net/rds/ib_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
  38 #include "rds.h"
  39 #include "ib.h"
  40
  41 /*
  42  * Set the selected protocol version
  43  */
  44 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  45 {
  46         conn->c_version = version;
  47 }
  48
  49 /*
  50  * Set up flow control
  51  */
  52 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  53 {
  54         struct rds_ib_connection *ic = conn->c_transport_data;
  55
  56         if (rds_ib_sysctl_flow_control && credits != 0) {
  57                 /* We're doing flow control */
  58                 ic->i_flowctl = 1;
  59                 rds_ib_send_add_credits(conn, credits);
  60         } else {
  61                 ic->i_flowctl = 0;
  62         }
  63 }
  64
  65 /*
  66  * Tune RNR behavior. Without flow control, we use a rather
  67  * low timeout, but not the absolute minimum - this should
  68  * be tunable.
  69  *
  70  * We already set the RNR retry count to 7 (which is the
  71  * smallest infinite number :-) above.
  72  * If flow control is off, we want to change this back to 0
  73  * so that we learn quickly when our credit accounting is
  74  * buggy.
  75  *
  76  * Caller passes in a qp_attr pointer - don't waste stack spacv
  77  * by allocation this twice.
  78  */
  79 static void
  80 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  81 {
  82         int ret;
  83
  84         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  85         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  86         if (ret)
  87                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  88 }
  89
  90 /*
  91  * Connection established.
  92  * We get here for both outgoing and incoming connection.
  93  */
  94 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  95 {
  96         const struct rds_ib_connect_private *dp = NULL;
  97         struct rds_ib_connection *ic = conn->c_transport_data;
  98         struct ib_qp_attr qp_attr;
  99         int err;
 100
 101         if (event->param.conn.private_data_len >= sizeof(*dp)) {
 102                 dp = event->param.conn.private_data;
 103
 104                 /* make sure it isn't empty data */
 105                 if (dp->dp_protocol_major) {
 106                         rds_ib_set_protocol(conn,
 107                                 RDS_PROTOCOL(dp->dp_protocol_major,
 108                                 dp->dp_protocol_minor));
 109                         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 110                 }
 111         }
 112
 113         if (conn->c_version < RDS_PROTOCOL(3,1)) {
 114                 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
 115                        " no longer supported\n",
 116                        &conn->c_faddr,
 117                        RDS_PROTOCOL_MAJOR(conn->c_version),
 118                        RDS_PROTOCOL_MINOR(conn->c_version));
 119                 rds_conn_destroy(conn);
 120                 return;
 121         } else {
 122                 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
 123                        &conn->c_faddr,
 124                        RDS_PROTOCOL_MAJOR(conn->c_version),
 125                        RDS_PROTOCOL_MINOR(conn->c_version),
 126                        ic->i_flowctl ? ", flow control" : "");
 127         }
 128
 129         /*
 130          * Init rings and fill recv. this needs to wait until protocol negotiation
 131          * is complete, since ring layout is different from 3.0 to 3.1.
 132          */
 133         rds_ib_send_init_ring(ic);
 134         rds_ib_recv_init_ring(ic);
 135         /* Post receive buffers - as a side effect, this will update
 136          * the posted credit count. */
 137         rds_ib_recv_refill(conn, 1);
 138
 139         /* Tune RNR behavior */
 140         rds_ib_tune_rnr(ic, &qp_attr);
 141
 142         qp_attr.qp_state = IB_QPS_RTS;
 143         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 144         if (err)
 145                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 146
 147         /* update ib_device with this local ipaddr */
 148         err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
 149         if (err)
 150                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 151                         err);
 152
 153         /* If the peer gave us the last packet it saw, process this as if
 154          * we had received a regular ACK. */
 155         if (dp && dp->dp_ack_seq)
 156                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 157
 158         rds_connect_complete(conn);
 159 }
 160
 161 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 162                         struct rdma_conn_param *conn_param,
 163                         struct rds_ib_connect_private *dp,
 164                         u32 protocol_version,
 165                         u32 max_responder_resources,
 166                         u32 max_initiator_depth)
 167 {
 168         struct rds_ib_connection *ic = conn->c_transport_data;
 169         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 170
 171         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 172
 173         conn_param->responder_resources =
 174                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 175         conn_param->initiator_depth =
 176                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 177         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 178         conn_param->rnr_retry_count = 7;
 179
 180         if (dp) {
 181                 memset(dp, 0, sizeof(*dp));
 182                 dp->dp_saddr = conn->c_laddr;
 183                 dp->dp_daddr = conn->c_faddr;
 184                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 185                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 186                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 187                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
 188
 189                 /* Advertise flow control */
 190                 if (ic->i_flowctl) {
 191                         unsigned int credits;
 192
 193                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 194                         dp->dp_credit = cpu_to_be32(credits);
 195                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 196                 }
 197
 198                 conn_param->private_data = dp;
 199                 conn_param->private_data_len = sizeof(*dp);
 200         }
 201 }
 202
 203 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 204 {
 205         rdsdebug("event %u data %p\n", event->event, data);
 206 }
 207
 208 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 209 {
 210         struct rds_connection *conn = data;
 211         struct rds_ib_connection *ic = conn->c_transport_data;
 212
 213         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
 214
 215         switch (event->event) {
 216         case IB_EVENT_COMM_EST:
 217                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 218                 break;
 219         default:
 220                 rdsdebug("Fatal QP Event %u "
 221                         "- connection %pI4->%pI4, reconnecting\n",
 222                         event->event, &conn->c_laddr, &conn->c_faddr);
 223                 rds_conn_drop(conn);
 224                 break;
 225         }
 226 }
 227
 228 /*
 229  * This needs to be very careful to not leave IS_ERR pointers around for
 230  * cleanup to trip over.
