg(msg);
2801
2802         return ret;
2803 }
2804
2805 static int cm_dreq_handler(struct cm_work *work)
2806 {
2807         struct cm_id_private *cm_id_priv;
2808         struct cm_dreq_msg *dreq_msg;
2809         struct ib_mad_send_buf *msg = NULL;
2810
2811         dreq_msg = (struct cm_dreq_msg *)work->mad_recv_wc->recv_buf.mad;
2812         cm_id_priv = cm_acquire_id(
2813                 cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg)),
2814                 cpu_to_be32(IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg)));
2815         if (!cm_id_priv) {
2816                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2817                                 counter[CM_DREQ_COUNTER]);
2818                 cm_issue_drep(work->port, work->mad_recv_wc);
2819                 trace_icm_no_priv_err(
2820                         IBA_GET(CM_DREQ_LOCAL_COMM_ID, dreq_msg),
2821                         IBA_GET(CM_DREQ_REMOTE_COMM_ID, dreq_msg));
2822                 return -EINVAL;
2823         }
2824
2825         work->cm_event.private_data =
2826                 IBA_GET_MEM_PTR(CM_DREQ_PRIVATE_DATA, dreq_msg);
2827
2828         spin_lock_irq(&cm_id_priv->lock);
2829         if (cm_id_priv->local_qpn !=
2830             cpu_to_be32(IBA_GET(CM_DREQ_REMOTE_QPN_EECN, dreq_msg)))
2831                 goto unlock;
2832
2833         switch (cm_id_priv->id.state) {
2834         case IB_CM_REP_SENT:
2835         case IB_CM_DREQ_SENT:
2836                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2837                 break;
2838         case IB_CM_ESTABLISHED:
2839                 if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT ||
2840                     cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
2841                         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2842                 break;
2843         case IB_CM_MRA_REP_RCVD:
2844                 break;
2845         case IB_CM_TIMEWAIT:
2846                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2847                                 counter[CM_DREQ_COUNTER]);
2848                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
2849                 if (IS_ERR(msg))
2850                         goto unlock;
2851
2852                 cm_format_drep((struct cm_drep_msg *) msg->mad, cm_id_priv,
2853                                cm_id_priv->private_data,
2854                                cm_id_priv->private_data_len);
2855                 spin_unlock_irq(&cm_id_priv->lock);
2856
2857                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
2858                     ib_post_send_mad(msg, NULL))
2859                         cm_free_msg(msg);
2860                 goto deref;
2861         case IB_CM_DREQ_RCVD:
2862                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
2863                                 counter[CM_DREQ_COUNTER]);
2864                 goto unlock;
2865         default:
2866                 trace_icm_dreq_unknown_err(&cm_id_priv->id);
2867                 goto unlock;
2868         }
2869         cm_id_priv->id.state = IB_CM_DREQ_RCVD;
2870         cm_id_priv->tid = dreq_msg->hdr.tid;
2871         cm_queue_work_unlock(cm_id_priv, work);
2872         return 0;
2873
2874 unlock: spin_unlock_irq(&cm_id_priv->lock);
2875 deref:  cm_deref_id(cm_id_priv);
2876         return -EINVAL;
2877 }
2878
2879 static int cm_drep_handler(struct cm_work *work)
2880 {
2881         struct cm_id_private *cm_id_priv;
2882         struct cm_drep_msg *drep_msg;
2883
2884         drep_msg = (struct cm_drep_msg *)work->mad_recv_wc->recv_buf.mad;
2885         cm_id_priv = cm_acquire_id(
2886                 cpu_to_be32(IBA_GET(CM_DREP_REMOTE_COMM_ID, drep_msg)),
2887                 cpu_to_be32(IBA_GET(CM_DREP_LOCAL_COMM_ID, drep_msg)));
2888         if (!cm_id_priv)
2889                 return -EINVAL;
2890
2891         work->cm_event.private_data =
2892                 IBA_GET_MEM_PTR(CM_DREP_PRIVATE_DATA, drep_msg);
2893
2894         spin_lock_irq(&cm_id_priv->lock);
2895         if (cm_id_priv->id.state != IB_CM_DREQ_SENT &&
2896             cm_id_priv->id.state != IB_CM_DREQ_RCVD) {
2897                 spin_unlock_irq(&cm_id_priv->lock);
2898                 goto out;
2899         }
2900         cm_enter_timewait(cm_id_priv);
2901
2902         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
2903         cm_queue_work_unlock(cm_id_priv, work);
2904         return 0;
2905 out:
2906         cm_deref_id(cm_id_priv);
2907         return -EINVAL;
2908 }
2909
2910 static int cm_send_rej_locked(struct cm_id_private *cm_id_priv,
2911                               enum ib_cm_rej_reason reason, void *ari,
2912                               u8 ari_length, const void *private_data,
2913                               u8 private_data_len)
2914 {
2915         enum ib_cm_state state = cm_id_priv->id.state;
2916         struct ib_mad_send_buf *msg;
2917         int ret;
2918
2919         lockdep_assert_held(&cm_id_priv->lock);
2920
2921         if ((private_data && private_data_len > IB_CM_REJ_PRIVATE_DATA_SIZE) ||
2922             (ari && ari_length > IB_CM_REJ_ARI_LENGTH))
2923                 return -EINVAL;
2924
2925         switch (state) {
2926         case IB_CM_REQ_SENT:
2927         case IB_CM_MRA_REQ_RCVD:
2928         case IB_CM_REQ_RCVD:
2929         case IB_CM_MRA_REQ_SENT:
2930         case IB_CM_REP_RCVD:
2931         case IB_CM_MRA_REP_SENT:
2932                 cm_reset_to_idle(cm_id_priv);
2933                 ret = cm_alloc_msg(cm_id_priv, &msg);
2934                 if (ret)
2935                         return ret;
2936                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2937                               ari, ari_length, private_data, private_data_len,
2938                               state);
2939                 break;
2940         case IB_CM_REP_SENT:
2941         case IB_CM_MRA_REP_RCVD:
2942                 cm_enter_timewait(cm_id_priv);
2943                 ret = cm_alloc_msg(cm_id_priv, &msg);
2944                 if (ret)
2945                         return ret;
2946                 cm_format_rej((struct cm_rej_msg *)msg->mad, cm_id_priv, reason,
2947                               ari, ari_length, private_data, private_data_len,
2948                               state);
2949                 break;
2950         default:
2951                 trace_icm_send_unknown_rej_err(&cm_id_priv->id);
2952                 return -EINVAL;
2953         }
2954
2955         trace_icm_send_rej(&cm_id_priv->id, reason);
2956         ret = ib_post_send_mad(msg, NULL);
2957         if (ret) {
2958                 cm_free_msg(msg);
2959                 return ret;
2960         }
2961
2962         return 0;
2963 }
2964
2965 int ib_send_cm_rej(struct ib_cm_id *cm_id, enum ib_cm_rej_reason reason,
2966                    void *ari, u8 ari_length, const void *private_data,
2967                    u8 private_data_len)
2968 {
2969         struct cm_id_private *cm_id_priv =
2970                 container_of(cm_id, struct cm_id_private, id);
2971         unsigned long flags;
2972         int ret;
2973
2974         spin_lock_irqsave(&cm_id_priv->lock, flags);
2975         ret = cm_send_rej_locked(cm_id_priv, reason, ari, ari_length,
2976                                  private_data, private_data_len);
2977         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
2978         return ret;
2979 }
2980 EXPORT_SYMBOL(ib_send_cm_rej);
2981
2982 static void cm_format_rej_event(struct cm_work *work)
2983 {
2984         struct cm_rej_msg *rej_msg;
2985         struct ib_cm_rej_event_param *param;
2986
2987         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
2988         param = &work->cm_event.param.rej_rcvd;
2989         param->ari = IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg);
2990         param->ari_length = IBA_GET(CM_REJ_REJECTED_INFO_LENGTH, rej_msg);
2991         param->reason = IBA_GET(CM_REJ_REASON, rej_msg);
2992         work->cm_event.private_data =
2993                 IBA_GET_MEM_PTR(CM_REJ_PRIVATE_DATA, rej_msg);
2994 }
2995
2996 static struct cm_id_private * cm_acquire_rejected_id(struct cm_rej_msg *rej_msg)
2997 {
2998         struct cm_id_private *cm_id_priv;
2999         __be32 remote_id;
3000
3001         remote_id = cpu_to_be32(IBA_GET(CM_REJ_LOCAL_COMM_ID, rej_msg));
3002
3003         if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_TIMEOUT) {
3004                 cm_id_priv = cm_find_remote_id(
3005                         *((__be64 *)IBA_GET_MEM_PTR(CM_REJ_ARI, rej_msg)),
3006                         remote_id);
3007         } else if (IBA_GET(CM_REJ_MESSAGE_REJECTED, rej_msg) ==
3008                    CM_MSG_RESPONSE_REQ)
3009                 cm_id_priv = cm_acquire_id(
