lustre/ptlrpc/events.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_RPC
  33
  34 #include <libcfs/libcfs.h>
  35 #include <linux/kernel.h>
  36 #include <linux/delay.h>
  37 #include <obd_class.h>
  38 #include <lustre_net.h>
  39 #include <lustre_sec.h>
  40 #include "ptlrpc_internal.h"
  41
  42 lnet_handler_t ptlrpc_handler;
  43 struct percpu_ref ptlrpc_pending;
  44
  45 /*
  46  *  Client's outgoing request callback
  47  */
  48 void request_out_callback(struct lnet_event *ev)
  49 {
  50         struct ptlrpc_cb_id   *cbid = ev->md_user_ptr;
  51         struct ptlrpc_request *req = cbid->cbid_arg;
  52         bool                   wakeup = false;
  53         ENTRY;
  54
  55         LASSERT(ev->type == LNET_EVENT_SEND || ev->type == LNET_EVENT_UNLINK);
  56         LASSERT(ev->unlinked);
  57
  58         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == cfs_fail_val &&
  59                      CFS_FAIL_CHECK_RESET(OBD_FAIL_NET_ERROR_RPC,
  60                                           OBD_FAIL_OSP_PRECREATE_PAUSE |
  61                                           CFS_FAIL_ONCE)))
  62                 ev->status = -ECONNABORTED;
  63
  64         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  65
  66         /* Do not update imp_next_ping for connection request */
  67         if (lustre_msg_get_opc(req->rq_reqmsg) !=
  68             req->rq_import->imp_connect_op)
  69                 ptlrpc_pinger_sending_on_import(req->rq_import);
  70
  71         sptlrpc_request_out_callback(req);
  72
  73         spin_lock(&req->rq_lock);
  74         req->rq_real_sent = ktime_get_real_seconds();
  75         req->rq_req_unlinked = 1;
  76         /* reply_in_callback happened before request_out_callback? */
  77         if (req->rq_reply_unlinked)
  78                 wakeup = true;
  79
  80         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  81                 /* Failed send: make it seem like the reply timed out, just
  82                  * like failing sends in client.c does currently...  */
  83                 req->rq_net_err = 1;
  84                 wakeup = true;
  85         }
  86
  87         if (wakeup)
  88                 ptlrpc_client_wake_req(req);
  89
  90         spin_unlock(&req->rq_lock);
  91
  92         ptlrpc_req_finished(req);
  93         EXIT;
  94 }
  95
  96 /*
  97  * Client's incoming reply callback
  98  */
  99 void reply_in_callback(struct lnet_event *ev)
 100 {
 101         struct ptlrpc_cb_id   *cbid = ev->md_user_ptr;
 102         struct ptlrpc_request *req = cbid->cbid_arg;
 103         ENTRY;
 104
 105         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
 106
 107         LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 108         LASSERT(ev->md_start == req->rq_repbuf);
 109         LASSERT(ev->offset + ev->mlength <= req->rq_repbuf_len);
 110         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 111          * for adaptive timeouts' early reply.
 112          */
 113         LASSERT((ev->md_options & LNET_MD_MANAGE_REMOTE) != 0);
 114
 115         spin_lock(&req->rq_lock);
 116
 117         req->rq_receiving_reply = 0;
 118         req->rq_early = 0;
 119         if (ev->unlinked)
 120                 req->rq_reply_unlinked = 1;
 121
 122         if (ev->status)
 123                 goto out_wake;
 124
 125         if (ev->type == LNET_EVENT_UNLINK) {
 126                 LASSERT(ev->unlinked);
 127                 DEBUG_REQ(D_NET, req, "unlink");
 128                 goto out_wake;
 129         }
 130
 131         if (ev->mlength < ev->rlength ) {
 132                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 133                        req->rq_replen, ev->rlength, ev->offset);
 134                 req->rq_reply_truncated = 1;
 135                 req->rq_replied = 1;
 136                 req->rq_status = -EOVERFLOW;
 137                 req->rq_nob_received = ev->rlength + ev->offset;
 138                 goto out_wake;
 139         }
 140
 141         if ((ev->offset == 0) &&
 142             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 143                 /* Early reply */
 144                 DEBUG_REQ(D_ADAPTTO, req,
 145                           "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
 146                           ev->mlength, ev->offset,
 147                           req->rq_replen, req->rq_replied, ev->unlinked);
 148
