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  * Copyright (c) 2012, 2013, Intel Corporation.
  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                 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(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         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                 /* Got reply, no resend required */
 157                 req->rq_resend = 0;
 158                 req->rq_reply_off = ev->offset;
 159                 req->rq_nob_received = ev->mlength;
 160                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 161                    so we must unlink in ptlrpc_unregister_reply */
 162                 DEBUG_REQ(D_INFO, req,
 163                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 164                           lustre_msg_get_flags(req->rq_reqmsg),
 165                           ev->mlength, ev->offset, req->rq_replen);
 166         }
 167
 168         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 169
 170 out_wake:
 171         /* NB don't unlock till after wakeup; req can disappear under us
 172          * since we don't have our own ref */
 173         ptlrpc_client_wake_req(req);
 174         spin_unlock(&req->rq_lock);
 175         EXIT;
 176 }
 177
 178 /*
 179  * Client's bulk has been written/read
 180  */
 181 void client_bulk_callback (lnet_event_t *ev)
 182 {
 183         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 184         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 185         struct ptlrpc_request   *req;
 186         ENTRY;
 187
 188         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 189                   ev->type == LNET_EVENT_PUT) ||
 190                  (desc->bd_type == BULK_GET_SOURCE &&
 191                   ev->type == LNET_EVENT_GET) ||
 192                  ev->type == LNET_EVENT_UNLINK);
 193         LASSERT (ev->unlinked);
 194
 195         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 196                 ev->status = -EIO;
 197
 198         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 199                 ev->status = -EIO;
 200
 201         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 202                "event type %d, status %d, desc %p\n",
 203                ev->type, ev->status, desc);
 204
 205         spin_lock(&desc->bd_lock);
 206         req = desc->bd_req;
 207         LASSERT(desc->bd_md_count > 0);
 208         desc->bd_md_count--;
 209
 210         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 211                 desc->bd_nob_transferred += ev->mlength;
 212                 desc->bd_sender = ev->sender;
 213         } else {
 214                 /* start reconnect and resend if network error hit */
 215                 spin_lock(&req->rq_lock);
 216                 req->rq_net_err = 1;
 217                 spin_unlock(&req->rq_lock);
 218         }
 219
 220         if (ev->status != 0)
 221                 desc->bd_failure = 1;
 222
 223         /* NB don't unlock till after wakeup; desc can disappear under us
 224          * otherwise */
 225         if (desc->bd_md_count == 0)
 226                 ptlrpc_client_wake_req(desc->bd_req);
 227
 228         spin_unlock(&desc->bd_lock);
 229         EXIT;
 230 }
 231
 232 /*
 233  * We will have percpt request history list for ptlrpc service in upcoming
 234  * patches because we don't want to be serialized by current per-service
 235  * history operations. So we require history ID can (somehow) show arriving
 236  * order w/o grabbing global lock, and user can sort them in userspace.
 237  *
 238  * This is how we generate history ID for ptlrpc_request:
 239  * ----------------------------------------------------
 240  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 241  * ----------------------------------------------------
 242  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 243  * ----------------------------------------------------
 244  *
 245  * it might not be precise but should be good enough.
