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
  44 /*
  45  * Set the selected protocol version
  46  */
  47 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  48 {
  49         conn->c_version = version;
  50 }
  51
  52 /*
  53  * Set up flow control
  54  */
  55 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  56 {
  57         struct rds_ib_connection *ic = conn->c_transport_data;
  58
  59         if (rds_ib_sysctl_flow_control && credits != 0) {
  60                 /* We're doing flow control */
  61                 ic->i_flowctl = 1;
  62                 rds_ib_send_add_credits(conn, credits);
  63         } else {
  64                 ic->i_flowctl = 0;
  65         }
  66 }
  67
  68 /*
  69  * Tune RNR behavior. Without flow control, we use a rather
  70  * low timeout, but not the absolute minimum - this should
  71  * be tunable.
  72  *
  73  * We already set the RNR retry count to 7 (which is the
  74  * smallest infinite number :-) above.
  75  * If flow control is off, we want to change this back to 0
  76  * so that we learn quickly when our credit accounting is
  77  * buggy.
  78  *
  79  * Caller passes in a qp_attr pointer - don't waste stack spacv
  80  * by allocation this twice.
  81  */
  82 static void
  83 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  84 {
  85         int ret;
  86
  87         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  88         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  89         if (ret)
  90                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  91 }
  92
  93 /*
  94  * Connection established.
  95  * We get here for both outgoing and incoming connection.
  96  */
  97 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  98 {
  99         struct rds_ib_connection *ic = conn->c_transport_data;
 100         const union rds_ib_conn_priv *dp = NULL;
 101         struct ib_qp_attr qp_attr;
 102         __be64 ack_seq = 0;
 103         __be32 credit = 0;
 104         u8 major = 0;
 105         u8 minor = 0;
 106         int err;
 107
 108         dp = event->param.conn.private_data;
 109         if (conn->c_isv6) {
 110                 if (event->param.conn.private_data_len >=
 111                     sizeof(struct rds6_ib_connect_private)) {
 112                         major = dp->ricp_v6.dp_protocol_major;
 113                         minor = dp->ricp_v6.dp_protocol_minor;
 114                         credit = dp->ricp_v6.dp_credit;
 115                         /* dp structure start is not guaranteed to be 8 bytes
 116                          * aligned.  Since dp_ack_seq is 64-bit extended load
 117                          * operations can be used so go through get_unaligned
 118                          * to avoid unaligned errors.
 119                          */
 120                         ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
 121                 }
 122         } else if (event->param.conn.private_data_len >=
 123                    sizeof(struct rds_ib_connect_private)) {
 124                 major = dp->ricp_v4.dp_protocol_major;
 125                 minor = dp->ricp_v4.dp_protocol_minor;
 126                 credit = dp->ricp_v4.dp_credit;
 127                 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
 128         }
 129
 130         /* make sure it isn't empty data */
 131         if (major) {
 132                 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
 133                 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
 134         }
 135
 136         if (conn->c_version < RDS_PROTOCOL(3, 1)) {
 137                 pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
 138                           &conn->c_laddr, &conn->c_faddr,
 139                           RDS_PROTOCOL_MAJOR(conn->c_version),
 140                           RDS_PROTOCOL_MINOR(conn->c_version));
 141                 set_bit(RDS_DESTROY_PENDING, &conn->c_path[0].cp_flags);
 142                 rds_conn_destroy(conn);
 143                 return;
 144         } else {
 145                 pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c> version %u.%u%s\n",
 146                           ic->i_active_side ? "Active" : "Passive",
 147                           &conn->c_laddr, &conn->c_faddr,
 148                           RDS_PROTOCOL_MAJOR(conn->c_version),
 149                           RDS_PROTOCOL_MINOR(conn->c_version),
 150                           ic->i_flowctl ? ", flow control" : "");
 151         }
 152
 153         atomic_set(&ic->i_cq_quiesce, 0);
 154
 155         /* Init rings and fill recv. this needs to wait until protocol
 156          * negotiation is complete, since ring layout is different
 157          * from 3.1 to 4.1.
