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 # ifdef __mips64__
  43 #  include <linux/kernel.h>
  44 # endif
  45 #endif
  46 #include <obd_class.h>
  47 #include <lustre_net.h>
  48 #include "ptlrpc_internal.h"
  49
  50 lnet_handle_eq_t   ptlrpc_eq_h;
  51
  52 /*
  53  *  Client's outgoing request callback
  54  */
  55 void request_out_callback(lnet_event_t *ev)
  56 {
  57         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  58         struct ptlrpc_request *req = cbid->cbid_arg;
  59         ENTRY;
  60
  61         LASSERT (ev->type == LNET_EVENT_SEND ||
  62                  ev->type == LNET_EVENT_UNLINK);
  63         LASSERT (ev->unlinked);
  64
  65         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  66
  67         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  68
  69                 /* Failed send: make it seem like the reply timed out, just
  70                  * like failing sends in client.c does currently...  */
  71
  72                 spin_lock(&req->rq_lock);
  73                 req->rq_net_err = 1;
  74                 spin_unlock(&req->rq_lock);
  75
  76                 ptlrpc_client_wake_req(req);
  77         }
  78
  79         ptlrpc_req_finished(req);
  80
  81         EXIT;
  82 }
  83
  84 /*
  85  * Client's incoming reply callback
  86  */
  87 void reply_in_callback(lnet_event_t *ev)
  88 {
  89         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  90         struct ptlrpc_request *req = cbid->cbid_arg;
  91         ENTRY;
  92
  93         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  94
  95         LASSERT(ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
  96         LASSERT(ev->md.start == req->rq_repbuf);
  97         LASSERT(ev->offset + ev->mlength <= req->rq_replen);
  98         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
  99            for adaptive timeouts' early reply. */
 100         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 101
 102         spin_lock(&req->rq_lock);
 103
 104         req->rq_receiving_reply = 0;
 105         req->rq_early = 0;
 106         if (ev->unlinked)
 107                 req->rq_must_unlink = 0;
 108
 109         if (ev->status)
 110                 goto out_wake;
 111
 112         if (ev->type == LNET_EVENT_UNLINK) {
 113                 LASSERT(ev->unlinked);
 114                 DEBUG_REQ(D_NET, req, "unlink");
 115                 goto out_wake;
 116         }
 117
 118         if (ev->mlength < ev->rlength ) {
 119                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 120                        req->rq_replen, ev->rlength, ev->offset);
 121                 req->rq_reply_truncate = 1;
 122                 req->rq_replied = 1;
 123                 req->rq_status = -EOVERFLOW;
 124                 req->rq_nob_received = ev->rlength + ev->offset;
 125                 goto out_wake;
 126         }
 127
 128         if ((ev->offset == 0) &&
 129             (lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 130                 /* Early reply */
 131                 DEBUG_REQ(D_ADAPTTO, req,
 132                           "Early reply received: mlen=%u offset=%d replen=%d "
 133                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 134                           req->rq_replen, req->rq_replied, ev->unlinked);
 135
 136                 if (unlikely(ev->mlength != lustre_msg_early_size(req)))
 137                         CERROR("early reply sized %u, expect %u\n",
 138                                ev->mlength, lustre_msg_early_size(req));
 139
 140                 req->rq_early_count++; /* number received, client side */
 141
 142                 if (req->rq_replied)   /* already got the real reply */
 143                         goto out_wake;
 144
 145                 req->rq_early = 1;
 146                 req->rq_nob_received = ev->mlength;
 147                 /* repmsg points to early reply */
 148                 req->rq_repmsg = req->rq_repbuf;
 149                 /* And we're still receiving */
 150                 req->rq_receiving_reply = 1;
 151         } else {
 152                 /* Real reply */
 153                 req->rq_rep_swab_mask = 0;
 154                 req->rq_replied = 1;
 155                 req->rq_nob_received = ev->mlength;
 156                 /* repmsg points to real reply */
 157                 req->rq_repmsg = (struct lustre_msg *)((char *)req->rq_repbuf +
 158                                                        ev->offset);
 159                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 160                    so we must unlink in ptlrpc_unregister_reply */
 161                 DEBUG_REQ(D_INFO, req,
 162                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 163                           lustre_msg_get_flags(req->rq_reqmsg),
