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         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  59
  60         /* Do not update imp_next_ping for connection request */
  61         if (lustre_msg_get_opc(req->rq_reqmsg) !=
  62             req->rq_import->imp_connect_op)
  63                 ptlrpc_pinger_sending_on_import(req->rq_import);
  64
  65         sptlrpc_request_out_callback(req);
  66
  67         spin_lock(&req->rq_lock);
  68         req->rq_real_sent = ktime_get_real_seconds();
  69         req->rq_req_unlinked = 1;
  70         /* reply_in_callback happened before request_out_callback? */
  71         if (req->rq_reply_unlinked)
  72                 wakeup = true;
  73
  74         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  75                 /* Failed send: make it seem like the reply timed out, just
  76                  * like failing sends in client.c does currently...  */
  77                 req->rq_net_err = 1;
  78                 wakeup = true;
  79         }
  80
  81         if (wakeup)
  82                 ptlrpc_client_wake_req(req);
  83
  84         spin_unlock(&req->rq_lock);
  85
  86         ptlrpc_req_finished(req);
  87         EXIT;
  88 }
  89
  90 /*
  91  * Client's incoming reply callback
  92  */
  93 void reply_in_callback(struct lnet_event *ev)
  94 {
  95         struct ptlrpc_cb_id   *cbid = ev->md_user_ptr;
  96         struct ptlrpc_request *req = cbid->cbid_arg;
  97         ENTRY;
  98
  99         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
 100
 101         LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 102         LASSERT(ev->md_start == req->rq_repbuf);
 103         LASSERT(ev->offset + ev->mlength <= req->rq_repbuf_len);
 104         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 105          * for adaptive timeouts' early reply.
 106          */
 107         LASSERT((ev->md_options & LNET_MD_MANAGE_REMOTE) != 0);
 108
 109         spin_lock(&req->rq_lock);
 110
 111         req->rq_receiving_reply = 0;
 112         req->rq_early = 0;
 113         if (ev->unlinked)
 114                 req->rq_reply_unlinked = 1;
 115
 116         if (ev->status)
 117                 goto out_wake;
 118
 119         if (ev->type == LNET_EVENT_UNLINK) {
 120                 LASSERT(ev->unlinked);
 121                 DEBUG_REQ(D_NET, req, "unlink");
 122                 goto out_wake;
 123         }
 124
 125         if (ev->mlength < ev->rlength) {
 126                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 127                        req->rq_replen, ev->rlength, ev->offset);
 128                 req->rq_reply_truncated = 1;
 129                 req->rq_replied = 1;
 130                 req->rq_status = -EOVERFLOW;
 131                 req->rq_nob_received = ev->rlength + ev->offset;
 132                 goto out_wake;
 133         }
 134
 135         if ((ev->offset == 0) &&
 136             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 137                 /* Early reply */
 138                 DEBUG_REQ(D_ADAPTTO, req,
 139                           "Early reply received, mlen=%u offset=%d replen=%d replied=%d unlinked=%d",
 140                           ev->mlength, ev->offset,
 141                           req->rq_replen, req->rq_replied, ev->unlinked);
 142
 143                 req->rq_early_count++; /* number received, client side */
 144
 145                 /* already got the real reply or buffers are already unlinked */
 146                 if (req->rq_replied ||
 147                     req->rq_reply_unlinked == 1)
 148                         goto out_wake;
 149
 150                 req->rq_early = 1;
 151                 req->rq_reply_off = ev->offset;
 152                 req->rq_nob_received = ev->mlength;
 153                 /* And we're still receiving */
 154                 req->rq_receiving_reply = 1;
 155         } else {
 156                 /* Real reply */
 157                 req->rq_rep_swab_mask = 0;