 231  */
 232 static int rds_ib_setup_qp(struct rds_connection *conn)
 233 {
 234         struct rds_ib_connection *ic = conn->c_transport_data;
 235         struct ib_device *dev = ic->i_cm_id->device;
 236         struct ib_qp_init_attr attr;
 237         struct rds_ib_device *rds_ibdev;
 238         int ret;
 239
 240         /*
 241          * It's normal to see a null device if an incoming connection races
 242          * with device removal, so we don't print a warning.
 243          */
 244         rds_ibdev = rds_ib_get_client_data(dev);
 245         if (!rds_ibdev)
 246                 return -EOPNOTSUPP;
 247
 248         /* add the conn now so that connection establishment has the dev */
 249         rds_ib_add_conn(rds_ibdev, conn);
 250
 251         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 252                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 253         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 254                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 255
 256         /* Protection domain and memory range */
 257         ic->i_pd = rds_ibdev->pd;
 258         ic->i_mr = rds_ibdev->mr;
 259
 260         ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
 261                                      rds_ib_cq_event_handler, conn,
 262                                      ic->i_send_ring.w_nr + 1, 0);
 263         if (IS_ERR(ic->i_send_cq)) {
 264                 ret = PTR_ERR(ic->i_send_cq);
 265                 ic->i_send_cq = NULL;
 266                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 267                 goto out;
 268         }
 269
 270         ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
 271                                      rds_ib_cq_event_handler, conn,
 272                                      ic->i_recv_ring.w_nr, 0);
 273         if (IS_ERR(ic->i_recv_cq)) {
 274                 ret = PTR_ERR(ic->i_recv_cq);
 275                 ic->i_recv_cq = NULL;
 276                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 277                 goto out;
 278         }
 279
 280         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 281         if (ret) {
 282                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 283                 goto out;
 284         }
 285
 286         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 287         if (ret) {
 288                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 289                 goto out;
 290         }
 291
 292         /* XXX negotiate max send/recv with remote? */
 293         memset(&attr, 0, sizeof(attr));
 294         attr.event_handler = rds_ib_qp_event_handler;
 295         attr.qp_context = conn;
 296         /* + 1 to allow for the single ack message */
 297         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
 298         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 299         attr.cap.max_send_sge = rds_ibdev->max_sge;
 300         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 301         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 302         attr.qp_type = IB_QPT_RC;
 303         attr.send_cq = ic->i_send_cq;
 304         attr.recv_cq = ic->i_recv_cq;
 305
 306         /*
 307          * XXX this can fail if max_*_wr is too large?  Are we supposed
 308          * to back off until we get a value that the hardware can support?