 158          */
 159         rds_ib_send_init_ring(ic);
 160         rds_ib_recv_init_ring(ic);
 161         /* Post receive buffers - as a side effect, this will update
 162          * the posted credit count. */
 163         rds_ib_recv_refill(conn, 1, GFP_KERNEL);
 164
 165         /* Tune RNR behavior */
 166         rds_ib_tune_rnr(ic, &qp_attr);
 167
 168         qp_attr.qp_state = IB_QPS_RTS;
 169         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 170         if (err)
 171                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 172
 173         /* update ib_device with this local ipaddr */
 174         err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
 175         if (err)
 176                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 177                         err);
 178
 179         /* If the peer gave us the last packet it saw, process this as if
 180          * we had received a regular ACK. */
 181         if (dp) {
 182                 if (ack_seq)
 183                         rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
 184                                             NULL);
 185         }
 186
 187         rds_connect_complete(conn);
 188 }
 189
 190 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 191                                       struct rdma_conn_param *conn_param,
 192                                       union rds_ib_conn_priv *dp,
 193                                       u32 protocol_version,
 194                                       u32 max_responder_resources,
 195                                       u32 max_initiator_depth,
 196                                       bool isv6)
 197 {
 198         struct rds_ib_connection *ic = conn->c_transport_data;
 199         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 200
 201         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 202
 203         conn_param->responder_resources =
 204                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 205         conn_param->initiator_depth =
 206                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 207         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 208         conn_param->rnr_retry_count = 7;
 209
 210         if (dp) {
 211                 memset(dp, 0, sizeof(*dp));
 212                 if (isv6) {
 213                         dp->ricp_v6.dp_saddr = conn->c_laddr;
 214                         dp->ricp_v6.dp_daddr = conn->c_faddr;
 215                         dp->ricp_v6.dp_protocol_major =
 216                             RDS_PROTOCOL_MAJOR(protocol_version);
 217                         dp->ricp_v6.dp_protocol_minor =
 218                             RDS_PROTOCOL_MINOR(protocol_version);
 219                         dp->ricp_v6.dp_protocol_minor_mask =
 220                             cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 221                         dp->ricp_v6.dp_ack_seq =
 222                             cpu_to_be64(rds_ib_piggyb_ack(ic));
 223
 224                         conn_param->private_data = &dp->ricp_v6;
 225                         conn_param->private_data_len = sizeof(dp->ricp_v6);
 226                 } else {
 227                         dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
 228                         dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
 229                         dp->ricp_v4.dp_protocol_major =
 230                             RDS_PROTOCOL_MAJOR(protocol_version);
 231                         dp->ricp_v4.dp_protocol_minor =
 232                             RDS_PROTOCOL_MINOR(protocol_version);
 233                         dp->ricp_v4.dp_protocol_minor_mask =
 234                             cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 235                         dp->ricp_v4.dp_ack_seq =
 236                             cpu_to_be64(rds_ib_piggyb_ack(ic));
 237
 238                         conn_param->private_data = &dp->ricp_v4;
 239                         conn_param->private_data_len = sizeof(dp->ricp_v4);
 240                 }
 241
 242                 /* Advertise flow control */
 243                 if (ic->i_flowctl) {
 244                         unsigned int credits;
 245
 246                         credits = IB_GET_POST_CREDITS
 247                                 (atomic_read(&ic->i_credits));
 248                         if (isv6)
 249                                 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
 250                         else
 251                                 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
 252                         atomic_sub(IB_SET_POST_CREDITS(credits),
 253                                    &ic->i_credits);
 254                 }
 255         }
 256 }
 257
 258 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 259 {
 260         rdsdebug("event %u (%s) data %p\n",
 261                  event->event, ib_event_msg(event->event), data);
 262 }
 263
 264 /* Plucking the oldest entry from the ring can be done concurrently with
 265  * the thread refilling the ring.  Each ring operation is protected by
 266  * spinlocks and the transient state of refilling doesn't change the
 267  * recording of which entry is oldest.
 268  *
 269  * This relies on IB only calling one cq comp_handler for each cq so that
 270  * there will only be one caller of rds_recv_incoming() per RDS connection.
 271  */
 272 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
 273 {
 274         struct rds_connection *conn = context;
 275         struct rds_ib_connection *ic = conn->c_transport_data;
 276
 277         rdsdebug("conn %p cq %p\n", conn, cq);
 278
 279         rds_ib_stats_inc(s_ib_evt_handler_call);
 280
 281         tasklet_schedule(&ic->i_recv_tasklet);
 282 }
 283
 284 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
 285                      struct ib_wc *wcs)
 286 {
 287         int nr, i;
 288         struct ib_wc *wc;
 289
 290         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 291                 for (i = 0; i < nr; i++) {
 292                         wc = wcs + i;
 293                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 294                                  (unsigned long long)wc->wr_id, wc->status,
