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 (c) 2002, 2010, Oracle and/or its affiliates. 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(D_NET, req, "type %d, status %d", ev->type, ev->status);
  69
  70         sptlrpc_request_out_callback(req);
  71         req->rq_real_sent = cfs_time_current_sec();
  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                 cfs_spin_lock(&req->rq_lock);
  79                 req->rq_net_err = 1;
  80                 cfs_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(D_NET, req, "type %d, status %d", ev->type, ev->status);
 100
 101         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 102         LASSERT (ev->md.start == req->rq_repbuf);
 103         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
 104         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 105            for adaptive timeouts' early reply. */
 106         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 107
 108         cfs_spin_lock(&req->rq_lock);
 109
 110         req->rq_receiving_reply = 0;
 111         req->rq_early = 0;
 112         if (ev->unlinked)
 113                 req->rq_must_unlink = 0;
 114
 115         if (ev->status)
 116                 goto out_wake;
 117
 118         if (ev->type == LNET_EVENT_UNLINK) {
 119                 LASSERT(ev->unlinked);
 120                 DEBUG_REQ(D_NET, req, "unlink");
 121                 goto out_wake;
 122         }
 123
 124         if (ev->mlength < ev->rlength ) {
 125                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 126                        req->rq_replen, ev->rlength, ev->offset);
 127                 req->rq_reply_truncate = 1;
 128                 req->rq_replied = 1;
 129                 req->rq_status = -EOVERFLOW;
 130                 req->rq_nob_received = ev->rlength + ev->offset;
 131                 goto out_wake;
 132         }
 133
 134         if ((ev->offset == 0) &&
 135             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 136                 /* Early reply */
 137                 DEBUG_REQ(D_ADAPTTO, req,
 138                           "Early reply received: mlen=%u offset=%d replen=%d "
 139                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 140                           req->rq_replen, req->rq_replied, ev->unlinked);
 141
 142                 req->rq_early_count++; /* number received, client side */
 143
 144                 if (req->rq_replied)   /* already got the real reply */
 145                         goto out_wake;
 146
 147                 req->rq_early = 1;
 148                 req->rq_reply_off = ev->offset;
 149                 req->rq_nob_received = ev->mlength;
 150                 /* And we're still receiving */
 151                 req->rq_receiving_reply = 1;
 152         } else {
 153                 /* Real reply */
 154                 req->rq_rep_swab_mask = 0;
 155                 req->rq_replied = 1;
 156                 req->rq_reply_off = ev->offset;
 157                 req->rq_nob_received = ev->mlength;
 158                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 159                    so we must unlink in ptlrpc_unregister_reply */
 160                 DEBUG_REQ(D_INFO, req,
 161                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 162                           lustre_msg_get_flags(req->rq_reqmsg),
 163                           ev->mlength, ev->offset, req->rq_replen);
 164         }
 165
 166         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 167
 168 out_wake:
 169         /* NB don't unlock till after wakeup; req can disappear under us
 170          * since we don't have our own ref */
 171         ptlrpc_client_wake_req(req);
 172         cfs_spin_unlock(&req->rq_lock);
 173         EXIT;
 174 }
 175
 176 /*
 177  * Client's bulk has been written/read
 178  */
 179 void client_bulk_callback (lnet_event_t *ev)
 180 {
 181         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 182         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 183         ENTRY;
 184
 185         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 186                   ev->type == LNET_EVENT_PUT) ||
 187                  (desc->bd_type == BULK_GET_SOURCE &&
 188                   ev->type == LNET_EVENT_GET) ||
 189                  ev->type == LNET_EVENT_UNLINK);
 190         LASSERT (ev->unlinked);
 191
 192         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 193                 ev->status = -EIO;
 194
 195         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 196                 ev->status = -EIO;
 197
 198         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 199                "event type %d, status %d, desc %p\n",
 200                ev->type, ev->status, desc);
 201
 202         cfs_spin_lock(&desc->bd_lock);
 203
 204         LASSERT(desc->bd_network_rw);
 205         desc->bd_network_rw = 0;
 206
 207         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 208                 desc->bd_success = 1;
 209                 desc->bd_nob_transferred = ev->mlength;
