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 <lustre_sec.h>
  38 #include "ptlrpc_internal.h"
  39
  40 lnet_handle_eq_t   ptlrpc_eq_h;
  41
  42 /*
  43  *  Client's outgoing request callback
  44  */
  45 void request_out_callback(lnet_event_t *ev)
  46 {
  47         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  48         struct ptlrpc_request *req = cbid->cbid_arg;
  49         ENTRY;
  50
  51         LASSERT (ev->type == LNET_EVENT_SEND ||
  52                  ev->type == LNET_EVENT_UNLINK);
  53         LASSERT (ev->unlinked);
  54
  55         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
  56                   "type %d, status %d", ev->type, ev->status);
  57
  58         sptlrpc_request_out_callback(req);
  59
  60         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  61
  62                 /* Failed send: make it seem like the reply timed out, just
  63                  * like failing sends in client.c does currently...  */
  64
  65                 spin_lock(&req->rq_lock);
  66                 req->rq_net_err = 1;
  67                 spin_unlock(&req->rq_lock);
  68
  69                 ptlrpc_wake_client_req(req);
  70         }
  71
  72         ptlrpc_req_finished(req);
  73
  74         EXIT;
  75 }
  76
  77 /*
  78  * Client's incoming reply callback
  79  */
  80 void reply_in_callback(lnet_event_t *ev)
  81 {
  82         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  83         struct ptlrpc_request *req = cbid->cbid_arg;
  84         ENTRY;
  85
  86         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
  87                   "type %d, status %d", ev->type, ev->status);
  88
  89         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
  90         LASSERT (ev->md.start == req->rq_repbuf);
  91         LASSERT (ev->mlength <= req->rq_repbuf_len);
  92         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
  93            for adaptive timeouts' early reply. */
  94         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
  95
  96         spin_lock(&req->rq_lock);
  97
  98         req->rq_receiving_reply = 0;
  99         req->rq_early = 0;
 100
 101         if (ev->status)
 102                 goto out_wake;
 103         if (ev->type == LNET_EVENT_UNLINK) {
 104                 req->rq_must_unlink = 0;
 105                 DEBUG_REQ(D_RPCTRACE, req, "unlink");
 106                 goto out_wake;
 107         }
 108
 109         if ((ev->offset == 0) &&
 110             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 111                 /* Early reply */
 112                 DEBUG_REQ(D_ADAPTTO, req,
 113                           "Early reply received: mlen=%u offset=%d replen=%d "
 114                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 115                           req->rq_replen, req->rq_replied, ev->unlinked);
 116
 117                 req->rq_early_count++; /* number received, client side */
 118                 if (req->rq_replied) {
 119                         /* If we already got the real reply, then we need to
 120                          * check if lnet_finalize() unlinked the md.  In that
 121                          * case, there will be no further callback of type
 122                          * LNET_EVENT_UNLINK.
 123                          */
 124                         if (ev->unlinked)
 125                                 req->rq_must_unlink = 0;
 126                         else
 127                                 DEBUG_REQ(D_RPCTRACE, req, "unlinked in reply");
 128                         goto out_wake;
 129                 }
 130                 req->rq_early = 1;
 131                 req->rq_reply_off = ev->offset;
 132                 req->rq_nob_received = ev->mlength;
 133                 /* And we're still receiving */
 134                 req->rq_receiving_reply = 1;
 135         } else {
 136                 /* Real reply */
 137                 req->rq_replied = 1;
 138                 req->rq_reply_off = ev->offset;
 139                 req->rq_nob_received = ev->mlength;
 140                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 141                    so we must unlink in ptlrpc_unregister_reply */
 142                 DEBUG_REQ(D_INFO, req,
 143                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 144                           lustre_msg_get_flags(req->rq_reqmsg),
 145                           ev->mlength, ev->offset, req->rq_replen);
 146         }
 147
 148         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 149
 150 out_wake:
 151         /* NB don't unlock till after wakeup; req can disappear under us
 152          * since we don't have our own ref */
 153         ptlrpc_wake_client_req(req);
 154         spin_unlock(&req->rq_lock);
 155         EXIT;
 156 }
 157
 158 /*
 159  * Client's bulk has been written/read
 160  */
 161 void client_bulk_callback (lnet_event_t *ev)
 162 {
 163         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 164         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 165         ENTRY;
 166
 167         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 168                   ev->type == LNET_EVENT_PUT) ||
