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 /*
 371  * Server's bulk completion callback
 372  */
 373 void server_bulk_callback (lnet_event_t *ev)
 374 {
 375         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 376         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 377         ENTRY;
 378
 379         LASSERT (ev->type == LNET_EVENT_SEND ||
 380                  ev->type == LNET_EVENT_UNLINK ||
 381                  (desc->bd_type == BULK_PUT_SOURCE &&
 382                   ev->type == LNET_EVENT_ACK) ||
 383                  (desc->bd_type == BULK_GET_SINK &&
 384                   ev->type == LNET_EVENT_REPLY));
 385
 386         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 387                "event type %d, status %d, desc %p\n",
 388                ev->type, ev->status, desc);
 389
 390         cfs_spin_lock(&desc->bd_lock);
 391
 392         if ((ev->type == LNET_EVENT_ACK ||
 393              ev->type == LNET_EVENT_REPLY) &&
 394             ev->status == 0) {
 395                 /* We heard back from the peer, so even if we get this
 396                  * before the SENT event (oh yes we can), we know we
 397                  * read/wrote the peer buffer and how much... */
 398                 desc->bd_success = 1;
 399                 desc->bd_nob_transferred = ev->mlength;
 400                 desc->bd_sender = ev->sender;
 401         }
 402
 403         if (ev->unlinked) {
 404                 /* This is the last callback no matter what... */
 405                 desc->bd_network_rw = 0;
 406                 cfs_waitq_signal(&desc->bd_waitq);
 407         }
 408
 409         cfs_spin_unlock(&desc->bd_lock);
 410         EXIT;
 411 }
 412
 413 static void ptlrpc_master_callback(lnet_event_t *ev)
 414 {
 415         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 416         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 417
 418         /* Honestly, it's best to find out early. */
 419         LASSERT (cbid->cbid_arg != LP_POISON);
 420         LASSERT (callback == request_out_callback ||
 421                  callback == reply_in_callback ||
 422                  callback == client_bulk_callback ||
 423                  callback == request_in_callback ||
 424                  callback == reply_out_callback ||
 425                  callback == server_bulk_callback);
 426
 427         callback (ev);
 428 }
 429
 430 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 431                          lnet_process_id_t *peer, lnet_nid_t *self)
 432 {
 433         int               best_dist = 0;
 434         __u32             best_order = 0;
 435         int               count = 0;
 436         int               rc = -ENOENT;
 437         int               portals_compatibility;
 438         int               dist;
 439         __u32             order;
 440         lnet_nid_t        dst_nid;
 441         lnet_nid_t        src_nid;
 442
 443         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 444
 445         peer->pid = LUSTRE_SRV_LNET_PID;
 446
 447         /* Choose the matching UUID that's closest */
 448         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 449                 dist = LNetDist(dst_nid, &src_nid, &order);
 450                 if (dist < 0)
 451                         continue;
 452
 453                 if (dist == 0) {                /* local! use loopback LND */
 454                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 455                         rc = 0;
 456                         break;
 457                 }
 458
 459                 if (rc < 0 ||
 460                     dist < best_dist ||
 461                     (dist == best_dist && order < best_order)) {
 462                         best_dist = dist;
 463                         best_order = order;
 464
 465                         if (portals_compatibility > 1) {
 466                                 /* Strong portals compatibility: Zero the nid's
 467                                  * NET, so if I'm reading new config logs, or
 468                                  * getting configured by (new) lconf I can
 469                                  * still talk to old servers. */
 470                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 471                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 472                         }
 473                         peer->nid = dst_nid;
 474                         *self = src_nid;
 475                         rc = 0;
 476                 }
 477         }
 478
