net/rds/iw_cm.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/slab.h>
  36 #include <linux/vmalloc.h>
  37 #include <linux/ratelimit.h>
  38
  39 #include "rds.h"
  40 #include "iw.h"
  41
  42 /*
  43  * Set the selected protocol version
  44  */
  45 static void rds_iw_set_protocol(struct rds_connection *conn, unsigned int version)
  46 {
  47         conn->c_version = version;
  48 }
  49
  50 /*
  51  * Set up flow control
  52  */
  53 static void rds_iw_set_flow_control(struct rds_connection *conn, u32 credits)
  54 {
  55         struct rds_iw_connection *ic = conn->c_transport_data;
  56
  57         if (rds_iw_sysctl_flow_control && credits != 0) {
  58                 /* We're doing flow control */
  59                 ic->i_flowctl = 1;
  60                 rds_iw_send_add_credits(conn, credits);
  61         } else {
  62                 ic->i_flowctl = 0;
  63         }
  64 }
  65
  66 /*
  67  * Connection established.
  68  * We get here for both outgoing and incoming connection.
  69  */
  70 void rds_iw_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  71 {
  72         const struct rds_iw_connect_private *dp = NULL;
  73         struct rds_iw_connection *ic = conn->c_transport_data;
  74         struct rds_iw_device *rds_iwdev;
  75         int err;
  76
  77         if (event->param.conn.private_data_len) {
  78                 dp = event->param.conn.private_data;
  79
  80                 rds_iw_set_protocol(conn,
  81                                 RDS_PROTOCOL(dp->dp_protocol_major,
  82                                         dp->dp_protocol_minor));
  83                 rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
  84         }
  85
  86         /* update ib_device with this local ipaddr & conn */
  87         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
  88         err = rds_iw_update_cm_id(rds_iwdev, ic->i_cm_id);
  89         if (err)
  90                 printk(KERN_ERR "rds_iw_update_ipaddr failed (%d)\n", err);
  91         rds_iw_add_conn(rds_iwdev, conn);
  92
  93         /* If the peer gave us the last packet it saw, process this as if
  94          * we had received a regular ACK. */
  95         if (dp && dp->dp_ack_seq)
  96                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
  97
  98         printk(KERN_NOTICE "RDS/IW: connected to %pI4<->%pI4 version %u.%u%s\n",
  99                         &conn->c_laddr, &conn->c_faddr,
 100                         RDS_PROTOCOL_MAJOR(conn->c_version),
 101                         RDS_PROTOCOL_MINOR(conn->c_version),
 102                         ic->i_flowctl ? ", flow control" : "");
 103
 104         rds_connect_complete(conn);
 105 }
 106
 107 static void rds_iw_cm_fill_conn_param(struct rds_connection *conn,
 108                         struct rdma_conn_param *conn_param,
 109                         struct rds_iw_connect_private *dp,
 110                         u32 protocol_version)
 111 {
 112         struct rds_iw_connection *ic = conn->c_transport_data;
 113
 114         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 115         /* XXX tune these? */
 116         conn_param->responder_resources = 1;
 117         conn_param->initiator_depth = 1;
 118
 119         if (dp) {
 120                 memset(dp, 0, sizeof(*dp));
 121                 dp->dp_saddr = conn->c_laddr;
 122                 dp->dp_daddr = conn->c_faddr;
 123                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 124                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 125                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IW_SUPPORTED_PROTOCOLS);
 126                 dp->dp_ack_seq = rds_iw_piggyb_ack(ic);
 127
 128                 /* Advertise flow control */
 129                 if (ic->i_flowctl) {
 130                         unsigned int credits;
 131
 132                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 133                         dp->dp_credit = cpu_to_be32(credits);
 134                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 135                 }
 136
 137                 conn_param->private_data = dp;
 138                 conn_param->private_data_len = sizeof(*dp);
 139         }
 140 }
 141
 142 static void rds_iw_cq_event_handler(struct ib_event *event, void *data)
 143 {
 144         rdsdebug("event %u data %p\n", event->event, data);
 145 }
 146
 147 static void rds_iw_qp_event_handler(struct ib_event *event, void *data)
 148 {
 149         struct rds_connection *conn = data;
