net/rds/ib_cm.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/in.h>
  35 #include <linux/slab.h>
  36 #include <linux/vmalloc.h>
  37 #include <linux/ratelimit.h>
  38
  39 #include "rds_single_path.h"
  40 #include "rds.h"
  41 #include "ib.h"
  42
  43 /*
  44  * Set the selected protocol version
  45  */
  46 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version)
  47 {
  48         conn->c_version = version;
  49 }
  50
  51 /*
  52  * Set up flow control
  53  */
  54 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits)
  55 {
  56         struct rds_ib_connection *ic = conn->c_transport_data;
  57
  58         if (rds_ib_sysctl_flow_control && credits != 0) {
  59                 /* We're doing flow control */
  60                 ic->i_flowctl = 1;
  61                 rds_ib_send_add_credits(conn, credits);
  62         } else {
  63                 ic->i_flowctl = 0;
  64         }
  65 }
  66
  67 /*
  68  * Tune RNR behavior. Without flow control, we use a rather
  69  * low timeout, but not the absolute minimum - this should
  70  * be tunable.
  71  *
  72  * We already set the RNR retry count to 7 (which is the
  73  * smallest infinite number :-) above.
  74  * If flow control is off, we want to change this back to 0
  75  * so that we learn quickly when our credit accounting is
  76  * buggy.
  77  *
  78  * Caller passes in a qp_attr pointer - don't waste stack spacv
  79  * by allocation this twice.
  80  */
  81 static void
  82 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr)
  83 {
  84         int ret;
  85
  86         attr->min_rnr_timer = IB_RNR_TIMER_000_32;
  87         ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER);
  88         if (ret)
  89                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret);
  90 }
  91
  92 /*
  93  * Connection established.
  94  * We get here for both outgoing and incoming connection.
  95  */
  96 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event)
  97 {
  98         const struct rds_ib_connect_private *dp = NULL;
  99         struct rds_ib_connection *ic = conn->c_transport_data;
 100         struct ib_qp_attr qp_attr;
 101         int err;
 102
 103         if (event->param.conn.private_data_len >= sizeof(*dp)) {
 104                 dp = event->param.conn.private_data;
 105
 106                 /* make sure it isn't empty data */
 107                 if (dp->dp_protocol_major) {
 108                         rds_ib_set_protocol(conn,
 109                                 RDS_PROTOCOL(dp->dp_protocol_major,
 110                                 dp->dp_protocol_minor));
 111                         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 112                 }
 113         }
 114
 115         if (conn->c_version < RDS_PROTOCOL(3, 1)) {
 116                 pr_notice("RDS/IB: Connection <%pI4,%pI4> version %u.%u no longer supported\n",
 117                           &conn->c_laddr, &conn->c_faddr,
 118                           RDS_PROTOCOL_MAJOR(conn->c_version),
 119                           RDS_PROTOCOL_MINOR(conn->c_version));
 120                 rds_conn_destroy(conn);
 121                 return;
 122         } else {
 123                 pr_notice("RDS/IB: %s conn connected <%pI4,%pI4> version %u.%u%s\n",
 124                           ic->i_active_side ? "Active" : "Passive",
 125                           &conn->c_laddr, &conn->c_faddr,
 126                           RDS_PROTOCOL_MAJOR(conn->c_version),
 127                           RDS_PROTOCOL_MINOR(conn->c_version),
 128                           ic->i_flowctl ? ", flow control" : "");
 129         }
 130
 131         atomic_set(&ic->i_cq_quiesce, 0);
 132
 133         /* Init rings and fill recv. this needs to wait until protocol
 134          * negotiation is complete, since ring layout is different
 135          * from 3.1 to 4.1.
 136          */
 137         rds_ib_send_init_ring(ic);
 138         rds_ib_recv_init_ring(ic);
 139         /* Post receive buffers - as a side effect, this will update
 140          * the posted credit count. */
 141         rds_ib_recv_refill(conn, 1, GFP_KERNEL);
 142
 143         /* Tune RNR behavior */
 144         rds_ib_tune_rnr(ic, &qp_attr);
 145
 146         qp_attr.qp_state = IB_QPS_RTS;
 147         err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE);
 148         if (err)
 149                 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err);
 150
 151         /* update ib_device with this local ipaddr */
 152         err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr);
 153         if (err)
 154                 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n",
 155                         err);
 156
 157         /* If the peer gave us the last packet it saw, process this as if
 158          * we had received a regular ACK. */
 159         if (dp) {
 160                 /* dp structure start is not guaranteed to be 8 bytes aligned.
