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         unsigned int send_size, recv_size;
 183         int ret;
 184
 185         /* The offset of 1 is to accommodate the additional ACK WR. */
 186         send_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_send_wr + 1);
 187         recv_size = min_t(unsigned int, rds_iwdev->max_wrs, rds_iw_sysctl_max_recv_wr + 1);
 188         rds_iw_ring_resize(send_ring, send_size - 1);
 189         rds_iw_ring_resize(recv_ring, recv_size - 1);
 190
 191         memset(attr, 0, sizeof(*attr));
 192         attr->event_handler = rds_iw_qp_event_handler;
 193         attr->qp_context = context;
 194         attr->cap.max_send_wr = send_size;
 195         attr->cap.max_recv_wr = recv_size;
 196         attr->cap.max_send_sge = rds_iwdev->max_sge;
 197         attr->cap.max_recv_sge = RDS_IW_RECV_SGE;
 198         attr->sq_sig_type = IB_SIGNAL_REQ_WR;
 199         attr->qp_type = IB_QPT_RC;
 200
 201         attr->send_cq = ib_create_cq(dev, send_cq_handler,
 202                                      rds_iw_cq_event_handler,
 203                                      context, send_size, 0);
 204         if (IS_ERR(attr->send_cq)) {
 205                 ret = PTR_ERR(attr->send_cq);
 206                 attr->send_cq = NULL;
 207                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 208                 goto out;
 209         }
 210
 211         attr->recv_cq = ib_create_cq(dev, recv_cq_handler,
 212                                      rds_iw_cq_event_handler,
 213                                      context, recv_size, 0);
 214         if (IS_ERR(attr->recv_cq)) {
 215                 ret = PTR_ERR(attr->recv_cq);
 216                 attr->recv_cq = NULL;
 217                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 218                 goto out;
 219         }
 220
 221         ret = ib_req_notify_cq(attr->send_cq, IB_CQ_NEXT_COMP);
 222         if (ret) {
 223                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 224                 goto out;
 225         }
 226
 227         ret = ib_req_notify_cq(attr->recv_cq, IB_CQ_SOLICITED);
 228         if (ret) {
 229                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 230                 goto out;
 231         }
 232
 233 out:
 234         if (ret) {
 235                 if (attr->send_cq)
 236                         ib_destroy_cq(attr->send_cq);
 237                 if (attr->recv_cq)
 238                         ib_destroy_cq(attr->recv_cq);
 239         }
 240         return ret;
 241 }
 242
 243 /*
 244  * This needs to be very careful to not leave IS_ERR pointers around for
 245  * cleanup to trip over.
 246  */
 247 static int rds_iw_setup_qp(struct rds_connection *conn)
 248 {
 249         struct rds_iw_connection *ic = conn->c_transport_data;
 250         struct ib_device *dev = ic->i_cm_id->device;
 251         struct ib_qp_init_attr attr;
 252         struct rds_iw_device *rds_iwdev;
 253         int ret;
 254
 255         /* rds_iw_add_one creates a rds_iw_device object per IB device,
 256          * and allocates a protection domain, memory range and MR pool
 257          * for each.  If that fails for any reason, it will not register
 258          * the rds_iwdev at all.
 259          */
 260         rds_iwdev = ib_get_client_data(dev, &rds_iw_client);
 261         if (!rds_iwdev) {
 262                 printk_ratelimited(KERN_NOTICE "RDS/IW: No client_data for device %s\n",
 263                                         dev->name);
 264                 return -EOPNOTSUPP;
 265         }
 266
 267         /* Protection domain and memory range */
 268         ic->i_pd = rds_iwdev->pd;
 269         ic->i_mr = rds_iwdev->mr;
 270
 271         ret = rds_iw_init_qp_attrs(&attr, rds_iwdev,
 272                         &ic->i_send_ring, rds_iw_send_cq_comp_handler,
 273                         &ic->i_recv_ring, rds_iw_recv_cq_comp_handler,
 274                         conn);
 275         if (ret < 0)
 276                 goto out;
 277
 278         ic->i_send_cq = attr.send_cq;
 279         ic->i_recv_cq = attr.recv_cq;
 280
 281         /*
 282          * XXX this can fail if max_*_wr is too large?  Are we supposed
 283          * to back off until we get a value that the hardware can support?