 246  */
 247
 248 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 249
 250 #define REQS_SEC_SHIFT          32
 251 #define REQS_USEC_SHIFT         16
 252 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 253
 254 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 255                                    struct ptlrpc_request *req)
 256 {
 257         __u64   sec = req->rq_arrival_time.tv_sec;
 258         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
 259         __u64   new_seq;
 260
 261         /* set sequence ID for request and add it to history list,
 262          * it must be called with hold svcpt::scp_lock */
 263
 264         new_seq = (sec << REQS_SEC_SHIFT) |
 265                   (usec << REQS_USEC_SHIFT) |
 266                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
 267
 268         if (new_seq > svcpt->scp_hist_seq) {
 269                 /* This handles the initial case of scp_hist_seq == 0 or
 270                  * we just jumped into a new time window */
 271                 svcpt->scp_hist_seq = new_seq;
 272         } else {
 273                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 274                 /* NB: increase sequence number in current usec bucket,
 275                  * however, it's possible that we used up all bits for
 276                  * sequence and jumped into the next usec bucket (future time),
 277                  * then we hope there will be less RPCs per bucket at some
 278                  * point, and sequence will catch up again */
 279                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
 280                 new_seq = svcpt->scp_hist_seq;
 281         }
 282
 283         req->rq_history_seq = new_seq;
 284
 285         cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 286 }
 287
 288 /*
 289  * Server's incoming request callback
 290  */
 291 void request_in_callback(lnet_event_t *ev)
 292 {
 293         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 294         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 295         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 296         struct ptlrpc_service             *service = svcpt->scp_service;
 297         struct ptlrpc_request             *req;
 298         ENTRY;
 299
 300         LASSERT (ev->type == LNET_EVENT_PUT ||
 301                  ev->type == LNET_EVENT_UNLINK);
 302         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 303         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 304                  rqbd->rqbd_buffer + service->srv_buf_size);
 305
 306         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 307                "event type %d, status %d, service %s\n",
 308                ev->type, ev->status, service->srv_name);
 309
 310         if (ev->unlinked) {
 311                 /* If this is the last request message to fit in the
 312                  * request buffer we can use the request object embedded in
 313                  * rqbd.  Note that if we failed to allocate a request,
 314                  * we'd have to re-post the rqbd, which we can't do in this
 315                  * context. */
 316                 req = &rqbd->rqbd_req;
 317                 memset(req, 0, sizeof (*req));
 318         } else {
 319                 LASSERT (ev->type == LNET_EVENT_PUT);
 320                 if (ev->status != 0) {
 321                         /* We moaned above already... */
 322                         return;
 323                 }
 324                 req = ptlrpc_request_cache_alloc(ALLOC_ATOMIC_TRY);
 325                 if (req == NULL) {
 326                         CERROR("Can't allocate incoming request descriptor: "
 327                                "Dropping %s RPC from %s\n",
 328                                service->srv_name,
 329                                libcfs_id2str(ev->initiator));
 330                         return;
 331                 }
 332         }
 333
 334         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 335          * flags are reset and scalars are zero.  We only set the message
 336          * size to non-zero if this was a successful receive. */
 337         req->rq_xid = ev->match_bits;
 338         req->rq_reqbuf = ev->md.start + ev->offset;
 339         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 340                 req->rq_reqdata_len = ev->mlength;
 341         do_gettimeofday(&req->rq_arrival_time);
 342         req->rq_peer = ev->initiator;
 343         req->rq_self = ev->target.nid;
 344         req->rq_rqbd = rqbd;
 345         req->rq_phase = RQ_PHASE_NEW;
 346         spin_lock_init(&req->rq_lock);
 347         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 348         CFS_INIT_LIST_HEAD(&req->rq_exp_list);
 349         atomic_set(&req->rq_refcount, 1);
 350         if (ev->type == LNET_EVENT_PUT)
 351                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
 352                        req, req->rq_xid, ev->mlength);
 353
 354         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 355
 356         spin_lock(&svcpt->scp_lock);
 357
 358         ptlrpc_req_add_history(svcpt, req);
 359
 360         if (ev->unlinked) {
 361                 svcpt->scp_nrqbds_posted--;
 362                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 363                        svcpt->scp_nrqbds_posted);
 364
 365                 /* Normally, don't complain about 0 buffers posted; LNET won't