 161                  * Since dp_ack_seq is 64-bit extended load operations can be
 162                  * used so go through get_unaligned to avoid unaligned errors.
 163                  */
 164                 __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq);
 165
 166                 if (dp_ack_seq)
 167                         rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq),
 168                                             NULL);
 169         }
 170
 171         rds_connect_complete(conn);
 172 }
 173
 174 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 175                         struct rdma_conn_param *conn_param,
 176                         struct rds_ib_connect_private *dp,
 177                         u32 protocol_version,
 178                         u32 max_responder_resources,
 179                         u32 max_initiator_depth)
 180 {
 181         struct rds_ib_connection *ic = conn->c_transport_data;
 182         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 183
 184         memset(conn_param, 0, sizeof(struct rdma_conn_param));
 185
 186         conn_param->responder_resources =
 187                 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources);
 188         conn_param->initiator_depth =
 189                 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth);
 190         conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7);
 191         conn_param->rnr_retry_count = 7;
 192
 193         if (dp) {
 194                 memset(dp, 0, sizeof(*dp));
 195                 dp->dp_saddr = conn->c_laddr;
 196                 dp->dp_daddr = conn->c_faddr;
 197                 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version);
 198                 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version);
 199                 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS);
 200                 dp->dp_ack_seq = cpu_to_be64(rds_ib_piggyb_ack(ic));
 201
 202                 /* Advertise flow control */
 203                 if (ic->i_flowctl) {
 204                         unsigned int credits;
 205
 206                         credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits));
 207                         dp->dp_credit = cpu_to_be32(credits);
 208                         atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits);
 209                 }
 210
 211                 conn_param->private_data = dp;
 212                 conn_param->private_data_len = sizeof(*dp);
 213         }
 214 }
 215
 216 static void rds_ib_cq_event_handler(struct ib_event *event, void *data)
 217 {
 218         rdsdebug("event %u (%s) data %p\n",
 219                  event->event, ib_event_msg(event->event), data);
 220 }
 221
 222 /* Plucking the oldest entry from the ring can be done concurrently with
 223  * the thread refilling the ring.  Each ring operation is protected by
 224  * spinlocks and the transient state of refilling doesn't change the
 225  * recording of which entry is oldest.
 226  *
 227  * This relies on IB only calling one cq comp_handler for each cq so that
 228  * there will only be one caller of rds_recv_incoming() per RDS connection.
 229  */
 230 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
 231 {
 232         struct rds_connection *conn = context;
 233         struct rds_ib_connection *ic = conn->c_transport_data;
 234
 235         rdsdebug("conn %p cq %p\n", conn, cq);
 236
 237         rds_ib_stats_inc(s_ib_evt_handler_call);
 238
 239         tasklet_schedule(&ic->i_recv_tasklet);
 240 }
 241
 242 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
 243                      struct ib_wc *wcs)
 244 {
 245         int nr, i;
 246         struct ib_wc *wc;
 247
 248         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 249                 for (i = 0; i < nr; i++) {
 250                         wc = wcs + i;
 251                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 252                                  (unsigned long long)wc->wr_id, wc->status,
 253                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 254
 255                         if (wc->wr_id <= ic->i_send_ring.w_nr ||
 256                             wc->wr_id == RDS_IB_ACK_WR_ID)
 257                                 rds_ib_send_cqe_handler(ic, wc);
 258                         else
 259                                 rds_ib_mr_cqe_handler(ic, wc);
 260
 261                 }
 262         }
 263 }
 264
 265 static void rds_ib_tasklet_fn_send(unsigned long data)
 266 {
 267         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 268         struct rds_connection *conn = ic->conn;
 269
 270         rds_ib_stats_inc(s_ib_tasklet_call);
 271
