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) {
 102                 dp = event->param.conn.private_data;
 103
 104                 rds_ib_set_protocol(conn,
 105                                 RDS_PROTOCOL(dp->dp_protocol_major,
 106                                         dp->dp_protocol_minor));
 107                 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 108         }
 109
 110         printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
 111                         &conn->c_laddr,
 112                         RDS_PROTOCOL_MAJOR(conn->c_version),
 113                         RDS_PROTOCOL_MINOR(conn->c_version),
 114                         ic->i_flowctl ? ", flow control" : "");
 115
 116         /* Tune RNR behavior */
 117         rds_ib_tune_rnr(ic, &qp_attr);
 118
 119         qp_attr.qp_state = IB_QPS_RTS;
 120         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 121         if (err)
 122                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 123
 124         /* update ib_device with this local ipaddr & conn */
 125         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
 126         err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
 127         if (err)
 128                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
 129         err = rds_ib_add_conn(rds_ibdev, conn);
 130         if (err)
 131                 printk(KERN_ERR "rds_ib_add_conn failed (%d)\n", err);
 132
 133         /* If the peer gave us the last packet it saw, process this as if
 134          * we had received a regular ACK. */
 135         if (dp && dp->dp_ack_seq)
 136                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 137
 138         rds_connect_complete(conn);
 139 }
 140
 141 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 142                         struct rdma_conn_param *conn_param,
 143                         struct rds_ib_connect_private *dp,
 144                         u32 protocol_version)
 145 {
 146         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 147         /* XXX tune these? */
 148         conn_param->responder_resources = 1;
 149         conn_param->initiator_depth = 1;
 150         conn_param->retry_count = 7;
 151         conn_param->rnr_retry_count = 7;
 152
 153         if (dp) {
 154                 struct rds_ib_connection *ic = conn->c_transport_data;
 155
 156                 memset(dp, 0, sizeof(*dp));
 157                 dp->dp_saddr = conn->c_laddr;
 158                 dp->dp_daddr = conn->c_faddr;
 159                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 160                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 161                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 162                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
 163
 164                 /* Advertise flow control */
 165                 if (ic->i_flowctl) {
 166                         unsigned int credits;
 167
 168                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 169                         dp->dp_credit = cpu_to_be32(credits);
 170                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 171                 }
 172
 173                 conn_param->private_data = dp;
 174                 conn_param->private_data_len = sizeof(*dp);
 175         }
 176 }
 177
 178 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 179 {
 180         rdsdebug("event %u data %p\n", event->event, data);
 181 }
 182
 183 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 184 {
 185         struct rds_connection *conn = data;
 186         struct rds_ib_connection *ic = conn->c_transport_data;
 187
 188         rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event);
 189
 190         switch (event->event) {
 191         case IB_EVENT_COMM_EST:
 192                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 193                 break;
 194         default:
 195                 printk(KERN_WARNING "RDS/ib: unhandled QP event %u "
 196                        "on connection to %pI4\n", event->event,
 197                        &conn->c_faddr);
 198                 break;
 199         }
 200 }
 201
 202 /*
 203  * This needs to be very careful to not leave IS_ERR pointers around for
 204  * cleanup to trip over.
 205  */
 206 static int rds_ib_setup_qp(struct rds_connection *conn)
 207 {
 208         struct rds_ib_connection *ic = conn->c_transport_data;
 209         struct ib_device *dev = ic->i_cm_id->device;
 210         struct ib_qp_init_attr attr;
 211         struct rds_ib_device *rds_ibdev;
 212         int ret;
 213
 214         /* rds_ib_add_one creates a rds_ib_device object per IB device,
 215          * and allocates a protection domain, memory range and FMR pool
 216          * for each.  If that fails for any reason, it will not register
 217          * the rds_ibdev at all.
