lustre/ptlrpc/events.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_RPC
  36
  37 #ifndef __KERNEL__
  38 # include <liblustre.h>
  39 #else
  40 # include <libcfs/libcfs.h>
  41 # ifdef __mips64__
  42 #  include <linux/kernel.h>
  43 # endif
  44 #endif
  45
  46 #include <obd_class.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ptlrpc_internal.h"
  50
  51 lnet_handle_eq_t   ptlrpc_eq_h;
  52
  53 /*
  54  *  Client's outgoing request callback
  55  */
  56 void request_out_callback(lnet_event_t *ev)
  57 {
  58         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  59         struct ptlrpc_request *req = cbid->cbid_arg;
  60         ENTRY;
  61
  62         LASSERT (ev->type == LNET_EVENT_SEND ||
  63                  ev->type == LNET_EVENT_UNLINK);
  64         LASSERT (ev->unlinked);
  65
  66         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  67
  68         sptlrpc_request_out_callback(req);
  69         req->rq_real_sent = cfs_time_current_sec();
  70
  71         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  72
  73                 /* Failed send: make it seem like the reply timed out, just
  74                  * like failing sends in client.c does currently...  */
  75
  76                 cfs_spin_lock(&req->rq_lock);
  77                 req->rq_net_err = 1;
  78                 cfs_spin_unlock(&req->rq_lock);
  79
  80                 ptlrpc_client_wake_req(req);
  81         }
  82
  83         ptlrpc_req_finished(req);
  84
  85         EXIT;
  86 }
  87
  88 /*
  89  * Client's incoming reply callback
  90  */
  91 void reply_in_callback(lnet_event_t *ev)
  92 {
  93         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  94         struct ptlrpc_request *req = cbid->cbid_arg;
  95         ENTRY;
  96
  97         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  98
  99         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 100         LASSERT (ev->md.start == req->rq_repbuf);
 101         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
 102         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 103            for adaptive timeouts' early reply. */
 104         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 105
 106         cfs_spin_lock(&req->rq_lock);
 107
 108         req->rq_receiving_reply = 0;
 109         req->rq_early = 0;
 110         if (ev->unlinked)
 111                 req->rq_must_unlink = 0;
 112
 113         if (ev->status)
 114                 goto out_wake;
 115
 116         if (ev->type == LNET_EVENT_UNLINK) {
 117                 LASSERT(ev->unlinked);
 118                 DEBUG_REQ(D_NET, req, "unlink");
 119                 goto out_wake;
 120         }
 121
 122         if (ev->mlength < ev->rlength ) {
 123                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 124                        req->rq_replen, ev->rlength, ev->offset);
 125                 req->rq_reply_truncate = 1;
 126                 req->rq_replied = 1;
 127                 req->rq_status = -EOVERFLOW;
 128                 req->rq_nob_received = ev->rlength + ev->offset;
 129                 goto out_wake;
 130         }
 131
 132         if ((ev->offset == 0) &&
 133             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 134                 /* Early reply */
 135                 DEBUG_REQ(D_ADAPTTO, req,
 136                           "Early reply received: mlen=%u offset=%d replen=%d "
 137                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 138                           req->rq_replen, req->rq_replied, ev->unlinked);
 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_reply_off = ev->offset;
 147                 req->rq_nob_received = ev->mlength;
 148                 /* And we're still receiving */
 149                 req->rq_receiving_reply = 1;
 150         } else {
 151                 /* Real reply */
 152                 req->rq_rep_swab_mask = 0;
 153                 req->rq_replied = 1;
 154                 req->rq_reply_off = ev->offset;
 155                 req->rq_nob_received = ev->mlength;
 156                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 157                    so we must unlink in ptlrpc_unregister_reply */
 158                 DEBUG_REQ(D_INFO, req,
 159                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 160                           lustre_msg_get_flags(req->rq_reqmsg),
 161                           ev->mlength, ev->offset, req->rq_replen);
 162         }
 163
 164         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 165
 166 out_wake:
 167         /* NB don't unlock till after wakeup; req can disappear under us
 168          * since we don't have our own ref */
 169         ptlrpc_client_wake_req(req);
 170         cfs_spin_unlock(&req->rq_lock);
