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         spin_lock(&req->rq_lock);
  72         req->rq_real_sent = cfs_time_current_sec();
  73         if (ev->unlinked)
  74                 req->rq_req_unlink = 0;
  75
  76         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  77
  78                 /* Failed send: make it seem like the reply timed out, just
  79                  * like failing sends in client.c does currently...  */
  80
  81                 req->rq_net_err = 1;
  82                 ptlrpc_client_wake_req(req);
  83         }
  84         spin_unlock(&req->rq_lock);
  85
  86         ptlrpc_req_finished(req);
  87
  88         EXIT;
  89 }
  90
  91 /*
  92  * Client's incoming reply callback
  93  */
  94 void reply_in_callback(lnet_event_t *ev)
  95 {
  96         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  97         struct ptlrpc_request *req = cbid->cbid_arg;
  98         ENTRY;
  99
 100         DEBUG_REQ(D_NET, req, "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->offset + 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         if (ev->unlinked)
 114                 req->rq_reply_unlink = 0;
 115
 116         if (ev->status)
 117                 goto out_wake;
 118
 119         if (ev->type == LNET_EVENT_UNLINK) {
 120                 LASSERT(ev->unlinked);
 121                 DEBUG_REQ(D_NET, req, "unlink");
 122                 goto out_wake;
 123         }
 124
 125         if (ev->mlength < ev->rlength ) {
 126                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 127                        req->rq_replen, ev->rlength, ev->offset);
 128                 req->rq_reply_truncate = 1;
 129                 req->rq_replied = 1;
 130                 req->rq_status = -EOVERFLOW;
 131                 req->rq_nob_received = ev->rlength + ev->offset;
 132                 goto out_wake;
 133         }
 134
 135         if ((ev->offset == 0) &&
 136             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 137                 /* Early reply */
 138                 DEBUG_REQ(D_ADAPTTO, req,
 139                           "Early reply received: mlen=%u offset=%d replen=%d "
 140                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 141                           req->rq_replen, req->rq_replied, ev->unlinked);
 142
 143                 req->rq_early_count++; /* number received, client side */
 144
 145                 if (req->rq_replied)   /* already got the real reply */
 146                         goto out_wake;
 147
 148                 req->rq_early = 1;
 149                 req->rq_reply_off = ev->offset;
 150                 req->rq_nob_received = ev->mlength;
 151                 /* And we're still receiving */
 152                 req->rq_receiving_reply = 1;
 153         } else {
 154                 /* Real reply */
 155                 req->rq_rep_swab_mask = 0;
 156                 req->rq_replied = 1;
 157                 /* Got reply, no resend required */
 158                 req->rq_resend = 0;
 159                 req->rq_reply_off = ev->offset;
 160                 req->rq_nob_received = ev->mlength;
 161                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 162                    so we must unlink in ptlrpc_unregister_reply */
 163                 DEBUG_REQ(D_INFO, req,
 164                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 165                           lustre_msg_get_flags(req->rq_reqmsg),
 166                           ev->mlength, ev->offset, req->rq_replen);
 167         }
 168
 169         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 170
 171 out_wake:
 172         /* NB don't unlock till after wakeup; req can disappear under us
 173          * since we don't have our own ref */
 174         ptlrpc_client_wake_req(req);
 175         spin_unlock(&req->rq_lock);
 176         EXIT;
 177 }
 178
 179 /*
 180  * Client's bulk has been written/read
 181  */
 182 void client_bulk_callback (lnet_event_t *ev)
 183 {
 184         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 185         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 186         struct ptlrpc_request   *req;
 187         ENTRY;
 188
 189         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 190                   ev->type == LNET_EVENT_PUT) ||
 191                  (desc->bd_type == BULK_GET_SOURCE &&
 192                   ev->type == LNET_EVENT_GET) ||
 193                  ev->type == LNET_EVENT_UNLINK);
 194         LASSERT (ev->unlinked);
 195
 196         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 197                 ev->status = -EIO;
 198
 199         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 200                 ev->status = -EIO;
 201
 202         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 203                "event type %d, status %d, desc %p\n",
 204                ev->type, ev->status, desc);
 205
 206         spin_lock(&desc->bd_lock);
 207         req = desc->bd_req;
 208         LASSERT(desc->bd_md_count > 0);
 209         desc->bd_md_count--;
 210
 211         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 212                 desc->bd_nob_transferred += ev->mlength;
 213                 desc->bd_sender = ev->sender;
 214         } else {
 215                 /* start reconnect and resend if network error hit */
 216                 spin_lock(&req->rq_lock);
 217                 req->rq_net_err = 1;
 218                 spin_unlock(&req->rq_lock);
 219         }
 220
 221         if (ev->status != 0)
 222                 desc->bd_failure = 1;
 223
 224         /* NB don't unlock till after wakeup; desc can disappear under us
 225          * otherwise */
 226         if (desc->bd_md_count == 0)
 227                 ptlrpc_client_wake_req(desc->bd_req);
 228
 229         spin_unlock(&desc->bd_lock);
 230         EXIT;
 231 }
 232
 233 /*
 234  * We will have percpt request history list for ptlrpc service in upcoming
 235  * patches because we don't want to be serialized by current per-service
 236  * history operations. So we require history ID can (somehow) show arriving
 237  * order w/o grabbing global lock, and user can sort them in userspace.