 149                 req->rq_early_count++; /* number received, client side */
 150
 151                 /* already got the real reply or buffers are already unlinked */
 152                 if (req->rq_replied ||
 153                     req->rq_reply_unlinked == 1)
 154                         goto out_wake;
 155
 156                 req->rq_early = 1;
 157                 req->rq_reply_off = ev->offset;
 158                 req->rq_nob_received = ev->mlength;
 159                 /* And we're still receiving */
 160                 req->rq_receiving_reply = 1;
 161         } else {
 162                 /* Real reply */
 163                 req->rq_rep_swab_mask = 0;
 164                 req->rq_replied = 1;
 165                 /* Got reply, no resend required */
 166                 req->rq_resend = 0;
 167                 req->rq_reply_off = ev->offset;
 168                 req->rq_nob_received = ev->mlength;
 169                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 170                    so we must unlink in ptlrpc_unregister_reply */
 171                 DEBUG_REQ(D_INFO, req,
 172                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 173                           lustre_msg_get_flags(req->rq_reqmsg),
 174                           ev->mlength, ev->offset, req->rq_replen);
 175         }
 176
 177         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
 178                 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
 179
 180 out_wake:
 181         /* NB don't unlock till after wakeup; req can disappear under us
 182          * since we don't have our own ref */
 183         ptlrpc_client_wake_req(req);
 184         spin_unlock(&req->rq_lock);
 185         EXIT;
 186 }
 187
 188 /*
 189  * Client's bulk has been written/read
 190  */
 191 void client_bulk_callback(struct lnet_event *ev)
 192 {
 193         struct ptlrpc_cb_id     *cbid = ev->md_user_ptr;
 194         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 195         struct ptlrpc_request   *req;
 196         ENTRY;
 197
 198         LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
 199                  ev->type == LNET_EVENT_PUT) ||
 200                 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
 201                  ev->type == LNET_EVENT_GET) ||
 202                 ev->type == LNET_EVENT_UNLINK);
 203         LASSERT(ev->unlinked);
 204
 205         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 206                 ev->status = -EIO;
 207
 208         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 209                 ev->status = -EIO;
 210
 211         CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
 212                      "event type %d, status %d, desc %p\n",
 213                      ev->type, ev->status, desc);
 214
 215         spin_lock(&desc->bd_lock);
 216         req = desc->bd_req;
 217         LASSERT(desc->bd_refs > 0);
 218         desc->bd_refs--;
 219
 220         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 221                 desc->bd_nob_transferred += ev->mlength;
 222                 desc->bd_sender = ev->sender;
 223         } else {
 224                 /* start reconnect and resend if network error hit */
 225                 spin_lock(&req->rq_lock);
 226                 req->rq_net_err = 1;
 227                 spin_unlock(&req->rq_lock);
 228                 desc->bd_failure = 1;
 229         }
 230
 231
 232         /* NB don't unlock till after wakeup; desc can disappear under us
 233          * otherwise */
 234         if (desc->bd_refs == 0)
 235                 ptlrpc_client_wake_req(desc->bd_req);
 236
 237         spin_unlock(&desc->bd_lock);
 238         EXIT;
 239 }
 240
 241 /*
 242  * We will have percpt request history list for ptlrpc service in upcoming
 243  * patches because we don't want to be serialized by current per-service
 244  * history operations. So we require history ID can (somehow) show arriving
 245  * order w/o grabbing global lock, and user can sort them in userspace.
 246  *
 247  * This is how we generate history ID for ptlrpc_request:
 248  * ----------------------------------------------------
 249  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 250  * ----------------------------------------------------
 251  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 252  * ----------------------------------------------------
 253  *
 254  * it might not be precise but should be good enough.
