net/rds/ib_cm.c

   1 /*
   2  * Copyright (c) 2006, 2018 Oracle and/or its affiliates. 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 #include <net/addrconf.h>
  39
  40 #include "rds_single_path.h"
  41 #include "rds.h"
  42 #include "ib.h"
  43 #include "ib_mr.h"
  44
  45 /*
  46  * Set the selected protocol version
  47  */
  48 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  49 {
  50         conn->c_version = version;
  51 }
  52
  53 /*
  54  * Set up flow control
  55  */
  56 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  57 {
  58         struct rds_ib_connection *ic = conn->c_transport_data;
  59
  60         if (rds_ib_sysctl_flow_control && credits != 0) {
  61                 /* We're doing flow control */
  62                 ic->i_flowctl = 1;
  63                 rds_ib_send_add_credits(conn, credits);
  64         } else {
  65                 ic->i_flowctl = 0;
  66         }
  67 }
  68
  69 /*
  70  * Tune RNR behavior. Without flow control, we use a rather
  71  * low timeout, but not the absolute minimum - this should
  72  * be tunable.
  73  *
  74  * We already set the RNR retry count to 7 (which is the
  75  * smallest infinite number :-) above.
  76  * If flow control is off, we want to change this back to 0
  77  * so that we learn quickly when our credit accounting is
  78  * buggy.
  79  *
  80  * Caller passes in a qp_attr pointer - don't waste stack spacv
  81  * by allocation this twice.
  82  */
  83 static void
  84 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  85 {
  86         int ret;
  87
  88         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  89         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  90         if (ret)
  91                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  92 }
  93
  94 /*
  95  * Connection established.
  96  * We get here for both outgoing and incoming connection.
  97  */
  98 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  99 {
 100         struct rds_ib_connection *ic = conn->c_transport_data;
 101         const union rds_ib_conn_priv *dp = NULL;
 102         struct ib_qp_attr qp_attr;
 103         __be64 ack_seq = 0;
 104         __be32 credit = 0;
 105         u8 major = 0;
 106         u8 minor = 0;
 107         int err;
 108
 109         dp = event->param.conn.private_data;
 110         if (conn->c_isv6) {
 111                 if (event->param.conn.private_data_len >=
 112                     sizeof(struct rds6_ib_connect_private)) {
 113                         major = dp->ricp_v6.dp_protocol_major;
 114                         minor = dp->ricp_v6.dp_protocol_minor;
 115                         credit = dp->ricp_v6.dp_credit;
 116                         /* dp structure start is not guaranteed to be 8 bytes
 117                          * aligned.  Since dp_ack_seq is 64-bit extended load
 118                          * operations can be used so go through get_unaligned
 119                          * to avoid unaligned errors.
 120                          */
 121                         ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
 122                 }
 123         } else if (event->param.conn.private_data_len >=
 124                    sizeof(struct rds_ib_connect_private)) {
 125                 major = dp->ricp_v4.dp_protocol_major;
 126                 minor = dp->ricp_v4.dp_protocol_minor;
 127                 credit = dp->ricp_v4.dp_credit;
 128                 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
 129         }
 130
 131         /* make sure it isn't empty data */
 132         if (major) {
 133                 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
 134                 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
 135         }
 136
 137         if (conn->c_version < RDS_PROTOCOL_VERSION) {
 138                 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
 139                         pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
 140                                   &conn->c_laddr, &conn->c_faddr,
 141                                   RDS_PROTOCOL_MAJOR(conn->c_version),
 142                                   RDS_PROTOCOL_MINOR(conn->c_version));
 143                         rds_conn_destroy(conn);
 144                         return;
 145                 }
 146         }
 147
 148         pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
 149                   ic->i_active_side ? "Active" : "Passive",
 150                   &conn->c_laddr, &conn->c_faddr, conn->c_tos,
 151                   RDS_PROTOCOL_MAJOR(conn->c_version),
 152                   RDS_PROTOCOL_MINOR(conn->c_version),
 153                   ic->i_flowctl ? ", flow control" : "");
 154
 155         /* receive sl from the peer */
 156         ic->i_sl = ic->i_cm_id->route.path_rec->sl;
 157
 158         atomic_set(&ic->i_cq_quiesce, 0);
 159
 160         /* Init rings and fill recv. this needs to wait until protocol
 161          * negotiation is complete, since ring layout is different
 162          * from 3.1 to 4.1.
