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