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 #include <linux/ratelimit.h>
  38
  39 #include "rds.h"
  40 #include "ib.h"
  41
  42 /*
  43  * Set the selected protocol version
  44  */
  45 static void rds_ib_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_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  54 {
  55         struct rds_ib_connection *ic = conn->c_transport_data;
  56
  57         if (rds_ib_sysctl_flow_control && credits != 0) {
  58                 /* We're doing flow control */
  59                 ic->i_flowctl = 1;
  60                 rds_ib_send_add_credits(conn, credits);
  61         } else {
  62                 ic->i_flowctl = 0;
  63         }
  64 }
  65
  66 /*
  67  * Tune RNR behavior. Without flow control, we use a rather
  68  * low timeout, but not the absolute minimum - this should
  69  * be tunable.
  70  *
  71  * We already set the RNR retry count to 7 (which is the
  72  * smallest infinite number :-) above.
  73  * If flow control is off, we want to change this back to 0
  74  * so that we learn quickly when our credit accounting is
  75  * buggy.
  76  *
  77  * Caller passes in a qp_attr pointer - don't waste stack spacv
  78  * by allocation this twice.
  79  */
  80 static void
  81 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  82 {
  83         int ret;
  84
  85         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  86         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  87         if (ret)
  88                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  89 }
  90
  91 /*
  92  * Connection established.
  93  * We get here for both outgoing and incoming connection.
  94  */
  95 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  96 {
  97         const struct rds_ib_connect_private *dp = NULL;
  98         struct rds_ib_connection *ic = conn->c_transport_data;
  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         if (conn->c_version < RDS_PROTOCOL(3,1)) {
 115                 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
 116                        " no longer supported\n",
 117                        &conn->c_faddr,
 118                        RDS_PROTOCOL_MAJOR(conn->c_version),
 119                        RDS_PROTOCOL_MINOR(conn->c_version));
 120                 rds_conn_destroy(conn);
 121                 return;
 122         } else {
 123                 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
 124                        &conn->c_faddr,
 125                        RDS_PROTOCOL_MAJOR(conn->c_version),
 126                        RDS_PROTOCOL_MINOR(conn->c_version),
 127                        ic->i_flowctl ? ", flow control" : "");
 128         }
 129
 130         /*
 131          * Init rings and fill recv. this needs to wait until protocol negotiation
 132          * is complete, since ring layout is different from 3.0 to 3.1.
 133          */
 134         rds_ib_send_init_ring(ic);
 135         rds_ib_recv_init_ring(ic);
 136         /* Post receive buffers - as a side effect, this will update
 137          * the posted credit count. */
 138         rds_ib_recv_refill(conn, 1);
 139
 140         /* Tune RNR behavior */
 141         rds_ib_tune_rnr(ic, &qp_attr);
 142
 143         qp_attr.qp_state = IB_QPS_RTS;
 144         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 145         if (err)
 146                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 147
 148         /* update ib_device with this local ipaddr */
 149         err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
 150         if (err)
 151                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 152                         err);
 153
 154         /* If the peer gave us the last packet it saw, process this as if
 155          * we had received a regular ACK. */
 156         if (dp) {
 157                 /* dp structure start is not guaranteed to be 8 bytes aligned.
 158                  * Since dp_ack_seq is 64-bit extended load operations can be
 159                  * used so go through get_unaligned to avoid unaligned errors.
