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                 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed,"
 117                        " no longer supported\n",
 118                        &conn->c_faddr,
 119                        RDS_PROTOCOL_MAJOR(conn->c_version),
 120                        RDS_PROTOCOL_MINOR(conn->c_version));
 121                 rds_conn_destroy(conn);
 122                 return;
 123         } else {
 124                 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n",
 125                        &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         /*
 132          * Init rings and fill recv. this needs to wait until protocol negotiation
 133          * is complete, since ring layout is different from 3.0 to 3.1.
 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 ? RDS_IB_DEFAULT_FR_WR : 0);
 387
 388         /* add the conn now so that connection establishment has the dev */
 389         rds_ib_add_conn(rds_ibdev, conn);
 390
 391         if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1)
 392                 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1);
 393         if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1)
 394                 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1);
 395
 396         /* Protection domain and memory range */
 397         ic->i_pd = rds_ibdev->pd;
 398
 399         cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
 400
 401         ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
 402                                      rds_ib_cq_event_handler, conn,
 403                                      &cq_attr);
 404         if (IS_ERR(ic->i_send_cq)) {
 405                 ret = PTR_ERR(ic->i_send_cq);
 406                 ic->i_send_cq = NULL;
 407                 rdsdebug("ib_create_cq send failed: %d\n", ret);
 408                 goto rds_ibdev_out;
 409         }
 410
 411         cq_attr.cqe = ic->i_recv_ring.w_nr;
 412         ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
 413                                      rds_ib_cq_event_handler, conn,
 414                                      &cq_attr);
 415         if (IS_ERR(ic->i_recv_cq)) {
 416                 ret = PTR_ERR(ic->i_recv_cq);
 417                 ic->i_recv_cq = NULL;
 418                 rdsdebug("ib_create_cq recv failed: %d\n", ret);
 419                 goto send_cq_out;
 420         }
 421
 422         ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP);
 423         if (ret) {
 424                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
 425                 goto recv_cq_out;
 426         }
 427
 428         ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED);
 429         if (ret) {
 430                 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret);
 431                 goto recv_cq_out;
 432         }
 433
 434         /* XXX negotiate max send/recv with remote? */
 435         memset(&attr, 0, sizeof(attr));
 436         attr.event_handler = rds_ib_qp_event_handler;
 437         attr.qp_context = conn;
 438         /* + 1 to allow for the single ack message */
 439         attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1;
 440         attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1;
 441         attr.cap.max_send_sge = rds_ibdev->max_sge;
 442         attr.cap.max_recv_sge = RDS_IB_RECV_SGE;
 443         attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 444         attr.qp_type = IB_QPT_RC;
 445         attr.send_cq = ic->i_send_cq;
 446         attr.recv_cq = ic->i_recv_cq;
 447         atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR);
 448
 449         /*
 450          * XXX this can fail if max_*_wr is too large?  Are we supposed
 451          * to back off until we get a value that the hardware can support?
 452          */
 453         ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr);
 454         if (ret) {
 455                 rdsdebug("rdma_create_qp failed: %d\n", ret);
 456                 goto recv_cq_out;
 457         }
 458
 459         ic->i_send_hdrs = ib_dma_alloc_coherent(dev,
 460                                            ic->i_send_ring.w_nr *
 461                                                 sizeof(struct rds_header),
 462                                            &ic->i_send_hdrs_dma, GFP_KERNEL);
 463         if (!ic->i_send_hdrs) {
 464                 ret = -ENOMEM;
 465                 rdsdebug("ib_dma_alloc_coherent send failed\n");
 466                 goto qp_out;
 467         }
 468
 469         ic->i_recv_hdrs = ib_dma_alloc_coherent(dev,
 470                                            ic->i_recv_ring.w_nr *
 471                                                 sizeof(struct rds_header),
 472                                            &ic->i_recv_hdrs_dma, GFP_KERNEL);
 473         if (!ic->i_recv_hdrs) {
 474                 ret = -ENOMEM;
 475                 rdsdebug("ib_dma_alloc_coherent recv failed\n");
 476                 goto send_hdrs_dma_out;
 477         }
 478
 479         ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header),
 480                                        &ic->i_ack_dma, GFP_KERNEL);
 481         if (!ic->i_ack) {
 482                 ret = -ENOMEM;
 483                 rdsdebug("ib_dma_alloc_coherent ack failed\n");
 484                 goto recv_hdrs_dma_out;
 485         }
 486
 487         ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work),
