net/rds/ib_cm.c

   1 /*
   2  * Copyright (c) 2006, 2019 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/dmapool.h>
  34 #include <linux/kernel.h>
  35 #include <linux/in.h>
  36 #include <linux/slab.h>
  37 #include <linux/vmalloc.h>
  38 #include <linux/ratelimit.h>
  39 #include <net/addrconf.h>
  40
  41 #include "rds_single_path.h"
  42 #include "rds.h"
  43 #include "ib.h"
  44 #include "ib_mr.h"
  45
  46 /*
  47  * Set the selected protocol version
  48  */
  49 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  50 {
  51         conn->c_version = version;
  52 }
  53
  54 /*
  55  * Set up flow control
  56  */
  57 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  58 {
  59         struct rds_ib_connection *ic = conn->c_transport_data;
  60
  61         if (rds_ib_sysctl_flow_control && credits != 0) {
  62                 /* We're doing flow control */
  63                 ic->i_flowctl = 1;
  64                 rds_ib_send_add_credits(conn, credits);
  65         } else {
  66                 ic->i_flowctl = 0;
  67         }
  68 }
  69
  70 /*
  71  * Tune RNR behavior. Without flow control, we use a rather
  72  * low timeout, but not the absolute minimum - this should
  73  * be tunable.
  74  *
  75  * We already set the RNR retry count to 7 (which is the
  76  * smallest infinite number :-) above.
  77  * If flow control is off, we want to change this back to 0
  78  * so that we learn quickly when our credit accounting is
  79  * buggy.
  80  *
  81  * Caller passes in a qp_attr pointer - don't waste stack spacv
  82  * by allocation this twice.
  83  */
  84 static void
  85 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  86 {
  87         int ret;
  88
  89         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  90         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  91         if (ret)
  92                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  93 }
  94
  95 /*
  96  * Connection established.
  97  * We get here for both outgoing and incoming connection.
  98  */
  99 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
 100 {
 101         struct rds_ib_connection *ic = conn->c_transport_data;
 102         const union rds_ib_conn_priv *dp = NULL;
 103         struct ib_qp_attr qp_attr;
 104         __be64 ack_seq = 0;
 105         __be32 credit = 0;
 106         u8 major = 0;
 107         u8 minor = 0;
 108         int err;
 109
 110         dp = event->param.conn.private_data;
 111         if (conn->c_isv6) {
 112                 if (event->param.conn.private_data_len >=
 113                     sizeof(struct rds6_ib_connect_private)) {
 114                         major = dp->ricp_v6.dp_protocol_major;
 115                         minor = dp->ricp_v6.dp_protocol_minor;
 116                         credit = dp->ricp_v6.dp_credit;
 117                         /* dp structure start is not guaranteed to be 8 bytes
 118                          * aligned.  Since dp_ack_seq is 64-bit extended load
 119                          * operations can be used so go through get_unaligned
 120                          * to avoid unaligned errors.
 121                          */
 122                         ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq);
 123                 }
 124         } else if (event->param.conn.private_data_len >=
 125                    sizeof(struct rds_ib_connect_private)) {
 126                 major = dp->ricp_v4.dp_protocol_major;
 127                 minor = dp->ricp_v4.dp_protocol_minor;
 128                 credit = dp->ricp_v4.dp_credit;
 129                 ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq);
 130         }
 131
 132         /* make sure it isn't empty data */
 133         if (major) {
 134                 rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor));
 135                 rds_ib_set_flow_control(conn, be32_to_cpu(credit));
 136         }
 137
 138         if (conn->c_version < RDS_PROTOCOL_VERSION) {
 139                 if (conn->c_version != RDS_PROTOCOL_COMPAT_VERSION) {
 140                         pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n",
 141                                   &conn->c_laddr, &conn->c_faddr,
 142                                   RDS_PROTOCOL_MAJOR(conn->c_version),
 143                                   RDS_PROTOCOL_MINOR(conn->c_version));
 144                         rds_conn_destroy(conn);
 145                         return;
 146                 }
 147         }
 148
 149         pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c,%d> version %u.%u%s\n",
 150                   ic->i_active_side ? "Active" : "Passive",
 151                   &conn->c_laddr, &conn->c_faddr, conn->c_tos,
 152                   RDS_PROTOCOL_MAJOR(conn->c_version),
 153                   RDS_PROTOCOL_MINOR(conn->c_version),
 154                   ic->i_flowctl ? ", flow control" : "");
 155
 156         /* receive sl from the peer */
 157         ic->i_sl = ic->i_cm_id->route.path_rec->sl;
 158
 159         atomic_set(&ic->i_cq_quiesce, 0);
 160
 161         /* Init rings and fill recv. this needs to wait until protocol
 162          * negotiation is complete, since ring layout is different
 163          * from 3.1 to 4.1.