 284          */
 285         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 286         if (ret) {
 287                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 288                 goto out;
 289         }
 290
 291         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 292                                            ic->i_send_ring.w_nr *
 293                                                 sizeof(struct rds_header),
 294                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 295         if (!ic->i_send_hdrs) {
 296                 ret = -ENOMEM;
 297                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 298                 goto out;
 299         }
 300
 301         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 302                                            ic->i_recv_ring.w_nr *
 303                                                 sizeof(struct rds_header),
 304                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 305         if (!ic->i_recv_hdrs) {
 306                 ret = -ENOMEM;
 307                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 308                 goto out;
 309         }
 310
 311         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 312                                        &ic->i_ack_dma, GFP_KERNEL);
 313         if (!ic->i_ack) {
 314                 ret = -ENOMEM;
 315                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 316                 goto out;
 317         }
 318
 319         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_iw_send_work));
 320         if (!ic->i_sends) {
 321                 ret = -ENOMEM;
 322                 rdsdebug("send allocation failed\n");
 323                 goto out;
 324         }
 325         rds_iw_send_init_ring(ic);
 326
 327         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_iw_recv_work));
 328         if (!ic->i_recvs) {
 329                 ret = -ENOMEM;
 330                 rdsdebug("recv allocation failed\n");
 331                 goto out;
 332         }
 333
 334         rds_iw_recv_init_ring(ic);
 335         rds_iw_recv_init_ack(ic);
 336
 337         /* Post receive buffers - as a side effect, this will update
 338          * the posted credit count. */
 339         rds_iw_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
 340
 341         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 342                  ic->i_send_cq, ic->i_recv_cq);
 343
 344 out:
 345         return ret;
 346 }
 347
 348 static u32 rds_iw_protocol_compatible(const struct rds_iw_connect_private *dp)
 349 {
 350         u16 common;
 351         u32 version = 0;
 352
 353         /* rdma_cm private data is odd - when there is any private data in the
 354          * request, we will be given a pretty large buffer without telling us the
 355          * original size. The only way to tell the difference is by looking at
 356          * the contents, which are initialized to zero.
 357          * If the protocol version fields aren't set, this is a connection attempt
 358          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 359          * We really should have changed this for OFED 1.3 :-( */
 360         if (dp->dp_protocol_major == 0)
 361                 return RDS_PROTOCOL_3_0;
 362
 363         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IW_SUPPORTED_PROTOCOLS;
 364         if (dp->dp_protocol_major == 3 && common) {
 365                 version = RDS_PROTOCOL_3_0;
 366                 while ((common >>= 1) != 0)
 367                         version++;
 368         }
 369         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using "
 370                         "incompatible protocol version %u.%u\n",
 371                         &dp->dp_saddr,
 372                         dp->dp_protocol_major,
 373                         dp->dp_protocol_minor);
 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                 mutex_unlock(&conn->c_cm_lock);
 455                 goto out;
 456         }
 457
 458         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
 459
 460         /* rdma_accept() calls rdma_reject() internally if it fails */
 461         err = rdma_accept(cm_id, &conn_param);
 462         mutex_unlock(&conn->c_cm_lock);
 463         if (err) {
 464                 rds_iw_conn_error(conn, "rdma_accept failed (%d)\n", err);
 465                 goto out;
 466         }
 467
 468         return 0;
 469
 470 out:
 471         rdma_reject(cm_id, NULL, 0);
 472         return destroy;
 473 }
 474
 475
 476 int rds_iw_cm_initiate_connect(struct rdma_cm_id *cm_id)
 477 {
 478         struct rds_connection *conn = cm_id->context;
 479         struct rds_iw_connection *ic = conn->c_transport_data;
 480         struct rdma_conn_param conn_param;
 481         struct rds_iw_connect_private dp;
 482         int ret;
 483
 484         /* If the peer doesn't do protocol negotiation, we must
 485          * default to RDSv3.0 */
 486         rds_iw_set_protocol(conn, RDS_PROTOCOL_3_0);
 487         ic->i_flowctl = rds_iw_sysctl_flow_control;     /* advertise flow control */
 488
 489         ret = rds_iw_setup_qp(conn);
 490         if (ret) {
 491                 rds_iw_conn_error(conn, "rds_iw_setup_qp failed (%d)\n", ret);
 492                 goto out;
 493         }
 494
 495         rds_iw_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
 496
 497         ret = rdma_connect(cm_id, &conn_param);
 498         if (ret)
 499                 rds_iw_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 500
 501 out:
 502         /* Beware - returning non-zero tells the rdma_cm to destroy
 503          * the cm_id. We should certainly not do it as long as we still
 504          * "own" the cm_id. */