 158                 req->rq_replied = 1;
 159                 /* Got reply, no resend required */
 160                 req->rq_resend = 0;
 161                 req->rq_reply_off = ev->offset;
 162                 req->rq_nob_received = ev->mlength;
 163                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 164                  * so we must unlink in ptlrpc_unregister_reply */
 165                 DEBUG_REQ(D_INFO, req,
 166                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 167                           lustre_msg_get_flags(req->rq_reqmsg), ev->mlength,
 168                           ev->offset, req->rq_replen);
 169         }
 170
 171         if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
 172                 req->rq_import->imp_last_reply_time = ktime_get_real_seconds();
 173
 174         if (req->rq_xid > req->rq_import->imp_highest_replied_xid)
 175                 req->rq_import->imp_highest_replied_xid = req->rq_xid;
 176
 177 out_wake:
 178         /* NB don't unlock till after wakeup; req can disappear under us
 179          * since we don't have our own ref */
 180         ptlrpc_client_wake_req(req);
 181         spin_unlock(&req->rq_lock);
 182         EXIT;
 183 }
 184
 185 /*
 186  * Client's bulk has been written/read
 187  */
 188 void client_bulk_callback(struct lnet_event *ev)
 189 {
 190         struct ptlrpc_cb_id     *cbid = ev->md_user_ptr;
 191         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 192         struct ptlrpc_request   *req;
 193         ENTRY;
 194
 195         LASSERT((ptlrpc_is_bulk_put_sink(desc->bd_type) &&
 196                  ev->type == LNET_EVENT_PUT) ||
 197                 (ptlrpc_is_bulk_get_source(desc->bd_type) &&
 198                  ev->type == LNET_EVENT_GET) ||
 199                 ev->type == LNET_EVENT_UNLINK);
 200         LASSERT(ev->unlinked);
 201
 202         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 203                 ev->status = -EIO;
 204
 205         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,
 206             CFS_FAIL_ONCE))
 207                 ev->status = -EIO;
 208
 209         CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
 210                      "event type %d, status %d, desc %p\n",
 211                      ev->type, ev->status, desc);
 212
 213         spin_lock(&desc->bd_lock);
 214         req = desc->bd_req;
 215         LASSERT(desc->bd_refs > 0);
 216         desc->bd_refs--;
 217
 218         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 219                 desc->bd_nob_transferred += ev->mlength;
 220                 desc->bd_sender = ev->sender;
 221         } else {
 222                 /* start reconnect and resend if network error hit */
 223                 spin_lock(&req->rq_lock);
 224                 req->rq_net_err = 1;
 225                 spin_unlock(&req->rq_lock);
 226                 desc->bd_failure = 1;
 227         }
 228
 229
 230         /* NB don't unlock till after wakeup; desc can disappear under us
 231          * otherwise */
 232         if (desc->bd_refs == 0)
 233                 ptlrpc_client_wake_req(desc->bd_req);
 234
 235         spin_unlock(&desc->bd_lock);
 236         EXIT;
 237 }
 238
 239 /*
 240  * We will have percpt request history list for ptlrpc service in upcoming
 241  * patches because we don't want to be serialized by current per-service
 242  * history operations. So we require history ID can (somehow) show arriving
 243  * order w/o grabbing global lock, and user can sort them in userspace.
 244  *
 245  * This is how we generate history ID for ptlrpc_request:
 246  * ----------------------------------------------------
 247  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 248  * ----------------------------------------------------
 249  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 250  * ----------------------------------------------------
 251  *
 252  * it might not be precise but should be good enough.
