lustre/ptlrpc/events.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  *
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27
  28 #ifndef __KERNEL__
  29 # include <liblustre.h>
  30 #else
  31 # ifdef __mips64__
  32 #  include <linux/kernel.h>
  33 # endif
  34 #endif
  35 #include <obd_class.h>
  36 #include <lustre_net.h>
  37 #include "ptlrpc_internal.h"
  38
  39 lnet_handle_eq_t   ptlrpc_eq_h;
  40
  41 /*
  42  *  Client's outgoing request callback
  43  */
  44 void request_out_callback(lnet_event_t *ev)
  45 {
  46         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  47         struct ptlrpc_request *req = cbid->cbid_arg;
  48         ENTRY;
  49
  50         LASSERT (ev->type == LNET_EVENT_SEND ||
  51                  ev->type == LNET_EVENT_UNLINK);
  52         LASSERT (ev->unlinked);
  53
  54         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
  55                   "type %d, status %d", ev->type, ev->status);
  56
  57         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  58
  59                 /* Failed send: make it seem like the reply timed out, just
  60                  * like failing sends in client.c does currently...  */
  61
  62                 spin_lock(&req->rq_lock);
  63                 req->rq_net_err = 1;
  64                 spin_unlock(&req->rq_lock);
  65
  66                 ptlrpc_wake_client_req(req);
  67         }
  68
  69         /* these balance the references in ptl_send_rpc() */
  70         atomic_dec(&req->rq_import->imp_inflight);
  71         ptlrpc_req_finished(req);
  72
  73         EXIT;
  74 }
  75
  76 /*
  77  * Client's incoming reply callback
  78  */
  79 void reply_in_callback(lnet_event_t *ev)
  80 {
  81         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  82         struct ptlrpc_request *req = cbid->cbid_arg;
  83         ENTRY;
  84
  85         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
  86                   "type %d, status %d", ev->type, ev->status);
  87
  88         LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
  89         LASSERT(ev->md.start == req->rq_repbuf);
  90         LASSERT(ev->mlength <= req->rq_replen);
  91
  92         spin_lock(&req->rq_lock);
  93
  94         req->rq_receiving_reply = 0;
  95         req->rq_early = 0;
  96
  97         if (ev->status)
  98                 goto out_wake;
  99         if (ev->type == LNET_EVENT_UNLINK) {
 100                 req->rq_must_unlink = 0;
 101                 DEBUG_REQ(D_RPCTRACE, req, "unlink");
 102                 goto out_wake;
 103         }
 104
 105         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 106            for adaptive timeouts' early reply. */
 107         LASSERT(ev->md.options & LNET_MD_MANAGE_REMOTE);
 108
 109         if ((ev->offset == 0) &&
 110             lustre_msg_get_flags(req->rq_reqmsg) & MSG_AT_SUPPORT) {
 111                 /* Early reply */
 112                 DEBUG_REQ(D_ADAPTTO, req,
 113                           "Early reply received: mlen=%u offset=%d replen=%d "
 114                           "replied=%d", ev->mlength, ev->offset, req->rq_replen,
 115                           req->rq_replied);
 116
 117                 LASSERT(ev->mlength == lustre_msg_early_size());
 118                 req->rq_early_count++; /* number received, client side */
 119                 if (req->rq_replied)
 120                         /* If we already got the real reply, ignore the early
 121                            reply, but signal so we can unlink */
 122                         goto out_wake;
 123                 req->rq_early = 1;
 124                 req->rq_nob_received = ev->mlength;
 125                 /* repmsg points to early reply */
 126                 req->rq_repmsg = req->rq_repbuf;
 127                 /* And we're still receiving */
 128                 req->rq_receiving_reply = 1;
 129         } else {
 130                 /* Real reply */
 131                 req->rq_replied = 1;
 132                 req->rq_nob_received = ev->mlength;
 133                 /* repmsg points to real reply */
 134                 req->rq_repmsg = (struct lustre_msg *)((char *)req->rq_repbuf +
 135                                                        ev->offset);
 136                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 137                    so we must unlink in ptlrpc_unregister_reply */