 164                           ev->mlength, ev->offset, req->rq_replen);
 165         }
 166
 167         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 168
 169 out_wake:
 170         /* NB don't unlock till after wakeup; req can disappear under us
 171          * since we don't have our own ref */
 172         ptlrpc_client_wake_req(req);
 173         spin_unlock(&req->rq_lock);
 174         EXIT;
 175 }
 176
 177 /*
 178  * Client's bulk has been written/read
 179  */
 180 void client_bulk_callback (lnet_event_t *ev)
 181 {
 182         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 183         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 184         ENTRY;
 185
 186         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB))
 187                 ev->status = -EIO;
 188
 189         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 190                   ev->type == LNET_EVENT_PUT) ||
 191                  (desc->bd_type == BULK_GET_SOURCE &&
 192                   ev->type == LNET_EVENT_GET) ||
 193                  ev->type == LNET_EVENT_UNLINK);
 194         LASSERT (ev->unlinked);
 195
 196         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 197                "event type %d, status %d, desc %p\n",
 198                ev->type, ev->status, desc);
 199
 200         spin_lock(&desc->bd_lock);
 201
 202         LASSERT(desc->bd_network_rw);
 203         desc->bd_network_rw = 0;
 204
 205         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 206                 desc->bd_success = 1;
 207                 desc->bd_nob_transferred = ev->mlength;
 208                 desc->bd_sender = ev->sender;
 209         }
 210
 211         /* NB don't unlock till after wakeup; desc can disappear under us
 212          * otherwise */
 213         ptlrpc_client_wake_req(desc->bd_req);
 214
 215         spin_unlock(&desc->bd_lock);
 216         EXIT;
 217 }
 218
 219 /*
 220  * Server's incoming request callback
 221  */
 222 void request_in_callback(lnet_event_t *ev)
 223 {
 224         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 225         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 226         struct ptlrpc_service             *service = rqbd->rqbd_service;
 227         struct ptlrpc_request             *req;
 228         ENTRY;
 229
 230         LASSERT (ev->type == LNET_EVENT_PUT ||
 231                  ev->type == LNET_EVENT_UNLINK);
 232         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 233         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 234                  rqbd->rqbd_buffer + service->srv_buf_size);
 235
 236         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 237                "event type %d, status %d, service %s\n",
 238                ev->type, ev->status, service->srv_name);
 239
 240         if (ev->unlinked) {
 241                 /* If this is the last request message to fit in the
 242                  * request buffer we can use the request object embedded in
 243                  * rqbd.  Note that if we failed to allocate a request,
 244                  * we'd have to re-post the rqbd, which we can't do in this
 245                  * context. */
 246                 req = &rqbd->rqbd_req;
 247                 memset(req, 0, sizeof (*req));
 248         } else {
 249                 LASSERT (ev->type == LNET_EVENT_PUT);
 250                 if (ev->status != 0) {
 251                         /* We moaned above already... */
 252                         return;
 253                 }
 254                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 255                 if (req == NULL) {
 256                         CERROR("Can't allocate incoming request descriptor: "
 257                                "Dropping %s RPC from %s\n",
 258                                service->srv_name,
 259                                libcfs_id2str(ev->initiator));
 260                         return;
 261                 }
 262         }
 263
 264         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 265          * flags are reset and scalars are zero.  We only set the message
 266          * size to non-zero if this was a successful receive. */
 267         req->rq_xid = ev->match_bits;
 268         req->rq_reqmsg = ev->md.start + ev->offset;
 269         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 270                 req->rq_reqlen = ev->mlength;
 271         do_gettimeofday(&req->rq_arrival_time);
 272         req->rq_peer = ev->initiator;
 273         req->rq_self = ev->target.nid;
 274         req->rq_rqbd = rqbd;
 275         req->rq_phase = RQ_PHASE_NEW;
 276 #ifdef CRAY_XT3
 277         req->rq_uid = ev->uid;
 278 #endif
 279         spin_lock_init(&req->rq_lock);
 280         CFS_INIT_LIST_HEAD(&req->rq_list);