 366                  * drop incoming reqs since we set the portal lazy */
 367                 if (test_req_buffer_pressure &&
 368                     ev->type != LNET_EVENT_UNLINK &&
 369                     svcpt->scp_nrqbds_posted == 0)
 370                         CWARN("All %s request buffers busy\n",
 371                               service->srv_name);
 372
 373                 /* req takes over the network's ref on rqbd */
 374         } else {
 375                 /* req takes a ref on rqbd */
 376                 rqbd->rqbd_refcount++;
 377         }
 378
 379         cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 380         svcpt->scp_nreqs_incoming++;
 381
 382         /* NB everything can disappear under us once the request
 383          * has been queued and we unlock, so do the wake now... */
 384         wake_up(&svcpt->scp_waitq);
 385
 386         spin_unlock(&svcpt->scp_lock);
 387         EXIT;
 388 }
 389
 390 /*
 391  *  Server's outgoing reply callback
 392  */
 393 void reply_out_callback(lnet_event_t *ev)
 394 {
 395         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 396         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 397         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 398         ENTRY;
 399
 400         LASSERT (ev->type == LNET_EVENT_SEND ||
 401                  ev->type == LNET_EVENT_ACK ||
 402                  ev->type == LNET_EVENT_UNLINK);
 403
 404         if (!rs->rs_difficult) {
 405                 /* 'Easy' replies have no further processing so I drop the
 406                  * net's ref on 'rs' */
 407                 LASSERT (ev->unlinked);
 408                 ptlrpc_rs_decref(rs);
 409                 EXIT;
 410                 return;
 411         }
 412
 413         LASSERT (rs->rs_on_net);
 414
 415         if (ev->unlinked) {
 416                 /* Last network callback. The net's ref on 'rs' stays put
 417                  * until ptlrpc_handle_rs() is done with it */
 418                 spin_lock(&svcpt->scp_rep_lock);
 419                 spin_lock(&rs->rs_lock);
 420
 421                 rs->rs_on_net = 0;
 422                 if (!rs->rs_no_ack ||
 423                     rs->rs_transno <=
 424                     rs->rs_export->exp_obd->obd_last_committed)
 425                         ptlrpc_schedule_difficult_reply(rs);
 426
 427                 spin_unlock(&rs->rs_lock);
 428                 spin_unlock(&svcpt->scp_rep_lock);
 429         }
 430         EXIT;
 431 }
 432
 433 #ifdef HAVE_SERVER_SUPPORT
 434 /*
 435  * Server's bulk completion callback
 436  */
 437 void server_bulk_callback (lnet_event_t *ev)
 438 {
 439         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 440         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 441         ENTRY;
 442
 443         LASSERT(ev->type == LNET_EVENT_SEND ||
 444                 ev->type == LNET_EVENT_UNLINK ||
 445                 (desc->bd_type == BULK_PUT_SOURCE &&
 446                  ev->type == LNET_EVENT_ACK) ||
 447                 (desc->bd_type == BULK_GET_SINK &&
 448                  ev->type == LNET_EVENT_REPLY));
 449
 450         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 451                "event type %d, status %d, desc %p\n",
 452                ev->type, ev->status, desc);
 453
 454         spin_lock(&desc->bd_lock);
 455
 456         LASSERT(desc->bd_md_count > 0);
 457
 458         if ((ev->type == LNET_EVENT_ACK ||
 459              ev->type == LNET_EVENT_REPLY) &&
 460             ev->status == 0) {
 461                 /* We heard back from the peer, so even if we get this
 462                  * before the SENT event (oh yes we can), we know we
 463                  * read/wrote the peer buffer and how much... */
 464                 desc->bd_nob_transferred += ev->mlength;
 465                 desc->bd_sender = ev->sender;
 466         }
 467
 468         if (ev->status != 0)
 469                 desc->bd_failure = 1;
 470
 471         if (ev->unlinked) {
 472                 desc->bd_md_count--;
 473                 /* This is the last callback no matter what... */
 474                 if (desc->bd_md_count == 0)
 475                         wake_up(&desc->bd_waitq);
 476         }
 477
 478         spin_unlock(&desc->bd_lock);
 479         EXIT;
 480 }
 481 #endif
 482
 483 static void ptlrpc_master_callback(lnet_event_t *ev)
 484 {
 485         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 486         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 487
 488         /* Honestly, it's best to find out early. */
 489         LASSERT (cbid->cbid_arg != LP_POISON);
 490         LASSERT (callback == request_out_callback ||
 491                  callback == reply_in_callback ||
 492                  callback == client_bulk_callback ||
 493                  callback == request_in_callback ||
 494                  callback == reply_out_callback
 495 #ifdef HAVE_SERVER_SUPPORT
 496                  || callback == server_bulk_callback
 497 #endif
 498                  );
 499
 500         callback (ev);
 501 }
 502
 503 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 504                          lnet_process_id_t *peer, lnet_nid_t *self)
 505 {
 506         int               best_dist = 0;
 507         __u32             best_order = 0;
 508         int               count = 0;
 509         int               rc = -ENOENT;
 510         int               portals_compatibility;
 511         int               dist;
 512         __u32             order;