 138                 DEBUG_REQ(D_INFO, req,
 139                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 140                           lustre_msg_get_flags(req->rq_reqmsg),
 141                           ev->mlength, ev->offset, req->rq_replen);
 142         }
 143
 144         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 145
 146 out_wake:
 147         /* NB don't unlock till after wakeup; req can disappear under us
 148          * since we don't have our own ref */
 149         ptlrpc_wake_client_req(req);
 150         spin_unlock(&req->rq_lock);
 151         EXIT;
 152 }
 153
 154 /*
 155  * Client's bulk has been written/read
 156  */
 157 void client_bulk_callback (lnet_event_t *ev)
 158 {
 159         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 160         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 161         ENTRY;
 162
 163         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB))
 164                 ev->status = -EIO;
 165
 166         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 167                   ev->type == LNET_EVENT_PUT) ||
 168                  (desc->bd_type == BULK_GET_SOURCE &&
 169                   ev->type == LNET_EVENT_GET) ||
 170                  ev->type == LNET_EVENT_UNLINK);
 171         LASSERT (ev->unlinked);
 172
 173         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 174                "event type %d, status %d, desc %p\n",
 175                ev->type, ev->status, desc);
 176
 177         spin_lock(&desc->bd_lock);
 178
 179         LASSERT(desc->bd_network_rw);
 180         desc->bd_network_rw = 0;
 181
 182         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 183                 desc->bd_success = 1;
 184                 desc->bd_nob_transferred = ev->mlength;
 185                 desc->bd_sender = ev->sender;
 186         }
 187
 188         /* NB don't unlock till after wakeup; desc can disappear under us
 189          * otherwise */
 190         ptlrpc_wake_client_req(desc->bd_req);
 191
 192         spin_unlock(&desc->bd_lock);
 193         EXIT;
 194 }
 195
 196 /*
 197  * Server's incoming request callback
 198  */
 199 void request_in_callback(lnet_event_t *ev)
 200 {
 201         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 202         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 203         struct ptlrpc_service             *service = rqbd->rqbd_service;
 204         struct ptlrpc_request             *req;
 205         ENTRY;
 206
 207         LASSERT (ev->type == LNET_EVENT_PUT ||
 208                  ev->type == LNET_EVENT_UNLINK);
 209         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 210         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 211                  rqbd->rqbd_buffer + service->srv_buf_size);
 212
 213         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 214                "event type %d, status %d, service %s\n",
 215                ev->type, ev->status, service->srv_name);
 216
 217         if (ev->unlinked) {
 218                 /* If this is the last request message to fit in the
 219                  * request buffer we can use the request object embedded in
 220                  * rqbd.  Note that if we failed to allocate a request,
 221                  * we'd have to re-post the rqbd, which we can't do in this
 222                  * context. */
 223                 req = &rqbd->rqbd_req;
 224                 memset(req, 0, sizeof (*req));
 225         } else {
 226                 LASSERT (ev->type == LNET_EVENT_PUT);
 227                 if (ev->status != 0) {
 228                         /* We moaned above already... */
 229                         return;
 230                 }
 231                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 232                 if (req == NULL) {
 233                         CERROR("Can't allocate incoming request descriptor: "
 234                                "Dropping %s RPC from %s\n",
 235                                service->srv_name,
 236                                libcfs_id2str(ev->initiator));
 237                         return;
 238                 }
 239         }
 240
 241         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 242          * flags are reset and scalars are zero.  We only set the message