3010                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3011                         0);
3012         else
3013                 cm_id_priv = cm_acquire_id(
3014                         cpu_to_be32(IBA_GET(CM_REJ_REMOTE_COMM_ID, rej_msg)),
3015                         remote_id);
3016
3017         return cm_id_priv;
3018 }
3019
3020 static int cm_rej_handler(struct cm_work *work)
3021 {
3022         struct cm_id_private *cm_id_priv;
3023         struct cm_rej_msg *rej_msg;
3024
3025         rej_msg = (struct cm_rej_msg *)work->mad_recv_wc->recv_buf.mad;
3026         cm_id_priv = cm_acquire_rejected_id(rej_msg);
3027         if (!cm_id_priv)
3028                 return -EINVAL;
3029
3030         cm_format_rej_event(work);
3031
3032         spin_lock_irq(&cm_id_priv->lock);
3033         switch (cm_id_priv->id.state) {
3034         case IB_CM_REQ_SENT:
3035         case IB_CM_MRA_REQ_RCVD:
3036         case IB_CM_REP_SENT:
3037         case IB_CM_MRA_REP_RCVD:
3038                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3039                 fallthrough;
3040         case IB_CM_REQ_RCVD:
3041         case IB_CM_MRA_REQ_SENT:
3042                 if (IBA_GET(CM_REJ_REASON, rej_msg) == IB_CM_REJ_STALE_CONN)
3043                         cm_enter_timewait(cm_id_priv);
3044                 else
3045                         cm_reset_to_idle(cm_id_priv);
3046                 break;
3047         case IB_CM_DREQ_SENT:
3048                 ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3049                 fallthrough;
3050         case IB_CM_REP_RCVD:
3051         case IB_CM_MRA_REP_SENT:
3052                 cm_enter_timewait(cm_id_priv);
3053                 break;
3054         case IB_CM_ESTABLISHED:
3055                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT ||
3056                     cm_id_priv->id.lap_state == IB_CM_LAP_SENT) {
3057                         if (cm_id_priv->id.lap_state == IB_CM_LAP_SENT)
3058                                 ib_cancel_mad(cm_id_priv->av.port->mad_agent,
3059                                               cm_id_priv->msg);
3060                         cm_enter_timewait(cm_id_priv);
3061                         break;
3062                 }
3063                 fallthrough;
3064         default:
3065                 trace_icm_rej_unknown_err(&cm_id_priv->id);
3066                 spin_unlock_irq(&cm_id_priv->lock);
3067                 goto out;
3068         }
3069
3070         cm_queue_work_unlock(cm_id_priv, work);
3071         return 0;
3072 out:
3073         cm_deref_id(cm_id_priv);
3074         return -EINVAL;
3075 }
3076
3077 int ib_send_cm_mra(struct ib_cm_id *cm_id,
3078                    u8 service_timeout,
3079                    const void *private_data,
3080                    u8 private_data_len)
3081 {
3082         struct cm_id_private *cm_id_priv;
3083         struct ib_mad_send_buf *msg;
3084         enum ib_cm_state cm_state;
3085         enum ib_cm_lap_state lap_state;
3086         enum cm_msg_response msg_response;
3087         void *data;
3088         unsigned long flags;
3089         int ret;
3090
3091         if (private_data && private_data_len > IB_CM_MRA_PRIVATE_DATA_SIZE)
3092                 return -EINVAL;
3093
3094         data = cm_copy_private_data(private_data, private_data_len);
3095         if (IS_ERR(data))
3096                 return PTR_ERR(data);
3097
3098         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3099
3100         spin_lock_irqsave(&cm_id_priv->lock, flags);
3101         switch(cm_id_priv->id.state) {
3102         case IB_CM_REQ_RCVD:
3103                 cm_state = IB_CM_MRA_REQ_SENT;
3104                 lap_state = cm_id->lap_state;
3105                 msg_response = CM_MSG_RESPONSE_REQ;
3106                 break;
3107         case IB_CM_REP_RCVD:
3108                 cm_state = IB_CM_MRA_REP_SENT;
3109                 lap_state = cm_id->lap_state;
3110                 msg_response = CM_MSG_RESPONSE_REP;
3111                 break;
3112         case IB_CM_ESTABLISHED:
3113                 if (cm_id->lap_state == IB_CM_LAP_RCVD) {
3114                         cm_state = cm_id->state;
3115                         lap_state = IB_CM_MRA_LAP_SENT;
3116                         msg_response = CM_MSG_RESPONSE_OTHER;
3117                         break;
3118                 }
3119                 fallthrough;
3120         default:
3121                 trace_icm_send_mra_unknown_err(&cm_id_priv->id);
3122                 ret = -EINVAL;
3123                 goto error1;
3124         }
3125
3126         if (!(service_timeout & IB_CM_MRA_FLAG_DELAY)) {
3127                 ret = cm_alloc_msg(cm_id_priv, &msg);
3128                 if (ret)
3129                         goto error1;
3130
3131                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3132                               msg_response, service_timeout,
3133                               private_data, private_data_len);
3134                 trace_icm_send_mra(cm_id);
3135                 ret = ib_post_send_mad(msg, NULL);
3136                 if (ret)
3137                         goto error2;
3138         }
3139
3140         cm_id->state = cm_state;
3141         cm_id->lap_state = lap_state;
3142         cm_id_priv->service_timeout = service_timeout;
3143         cm_set_private_data(cm_id_priv, data, private_data_len);
3144         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3145         return 0;
3146
3147 error1: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3148         kfree(data);
3149         return ret;
3150
3151 error2: spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3152         kfree(data);
3153         cm_free_msg(msg);
3154         return ret;
3155 }
3156 EXPORT_SYMBOL(ib_send_cm_mra);
3157
3158 static struct cm_id_private * cm_acquire_mraed_id(struct cm_mra_msg *mra_msg)
3159 {
3160         switch (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg)) {
3161         case CM_MSG_RESPONSE_REQ:
3162                 return cm_acquire_id(
3163                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3164                         0);
3165         case CM_MSG_RESPONSE_REP:
3166         case CM_MSG_RESPONSE_OTHER:
3167                 return cm_acquire_id(
3168                         cpu_to_be32(IBA_GET(CM_MRA_REMOTE_COMM_ID, mra_msg)),
3169                         cpu_to_be32(IBA_GET(CM_MRA_LOCAL_COMM_ID, mra_msg)));
3170         default:
3171                 return NULL;
3172         }
3173 }
3174
3175 static int cm_mra_handler(struct cm_work *work)
3176 {
3177         struct cm_id_private *cm_id_priv;
3178         struct cm_mra_msg *mra_msg;
3179         int timeout;
3180
3181         mra_msg = (struct cm_mra_msg *)work->mad_recv_wc->recv_buf.mad;
3182         cm_id_priv = cm_acquire_mraed_id(mra_msg);
3183         if (!cm_id_priv)
3184                 return -EINVAL;
3185
3186         work->cm_event.private_data =
3187                 IBA_GET_MEM_PTR(CM_MRA_PRIVATE_DATA, mra_msg);
3188         work->cm_event.param.mra_rcvd.service_timeout =
3189                 IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg);
3190         timeout = cm_convert_to_ms(IBA_GET(CM_MRA_SERVICE_TIMEOUT, mra_msg)) +
3191                   cm_convert_to_ms(cm_id_priv->av.timeout);
3192
3193         spin_lock_irq(&cm_id_priv->lock);
3194         switch (cm_id_priv->id.state) {
3195         case IB_CM_REQ_SENT:
3196                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3197                             CM_MSG_RESPONSE_REQ ||
3198                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3199                                   cm_id_priv->msg, timeout))
3200                         goto out;
3201                 cm_id_priv->id.state = IB_CM_MRA_REQ_RCVD;
3202                 break;
3203         case IB_CM_REP_SENT:
3204                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3205                             CM_MSG_RESPONSE_REP ||
3206                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3207                                   cm_id_priv->msg, timeout))
3208                         goto out;
3209                 cm_id_priv->id.state = IB_CM_MRA_REP_RCVD;
3210                 break;
3211         case IB_CM_ESTABLISHED:
3212                 if (IBA_GET(CM_MRA_MESSAGE_MRAED, mra_msg) !=
3213                             CM_MSG_RESPONSE_OTHER ||
3214                     cm_id_priv->id.lap_state != IB_CM_LAP_SENT ||
3215                     ib_modify_mad(cm_id_priv->av.port->mad_agent,
3216                                   cm_id_priv->msg, timeout)) {
3217                         if (cm_id_priv->id.lap_state == IB_CM_MRA_LAP_RCVD)
3218                                 atomic_long_inc(&work->port->
3219                                                 counter_group[CM_RECV_DUPLICATES].