 163          */
 164         rds_ib_send_init_ring(ic);
 165         rds_ib_recv_init_ring(ic);
 166         /* Post receive buffers - as a side effect, this will update
 167          * the posted credit count. */
 168         rds_ib_recv_refill(conn, 1, GFP_KERNEL);
 169
 170         /* Tune RNR behavior */
 171         rds_ib_tune_rnr(ic, &qp_attr);
 172
 173         qp_attr.qp_state = IB_QPS_RTS;
 174         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 175         if (err)
 176                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 177
 178         /* update ib_device with this local ipaddr */
 179         err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
 180         if (err)
 181                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 182                         err);
 183
 184         /* If the peer gave us the last packet it saw, process this as if
 185          * we had received a regular ACK. */
 186         if (dp) {
 187                 if (ack_seq)
 188                         rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
 189                                             NULL);
 190         }
 191
 192         conn->c_proposed_version = conn->c_version;
 193         rds_connect_complete(conn);
 194 }
 195
 196 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 197                                       struct rdma_conn_param *conn_param,
 198                                       union rds_ib_conn_priv *dp,
 199                                       u32 protocol_version,
 200                                       u32 max_responder_resources,
 201                                       u32 max_initiator_depth,
 202                                       bool isv6)
 203 {
 204         struct rds_ib_connection *ic = conn->c_transport_data;
 205         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 206
 207         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 208
 209         conn_param->responder_resources =
 210                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 211         conn_param->initiator_depth =
 212                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 213         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 214         conn_param->rnr_retry_count = 7;
 215
 216         if (dp) {
 217                 memset(dp, 0, sizeof(*dp));
 218                 if (isv6) {
 219                         dp->ricp_v6.dp_saddr = conn->c_laddr;
 220                         dp->ricp_v6.dp_daddr = conn->c_faddr;
 221                         dp->ricp_v6.dp_protocol_major =
 222                             RDS_PROTOCOL_MAJOR(protocol_version);
 223                         dp->ricp_v6.dp_protocol_minor =
 224                             RDS_PROTOCOL_MINOR(protocol_version);
 225                         dp->ricp_v6.dp_protocol_minor_mask =
 226                             cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 227                         dp->ricp_v6.dp_ack_seq =
 228                             cpu_to_be64(rds_ib_piggyb_ack(ic));
 229                         dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
 230
 231                         conn_param->private_data = &dp->ricp_v6;
 232                         conn_param->private_data_len = sizeof(dp->ricp_v6);
 233                 } else {
 234                         dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
 235                         dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
 236                         dp->ricp_v4.dp_protocol_major =
 237                             RDS_PROTOCOL_MAJOR(protocol_version);
 238                         dp->ricp_v4.dp_protocol_minor =
 239                             RDS_PROTOCOL_MINOR(protocol_version);
 240                         dp->ricp_v4.dp_protocol_minor_mask =
 241                             cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 242                         dp->ricp_v4.dp_ack_seq =
 243                             cpu_to_be64(rds_ib_piggyb_ack(ic));
 244                         dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
 245
 246                         conn_param->private_data = &dp->ricp_v4;
 247                         conn_param->private_data_len = sizeof(dp->ricp_v4);
 248                 }
 249
 250                 /* Advertise flow control */
 251                 if (ic->i_flowctl) {
 252                         unsigned int credits;
 253
 254                         credits = IB_GET_POST_CREDITS
 255                                 (atomic_read(&ic->i_credits));
 256                         if (isv6)
 257                                 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
 258                         else
 259                                 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
 260                         atomic_sub(IB_SET_POST_CREDITS(credits),
 261                                    &ic->i_credits);
 262                 }
 263         }
 264 }
 265
 266 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 267 {
 268         rdsdebug("event %u (%s) data %p\n",
 269                  event->event, ib_event_msg(event->event), data);
 270 }
 271
 272 /* Plucking the oldest entry from the ring can be done concurrently with
 273  * the thread refilling the ring.  Each ring operation is protected by
 274  * spinlocks and the transient state of refilling doesn't change the
 275  * recording of which entry is oldest.
 276  *
 277  * This relies on IB only calling one cq comp_handler for each cq so that
 278  * there will only be one caller of rds_recv_incoming() per RDS connection.