 513         lnet_nid_t        dst_nid;
 514         lnet_nid_t        src_nid;
 515
 516         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 517
 518         peer->pid = LUSTRE_SRV_LNET_PID;
 519
 520         /* Choose the matching UUID that's closest */
 521         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 522                 dist = LNetDist(dst_nid, &src_nid, &order);
 523                 if (dist < 0)
 524                         continue;
 525
 526                 if (dist == 0) {                /* local! use loopback LND */
 527                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 528                         rc = 0;
 529                         break;
 530                 }
 531
 532                 if (rc < 0 ||
 533                     dist < best_dist ||
 534                     (dist == best_dist && order < best_order)) {
 535                         best_dist = dist;
 536                         best_order = order;
 537
 538                         if (portals_compatibility > 1) {
 539                                 /* Strong portals compatibility: Zero the nid's
 540                                  * NET, so if I'm reading new config logs, or
 541                                  * getting configured by (new) lconf I can
 542                                  * still talk to old servers. */
 543                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 544                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 545                         }
 546                         peer->nid = dst_nid;
 547                         *self = src_nid;
 548                         rc = 0;
 549                 }
 550         }
 551
 552         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 553         return rc;
 554 }
 555
 556 void ptlrpc_ni_fini(void)
 557 {
 558         wait_queue_head_t         waitq;
 559         struct l_wait_info  lwi;
 560         int                 rc;
 561         int                 retries;
 562
 563         /* Wait for the event queue to become idle since there may still be
 564          * messages in flight with pending events (i.e. the fire-and-forget
 565          * messages == client requests and "non-difficult" server
 566          * replies */
 567
 568         for (retries = 0;; retries++) {
 569                 rc = LNetEQFree(ptlrpc_eq_h);
 570                 switch (rc) {
 571                 default:
 572                         LBUG();
 573
 574                 case 0:
 575                         LNetNIFini();
 576                         return;
 577
 578                 case -EBUSY:
 579                         if (retries != 0)
 580                                 CWARN("Event queue still busy\n");
 581
 582                         /* Wait for a bit */
 583                         init_waitqueue_head(&waitq);
 584                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 585                         l_wait_event(waitq, 0, &lwi);
 586                         break;
 587                 }
 588         }
 589         /* notreached */
 590 }
 591
 592 lnet_pid_t ptl_get_pid(void)
 593 {
 594         lnet_pid_t        pid;
 595
 596 #ifndef  __KERNEL__
 597         pid = getpid();
 598 #else
 599         pid = LUSTRE_SRV_LNET_PID;
 600 #endif
 601         return pid;
 602 }
 603
 604 int ptlrpc_ni_init(void)
 605 {
 606         int              rc;
 607         lnet_pid_t       pid;
 608
 609         pid = ptl_get_pid();
 610         CDEBUG(D_NET, "My pid is: %x\n", pid);
 611
 612         /* We're not passing any limits yet... */
 613         rc = LNetNIInit(pid);
 614         if (rc < 0) {
 615                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 616                 return (-ENOENT);
 617         }
 618
 619         /* CAVEAT EMPTOR: how we process portals events is _radically_
 620          * different depending on... */
 621 #ifdef __KERNEL__
 622         /* kernel LNet calls our master callback when there are new event,
 623          * because we are guaranteed to get every event via callback,
 624          * so we just set EQ size to 0 to avoid overhread of serializing
 625          * enqueue/dequeue operations in LNet. */
 626         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 627 #else
 628         /* liblustre calls the master callback when it removes events from the
 629          * event queue.  The event queue has to be big enough not to drop
 630          * anything */
 631         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 632 #endif
 633         if (rc == 0)
 634                 return 0;
 635
 636         CERROR ("Failed to allocate event queue: %d\n", rc);
 637         LNetNIFini();
 638
 639         return (-ENOMEM);
 640 }
 641
 642 #ifndef __KERNEL__
 643 CFS_LIST_HEAD(liblustre_wait_callbacks);
 644 CFS_LIST_HEAD(liblustre_idle_callbacks);
 645 void *liblustre_services_callback;
 646
 647 void *
 648 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 649                                       const char *name,
 650                                       int (*fn)(void *arg), void *arg)
 651 {
 652         struct liblustre_wait_callback *llwc;
 653
 654         OBD_ALLOC(llwc, sizeof(*llwc));
 655         LASSERT (llwc != NULL);
 656
 657         llwc->llwc_name = name;
 658         llwc->llwc_fn = fn;
 659         llwc->llwc_arg = arg;
 660         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 661
 662         return (llwc);
 663 }
 664
 665 void