 169                  (desc->bd_type == BULK_GET_SOURCE &&
 170                   ev->type == LNET_EVENT_GET) ||
 171                  ev->type == LNET_EVENT_UNLINK);
 172         LASSERT (ev->unlinked);
 173
 174         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 175                "event type %d, status %d, desc %p\n",
 176                ev->type, ev->status, desc);
 177
 178         spin_lock(&desc->bd_lock);
 179
 180         LASSERT(desc->bd_network_rw);
 181         desc->bd_network_rw = 0;
 182
 183         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 184                 desc->bd_success = 1;
 185                 desc->bd_nob_transferred = ev->mlength;
 186                 desc->bd_sender = ev->sender;
 187         }
 188
 189         sptlrpc_enc_pool_put_pages(desc);
 190
 191         /* NB don't unlock till after wakeup; desc can disappear under us
 192          * otherwise */
 193         ptlrpc_wake_client_req(desc->bd_req);
 194
 195         spin_unlock(&desc->bd_lock);
 196         EXIT;
 197 }
 198
 199 /*
 200  * Server's incoming request callback
 201  */
 202 void request_in_callback(lnet_event_t *ev)
 203 {
 204         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 205         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 206         struct ptlrpc_service             *service = rqbd->rqbd_service;
 207         struct ptlrpc_request             *req;
 208         ENTRY;
 209
 210         LASSERT (ev->type == LNET_EVENT_PUT ||
 211                  ev->type == LNET_EVENT_UNLINK);
 212         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 213         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 214                  rqbd->rqbd_buffer + service->srv_buf_size);
 215
 216         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 217                "event type %d, status %d, service %s\n",
 218                ev->type, ev->status, service->srv_name);
 219
 220         if (ev->unlinked) {
 221                 /* If this is the last request message to fit in the
 222                  * request buffer we can use the request object embedded in
 223                  * rqbd.  Note that if we failed to allocate a request,
 224                  * we'd have to re-post the rqbd, which we can't do in this
 225                  * context. */
 226                 req = &rqbd->rqbd_req;
 227                 memset(req, 0, sizeof (*req));
 228         } else {
 229                 LASSERT (ev->type == LNET_EVENT_PUT);
 230                 if (ev->status != 0) {
 231                         /* We moaned above already... */
 232                         return;
 233                 }
 234                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 235                 if (req == NULL) {
 236                         CERROR("Can't allocate incoming request descriptor: "
 237                                "Dropping %s RPC from %s\n",
 238                                service->srv_name,
 239                                libcfs_id2str(ev->initiator));
 240                         return;
 241                 }
 242         }
 243
 244         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 245          * flags are reset and scalars are zero.  We only set the message
 246          * size to non-zero if this was a successful receive. */
 247         req->rq_xid = ev->match_bits;
 248         req->rq_reqbuf = ev->md.start + ev->offset;
 249         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 250                 req->rq_reqdata_len = ev->mlength;
 251         do_gettimeofday(&req->rq_arrival_time);
 252         req->rq_peer = ev->initiator;
 253         req->rq_self = ev->target.nid;
 254         req->rq_rqbd = rqbd;
 255         req->rq_phase = RQ_PHASE_NEW;
 256 #ifdef CRAY_XT3
 257         req->rq_uid = ev->uid;
 258 #endif
 259         spin_lock_init(&req->rq_lock);
 260         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 261         atomic_set(&req->rq_refcount, 1);
 262         if (ev->type == LNET_EVENT_PUT)
 263                 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
 264
 265         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 266
 267         spin_lock(&service->srv_lock);
 268
 269         req->rq_history_seq = service->srv_request_seq++;
 270         list_add_tail(&req->rq_history_list, &service->srv_request_history);
 271
 272         if (ev->unlinked) {
 273                 service->srv_nrqbd_receiving--;
 274                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 275                        service->srv_nrqbd_receiving);
 276
 277                 /* Normally, don't complain about 0 buffers posted; LNET won't
 278                  * drop incoming reqs since we set the portal lazy */
 279                 if (test_req_buffer_pressure &&
 280                     ev->type != LNET_EVENT_UNLINK &&
 281                     service->srv_nrqbd_receiving == 0)
 282                         CWARN("All %s request buffers busy\n",
 283                               service->srv_name);
 284
 285                 /* req takes over the network's ref on rqbd */