 164          */
 165         rds_ib_send_init_ring(ic);
 166         rds_ib_recv_init_ring(ic);
 167         /* Post receive buffers - as a side effect, this will update
 168          * the posted credit count. */
 169         rds_ib_recv_refill(conn, 1, GFP_KERNEL);
 170
 171         /* Tune RNR behavior */
 172         rds_ib_tune_rnr(ic, &qp_attr);
 173
 174         qp_attr.qp_state = IB_QPS_RTS;
 175         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 176         if (err)
 177                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 178
 179         /* update ib_device with this local ipaddr */
 180         err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr);
 181         if (err)
 182                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 183                         err);
 184
 185         /* If the peer gave us the last packet it saw, process this as if
 186          * we had received a regular ACK. */
 187         if (dp) {
 188                 if (ack_seq)
 189                         rds_send_drop_acked(conn, be64_to_cpu(ack_seq),
 190                                             NULL);
 191         }
 192
 193         conn->c_proposed_version = conn->c_version;
 194         rds_connect_complete(conn);
 195 }
 196
 197 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 198                                       struct rdma_conn_param *conn_param,
 199                                       union rds_ib_conn_priv *dp,
 200                                       u32 protocol_version,
 201                                       u32 max_responder_resources,
 202                                       u32 max_initiator_depth,
 203                                       bool isv6)
 204 {
 205         struct rds_ib_connection *ic = conn->c_transport_data;
 206         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 207
 208         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 209
 210         conn_param->responder_resources =
 211                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 212         conn_param->initiator_depth =
 213                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 214         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 215         conn_param->rnr_retry_count = 7;
 216
 217         if (dp) {
 218                 memset(dp, 0, sizeof(*dp));
 219                 if (isv6) {
 220                         dp->ricp_v6.dp_saddr = conn->c_laddr;
 221                         dp->ricp_v6.dp_daddr = conn->c_faddr;
 222                         dp->ricp_v6.dp_protocol_major =
 223                             RDS_PROTOCOL_MAJOR(protocol_version);
 224                         dp->ricp_v6.dp_protocol_minor =
 225                             RDS_PROTOCOL_MINOR(protocol_version);
 226                         dp->ricp_v6.dp_protocol_minor_mask =
 227                             cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 228                         dp->ricp_v6.dp_ack_seq =
 229                             cpu_to_be64(rds_ib_piggyb_ack(ic));
 230                         dp->ricp_v6.dp_cmn.ricpc_dp_toss = conn->c_tos;
 231
 232                         conn_param->private_data = &dp->ricp_v6;
 233                         conn_param->private_data_len = sizeof(dp->ricp_v6);
 234                 } else {
 235                         dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3];
 236                         dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3];
 237                         dp->ricp_v4.dp_protocol_major =
 238                             RDS_PROTOCOL_MAJOR(protocol_version);
 239                         dp->ricp_v4.dp_protocol_minor =
 240                             RDS_PROTOCOL_MINOR(protocol_version);
 241                         dp->ricp_v4.dp_protocol_minor_mask =
 242                             cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 243                         dp->ricp_v4.dp_ack_seq =
 244                             cpu_to_be64(rds_ib_piggyb_ack(ic));
 245                         dp->ricp_v4.dp_cmn.ricpc_dp_toss = conn->c_tos;
 246
 247                         conn_param->private_data = &dp->ricp_v4;
 248                         conn_param->private_data_len = sizeof(dp->ricp_v4);
 249                 }
 250
 251                 /* Advertise flow control */
 252                 if (ic->i_flowctl) {
 253                         unsigned int credits;
 254
 255                         credits = IB_GET_POST_CREDITS
 256                                 (atomic_read(&ic->i_credits));
 257                         if (isv6)
 258                                 dp->ricp_v6.dp_credit = cpu_to_be32(credits);
 259                         else
 260                                 dp->ricp_v4.dp_credit = cpu_to_be32(credits);
 261                         atomic_sub(IB_SET_POST_CREDITS(credits),
 262                                    &ic->i_credits);
 263                 }
 264         }
 265 }
 266
 267 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 268 {
 269         rdsdebug("event %u (%s) data %p\n",
 270                  event->event, ib_event_msg(event->event), data);
 271 }
 272
 273 /* Plucking the oldest entry from the ring can be done concurrently with
 274  * the thread refilling the ring.  Each ring operation is protected by
 275  * spinlocks and the transient state of refilling doesn't change the
 276  * recording of which entry is oldest.
 277  *
 278  * This relies on IB only calling one cq comp_handler for each cq so that
 279  * there will only be one caller of rds_recv_incoming() per RDS connection.