 295                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 296
 297                         if (wc->wr_id <= ic->i_send_ring.w_nr ||
 298                             wc->wr_id == RDS_IB_ACK_WR_ID)
 299                                 rds_ib_send_cqe_handler(ic, wc);
 300                         else
 301                                 rds_ib_mr_cqe_handler(ic, wc);
 302
 303                 }
 304         }
 305 }
 306
 307 static void rds_ib_tasklet_fn_send(unsigned long data)
 308 {
 309         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 310         struct rds_connection *conn = ic->conn;
 311
 312         rds_ib_stats_inc(s_ib_tasklet_call);
 313
 314         /* if cq has been already reaped, ignore incoming cq event */
 315         if (atomic_read(&ic->i_cq_quiesce))
 316                 return;
 317
 318         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 319         ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 320         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 321
 322         if (rds_conn_up(conn) &&
 323             (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 324             test_bit(0, &conn->c_map_queued)))
 325                 rds_send_xmit(&ic->conn->c_path[0]);
 326 }
 327
 328 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
 329                      struct ib_wc *wcs,
 330                      struct rds_ib_ack_state *ack_state)
 331 {
 332         int nr, i;
 333         struct ib_wc *wc;
 334
 335         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 336                 for (i = 0; i < nr; i++) {
 337                         wc = wcs + i;
 338                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 339                                  (unsigned long long)wc->wr_id, wc->status,
 340                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 341
 342                         rds_ib_recv_cqe_handler(ic, wc, ack_state);
 343                 }
 344         }
 345 }
 346
 347 static void rds_ib_tasklet_fn_recv(unsigned long data)
 348 {
 349         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 350         struct rds_connection *conn = ic->conn;
 351         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 352         struct rds_ib_ack_state state;
 353
 354         if (!rds_ibdev)
 355                 rds_conn_drop(conn);
 356
 357         rds_ib_stats_inc(s_ib_tasklet_call);
 358
 359         /* if cq has been already reaped, ignore incoming cq event */
 360         if (atomic_read(&ic->i_cq_quiesce))
 361                 return;
 362
 363         memset(&state, 0, sizeof(state));
 364         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 365         ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 366         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 367
 368         if (state.ack_next_valid)
 369                 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
 370         if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
 371                 rds_send_drop_acked(conn, state.ack_recv, NULL);
 372                 ic->i_ack_recv = state.ack_recv;
 373         }
 374
 375         if (rds_conn_up(conn))
 376                 rds_ib_attempt_ack(ic);
 377 }
 378
 379 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 380 {
 381         struct rds_connection *conn = data;
 382         struct rds_ib_connection *ic = conn->c_transport_data;
 383
 384         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 385                  ib_event_msg(event->event));
 386
 387         switch (event->event) {
 388         case IB_EVENT_COMM_EST:
 389                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 390                 break;
 391         default:
 392                 rdsdebug("Fatal QP Event %u (%s) "
 393                         "- connection %pI6c->%pI6c, reconnecting\n",
 394                         event->event, ib_event_msg(event->event),
 395                         &conn->c_laddr, &conn->c_faddr);
 396                 rds_conn_drop(conn);
 397                 break;
 398         }
 399 }
 400
 401 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
 402 {
 403         struct rds_connection *conn = context;
 404         struct rds_ib_connection *ic = conn->c_transport_data;
 405
 406         rdsdebug("conn %p cq %p\n", conn, cq);
 407
 408         rds_ib_stats_inc(s_ib_evt_handler_call);
 409
 410         tasklet_schedule(&ic->i_send_tasklet);
 411 }
 412
 413 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
 414 {
 415         int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
 416         int index = rds_ibdev->dev->num_comp_vectors - 1;
 417         int i;
 418
 419         for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
 420                 if (rds_ibdev->vector_load[i] < min) {
 421                         index = i;
 422                         min = rds_ibdev->vector_load[i];
 423                 }
 424         }
 425
 426         rds_ibdev->vector_load[index]++;
 427         return index;
 428 }
 429
 430 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
 431 {
 432         rds_ibdev->vector_load[index]--;
 433 }
 434
 435 /*
 436  * This needs to be very careful to not leave IS_ERR pointers around for
 437  * cleanup to trip over.
 438  */
 439 static int rds_ib_setup_qp(struct rds_connection *conn)
 440 {
 441         struct rds_ib_connection *ic = conn->c_transport_data;
 442         struct ib_device *dev = ic->i_cm_id->device;
 443         struct ib_qp_init_attr attr;
 444         struct ib_cq_init_attr cq_attr = {};
 445         struct rds_ib_device *rds_ibdev;
 446         int ret, fr_queue_space;
 447
 448         /*
 449          * It's normal to see a null device if an incoming connection races
 450          * with device removal, so we don't print a warning.
 451          */
 452         rds_ibdev = rds_ib_get_client_data(dev);
 453         if (!rds_ibdev)
 454                 return -EOPNOTSUPP;
 455
 456         /* The fr_queue_space is currently set to 512, to add extra space on
 457          * completion queue and send queue. This extra space is used for FRMR
 458          * registration and invalidation work requests
 459          */
 460         fr_queue_space = rds_ibdev->use_fastreg ?