 281         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 282         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
 283         CFS_INIT_LIST_HEAD(&req->rq_set_chain);
 284         CFS_INIT_LIST_HEAD(&req->rq_history_list);
 285         CFS_INIT_LIST_HEAD(&req->rq_exp_list);
 286         atomic_set(&req->rq_refcount, 1);
 287         if (ev->type == LNET_EVENT_PUT)
 288                 DEBUG_REQ(D_NET, req, "incoming req");
 289
 290         spin_lock(&service->srv_lock);
 291
 292         req->rq_history_seq = service->srv_request_seq++;
 293         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         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         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                 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_server_handle_reply() is done with it */
 354                 spin_lock(&svc->srv_lock);
 355                 rs->rs_on_net = 0;
 356                 ptlrpc_schedule_difficult_reply (rs);
 357                 spin_unlock(&svc->srv_lock);
 358         }
 359
 360         EXIT;
 361 }
 362
 363 /*
 364  * Server's bulk completion callback
 365  */
 366 void server_bulk_callback (lnet_event_t *ev)
 367 {
 368         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 369         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 370         ENTRY;
 371
 372         LASSERT (ev->type == LNET_EVENT_SEND ||
 373                  ev->type == LNET_EVENT_UNLINK ||
 374                  (desc->bd_type == BULK_PUT_SOURCE &&
 375                   ev->type == LNET_EVENT_ACK) ||
 376                  (desc->bd_type == BULK_GET_SINK &&
 377                   ev->type == LNET_EVENT_REPLY));
 378
 379         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 380                "event type %d, status %d, desc %p\n",
 381                ev->type, ev->status, desc);
 382
 383         spin_lock(&desc->bd_lock);
 384
 385         if ((ev->type == LNET_EVENT_ACK ||
 386              ev->type == LNET_EVENT_REPLY) &&
 387             ev->status == 0) {
 388                 /* We heard back from the peer, so even if we get this
 389                  * before the SENT event (oh yes we can), we know we
 390                  * read/wrote the peer buffer and how much... */
 391                 desc->bd_success = 1;
 392                 desc->bd_nob_transferred = ev->mlength;
 393                 desc->bd_sender = ev->sender;
 394         }
 395
 396         if (ev->unlinked) {
 397                 /* This is the last callback no matter what... */
 398                 desc->bd_network_rw = 0;
 399                 cfs_waitq_signal(&desc->bd_waitq);
 400         }
 401
 402         spin_unlock(&desc->bd_lock);
 403         EXIT;
 404 }
 405
 406 static void ptlrpc_master_callback(lnet_event_t *ev)
 407 {
 408         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 409         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 410
 411         /* Honestly, it's best to find out early. */
 412         LASSERT (cbid->cbid_arg != LP_POISON);
 413         LASSERT (callback == request_out_callback ||
 414                  callback == reply_in_callback ||
 415                  callback == client_bulk_callback ||
 416                  callback == request_in_callback ||
 417                  callback == reply_out_callback ||
 418                  callback == server_bulk_callback);
 419
 420         callback (ev);
 421 }
 422
 423 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 424                          lnet_process_id_t *peer, lnet_nid_t *self)
 425 {
 426         int               best_dist = 0;
 427         __u32             best_order = 0;
 428         int               count = 0;
 429         int               rc = -ENOENT;
 430         int               portals_compatibility;
 431         int               dist;
 432         __u32             order;
 433         lnet_nid_t        dst_nid;
 434         lnet_nid_t        src_nid;
 435
 436         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 437
 438         peer->pid = LUSTRE_SRV_LNET_PID;
 439
 440         /* Choose the matching UUID that's closest */
 441         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 442                 dist = LNetDist(dst_nid, &src_nid, &order);
 443                 if (dist < 0)
 444                         continue;
 445
 446                 if (dist == 0) {                /* local! use loopback LND */
 447                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 448                         rc = 0;
 449                         break;
 450                 }
 451
 452                 if (rc < 0 ||
 453                     dist < best_dist ||
 454                     (dist == best_dist && order < best_order)) {
 455                         best_dist = dist;
 456                         best_order = order;
 457
