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_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
 240                      struct ib_wc *wcs)
 241 {
 242         int nr, i;
 243         struct ib_wc *wc;
 244
 245         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 246                 for (i = 0; i < nr; i++) {
 247                         wc = wcs + i;
 248                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 249                                  (unsigned long long)wc->wr_id, wc->status,
 250                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 251
 252                         rds_ib_send_cqe_handler(ic, wc);
 253                 }
 254         }
 255 }
 256
 257 static void rds_ib_tasklet_fn_send(unsigned long data)
 258 {
 259         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 260         struct rds_connection *conn = ic->conn;
 261
 262         rds_ib_stats_inc(s_ib_tasklet_call);
 263
 264         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 265         ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 266         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 267
 268         if (rds_conn_up(conn) &&
 269             (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 270             test_bit(0, &conn->c_map_queued)))
 271                 rds_send_xmit(ic->conn);
 272 }
 273
 274 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
 275                      struct ib_wc *wcs,
 276                      struct rds_ib_ack_state *ack_state)
 277 {
 278         int nr, i;
 279         struct ib_wc *wc;
 280
 281         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 282                 for (i = 0; i < nr; i++) {
 283                         wc = wcs + i;
 284                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 285                                  (unsigned long long)wc->wr_id, wc->status,
 286                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 287
 288                         rds_ib_recv_cqe_handler(ic, wc, ack_state);
 289                 }
 290         }
 291 }
 292
 293 static void rds_ib_tasklet_fn_recv(unsigned long data)
 294 {
 295         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 296         struct rds_connection *conn = ic->conn;
 297         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 298         struct rds_ib_ack_state state;
 299
 300         if (!rds_ibdev)
 301                 rds_conn_drop(conn);
 302
 303         rds_ib_stats_inc(s_ib_tasklet_call);
 304
 305         memset(&state, 0, sizeof(state));
 306         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 307         ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 308         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 309
 310         if (state.ack_next_valid)
 311                 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
 312         if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
 313                 rds_send_drop_acked(conn, state.ack_recv, NULL);
 314                 ic->i_ack_recv = state.ack_recv;
 315         }
 316
 317         if (rds_conn_up(conn))
 318                 rds_ib_attempt_ack(ic);
 319 }
 320
 321 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 322 {
 323         struct rds_connection *conn = data;
 324         struct rds_ib_connection *ic = conn->c_transport_data;
 325
 326         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 327                  ib_event_msg(event->event));
 328
 329         switch (event->event) {
 330         case IB_EVENT_COMM_EST:
 331                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 332                 break;
 333         default:
 334                 rdsdebug("Fatal QP Event %u (%s) "
 335                         "- connection %pI4->%pI4, reconnecting\n",
 336                         event->event, ib_event_msg(event->event),
 337                         &conn->c_laddr, &conn->c_faddr);
 338                 rds_conn_drop(conn);
 339                 break;
 340         }
 341 }
 342
 343 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
 344 {
 345         struct rds_connection *conn = context;
 346         struct rds_ib_connection *ic = conn->c_transport_data;
 347
 348         rdsdebug("conn %p cq %p\n", conn, cq);
 349
 350         rds_ib_stats_inc(s_ib_evt_handler_call);
 351
 352         tasklet_schedule(&ic->i_send_tasklet);
 353 }
 354
 355 /*
 356  * This needs to be very careful to not leave IS_ERR pointers around for
 357  * cleanup to trip over.
 358  */
 359 static int rds_ib_setup_qp(struct rds_connection *conn)
 360 {
 361         struct rds_ib_connection *ic = conn->c_transport_data;
 362         struct ib_device *dev = ic->i_cm_id->device;
 363         struct ib_qp_init_attr attr;
 364         struct ib_cq_init_attr cq_attr = {};
 365         struct rds_ib_device *rds_ibdev;
 366         int ret, fr_queue_space;
 367
 368         /*
 369          * It's normal to see a null device if an incoming connection races
 370          * with device removal, so we don't print a warning.
