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((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         req->rq_real_sent = cfs_time_current_sec();
  73
  74         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  75
  76                 /* Failed send: make it seem like the reply timed out, just
  77                  * like failing sends in client.c does currently...  */
  78
  79                 cfs_spin_lock(&req->rq_lock);
  80                 req->rq_net_err = 1;
  81                 cfs_spin_unlock(&req->rq_lock);
  82
  83                 ptlrpc_client_wake_req(req);
  84         }
  85
  86         ptlrpc_req_finished(req);
  87
  88         EXIT;
  89 }
  90
  91 /*
  92  * Client's incoming reply callback
  93  */
  94 void reply_in_callback(lnet_event_t *ev)
  95 {
  96         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  97         struct ptlrpc_request *req = cbid->cbid_arg;
  98         ENTRY;
  99
 100         DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
 101                   "type %d, status %d", ev->type, ev->status);
 102
 103         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 104         LASSERT (ev->md.start == req->rq_repbuf);
 105         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
 106         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 107            for adaptive timeouts' early reply. */
 108         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 109
 110         cfs_spin_lock(&req->rq_lock);
 111
 112         req->rq_receiving_reply = 0;
 113         req->rq_early = 0;
 114         if (ev->unlinked)
 115                 req->rq_must_unlink = 0;
 116
 117         if (ev->status)
 118                 goto out_wake;
 119
 120         if (ev->type == LNET_EVENT_UNLINK) {
 121                 LASSERT(ev->unlinked);
 122                 DEBUG_REQ(D_RPCTRACE, req, "unlink");
 123                 goto out_wake;
 124         }
 125
 126         if (ev->mlength < ev->rlength ) {
 127                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 128                        req->rq_replen, ev->rlength, ev->offset);
 129                 req->rq_reply_truncate = 1;
 130                 req->rq_replied = 1;
 131                 req->rq_status = -EOVERFLOW;
 132                 req->rq_nob_received = ev->rlength + ev->offset;
 133                 goto out_wake;
 134         }
 135
 136         if ((ev->offset == 0) &&
 137             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 138                 /* Early reply */
 139                 DEBUG_REQ(D_ADAPTTO, req,
 140                           "Early reply received: mlen=%u offset=%d replen=%d "
 141                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 142                           req->rq_replen, req->rq_replied, ev->unlinked);
 143
 144                 req->rq_early_count++; /* number received, client side */
 145
 146                 if (req->rq_replied)   /* already got the real reply */
 147                         goto out_wake;
 148
 149                 req->rq_early = 1;
 150                 req->rq_reply_off = ev->offset;
 151                 req->rq_nob_received = ev->mlength;
 152                 /* And we're still receiving */
 153                 req->rq_receiving_reply = 1;
 154         } else {
 155                 /* Real reply */
 156                 req->rq_rep_swab_mask = 0;
 157                 req->rq_replied = 1;
 158                 req->rq_reply_off = ev->offset;
 159                 req->rq_nob_received = ev->mlength;
 160                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 161                    so we must unlink in ptlrpc_unregister_reply */
 162                 DEBUG_REQ(D_INFO, req,
 163                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 164                           lustre_msg_get_flags(req->rq_reqmsg),
 165                           ev->mlength, ev->offset, req->rq_replen);
 166         }
 167
 168         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 169
 170 out_wake:
 171         /* NB don't unlock till after wakeup; req can disappear under us
 172          * since we don't have our own ref */
 173         ptlrpc_client_wake_req(req);
 174         cfs_spin_unlock(&req->rq_lock);
 175         EXIT;
 176 }
 177
 178 /*
 179  * Client's bulk has been written/read
 180  */
 181 void client_bulk_callback (lnet_event_t *ev)
 182 {
 183         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 184         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 185         ENTRY;
 186
 187         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 188                   ev->type == LNET_EVENT_PUT) ||
 189                  (desc->bd_type == BULK_GET_SOURCE &&
 190                   ev->type == LNET_EVENT_GET) ||
 191                  ev->type == LNET_EVENT_UNLINK);
 192         LASSERT (ev->unlinked);
 193
 194         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 195                "event type %d, status %d, desc %p\n",
 196                ev->type, ev->status, desc);
 197
 198         cfs_spin_lock(&desc->bd_lock);
 199
 200         LASSERT(desc->bd_network_rw);