 255  */
 256
 257 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 258
 259 #define REQS_SEC_SHIFT          32
 260 #define REQS_USEC_SHIFT         16
 261 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 262
 263 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 264                                    struct ptlrpc_request *req)
 265 {
 266         u64 sec = req->rq_arrival_time.tv_sec;
 267         u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
 268         u64 new_seq;
 269
 270         /* set sequence ID for request and add it to history list,
 271          * it must be called with hold svcpt::scp_lock */
 272
 273         new_seq = (sec << REQS_SEC_SHIFT) |
 274                   (usec << REQS_USEC_SHIFT) |
 275                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
 276
 277         if (new_seq > svcpt->scp_hist_seq) {
 278                 /* This handles the initial case of scp_hist_seq == 0 or
 279                  * we just jumped into a new time window */
 280                 svcpt->scp_hist_seq = new_seq;
 281         } else {
 282                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 283                 /* NB: increase sequence number in current usec bucket,
 284                  * however, it's possible that we used up all bits for
 285                  * sequence and jumped into the next usec bucket (future time),
 286                  * then we hope there will be less RPCs per bucket at some
 287                  * point, and sequence will catch up again */
 288                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
 289                 new_seq = svcpt->scp_hist_seq;
 290         }
 291
 292         req->rq_history_seq = new_seq;
 293
 294         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 295 }
 296
 297 /*
 298  * Server's incoming request callback
 299  */
 300 void request_in_callback(struct lnet_event *ev)
 301 {
 302         struct ptlrpc_cb_id               *cbid = ev->md_user_ptr;
 303         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 304         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 305         struct ptlrpc_service             *service = svcpt->scp_service;
 306         struct ptlrpc_request             *req;
 307         ENTRY;
 308
 309         LASSERT(ev->type == LNET_EVENT_PUT ||
 310                 ev->type == LNET_EVENT_UNLINK);
 311         LASSERT((char *)ev->md_start >= rqbd->rqbd_buffer);
 312         LASSERT((char *)ev->md_start + ev->offset + ev->mlength <=
 313                 rqbd->rqbd_buffer + service->srv_buf_size);
 314
 315         CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
 316                      "event type %d, status %d, service %s\n",
 317                      ev->type, ev->status, service->srv_name);
 318
 319         if (ev->unlinked) {
 320                 /* If this is the last request message to fit in the
 321                  * request buffer we can use the request object embedded in
 322                  * rqbd.  Note that if we failed to allocate a request,
 323                  * we'd have to re-post the rqbd, which we can't do in this
 324                  * context.
 325                  */
 326                 req = &rqbd->rqbd_req;
 327                 memset(req, 0, sizeof(*req));
 328         } else {
 329                 LASSERT(ev->type == LNET_EVENT_PUT);
 330                 if (ev->status != 0) /* We moaned above already... */
 331                         return;
 332                 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
 333                 if (req == NULL) {
 334                         CERROR("Can't allocate incoming request descriptor: "
 335                                "Dropping %s RPC from %s\n",
 336                                service->srv_name,
 337                                libcfs_id2str(ev->initiator));
 338                         return;
 339                 }
 340         }
 341
 342         ptlrpc_srv_req_init(req);
 343         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 344          * flags are reset and scalars are zero.  We only set the message
 345          * size to non-zero if this was a successful receive. */
 346         req->rq_xid = ev->match_bits;
 347         req->rq_reqbuf = ev->md_start + ev->offset;
 348         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 349                 req->rq_reqdata_len = ev->mlength;
 350         ktime_get_real_ts64(&req->rq_arrival_time);
 351         /* Multi-Rail: keep track of both initiator and source NID. */
 352         req->rq_peer = ev->initiator;
 353         req->rq_source = ev->source;
 354         req->rq_self = ev->target.nid;
 355         req->rq_rqbd = rqbd;
 356         req->rq_phase = RQ_PHASE_NEW;
 357         if (ev->type == LNET_EVENT_PUT)
 358                 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
 359                        req, req->rq_xid, ev->mlength);
 360
 361         CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
 362                 libcfs_id2str(req->rq_peer), libcfs_id2str(req->rq_source));
 363
 364         spin_lock(&svcpt->scp_lock);
 365
 366         ptlrpc_req_add_history(svcpt, req);
 367
 368         if (ev->unlinked) {
 369                 svcpt->scp_nrqbds_posted--;
 370                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 371                        svcpt->scp_nrqbds_posted);
 372
 373                 /* Normally, don't complain about 0 buffers posted; LNET won't
 374                  * drop incoming reqs since we set the portal lazy */
 375                 if (test_req_buffer_pressure &&
 376                     ev->type != LNET_EVENT_UNLINK &&
 377                     svcpt->scp_nrqbds_posted == 0)
 378                         CWARN("All %s request buffers busy\n",
 379                               service->srv_name);
 380
 381                 /* req takes over the network's ref on rqbd */
 382         } else {
 383                 /* req takes a ref on rqbd */
 384                 rqbd->rqbd_refcount++;
 385         }
 386
 387         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 388         svcpt->scp_nreqs_incoming++;
 389
 390         /* NB everything can disappear under us once the request
 391          * has been queued and we unlock, so do the wake now... */
 392         wake_up(&svcpt->scp_waitq);
 393
 394         spin_unlock(&svcpt->scp_lock);
 395         EXIT;
 396 }
 397
 398 /*
 399  *  Server's outgoing reply callback
 400  */
 401 void reply_out_callback(struct lnet_event *ev)
 402 {
 403         struct ptlrpc_cb_id       *cbid = ev->md_user_ptr;
 404         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 405         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 406         bool need_schedule = false;
 407
 408         ENTRY;
 409
 410         LASSERT(ev->type == LNET_EVENT_SEND ||
 411                 ev->type == LNET_EVENT_ACK ||
 412                 ev->type == LNET_EVENT_UNLINK);
 413
 414         if (!rs->rs_difficult) {
 415                 /* 'Easy' replies have no further processing so I drop the
 416                  * net's ref on 'rs'
 417                  */
 418                 LASSERT(ev->unlinked);
 419                 ptlrpc_rs_decref(rs);
 420                 EXIT;
 421                 return;
 422         }
 423
 424         if (ev->type == LNET_EVENT_SEND) {
 425                 spin_lock(&rs->rs_lock);
 426                 rs->rs_sent = 1;
 427                 /* If transaction was committed before the SEND, and the ACK
 428                  * is lost, then we need to schedule so ptlrpc_hr can unlink
 429                  * the MD.