3220                                                 counter[CM_MRA_COUNTER]);
3221                         goto out;
3222                 }
3223                 cm_id_priv->id.lap_state = IB_CM_MRA_LAP_RCVD;
3224                 break;
3225         case IB_CM_MRA_REQ_RCVD:
3226         case IB_CM_MRA_REP_RCVD:
3227                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3228                                 counter[CM_MRA_COUNTER]);
3229                 fallthrough;
3230         default:
3231                 trace_icm_mra_unknown_err(&cm_id_priv->id);
3232                 goto out;
3233         }
3234
3235         cm_id_priv->msg->context[1] = (void *) (unsigned long)
3236                                       cm_id_priv->id.state;
3237         cm_queue_work_unlock(cm_id_priv, work);
3238         return 0;
3239 out:
3240         spin_unlock_irq(&cm_id_priv->lock);
3241         cm_deref_id(cm_id_priv);
3242         return -EINVAL;
3243 }
3244
3245 static void cm_format_path_lid_from_lap(struct cm_lap_msg *lap_msg,
3246                                         struct sa_path_rec *path)
3247 {
3248         u32 lid;
3249
3250         if (path->rec_type != SA_PATH_REC_TYPE_OPA) {
3251                 sa_path_set_dlid(path, IBA_GET(CM_LAP_ALTERNATE_LOCAL_PORT_LID,
3252                                                lap_msg));
3253                 sa_path_set_slid(path, IBA_GET(CM_LAP_ALTERNATE_REMOTE_PORT_LID,
3254                                                lap_msg));
3255         } else {
3256                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3257                         CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg));
3258                 sa_path_set_dlid(path, lid);
3259
3260                 lid = opa_get_lid_from_gid(IBA_GET_MEM_PTR(
3261                         CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg));
3262                 sa_path_set_slid(path, lid);
3263         }
3264 }
3265
3266 static void cm_format_path_from_lap(struct cm_id_private *cm_id_priv,
3267                                     struct sa_path_rec *path,
3268                                     struct cm_lap_msg *lap_msg)
3269 {
3270         path->dgid = *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID, lap_msg);
3271         path->sgid =
3272                 *IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_REMOTE_PORT_GID, lap_msg);
3273         path->flow_label =
3274                 cpu_to_be32(IBA_GET(CM_LAP_ALTERNATE_FLOW_LABEL, lap_msg));
3275         path->hop_limit = IBA_GET(CM_LAP_ALTERNATE_HOP_LIMIT, lap_msg);
3276         path->traffic_class = IBA_GET(CM_LAP_ALTERNATE_TRAFFIC_CLASS, lap_msg);
3277         path->reversible = 1;
3278         path->pkey = cm_id_priv->pkey;
3279         path->sl = IBA_GET(CM_LAP_ALTERNATE_SL, lap_msg);
3280         path->mtu_selector = IB_SA_EQ;
3281         path->mtu = cm_id_priv->path_mtu;
3282         path->rate_selector = IB_SA_EQ;
3283         path->rate = IBA_GET(CM_LAP_ALTERNATE_PACKET_RATE, lap_msg);
3284         path->packet_life_time_selector = IB_SA_EQ;
3285         path->packet_life_time =
3286                 IBA_GET(CM_LAP_ALTERNATE_LOCAL_ACK_TIMEOUT, lap_msg);
3287         path->packet_life_time -= (path->packet_life_time > 0);
3288         cm_format_path_lid_from_lap(lap_msg, path);
3289 }
3290
3291 static int cm_lap_handler(struct cm_work *work)
3292 {
3293         struct cm_id_private *cm_id_priv;
3294         struct cm_lap_msg *lap_msg;
3295         struct ib_cm_lap_event_param *param;
3296         struct ib_mad_send_buf *msg = NULL;
3297         int ret;
3298
3299         /* Currently Alternate path messages are not supported for
3300          * RoCE link layer.
3301          */
3302         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3303                                work->port->port_num))
3304                 return -EINVAL;
3305
3306         /* todo: verify LAP request and send reject APR if invalid. */
3307         lap_msg = (struct cm_lap_msg *)work->mad_recv_wc->recv_buf.mad;
3308         cm_id_priv = cm_acquire_id(
3309                 cpu_to_be32(IBA_GET(CM_LAP_REMOTE_COMM_ID, lap_msg)),
3310                 cpu_to_be32(IBA_GET(CM_LAP_LOCAL_COMM_ID, lap_msg)));
3311         if (!cm_id_priv)
3312                 return -EINVAL;
3313
3314         param = &work->cm_event.param.lap_rcvd;
3315         memset(&work->path[0], 0, sizeof(work->path[1]));
3316         cm_path_set_rec_type(work->port->cm_dev->ib_device,
3317                              work->port->port_num, &work->path[0],
3318                              IBA_GET_MEM_PTR(CM_LAP_ALTERNATE_LOCAL_PORT_GID,
3319                                              lap_msg));
3320         param->alternate_path = &work->path[0];
3321         cm_format_path_from_lap(cm_id_priv, param->alternate_path, lap_msg);
3322         work->cm_event.private_data =
3323                 IBA_GET_MEM_PTR(CM_LAP_PRIVATE_DATA, lap_msg);
3324
3325         spin_lock_irq(&cm_id_priv->lock);
3326         if (cm_id_priv->id.state != IB_CM_ESTABLISHED)
3327                 goto unlock;
3328
3329         switch (cm_id_priv->id.lap_state) {
3330         case IB_CM_LAP_UNINIT:
3331         case IB_CM_LAP_IDLE:
3332                 break;
3333         case IB_CM_MRA_LAP_SENT:
3334                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3335                                 counter[CM_LAP_COUNTER]);
3336                 msg = cm_alloc_response_msg_no_ah(work->port, work->mad_recv_wc);
3337                 if (IS_ERR(msg))
3338                         goto unlock;
3339
3340                 cm_format_mra((struct cm_mra_msg *) msg->mad, cm_id_priv,
3341                               CM_MSG_RESPONSE_OTHER,
3342                               cm_id_priv->service_timeout,
3343                               cm_id_priv->private_data,
3344                               cm_id_priv->private_data_len);
3345                 spin_unlock_irq(&cm_id_priv->lock);
3346
3347                 if (cm_create_response_msg_ah(work->port, work->mad_recv_wc, msg) ||
3348                     ib_post_send_mad(msg, NULL))
3349                         cm_free_msg(msg);
3350                 goto deref;
3351         case IB_CM_LAP_RCVD:
3352                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3353                                 counter[CM_LAP_COUNTER]);
3354                 goto unlock;
3355         default:
3356                 goto unlock;
3357         }
3358
3359         ret = cm_init_av_for_lap(work->port, work->mad_recv_wc->wc,
3360                                  work->mad_recv_wc->recv_buf.grh,
3361                                  &cm_id_priv->av);
3362         if (ret)
3363                 goto unlock;
3364
3365         ret = cm_init_av_by_path(param->alternate_path, NULL,
3366                                  &cm_id_priv->alt_av, cm_id_priv);
3367         if (ret)
3368                 goto unlock;
3369
3370         cm_id_priv->id.lap_state = IB_CM_LAP_RCVD;
3371         cm_id_priv->tid = lap_msg->hdr.tid;
3372         cm_queue_work_unlock(cm_id_priv, work);
3373         return 0;
3374
3375 unlock: spin_unlock_irq(&cm_id_priv->lock);
3376 deref:  cm_deref_id(cm_id_priv);
3377         return -EINVAL;
3378 }
3379
3380 static int cm_apr_handler(struct cm_work *work)
3381 {
3382         struct cm_id_private *cm_id_priv;
3383         struct cm_apr_msg *apr_msg;
3384
3385         /* Currently Alternate path messages are not supported for
3386          * RoCE link layer.