 461                          (RDS_IB_DEFAULT_FR_WR + 1) +
 462                          (RDS_IB_DEFAULT_FR_INV_WR + 1)
 463                          : 0;
 464
 465         /* add the conn now so that connection establishment has the dev */
 466         rds_ib_add_conn(rds_ibdev, conn);
 467
 468         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 469                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 470         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 471                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 472
 473         /* Protection domain and memory range */
 474         ic->i_pd = rds_ibdev->pd;
 475
 476         ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
 477         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 478         cq_attr.comp_vector = ic->i_scq_vector;
 479         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 480                                      rds_ib_cq_event_handler, conn,
 481                                      &cq_attr);
 482         if (IS_ERR(ic->i_send_cq)) {
 483                 ret = PTR_ERR(ic->i_send_cq);
 484                 ic->i_send_cq = NULL;
 485                 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
 486                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 487                 goto rds_ibdev_out;
 488         }
 489
 490         ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
 491         cq_attr.cqe = ic->i_recv_ring.w_nr;
 492         cq_attr.comp_vector = ic->i_rcq_vector;
 493         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 494                                      rds_ib_cq_event_handler, conn,
 495                                      &cq_attr);
 496         if (IS_ERR(ic->i_recv_cq)) {
 497                 ret = PTR_ERR(ic->i_recv_cq);
 498                 ic->i_recv_cq = NULL;
 499                 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
 500                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 501                 goto send_cq_out;
 502         }
 503
 504         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 505         if (ret) {
 506                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 507                 goto recv_cq_out;
 508         }
 509
 510         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 511         if (ret) {
 512                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 513                 goto recv_cq_out;
 514         }
 515
 516         /* XXX negotiate max send/recv with remote? */
 517         memset(&attr, 0, sizeof(attr));
 518         attr.event_handler = rds_ib_qp_event_handler;
 519         attr.qp_context = conn;
 520         /* + 1 to allow for the single ack message */
 521         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 522         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 523         attr.cap.max_send_sge = rds_ibdev->max_sge;
 524         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 525         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 526         attr.qp_type = IB_QPT_RC;
 527         attr.send_cq = ic->i_send_cq;
 528         attr.recv_cq = ic->i_recv_cq;
 529         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
 530         atomic_set(&ic->i_fastunreg_wrs, RDS_IB_DEFAULT_FR_INV_WR);
 531
 532         /*
 533          * XXX this can fail if max_*_wr is too large?  Are we supposed
 534          * to back off until we get a value that the hardware can support?
 535          */
 536         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 537         if (ret) {
 538                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 539                 goto recv_cq_out;
 540         }
 541
 542         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 543                                            ic->i_send_ring.w_nr *
 544                                                 sizeof(struct rds_header),
 545                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 546         if (!ic->i_send_hdrs) {
 547                 ret = -ENOMEM;
 548                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 549                 goto qp_out;
 550         }
 551
 552         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 553                                            ic->i_recv_ring.w_nr *
 554                                                 sizeof(struct rds_header),
 555                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 556         if (!ic->i_recv_hdrs) {
 557                 ret = -ENOMEM;
 558                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 559                 goto send_hdrs_dma_out;
 560         }
 561
 562         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 563                                        &ic->i_ack_dma, GFP_KERNEL);
 564         if (!ic->i_ack) {
 565                 ret = -ENOMEM;
 566                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 567                 goto recv_hdrs_dma_out;
 568         }
 569
 570         ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
 571                                               ic->i_send_ring.w_nr),
 572                                    ibdev_to_node(dev));
 573         if (!ic->i_sends) {
 574                 ret = -ENOMEM;
 575                 rdsdebug("send allocation failed\n");
 576                 goto ack_dma_out;
 577         }
 578
 579         ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
 580                                               ic->i_recv_ring.w_nr),
 581                                    ibdev_to_node(dev));
 582         if (!ic->i_recvs) {
 583                 ret = -ENOMEM;
 584                 rdsdebug("recv allocation failed\n");
 585                 goto sends_out;
 586         }
 587
 588         rds_ib_recv_init_ack(ic);
 589
 590         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 591                  ic->i_send_cq, ic->i_recv_cq);
 592
 593         goto out;
 594
 595 sends_out:
 596         vfree(ic->i_sends);
 597 ack_dma_out:
 598         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 599                              ic->i_ack, ic->i_ack_dma);
 600 recv_hdrs_dma_out:
 601         ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr *
 602                                         sizeof(struct rds_header),
 603                                         ic->i_recv_hdrs, ic->i_recv_hdrs_dma);
 604 send_hdrs_dma_out:
 605         ib_dma_free_coherent(dev, ic->i_send_ring.w_nr *
 606                                         sizeof(struct rds_header),
 607                                         ic->i_send_hdrs, ic->i_send_hdrs_dma);
 608 qp_out:
 609         rdma_destroy_qp(ic->i_cm_id);
 610 recv_cq_out:
 611         if (!ib_destroy_cq(ic->i_recv_cq))
 612                 ic->i_recv_cq = NULL;
 613 send_cq_out:
 614         if (!ib_destroy_cq(ic->i_send_cq))
 615                 ic->i_send_cq = NULL;
 616 rds_ibdev_out:
 617         rds_ib_remove_conn(rds_ibdev, conn);
 618 out:
 619         rds_ib_dev_put(rds_ibdev);
 620
 621         return ret;
 622 }
 623
 624 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
 625 {
 626         const union rds_ib_conn_priv *dp = event->param.conn.private_data;
 627         u8 data_len, major, minor;
 628         u32 version = 0;
 629         __be16 mask;
 630         u16 common;
 631
 632         /*
 633          * rdma_cm private data is odd - when there is any private data in the
 634          * request, we will be given a pretty large buffer without telling us the
 635          * original size. The only way to tell the difference is by looking at
 636          * the contents, which are initialized to zero.