 666 liblustre_deregister_waitidle_callback (void *opaque)
 667 {
 668         struct liblustre_wait_callback *llwc = opaque;
 669
 670         cfs_list_del(&llwc->llwc_list);
 671         OBD_FREE(llwc, sizeof(*llwc));
 672 }
 673
 674 void *
 675 liblustre_register_wait_callback (const char *name,
 676                                   int (*fn)(void *arg), void *arg)
 677 {
 678         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 679                                                     name, fn, arg);
 680 }
 681
 682 void
 683 liblustre_deregister_wait_callback (void *opaque)
 684 {
 685         liblustre_deregister_waitidle_callback(opaque);
 686 }
 687
 688 void *
 689 liblustre_register_idle_callback (const char *name,
 690                                   int (*fn)(void *arg), void *arg)
 691 {
 692         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 693                                                     name, fn, arg);
 694 }
 695
 696 void
 697 liblustre_deregister_idle_callback (void *opaque)
 698 {
 699         liblustre_deregister_waitidle_callback(opaque);
 700 }
 701
 702 int
 703 liblustre_check_events (int timeout)
 704 {
 705         lnet_event_t ev;
 706         int         rc;
 707         int         i;
 708         ENTRY;
 709
 710         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 711         if (rc == 0)
 712                 RETURN(0);
 713
 714         LASSERT (rc == -EOVERFLOW || rc == 1);
 715
 716         /* liblustre: no asynch callback so we can't afford to miss any
 717          * events... */
 718         if (rc == -EOVERFLOW) {
 719                 CERROR ("Dropped an event!!!\n");
 720                 abort();
 721         }
 722
 723         ptlrpc_master_callback (&ev);
 724         RETURN(1);
 725 }
 726
 727 int liblustre_waiting = 0;
 728
 729 int
 730 liblustre_wait_event (int timeout)
 731 {
 732         cfs_list_t                     *tmp;
 733         struct liblustre_wait_callback *llwc;
 734         int                             found_something = 0;
 735
 736         /* single threaded recursion check... */
 737         liblustre_waiting = 1;
 738
 739         for (;;) {
 740                 /* Deal with all pending events */
 741                 while (liblustre_check_events(0))
 742                         found_something = 1;
 743
 744                 /* Give all registered callbacks a bite at the cherry */
 745                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 746                         llwc = cfs_list_entry(tmp,
 747                                               struct liblustre_wait_callback,
 748                                               llwc_list);
 749
 750                         if (llwc->llwc_fn(llwc->llwc_arg))
 751                                 found_something = 1;
 752                 }
 753
 754                 if (found_something || timeout == 0)
 755                         break;
 756
 757                 /* Nothing so far, but I'm allowed to block... */
 758                 found_something = liblustre_check_events(timeout);
 759                 if (!found_something)           /* still nothing */
 760                         break;                  /* I timed out */
 761         }
 762
 763         liblustre_waiting = 0;
 764
 765         return found_something;
 766 }
 767
 768 void
 769 liblustre_wait_idle(void)
 770 {
 771         static int recursed = 0;
 772
 773         cfs_list_t                     *tmp;
 774         struct liblustre_wait_callback *llwc;
 775         int                             idle = 0;
 776
 777         LASSERT(!recursed);
 778         recursed = 1;
 779
 780         do {
 781                 liblustre_wait_event(0);
 782
 783                 idle = 1;
 784
 785                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 786                         llwc = cfs_list_entry(tmp,
 787                                               struct liblustre_wait_callback,
 788                                               llwc_list);
 789
 790                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 791                                 idle = 0;
 792                                 break;
 793                         }
 794                 }
 795
 796         } while (!idle);
 797
 798         recursed = 0;
 799 }
 800
 801 #endif /* __KERNEL__ */
 802
 803 int ptlrpc_init_portals(void)
 804 {
 805         int   rc = ptlrpc_ni_init();
 806
 807         if (rc != 0) {
 808                 CERROR("network initialisation failed\n");
 809                 return -EIO;
 810         }
 811 #ifndef __KERNEL__
 812         liblustre_services_callback =
 813                 liblustre_register_wait_callback("liblustre_check_services",
 814                                                  &liblustre_check_services,
 815                                                  NULL);
 816         init_completion_module(liblustre_wait_event);
 817 #endif
 818         rc = ptlrpcd_addref();
 819         if (rc == 0)
 820                 return 0;
 821
 822         CERROR("rpcd initialisation failed\n");
 823 #ifndef __KERNEL__
 824         liblustre_deregister_wait_callback(liblustre_services_callback);
 825 #endif
 826         ptlrpc_ni_fini();
 827         return rc;
 828 }
 829
 830 void ptlrpc_exit_portals(void)
 831 {
 832 #ifndef __KERNEL__
 833         liblustre_deregister_wait_callback(liblustre_services_callback);
 834 #endif
 835         ptlrpcd_decref();
 836         ptlrpc_ni_fini();
 837 }