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