net/rds/ib_cm.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/slab.h>
  36 #include <linux/vmalloc.h>
  37 #include <linux/ratelimit.h>
  38
  39 #include "rds.h"
  40 #include "ib.h"
  41
  42 /*
  43  * Set the selected protocol version
  44  */
  45 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  46 {
  47         conn->c_version = version;
  48 }
  49
  50 /*
  51  * Set up flow control
  52  */
  53 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  54 {
  55         struct rds_ib_connection *ic = conn->c_transport_data;
  56
  57         if (rds_ib_sysctl_flow_control && credits != 0) {
  58                 /* We're doing flow control */
  59                 ic->i_flowctl = 1;
  60                 rds_ib_send_add_credits(conn, credits);
  61         } else {
  62                 ic->i_flowctl = 0;
  63         }
  64 }
  65
  66 /*
  67  * Tune RNR behavior. Without flow control, we use a rather
  68  * low timeout, but not the absolute minimum - this should
  69  * be tunable.
  70  *
  71  * We already set the RNR retry count to 7 (which is the
  72  * smallest infinite number :-) above.
  73  * If flow control is off, we want to change this back to 0
  74  * so that we learn quickly when our credit accounting is
  75  * buggy.
  76  *
  77  * Caller passes in a qp_attr pointer - don't waste stack spacv
  78  * by allocation this twice.
  79  */
  80 static void
  81 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  82 {
  83         int ret;
  84
  85         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  86         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  87         if (ret)
  88                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  89 }
  90
  91 /*
  92  * Connection established.
  93  * We get here for both outgoing and incoming connection.
  94  */
  95 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  96 {
  97         const struct rds_ib_connect_private *dp = NULL;
  98         struct rds_ib_connection *ic = conn->c_transport_data;
  99         struct ib_qp_attr qp_attr;
 100         int err;
 101
 102         if (event->param.conn.private_data_len >= sizeof(*dp)) {
 103                 dp = event->param.conn.private_data;
 104
 105                 /* make sure it isn't empty data */
 106                 if (dp->dp_protocol_major) {
 107                         rds_ib_set_protocol(conn,
 108                                 RDS_PROTOCOL(dp->dp_protocol_major,
 109                                 dp->dp_protocol_minor));
 110                         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 111                 }
 112         }
 113
 114         if (conn->c_version < RDS_PROTOCOL(3,1)) {
 115                 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
 116                        " no longer supported\n",
 117                        &conn->c_faddr,
 118                        RDS_PROTOCOL_MAJOR(conn->c_version),
 119                        RDS_PROTOCOL_MINOR(conn->c_version));
 120                 rds_conn_destroy(conn);
 121                 return;
 122         } else {
 123                 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
 124                        &conn->c_faddr,
 125                        RDS_PROTOCOL_MAJOR(conn->c_version),
 126                        RDS_PROTOCOL_MINOR(conn->c_version),
 127                        ic->i_flowctl ? ", flow control" : "");
 128         }
 129
 130         /*
 131          * Init rings and fill recv. this needs to wait until protocol negotiation
 132          * is complete, since ring layout is different from 3.0 to 3.1.
 133          */
 134         rds_ib_send_init_ring(ic);
 135         rds_ib_recv_init_ring(ic);
 136         /* Post receive buffers - as a side effect, this will update
 137          * the posted credit count. */
 138         rds_ib_recv_refill(conn, 1, GFP_KERNEL);
 139
 140         /* Tune RNR behavior */
 141         rds_ib_tune_rnr(ic, &qp_attr);
 142
 143         qp_attr.qp_state = IB_QPS_RTS;
 144         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 145         if (err)
 146                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 147
 148         /* update ib_device with this local ipaddr */
 149         err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
 150         if (err)
 151                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 152                         err);
 153
 154         /* If the peer gave us the last packet it saw, process this as if
 155          * we had received a regular ACK. */
 156         if (dp) {
 157                 /* dp structure start is not guaranteed to be 8 bytes aligned.