 505         if (ret) {
 506                 struct rds_iw_connection *ic = conn->c_transport_data;
 507
 508                 if (ic->i_cm_id == cm_id)
 509                         ret = 0;
 510         }
 511         return ret;
 512 }
 513
 514 int rds_iw_conn_connect(struct rds_connection *conn)
 515 {
 516         struct rds_iw_connection *ic = conn->c_transport_data;
 517         struct rds_iw_device *rds_iwdev;
 518         struct sockaddr_in src, dest;
 519         int ret;
 520
 521         /* XXX I wonder what affect the port space has */
 522         /* delegate cm event handler to rdma_transport */
 523         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 524                                      RDMA_PS_TCP, IB_QPT_RC);
 525         if (IS_ERR(ic->i_cm_id)) {
 526                 ret = PTR_ERR(ic->i_cm_id);
 527                 ic->i_cm_id = NULL;
 528                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 529                 goto out;
 530         }
 531
 532         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 533
 534         src.sin_family = AF_INET;
 535         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 536         src.sin_port = (__force u16)htons(0);
 537
 538         /* First, bind to the local address and device. */
 539         ret = rdma_bind_addr(ic->i_cm_id, (struct sockaddr *) &src);
 540         if (ret) {
 541                 rdsdebug("rdma_bind_addr(%pI4) failed: %d\n",
 542                                 &conn->c_laddr, ret);
 543                 rdma_destroy_id(ic->i_cm_id);
 544                 ic->i_cm_id = NULL;
 545                 goto out;
 546         }
 547
 548         rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
 549         ic->i_dma_local_lkey = rds_iwdev->dma_local_lkey;
 550
 551         dest.sin_family = AF_INET;
 552         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 553         dest.sin_port = (__force u16)htons(RDS_PORT);
 554
 555         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 556                                 (struct sockaddr *)&dest,
 557                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 558         if (ret) {
 559                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 560                          ret);
 561                 rdma_destroy_id(ic->i_cm_id);
 562                 ic->i_cm_id = NULL;
 563         }
 564
 565 out:
 566         return ret;
 567 }
 568
 569 /*
 570  * This is so careful about only cleaning up resources that were built up
 571  * so that it can be called at any point during startup.  In fact it
 572  * can be called multiple times for a given connection.
 573  */
 574 void rds_iw_conn_shutdown(struct rds_connection *conn)
 575 {
 576         struct rds_iw_connection *ic = conn->c_transport_data;
 577         int err = 0;
 578         struct ib_qp_attr qp_attr;
 579
 580         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 581                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 582                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 583
 584         if (ic->i_cm_id) {
 585                 struct ib_device *dev = ic->i_cm_id->device;
 586
 587                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 588                 err = rdma_disconnect(ic->i_cm_id);
 589                 if (err) {
 590                         /* Actually this may happen quite frequently, when
 591                          * an outgoing connect raced with an incoming connect.
 592                          */
 593                         rdsdebug("rds_iw_conn_shutdown: failed to disconnect,"
 594                                    " cm: %p err %d\n", ic->i_cm_id, err);
 595                 }
 596
 597                 if (ic->i_cm_id->qp) {
 598                         qp_attr.qp_state = IB_QPS_ERR;
 599                         ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 600                 }
 601
 602                 wait_event(rds_iw_ring_empty_wait,
 603                         rds_iw_ring_empty(&ic->i_send_ring) &&
 604                         rds_iw_ring_empty(&ic->i_recv_ring));
 605
 606                 if (ic->i_send_hdrs)
 607                         ib_dma_free_coherent(dev,
 608                                            ic->i_send_ring.w_nr *
 609                                                 sizeof(struct rds_header),
 610                                            ic->i_send_hdrs,
 611                                            ic->i_send_hdrs_dma);
 612
 613                 if (ic->i_recv_hdrs)
 614                         ib_dma_free_coherent(dev,
 615                                            ic->i_recv_ring.w_nr *
 616                                                 sizeof(struct rds_header),
 617                                            ic->i_recv_hdrs,
 618                                            ic->i_recv_hdrs_dma);
 619
 620                 if (ic->i_ack)
 621                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 622                                              ic->i_ack, ic->i_ack_dma);
 623
 624                 if (ic->i_sends)
 625                         rds_iw_send_clear_ring(ic);
 626                 if (ic->i_recvs)
 627                         rds_iw_recv_clear_ring(ic);
 628
 629                 if (ic->i_cm_id->qp)
 630                         rdma_destroy_qp(ic->i_cm_id);
 631                 if (ic->i_send_cq)
 632                         ib_destroy_cq(ic->i_send_cq);
 633                 if (ic->i_recv_cq)
 634                         ib_destroy_cq(ic->i_recv_cq);
 635
 636                 /*
 637                  * If associated with an rds_iw_device:
 638                  *      Move connection back to the nodev list.