 280  */
 281 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
 282 {
 283         struct rds_connection *conn = context;
 284         struct rds_ib_connection *ic = conn->c_transport_data;
 285
 286         rdsdebug("conn %p cq %p\n", conn, cq);
 287
 288         rds_ib_stats_inc(s_ib_evt_handler_call);
 289
 290         tasklet_schedule(&ic->i_recv_tasklet);
 291 }
 292
 293 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
 294                      struct ib_wc *wcs)
 295 {
 296         int nr, i;
 297         struct ib_wc *wc;
 298
 299         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 300                 for (i = 0; i < nr; i++) {
 301                         wc = wcs + i;
 302                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 303                                  (unsigned long long)wc->wr_id, wc->status,
 304                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 305
 306                         if (wc->wr_id <= ic->i_send_ring.w_nr ||
 307                             wc->wr_id == RDS_IB_ACK_WR_ID)
 308                                 rds_ib_send_cqe_handler(ic, wc);
 309                         else
 310                                 rds_ib_mr_cqe_handler(ic, wc);
 311
 312                 }
 313         }
 314 }
 315
 316 static void rds_ib_tasklet_fn_send(unsigned long data)
 317 {
 318         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 319         struct rds_connection *conn = ic->conn;
 320
 321         rds_ib_stats_inc(s_ib_tasklet_call);
 322
 323         /* if cq has been already reaped, ignore incoming cq event */
 324         if (atomic_read(&ic->i_cq_quiesce))
 325                 return;
 326
 327         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 328         ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 329         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 330
 331         if (rds_conn_up(conn) &&
 332             (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 333             test_bit(0, &conn->c_map_queued)))
 334                 rds_send_xmit(&ic->conn->c_path[0]);
 335 }
 336
 337 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
 338                      struct ib_wc *wcs,
 339                      struct rds_ib_ack_state *ack_state)
 340 {
 341         int nr, i;
 342         struct ib_wc *wc;
 343
 344         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 345                 for (i = 0; i < nr; i++) {
 346                         wc = wcs + i;
 347                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 348                                  (unsigned long long)wc->wr_id, wc->status,
 349                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 350
 351                         rds_ib_recv_cqe_handler(ic, wc, ack_state);
 352                 }
 353         }
 354 }
 355
 356 static void rds_ib_tasklet_fn_recv(unsigned long data)
 357 {
 358         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 359         struct rds_connection *conn = ic->conn;
 360         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 361         struct rds_ib_ack_state state;
 362
 363         if (!rds_ibdev)
 364                 rds_conn_drop(conn);
 365
 366         rds_ib_stats_inc(s_ib_tasklet_call);
 367
 368         /* if cq has been already reaped, ignore incoming cq event */
 369         if (atomic_read(&ic->i_cq_quiesce))
 370                 return;
 371
 372         memset(&state, 0, sizeof(state));
 373         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 374         ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 375         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 376
 377         if (state.ack_next_valid)
 378                 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
 379         if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
 380                 rds_send_drop_acked(conn, state.ack_recv, NULL);
 381                 ic->i_ack_recv = state.ack_recv;
 382         }
 383
 384         if (rds_conn_up(conn))
 385                 rds_ib_attempt_ack(ic);
 386 }
 387
 388 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 389 {
 390         struct rds_connection *conn = data;
 391         struct rds_ib_connection *ic = conn->c_transport_data;
 392
 393         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 394                  ib_event_msg(event->event));
 395
 396         switch (event->event) {
 397         case IB_EVENT_COMM_EST:
 398                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 399                 break;
 400         default:
 401                 rdsdebug("Fatal QP Event %u (%s) - connection %pI6c->%pI6c, reconnecting\n",
 402                          event->event, ib_event_msg(event->event),
 403                          &conn->c_laddr, &conn->c_faddr);
 404                 rds_conn_drop(conn);
 405                 break;
 406         }
 407 }
 408
 409 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
 410 {
 411         struct rds_connection *conn = context;
 412         struct rds_ib_connection *ic = conn->c_transport_data;
 413
 414         rdsdebug("conn %p cq %p\n", conn, cq);
 415
 416         rds_ib_stats_inc(s_ib_evt_handler_call);
 417
 418         tasklet_schedule(&ic->i_send_tasklet);
 419 }
 420
 421 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
 422 {
 423         int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
 424         int index = rds_ibdev->dev->num_comp_vectors - 1;
 425         int i;
 426
 427         for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
 428                 if (rds_ibdev->vector_load[i] < min) {
 429                         index = i;
 430                         min = rds_ibdev->vector_load[i];
 431                 }
 432         }
 433
 434         rds_ibdev->vector_load[index]++;
 435         return index;
 436 }
 437
 438 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
 439 {
 440         rds_ibdev->vector_load[index]--;
 441 }
 442
 443 /* Allocate DMA coherent memory to be used to store struct rds_header for
 444  * sending/receiving packets.  The pointers to the DMA memory and the
 445  * associated DMA addresses are stored in two arrays.
 446  *
 447  * @ibdev: the IB device
 448  * @pool: the DMA memory pool
 449  * @dma_addrs: pointer to the array for storing DMA addresses
 450  * @num_hdrs: number of headers to allocate
 451  *
 452  * It returns the pointer to the array storing the DMA memory pointers.  On
 453  * error, NULL pointer is returned.
 454  */
 455 struct rds_header **rds_dma_hdrs_alloc(struct ib_device *ibdev,
 456                                        struct dma_pool *pool,
 457                                        dma_addr_t **dma_addrs, u32 num_hdrs)
 458 {
 459         struct rds_header **hdrs;
 460         dma_addr_t *hdr_daddrs;
 461         u32 i;
 462
 463         hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
 464                              ibdev_to_node(ibdev));
 465         if (!hdrs)
 466                 return NULL;
 467
 468         hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
 469                                    ibdev_to_node(ibdev));
 470         if (!hdr_daddrs) {
 471                 kvfree(hdrs);
 472                 return NULL;
 473         }
 474
 475         for (i = 0; i < num_hdrs; i++) {
 476                 hdrs[i] = dma_pool_zalloc(pool, GFP_KERNEL, &hdr_daddrs[i]);
 477                 if (!hdrs[i]) {
 478                         rds_dma_hdrs_free(pool, hdrs, hdr_daddrs, i);
 479                         return NULL;
 480                 }
 481         }
 482
 483         *dma_addrs = hdr_daddrs;
 484         return hdrs;
 485 }
 486
 487 /* Free the DMA memory used to store struct rds_header.
 488  *
 489  * @pool: the DMA memory pool
 490  * @hdrs: pointer to the array storing DMA memory pointers
 491  * @dma_addrs: pointer to the array storing DMA addresses
 492  * @num_hdars: number of headers to free.
 493  */
 494 void rds_dma_hdrs_free(struct dma_pool *pool, struct rds_header **hdrs,
 495                        dma_addr_t *dma_addrs, u32 num_hdrs)
 496 {
 497         u32 i;
 498
 499         for (i = 0; i < num_hdrs; i++)
 500                 dma_pool_free(pool, hdrs[i], dma_addrs[i]);
 501         kvfree(hdrs);
 502         kvfree(dma_addrs);
 503 }
 504
 505 /*
 506  * This needs to be very careful to not leave IS_ERR pointers around for
 507  * cleanup to trip over.