 158                  * Since dp_ack_seq is 64-bit extended load operations can be
 159                  * used so go through get_unaligned to avoid unaligned errors.
 160                  */
 161                 __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
 162
 163                 if (dp_ack_seq)
 164                         rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
 165                                             NULL);
 166         }
 167
 168         rds_connect_complete(conn);
 169 }
 170
 171 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 172                         struct rdma_conn_param *conn_param,
 173                         struct rds_ib_connect_private *dp,
 174                         u32 protocol_version,
 175                         u32 max_responder_resources,
 176                         u32 max_initiator_depth)
 177 {
 178         struct rds_ib_connection *ic = conn->c_transport_data;
 179         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 180
 181         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 182
 183         conn_param->responder_resources =
 184                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 185         conn_param->initiator_depth =
 186                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 187         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 188         conn_param->rnr_retry_count = 7;
 189
 190         if (dp) {
 191                 memset(dp, 0, sizeof(*dp));
 192                 dp->dp_saddr = conn->c_laddr;
 193                 dp->dp_daddr = conn->c_faddr;
 194                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 195                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 196                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 197                 dp->dp_ack_seq = rds_ib_piggyb_ack(ic);
 198
 199                 /* Advertise flow control */
 200                 if (ic->i_flowctl) {
 201                         unsigned int credits;
 202
 203                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 204                         dp->dp_credit = cpu_to_be32(credits);
 205                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 206                 }
 207
 208                 conn_param->private_data = dp;
 209                 conn_param->private_data_len = sizeof(*dp);
 210         }
 211 }
 212
 213 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 214 {
 215         rdsdebug("event %u (%s) data %p\n",
 216                  event->event, ib_event_msg(event->event), data);
 217 }
 218
 219 /* Plucking the oldest entry from the ring can be done concurrently with
 220  * the thread refilling the ring.  Each ring operation is protected by
 221  * spinlocks and the transient state of refilling doesn't change the
 222  * recording of which entry is oldest.
 223  *
 224  * This relies on IB only calling one cq comp_handler for each cq so that
 225  * there will only be one caller of rds_recv_incoming() per RDS connection.
 226  */
 227 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
 228 {
 229         struct rds_connection *conn = context;
 230         struct rds_ib_connection *ic = conn->c_transport_data;
 231
 232         rdsdebug("conn %p cq %p\n", conn, cq);
 233
 234         rds_ib_stats_inc(s_ib_evt_handler_call);
 235
 236         tasklet_schedule(&ic->i_recv_tasklet);
 237 }
 238
 239 static void poll_cq(struct rds_ib_connection *ic, struct ib_cq *cq,
 240                     struct ib_wc *wcs,
 241                     struct rds_ib_ack_state *ack_state)
 242 {
 243         int nr;
 244         int i;
 245         struct ib_wc *wc;
 246
 247         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 248                 for (i = 0; i < nr; i++) {
 249                         wc = wcs + i;
 250                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 251                                  (unsigned long long)wc->wr_id, wc->status,
 252                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 253
 254                         if (wc->wr_id & RDS_IB_SEND_OP)
 255                                 rds_ib_send_cqe_handler(ic, wc);
 256                         else
 257                                 rds_ib_recv_cqe_handler(ic, wc, ack_state);
 258                 }
 259         }
 260 }
 261
 262 static void rds_ib_tasklet_fn_send(unsigned long data)
 263 {
 264         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 265         struct rds_connection *conn = ic->conn;
 266         struct rds_ib_ack_state state;
 267
 268         rds_ib_stats_inc(s_ib_tasklet_call);
 269
 270         memset(&state, 0, sizeof(state));
 271         poll_cq(ic, ic->i_send_cq, ic->i_send_wc, &state);
 272         ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 273         poll_cq(ic, ic->i_send_cq, ic->i_send_wc, &state);
 274
 275         if (rds_conn_up(conn) &&
 276             (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 277             test_bit(0, &conn->c_map_queued)))
 278                 rds_send_xmit(ic->conn);
 279 }
 280
 281 static void rds_ib_tasklet_fn_recv(unsigned long data)
 282 {
 283         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 284         struct rds_connection *conn = ic->conn;
 285         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 286         struct rds_ib_ack_state state;
 287
 288         BUG_ON(!rds_ibdev);
 289
 290         rds_ib_stats_inc(s_ib_tasklet_call);
 291