 171         EXIT;
 172 }
 173
 174 /*
 175  * Client's bulk has been written/read
 176  */
 177 void client_bulk_callback (lnet_event_t *ev)
 178 {
 179         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 180         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 181         ENTRY;
 182
 183         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 184                   ev->type == LNET_EVENT_PUT) ||
 185                  (desc->bd_type == BULK_GET_SOURCE &&
 186                   ev->type == LNET_EVENT_GET) ||
 187                  ev->type == LNET_EVENT_UNLINK);
 188         LASSERT (ev->unlinked);
 189
 190         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 191                 ev->status = -EIO;
 192
 193         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 194                 ev->status = -EIO;
 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         cfs_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         /* release the encrypted pages for write */
 212         if (desc->bd_req->rq_bulk_write)
 213                 sptlrpc_enc_pool_put_pages(desc);
 214
 215         /* NB don't unlock till after wakeup; desc can disappear under us
 216          * otherwise */
 217         ptlrpc_client_wake_req(desc->bd_req);
 218
 219         cfs_spin_unlock(&desc->bd_lock);
 220         EXIT;
 221 }
 222
 223 /*
 224  * We will have percpt request history list for ptlrpc service in upcoming
 225  * patches because we don't want to be serialized by current per-service
 226  * history operations. So we require history ID can (somehow) show arriving
 227  * order w/o grabbing global lock, and user can sort them in userspace.
 228  *
 229  * This is how we generate history ID for ptlrpc_request:
 230  * ----------------------------------------------------
 231  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 232  * ----------------------------------------------------
 233  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 234  * ----------------------------------------------------
 235  *
 236  * it might not be precise but should be good enough.
 237  */
 238
 239 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 240
 241 #define REQS_SEC_SHIFT          32
 242 #define REQS_USEC_SHIFT         16
 243 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 244
 245 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 246                                    struct ptlrpc_request *req)
 247 {
 248         __u64   sec = req->rq_arrival_time.tv_sec;
 249         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
 250         __u64   new_seq;
 251
 252         /* set sequence ID for request and add it to history list,
 253          * it must be called with hold svcpt::scp_lock */
 254
 255         new_seq = (sec << REQS_SEC_SHIFT) |
 256                   (usec << REQS_USEC_SHIFT) | svcpt->scp_cpt;
 257         if (new_seq > svcpt->scp_hist_seq) {
 258                 /* This handles the initial case of scp_hist_seq == 0 or
 259                  * we just jumped into a new time window */
 260                 svcpt->scp_hist_seq = new_seq;
 261         } else {
 262                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 263                 /* NB: increase sequence number in current usec bucket,
 264                  * however, it's possible that we used up all bits for
 265                  * sequence and jumped into the next usec bucket (future time),
 266                  * then we hope there will be less RPCs per bucket at some
 267                  * point, and sequence will catch up again */
 268                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
 269                 new_seq = svcpt->scp_hist_seq;
 270         }
 271
 272         req->rq_history_seq = new_seq;
 273
 274         cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 275 }
 276
 277 /*
 278  * Server's incoming request callback
 279  */
 280 void request_in_callback(lnet_event_t *ev)
 281 {
 282         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 283         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 284         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 285         struct ptlrpc_service             *service = svcpt->scp_service;
 286         struct ptlrpc_request             *req;
 287         ENTRY;
 288
 289         LASSERT (ev->type == LNET_EVENT_PUT ||
 290                  ev->type == LNET_EVENT_UNLINK);
 291         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 292         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 293                  rqbd->rqbd_buffer + service->srv_buf_size);
 294
 295         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 296                "event type %d, status %d, service %s\n",