 286         } else {
 287                 /* req takes a ref on rqbd */
 288                 rqbd->rqbd_refcount++;
 289         }
 290
 291         list_add_tail(&req->rq_list, &service->srv_req_in_queue);
 292         service->srv_n_queued_reqs++;
 293
 294         /* NB everything can disappear under us once the request
 295          * has been queued and we unlock, so do the wake now... */
 296         cfs_waitq_signal(&service->srv_waitq);
 297
 298         spin_unlock(&service->srv_lock);
 299         EXIT;
 300 }
 301
 302 /*
 303  *  Server's outgoing reply callback
 304  */
 305 void reply_out_callback(lnet_event_t *ev)
 306 {
 307         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 308         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 309         struct ptlrpc_service     *svc = rs->rs_service;
 310         ENTRY;
 311
 312         LASSERT (ev->type == LNET_EVENT_SEND ||
 313                  ev->type == LNET_EVENT_ACK ||
 314                  ev->type == LNET_EVENT_UNLINK);
 315
 316         if (!rs->rs_difficult) {
 317                 /* 'Easy' replies have no further processing so I drop the
 318                  * net's ref on 'rs' */
 319                 LASSERT (ev->unlinked);
 320                 ptlrpc_rs_decref(rs);
 321                 atomic_dec (&svc->srv_outstanding_replies);
 322                 EXIT;
 323                 return;
 324         }
 325
 326         LASSERT (rs->rs_on_net);
 327
 328         if (ev->unlinked) {
 329                 /* Last network callback.  The net's ref on 'rs' stays put
 330                  * until ptlrpc_server_handle_reply() is done with it */
 331                 spin_lock(&svc->srv_lock);
 332                 rs->rs_on_net = 0;
 333                 ptlrpc_schedule_difficult_reply (rs);
 334                 spin_unlock(&svc->srv_lock);
 335         }
 336
 337         EXIT;
 338 }
 339
 340 /*
 341  * Server's bulk completion callback
 342  */
 343 void server_bulk_callback (lnet_event_t *ev)
 344 {
 345         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 346         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 347         ENTRY;
 348
 349         LASSERT (ev->type == LNET_EVENT_SEND ||
 350                  ev->type == LNET_EVENT_UNLINK ||
 351                  (desc->bd_type == BULK_PUT_SOURCE &&
 352                   ev->type == LNET_EVENT_ACK) ||
 353                  (desc->bd_type == BULK_GET_SINK &&
 354                   ev->type == LNET_EVENT_REPLY));
 355
 356         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 357                "event type %d, status %d, desc %p\n",
 358                ev->type, ev->status, desc);
 359
 360         spin_lock(&desc->bd_lock);
 361
 362         if ((ev->type == LNET_EVENT_ACK ||
 363              ev->type == LNET_EVENT_REPLY) &&
 364             ev->status == 0) {
 365                 /* We heard back from the peer, so even if we get this
 366                  * before the SENT event (oh yes we can), we know we
 367                  * read/wrote the peer buffer and how much... */
 368                 desc->bd_success = 1;
 369                 desc->bd_nob_transferred = ev->mlength;
 370                 desc->bd_sender = ev->sender;
 371         }
 372
 373         if (ev->unlinked) {
 374                 /* This is the last callback no matter what... */
 375                 desc->bd_network_rw = 0;
 376                 cfs_waitq_signal(&desc->bd_waitq);
 377         }
 378
 379         spin_unlock(&desc->bd_lock);
 380         EXIT;
 381 }
 382
 383 static void ptlrpc_master_callback(lnet_event_t *ev)
 384 {
 385         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 386         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 387
 388         /* Honestly, it's best to find out early. */
 389         LASSERT (cbid->cbid_arg != LP_POISON);
 390         LASSERT (callback == request_out_callback ||
 391                  callback == reply_in_callback ||
 392                  callback == client_bulk_callback ||
 393                  callback == request_in_callback ||
 394                  callback == reply_out_callback ||
 395                  callback == server_bulk_callback);
 396
 397         callback (ev);
 398 }
 399
 400 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 401                          lnet_process_id_t *peer, lnet_nid_t *self)
 402 {
 403         int               best_dist = 0;
 404         __u32             best_order = 0;
 405         int               count = 0;
 406         int               rc = -ENOENT;
 407         int               portals_compatibility;
 408         int               dist;
 409         __u32             order;
 410         lnet_nid_t        dst_nid;
 411         lnet_nid_t        src_nid;
 412
 413         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 414
 415         peer->pid = LUSTRE_SRV_LNET_PID;
 416
 417         /* Choose the matching UUID that's closest */
 418         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 419                 dist = LNetDist(dst_nid, &src_nid, &order);
 420                 if (dist < 0)
 421                         continue;
 422
