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/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 #include <rdma/ib_cm.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 static void rds_dma_hdr_free(struct ib_device *dev, struct rds_header *hdr,
 444                 dma_addr_t dma_addr, enum dma_data_direction dir)
 445 {
 446         ib_dma_unmap_single(dev, dma_addr, sizeof(*hdr), dir);
 447         kfree(hdr);
 448 }
 449
 450 static struct rds_header *rds_dma_hdr_alloc(struct ib_device *dev,
 451                 dma_addr_t *dma_addr, enum dma_data_direction dir)
 452 {
 453         struct rds_header *hdr;
 454
 455         hdr = kzalloc_node(sizeof(*hdr), GFP_KERNEL, ibdev_to_node(dev));
 456         if (!hdr)
 457                 return NULL;
 458
 459         *dma_addr = ib_dma_map_single(dev, hdr, sizeof(*hdr),
 460                                       DMA_BIDIRECTIONAL);
 461         if (ib_dma_mapping_error(dev, *dma_addr)) {
 462                 kfree(hdr);
 463                 return NULL;
 464         }
 465
 466         return hdr;
 467 }
 468
 469 /* Free the DMA memory used to store struct rds_header.
 470  *
 471  * @dev: the RDS IB device
 472  * @hdrs: pointer to the array storing DMA memory pointers
 473  * @dma_addrs: pointer to the array storing DMA addresses
 474  * @num_hdars: number of headers to free.
 475  */
 476 static void rds_dma_hdrs_free(struct rds_ib_device *dev,
 477                 struct rds_header **hdrs, dma_addr_t *dma_addrs, u32 num_hdrs,
 478                 enum dma_data_direction dir)
 479 {
 480         u32 i;
 481
 482         for (i = 0; i < num_hdrs; i++)
 483                 rds_dma_hdr_free(dev->dev, hdrs[i], dma_addrs[i], dir);
 484         kvfree(hdrs);
 485         kvfree(dma_addrs);
 486 }
 487
 488
 489 /* Allocate DMA coherent memory to be used to store struct rds_header for
 490  * sending/receiving packets.  The pointers to the DMA memory and the
 491  * associated DMA addresses are stored in two arrays.
 492  *
 493  * @dev: the RDS IB device
 494  * @dma_addrs: pointer to the array for storing DMA addresses
 495  * @num_hdrs: number of headers to allocate
 496  *
 497  * It returns the pointer to the array storing the DMA memory pointers.  On
 498  * error, NULL pointer is returned.
 499  */
 500 static struct rds_header **rds_dma_hdrs_alloc(struct rds_ib_device *dev,
 501                 dma_addr_t **dma_addrs, u32 num_hdrs,
 502                 enum dma_data_direction dir)
 503 {
 504         struct rds_header **hdrs;
 505         dma_addr_t *hdr_daddrs;
 506         u32 i;
 507
 508         hdrs = kvmalloc_node(sizeof(*hdrs) * num_hdrs, GFP_KERNEL,
 509                              ibdev_to_node(dev->dev));
 510         if (!hdrs)
 511                 return NULL;
 512
 513         hdr_daddrs = kvmalloc_node(sizeof(*hdr_daddrs) * num_hdrs, GFP_KERNEL,
 514                                    ibdev_to_node(dev->dev));
 515         if (!hdr_daddrs) {
 516                 kvfree(hdrs);
 517                 return NULL;
 518         }
 519
 520         for (i = 0; i < num_hdrs; i++) {
 521                 hdrs[i] = rds_dma_hdr_alloc(dev->dev, &hdr_daddrs[i], dir);
 522                 if (!hdrs[i]) {
 523                         rds_dma_hdrs_free(dev, hdrs, hdr_daddrs, i, dir);
 524                         return NULL;
 525                 }
 526         }
 527
 528         *dma_addrs = hdr_daddrs;
 529         return hdrs;
 530 }
 531
 532 /*
 533  * This needs to be very careful to not leave IS_ERR pointers around for
 534  * cleanup to trip over.