 253  */
 254
 255 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 256
 257 #define REQS_SEC_SHIFT          32
 258 #define REQS_USEC_SHIFT         16
 259 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 260
 261 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 262                                    struct ptlrpc_request *req)
 263 {
 264         u64 sec = req->rq_arrival_time.tv_sec;
 265         u32 usec = req->rq_arrival_time.tv_nsec / NSEC_PER_USEC / 16; /* usec / 16 */
 266         u64 new_seq;
 267
 268         /* set sequence ID for request and add it to history list,
 269          * it must be called with hold svcpt::scp_lock */
 270
 271         new_seq = (sec << REQS_SEC_SHIFT) |
 272                   (usec << REQS_USEC_SHIFT) |
 273                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
 274
 275         if (new_seq > svcpt->scp_hist_seq) {
 276                 /* This handles the initial case of scp_hist_seq == 0 or
 277                  * we just jumped into a new time window */
 278                 svcpt->scp_hist_seq = new_seq;
 279         } else {
 280                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 281                 /* NB: increase sequence number in current usec bucket,
 282                  * however, it's possible that we used up all bits for
 283                  * sequence and jumped into the next usec bucket (future time),
 284                  * then we hope there will be less RPCs per bucket at some
 285                  * point, and sequence will catch up again */
 286                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
 287                 new_seq = svcpt->scp_hist_seq;
 288         }
 289
 290         req->rq_history_seq = new_seq;
 291
 292         list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 293 }
 294
 295 /*
 296  * Server's incoming request callback
 297  */
 298 void request_in_callback(struct lnet_event *ev)
 299 {
 300         struct ptlrpc_cb_id               *cbid = ev->md_user_ptr;
 301         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 302         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 303         struct ptlrpc_service             *service = svcpt->scp_service;
 304         struct ptlrpc_request             *req;
 305         ENTRY;
 306
 307         LASSERT(ev->type == LNET_EVENT_PUT ||
 308                 ev->type == LNET_EVENT_UNLINK);
 309         LASSERT((char *)ev->md_start >= rqbd->rqbd_buffer);
 310         LASSERT((char *)ev->md_start + ev->offset + ev->mlength <=
 311                 rqbd->rqbd_buffer + service->srv_buf_size);
 312
 313         CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
 314                      "event type %d, status %d, service %s\n",
 315                      ev->type, ev->status, service->srv_name);
 316
 317         if (ev->unlinked) {
 318                 /* If this is the last request message to fit in the
 319                  * request buffer we can use the request object embedded in
 320                  * rqbd.  Note that if we failed to allocate a request,
 321                  * we'd have to re-post the rqbd, which we can't do in this
 322                  * context.
 323                  */
 324                 req = &rqbd->rqbd_req;
 325                 memset(req, 0, sizeof(*req));
 326         } else {
 327                 LASSERT(ev->type == LNET_EVENT_PUT);
 328                 if (ev->status != 0) /* We moaned above already... */
 329                         return;
 330                 req = ptlrpc_request_cache_alloc(GFP_ATOMIC);
 331                 if (req == NULL) {
 332                         CERROR("Can't allocate incoming request descriptor: Dropping %s RPC from %s\n",
 333                                 service->srv_name,
 334                                 libcfs_idstr(&ev->initiator));
 335                         return;
 336                 }
 337         }
 338
 339         ptlrpc_srv_req_init(req);
 340         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 341          * flags are reset and scalars are zero.  We only set the message
 342          * size to non-zero if this was a successful receive. */
 343         req->rq_xid = ev->match_bits;
 344         req->rq_reqbuf = ev->md_start + ev->offset;
 345         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 346                 req->rq_reqdata_len = ev->mlength;