 508  */
 509 static int rds_ib_setup_qp(struct rds_connection *conn)
 510 {
 511         struct rds_ib_connection *ic = conn->c_transport_data;
 512         struct ib_device *dev = ic->i_cm_id->device;
 513         struct ib_qp_init_attr attr;
 514         struct ib_cq_init_attr cq_attr = {};
 515         struct rds_ib_device *rds_ibdev;
 516         unsigned long max_wrs;
 517         int ret, fr_queue_space;
 518         struct dma_pool *pool;
 519
 520         /*
 521          * It's normal to see a null device if an incoming connection races
 522          * with device removal, so we don't print a warning.
 523          */
 524         rds_ibdev = rds_ib_get_client_data(dev);
 525         if (!rds_ibdev)
 526                 return -EOPNOTSUPP;
 527
 528         /* The fr_queue_space is currently set to 512, to add extra space on
 529          * completion queue and send queue. This extra space is used for FRMR
 530          * registration and invalidation work requests
 531          */
 532         fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0);
 533
 534         /* add the conn now so that connection establishment has the dev */
 535         rds_ib_add_conn(rds_ibdev, conn);
 536
 537         max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
 538                 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
 539         if (ic->i_send_ring.w_nr != max_wrs)
 540                 rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
 541
 542         max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
 543                 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
 544         if (ic->i_recv_ring.w_nr != max_wrs)
 545                 rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
 546
 547         /* Protection domain and memory range */
 548         ic->i_pd = rds_ibdev->pd;
 549
 550         ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
 551         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 552         cq_attr.comp_vector = ic->i_scq_vector;
 553         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 554                                      rds_ib_cq_event_handler, conn,
 555                                      &cq_attr);
 556         if (IS_ERR(ic->i_send_cq)) {
 557                 ret = PTR_ERR(ic->i_send_cq);
 558                 ic->i_send_cq = NULL;
 559                 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
 560                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 561                 goto rds_ibdev_out;
 562         }
 563
 564         ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
 565         cq_attr.cqe = ic->i_recv_ring.w_nr;
 566         cq_attr.comp_vector = ic->i_rcq_vector;
 567         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 568                                      rds_ib_cq_event_handler, conn,
 569                                      &cq_attr);
 570         if (IS_ERR(ic->i_recv_cq)) {
 571                 ret = PTR_ERR(ic->i_recv_cq);
 572                 ic->i_recv_cq = NULL;
 573                 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
 574                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 575                 goto send_cq_out;
 576         }
 577
 578         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 579         if (ret) {
 580                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 581                 goto recv_cq_out;
 582         }
 583
 584         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 585         if (ret) {
 586                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 587                 goto recv_cq_out;
 588         }
 589
 590         /* XXX negotiate max send/recv with remote? */
 591         memset(&attr, 0, sizeof(attr));
 592         attr.event_handler = rds_ib_qp_event_handler;
 593         attr.qp_context = conn;
 594         /* + 1 to allow for the single ack message */
 595         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 596         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 597         attr.cap.max_send_sge = rds_ibdev->max_sge;
 598         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 599         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 600         attr.qp_type = IB_QPT_RC;
 601         attr.send_cq = ic->i_send_cq;
 602         attr.recv_cq = ic->i_recv_cq;
 603
 604         /*
 605          * XXX this can fail if max_*_wr is too large?  Are we supposed
 606          * to back off until we get a value that the hardware can support?
 607          */
 608         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 609         if (ret) {
 610                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 611                 goto recv_cq_out;
 612         }
 613
 614         pool = rds_ibdev->rid_hdrs_pool;
 615         ic->i_send_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_send_hdrs_dma,
 616                                              ic->i_send_ring.w_nr);
 617         if (!ic->i_send_hdrs) {
 618                 ret = -ENOMEM;
 619                 rdsdebug("DMA send hdrs alloc failed\n");
 620                 goto qp_out;
 621         }
 622
 623         ic->i_recv_hdrs = rds_dma_hdrs_alloc(dev, pool, &ic->i_recv_hdrs_dma,
 624                                              ic->i_recv_ring.w_nr);
 625         if (!ic->i_recv_hdrs) {
 626                 ret = -ENOMEM;
 627                 rdsdebug("DMA recv hdrs alloc failed\n");
 628                 goto send_hdrs_dma_out;
 629         }
 630
 631         ic->i_ack = dma_pool_zalloc(pool, GFP_KERNEL,
 632                                     &ic->i_ack_dma);
 633         if (!ic->i_ack) {
 634                 ret = -ENOMEM;
 635                 rdsdebug("DMA ack header alloc failed\n");
 636                 goto recv_hdrs_dma_out;
 637         }
 638
 639         ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
 640                                               ic->i_send_ring.w_nr),
 641                                    ibdev_to_node(dev));
 642         if (!ic->i_sends) {
 643                 ret = -ENOMEM;
 644                 rdsdebug("send allocation failed\n");
 645                 goto ack_dma_out;
 646         }
 647
 648         ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
 649                                               ic->i_recv_ring.w_nr),
 650                                    ibdev_to_node(dev));
 651         if (!ic->i_recvs) {
 652                 ret = -ENOMEM;
 653                 rdsdebug("recv allocation failed\n");
 654                 goto sends_out;
 655         }
 656