 535  */
 536 static int rds_ib_setup_qp(struct rds_connection *conn)
 537 {
 538         struct rds_ib_connection *ic = conn->c_transport_data;
 539         struct ib_device *dev = ic->i_cm_id->device;
 540         struct ib_qp_init_attr attr;
 541         struct ib_cq_init_attr cq_attr = {};
 542         struct rds_ib_device *rds_ibdev;
 543         unsigned long max_wrs;
 544         int ret, fr_queue_space;
 545
 546         /*
 547          * It's normal to see a null device if an incoming connection races
 548          * with device removal, so we don't print a warning.
 549          */
 550         rds_ibdev = rds_ib_get_client_data(dev);
 551         if (!rds_ibdev)
 552                 return -EOPNOTSUPP;
 553
 554         /* The fr_queue_space is currently set to 512, to add extra space on
 555          * completion queue and send queue. This extra space is used for FRWR
 556          * registration and invalidation work requests
 557          */
 558         fr_queue_space = RDS_IB_DEFAULT_FR_WR;
 559
 560         /* add the conn now so that connection establishment has the dev */
 561         rds_ib_add_conn(rds_ibdev, conn);
 562
 563         max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_send_wr + 1 ?
 564                 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_send_wr;
 565         if (ic->i_send_ring.w_nr != max_wrs)
 566                 rds_ib_ring_resize(&ic->i_send_ring, max_wrs);
 567
 568         max_wrs = rds_ibdev->max_wrs < rds_ib_sysctl_max_recv_wr + 1 ?
 569                 rds_ibdev->max_wrs - 1 : rds_ib_sysctl_max_recv_wr;
 570         if (ic->i_recv_ring.w_nr != max_wrs)
 571                 rds_ib_ring_resize(&ic->i_recv_ring, max_wrs);
 572
 573         /* Protection domain and memory range */
 574         ic->i_pd = rds_ibdev->pd;
 575
 576         ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
 577         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 578         cq_attr.comp_vector = ic->i_scq_vector;
 579         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 580                                      rds_ib_cq_event_handler, conn,
 581                                      &cq_attr);
 582         if (IS_ERR(ic->i_send_cq)) {
 583                 ret = PTR_ERR(ic->i_send_cq);
 584                 ic->i_send_cq = NULL;
 585                 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
 586                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 587                 goto rds_ibdev_out;
 588         }
 589
 590         ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
 591         cq_attr.cqe = ic->i_recv_ring.w_nr;
 592         cq_attr.comp_vector = ic->i_rcq_vector;
 593         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 594                                      rds_ib_cq_event_handler, conn,
 595                                      &cq_attr);
 596         if (IS_ERR(ic->i_recv_cq)) {
 597                 ret = PTR_ERR(ic->i_recv_cq);
 598                 ic->i_recv_cq = NULL;
 599                 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
 600                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 601                 goto send_cq_out;
 602         }
 603
 604         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 605         if (ret) {
 606                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 607                 goto recv_cq_out;
 608         }
 609
 610         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 611         if (ret) {
 612                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 613                 goto recv_cq_out;
 614         }
 615
 616         /* XXX negotiate max send/recv with remote? */
 617         memset(&attr, 0, sizeof(attr));
 618         attr.event_handler = rds_ib_qp_event_handler;
 619         attr.qp_context = conn;
 620         /* + 1 to allow for the single ack message */
 621         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 622         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 623         attr.cap.max_send_sge = rds_ibdev->max_sge;
 624         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 625         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 626         attr.qp_type = IB_QPT_RC;
 627         attr.send_cq = ic->i_send_cq;
 628         attr.recv_cq = ic->i_recv_cq;
 629
 630         /*
 631          * XXX this can fail if max_*_wr is too large?  Are we supposed
 632          * to back off until we get a value that the hardware can support?