 347         ktime_get_real_ts64(&req->rq_arrival_time);
 348         /* Multi-Rail: keep track of both initiator and source NID. */
 349         req->rq_peer = ev->initiator;
 350         req->rq_source = ev->source;
 351         req->rq_self = ev->target.nid;
 352         req->rq_rqbd = rqbd;
 353         req->rq_phase = RQ_PHASE_NEW;
 354         if (ev->type == LNET_EVENT_PUT)
 355                 CDEBUG(D_INFO, "incoming req@%p x%llu msgsize %u\n",
 356                        req, req->rq_xid, ev->mlength);
 357
 358         CDEBUG(D_RPCTRACE, "peer: %s (source: %s)\n",
 359                 libcfs_idstr(&req->rq_peer), libcfs_idstr(&req->rq_source));
 360
 361         spin_lock(&svcpt->scp_lock);
 362
 363         ptlrpc_req_add_history(svcpt, req);
 364
 365         if (ev->unlinked) {
 366                 svcpt->scp_nrqbds_posted--;
 367                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 368                        svcpt->scp_nrqbds_posted);
 369
 370                 /* Normally, don't complain about 0 buffers posted; LNET won't
 371                  * drop incoming reqs since we set the portal lazy */
 372                 if (test_req_buffer_pressure &&
 373                     ev->type != LNET_EVENT_UNLINK &&
 374                     svcpt->scp_nrqbds_posted == 0)
 375                         CWARN("All %s request buffers busy\n",
 376                               service->srv_name);
 377
 378                 /* req takes over the network's ref on rqbd */
 379         } else {
 380                 /* req takes a ref on rqbd */
 381                 rqbd->rqbd_refcount++;
 382         }
 383
 384         list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 385         svcpt->scp_nreqs_incoming++;
 386
 387         /* NB everything can disappear under us once the request
 388          * has been queued and we unlock, so do the wake now... */
 389         wake_up(&svcpt->scp_waitq);
 390
 391         spin_unlock(&svcpt->scp_lock);
 392         EXIT;
 393 }
 394
 395 /*
 396  *  Server's outgoing reply callback
 397  */
 398 void reply_out_callback(struct lnet_event *ev)
 399 {
 400         struct ptlrpc_cb_id       *cbid = ev->md_user_ptr;
 401         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 402         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 403         bool need_schedule = false;
 404
 405         ENTRY;
 406
 407         LASSERT(ev->type == LNET_EVENT_SEND ||
 408                 ev->type == LNET_EVENT_ACK ||
 409                 ev->type == LNET_EVENT_UNLINK);
 410
 411         if (!rs->rs_difficult) {
 412                 /* 'Easy' replies have no further processing so I drop the
 413                  * net's ref on 'rs'
 414                  */
 415                 LASSERT(ev->unlinked);
 416                 ptlrpc_rs_decref(rs);
 417                 EXIT;
 418                 return;
 419         }
 420
 421         if (ev->type == LNET_EVENT_SEND) {
 422                 spin_lock(&rs->rs_lock);
 423                 rs->rs_sent = 1;
 424                 /* If transaction was committed before the SEND, and the ACK
 425                  * is lost, then we need to schedule so ptlrpc_hr can unlink
 426                  * the MD.
 427                  */
 428                 if (rs->rs_handled)
 429                         need_schedule = true;
 430                 spin_unlock(&rs->rs_lock);
 431         }
 432
 433         if (ev->unlinked || need_schedule) {
 434                 LASSERT(rs->rs_sent);
 435
 436                 /* Last network callback. The net's ref on 'rs' stays put
 437                  * until ptlrpc_handle_rs() is done with it
 438                  */
 439                 spin_lock(&svcpt->scp_rep_lock);
 440                 spin_lock(&rs->rs_lock);
 441
 442                 rs->rs_unlinked = ev->unlinked;
 443                 if (!rs->rs_no_ack ||
 444                     rs->rs_transno <=
 445                     rs->rs_export->exp_obd->obd_last_committed ||
 446                     list_empty(&rs->rs_obd_list))
 447                         ptlrpc_schedule_difficult_reply(rs);
 448
 449                 spin_unlock(&rs->rs_lock);
 450                 spin_unlock(&svcpt->scp_rep_lock);
 451         }
 452         EXIT;
 453 }
 454
 455 #ifdef HAVE_SERVER_SUPPORT
 456 /*
 457  * Server's bulk completion callback
 458  */
 459 void server_bulk_callback(struct lnet_event *ev)