 279  */
 280 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
 281 {
 282         struct rds_connection *conn = context;
 283         struct rds_ib_connection *ic = conn->c_transport_data;
 284
 285         rdsdebug("conn %p cq %p\n", conn, cq);
 286
 287         rds_ib_stats_inc(s_ib_evt_handler_call);
 288
 289         tasklet_schedule(&ic->i_recv_tasklet);
 290 }
 291
 292 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
 293                      struct ib_wc *wcs)
 294 {
 295         int nr, i;
 296         struct ib_wc *wc;
 297
 298         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 299                 for (i = 0; i < nr; i++) {
 300                         wc = wcs + i;
 301                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 302                                  (unsigned long long)wc->wr_id, wc->status,
 303                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 304
 305                         if (wc->wr_id <= ic->i_send_ring.w_nr ||
 306                             wc->wr_id == RDS_IB_ACK_WR_ID)
 307                                 rds_ib_send_cqe_handler(ic, wc);
 308                         else
 309                                 rds_ib_mr_cqe_handler(ic, wc);
 310
 311                 }
 312         }
 313 }
 314
 315 static void rds_ib_tasklet_fn_send(unsigned long data)
 316 {
 317         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 318         struct rds_connection *conn = ic->conn;
 319
 320         rds_ib_stats_inc(s_ib_tasklet_call);
 321
 322         /* if cq has been already reaped, ignore incoming cq event */
 323         if (atomic_read(&ic->i_cq_quiesce))
 324                 return;
 325
 326         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 327         ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 328         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 329
 330         if (rds_conn_up(conn) &&
 331             (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 332             test_bit(0, &conn->c_map_queued)))
 333                 rds_send_xmit(&ic->conn->c_path[0]);
 334 }
 335
 336 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
 337                      struct ib_wc *wcs,
 338                      struct rds_ib_ack_state *ack_state)
 339 {
 340         int nr, i;
 341         struct ib_wc *wc;
 342
 343         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 344                 for (i = 0; i < nr; i++) {
 345                         wc = wcs + i;
 346                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 347                                  (unsigned long long)wc->wr_id, wc->status,
 348                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 349
 350                         rds_ib_recv_cqe_handler(ic, wc, ack_state);
 351                 }
 352         }
 353 }
 354
 355 static void rds_ib_tasklet_fn_recv(unsigned long data)
 356 {
 357         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 358         struct rds_connection *conn = ic->conn;
 359         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 360         struct rds_ib_ack_state state;
 361
 362         if (!rds_ibdev)
 363                 rds_conn_drop(conn);
 364
 365         rds_ib_stats_inc(s_ib_tasklet_call);
 366
 367         /* if cq has been already reaped, ignore incoming cq event */
 368         if (atomic_read(&ic->i_cq_quiesce))
 369                 return;
 370
 371         memset(&state, 0, sizeof(state));
 372         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 373         ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 374         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 375
 376         if (state.ack_next_valid)
 377                 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
 378         if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
 379                 rds_send_drop_acked(conn, state.ack_recv, NULL);
 380                 ic->i_ack_recv = state.ack_recv;
 381         }
 382
 383         if (rds_conn_up(conn))
 384                 rds_ib_attempt_ack(ic);
 385 }
 386
 387 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 388 {
 389         struct rds_connection *conn = data;
 390         struct rds_ib_connection *ic = conn->c_transport_data;
 391
 392         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 393                  ib_event_msg(event->event));
 394
 395         switch (event->event) {
 396         case IB_EVENT_COMM_EST:
 397                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 398                 break;
 399         default:
 400                 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
 401                          event->event, ib_event_msg(event->event),
 402                          &conn->c_laddr, &conn->c_faddr);
 403                 rds_conn_drop(conn);
 404                 break;
 405         }
 406 }
 407
 408 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
 409 {
 410         struct rds_connection *conn = context;
 411         struct rds_ib_connection *ic = conn->c_transport_data;
 412
 413         rdsdebug("conn %p cq %p\n", conn, cq);
 414
 415         rds_ib_stats_inc(s_ib_evt_handler_call);
 416
 417         tasklet_schedule(&ic->i_send_tasklet);
 418 }
 419
 420 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
 421 {
 422         int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
 423         int index = rds_ibdev->dev->num_comp_vectors - 1;
 424         int i;
 425
 426         for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
 427                 if (rds_ibdev->vector_load[i] < min) {
 428                         index = i;
 429                         min = rds_ibdev->vector_load[i];
 430                 }
 431         }
 432
 433         rds_ibdev->vector_load[index]++;
 434         return index;
 435 }
 436
 437 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
 438 {
 439         rds_ibdev->vector_load[index]--;
 440 }
 441
 442 /*
 443  * This needs to be very careful to not leave IS_ERR pointers around for
 444  * cleanup to trip over.
 445  */
 446 static int rds_ib_setup_qp(struct rds_connection *conn)
 447 {
 448         struct rds_ib_connection *ic = conn->c_transport_data;
 449         struct ib_device *dev = ic->i_cm_id->device;
 450         struct ib_qp_init_attr attr;
 451         struct ib_cq_init_attr cq_attr = {};
 452         struct rds_ib_device *rds_ibdev;
 453         int ret, fr_queue_space;
 454
 455         /*
 456          * It's normal to see a null device if an incoming connection races
 457          * with device removal, so we don't print a warning.
