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         err = rds_iw_add_conn(rds_iwdev, conn);
  90         if (err)
  91                 printk(KERN_ERR "rds_iw_add_conn failed (%d)\n", err);
  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                 rds_iw_conn_error(conn, "RDS/IW: Fatal QP Event %u - connection %pI4->%pI4...reconnecting\n",
 162                         event->event, &conn->c_laddr,
 163                         &conn->c_faddr);
 164                 break;
 165         }
 166 }
 167
 168 /*
 169  * Create a QP
 170  */
 171 static int rds_iw_init_qp_attrs(struct ib_qp_init_attr *attr,
 172                 struct rds_iw_device *rds_iwdev,
 173                 struct rds_iw_work_ring *send_ring,
 174                 void (*send_cq_handler)(struct ib_cq *, void *),
 175                 struct rds_iw_work_ring *recv_ring,
 176                 void (*recv_cq_handler)(struct ib_cq *, void *),
 177                 void *context)
 178 {
 179         struct ib_device *dev = rds_iwdev->dev;
 180         unsigned int send_size, recv_size;
 181         int ret;
 182
 183         /* The offset of 1 is to accomodate the additional ACK WR. */
 184         send_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_send_wr + 1);
 185         recv_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_recv_wr + 1);
 186         rds_iw_ring_resize(send_ring, send_size - 1);
 187         rds_iw_ring_resize(recv_ring, recv_size - 1);
 188
 189         memset(attr, 0, sizeof(*attr));
 190         attr->event_handler = rds_iw_qp_event_handler;
 191         attr->qp_context = context;
 192         attr->cap.max_send_wr = send_size;
 193         attr->cap.max_recv_wr = recv_size;
 194         attr->cap.max_send_sge = rds_iwdev->max_sge;
 195         attr->cap.max_recv_sge = RDS_IW_RECV_SGE;
 196         attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 197         attr->qp_type = IB_QPT_RC;
 198
 199         attr->send_cq = ib_create_cq(dev, send_cq_handler,
 200                                      rds_iw_cq_event_handler,
 201                                      context, send_size, 0);
 202         if (IS_ERR(attr->send_cq)) {
 203                 ret = PTR_ERR(attr->send_cq);
 204                 attr->send_cq = NULL;
 205                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 206                 goto out;
 207         }
 208
 209         attr->recv_cq = ib_create_cq(dev, recv_cq_handler,
 210                                      rds_iw_cq_event_handler,
 211                                      context, recv_size, 0);
 212         if (IS_ERR(attr->recv_cq)) {
 213                 ret = PTR_ERR(attr->recv_cq);
 214                 attr->recv_cq = NULL;
 215                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 216                 goto out;
 217         }
 218
 219         ret = ib_req_notify_cq(attr->send_cq, IB_CQ_NEXT_COMP);
 220         if (ret) {
 221                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 222                 goto out;
 223         }
 224
 225         ret = ib_req_notify_cq(attr->recv_cq, IB_CQ_SOLICITED);
 226         if (ret) {
 227                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 228                 goto out;
 229         }
 230
 231 out:
 232         if (ret) {
 233                 if (attr->send_cq)
 234                         ib_destroy_cq(attr->send_cq);
 235                 if (attr->recv_cq)
 236                         ib_destroy_cq(attr->recv_cq);
 237         }
 238         return ret;
 239 }
 240
 241 /*
 242  * This needs to be very careful to not leave IS_ERR pointers around for
 243  * cleanup to trip over.
 244  */
 245 static int rds_iw_setup_qp(struct rds_connection *conn)
 246 {
 247         struct rds_iw_connection *ic = conn->c_transport_data;
 248         struct ib_device *dev = ic->i_cm_id->device;
 249         struct ib_qp_init_attr attr;
 250         struct rds_iw_device *rds_iwdev;
 251         int ret;
 252
 253         /* rds_iw_add_one creates a rds_iw_device object per IB device,
 254          * and allocates a protection domain, memory range and MR pool
 255          * for each.  If that fails for any reason, it will not register
 256          * the rds_iwdev at all.