 639                  *      Remove cm_id from the device cm_id list.
 640                  */
 641                 if (ic->rds_iwdev)
 642                         rds_iw_remove_conn(ic->rds_iwdev, conn);
 643
 644                 rdma_destroy_id(ic->i_cm_id);
 645
 646                 ic->i_cm_id = NULL;
 647                 ic->i_pd = NULL;
 648                 ic->i_mr = NULL;
 649                 ic->i_send_cq = NULL;
 650                 ic->i_recv_cq = NULL;
 651                 ic->i_send_hdrs = NULL;
 652                 ic->i_recv_hdrs = NULL;
 653                 ic->i_ack = NULL;
 654         }
 655         BUG_ON(ic->rds_iwdev);
 656
 657         /* Clear pending transmit */
 658         if (ic->i_rm) {
 659                 rds_message_put(ic->i_rm);
 660                 ic->i_rm = NULL;
 661         }
 662
 663         /* Clear the ACK state */
 664         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 665 #ifdef KERNEL_HAS_ATOMIC64
 666         atomic64_set(&ic->i_ack_next, 0);
 667 #else
 668         ic->i_ack_next = 0;
 669 #endif
 670         ic->i_ack_recv = 0;
 671
 672         /* Clear flow control state */
 673         ic->i_flowctl = 0;
 674         atomic_set(&ic->i_credits, 0);
 675
 676         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
 677         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
 678
 679         if (ic->i_iwinc) {
 680                 rds_inc_put(&ic->i_iwinc->ii_inc);
 681                 ic->i_iwinc = NULL;
 682         }
 683
 684         vfree(ic->i_sends);
 685         ic->i_sends = NULL;
 686         vfree(ic->i_recvs);
 687         ic->i_recvs = NULL;
 688         rdsdebug("shutdown complete\n");
 689 }
 690
 691 int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 692 {
 693         struct rds_iw_connection *ic;
 694         unsigned long flags;
 695
 696         /* XXX too lazy? */
 697         ic = kzalloc(sizeof(struct rds_iw_connection), gfp);
 698         if (!ic)
 699                 return -ENOMEM;
 700
 701         INIT_LIST_HEAD(&ic->iw_node);
 702         tasklet_init(&ic->i_recv_tasklet, rds_iw_recv_tasklet_fn,
 703                      (unsigned long) ic);
 704         mutex_init(&ic->i_recv_mutex);
 705 #ifndef KERNEL_HAS_ATOMIC64
 706         spin_lock_init(&ic->i_ack_lock);
 707 #endif
 708
 709         /*
 710          * rds_iw_conn_shutdown() waits for these to be emptied so they
 711          * must be initialized before it can be called.
 712          */
 713         rds_iw_ring_init(&ic->i_send_ring, rds_iw_sysctl_max_send_wr);
 714         rds_iw_ring_init(&ic->i_recv_ring, rds_iw_sysctl_max_recv_wr);
 715
 716         ic->conn = conn;
 717         conn->c_transport_data = ic;
 718
 719         spin_lock_irqsave(&iw_nodev_conns_lock, flags);
 720         list_add_tail(&ic->iw_node, &iw_nodev_conns);
 721         spin_unlock_irqrestore(&iw_nodev_conns_lock, flags);
 722
 723
 724         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 725         return 0;
 726 }
 727
 728 /*
 729  * Free a connection. Connection must be shut down and not set for reconnect.
 730  */
 731 void rds_iw_conn_free(void *arg)
 732 {
 733         struct rds_iw_connection *ic = arg;
 734         spinlock_t      *lock_ptr;
 735
 736         rdsdebug("ic %p\n", ic);
 737
 738         /*
 739          * Conn is either on a dev's list or on the nodev list.
 740          * A race with shutdown() or connect() would cause problems
 741          * (since rds_iwdev would change) but that should never happen.
 742          */
 743         lock_ptr = ic->rds_iwdev ? &ic->rds_iwdev->spinlock : &iw_nodev_conns_lock;
 744
 745         spin_lock_irq(lock_ptr);
 746         list_del(&ic->iw_node);
 747         spin_unlock_irq(lock_ptr);
 748
 749         kfree(ic);
 750 }
 751
 752 /*
 753  * An error occurred on the connection
 754  */
 755 void
 756 __rds_iw_conn_error(struct rds_connection *conn, const char *fmt, ...)
 757 {
 758         va_list ap;
 759
 760         rds_conn_drop(conn);
 761
 762         va_start(ap, fmt);
 763         vprintk(fmt, ap);
 764         va_end(ap);
 765 }