 657         rds_ib_recv_init_ack(ic);
 658
 659         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 660                  ic->i_send_cq, ic->i_recv_cq);
 661
 662         goto out;
 663
 664 sends_out:
 665         vfree(ic->i_sends);
 666
 667 ack_dma_out:
 668         dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
 669         ic->i_ack = NULL;
 670
 671 recv_hdrs_dma_out:
 672         rds_dma_hdrs_free(pool, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
 673                           ic->i_recv_ring.w_nr);
 674         ic->i_recv_hdrs = NULL;
 675         ic->i_recv_hdrs_dma = NULL;
 676
 677 send_hdrs_dma_out:
 678         rds_dma_hdrs_free(pool, ic->i_send_hdrs, ic->i_send_hdrs_dma,
 679                           ic->i_send_ring.w_nr);
 680         ic->i_send_hdrs = NULL;
 681         ic->i_send_hdrs_dma = NULL;
 682
 683 qp_out:
 684         rdma_destroy_qp(ic->i_cm_id);
 685 recv_cq_out:
 686         ib_destroy_cq(ic->i_recv_cq);
 687         ic->i_recv_cq = NULL;
 688 send_cq_out:
 689         ib_destroy_cq(ic->i_send_cq);
 690         ic->i_send_cq = NULL;
 691 rds_ibdev_out:
 692         rds_ib_remove_conn(rds_ibdev, conn);
 693 out:
 694         rds_ib_dev_put(rds_ibdev);
 695
 696         return ret;
 697 }
 698
 699 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
 700 {
 701         const union rds_ib_conn_priv *dp = event->param.conn.private_data;
 702         u8 data_len, major, minor;
 703         u32 version = 0;
 704         __be16 mask;
 705         u16 common;
 706
 707         /*
 708          * rdma_cm private data is odd - when there is any private data in the
 709          * request, we will be given a pretty large buffer without telling us the
 710          * original size. The only way to tell the difference is by looking at
 711          * the contents, which are initialized to zero.
 712          * If the protocol version fields aren't set, this is a connection attempt
 713          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 714          * We really should have changed this for OFED 1.3 :-(
 715          */
 716
 717         /* Be paranoid. RDS always has privdata */
 718         if (!event->param.conn.private_data_len) {
 719                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 720                         "rejecting\n");
 721                 return 0;
 722         }
 723
 724         if (isv6) {
 725                 data_len = sizeof(struct rds6_ib_connect_private);
 726                 major = dp->ricp_v6.dp_protocol_major;
 727                 minor = dp->ricp_v6.dp_protocol_minor;
 728                 mask = dp->ricp_v6.dp_protocol_minor_mask;
 729         } else {
 730                 data_len = sizeof(struct rds_ib_connect_private);
 731                 major = dp->ricp_v4.dp_protocol_major;
 732                 minor = dp->ricp_v4.dp_protocol_minor;
 733                 mask = dp->ricp_v4.dp_protocol_minor_mask;
 734         }
 735
 736         /* Even if len is crap *now* I still want to check it. -ASG */
 737         if (event->param.conn.private_data_len < data_len || major == 0)
 738                 return RDS_PROTOCOL_4_0;
 739
 740         common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 741         if (major == 4 && common) {
 742                 version = RDS_PROTOCOL_4_0;
 743                 while ((common >>= 1) != 0)
 744                         version++;
 745         } else if (RDS_PROTOCOL_COMPAT_VERSION ==
 746                    RDS_PROTOCOL(major, minor)) {
 747                 version = RDS_PROTOCOL_COMPAT_VERSION;
 748         } else {
 749                 if (isv6)
 750                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
 751                                            &dp->ricp_v6.dp_saddr, major, minor);
 752                 else
 753                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 754                                            &dp->ricp_v4.dp_saddr, major, minor);
 755         }
 756         return version;
 757 }
 758
 759 #if IS_ENABLED(CONFIG_IPV6)
 760 /* Given an IPv6 address, find the net_device which hosts that address and
 761  * return its index.  This is used by the rds_ib_cm_handle_connect() code to
 762  * find the interface index of where an incoming request comes from when
 763  * the request is using a link local address.
 764  *
 765  * Note one problem in this search.  It is possible that two interfaces have
 766  * the same link local address.  Unfortunately, this cannot be solved unless
 767  * the underlying layer gives us the interface which an incoming RDMA connect
 768  * request comes from.
 769  */
 770 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
 771 {
 772         struct net_device *dev;
 773         int idx = 0;
 774
 775         rcu_read_lock();
 776         for_each_netdev_rcu(net, dev) {
 777                 if (ipv6_chk_addr(net, addr, dev, 1)) {
 778                         idx = dev->ifindex;
 779                         break;
 780                 }
 781         }
 782         rcu_read_unlock();
 783
 784         return idx;
 785 }
 786 #endif
 787
 788 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 789                              struct rdma_cm_event *event, bool isv6)
 790 {
 791         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 792         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 793         const struct rds_ib_conn_priv_cmn *dp_cmn;
 794         struct rds_connection *conn = NULL;
 795         struct rds_ib_connection *ic = NULL;
 796         struct rdma_conn_param conn_param;
 797         const union rds_ib_conn_priv *dp;
 798         union rds_ib_conn_priv dp_rep;
 799         struct in6_addr s_mapped_addr;
 800         struct in6_addr d_mapped_addr;
 801         const struct in6_addr *saddr6;
 802         const struct in6_addr *daddr6;
 803         int destroy = 1;
 804         u32 ifindex = 0;
 805         u32 version;
 806         int err = 1;
 807
 808         /* Check whether the remote protocol version matches ours. */
 809         version = rds_ib_protocol_compatible(event, isv6);
 810         if (!version) {
 811                 err = RDS_RDMA_REJ_INCOMPAT;
 812                 goto out;
 813         }
 814
 815         dp = event->param.conn.private_data;
 816         if (isv6) {
 817 #if IS_ENABLED(CONFIG_IPV6)
 818                 dp_cmn = &dp->ricp_v6.dp_cmn;
 819                 saddr6 = &dp->ricp_v6.dp_saddr;
 820                 daddr6 = &dp->ricp_v6.dp_daddr;
 821                 /* If either address is link local, need to find the
 822                  * interface index in order to create a proper RDS
 823                  * connection.