 458          */
 459         rds_ibdev = rds_ib_get_client_data(dev);
 460         if (!rds_ibdev)
 461                 return -EOPNOTSUPP;
 462
 463         /* The fr_queue_space is currently set to 512, to add extra space on
 464          * completion queue and send queue. This extra space is used for FRMR
 465          * registration and invalidation work requests
 466          */
 467         fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
 468
 469         /* add the conn now so that connection establishment has the dev */
 470         rds_ib_add_conn(rds_ibdev, conn);
 471
 472         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 473                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 474         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 475                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 476
 477         /* Protection domain and memory range */
 478         ic->i_pd = rds_ibdev->pd;
 479
 480         ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
 481         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 482         cq_attr.comp_vector = ic->i_scq_vector;
 483         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 484                                      rds_ib_cq_event_handler, conn,
 485                                      &cq_attr);
 486         if (IS_ERR(ic->i_send_cq)) {
 487                 ret = PTR_ERR(ic->i_send_cq);
 488                 ic->i_send_cq = NULL;
 489                 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
 490                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 491                 goto rds_ibdev_out;
 492         }
 493
 494         ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
 495         cq_attr.cqe = ic->i_recv_ring.w_nr;
 496         cq_attr.comp_vector = ic->i_rcq_vector;
 497         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 498                                      rds_ib_cq_event_handler, conn,
 499                                      &cq_attr);
 500         if (IS_ERR(ic->i_recv_cq)) {
 501                 ret = PTR_ERR(ic->i_recv_cq);
 502                 ic->i_recv_cq = NULL;
 503                 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
 504                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 505                 goto send_cq_out;
 506         }
 507
 508         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 509         if (ret) {
 510                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 511                 goto recv_cq_out;
 512         }
 513
 514         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 515         if (ret) {
 516                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 517                 goto recv_cq_out;
 518         }
 519
 520         /* XXX negotiate max send/recv with remote? */
 521         memset(&attr, 0, sizeof(attr));
 522         attr.event_handler = rds_ib_qp_event_handler;
 523         attr.qp_context = conn;
 524         /* + 1 to allow for the single ack message */
 525         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 526         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 527         attr.cap.max_send_sge = rds_ibdev->max_sge;
 528         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 529         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 530         attr.qp_type = IB_QPT_RC;
 531         attr.send_cq = ic->i_send_cq;
 532         attr.recv_cq = ic->i_recv_cq;
 533
 534         /*
 535          * XXX this can fail if max_*_wr is too large?  Are we supposed
 536          * to back off until we get a value that the hardware can support?
 537          */
 538         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 539         if (ret) {
 540                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 541                 goto recv_cq_out;
 542         }
 543
 544         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 545                                            ic->i_send_ring.w_nr *
 546                                                 sizeof(struct rds_header),
 547                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 548         if (!ic->i_send_hdrs) {
 549                 ret = -ENOMEM;
 550                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 551                 goto qp_out;
 552         }
 553
 554         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 555                                            ic->i_recv_ring.w_nr *
 556                                                 sizeof(struct rds_header),
 557                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 558         if (!ic->i_recv_hdrs) {
 559                 ret = -ENOMEM;
 560                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 561                 goto send_hdrs_dma_out;
 562         }
 563
 564         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 565                                        &ic->i_ack_dma, GFP_KERNEL);
 566         if (!ic->i_ack) {
 567                 ret = -ENOMEM;
 568                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 569                 goto recv_hdrs_dma_out;
 570         }
 571
 572         ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
 573                                               ic->i_send_ring.w_nr),
 574                                    ibdev_to_node(dev));
 575         if (!ic->i_sends) {
 576                 ret = -ENOMEM;
 577                 rdsdebug("send allocation failed\n");
 578                 goto ack_dma_out;
 579         }
 580
 581         ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
 582                                               ic->i_recv_ring.w_nr),
 583                                    ibdev_to_node(dev));
 584         if (!ic->i_recvs) {
 585                 ret = -ENOMEM;
 586                 rdsdebug("recv allocation failed\n");
 587                 goto sends_out;
 588         }
 589
 590         rds_ib_recv_init_ack(ic);
 591
 592         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 593                  ic->i_send_cq, ic->i_recv_cq);
 594
 595         goto out;
 596
 597 sends_out:
 598         vfree(ic->i_sends);
 599 ack_dma_out:
 600         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 601                              ic->i_ack, ic->i_ack_dma);
 602 recv_hdrs_dma_out:
 603         ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr *
 604                                         sizeof(struct rds_header),
 605                                         ic->i_recv_hdrs, ic->i_recv_hdrs_dma);
 606 send_hdrs_dma_out:
 607         ib_dma_free_coherent(dev, ic->i_send_ring.w_nr *
 608                                         sizeof(struct rds_header),
 609                                         ic->i_send_hdrs, ic->i_send_hdrs_dma);
 610 qp_out:
 611         rdma_destroy_qp(ic->i_cm_id);
 612 recv_cq_out:
 613         ib_destroy_cq(ic->i_recv_cq);
 614         ic->i_recv_cq = NULL;
 615 send_cq_out:
 616         ib_destroy_cq(ic->i_send_cq);
 617         ic->i_send_cq = NULL;
 618 rds_ibdev_out:
 619         rds_ib_remove_conn(rds_ibdev, conn);
 620 out:
 621         rds_ib_dev_put(rds_ibdev);
 622
 623         return ret;
 624 }
 625
 626 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
 627 {
 628         const union rds_ib_conn_priv *dp = event->param.conn.private_data;
 629         u8 data_len, major, minor;
 630         u32 version = 0;
 631         __be16 mask;
 632         u16 common;
 633
 634         /*
 635          * rdma_cm private data is odd - when there is any private data in the
 636          * request, we will be given a pretty large buffer without telling us the
 637          * original size. The only way to tell the difference is by looking at
 638          * the contents, which are initialized to zero.
