lustre/ptlrpc/events.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38
  39 #ifndef __KERNEL__
  40 # include <liblustre.h>
  41 #else
  42 # include <libcfs/libcfs.h>
  43 # ifdef __mips64__
  44 #  include <linux/kernel.h>
  45 # endif
  46 #endif
  47
  48 #include <obd_class.h>
  49 #include <lustre_net.h>
  50 #include <lustre_sec.h>
  51 #include "ptlrpc_internal.h"
  52
  53 lnet_handle_eq_t   ptlrpc_eq_h;
  54
  55 /*
  56  *  Client's outgoing request callback
  57  */
  58 void request_out_callback(lnet_event_t *ev)
  59 {
  60         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  61         struct ptlrpc_request *req = cbid->cbid_arg;
  62         ENTRY;
  63
  64         LASSERT (ev->type == LNET_EVENT_SEND ||
  65                  ev->type == LNET_EVENT_UNLINK);
  66         LASSERT (ev->unlinked);
  67
  68         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
  69                   "type %d, status %d", ev->type, ev->status);
  70
  71         sptlrpc_request_out_callback(req);
  72
  73         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  74
  75                 /* Failed send: make it seem like the reply timed out, just
  76                  * like failing sends in client.c does currently...  */
  77
  78                 spin_lock(&req->rq_lock);
  79                 req->rq_net_err = 1;
  80                 spin_unlock(&req->rq_lock);
  81
  82                 ptlrpc_client_wake_req(req);
  83         }
  84
  85         ptlrpc_req_finished(req);
  86
  87         EXIT;
  88 }
  89
  90 /*
  91  * Client's incoming reply callback
  92  */
  93 void reply_in_callback(lnet_event_t *ev)
  94 {
  95         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  96         struct ptlrpc_request *req = cbid->cbid_arg;
  97         ENTRY;
  98
  99         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
 100                   "type %d, status %d", ev->type, ev->status);
 101
 102         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 103         LASSERT (ev->md.start == req->rq_repbuf);
 104         LASSERT (ev->mlength <= req->rq_repbuf_len);
 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) != 0);
 108
 109         spin_lock(&req->rq_lock);
 110
 111         req->rq_receiving_reply = 0;
 112         req->rq_early = 0;
 113         if (ev->unlinked)
 114                 req->rq_must_unlink = 0;
 115
 116         if (ev->status)
 117                 goto out_wake;
 118         if (ev->type == LNET_EVENT_UNLINK) {
 119                 LASSERT(ev->unlinked);
 120                 DEBUG_REQ(D_RPCTRACE, req, "unlink");
 121                 goto out_wake;
 122         }
 123
 124         if ((ev->offset == 0) &&
 125             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 126                 /* Early reply */
 127                 DEBUG_REQ(D_ADAPTTO, req,
 128                           "Early reply received: mlen=%u offset=%d replen=%d "
 129                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 130                           req->rq_replen, req->rq_replied, ev->unlinked);
 131
 132                 req->rq_early_count++; /* number received, client side */
 133
 134                 if (req->rq_replied)   /* already got the real reply */
 135                         goto out_wake;
 136
 137                 req->rq_early = 1;
 138                 req->rq_reply_off = ev->offset;
 139                 req->rq_nob_received = ev->mlength;
 140                 /* And we're still receiving */
 141                 req->rq_receiving_reply = 1;
 142         } else {
 143                 /* Real reply */
 144                 req->rq_rep_swab_mask = 0;
 145                 req->rq_replied = 1;
 146                 req->rq_reply_off = ev->offset;
 147                 req->rq_nob_received = ev->mlength;
 148                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 149                    so we must unlink in ptlrpc_unregister_reply */
 150                 DEBUG_REQ(D_INFO, req,
 151                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 152                           lustre_msg_get_flags(req->rq_reqmsg),
 153                           ev->mlength, ev->offset, req->rq_replen);
 154         }
 155
 156         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 157
 158 out_wake:
 159         /* NB don't unlock till after wakeup; req can disappear under us
 160          * since we don't have our own ref */
 161         ptlrpc_client_wake_req(req);
 162         spin_unlock(&req->rq_lock);
 163         EXIT;
 164 }
 165
 166 /*
 167  * Client's bulk has been written/read
 168  */
 169 void client_bulk_callback (lnet_event_t *ev)
 170 {
 171         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 172         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 173         ENTRY;
 174
 175         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 176                   ev->type == LNET_EVENT_PUT) ||
 177                  (desc->bd_type == BULK_GET_SOURCE &&
 178                   ev->type == LNET_EVENT_GET) ||
 179                  ev->type == LNET_EVENT_UNLINK);
 180         LASSERT (ev->unlinked);
 181