 210                 desc->bd_sender = ev->sender;
 211         }
 212
 213         /* release the encrypted pages for write */
 214         if (desc->bd_req->rq_bulk_write)
 215                 sptlrpc_enc_pool_put_pages(desc);
 216
 217         /* NB don't unlock till after wakeup; desc can disappear under us
 218          * otherwise */
 219         ptlrpc_client_wake_req(desc->bd_req);
 220
 221         cfs_spin_unlock(&desc->bd_lock);
 222         EXIT;
 223 }
 224
 225 /*
 226  * Server's incoming request callback
 227  */
 228 void request_in_callback(lnet_event_t *ev)
 229 {
 230         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 231         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 232         struct ptlrpc_service             *service = rqbd->rqbd_service;
 233         struct ptlrpc_request             *req;
 234         ENTRY;
 235
 236         LASSERT (ev->type == LNET_EVENT_PUT ||
 237                  ev->type == LNET_EVENT_UNLINK);
 238         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 239         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 240                  rqbd->rqbd_buffer + service->srv_buf_size);
 241
 242         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 243                "event type %d, status %d, service %s\n",
 244                ev->type, ev->status, service->srv_name);
 245
 246         if (ev->unlinked) {
 247                 /* If this is the last request message to fit in the
 248                  * request buffer we can use the request object embedded in
 249                  * rqbd.  Note that if we failed to allocate a request,
 250                  * we'd have to re-post the rqbd, which we can't do in this
 251                  * context. */
 252                 req = &rqbd->rqbd_req;
 253                 memset(req, 0, sizeof (*req));
 254         } else {
 255                 LASSERT (ev->type == LNET_EVENT_PUT);
 256                 if (ev->status != 0) {
 257                         /* We moaned above already... */
 258                         return;
 259                 }
 260                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 261                 if (req == NULL) {
 262                         CERROR("Can't allocate incoming request descriptor: "
 263                                "Dropping %s RPC from %s\n",
 264                                service->srv_name,
 265                                libcfs_id2str(ev->initiator));
 266                         return;
 267                 }
 268         }
 269
 270         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 271          * flags are reset and scalars are zero.  We only set the message
 272          * size to non-zero if this was a successful receive. */
 273         req->rq_xid = ev->match_bits;
 274         req->rq_reqbuf = ev->md.start + ev->offset;
 275         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 276                 req->rq_reqdata_len = ev->mlength;
 277         cfs_gettimeofday(&req->rq_arrival_time);
 278         req->rq_peer = ev->initiator;
 279         req->rq_self = ev->target.nid;
 280         req->rq_rqbd = rqbd;
 281         req->rq_phase = RQ_PHASE_NEW;
 282 #ifdef CRAY_XT3
 283         req->rq_uid = ev->uid;
 284 #endif
 285         cfs_spin_lock_init(&req->rq_lock);
 286         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 287         cfs_atomic_set(&req->rq_refcount, 1);
 288         if (ev->type == LNET_EVENT_PUT)
 289                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
 290                        req, req->rq_xid, ev->mlength);
 291
 292         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 293
 294         cfs_spin_lock(&service->srv_lock);
 295
 296         req->rq_history_seq = service->srv_request_seq++;
 297         cfs_list_add_tail(&req->rq_history_list, &service->srv_request_history);
 298
 299         if (ev->unlinked) {
 300                 service->srv_nrqbd_receiving--;
 301                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 302                        service->srv_nrqbd_receiving);
 303
 304                 /* Normally, don't complain about 0 buffers posted; LNET won't
 305                  * drop incoming reqs since we set the portal lazy */
 306                 if (test_req_buffer_pressure &&
 307                     ev->type != LNET_EVENT_UNLINK &&
 308                     service->srv_nrqbd_receiving == 0)
 309                         CWARN("All %s request buffers busy\n",
 310                               service->srv_name);
 311
 312                 /* req takes over the network's ref on rqbd */
 313         } else {
 314                 /* req takes a ref on rqbd */
 315                 rqbd->rqbd_refcount++;
 316         }
 317
 318         cfs_list_add_tail(&req->rq_list, &service->srv_req_in_queue);
 319         service->srv_n_queued_reqs++;
 320
 321         /* NB everything can disappear under us once the request
 322          * has been queued and we unlock, so do the wake now... */
 323         cfs_waitq_signal(&service->srv_waitq);
 324
 325         cfs_spin_unlock(&service->srv_lock);