 272         /* if cq has been already reaped, ignore incoming cq event */
 273         if (atomic_read(&ic->i_cq_quiesce))
 274                 return;
 275
 276         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 277         ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 278         poll_scq(ic, ic->i_send_cq, ic->i_send_wc);
 279
 280         if (rds_conn_up(conn) &&
 281             (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
 282             test_bit(0, &conn->c_map_queued)))
 283                 rds_send_xmit(&ic->conn->c_path[0]);
 284 }
 285
 286 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
 287                      struct ib_wc *wcs,
 288                      struct rds_ib_ack_state *ack_state)
 289 {
 290         int nr, i;
 291         struct ib_wc *wc;
 292
 293         while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) {
 294                 for (i = 0; i < nr; i++) {
 295                         wc = wcs + i;
 296                         rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
 297                                  (unsigned long long)wc->wr_id, wc->status,
 298                                  wc->byte_len, be32_to_cpu(wc->ex.imm_data));
 299
 300                         rds_ib_recv_cqe_handler(ic, wc, ack_state);
 301                 }
 302         }
 303 }
 304
 305 static void rds_ib_tasklet_fn_recv(unsigned long data)
 306 {
 307         struct rds_ib_connection *ic = (struct rds_ib_connection *)data;
 308         struct rds_connection *conn = ic->conn;
 309         struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
 310         struct rds_ib_ack_state state;
 311
 312         if (!rds_ibdev)
 313                 rds_conn_drop(conn);
 314
 315         rds_ib_stats_inc(s_ib_tasklet_call);
 316
 317         /* if cq has been already reaped, ignore incoming cq event */
 318         if (atomic_read(&ic->i_cq_quiesce))
 319                 return;
 320
 321         memset(&state, 0, sizeof(state));
 322         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 323         ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 324         poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state);
 325
 326         if (state.ack_next_valid)
 327                 rds_ib_set_ack(ic, state.ack_next, state.ack_required);
 328         if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) {
 329                 rds_send_drop_acked(conn, state.ack_recv, NULL);
 330                 ic->i_ack_recv = state.ack_recv;
 331         }
 332
 333         if (rds_conn_up(conn))
 334                 rds_ib_attempt_ack(ic);
 335 }
 336
 337 static void rds_ib_qp_event_handler(struct ib_event *event, void *data)
 338 {
 339         struct rds_connection *conn = data;
 340         struct rds_ib_connection *ic = conn->c_transport_data;
 341
 342         rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event,
 343                  ib_event_msg(event->event));
 344
 345         switch (event->event) {
 346         case IB_EVENT_COMM_EST:
 347                 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST);
 348                 break;
 349         default:
 350                 rdsdebug("Fatal QP Event %u (%s) "
 351                         "- connection %pI4->%pI4, reconnecting\n",
 352                         event->event, ib_event_msg(event->event),
 353                         &conn->c_laddr, &conn->c_faddr);
 354                 rds_conn_drop(conn);
 355                 break;
 356         }
 357 }
 358
 359 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context)
 360 {
 361         struct rds_connection *conn = context;
 362         struct rds_ib_connection *ic = conn->c_transport_data;
 363
 364         rdsdebug("conn %p cq %p\n", conn, cq);
 365
 366         rds_ib_stats_inc(s_ib_evt_handler_call);
 367
 368         tasklet_schedule(&ic->i_send_tasklet);
 369 }
 370
 371 static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev)
 372 {
 373         int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1];
 374         int index = rds_ibdev->dev->num_comp_vectors - 1;
 375         int i;
 376
 377         for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) {
 378                 if (rds_ibdev->vector_load[i] < min) {
 379                         index = i;
 380                         min = rds_ibdev->vector_load[i];
 381                 }
 382         }
 383
 384         rds_ibdev->vector_load[index]++;
 385         return index;
 386 }
 387
 388 static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index)
 389 {
 390         rds_ibdev->vector_load[index]--;
 391 }
 392
 393 /*
 394  * This needs to be very careful to not leave IS_ERR pointers around for
 395  * cleanup to trip over.