 309          */
 310         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 311         if (ret) {
 312                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 313                 goto out;
 314         }
 315
 316         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 317                                            ic->i_send_ring.w_nr *
 318                                                 sizeof(struct rds_header),
 319                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 320         if (!ic->i_send_hdrs) {
 321                 ret = -ENOMEM;
 322                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 323                 goto out;
 324         }
 325
 326         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 327                                            ic->i_recv_ring.w_nr *
 328                                                 sizeof(struct rds_header),
 329                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 330         if (!ic->i_recv_hdrs) {
 331                 ret = -ENOMEM;
 332                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 333                 goto out;
 334         }
 335
 336         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 337                                        &ic->i_ack_dma, GFP_KERNEL);
 338         if (!ic->i_ack) {
 339                 ret = -ENOMEM;
 340                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 341                 goto out;
 342         }
 343
 344         ic->i_sends = vmalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 345                                    ibdev_to_node(dev));
 346         if (!ic->i_sends) {
 347                 ret = -ENOMEM;
 348                 rdsdebug("send allocation failed\n");
 349                 goto out;
 350         }
 351         memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
 352
 353         ic->i_recvs = vmalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 354                                    ibdev_to_node(dev));
 355         if (!ic->i_recvs) {
 356                 ret = -ENOMEM;
 357                 rdsdebug("recv allocation failed\n");
 358                 goto out;
 359         }
 360         memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
 361
 362         rds_ib_recv_init_ack(ic);
 363
 364         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 365                  ic->i_send_cq, ic->i_recv_cq);
 366
 367 out:
 368         rds_ib_dev_put(rds_ibdev);
 369         return ret;
 370 }
 371
 372 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 373 {
 374         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 375         u16 common;
 376         u32 version = 0;
 377
 378         /*
 379          * rdma_cm private data is odd - when there is any private data in the
 380          * request, we will be given a pretty large buffer without telling us the
 381          * original size. The only way to tell the difference is by looking at
 382          * the contents, which are initialized to zero.
 383          * If the protocol version fields aren't set, this is a connection attempt
 384          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 385          * We really should have changed this for OFED 1.3 :-(
 386          */
 387
 388         /* Be paranoid. RDS always has privdata */
 389         if (!event->param.conn.private_data_len) {
 390                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 391                         "rejecting\n");
 392                 return 0;
 393         }
 394
 395         /* Even if len is crap *now* I still want to check it. -ASG */
 396         if (event->param.conn.private_data_len < sizeof (*dp) ||
 397             dp->dp_protocol_major == 0)
 398                 return RDS_PROTOCOL_3_0;
 399
 400         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 401         if (dp->dp_protocol_major == 3 && common) {
 402                 version = RDS_PROTOCOL_3_0;
 403                 while ((common >>= 1) != 0)
 404                         version++;
 405         } else if (printk_ratelimit()) {
 406                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
 407                         "incompatible protocol version %u.%u\n",
 408                         &dp->dp_saddr,
 409                         dp->dp_protocol_major,
 410                         dp->dp_protocol_minor);
 411         }
 412         return version;
 413 }
 414
 415 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 416                                     struct rdma_cm_event *event)
 417 {
 418         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 419         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 420         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 421         struct rds_ib_connect_private dp_rep;
 422         struct rds_connection *conn = NULL;
 423         struct rds_ib_connection *ic = NULL;
 424         struct rdma_conn_param conn_param;
 425         u32 version;
 426         int err = 1, destroy = 1;
 427
 428         /* Check whether the remote protocol version matches ours. */
 429         version = rds_ib_protocol_compatible(event);
 430         if (!version)
 431                 goto out;
 432
 433         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 434                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 435                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 436                  (unsigned long long)be64_to_cpu(lguid),
 437                  (unsigned long long)be64_to_cpu(fguid));
 438
 439         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
 440                                GFP_KERNEL);
 441         if (IS_ERR(conn)) {
 442                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 443                 conn = NULL;
 444                 goto out;
 445         }
 446
 447         /*
 448          * The connection request may occur while the
 449          * previous connection exist, e.g. in case of failover.