 824                  */
 825                 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
 826                         /* Using init_net for now ..  */
 827                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 828                         /* No index found...  Need to bail out. */
 829                         if (ifindex == 0) {
 830                                 err = -EOPNOTSUPP;
 831                                 goto out;
 832                         }
 833                 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
 834                         /* Use our address to find the correct index. */
 835                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 836                         /* No index found...  Need to bail out. */
 837                         if (ifindex == 0) {
 838                                 err = -EOPNOTSUPP;
 839                                 goto out;
 840                         }
 841                 }
 842 #else
 843                 err = -EOPNOTSUPP;
 844                 goto out;
 845 #endif
 846         } else {
 847                 dp_cmn = &dp->ricp_v4.dp_cmn;
 848                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
 849                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
 850                 saddr6 = &s_mapped_addr;
 851                 daddr6 = &d_mapped_addr;
 852         }
 853
 854         rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
 855                  saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
 856                  RDS_PROTOCOL_MINOR(version),
 857                  (unsigned long long)be64_to_cpu(lguid),
 858                  (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
 859
 860         /* RDS/IB is not currently netns aware, thus init_net */
 861         conn = rds_conn_create(&init_net, daddr6, saddr6,
 862                                &rds_ib_transport, dp_cmn->ricpc_dp_toss,
 863                                GFP_KERNEL, ifindex);
 864         if (IS_ERR(conn)) {
 865                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 866                 conn = NULL;
 867                 goto out;
 868         }
 869
 870         /*
 871          * The connection request may occur while the
 872          * previous connection exist, e.g. in case of failover.
 873          * But as connections may be initiated simultaneously
 874          * by both hosts, we have a random backoff mechanism -
 875          * see the comment above rds_queue_reconnect()
 876          */
 877         mutex_lock(&conn->c_cm_lock);
 878         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 879                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 880                         rdsdebug("incoming connect while connecting\n");
 881                         rds_conn_drop(conn);
 882                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 883                 } else
 884                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 885                         /* Wait and see - our connect may still be succeeding */
 886                         rds_ib_stats_inc(s_ib_connect_raced);
 887                 }
 888                 goto out;
 889         }
 890
 891         ic = conn->c_transport_data;
 892
 893         rds_ib_set_protocol(conn, version);
 894         rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
 895
 896         /* If the peer gave us the last packet it saw, process this as if
 897          * we had received a regular ACK. */
 898         if (dp_cmn->ricpc_ack_seq)
 899                 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
 900                                     NULL);
 901
 902         BUG_ON(cm_id->context);
 903         BUG_ON(ic->i_cm_id);
 904
 905         ic->i_cm_id = cm_id;
 906         cm_id->context = conn;
 907
 908         /* We got halfway through setting up the ib_connection, if we
 909          * fail now, we have to take the long route out of this mess. */
 910         destroy = 0;
 911
 912         err = rds_ib_setup_qp(conn);
 913         if (err) {
 914                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 915                 goto out;
 916         }
 917
 918         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 919                                   event->param.conn.responder_resources,
 920                                   event->param.conn.initiator_depth, isv6);
 921
 922         /* rdma_accept() calls rdma_reject() internally if it fails */
 923         if (rdma_accept(cm_id, &conn_param))
 924                 rds_ib_conn_error(conn, "rdma_accept failed\n");
 925
 926 out:
 927         if (conn)
 928                 mutex_unlock(&conn->c_cm_lock);
 929         if (err)
 930                 rdma_reject(cm_id, &err, sizeof(int));
 931         return destroy;
 932 }
 933
 934
 935 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
 936 {
 937         struct rds_connection *conn = cm_id->context;
 938         struct rds_ib_connection *ic = conn->c_transport_data;
 939         struct rdma_conn_param conn_param;
 940         union rds_ib_conn_priv dp;
 941         int ret;
 942
 943         /* If the peer doesn't do protocol negotiation, we must
 944          * default to RDSv3.0 */
 945         rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
 946         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 947
 948         ret = rds_ib_setup_qp(conn);
 949         if (ret) {
 950                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 951                 goto out;
 952         }
 953
 954         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
 955                                   conn->c_proposed_version,
 956                                   UINT_MAX, UINT_MAX, isv6);
 957         ret = rdma_connect(cm_id, &conn_param);
 958         if (ret)
 959                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 960
 961 out:
 962         /* Beware - returning non-zero tells the rdma_cm to destroy
 963          * the cm_id. We should certainly not do it as long as we still
 964          * "own" the cm_id. */
 965         if (ret) {
 966                 if (ic->i_cm_id == cm_id)
 967                         ret = 0;
 968         }
 969         ic->i_active_side = true;
 970         return ret;
 971 }
 972
 973 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
 974 {
 975         struct rds_connection *conn = cp->cp_conn;
 976         struct sockaddr_storage src, dest;
 977         rdma_cm_event_handler handler;
 978         struct rds_ib_connection *ic;
 979         int ret;
 980
 981         ic = conn->c_transport_data;
 982
 983         /* XXX I wonder what affect the port space has */
 984         /* delegate cm event handler to rdma_transport */
 985 #if IS_ENABLED(CONFIG_IPV6)
 986         if (conn->c_isv6)
 987                 handler = rds6_rdma_cm_event_handler;
 988         else
 989 #endif
 990                 handler = rds_rdma_cm_event_handler;
 991         ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
 992                                      RDMA_PS_TCP, IB_QPT_RC);
 993         if (IS_ERR(ic->i_cm_id)) {