 182         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 183                "event type %d, status %d, desc %p\n",
 184                ev->type, ev->status, desc);
 185
 186         spin_lock(&desc->bd_lock);
 187
 188         LASSERT(desc->bd_network_rw);
 189         desc->bd_network_rw = 0;
 190
 191         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 192                 desc->bd_success = 1;
 193                 desc->bd_nob_transferred = ev->mlength;
 194                 desc->bd_sender = ev->sender;
 195         }
 196
 197         /* release the encrypted pages for write */
 198         if (desc->bd_req->rq_bulk_write)
 199                 sptlrpc_enc_pool_put_pages(desc);
 200
 201         /* NB don't unlock till after wakeup; desc can disappear under us
 202          * otherwise */
 203         ptlrpc_client_wake_req(desc->bd_req);
 204
 205         spin_unlock(&desc->bd_lock);
 206         EXIT;
 207 }
 208
 209 /*
 210  * Server's incoming request callback
 211  */
 212 void request_in_callback(lnet_event_t *ev)
 213 {
 214         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 215         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 216         struct ptlrpc_service             *service = rqbd->rqbd_service;
 217         struct ptlrpc_request             *req;
 218         ENTRY;
 219
 220         LASSERT (ev->type == LNET_EVENT_PUT ||
 221                  ev->type == LNET_EVENT_UNLINK);
 222         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 223         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 224                  rqbd->rqbd_buffer + service->srv_buf_size);
 225
 226         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 227                "event type %d, status %d, service %s\n",
 228                ev->type, ev->status, service->srv_name);
 229
 230         if (ev->unlinked) {
 231                 /* If this is the last request message to fit in the
 232                  * request buffer we can use the request object embedded in
 233                  * rqbd.  Note that if we failed to allocate a request,
 234                  * we'd have to re-post the rqbd, which we can't do in this
 235                  * context. */
 236                 req = &rqbd->rqbd_req;
 237                 memset(req, 0, sizeof (*req));
 238         } else {
 239                 LASSERT (ev->type == LNET_EVENT_PUT);
 240                 if (ev->status != 0) {
 241                         /* We moaned above already... */
 242                         return;
 243                 }
 244                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 245                 if (req == NULL) {
 246                         CERROR("Can't allocate incoming request descriptor: "
 247                                "Dropping %s RPC from %s\n",
 248                                service->srv_name,
 249                                libcfs_id2str(ev->initiator));
 250                         return;
 251                 }
 252         }
 253
 254         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 255          * flags are reset and scalars are zero.  We only set the message
 256          * size to non-zero if this was a successful receive. */
 257         req->rq_xid = ev->match_bits;
 258         req->rq_reqbuf = ev->md.start + ev->offset;
 259         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 260                 req->rq_reqdata_len = ev->mlength;
 261         do_gettimeofday(&req->rq_arrival_time);
 262         req->rq_peer = ev->initiator;
 263         req->rq_self = ev->target.nid;
 264         req->rq_rqbd = rqbd;
 265         req->rq_phase = RQ_PHASE_NEW;
 266 #ifdef CRAY_XT3
 267         req->rq_uid = ev->uid;
 268 #endif
 269         spin_lock_init(&req->rq_lock);
 270         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 271         atomic_set(&req->rq_refcount, 1);
 272         if (ev->type == LNET_EVENT_PUT)
 273                 DEBUG_REQ(D_RPCTRACE, req, "incoming req");
 274
 275         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 276
 277         spin_lock(&service->srv_lock);
 278
 279         req->rq_history_seq = service->srv_request_seq++;
 280         list_add_tail(&req->rq_history_list, &service->srv_request_history);
 281
 282         if (ev->unlinked) {
 283                 service->srv_nrqbd_receiving--;
 284                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 285                        service->srv_nrqbd_receiving);
 286
 287                 /* Normally, don't complain about 0 buffers posted; LNET won't
 288                  * drop incoming reqs since we set the portal lazy */
 289                 if (test_req_buffer_pressure &&
 290                     ev->type != LNET_EVENT_UNLINK &&
 291                     service->srv_nrqbd_receiving == 0)
 292                         CWARN("All %s request buffers busy\n",
 293                               service->srv_name);
 294
 295                 /* req takes over the network's ref on rqbd */
 296         } else {
 297                 /* req takes a ref on rqbd */
 298                 rqbd->rqbd_refcount++;
 299         }
 300
 301         list_add_tail(&req->rq_list, &service->srv_req_in_queue);