 488                                    ibdev_to_node(dev));
 489         if (!ic->i_sends) {
 490                 ret = -ENOMEM;
 491                 rdsdebug("send allocation failed\n");
 492                 goto ack_dma_out;
 493         }
 494
 495         ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work),
 496                                    ibdev_to_node(dev));
 497         if (!ic->i_recvs) {
 498                 ret = -ENOMEM;
 499                 rdsdebug("recv allocation failed\n");
 500                 goto sends_out;
 501         }
 502
 503         rds_ib_recv_init_ack(ic);
 504
 505         rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd,
 506                  ic->i_send_cq, ic->i_recv_cq);
 507
 508         goto out;
 509
 510 sends_out:
 511         vfree(ic->i_sends);
 512 ack_dma_out:
 513         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 514                              ic->i_ack, ic->i_ack_dma);
 515 recv_hdrs_dma_out:
 516         ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr *
 517                                         sizeof(struct rds_header),
 518                                         ic->i_recv_hdrs, ic->i_recv_hdrs_dma);
 519 send_hdrs_dma_out:
 520         ib_dma_free_coherent(dev, ic->i_send_ring.w_nr *
 521                                         sizeof(struct rds_header),
 522                                         ic->i_send_hdrs, ic->i_send_hdrs_dma);
 523 qp_out:
 524         rdma_destroy_qp(ic->i_cm_id);
 525 recv_cq_out:
 526         if (!ib_destroy_cq(ic->i_recv_cq))
 527                 ic->i_recv_cq = NULL;
 528 send_cq_out:
 529         if (!ib_destroy_cq(ic->i_send_cq))
 530                 ic->i_send_cq = NULL;
 531 rds_ibdev_out:
 532         rds_ib_remove_conn(rds_ibdev, conn);
 533 out:
 534         rds_ib_dev_put(rds_ibdev);
 535
 536         return ret;
 537 }
 538
 539 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event)
 540 {
 541         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 542         u16 common;
 543         u32 version = 0;
 544
 545         /*
 546          * rdma_cm private data is odd - when there is any private data in the
 547          * request, we will be given a pretty large buffer without telling us the
 548          * original size. The only way to tell the difference is by looking at
 549          * the contents, which are initialized to zero.
 550          * If the protocol version fields aren't set, this is a connection attempt
 551          * from an older version. This could could be 3.0 or 2.0 - we can't tell.
 552          * We really should have changed this for OFED 1.3 :-(
 553          */
 554
 555         /* Be paranoid. RDS always has privdata */
 556         if (!event->param.conn.private_data_len) {
 557                 printk(KERN_NOTICE "RDS incoming connection has no private data, "
 558                         "rejecting\n");
 559                 return 0;
 560         }
 561
 562         /* Even if len is crap *now* I still want to check it. -ASG */
 563         if (event->param.conn.private_data_len < sizeof (*dp) ||
 564             dp->dp_protocol_major == 0)
 565                 return RDS_PROTOCOL_3_0;
 566
 567         common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS;
 568         if (dp->dp_protocol_major == 3 && common) {
 569                 version = RDS_PROTOCOL_3_0;
 570                 while ((common >>= 1) != 0)
 571                         version++;
 572         } else
 573                 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n",
 574                                 &dp->dp_saddr,
 575                                 dp->dp_protocol_major,
 576                                 dp->dp_protocol_minor);
 577         return version;
 578 }
 579
 580 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 581                                     struct rdma_cm_event *event)
 582 {
 583         __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id;
 584         __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id;
 585         const struct rds_ib_connect_private *dp = event->param.conn.private_data;
 586         struct rds_ib_connect_private dp_rep;
 587         struct rds_connection *conn = NULL;
 588         struct rds_ib_connection *ic = NULL;
 589         struct rdma_conn_param conn_param;
 590         u32 version;
 591         int err = 1, destroy = 1;
 592
 593         /* Check whether the remote protocol version matches ours. */
 594         version = rds_ib_protocol_compatible(event);
 595         if (!version)
 596                 goto out;
 597
 598         rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid "
 599                  "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr,
 600                  RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version),
 601                  (unsigned long long)be64_to_cpu(lguid),
 602                  (unsigned long long)be64_to_cpu(fguid));
 603
 604         /* RDS/IB is not currently netns aware, thus init_net */
 605         conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr,
 606                                &rds_ib_transport, GFP_KERNEL);
 607         if (IS_ERR(conn)) {
 608                 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn));
 609                 conn = NULL;
 610                 goto out;
 611         }
 612
 613         /*
 614          * The connection request may occur while the
 615          * previous connection exist, e.g. in case of failover.