3387          */
3388         if (rdma_protocol_roce(work->port->cm_dev->ib_device,
3389                                work->port->port_num))
3390                 return -EINVAL;
3391
3392         apr_msg = (struct cm_apr_msg *)work->mad_recv_wc->recv_buf.mad;
3393         cm_id_priv = cm_acquire_id(
3394                 cpu_to_be32(IBA_GET(CM_APR_REMOTE_COMM_ID, apr_msg)),
3395                 cpu_to_be32(IBA_GET(CM_APR_LOCAL_COMM_ID, apr_msg)));
3396         if (!cm_id_priv)
3397                 return -EINVAL; /* Unmatched reply. */
3398
3399         work->cm_event.param.apr_rcvd.ap_status =
3400                 IBA_GET(CM_APR_AR_STATUS, apr_msg);
3401         work->cm_event.param.apr_rcvd.apr_info =
3402                 IBA_GET_MEM_PTR(CM_APR_ADDITIONAL_INFORMATION, apr_msg);
3403         work->cm_event.param.apr_rcvd.info_len =
3404                 IBA_GET(CM_APR_ADDITIONAL_INFORMATION_LENGTH, apr_msg);
3405         work->cm_event.private_data =
3406                 IBA_GET_MEM_PTR(CM_APR_PRIVATE_DATA, apr_msg);
3407
3408         spin_lock_irq(&cm_id_priv->lock);
3409         if (cm_id_priv->id.state != IB_CM_ESTABLISHED ||
3410             (cm_id_priv->id.lap_state != IB_CM_LAP_SENT &&
3411              cm_id_priv->id.lap_state != IB_CM_MRA_LAP_RCVD)) {
3412                 spin_unlock_irq(&cm_id_priv->lock);
3413                 goto out;
3414         }
3415         cm_id_priv->id.lap_state = IB_CM_LAP_IDLE;
3416         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3417         cm_id_priv->msg = NULL;
3418         cm_queue_work_unlock(cm_id_priv, work);
3419         return 0;
3420 out:
3421         cm_deref_id(cm_id_priv);
3422         return -EINVAL;
3423 }
3424
3425 static int cm_timewait_handler(struct cm_work *work)
3426 {
3427         struct cm_timewait_info *timewait_info;
3428         struct cm_id_private *cm_id_priv;
3429
3430         timewait_info = container_of(work, struct cm_timewait_info, work);
3431         spin_lock_irq(&cm.lock);
3432         list_del(&timewait_info->list);
3433         spin_unlock_irq(&cm.lock);
3434
3435         cm_id_priv = cm_acquire_id(timewait_info->work.local_id,
3436                                    timewait_info->work.remote_id);
3437         if (!cm_id_priv)
3438                 return -EINVAL;
3439
3440         spin_lock_irq(&cm_id_priv->lock);
3441         if (cm_id_priv->id.state != IB_CM_TIMEWAIT ||
3442             cm_id_priv->remote_qpn != timewait_info->remote_qpn) {
3443                 spin_unlock_irq(&cm_id_priv->lock);
3444                 goto out;
3445         }
3446         cm_id_priv->id.state = IB_CM_IDLE;
3447         cm_queue_work_unlock(cm_id_priv, work);
3448         return 0;
3449 out:
3450         cm_deref_id(cm_id_priv);
3451         return -EINVAL;
3452 }
3453
3454 static void cm_format_sidr_req(struct cm_sidr_req_msg *sidr_req_msg,
3455                                struct cm_id_private *cm_id_priv,
3456                                struct ib_cm_sidr_req_param *param)
3457 {
3458         cm_format_mad_hdr(&sidr_req_msg->hdr, CM_SIDR_REQ_ATTR_ID,
3459                           cm_form_tid(cm_id_priv));
3460         IBA_SET(CM_SIDR_REQ_REQUESTID, sidr_req_msg,
3461                 be32_to_cpu(cm_id_priv->id.local_id));
3462         IBA_SET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg,
3463                 be16_to_cpu(param->path->pkey));
3464         IBA_SET(CM_SIDR_REQ_SERVICEID, sidr_req_msg,
3465                 be64_to_cpu(param->service_id));
3466
3467         if (param->private_data && param->private_data_len)
3468                 IBA_SET_MEM(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg,
3469                             param->private_data, param->private_data_len);
3470 }
3471
3472 int ib_send_cm_sidr_req(struct ib_cm_id *cm_id,
3473                         struct ib_cm_sidr_req_param *param)
3474 {
3475         struct cm_id_private *cm_id_priv;
3476         struct ib_mad_send_buf *msg;
3477         unsigned long flags;
3478         int ret;
3479
3480         if (!param->path || (param->private_data &&
3481              param->private_data_len > IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE))
3482                 return -EINVAL;
3483
3484         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3485         ret = cm_init_av_by_path(param->path, param->sgid_attr,
3486                                  &cm_id_priv->av,
3487                                  cm_id_priv);
3488         if (ret)
3489                 goto out;
3490
3491         cm_id->service_id = param->service_id;
3492         cm_id->service_mask = ~cpu_to_be64(0);
3493         cm_id_priv->timeout_ms = param->timeout_ms;
3494         cm_id_priv->max_cm_retries = param->max_cm_retries;
3495         ret = cm_alloc_msg(cm_id_priv, &msg);
3496         if (ret)
3497                 goto out;
3498
3499         cm_format_sidr_req((struct cm_sidr_req_msg *) msg->mad, cm_id_priv,
3500                            param);
3501         msg->timeout_ms = cm_id_priv->timeout_ms;
3502         msg->context[1] = (void *) (unsigned long) IB_CM_SIDR_REQ_SENT;
3503
3504         spin_lock_irqsave(&cm_id_priv->lock, flags);
3505         if (cm_id->state == IB_CM_IDLE) {
3506                 trace_icm_send_sidr_req(&cm_id_priv->id);
3507                 ret = ib_post_send_mad(msg, NULL);
3508         } else {
3509                 ret = -EINVAL;
3510         }
3511
3512         if (ret) {
3513                 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3514                 cm_free_msg(msg);
3515                 goto out;
3516         }
3517         cm_id->state = IB_CM_SIDR_REQ_SENT;
3518         cm_id_priv->msg = msg;
3519         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3520 out:
3521         return ret;
3522 }
3523 EXPORT_SYMBOL(ib_send_cm_sidr_req);
3524
3525 static void cm_format_sidr_req_event(struct cm_work *work,
3526                                      const struct cm_id_private *rx_cm_id,
3527                                      struct ib_cm_id *listen_id)
3528 {
3529         struct cm_sidr_req_msg *sidr_req_msg;
3530         struct ib_cm_sidr_req_event_param *param;
3531
3532         sidr_req_msg = (struct cm_sidr_req_msg *)
3533                                 work->mad_recv_wc->recv_buf.mad;
3534         param = &work->cm_event.param.sidr_req_rcvd;
3535         param->pkey = IBA_GET(CM_SIDR_REQ_PARTITION_KEY, sidr_req_msg);
3536         param->listen_id = listen_id;
3537         param->service_id =
3538                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3539         param->bth_pkey = cm_get_bth_pkey(work);
3540         param->port = work->port->port_num;
3541         param->sgid_attr = rx_cm_id->av.ah_attr.grh.sgid_attr;
3542         work->cm_event.private_data =
3543                 IBA_GET_MEM_PTR(CM_SIDR_REQ_PRIVATE_DATA, sidr_req_msg);
3544 }
3545
3546 static int cm_sidr_req_handler(struct cm_work *work)
3547 {
3548         struct cm_id_private *cm_id_priv, *listen_cm_id_priv;
3549         struct cm_sidr_req_msg *sidr_req_msg;
3550         struct ib_wc *wc;
3551         int ret;
3552
3553         cm_id_priv =
3554                 cm_alloc_id_priv(work->port->cm_dev->ib_device, NULL, NULL);
3555         if (IS_ERR(cm_id_priv))
3556                 return PTR_ERR(cm_id_priv);
3557
3558         /* Record SGID/SLID and request ID for lookup. */
3559         sidr_req_msg = (struct cm_sidr_req_msg *)
3560                                 work->mad_recv_wc->recv_buf.mad;
3561
3562         cm_id_priv->id.remote_id =
3563                 cpu_to_be32(IBA_GET(CM_SIDR_REQ_REQUESTID, sidr_req_msg));
3564         cm_id_priv->id.service_id =
3565                 cpu_to_be64(IBA_GET(CM_SIDR_REQ_SERVICEID, sidr_req_msg));
3566         cm_id_priv->id.service_mask = ~cpu_to_be64(0);
3567         cm_id_priv->tid = sidr_req_msg->hdr.tid;
3568
3569         wc = work->mad_recv_wc->wc;
3570         cm_id_priv->av.dgid.global.subnet_prefix = cpu_to_be64(wc->slid);
3571         cm_id_priv->av.dgid.global.interface_id = 0;
3572         ret = cm_init_av_for_response(work->port, work->mad_recv_wc->wc,
3573                                       work->mad_recv_wc->recv_buf.grh,
3574                                       &cm_id_priv->av);
3575         if (ret)
3576                 goto out;
3577
3578         spin_lock_irq(&cm.lock);
3579         listen_cm_id_priv = cm_insert_remote_sidr(cm_id_priv);
3580         if (listen_cm_id_priv) {
3581                 spin_unlock_irq(&cm.lock);
3582                 atomic_long_inc(&work->port->counter_group[CM_RECV_DUPLICATES].
3583                                 counter[CM_SIDR_REQ_COUNTER]);
3584                 goto out; /* Duplicate message. */
3585         }
3586         cm_id_priv->id.state = IB_CM_SIDR_REQ_RCVD;
3587         listen_cm_id_priv = cm_find_listen(cm_id_priv->id.device,
3588                                            cm_id_priv->id.service_id);
3589         if (!listen_cm_id_priv) {
3590                 spin_unlock_irq(&cm.lock);
3591                 ib_send_cm_sidr_rep(&cm_id_priv->id,
3592                                     &(struct ib_cm_sidr_rep_param){
3593                                             .status = IB_SIDR_UNSUPPORTED });
3594                 goto out; /* No match. */
3595         }
3596         spin_unlock_irq(&cm.lock);
3597
3598         cm_id_priv->id.cm_handler = listen_cm_id_priv->id.cm_handler;
3599         cm_id_priv->id.context = listen_cm_id_priv->id.context;
3600
3601         /*
3602          * A SIDR ID does not need to be in the xarray since it does not receive
3603          * mads, is not placed in the remote_id or remote_qpn rbtree, and does
3604          * not enter timewait.
3605          */
3606
3607         cm_format_sidr_req_event(work, cm_id_priv, &listen_cm_id_priv->id);
3608         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &work->cm_event);
3609         cm_free_work(work);
3610         /*
3611          * A pointer to the listen_cm_id is held in the event, so this deref
3612          * must be after the event is delivered above.