 302         service->srv_n_queued_reqs++;
 303
 304         /* NB everything can disappear under us once the request
 305          * has been queued and we unlock, so do the wake now... */
 306         cfs_waitq_signal(&service->srv_waitq);
 307
 308         spin_unlock(&service->srv_lock);
 309         EXIT;
 310 }
 311
 312 /*
 313  *  Server's outgoing reply callback
 314  */
 315 void reply_out_callback(lnet_event_t *ev)
 316 {
 317         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 318         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 319         struct ptlrpc_service     *svc = rs->rs_service;
 320         ENTRY;
 321
 322         LASSERT (ev->type == LNET_EVENT_SEND ||
 323                  ev->type == LNET_EVENT_ACK ||
 324                  ev->type == LNET_EVENT_UNLINK);
 325
 326         if (!rs->rs_difficult) {
 327                 /* 'Easy' replies have no further processing so I drop the
 328                  * net's ref on 'rs' */
 329                 LASSERT (ev->unlinked);
 330                 ptlrpc_rs_decref(rs);
 331                 atomic_dec (&svc->srv_outstanding_replies);
 332                 EXIT;
 333                 return;
 334         }
 335
 336         LASSERT (rs->rs_on_net);
 337
 338         if (ev->unlinked) {
 339                 /* Last network callback.  The net's ref on 'rs' stays put
 340                  * until ptlrpc_server_handle_reply() is done with it */
 341                 spin_lock(&svc->srv_lock);
 342                 spin_lock(&rs->rs_lock);
 343                 rs->rs_on_net = 0;
 344                 if (!rs->rs_no_ack ||
 345                     rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
 346                         ptlrpc_schedule_difficult_reply (rs);
 347                 spin_unlock(&rs->rs_lock);
 348                 spin_unlock(&svc->srv_lock);
 349         }
 350
 351         EXIT;
 352 }
 353
 354 /*
 355  * Server's bulk completion callback
 356  */
 357 void server_bulk_callback (lnet_event_t *ev)
 358 {
 359         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 360         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 361         ENTRY;
 362
 363         LASSERT (ev->type == LNET_EVENT_SEND ||
 364                  ev->type == LNET_EVENT_UNLINK ||
 365                  (desc->bd_type == BULK_PUT_SOURCE &&
 366                   ev->type == LNET_EVENT_ACK) ||
 367                  (desc->bd_type == BULK_GET_SINK &&
 368                   ev->type == LNET_EVENT_REPLY));
 369
 370         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 371                "event type %d, status %d, desc %p\n",
 372                ev->type, ev->status, desc);
 373
 374         spin_lock(&desc->bd_lock);
 375
 376         if ((ev->type == LNET_EVENT_ACK ||
 377              ev->type == LNET_EVENT_REPLY) &&
 378             ev->status == 0) {
 379                 /* We heard back from the peer, so even if we get this
 380                  * before the SENT event (oh yes we can), we know we
 381                  * read/wrote the peer buffer and how much... */
 382                 desc->bd_success = 1;
 383                 desc->bd_nob_transferred = ev->mlength;
 384                 desc->bd_sender = ev->sender;
 385         }
 386
 387         if (ev->unlinked) {
 388                 /* This is the last callback no matter what... */
 389                 desc->bd_network_rw = 0;
 390                 cfs_waitq_signal(&desc->bd_waitq);
 391         }
 392
 393         spin_unlock(&desc->bd_lock);
 394         EXIT;
 395 }
 396
 397 static void ptlrpc_master_callback(lnet_event_t *ev)
 398 {
 399         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 400         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 401
 402         /* Honestly, it's best to find out early. */
 403         LASSERT (cbid->cbid_arg != LP_POISON);
 404         LASSERT (callback == request_out_callback ||
 405                  callback == reply_in_callback ||
 406                  callback == client_bulk_callback ||
 407                  callback == request_in_callback ||
 408                  callback == reply_out_callback ||
 409                  callback == server_bulk_callback);
 410
 411         callback (ev);
 412 }
 413
 414 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 415                          lnet_process_id_t *peer, lnet_nid_t *self)
 416 {
 417         int               best_dist = 0;
 418         __u32             best_order = 0;
 419         int               count = 0;
 420         int               rc = -ENOENT;
 421         int               portals_compatibility;
 422         int               dist;
 423         __u32             order;
 424         lnet_nid_t        dst_nid;
 425         lnet_nid_t        src_nid;
 426
 427         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 428
 429         peer->pid = LUSTRE_SRV_LNET_PID;
 430
 431         /* Choose the matching UUID that's closest */
 432         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 433                 dist = LNetDist(dst_nid, &src_nid, &order);
 434                 if (dist < 0)
 435                         continue;
 436