 396  */
 397 static int rds_ib_setup_qp(struct rds_connection *conn)
 398 {
 399         struct rds_ib_connection *ic = conn->c_transport_data;
 400         struct ib_device *dev = ic->i_cm_id->device;
 401         struct ib_qp_init_attr attr;
 402         struct ib_cq_init_attr cq_attr = {};
 403         struct rds_ib_device *rds_ibdev;
 404         int ret, fr_queue_space;
 405
 406         /*
 407          * It's normal to see a null device if an incoming connection races
 408          * with device removal, so we don't print a warning.
 409          */
 410         rds_ibdev = rds_ib_get_client_data(dev);
 411         if (!rds_ibdev)
 412                 return -EOPNOTSUPP;
 413
 414         /* The fr_queue_space is currently set to 512, to add extra space on
 415          * completion queue and send queue. This extra space is used for FRMR
 416          * registration and invalidation work requests
 417          */
 418         fr_queue_space = rds_ibdev->use_fastreg ?
 419                          (RDS_IB_DEFAULT_FR_WR + 1) +
 420                          (RDS_IB_DEFAULT_FR_INV_WR + 1)
 421                          : 0;
 422
 423         /* add the conn now so that connection establishment has the dev */
 424         rds_ib_add_conn(rds_ibdev, conn);
 425
 426         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 427                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 428         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 429                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 430
 431         /* Protection domain and memory range */
 432         ic->i_pd = rds_ibdev->pd;
 433
 434         ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
 435         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 436         cq_attr.comp_vector = ic->i_scq_vector;
 437         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 438                                      rds_ib_cq_event_handler, conn,
 439                                      &cq_attr);
 440         if (IS_ERR(ic->i_send_cq)) {
 441                 ret = PTR_ERR(ic->i_send_cq);
 442                 ic->i_send_cq = NULL;
 443                 ibdev_put_vector(rds_ibdev, ic->i_scq_vector);
 444                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 445                 goto rds_ibdev_out;
 446         }
 447
 448         ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
 449         cq_attr.cqe = ic->i_recv_ring.w_nr;
 450         cq_attr.comp_vector = ic->i_rcq_vector;
 451         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 452                                      rds_ib_cq_event_handler, conn,
 453                                      &cq_attr);
 454         if (IS_ERR(ic->i_recv_cq)) {
 455                 ret = PTR_ERR(ic->i_recv_cq);
 456                 ic->i_recv_cq = NULL;
 457                 ibdev_put_vector(rds_ibdev, ic->i_rcq_vector);
 458                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 459                 goto send_cq_out;
 460         }
 461
 462         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 463         if (ret) {
 464                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 465                 goto recv_cq_out;
 466         }
 467
 468         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 469         if (ret) {
 470                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 471                 goto recv_cq_out;
 472         }
 473
 474         /* XXX negotiate max send/recv with remote? */
 475         memset(&attr, 0, sizeof(attr));
 476         attr.event_handler = rds_ib_qp_event_handler;
 477         attr.qp_context = conn;
 478         /* + 1 to allow for the single ack message */
 479         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 480         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 481         attr.cap.max_send_sge = rds_ibdev->max_sge;
 482         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 483         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 484         attr.qp_type = IB_QPT_RC;
 485         attr.send_cq = ic->i_send_cq;
 486         attr.recv_cq = ic->i_recv_cq;
 487         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
 488         atomic_set(&ic->i_fastunreg_wrs, RDS_IB_DEFAULT_FR_INV_WR);
 489
 490         /*
 491          * XXX this can fail if max_*_wr is too large?  Are we supposed
 492          * to back off until we get a value that the hardware can support?