 238  *
 239  * This is how we generate history ID for ptlrpc_request:
 240  * ----------------------------------------------------
 241  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 242  * ----------------------------------------------------
 243  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 244  * ----------------------------------------------------
 245  *
 246  * it might not be precise but should be good enough.
 247  */
 248
 249 #define REQS_CPT_BITS(svcpt)    ((svcpt)->scp_service->srv_cpt_bits)
 250
 251 #define REQS_SEC_SHIFT          32
 252 #define REQS_USEC_SHIFT         16
 253 #define REQS_SEQ_SHIFT(svcpt)   REQS_CPT_BITS(svcpt)
 254
 255 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 256                                    struct ptlrpc_request *req)
 257 {
 258         __u64   sec = req->rq_arrival_time.tv_sec;
 259         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
 260         __u64   new_seq;
 261
 262         /* set sequence ID for request and add it to history list,
 263          * it must be called with hold svcpt::scp_lock */
 264
 265         new_seq = (sec << REQS_SEC_SHIFT) |
 266                   (usec << REQS_USEC_SHIFT) |
 267                   (svcpt->scp_cpt < 0 ? 0 : svcpt->scp_cpt);
 268
 269         if (new_seq > svcpt->scp_hist_seq) {
 270                 /* This handles the initial case of scp_hist_seq == 0 or
 271                  * we just jumped into a new time window */
 272                 svcpt->scp_hist_seq = new_seq;
 273         } else {
 274                 LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
 275                 /* NB: increase sequence number in current usec bucket,
 276                  * however, it's possible that we used up all bits for
 277                  * sequence and jumped into the next usec bucket (future time),
 278                  * then we hope there will be less RPCs per bucket at some
 279                  * point, and sequence will catch up again */
 280                 svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
 281                 new_seq = svcpt->scp_hist_seq;
 282         }
 283
 284         req->rq_history_seq = new_seq;
 285
 286         cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 287 }
 288
 289 /*
 290  * Server's incoming request callback
 291  */
 292 void request_in_callback(lnet_event_t *ev)
 293 {
 294         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 295         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 296         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 297         struct ptlrpc_service             *service = svcpt->scp_service;
 298         struct ptlrpc_request             *req;
 299         ENTRY;
 300
 301         LASSERT (ev->type == LNET_EVENT_PUT ||
 302                  ev->type == LNET_EVENT_UNLINK);
 303         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 304         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 305                  rqbd->rqbd_buffer + service->srv_buf_size);
 306
 307         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 308                "event type %d, status %d, service %s\n",
 309                ev->type, ev->status, service->srv_name);
 310
 311         if (ev->unlinked) {
 312                 /* If this is the last request message to fit in the
 313                  * request buffer we can use the request object embedded in
 314                  * rqbd.  Note that if we failed to allocate a request,
 315                  * we'd have to re-post the rqbd, which we can't do in this
 316                  * context. */
 317                 req = &rqbd->rqbd_req;
 318                 memset(req, 0, sizeof (*req));
 319         } else {
 320                 LASSERT (ev->type == LNET_EVENT_PUT);
 321                 if (ev->status != 0) {
 322                         /* We moaned above already... */
 323                         return;
 324                 }
 325                 req = ptlrpc_request_cache_alloc(ALLOC_ATOMIC_TRY);
 326                 if (req == NULL) {
 327                         CERROR("Can't allocate incoming request descriptor: "
 328                                "Dropping %s RPC from %s\n",
 329                                service->srv_name,
 330                                libcfs_id2str(ev->initiator));
 331                         return;
 332                 }
 333         }
 334
 335         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 336          * flags are reset and scalars are zero.  We only set the message
 337          * size to non-zero if this was a successful receive. */
 338         req->rq_xid = ev->match_bits;
 339         req->rq_reqbuf = ev->md.start + ev->offset;
 340         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 341                 req->rq_reqdata_len = ev->mlength;
 342         do_gettimeofday(&req->rq_arrival_time);
 343         req->rq_peer = ev->initiator;
 344         req->rq_self = ev->target.nid;