 201         desc->bd_network_rw = 0;
 202
 203         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 204                 desc->bd_success = 1;
 205                 desc->bd_nob_transferred = ev->mlength;
 206                 desc->bd_sender = ev->sender;
 207         }
 208
 209         /* release the encrypted pages for write */
 210         if (desc->bd_req->rq_bulk_write)
 211                 sptlrpc_enc_pool_put_pages(desc);
 212
 213         /* NB don't unlock till after wakeup; desc can disappear under us
 214          * otherwise */
 215         ptlrpc_client_wake_req(desc->bd_req);
 216
 217         cfs_spin_unlock(&desc->bd_lock);
 218         EXIT;
 219 }
 220
 221 /*
 222  * Server's incoming request callback
 223  */
 224 void request_in_callback(lnet_event_t *ev)
 225 {
 226         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 227         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 228         struct ptlrpc_service             *service = rqbd->rqbd_service;
 229         struct ptlrpc_request             *req;
 230         ENTRY;
 231
 232         LASSERT (ev->type == LNET_EVENT_PUT ||
 233                  ev->type == LNET_EVENT_UNLINK);
 234         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 235         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 236                  rqbd->rqbd_buffer + service->srv_buf_size);
 237
 238         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 239                "event type %d, status %d, service %s\n",
 240                ev->type, ev->status, service->srv_name);
 241
 242         if (ev->unlinked) {
 243                 /* If this is the last request message to fit in the
 244                  * request buffer we can use the request object embedded in
 245                  * rqbd.  Note that if we failed to allocate a request,
 246                  * we'd have to re-post the rqbd, which we can't do in this
 247                  * context. */
 248                 req = &rqbd->rqbd_req;
 249                 memset(req, 0, sizeof (*req));
 250         } else {
 251                 LASSERT (ev->type == LNET_EVENT_PUT);
 252                 if (ev->status != 0) {
 253                         /* We moaned above already... */
 254                         return;
 255                 }
 256                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 257                 if (req == NULL) {
 258                         CERROR("Can't allocate incoming request descriptor: "
 259                                "Dropping %s RPC from %s\n",
 260                                service->srv_name,
 261                                libcfs_id2str(ev->initiator));
 262                         return;
 263                 }
 264         }
 265
 266         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 267          * flags are reset and scalars are zero.  We only set the message
 268          * size to non-zero if this was a successful receive. */
 269         req->rq_xid = ev->match_bits;
 270         req->rq_reqbuf = ev->md.start + ev->offset;
 271         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 272                 req->rq_reqdata_len = ev->mlength;
 273         cfs_gettimeofday(&req->rq_arrival_time);
 274         req->rq_peer = ev->initiator;
 275         req->rq_self = ev->target.nid;
 276         req->rq_rqbd = rqbd;
 277         req->rq_phase = RQ_PHASE_NEW;
 278 #ifdef CRAY_XT3
 279         req->rq_uid = ev->uid;
 280 #endif
 281         cfs_spin_lock_init(&req->rq_lock);
 282         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 283         cfs_atomic_set(&req->rq_refcount, 1);
 284         if (ev->type == LNET_EVENT_PUT)
 285                 CDEBUG(D_RPCTRACE, "incoming req@%p x"LPU64" msgsize %u\n",
 286                        req, req->rq_xid, ev->mlength);
 287
 288         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 289
 290         cfs_spin_lock(&service->srv_lock);
 291
 292         req->rq_history_seq = service->srv_request_seq++;
 293         cfs_list_add_tail(&req->rq_history_list, &service->srv_request_history);
 294
 295         if (ev->unlinked) {
 296                 service->srv_nrqbd_receiving--;
 297                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 298                        service->srv_nrqbd_receiving);
 299
 300                 /* Normally, don't complain about 0 buffers posted; LNET won't
 301                  * drop incoming reqs since we set the portal lazy */
 302                 if (test_req_buffer_pressure &&
 303                     ev->type != LNET_EVENT_UNLINK &&
 304                     service->srv_nrqbd_receiving == 0)
 305                         CWARN("All %s request buffers busy\n",
 306                               service->srv_name);
 307
 308                 /* req takes over the network's ref on rqbd */
 309         } else {
 310                 /* req takes a ref on rqbd */
 311                 rqbd->rqbd_refcount++;
 312         }
 313
 314         cfs_list_add_tail(&req->rq_list, &service->srv_req_in_queue);
 315         service->srv_n_queued_reqs++;
 316
 317         /* NB everything can disappear under us once the request
