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/vmalloc.h>
  36
  37 #include "rds.h"
  38 #include "ib.h"
  39
  40 /*
  41  * Set the selected protocol version
  42  */
  43 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  44 {
  45         conn->c_version = version;
  46 }
  47
  48 /*
  49  * Set up flow control
  50  */
  51 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  52 {
  53         struct rds_ib_connection *ic = conn->c_transport_data;
  54
  55         if (rds_ib_sysctl_flow_control && credits != 0) {
  56                 /* We're doing flow control */
  57                 ic->i_flowctl = 1;
  58                 rds_ib_send_add_credits(conn, credits);
  59         } else {
  60                 ic->i_flowctl = 0;
  61         }
  62 }
  63
  64 /*
  65  * Tune RNR behavior. Without flow control, we use a rather
  66  * low timeout, but not the absolute minimum - this should
  67  * be tunable.
  68  *
  69  * We already set the RNR retry count to 7 (which is the
  70  * smallest infinite number :-) above.
  71  * If flow control is off, we want to change this back to 0
  72  * so that we learn quickly when our credit accounting is
  73  * buggy.
  74  *
  75  * Caller passes in a qp_attr pointer - don't waste stack spacv
  76  * by allocation this twice.
  77  */
  78 static void
  79 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  80 {
  81         int ret;
  82
  83         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  84         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  85         if (ret)
  86                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  87 }
  88
  89 /*
  90  * Connection established.
  91  * We get here for both outgoing and incoming connection.
  92  */
  93 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  94 {
  95         const struct rds_ib_connect_private *dp = NULL;
  96         struct rds_ib_connection *ic = conn->c_transport_data;
  97         struct rds_ib_device *rds_ibdev;
  98         struct ib_qp_attr qp_attr;
  99         int err;
 100
 101         if (event->param.conn.private_data_len >= sizeof(*dp)) {
 102                 dp = event->param.conn.private_data;
 103
 104                 /* make sure it isn't empty data */
 105                 if (dp->dp_protocol_major) {
 106                         rds_ib_set_protocol(conn,
 107                                 RDS_PROTOCOL(dp->dp_protocol_major,
 108                                 dp->dp_protocol_minor));
 109                         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 110                 }
 111         }
 112
 113         printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
 114                         &conn->c_faddr,
 115                         RDS_PROTOCOL_MAJOR(conn->c_version),
 116                         RDS_PROTOCOL_MINOR(conn->c_version),
 117                         ic->i_flowctl ? ", flow control" : "");
 118
 119         /*
 120          * Init rings and fill recv. this needs to wait until protocol negotiation
 121          * is complete, since ring layout is different from 3.0 to 3.1.
 122          */
 123         rds_ib_send_init_ring(ic);
 124         rds_ib_recv_init_ring(ic);
 125         /* Post receive buffers - as a side effect, this will update
 126          * the posted credit count. */
 127         rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
 128
 129         /* Tune RNR behavior */
 130         rds_ib_tune_rnr(ic, &qp_attr);
 131
 132         qp_attr.qp_state = IB_QPS_RTS;
 133         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 134         if (err)
 135                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 136
 137         /* update ib_device with this local ipaddr & conn */
 138         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 139         err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
 140         if (err)
 141                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
 142         rds_ib_add_conn(rds_ibdev, conn);
 143
 144         /* If the peer gave us the last packet it saw, process this as if
 145          * we had received a regular ACK. */
 146         if (dp && dp->dp_ack_seq)
 147                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 148
 149         rds_connect_complete(conn);
 150 }
 151
 152 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 153                         struct rdma_conn_param *conn_param,
 154                         struct rds_ib_connect_private *dp,
 155                         u32 protocol_version)
 156 {
 157         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 158         /* XXX tune these? */
 159         conn_param->responder_resources = 1;
 160         conn_param->initiator_depth = 1;
 161         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 162         conn_param->rnr_retry_count = 7;
 163
 164         if (dp) {
 165                 struct rds_ib_connection *ic = conn->c_transport_data;
 166
 167                 memset(dp, 0, sizeof(*dp));
 168                 dp->dp_saddr = conn->c_laddr;
 169                 dp->dp_daddr = conn->c_faddr;
 170                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 171                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 172                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 173                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
 174
 175                 /* Advertise flow control */
 176                 if (ic->i_flowctl) {
 177                         unsigned int credits;
 178
 179                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 180                         dp->dp_credit = cpu_to_be32(credits);
 181                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 182                 }
 183
 184                 conn_param->private_data = dp;
 185                 conn_param->private_data_len = sizeof(*dp);
 186         }
 187 }
 188
 189 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 190 {
 191         rdsdebug("event %u data %p\n", event->event, data);
 192 }
 193
 194 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 195 {
 196         struct rds_connection *conn = data;
 197         struct rds_ib_connection *ic = conn->c_transport_data;
 198
 199         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
 200
 201         switch (event->event) {
 202         case IB_EVENT_COMM_EST:
 203                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 204                 break;
 205         default:
 206                 rdsdebug("Fatal QP Event %u "
 207                         "- connection %pI4->%pI4, reconnecting\n",
 208                         event->event, &conn->c_laddr, &conn->c_faddr);
 209                 rds_conn_drop(conn);
 210                 break;
 211         }
 212 }
 213
 214 /*
 215  * This needs to be very careful to not leave IS_ERR pointers around for
 216  * cleanup to trip over.