 493          */
 494         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 495         if (ret) {
 496                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 497                 goto recv_cq_out;
 498         }
 499
 500         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 501                                            ic->i_send_ring.w_nr *
 502                                                 sizeof(struct rds_header),
 503                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 504         if (!ic->i_send_hdrs) {
 505                 ret = -ENOMEM;
 506                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 507                 goto qp_out;
 508         }
 509
 510         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 511                                            ic->i_recv_ring.w_nr *
 512                                                 sizeof(struct rds_header),
 513                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 514         if (!ic->i_recv_hdrs) {
 515                 ret = -ENOMEM;
 516                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 517                 goto send_hdrs_dma_out;
 518         }
 519
 520         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 521                                        &ic->i_ack_dma, GFP_KERNEL);
 522         if (!ic->i_ack) {
 523                 ret = -ENOMEM;
 524                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 525                 goto recv_hdrs_dma_out;
 526         }
 527
 528         ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 529                                    ibdev_to_node(dev));
 530         if (!ic->i_sends) {
 531                 ret = -ENOMEM;
 532                 rdsdebug("send allocation failed\n");
 533                 goto ack_dma_out;
 534         }
 535
 536         ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 537                                    ibdev_to_node(dev));
 538         if (!ic->i_recvs) {
 539                 ret = -ENOMEM;
 540                 rdsdebug("recv allocation failed\n");
 541                 goto sends_out;
 542         }
 543
 544         rds_ib_recv_init_ack(ic);
 545
 546         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 547                  ic->i_send_cq, ic->i_recv_cq);
 548
 549         return ret;
 550
 551 sends_out:
 552         vfree(ic->i_sends);
 553 ack_dma_out:
 554         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 555                              ic->i_ack, ic->i_ack_dma);
 556 recv_hdrs_dma_out:
 557         ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr *
 558                                         sizeof(struct rds_header),
 559                                         ic->i_recv_hdrs, ic->i_recv_hdrs_dma);
 560 send_hdrs_dma_out:
 561         ib_dma_free_coherent(dev, ic->i_send_ring.w_nr *
 562                                         sizeof(struct rds_header),
 563                                         ic->i_send_hdrs, ic->i_send_hdrs_dma);
 564 qp_out:
 565         rdma_destroy_qp(ic->i_cm_id);
 566 recv_cq_out:
 567         if (!ib_destroy_cq(ic->i_recv_cq))
 568                 ic->i_recv_cq = NULL;
 569 send_cq_out:
 570         if (!ib_destroy_cq(ic->i_send_cq))
 571                 ic->i_send_cq = NULL;
 572 rds_ibdev_out:
 573         rds_ib_remove_conn(rds_ibdev, conn);
 574         rds_ib_dev_put(rds_ibdev);
 575
 576         return ret;
 577 }
 578
 579 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 580 {
 581         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 582         u16 common;
 583         u32 version = 0;
 584
 585         /*
 586          * rdma_cm private data is odd - when there is any private data in the
 587          * request, we will be given a pretty large buffer without telling us the
 588          * original size. The only way to tell the difference is by looking at
 589          * the contents, which are initialized to zero.
 590          * If the protocol version fields aren't set, this is a connection attempt
 591          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 592          * We really should have changed this for OFED 1.3 :-(
 593          */
 594
 595         /* Be paranoid. RDS always has privdata */
 596         if (!event->param.conn.private_data_len) {
 597                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 598                         "rejecting\n");
 599                 return 0;
 600         }
 601
 602         /* Even if len is crap *now* I still want to check it. -ASG */
 603         if (event->param.conn.private_data_len < sizeof (*dp) ||
 604             dp->dp_protocol_major == 0)
 605                 return RDS_PROTOCOL_3_0;
 606
 607         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 608         if (dp->dp_protocol_major == 3 && common) {
 609                 version = RDS_PROTOCOL_3_0;
 610                 while ((common >>= 1) != 0)
 611                         version++;
 612         } else
 613                 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 614                                 &dp->dp_saddr,
 615                                 dp->dp_protocol_major,
 616                                 dp->dp_protocol_minor);
 617         return version;
 618 }
 619
 620 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 621                                     struct rdma_cm_event *event)
 622 {
 623         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 624         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 625         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 626         struct rds_ib_connect_private dp_rep;
 627         struct rds_connection *conn = NULL;
 628         struct rds_ib_connection *ic = NULL;
 629         struct rdma_conn_param conn_param;
 630         u32 version;
 631         int err = 1, destroy = 1;
 632
 633         /* Check whether the remote protocol version matches ours. */
 634         version = rds_ib_protocol_compatible(event);
 635         if (!version)
 636                 goto out;
 637
 638         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 639                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 640                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 641                  (unsigned long long)be64_to_cpu(lguid),
 642                  (unsigned long long)be64_to_cpu(fguid));
 643
 644         /* RDS/IB is not currently netns aware, thus init_net */
 645         conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
 646                                &rds_ib_transport, GFP_KERNEL);
 647         if (IS_ERR(conn)) {
 648                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 649                 conn = NULL;
 650                 goto out;
 651         }
 652
 653         /*
 654          * The connection request may occur while the
 655          * previous connection exist, e.g. in case of failover.