 479         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 480         return rc;
 481 }
 482
 483 void ptlrpc_ni_fini(void)
 484 {
 485         cfs_waitq_t         waitq;
 486         struct l_wait_info  lwi;
 487         int                 rc;
 488         int                 retries;
 489
 490         /* Wait for the event queue to become idle since there may still be
 491          * messages in flight with pending events (i.e. the fire-and-forget
 492          * messages == client requests and "non-difficult" server
 493          * replies */
 494
 495         for (retries = 0;; retries++) {
 496                 rc = LNetEQFree(ptlrpc_eq_h);
 497                 switch (rc) {
 498                 default:
 499                         LBUG();
 500
 501                 case 0:
 502                         LNetNIFini();
 503                         return;
 504
 505                 case -EBUSY:
 506                         if (retries != 0)
 507                                 CWARN("Event queue still busy\n");
 508
 509                         /* Wait for a bit */
 510                         cfs_waitq_init(&waitq);
 511                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 512                         l_wait_event(waitq, 0, &lwi);
 513                         break;
 514                 }
 515         }
 516         /* notreached */
 517 }
 518
 519 lnet_pid_t ptl_get_pid(void)
 520 {
 521         lnet_pid_t        pid;
 522
 523 #ifndef  __KERNEL__
 524         pid = getpid();
 525 #else
 526         pid = LUSTRE_SRV_LNET_PID;
 527 #endif
 528         return pid;
 529 }
 530
 531 int ptlrpc_ni_init(void)
 532 {
 533         int              rc;
 534         lnet_pid_t       pid;
 535
 536         pid = ptl_get_pid();
 537         CDEBUG(D_NET, "My pid is: %x\n", pid);
 538
 539         /* We're not passing any limits yet... */
 540         rc = LNetNIInit(pid);
 541         if (rc < 0) {
 542                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 543                 return (-ENOENT);
 544         }
 545
 546         /* CAVEAT EMPTOR: how we process portals events is _radically_
 547          * different depending on... */
 548 #ifdef __KERNEL__
 549         /* kernel portals calls our master callback when events are added to
 550          * the event queue.  In fact lustre never pulls events off this queue,
 551          * so it's only sized for some debug history. */
 552         rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
 553 #else
 554         /* liblustre calls the master callback when it removes events from the
 555          * event queue.  The event queue has to be big enough not to drop
 556          * anything */
 557         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 558 #endif
 559         if (rc == 0)
 560                 return 0;
 561
 562         CERROR ("Failed to allocate event queue: %d\n", rc);
 563         LNetNIFini();
 564
 565         return (-ENOMEM);
 566 }
 567
 568 #ifndef __KERNEL__
 569 CFS_LIST_HEAD(liblustre_wait_callbacks);
 570 CFS_LIST_HEAD(liblustre_idle_callbacks);
 571 void *liblustre_services_callback;
 572
 573 void *
 574 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 575                                       const char *name,
 576                                       int (*fn)(void *arg), void *arg)
 577 {
 578         struct liblustre_wait_callback *llwc;
 579
 580         OBD_ALLOC(llwc, sizeof(*llwc));
 581         LASSERT (llwc != NULL);
 582
 583         llwc->llwc_name = name;
 584         llwc->llwc_fn = fn;
 585         llwc->llwc_arg = arg;
 586         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 587
 588         return (llwc);
 589 }
 590
 591 void
 592 liblustre_deregister_waitidle_callback (void *opaque)
 593 {
 594         struct liblustre_wait_callback *llwc = opaque;
 595
 596         cfs_list_del(&llwc->llwc_list);
 597         OBD_FREE(llwc, sizeof(*llwc));
 598 }
 599
 600 void *
 601 liblustre_register_wait_callback (const char *name,
 602                                   int (*fn)(void *arg), void *arg)
 603 {
 604         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 605                                                     name, fn, arg);
 606 }
 607
 608 void
 609 liblustre_deregister_wait_callback (void *opaque)
 610 {
 611         liblustre_deregister_waitidle_callback(opaque);
 612 }
 613
 614 void *
 615 liblustre_register_idle_callback (const char *name,
 616                                   int (*fn)(void *arg), void *arg)
 617 {
 618         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 619                                                     name, fn, arg);