 243          * size to non-zero if this was a successful receive. */
 244         req->rq_xid = ev->match_bits;
 245         req->rq_reqmsg = ev->md.start + ev->offset;
 246         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 247                 req->rq_reqlen = ev->mlength;
 248         do_gettimeofday(&req->rq_arrival_time);
 249         req->rq_peer = ev->initiator;
 250         req->rq_self = ev->target.nid;
 251         req->rq_rqbd = rqbd;
 252         req->rq_phase = RQ_PHASE_NEW;
 253 #ifdef CRAY_XT3
 254         req->rq_uid = ev->uid;
 255 #endif
 256         spin_lock_init(&req->rq_lock);
 257         sema_init(&req->rq_rs_sem, 1);
 258         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 259         atomic_set(&req->rq_refcount, 1);
 260         if (ev->type == LNET_EVENT_PUT)
 261                 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
 262
 263         spin_lock(&service->srv_lock);
 264
 265         req->rq_history_seq = service->srv_request_seq++;
 266         list_add_tail(&req->rq_history_list, &service->srv_request_history);
 267
 268         if (ev->unlinked) {
 269                 service->srv_nrqbd_receiving--;
 270                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 271                        service->srv_nrqbd_receiving);
 272
 273                 /* Normally, don't complain about 0 buffers posted; LNET won't
 274                  * drop incoming reqs since we set the portal lazy */
 275                 if (test_req_buffer_pressure &&
 276                     ev->type != LNET_EVENT_UNLINK &&
 277                     service->srv_nrqbd_receiving == 0)
 278                         CWARN("All %s request buffers busy\n",
 279                               service->srv_name);
 280
 281                 /* req takes over the network's ref on rqbd */
 282         } else {
 283                 /* req takes a ref on rqbd */
 284                 rqbd->rqbd_refcount++;
 285         }
 286
 287         list_add_tail(&req->rq_list, &service->srv_req_in_queue);
 288         service->srv_n_queued_reqs++;
 289
 290         /* NB everything can disappear under us once the request
 291          * has been queued and we unlock, so do the wake now... */
 292         cfs_waitq_signal(&service->srv_waitq);
 293
 294         spin_unlock(&service->srv_lock);
 295         EXIT;
 296 }
 297
 298 /*
 299  *  Server's outgoing reply callback
 300  */
 301 void reply_out_callback(lnet_event_t *ev)
 302 {
 303         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 304         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 305         struct ptlrpc_service     *svc = rs->rs_service;
 306         ENTRY;
 307
 308         LASSERT (ev->type == LNET_EVENT_SEND ||
 309                  ev->type == LNET_EVENT_ACK ||
 310                  ev->type == LNET_EVENT_UNLINK);
 311
 312         if (!rs->rs_difficult) {
 313                 /* 'Easy' replies have no further processing so I drop the
 314                  * net's ref on 'rs' */
 315                 LASSERT (ev->unlinked);
 316                 ptlrpc_rs_decref(rs);
 317                 atomic_dec (&svc->srv_outstanding_replies);
 318                 EXIT;
 319                 return;
 320         }
 321
 322         LASSERT (rs->rs_on_net);
 323
 324         if (ev->unlinked) {
 325                 /* Last network callback.  The net's ref on 'rs' stays put
 326                  * until ptlrpc_server_handle_reply() is done with it */
 327                 spin_lock(&svc->srv_lock);
 328                 rs->rs_on_net = 0;
 329                 ptlrpc_schedule_difficult_reply (rs);
 330                 spin_unlock(&svc->srv_lock);
 331         }
 332
 333         EXIT;
 334 }
 335
 336 /*
 337  * Server's bulk completion callback
 338  */
 339 void server_bulk_callback (lnet_event_t *ev)
 340 {
 341         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 342         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 343         ENTRY;
 344
 345         LASSERT (ev->type == LNET_EVENT_SEND ||
 346                  ev->type == LNET_EVENT_UNLINK ||
 347                  (desc->bd_type == BULK_PUT_SOURCE &&
 348                   ev->type == LNET_EVENT_ACK) ||
 349                  (desc->bd_type == BULK_GET_SINK &&
 350                   ev->type == LNET_EVENT_REPLY));