 616          * But as connections may be initiated simultaneously
 617          * by both hosts, we have a random backoff mechanism -
 618          * see the comment above rds_queue_reconnect()
 619          */
 620         mutex_lock(&conn->c_cm_lock);
 621         if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 622                 if (rds_conn_state(conn) == RDS_CONN_UP) {
 623                         rdsdebug("incoming connect while connecting\n");
 624                         rds_conn_drop(conn);
 625                         rds_ib_stats_inc(s_ib_listen_closed_stale);
 626                 } else
 627                 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) {
 628                         /* Wait and see - our connect may still be succeeding */
 629                         rds_ib_stats_inc(s_ib_connect_raced);
 630                 }
 631                 goto out;
 632         }
 633
 634         ic = conn->c_transport_data;
 635
 636         rds_ib_set_protocol(conn, version);
 637         rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit));
 638
 639         /* If the peer gave us the last packet it saw, process this as if
 640          * we had received a regular ACK. */
 641         if (dp->dp_ack_seq)
 642                 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL);
 643
 644         BUG_ON(cm_id->context);
 645         BUG_ON(ic->i_cm_id);
 646
 647         ic->i_cm_id = cm_id;
 648         cm_id->context = conn;
 649
 650         /* We got halfway through setting up the ib_connection, if we
 651          * fail now, we have to take the long route out of this mess. */
 652         destroy = 0;
 653
 654         err = rds_ib_setup_qp(conn);
 655         if (err) {
 656                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err);
 657                 goto out;
 658         }
 659
 660         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 661                 event->param.conn.responder_resources,
 662                 event->param.conn.initiator_depth);
 663
 664         /* rdma_accept() calls rdma_reject() internally if it fails */
 665         err = rdma_accept(cm_id, &conn_param);
 666         if (err)
 667                 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err);
 668
 669 out:
 670         if (conn)
 671                 mutex_unlock(&conn->c_cm_lock);
 672         if (err)
 673                 rdma_reject(cm_id, NULL, 0);
 674         return destroy;
 675 }
 676
 677
 678 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id)
 679 {
 680         struct rds_connection *conn = cm_id->context;
 681         struct rds_ib_connection *ic = conn->c_transport_data;
 682         struct rdma_conn_param conn_param;
 683         struct rds_ib_connect_private dp;
 684         int ret;
 685
 686         /* If the peer doesn't do protocol negotiation, we must
 687          * default to RDSv3.0 */
 688         rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0);
 689         ic->i_flowctl = rds_ib_sysctl_flow_control;     /* advertise flow control */
 690
 691         ret = rds_ib_setup_qp(conn);
 692         if (ret) {
 693                 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret);
 694                 goto out;
 695         }
 696
 697         rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION,
 698                 UINT_MAX, UINT_MAX);
 699         ret = rdma_connect(cm_id, &conn_param);
 700         if (ret)
 701                 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret);
 702
 703 out:
 704         /* Beware - returning non-zero tells the rdma_cm to destroy
 705          * the cm_id. We should certainly not do it as long as we still
 706          * "own" the cm_id. */
 707         if (ret) {
 708                 if (ic->i_cm_id == cm_id)
 709                         ret = 0;
 710         }
 711         return ret;
 712 }
 713
 714 int rds_ib_conn_path_connect(struct rds_conn_path *cp)
 715 {
 716         struct rds_connection *conn = cp->cp_conn;
 717         struct rds_ib_connection *ic = conn->c_transport_data;
 718         struct sockaddr_in src, dest;
 719         int ret;
 720
 721         /* XXX I wonder what affect the port space has */
 722         /* delegate cm event handler to rdma_transport */
 723         ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn,
 724                                      RDMA_PS_TCP, IB_QPT_RC);
 725         if (IS_ERR(ic->i_cm_id)) {
 726                 ret = PTR_ERR(ic->i_cm_id);
 727                 ic->i_cm_id = NULL;
 728                 rdsdebug("rdma_create_id() failed: %d\n", ret);
 729                 goto out;
 730         }
 731
 732         rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn);
 733
 734         src.sin_family = AF_INET;
 735         src.sin_addr.s_addr = (__force u32)conn->c_laddr;
 736         src.sin_port = (__force u16)htons(0);
 737
 738         dest.sin_family = AF_INET;
 739         dest.sin_addr.s_addr = (__force u32)conn->c_faddr;
 740         dest.sin_port = (__force u16)htons(RDS_PORT);
 741
 742         ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src,
 743                                 (struct sockaddr *)&dest,
 744                                 RDS_RDMA_RESOLVE_TIMEOUT_MS);
 745         if (ret) {
 746                 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id,
 747                          ret);
 748                 rdma_destroy_id(ic->i_cm_id);
 749                 ic->i_cm_id = NULL;
 750         }
 751
 752 out:
 753         return ret;
 754 }
 755
 756 /*
 757  * This is so careful about only cleaning up resources that were built up
 758  * so that it can be called at any point during startup.  In fact it
 759  * can be called multiple times for a given connection.