 297                ev->type, ev->status, service->srv_name);
 298
 299         if (ev->unlinked) {
 300                 /* If this is the last request message to fit in the
 301                  * request buffer we can use the request object embedded in
 302                  * rqbd.  Note that if we failed to allocate a request,
 303                  * we'd have to re-post the rqbd, which we can't do in this
 304                  * context. */
 305                 req = &rqbd->rqbd_req;
 306                 memset(req, 0, sizeof (*req));
 307         } else {
 308                 LASSERT (ev->type == LNET_EVENT_PUT);
 309                 if (ev->status != 0) {
 310                         /* We moaned above already... */
 311                         return;
 312                 }
 313                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 314                 if (req == NULL) {
 315                         CERROR("Can't allocate incoming request descriptor: "
 316                                "Dropping %s RPC from %s\n",
 317                                service->srv_name,
 318                                libcfs_id2str(ev->initiator));
 319                         return;
 320                 }
 321         }
 322
 323         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 324          * flags are reset and scalars are zero.  We only set the message
 325          * size to non-zero if this was a successful receive. */
 326         req->rq_xid = ev->match_bits;
 327         req->rq_reqbuf = ev->md.start + ev->offset;
 328         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 329                 req->rq_reqdata_len = ev->mlength;
 330         cfs_gettimeofday(&req->rq_arrival_time);
 331         req->rq_peer = ev->initiator;
 332         req->rq_self = ev->target.nid;
 333         req->rq_rqbd = rqbd;
 334         req->rq_phase = RQ_PHASE_NEW;
 335         cfs_spin_lock_init(&req->rq_lock);
 336         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 337         cfs_atomic_set(&req->rq_refcount, 1);
 338         if (ev->type == LNET_EVENT_PUT)
 339                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
 340                        req, req->rq_xid, ev->mlength);
 341
 342         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 343
 344         cfs_spin_lock(&svcpt->scp_lock);
 345
 346         ptlrpc_req_add_history(svcpt, req);
 347
 348         if (ev->unlinked) {
 349                 svcpt->scp_nrqbds_posted--;
 350                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 351                        svcpt->scp_nrqbds_posted);
 352
 353                 /* Normally, don't complain about 0 buffers posted; LNET won't
 354                  * drop incoming reqs since we set the portal lazy */
 355                 if (test_req_buffer_pressure &&
 356                     ev->type != LNET_EVENT_UNLINK &&
 357                     svcpt->scp_nrqbds_posted == 0)
 358                         CWARN("All %s request buffers busy\n",
 359                               service->srv_name);
 360
 361                 /* req takes over the network's ref on rqbd */
 362         } else {
 363                 /* req takes a ref on rqbd */
 364                 rqbd->rqbd_refcount++;
 365         }
 366
 367         cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 368         svcpt->scp_nreqs_incoming++;
 369
 370         /* NB everything can disappear under us once the request
 371          * has been queued and we unlock, so do the wake now... */
 372         cfs_waitq_signal(&svcpt->scp_waitq);
 373
 374         cfs_spin_unlock(&svcpt->scp_lock);
 375         EXIT;
 376 }
 377
 378 /*
 379  *  Server's outgoing reply callback
 380  */
 381 void reply_out_callback(lnet_event_t *ev)
 382 {
 383         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 384         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 385         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 386         ENTRY;
 387
 388         LASSERT (ev->type == LNET_EVENT_SEND ||
 389                  ev->type == LNET_EVENT_ACK ||
 390                  ev->type == LNET_EVENT_UNLINK);
 391
 392         if (!rs->rs_difficult) {
 393                 /* 'Easy' replies have no further processing so I drop the
 394                  * net's ref on 'rs' */
 395                 LASSERT (ev->unlinked);
 396                 ptlrpc_rs_decref(rs);
 397                 EXIT;
 398                 return;
 399         }
 400
 401         LASSERT (rs->rs_on_net);
 402
 403         if (ev->unlinked) {
 404                 /* Last network callback. The net's ref on 'rs' stays put
 405                  * until ptlrpc_handle_rs() is done with it */
 406                 cfs_spin_lock(&svcpt->scp_rep_lock);