 217  */
 218 static int rds_ib_setup_qp(struct rds_connection *conn)
 219 {
 220         struct rds_ib_connection *ic = conn->c_transport_data;
 221         struct ib_device *dev = ic->i_cm_id->device;
 222         struct ib_qp_init_attr attr;
 223         struct rds_ib_device *rds_ibdev;
 224         int ret;
 225
 226         /* rds_ib_add_one creates a rds_ib_device object per IB device,
 227          * and allocates a protection domain, memory range and FMR pool
 228          * for each.  If that fails for any reason, it will not register
 229          * the rds_ibdev at all.
 230          */
 231         rds_ibdev = ib_get_client_data(dev, &rds_ib_client);
 232         if (rds_ibdev == NULL) {
 233                 if (printk_ratelimit())
 234                         printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
 235                                         dev->name);
 236                 return -EOPNOTSUPP;
 237         }
 238
 239         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 240                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 241         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 242                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 243
 244         /* Protection domain and memory range */
 245         ic->i_pd = rds_ibdev->pd;
 246         ic->i_mr = rds_ibdev->mr;
 247
 248         ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
 249                                      rds_ib_cq_event_handler, conn,
 250                                      ic->i_send_ring.w_nr + 1, 0);
 251         if (IS_ERR(ic->i_send_cq)) {
 252                 ret = PTR_ERR(ic->i_send_cq);
 253                 ic->i_send_cq = NULL;
 254                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 255                 goto out;
 256         }
 257
 258         ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
 259                                      rds_ib_cq_event_handler, conn,
 260                                      ic->i_recv_ring.w_nr, 0);
 261         if (IS_ERR(ic->i_recv_cq)) {
 262                 ret = PTR_ERR(ic->i_recv_cq);
 263                 ic->i_recv_cq = NULL;
 264                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 265                 goto out;
 266         }
 267
 268         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 269         if (ret) {
 270                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 271                 goto out;
 272         }
 273
 274         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 275         if (ret) {
 276                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 277                 goto out;
 278         }
 279
 280         /* XXX negotiate max send/recv with remote? */
 281         memset(&attr, 0, sizeof(attr));
 282         attr.event_handler = rds_ib_qp_event_handler;
 283         attr.qp_context = conn;
 284         /* + 1 to allow for the single ack message */
 285         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
 286         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 287         attr.cap.max_send_sge = rds_ibdev->max_sge;
 288         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 289         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 290         attr.qp_type = IB_QPT_RC;
 291         attr.send_cq = ic->i_send_cq;
 292         attr.recv_cq = ic->i_recv_cq;
 293
 294         /*
 295          * XXX this can fail if max_*_wr is too large?  Are we supposed
 296          * to back off until we get a value that the hardware can support?
