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