 292         memset(&state, 0, sizeof(state));
 293         poll_cq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 294         ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 295         poll_cq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 296
 297         if (state.ack_next_valid)
 298                 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
 299         if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
 300                 rds_send_drop_acked(conn, state.ack_recv, NULL);
 301                 ic->i_ack_recv = state.ack_recv;
 302         }
 303
 304         if (rds_conn_up(conn))
 305                 rds_ib_attempt_ack(ic);
 306 }
 307
 308 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 309 {
 310         struct rds_connection *conn = data;
 311         struct rds_ib_connection *ic = conn->c_transport_data;
 312
 313         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 314                  ib_event_msg(event->event));
 315
 316         switch (event->event) {
 317         case IB_EVENT_COMM_EST:
 318                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 319                 break;
 320         default:
 321                 rdsdebug("Fatal QP Event %u (%s) "
 322                         "- connection %pI4->%pI4, reconnecting\n",
 323                         event->event, ib_event_msg(event->event),
 324                         &conn->c_laddr, &conn->c_faddr);
 325                 rds_conn_drop(conn);
 326                 break;
 327         }
 328 }
 329
 330 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
 331 {
 332         struct rds_connection *conn = context;
 333         struct rds_ib_connection *ic = conn->c_transport_data;
 334
 335         rdsdebug("conn %p cq %p\n", conn, cq);
 336
 337         rds_ib_stats_inc(s_ib_evt_handler_call);
 338
 339         tasklet_schedule(&ic->i_send_tasklet);
 340 }
 341
 342 /*
 343  * This needs to be very careful to not leave IS_ERR pointers around for
 344  * cleanup to trip over.
 345  */
 346 static int rds_ib_setup_qp(struct rds_connection *conn)
 347 {
 348         struct rds_ib_connection *ic = conn->c_transport_data;
 349         struct ib_device *dev = ic->i_cm_id->device;
 350         struct ib_qp_init_attr attr;
 351         struct ib_cq_init_attr cq_attr = {};
 352         struct rds_ib_device *rds_ibdev;
 353         int ret;
 354
 355         /*
 356          * It's normal to see a null device if an incoming connection races
 357          * with device removal, so we don't print a warning.
 358          */
 359         rds_ibdev = rds_ib_get_client_data(dev);
 360         if (!rds_ibdev)
 361                 return -EOPNOTSUPP;
 362
 363         /* add the conn now so that connection establishment has the dev */
 364         rds_ib_add_conn(rds_ibdev, conn);
 365
 366         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 367                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 368         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 369                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 370
 371         /* Protection domain and memory range */
 372         ic->i_pd = rds_ibdev->pd;
 373
 374         cq_attr.cqe = ic->i_send_ring.w_nr + 1;
 375
 376         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 377                                      rds_ib_cq_event_handler, conn,
 378                                      &cq_attr);
 379         if (IS_ERR(ic->i_send_cq)) {
 380                 ret = PTR_ERR(ic->i_send_cq);
 381                 ic->i_send_cq = NULL;
 382                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 383                 goto out;
 384         }
 385
 386         cq_attr.cqe = ic->i_recv_ring.w_nr;
 387         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 388                                      rds_ib_cq_event_handler, conn,
 389                                      &cq_attr);
 390         if (IS_ERR(ic->i_recv_cq)) {
 391                 ret = PTR_ERR(ic->i_recv_cq);
 392                 ic->i_recv_cq = NULL;
 393                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 394                 goto out;
 395         }
 396
 397         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 398         if (ret) {
 399                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 400                 goto out;
 401         }
 402
 403         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 404         if (ret) {
 405                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 406                 goto out;
 407         }
 408
 409         /* XXX negotiate max send/recv with remote? */
 410         memset(&attr, 0, sizeof(attr));
 411         attr.event_handler = rds_ib_qp_event_handler;
 412         attr.qp_context = conn;
 413         /* + 1 to allow for the single ack message */
 414         attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1;
 415         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 416         attr.cap.max_send_sge = rds_ibdev->max_sge;
 417         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 418         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 419         attr.qp_type = IB_QPT_RC;
 420         attr.send_cq = ic->i_send_cq;
 421         attr.recv_cq = ic->i_recv_cq;
 422
 423         /*
 424          * XXX this can fail if max_*_wr is too large?  Are we supposed
 425          * to back off until we get a value that the hardware can support?