 297          */
 298         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 299         if (ret) {
 300                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 301                 goto out;
 302         }
 303
 304         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 305                                            ic->i_send_ring.w_nr *
 306                                                 sizeof(struct rds_header),
 307                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 308         if (ic->i_send_hdrs == NULL) {
 309                 ret = -ENOMEM;
 310                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 311                 goto out;
 312         }
 313
 314         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 315                                            ic->i_recv_ring.w_nr *
 316                                                 sizeof(struct rds_header),
 317                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 318         if (ic->i_recv_hdrs == NULL) {
 319                 ret = -ENOMEM;
 320                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 321                 goto out;
 322         }
 323
 324         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 325                                        &ic->i_ack_dma, GFP_KERNEL);
 326         if (ic->i_ack == NULL) {
 327                 ret = -ENOMEM;
 328                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 329                 goto out;
 330         }
 331
 332         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
 333         if (ic->i_sends == NULL) {
 334                 ret = -ENOMEM;
 335                 rdsdebug("send allocation failed\n");
 336                 goto out;
 337         }
 338         memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
 339
 340         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
 341         if (ic->i_recvs == NULL) {
 342                 ret = -ENOMEM;
 343                 rdsdebug("recv allocation failed\n");
 344                 goto out;
 345         }
 346         memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
 347
 348         rds_ib_recv_init_ack(ic);
 349
 350         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 351                  ic->i_send_cq, ic->i_recv_cq);
 352
 353 out:
 354         return ret;
 355 }
 356
 357 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 358 {
 359         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 360         u16 common;
 361         u32 version = 0;
 362
 363         /*
 364          * rdma_cm private data is odd - when there is any private data in the
 365          * request, we will be given a pretty large buffer without telling us the
 366          * original size. The only way to tell the difference is by looking at
 367          * the contents, which are initialized to zero.
 368          * If the protocol version fields aren't set, this is a connection attempt
 369          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 370          * We really should have changed this for OFED 1.3 :-(
 371          */
 372
 373         /* Be paranoid. RDS always has privdata */
 374         if (!event->param.conn.private_data_len) {
 375                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 376                         "rejecting\n");
 377                 return 0;
 378         }
 379
 380         /* Even if len is crap *now* I still want to check it. -ASG */
 381         if (event->param.conn.private_data_len < sizeof (*dp) ||
 382             dp->dp_protocol_major == 0)
 383                 return RDS_PROTOCOL_3_0;
 384
 385         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 386         if (dp->dp_protocol_major == 3 && common) {
 387                 version = RDS_PROTOCOL_3_0;
 388                 while ((common >>= 1) != 0)
 389                         version++;
 390         } else if (printk_ratelimit()) {
 391                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
 392                         "incompatible protocol version %u.%u\n",
 393                         &dp->dp_saddr,
 394                         dp->dp_protocol_major,
 395                         dp->dp_protocol_minor);
 396         }
 397         return version;
 398 }
 399
 400 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 401                                     struct rdma_cm_event *event)
 402 {
 403         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 404         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 405         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 406         struct rds_ib_connect_private dp_rep;
 407         struct rds_connection *conn = NULL;
 408         struct rds_ib_connection *ic = NULL;
 409         struct rdma_conn_param conn_param;
 410         u32 version;
 411         int err, destroy = 1;
 412
 413         /* Check whether the remote protocol version matches ours. */
 414         version = rds_ib_protocol_compatible(event);
 415         if (!version)
 416                 goto out;
 417
 418         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 419                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 420                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 421                  (unsigned long long)be64_to_cpu(lguid),
 422                  (unsigned long long)be64_to_cpu(fguid));
 423
 424         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
 425                                GFP_KERNEL);
 426         if (IS_ERR(conn)) {
 427                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 428                 conn = NULL;
 429                 goto out;
 430         }
 431
 432         /*
 433          * The connection request may occur while the
 434          * previous connection exist, e.g. in case of failover.