 326         EXIT;
 327 }
 328
 329 /*
 330  *  Server's outgoing reply callback
 331  */
 332 void reply_out_callback(lnet_event_t *ev)
 333 {
 334         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 335         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 336         struct ptlrpc_service     *svc = rs->rs_service;
 337         ENTRY;
 338
 339         LASSERT (ev->type == LNET_EVENT_SEND ||
 340                  ev->type == LNET_EVENT_ACK ||
 341                  ev->type == LNET_EVENT_UNLINK);
 342
 343         if (!rs->rs_difficult) {
 344                 /* 'Easy' replies have no further processing so I drop the
 345                  * net's ref on 'rs' */
 346                 LASSERT (ev->unlinked);
 347                 ptlrpc_rs_decref(rs);
 348                 EXIT;
 349                 return;
 350         }
 351
 352         LASSERT (rs->rs_on_net);
 353
 354         if (ev->unlinked) {
 355                 /* Last network callback. The net's ref on 'rs' stays put
 356                  * until ptlrpc_handle_rs() is done with it */
 357                 cfs_spin_lock(&svc->srv_rs_lock);
 358                 cfs_spin_lock(&rs->rs_lock);
 359                 rs->rs_on_net = 0;
 360                 if (!rs->rs_no_ack ||
 361                     rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
 362                         ptlrpc_schedule_difficult_reply (rs);
 363                 cfs_spin_unlock(&rs->rs_lock);
 364                 cfs_spin_unlock(&svc->srv_rs_lock);
 365         }
 366
 367         EXIT;
 368 }
 369
 370 #ifdef HAVE_SERVER_SUPPORT
 371 /*
 372  * Server's bulk completion callback
 373  */
 374 void server_bulk_callback (lnet_event_t *ev)
 375 {
 376         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 377         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 378         ENTRY;
 379
 380         LASSERT (ev->type == LNET_EVENT_SEND ||
 381                  ev->type == LNET_EVENT_UNLINK ||
 382                  (desc->bd_type == BULK_PUT_SOURCE &&
 383                   ev->type == LNET_EVENT_ACK) ||
 384                  (desc->bd_type == BULK_GET_SINK &&
 385                   ev->type == LNET_EVENT_REPLY));
 386
 387         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 388                "event type %d, status %d, desc %p\n",
 389                ev->type, ev->status, desc);
 390
 391         cfs_spin_lock(&desc->bd_lock);
 392
 393         if ((ev->type == LNET_EVENT_ACK ||
 394              ev->type == LNET_EVENT_REPLY) &&
 395             ev->status == 0) {
 396                 /* We heard back from the peer, so even if we get this
 397                  * before the SENT event (oh yes we can), we know we
 398                  * read/wrote the peer buffer and how much... */
 399                 desc->bd_success = 1;
 400                 desc->bd_nob_transferred = ev->mlength;
 401                 desc->bd_sender = ev->sender;
 402         }
 403
 404         if (ev->unlinked) {
 405                 /* This is the last callback no matter what... */
 406                 desc->bd_network_rw = 0;
 407                 cfs_waitq_signal(&desc->bd_waitq);
 408         }
 409
 410         cfs_spin_unlock(&desc->bd_lock);
 411         EXIT;
 412 }
 413 #endif
 414
 415 static void ptlrpc_master_callback(lnet_event_t *ev)
 416 {
 417         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 418         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 419
 420         /* Honestly, it's best to find out early. */
 421         LASSERT (cbid->cbid_arg != LP_POISON);
 422         LASSERT (callback == request_out_callback ||
 423                  callback == reply_in_callback ||
 424                  callback == client_bulk_callback ||
 425                  callback == request_in_callback ||
 426                  callback == reply_out_callback
 427 #ifdef HAVE_SERVER_SUPPORT
 428                  || callback == server_bulk_callback
 429 #endif
 430                  );
 431
 432         callback (ev);
 433 }
 434
 435 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 436                          lnet_process_id_t *peer, lnet_nid_t *self)
 437 {
 438         int               best_dist = 0;
 439         __u32             best_order = 0;
 440         int               count = 0;
 441         int               rc = -ENOENT;
 442         int               portals_compatibility;
 443         int               dist;
 444         __u32             order;
 445         lnet_nid_t        dst_nid;
 446         lnet_nid_t        src_nid;
 447
 448         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 449
 450         peer->pid = LUSTRE_SRV_LNET_PID;
 451
 452         /* Choose the matching UUID that's closest */
 453         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 454                 dist = LNetDist(dst_nid, &src_nid, &order);
 455                 if (dist < 0)
 456                         continue;
 457
 458                 if (dist == 0) {                /* local! use loopback LND */
 459                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 460                         rc = 0;