 426          */
 427         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 428         if (ret) {
 429                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 430                 goto out;
 431         }
 432
 433         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 434                                            ic->i_send_ring.w_nr *
 435                                                 sizeof(struct rds_header),
 436                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 437         if (!ic->i_send_hdrs) {
 438                 ret = -ENOMEM;
 439                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 440                 goto out;
 441         }
 442
 443         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 444                                            ic->i_recv_ring.w_nr *
 445                                                 sizeof(struct rds_header),
 446                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 447         if (!ic->i_recv_hdrs) {
 448                 ret = -ENOMEM;
 449                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 450                 goto out;
 451         }
 452
 453         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 454                                        &ic->i_ack_dma, GFP_KERNEL);
 455         if (!ic->i_ack) {
 456                 ret = -ENOMEM;
 457                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 458                 goto out;
 459         }
 460
 461         ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 462                                    ibdev_to_node(dev));
 463         if (!ic->i_sends) {
 464                 ret = -ENOMEM;
 465                 rdsdebug("send allocation failed\n");
 466                 goto out;
 467         }
 468
 469         ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 470                                    ibdev_to_node(dev));
 471         if (!ic->i_recvs) {
 472                 ret = -ENOMEM;
 473                 rdsdebug("recv allocation failed\n");
 474                 goto out;
 475         }
 476
 477         rds_ib_recv_init_ack(ic);
 478
 479         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 480                  ic->i_send_cq, ic->i_recv_cq);
 481
 482 out:
 483         rds_ib_dev_put(rds_ibdev);
 484         return ret;
 485 }
 486
 487 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 488 {
 489         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 490         u16 common;
 491         u32 version = 0;
 492
 493         /*
 494          * rdma_cm private data is odd - when there is any private data in the
 495          * request, we will be given a pretty large buffer without telling us the
 496          * original size. The only way to tell the difference is by looking at
 497          * the contents, which are initialized to zero.
 498          * If the protocol version fields aren't set, this is a connection attempt
 499          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 500          * We really should have changed this for OFED 1.3 :-(
 501          */
 502
 503         /* Be paranoid. RDS always has privdata */
 504         if (!event->param.conn.private_data_len) {
 505                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 506                         "rejecting\n");
 507                 return 0;
 508         }
 509
 510         /* Even if len is crap *now* I still want to check it. -ASG */
 511         if (event->param.conn.private_data_len < sizeof (*dp) ||
 512             dp->dp_protocol_major == 0)
 513                 return RDS_PROTOCOL_3_0;
 514
 515         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 516         if (dp->dp_protocol_major == 3 && common) {
 517                 version = RDS_PROTOCOL_3_0;
 518                 while ((common >>= 1) != 0)
 519                         version++;
 520         } else
 521                 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 522                                 &dp->dp_saddr,
 523                                 dp->dp_protocol_major,
 524                                 dp->dp_protocol_minor);
 525         return version;
 526 }
 527
 528 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 529                                     struct rdma_cm_event *event)
 530 {
 531         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 532         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 533         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 534         struct rds_ib_connect_private dp_rep;
 535         struct rds_connection *conn = NULL;
 536         struct rds_ib_connection *ic = NULL;
 537         struct rdma_conn_param conn_param;
 538         u32 version;
 539         int err = 1, destroy = 1;
 540
 541         /* Check whether the remote protocol version matches ours. */
 542         version = rds_ib_protocol_compatible(event);
 543         if (!version)
 544                 goto out;
 545
 546         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 547                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 548                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 549                  (unsigned long long)be64_to_cpu(lguid),
 550                  (unsigned long long)be64_to_cpu(fguid));
 551
 552         /* RDS/IB is not currently netns aware, thus init_net */
 553         conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
 554                                &rds_ib_transport, GFP_KERNEL);
 555         if (IS_ERR(conn)) {
 556                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 557                 conn = NULL;
 558                 goto out;
 559         }
 560
 561         /*
 562          * The connection request may occur while the
 563          * previous connection exist, e.g. in case of failover.