 435          * But as connections may be initiated simultaneously
 436          * by both hosts, we have a random backoff mechanism -
 437          * see the comment above rds_queue_reconnect()
 438          */
 439         mutex_lock(&conn->c_cm_lock);
 440         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 441                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 442                         rdsdebug("incoming connect while connecting\n");
 443                         rds_conn_drop(conn);
 444                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 445                 } else
 446                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 447                         /* Wait and see - our connect may still be succeeding */
 448                         rds_ib_stats_inc(s_ib_connect_raced);
 449                 }
 450                 mutex_unlock(&conn->c_cm_lock);
 451                 goto out;
 452         }
 453
 454         ic = conn->c_transport_data;
 455
 456         rds_ib_set_protocol(conn, version);
 457         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 458
 459         /* If the peer gave us the last packet it saw, process this as if
 460          * we had received a regular ACK. */
 461         if (dp->dp_ack_seq)
 462                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 463
 464         BUG_ON(cm_id->context);
 465         BUG_ON(ic->i_cm_id);
 466
 467         ic->i_cm_id = cm_id;
 468         cm_id->context = conn;
 469
 470         /* We got halfway through setting up the ib_connection, if we
 471          * fail now, we have to take the long route out of this mess. */
 472         destroy = 0;
 473
 474         err = rds_ib_setup_qp(conn);
 475         if (err) {
 476                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 477                 goto out;
 478         }
 479
 480         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
 481
 482         /* rdma_accept() calls rdma_reject() internally if it fails */
 483         err = rdma_accept(cm_id, &conn_param);
 484         mutex_unlock(&conn->c_cm_lock);
 485         if (err) {
 486                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 487                 goto out;
 488         }
 489
 490         return 0;
 491
 492 out:
 493         rdma_reject(cm_id, NULL, 0);
 494         return destroy;
 495 }
 496
 497
 498 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 499 {
 500         struct rds_connection *conn = cm_id->context;
 501         struct rds_ib_connection *ic = conn->c_transport_data;
 502         struct rdma_conn_param conn_param;
 503         struct rds_ib_connect_private dp;
 504         int ret;
 505
 506         /* If the peer doesn't do protocol negotiation, we must
 507          * default to RDSv3.0 */
 508         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 509         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 510
 511         ret = rds_ib_setup_qp(conn);
 512         if (ret) {
 513                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 514                 goto out;
 515         }
 516
 517         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
 518
 519         ret = rdma_connect(cm_id, &conn_param);
 520         if (ret)
 521                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 522
 523 out:
 524         /* Beware - returning non-zero tells the rdma_cm to destroy
 525          * the cm_id. We should certainly not do it as long as we still
 526          * "own" the cm_id. */
 527         if (ret) {
 528                 if (ic->i_cm_id == cm_id)
 529                         ret = 0;
 530         }
 531         return ret;
 532 }
 533
 534 int rds_ib_conn_connect(struct rds_connection *conn)
 535 {
 536         struct rds_ib_connection *ic = conn->c_transport_data;
 537         struct sockaddr_in src, dest;
 538         int ret;
 539
 540         /* XXX I wonder what affect the port space has */
 541         /* delegate cm event handler to rdma_transport */
 542         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 543                                      RDMA_PS_TCP);
 544         if (IS_ERR(ic->i_cm_id)) {
 545                 ret = PTR_ERR(ic->i_cm_id);
 546                 ic->i_cm_id = NULL;
 547                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 548                 goto out;
 549         }
 550
 551         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 552
 553         src.sin_family = AF_INET;
 554         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 555         src.sin_port = (__force u16)htons(0);
 556
 557         dest.sin_family = AF_INET;
 558         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 559         dest.sin_port = (__force u16)htons(RDS_PORT);
 560
 561         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 562                                 (struct sockaddr *)&dest,
 563                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 564         if (ret) {
 565                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 566                          ret);
 567                 rdma_destroy_id(ic->i_cm_id);
 568                 ic->i_cm_id = NULL;
 569         }
 570
 571 out:
 572         return ret;
 573 }
 574
 575 /*
 576  * This is so careful about only cleaning up resources that were built up
 577  * so that it can be called at any point during startup.  In fact it
 578  * can be called multiple times for a given connection.
 579  */
 580 void rds_ib_conn_shutdown(struct rds_connection *conn)
 581 {
 582         struct rds_ib_connection *ic = conn->c_transport_data;
 583         int err = 0;
 584
 585         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 586                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 587                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 588
 589         if (ic->i_cm_id) {
 590                 struct ib_device *dev = ic->i_cm_id->device;
 591
 592                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 593                 err = rdma_disconnect(ic->i_cm_id);
 594                 if (err) {
 595                         /* Actually this may happen quite frequently, when
 596                          * an outgoing connect raced with an incoming connect.