 423                 if (dist == 0) {                /* local! use loopback LND */
 424                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 425                         rc = 0;
 426                         break;
 427                 }
 428
 429                 if (rc < 0 ||
 430                     dist < best_dist ||
 431                     (dist == best_dist && order < best_order)) {
 432                         best_dist = dist;
 433                         best_order = order;
 434
 435                         if (portals_compatibility > 1) {
 436                                 /* Strong portals compatibility: Zero the nid's
 437                                  * NET, so if I'm reading new config logs, or
 438                                  * getting configured by (new) lconf I can
 439                                  * still talk to old servers. */
 440                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 441                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 442                         }
 443                         peer->nid = dst_nid;
 444                         *self = src_nid;
 445                         rc = 0;
 446                 }
 447         }
 448
 449         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 450         if (rc != 0)
 451                 CERROR("No NID found for %s\n", uuid->uuid);
 452         return rc;
 453 }
 454
 455 void ptlrpc_ni_fini(void)
 456 {
 457         cfs_waitq_t         waitq;
 458         struct l_wait_info  lwi;
 459         int                 rc;
 460         int                 retries;
 461
 462         /* Wait for the event queue to become idle since there may still be
 463          * messages in flight with pending events (i.e. the fire-and-forget
 464          * messages == client requests and "non-difficult" server
 465          * replies */
 466
 467         for (retries = 0;; retries++) {
 468                 rc = LNetEQFree(ptlrpc_eq_h);
 469                 switch (rc) {
 470                 default:
 471                         LBUG();
 472
 473                 case 0:
 474                         LNetNIFini();
 475                         return;
 476
 477                 case -EBUSY:
 478                         if (retries != 0)
 479                                 CWARN("Event queue still busy\n");
 480
 481                         /* Wait for a bit */
 482                         cfs_waitq_init(&waitq);
 483                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 484                         l_wait_event(waitq, 0, &lwi);
 485                         break;
 486                 }
 487         }
 488         /* notreached */
 489 }
 490
 491 lnet_pid_t ptl_get_pid(void)
 492 {
 493         lnet_pid_t        pid;
 494
 495 #ifndef  __KERNEL__
 496         pid = getpid();
 497 #else
 498         pid = LUSTRE_SRV_LNET_PID;
 499 #endif
 500         return pid;
 501 }
 502
 503 int ptlrpc_ni_init(void)
 504 {
 505         int              rc;
 506         lnet_pid_t       pid;
 507
 508         pid = ptl_get_pid();
 509         CDEBUG(D_NET, "My pid is: %x\n", pid);
 510
 511         /* We're not passing any limits yet... */
 512         rc = LNetNIInit(pid);
 513         if (rc < 0) {
 514                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 515                 return (-ENOENT);
 516         }
 517
 518         /* CAVEAT EMPTOR: how we process portals events is _radically_
 519          * different depending on... */
 520 #ifdef __KERNEL__
 521         /* kernel portals calls our master callback when events are added to
 522          * the event queue.  In fact lustre never pulls events off this queue,
 523          * so it's only sized for some debug history. */
 524         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 525 #else
 526         /* liblustre calls the master callback when it removes events from the
 527          * event queue.  The event queue has to be big enough not to drop
 528          * anything */
 529         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 530 #endif
 531         if (rc == 0)
 532                 return 0;
 533
 534         CERROR ("Failed to allocate event queue: %d\n", rc);
 535         LNetNIFini();
 536
 537         return (-ENOMEM);
 538 }
 539
 540 #ifndef __KERNEL__
 541 CFS_LIST_HEAD(liblustre_wait_callbacks);
 542 CFS_LIST_HEAD(liblustre_idle_callbacks);
 543 void *liblustre_services_callback;
 544
 545 void *
 546 liblustre_register_waitidle_callback (struct list_head *callback_list,
 547                                       const char *name,
 548                                       int (*fn)(void *arg), void *arg)
 549 {
 550         struct liblustre_wait_callback *llwc;
 551
 552         OBD_ALLOC(llwc, sizeof(*llwc));
 553         LASSERT (llwc != NULL);
 554
 555         llwc->llwc_name = name;
 556         llwc->llwc_fn = fn;
 557         llwc->llwc_arg = arg;
 558         list_add_tail(&llwc->llwc_list, callback_list);
 559
 560         return (llwc);
 561 }
 562
 563 void
 564 liblustre_deregister_waitidle_callback (void *opaque)
 565 {
 566         struct liblustre_wait_callback *llwc = opaque;