 160                  */
 161                 __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
 162
 163                 if (dp_ack_seq)
 164                         rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
 165                                             NULL);
 166         }
 167
 168         rds_connect_complete(conn);
 169 }
 170
 171 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 172                         struct rdma_conn_param *conn_param,
 173                         struct rds_ib_connect_private *dp,
 174                         u32 protocol_version,
 175                         u32 max_responder_resources,
 176                         u32 max_initiator_depth)
 177 {
 178         struct rds_ib_connection *ic = conn->c_transport_data;
 179         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 180
 181         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 182
 183         conn_param->responder_resources =
 184                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 185         conn_param->initiator_depth =
 186                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 187         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 188         conn_param->rnr_retry_count = 7;
 189
 190         if (dp) {
 191                 memset(dp, 0, sizeof(*dp));
 192                 dp->dp_saddr = conn->c_laddr;
 193                 dp->dp_daddr = conn->c_faddr;
 194                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 195                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 196                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 197                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
 198
 199                 /* Advertise flow control */
 200                 if (ic->i_flowctl) {
 201                         unsigned int credits;
 202
 203                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 204                         dp->dp_credit = cpu_to_be32(credits);
 205                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 206                 }
 207
 208                 conn_param->private_data = dp;
 209                 conn_param->private_data_len = sizeof(*dp);
 210         }
 211 }
 212
 213 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 214 {
 215         rdsdebug("event %u (%s) data %p\n",
 216                  event->event, ib_event_msg(event->event), data);
 217 }
 218
 219 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 220 {
 221         struct rds_connection *conn = data;
 222         struct rds_ib_connection *ic = conn->c_transport_data;
 223
 224         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 225                  ib_event_msg(event->event));
 226
 227         switch (event->event) {
 228         case IB_EVENT_COMM_EST:
 229                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 230                 break;
 231         default:
 232                 rdsdebug("Fatal QP Event %u (%s) "
 233                         "- connection %pI4->%pI4, reconnecting\n",
 234                         event->event, ib_event_msg(event->event),
 235                         &conn->c_laddr, &conn->c_faddr);
 236                 rds_conn_drop(conn);
 237                 break;
 238         }
 239 }
 240
 241 /*
 242  * This needs to be very careful to not leave IS_ERR pointers around for
 243  * cleanup to trip over.
 244  */
 245 static int rds_ib_setup_qp(struct rds_connection *conn)
 246 {
 247         struct rds_ib_connection *ic = conn->c_transport_data;
 248         struct ib_device *dev = ic->i_cm_id->device;
 249         struct ib_qp_init_attr attr;
 250         struct rds_ib_device *rds_ibdev;
 251         int ret;
 252
 253         /*
 254          * It's normal to see a null device if an incoming connection races
 255          * with device removal, so we don't print a warning.
 256          */
 257         rds_ibdev = rds_ib_get_client_data(dev);
 258         if (!rds_ibdev)
 259                 return -EOPNOTSUPP;
 260
 261         /* add the conn now so that connection establishment has the dev */
 262         rds_ib_add_conn(rds_ibdev, conn);
 263
 264         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 265                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 266         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 267                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 268
 269         /* Protection domain and memory range */
 270         ic->i_pd = rds_ibdev->pd;
 271         ic->i_mr = rds_ibdev->mr;
 272
 273         ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
 274                                      rds_ib_cq_event_handler, conn,
 275                                      ic->i_send_ring.w_nr + 1, 0);
 276         if (IS_ERR(ic->i_send_cq)) {
 277                 ret = PTR_ERR(ic->i_send_cq);
 278                 ic->i_send_cq = NULL;
 279                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 280                 goto out;
 281         }
 282
 283         ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
 284                                      rds_ib_cq_event_handler, conn,
 285                                      ic->i_recv_ring.w_nr, 0);
 286         if (IS_ERR(ic->i_recv_cq)) {
 287                 ret = PTR_ERR(ic->i_recv_cq);
 288                 ic->i_recv_cq = NULL;
 289                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 290                 goto out;
 291         }
 292
 293         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 294         if (ret) {
 295                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 296                 goto out;
 297         }
 298
 299         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 300         if (ret) {
 301                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 302                 goto out;
 303         }
 304
 305         /* XXX negotiate max send/recv with remote? */
 306         memset(&attr, 0, sizeof(attr));
 307         attr.event_handler = rds_ib_qp_event_handler;
 308         attr.qp_context = conn;
 309         /* + 1 to allow for the single ack message */
 310         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
 311         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 312         attr.cap.max_send_sge = rds_ibdev->max_sge;
 313         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 314         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 315         attr.qp_type = IB_QPT_RC;
 316         attr.send_cq = ic->i_send_cq;
 317         attr.recv_cq = ic->i_recv_cq;
 318
 319         /*
 320          * XXX this can fail if max_*_wr is too large?  Are we supposed
 321          * to back off until we get a value that the hardware can support?