 656          * But as connections may be initiated simultaneously
 657          * by both hosts, we have a random backoff mechanism -
 658          * see the comment above rds_queue_reconnect()
 659          */
 660         mutex_lock(&conn->c_cm_lock);
 661         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 662                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 663                         rdsdebug("incoming connect while connecting\n");
 664                         rds_conn_drop(conn);
 665                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 666                 } else
 667                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 668                         /* Wait and see - our connect may still be succeeding */
 669                         rds_ib_stats_inc(s_ib_connect_raced);
 670                 }
 671                 goto out;
 672         }
 673
 674         ic = conn->c_transport_data;
 675
 676         rds_ib_set_protocol(conn, version);
 677         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 678
 679         /* If the peer gave us the last packet it saw, process this as if
 680          * we had received a regular ACK. */
 681         if (dp->dp_ack_seq)
 682                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 683
 684         BUG_ON(cm_id->context);
 685         BUG_ON(ic->i_cm_id);
 686
 687         ic->i_cm_id = cm_id;
 688         cm_id->context = conn;
 689
 690         /* We got halfway through setting up the ib_connection, if we
 691          * fail now, we have to take the long route out of this mess. */
 692         destroy = 0;
 693
 694         err = rds_ib_setup_qp(conn);
 695         if (err) {
 696                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 697                 goto out;
 698         }
 699
 700         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 701                 event->param.conn.responder_resources,
 702                 event->param.conn.initiator_depth);
 703
 704         /* rdma_accept() calls rdma_reject() internally if it fails */
 705         if (rdma_accept(cm_id, &conn_param))
 706                 rds_ib_conn_error(conn, "rdma_accept failed\n");
 707
 708 out:
 709         if (conn)
 710                 mutex_unlock(&conn->c_cm_lock);
 711         if (err)
 712                 rdma_reject(cm_id, NULL, 0);
 713         return destroy;
 714 }
 715
 716
 717 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 718 {
 719         struct rds_connection *conn = cm_id->context;
 720         struct rds_ib_connection *ic = conn->c_transport_data;
 721         struct rdma_conn_param conn_param;
 722         struct rds_ib_connect_private dp;
 723         int ret;
 724
 725         /* If the peer doesn't do protocol negotiation, we must
 726          * default to RDSv3.0 */
 727         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 728         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 729
 730         ret = rds_ib_setup_qp(conn);
 731         if (ret) {
 732                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 733                 goto out;
 734         }
 735
 736         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 737                 UINT_MAX, UINT_MAX);
 738         ret = rdma_connect(cm_id, &conn_param);
 739         if (ret)
 740                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 741
 742 out:
 743         /* Beware - returning non-zero tells the rdma_cm to destroy
 744          * the cm_id. We should certainly not do it as long as we still
 745          * "own" the cm_id. */
 746         if (ret) {
 747                 if (ic->i_cm_id == cm_id)
 748                         ret = 0;
 749         }
 750         ic->i_active_side = true;
 751         return ret;
 752 }
 753
 754 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
 755 {
 756         struct rds_connection *conn = cp->cp_conn;
 757         struct rds_ib_connection *ic = conn->c_transport_data;
 758         struct sockaddr_in src, dest;
 759         int ret;
 760
 761         /* XXX I wonder what affect the port space has */
 762         /* delegate cm event handler to rdma_transport */
 763         ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn,
 764                                      RDMA_PS_TCP, IB_QPT_RC);
 765         if (IS_ERR(ic->i_cm_id)) {
 766                 ret = PTR_ERR(ic->i_cm_id);
 767                 ic->i_cm_id = NULL;
 768                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 769                 goto out;
 770         }
 771
 772         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 773
 774         src.sin_family = AF_INET;
 775         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 776         src.sin_port = (__force u16)htons(0);
 777
 778         dest.sin_family = AF_INET;
 779         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 780         dest.sin_port = (__force u16)htons(RDS_PORT);
 781
 782         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 783                                 (struct sockaddr *)&dest,
 784                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 785         if (ret) {
 786                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 787                          ret);
 788                 rdma_destroy_id(ic->i_cm_id);
 789                 ic->i_cm_id = NULL;
 790         }
 791
 792 out:
 793         return ret;
 794 }
 795
 796 /*
 797  * This is so careful about only cleaning up resources that were built up
 798  * so that it can be called at any point during startup.  In fact it
 799  * can be called multiple times for a given connection.
