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         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 & conn */
 149         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 150         err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
 151         if (err)
 152                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
 153         rds_ib_add_conn(rds_ibdev, conn);
 154
 155         /* If the peer gave us the last packet it saw, process this as if
 156          * we had received a regular ACK. */
 157         if (dp && dp->dp_ack_seq)
 158                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 159
 160         rds_connect_complete(conn);
 161 }
 162
 163 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 164                         struct rdma_conn_param *conn_param,
 165                         struct rds_ib_connect_private *dp,
 166                         u32 protocol_version,
 167                         u32 max_responder_resources,
 168                         u32 max_initiator_depth)
 169 {
 170         struct rds_ib_connection *ic = conn->c_transport_data;
 171         struct rds_ib_device *rds_ibdev;
 172
 173         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 174
 175         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 176
 177         conn_param->responder_resources =
 178                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 179         conn_param->initiator_depth =
 180                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 181         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 182         conn_param->rnr_retry_count = 7;
 183
 184         if (dp) {
 185                 memset(dp, 0, sizeof(*dp));
 186                 dp->dp_saddr = conn->c_laddr;
 187                 dp->dp_daddr = conn->c_faddr;
 188                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 189                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 190                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 191                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
 192
 193                 /* Advertise flow control */
 194                 if (ic->i_flowctl) {
 195                         unsigned int credits;
 196
 197                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 198                         dp->dp_credit = cpu_to_be32(credits);
 199                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 200                 }
 201
 202                 conn_param->private_data = dp;
 203                 conn_param->private_data_len = sizeof(*dp);
 204         }
 205 }
 206
 207 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 208 {
 209         rdsdebug("event %u data %p\n", event->event, data);
 210 }
 211
 212 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 213 {
 214         struct rds_connection *conn = data;
 215         struct rds_ib_connection *ic = conn->c_transport_data;
 216
 217         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
 218
 219         switch (event->event) {
 220         case IB_EVENT_COMM_EST:
 221                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 222                 break;
 223         default:
 224                 rdsdebug("Fatal QP Event %u "
 225                         "- connection %pI4->%pI4, reconnecting\n",
 226                         event->event, &conn->c_laddr, &conn->c_faddr);
 227                 rds_conn_drop(conn);
 228                 break;
 229         }
 230 }
 231
 232 /*
 233  * This needs to be very careful to not leave IS_ERR pointers around for
 234  * cleanup to trip over.
 235  */
 236 static int rds_ib_setup_qp(struct rds_connection *conn)
 237 {
 238         struct rds_ib_connection *ic = conn->c_transport_data;
 239         struct ib_device *dev = ic->i_cm_id->device;
 240         struct ib_qp_init_attr attr;
 241         struct rds_ib_device *rds_ibdev;
 242         int ret;
 243
 244         /* rds_ib_add_one creates a rds_ib_device object per IB device,
 245          * and allocates a protection domain, memory range and FMR pool
 246          * for each.  If that fails for any reason, it will not register
 247          * the rds_ibdev at all.
 248          */
 249         rds_ibdev = ib_get_client_data(dev, &rds_ib_client);
 250         if (!rds_ibdev) {
 251                 if (printk_ratelimit())
 252                         printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
 253                                         dev->name);
 254                 return -EOPNOTSUPP;
 255         }
 256
 257         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 258                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 259         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 260                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 261
 262         /* Protection domain and memory range */
 263         ic->i_pd = rds_ibdev->pd;
 264         ic->i_mr = rds_ibdev->mr;
 265
 266         ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
 267                                      rds_ib_cq_event_handler, conn,
 268                                      ic->i_send_ring.w_nr + 1, 0);
 269         if (IS_ERR(ic->i_send_cq)) {
 270                 ret = PTR_ERR(ic->i_send_cq);
 271                 ic->i_send_cq = NULL;
 272                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 273                 goto out;
 274         }
 275
 276         ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
 277                                      rds_ib_cq_event_handler, conn,
 278                                      ic->i_recv_ring.w_nr, 0);
 279         if (IS_ERR(ic->i_recv_cq)) {
 280                 ret = PTR_ERR(ic->i_recv_cq);
 281                 ic->i_recv_cq = NULL;
 282                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 283                 goto out;
 284         }
 285
 286         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 287         if (ret) {
 288                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 289                 goto out;
 290         }
 291
 292         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 293         if (ret) {
 294                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 295                 goto out;
 296         }
 297
 298         /* XXX negotiate max send/recv with remote? */
 299         memset(&attr, 0, sizeof(attr));
 300         attr.event_handler = rds_ib_qp_event_handler;
 301         attr.qp_context = conn;
 302         /* + 1 to allow for the single ack message */
 303         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
 304         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 305         attr.cap.max_send_sge = rds_ibdev->max_sge;
 306         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 307         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 308         attr.qp_type = IB_QPT_RC;
 309         attr.send_cq = ic->i_send_cq;
 310         attr.recv_cq = ic->i_recv_cq;
 311
 312         /*
 313          * XXX this can fail if max_*_wr is too large?  Are we supposed
 314          * to back off until we get a value that the hardware can support?