 257          */
 258         rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
 259         if (rds_iwdev == NULL) {
 260                 if (printk_ratelimit())
 261                         printk(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
 262                                         dev->name);
 263                 return -EOPNOTSUPP;
 264         }
 265
 266         /* Protection domain and memory range */
 267         ic->i_pd = rds_iwdev->pd;
 268         ic->i_mr = rds_iwdev->mr;
 269
 270         ret = rds_iw_init_qp_attrs(&attr, rds_iwdev,
 271                         &ic->i_send_ring, rds_iw_send_cq_comp_handler,
 272                         &ic->i_recv_ring, rds_iw_recv_cq_comp_handler,
 273                         conn);
 274         if (ret < 0)
 275                 goto out;
 276
 277         ic->i_send_cq = attr.send_cq;
 278         ic->i_recv_cq = attr.recv_cq;
 279
 280         /*
 281          * XXX this can fail if max_*_wr is too large?  Are we supposed
 282          * to back off until we get a value that the hardware can support?
 283          */
 284         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 285         if (ret) {
 286                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 287                 goto out;
 288         }
 289
 290         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 291                                            ic->i_send_ring.w_nr *
 292                                                 sizeof(struct rds_header),
 293                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 294         if (ic->i_send_hdrs == NULL) {
 295                 ret = -ENOMEM;
 296                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 297                 goto out;
 298         }
 299
 300         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 301                                            ic->i_recv_ring.w_nr *
 302                                                 sizeof(struct rds_header),
 303                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 304         if (ic->i_recv_hdrs == NULL) {
 305                 ret = -ENOMEM;
 306                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 307                 goto out;
 308         }
 309
 310         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 311                                        &ic->i_ack_dma, GFP_KERNEL);
 312         if (ic->i_ack == NULL) {
 313                 ret = -ENOMEM;
 314                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 315                 goto out;
 316         }
 317
 318         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
 319         if (ic->i_sends == NULL) {
 320                 ret = -ENOMEM;
 321                 rdsdebug("send allocation failed\n");
 322                 goto out;
 323         }
 324         rds_iw_send_init_ring(ic);
 325
 326         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
 327         if (ic->i_recvs == NULL) {
 328                 ret = -ENOMEM;
 329                 rdsdebug("recv allocation failed\n");
 330                 goto out;
 331         }
 332
 333         rds_iw_recv_init_ring(ic);
 334         rds_iw_recv_init_ack(ic);
 335
 336         /* Post receive buffers - as a side effect, this will update
 337          * the posted credit count. */
 338         rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
 339
 340         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 341                  ic->i_send_cq, ic->i_recv_cq);
 342
 343 out:
 344         return ret;
 345 }
 346
 347 static u32 rds_iw_protocol_compatible(const struct rds_iw_connect_private *dp)
 348 {
 349         u16 common;
 350         u32 version = 0;
 351
 352         /* rdma_cm private data is odd - when there is any private data in the
 353          * request, we will be given a pretty large buffer without telling us the
 354          * original size. The only way to tell the difference is by looking at
 355          * the contents, which are initialized to zero.
 356          * If the protocol version fields aren't set, this is a connection attempt
 357          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 358          * We really should have changed this for OFED 1.3 :-( */
 359         if (dp->dp_protocol_major == 0)
 360                 return RDS_PROTOCOL_3_0;
 361
 362         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IW_SUPPORTED_PROTOCOLS;
 363         if (dp->dp_protocol_major == 3 && common) {
 364                 version = RDS_PROTOCOL_3_0;
 365                 while ((common >>= 1) != 0)
 366                         version++;
 367         } else if (printk_ratelimit()) {
 368                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
 369                         "incompatible protocol version %u.%u\n",
 370                         &dp->dp_saddr,
 371                         dp->dp_protocol_major,
 372                         dp->dp_protocol_minor);
 373         }
 374         return version;
 375 }
 376
 377 int rds_iw_cm_handle_connect(struct rdma_cm_id *cm_id,
 378                                     struct rdma_cm_event *event)
 379 {
 380         const struct rds_iw_connect_private *dp = event->param.conn.private_data;
 381         struct rds_iw_connect_private dp_rep;
 382         struct rds_connection *conn = NULL;
 383         struct rds_iw_connection *ic = NULL;
 384         struct rdma_conn_param conn_param;
 385         struct rds_iw_device *rds_iwdev;
 386         u32 version;
 387         int err, destroy = 1;
 388
 389         /* Check whether the remote protocol version matches ours. */
 390         version = rds_iw_protocol_compatible(dp);
 391         if (!version)
 392                 goto out;
 393
 394         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u\n",
 395                  &dp->dp_saddr, &dp->dp_daddr,
 396                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version));
 397
 398         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_iw_transport,
 399                                GFP_KERNEL);
 400         if (IS_ERR(conn)) {
 401                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 402                 conn = NULL;
 403                 goto out;
 404         }
 405
 406         /*
 407          * The connection request may occur while the
 408          * previous connection exist, e.g. in case of failover.