 633          */
 634         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 635         if (ret) {
 636                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 637                 goto recv_cq_out;
 638         }
 639
 640         ic->i_send_hdrs = rds_dma_hdrs_alloc(rds_ibdev, &ic->i_send_hdrs_dma,
 641                                              ic->i_send_ring.w_nr,
 642                                              DMA_TO_DEVICE);
 643         if (!ic->i_send_hdrs) {
 644                 ret = -ENOMEM;
 645                 rdsdebug("DMA send hdrs alloc failed\n");
 646                 goto qp_out;
 647         }
 648
 649         ic->i_recv_hdrs = rds_dma_hdrs_alloc(rds_ibdev, &ic->i_recv_hdrs_dma,
 650                                              ic->i_recv_ring.w_nr,
 651                                              DMA_FROM_DEVICE);
 652         if (!ic->i_recv_hdrs) {
 653                 ret = -ENOMEM;
 654                 rdsdebug("DMA recv hdrs alloc failed\n");
 655                 goto send_hdrs_dma_out;
 656         }
 657
 658         ic->i_ack = rds_dma_hdr_alloc(rds_ibdev->dev, &ic->i_ack_dma,
 659                                       DMA_TO_DEVICE);
 660         if (!ic->i_ack) {
 661                 ret = -ENOMEM;
 662                 rdsdebug("DMA ack header alloc failed\n");
 663                 goto recv_hdrs_dma_out;
 664         }
 665
 666         ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work),
 667                                               ic->i_send_ring.w_nr),
 668                                    ibdev_to_node(dev));
 669         if (!ic->i_sends) {
 670                 ret = -ENOMEM;
 671                 rdsdebug("send allocation failed\n");
 672                 goto ack_dma_out;
 673         }
 674
 675         ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work),
 676                                               ic->i_recv_ring.w_nr),
 677                                    ibdev_to_node(dev));
 678         if (!ic->i_recvs) {
 679                 ret = -ENOMEM;
 680                 rdsdebug("recv allocation failed\n");
 681                 goto sends_out;
 682         }
 683
 684         rds_ib_recv_init_ack(ic);
 685
 686         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 687                  ic->i_send_cq, ic->i_recv_cq);
 688
 689         goto out;
 690
 691 sends_out:
 692         vfree(ic->i_sends);
 693
 694 ack_dma_out:
 695         rds_dma_hdr_free(rds_ibdev->dev, ic->i_ack, ic->i_ack_dma,
 696                          DMA_TO_DEVICE);
 697         ic->i_ack = NULL;
 698
 699 recv_hdrs_dma_out:
 700         rds_dma_hdrs_free(rds_ibdev, ic->i_recv_hdrs, ic->i_recv_hdrs_dma,
 701                           ic->i_recv_ring.w_nr, DMA_FROM_DEVICE);
 702         ic->i_recv_hdrs = NULL;
 703         ic->i_recv_hdrs_dma = NULL;
 704
 705 send_hdrs_dma_out:
 706         rds_dma_hdrs_free(rds_ibdev, ic->i_send_hdrs, ic->i_send_hdrs_dma,
 707                           ic->i_send_ring.w_nr, DMA_TO_DEVICE);
 708         ic->i_send_hdrs = NULL;
 709         ic->i_send_hdrs_dma = NULL;
 710
 711 qp_out:
 712         rdma_destroy_qp(ic->i_cm_id);
 713 recv_cq_out:
 714         ib_destroy_cq(ic->i_recv_cq);
 715         ic->i_recv_cq = NULL;
 716 send_cq_out:
 717         ib_destroy_cq(ic->i_send_cq);
 718         ic->i_send_cq = NULL;
 719 rds_ibdev_out:
 720         rds_ib_remove_conn(rds_ibdev, conn);
 721 out:
 722         rds_ib_dev_put(rds_ibdev);
 723
 724         return ret;
 725 }
 726
 727 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6)
 728 {
 729         const union rds_ib_conn_priv *dp = event->param.conn.private_data;
 730         u8 data_len, major, minor;
 731         u32 version = 0;
 732         __be16 mask;
 733         u16 common;
 734
 735         /*
 736          * rdma_cm private data is odd - when there is any private data in the
 737          * request, we will be given a pretty large buffer without telling us the
 738          * original size. The only way to tell the difference is by looking at
 739          * the contents, which are initialized to zero.
 740          * If the protocol version fields aren't set, this is a connection attempt
 741          * from an older version. This could be 3.0 or 2.0 - we can't tell.