 461                         break;
 462                 }
 463
 464                 if (rc < 0 ||
 465                     dist < best_dist ||
 466                     (dist == best_dist && order < best_order)) {
 467                         best_dist = dist;
 468                         best_order = order;
 469
 470                         if (portals_compatibility > 1) {
 471                                 /* Strong portals compatibility: Zero the nid's
 472                                  * NET, so if I'm reading new config logs, or
 473                                  * getting configured by (new) lconf I can
 474                                  * still talk to old servers. */
 475                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 476                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 477                         }
 478                         peer->nid = dst_nid;
 479                         *self = src_nid;
 480                         rc = 0;
 481                 }
 482         }
 483
 484         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 485         return rc;
 486 }
 487
 488 void ptlrpc_ni_fini(void)
 489 {
 490         cfs_waitq_t         waitq;
 491         struct l_wait_info  lwi;
 492         int                 rc;
 493         int                 retries;
 494
 495         /* Wait for the event queue to become idle since there may still be
 496          * messages in flight with pending events (i.e. the fire-and-forget
 497          * messages == client requests and "non-difficult" server
 498          * replies */
 499
 500         for (retries = 0;; retries++) {
 501                 rc = LNetEQFree(ptlrpc_eq_h);
 502                 switch (rc) {
 503                 default:
 504                         LBUG();
 505
 506                 case 0:
 507                         LNetNIFini();
 508                         return;
 509
 510                 case -EBUSY:
 511                         if (retries != 0)
 512                                 CWARN("Event queue still busy\n");
 513
 514                         /* Wait for a bit */
 515                         cfs_waitq_init(&waitq);
 516                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 517                         l_wait_event(waitq, 0, &lwi);
 518                         break;
 519                 }
 520         }
 521         /* notreached */
 522 }
 523
 524 lnet_pid_t ptl_get_pid(void)
 525 {
 526         lnet_pid_t        pid;
 527
 528 #ifndef  __KERNEL__
 529         pid = getpid();
 530 #else
 531         pid = LUSTRE_SRV_LNET_PID;
 532 #endif
 533         return pid;
 534 }
 535
 536 int ptlrpc_ni_init(void)
 537 {
 538         int              rc;
 539         lnet_pid_t       pid;
 540
 541         pid = ptl_get_pid();
 542         CDEBUG(D_NET, "My pid is: %x\n", pid);
 543
 544         /* We're not passing any limits yet... */
 545         rc = LNetNIInit(pid);
 546         if (rc < 0) {
 547                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 548                 return (-ENOENT);
 549         }
 550
 551         /* CAVEAT EMPTOR: how we process portals events is _radically_
 552          * different depending on... */
 553 #ifdef __KERNEL__
 554         /* kernel portals calls our master callback when events are added to
 555          * the event queue.  In fact lustre never pulls events off this queue,
 556          * so it's only sized for some debug history. */
 557         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 558 #else
 559         /* liblustre calls the master callback when it removes events from the
 560          * event queue.  The event queue has to be big enough not to drop
 561          * anything */
 562         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 563 #endif
 564         if (rc == 0)
 565                 return 0;
 566
 567         CERROR ("Failed to allocate event queue: %d\n", rc);
 568         LNetNIFini();
 569
 570         return (-ENOMEM);
 571 }
 572
 573 #ifndef __KERNEL__
 574 CFS_LIST_HEAD(liblustre_wait_callbacks);
 575 CFS_LIST_HEAD(liblustre_idle_callbacks);
 576 void *liblustre_services_callback;
 577
 578 void *
 579 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 580                                       const char *name,
 581                                       int (*fn)(void *arg), void *arg)
 582 {
 583         struct liblustre_wait_callback *llwc;
 584
 585         OBD_ALLOC(llwc, sizeof(*llwc));
 586         LASSERT (llwc != NULL);
 587
 588         llwc->llwc_name = name;
 589         llwc->llwc_fn = fn;
 590         llwc->llwc_arg = arg;
 591         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 592
 593         return (llwc);
 594 }
 595
 596 void
 597 liblustre_deregister_waitidle_callback (void *opaque)
 598 {
 599         struct liblustre_wait_callback *llwc = opaque;
 600
 601         cfs_list_del(&llwc->llwc_list);
 602         OBD_FREE(llwc, sizeof(*llwc));
 603 }
 604
 605 void *
 606 liblustre_register_wait_callback (const char *name,
 607                                   int (*fn)(void *arg), void *arg)
 608 {