 371          */
 372         rds_ibdev = rds_ib_get_client_data(dev);
 373         if (!rds_ibdev)
 374                 return -EOPNOTSUPP;
 375
 376         /* The fr_queue_space is currently set to 512, to add extra space on
 377          * completion queue and send queue. This extra space is used for FRMR
 378          * registration and invalidation work requests
 379          */
 380         fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
 381
 382         /* add the conn now so that connection establishment has the dev */
 383         rds_ib_add_conn(rds_ibdev, conn);
 384
 385         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 386                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 387         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 388                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 389
 390         /* Protection domain and memory range */
 391         ic->i_pd = rds_ibdev->pd;
 392
 393         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 394
 395         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 396                                      rds_ib_cq_event_handler, conn,
 397                                      &cq_attr);
 398         if (IS_ERR(ic->i_send_cq)) {
 399                 ret = PTR_ERR(ic->i_send_cq);
 400                 ic->i_send_cq = NULL;
 401                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 402                 goto out;
 403         }
 404
 405         cq_attr.cqe = ic->i_recv_ring.w_nr;
 406         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 407                                      rds_ib_cq_event_handler, conn,
 408                                      &cq_attr);
 409         if (IS_ERR(ic->i_recv_cq)) {
 410                 ret = PTR_ERR(ic->i_recv_cq);
 411                 ic->i_recv_cq = NULL;
 412                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 413                 goto out;
 414         }
 415
 416         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 417         if (ret) {
 418                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 419                 goto out;
 420         }
 421
 422         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 423         if (ret) {
 424                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 425                 goto out;
 426         }
 427
 428         /* XXX negotiate max send/recv with remote? */
 429         memset(&attr, 0, sizeof(attr));
 430         attr.event_handler = rds_ib_qp_event_handler;
 431         attr.qp_context = conn;
 432         /* + 1 to allow for the single ack message */
 433         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 434         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 435         attr.cap.max_send_sge = rds_ibdev->max_sge;
 436         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 437         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 438         attr.qp_type = IB_QPT_RC;
 439         attr.send_cq = ic->i_send_cq;
 440         attr.recv_cq = ic->i_recv_cq;
 441         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
 442
 443         /*
 444          * XXX this can fail if max_*_wr is too large?  Are we supposed
 445          * to back off until we get a value that the hardware can support?
 446          */
 447         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 448         if (ret) {
 449                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 450                 goto out;
 451         }
 452
 453         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 454                                            ic->i_send_ring.w_nr *
 455                                                 sizeof(struct rds_header),
 456                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 457         if (!ic->i_send_hdrs) {
 458                 ret = -ENOMEM;
 459                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 460                 goto out;
 461         }
 462
 463         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 464                                            ic->i_recv_ring.w_nr *
 465                                                 sizeof(struct rds_header),
 466                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 467         if (!ic->i_recv_hdrs) {
 468                 ret = -ENOMEM;
 469                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 470                 goto out;
 471         }
 472
 473         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 474                                        &ic->i_ack_dma, GFP_KERNEL);
 475         if (!ic->i_ack) {
 476                 ret = -ENOMEM;
 477                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 478                 goto out;
 479         }
 480
 481         ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 482                                    ibdev_to_node(dev));
 483         if (!ic->i_sends) {
 484                 ret = -ENOMEM;
 485                 rdsdebug("send allocation failed\n");
 486                 goto out;
 487         }
 488
 489         ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 490                                    ibdev_to_node(dev));
 491         if (!ic->i_recvs) {
 492                 ret = -ENOMEM;
 493                 rdsdebug("recv allocation failed\n");
 494                 goto out;
 495         }
 496
 497         rds_ib_recv_init_ack(ic);
 498
 499         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 500                  ic->i_send_cq, ic->i_recv_cq);
 501
 502 out:
 503         rds_ib_dev_put(rds_ibdev);
 504         return ret;
 505 }
 506
 507 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 508 {
 509         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 510         u16 common;
 511         u32 version = 0;
 512
 513         /*
 514          * rdma_cm private data is odd - when there is any private data in the
 515          * request, we will be given a pretty large buffer without telling us the
 516          * original size. The only way to tell the difference is by looking at
 517          * the contents, which are initialized to zero.
