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