 351
 352         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 353                "event type %d, status %d, desc %p\n",
 354                ev->type, ev->status, desc);
 355
 356         spin_lock(&desc->bd_lock);
 357
 358         if ((ev->type == LNET_EVENT_ACK ||
 359              ev->type == LNET_EVENT_REPLY) &&
 360             ev->status == 0) {
 361                 /* We heard back from the peer, so even if we get this
 362                  * before the SENT event (oh yes we can), we know we
 363                  * read/wrote the peer buffer and how much... */
 364                 desc->bd_success = 1;
 365                 desc->bd_nob_transferred = ev->mlength;
 366                 desc->bd_sender = ev->sender;
 367         }
 368
 369         if (ev->unlinked) {
 370                 /* This is the last callback no matter what... */
 371                 desc->bd_network_rw = 0;
 372                 cfs_waitq_signal(&desc->bd_waitq);
 373         }
 374
 375         spin_unlock(&desc->bd_lock);
 376         EXIT;
 377 }
 378
 379 static void ptlrpc_master_callback(lnet_event_t *ev)
 380 {
 381         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 382         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 383
 384         /* Honestly, it's best to find out early. */
 385         LASSERT (cbid->cbid_arg != LP_POISON);
 386         LASSERT (callback == request_out_callback ||
 387                  callback == reply_in_callback ||
 388                  callback == client_bulk_callback ||
 389                  callback == request_in_callback ||
 390                  callback == reply_out_callback ||
 391                  callback == server_bulk_callback);
 392
 393         callback (ev);
 394 }
 395
 396 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 397                          lnet_process_id_t *peer, lnet_nid_t *self)
 398 {
 399         int               best_dist = 0;
 400         int               best_order = 0;
 401         int               count = 0;
 402         int               rc = -ENOENT;
 403         int               portals_compatibility;
 404         int               dist;
 405         int               order;
 406         lnet_nid_t        dst_nid;
 407         lnet_nid_t        src_nid;
 408
 409         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 410
 411         peer->pid = LUSTRE_SRV_LNET_PID;
 412
 413         /* Choose the matching UUID that's closest */
 414         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 415                 dist = LNetDist(dst_nid, &src_nid, &order);
 416                 if (dist < 0)
 417                         continue;
 418
 419                 if (dist == 0) {                /* local! use loopback LND */
 420                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 421                         rc = 0;
 422                         break;
 423                 }
 424
 425                 LASSERT (order >= 0);
 426                 if (rc < 0 ||
 427                     dist < best_dist ||
 428                     (dist == best_dist && order < best_order)) {
 429                         best_dist = dist;
 430                         best_order = order;
 431
 432                         if (portals_compatibility > 1) {
 433                                 /* Strong portals compatibility: Zero the nid's
 434                                  * NET, so if I'm reading new config logs, or
 435                                  * getting configured by (new) lconf I can
 436                                  * still talk to old servers. */
 437                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 438                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 439                         }
 440                         peer->nid = dst_nid;
 441                         *self = src_nid;
 442                         rc = 0;
 443                 }
 444         }
 445
 446         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 447         if (rc != 0)
 448                 CERROR("No NID found for %s\n", uuid->uuid);
 449         return rc;
 450 }
 451
 452 void ptlrpc_ni_fini(void)
 453 {
 454         cfs_waitq_t         waitq;
 455         struct l_wait_info  lwi;
 456         int                 rc;
 457         int                 retries;