 407                 cfs_spin_lock(&rs->rs_lock);
 408
 409                 rs->rs_on_net = 0;
 410                 if (!rs->rs_no_ack ||
 411                     rs->rs_transno <=
 412                     rs->rs_export->exp_obd->obd_last_committed)
 413                         ptlrpc_schedule_difficult_reply(rs);
 414
 415                 cfs_spin_unlock(&rs->rs_lock);
 416                 cfs_spin_unlock(&svcpt->scp_rep_lock);
 417         }
 418         EXIT;
 419 }
 420
 421 #ifdef HAVE_SERVER_SUPPORT
 422 /*
 423  * Server's bulk completion callback
 424  */
 425 void server_bulk_callback (lnet_event_t *ev)
 426 {
 427         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 428         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 429         ENTRY;
 430
 431         LASSERT (ev->type == LNET_EVENT_SEND ||
 432                  ev->type == LNET_EVENT_UNLINK ||
 433                  (desc->bd_type == BULK_PUT_SOURCE &&
 434                   ev->type == LNET_EVENT_ACK) ||
 435                  (desc->bd_type == BULK_GET_SINK &&
 436                   ev->type == LNET_EVENT_REPLY));
 437
 438         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 439                "event type %d, status %d, desc %p\n",
 440                ev->type, ev->status, desc);
 441
 442         cfs_spin_lock(&desc->bd_lock);
 443
 444         if ((ev->type == LNET_EVENT_ACK ||
 445              ev->type == LNET_EVENT_REPLY) &&
 446             ev->status == 0) {
 447                 /* We heard back from the peer, so even if we get this
 448                  * before the SENT event (oh yes we can), we know we
 449                  * read/wrote the peer buffer and how much... */
 450                 desc->bd_success = 1;
 451                 desc->bd_nob_transferred = ev->mlength;
 452                 desc->bd_sender = ev->sender;
 453         }
 454
 455         if (ev->unlinked) {
 456                 /* This is the last callback no matter what... */
 457                 desc->bd_network_rw = 0;
 458                 cfs_waitq_signal(&desc->bd_waitq);
 459         }
 460
 461         cfs_spin_unlock(&desc->bd_lock);
 462         EXIT;
 463 }
 464 #endif
 465
 466 static void ptlrpc_master_callback(lnet_event_t *ev)
 467 {
 468         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 469         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 470
 471         /* Honestly, it's best to find out early. */
 472         LASSERT (cbid->cbid_arg != LP_POISON);
 473         LASSERT (callback == request_out_callback ||
 474                  callback == reply_in_callback ||
 475                  callback == client_bulk_callback ||
 476                  callback == request_in_callback ||
 477                  callback == reply_out_callback
 478 #ifdef HAVE_SERVER_SUPPORT
 479                  || callback == server_bulk_callback
 480 #endif
 481                  );
 482
 483         callback (ev);
 484 }
 485
 486 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 487                          lnet_process_id_t *peer, lnet_nid_t *self)
 488 {
 489         int               best_dist = 0;
 490         __u32             best_order = 0;
 491         int               count = 0;
 492         int               rc = -ENOENT;
 493         int               portals_compatibility;
 494         int               dist;
 495         __u32             order;
 496         lnet_nid_t        dst_nid;
 497         lnet_nid_t        src_nid;
 498
 499         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 500
 501         peer->pid = LUSTRE_SRV_LNET_PID;
 502
 503         /* Choose the matching UUID that's closest */
 504         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 505                 dist = LNetDist(dst_nid, &src_nid, &order);
 506                 if (dist < 0)
 507                         continue;
 508
 509                 if (dist == 0) {                /* local! use loopback LND */
 510                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 511                         rc = 0;
 512                         break;
 513                 }
 514
 515                 if (rc < 0 ||
 516                     dist < best_dist ||
 517                     (dist == best_dist && order < best_order)) {
 518                         best_dist = dist;
 519                         best_order = order;
 520
 521                         if (portals_compatibility > 1) {
 522                                 /* Strong portals compatibility: Zero the nid's
 523                                  * NET, so if I'm reading new config logs, or
 524                                  * getting configured by (new) lconf I can
 525                                  * still talk to old servers. */