 637          * If the protocol version fields aren't set, this is a connection attempt
 638          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 639          * We really should have changed this for OFED 1.3 :-(
 640          */
 641
 642         /* Be paranoid. RDS always has privdata */
 643         if (!event->param.conn.private_data_len) {
 644                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 645                         "rejecting\n");
 646                 return 0;
 647         }
 648
 649         if (isv6) {
 650                 data_len = sizeof(struct rds6_ib_connect_private);
 651                 major = dp->ricp_v6.dp_protocol_major;
 652                 minor = dp->ricp_v6.dp_protocol_minor;
 653                 mask = dp->ricp_v6.dp_protocol_minor_mask;
 654         } else {
 655                 data_len = sizeof(struct rds_ib_connect_private);
 656                 major = dp->ricp_v4.dp_protocol_major;
 657                 minor = dp->ricp_v4.dp_protocol_minor;
 658                 mask = dp->ricp_v4.dp_protocol_minor_mask;
 659         }
 660
 661         /* Even if len is crap *now* I still want to check it. -ASG */
 662         if (event->param.conn.private_data_len < data_len || major == 0)
 663                 return RDS_PROTOCOL_3_0;
 664
 665         common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 666         if (major == 3 && common) {
 667                 version = RDS_PROTOCOL_3_0;
 668                 while ((common >>= 1) != 0)
 669                         version++;
 670         } else {
 671                 if (isv6)
 672                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
 673                                            &dp->ricp_v6.dp_saddr, major, minor);
 674                 else
 675                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 676                                            &dp->ricp_v4.dp_saddr, major, minor);
 677         }
 678         return version;
 679 }
 680
 681 #if IS_ENABLED(CONFIG_IPV6)
 682 /* Given an IPv6 address, find the net_device which hosts that address and
 683  * return its index.  This is used by the rds_ib_cm_handle_connect() code to
 684  * find the interface index of where an incoming request comes from when
 685  * the request is using a link local address.
 686  *
 687  * Note one problem in this search.  It is possible that two interfaces have
 688  * the same link local address.  Unfortunately, this cannot be solved unless
 689  * the underlying layer gives us the interface which an incoming RDMA connect
 690  * request comes from.
 691  */
 692 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
 693 {
 694         struct net_device *dev;
 695         int idx = 0;
 696
 697         rcu_read_lock();
 698         for_each_netdev_rcu(net, dev) {
 699                 if (ipv6_chk_addr(net, addr, dev, 1)) {
 700                         idx = dev->ifindex;
 701                         break;
 702                 }
 703         }
 704         rcu_read_unlock();
 705
 706         return idx;
 707 }
 708 #endif
 709
 710 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 711                              struct rdma_cm_event *event, bool isv6)
 712 {
 713         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 714         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 715         const struct rds_ib_conn_priv_cmn *dp_cmn;
 716         struct rds_connection *conn = NULL;
 717         struct rds_ib_connection *ic = NULL;
 718         struct rdma_conn_param conn_param;
 719         const union rds_ib_conn_priv *dp;
 720         union rds_ib_conn_priv dp_rep;
 721         struct in6_addr s_mapped_addr;
 722         struct in6_addr d_mapped_addr;
 723         const struct in6_addr *saddr6;
 724         const struct in6_addr *daddr6;
 725         int destroy = 1;
 726         u32 ifindex = 0;
 727         u32 version;
 728         int err = 1;
 729
 730         /* Check whether the remote protocol version matches ours. */
 731         version = rds_ib_protocol_compatible(event, isv6);
 732         if (!version)
 733                 goto out;
 734
 735         dp = event->param.conn.private_data;
 736         if (isv6) {
 737 #if IS_ENABLED(CONFIG_IPV6)
 738                 dp_cmn = &dp->ricp_v6.dp_cmn;
 739                 saddr6 = &dp->ricp_v6.dp_saddr;
 740                 daddr6 = &dp->ricp_v6.dp_daddr;
 741                 /* If either address is link local, need to find the
 742                  * interface index in order to create a proper RDS
 743                  * connection.