 437                 if (dist == 0) {                /* local! use loopback LND */
 438                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 439                         rc = 0;
 440                         break;
 441                 }
 442
 443                 if (rc < 0 ||
 444                     dist < best_dist ||
 445                     (dist == best_dist && order < best_order)) {
 446                         best_dist = dist;
 447                         best_order = order;
 448
 449                         if (portals_compatibility > 1) {
 450                                 /* Strong portals compatibility: Zero the nid's
 451                                  * NET, so if I'm reading new config logs, or
 452                                  * getting configured by (new) lconf I can
 453                                  * still talk to old servers. */
 454                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 455                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 456                         }
 457                         peer->nid = dst_nid;
 458                         *self = src_nid;
 459                         rc = 0;
 460                 }
 461         }
 462
 463         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 464         if (rc != 0)
 465                 CERROR("No NID found for %s\n", uuid->uuid);
 466         return rc;
 467 }
 468
 469 void ptlrpc_ni_fini(void)
 470 {
 471         cfs_waitq_t         waitq;
 472         struct l_wait_info  lwi;
 473         int                 rc;
 474         int                 retries;
 475
 476         /* Wait for the event queue to become idle since there may still be
 477          * messages in flight with pending events (i.e. the fire-and-forget
 478          * messages == client requests and "non-difficult" server
 479          * replies */
 480
 481         for (retries = 0;; retries++) {
 482                 rc = LNetEQFree(ptlrpc_eq_h);
 483                 switch (rc) {
 484                 default:
 485                         LBUG();
 486
 487                 case 0:
 488                         LNetNIFini();
 489                         return;
 490
 491                 case -EBUSY:
 492                         if (retries != 0)
 493                                 CWARN("Event queue still busy\n");
 494
 495                         /* Wait for a bit */
 496                         cfs_waitq_init(&waitq);
 497                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 498                         l_wait_event(waitq, 0, &lwi);
 499                         break;
 500                 }
 501         }
 502         /* notreached */
 503 }
 504
 505 lnet_pid_t ptl_get_pid(void)
 506 {
 507         lnet_pid_t        pid;
 508
 509 #ifndef  __KERNEL__
 510         pid = getpid();
 511 #else
 512         pid = LUSTRE_SRV_LNET_PID;
 513 #endif
 514         return pid;
 515 }
 516
 517 int ptlrpc_ni_init(void)
 518 {
 519         int              rc;
 520         lnet_pid_t       pid;
 521
 522         pid = ptl_get_pid();
 523         CDEBUG(D_NET, "My pid is: %x\n", pid);
 524
 525         /* We're not passing any limits yet... */
 526         rc = LNetNIInit(pid);
 527         if (rc < 0) {
 528                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 529                 return (-ENOENT);
 530         }
 531
 532         /* CAVEAT EMPTOR: how we process portals events is _radically_
 533          * different depending on... */
 534 #ifdef __KERNEL__
 535         /* kernel portals calls our master callback when events are added to
 536          * the event queue.  In fact lustre never pulls events off this queue,
 537          * so it's only sized for some debug history. */
 538         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 539 #else
 540         /* liblustre calls the master callback when it removes events from the
 541          * event queue.  The event queue has to be big enough not to drop
 542          * anything */
 543         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 544 #endif
 545         if (rc == 0)
 546                 return 0;
 547
 548         CERROR ("Failed to allocate event queue: %d\n", rc);
 549         LNetNIFini();
 550
 551         return (-ENOMEM);
 552 }
 553
 554 #ifndef __KERNEL__
 555 CFS_LIST_HEAD(liblustre_wait_callbacks);
 556 CFS_LIST_HEAD(liblustre_idle_callbacks);
 557 void *liblustre_services_callback;
 558
 559 void *
 560 liblustre_register_waitidle_callback (struct list_head *callback_list,
 561                                       const char *name,
 562                                       int (*fn)(void *arg), void *arg)
 563 {
 564         struct liblustre_wait_callback *llwc;
 565
 566         OBD_ALLOC(llwc, sizeof(*llwc));
 567         LASSERT (llwc != NULL);
 568
 569         llwc->llwc_name = name;
 570         llwc->llwc_fn = fn;
 571         llwc->llwc_arg = arg;
 572         list_add_tail(&llwc->llwc_list, callback_list);
 573
 574         return (llwc);
 575 }
 576
 577 void
 578 liblustre_deregister_waitidle_callback (void *opaque)
 579 {
 580         struct liblustre_wait_callback *llwc = opaque;
 581
 582         list_del(&llwc->llwc_list);