 742          * We really should have changed this for OFED 1.3 :-(
 743          */
 744
 745         /* Be paranoid. RDS always has privdata */
 746         if (!event->param.conn.private_data_len) {
 747                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 748                         "rejecting\n");
 749                 return 0;
 750         }
 751
 752         if (isv6) {
 753                 data_len = sizeof(struct rds6_ib_connect_private);
 754                 major = dp->ricp_v6.dp_protocol_major;
 755                 minor = dp->ricp_v6.dp_protocol_minor;
 756                 mask = dp->ricp_v6.dp_protocol_minor_mask;
 757         } else {
 758                 data_len = sizeof(struct rds_ib_connect_private);
 759                 major = dp->ricp_v4.dp_protocol_major;
 760                 minor = dp->ricp_v4.dp_protocol_minor;
 761                 mask = dp->ricp_v4.dp_protocol_minor_mask;
 762         }
 763
 764         /* Even if len is crap *now* I still want to check it. -ASG */
 765         if (event->param.conn.private_data_len < data_len || major == 0)
 766                 return RDS_PROTOCOL_4_0;
 767
 768         common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 769         if (major == 4 && common) {
 770                 version = RDS_PROTOCOL_4_0;
 771                 while ((common >>= 1) != 0)
 772                         version++;
 773         } else if (RDS_PROTOCOL_COMPAT_VERSION ==
 774                    RDS_PROTOCOL(major, minor)) {
 775                 version = RDS_PROTOCOL_COMPAT_VERSION;
 776         } else {
 777                 if (isv6)
 778                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n",
 779                                            &dp->ricp_v6.dp_saddr, major, minor);
 780                 else
 781                         printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 782                                            &dp->ricp_v4.dp_saddr, major, minor);
 783         }
 784         return version;
 785 }
 786
 787 #if IS_ENABLED(CONFIG_IPV6)
 788 /* Given an IPv6 address, find the net_device which hosts that address and
 789  * return its index.  This is used by the rds_ib_cm_handle_connect() code to
 790  * find the interface index of where an incoming request comes from when
 791  * the request is using a link local address.
 792  *
 793  * Note one problem in this search.  It is possible that two interfaces have
 794  * the same link local address.  Unfortunately, this cannot be solved unless
 795  * the underlying layer gives us the interface which an incoming RDMA connect
 796  * request comes from.
 797  */
 798 static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr)
 799 {
 800         struct net_device *dev;
 801         int idx = 0;
 802
 803         rcu_read_lock();
 804         for_each_netdev_rcu(net, dev) {
 805                 if (ipv6_chk_addr(net, addr, dev, 1)) {
 806                         idx = dev->ifindex;
 807                         break;
 808                 }
 809         }
 810         rcu_read_unlock();
 811
 812         return idx;
 813 }
 814 #endif
 815
 816 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 817                              struct rdma_cm_event *event, bool isv6)
 818 {
 819         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 820         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 821         const struct rds_ib_conn_priv_cmn *dp_cmn;
 822         struct rds_connection *conn = NULL;
 823         struct rds_ib_connection *ic = NULL;
 824         struct rdma_conn_param conn_param;
 825         const union rds_ib_conn_priv *dp;
 826         union rds_ib_conn_priv dp_rep;
 827         struct in6_addr s_mapped_addr;
 828         struct in6_addr d_mapped_addr;
 829         const struct in6_addr *saddr6;
 830         const struct in6_addr *daddr6;
 831         int destroy = 1;
 832         u32 ifindex = 0;
 833         u32 version;
 834         int err = 1;
 835
 836         /* Check whether the remote protocol version matches ours. */
 837         version = rds_ib_protocol_compatible(event, isv6);
 838         if (!version) {
 839                 err = RDS_RDMA_REJ_INCOMPAT;
 840                 goto out;
 841         }
 842
 843         dp = event->param.conn.private_data;
 844         if (isv6) {
 845 #if IS_ENABLED(CONFIG_IPV6)
 846                 dp_cmn = &dp->ricp_v6.dp_cmn;
 847                 saddr6 = &dp->ricp_v6.dp_saddr;
 848                 daddr6 = &dp->ricp_v6.dp_daddr;
 849                 /* If either address is link local, need to find the
 850                  * interface index in order to create a proper RDS
 851                  * connection.