 639          * If the protocol version fields aren't set, this is a connection attempt
 640          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 641          * We really should have changed this for OFED 1.3 :-(
 642          */
 643
 644         /* Be paranoid. RDS always has privdata */
 645         if (!event->param.conn.private_data_len) {
 646                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 647                         "rejecting\n");
 648                 return 0;
 649         }
 650
 651         if (isv6) {
 652                 data_len = sizeof(struct rds6_ib_connect_private);
 653                 major = dp->ricp_v6.dp_protocol_major;
 654                 minor = dp->ricp_v6.dp_protocol_minor;
 655                 mask = dp->ricp_v6.dp_protocol_minor_mask;
 656         } else {
 657                 data_len = sizeof(struct rds_ib_connect_private);
 658                 major = dp->ricp_v4.dp_protocol_major;
 659                 minor = dp->ricp_v4.dp_protocol_minor;
 660                 mask = dp->ricp_v4.dp_protocol_minor_mask;
 661         }
 662
 663         /* Even if len is crap *now* I still want to check it. -ASG */
 664         if (event->param.conn.private_data_len < data_len || major == 0)
 665                 return RDS_PROTOCOL_4_0;
 666
 667         common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 668         if (major == 4 && common) {
 669                 version = RDS_PROTOCOL_4_0;
 670                 while ((common >>= 1) != 0)
 671                         version++;
 672         } else if (RDS_PROTOCOL_COMPAT_VERSION ==
 673                    RDS_PROTOCOL(major, minor)) {
 674                 version = RDS_PROTOCOL_COMPAT_VERSION;
 675         } else {
 676                 if (isv6)
 677                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
 678                                            &dp->ricp_v6.dp_saddr, major, minor);
 679                 else
 680                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 681                                            &dp->ricp_v4.dp_saddr, major, minor);
 682         }
 683         return version;
 684 }
 685
 686 #if IS_ENABLED(CONFIG_IPV6)
 687 /* Given an IPv6 address, find the net_device which hosts that address and
 688  * return its index.  This is used by the rds_ib_cm_handle_connect() code to
 689  * find the interface index of where an incoming request comes from when
 690  * the request is using a link local address.
 691  *
 692  * Note one problem in this search.  It is possible that two interfaces have
 693  * the same link local address.  Unfortunately, this cannot be solved unless
 694  * the underlying layer gives us the interface which an incoming RDMA connect
 695  * request comes from.
 696  */
 697 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
 698 {
 699         struct net_device *dev;
 700         int idx = 0;
 701
 702         rcu_read_lock();
 703         for_each_netdev_rcu(net, dev) {
 704                 if (ipv6_chk_addr(net, addr, dev, 1)) {
 705                         idx = dev->ifindex;
 706                         break;
 707                 }
 708         }
 709         rcu_read_unlock();
 710
 711         return idx;
 712 }
 713 #endif
 714
 715 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 716                              struct rdma_cm_event *event, bool isv6)
 717 {
 718         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 719         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 720         const struct rds_ib_conn_priv_cmn *dp_cmn;
 721         struct rds_connection *conn = NULL;
 722         struct rds_ib_connection *ic = NULL;
 723         struct rdma_conn_param conn_param;
 724         const union rds_ib_conn_priv *dp;
 725         union rds_ib_conn_priv dp_rep;
 726         struct in6_addr s_mapped_addr;
 727         struct in6_addr d_mapped_addr;
 728         const struct in6_addr *saddr6;
 729         const struct in6_addr *daddr6;
 730         int destroy = 1;
 731         u32 ifindex = 0;
 732         u32 version;
 733         int err = 1;
 734
 735         /* Check whether the remote protocol version matches ours. */
 736         version = rds_ib_protocol_compatible(event, isv6);
 737         if (!version) {
 738                 err = RDS_RDMA_REJ_INCOMPAT;
 739                 goto out;
 740         }
 741
 742         dp = event->param.conn.private_data;
 743         if (isv6) {
 744 #if IS_ENABLED(CONFIG_IPV6)
 745                 dp_cmn = &dp->ricp_v6.dp_cmn;
 746                 saddr6 = &dp->ricp_v6.dp_saddr;
 747                 daddr6 = &dp->ricp_v6.dp_daddr;
 748                 /* If either address is link local, need to find the
 749                  * interface index in order to create a proper RDS
 750                  * connection.