 450          * But as connections may be initiated simultaneously
 451          * by both hosts, we have a random backoff mechanism -
 452          * see the comment above rds_queue_reconnect()
 453          */
 454         mutex_lock(&conn->c_cm_lock);
 455         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 456                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 457                         rdsdebug("incoming connect while connecting\n");
 458                         rds_conn_drop(conn);
 459                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 460                 } else
 461                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 462                         /* Wait and see - our connect may still be succeeding */
 463                         rds_ib_stats_inc(s_ib_connect_raced);
 464                 }
 465                 goto out;
 466         }
 467
 468         ic = conn->c_transport_data;
 469
 470         rds_ib_set_protocol(conn, version);
 471         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 472
 473         /* If the peer gave us the last packet it saw, process this as if
 474          * we had received a regular ACK. */
 475         if (dp->dp_ack_seq)
 476                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 477
 478         BUG_ON(cm_id->context);
 479         BUG_ON(ic->i_cm_id);
 480
 481         ic->i_cm_id = cm_id;
 482         cm_id->context = conn;
 483
 484         /* We got halfway through setting up the ib_connection, if we
 485          * fail now, we have to take the long route out of this mess. */
 486         destroy = 0;
 487
 488         err = rds_ib_setup_qp(conn);
 489         if (err) {
 490                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 491                 mutex_unlock(&conn->c_cm_lock);
 492                 goto out;
 493         }
 494
 495         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 496                 event->param.conn.responder_resources,
 497                 event->param.conn.initiator_depth);
 498
 499         /* rdma_accept() calls rdma_reject() internally if it fails */
 500         err = rdma_accept(cm_id, &conn_param);
 501         if (err)
 502                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 503
 504 out:
 505         if (conn)
 506                 mutex_unlock(&conn->c_cm_lock);
 507         if (err)
 508                 rdma_reject(cm_id, NULL, 0);
 509         return destroy;
 510 }
 511
 512
 513 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 514 {
 515         struct rds_connection *conn = cm_id->context;
 516         struct rds_ib_connection *ic = conn->c_transport_data;
 517         struct rdma_conn_param conn_param;
 518         struct rds_ib_connect_private dp;
 519         int ret;
 520
 521         /* If the peer doesn't do protocol negotiation, we must
 522          * default to RDSv3.0 */
 523         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 524         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 525
 526         ret = rds_ib_setup_qp(conn);
 527         if (ret) {
 528                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 529                 goto out;
 530         }
 531
 532         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 533                 UINT_MAX, UINT_MAX);
 534         ret = rdma_connect(cm_id, &conn_param);
 535         if (ret)
 536                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 537
 538 out:
 539         /* Beware - returning non-zero tells the rdma_cm to destroy
 540          * the cm_id. We should certainly not do it as long as we still
 541          * "own" the cm_id. */
 542         if (ret) {
 543                 if (ic->i_cm_id == cm_id)
 544                         ret = 0;
 545         }
 546         return ret;
 547 }
 548
 549 int rds_ib_conn_connect(struct rds_connection *conn)
 550 {
 551         struct rds_ib_connection *ic = conn->c_transport_data;
 552         struct sockaddr_in src, dest;
 553         int ret;
 554
 555         /* XXX I wonder what affect the port space has */
 556         /* delegate cm event handler to rdma_transport */
 557         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 558                                      RDMA_PS_TCP);
 559         if (IS_ERR(ic->i_cm_id)) {
 560                 ret = PTR_ERR(ic->i_cm_id);
 561                 ic->i_cm_id = NULL;
 562                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 563                 goto out;
 564         }
 565
 566         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 567
 568         src.sin_family = AF_INET;
 569         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 570         src.sin_port = (__force u16)htons(0);
 571
 572         dest.sin_family = AF_INET;
 573         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 574         dest.sin_port = (__force u16)htons(RDS_PORT);
 575
 576         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 577                                 (struct sockaddr *)&dest,
 578                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 579         if (ret) {
 580                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 581                          ret);
 582                 rdma_destroy_id(ic->i_cm_id);
 583                 ic->i_cm_id = NULL;
 584         }
 585
 586 out:
 587         return ret;
 588 }
 589
 590 /*
 591  * This is so careful about only cleaning up resources that were built up
 592  * so that it can be called at any point during startup.  In fact it
 593  * can be called multiple times for a given connection.
 594  */
 595 void rds_ib_conn_shutdown(struct rds_connection *conn)
 596 {
 597         struct rds_ib_connection *ic = conn->c_transport_data;
 598         int err = 0;
 599
 600         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 601                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 602                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 603
 604         if (ic->i_cm_id) {
 605                 struct ib_device *dev = ic->i_cm_id->device;
 606
 607                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 608                 err = rdma_disconnect(ic->i_cm_id);
 609                 if (err) {
 610                         /* Actually this may happen quite frequently, when
 611                          * an outgoing connect raced with an incoming connect.
 612                          */
 613                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 614                                 ic->i_cm_id, err);
 615                 }
 616
 617                 /*
 618                  * Don't wait for the send ring to be empty -- there may be completed
 619                  * non-signaled entries sitting on there. We unmap these below.