 458                         if (portals_compatibility > 1) {
 459                                 /* Strong portals compatibility: Zero the nid's
 460                                  * NET, so if I'm reading new config logs, or
 461                                  * getting configured by (new) lconf I can
 462                                  * still talk to old servers. */
 463                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 464                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 465                         }
 466                         peer->nid = dst_nid;
 467                         *self = src_nid;
 468                         rc = 0;
 469                 }
 470         }
 471
 472         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 473         return rc;
 474 }
 475
 476 void ptlrpc_ni_fini(void)
 477 {
 478         cfs_waitq_t         waitq;
 479         struct l_wait_info  lwi;
 480         int                 rc;
 481         int                 retries;
 482
 483         /* Wait for the event queue to become idle since there may still be
 484          * messages in flight with pending events (i.e. the fire-and-forget
 485          * messages == client requests and "non-difficult" server
 486          * replies */
 487
 488         for (retries = 0;; retries++) {
 489                 rc = LNetEQFree(ptlrpc_eq_h);
 490                 switch (rc) {
 491                 default:
 492                         LBUG();
 493
 494                 case 0:
 495                         LNetNIFini();
 496                         return;
 497
 498                 case -EBUSY:
 499                         if (retries != 0)
 500                                 CWARN("Event queue still busy\n");
 501
 502                         /* Wait for a bit */
 503                         cfs_waitq_init(&waitq);
 504                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 505                         l_wait_event(waitq, 0, &lwi);
 506                         break;
 507                 }
 508         }
 509         /* notreached */
 510 }
 511
 512 lnet_pid_t ptl_get_pid(void)
 513 {
 514         lnet_pid_t        pid;
 515
 516 #ifndef  __KERNEL__
 517         pid = getpid();
 518 #else
 519         pid = LUSTRE_SRV_LNET_PID;
 520 #endif
 521         return pid;
 522 }
 523
 524 int ptlrpc_ni_init(void)
 525 {
 526         int              rc;
 527         lnet_pid_t       pid;
 528
 529         pid = ptl_get_pid();
 530         CDEBUG(D_NET, "My pid is: %x\n", pid);
 531
 532         /* We're not passing any limits yet... */
 533         rc = LNetNIInit(pid);
 534         if (rc < 0) {
 535                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 536                 return (-ENOENT);
 537         }
 538
 539         /* CAVEAT EMPTOR: how we process portals events is _radically_
 540          * different depending on... */
 541 #ifdef __KERNEL__
 542         /* kernel portals calls our master callback when events are added to
 543          * the event queue.  In fact lustre never pulls events off this queue,
 544          * so it's only sized for some debug history. */
 545         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 546 #else
 547         /* liblustre calls the master callback when it removes events from the
 548          * event queue.  The event queue has to be big enough not to drop
 549          * anything */
 550         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 551 #endif
 552         if (rc == 0)
 553                 return 0;
 554
 555         CERROR ("Failed to allocate event queue: %d\n", rc);
 556         LNetNIFini();
 557
 558         return (-ENOMEM);
 559 }
 560
 561 #ifndef __KERNEL__
 562 CFS_LIST_HEAD(liblustre_wait_callbacks);
 563 CFS_LIST_HEAD(liblustre_idle_callbacks);
 564 void *liblustre_services_callback;
 565
 566 void *
 567 liblustre_register_waitidle_callback (struct list_head *callback_list,
 568                                       const char *name,
 569                                       int (*fn)(void *arg), void *arg)
 570 {
 571         struct liblustre_wait_callback *llwc;
 572
 573         OBD_ALLOC(llwc, sizeof(*llwc));
 574         LASSERT (llwc != NULL);
 575
 576         llwc->llwc_name = name;
 577         llwc->llwc_fn = fn;
 578         llwc->llwc_arg = arg;
 579         list_add_tail(&llwc->llwc_list, callback_list);
 580
 581         return (llwc);
 582 }
 583
 584 void
 585 liblustre_deregister_waitidle_callback (void *opaque)
 586 {
 587         struct liblustre_wait_callback *llwc = opaque;
 588
 589         list_del(&llwc->llwc_list);
 590         OBD_FREE(llwc, sizeof(*llwc));
 591 }
 592
 593 void *
 594 liblustre_register_wait_callback (const char *name,
 595                                   int (*fn)(void *arg), void *arg)
 596 {