3613          */
3614         cm_deref_id(listen_cm_id_priv);
3615         if (ret)
3616                 cm_destroy_id(&cm_id_priv->id, ret);
3617         return 0;
3618 out:
3619         ib_destroy_cm_id(&cm_id_priv->id);
3620         return -EINVAL;
3621 }
3622
3623 static void cm_format_sidr_rep(struct cm_sidr_rep_msg *sidr_rep_msg,
3624                                struct cm_id_private *cm_id_priv,
3625                                struct ib_cm_sidr_rep_param *param)
3626 {
3627         cm_format_mad_ece_hdr(&sidr_rep_msg->hdr, CM_SIDR_REP_ATTR_ID,
3628                               cm_id_priv->tid, param->ece.attr_mod);
3629         IBA_SET(CM_SIDR_REP_REQUESTID, sidr_rep_msg,
3630                 be32_to_cpu(cm_id_priv->id.remote_id));
3631         IBA_SET(CM_SIDR_REP_STATUS, sidr_rep_msg, param->status);
3632         IBA_SET(CM_SIDR_REP_QPN, sidr_rep_msg, param->qp_num);
3633         IBA_SET(CM_SIDR_REP_SERVICEID, sidr_rep_msg,
3634                 be64_to_cpu(cm_id_priv->id.service_id));
3635         IBA_SET(CM_SIDR_REP_Q_KEY, sidr_rep_msg, param->qkey);
3636         IBA_SET(CM_SIDR_REP_VENDOR_ID_L, sidr_rep_msg,
3637                 param->ece.vendor_id & 0xFF);
3638         IBA_SET(CM_SIDR_REP_VENDOR_ID_H, sidr_rep_msg,
3639                 (param->ece.vendor_id >> 8) & 0xFF);
3640
3641         if (param->info && param->info_length)
3642                 IBA_SET_MEM(CM_SIDR_REP_ADDITIONAL_INFORMATION, sidr_rep_msg,
3643                             param->info, param->info_length);
3644
3645         if (param->private_data && param->private_data_len)
3646                 IBA_SET_MEM(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg,
3647                             param->private_data, param->private_data_len);
3648 }
3649
3650 static int cm_send_sidr_rep_locked(struct cm_id_private *cm_id_priv,
3651                                    struct ib_cm_sidr_rep_param *param)
3652 {
3653         struct ib_mad_send_buf *msg;
3654         int ret;
3655
3656         lockdep_assert_held(&cm_id_priv->lock);
3657
3658         if ((param->info && param->info_length > IB_CM_SIDR_REP_INFO_LENGTH) ||
3659             (param->private_data &&
3660              param->private_data_len > IB_CM_SIDR_REP_PRIVATE_DATA_SIZE))
3661                 return -EINVAL;
3662
3663         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_RCVD)
3664                 return -EINVAL;
3665
3666         ret = cm_alloc_msg(cm_id_priv, &msg);
3667         if (ret)
3668                 return ret;
3669
3670         cm_format_sidr_rep((struct cm_sidr_rep_msg *) msg->mad, cm_id_priv,
3671                            param);
3672         trace_icm_send_sidr_rep(&cm_id_priv->id);
3673         ret = ib_post_send_mad(msg, NULL);
3674         if (ret) {
3675                 cm_free_msg(msg);
3676                 return ret;
3677         }
3678         cm_id_priv->id.state = IB_CM_IDLE;
3679         spin_lock_irq(&cm.lock);
3680         if (!RB_EMPTY_NODE(&cm_id_priv->sidr_id_node)) {
3681                 rb_erase(&cm_id_priv->sidr_id_node, &cm.remote_sidr_table);
3682                 RB_CLEAR_NODE(&cm_id_priv->sidr_id_node);
3683         }
3684         spin_unlock_irq(&cm.lock);
3685         return 0;
3686 }
3687
3688 int ib_send_cm_sidr_rep(struct ib_cm_id *cm_id,
3689                         struct ib_cm_sidr_rep_param *param)
3690 {
3691         struct cm_id_private *cm_id_priv =
3692                 container_of(cm_id, struct cm_id_private, id);
3693         unsigned long flags;
3694         int ret;
3695
3696         spin_lock_irqsave(&cm_id_priv->lock, flags);
3697         ret = cm_send_sidr_rep_locked(cm_id_priv, param);
3698         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3699         return ret;
3700 }
3701 EXPORT_SYMBOL(ib_send_cm_sidr_rep);
3702
3703 static void cm_format_sidr_rep_event(struct cm_work *work,
3704                                      const struct cm_id_private *cm_id_priv)
3705 {
3706         struct cm_sidr_rep_msg *sidr_rep_msg;
3707         struct ib_cm_sidr_rep_event_param *param;
3708
3709         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3710                                 work->mad_recv_wc->recv_buf.mad;
3711         param = &work->cm_event.param.sidr_rep_rcvd;
3712         param->status = IBA_GET(CM_SIDR_REP_STATUS, sidr_rep_msg);
3713         param->qkey = IBA_GET(CM_SIDR_REP_Q_KEY, sidr_rep_msg);
3714         param->qpn = IBA_GET(CM_SIDR_REP_QPN, sidr_rep_msg);
3715         param->info = IBA_GET_MEM_PTR(CM_SIDR_REP_ADDITIONAL_INFORMATION,
3716                                       sidr_rep_msg);
3717         param->info_len = IBA_GET(CM_SIDR_REP_ADDITIONAL_INFORMATION_LENGTH,
3718                                   sidr_rep_msg);
3719         param->sgid_attr = cm_id_priv->av.ah_attr.grh.sgid_attr;
3720         work->cm_event.private_data =
3721                 IBA_GET_MEM_PTR(CM_SIDR_REP_PRIVATE_DATA, sidr_rep_msg);
3722 }
3723
3724 static int cm_sidr_rep_handler(struct cm_work *work)
3725 {
3726         struct cm_sidr_rep_msg *sidr_rep_msg;
3727         struct cm_id_private *cm_id_priv;
3728
3729         sidr_rep_msg = (struct cm_sidr_rep_msg *)
3730                                 work->mad_recv_wc->recv_buf.mad;
3731         cm_id_priv = cm_acquire_id(
3732                 cpu_to_be32(IBA_GET(CM_SIDR_REP_REQUESTID, sidr_rep_msg)), 0);
3733         if (!cm_id_priv)
3734                 return -EINVAL; /* Unmatched reply. */
3735
3736         spin_lock_irq(&cm_id_priv->lock);
3737         if (cm_id_priv->id.state != IB_CM_SIDR_REQ_SENT) {
3738                 spin_unlock_irq(&cm_id_priv->lock);
3739                 goto out;
3740         }
3741         cm_id_priv->id.state = IB_CM_IDLE;
3742         ib_cancel_mad(cm_id_priv->av.port->mad_agent, cm_id_priv->msg);
3743         spin_unlock_irq(&cm_id_priv->lock);
3744
3745         cm_format_sidr_rep_event(work, cm_id_priv);
3746         cm_process_work(cm_id_priv, work);
3747         return 0;
3748 out:
3749         cm_deref_id(cm_id_priv);
3750         return -EINVAL;
3751 }
3752
3753 static void cm_process_send_error(struct ib_mad_send_buf *msg,
3754                                   enum ib_wc_status wc_status)
3755 {
3756         struct cm_id_private *cm_id_priv;
3757         struct ib_cm_event cm_event;
3758         enum ib_cm_state state;
3759         int ret;
3760
3761         memset(&cm_event, 0, sizeof cm_event);
3762         cm_id_priv = msg->context[0];
3763
3764         /* Discard old sends or ones without a response. */
3765         spin_lock_irq(&cm_id_priv->lock);
3766         state = (enum ib_cm_state) (unsigned long) msg->context[1];
3767         if (msg != cm_id_priv->msg || state != cm_id_priv->id.state)
3768                 goto discard;
3769
3770         trace_icm_mad_send_err(state, wc_status);
3771         switch (state) {
3772         case IB_CM_REQ_SENT:
3773         case IB_CM_MRA_REQ_RCVD:
3774                 cm_reset_to_idle(cm_id_priv);
3775                 cm_event.event = IB_CM_REQ_ERROR;
3776                 break;
3777         case IB_CM_REP_SENT:
3778         case IB_CM_MRA_REP_RCVD:
3779                 cm_reset_to_idle(cm_id_priv);
3780                 cm_event.event = IB_CM_REP_ERROR;
3781                 break;
3782         case IB_CM_DREQ_SENT:
3783                 cm_enter_timewait(cm_id_priv);
3784                 cm_event.event = IB_CM_DREQ_ERROR;
3785                 break;
3786         case IB_CM_SIDR_REQ_SENT:
3787                 cm_id_priv->id.state = IB_CM_IDLE;
3788                 cm_event.event = IB_CM_SIDR_REQ_ERROR;
3789                 break;
3790         default:
3791                 goto discard;
3792         }
3793         spin_unlock_irq(&cm_id_priv->lock);
3794         cm_event.param.send_status = wc_status;
3795
3796         /* No other events can occur on the cm_id at this point. */
3797         ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, &cm_event);
3798         cm_free_msg(msg);
3799         if (ret)
3800                 ib_destroy_cm_id(&cm_id_priv->id);
3801         return;
3802 discard:
3803         spin_unlock_irq(&cm_id_priv->lock);
3804         cm_free_msg(msg);
3805 }
3806
3807 static void cm_send_handler(struct ib_mad_agent *mad_agent,
3808                             struct ib_mad_send_wc *mad_send_wc)
3809 {
3810         struct ib_mad_send_buf *msg = mad_send_wc->send_buf;
3811         struct cm_port *port;
3812         u16 attr_index;
3813
3814         port = mad_agent->context;
3815         attr_index = be16_to_cpu(((struct ib_mad_hdr *)
3816                                   msg->mad)->attr_id) - CM_ATTR_ID_OFFSET;
3817
3818         /*
3819          * If the send was in response to a received message (context[0] is not
3820          * set to a cm_id), and is not a REJ, then it is a send that was
3821          * manually retried.
3822          */
3823         if (!msg->context[0] && (attr_index != CM_REJ_COUNTER))
3824                 msg->retries = 1;
3825
3826         atomic_long_add(1 + msg->retries,
3827                         &port->counter_group[CM_XMIT].counter[attr_index]);
3828         if (msg->retries)
3829                 atomic_long_add(msg->retries,
3830                                 &port->counter_group[CM_XMIT_RETRIES].