 150         struct rds_iw_connection *ic = conn->c_transport_data;
 151
 152         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
 153
 154         switch (event->event) {
 155         case IB_EVENT_COMM_EST:
 156                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 157                 break;
 158         case IB_EVENT_QP_REQ_ERR:
 159         case IB_EVENT_QP_FATAL:
 160         default:
 161                 rdsdebug("Fatal QP Event %u "
 162                         "- connection %pI4->%pI4, reconnecting\n",
 163                         event->event, &conn->c_laddr,
 164                         &conn->c_faddr);
 165                 rds_conn_drop(conn);
 166                 break;
 167         }
 168 }
 169
 170 /*
 171  * Create a QP
 172  */
 173 static int rds_iw_init_qp_attrs(struct ib_qp_init_attr *attr,
 174                 struct rds_iw_device *rds_iwdev,
 175                 struct rds_iw_work_ring *send_ring,
 176                 void (*send_cq_handler)(struct ib_cq *, void *),
 177                 struct rds_iw_work_ring *recv_ring,
 178                 void (*recv_cq_handler)(struct ib_cq *, void *),
 179                 void *context)
 180 {
 181         struct ib_device *dev = rds_iwdev->dev;
 182         struct ib_cq_init_attr cq_attr = {};
 183         unsigned int send_size, recv_size;
 184         int ret;
 185
 186         /* The offset of 1 is to accommodate the additional ACK WR. */
 187         send_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_send_wr + 1);
 188         recv_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_recv_wr + 1);
 189         rds_iw_ring_resize(send_ring, send_size - 1);
 190         rds_iw_ring_resize(recv_ring, recv_size - 1);
 191
 192         memset(attr, 0, sizeof(*attr));
 193         attr->event_handler = rds_iw_qp_event_handler;
 194         attr->qp_context = context;
 195         attr->cap.max_send_wr = send_size;
 196         attr->cap.max_recv_wr = recv_size;
 197         attr->cap.max_send_sge = rds_iwdev->max_sge;
 198         attr->cap.max_recv_sge = RDS_IW_RECV_SGE;
 199         attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 200         attr->qp_type = IB_QPT_RC;
 201
 202         cq_attr.cqe = send_size;
 203         attr->send_cq = ib_create_cq(dev, send_cq_handler,
 204                                      rds_iw_cq_event_handler,
 205                                      context, &cq_attr);
 206         if (IS_ERR(attr->send_cq)) {
 207                 ret = PTR_ERR(attr->send_cq);
 208                 attr->send_cq = NULL;
 209                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 210                 goto out;
 211         }
 212
 213         cq_attr.cqe = recv_size;
 214         attr->recv_cq = ib_create_cq(dev, recv_cq_handler,
 215                                      rds_iw_cq_event_handler,
 216                                      context, &cq_attr);
 217         if (IS_ERR(attr->recv_cq)) {
 218                 ret = PTR_ERR(attr->recv_cq);
 219                 attr->recv_cq = NULL;
 220                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 221                 goto out;
 222         }
 223
 224         ret = ib_req_notify_cq(attr->send_cq, IB_CQ_NEXT_COMP);
 225         if (ret) {
 226                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 227                 goto out;
 228         }
 229
 230         ret = ib_req_notify_cq(attr->recv_cq, IB_CQ_SOLICITED);
 231         if (ret) {
 232                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 233                 goto out;
 234         }
 235
 236 out:
 237         if (ret) {
 238                 if (attr->send_cq)
 239                         ib_destroy_cq(attr->send_cq);
 240                 if (attr->recv_cq)
 241                         ib_destroy_cq(attr->recv_cq);
 242         }
 243         return ret;
 244 }
 245
 246 /*
 247  * This needs to be very careful to not leave IS_ERR pointers around for
 248  * cleanup to trip over.
 249  */
 250 static int rds_iw_setup_qp(struct rds_connection *conn)
 251 {
 252         struct rds_iw_connection *ic = conn->c_transport_data;
 253         struct ib_device *dev = ic->i_cm_id->device;
 254         struct ib_qp_init_attr attr;
 255         struct rds_iw_device *rds_iwdev;
 256         int ret;
 257
 258         /* rds_iw_add_one creates a rds_iw_device object per IB device,
 259          * and allocates a protection domain, memory range and MR pool
 260          * for each.  If that fails for any reason, it will not register
 261          * the rds_iwdev at all.