 322          */
 323         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 324         if (ret) {
 325                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 326                 goto out;
 327         }
 328
 329         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 330                                            ic->i_send_ring.w_nr *
 331                                                 sizeof(struct rds_header),
 332                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 333         if (!ic->i_send_hdrs) {
 334                 ret = -ENOMEM;
 335                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 336                 goto out;
 337         }
 338
 339         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 340                                            ic->i_recv_ring.w_nr *
 341                                                 sizeof(struct rds_header),
 342                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 343         if (!ic->i_recv_hdrs) {
 344                 ret = -ENOMEM;
 345                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 346                 goto out;
 347         }
 348
 349         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 350                                        &ic->i_ack_dma, GFP_KERNEL);
 351         if (!ic->i_ack) {
 352                 ret = -ENOMEM;
 353                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 354                 goto out;
 355         }
 356
 357         ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 358                                    ibdev_to_node(dev));
 359         if (!ic->i_sends) {
 360                 ret = -ENOMEM;
 361                 rdsdebug("send allocation failed\n");
 362                 goto out;
 363         }
 364
 365         ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 366                                    ibdev_to_node(dev));
 367         if (!ic->i_recvs) {
 368                 ret = -ENOMEM;
 369                 rdsdebug("recv allocation failed\n");
 370                 goto out;
 371         }
 372
 373         rds_ib_recv_init_ack(ic);
 374
 375         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 376                  ic->i_send_cq, ic->i_recv_cq);
 377
 378 out:
 379         rds_ib_dev_put(rds_ibdev);
 380         return ret;
 381 }
 382
 383 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 384 {
 385         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 386         u16 common;
 387         u32 version = 0;
 388
 389         /*
 390          * rdma_cm private data is odd - when there is any private data in the
 391          * request, we will be given a pretty large buffer without telling us the
 392          * original size. The only way to tell the difference is by looking at
 393          * the contents, which are initialized to zero.
 394          * If the protocol version fields aren't set, this is a connection attempt
 395          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 396          * We really should have changed this for OFED 1.3 :-(
 397          */
 398
 399         /* Be paranoid. RDS always has privdata */
 400         if (!event->param.conn.private_data_len) {
 401                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 402                         "rejecting\n");
 403                 return 0;
 404         }
 405
 406         /* Even if len is crap *now* I still want to check it. -ASG */
 407         if (event->param.conn.private_data_len < sizeof (*dp) ||
 408             dp->dp_protocol_major == 0)
 409                 return RDS_PROTOCOL_3_0;
 410
 411         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 412         if (dp->dp_protocol_major == 3 && common) {
 413                 version = RDS_PROTOCOL_3_0;
 414                 while ((common >>= 1) != 0)
 415                         version++;
 416         } else
 417                 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 418                                 &dp->dp_saddr,
 419                                 dp->dp_protocol_major,
 420                                 dp->dp_protocol_minor);
 421         return version;
 422 }
 423
 424 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 425                                     struct rdma_cm_event *event)
 426 {
 427         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 428         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 429         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 430         struct rds_ib_connect_private dp_rep;
 431         struct rds_connection *conn = NULL;
 432         struct rds_ib_connection *ic = NULL;
 433         struct rdma_conn_param conn_param;
 434         u32 version;
 435         int err = 1, destroy = 1;
 436
 437         /* Check whether the remote protocol version matches ours. */
 438         version = rds_ib_protocol_compatible(event);
 439         if (!version)
 440                 goto out;
 441
 442         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 443                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 444                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 445                  (unsigned long long)be64_to_cpu(lguid),
 446                  (unsigned long long)be64_to_cpu(fguid));
 447
 448         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
 449                                GFP_KERNEL);
 450         if (IS_ERR(conn)) {
 451                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 452                 conn = NULL;
 453                 goto out;
 454         }
 455
 456         /*
 457          * The connection request may occur while the
 458          * previous connection exist, e.g. in case of failover.