3831                                 counter[attr_index]);
3832
3833         switch (mad_send_wc->status) {
3834         case IB_WC_SUCCESS:
3835         case IB_WC_WR_FLUSH_ERR:
3836                 cm_free_msg(msg);
3837                 break;
3838         default:
3839                 if (msg->context[0] && msg->context[1])
3840                         cm_process_send_error(msg, mad_send_wc->status);
3841                 else
3842                         cm_free_msg(msg);
3843                 break;
3844         }
3845 }
3846
3847 static void cm_work_handler(struct work_struct *_work)
3848 {
3849         struct cm_work *work = container_of(_work, struct cm_work, work.work);
3850         int ret;
3851
3852         switch (work->cm_event.event) {
3853         case IB_CM_REQ_RECEIVED:
3854                 ret = cm_req_handler(work);
3855                 break;
3856         case IB_CM_MRA_RECEIVED:
3857                 ret = cm_mra_handler(work);
3858                 break;
3859         case IB_CM_REJ_RECEIVED:
3860                 ret = cm_rej_handler(work);
3861                 break;
3862         case IB_CM_REP_RECEIVED:
3863                 ret = cm_rep_handler(work);
3864                 break;
3865         case IB_CM_RTU_RECEIVED:
3866                 ret = cm_rtu_handler(work);
3867                 break;
3868         case IB_CM_USER_ESTABLISHED:
3869                 ret = cm_establish_handler(work);
3870                 break;
3871         case IB_CM_DREQ_RECEIVED:
3872                 ret = cm_dreq_handler(work);
3873                 break;
3874         case IB_CM_DREP_RECEIVED:
3875                 ret = cm_drep_handler(work);
3876                 break;
3877         case IB_CM_SIDR_REQ_RECEIVED:
3878                 ret = cm_sidr_req_handler(work);
3879                 break;
3880         case IB_CM_SIDR_REP_RECEIVED:
3881                 ret = cm_sidr_rep_handler(work);
3882                 break;
3883         case IB_CM_LAP_RECEIVED:
3884                 ret = cm_lap_handler(work);
3885                 break;
3886         case IB_CM_APR_RECEIVED:
3887                 ret = cm_apr_handler(work);
3888                 break;
3889         case IB_CM_TIMEWAIT_EXIT:
3890                 ret = cm_timewait_handler(work);
3891                 break;
3892         default:
3893                 trace_icm_handler_err(work->cm_event.event);
3894                 ret = -EINVAL;
3895                 break;
3896         }
3897         if (ret)
3898                 cm_free_work(work);
3899 }
3900
3901 static int cm_establish(struct ib_cm_id *cm_id)
3902 {
3903         struct cm_id_private *cm_id_priv;
3904         struct cm_work *work;
3905         unsigned long flags;
3906         int ret = 0;
3907         struct cm_device *cm_dev;
3908
3909         cm_dev = ib_get_client_data(cm_id->device, &cm_client);
3910         if (!cm_dev)
3911                 return -ENODEV;
3912
3913         work = kmalloc(sizeof *work, GFP_ATOMIC);
3914         if (!work)
3915                 return -ENOMEM;
3916
3917         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3918         spin_lock_irqsave(&cm_id_priv->lock, flags);
3919         switch (cm_id->state)
3920         {
3921         case IB_CM_REP_SENT:
3922         case IB_CM_MRA_REP_RCVD:
3923                 cm_id->state = IB_CM_ESTABLISHED;
3924                 break;
3925         case IB_CM_ESTABLISHED:
3926                 ret = -EISCONN;
3927                 break;
3928         default:
3929                 trace_icm_establish_err(cm_id);
3930                 ret = -EINVAL;
3931                 break;
3932         }
3933         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3934
3935         if (ret) {
3936                 kfree(work);
3937                 goto out;
3938         }
3939
3940         /*
3941          * The CM worker thread may try to destroy the cm_id before it
3942          * can execute this work item.  To prevent potential deadlock,
3943          * we need to find the cm_id once we're in the context of the
3944          * worker thread, rather than holding a reference on it.
3945          */
3946         INIT_DELAYED_WORK(&work->work, cm_work_handler);
3947         work->local_id = cm_id->local_id;
3948         work->remote_id = cm_id->remote_id;
3949         work->mad_recv_wc = NULL;
3950         work->cm_event.event = IB_CM_USER_ESTABLISHED;
3951
3952         /* Check if the device started its remove_one */
3953         spin_lock_irqsave(&cm.lock, flags);
3954         if (!cm_dev->going_down) {
3955                 queue_delayed_work(cm.wq, &work->work, 0);
3956         } else {
3957                 kfree(work);
3958                 ret = -ENODEV;
3959         }
3960         spin_unlock_irqrestore(&cm.lock, flags);
3961
3962 out:
3963         return ret;
3964 }
3965
3966 static int cm_migrate(struct ib_cm_id *cm_id)
3967 {
3968         struct cm_id_private *cm_id_priv;
3969         struct cm_av tmp_av;
3970         unsigned long flags;
3971         int tmp_send_port_not_ready;
3972         int ret = 0;
3973
3974         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
3975         spin_lock_irqsave(&cm_id_priv->lock, flags);
3976         if (cm_id->state == IB_CM_ESTABLISHED &&
3977             (cm_id->lap_state == IB_CM_LAP_UNINIT ||
3978              cm_id->lap_state == IB_CM_LAP_IDLE)) {
3979                 cm_id->lap_state = IB_CM_LAP_IDLE;
3980                 /* Swap address vector */
3981                 tmp_av = cm_id_priv->av;
3982                 cm_id_priv->av = cm_id_priv->alt_av;
3983                 cm_id_priv->alt_av = tmp_av;
3984                 /* Swap port send ready state */
3985                 tmp_send_port_not_ready = cm_id_priv->prim_send_port_not_ready;
3986                 cm_id_priv->prim_send_port_not_ready = cm_id_priv->altr_send_port_not_ready;
3987                 cm_id_priv->altr_send_port_not_ready = tmp_send_port_not_ready;
3988         } else
3989                 ret = -EINVAL;
3990         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
3991
3992         return ret;
3993 }
3994
3995 int ib_cm_notify(struct ib_cm_id *cm_id, enum ib_event_type event)
3996 {
3997         int ret;
3998
3999         switch (event) {
4000         case IB_EVENT_COMM_EST:
4001                 ret = cm_establish(cm_id);
4002                 break;
4003         case IB_EVENT_PATH_MIG:
4004                 ret = cm_migrate(cm_id);
4005                 break;
4006         default:
4007                 ret = -EINVAL;
4008         }
4009         return ret;
4010 }
4011 EXPORT_SYMBOL(ib_cm_notify);
4012
4013 static void cm_recv_handler(struct ib_mad_agent *mad_agent,
4014                             struct ib_mad_send_buf *send_buf,
4015                             struct ib_mad_recv_wc *mad_recv_wc)
4016 {
4017         struct cm_port *port = mad_agent->context;
4018         struct cm_work *work;
4019         enum ib_cm_event_type event;
4020         bool alt_path = false;
4021         u16 attr_id;
4022         int paths = 0;
4023         int going_down = 0;
4024
4025         switch (mad_recv_wc->recv_buf.mad->mad_hdr.attr_id) {
4026         case CM_REQ_ATTR_ID:
4027                 alt_path = cm_req_has_alt_path((struct cm_req_msg *)
4028                                                 mad_recv_wc->recv_buf.mad);
4029                 paths = 1 + (alt_path != 0);
4030                 event = IB_CM_REQ_RECEIVED;
4031                 break;
4032         case CM_MRA_ATTR_ID:
4033                 event = IB_CM_MRA_RECEIVED;
4034                 break;
4035         case CM_REJ_ATTR_ID:
4036                 event = IB_CM_REJ_RECEIVED;
4037                 break;
4038         case CM_REP_ATTR_ID:
4039                 event = IB_CM_REP_RECEIVED;
4040                 break;
4041         case CM_RTU_ATTR_ID:
4042                 event = IB_CM_RTU_RECEIVED;
4043                 break;
4044         case CM_DREQ_ATTR_ID:
4045                 event = IB_CM_DREQ_RECEIVED;
4046                 break;
4047         case CM_DREP_ATTR_ID:
4048                 event = IB_CM_DREP_RECEIVED;
4049                 break;
4050         case CM_SIDR_REQ_ATTR_ID:
4051                 event = IB_CM_SIDR_REQ_RECEIVED;
4052                 break;
4053         case CM_SIDR_REP_ATTR_ID:
4054                 event = IB_CM_SIDR_REP_RECEIVED;
4055                 break;
4056         case CM_LAP_ATTR_ID:
4057                 paths = 1;
4058                 event = IB_CM_LAP_RECEIVED;
4059                 break;
4060         case CM_APR_ATTR_ID:
4061                 event = IB_CM_APR_RECEIVED;
4062                 break;
4063         default:
4064                 ib_free_recv_mad(mad_recv_wc);
4065                 return;
4066         }
4067
4068         attr_id = be16_to_cpu(mad_recv_wc->recv_buf.mad->mad_hdr.attr_id);
4069         atomic_long_inc(&port->counter_group[CM_RECV].