 218          */
 219         rds_ibdev = ib_get_client_data(dev, &rds_ib_client);
 220         if (rds_ibdev == NULL) {
 221                 if (printk_ratelimit())
 222                         printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n",
 223                                         dev->name);
 224                 return -EOPNOTSUPP;
 225         }
 226
 227         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 228                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 229         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 230                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 231
 232         /* Protection domain and memory range */
 233         ic->i_pd = rds_ibdev->pd;
 234         ic->i_mr = rds_ibdev->mr;
 235
 236         ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler,
 237                                      rds_ib_cq_event_handler, conn,
 238                                      ic->i_send_ring.w_nr + 1, 0);
 239         if (IS_ERR(ic->i_send_cq)) {
 240                 ret = PTR_ERR(ic->i_send_cq);
 241                 ic->i_send_cq = NULL;
 242                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 243                 goto out;
 244         }
 245
 246         ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler,
 247                                      rds_ib_cq_event_handler, conn,
 248                                      ic->i_recv_ring.w_nr, 0);
 249         if (IS_ERR(ic->i_recv_cq)) {
 250                 ret = PTR_ERR(ic->i_recv_cq);
 251                 ic->i_recv_cq = NULL;
 252                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 253                 goto out;
 254         }
 255
 256         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 257         if (ret) {
 258                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 259                 goto out;
 260         }
 261
 262         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 263         if (ret) {
 264                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 265                 goto out;
 266         }
 267
 268         /* XXX negotiate max send/recv with remote? */
 269         memset(&attr, 0, sizeof(attr));
 270         attr.event_handler = rds_ib_qp_event_handler;
 271         attr.qp_context = conn;
 272         /* + 1 to allow for the single ack message */
 273         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
 274         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 275         attr.cap.max_send_sge = rds_ibdev->max_sge;
 276         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 277         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 278         attr.qp_type = IB_QPT_RC;
 279         attr.send_cq = ic->i_send_cq;
 280         attr.recv_cq = ic->i_recv_cq;
 281
 282         /*
 283          * XXX this can fail if max_*_wr is too large?  Are we supposed
 284          * to back off until we get a value that the hardware can support?
 285          */
 286         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 287         if (ret) {
 288                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 289                 goto out;
 290         }
 291
 292         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 293                                            ic->i_send_ring.w_nr *
 294                                                 sizeof(struct rds_header),
 295                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 296         if (ic->i_send_hdrs == NULL) {
 297                 ret = -ENOMEM;
 298                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 299                 goto out;
 300         }
 301
 302         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 303                                            ic->i_recv_ring.w_nr *
 304                                                 sizeof(struct rds_header),
 305                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 306         if (ic->i_recv_hdrs == NULL) {
 307                 ret = -ENOMEM;
 308                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 309                 goto out;
 310         }
 311
 312         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 313                                        &ic->i_ack_dma, GFP_KERNEL);
 314         if (ic->i_ack == NULL) {
 315                 ret = -ENOMEM;
 316                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 317                 goto out;
 318         }
 319
 320         ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work));
 321         if (ic->i_sends == NULL) {
 322                 ret = -ENOMEM;
 323                 rdsdebug("send allocation failed\n");
 324                 goto out;
 325         }
 326         rds_ib_send_init_ring(ic);
 327
 328         ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work));
 329         if (ic->i_recvs == NULL) {
 330                 ret = -ENOMEM;
 331                 rdsdebug("recv allocation failed\n");
 332                 goto out;
 333         }
 334
 335         rds_ib_recv_init_ring(ic);
 336         rds_ib_recv_init_ack(ic);
 337
 338         /* Post receive buffers - as a side effect, this will update
 339          * the posted credit count. */
 340         rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1);
 341
 342         rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr,
 343                  ic->i_send_cq, ic->i_recv_cq);
 344
 345 out:
 346         return ret;
 347 }
 348
 349 static u32 rds_ib_protocol_compatible(const struct rds_ib_connect_private *dp)
 350 {
 351         u16 common;
 352         u32 version = 0;
 353
 354         /* rdma_cm private data is odd - when there is any private data in the
 355          * request, we will be given a pretty large buffer without telling us the
 356          * original size. The only way to tell the difference is by looking at
 357          * the contents, which are initialized to zero.