 409          * But as connections may be initiated simultaneously
 410          * by both hosts, we have a random backoff mechanism -
 411          * see the comment above rds_queue_reconnect()
 412          */
 413         mutex_lock(&conn->c_cm_lock);
 414         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 415                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 416                         rdsdebug("incoming connect while connecting\n");
 417                         rds_conn_drop(conn);
 418                         rds_iw_stats_inc(s_iw_listen_closed_stale);
 419                 } else
 420                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 421                         /* Wait and see - our connect may still be succeeding */
 422                         rds_iw_stats_inc(s_iw_connect_raced);
 423                 }
 424                 mutex_unlock(&conn->c_cm_lock);
 425                 goto out;
 426         }
 427
 428         ic = conn->c_transport_data;
 429
 430         rds_iw_set_protocol(conn, version);
 431         rds_iw_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 432
 433         /* If the peer gave us the last packet it saw, process this as if
 434          * we had received a regular ACK. */
 435         if (dp->dp_ack_seq)
 436                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 437
 438         BUG_ON(cm_id->context);
 439         BUG_ON(ic->i_cm_id);
 440
 441         ic->i_cm_id = cm_id;
 442         cm_id->context = conn;
 443
 444         rds_iwdev = ib_get_client_data(cm_id->device, &rds_iw_client);
 445         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
 446
 447         /* We got halfway through setting up the ib_connection, if we
 448          * fail now, we have to take the long route out of this mess. */
 449         destroy = 0;
 450
 451         err = rds_iw_setup_qp(conn);
 452         if (err) {
 453                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", err);
 454                 goto out;
 455         }
 456
 457         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
 458
 459         /* rdma_accept() calls rdma_reject() internally if it fails */
 460         err = rdma_accept(cm_id, &conn_param);
 461         mutex_unlock(&conn->c_cm_lock);
 462         if (err) {
 463                 rds_iw_conn_error(conn, "rdma_accept failed (%d)\n", err);
 464                 goto out;
 465         }
 466
 467         return 0;
 468
 469 out:
 470         rdma_reject(cm_id, NULL, 0);
 471         return destroy;
 472 }
 473
 474
 475 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id)
 476 {
 477         struct rds_connection *conn = cm_id->context;
 478         struct rds_iw_connection *ic = conn->c_transport_data;
 479         struct rdma_conn_param conn_param;
 480         struct rds_iw_connect_private dp;
 481         int ret;
 482
 483         /* If the peer doesn't do protocol negotiation, we must
 484          * default to RDSv3.0 */
 485         rds_iw_set_protocol(conn, RDS_PROTOCOL_3_0);
 486         ic->i_flowctl = rds_iw_sysctl_flow_control;     /* advertise flow control */
 487
 488         ret = rds_iw_setup_qp(conn);
 489         if (ret) {
 490                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", ret);
 491                 goto out;
 492         }
 493
 494         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
 495
 496         ret = rdma_connect(cm_id, &conn_param);
 497         if (ret)
 498                 rds_iw_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 499
 500 out:
 501         /* Beware - returning non-zero tells the rdma_cm to destroy
 502          * the cm_id. We should certainly not do it as long as we still
 503          * "own" the cm_id. */
 504         if (ret) {
 505                 struct rds_iw_connection *ic = conn->c_transport_data;
 506
 507                 if (ic->i_cm_id == cm_id)
 508                         ret = 0;
 509         }
 510         return ret;
 511 }
 512
 513 int rds_iw_conn_connect(struct rds_connection *conn)
 514 {
 515         struct rds_iw_connection *ic = conn->c_transport_data;
 516         struct rds_iw_device *rds_iwdev;
 517         struct sockaddr_in src, dest;
 518         int ret;
 519
 520         /* XXX I wonder what affect the port space has */
 521         /* delegate cm event handler to rdma_transport */
 522         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 523                                      RDMA_PS_TCP);
 524         if (IS_ERR(ic->i_cm_id)) {
 525                 ret = PTR_ERR(ic->i_cm_id);
 526                 ic->i_cm_id = NULL;
 527                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 528                 goto out;
 529         }
 530