4070                         counter[attr_id - CM_ATTR_ID_OFFSET]);
4071
4072         work = kmalloc(struct_size(work, path, paths), GFP_KERNEL);
4073         if (!work) {
4074                 ib_free_recv_mad(mad_recv_wc);
4075                 return;
4076         }
4077
4078         INIT_DELAYED_WORK(&work->work, cm_work_handler);
4079         work->cm_event.event = event;
4080         work->mad_recv_wc = mad_recv_wc;
4081         work->port = port;
4082
4083         /* Check if the device started its remove_one */
4084         spin_lock_irq(&cm.lock);
4085         if (!port->cm_dev->going_down)
4086                 queue_delayed_work(cm.wq, &work->work, 0);
4087         else
4088                 going_down = 1;
4089         spin_unlock_irq(&cm.lock);
4090
4091         if (going_down) {
4092                 kfree(work);
4093                 ib_free_recv_mad(mad_recv_wc);
4094         }
4095 }
4096
4097 static int cm_init_qp_init_attr(struct cm_id_private *cm_id_priv,
4098                                 struct ib_qp_attr *qp_attr,
4099                                 int *qp_attr_mask)
4100 {
4101         unsigned long flags;
4102         int ret;
4103
4104         spin_lock_irqsave(&cm_id_priv->lock, flags);
4105         switch (cm_id_priv->id.state) {
4106         case IB_CM_REQ_SENT:
4107         case IB_CM_MRA_REQ_RCVD:
4108         case IB_CM_REQ_RCVD:
4109         case IB_CM_MRA_REQ_SENT:
4110         case IB_CM_REP_RCVD:
4111         case IB_CM_MRA_REP_SENT:
4112         case IB_CM_REP_SENT:
4113         case IB_CM_MRA_REP_RCVD:
4114         case IB_CM_ESTABLISHED:
4115                 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS |
4116                                 IB_QP_PKEY_INDEX | IB_QP_PORT;
4117                 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE;
4118                 if (cm_id_priv->responder_resources)
4119                         qp_attr->qp_access_flags |= IB_ACCESS_REMOTE_READ |
4120                                                     IB_ACCESS_REMOTE_ATOMIC;
4121                 qp_attr->pkey_index = cm_id_priv->av.pkey_index;
4122                 qp_attr->port_num = cm_id_priv->av.port->port_num;
4123                 ret = 0;
4124                 break;
4125         default:
4126                 trace_icm_qp_init_err(&cm_id_priv->id);
4127                 ret = -EINVAL;
4128                 break;
4129         }
4130         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4131         return ret;
4132 }
4133
4134 static int cm_init_qp_rtr_attr(struct cm_id_private *cm_id_priv,
4135                                struct ib_qp_attr *qp_attr,
4136                                int *qp_attr_mask)
4137 {
4138         unsigned long flags;
4139         int ret;
4140
4141         spin_lock_irqsave(&cm_id_priv->lock, flags);
4142         switch (cm_id_priv->id.state) {
4143         case IB_CM_REQ_RCVD:
4144         case IB_CM_MRA_REQ_SENT:
4145         case IB_CM_REP_RCVD:
4146         case IB_CM_MRA_REP_SENT:
4147         case IB_CM_REP_SENT:
4148         case IB_CM_MRA_REP_RCVD:
4149         case IB_CM_ESTABLISHED:
4150                 *qp_attr_mask = IB_QP_STATE | IB_QP_AV | IB_QP_PATH_MTU |
4151                                 IB_QP_DEST_QPN | IB_QP_RQ_PSN;
4152                 qp_attr->ah_attr = cm_id_priv->av.ah_attr;
4153                 qp_attr->path_mtu = cm_id_priv->path_mtu;
4154                 qp_attr->dest_qp_num = be32_to_cpu(cm_id_priv->remote_qpn);
4155                 qp_attr->rq_psn = be32_to_cpu(cm_id_priv->rq_psn);
4156                 if (cm_id_priv->qp_type == IB_QPT_RC ||
4157                     cm_id_priv->qp_type == IB_QPT_XRC_TGT) {
4158                         *qp_attr_mask |= IB_QP_MAX_DEST_RD_ATOMIC |
4159                                          IB_QP_MIN_RNR_TIMER;
4160                         qp_attr->max_dest_rd_atomic =
4161                                         cm_id_priv->responder_resources;
4162                         qp_attr->min_rnr_timer = 0;
4163                 }
4164                 if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4165                         *qp_attr_mask |= IB_QP_ALT_PATH;
4166                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4167                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4168                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4169                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4170                 }
4171                 ret = 0;
4172                 break;
4173         default:
4174                 trace_icm_qp_rtr_err(&cm_id_priv->id);
4175                 ret = -EINVAL;
4176                 break;
4177         }
4178         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4179         return ret;
4180 }
4181
4182 static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
4183                                struct ib_qp_attr *qp_attr,
4184                                int *qp_attr_mask)
4185 {
4186         unsigned long flags;
4187         int ret;
4188
4189         spin_lock_irqsave(&cm_id_priv->lock, flags);
4190         switch (cm_id_priv->id.state) {
4191         /* Allow transition to RTS before sending REP */
4192         case IB_CM_REQ_RCVD:
4193         case IB_CM_MRA_REQ_SENT:
4194
4195         case IB_CM_REP_RCVD:
4196         case IB_CM_MRA_REP_SENT:
4197         case IB_CM_REP_SENT:
4198         case IB_CM_MRA_REP_RCVD:
4199         case IB_CM_ESTABLISHED:
4200                 if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
4201                         *qp_attr_mask = IB_QP_STATE | IB_QP_SQ_PSN;
4202                         qp_attr->sq_psn = be32_to_cpu(cm_id_priv->sq_psn);
4203                         switch (cm_id_priv->qp_type) {
4204                         case IB_QPT_RC:
4205                         case IB_QPT_XRC_INI:
4206                                 *qp_attr_mask |= IB_QP_RETRY_CNT | IB_QP_RNR_RETRY |
4207                                                  IB_QP_MAX_QP_RD_ATOMIC;
4208                                 qp_attr->retry_cnt = cm_id_priv->retry_count;
4209                                 qp_attr->rnr_retry = cm_id_priv->rnr_retry_count;
4210                                 qp_attr->max_rd_atomic = cm_id_priv->initiator_depth;
4211                                 fallthrough;
4212                         case IB_QPT_XRC_TGT:
4213                                 *qp_attr_mask |= IB_QP_TIMEOUT;
4214                                 qp_attr->timeout = cm_id_priv->av.timeout;
4215                                 break;
4216                         default:
4217                                 break;
4218                         }
4219                         if (rdma_ah_get_dlid(&cm_id_priv->alt_av.ah_attr)) {
4220                                 *qp_attr_mask |= IB_QP_PATH_MIG_STATE;
4221                                 qp_attr->path_mig_state = IB_MIG_REARM;
4222                         }
4223                 } else {
4224                         *qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
4225                         qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num;
4226                         qp_attr->alt_pkey_index = cm_id_priv->alt_av.pkey_index;
4227                         qp_attr->alt_timeout = cm_id_priv->alt_av.timeout;
4228                         qp_attr->alt_ah_attr = cm_id_priv->alt_av.ah_attr;
4229                         qp_attr->path_mig_state = IB_MIG_REARM;
4230                 }
4231                 ret = 0;
4232                 break;
4233         default:
4234                 trace_icm_qp_rts_err(&cm_id_priv->id);
4235                 ret = -EINVAL;
4236                 break;
4237         }
4238         spin_unlock_irqrestore(&cm_id_priv->lock, flags);
4239         return ret;
4240 }
4241
4242 int ib_cm_init_qp_attr(struct ib_cm_id *cm_id,
4243                        struct ib_qp_attr *qp_attr,
4244                        int *qp_attr_mask)
4245 {
4246         struct cm_id_private *cm_id_priv;
4247         int ret;
4248
4249         cm_id_priv = container_of(cm_id, struct cm_id_private, id);
4250         switch (qp_attr->qp_state) {
4251         case IB_QPS_INIT:
4252                 ret = cm_init_qp_init_attr(cm_id_priv, qp_attr, qp_attr_mask);
4253                 break;
4254         case IB_QPS_RTR:
4255                 ret = cm_init_qp_rtr_attr(cm_id_priv, qp_attr, qp_attr_mask);
4256                 break;
4257         case IB_QPS_RTS:
4258                 ret = cm_init_qp_rts_attr(cm_id_priv, qp_attr, qp_attr_mask);
4259                 break;
4260         default:
4261                 ret = -EINVAL;
4262                 break;
4263         }
4264         return ret;
4265 }
4266 EXPORT_SYMBOL(ib_cm_init_qp_attr);
4267
4268 static ssize_t cm_show_counter(struct kobject *obj, struct attribute *attr,
4269                                char *buf)
4270 {
4271         struct cm_counter_group *group;
4272         struct cm_counter_attribute *cm_attr;
4273
4274         group = container_of(obj, struct cm_counter_group, obj);
4275         cm_attr = container_of(attr, struct cm_counter_attribute, attr);
4276
4277         return sysfs_emit(buf, "%ld\n",
4278                           atomic_long_read(&group->counter[cm_attr->index]));