 852                  */
 853                 if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) {
 854                         /* Using init_net for now ..  */
 855                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 856                         /* No index found...  Need to bail out. */
 857                         if (ifindex == 0) {
 858                                 err = -EOPNOTSUPP;
 859                                 goto out;
 860                         }
 861                 } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) {
 862                         /* Use our address to find the correct index. */
 863                         ifindex = __rds_find_ifindex(&init_net, daddr6);
 864                         /* No index found...  Need to bail out. */
 865                         if (ifindex == 0) {
 866                                 err = -EOPNOTSUPP;
 867                                 goto out;
 868                         }
 869                 }
 870 #else
 871                 err = -EOPNOTSUPP;
 872                 goto out;
 873 #endif
 874         } else {
 875                 dp_cmn = &dp->ricp_v4.dp_cmn;
 876                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr);
 877                 ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
 878                 saddr6 = &s_mapped_addr;
 879                 daddr6 = &d_mapped_addr;
 880         }
 881
 882         rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid 0x%llx, tos:%d\n",
 883                  saddr6, daddr6, RDS_PROTOCOL_MAJOR(version),
 884                  RDS_PROTOCOL_MINOR(version),
 885                  (unsigned long long)be64_to_cpu(lguid),
 886                  (unsigned long long)be64_to_cpu(fguid), dp_cmn->ricpc_dp_toss);
 887
 888         /* RDS/IB is not currently netns aware, thus init_net */
 889         conn = rds_conn_create(&init_net, daddr6, saddr6,
 890                                &rds_ib_transport, dp_cmn->ricpc_dp_toss,
 891                                GFP_KERNEL, ifindex);
 892         if (IS_ERR(conn)) {
 893                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 894                 conn = NULL;
 895                 goto out;
 896         }
 897
 898         /*
 899          * The connection request may occur while the
 900          * previous connection exist, e.g. in case of failover.
 901          * But as connections may be initiated simultaneously
 902          * by both hosts, we have a random backoff mechanism -
 903          * see the comment above rds_queue_reconnect()
 904          */
 905         mutex_lock(&conn->c_cm_lock);
 906         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 907                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 908                         rdsdebug("incoming connect while connecting\n");
 909                         rds_conn_drop(conn);
 910                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 911                 } else
 912                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 913                         /* Wait and see - our connect may still be succeeding */
 914                         rds_ib_stats_inc(s_ib_connect_raced);
 915                 }
 916                 goto out;
 917         }
 918
 919         ic = conn->c_transport_data;
 920
 921         rds_ib_set_protocol(conn, version);
 922         rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit));
 923
 924         /* If the peer gave us the last packet it saw, process this as if
 925          * we had received a regular ACK. */
 926         if (dp_cmn->ricpc_ack_seq)
 927                 rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq),
 928                                     NULL);
 929
 930         BUG_ON(cm_id->context);
 931         BUG_ON(ic->i_cm_id);
 932
 933         ic->i_cm_id = cm_id;
 934         cm_id->context = conn;
 935
 936         /* We got halfway through setting up the ib_connection, if we
 937          * fail now, we have to take the long route out of this mess. */
 938         destroy = 0;
 939
 940         err = rds_ib_setup_qp(conn);
 941         if (err) {
 942                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 943                 goto out;
 944         }
 945
 946         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 947                                   event->param.conn.responder_resources,
 948                                   event->param.conn.initiator_depth, isv6);
 949
 950         /* rdma_accept() calls rdma_reject() internally if it fails */
 951         if (rdma_accept(cm_id, &conn_param))
 952                 rds_ib_conn_error(conn, "rdma_accept failed\n");
 953
 954 out:
 955         if (conn)
 956                 mutex_unlock(&conn->c_cm_lock);
 957         if (err)
 958                 rdma_reject(cm_id, &err, sizeof(int),
 959                             IB_CM_REJ_CONSUMER_DEFINED);
 960         return destroy;
 961 }
 962
 963
 964 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
 965 {
 966         struct rds_connection *conn = cm_id->context;
 967         struct rds_ib_connection *ic = conn->c_transport_data;
