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