 459          * But as connections may be initiated simultaneously
 460          * by both hosts, we have a random backoff mechanism -
 461          * see the comment above rds_queue_reconnect()
 462          */
 463         mutex_lock(&conn->c_cm_lock);
 464         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 465                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 466                         rdsdebug("incoming connect while connecting\n");
 467                         rds_conn_drop(conn);
 468                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 469                 } else
 470                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 471                         /* Wait and see - our connect may still be succeeding */
 472                         rds_ib_stats_inc(s_ib_connect_raced);
 473                 }
 474                 goto out;
 475         }
 476
 477         ic = conn->c_transport_data;
 478
 479         rds_ib_set_protocol(conn, version);
 480         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 481
 482         /* If the peer gave us the last packet it saw, process this as if
 483          * we had received a regular ACK. */
 484         if (dp->dp_ack_seq)
 485                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 486
 487         BUG_ON(cm_id->context);
 488         BUG_ON(ic->i_cm_id);
 489
 490         ic->i_cm_id = cm_id;
 491         cm_id->context = conn;
 492
 493         /* We got halfway through setting up the ib_connection, if we
 494          * fail now, we have to take the long route out of this mess. */
 495         destroy = 0;
 496
 497         err = rds_ib_setup_qp(conn);
 498         if (err) {
 499                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 500                 goto out;
 501         }
 502
 503         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 504                 event->param.conn.responder_resources,
 505                 event->param.conn.initiator_depth);
 506
 507         /* rdma_accept() calls rdma_reject() internally if it fails */
 508         err = rdma_accept(cm_id, &conn_param);
 509         if (err)
 510                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 511
 512 out:
 513         if (conn)
 514                 mutex_unlock(&conn->c_cm_lock);
 515         if (err)
 516                 rdma_reject(cm_id, NULL, 0);
 517         return destroy;
 518 }
 519
 520
 521 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 522 {
 523         struct rds_connection *conn = cm_id->context;
 524         struct rds_ib_connection *ic = conn->c_transport_data;
 525         struct rdma_conn_param conn_param;
 526         struct rds_ib_connect_private dp;
 527         int ret;
 528
 529         /* If the peer doesn't do protocol negotiation, we must
 530          * default to RDSv3.0 */
 531         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 532         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 533
 534         ret = rds_ib_setup_qp(conn);
 535         if (ret) {
 536                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 537                 goto out;
 538         }
 539
 540         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 541                 UINT_MAX, UINT_MAX);
 542         ret = rdma_connect(cm_id, &conn_param);
 543         if (ret)
 544                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 545
 546 out:
 547         /* Beware - returning non-zero tells the rdma_cm to destroy
 548          * the cm_id. We should certainly not do it as long as we still
 549          * "own" the cm_id. */
 550         if (ret) {
 551                 if (ic->i_cm_id == cm_id)
 552                         ret = 0;
 553         }
 554         return ret;
 555 }
 556
 557 int rds_ib_conn_connect(struct rds_connection *conn)
 558 {
 559         struct rds_ib_connection *ic = conn->c_transport_data;
 560         struct sockaddr_in src, dest;
 561         int ret;
 562
 563         /* XXX I wonder what affect the port space has */
 564         /* delegate cm event handler to rdma_transport */
 565         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 566                                      RDMA_PS_TCP, IB_QPT_RC);
 567         if (IS_ERR(ic->i_cm_id)) {
 568                 ret = PTR_ERR(ic->i_cm_id);
 569                 ic->i_cm_id = NULL;
 570                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 571                 goto out;
 572         }
 573
 574         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 575
 576         src.sin_family = AF_INET;
 577         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 578         src.sin_port = (__force u16)htons(0);
 579
 580         dest.sin_family = AF_INET;
 581         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 582         dest.sin_port = (__force u16)htons(RDS_PORT);
 583
 584         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 585                                 (struct sockaddr *)&dest,
 586                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 587         if (ret) {
 588                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 589                          ret);
 590                 rdma_destroy_id(ic->i_cm_id);
 591                 ic->i_cm_id = NULL;
 592         }
 593
 594 out:
 595         return ret;
 596 }
 597
 598 /*
 599  * This is so careful about only cleaning up resources that were built up
 600  * so that it can be called at any point during startup.  In fact it
 601  * can be called multiple times for a given connection.