 318          * has been queued and we unlock, so do the wake now... */
 319         cfs_waitq_signal(&service->srv_waitq);
 320
 321         cfs_spin_unlock(&service->srv_lock);
 322         EXIT;
 323 }
 324
 325 /*
 326  *  Server's outgoing reply callback
 327  */
 328 void reply_out_callback(lnet_event_t *ev)
 329 {
 330         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 331         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 332         struct ptlrpc_service     *svc = rs->rs_service;
 333         ENTRY;
 334
 335         LASSERT (ev->type == LNET_EVENT_SEND ||
 336                  ev->type == LNET_EVENT_ACK ||
 337                  ev->type == LNET_EVENT_UNLINK);
 338
 339         if (!rs->rs_difficult) {
 340                 /* 'Easy' replies have no further processing so I drop the
 341                  * net's ref on 'rs' */
 342                 LASSERT (ev->unlinked);
 343                 ptlrpc_rs_decref(rs);
 344                 cfs_atomic_dec (&svc->srv_outstanding_replies);
 345                 EXIT;
 346                 return;
 347         }
 348
 349         LASSERT (rs->rs_on_net);
 350
 351         if (ev->unlinked) {
 352                 /* Last network callback. The net's ref on 'rs' stays put
 353                  * until ptlrpc_handle_rs() is done with it */
 354                 cfs_spin_lock(&svc->srv_lock);
 355                 cfs_spin_lock(&rs->rs_lock);
 356                 rs->rs_on_net = 0;
 357                 if (!rs->rs_no_ack ||
 358                     rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
 359                         ptlrpc_schedule_difficult_reply (rs);
 360                 cfs_spin_unlock(&rs->rs_lock);
 361                 cfs_spin_unlock(&svc->srv_lock);
 362         }
 363
 364         EXIT;
 365 }
 366
 367 /*
 368  * Server's bulk completion callback
 369  */
 370 void server_bulk_callback (lnet_event_t *ev)
 371 {
 372         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 373         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 374         ENTRY;
 375
 376         LASSERT (ev->type == LNET_EVENT_SEND ||
 377                  ev->type == LNET_EVENT_UNLINK ||
 378                  (desc->bd_type == BULK_PUT_SOURCE &&
 379                   ev->type == LNET_EVENT_ACK) ||
 380                  (desc->bd_type == BULK_GET_SINK &&
 381                   ev->type == LNET_EVENT_REPLY));
 382
 383         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 384                "event type %d, status %d, desc %p\n",
 385                ev->type, ev->status, desc);
 386
 387         cfs_spin_lock(&desc->bd_lock);
 388
 389         if ((ev->type == LNET_EVENT_ACK ||
 390              ev->type == LNET_EVENT_REPLY) &&
 391             ev->status == 0) {
 392                 /* We heard back from the peer, so even if we get this
 393                  * before the SENT event (oh yes we can), we know we
 394                  * read/wrote the peer buffer and how much... */
 395                 desc->bd_success = 1;
 396                 desc->bd_nob_transferred = ev->mlength;
 397                 desc->bd_sender = ev->sender;
 398         }
 399
 400         if (ev->unlinked) {
 401                 /* This is the last callback no matter what... */
 402                 desc->bd_network_rw = 0;
 403                 cfs_waitq_signal(&desc->bd_waitq);
 404         }
 405
 406         cfs_spin_unlock(&desc->bd_lock);
 407         EXIT;
 408 }
 409
 410 static void ptlrpc_master_callback(lnet_event_t *ev)
 411 {
 412         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 413         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 414
 415         /* Honestly, it's best to find out early. */
 416         LASSERT (cbid->cbid_arg != LP_POISON);
 417         LASSERT (callback == request_out_callback ||
 418                  callback == reply_in_callback ||
 419                  callback == client_bulk_callback ||
 420                  callback == request_in_callback ||
 421                  callback == reply_out_callback ||
 422                  callback == server_bulk_callback);
 423
 424         callback (ev);
 425 }
 426
 427 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 428                          lnet_process_id_t *peer, lnet_nid_t *self)
 429 {
 430         int               best_dist = 0;
 431         __u32             best_order = 0;
 432         int               count = 0;
 433         int               rc = -ENOENT;
 434         int               portals_compatibility;
 435         int               dist;
 436         __u32             order;
 437         lnet_nid_t        dst_nid;
 438         lnet_nid_t        src_nid;
 439
 440         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 441
 442         peer->pid = LUSTRE_SRV_LNET_PID;
 443
 444         /* Choose the matching UUID that's closest */
 445         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 446                 dist = LNetDist(dst_nid, &src_nid, &order);
 447                 if (dist < 0)
 448                         continue;
 449
 450                 if (dist == 0) {                /* local! use loopback LND */