 430                  */
 431                 if (rs->rs_handled)
 432                         need_schedule = true;
 433                 spin_unlock(&rs->rs_lock);
 434         }
 435
 436         if (ev->unlinked || need_schedule) {
 437                 LASSERT(rs->rs_sent);
 438
 439                 /* Last network callback. The net's ref on 'rs' stays put
 440                  * until ptlrpc_handle_rs() is done with it
 441                  */
 442                 spin_lock(&svcpt->scp_rep_lock);
 443                 spin_lock(&rs->rs_lock);
 444
 445                 rs->rs_unlinked = ev->unlinked;
 446                 if (!rs->rs_no_ack ||
 447                     rs->rs_transno <=
 448                     rs->rs_export->exp_obd->obd_last_committed ||
 449                     list_empty(&rs->rs_obd_list))
 450                         ptlrpc_schedule_difficult_reply(rs);
 451
 452                 spin_unlock(&rs->rs_lock);
 453                 spin_unlock(&svcpt->scp_rep_lock);
 454         }
 455         EXIT;
 456 }
 457
 458 #ifdef HAVE_SERVER_SUPPORT
 459 /*
 460  * Server's bulk completion callback
 461  */
 462 void server_bulk_callback(struct lnet_event *ev)
 463 {
 464         struct ptlrpc_cb_id     *cbid = ev->md_user_ptr;
 465         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 466         ENTRY;
 467
 468         LASSERT(ev->type == LNET_EVENT_SEND ||
 469                 ev->type == LNET_EVENT_UNLINK ||
 470                 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
 471                  ev->type == LNET_EVENT_ACK) ||
 472                 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
 473                  ev->type == LNET_EVENT_REPLY));
 474
 475         CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
 476                      "event type %d, status %d, desc %p\n",
 477                      ev->type, ev->status, desc);
 478
 479         spin_lock(&desc->bd_lock);
 480
 481         LASSERT(desc->bd_refs > 0);
 482
 483         if ((ev->type == LNET_EVENT_ACK ||
 484              ev->type == LNET_EVENT_REPLY) &&
 485             ev->status == 0) {
 486                 /* We heard back from the peer, so even if we get this
 487                  * before the SENT event (oh yes we can), we know we
 488                  * read/wrote the peer buffer and how much... */
 489                 desc->bd_nob_transferred += ev->mlength;
 490                 desc->bd_sender = ev->sender;
 491         }
 492
 493         if (ev->status != 0)
 494                 desc->bd_failure = 1;
 495
 496         if (ev->unlinked) {
 497                 desc->bd_refs--;
 498                 /* This is the last callback no matter what... */
 499                 if (desc->bd_refs == 0)
 500                         wake_up(&desc->bd_waitq);
 501         }
 502
 503         spin_unlock(&desc->bd_lock);
 504         EXIT;
 505 }
 506 #endif
 507
 508 static void ptlrpc_master_callback(struct lnet_event *ev)
 509 {
 510         struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
 511         void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
 512
 513         /* Honestly, it's best to find out early. */
 514         LASSERT(cbid->cbid_arg != LP_POISON);
 515         LASSERT(callback == request_out_callback ||
 516                 callback == reply_in_callback ||
 517                 callback == client_bulk_callback ||
 518                 callback == request_in_callback ||
 519                 callback == reply_out_callback
 520 #ifdef HAVE_SERVER_SUPPORT
 521                 || callback == server_bulk_callback
 522 #endif
 523                 );
 524
 525         callback(ev);
 526         if (ev->unlinked)
 527                 percpu_ref_put(&ptlrpc_pending);
 528 }
 529
 530 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
 531                         struct lnet_process_id *peer, lnet_nid_t *self)
 532 {
 533         int best_dist = 0;
 534         __u32 best_order = 0;
 535         int count = 0;