 994                 ret = PTR_ERR(ic->i_cm_id);
 995                 ic->i_cm_id = NULL;
 996                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 997                 goto out;
 998         }
 999
1000         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
1001
1002         if (ipv6_addr_v4mapped(&conn->c_faddr)) {
1003                 struct sockaddr_in *sin;
1004
1005                 sin = (struct sockaddr_in *)&src;
1006                 sin->sin_family = AF_INET;
1007                 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1008                 sin->sin_port = 0;
1009
1010                 sin = (struct sockaddr_in *)&dest;
1011                 sin->sin_family = AF_INET;
1012                 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1013                 sin->sin_port = htons(RDS_PORT);
1014         } else {
1015                 struct sockaddr_in6 *sin6;
1016
1017                 sin6 = (struct sockaddr_in6 *)&src;
1018                 sin6->sin6_family = AF_INET6;
1019                 sin6->sin6_addr = conn->c_laddr;
1020                 sin6->sin6_port = 0;
1021                 sin6->sin6_scope_id = conn->c_dev_if;
1022
1023                 sin6 = (struct sockaddr_in6 *)&dest;
1024                 sin6->sin6_family = AF_INET6;
1025                 sin6->sin6_addr = conn->c_faddr;
1026                 sin6->sin6_port = htons(RDS_CM_PORT);
1027                 sin6->sin6_scope_id = conn->c_dev_if;
1028         }
1029
1030         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1031                                 (struct sockaddr *)&dest,
1032                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1033         if (ret) {
1034                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1035                          ret);
1036                 rdma_destroy_id(ic->i_cm_id);
1037                 ic->i_cm_id = NULL;
1038         }
1039
1040 out:
1041         return ret;
1042 }
1043
1044 /*
1045  * This is so careful about only cleaning up resources that were built up
1046  * so that it can be called at any point during startup.  In fact it
1047  * can be called multiple times for a given connection.
1048  */
1049 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1050 {
1051         struct rds_connection *conn = cp->cp_conn;
1052         struct rds_ib_connection *ic = conn->c_transport_data;
1053         int err = 0;
1054
1055         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1056                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1057                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1058
1059         if (ic->i_cm_id) {
1060                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1061                 err = rdma_disconnect(ic->i_cm_id);
1062                 if (err) {
1063                         /* Actually this may happen quite frequently, when
1064                          * an outgoing connect raced with an incoming connect.
1065                          */
1066                         rdsdebug("failed to disconnect, cm: %p err %d\n",
1067                                 ic->i_cm_id, err);
1068                 }
1069
1070                 /* kick off "flush_worker" for all pools in order to reap
1071                  * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1072                  */
1073                 rds_ib_flush_mrs();
1074
1075                 /*
1076                  * We want to wait for tx and rx completion to finish
1077                  * before we tear down the connection, but we have to be
1078                  * careful not to get stuck waiting on a send ring that
1079                  * only has unsignaled sends in it.  We've shutdown new
1080                  * sends before getting here so by waiting for signaled
1081                  * sends to complete we're ensured that there will be no
1082                  * more tx processing.
1083                  */
1084                 wait_event(rds_ib_ring_empty_wait,
1085                            rds_ib_ring_empty(&ic->i_recv_ring) &&
1086                            (atomic_read(&ic->i_signaled_sends) == 0) &&
1087                            (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1088                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1089                 tasklet_kill(&ic->i_send_tasklet);
1090                 tasklet_kill(&ic->i_recv_tasklet);
1091
1092                 atomic_set(&ic->i_cq_quiesce, 1);
1093
1094                 /* first destroy the ib state that generates callbacks */
1095                 if (ic->i_cm_id->qp)
1096                         rdma_destroy_qp(ic->i_cm_id);
1097                 if (ic->i_send_cq) {
1098                         if (ic->rds_ibdev)
1099                                 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1100                         ib_destroy_cq(ic->i_send_cq);
1101                 }
1102
1103                 if (ic->i_recv_cq) {
1104                         if (ic->rds_ibdev)
1105                                 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1106                         ib_destroy_cq(ic->i_recv_cq);
1107                 }
1108
1109                 if (ic->rds_ibdev) {
1110                         struct dma_pool *pool;
1111
1112                         pool = ic->rds_ibdev->rid_hdrs_pool;
1113
1114                         /* then free the resources that ib callbacks use */
1115                         if (ic->i_send_hdrs) {
1116                                 rds_dma_hdrs_free(pool, ic->i_send_hdrs,
1117                                                   ic->i_send_hdrs_dma,
1118                                                   ic->i_send_ring.w_nr);
1119                                 ic->i_send_hdrs = NULL;
1120                                 ic->i_send_hdrs_dma = NULL;
1121                         }
1122
1123                         if (ic->i_recv_hdrs) {
1124                                 rds_dma_hdrs_free(pool, ic->i_recv_hdrs,
1125                                                   ic->i_recv_hdrs_dma,
1126                                                   ic->i_recv_ring.w_nr);
1127                                 ic->i_recv_hdrs = NULL;
1128                                 ic->i_recv_hdrs_dma = NULL;
1129                         }
1130
1131                         if (ic->i_ack) {
1132                                 dma_pool_free(pool, ic->i_ack, ic->i_ack_dma);
1133                                 ic->i_ack = NULL;
1134                         }
1135                 } else {
1136                         WARN_ON(ic->i_send_hdrs);
1137                         WARN_ON(ic->i_send_hdrs_dma);
1138                         WARN_ON(ic->i_recv_hdrs);
1139                         WARN_ON(ic->i_recv_hdrs_dma);
1140                         WARN_ON(ic->i_ack);
1141                 }
1142
1143                 if (ic->i_sends)
1144                         rds_ib_send_clear_ring(ic);
1145                 if (ic->i_recvs)
1146                         rds_ib_recv_clear_ring(ic);
1147
1148                 rdma_destroy_id(ic->i_cm_id);
1149
1150                 /*
1151                  * Move connection back to the nodev list.