 345         req->rq_rqbd = rqbd;
 346         req->rq_phase = RQ_PHASE_NEW;
 347         spin_lock_init(&req->rq_lock);
 348         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 349         CFS_INIT_LIST_HEAD(&req->rq_exp_list);
 350         atomic_set(&req->rq_refcount, 1);
 351         if (ev->type == LNET_EVENT_PUT)
 352                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
 353                        req, req->rq_xid, ev->mlength);
 354
 355         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 356
 357         spin_lock(&svcpt->scp_lock);
 358
 359         ptlrpc_req_add_history(svcpt, req);
 360
 361         if (ev->unlinked) {
 362                 svcpt->scp_nrqbds_posted--;
 363                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 364                        svcpt->scp_nrqbds_posted);
 365
 366                 /* Normally, don't complain about 0 buffers posted; LNET won't
 367                  * drop incoming reqs since we set the portal lazy */
 368                 if (test_req_buffer_pressure &&
 369                     ev->type != LNET_EVENT_UNLINK &&
 370                     svcpt->scp_nrqbds_posted == 0)
 371                         CWARN("All %s request buffers busy\n",
 372                               service->srv_name);
 373
 374                 /* req takes over the network's ref on rqbd */
 375         } else {
 376                 /* req takes a ref on rqbd */
 377                 rqbd->rqbd_refcount++;
 378         }
 379
 380         cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 381         svcpt->scp_nreqs_incoming++;
 382
 383         /* NB everything can disappear under us once the request
 384          * has been queued and we unlock, so do the wake now... */
 385         wake_up(&svcpt->scp_waitq);
 386
 387         spin_unlock(&svcpt->scp_lock);
 388         EXIT;
 389 }
 390
 391 /*
 392  *  Server's outgoing reply callback
 393  */
 394 void reply_out_callback(lnet_event_t *ev)
 395 {
 396         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 397         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 398         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 399         ENTRY;
 400
 401         LASSERT (ev->type == LNET_EVENT_SEND ||
 402                  ev->type == LNET_EVENT_ACK ||
 403                  ev->type == LNET_EVENT_UNLINK);
 404
 405         if (!rs->rs_difficult) {
 406                 /* 'Easy' replies have no further processing so I drop the
 407                  * net's ref on 'rs' */
 408                 LASSERT (ev->unlinked);
 409                 ptlrpc_rs_decref(rs);
 410                 EXIT;
 411                 return;
 412         }
 413
 414         LASSERT (rs->rs_on_net);
 415
 416         if (ev->unlinked) {
 417                 /* Last network callback. The net's ref on 'rs' stays put
 418                  * until ptlrpc_handle_rs() is done with it */
 419                 spin_lock(&svcpt->scp_rep_lock);
 420                 spin_lock(&rs->rs_lock);
 421
 422                 rs->rs_on_net = 0;
 423                 if (!rs->rs_no_ack ||
 424                     rs->rs_transno <=
 425                     rs->rs_export->exp_obd->obd_last_committed)
 426                         ptlrpc_schedule_difficult_reply(rs);
 427
 428                 spin_unlock(&rs->rs_lock);
 429                 spin_unlock(&svcpt->scp_rep_lock);
 430         }
 431         EXIT;
 432 }
 433
 434 #ifdef HAVE_SERVER_SUPPORT
 435 /*
 436  * Server's bulk completion callback
 437  */
 438 void server_bulk_callback (lnet_event_t *ev)
 439 {
 440         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 441         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 442         ENTRY;
 443
 444         LASSERT(ev->type == LNET_EVENT_SEND ||
 445                 ev->type == LNET_EVENT_UNLINK ||
 446                 (desc->bd_type == BULK_PUT_SOURCE &&
 447                  ev->type == LNET_EVENT_ACK) ||
 448                 (desc->bd_type == BULK_GET_SINK &&
 449                  ev->type == LNET_EVENT_REPLY));
 450
 451         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 452                "event type %d, status %d, desc %p\n",
 453                ev->type, ev->status, desc);
 454
 455         spin_lock(&desc->bd_lock);
 456
 457         LASSERT(desc->bd_md_count > 0);
 458
 459         if ((ev->type == LNET_EVENT_ACK ||
 460              ev->type == LNET_EVENT_REPLY) &&
 461             ev->status == 0) {
 462                 /* We heard back from the peer, so even if we get this
 463                  * before the SENT event (oh yes we can), we know we
 464                  * read/wrote the peer buffer and how much... */
 465                 desc->bd_nob_transferred += ev->mlength;
 466                 desc->bd_sender = ev->sender;
 467         }
 468
 469         if (ev->status != 0)
 470                 desc->bd_failure = 1;
 471
 472         if (ev->unlinked) {