 536         int rc = -ENOENT;
 537         int dist;
 538         __u32 order;
 539         lnet_nid_t dst_nid;
 540         lnet_nid_t src_nid;
 541
 542         peer->pid = LNET_PID_LUSTRE;
 543
 544         /* Choose the matching UUID that's closest */
 545         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 546                 if (peer->nid != LNET_NID_ANY && LNET_NIDADDR(peer->nid) == 0 &&
 547                     LNET_NIDNET(dst_nid) != LNET_NIDNET(peer->nid))
 548                         continue;
 549
 550                 dist = LNetDist(dst_nid, &src_nid, &order);
 551                 if (dist < 0)
 552                         continue;
 553
 554                 if (dist == 0) {                /* local! use loopback LND */
 555                         peer->nid = *self = LNET_NID_LO_0;
 556                         rc = 0;
 557                         break;
 558                 }
 559
 560                 if (rc < 0 ||
 561                     dist < best_dist ||
 562                     (dist == best_dist && order < best_order)) {
 563                         best_dist = dist;
 564                         best_order = order;
 565
 566                         peer->nid = dst_nid;
 567                         *self = src_nid;
 568                         rc = 0;
 569                 }
 570         }
 571
 572         CDEBUG(D_NET, "%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 573         return rc;
 574 }
 575
 576 static struct completion ptlrpc_done;
 577
 578 static void ptlrpc_release(struct percpu_ref *ref)
 579 {
 580         complete(&ptlrpc_done);
 581 }
 582
 583 static void ptlrpc_ni_fini(void)
 584 {
 585         /* Wait for the event queue to become idle since there may still be
 586          * messages in flight with pending events (i.e. the fire-and-forget
 587          * messages == client requests and "non-difficult" server
 588          * replies */
 589
 590         init_completion(&ptlrpc_done);
 591         percpu_ref_kill(&ptlrpc_pending);
 592         wait_for_completion(&ptlrpc_done);
 593
 594         lnet_assert_handler_unused(ptlrpc_handler);
 595         LNetNIFini();
 596 }
 597
 598 lnet_pid_t ptl_get_pid(void)
 599 {
 600         return LNET_PID_LUSTRE;
 601 }
 602
 603 int ptlrpc_ni_init(void)
 604 {
 605         int rc;
 606         lnet_pid_t pid;
 607
 608         pid = ptl_get_pid();
 609         CDEBUG(D_NET, "My pid is: %x\n", pid);
 610
 611         /* We're not passing any limits yet... */
 612         rc = LNetNIInit(pid);
 613         if (rc < 0) {
 614                 CDEBUG(D_NET, "ptlrpc: Can't init network interface: rc = %d\n",
 615                        rc);
 616                 return rc;
 617         }
 618
 619         rc = percpu_ref_init(&ptlrpc_pending, ptlrpc_release, 0, GFP_KERNEL);
 620         if (rc) {
 621                 CERROR("ptlrpc: Can't init percpu refcount: rc = %d\n", rc);
 622                 return rc;
 623         }
 624         /* CAVEAT EMPTOR: how we process portals events is _radically_
 625          * different depending on...
 626          */
 627         /* kernel LNet calls our master callback when there are new event,
 628          * because we are guaranteed to get every event via callback,
 629          * so we just set EQ size to 0 to avoid overhread of serializing
 630          * enqueue/dequeue operations in LNet. */
 631         ptlrpc_handler = ptlrpc_master_callback;
 632         return 0;
 633 }
 634
 635 int ptlrpc_init_portals(void)
 636 {
 637         int   rc = ptlrpc_ni_init();
 638
 639         if (rc != 0) {
 640                 CERROR("network initialisation failed\n");
 641                 return rc;
 642         }
 643         rc = ptlrpcd_addref();
 644         if (rc == 0)
 645                 return 0;
 646
 647         CERROR("rpcd initialisation failed\n");
 648         ptlrpc_ni_fini();
 649         return rc;
 650 }
 651
 652 void ptlrpc_exit_portals(void)
 653 {
 654         ptlrpcd_decref();
 655         ptlrpc_ni_fini();
 656 }