 526                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 527                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 528                         }
 529                         peer->nid = dst_nid;
 530                         *self = src_nid;
 531                         rc = 0;
 532                 }
 533         }
 534
 535         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 536         return rc;
 537 }
 538
 539 void ptlrpc_ni_fini(void)
 540 {
 541         cfs_waitq_t         waitq;
 542         struct l_wait_info  lwi;
 543         int                 rc;
 544         int                 retries;
 545
 546         /* Wait for the event queue to become idle since there may still be
 547          * messages in flight with pending events (i.e. the fire-and-forget
 548          * messages == client requests and "non-difficult" server
 549          * replies */
 550
 551         for (retries = 0;; retries++) {
 552                 rc = LNetEQFree(ptlrpc_eq_h);
 553                 switch (rc) {
 554                 default:
 555                         LBUG();
 556
 557                 case 0:
 558                         LNetNIFini();
 559                         return;
 560
 561                 case -EBUSY:
 562                         if (retries != 0)
 563                                 CWARN("Event queue still busy\n");
 564
 565                         /* Wait for a bit */
 566                         cfs_waitq_init(&waitq);
 567                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 568                         l_wait_event(waitq, 0, &lwi);
 569                         break;
 570                 }
 571         }
 572         /* notreached */
 573 }
 574
 575 lnet_pid_t ptl_get_pid(void)
 576 {
 577         lnet_pid_t        pid;
 578
 579 #ifndef  __KERNEL__
 580         pid = getpid();
 581 #else
 582         pid = LUSTRE_SRV_LNET_PID;
 583 #endif
 584         return pid;
 585 }
 586
 587 int ptlrpc_ni_init(void)
 588 {
 589         int              rc;
 590         lnet_pid_t       pid;
 591
 592         pid = ptl_get_pid();
 593         CDEBUG(D_NET, "My pid is: %x\n", pid);
 594
 595         /* We're not passing any limits yet... */
 596         rc = LNetNIInit(pid);
 597         if (rc < 0) {
 598                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 599                 return (-ENOENT);
 600         }
 601
 602         /* CAVEAT EMPTOR: how we process portals events is _radically_
 603          * different depending on... */
 604 #ifdef __KERNEL__
 605         /* kernel LNet calls our master callback when there are new event,
 606          * because we are guaranteed to get every event via callback,
 607          * so we just set EQ size to 0 to avoid overhread of serializing
 608          * enqueue/dequeue operations in LNet. */
 609         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 610 #else
 611         /* liblustre calls the master callback when it removes events from the
 612          * event queue.  The event queue has to be big enough not to drop
 613          * anything */
 614         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 615 #endif
 616         if (rc == 0)
 617                 return 0;
 618
 619         CERROR ("Failed to allocate event queue: %d\n", rc);
 620         LNetNIFini();
 621
 622         return (-ENOMEM);
 623 }
 624
 625 #ifndef __KERNEL__
 626 CFS_LIST_HEAD(liblustre_wait_callbacks);
 627 CFS_LIST_HEAD(liblustre_idle_callbacks);
 628 void *liblustre_services_callback;
 629
 630 void *
 631 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 632                                       const char *name,
 633                                       int (*fn)(void *arg), void *arg)
 634 {
 635         struct liblustre_wait_callback *llwc;
 636
 637         OBD_ALLOC(llwc, sizeof(*llwc));
 638         LASSERT (llwc != NULL);
 639
 640         llwc->llwc_name = name;
 641         llwc->llwc_fn = fn;
 642         llwc->llwc_arg = arg;
 643         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 644
 645         return (llwc);
 646 }
 647
 648 void
 649 liblustre_deregister_waitidle_callback (void *opaque)
 650 {
 651         struct liblustre_wait_callback *llwc = opaque;
 652
 653         cfs_list_del(&llwc->llwc_list);
 654         OBD_FREE(llwc, sizeof(*llwc));
 655 }
 656
 657 void *
 658 liblustre_register_wait_callback (const char *name,
 659                                   int (*fn)(void *arg), void *arg)
 660 {
 661         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 662                                                     name, fn, arg);
 663 }
 664
 665 void
 666 liblustre_deregister_wait_callback (void *opaque)
 667 {
 668         liblustre_deregister_waitidle_callback(opaque);
 669 }
 670
 671 void *
 672 liblustre_register_idle_callback (const char *name,