 609         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 610                                                     name, fn, arg);
 611 }
 612
 613 void
 614 liblustre_deregister_wait_callback (void *opaque)
 615 {
 616         liblustre_deregister_waitidle_callback(opaque);
 617 }
 618
 619 void *
 620 liblustre_register_idle_callback (const char *name,
 621                                   int (*fn)(void *arg), void *arg)
 622 {
 623         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 624                                                     name, fn, arg);
 625 }
 626
 627 void
 628 liblustre_deregister_idle_callback (void *opaque)
 629 {
 630         liblustre_deregister_waitidle_callback(opaque);
 631 }
 632
 633 int
 634 liblustre_check_events (int timeout)
 635 {
 636         lnet_event_t ev;
 637         int         rc;
 638         int         i;
 639         ENTRY;
 640
 641         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 642         if (rc == 0)
 643                 RETURN(0);
 644
 645         LASSERT (rc == -EOVERFLOW || rc == 1);
 646
 647         /* liblustre: no asynch callback so we can't afford to miss any
 648          * events... */
 649         if (rc == -EOVERFLOW) {
 650                 CERROR ("Dropped an event!!!\n");
 651                 abort();
 652         }
 653
 654         ptlrpc_master_callback (&ev);
 655         RETURN(1);
 656 }
 657
 658 int liblustre_waiting = 0;
 659
 660 int
 661 liblustre_wait_event (int timeout)
 662 {
 663         cfs_list_t                     *tmp;
 664         struct liblustre_wait_callback *llwc;
 665         int                             found_something = 0;
 666
 667         /* single threaded recursion check... */
 668         liblustre_waiting = 1;
 669
 670         for (;;) {
 671                 /* Deal with all pending events */
 672                 while (liblustre_check_events(0))
 673                         found_something = 1;
 674
 675                 /* Give all registered callbacks a bite at the cherry */
 676                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 677                         llwc = cfs_list_entry(tmp,
 678                                               struct liblustre_wait_callback,
 679                                               llwc_list);
 680
 681                         if (llwc->llwc_fn(llwc->llwc_arg))
 682                                 found_something = 1;
 683                 }
 684
 685                 if (found_something || timeout == 0)
 686                         break;
 687
 688                 /* Nothing so far, but I'm allowed to block... */
 689                 found_something = liblustre_check_events(timeout);
 690                 if (!found_something)           /* still nothing */
 691                         break;                  /* I timed out */
 692         }
 693
 694         liblustre_waiting = 0;
 695
 696         return found_something;
 697 }
 698
 699 void
 700 liblustre_wait_idle(void)
 701 {
 702         static int recursed = 0;
 703
 704         cfs_list_t                     *tmp;
 705         struct liblustre_wait_callback *llwc;
 706         int                             idle = 0;
 707
 708         LASSERT(!recursed);
 709         recursed = 1;
 710
 711         do {
 712                 liblustre_wait_event(0);
 713
 714                 idle = 1;
 715
 716                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 717                         llwc = cfs_list_entry(tmp,
 718                                               struct liblustre_wait_callback,
 719                                               llwc_list);
 720
 721                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 722                                 idle = 0;
 723                                 break;
 724                         }
 725                 }
 726
 727         } while (!idle);
 728
 729         recursed = 0;
 730 }
 731
 732 #endif /* __KERNEL__ */
 733
 734 int ptlrpc_init_portals(void)
 735 {
 736         int   rc = ptlrpc_ni_init();
 737
 738         if (rc != 0) {
 739                 CERROR("network initialisation failed\n");
 740                 return -EIO;
 741         }
 742 #ifndef __KERNEL__
 743         liblustre_services_callback =
 744                 liblustre_register_wait_callback("liblustre_check_services",
 745                                                  &liblustre_check_services,
 746                                                  NULL);
 747         cfs_init_completion_module(liblustre_wait_event);
 748 #endif
 749         rc = ptlrpcd_addref();
 750         if (rc == 0)
 751                 return 0;
 752
 753         CERROR("rpcd initialisation failed\n");
 754 #ifndef __KERNEL__
 755         liblustre_deregister_wait_callback(liblustre_services_callback);
 756 #endif
 757         ptlrpc_ni_fini();
 758         return rc;
 759 }
 760
 761 void ptlrpc_exit_portals(void)
 762 {
 763 #ifndef __KERNEL__
 764         liblustre_deregister_wait_callback(liblustre_services_callback);
 765 #endif
 766         ptlrpcd_decref();
 767         ptlrpc_ni_fini();
 768 }