 564          * But as connections may be initiated simultaneously
 565          * by both hosts, we have a random backoff mechanism -
 566          * see the comment above rds_queue_reconnect()
 567          */
 568         mutex_lock(&conn->c_cm_lock);
 569         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 570                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 571                         rdsdebug("incoming connect while connecting\n");
 572                         rds_conn_drop(conn);
 573                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 574                 } else
 575                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 576                         /* Wait and see - our connect may still be succeeding */
 577                         rds_ib_stats_inc(s_ib_connect_raced);
 578                 }
 579                 goto out;
 580         }
 581
 582         ic = conn->c_transport_data;
 583
 584         rds_ib_set_protocol(conn, version);
 585         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 586
 587         /* If the peer gave us the last packet it saw, process this as if
 588          * we had received a regular ACK. */
 589         if (dp->dp_ack_seq)
 590                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 591
 592         BUG_ON(cm_id->context);
 593         BUG_ON(ic->i_cm_id);
 594
 595         ic->i_cm_id = cm_id;
 596         cm_id->context = conn;
 597
 598         /* We got halfway through setting up the ib_connection, if we
 599          * fail now, we have to take the long route out of this mess. */
 600         destroy = 0;
 601
 602         err = rds_ib_setup_qp(conn);
 603         if (err) {
 604                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 605                 goto out;
 606         }
 607
 608         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 609                 event->param.conn.responder_resources,
 610                 event->param.conn.initiator_depth);
 611
 612         /* rdma_accept() calls rdma_reject() internally if it fails */
 613         err = rdma_accept(cm_id, &conn_param);
 614         if (err)
 615                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 616
 617 out:
 618         if (conn)
 619                 mutex_unlock(&conn->c_cm_lock);
 620         if (err)
 621                 rdma_reject(cm_id, NULL, 0);
 622         return destroy;
 623 }
 624
 625
 626 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 627 {
 628         struct rds_connection *conn = cm_id->context;
 629         struct rds_ib_connection *ic = conn->c_transport_data;
 630         struct rdma_conn_param conn_param;
 631         struct rds_ib_connect_private dp;
 632         int ret;
 633
 634         /* If the peer doesn't do protocol negotiation, we must
 635          * default to RDSv3.0 */
 636         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 637         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 638
 639         ret = rds_ib_setup_qp(conn);
 640         if (ret) {
 641                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 642                 goto out;
 643         }
 644
 645         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 646                 UINT_MAX, UINT_MAX);
 647         ret = rdma_connect(cm_id, &conn_param);
 648         if (ret)
 649                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 650
 651 out:
 652         /* Beware - returning non-zero tells the rdma_cm to destroy
 653          * the cm_id. We should certainly not do it as long as we still
 654          * "own" the cm_id. */
 655         if (ret) {
 656                 if (ic->i_cm_id == cm_id)
 657                         ret = 0;
 658         }
 659         return ret;
 660 }
 661
 662 int rds_ib_conn_connect(struct rds_connection *conn)
 663 {
 664         struct rds_ib_connection *ic = conn->c_transport_data;
 665         struct sockaddr_in src, dest;
 666         int ret;
 667
 668         /* XXX I wonder what affect the port space has */
 669         /* delegate cm event handler to rdma_transport */
 670         ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn,
 671                                      RDMA_PS_TCP, IB_QPT_RC);
 672         if (IS_ERR(ic->i_cm_id)) {
 673                 ret = PTR_ERR(ic->i_cm_id);
 674                 ic->i_cm_id = NULL;
 675                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 676                 goto out;
 677         }
 678
 679         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 680
 681         src.sin_family = AF_INET;
 682         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 683         src.sin_port = (__force u16)htons(0);
 684
 685         dest.sin_family = AF_INET;
 686         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 687         dest.sin_port = (__force u16)htons(RDS_PORT);
 688
 689         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 690                                 (struct sockaddr *)&dest,
 691                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 692         if (ret) {
 693                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 694                          ret);
 695                 rdma_destroy_id(ic->i_cm_id);
 696                 ic->i_cm_id = NULL;
 697         }
 698
 699 out:
 700         return ret;
 701 }
 702
 703 /*
 704  * This is so careful about only cleaning up resources that were built up
 705  * so that it can be called at any point during startup.  In fact it
 706  * can be called multiple times for a given connection.