 358          * If the protocol version fields aren't set, this is a connection attempt
 359          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 360          * We really should have changed this for OFED 1.3 :-( */
 361         if (dp->dp_protocol_major == 0)
 362                 return RDS_PROTOCOL_3_0;
 363
 364         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 365         if (dp->dp_protocol_major == 3 && common) {
 366                 version = RDS_PROTOCOL_3_0;
 367                 while ((common >>= 1) != 0)
 368                         version++;
 369         } else if (printk_ratelimit()) {
 370                 printk(KERN_NOTICE "RDS: Connection from %pI4 using "
 371                         "incompatible protocol version %u.%u\n",
 372                         &dp->dp_saddr,
 373                         dp->dp_protocol_major,
 374                         dp->dp_protocol_minor);
 375         }
 376         return version;
 377 }
 378
 379 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 380                                     struct rdma_cm_event *event)
 381 {
 382         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 383         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 384         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 385         struct rds_ib_connect_private dp_rep;
 386         struct rds_connection *conn = NULL;
 387         struct rds_ib_connection *ic = NULL;
 388         struct rdma_conn_param conn_param;
 389         u32 version;
 390         int err, destroy = 1;
 391
 392         /* Check whether the remote protocol version matches ours. */
 393         version = rds_ib_protocol_compatible(dp);
 394         if (!version)
 395                 goto out;
 396
 397         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 398                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 399                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 400                  (unsigned long long)be64_to_cpu(lguid),
 401                  (unsigned long long)be64_to_cpu(fguid));
 402
 403         conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport,
 404                                GFP_KERNEL);
 405         if (IS_ERR(conn)) {
 406                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 407                 conn = NULL;
 408                 goto out;
 409         }
 410
 411         /*
 412          * The connection request may occur while the
 413          * previous connection exist, e.g. in case of failover.
 414          * But as connections may be initiated simultaneously
 415          * by both hosts, we have a random backoff mechanism -
 416          * see the comment above rds_queue_reconnect()
 417          */
 418         mutex_lock(&conn->c_cm_lock);
 419         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 420                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 421                         rdsdebug("incoming connect while connecting\n");
 422                         rds_conn_drop(conn);
 423                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 424                 } else
 425                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 426                         /* Wait and see - our connect may still be succeeding */
 427                         rds_ib_stats_inc(s_ib_connect_raced);
 428                 }
 429                 mutex_unlock(&conn->c_cm_lock);
 430                 goto out;
 431         }
 432
 433         ic = conn->c_transport_data;
 434
 435         rds_ib_set_protocol(conn, version);
 436         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 437
 438         /* If the peer gave us the last packet it saw, process this as if
 439          * we had received a regular ACK. */
 440         if (dp->dp_ack_seq)
 441                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 442
 443         BUG_ON(cm_id->context);
 444         BUG_ON(ic->i_cm_id);
 445
 446         ic->i_cm_id = cm_id;
 447         cm_id->context = conn;
 448
 449         /* We got halfway through setting up the ib_connection, if we
 450          * fail now, we have to take the long route out of this mess. */
 451         destroy = 0;
 452
 453         err = rds_ib_setup_qp(conn);
 454         if (err) {
 455                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 456                 goto out;
 457         }
 458
 459         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version);
 460
 461         /* rdma_accept() calls rdma_reject() internally if it fails */
 462         err = rdma_accept(cm_id, &conn_param);
 463         mutex_unlock(&conn->c_cm_lock);
 464         if (err) {
 465                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 466                 goto out;
 467         }
 468
 469         return 0;
 470
 471 out:
 472         rdma_reject(cm_id, NULL, 0);
 473         return destroy;
 474 }
 475
 476
 477 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 478 {
 479         struct rds_connection *conn = cm_id->context;
 480         struct rds_ib_connection *ic = conn->c_transport_data;
 481         struct rdma_conn_param conn_param;
 482         struct rds_ib_connect_private dp;
 483         int ret;
 484
 485         /* If the peer doesn't do protocol negotiation, we must
 486          * default to RDSv3.0 */