 620 }
 621
 622 void
 623 liblustre_deregister_idle_callback (void *opaque)
 624 {
 625         liblustre_deregister_waitidle_callback(opaque);
 626 }
 627
 628 int
 629 liblustre_check_events (int timeout)
 630 {
 631         lnet_event_t ev;
 632         int         rc;
 633         int         i;
 634         ENTRY;
 635
 636         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 637         if (rc == 0)
 638                 RETURN(0);
 639
 640         LASSERT (rc == -EOVERFLOW || rc == 1);
 641
 642         /* liblustre: no asynch callback so we can't afford to miss any
 643          * events... */
 644         if (rc == -EOVERFLOW) {
 645                 CERROR ("Dropped an event!!!\n");
 646                 abort();
 647         }
 648
 649         ptlrpc_master_callback (&ev);
 650         RETURN(1);
 651 }
 652
 653 int liblustre_waiting = 0;
 654
 655 int
 656 liblustre_wait_event (int timeout)
 657 {
 658         cfs_list_t                     *tmp;
 659         struct liblustre_wait_callback *llwc;
 660         int                             found_something = 0;
 661
 662         /* single threaded recursion check... */
 663         liblustre_waiting = 1;
 664
 665         for (;;) {
 666                 /* Deal with all pending events */
 667                 while (liblustre_check_events(0))
 668                         found_something = 1;
 669
 670                 /* Give all registered callbacks a bite at the cherry */
 671                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 672                         llwc = cfs_list_entry(tmp,
 673                                               struct liblustre_wait_callback,
 674                                               llwc_list);
 675
 676                         if (llwc->llwc_fn(llwc->llwc_arg))
 677                                 found_something = 1;
 678                 }
 679
 680                 if (found_something || timeout == 0)
 681                         break;
 682
 683                 /* Nothing so far, but I'm allowed to block... */
 684                 found_something = liblustre_check_events(timeout);
 685                 if (!found_something)           /* still nothing */
 686                         break;                  /* I timed out */
 687         }
 688
 689         liblustre_waiting = 0;
 690
 691         return found_something;
 692 }
 693
 694 void
 695 liblustre_wait_idle(void)
 696 {
 697         static int recursed = 0;
 698
 699         cfs_list_t                     *tmp;
 700         struct liblustre_wait_callback *llwc;
 701         int                             idle = 0;
 702
 703         LASSERT(!recursed);
 704         recursed = 1;
 705
 706         do {
 707                 liblustre_wait_event(0);
 708
 709                 idle = 1;
 710
 711                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 712                         llwc = cfs_list_entry(tmp,
 713                                               struct liblustre_wait_callback,
 714                                               llwc_list);
 715
 716                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 717                                 idle = 0;
 718                                 break;
 719                         }
 720                 }
 721
 722         } while (!idle);
 723
 724         recursed = 0;
 725 }
 726
 727 #endif /* __KERNEL__ */
 728
 729 int ptlrpc_init_portals(void)
 730 {
 731         int   rc = ptlrpc_ni_init();
 732
 733         if (rc != 0) {
 734                 CERROR("network initialisation failed\n");
 735                 return -EIO;
 736         }
 737 #ifndef __KERNEL__
 738         liblustre_services_callback =
 739                 liblustre_register_wait_callback("liblustre_check_services",
 740                                                  &liblustre_check_services,
 741                                                  NULL);
 742         cfs_init_completion_module(liblustre_wait_event);
 743 #endif
 744         rc = ptlrpcd_addref();
 745         if (rc == 0)
 746                 return 0;
 747
 748         CERROR("rpcd initialisation failed\n");
 749 #ifndef __KERNEL__
 750         liblustre_deregister_wait_callback(liblustre_services_callback);
 751 #endif
 752         ptlrpc_ni_fini();
 753         return rc;
 754 }
 755
 756 void ptlrpc_exit_portals(void)
 757 {
 758 #ifndef __KERNEL__
 759         liblustre_deregister_wait_callback(liblustre_services_callback);
 760 #endif
 761         ptlrpcd_decref();
 762         ptlrpc_ni_fini();
 763 }