 597                          */
 598                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 599                                 ic->i_cm_id, err);
 600                 }
 601
 602                 wait_event(rds_ib_ring_empty_wait,
 603                         rds_ib_ring_empty(&ic->i_send_ring) &&
 604                         rds_ib_ring_empty(&ic->i_recv_ring));
 605
 606                 if (ic->i_send_hdrs)
 607                         ib_dma_free_coherent(dev,
 608                                            ic->i_send_ring.w_nr *
 609                                                 sizeof(struct rds_header),
 610                                            ic->i_send_hdrs,
 611                                            ic->i_send_hdrs_dma);
 612
 613                 if (ic->i_recv_hdrs)
 614                         ib_dma_free_coherent(dev,
 615                                            ic->i_recv_ring.w_nr *
 616                                                 sizeof(struct rds_header),
 617                                            ic->i_recv_hdrs,
 618                                            ic->i_recv_hdrs_dma);
 619
 620                 if (ic->i_ack)
 621                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 622                                              ic->i_ack, ic->i_ack_dma);
 623
 624                 if (ic->i_sends)
 625                         rds_ib_send_clear_ring(ic);
 626                 if (ic->i_recvs)
 627                         rds_ib_recv_clear_ring(ic);
 628
 629                 if (ic->i_cm_id->qp)
 630                         rdma_destroy_qp(ic->i_cm_id);
 631                 if (ic->i_send_cq)
 632                         ib_destroy_cq(ic->i_send_cq);
 633                 if (ic->i_recv_cq)
 634                         ib_destroy_cq(ic->i_recv_cq);
 635                 rdma_destroy_id(ic->i_cm_id);
 636
 637                 /*
 638                  * Move connection back to the nodev list.
 639                  */
 640                 if (ic->rds_ibdev)
 641                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 642
 643                 ic->i_cm_id = NULL;
 644                 ic->i_pd = NULL;
 645                 ic->i_mr = NULL;
 646                 ic->i_send_cq = NULL;
 647                 ic->i_recv_cq = NULL;
 648                 ic->i_send_hdrs = NULL;
 649                 ic->i_recv_hdrs = NULL;
 650                 ic->i_ack = NULL;
 651         }
 652         BUG_ON(ic->rds_ibdev);
 653
 654         /* Clear pending transmit */
 655         if (ic->i_rm) {
 656                 rds_message_put(ic->i_rm);
 657                 ic->i_rm = NULL;
 658         }
 659
 660         /* Clear the ACK state */
 661         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 662 #ifdef KERNEL_HAS_ATOMIC64
 663         atomic64_set(&ic->i_ack_next, 0);
 664 #else
 665         ic->i_ack_next = 0;
 666 #endif
 667         ic->i_ack_recv = 0;
 668
 669         /* Clear flow control state */
 670         ic->i_flowctl = 0;
 671         atomic_set(&ic->i_credits, 0);
 672
 673         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 674         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 675
 676         if (ic->i_ibinc) {
 677                 rds_inc_put(&ic->i_ibinc->ii_inc);
 678                 ic->i_ibinc = NULL;
 679         }
 680
 681         vfree(ic->i_sends);
 682         ic->i_sends = NULL;
 683         vfree(ic->i_recvs);
 684         ic->i_recvs = NULL;
 685 }
 686
 687 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 688 {
 689         struct rds_ib_connection *ic;
 690         unsigned long flags;
 691
 692         /* XXX too lazy? */
 693         ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
 694         if (ic == NULL)
 695                 return -ENOMEM;
 696
 697         INIT_LIST_HEAD(&ic->ib_node);
 698         tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
 699                      (unsigned long) ic);
 700         mutex_init(&ic->i_recv_mutex);
 701 #ifndef KERNEL_HAS_ATOMIC64
 702         spin_lock_init(&ic->i_ack_lock);
 703 #endif
 704
 705         /*
 706          * rds_ib_conn_shutdown() waits for these to be emptied so they
 707          * must be initialized before it can be called.
 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         ic->conn = conn;
 713         conn->c_transport_data = ic;
 714
 715         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 716         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 717         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 718
 719
 720         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 721         return 0;
 722 }
 723
 724 /*
 725  * Free a connection. Connection must be shut down and not set for reconnect.
 726  */
 727 void rds_ib_conn_free(void *arg)
 728 {
 729         struct rds_ib_connection *ic = arg;
 730         spinlock_t      *lock_ptr;
 731
 732         rdsdebug("ic %p\n", ic);
 733
 734         /*
 735          * Conn is either on a dev's list or on the nodev list.
 736          * A race with shutdown() or connect() would cause problems
 737          * (since rds_ibdev would change) but that should never happen.
 738          */
 739         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 740
 741         spin_lock_irq(lock_ptr);
 742         list_del(&ic->ib_node);
 743         spin_unlock_irq(lock_ptr);
 744
 745         kfree(ic);
 746 }
 747
 748
 749 /*
 750  * An error occurred on the connection
 751  */
 752 void
 753 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 754 {
 755         va_list ap;
 756
 757         rds_conn_drop(conn);
 758
 759         va_start(ap, fmt);
 760         vprintk(fmt, ap);
 761         va_end(ap);
 762 }