 583         OBD_FREE(llwc, sizeof(*llwc));
 584 }
 585
 586 void *
 587 liblustre_register_wait_callback (const char *name,
 588                                   int (*fn)(void *arg), void *arg)
 589 {
 590         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 591                                                     name, fn, arg);
 592 }
 593
 594 void
 595 liblustre_deregister_wait_callback (void *opaque)
 596 {
 597         liblustre_deregister_waitidle_callback(opaque);
 598 }
 599
 600 void *
 601 liblustre_register_idle_callback (const char *name,
 602                                   int (*fn)(void *arg), void *arg)
 603 {
 604         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 605                                                     name, fn, arg);
 606 }
 607
 608 void
 609 liblustre_deregister_idle_callback (void *opaque)
 610 {
 611         liblustre_deregister_waitidle_callback(opaque);
 612 }
 613
 614 int
 615 liblustre_check_events (int timeout)
 616 {
 617         lnet_event_t ev;
 618         int         rc;
 619         int         i;
 620         ENTRY;
 621
 622         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 623         if (rc == 0)
 624                 RETURN(0);
 625
 626         LASSERT (rc == -EOVERFLOW || rc == 1);
 627
 628         /* liblustre: no asynch callback so we can't affort to miss any
 629          * events... */
 630         if (rc == -EOVERFLOW) {
 631                 CERROR ("Dropped an event!!!\n");
 632                 abort();
 633         }
 634
 635         ptlrpc_master_callback (&ev);
 636         RETURN(1);
 637 }
 638
 639 int liblustre_waiting = 0;
 640
 641 int
 642 liblustre_wait_event (int timeout)
 643 {
 644         struct list_head               *tmp;
 645         struct liblustre_wait_callback *llwc;
 646         int                             found_something = 0;
 647
 648         /* single threaded recursion check... */
 649         liblustre_waiting = 1;
 650
 651         for (;;) {
 652                 /* Deal with all pending events */
 653                 while (liblustre_check_events(0))
 654                         found_something = 1;
 655
 656                 /* Give all registered callbacks a bite at the cherry */
 657                 list_for_each(tmp, &liblustre_wait_callbacks) {
 658                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 659                                           llwc_list);
 660
 661                         if (llwc->llwc_fn(llwc->llwc_arg))
 662                                 found_something = 1;
 663                 }
 664
 665                 if (found_something || timeout == 0)
 666                         break;
 667
 668                 /* Nothing so far, but I'm allowed to block... */
 669                 found_something = liblustre_check_events(timeout);
 670                 if (!found_something)           /* still nothing */
 671                         break;                  /* I timed out */
 672         }
 673
 674         liblustre_waiting = 0;
 675
 676         return found_something;
 677 }
 678
 679 void
 680 liblustre_wait_idle(void)
 681 {
 682         static int recursed = 0;
 683
 684         struct list_head               *tmp;
 685         struct liblustre_wait_callback *llwc;
 686         int                             idle = 0;
 687
 688         LASSERT(!recursed);
 689         recursed = 1;
 690
 691         do {
 692                 liblustre_wait_event(0);
 693
 694                 idle = 1;
 695
 696                 list_for_each(tmp, &liblustre_idle_callbacks) {
 697                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 698                                           llwc_list);
 699
 700                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 701                                 idle = 0;
 702                                 break;
 703                         }
 704                 }
 705
 706         } while (!idle);
 707
 708         recursed = 0;
 709 }
 710
 711 #endif /* __KERNEL__ */
 712
 713 int ptlrpc_init_portals(void)
 714 {
 715         int   rc = ptlrpc_ni_init();
 716
 717         if (rc != 0) {
 718                 CERROR("network initialisation failed\n");
 719                 return -EIO;
 720         }
 721 #ifndef __KERNEL__
 722         liblustre_services_callback =
 723                 liblustre_register_wait_callback("liblustre_check_services",
 724                                                  &liblustre_check_services,
 725                                                  NULL);
 726         init_completion_module(liblustre_wait_event);
 727 #endif
 728         rc = ptlrpcd_addref();
 729         if (rc == 0)
 730                 return 0;
 731
 732         CERROR("rpcd initialisation failed\n");
 733 #ifndef __KERNEL__
 734         liblustre_deregister_wait_callback(liblustre_services_callback);
 735 #endif
 736         ptlrpc_ni_fini();
 737         return rc;
 738 }
 739
 740 void ptlrpc_exit_portals(void)
 741 {
 742 #ifndef __KERNEL__
 743         liblustre_deregister_wait_callback(liblustre_services_callback);
 744 #endif
 745         ptlrpcd_decref();
 746         ptlrpc_ni_fini();
 747 }