 760  */
 761 void rds_ib_conn_path_shutdown(struct rds_conn_path *cp)
 762 {
 763         struct rds_connection *conn = cp->cp_conn;
 764         struct rds_ib_connection *ic = conn->c_transport_data;
 765         int err = 0;
 766
 767         rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id,
 768                  ic->i_pd, ic->i_send_cq, ic->i_recv_cq,
 769                  ic->i_cm_id ? ic->i_cm_id->qp : NULL);
 770
 771         if (ic->i_cm_id) {
 772                 struct ib_device *dev = ic->i_cm_id->device;
 773
 774                 rdsdebug("disconnecting cm %p\n", ic->i_cm_id);
 775                 err = rdma_disconnect(ic->i_cm_id);
 776                 if (err) {
 777                         /* Actually this may happen quite frequently, when
 778                          * an outgoing connect raced with an incoming connect.
 779                          */
 780                         rdsdebug("failed to disconnect, cm: %p err %d\n",
 781                                 ic->i_cm_id, err);
 782                 }
 783
 784                 /*
 785                  * We want to wait for tx and rx completion to finish
 786                  * before we tear down the connection, but we have to be
 787                  * careful not to get stuck waiting on a send ring that
 788                  * only has unsignaled sends in it.  We've shutdown new
 789                  * sends before getting here so by waiting for signaled
 790                  * sends to complete we're ensured that there will be no
 791                  * more tx processing.
 792                  */
 793                 wait_event(rds_ib_ring_empty_wait,
 794                            rds_ib_ring_empty(&ic->i_recv_ring) &&
 795                            (atomic_read(&ic->i_signaled_sends) == 0) &&
 796                            (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR));
 797                 tasklet_kill(&ic->i_send_tasklet);
 798                 tasklet_kill(&ic->i_recv_tasklet);
 799
 800                 /* first destroy the ib state that generates callbacks */
 801                 if (ic->i_cm_id->qp)
 802                         rdma_destroy_qp(ic->i_cm_id);
 803                 if (ic->i_send_cq)
 804                         ib_destroy_cq(ic->i_send_cq);
 805                 if (ic->i_recv_cq)
 806                         ib_destroy_cq(ic->i_recv_cq);
 807
 808                 /* then free the resources that ib callbacks use */
 809                 if (ic->i_send_hdrs)
 810                         ib_dma_free_coherent(dev,
 811                                            ic->i_send_ring.w_nr *
 812                                                 sizeof(struct rds_header),
 813                                            ic->i_send_hdrs,
 814                                            ic->i_send_hdrs_dma);
 815
 816                 if (ic->i_recv_hdrs)
 817                         ib_dma_free_coherent(dev,
 818                                            ic->i_recv_ring.w_nr *
 819                                                 sizeof(struct rds_header),
 820                                            ic->i_recv_hdrs,
 821                                            ic->i_recv_hdrs_dma);
 822
 823                 if (ic->i_ack)
 824                         ib_dma_free_coherent(dev, sizeof(struct rds_header),
 825                                              ic->i_ack, ic->i_ack_dma);
 826
 827                 if (ic->i_sends)
 828                         rds_ib_send_clear_ring(ic);
 829                 if (ic->i_recvs)
 830                         rds_ib_recv_clear_ring(ic);
 831
 832                 rdma_destroy_id(ic->i_cm_id);
 833
 834                 /*
 835                  * Move connection back to the nodev list.