 460 {
 461         struct ptlrpc_cb_id     *cbid = ev->md_user_ptr;
 462         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 463         ENTRY;
 464
 465         LASSERT(ev->type == LNET_EVENT_SEND ||
 466                 ev->type == LNET_EVENT_UNLINK ||
 467                 (ptlrpc_is_bulk_put_source(desc->bd_type) &&
 468                  ev->type == LNET_EVENT_ACK) ||
 469                 (ptlrpc_is_bulk_get_sink(desc->bd_type) &&
 470                  ev->type == LNET_EVENT_REPLY));
 471
 472         CDEBUG_LIMIT((ev->status == 0) ? D_NET : D_ERROR,
 473                      "event type %d, status %d, desc %p\n",
 474                      ev->type, ev->status, desc);
 475
 476         spin_lock(&desc->bd_lock);
 477
 478         LASSERT(desc->bd_refs > 0);
 479
 480         if ((ev->type == LNET_EVENT_ACK ||
 481              ev->type == LNET_EVENT_REPLY) &&
 482             ev->status == 0) {
 483                 /* We heard back from the peer, so even if we get this
 484                  * before the SENT event (oh yes we can), we know we
 485                  * read/wrote the peer buffer and how much... */
 486                 desc->bd_nob_transferred += ev->mlength;
 487                 desc->bd_sender = ev->sender;
 488         }
 489
 490         if (ev->status != 0)
 491                 desc->bd_failure = 1;
 492
 493         if (ev->unlinked) {
 494                 desc->bd_refs--;
 495                 /* This is the last callback no matter what... */
 496                 if (desc->bd_refs == 0)
 497                         wake_up(&desc->bd_waitq);
 498         }
 499
 500         spin_unlock(&desc->bd_lock);
 501         EXIT;
 502 }
 503 #endif
 504
 505 static void ptlrpc_master_callback(struct lnet_event *ev)
 506 {
 507         struct ptlrpc_cb_id *cbid = ev->md_user_ptr;
 508         void (*callback)(struct lnet_event *ev) = cbid->cbid_fn;
 509
 510         /* Honestly, it's best to find out early. */
 511         LASSERT(cbid->cbid_arg != LP_POISON);
 512         LASSERT(callback == request_out_callback ||
 513                 callback == reply_in_callback ||
 514                 callback == client_bulk_callback ||
 515                 callback == request_in_callback ||
 516                 callback == reply_out_callback
 517 #ifdef HAVE_SERVER_SUPPORT
 518                 || callback == server_bulk_callback
 519 #endif
 520                 );
 521
 522         callback(ev);
 523         if (ev->unlinked)
 524                 percpu_ref_put(&ptlrpc_pending);
 525 }
 526
 527 int ptlrpc_uuid_to_peer(struct obd_uuid *uuid,
 528                         struct lnet_processid *peer,
 529                         struct lnet_nid *self,
 530                         u32 refnet)
 531 {
 532         int best_dist = 0;
 533         __u32 best_order = 0;
 534         int count = 0;
 535         int rc = -ENOENT;
 536         int dist;
 537         __u32 order;
 538         struct lnet_nid dst_nid;
 539         struct lnet_nid src_nid;
 540
 541         peer->pid = LNET_PID_LUSTRE;
 542
 543         /* Choose the matching UUID that's closest */
 544         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 545                 if (refnet != LNET_NET_ANY &&
 546                     LNET_NID_NET(&dst_nid) != refnet)
 547                         continue;
 548
 549                 dist = LNetDist(&dst_nid, &src_nid, &order);
 550                 if (dist < 0)
 551                         continue;
 552
 553                 if (dist == 0) {                /* local! use loopback LND */
 554                         lnet_nid4_to_nid(LNET_NID_LO_0, self);
 555                         peer->nid = *self;
 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_idstr(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, true);
 595         LNetNIFini();
 596 }
 597
 598 lnet_pid_t ptl_get_pid(void)
 599 {
 600         return LNET_PID_LUSTRE;
 601 }
 602
 603 static 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: rc = %d\n", rc);
 641                 return rc;
 642         }
 643         rc = ptlrpcd_addref();
 644         if (rc == 0)
 645                 return 0;
 646
 647         CERROR("rpcd initialisation failed: rc = %d\n", rc);
 648         ptlrpc_ni_fini();
 649         return rc;
 650 }
 651
 652 void ptlrpc_exit_portals(void)
 653 {
 654         ptlrpcd_decref();
 655         ptlrpc_ni_fini();
 656 }