 518          * If the protocol version fields aren't set, this is a connection attempt
 519          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 520          * We really should have changed this for OFED 1.3 :-(
 521          */
 522
 523         /* Be paranoid. RDS always has privdata */
 524         if (!event->param.conn.private_data_len) {
 525                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 526                         "rejecting\n");
 527                 return 0;
 528         }
 529
 530         /* Even if len is crap *now* I still want to check it. -ASG */
 531         if (event->param.conn.private_data_len < sizeof (*dp) ||
 532             dp->dp_protocol_major == 0)
 533                 return RDS_PROTOCOL_3_0;
 534
 535         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 536         if (dp->dp_protocol_major == 3 && common) {
 537                 version = RDS_PROTOCOL_3_0;
 538                 while ((common >>= 1) != 0)
 539                         version++;
 540         } else
 541                 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 542                                 &dp->dp_saddr,
 543                                 dp->dp_protocol_major,
 544                                 dp->dp_protocol_minor);
 545         return version;
 546 }
 547
 548 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 549                                     struct rdma_cm_event *event)
 550 {
 551         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 552         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 553         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 554         struct rds_ib_connect_private dp_rep;
 555         struct rds_connection *conn = NULL;
 556         struct rds_ib_connection *ic = NULL;
 557         struct rdma_conn_param conn_param;
 558         u32 version;
 559         int err = 1, destroy = 1;
 560
 561         /* Check whether the remote protocol version matches ours. */
 562         version = rds_ib_protocol_compatible(event);
 563         if (!version)
 564                 goto out;
 565
 566         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 567                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 568                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 569                  (unsigned long long)be64_to_cpu(lguid),
 570                  (unsigned long long)be64_to_cpu(fguid));
 571
 572         /* RDS/IB is not currently netns aware, thus init_net */
 573         conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
 574                                &rds_ib_transport, GFP_KERNEL);
 575         if (IS_ERR(conn)) {
 576                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 577                 conn = NULL;
 578                 goto out;
 579         }
 580
 581         /*
 582          * The connection request may occur while the
 583          * previous connection exist, e.g. in case of failover.
 584          * But as connections may be initiated simultaneously
 585          * by both hosts, we have a random backoff mechanism -
 586          * see the comment above rds_queue_reconnect()
 587          */
 588         mutex_lock(&conn->c_cm_lock);
 589         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 590                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 591                         rdsdebug("incoming connect while connecting\n");
 592                         rds_conn_drop(conn);
 593                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 594                 } else
 595                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 596                         /* Wait and see - our connect may still be succeeding */
 597                         rds_ib_stats_inc(s_ib_connect_raced);
 598                 }
 599                 goto out;
 600         }
 601
 602         ic = conn->c_transport_data;
 603
 604         rds_ib_set_protocol(conn, version);
 605         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 606
 607         /* If the peer gave us the last packet it saw, process this as if
 608          * we had received a regular ACK. */
 609         if (dp->dp_ack_seq)
 610                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 611
 612         BUG_ON(cm_id->context);
 613         BUG_ON(ic->i_cm_id);
 614
 615         ic->i_cm_id = cm_id;
 616         cm_id->context = conn;
 617
 618         /* We got halfway through setting up the ib_connection, if we
 619          * fail now, we have to take the long route out of this mess. */
 620         destroy = 0;
 621
 622         err = rds_ib_setup_qp(conn);
 623         if (err) {
 624                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 625                 goto out;
 626         }
 627
 628         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 629                 event->param.conn.responder_resources,
 630                 event->param.conn.initiator_depth);
 631
 632         /* rdma_accept() calls rdma_reject() internally if it fails */
 633         err = rdma_accept(cm_id, &conn_param);
 634         if (err)
 635                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 636
 637 out:
 638         if (conn)
 639                 mutex_unlock(&conn->c_cm_lock);
 640         if (err)
 641                 rdma_reject(cm_id, NULL, 0);
 642         return destroy;
 643 }
 644
 645
 646 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 647 {
 648         struct rds_connection *conn = cm_id->context;
 649         struct rds_ib_connection *ic = conn->c_transport_data;
 650         struct rdma_conn_param conn_param;
 651         struct rds_ib_connect_private dp;
 652         int ret;
 653
 654         /* If the peer doesn't do protocol negotiation, we must
 655          * default to RDSv3.0 */
 656         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 657         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 658
 659         ret = rds_ib_setup_qp(conn);
 660         if (ret) {
 661                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 662                 goto out;