 487         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 488         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 489
 490         ret = rds_ib_setup_qp(conn);
 491         if (ret) {
 492                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 493                 goto out;
 494         }
 495
 496         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION);
 497
 498         ret = rdma_connect(cm_id, &conn_param);
 499         if (ret)
 500                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 501
 502 out:
 503         /* Beware - returning non-zero tells the rdma_cm to destroy
 504          * the cm_id. We should certainly not do it as long as we still
 505          * "own" the cm_id. */
 506         if (ret) {
 507                 if (ic->i_cm_id == cm_id)
 508                         ret = 0;
 509         }
 510         return ret;
 511 }
 512
 513 int rds_ib_conn_connect(struct rds_connection *conn)
 514 {
 515         struct rds_ib_connection *ic = conn->c_transport_data;
 516         struct sockaddr_in src, dest;
 517         int ret;
 518
 519         /* XXX I wonder what affect the port space has */
 520         /* delegate cm event handler to rdma_transport */
 521         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 522                                      RDMA_PS_TCP);
 523         if (IS_ERR(ic->i_cm_id)) {
 524                 ret = PTR_ERR(ic->i_cm_id);
 525                 ic->i_cm_id = NULL;
 526                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 527                 goto out;
 528         }
 529
 530         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 531
 532         src.sin_family = AF_INET;
 533         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 534         src.sin_port = (__force u16)htons(0);
 535
 536         dest.sin_family = AF_INET;
 537         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 538         dest.sin_port = (__force u16)htons(RDS_PORT);
 539
 540         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 541                                 (struct sockaddr *)&dest,
 542                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 543         if (ret) {
 544                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 545                          ret);
 546                 rdma_destroy_id(ic->i_cm_id);
 547                 ic->i_cm_id = NULL;
 548         }
 549
 550 out:
 551         return ret;
 552 }
 553
 554 /*
 555  * This is so careful about only cleaning up resources that were built up
 556  * so that it can be called at any point during startup.  In fact it
 557  * can be called multiple times for a given connection.
 558  */
 559 void rds_ib_conn_shutdown(struct rds_connection *conn)
 560 {
 561         struct rds_ib_connection *ic = conn->c_transport_data;
 562         int err = 0;
 563
 564         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 565                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 566                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 567
 568         if (ic->i_cm_id) {
 569                 struct ib_device *dev = ic->i_cm_id->device;
 570
 571                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 572                 err = rdma_disconnect(ic->i_cm_id);
 573                 if (err) {
 574                         /* Actually this may happen quite frequently, when
 575                          * an outgoing connect raced with an incoming connect.
 576                          */
 577                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 578                                 ic->i_cm_id, err);
 579                 }
 580
 581                 wait_event(rds_ib_ring_empty_wait,
 582                         rds_ib_ring_empty(&ic->i_send_ring) &&
 583                         rds_ib_ring_empty(&ic->i_recv_ring));
 584
 585                 if (ic->i_send_hdrs)
 586                         ib_dma_free_coherent(dev,
 587                                            ic->i_send_ring.w_nr *
 588                                                 sizeof(struct rds_header),
 589                                            ic->i_send_hdrs,
 590                                            ic->i_send_hdrs_dma);
 591
 592                 if (ic->i_recv_hdrs)
 593                         ib_dma_free_coherent(dev,
 594                                            ic->i_recv_ring.w_nr *
 595                                                 sizeof(struct rds_header),
 596                                            ic->i_recv_hdrs,
 597                                            ic->i_recv_hdrs_dma);
 598
 599                 if (ic->i_ack)
 600                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 601                                              ic->i_ack, ic->i_ack_dma);
 602
 603                 if (ic->i_sends)
 604                         rds_ib_send_clear_ring(ic);
 605                 if (ic->i_recvs)
 606                         rds_ib_recv_clear_ring(ic);
 607
 608                 if (ic->i_cm_id->qp)
 609                         rdma_destroy_qp(ic->i_cm_id);
 610                 if (ic->i_send_cq)
 611                         ib_destroy_cq(ic->i_send_cq);
 612                 if (ic->i_recv_cq)
 613                         ib_destroy_cq(ic->i_recv_cq);
 614                 rdma_destroy_id(ic->i_cm_id);
 615
 616                 /*
 617                  * Move connection back to the nodev list.