 473                 desc->bd_md_count--;
 474                 /* This is the last callback no matter what... */
 475                 if (desc->bd_md_count == 0)
 476                         wake_up(&desc->bd_waitq);
 477         }
 478
 479         spin_unlock(&desc->bd_lock);
 480         EXIT;
 481 }
 482 #endif
 483
 484 static void ptlrpc_master_callback(lnet_event_t *ev)
 485 {
 486         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 487         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 488
 489         /* Honestly, it's best to find out early. */
 490         LASSERT (cbid->cbid_arg != LP_POISON);
 491         LASSERT (callback == request_out_callback ||
 492                  callback == reply_in_callback ||
 493                  callback == client_bulk_callback ||
 494                  callback == request_in_callback ||
 495                  callback == reply_out_callback
 496 #ifdef HAVE_SERVER_SUPPORT
 497                  || callback == server_bulk_callback
 498 #endif
 499                  );
 500
 501         callback (ev);
 502 }
 503
 504 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 505                          lnet_process_id_t *peer, lnet_nid_t *self)
 506 {
 507         int               best_dist = 0;
 508         __u32             best_order = 0;
 509         int               count = 0;
 510         int               rc = -ENOENT;
 511         int               portals_compatibility;
 512         int               dist;
 513         __u32             order;
 514         lnet_nid_t        dst_nid;
 515         lnet_nid_t        src_nid;
 516
 517         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 518
 519         peer->pid = LUSTRE_SRV_LNET_PID;
 520
 521         /* Choose the matching UUID that's closest */
 522         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 523                 dist = LNetDist(dst_nid, &src_nid, &order);
 524                 if (dist < 0)
 525                         continue;
 526
 527                 if (dist == 0) {                /* local! use loopback LND */
 528                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 529                         rc = 0;
 530                         break;
 531                 }
 532
 533                 if (rc < 0 ||
 534                     dist < best_dist ||
 535                     (dist == best_dist && order < best_order)) {
 536                         best_dist = dist;
 537                         best_order = order;
 538
 539                         if (portals_compatibility > 1) {
 540                                 /* Strong portals compatibility: Zero the nid's
 541                                  * NET, so if I'm reading new config logs, or
 542                                  * getting configured by (new) lconf I can
 543                                  * still talk to old servers. */
 544                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 545                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 546                         }
 547                         peer->nid = dst_nid;
 548                         *self = src_nid;
 549                         rc = 0;
 550                 }
 551         }
 552
 553         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 554         return rc;
 555 }
 556
 557 void ptlrpc_ni_fini(void)
 558 {
 559         wait_queue_head_t         waitq;
 560         struct l_wait_info  lwi;
 561         int                 rc;
 562         int                 retries;
 563
 564         /* Wait for the event queue to become idle since there may still be
 565          * messages in flight with pending events (i.e. the fire-and-forget
 566          * messages == client requests and "non-difficult" server
 567          * replies */
 568
 569         for (retries = 0;; retries++) {
 570                 rc = LNetEQFree(ptlrpc_eq_h);
 571                 switch (rc) {
 572                 default:
 573                         LBUG();
 574
 575                 case 0:
 576                         LNetNIFini();
 577                         return;
 578
 579                 case -EBUSY:
 580                         if (retries != 0)
 581                                 CWARN("Event queue still busy\n");
 582
 583                         /* Wait for a bit */
 584                         init_waitqueue_head(&waitq);
 585                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 586                         l_wait_event(waitq, 0, &lwi);
 587                         break;
 588                 }
 589         }
 590         /* notreached */
 591 }
 592
 593 lnet_pid_t ptl_get_pid(void)
 594 {
 595         lnet_pid_t        pid;
 596
 597 #ifndef  __KERNEL__
 598         pid = getpid();
 599 #else
 600         pid = LUSTRE_SRV_LNET_PID;
 601 #endif
 602         return pid;
 603 }
 604
 605 int ptlrpc_ni_init(void)
 606 {
 607         int              rc;
 608         lnet_pid_t       pid;
 609
 610         pid = ptl_get_pid();