 751                  */
 752                 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
 753                         /* Using init_net for now ..  */
 754                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 755                         /* No index found...  Need to bail out. */
 756                         if (ifindex == 0) {
 757                                 err = -EOPNOTSUPP;
 758                                 goto out;
 759                         }
 760                 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
 761                         /* Use our address to find the correct index. */
 762                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 763                         /* No index found...  Need to bail out. */
 764                         if (ifindex == 0) {
 765                                 err = -EOPNOTSUPP;
 766                                 goto out;
 767                         }
 768                 }
 769 #else
 770                 err = -EOPNOTSUPP;
 771                 goto out;
 772 #endif
 773         } else {
 774                 dp_cmn = &dp->ricp_v4.dp_cmn;
 775                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
 776                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
 777                 saddr6 = &s_mapped_addr;
 778                 daddr6 = &d_mapped_addr;
 779         }
 780
 781         rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
 782                  saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
 783                  RDS_PROTOCOL_MINOR(version),
 784                  (unsigned long long)be64_to_cpu(lguid),
 785                  (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
 786
 787         /* RDS/IB is not currently netns aware, thus init_net */
 788         conn = rds_conn_create(&init_net, daddr6, saddr6,
 789                                &rds_ib_transport, dp_cmn->ricpc_dp_toss,
 790                                GFP_KERNEL, ifindex);
 791         if (IS_ERR(conn)) {
 792                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 793                 conn = NULL;
 794                 goto out;
 795         }
 796
 797         /*
 798          * The connection request may occur while the
 799          * previous connection exist, e.g. in case of failover.
 800          * But as connections may be initiated simultaneously
 801          * by both hosts, we have a random backoff mechanism -
 802          * see the comment above rds_queue_reconnect()
 803          */
 804         mutex_lock(&conn->c_cm_lock);
 805         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 806                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 807                         rdsdebug("incoming connect while connecting\n");
 808                         rds_conn_drop(conn);
 809                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 810                 } else
 811                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 812                         /* Wait and see - our connect may still be succeeding */
 813                         rds_ib_stats_inc(s_ib_connect_raced);
 814                 }
 815                 goto out;
 816         }
 817
 818         ic = conn->c_transport_data;
 819
 820         rds_ib_set_protocol(conn, version);
 821         rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
 822
 823         /* If the peer gave us the last packet it saw, process this as if
 824          * we had received a regular ACK. */
 825         if (dp_cmn->ricpc_ack_seq)
 826                 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
 827                                     NULL);
 828
 829         BUG_ON(cm_id->context);
 830         BUG_ON(ic->i_cm_id);
 831
 832         ic->i_cm_id = cm_id;
 833         cm_id->context = conn;
 834
 835         /* We got halfway through setting up the ib_connection, if we
 836          * fail now, we have to take the long route out of this mess. */
 837         destroy = 0;
 838
 839         err = rds_ib_setup_qp(conn);
 840         if (err) {
 841                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 842                 goto out;
 843         }
 844
 845         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 846                                   event->param.conn.responder_resources,
 847                                   event->param.conn.initiator_depth, isv6);
 848
 849         /* rdma_accept() calls rdma_reject() internally if it fails */
 850         if (rdma_accept(cm_id, &conn_param))
 851                 rds_ib_conn_error(conn, "rdma_accept failed\n");
 852
 853 out:
 854         if (conn)
 855                 mutex_unlock(&conn->c_cm_lock);
 856         if (err)
 857                 rdma_reject(cm_id, &err, sizeof(int));
 858         return destroy;
 859 }
 860
 861
 862 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
 863 {
 864         struct rds_connection *conn = cm_id->context;
 865         struct rds_ib_connection *ic = conn->c_transport_data;
 866         struct rdma_conn_param conn_param;
 867         union rds_ib_conn_priv dp;
 868         int ret;
 869
 870         /* If the peer doesn't do protocol negotiation, we must
 871          * default to RDSv3.0 */
 872         rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
 873         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 874
 875         ret = rds_ib_setup_qp(conn);
 876         if (ret) {
 877                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 878                 goto out;
 879         }
 880
 881         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
 882                                   conn->c_proposed_version,
 883                                   UINT_MAX, UINT_MAX, isv6);
 884         ret = rdma_connect(cm_id, &conn_param);
 885         if (ret)
 886                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 887
 888 out:
 889         /* Beware - returning non-zero tells the rdma_cm to destroy
 890          * the cm_id. We should certainly not do it as long as we still
 891          * "own" the cm_id. */
 892         if (ret) {
 893                 if (ic->i_cm_id == cm_id)
 894                         ret = 0;
 895         }
 896         ic->i_active_side = true;
 897         return ret;
 898 }
 899
 900 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
 901 {
 902         struct rds_connection *conn = cp->cp_conn;
 903         struct sockaddr_storage src, dest;
 904         rdma_cm_event_handler handler;
 905         struct rds_ib_connection *ic;
 906         int ret;
 907
 908         ic = conn->c_transport_data;
 909
 910         /* XXX I wonder what affect the port space has */
 911         /* delegate cm event handler to rdma_transport */
 912 #if IS_ENABLED(CONFIG_IPV6)
 913         if (conn->c_isv6)
 914                 handler = rds6_rdma_cm_event_handler;