 618                  */
 619                 if (ic->rds_ibdev) {
 620
 621                         spin_lock_irq(&ic->rds_ibdev->spinlock);
 622                         BUG_ON(list_empty(&ic->ib_node));
 623                         list_del(&ic->ib_node);
 624                         spin_unlock_irq(&ic->rds_ibdev->spinlock);
 625
 626                         spin_lock_irq(&ib_nodev_conns_lock);
 627                         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 628                         spin_unlock_irq(&ib_nodev_conns_lock);
 629                         ic->rds_ibdev = NULL;
 630                 }
 631
 632                 ic->i_cm_id = NULL;
 633                 ic->i_pd = NULL;
 634                 ic->i_mr = NULL;
 635                 ic->i_send_cq = NULL;
 636                 ic->i_recv_cq = NULL;
 637                 ic->i_send_hdrs = NULL;
 638                 ic->i_recv_hdrs = NULL;
 639                 ic->i_ack = NULL;
 640         }
 641         BUG_ON(ic->rds_ibdev);
 642
 643         /* Clear pending transmit */
 644         if (ic->i_rm) {
 645                 rds_message_put(ic->i_rm);
 646                 ic->i_rm = NULL;
 647         }
 648
 649         /* Clear the ACK state */
 650         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 651         rds_ib_set_64bit(&ic->i_ack_next, 0);
 652         ic->i_ack_recv = 0;
 653
 654         /* Clear flow control state */
 655         ic->i_flowctl = 0;
 656         atomic_set(&ic->i_credits, 0);
 657
 658         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 659         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 660
 661         if (ic->i_ibinc) {
 662                 rds_inc_put(&ic->i_ibinc->ii_inc);
 663                 ic->i_ibinc = NULL;
 664         }
 665
 666         vfree(ic->i_sends);
 667         ic->i_sends = NULL;
 668         vfree(ic->i_recvs);
 669         ic->i_recvs = NULL;
 670 }
 671
 672 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 673 {
 674         struct rds_ib_connection *ic;
 675         unsigned long flags;
 676
 677         /* XXX too lazy? */
 678         ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL);
 679         if (ic == NULL)
 680                 return -ENOMEM;
 681
 682         INIT_LIST_HEAD(&ic->ib_node);
 683         mutex_init(&ic->i_recv_mutex);
 684
 685         /*
 686          * rds_ib_conn_shutdown() waits for these to be emptied so they
 687          * must be initialized before it can be called.
 688          */
 689         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 690         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 691
 692         ic->conn = conn;
 693         conn->c_transport_data = ic;
 694
 695         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 696         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 697         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 698
 699
 700         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 701         return 0;
 702 }
 703
 704 void rds_ib_conn_free(void *arg)
 705 {
 706         struct rds_ib_connection *ic = arg;
 707         rdsdebug("ic %p\n", ic);
 708         list_del(&ic->ib_node);
 709         kfree(ic);
 710 }
 711
 712
 713 /*
 714  * An error occurred on the connection
 715  */
 716 void
 717 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 718 {
 719         va_list ap;
 720
 721         rds_conn_drop(conn);
 722
 723         va_start(ap, fmt);
 724         vprintk(fmt, ap);
 725         va_end(ap);
 726 }