 673                                   int (*fn)(void *arg), void *arg)
 674 {
 675         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 676                                                     name, fn, arg);
 677 }
 678
 679 void
 680 liblustre_deregister_idle_callback (void *opaque)
 681 {
 682         liblustre_deregister_waitidle_callback(opaque);
 683 }
 684
 685 int
 686 liblustre_check_events (int timeout)
 687 {
 688         lnet_event_t ev;
 689         int         rc;
 690         int         i;
 691         ENTRY;
 692
 693         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 694         if (rc == 0)
 695                 RETURN(0);
 696
 697         LASSERT (rc == -EOVERFLOW || rc == 1);
 698
 699         /* liblustre: no asynch callback so we can't afford to miss any
 700          * events... */
 701         if (rc == -EOVERFLOW) {
 702                 CERROR ("Dropped an event!!!\n");
 703                 abort();
 704         }
 705
 706         ptlrpc_master_callback (&ev);
 707         RETURN(1);
 708 }
 709
 710 int liblustre_waiting = 0;
 711
 712 int
 713 liblustre_wait_event (int timeout)
 714 {
 715         cfs_list_t                     *tmp;
 716         struct liblustre_wait_callback *llwc;
 717         int                             found_something = 0;
 718
 719         /* single threaded recursion check... */
 720         liblustre_waiting = 1;
 721
 722         for (;;) {
 723                 /* Deal with all pending events */
 724                 while (liblustre_check_events(0))
 725                         found_something = 1;
 726
 727                 /* Give all registered callbacks a bite at the cherry */
 728                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 729                         llwc = cfs_list_entry(tmp,
 730                                               struct liblustre_wait_callback,
 731                                               llwc_list);
 732
 733                         if (llwc->llwc_fn(llwc->llwc_arg))
 734                                 found_something = 1;
 735                 }
 736
 737                 if (found_something || timeout == 0)
 738                         break;
 739
 740                 /* Nothing so far, but I'm allowed to block... */
 741                 found_something = liblustre_check_events(timeout);
 742                 if (!found_something)           /* still nothing */
 743                         break;                  /* I timed out */
 744         }
 745
 746         liblustre_waiting = 0;
 747
 748         return found_something;
 749 }
 750
 751 void
 752 liblustre_wait_idle(void)
 753 {
 754         static int recursed = 0;
 755
 756         cfs_list_t                     *tmp;
 757         struct liblustre_wait_callback *llwc;
 758         int                             idle = 0;
 759
 760         LASSERT(!recursed);
 761         recursed = 1;
 762
 763         do {
 764                 liblustre_wait_event(0);
 765
 766                 idle = 1;
 767
 768                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 769                         llwc = cfs_list_entry(tmp,
 770                                               struct liblustre_wait_callback,
 771                                               llwc_list);
 772
 773                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 774                                 idle = 0;
 775                                 break;
 776                         }
 777                 }
 778
 779         } while (!idle);
 780
 781         recursed = 0;
 782 }
 783
 784 #endif /* __KERNEL__ */
 785
 786 int ptlrpc_init_portals(void)
 787 {
 788         int   rc = ptlrpc_ni_init();
 789
 790         if (rc != 0) {
 791                 CERROR("network initialisation failed\n");
 792                 return -EIO;
 793         }
 794 #ifndef __KERNEL__
 795         liblustre_services_callback =
 796                 liblustre_register_wait_callback("liblustre_check_services",
 797                                                  &liblustre_check_services,
 798                                                  NULL);
 799         cfs_init_completion_module(liblustre_wait_event);
 800 #endif
 801         rc = ptlrpcd_addref();
 802         if (rc == 0)
 803                 return 0;
 804
 805         CERROR("rpcd initialisation failed\n");
 806 #ifndef __KERNEL__
 807         liblustre_deregister_wait_callback(liblustre_services_callback);
 808 #endif
 809         ptlrpc_ni_fini();
 810         return rc;
 811 }
 812
 813 void ptlrpc_exit_portals(void)
 814 {
 815 #ifndef __KERNEL__
 816         liblustre_deregister_wait_callback(liblustre_services_callback);
 817 #endif
 818         ptlrpcd_decref();
 819         ptlrpc_ni_fini();
 820 }