 597         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 598                                                     name, fn, arg);
 599 }
 600
 601 void
 602 liblustre_deregister_wait_callback (void *opaque)
 603 {
 604         liblustre_deregister_waitidle_callback(opaque);
 605 }
 606
 607 void *
 608 liblustre_register_idle_callback (const char *name,
 609                                   int (*fn)(void *arg), void *arg)
 610 {
 611         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 612                                                     name, fn, arg);
 613 }
 614
 615 void
 616 liblustre_deregister_idle_callback (void *opaque)
 617 {
 618         liblustre_deregister_waitidle_callback(opaque);
 619 }
 620
 621 int
 622 liblustre_check_events (int timeout)
 623 {
 624         lnet_event_t ev;
 625         int         rc;
 626         int         i;
 627         ENTRY;
 628
 629         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 630         if (rc == 0)
 631                 RETURN(0);
 632
 633         LASSERT (rc == -EOVERFLOW || rc == 1);
 634
 635         /* liblustre: no asynch callback so we can't affort to miss any
 636          * events... */
 637         if (rc == -EOVERFLOW) {
 638                 CERROR ("Dropped an event!!!\n");
 639                 abort();
 640         }
 641
 642         ptlrpc_master_callback (&ev);
 643         RETURN(1);
 644 }
 645
 646 int liblustre_waiting = 0;
 647
 648 int
 649 liblustre_wait_event (int timeout)
 650 {
 651         struct list_head               *tmp;
 652         struct liblustre_wait_callback *llwc;
 653         int                             found_something = 0;
 654
 655         /* single threaded recursion check... */
 656         liblustre_waiting = 1;
 657
 658         for (;;) {
 659                 /* Deal with all pending events */
 660                 while (liblustre_check_events(0))
 661                         found_something = 1;
 662
 663                 /* Give all registered callbacks a bite at the cherry */
 664                 list_for_each(tmp, &liblustre_wait_callbacks) {
 665                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 666                                           llwc_list);
 667
 668                         if (llwc->llwc_fn(llwc->llwc_arg))
 669                                 found_something = 1;
 670                 }
 671
 672                 if (found_something || timeout == 0)
 673                         break;
 674
 675                 /* Nothing so far, but I'm allowed to block... */
 676                 found_something = liblustre_check_events(timeout);
 677                 if (!found_something)           /* still nothing */
 678                         break;                  /* I timed out */
 679         }
 680
 681         liblustre_waiting = 0;
 682
 683         return found_something;
 684 }
 685
 686 void
 687 liblustre_wait_idle(void)
 688 {
 689         static int recursed = 0;
 690
 691         struct list_head               *tmp;
 692         struct liblustre_wait_callback *llwc;
 693         int                             idle = 0;
 694
 695         LASSERT(!recursed);
 696         recursed = 1;
 697
 698         do {
 699                 liblustre_wait_event(0);
 700
 701                 idle = 1;
 702
 703                 list_for_each(tmp, &liblustre_idle_callbacks) {
 704                         llwc = list_entry(tmp, struct liblustre_wait_callback,
 705                                           llwc_list);
 706
 707                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 708                                 idle = 0;
 709                                 break;
 710                         }
 711                 }
 712
 713         } while (!idle);
 714
 715         recursed = 0;
 716 }
 717
 718 #endif /* __KERNEL__ */
 719
 720 int ptlrpc_init_portals(void)
 721 {
 722         int   rc = ptlrpc_ni_init();
 723
 724         if (rc != 0) {
 725                 CERROR("network initialisation failed\n");
 726                 return -EIO;
 727         }
 728 #ifndef __KERNEL__
 729         liblustre_services_callback =
 730                 liblustre_register_wait_callback("liblustre_check_services",
 731                                                  &liblustre_check_services, NULL);
 732 #endif
 733         rc = ptlrpcd_addref();
 734         if (rc == 0)
 735                 return 0;
 736
 737         CERROR("rpcd initialisation failed\n");
 738 #ifndef __KERNEL__
 739         liblustre_deregister_wait_callback(liblustre_services_callback);
 740 #endif
 741         ptlrpc_ni_fini();
 742         return rc;
 743 }
 744
 745 void ptlrpc_exit_portals(void)
 746 {
 747 #ifndef __KERNEL__
 748         liblustre_deregister_wait_callback(liblustre_services_callback);
 749 #endif
 750         ptlrpcd_decref();
 751         ptlrpc_ni_fini();
 752 }