 707  */
 708 void rds_ib_conn_shutdown(struct rds_connection *conn)
 709 {
 710         struct rds_ib_connection *ic = conn->c_transport_data;
 711         int err = 0;
 712
 713         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 714                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 715                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 716
 717         if (ic->i_cm_id) {
 718                 struct ib_device *dev = ic->i_cm_id->device;
 719
 720                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 721                 err = rdma_disconnect(ic->i_cm_id);
 722                 if (err) {
 723                         /* Actually this may happen quite frequently, when
 724                          * an outgoing connect raced with an incoming connect.
 725                          */
 726                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 727                                 ic->i_cm_id, err);
 728                 }
 729
 730                 /*
 731                  * We want to wait for tx and rx completion to finish
 732                  * before we tear down the connection, but we have to be
 733                  * careful not to get stuck waiting on a send ring that
 734                  * only has unsignaled sends in it.  We've shutdown new
 735                  * sends before getting here so by waiting for signaled
 736                  * sends to complete we're ensured that there will be no
 737                  * more tx processing.
 738                  */
 739                 wait_event(rds_ib_ring_empty_wait,
 740                            rds_ib_ring_empty(&ic->i_recv_ring) &&
 741                            (atomic_read(&ic->i_signaled_sends) == 0));
 742                 tasklet_kill(&ic->i_send_tasklet);
 743                 tasklet_kill(&ic->i_recv_tasklet);
 744
 745                 /* first destroy the ib state that generates callbacks */
 746                 if (ic->i_cm_id->qp)
 747                         rdma_destroy_qp(ic->i_cm_id);
 748                 if (ic->i_send_cq)
 749                         ib_destroy_cq(ic->i_send_cq);
 750                 if (ic->i_recv_cq)
 751                         ib_destroy_cq(ic->i_recv_cq);
 752
 753                 /* then free the resources that ib callbacks use */
 754                 if (ic->i_send_hdrs)
 755                         ib_dma_free_coherent(dev,
 756                                            ic->i_send_ring.w_nr *
 757                                                 sizeof(struct rds_header),
 758                                            ic->i_send_hdrs,
 759                                            ic->i_send_hdrs_dma);
 760
 761                 if (ic->i_recv_hdrs)
 762                         ib_dma_free_coherent(dev,
 763                                            ic->i_recv_ring.w_nr *
 764                                                 sizeof(struct rds_header),
 765                                            ic->i_recv_hdrs,
 766                                            ic->i_recv_hdrs_dma);
 767
 768                 if (ic->i_ack)
 769                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 770                                              ic->i_ack, ic->i_ack_dma);
 771
 772                 if (ic->i_sends)
 773                         rds_ib_send_clear_ring(ic);
 774                 if (ic->i_recvs)
 775                         rds_ib_recv_clear_ring(ic);
 776
 777                 rdma_destroy_id(ic->i_cm_id);
 778
 779                 /*
 780                  * Move connection back to the nodev list.