 744                  */
 745                 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
 746                         /* Using init_net for now ..  */
 747                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 748                         /* No index found...  Need to bail out. */
 749                         if (ifindex == 0) {
 750                                 err = -EOPNOTSUPP;
 751                                 goto out;
 752                         }
 753                 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
 754                         /* Use our address to find the correct index. */
 755                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 756                         /* No index found...  Need to bail out. */
 757                         if (ifindex == 0) {
 758                                 err = -EOPNOTSUPP;
 759                                 goto out;
 760                         }
 761                 }
 762 #else
 763                 err = -EOPNOTSUPP;
 764                 goto out;
 765 #endif
 766         } else {
 767                 dp_cmn = &dp->ricp_v4.dp_cmn;
 768                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
 769                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
 770                 saddr6 = &s_mapped_addr;
 771                 daddr6 = &d_mapped_addr;
 772         }
 773
 774         rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid "
 775                  "0x%llx\n", saddr6, daddr6,
 776                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 777                  (unsigned long long)be64_to_cpu(lguid),
 778                  (unsigned long long)be64_to_cpu(fguid));
 779
 780         /* RDS/IB is not currently netns aware, thus init_net */
 781         conn = rds_conn_create(&init_net, daddr6, saddr6,
 782                                &rds_ib_transport, GFP_KERNEL, ifindex);
 783         if (IS_ERR(conn)) {
 784                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 785                 conn = NULL;
 786                 goto out;
 787         }
 788
 789         /*
 790          * The connection request may occur while the
 791          * previous connection exist, e.g. in case of failover.
 792          * But as connections may be initiated simultaneously
 793          * by both hosts, we have a random backoff mechanism -
 794          * see the comment above rds_queue_reconnect()
 795          */
 796         mutex_lock(&conn->c_cm_lock);
 797         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 798                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 799                         rdsdebug("incoming connect while connecting\n");
 800                         rds_conn_drop(conn);
 801                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 802                 } else
 803                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 804                         /* Wait and see - our connect may still be succeeding */
 805                         rds_ib_stats_inc(s_ib_connect_raced);
 806                 }
 807                 goto out;
 808         }
 809
 810         ic = conn->c_transport_data;
 811
 812         rds_ib_set_protocol(conn, version);
 813         rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
 814
 815         /* If the peer gave us the last packet it saw, process this as if
 816          * we had received a regular ACK. */
 817         if (dp_cmn->ricpc_ack_seq)
 818                 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
 819                                     NULL);
 820
 821         BUG_ON(cm_id->context);
 822         BUG_ON(ic->i_cm_id);
 823
 824         ic->i_cm_id = cm_id;
 825         cm_id->context = conn;
 826
 827         /* We got halfway through setting up the ib_connection, if we
 828          * fail now, we have to take the long route out of this mess. */
 829         destroy = 0;
 830
 831         err = rds_ib_setup_qp(conn);
 832         if (err) {
 833                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 834                 goto out;
 835         }
 836
 837         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 838                                   event->param.conn.responder_resources,
 839                                   event->param.conn.initiator_depth, isv6);
 840
 841         /* rdma_accept() calls rdma_reject() internally if it fails */
 842         if (rdma_accept(cm_id, &conn_param))
 843                 rds_ib_conn_error(conn, "rdma_accept failed\n");
 844
 845 out:
 846         if (conn)
 847                 mutex_unlock(&conn->c_cm_lock);
 848         if (err)
 849                 rdma_reject(cm_id, NULL, 0);
 850         return destroy;
 851 }
 852
 853
 854 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
 855 {
 856         struct rds_connection *conn = cm_id->context;
 857         struct rds_ib_connection *ic = conn->c_transport_data;
 858         struct rdma_conn_param conn_param;
 859         union rds_ib_conn_priv dp;
 860         int ret;
 861
 862         /* If the peer doesn't do protocol negotiation, we must
 863          * default to RDSv3.0 */
 864         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 865         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 866
 867         ret = rds_ib_setup_qp(conn);
 868         if (ret) {
 869                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 870                 goto out;
 871         }
 872
 873         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 874                                   UINT_MAX, UINT_MAX, isv6);
 875         ret = rdma_connect(cm_id, &conn_param);
 876         if (ret)
 877                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 878
 879 out:
 880         /* Beware - returning non-zero tells the rdma_cm to destroy
 881          * the cm_id. We should certainly not do it as long as we still
 882          * "own" the cm_id. */
 883         if (ret) {
 884                 if (ic->i_cm_id == cm_id)
 885                         ret = 0;
 886         }
 887         ic->i_active_side = true;
 888         return ret;
 889 }
 890
 891 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
 892 {
 893         struct rds_connection *conn = cp->cp_conn;
 894         struct sockaddr_storage src, dest;
 895         rdma_cm_event_handler handler;
 896         struct rds_ib_connection *ic;
 897         int ret;
 898
 899         ic = conn->c_transport_data;
 900
 901         /* XXX I wonder what affect the port space has */
 902         /* delegate cm event handler to rdma_transport */
 903 #if IS_ENABLED(CONFIG_IPV6)
 904         if (conn->c_isv6)