 968         struct rdma_conn_param conn_param;
 969         union rds_ib_conn_priv dp;
 970         int ret;
 971
 972         /* If the peer doesn't do protocol negotiation, we must
 973          * default to RDSv3.0 */
 974         rds_ib_set_protocol(conn, RDS_PROTOCOL_4_1);
 975         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 976
 977         ret = rds_ib_setup_qp(conn);
 978         if (ret) {
 979                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 980                 goto out;
 981         }
 982
 983         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
 984                                   conn->c_proposed_version,
 985                                   UINT_MAX, UINT_MAX, isv6);
 986         ret = rdma_connect_locked(cm_id, &conn_param);
 987         if (ret)
 988                 rds_ib_conn_error(conn, "rdma_connect_locked failed (%d)\n",
 989                                   ret);
 990
 991 out:
 992         /* Beware - returning non-zero tells the rdma_cm to destroy
 993          * the cm_id. We should certainly not do it as long as we still
 994          * "own" the cm_id. */
 995         if (ret) {
 996                 if (ic->i_cm_id == cm_id)
 997                         ret = 0;
 998         }
 999         ic->i_active_side = true;
1000         return ret;
1001 }
1002
1003 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
1004 {
1005         struct rds_connection *conn = cp->cp_conn;
1006         struct sockaddr_storage src, dest;
1007         rdma_cm_event_handler handler;
1008         struct rds_ib_connection *ic;
1009         int ret;
1010
1011         ic = conn->c_transport_data;
1012
1013         /* XXX I wonder what affect the port space has */
1014         /* delegate cm event handler to rdma_transport */
1015 #if IS_ENABLED(CONFIG_IPV6)
1016         if (conn->c_isv6)
1017                 handler = rds6_rdma_cm_event_handler;
1018         else
1019 #endif
1020                 handler = rds_rdma_cm_event_handler;
1021         ic->i_cm_id = rdma_create_id(&init_net, handler, conn,
1022                                      RDMA_PS_TCP, IB_QPT_RC);
1023         if (IS_ERR(ic->i_cm_id)) {
1024                 ret = PTR_ERR(ic->i_cm_id);
1025                 ic->i_cm_id = NULL;
1026                 rdsdebug("rdma_create_id() failed: %d\n", ret);
1027                 goto out;
1028         }
1029
1030         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
1031
1032         if (ipv6_addr_v4mapped(&conn->c_faddr)) {
1033                 struct sockaddr_in *sin;
1034
1035                 sin = (struct sockaddr_in *)&src;
1036                 sin->sin_family = AF_INET;
1037                 sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3];
1038                 sin->sin_port = 0;
1039
1040                 sin = (struct sockaddr_in *)&dest;
1041                 sin->sin_family = AF_INET;
1042                 sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3];
1043                 sin->sin_port = htons(RDS_PORT);
1044         } else {
1045                 struct sockaddr_in6 *sin6;
1046
1047                 sin6 = (struct sockaddr_in6 *)&src;
1048                 sin6->sin6_family = AF_INET6;
1049                 sin6->sin6_addr = conn->c_laddr;
1050                 sin6->sin6_port = 0;
1051                 sin6->sin6_scope_id = conn->c_dev_if;
1052
1053                 sin6 = (struct sockaddr_in6 *)&dest;
1054                 sin6->sin6_family = AF_INET6;
1055                 sin6->sin6_addr = conn->c_faddr;
1056                 sin6->sin6_port = htons(RDS_CM_PORT);
1057                 sin6->sin6_scope_id = conn->c_dev_if;
1058         }
1059
1060         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
1061                                 (struct sockaddr *)&dest,
1062                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
1063         if (ret) {
1064                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
1065                          ret);
1066                 rdma_destroy_id(ic->i_cm_id);
1067                 ic->i_cm_id = NULL;
1068         }
1069
1070 out:
1071         return ret;
1072 }
1073
1074 /*
1075  * This is so careful about only cleaning up resources that were built up
1076  * so that it can be called at any point during startup.  In fact it
1077  * can be called multiple times for a given connection.
1078  */
1079 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
1080 {
1081         struct rds_connection *conn = cp->cp_conn;
1082         struct rds_ib_connection *ic = conn->c_transport_data;
1083         int err = 0;
1084
1085         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
1086                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
1087                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
1088
1089         if (ic->i_cm_id) {
1090                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
1091                 err = rdma_disconnect(ic->i_cm_id);
1092                 if (err) {
1093                         /* Actually this may happen quite frequently, when
1094                          * an outgoing connect raced with an incoming connect.