 531         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 532
 533         src.sin_family = AF_INET;
 534         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 535         src.sin_port = (__force u16)htons(0);
 536
 537         /* First, bind to the local address and device. */
 538         ret = rdma_bind_addr(ic->i_cm_id, (struct sockaddr *) &src);
 539         if (ret) {
 540                 rdsdebug("rdma_bind_addr(%pI4) failed: %d\n",
 541                                 &conn->c_laddr, ret);
 542                 rdma_destroy_id(ic->i_cm_id);
 543                 ic->i_cm_id = NULL;
 544                 goto out;
 545         }
 546
 547         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 548         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
 549
 550         dest.sin_family = AF_INET;
 551         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 552         dest.sin_port = (__force u16)htons(RDS_PORT);
 553
 554         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 555                                 (struct sockaddr *)&dest,
 556                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 557         if (ret) {
 558                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 559                          ret);
 560                 rdma_destroy_id(ic->i_cm_id);
 561                 ic->i_cm_id = NULL;
 562         }
 563
 564 out:
 565         return ret;
 566 }
 567
 568 /*
 569  * This is so careful about only cleaning up resources that were built up
 570  * so that it can be called at any point during startup.  In fact it
 571  * can be called multiple times for a given connection.
 572  */
 573 void rds_iw_conn_shutdown(struct rds_connection *conn)
 574 {
 575         struct rds_iw_connection *ic = conn->c_transport_data;
 576         int err = 0;
 577         struct ib_qp_attr qp_attr;
 578
 579         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 580                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 581                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 582
 583         if (ic->i_cm_id) {
 584                 struct ib_device *dev = ic->i_cm_id->device;
 585
 586                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 587                 err = rdma_disconnect(ic->i_cm_id);
 588                 if (err) {
 589                         /* Actually this may happen quite frequently, when
 590                          * an outgoing connect raced with an incoming connect.
 591                          */
 592                         rdsdebug("rds_iw_conn_shutdown: failed to disconnect,"
 593                                    " cm: %p err %d\n", ic->i_cm_id, err);
 594                 }
 595
 596                 if (ic->i_cm_id->qp) {
 597                         qp_attr.qp_state = IB_QPS_ERR;
 598                         ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 599                 }
 600
 601                 wait_event(rds_iw_ring_empty_wait,
 602                         rds_iw_ring_empty(&ic->i_send_ring) &&
 603                         rds_iw_ring_empty(&ic->i_recv_ring));
 604
 605                 if (ic->i_send_hdrs)
 606                         ib_dma_free_coherent(dev,
 607                                            ic->i_send_ring.w_nr *
 608                                                 sizeof(struct rds_header),
 609                                            ic->i_send_hdrs,
 610                                            ic->i_send_hdrs_dma);
 611
 612                 if (ic->i_recv_hdrs)
 613                         ib_dma_free_coherent(dev,
 614                                            ic->i_recv_ring.w_nr *
 615                                                 sizeof(struct rds_header),
 616                                            ic->i_recv_hdrs,
 617                                            ic->i_recv_hdrs_dma);
 618
 619                 if (ic->i_ack)
 620                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 621                                              ic->i_ack, ic->i_ack_dma);
 622
 623                 if (ic->i_sends)
 624                         rds_iw_send_clear_ring(ic);
 625                 if (ic->i_recvs)
 626                         rds_iw_recv_clear_ring(ic);
 627
 628                 if (ic->i_cm_id->qp)
 629                         rdma_destroy_qp(ic->i_cm_id);
 630                 if (ic->i_send_cq)
 631                         ib_destroy_cq(ic->i_send_cq);
 632                 if (ic->i_recv_cq)
 633                         ib_destroy_cq(ic->i_recv_cq);
 634
 635                 /*
 636                  * If associated with an rds_iw_device:
 637                  *      Move connection back to the nodev list.