 451                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 452                         rc = 0;
 453                         break;
 454                 }
 455
 456                 if (rc < 0 ||
 457                     dist < best_dist ||
 458                     (dist == best_dist && order < best_order)) {
 459                         best_dist = dist;
 460                         best_order = order;
 461
 462                         if (portals_compatibility > 1) {
 463                                 /* Strong portals compatibility: Zero the nid's
 464                                  * NET, so if I'm reading new config logs, or
 465                                  * getting configured by (new) lconf I can
 466                                  * still talk to old servers. */
 467                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 468                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 469                         }
 470                         peer->nid = dst_nid;
 471                         *self = src_nid;
 472                         rc = 0;
 473                 }
 474         }
 475
 476         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 477         return rc;
 478 }
 479
 480 void ptlrpc_ni_fini(void)
 481 {
 482         cfs_waitq_t         waitq;
 483         struct l_wait_info  lwi;
 484         int                 rc;
 485         int                 retries;
 486
 487         /* Wait for the event queue to become idle since there may still be
 488          * messages in flight with pending events (i.e. the fire-and-forget
 489          * messages == client requests and "non-difficult" server
 490          * replies */
 491
 492         for (retries = 0;; retries++) {
 493                 rc = LNetEQFree(ptlrpc_eq_h);
 494                 switch (rc) {
 495                 default:
 496                         LBUG();
 497
 498                 case 0:
 499                         LNetNIFini();
 500                         return;
 501
 502                 case -EBUSY:
 503                         if (retries != 0)
 504                                 CWARN("Event queue still busy\n");
 505
 506                         /* Wait for a bit */
 507                         cfs_waitq_init(&waitq);
 508                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 509                         l_wait_event(waitq, 0, &lwi);
 510                         break;
 511                 }
 512         }
 513         /* notreached */
 514 }
 515
 516 lnet_pid_t ptl_get_pid(void)
 517 {
 518         lnet_pid_t        pid;
 519
 520 #ifndef  __KERNEL__
 521         pid = getpid();
 522 #else
 523         pid = LUSTRE_SRV_LNET_PID;
 524 #endif
 525         return pid;
 526 }
 527
 528 int ptlrpc_ni_init(void)
 529 {
 530         int              rc;
 531         lnet_pid_t       pid;
 532
 533         pid = ptl_get_pid();
 534         CDEBUG(D_NET, "My pid is: %x\n", pid);
 535
 536         /* We're not passing any limits yet... */
 537         rc = LNetNIInit(pid);
 538         if (rc < 0) {
 539                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 540                 return (-ENOENT);
 541         }
 542
 543         /* CAVEAT EMPTOR: how we process portals events is _radically_
 544          * different depending on... */
 545 #ifdef __KERNEL__
 546         /* kernel portals calls our master callback when events are added to
 547          * the event queue.  In fact lustre never pulls events off this queue,
 548          * so it's only sized for some debug history. */
 549         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 550 #else
 551         /* liblustre calls the master callback when it removes events from the
 552          * event queue.  The event queue has to be big enough not to drop
 553          * anything */
 554         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 555 #endif
 556         if (rc == 0)
 557                 return 0;
 558
 559         CERROR ("Failed to allocate event queue: %d\n", rc);
 560         LNetNIFini();
 561
 562         return (-ENOMEM);
 563 }
 564
 565 #ifndef __KERNEL__
 566 CFS_LIST_HEAD(liblustre_wait_callbacks);
 567 CFS_LIST_HEAD(liblustre_idle_callbacks);
 568 void *liblustre_services_callback;
 569
 570 void *
 571 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 572                                       const char *name,
 573                                       int (*fn)(void *arg), void *arg)
 574 {
 575         struct liblustre_wait_callback *llwc;
 576
 577         OBD_ALLOC(llwc, sizeof(*llwc));
 578         LASSERT (llwc != NULL);
 579
 580         llwc->llwc_name = name;
 581         llwc->llwc_fn = fn;
 582         llwc->llwc_arg = arg;
 583         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 584
 585         return (llwc);
 586 }
 587
 588 void
 589 liblustre_deregister_waitidle_callback (void *opaque)
 590 {
 591         struct liblustre_wait_callback *llwc = opaque;
 592
 593         cfs_list_del(&llwc->llwc_list);
 594         OBD_FREE(llwc, sizeof(*llwc));
 595 }
 596
 597 void *
 598 liblustre_register_wait_callback (const char *name,