 915         else
 916 #endif
 917                 handler = rds_rdma_cm_event_handler;
 918         ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
 919                                      RDMA_PS_TCP, IB_QPT_RC);
 920         if (IS_ERR(ic->i_cm_id)) {
 921                 ret = PTR_ERR(ic->i_cm_id);
 922                 ic->i_cm_id = NULL;
 923                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 924                 goto out;
 925         }
 926
 927         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 928
 929         if (ipv6_addr_v4mapped(&conn->c_faddr)) {
 930                 struct sockaddr_in *sin;
 931
 932                 sin = (struct sockaddr_in *)&src;
 933                 sin->sin_family = AF_INET;
 934                 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
 935                 sin->sin_port = 0;
 936
 937                 sin = (struct sockaddr_in *)&dest;
 938                 sin->sin_family = AF_INET;
 939                 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
 940                 sin->sin_port = htons(RDS_PORT);
 941         } else {
 942                 struct sockaddr_in6 *sin6;
 943
 944                 sin6 = (struct sockaddr_in6 *)&src;
 945                 sin6->sin6_family = AF_INET6;
 946                 sin6->sin6_addr = conn->c_laddr;
 947                 sin6->sin6_port = 0;
 948                 sin6->sin6_scope_id = conn->c_dev_if;
 949
 950                 sin6 = (struct sockaddr_in6 *)&dest;
 951                 sin6->sin6_family = AF_INET6;
 952                 sin6->sin6_addr = conn->c_faddr;
 953                 sin6->sin6_port = htons(RDS_CM_PORT);
 954                 sin6->sin6_scope_id = conn->c_dev_if;
 955         }
 956
 957         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 958                                 (struct sockaddr *)&dest,
 959                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 960         if (ret) {
 961                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 962                          ret);
 963                 rdma_destroy_id(ic->i_cm_id);
 964                 ic->i_cm_id = NULL;
 965         }
 966
 967 out:
 968         return ret;
 969 }
 970
 971 /*
 972  * This is so careful about only cleaning up resources that were built up
 973  * so that it can be called at any point during startup.  In fact it
 974  * can be called multiple times for a given connection.
 975  */
 976 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
 977 {
 978         struct rds_connection *conn = cp->cp_conn;
 979         struct rds_ib_connection *ic = conn->c_transport_data;
 980         int err = 0;
 981
 982         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 983                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 984                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 985
 986         if (ic->i_cm_id) {
 987                 struct ib_device *dev = ic->i_cm_id->device;
 988
 989                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 990                 err = rdma_disconnect(ic->i_cm_id);
 991                 if (err) {
 992                         /* Actually this may happen quite frequently, when
 993                          * an outgoing connect raced with an incoming connect.
 994                          */
 995                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 996                                 ic->i_cm_id, err);
 997                 }
 998
 999                 /* kick off "flush_worker" for all pools in order to reap
1000                  * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1001                  */
1002                 rds_ib_flush_mrs();
1003
1004                 /*
1005                  * We want to wait for tx and rx completion to finish
1006                  * before we tear down the connection, but we have to be
1007                  * careful not to get stuck waiting on a send ring that
1008                  * only has unsignaled sends in it.  We've shutdown new
1009                  * sends before getting here so by waiting for signaled
1010                  * sends to complete we're ensured that there will be no
1011                  * more tx processing.
1012                  */
1013                 wait_event(rds_ib_ring_empty_wait,
1014                            rds_ib_ring_empty(&ic->i_recv_ring) &&
1015                            (atomic_read(&ic->i_signaled_sends) == 0) &&
1016                            (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1017                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1018                 tasklet_kill(&ic->i_send_tasklet);
1019                 tasklet_kill(&ic->i_recv_tasklet);
1020
1021                 atomic_set(&ic->i_cq_quiesce, 1);
1022
1023                 /* first destroy the ib state that generates callbacks */
1024                 if (ic->i_cm_id->qp)
1025                         rdma_destroy_qp(ic->i_cm_id);
1026                 if (ic->i_send_cq) {
1027                         if (ic->rds_ibdev)
1028                                 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1029                         ib_destroy_cq(ic->i_send_cq);
1030                 }
1031
1032                 if (ic->i_recv_cq) {
1033                         if (ic->rds_ibdev)
1034                                 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1035                         ib_destroy_cq(ic->i_recv_cq);
1036                 }
1037
1038                 /* then free the resources that ib callbacks use */
1039                 if (ic->i_send_hdrs)
1040                         ib_dma_free_coherent(dev,
1041                                            ic->i_send_ring.w_nr *
1042                                                 sizeof(struct rds_header),
1043                                            ic->i_send_hdrs,
1044                                            ic->i_send_hdrs_dma);
1045
1046                 if (ic->i_recv_hdrs)
1047                         ib_dma_free_coherent(dev,
1048                                            ic->i_recv_ring.w_nr *
1049                                                 sizeof(struct rds_header),
1050                                            ic->i_recv_hdrs,
1051                                            ic->i_recv_hdrs_dma);
1052
1053                 if (ic->i_ack)
1054                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
1055                                              ic->i_ack, ic->i_ack_dma);
1056
1057                 if (ic->i_sends)
1058                         rds_ib_send_clear_ring(ic);
1059                 if (ic->i_recvs)
1060                         rds_ib_recv_clear_ring(ic);
1061
1062                 rdma_destroy_id(ic->i_cm_id);
1063
1064                 /*
1065                  * Move connection back to the nodev list.