 638                  *      Remove cm_id from the device cm_id list.
 639                  */
 640                 if (ic->rds_iwdev) {
 641
 642                         spin_lock_irq(&ic->rds_iwdev->spinlock);
 643                         BUG_ON(list_empty(&ic->iw_node));
 644                         list_del(&ic->iw_node);
 645                         spin_unlock_irq(&ic->rds_iwdev->spinlock);
 646
 647                         spin_lock_irq(&iw_nodev_conns_lock);
 648                         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 649                         spin_unlock_irq(&iw_nodev_conns_lock);
 650                         rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
 651                         ic->rds_iwdev = NULL;
 652                 }
 653
 654                 rdma_destroy_id(ic->i_cm_id);
 655
 656                 ic->i_cm_id = NULL;
 657                 ic->i_pd = NULL;
 658                 ic->i_mr = NULL;
 659                 ic->i_send_cq = NULL;
 660                 ic->i_recv_cq = NULL;
 661                 ic->i_send_hdrs = NULL;
 662                 ic->i_recv_hdrs = NULL;
 663                 ic->i_ack = NULL;
 664         }
 665         BUG_ON(ic->rds_iwdev);
 666
 667         /* Clear pending transmit */
 668         if (ic->i_rm) {
 669                 rds_message_put(ic->i_rm);
 670                 ic->i_rm = NULL;
 671         }
 672
 673         /* Clear the ACK state */
 674         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 675         rds_iw_set_64bit(&ic->i_ack_next, 0);
 676         ic->i_ack_recv = 0;
 677
 678         /* Clear flow control state */
 679         ic->i_flowctl = 0;
 680         atomic_set(&ic->i_credits, 0);
 681
 682         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
 683         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
 684
 685         if (ic->i_iwinc) {
 686                 rds_inc_put(&ic->i_iwinc->ii_inc);
 687                 ic->i_iwinc = NULL;
 688         }
 689
 690         vfree(ic->i_sends);
 691         ic->i_sends = NULL;
 692         vfree(ic->i_recvs);
 693         ic->i_recvs = NULL;
 694         rdsdebug("shutdown complete\n");
 695 }
 696
 697 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 698 {
 699         struct rds_iw_connection *ic;
 700         unsigned long flags;
 701
 702         /* XXX too lazy? */
 703         ic = kzalloc(sizeof(struct rds_iw_connection), GFP_KERNEL);
 704         if (ic == NULL)
 705                 return -ENOMEM;
 706
 707         INIT_LIST_HEAD(&ic->iw_node);
 708         mutex_init(&ic->i_recv_mutex);
 709
 710         /*
 711          * rds_iw_conn_shutdown() waits for these to be emptied so they
 712          * must be initialized before it can be called.
 713          */
 714         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
 715         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
 716
 717         ic->conn = conn;
 718         conn->c_transport_data = ic;
 719
 720         spin_lock_irqsave(&iw_nodev_conns_lock, flags);
 721         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 722         spin_unlock_irqrestore(&iw_nodev_conns_lock, flags);
 723
 724
 725         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 726         return 0;
 727 }
 728
 729 void rds_iw_conn_free(void *arg)
 730 {
 731         struct rds_iw_connection *ic = arg;
 732         rdsdebug("ic %p\n", ic);
 733         list_del(&ic->iw_node);
 734         kfree(ic);
 735 }
 736
 737 /*
 738  * An error occurred on the connection
 739  */
 740 void
 741 __rds_iw_conn_error(struct rds_connection *conn, const char *fmt, ...)
 742 {
 743         va_list ap;
 744
 745         rds_conn_drop(conn);
 746
 747         va_start(ap, fmt);
 748         vprintk(fmt, ap);
 749         va_end(ap);
 750 }