1152                  */
1153                 if (ic->rds_ibdev)
1154                         rds_ib_remove_conn(ic->rds_ibdev, conn);
1155
1156                 ic->i_cm_id = NULL;
1157                 ic->i_pd = NULL;
1158                 ic->i_send_cq = NULL;
1159                 ic->i_recv_cq = NULL;
1160         }
1161         BUG_ON(ic->rds_ibdev);
1162
1163         /* Clear pending transmit */
1164         if (ic->i_data_op) {
1165                 struct rds_message *rm;
1166
1167                 rm = container_of(ic->i_data_op, struct rds_message, data);
1168                 rds_message_put(rm);
1169                 ic->i_data_op = NULL;
1170         }
1171
1172         /* Clear the ACK state */
1173         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1174 #ifdef KERNEL_HAS_ATOMIC64
1175         atomic64_set(&ic->i_ack_next, 0);
1176 #else
1177         ic->i_ack_next = 0;
1178 #endif
1179         ic->i_ack_recv = 0;
1180
1181         /* Clear flow control state */
1182         ic->i_flowctl = 0;
1183         atomic_set(&ic->i_credits, 0);
1184
1185         /* Re-init rings, but retain sizes. */
1186         rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1187         rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1188
1189         if (ic->i_ibinc) {
1190                 rds_inc_put(&ic->i_ibinc->ii_inc);
1191                 ic->i_ibinc = NULL;
1192         }
1193
1194         vfree(ic->i_sends);
1195         ic->i_sends = NULL;
1196         vfree(ic->i_recvs);
1197         ic->i_recvs = NULL;
1198         ic->i_active_side = false;
1199 }
1200
1201 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1202 {
1203         struct rds_ib_connection *ic;
1204         unsigned long flags;
1205         int ret;
1206
1207         /* XXX too lazy? */
1208         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1209         if (!ic)
1210                 return -ENOMEM;
1211
1212         ret = rds_ib_recv_alloc_caches(ic, gfp);
1213         if (ret) {
1214                 kfree(ic);
1215                 return ret;
1216         }
1217
1218         INIT_LIST_HEAD(&ic->ib_node);
1219         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1220                      (unsigned long)ic);
1221         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1222                      (unsigned long)ic);
1223         mutex_init(&ic->i_recv_mutex);
1224 #ifndef KERNEL_HAS_ATOMIC64
1225         spin_lock_init(&ic->i_ack_lock);
1226 #endif
1227         atomic_set(&ic->i_signaled_sends, 0);
1228         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1229
1230         /*
1231          * rds_ib_conn_shutdown() waits for these to be emptied so they
1232          * must be initialized before it can be called.
1233          */
1234         rds_ib_ring_init(&ic->i_send_ring, 0);
1235         rds_ib_ring_init(&ic->i_recv_ring, 0);
1236
1237         ic->conn = conn;
1238         conn->c_transport_data = ic;
1239
1240         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1241         list_add_tail(&ic->ib_node, &ib_nodev_conns);
1242         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1243
1244
1245         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1246         return 0;
1247 }
1248
1249 /*
1250  * Free a connection. Connection must be shut down and not set for reconnect.
1251  */
1252 void rds_ib_conn_free(void *arg)
1253 {
1254         struct rds_ib_connection *ic = arg;
1255         spinlock_t      *lock_ptr;
1256
1257         rdsdebug("ic %p\n", ic);
1258
1259         /*
1260          * Conn is either on a dev's list or on the nodev list.
1261          * A race with shutdown() or connect() would cause problems
1262          * (since rds_ibdev would change) but that should never happen.
1263          */
1264         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1265
1266         spin_lock_irq(lock_ptr);
1267         list_del(&ic->ib_node);
1268         spin_unlock_irq(lock_ptr);
1269
1270         rds_ib_recv_free_caches(ic);
1271
1272         kfree(ic);
1273 }
1274
1275
1276 /*
1277  * An error occurred on the connection
1278  */
1279 void
1280 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1281 {
1282         va_list ap;
1283
1284         rds_conn_drop(conn);
1285
1286         va_start(ap, fmt);
1287         vprintk(fmt, ap);
1288         va_end(ap);
1289 }