 663         }
 664
 665         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 666                 UINT_MAX, UINT_MAX);
 667         ret = rdma_connect(cm_id, &conn_param);
 668         if (ret)
 669                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 670
 671 out:
 672         /* Beware - returning non-zero tells the rdma_cm to destroy
 673          * the cm_id. We should certainly not do it as long as we still
 674          * "own" the cm_id. */
 675         if (ret) {
 676                 if (ic->i_cm_id == cm_id)
 677                         ret = 0;
 678         }
 679         return ret;
 680 }
 681
 682 int rds_ib_conn_connect(struct rds_connection *conn)
 683 {
 684         struct rds_ib_connection *ic = conn->c_transport_data;
 685         struct sockaddr_in src, dest;
 686         int ret;
 687
 688         /* XXX I wonder what affect the port space has */
 689         /* delegate cm event handler to rdma_transport */
 690         ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn,
 691                                      RDMA_PS_TCP, IB_QPT_RC);
 692         if (IS_ERR(ic->i_cm_id)) {
 693                 ret = PTR_ERR(ic->i_cm_id);
 694                 ic->i_cm_id = NULL;
 695                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 696                 goto out;
 697         }
 698
 699         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 700
 701         src.sin_family = AF_INET;
 702         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 703         src.sin_port = (__force u16)htons(0);
 704
 705         dest.sin_family = AF_INET;
 706         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 707         dest.sin_port = (__force u16)htons(RDS_PORT);
 708
 709         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 710                                 (struct sockaddr *)&dest,
 711                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 712         if (ret) {
 713                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 714                          ret);
 715                 rdma_destroy_id(ic->i_cm_id);
 716                 ic->i_cm_id = NULL;
 717         }
 718
 719 out:
 720         return ret;
 721 }
 722
 723 /*
 724  * This is so careful about only cleaning up resources that were built up
 725  * so that it can be called at any point during startup.  In fact it
 726  * can be called multiple times for a given connection.
 727  */
 728 void rds_ib_conn_shutdown(struct rds_connection *conn)
 729 {
 730         struct rds_ib_connection *ic = conn->c_transport_data;
 731         int err = 0;
 732
 733         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 734                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 735                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 736
 737         if (ic->i_cm_id) {
 738                 struct ib_device *dev = ic->i_cm_id->device;
 739
 740                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 741                 err = rdma_disconnect(ic->i_cm_id);
 742                 if (err) {
 743                         /* Actually this may happen quite frequently, when
 744                          * an outgoing connect raced with an incoming connect.
 745                          */
 746                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 747                                 ic->i_cm_id, err);
 748                 }
 749
 750                 /*
 751                  * We want to wait for tx and rx completion to finish
 752                  * before we tear down the connection, but we have to be
 753                  * careful not to get stuck waiting on a send ring that
 754                  * only has unsignaled sends in it.  We've shutdown new
 755                  * sends before getting here so by waiting for signaled
 756                  * sends to complete we're ensured that there will be no
 757                  * more tx processing.
 758                  */
 759                 wait_event(rds_ib_ring_empty_wait,
 760                            rds_ib_ring_empty(&ic->i_recv_ring) &&
 761                            (atomic_read(&ic->i_signaled_sends) == 0) &&
 762                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
 763                 tasklet_kill(&ic->i_send_tasklet);
 764                 tasklet_kill(&ic->i_recv_tasklet);
 765
 766                 /* first destroy the ib state that generates callbacks */
 767                 if (ic->i_cm_id->qp)
 768                         rdma_destroy_qp(ic->i_cm_id);
 769                 if (ic->i_send_cq)
 770                         ib_destroy_cq(ic->i_send_cq);
 771                 if (ic->i_recv_cq)
 772                         ib_destroy_cq(ic->i_recv_cq);
 773
 774                 /* then free the resources that ib callbacks use */
 775                 if (ic->i_send_hdrs)
 776                         ib_dma_free_coherent(dev,
 777                                            ic->i_send_ring.w_nr *
 778                                                 sizeof(struct rds_header),
 779                                            ic->i_send_hdrs,
 780                                            ic->i_send_hdrs_dma);
 781
 782                 if (ic->i_recv_hdrs)
 783                         ib_dma_free_coherent(dev,
 784                                            ic->i_recv_ring.w_nr *
 785                                                 sizeof(struct rds_header),
 786                                            ic->i_recv_hdrs,
 787                                            ic->i_recv_hdrs_dma);
 788
 789                 if (ic->i_ack)
 790                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 791                                              ic->i_ack, ic->i_ack_dma);
 792
 793                 if (ic->i_sends)
 794                         rds_ib_send_clear_ring(ic);
 795                 if (ic->i_recvs)
 796                         rds_ib_recv_clear_ring(ic);
 797
 798                 rdma_destroy_id(ic->i_cm_id);
 799
 800                 /*
 801                  * Move connection back to the nodev list.