 262          */
 263         rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
 264         if (!rds_iwdev) {
 265                 printk_ratelimited(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
 266                                         dev->name);
 267                 return -EOPNOTSUPP;
 268         }
 269
 270         /* Protection domain and memory range */
 271         ic->i_pd = rds_iwdev->pd;
 272         ic->i_mr = rds_iwdev->mr;
 273
 274         ret = rds_iw_init_qp_attrs(&attr, rds_iwdev,
 275                         &ic->i_send_ring, rds_iw_send_cq_comp_handler,
 276                         &ic->i_recv_ring, rds_iw_recv_cq_comp_handler,
 277                         conn);
 278         if (ret < 0)
 279                 goto out;
 280
 281         ic->i_send_cq = attr.send_cq;
 282         ic->i_recv_cq = attr.recv_cq;
 283
 284         /*
 285          * XXX this can fail if max_*_wr is too large?  Are we supposed
 286          * to back off until we get a value that the hardware can support?
 287          */
 288         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 289         if (ret) {
 290                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 291                 goto out;
 292         }
 293
 294         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 295                                            ic->i_send_ring.w_nr *
 296                                                 sizeof(struct rds_header),
 297                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 298         if (!ic->i_send_hdrs) {
 299                 ret = -ENOMEM;
 300                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 301                 goto out;
 302         }
 303
 304         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 305                                            ic->i_recv_ring.w_nr *
 306                                                 sizeof(struct rds_header),
 307                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 308         if (!ic->i_recv_hdrs) {
 309                 ret = -ENOMEM;
 310                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 311                 goto out;
 312         }
 313
 314         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 315                                        &ic->i_ack_dma, GFP_KERNEL);
 316         if (!ic->i_ack) {
 317                 ret = -ENOMEM;
 318                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 319                 goto out;
 320         }
 321
 322         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
 323         if (!ic->i_sends) {
 324                 ret = -ENOMEM;
 325                 rdsdebug("send allocation failed\n");
 326                 goto out;
 327         }
 328         rds_iw_send_init_ring(ic);
 329
 330         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
 331         if (!ic->i_recvs) {
 332                 ret = -ENOMEM;
 333                 rdsdebug("recv allocation failed\n");
 334                 goto out;
 335         }
 336
 337         rds_iw_recv_init_ring(ic);
 338         rds_iw_recv_init_ack(ic);
 339
 340         /* Post receive buffers - as a side effect, this will update
 341          * the posted credit count. */
 342         rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
 343
 344         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 345                  ic->i_send_cq, ic->i_recv_cq);
 346
 347 out:
 348         return ret;
 349 }
 350
 351 static u32 rds_iw_protocol_compatible(const struct rds_iw_connect_private *dp)
 352 {
 353         u16 common;
 354         u32 version = 0;
 355
 356         /* rdma_cm private data is odd - when there is any private data in the
 357          * request, we will be given a pretty large buffer without telling us the
 358          * original size. The only way to tell the difference is by looking at
 359          * the contents, which are initialized to zero.
 360          * If the protocol version fields aren't set, this is a connection attempt
 361          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 362          * We really should have changed this for OFED 1.3 :-( */
 363         if (dp->dp_protocol_major == 0)
 364                 return RDS_PROTOCOL_3_0;
 365
 366         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IW_SUPPORTED_PROTOCOLS;
 367         if (dp->dp_protocol_major == 3 && common) {
 368                 version = RDS_PROTOCOL_3_0;
 369                 while ((common >>= 1) != 0)
 370                         version++;
 371         }
 372         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using "
 373                         "incompatible protocol version %u.%u\n",
 374                         &dp->dp_saddr,
 375                         dp->dp_protocol_major,
 376                         dp->dp_protocol_minor);
 377         return version;
 378 }
 379
 380 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
 381                                     struct rdma_cm_event *event)
 382 {
 383         const struct rds_iw_connect_private *dp = event->param.conn.private_data;
 384         struct rds_iw_connect_private dp_rep;
 385         struct rds_connection *conn = NULL;
 386         struct rds_iw_connection *ic = NULL;
 387         struct rdma_conn_param conn_param;
 388         struct rds_iw_device *rds_iwdev;
 389         u32 version;
 390         int err, destroy = 1;
 391
 392         /* Check whether the remote protocol version matches ours. */
 393         version = rds_iw_protocol_compatible(dp);
 394         if (!version)
 395                 goto out;
 396
 397         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u\n",
 398                  &dp->dp_saddr, &dp->dp_daddr,
 399                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version));
 400
 401         /* RDS/IW is not currently netns aware, thus init_net */
 402         conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
 403                                &rds_iw_transport, GFP_KERNEL);
 404         if (IS_ERR(conn)) {
 405                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 406                 conn = NULL;
 407                 goto out;
 408         }
 409
 410         /*
 411          * The connection request may occur while the
 412          * previous connection exist, e.g. in case of failover.