 458
 459         /* Wait for the event queue to become idle since there may still be
 460          * messages in flight with pending events (i.e. the fire-and-forget
 461          * messages == client requests and "non-difficult" server
 462          * replies */
 463
 464         for (retries = 0;; retries++) {
 465                 rc = LNetEQFree(ptlrpc_eq_h);
 466                 switch (rc) {
 467                 default:
 468                         LBUG();
 469
 470                 case 0:
 471                         LNetNIFini();
 472                         return;
 473
 474                 case -EBUSY:
 475                         if (retries != 0)
 476                                 CWARN("Event queue still busy\n");
 477
 478                         /* Wait for a bit */
 479                         cfs_waitq_init(&waitq);
 480                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 481                         l_wait_event(waitq, 0, &lwi);
 482                         break;
 483                 }
 484         }
 485         /* notreached */
 486 }
 487
 488 lnet_pid_t ptl_get_pid(void)
 489 {
 490         lnet_pid_t        pid;
 491
 492 #ifndef  __KERNEL__
 493         pid = getpid();
 494 #else
 495         pid = LUSTRE_SRV_LNET_PID;
 496 #endif
 497         return pid;
 498 }
 499
 500 int ptlrpc_ni_init(void)
 501 {
 502         int              rc;
 503         lnet_pid_t       pid;
 504
 505         pid = ptl_get_pid();
 506         CDEBUG(D_NET, "My pid is: %x\n", pid);
 507
 508         /* We're not passing any limits yet... */
 509         rc = LNetNIInit(pid);
 510         if (rc < 0) {
 511                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 512                 return (-ENOENT);
 513         }
 514
 515         /* CAVEAT EMPTOR: how we process portals events is _radically_
 516          * different depending on... */
 517 #ifdef __KERNEL__
 518         /* kernel portals calls our master callback when events are added to
 519          * the event queue.  In fact lustre never pulls events off this queue,
 520          * so it's only sized for some debug history. */
 521         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 522 #else
 523         /* liblustre calls the master callback when it removes events from the
 524          * event queue.  The event queue has to be big enough not to drop
 525          * anything */
 526         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 527 #endif
 528         if (rc == 0)
 529                 return 0;
 530
 531         CERROR ("Failed to allocate event queue: %d\n", rc);
 532         LNetNIFini();
 533
 534         return (-ENOMEM);
 535 }
 536
 537 #ifndef __KERNEL__
 538 CFS_LIST_HEAD(liblustre_wait_callbacks);
 539 CFS_LIST_HEAD(liblustre_idle_callbacks);
 540 void *liblustre_services_callback;
 541
 542 void *
 543 liblustre_register_waitidle_callback (struct list_head *callback_list,
 544                                       const char *name,
 545                                       int (*fn)(void *arg), void *arg)
 546 {
 547         struct liblustre_wait_callback *llwc;
 548
 549         OBD_ALLOC(llwc, sizeof(*llwc));
 550         LASSERT (llwc != NULL);
 551
 552         llwc->llwc_name = name;
 553         llwc->llwc_fn = fn;
 554         llwc->llwc_arg = arg;
 555         list_add_tail(&llwc->llwc_list, callback_list);
 556
 557         return (llwc);
 558 }
 559
 560 void
 561 liblustre_deregister_waitidle_callback (void *opaque)
 562 {
 563         struct liblustre_wait_callback *llwc = opaque;
 564
 565         list_del(&llwc->llwc_list);
 566         OBD_FREE(llwc, sizeof(*llwc));
 567 }
 568
 569 void *
 570 liblustre_register_wait_callback (const char *name,
 571                                   int (*fn)(void *arg), void *arg)
 572 {
 573         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 574                                                     name, fn, arg);
 575 }
 576
 577 void
 578 liblustre_deregister_wait_callback (void *opaque)
 579 {
 580         liblustre_deregister_waitidle_callback(opaque);
 581 }
 582
 583 void *
 584 liblustre_register_idle_callback (const char *name,
 585                                   int (*fn)(void *arg), void *arg)
 586 {
 587         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 588                                                     name, fn, arg);