 620                  */
 621                 wait_event(rds_ib_ring_empty_wait,
 622                         rds_ib_ring_empty(&ic->i_recv_ring));
 623
 624                 if (ic->i_send_hdrs)
 625                         ib_dma_free_coherent(dev,
 626                                            ic->i_send_ring.w_nr *
 627                                                 sizeof(struct rds_header),
 628                                            ic->i_send_hdrs,
 629                                            ic->i_send_hdrs_dma);
 630
 631                 if (ic->i_recv_hdrs)
 632                         ib_dma_free_coherent(dev,
 633                                            ic->i_recv_ring.w_nr *
 634                                                 sizeof(struct rds_header),
 635                                            ic->i_recv_hdrs,
 636                                            ic->i_recv_hdrs_dma);
 637
 638                 if (ic->i_ack)
 639                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 640                                              ic->i_ack, ic->i_ack_dma);
 641
 642                 if (ic->i_sends)
 643                         rds_ib_send_clear_ring(ic);
 644                 if (ic->i_recvs)
 645                         rds_ib_recv_clear_ring(ic);
 646
 647                 if (ic->i_cm_id->qp)
 648                         rdma_destroy_qp(ic->i_cm_id);
 649                 if (ic->i_send_cq)
 650                         ib_destroy_cq(ic->i_send_cq);
 651                 if (ic->i_recv_cq)
 652                         ib_destroy_cq(ic->i_recv_cq);
 653                 rdma_destroy_id(ic->i_cm_id);
 654
 655                 /*
 656                  * Move connection back to the nodev list.
 657                  */
 658                 if (ic->rds_ibdev)
 659                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 660
 661                 ic->i_cm_id = NULL;
 662                 ic->i_pd = NULL;
 663                 ic->i_mr = NULL;
 664                 ic->i_send_cq = NULL;
 665                 ic->i_recv_cq = NULL;
 666                 ic->i_send_hdrs = NULL;
 667                 ic->i_recv_hdrs = NULL;
 668                 ic->i_ack = NULL;
 669         }
 670         BUG_ON(ic->rds_ibdev);
 671
 672         /* Clear pending transmit */
 673         if (ic->i_data_op) {
 674                 struct rds_message *rm;
 675
 676                 rm = container_of(ic->i_data_op, struct rds_message, data);
 677                 rds_message_put(rm);
 678                 ic->i_data_op = NULL;
 679         }
 680
 681         /* Clear the ACK state */
 682         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 683 #ifdef KERNEL_HAS_ATOMIC64
 684         atomic64_set(&ic->i_ack_next, 0);
 685 #else
 686         ic->i_ack_next = 0;
 687 #endif
 688         ic->i_ack_recv = 0;
 689
 690         /* Clear flow control state */
 691         ic->i_flowctl = 0;
 692         atomic_set(&ic->i_credits, 0);
 693
 694         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 695         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 696
 697         if (ic->i_ibinc) {
 698                 rds_inc_put(&ic->i_ibinc->ii_inc);
 699                 ic->i_ibinc = NULL;
 700         }
 701
 702         vfree(ic->i_sends);
 703         ic->i_sends = NULL;
 704         vfree(ic->i_recvs);
 705         ic->i_recvs = NULL;
 706 }
 707
 708 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 709 {
 710         struct rds_ib_connection *ic;
 711         unsigned long flags;
 712
 713         /* XXX too lazy? */
 714         ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
 715         if (!ic)
 716                 return -ENOMEM;
 717
 718         INIT_LIST_HEAD(&ic->ib_node);
 719         tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
 720                      (unsigned long) ic);
 721         mutex_init(&ic->i_recv_mutex);
 722 #ifndef KERNEL_HAS_ATOMIC64
 723         spin_lock_init(&ic->i_ack_lock);
 724 #endif
 725
 726         /*
 727          * rds_ib_conn_shutdown() waits for these to be emptied so they
 728          * must be initialized before it can be called.
 729          */
 730         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 731         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 732
 733         ic->conn = conn;
 734         conn->c_transport_data = ic;
 735
 736         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 737         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 738         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 739
 740
 741         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 742         return 0;
 743 }
 744
 745 /*
 746  * Free a connection. Connection must be shut down and not set for reconnect.
 747  */
 748 void rds_ib_conn_free(void *arg)
 749 {
 750         struct rds_ib_connection *ic = arg;
 751         spinlock_t      *lock_ptr;
 752
 753         rdsdebug("ic %p\n", ic);
 754
 755         /*
 756          * Conn is either on a dev's list or on the nodev list.
 757          * A race with shutdown() or connect() would cause problems
 758          * (since rds_ibdev would change) but that should never happen.
 759          */
 760         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 761
 762         spin_lock_irq(lock_ptr);
 763         list_del(&ic->ib_node);
 764         spin_unlock_irq(lock_ptr);
 765
 766         kfree(ic);
 767 }
 768
 769
 770 /*
 771  * An error occurred on the connection
 772  */
 773 void
 774 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 775 {
 776         va_list ap;
 777
 778         rds_conn_drop(conn);
 779
 780         va_start(ap, fmt);
 781         vprintk(fmt, ap);
 782         va_end(ap);
 783 }