 800  */
 801 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
 802 {
 803         struct rds_connection *conn = cp->cp_conn;
 804         struct rds_ib_connection *ic = conn->c_transport_data;
 805         int err = 0;
 806
 807         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 808                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 809                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 810
 811         if (ic->i_cm_id) {
 812                 struct ib_device *dev = ic->i_cm_id->device;
 813
 814                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 815                 err = rdma_disconnect(ic->i_cm_id);
 816                 if (err) {
 817                         /* Actually this may happen quite frequently, when
 818                          * an outgoing connect raced with an incoming connect.
 819                          */
 820                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 821                                 ic->i_cm_id, err);
 822                 }
 823
 824                 /*
 825                  * We want to wait for tx and rx completion to finish
 826                  * before we tear down the connection, but we have to be
 827                  * careful not to get stuck waiting on a send ring that
 828                  * only has unsignaled sends in it.  We've shutdown new
 829                  * sends before getting here so by waiting for signaled
 830                  * sends to complete we're ensured that there will be no
 831                  * more tx processing.
 832                  */
 833                 wait_event(rds_ib_ring_empty_wait,
 834                            rds_ib_ring_empty(&ic->i_recv_ring) &&
 835                            (atomic_read(&ic->i_signaled_sends) == 0) &&
 836                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR) &&
 837                            (atomic_read(&ic->i_fastunreg_wrs) == RDS_IB_DEFAULT_FR_INV_WR));
 838                 tasklet_kill(&ic->i_send_tasklet);
 839                 tasklet_kill(&ic->i_recv_tasklet);
 840
 841                 atomic_set(&ic->i_cq_quiesce, 1);
 842
 843                 /* first destroy the ib state that generates callbacks */
 844                 if (ic->i_cm_id->qp)
 845                         rdma_destroy_qp(ic->i_cm_id);
 846                 if (ic->i_send_cq) {
 847                         if (ic->rds_ibdev)
 848                                 ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector);
 849                         ib_destroy_cq(ic->i_send_cq);
 850                 }
 851
 852                 if (ic->i_recv_cq) {
 853                         if (ic->rds_ibdev)
 854                                 ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector);
 855                         ib_destroy_cq(ic->i_recv_cq);
 856                 }
 857
 858                 /* then free the resources that ib callbacks use */
 859                 if (ic->i_send_hdrs)
 860                         ib_dma_free_coherent(dev,
 861                                            ic->i_send_ring.w_nr *
 862                                                 sizeof(struct rds_header),
 863                                            ic->i_send_hdrs,
 864                                            ic->i_send_hdrs_dma);
 865
 866                 if (ic->i_recv_hdrs)
 867                         ib_dma_free_coherent(dev,
 868                                            ic->i_recv_ring.w_nr *
 869                                                 sizeof(struct rds_header),
 870                                            ic->i_recv_hdrs,
 871                                            ic->i_recv_hdrs_dma);
 872
 873                 if (ic->i_ack)
 874                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 875                                              ic->i_ack, ic->i_ack_dma);
 876
 877                 if (ic->i_sends)
 878                         rds_ib_send_clear_ring(ic);
 879                 if (ic->i_recvs)
 880                         rds_ib_recv_clear_ring(ic);
 881
 882                 rdma_destroy_id(ic->i_cm_id);
 883
 884                 /*
 885                  * Move connection back to the nodev list.