1095                          */
1096                         rdsdebug("failed to disconnect, cm: %p err %d\n",
1097                                 ic->i_cm_id, err);
1098                 }
1099
1100                 /* kick off "flush_worker" for all pools in order to reap
1101                  * all FRMR registrations that are still marked "FRMR_IS_INUSE"
1102                  */
1103                 rds_ib_flush_mrs();
1104
1105                 /*
1106                  * We want to wait for tx and rx completion to finish
1107                  * before we tear down the connection, but we have to be
1108                  * careful not to get stuck waiting on a send ring that
1109                  * only has unsignaled sends in it.  We've shutdown new
1110                  * sends before getting here so by waiting for signaled
1111                  * sends to complete we're ensured that there will be no
1112                  * more tx processing.
1113                  */
1114                 wait_event(rds_ib_ring_empty_wait,
1115                            rds_ib_ring_empty(&ic->i_recv_ring) &&
1116                            (atomic_read(&ic->i_signaled_sends) == 0) &&
1117                            (atomic_read(&ic->i_fastreg_inuse_count) == 0) &&
1118                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
1119                 tasklet_kill(&ic->i_send_tasklet);
1120                 tasklet_kill(&ic->i_recv_tasklet);
1121
1122                 atomic_set(&ic->i_cq_quiesce, 1);
1123
1124                 /* first destroy the ib state that generates callbacks */
1125                 if (ic->i_cm_id->qp)
1126                         rdma_destroy_qp(ic->i_cm_id);
1127                 if (ic->i_send_cq) {
1128                         if (ic->rds_ibdev)
1129                                 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
1130                         ib_destroy_cq(ic->i_send_cq);
1131                 }
1132
1133                 if (ic->i_recv_cq) {
1134                         if (ic->rds_ibdev)
1135                                 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
1136                         ib_destroy_cq(ic->i_recv_cq);
1137                 }
1138
1139                 if (ic->rds_ibdev) {
1140                         /* then free the resources that ib callbacks use */
1141                         if (ic->i_send_hdrs) {
1142                                 rds_dma_hdrs_free(ic->rds_ibdev,
1143                                                   ic->i_send_hdrs,
1144                                                   ic->i_send_hdrs_dma,
1145                                                   ic->i_send_ring.w_nr,
1146                                                   DMA_TO_DEVICE);
1147                                 ic->i_send_hdrs = NULL;
1148                                 ic->i_send_hdrs_dma = NULL;
1149                         }
1150
1151                         if (ic->i_recv_hdrs) {
1152                                 rds_dma_hdrs_free(ic->rds_ibdev,
1153                                                   ic->i_recv_hdrs,
1154                                                   ic->i_recv_hdrs_dma,
1155                                                   ic->i_recv_ring.w_nr,
1156                                                   DMA_FROM_DEVICE);
1157                                 ic->i_recv_hdrs = NULL;
1158                                 ic->i_recv_hdrs_dma = NULL;
1159                         }
1160
1161                         if (ic->i_ack) {
1162                                 rds_dma_hdr_free(ic->rds_ibdev->dev, ic->i_ack,
1163                                                  ic->i_ack_dma, DMA_TO_DEVICE);
1164                                 ic->i_ack = NULL;
1165                         }
1166                 } else {
1167                         WARN_ON(ic->i_send_hdrs);
1168                         WARN_ON(ic->i_send_hdrs_dma);
1169                         WARN_ON(ic->i_recv_hdrs);
1170                         WARN_ON(ic->i_recv_hdrs_dma);
1171                         WARN_ON(ic->i_ack);
1172                 }
1173
1174                 if (ic->i_sends)
1175                         rds_ib_send_clear_ring(ic);
1176                 if (ic->i_recvs)
1177                         rds_ib_recv_clear_ring(ic);
1178
1179                 rdma_destroy_id(ic->i_cm_id);
1180
1181                 /*
1182                  * Move connection back to the nodev list.