 602  */
 603 void rds_ib_conn_shutdown(struct rds_connection *conn)
 604 {
 605         struct rds_ib_connection *ic = conn->c_transport_data;
 606         int err = 0;
 607
 608         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 609                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 610                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 611
 612         if (ic->i_cm_id) {
 613                 struct ib_device *dev = ic->i_cm_id->device;
 614
 615                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 616                 err = rdma_disconnect(ic->i_cm_id);
 617                 if (err) {
 618                         /* Actually this may happen quite frequently, when
 619                          * an outgoing connect raced with an incoming connect.
 620                          */
 621                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 622                                 ic->i_cm_id, err);
 623                 }
 624
 625                 /*
 626                  * We want to wait for tx and rx completion to finish
 627                  * before we tear down the connection, but we have to be
 628                  * careful not to get stuck waiting on a send ring that
 629                  * only has unsignaled sends in it.  We've shutdown new
 630                  * sends before getting here so by waiting for signaled
 631                  * sends to complete we're ensured that there will be no
 632                  * more tx processing.
 633                  */
 634                 wait_event(rds_ib_ring_empty_wait,
 635                            rds_ib_ring_empty(&ic->i_recv_ring) &&
 636                            (atomic_read(&ic->i_signaled_sends) == 0));
 637                 tasklet_kill(&ic->i_recv_tasklet);
 638
 639                 if (ic->i_send_hdrs)
 640                         ib_dma_free_coherent(dev,
 641                                            ic->i_send_ring.w_nr *
 642                                                 sizeof(struct rds_header),
 643                                            ic->i_send_hdrs,
 644                                            ic->i_send_hdrs_dma);
 645
 646                 if (ic->i_recv_hdrs)
 647                         ib_dma_free_coherent(dev,
 648                                            ic->i_recv_ring.w_nr *
 649                                                 sizeof(struct rds_header),
 650                                            ic->i_recv_hdrs,
 651                                            ic->i_recv_hdrs_dma);
 652
 653                 if (ic->i_ack)
 654                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 655                                              ic->i_ack, ic->i_ack_dma);
 656
 657                 if (ic->i_sends)
 658                         rds_ib_send_clear_ring(ic);
 659                 if (ic->i_recvs)
 660                         rds_ib_recv_clear_ring(ic);
 661
 662                 if (ic->i_cm_id->qp)
 663                         rdma_destroy_qp(ic->i_cm_id);
 664                 if (ic->i_send_cq)
 665                         ib_destroy_cq(ic->i_send_cq);
 666                 if (ic->i_recv_cq)
 667                         ib_destroy_cq(ic->i_recv_cq);
 668                 rdma_destroy_id(ic->i_cm_id);
 669
 670                 /*
 671                  * Move connection back to the nodev list.