 886                  */
 887                 if (ic->rds_ibdev)
 888                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 889
 890                 ic->i_cm_id = NULL;
 891                 ic->i_pd = NULL;
 892                 ic->i_send_cq = NULL;
 893                 ic->i_recv_cq = NULL;
 894                 ic->i_send_hdrs = NULL;
 895                 ic->i_recv_hdrs = NULL;
 896                 ic->i_ack = NULL;
 897         }
 898         BUG_ON(ic->rds_ibdev);
 899
 900         /* Clear pending transmit */
 901         if (ic->i_data_op) {
 902                 struct rds_message *rm;
 903
 904                 rm = container_of(ic->i_data_op, struct rds_message, data);
 905                 rds_message_put(rm);
 906                 ic->i_data_op = NULL;
 907         }
 908
 909         /* Clear the ACK state */
 910         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 911 #ifdef KERNEL_HAS_ATOMIC64
 912         atomic64_set(&ic->i_ack_next, 0);
 913 #else
 914         ic->i_ack_next = 0;
 915 #endif
 916         ic->i_ack_recv = 0;
 917
 918         /* Clear flow control state */
 919         ic->i_flowctl = 0;
 920         atomic_set(&ic->i_credits, 0);
 921
 922         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 923         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 924
 925         if (ic->i_ibinc) {
 926                 rds_inc_put(&ic->i_ibinc->ii_inc);
 927                 ic->i_ibinc = NULL;
 928         }
 929
 930         vfree(ic->i_sends);
 931         ic->i_sends = NULL;
 932         vfree(ic->i_recvs);
 933         ic->i_recvs = NULL;
 934         ic->i_active_side = false;
 935 }
 936
 937 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 938 {
 939         struct rds_ib_connection *ic;
 940         unsigned long flags;
 941         int ret;
 942
 943         /* XXX too lazy? */
 944         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
 945         if (!ic)
 946                 return -ENOMEM;
 947
 948         ret = rds_ib_recv_alloc_caches(ic);
 949         if (ret) {
 950                 kfree(ic);
 951                 return ret;
 952         }
 953
 954         INIT_LIST_HEAD(&ic->ib_node);
 955         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
 956                      (unsigned long)ic);
 957         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
 958                      (unsigned long)ic);
 959         mutex_init(&ic->i_recv_mutex);
 960 #ifndef KERNEL_HAS_ATOMIC64
 961         spin_lock_init(&ic->i_ack_lock);
 962 #endif
 963         atomic_set(&ic->i_signaled_sends, 0);
 964
 965         /*
 966          * rds_ib_conn_shutdown() waits for these to be emptied so they
 967          * must be initialized before it can be called.
 968          */
 969         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 970         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 971
 972         ic->conn = conn;
 973         conn->c_transport_data = ic;
 974
 975         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 976         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 977         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 978
 979
 980         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 981         return 0;
 982 }
 983
 984 /*
 985  * Free a connection. Connection must be shut down and not set for reconnect.
 986  */
 987 void rds_ib_conn_free(void *arg)
 988 {
 989         struct rds_ib_connection *ic = arg;
 990         spinlock_t      *lock_ptr;
 991
 992         rdsdebug("ic %p\n", ic);
 993
 994         /*
 995          * Conn is either on a dev's list or on the nodev list.
 996          * A race with shutdown() or connect() would cause problems
 997          * (since rds_ibdev would change) but that should never happen.
 998          */
 999         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
1000
1001         spin_lock_irq(lock_ptr);
1002         list_del(&ic->ib_node);
1003         spin_unlock_irq(lock_ptr);
1004
1005         rds_ib_recv_free_caches(ic);
1006
1007         kfree(ic);
1008 }
1009
1010
1011 /*
1012  * An error occurred on the connection
1013  */
1014 void
1015 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
1016 {
1017         va_list ap;
1018
1019         rds_conn_drop(conn);
1020
1021         va_start(ap, fmt);
1022         vprintk(fmt, ap);
1023         va_end(ap);
1024 }