 599                                   int (*fn)(void *arg), void *arg)
 600 {
 601         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 602                                                     name, fn, arg);
 603 }
 604
 605 void
 606 liblustre_deregister_wait_callback (void *opaque)
 607 {
 608         liblustre_deregister_waitidle_callback(opaque);
 609 }
 610
 611 void *
 612 liblustre_register_idle_callback (const char *name,
 613                                   int (*fn)(void *arg), void *arg)
 614 {
 615         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 616                                                     name, fn, arg);
 617 }
 618
 619 void
 620 liblustre_deregister_idle_callback (void *opaque)
 621 {
 622         liblustre_deregister_waitidle_callback(opaque);
 623 }
 624
 625 int
 626 liblustre_check_events (int timeout)
 627 {
 628         lnet_event_t ev;
 629         int         rc;
 630         int         i;
 631         ENTRY;
 632
 633         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 634         if (rc == 0)
 635                 RETURN(0);
 636
 637         LASSERT (rc == -EOVERFLOW || rc == 1);
 638
 639         /* liblustre: no asynch callback so we can't afford to miss any
 640          * events... */
 641         if (rc == -EOVERFLOW) {
 642                 CERROR ("Dropped an event!!!\n");
 643                 abort();
 644         }
 645
 646         ptlrpc_master_callback (&ev);
 647         RETURN(1);
 648 }
 649
 650 int liblustre_waiting = 0;
 651
 652 int
 653 liblustre_wait_event (int timeout)
 654 {
 655         cfs_list_t                     *tmp;
 656         struct liblustre_wait_callback *llwc;
 657         int                             found_something = 0;
 658
 659         /* single threaded recursion check... */
 660         liblustre_waiting = 1;
 661
 662         for (;;) {
 663                 /* Deal with all pending events */
 664                 while (liblustre_check_events(0))
 665                         found_something = 1;
 666
 667                 /* Give all registered callbacks a bite at the cherry */
 668                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 669                         llwc = cfs_list_entry(tmp,
 670                                               struct liblustre_wait_callback,
 671                                               llwc_list);
 672
 673                         if (llwc->llwc_fn(llwc->llwc_arg))
 674                                 found_something = 1;
 675                 }
 676
 677                 if (found_something || timeout == 0)
 678                         break;
 679
 680                 /* Nothing so far, but I'm allowed to block... */
 681                 found_something = liblustre_check_events(timeout);
 682                 if (!found_something)           /* still nothing */
 683                         break;                  /* I timed out */
 684         }
 685
 686         liblustre_waiting = 0;
 687
 688         return found_something;
 689 }
 690
 691 void
 692 liblustre_wait_idle(void)
 693 {
 694         static int recursed = 0;
 695
 696         cfs_list_t                     *tmp;
 697         struct liblustre_wait_callback *llwc;
 698         int                             idle = 0;
 699
 700         LASSERT(!recursed);
 701         recursed = 1;
 702
 703         do {
 704                 liblustre_wait_event(0);
 705
 706                 idle = 1;
 707
 708                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 709                         llwc = cfs_list_entry(tmp,
 710                                               struct liblustre_wait_callback,
 711                                               llwc_list);
 712
 713                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 714                                 idle = 0;
 715                                 break;
 716                         }
 717                 }
 718
 719         } while (!idle);
 720
 721         recursed = 0;
 722 }
 723
 724 #endif /* __KERNEL__ */
 725
 726 int ptlrpc_init_portals(void)
 727 {
 728         int   rc = ptlrpc_ni_init();
 729
 730         if (rc != 0) {
 731                 CERROR("network initialisation failed\n");
 732                 return -EIO;
 733         }
 734 #ifndef __KERNEL__
 735         liblustre_services_callback =
 736                 liblustre_register_wait_callback("liblustre_check_services",
 737                                                  &liblustre_check_services,
 738                                                  NULL);
 739         cfs_init_completion_module(liblustre_wait_event);
 740 #endif
 741         rc = ptlrpcd_addref();
 742         if (rc == 0)
 743                 return 0;
 744
 745         CERROR("rpcd initialisation failed\n");
 746 #ifndef __KERNEL__
 747         liblustre_deregister_wait_callback(liblustre_services_callback);
 748 #endif
 749         ptlrpc_ni_fini();
 750         return rc;
 751 }
 752
 753 void ptlrpc_exit_portals(void)
 754 {
 755 #ifndef __KERNEL__
 756         liblustre_deregister_wait_callback(liblustre_services_callback);
 757 #endif
 758         ptlrpcd_decref();
 759         ptlrpc_ni_fini();
 760 }