 672                  */
 673                 if (ic->rds_ibdev)
 674                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 675
 676                 ic->i_cm_id = NULL;
 677                 ic->i_pd = NULL;
 678                 ic->i_mr = NULL;
 679                 ic->i_send_cq = NULL;
 680                 ic->i_recv_cq = NULL;
 681                 ic->i_send_hdrs = NULL;
 682                 ic->i_recv_hdrs = NULL;
 683                 ic->i_ack = NULL;
 684         }
 685         BUG_ON(ic->rds_ibdev);
 686
 687         /* Clear pending transmit */
 688         if (ic->i_data_op) {
 689                 struct rds_message *rm;
 690
 691                 rm = container_of(ic->i_data_op, struct rds_message, data);
 692                 rds_message_put(rm);
 693                 ic->i_data_op = NULL;
 694         }
 695
 696         /* Clear the ACK state */
 697         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 698 #ifdef KERNEL_HAS_ATOMIC64
 699         atomic64_set(&ic->i_ack_next, 0);
 700 #else
 701         ic->i_ack_next = 0;
 702 #endif
 703         ic->i_ack_recv = 0;
 704
 705         /* Clear flow control state */
 706         ic->i_flowctl = 0;
 707         atomic_set(&ic->i_credits, 0);
 708
 709         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 710         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 711
 712         if (ic->i_ibinc) {
 713                 rds_inc_put(&ic->i_ibinc->ii_inc);
 714                 ic->i_ibinc = NULL;
 715         }
 716
 717         vfree(ic->i_sends);
 718         ic->i_sends = NULL;
 719         vfree(ic->i_recvs);
 720         ic->i_recvs = NULL;
 721 }
 722
 723 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 724 {
 725         struct rds_ib_connection *ic;
 726         unsigned long flags;
 727         int ret;
 728
 729         /* XXX too lazy? */
 730         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
 731         if (!ic)
 732                 return -ENOMEM;
 733
 734         ret = rds_ib_recv_alloc_caches(ic);
 735         if (ret) {
 736                 kfree(ic);
 737                 return ret;
 738         }
 739
 740         INIT_LIST_HEAD(&ic->ib_node);
 741         tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
 742                      (unsigned long) ic);
 743         mutex_init(&ic->i_recv_mutex);
 744 #ifndef KERNEL_HAS_ATOMIC64
 745         spin_lock_init(&ic->i_ack_lock);
 746 #endif
 747         atomic_set(&ic->i_signaled_sends, 0);
 748
 749         /*
 750          * rds_ib_conn_shutdown() waits for these to be emptied so they
 751          * must be initialized before it can be called.
 752          */
 753         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 754         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 755
 756         ic->conn = conn;
 757         conn->c_transport_data = ic;
 758
 759         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 760         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 761         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 762
 763
 764         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 765         return 0;
 766 }
 767
 768 /*
 769  * Free a connection. Connection must be shut down and not set for reconnect.
 770  */
 771 void rds_ib_conn_free(void *arg)
 772 {
 773         struct rds_ib_connection *ic = arg;
 774         spinlock_t      *lock_ptr;
 775
 776         rdsdebug("ic %p\n", ic);
 777
 778         /*
 779          * Conn is either on a dev's list or on the nodev list.
 780          * A race with shutdown() or connect() would cause problems
 781          * (since rds_ibdev would change) but that should never happen.
 782          */
 783         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 784
 785         spin_lock_irq(lock_ptr);
 786         list_del(&ic->ib_node);
 787         spin_unlock_irq(lock_ptr);
 788
 789         rds_ib_recv_free_caches(ic);
 790
 791         kfree(ic);
 792 }
 793
 794
 795 /*
 796  * An error occurred on the connection
 797  */
 798 void
 799 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 800 {
 801         va_list ap;
 802
 803         rds_conn_drop(conn);
 804
 805         va_start(ap, fmt);
 806         vprintk(fmt, ap);
 807         va_end(ap);
 808 }