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