 567
 568         list_del(&llwc->llwc_list);
 569         OBD_FREE(llwc, sizeof(*llwc));
 570 }
 571
 572 void *
 573 liblustre_register_wait_callback (const char *name,
 574                                   int (*fn)(void *arg), void *arg)
 575 {
 576         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 577                                                     name, fn, arg);
 578 }
 579
 580 void
 581 liblustre_deregister_wait_callback (void *opaque)
 582 {
 583         liblustre_deregister_waitidle_callback(opaque);
 584 }
 585
 586 void *
 587 liblustre_register_idle_callback (const char *name,
 588                                   int (*fn)(void *arg), void *arg)
 589 {
 590         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 591                                                     name, fn, arg);
 592 }
 593
 594 void
 595 liblustre_deregister_idle_callback (void *opaque)
 596 {
 597         liblustre_deregister_waitidle_callback(opaque);
 598 }
 599
 600 int
 601 liblustre_check_events (int timeout)
 602 {
 603         lnet_event_t ev;
 604         int         rc;
 605         int         i;
 606         ENTRY;
 607
 608         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 609         if (rc == 0)
 610                 RETURN(0);
 611
 612         LASSERT (rc == -EOVERFLOW || rc == 1);
 613
 614         /* liblustre: no asynch callback so we can't affort to miss any
 615          * events... */
 616         if (rc == -EOVERFLOW) {
 617                 CERROR ("Dropped an event!!!\n");
 618                 abort();
 619         }
 620
 621         ptlrpc_master_callback (&ev);
 622         RETURN(1);
 623 }
 624
 625 int liblustre_waiting = 0;
 626
 627 int
 628 liblustre_wait_event (int timeout)
 629 {
 630         struct list_head               *tmp;
 631         struct liblustre_wait_callback *llwc;
 632         int                             found_something = 0;
 633
 634         /* single threaded recursion check... */
 635         liblustre_waiting = 1;
 636
 637         for (;;) {
 638                 /* Deal with all pending events */
 639                 while (liblustre_check_events(0))
 640                         found_something = 1;
 641
 642                 /* Give all registered callbacks a bite at the cherry */
 643                 list_for_each(tmp, &liblustre_wait_callbacks) {
 644                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 645                                           llwc_list);
 646
 647                         if (llwc->llwc_fn(llwc->llwc_arg))
 648                                 found_something = 1;
 649                 }
 650
 651                 if (found_something || timeout == 0)
 652                         break;
 653
 654                 /* Nothing so far, but I'm allowed to block... */
 655                 found_something = liblustre_check_events(timeout);
 656                 if (!found_something)           /* still nothing */
 657                         break;                  /* I timed out */
 658         }
 659
 660         liblustre_waiting = 0;
 661
 662         return found_something;
 663 }
 664
 665 void
 666 liblustre_wait_idle(void)
 667 {
 668         static int recursed = 0;
 669
 670         struct list_head               *tmp;
 671         struct liblustre_wait_callback *llwc;
 672         int                             idle = 0;
 673
 674         LASSERT(!recursed);
 675         recursed = 1;
 676
 677         do {
 678                 liblustre_wait_event(0);
 679
 680                 idle = 1;
 681
 682                 list_for_each(tmp, &liblustre_idle_callbacks) {
 683                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 684                                           llwc_list);
 685
 686                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 687                                 idle = 0;
 688                                 break;
 689                         }
 690                 }
 691
 692         } while (!idle);
 693
 694         recursed = 0;
 695 }
 696
 697 #endif /* __KERNEL__ */
 698
 699 int ptlrpc_init_portals(void)
 700 {
 701         int   rc = ptlrpc_ni_init();
 702
 703         if (rc != 0) {
 704                 CERROR("network initialisation failed\n");
 705                 return -EIO;
 706         }
 707 #ifndef __KERNEL__
 708         liblustre_services_callback =
 709                 liblustre_register_wait_callback("liblustre_check_services",
 710                                                  &liblustre_check_services,
 711                                                  NULL);
 712 #endif
 713         rc = ptlrpcd_addref();
 714         if (rc == 0)
 715                 return 0;
 716
 717         CERROR("rpcd initialisation failed\n");
 718 #ifndef __KERNEL__
 719         liblustre_deregister_wait_callback(liblustre_services_callback);
 720 #endif
 721         ptlrpc_ni_fini();
 722         return rc;
 723 }
 724
 725 void ptlrpc_exit_portals(void)
 726 {
 727 #ifndef __KERNEL__
 728         liblustre_deregister_wait_callback(liblustre_services_callback);
 729 #endif
 730         ptlrpcd_decref();
 731         ptlrpc_ni_fini();
 732 }