 413          * But as connections may be initiated simultaneously
 414          * by both hosts, we have a random backoff mechanism -
 415          * see the comment above rds_queue_reconnect()
 416          */
 417         mutex_lock(&conn->c_cm_lock);
 418         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 419                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 420                         rdsdebug("incoming connect while connecting\n");
 421                         rds_conn_drop(conn);
 422                         rds_iw_stats_inc(s_iw_listen_closed_stale);
 423                 } else
 424                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 425                         /* Wait and see - our connect may still be succeeding */
 426                         rds_iw_stats_inc(s_iw_connect_raced);
 427                 }
 428                 mutex_unlock(&conn->c_cm_lock);
 429                 goto out;
 430         }
 431
 432         ic = conn->c_transport_data;
 433
 434         rds_iw_set_protocol(conn, version);
 435         rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 436
 437         /* If the peer gave us the last packet it saw, process this as if
 438          * we had received a regular ACK. */
 439         if (dp->dp_ack_seq)
 440                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 441
 442         BUG_ON(cm_id->context);
 443         BUG_ON(ic->i_cm_id);
 444
 445         ic->i_cm_id = cm_id;
 446         cm_id->context = conn;
 447
 448         rds_iwdev = ib_get_client_data(cm_id->device, &rds_iw_client);
 449         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
 450
 451         /* We got halfway through setting up the ib_connection, if we
 452          * fail now, we have to take the long route out of this mess. */
 453         destroy = 0;
 454
 455         err = rds_iw_setup_qp(conn);
 456         if (err) {
 457                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", err);
 458                 mutex_unlock(&conn->c_cm_lock);
 459                 goto out;
 460         }
 461
 462         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
 463
 464         /* rdma_accept() calls rdma_reject() internally if it fails */
 465         err = rdma_accept(cm_id, &conn_param);
 466         mutex_unlock(&conn->c_cm_lock);
 467         if (err) {
 468                 rds_iw_conn_error(conn, "rdma_accept failed (%d)\n", err);
 469                 goto out;
 470         }
 471
 472         return 0;
 473
 474 out:
 475         rdma_reject(cm_id, NULL, 0);
 476         return destroy;
 477 }
 478
 479
 480 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id)
 481 {
 482         struct rds_connection *conn = cm_id->context;
 483         struct rds_iw_connection *ic = conn->c_transport_data;
 484         struct rdma_conn_param conn_param;
 485         struct rds_iw_connect_private dp;
 486         int ret;
 487
 488         /* If the peer doesn't do protocol negotiation, we must
 489          * default to RDSv3.0 */
 490         rds_iw_set_protocol(conn, RDS_PROTOCOL_3_0);
 491         ic->i_flowctl = rds_iw_sysctl_flow_control;     /* advertise flow control */
 492
 493         ret = rds_iw_setup_qp(conn);
 494         if (ret) {
 495                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", ret);
 496                 goto out;
 497         }
 498
 499         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
 500
 501         ret = rdma_connect(cm_id, &conn_param);
 502         if (ret)
 503                 rds_iw_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 504
 505 out:
 506         /* Beware - returning non-zero tells the rdma_cm to destroy
 507          * the cm_id. We should certainly not do it as long as we still
 508          * "own" the cm_id. */
 509         if (ret) {
 510                 struct rds_iw_connection *ic = conn->c_transport_data;
 511
 512                 if (ic->i_cm_id == cm_id)
 513                         ret = 0;
 514         }
 515         return ret;
 516 }
 517
 518 int rds_iw_conn_connect(struct rds_connection *conn)
 519 {
 520         struct rds_iw_connection *ic = conn->c_transport_data;
 521         struct rds_iw_device *rds_iwdev;
 522         struct sockaddr_in src, dest;
 523         int ret;
 524
 525         /* XXX I wonder what affect the port space has */
 526         /* delegate cm event handler to rdma_transport */
 527         ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn,
 528                                      RDMA_PS_TCP, IB_QPT_RC);
 529         if (IS_ERR(ic->i_cm_id)) {
 530                 ret = PTR_ERR(ic->i_cm_id);
 531                 ic->i_cm_id = NULL;
 532                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 533                 goto out;