 315          */
 316         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 317         if (ret) {
 318                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 319                 goto out;
 320         }
 321
 322         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 323                                            ic->i_send_ring.w_nr *
 324                                                 sizeof(struct rds_header),
 325                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 326         if (!ic->i_send_hdrs) {
 327                 ret = -ENOMEM;
 328                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 329                 goto out;
 330         }
 331
 332         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 333                                            ic->i_recv_ring.w_nr *
 334                                                 sizeof(struct rds_header),
 335                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 336         if (!ic->i_recv_hdrs) {
 337                 ret = -ENOMEM;
 338                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 339                 goto out;
 340         }
 341
 342         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 343                                        &ic->i_ack_dma, GFP_KERNEL);
 344         if (!ic->i_ack) {
 345                 ret = -ENOMEM;
 346                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 347                 goto out;
 348         }
 349
 350         ic->i_sends = vmalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 351                                    ibdev_to_node(dev));
 352         if (!ic->i_sends) {
 353                 ret = -ENOMEM;
 354                 rdsdebug("send allocation failed\n");
 355                 goto out;
 356         }
 357         memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
 358
 359         ic->i_recvs = vmalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 360                                    ibdev_to_node(dev));
 361         if (!ic->i_recvs) {
 362                 ret = -ENOMEM;
 363                 rdsdebug("recv allocation failed\n");
 364                 goto out;
 365         }
 366         memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
 367
 368         rds_ib_recv_init_ack(ic);
 369
 370         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 371                  ic->i_send_cq, ic->i_recv_cq);
 372
 373 out:
 374         return ret;
 375 }
 376
 377 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 378 {
 379         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 380         u16 common;
 381         u32 version = 0;
 382
 383         /*
 384          * rdma_cm private data is odd - when there is any private data in the
 385          * request, we will be given a pretty large buffer without telling us the
 386          * original size. The only way to tell the difference is by looking at
 387          * the contents, which are initialized to zero.
 388          * If the protocol version fields aren't set, this is a connection attempt
 389          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 390          * We really should have changed this for OFED 1.3 :-(
 391          */
 392
 393         /* Be paranoid. RDS always has privdata */
 394         if (!event->param.conn.private_data_len) {
 395                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 396                         "rejecting\n");
 397                 return 0;
 398         }
 399
 400         /* Even if len is crap *now* I still want to check it. -ASG */
 401         if (event->param.conn.private_data_len < sizeof (*dp) ||
 402             dp->dp_protocol_major == 0)
 403                 return RDS_PROTOCOL_3_0;
 404
 405         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 406         if (dp->dp_protocol_major == 3 && common) {
 407                 version = RDS_PROTOCOL_3_0;
 408                 while ((common >>= 1) != 0)
 409                         version++;
 410         } else if (printk_ratelimit()) {
 411                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
 412                         "incompatible protocol version %u.%u\n",
 413                         &dp->dp_saddr,
 414                         dp->dp_protocol_major,
 415                         dp->dp_protocol_minor);
 416         }
 417         return version;
 418 }
 419
 420 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 421                                     struct rdma_cm_event *event)
 422 {
 423         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 424         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 425         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 426         struct rds_ib_connect_private dp_rep;
 427         struct rds_connection *conn = NULL;
 428         struct rds_ib_connection *ic = NULL;
 429         struct rdma_conn_param conn_param;
 430         u32 version;
 431         int err, destroy = 1;
 432
 433         /* Check whether the remote protocol version matches ours. */
 434         version = rds_ib_protocol_compatible(event);
 435         if (!version)
 436                 goto out;
 437
 438         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 439                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 440                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 441                  (unsigned long long)be64_to_cpu(lguid),
 442                  (unsigned long long)be64_to_cpu(fguid));
 443
 444         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
 445                                GFP_KERNEL);
 446         if (IS_ERR(conn)) {
 447                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 448                 conn = NULL;
 449                 goto out;
 450         }
 451
 452         /*
 453          * The connection request may occur while the
 454          * previous connection exist, e.g. in case of failover.