 905                 handler = rds6_rdma_cm_event_handler;
 906         else
 907 #endif
 908                 handler = rds_rdma_cm_event_handler;
 909         ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
 910                                      RDMA_PS_TCP, IB_QPT_RC);
 911         if (IS_ERR(ic->i_cm_id)) {
 912                 ret = PTR_ERR(ic->i_cm_id);
 913                 ic->i_cm_id = NULL;
 914                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 915                 goto out;
 916         }
 917
 918         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 919
 920         if (ipv6_addr_v4mapped(&conn->c_faddr)) {
 921                 struct sockaddr_in *sin;
 922
 923                 sin = (struct sockaddr_in *)&src;
 924                 sin->sin_family = AF_INET;
 925                 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
 926                 sin->sin_port = 0;
 927
 928                 sin = (struct sockaddr_in *)&dest;
 929                 sin->sin_family = AF_INET;
 930                 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
 931                 sin->sin_port = htons(RDS_PORT);
 932         } else {
 933                 struct sockaddr_in6 *sin6;
 934
 935                 sin6 = (struct sockaddr_in6 *)&src;
 936                 sin6->sin6_family = AF_INET6;
 937                 sin6->sin6_addr = conn->c_laddr;
 938                 sin6->sin6_port = 0;
 939                 sin6->sin6_scope_id = conn->c_dev_if;
 940
 941                 sin6 = (struct sockaddr_in6 *)&dest;
 942                 sin6->sin6_family = AF_INET6;
 943                 sin6->sin6_addr = conn->c_faddr;
 944                 sin6->sin6_port = htons(RDS_CM_PORT);
 945                 sin6->sin6_scope_id = conn->c_dev_if;
 946         }
 947
 948         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 949                                 (struct sockaddr *)&dest,
 950                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 951         if (ret) {
 952                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 953                          ret);
 954                 rdma_destroy_id(ic->i_cm_id);
 955                 ic->i_cm_id = NULL;
 956         }
 957
 958 out:
 959         return ret;
 960 }
 961
 962 /*
 963  * This is so careful about only cleaning up resources that were built up
 964  * so that it can be called at any point during startup.  In fact it
 965  * can be called multiple times for a given connection.
 966  */
 967 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
 968 {
 969         struct rds_connection *conn = cp->cp_conn;
 970         struct rds_ib_connection *ic = conn->c_transport_data;
 971         int err = 0;
 972
 973         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 974                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 975                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 976
 977         if (ic->i_cm_id) {
 978                 struct ib_device *dev = ic->i_cm_id->device;
 979
 980                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 981                 err = rdma_disconnect(ic->i_cm_id);
 982                 if (err) {
 983                         /* Actually this may happen quite frequently, when
 984                          * an outgoing connect raced with an incoming connect.
 985                          */
 986                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 987                                 ic->i_cm_id, err);
 988                 }
 989
 990                 /*
 991                  * We want to wait for tx and rx completion to finish
 992                  * before we tear down the connection, but we have to be
 993                  * careful not to get stuck waiting on a send ring that
 994                  * only has unsignaled sends in it.  We've shutdown new
 995                  * sends before getting here so by waiting for signaled
 996                  * sends to complete we're ensured that there will be no
 997                  * more tx processing.
 998                  */
 999                 wait_event(rds_ib_ring_empty_wait,
1000                            rds_ib_ring_empty(&ic->i_recv_ring) &&
1001                            (atomic_read(&ic->i_signaled_sends) == 0) &&
1002                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR) &&
1003                            (atomic_read(&ic->i_fastunreg_wrs) == RDS_IB_DEFAULT_FR_INV_WR));
1004                 tasklet_kill(&ic->i_send_tasklet);
1005                 tasklet_kill(&ic->i_recv_tasklet);
1006
1007                 atomic_set(&ic->i_cq_quiesce, 1);
1008
1009                 /* first destroy the ib state that generates callbacks */
1010                 if (ic->i_cm_id->qp)
1011                         rdma_destroy_qp(ic->i_cm_id);
1012                 if (ic->i_send_cq) {
1013                         if (ic->rds_ibdev)
1014                                 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1015                         ib_destroy_cq(ic->i_send_cq);
1016                 }
1017
1018                 if (ic->i_recv_cq) {
1019                         if (ic->rds_ibdev)
1020                                 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1021                         ib_destroy_cq(ic->i_recv_cq);
1022                 }
1023
1024                 /* then free the resources that ib callbacks use */
1025                 if (ic->i_send_hdrs)
1026                         ib_dma_free_coherent(dev,
1027                                            ic->i_send_ring.w_nr *
1028                                                 sizeof(struct rds_header),
1029                                            ic->i_send_hdrs,
1030                                            ic->i_send_hdrs_dma);
1031
1032                 if (ic->i_recv_hdrs)
1033                         ib_dma_free_coherent(dev,
1034                                            ic->i_recv_ring.w_nr *
1035                                                 sizeof(struct rds_header),
1036                                            ic->i_recv_hdrs,
1037                                            ic->i_recv_hdrs_dma);
1038
1039                 if (ic->i_ack)
1040                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
1041                                              ic->i_ack, ic->i_ack_dma);
1042
1043                 if (ic->i_sends)
1044                         rds_ib_send_clear_ring(ic);
1045                 if (ic->i_recvs)
1046                         rds_ib_recv_clear_ring(ic);
1047
1048                 rdma_destroy_id(ic->i_cm_id);
1049
1050                 /*
1051                  * Move connection back to the nodev list.