 589 }
 590
 591 void
 592 liblustre_deregister_idle_callback (void *opaque)
 593 {
 594         liblustre_deregister_waitidle_callback(opaque);
 595 }
 596
 597 int
 598 liblustre_check_events (int timeout)
 599 {
 600         lnet_event_t ev;
 601         int         rc;
 602         int         i;
 603         ENTRY;
 604
 605         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 606         if (rc == 0)
 607                 RETURN(0);
 608
 609         LASSERT (rc == -EOVERFLOW || rc == 1);
 610
 611         /* liblustre: no asynch callback so we can't affort to miss any
 612          * events... */
 613         if (rc == -EOVERFLOW) {
 614                 CERROR ("Dropped an event!!!\n");
 615                 abort();
 616         }
 617
 618         ptlrpc_master_callback (&ev);
 619         RETURN(1);
 620 }
 621
 622 int liblustre_waiting = 0;
 623
 624 int
 625 liblustre_wait_event (int timeout)
 626 {
 627         struct list_head               *tmp;
 628         struct liblustre_wait_callback *llwc;
 629         int                             found_something = 0;
 630
 631         /* single threaded recursion check... */
 632         liblustre_waiting = 1;
 633
 634         for (;;) {
 635                 /* Deal with all pending events */
 636                 while (liblustre_check_events(0))
 637                         found_something = 1;
 638
 639                 /* Give all registered callbacks a bite at the cherry */
 640                 list_for_each(tmp, &liblustre_wait_callbacks) {
 641                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 642                                           llwc_list);
 643
 644                         if (llwc->llwc_fn(llwc->llwc_arg))
 645                                 found_something = 1;
 646                 }
 647
 648                 if (found_something || timeout == 0)
 649                         break;
 650
 651                 /* Nothing so far, but I'm allowed to block... */
 652                 found_something = liblustre_check_events(timeout);
 653                 if (!found_something)           /* still nothing */
 654                         break;                  /* I timed out */
 655         }
 656
 657         liblustre_waiting = 0;
 658
 659         return found_something;
 660 }
 661
 662 void
 663 liblustre_wait_idle(void)
 664 {
 665         static int recursed = 0;
 666
 667         struct list_head               *tmp;
 668         struct liblustre_wait_callback *llwc;
 669         int                             idle = 0;
 670
 671         LASSERT(!recursed);
 672         recursed = 1;
 673
 674         do {
 675                 liblustre_wait_event(0);
 676
 677                 idle = 1;
 678
 679                 list_for_each(tmp, &liblustre_idle_callbacks) {
 680                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 681                                           llwc_list);
 682
 683                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 684                                 idle = 0;
 685                                 break;
 686                         }
 687                 }
 688
 689         } while (!idle);
 690
 691         recursed = 0;
 692 }
 693
 694 #endif /* __KERNEL__ */
 695
 696 int ptlrpc_init_portals(void)
 697 {
 698         int   rc = ptlrpc_ni_init();
 699
 700         if (rc != 0) {
 701                 CERROR("network initialisation failed\n");
 702                 return -EIO;
 703         }
 704 #ifndef __KERNEL__
 705         liblustre_services_callback =
 706                 liblustre_register_wait_callback("liblustre_check_services",
 707                                                  &liblustre_check_services, NULL);
 708 #endif
 709         rc = ptlrpcd_addref();
 710         if (rc == 0)
 711                 return 0;
 712
 713         CERROR("rpcd initialisation failed\n");
 714 #ifndef __KERNEL__
 715         liblustre_deregister_wait_callback(liblustre_services_callback);
 716 #endif
 717         ptlrpc_ni_fini();
 718         return rc;
 719 }
 720
 721 void ptlrpc_exit_portals(void)
 722 {
 723 #ifndef __KERNEL__
 724         liblustre_deregister_wait_callback(liblustre_services_callback);
 725 #endif
 726         ptlrpcd_decref();
 727         ptlrpc_ni_fini();
 728 }