1066                  */
1067                 if (ic->rds_ibdev)
1068                         rds_ib_remove_conn(ic->rds_ibdev, conn);
1069
1070                 ic->i_cm_id = NULL;
1071                 ic->i_pd = NULL;
1072                 ic->i_send_cq = NULL;
1073                 ic->i_recv_cq = NULL;
1074                 ic->i_send_hdrs = NULL;
1075                 ic->i_recv_hdrs = NULL;
1076                 ic->i_ack = NULL;
1077         }
1078         BUG_ON(ic->rds_ibdev);
1079
1080         /* Clear pending transmit */
1081         if (ic->i_data_op) {
1082                 struct rds_message *rm;
1083
1084                 rm = container_of(ic->i_data_op, struct rds_message, data);
1085                 rds_message_put(rm);
1086                 ic->i_data_op = NULL;
1087         }
1088
1089         /* Clear the ACK state */
1090         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1091 #ifdef KERNEL_HAS_ATOMIC64
1092         atomic64_set(&ic->i_ack_next, 0);
1093 #else
1094         ic->i_ack_next = 0;
1095 #endif
1096         ic->i_ack_recv = 0;
1097
1098         /* Clear flow control state */
1099         ic->i_flowctl = 0;
1100         atomic_set(&ic->i_credits, 0);
1101
1102         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1103         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1104
1105         if (ic->i_ibinc) {
1106                 rds_inc_put(&ic->i_ibinc->ii_inc);
1107                 ic->i_ibinc = NULL;
1108         }
1109
1110         vfree(ic->i_sends);
1111         ic->i_sends = NULL;
1112         vfree(ic->i_recvs);
1113         ic->i_recvs = NULL;
1114         ic->i_active_side = false;
1115 }
1116
1117 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1118 {
1119         struct rds_ib_connection *ic;
1120         unsigned long flags;
1121         int ret;
1122
1123         /* XXX too lazy? */
1124         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1125         if (!ic)
1126                 return -ENOMEM;
1127
1128         ret = rds_ib_recv_alloc_caches(ic, gfp);
1129         if (ret) {
1130                 kfree(ic);
1131                 return ret;
1132         }
1133
1134         INIT_LIST_HEAD(&ic->ib_node);
1135         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1136                      (unsigned long)ic);
1137         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1138                      (unsigned long)ic);
1139         mutex_init(&ic->i_recv_mutex);
1140 #ifndef KERNEL_HAS_ATOMIC64
1141         spin_lock_init(&ic->i_ack_lock);
1142 #endif
1143         atomic_set(&ic->i_signaled_sends, 0);
1144         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1145
1146         /*
1147          * rds_ib_conn_shutdown() waits for these to be emptied so they
1148          * must be initialized before it can be called.
1149          */
1150         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1151         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1152
1153         ic->conn = conn;
1154         conn->c_transport_data = ic;
1155
1156         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1157         list_add_tail(&ic->ib_node, &ib_nodev_conns);
1158         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1159
1160
1161         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1162         return 0;
1163 }
1164
1165 /*
1166  * Free a connection. Connection must be shut down and not set for reconnect.
1167  */
1168 void rds_ib_conn_free(void *arg)
1169 {
1170         struct rds_ib_connection *ic = arg;
1171         spinlock_t      *lock_ptr;
1172
1173         rdsdebug("ic %p\n", ic);
1174
1175         /*
1176          * Conn is either on a dev's list or on the nodev list.
1177          * A race with shutdown() or connect() would cause problems
1178          * (since rds_ibdev would change) but that should never happen.
1179          */
1180         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1181
1182         spin_lock_irq(lock_ptr);
1183         list_del(&ic->ib_node);
1184         spin_unlock_irq(lock_ptr);
1185
1186         rds_ib_recv_free_caches(ic);
1187
1188         kfree(ic);
1189 }
1190
1191
1192 /*
1193  * An error occurred on the connection
1194  */
1195 void
1196 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1197 {
1198         va_list ap;
1199
1200         rds_conn_drop(conn);
1201
1202         va_start(ap, fmt);
1203         vprintk(fmt, ap);
1204         va_end(ap);
1205 }