4279 }
4280
4281 static const struct sysfs_ops cm_counter_ops = {
4282         .show = cm_show_counter
4283 };
4284
4285 static struct kobj_type cm_counter_obj_type = {
4286         .sysfs_ops = &cm_counter_ops,
4287         .default_attrs = cm_counter_default_attrs
4288 };
4289
4290 static int cm_create_port_fs(struct cm_port *port)
4291 {
4292         int i, ret;
4293
4294         for (i = 0; i < CM_COUNTER_GROUPS; i++) {
4295                 ret = ib_port_register_module_stat(port->cm_dev->ib_device,
4296                                                    port->port_num,
4297                                                    &port->counter_group[i].obj,
4298                                                    &cm_counter_obj_type,
4299                                                    counter_group_names[i]);
4300                 if (ret)
4301                         goto error;
4302         }
4303
4304         return 0;
4305
4306 error:
4307         while (i--)
4308                 ib_port_unregister_module_stat(&port->counter_group[i].obj);
4309         return ret;
4310
4311 }
4312
4313 static void cm_remove_port_fs(struct cm_port *port)
4314 {
4315         int i;
4316
4317         for (i = 0; i < CM_COUNTER_GROUPS; i++)
4318                 ib_port_unregister_module_stat(&port->counter_group[i].obj);
4319
4320 }
4321
4322 static int cm_add_one(struct ib_device *ib_device)
4323 {
4324         struct cm_device *cm_dev;
4325         struct cm_port *port;
4326         struct ib_mad_reg_req reg_req = {
4327                 .mgmt_class = IB_MGMT_CLASS_CM,
4328                 .mgmt_class_version = IB_CM_CLASS_VERSION,
4329         };
4330         struct ib_port_modify port_modify = {
4331                 .set_port_cap_mask = IB_PORT_CM_SUP
4332         };
4333         unsigned long flags;
4334         int ret;
4335         int count = 0;
4336         u8 i;
4337
4338         cm_dev = kzalloc(struct_size(cm_dev, port, ib_device->phys_port_cnt),
4339                          GFP_KERNEL);
4340         if (!cm_dev)
4341                 return -ENOMEM;
4342
4343         cm_dev->ib_device = ib_device;
4344         cm_dev->ack_delay = ib_device->attrs.local_ca_ack_delay;
4345         cm_dev->going_down = 0;
4346
4347         set_bit(IB_MGMT_METHOD_SEND, reg_req.method_mask);
4348         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4349                 if (!rdma_cap_ib_cm(ib_device, i))
4350                         continue;
4351
4352                 port = kzalloc(sizeof *port, GFP_KERNEL);
4353                 if (!port) {
4354                         ret = -ENOMEM;
4355                         goto error1;
4356                 }
4357
4358                 cm_dev->port[i-1] = port;
4359                 port->cm_dev = cm_dev;
4360                 port->port_num = i;
4361
4362                 INIT_LIST_HEAD(&port->cm_priv_prim_list);
4363                 INIT_LIST_HEAD(&port->cm_priv_altr_list);
4364
4365                 ret = cm_create_port_fs(port);
4366                 if (ret)
4367                         goto error1;
4368
4369                 port->mad_agent = ib_register_mad_agent(ib_device, i,
4370                                                         IB_QPT_GSI,
4371                                                         &reg_req,
4372                                                         0,
4373                                                         cm_send_handler,
4374                                                         cm_recv_handler,
4375                                                         port,
4376                                                         0);
4377                 if (IS_ERR(port->mad_agent)) {
4378                         ret = PTR_ERR(port->mad_agent);
4379                         goto error2;
4380                 }
4381
4382                 ret = ib_modify_port(ib_device, i, 0, &port_modify);
4383                 if (ret)
4384                         goto error3;
4385
4386                 count++;
4387         }
4388
4389         if (!count) {
4390                 ret = -EOPNOTSUPP;
4391                 goto free;
4392         }
4393
4394         ib_set_client_data(ib_device, &cm_client, cm_dev);
4395
4396         write_lock_irqsave(&cm.device_lock, flags);
4397         list_add_tail(&cm_dev->list, &cm.device_list);
4398         write_unlock_irqrestore(&cm.device_lock, flags);
4399         return 0;
4400
4401 error3:
4402         ib_unregister_mad_agent(port->mad_agent);
4403 error2:
4404         cm_remove_port_fs(port);
4405 error1:
4406         port_modify.set_port_cap_mask = 0;
4407         port_modify.clr_port_cap_mask = IB_PORT_CM_SUP;
4408         kfree(port);
4409         while (--i) {
4410                 if (!rdma_cap_ib_cm(ib_device, i))
4411                         continue;
4412
4413                 port = cm_dev->port[i-1];
4414                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4415                 ib_unregister_mad_agent(port->mad_agent);
4416                 cm_remove_port_fs(port);
4417                 kfree(port);
4418         }
4419 free:
4420         kfree(cm_dev);
4421         return ret;
4422 }
4423
4424 static void cm_remove_one(struct ib_device *ib_device, void *client_data)
4425 {
4426         struct cm_device *cm_dev = client_data;
4427         struct cm_port *port;
4428         struct cm_id_private *cm_id_priv;
4429         struct ib_mad_agent *cur_mad_agent;
4430         struct ib_port_modify port_modify = {
4431                 .clr_port_cap_mask = IB_PORT_CM_SUP
4432         };
4433         unsigned long flags;
4434         int i;
4435
4436         write_lock_irqsave(&cm.device_lock, flags);
4437         list_del(&cm_dev->list);
4438         write_unlock_irqrestore(&cm.device_lock, flags);
4439
4440         spin_lock_irq(&cm.lock);
4441         cm_dev->going_down = 1;
4442         spin_unlock_irq(&cm.lock);
4443
4444         for (i = 1; i <= ib_device->phys_port_cnt; i++) {
4445                 if (!rdma_cap_ib_cm(ib_device, i))
4446                         continue;
4447
4448                 port = cm_dev->port[i-1];
4449                 ib_modify_port(ib_device, port->port_num, 0, &port_modify);
4450                 /* Mark all the cm_id's as not valid */
4451                 spin_lock_irq(&cm.lock);
4452                 list_for_each_entry(cm_id_priv, &port->cm_priv_altr_list, altr_list)
4453                         cm_id_priv->altr_send_port_not_ready = 1;
4454                 list_for_each_entry(cm_id_priv, &port->cm_priv_prim_list, prim_list)
4455                         cm_id_priv->prim_send_port_not_ready = 1;
4456                 spin_unlock_irq(&cm.lock);
4457                 /*
4458                  * We flush the queue here after the going_down set, this
4459                  * verify that no new works will be queued in the recv handler,
4460                  * after that we can call the unregister_mad_agent
4461                  */
4462                 flush_workqueue(cm.wq);
4463                 spin_lock_irq(&cm.state_lock);
4464                 cur_mad_agent = port->mad_agent;
4465                 port->mad_agent = NULL;
4466                 spin_unlock_irq(&cm.state_lock);
4467                 ib_unregister_mad_agent(cur_mad_agent);
4468                 cm_remove_port_fs(port);
4469                 kfree(port);
4470         }
4471
4472         kfree(cm_dev);
4473 }
4474
4475 static int __init ib_cm_init(void)
4476 {
4477         int ret;
4478
4479         INIT_LIST_HEAD(&cm.device_list);
4480         rwlock_init(&cm.device_lock);
4481         spin_lock_init(&cm.lock);
4482         spin_lock_init(&cm.state_lock);
4483         cm.listen_service_table = RB_ROOT;
4484         cm.listen_service_id = be64_to_cpu(IB_CM_ASSIGN_SERVICE_ID);
4485         cm.remote_id_table = RB_ROOT;
4486         cm.remote_qp_table = RB_ROOT;
4487         cm.remote_sidr_table = RB_ROOT;
4488         xa_init_flags(&cm.local_id_table, XA_FLAGS_ALLOC);
4489         get_random_bytes(&cm.random_id_operand, sizeof cm.random_id_operand);
4490         INIT_LIST_HEAD(&cm.timewait_list);
4491
4492         cm.wq = alloc_workqueue("ib_cm", 0, 1);
4493         if (!cm.wq) {
4494                 ret = -ENOMEM;
4495                 goto error2;
4496         }
4497
4498         ret = ib_register_client(&cm_client);
4499         if (ret)
4500                 goto error3;
4501
4502         return 0;
4503 error3:
4504         destroy_workqueue(cm.wq);
4505 error2:
4506         return ret;
4507 }
4508
4509 static void __exit ib_cm_cleanup(void)
4510 {
4511         struct cm_timewait_info *timewait_info, *tmp;
4512
4513         spin_lock_irq(&cm.lock);
4514         list_for_each_entry(timewait_info, &cm.timewait_list, list)
4515                 cancel_delayed_work(&timewait_info->work.work);
4516         spin_unlock_irq(&cm.lock);
4517
4518         ib_unregister_client(&cm_client);
4519         destroy_workqueue(cm.wq);
4520
4521         list_for_each_entry_safe(timewait_info, tmp, &cm.timewait_list, list) {
4522                 list_del(&timewait_info->list);
4523                 kfree(timewait_info);
4524         }
4525
4526         WARN_ON(!xa_empty(&cm.local_id_table));
4527 }
4528
4529 module_init(ib_cm_init);
4530 module_exit(ib_cm_cleanup);