 802                  */
 803                 if (ic->rds_ibdev)
 804                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 805
 806                 ic->i_cm_id = NULL;
 807                 ic->i_pd = NULL;
 808                 ic->i_send_cq = NULL;
 809                 ic->i_recv_cq = NULL;
 810                 ic->i_send_hdrs = NULL;
 811                 ic->i_recv_hdrs = NULL;
 812                 ic->i_ack = NULL;
 813         }
 814         BUG_ON(ic->rds_ibdev);
 815
 816         /* Clear pending transmit */
 817         if (ic->i_data_op) {
 818                 struct rds_message *rm;
 819
 820                 rm = container_of(ic->i_data_op, struct rds_message, data);
 821                 rds_message_put(rm);
 822                 ic->i_data_op = NULL;
 823         }
 824
 825         /* Clear the ACK state */
 826         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 827 #ifdef KERNEL_HAS_ATOMIC64
 828         atomic64_set(&ic->i_ack_next, 0);
 829 #else
 830         ic->i_ack_next = 0;
 831 #endif
 832         ic->i_ack_recv = 0;
 833
 834         /* Clear flow control state */
 835         ic->i_flowctl = 0;
 836         atomic_set(&ic->i_credits, 0);
 837
 838         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 839         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 840
 841         if (ic->i_ibinc) {
 842                 rds_inc_put(&ic->i_ibinc->ii_inc);
 843                 ic->i_ibinc = NULL;
 844         }
 845
 846         vfree(ic->i_sends);
 847         ic->i_sends = NULL;
 848         vfree(ic->i_recvs);
 849         ic->i_recvs = NULL;
 850 }
 851
 852 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 853 {
 854         struct rds_ib_connection *ic;
 855         unsigned long flags;
 856         int ret;
 857
 858         /* XXX too lazy? */
 859         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
 860         if (!ic)
 861                 return -ENOMEM;
 862
 863         ret = rds_ib_recv_alloc_caches(ic);
 864         if (ret) {
 865                 kfree(ic);
 866                 return ret;
 867         }
 868
 869         INIT_LIST_HEAD(&ic->ib_node);
 870         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
 871                      (unsigned long)ic);
 872         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
 873                      (unsigned long)ic);
 874         mutex_init(&ic->i_recv_mutex);
 875 #ifndef KERNEL_HAS_ATOMIC64
 876         spin_lock_init(&ic->i_ack_lock);
 877 #endif
 878         atomic_set(&ic->i_signaled_sends, 0);
 879
 880         /*
 881          * rds_ib_conn_shutdown() waits for these to be emptied so they
 882          * must be initialized before it can be called.
 883          */
 884         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 885         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 886
 887         ic->conn = conn;
 888         conn->c_transport_data = ic;
 889
 890         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 891         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 892         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 893
 894
 895         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 896         return 0;
 897 }
 898
 899 /*
 900  * Free a connection. Connection must be shut down and not set for reconnect.
 901  */
 902 void rds_ib_conn_free(void *arg)
 903 {
 904         struct rds_ib_connection *ic = arg;
 905         spinlock_t      *lock_ptr;
 906
 907         rdsdebug("ic %p\n", ic);
 908
 909         /*
 910          * Conn is either on a dev's list or on the nodev list.
 911          * A race with shutdown() or connect() would cause problems
 912          * (since rds_ibdev would change) but that should never happen.
 913          */
 914         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 915
 916         spin_lock_irq(lock_ptr);
 917         list_del(&ic->ib_node);
 918         spin_unlock_irq(lock_ptr);
 919
 920         rds_ib_recv_free_caches(ic);
 921
 922         kfree(ic);
 923 }
 924
 925
 926 /*
 927  * An error occurred on the connection
 928  */
 929 void
 930 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 931 {
 932         va_list ap;
 933
 934         rds_conn_drop(conn);
 935
 936         va_start(ap, fmt);
 937         vprintk(fmt, ap);
 938         va_end(ap);
 939 }