 534         }
 535
 536         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 537
 538         src.sin_family = AF_INET;
 539         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 540         src.sin_port = (__force u16)htons(0);
 541
 542         /* First, bind to the local address and device. */
 543         ret = rdma_bind_addr(ic->i_cm_id, (struct sockaddr *) &src);
 544         if (ret) {
 545                 rdsdebug("rdma_bind_addr(%pI4) failed: %d\n",
 546                                 &conn->c_laddr, ret);
 547                 rdma_destroy_id(ic->i_cm_id);
 548                 ic->i_cm_id = NULL;
 549                 goto out;
 550         }
 551
 552         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 553         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
 554
 555         dest.sin_family = AF_INET;
 556         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 557         dest.sin_port = (__force u16)htons(RDS_PORT);
 558
 559         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 560                                 (struct sockaddr *)&dest,
 561                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 562         if (ret) {
 563                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 564                          ret);
 565                 rdma_destroy_id(ic->i_cm_id);
 566                 ic->i_cm_id = NULL;
 567         }
 568
 569 out:
 570         return ret;
 571 }
 572
 573 /*
 574  * This is so careful about only cleaning up resources that were built up
 575  * so that it can be called at any point during startup.  In fact it
 576  * can be called multiple times for a given connection.
 577  */
 578 void rds_iw_conn_shutdown(struct rds_connection *conn)
 579 {
 580         struct rds_iw_connection *ic = conn->c_transport_data;
 581         int err = 0;
 582         struct ib_qp_attr qp_attr;
 583
 584         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 585                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 586                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 587
 588         if (ic->i_cm_id) {
 589                 struct ib_device *dev = ic->i_cm_id->device;
 590
 591                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 592                 err = rdma_disconnect(ic->i_cm_id);
 593                 if (err) {
 594                         /* Actually this may happen quite frequently, when
 595                          * an outgoing connect raced with an incoming connect.
 596                          */
 597                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 598                                  ic->i_cm_id, err);
 599                 }
 600
 601                 if (ic->i_cm_id->qp) {
 602                         qp_attr.qp_state = IB_QPS_ERR;
 603                         ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 604                 }
 605
 606                 wait_event(rds_iw_ring_empty_wait,
 607                         rds_iw_ring_empty(&ic->i_send_ring) &&
 608                         rds_iw_ring_empty(&ic->i_recv_ring));
 609
 610                 if (ic->i_send_hdrs)
 611                         ib_dma_free_coherent(dev,
 612                                            ic->i_send_ring.w_nr *
 613                                                 sizeof(struct rds_header),
 614                                            ic->i_send_hdrs,
 615                                            ic->i_send_hdrs_dma);
 616
 617                 if (ic->i_recv_hdrs)
 618                         ib_dma_free_coherent(dev,
 619                                            ic->i_recv_ring.w_nr *
 620                                                 sizeof(struct rds_header),
 621                                            ic->i_recv_hdrs,
 622                                            ic->i_recv_hdrs_dma);
 623
 624                 if (ic->i_ack)
 625                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 626                                              ic->i_ack, ic->i_ack_dma);
 627
 628                 if (ic->i_sends)
 629                         rds_iw_send_clear_ring(ic);
 630                 if (ic->i_recvs)
 631                         rds_iw_recv_clear_ring(ic);
 632
 633                 if (ic->i_cm_id->qp)
 634                         rdma_destroy_qp(ic->i_cm_id);
 635                 if (ic->i_send_cq)
 636                         ib_destroy_cq(ic->i_send_cq);
 637                 if (ic->i_recv_cq)
 638                         ib_destroy_cq(ic->i_recv_cq);
 639
 640                 /*
 641                  * If associated with an rds_iw_device:
 642                  *      Move connection back to the nodev list.