 836                  */
 837                 if (ic->rds_ibdev)
 838                         rds_ib_remove_conn(ic->rds_ibdev, conn);
 839
 840                 ic->i_cm_id = NULL;
 841                 ic->i_pd = NULL;
 842                 ic->i_send_cq = NULL;
 843                 ic->i_recv_cq = NULL;
 844                 ic->i_send_hdrs = NULL;
 845                 ic->i_recv_hdrs = NULL;
 846                 ic->i_ack = NULL;
 847         }
 848         BUG_ON(ic->rds_ibdev);
 849
 850         /* Clear pending transmit */
 851         if (ic->i_data_op) {
 852                 struct rds_message *rm;
 853
 854                 rm = container_of(ic->i_data_op, struct rds_message, data);
 855                 rds_message_put(rm);
 856                 ic->i_data_op = NULL;
 857         }
 858
 859         /* Clear the ACK state */
 860         clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
 861 #ifdef KERNEL_HAS_ATOMIC64
 862         atomic64_set(&ic->i_ack_next, 0);
 863 #else
 864         ic->i_ack_next = 0;
 865 #endif
 866         ic->i_ack_recv = 0;
 867
 868         /* Clear flow control state */
 869         ic->i_flowctl = 0;
 870         atomic_set(&ic->i_credits, 0);
 871
 872         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 873         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 874
 875         if (ic->i_ibinc) {
 876                 rds_inc_put(&ic->i_ibinc->ii_inc);
 877                 ic->i_ibinc = NULL;
 878         }
 879
 880         vfree(ic->i_sends);
 881         ic->i_sends = NULL;
 882         vfree(ic->i_recvs);
 883         ic->i_recvs = NULL;
 884 }
 885
 886 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 887 {
 888         struct rds_ib_connection *ic;
 889         unsigned long flags;
 890         int ret;
 891
 892         /* XXX too lazy? */
 893         ic = kzalloc(sizeof(struct rds_ib_connection), gfp);
 894         if (!ic)
 895                 return -ENOMEM;
 896
 897         ret = rds_ib_recv_alloc_caches(ic);
 898         if (ret) {
 899                 kfree(ic);
 900                 return ret;
 901         }
 902
 903         INIT_LIST_HEAD(&ic->ib_node);
 904         tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send,
 905                      (unsigned long)ic);
 906         tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv,
 907                      (unsigned long)ic);
 908         mutex_init(&ic->i_recv_mutex);
 909 #ifndef KERNEL_HAS_ATOMIC64
 910         spin_lock_init(&ic->i_ack_lock);
 911 #endif
 912         atomic_set(&ic->i_signaled_sends, 0);
 913
 914         /*
 915          * rds_ib_conn_shutdown() waits for these to be emptied so they
 916          * must be initialized before it can be called.
 917          */
 918         rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr);
 919         rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr);
 920
 921         ic->conn = conn;
 922         conn->c_transport_data = ic;
 923
 924         spin_lock_irqsave(&ib_nodev_conns_lock, flags);
 925         list_add_tail(&ic->ib_node, &ib_nodev_conns);
 926         spin_unlock_irqrestore(&ib_nodev_conns_lock, flags);
 927
 928
 929         rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data);
 930         return 0;
 931 }
 932
 933 /*
 934  * Free a connection. Connection must be shut down and not set for reconnect.
 935  */
 936 void rds_ib_conn_free(void *arg)
 937 {
 938         struct rds_ib_connection *ic = arg;
 939         spinlock_t      *lock_ptr;
 940
 941         rdsdebug("ic %p\n", ic);
 942
 943         /*
 944          * Conn is either on a dev's list or on the nodev list.
 945          * A race with shutdown() or connect() would cause problems
 946          * (since rds_ibdev would change) but that should never happen.
 947          */
 948         lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
 949
 950         spin_lock_irq(lock_ptr);
 951         list_del(&ic->ib_node);
 952         spin_unlock_irq(lock_ptr);
 953
 954         rds_ib_recv_free_caches(ic);
 955
 956         kfree(ic);
 957 }
 958
 959
 960 /*
 961  * An error occurred on the connection
 962  */
 963 void
 964 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...)
 965 {
 966         va_list ap;
 967
 968         rds_conn_drop(conn);
 969
 970         va_start(ap, fmt);
 971         vprintk(fmt, ap);
 972         va_end(ap);
 973 }