 455          * But as connections may be initiated simultaneously
 456          * by both hosts, we have a random backoff mechanism -
 457          * see the comment above rds_queue_reconnect()
 458          */
 459         mutex_lock(&conn->c_cm_lock);
 460         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 461                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 462                         rdsdebug("incoming connect while connecting\n");
 463                         rds_conn_drop(conn);
 464                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 465                 } else
 466                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 467                         /* Wait and see - our connect may still be succeeding */
 468                         rds_ib_stats_inc(s_ib_connect_raced);
 469                 }
 470                 mutex_unlock(&conn->c_cm_lock);
 471                 goto out;
 472         }
 473
 474         ic = conn->c_transport_data;
 475
 476         rds_ib_set_protocol(conn, version);
 477         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 478
 479         /* If the peer gave us the last packet it saw, process this as if
 480          * we had received a regular ACK. */
 481         if (dp->dp_ack_seq)
 482                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 483
 484         BUG_ON(cm_id->context);
 485         BUG_ON(ic->i_cm_id);
 486
 487         ic->i_cm_id = cm_id;
 488         cm_id->context = conn;
 489
 490         /* We got halfway through setting up the ib_connection, if we
 491          * fail now, we have to take the long route out of this mess. */
 492         destroy = 0;
 493
 494         err = rds_ib_setup_qp(conn);
 495         if (err) {
 496                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 497                 mutex_unlock(&conn->c_cm_lock);
 498                 goto out;
 499         }
 500
 501         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 502                 event->param.conn.responder_resources,
 503                 event->param.conn.initiator_depth);
 504
 505         /* rdma_accept() calls rdma_reject() internally if it fails */
 506         err = rdma_accept(cm_id, &conn_param);
 507         mutex_unlock(&conn->c_cm_lock);
 508         if (err) {
 509                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 510                 goto out;
 511         }
 512
 513         return 0;
 514
 515 out:
 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);
 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                  * Don't wait for the send ring to be empty -- there may be completed
 627                  * non-signaled entries sitting on there. We unmap these below.
 628                  */
 629                 wait_event(rds_ib_ring_empty_wait,
 630                         rds_ib_ring_empty(&ic->i_recv_ring));
 631
 632                 if (ic->i_send_hdrs)
 633                         ib_dma_free_coherent(dev,
 634                                            ic->i_send_ring.w_nr *
 635                                                 sizeof(struct rds_header),
 636                                            ic->i_send_hdrs,
 637                                            ic->i_send_hdrs_dma);
 638
 639                 if (ic->i_recv_hdrs)
 640                         ib_dma_free_coherent(dev,
 641                                            ic->i_recv_ring.w_nr *
 642                                                 sizeof(struct rds_header),
 643                                            ic->i_recv_hdrs,
 644                                            ic->i_recv_hdrs_dma);
 645
 646                 if (ic->i_ack)
 647                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 648                                              ic->i_ack, ic->i_ack_dma);
 649
 650                 if (ic->i_sends)
 651                         rds_ib_send_clear_ring(ic);
 652                 if (ic->i_recvs)
 653                         rds_ib_recv_clear_ring(ic);
 654
 655                 if (ic->i_cm_id->qp)
 656                         rdma_destroy_qp(ic->i_cm_id);
 657                 if (ic->i_send_cq)
 658                         ib_destroy_cq(ic->i_send_cq);
 659                 if (ic->i_recv_cq)
 660                         ib_destroy_cq(ic->i_recv_cq);
 661                 rdma_destroy_id(ic->i_cm_id);
 662
 663                 /*
 664                  * Move connection back to the nodev list.