 643                  *      Remove cm_id from the device cm_id list.
 644                  */
 645                 if (ic->rds_iwdev)
 646                         rds_iw_remove_conn(ic->rds_iwdev, conn);
 647
 648                 rdma_destroy_id(ic->i_cm_id);
 649
 650                 ic->i_cm_id = NULL;
 651                 ic->i_pd = NULL;
 652                 ic->i_mr = NULL;
 653                 ic->i_send_cq = NULL;
 654                 ic->i_recv_cq = NULL;
 655                 ic->i_send_hdrs = NULL;
 656                 ic->i_recv_hdrs = NULL;
 657                 ic->i_ack = NULL;
 658         }
 659         BUG_ON(ic->rds_iwdev);
 660
 661         /* Clear pending transmit */
 662         if (ic->i_rm) {
 663                 rds_message_put(ic->i_rm);
 664                 ic->i_rm = NULL;
 665         }
 666
 667         /* Clear the ACK state */
 668         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 669 #ifdef KERNEL_HAS_ATOMIC64
 670         atomic64_set(&ic->i_ack_next, 0);
 671 #else
 672         ic->i_ack_next = 0;
 673 #endif
 674         ic->i_ack_recv = 0;
 675
 676         /* Clear flow control state */
 677         ic->i_flowctl = 0;
 678         atomic_set(&ic->i_credits, 0);
 679
 680         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
 681         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
 682
 683         if (ic->i_iwinc) {
 684                 rds_inc_put(&ic->i_iwinc->ii_inc);
 685                 ic->i_iwinc = NULL;
 686         }
 687
 688         vfree(ic->i_sends);
 689         ic->i_sends = NULL;
 690         vfree(ic->i_recvs);
 691         ic->i_recvs = NULL;
 692         rdsdebug("shutdown complete\n");
 693 }
 694
 695 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 696 {
 697         struct rds_iw_connection *ic;
 698         unsigned long flags;
 699
 700         /* XXX too lazy? */
 701         ic = kzalloc(sizeof(struct rds_iw_connection), gfp);
 702         if (!ic)
 703                 return -ENOMEM;
 704
 705         INIT_LIST_HEAD(&ic->iw_node);
 706         tasklet_init(&ic->i_recv_tasklet, rds_iw_recv_tasklet_fn,
 707                      (unsigned long) ic);
 708         mutex_init(&ic->i_recv_mutex);
 709 #ifndef KERNEL_HAS_ATOMIC64
 710         spin_lock_init(&ic->i_ack_lock);
 711 #endif
 712
 713         /*
 714          * rds_iw_conn_shutdown() waits for these to be emptied so they
 715          * must be initialized before it can be called.
 716          */
 717         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
 718         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
 719
 720         ic->conn = conn;
 721         conn->c_transport_data = ic;
 722
 723         spin_lock_irqsave(&iw_nodev_conns_lock, flags);
 724         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 725         spin_unlock_irqrestore(&iw_nodev_conns_lock, flags);
 726
 727
 728         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 729         return 0;
 730 }
 731
 732 /*
 733  * Free a connection. Connection must be shut down and not set for reconnect.
 734  */
 735 void rds_iw_conn_free(void *arg)
 736 {
 737         struct rds_iw_connection *ic = arg;
 738         spinlock_t      *lock_ptr;
 739
 740         rdsdebug("ic %p\n", ic);
 741
 742         /*
 743          * Conn is either on a dev's list or on the nodev list.
 744          * A race with shutdown() or connect() would cause problems
 745          * (since rds_iwdev would change) but that should never happen.
 746          */
 747         lock_ptr = ic->rds_iwdev ? &ic->rds_iwdev->spinlock : &iw_nodev_conns_lock;
 748
 749         spin_lock_irq(lock_ptr);
 750         list_del(&ic->iw_node);
 751         spin_unlock_irq(lock_ptr);
 752
 753         kfree(ic);
 754 }
 755
 756 /*
 757  * An error occurred on the connection
 758  */
 759 void
 760 __rds_iw_conn_error(struct rds_connection *conn, const char *fmt, ...)
 761 {
 762         va_list ap;
 763
 764         rds_conn_drop(conn);
 765
 766         va_start(ap, fmt);
 767         vprintk(fmt, ap);
 768         va_end(ap);
 769 }