1052                  */
1053                 if (ic->rds_ibdev)
1054                         rds_ib_remove_conn(ic->rds_ibdev, conn);
1055
1056                 ic->i_cm_id = NULL;
1057                 ic->i_pd = NULL;
1058                 ic->i_send_cq = NULL;
1059                 ic->i_recv_cq = NULL;
1060                 ic->i_send_hdrs = NULL;
1061                 ic->i_recv_hdrs = NULL;
1062                 ic->i_ack = NULL;
1063         }
1064         BUG_ON(ic->rds_ibdev);
1065
1066         /* Clear pending transmit */
1067         if (ic->i_data_op) {
1068                 struct rds_message *rm;
1069
1070                 rm = container_of(ic->i_data_op, struct rds_message, data);
1071                 rds_message_put(rm);
1072                 ic->i_data_op = NULL;
1073         }
1074
1075         /* Clear the ACK state */
1076         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1077 #ifdef KERNEL_HAS_ATOMIC64
1078         atomic64_set(&ic->i_ack_next, 0);
1079 #else
1080         ic->i_ack_next = 0;
1081 #endif
1082         ic->i_ack_recv = 0;
1083
1084         /* Clear flow control state */
1085         ic->i_flowctl = 0;
1086         atomic_set(&ic->i_credits, 0);
1087
1088         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1089         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1090
1091         if (ic->i_ibinc) {
1092                 rds_inc_put(&ic->i_ibinc->ii_inc);
1093                 ic->i_ibinc = NULL;
1094         }
1095
1096         vfree(ic->i_sends);
1097         ic->i_sends = NULL;
1098         vfree(ic->i_recvs);
1099         ic->i_recvs = NULL;
1100         ic->i_active_side = false;
1101 }
1102
1103 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1104 {
1105         struct rds_ib_connection *ic;
1106         unsigned long flags;
1107         int ret;
1108
1109         /* XXX too lazy? */
1110         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1111         if (!ic)
1112                 return -ENOMEM;
1113
1114         ret = rds_ib_recv_alloc_caches(ic, gfp);
1115         if (ret) {
1116                 kfree(ic);
1117                 return ret;
1118         }
1119
1120         INIT_LIST_HEAD(&ic->ib_node);
1121         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1122                      (unsigned long)ic);
1123         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1124                      (unsigned long)ic);
1125         mutex_init(&ic->i_recv_mutex);
1126 #ifndef KERNEL_HAS_ATOMIC64
1127         spin_lock_init(&ic->i_ack_lock);
1128 #endif
1129         atomic_set(&ic->i_signaled_sends, 0);
1130
1131         /*
1132          * rds_ib_conn_shutdown() waits for these to be emptied so they
1133          * must be initialized before it can be called.
1134          */
1135         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
1136         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
1137
1138         ic->conn = conn;
1139         conn->c_transport_data = ic;
1140
1141         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1142         list_add_tail(&ic->ib_node, &ib_nodev_conns);
1143         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1144
1145
1146         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1147         return 0;
1148 }
1149
1150 /*
1151  * Free a connection. Connection must be shut down and not set for reconnect.
1152  */
1153 void rds_ib_conn_free(void *arg)
1154 {
1155         struct rds_ib_connection *ic = arg;
1156         spinlock_t      *lock_ptr;
1157
1158         rdsdebug("ic %p\n", ic);
1159
1160         /*
1161          * Conn is either on a dev's list or on the nodev list.
1162          * A race with shutdown() or connect() would cause problems
1163          * (since rds_ibdev would change) but that should never happen.
1164          */
1165         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1166
1167         spin_lock_irq(lock_ptr);
1168         list_del(&ic->ib_node);
1169         spin_unlock_irq(lock_ptr);
1170
1171         rds_ib_recv_free_caches(ic);
1172
1173         kfree(ic);
1174 }
1175
1176
1177 /*
1178  * An error occurred on the connection
1179  */
1180 void
1181 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1182 {
1183         va_list ap;
1184
1185         rds_conn_drop(conn);
1186
1187         va_start(ap, fmt);
1188         vprintk(fmt, ap);
1189         va_end(ap);
1190 }