1183                  */
1184                 if (ic->rds_ibdev)
1185                         rds_ib_remove_conn(ic->rds_ibdev, conn);
1186
1187                 ic->i_cm_id = NULL;
1188                 ic->i_pd = NULL;
1189                 ic->i_send_cq = NULL;
1190                 ic->i_recv_cq = NULL;
1191         }
1192         BUG_ON(ic->rds_ibdev);
1193
1194         /* Clear pending transmit */
1195         if (ic->i_data_op) {
1196                 struct rds_message *rm;
1197
1198                 rm = container_of(ic->i_data_op, struct rds_message, data);
1199                 rds_message_put(rm);
1200                 ic->i_data_op = NULL;
1201         }
1202
1203         /* Clear the ACK state */
1204         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
1205 #ifdef KERNEL_HAS_ATOMIC64
1206         atomic64_set(&ic->i_ack_next, 0);
1207 #else
1208         ic->i_ack_next = 0;
1209 #endif
1210         ic->i_ack_recv = 0;
1211
1212         /* Clear flow control state */
1213         ic->i_flowctl = 0;
1214         atomic_set(&ic->i_credits, 0);
1215
1216         /* Re-init rings, but retain sizes. */
1217         rds_ib_ring_init(&ic->i_send_ring, ic->i_send_ring.w_nr);
1218         rds_ib_ring_init(&ic->i_recv_ring, ic->i_recv_ring.w_nr);
1219
1220         if (ic->i_ibinc) {
1221                 rds_inc_put(&ic->i_ibinc->ii_inc);
1222                 ic->i_ibinc = NULL;
1223         }
1224
1225         vfree(ic->i_sends);
1226         ic->i_sends = NULL;
1227         vfree(ic->i_recvs);
1228         ic->i_recvs = NULL;
1229         ic->i_active_side = false;
1230 }
1231
1232 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
1233 {
1234         struct rds_ib_connection *ic;
1235         unsigned long flags;
1236         int ret;
1237
1238         /* XXX too lazy? */
1239         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
1240         if (!ic)
1241                 return -ENOMEM;
1242
1243         ret = rds_ib_recv_alloc_caches(ic, gfp);
1244         if (ret) {
1245                 kfree(ic);
1246                 return ret;
1247         }
1248
1249         INIT_LIST_HEAD(&ic->ib_node);
1250         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
1251                      (unsigned long)ic);
1252         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
1253                      (unsigned long)ic);
1254         mutex_init(&ic->i_recv_mutex);
1255 #ifndef KERNEL_HAS_ATOMIC64
1256         spin_lock_init(&ic->i_ack_lock);
1257 #endif
1258         atomic_set(&ic->i_signaled_sends, 0);
1259         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
1260
1261         /*
1262          * rds_ib_conn_shutdown() waits for these to be emptied so they
1263          * must be initialized before it can be called.
1264          */
1265         rds_ib_ring_init(&ic->i_send_ring, 0);
1266         rds_ib_ring_init(&ic->i_recv_ring, 0);
1267
1268         ic->conn = conn;
1269         conn->c_transport_data = ic;
1270
1271         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
1272         list_add_tail(&ic->ib_node, &ib_nodev_conns);
1273         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
1274
1275
1276         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
1277         return 0;
1278 }
1279
1280 /*
1281  * Free a connection. Connection must be shut down and not set for reconnect.
1282  */
1283 void rds_ib_conn_free(void *arg)
1284 {
1285         struct rds_ib_connection *ic = arg;
1286         spinlock_t      *lock_ptr;
1287
1288         rdsdebug("ic %p\n", ic);
1289
1290         /*
1291          * Conn is either on a dev's list or on the nodev list.
1292          * A race with shutdown() or connect() would cause problems
1293          * (since rds_ibdev would change) but that should never happen.
1294          */
1295         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1296
1297         spin_lock_irq(lock_ptr);
1298         list_del(&ic->ib_node);
1299         spin_unlock_irq(lock_ptr);
1300
1301         rds_ib_recv_free_caches(ic);
1302
1303         kfree(ic);
1304 }
1305
1306
1307 /*
1308  * An error occurred on the connection
1309  */
1310 void
1311 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1312 {
1313         va_list ap;
1314
1315         rds_conn_drop(conn);
1316
1317         va_start(ap, fmt);
1318         vprintk(fmt, ap);
1319         va_end(ap);
1320 }