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