 781                  */
 782                 if (ic->rds_ibdev)
 783                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 784
 785                 ic->i_cm_id = NULL;
 786                 ic->i_pd = NULL;
 787                 ic->i_send_cq = NULL;
 788                 ic->i_recv_cq = NULL;
 789                 ic->i_send_hdrs = NULL;
 790                 ic->i_recv_hdrs = NULL;
 791                 ic->i_ack = NULL;
 792         }
 793         BUG_ON(ic->rds_ibdev);
 794
 795         /* Clear pending transmit */
 796         if (ic->i_data_op) {
 797                 struct rds_message *rm;
 798
 799                 rm = container_of(ic->i_data_op, struct rds_message, data);
 800                 rds_message_put(rm);
 801                 ic->i_data_op = NULL;
 802         }
 803
 804         /* Clear the ACK state */
 805         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 806 #ifdef KERNEL_HAS_ATOMIC64
 807         atomic64_set(&ic->i_ack_next, 0);
 808 #else
 809         ic->i_ack_next = 0;
 810 #endif
 811         ic->i_ack_recv = 0;
 812
 813         /* Clear flow control state */
 814         ic->i_flowctl = 0;
 815         atomic_set(&ic->i_credits, 0);
 816
 817         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 818         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 819
 820         if (ic->i_ibinc) {
 821                 rds_inc_put(&ic->i_ibinc->ii_inc);
 822                 ic->i_ibinc = NULL;
 823         }
 824
 825         vfree(ic->i_sends);
 826         ic->i_sends = NULL;
 827         vfree(ic->i_recvs);
 828         ic->i_recvs = NULL;
 829 }
 830
 831 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 832 {
 833         struct rds_ib_connection *ic;
 834         unsigned long flags;
 835         int ret;
 836
 837         /* XXX too lazy? */
 838         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
 839         if (!ic)
 840                 return -ENOMEM;
 841
 842         ret = rds_ib_recv_alloc_caches(ic);
 843         if (ret) {
 844                 kfree(ic);
 845                 return ret;
 846         }
 847
 848         INIT_LIST_HEAD(&ic->ib_node);
 849         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
 850                      (unsigned long)ic);
 851         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
 852                      (unsigned long)ic);
 853         mutex_init(&ic->i_recv_mutex);
 854 #ifndef KERNEL_HAS_ATOMIC64
 855         spin_lock_init(&ic->i_ack_lock);
 856 #endif
 857         atomic_set(&ic->i_signaled_sends, 0);
 858
 859         /*
 860          * rds_ib_conn_shutdown() waits for these to be emptied so they
 861          * must be initialized before it can be called.
 862          */
 863         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 864         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 865
 866         ic->conn = conn;
 867         conn->c_transport_data = ic;
 868
 869         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 870         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 871         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 872
 873
 874         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 875         return 0;
 876 }
 877
 878 /*
 879  * Free a connection. Connection must be shut down and not set for reconnect.
 880  */
 881 void rds_ib_conn_free(void *arg)
 882 {
 883         struct rds_ib_connection *ic = arg;
 884         spinlock_t      *lock_ptr;
 885
 886         rdsdebug("ic %p\n", ic);
 887
 888         /*
 889          * Conn is either on a dev's list or on the nodev list.
 890          * A race with shutdown() or connect() would cause problems
 891          * (since rds_ibdev would change) but that should never happen.
 892          */
 893         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 894
 895         spin_lock_irq(lock_ptr);
 896         list_del(&ic->ib_node);
 897         spin_unlock_irq(lock_ptr);
 898
 899         rds_ib_recv_free_caches(ic);
 900
 901         kfree(ic);
 902 }
 903
 904
 905 /*
 906  * An error occurred on the connection
 907  */
 908 void
 909 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 910 {
 911         va_list ap;
 912
 913         rds_conn_drop(conn);
 914
 915         va_start(ap, fmt);
 916         vprintk(fmt, ap);
 917         va_end(ap);
 918 }