 665                  */
 666                 if (ic->rds_ibdev)
 667                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 668
 669                 ic->i_cm_id = NULL;
 670                 ic->i_pd = NULL;
 671                 ic->i_mr = NULL;
 672                 ic->i_send_cq = NULL;
 673                 ic->i_recv_cq = NULL;
 674                 ic->i_send_hdrs = NULL;
 675                 ic->i_recv_hdrs = NULL;
 676                 ic->i_ack = NULL;
 677         }
 678         BUG_ON(ic->rds_ibdev);
 679
 680         /* Clear pending transmit */
 681         if (ic->i_data_op) {
 682                 struct rds_message *rm;
 683
 684                 rm = container_of(ic->i_data_op, struct rds_message, data);
 685                 rds_message_put(rm);
 686                 ic->i_data_op = NULL;
 687         }
 688
 689         /* Clear the ACK state */
 690         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 691 #ifdef KERNEL_HAS_ATOMIC64
 692         atomic64_set(&ic->i_ack_next, 0);
 693 #else
 694         ic->i_ack_next = 0;
 695 #endif
 696         ic->i_ack_recv = 0;
 697
 698         /* Clear flow control state */
 699         ic->i_flowctl = 0;
 700         atomic_set(&ic->i_credits, 0);
 701
 702         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 703         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 704
 705         if (ic->i_ibinc) {
 706                 rds_inc_put(&ic->i_ibinc->ii_inc);
 707                 ic->i_ibinc = NULL;
 708         }
 709
 710         vfree(ic->i_sends);
 711         ic->i_sends = NULL;
 712         vfree(ic->i_recvs);
 713         ic->i_recvs = NULL;
 714 }
 715
 716 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 717 {
 718         struct rds_ib_connection *ic;
 719         unsigned long flags;
 720
 721         /* XXX too lazy? */
 722         ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
 723         if (!ic)
 724                 return -ENOMEM;
 725
 726         INIT_LIST_HEAD(&ic->ib_node);
 727         tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn,
 728                      (unsigned long) ic);
 729         mutex_init(&ic->i_recv_mutex);
 730 #ifndef KERNEL_HAS_ATOMIC64
 731         spin_lock_init(&ic->i_ack_lock);
 732 #endif
 733
 734         /*
 735          * rds_ib_conn_shutdown() waits for these to be emptied so they
 736          * must be initialized before it can be called.
 737          */
 738         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 739         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 740
 741         ic->conn = conn;
 742         conn->c_transport_data = ic;
 743
 744         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 745         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 746         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 747
 748
 749         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 750         return 0;
 751 }
 752
 753 /*
 754  * Free a connection. Connection must be shut down and not set for reconnect.
 755  */
 756 void rds_ib_conn_free(void *arg)
 757 {
 758         struct rds_ib_connection *ic = arg;
 759         spinlock_t      *lock_ptr;
 760
 761         rdsdebug("ic %p\n", ic);
 762
 763         /*
 764          * Conn is either on a dev's list or on the nodev list.
 765          * A race with shutdown() or connect() would cause problems
 766          * (since rds_ibdev would change) but that should never happen.
 767          */
 768         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 769
 770         spin_lock_irq(lock_ptr);
 771         list_del(&ic->ib_node);
 772         spin_unlock_irq(lock_ptr);
 773
 774         kfree(ic);
 775 }
 776
 777
 778 /*
 779  * An error occurred on the connection
 780  */
 781 void
 782 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 783 {
 784         va_list ap;
 785
 786         rds_conn_drop(conn);
 787
 788         va_start(ap, fmt);
 789         vprintk(fmt, ap);
 790         va_end(ap);
 791 }