 611         CDEBUG(D_NET, "My pid is: %x\n", pid);
 612
 613         /* We're not passing any limits yet... */
 614         rc = LNetNIInit(pid);
 615         if (rc < 0) {
 616                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 617                 return (-ENOENT);
 618         }
 619
 620         /* CAVEAT EMPTOR: how we process portals events is _radically_
 621          * different depending on... */
 622 #ifdef __KERNEL__
 623         /* kernel LNet calls our master callback when there are new event,
 624          * because we are guaranteed to get every event via callback,
 625          * so we just set EQ size to 0 to avoid overhread of serializing
 626          * enqueue/dequeue operations in LNet. */
 627         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 628 #else
 629         /* liblustre calls the master callback when it removes events from the
 630          * event queue.  The event queue has to be big enough not to drop
 631          * anything */
 632         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 633 #endif
 634         if (rc == 0)
 635                 return 0;
 636
 637         CERROR ("Failed to allocate event queue: %d\n", rc);
 638         LNetNIFini();
 639
 640         return (-ENOMEM);
 641 }
 642
 643 #ifndef __KERNEL__
 644 CFS_LIST_HEAD(liblustre_wait_callbacks);
 645 CFS_LIST_HEAD(liblustre_idle_callbacks);
 646 void *liblustre_services_callback;
 647
 648 void *
 649 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 650                                       const char *name,
 651                                       int (*fn)(void *arg), void *arg)
 652 {
 653         struct liblustre_wait_callback *llwc;
 654
 655         OBD_ALLOC(llwc, sizeof(*llwc));
 656         LASSERT (llwc != NULL);
 657
 658         llwc->llwc_name = name;
 659         llwc->llwc_fn = fn;
 660         llwc->llwc_arg = arg;
 661         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 662
 663         return (llwc);
 664 }
 665
 666 void
 667 liblustre_deregister_waitidle_callback (void *opaque)
 668 {
 669         struct liblustre_wait_callback *llwc = opaque;
 670
 671         cfs_list_del(&llwc->llwc_list);
 672         OBD_FREE(llwc, sizeof(*llwc));
 673 }
 674
 675 void *
 676 liblustre_register_wait_callback (const char *name,
 677                                   int (*fn)(void *arg), void *arg)
 678 {
 679         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 680                                                     name, fn, arg);
 681 }
 682
 683 void
 684 liblustre_deregister_wait_callback (void *opaque)
 685 {
 686         liblustre_deregister_waitidle_callback(opaque);
 687 }
 688
 689 void *
 690 liblustre_register_idle_callback (const char *name,
 691                                   int (*fn)(void *arg), void *arg)
 692 {
 693         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 694                                                     name, fn, arg);
 695 }
 696
 697 void
 698 liblustre_deregister_idle_callback (void *opaque)
 699 {
 700         liblustre_deregister_waitidle_callback(opaque);
 701 }
 702
 703 int
 704 liblustre_check_events (int timeout)
 705 {
 706         lnet_event_t ev;
 707         int         rc;
 708         int         i;
 709         ENTRY;
 710
 711         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 712         if (rc == 0)
 713                 RETURN(0);
 714
 715         LASSERT (rc == -EOVERFLOW || rc == 1);
 716
 717         /* liblustre: no asynch callback so we can't afford to miss any
 718          * events... */
 719         if (rc == -EOVERFLOW) {
 720                 CERROR ("Dropped an event!!!\n");
 721                 abort();
 722         }
 723
 724         ptlrpc_master_callback (&ev);
 725         RETURN(1);
 726 }
 727
 728 int liblustre_waiting = 0;
 729
 730 int
 731 liblustre_wait_event (int timeout)
 732 {
 733         cfs_list_t                     *tmp;
 734         struct liblustre_wait_callback *llwc;
 735         int                             found_something = 0;
 736
 737         /* single threaded recursion check... */
 738         liblustre_waiting = 1;
 739
 740         for (;;) {
 741                 /* Deal with all pending events */
 742                 while (liblustre_check_events(0))
 743                         found_something = 1;
 744
 745                 /* Give all registered callbacks a bite at the cherry */
 746                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 747                         llwc = cfs_list_entry(tmp,
 748                                               struct liblustre_wait_callback,
 749                                               llwc_list);
 750
 751                         if (llwc->llwc_fn(llwc->llwc_arg))
 752                                 found_something = 1;
 753                 }
 754
 755                 if (found_something || timeout == 0)
 756                         break;
 757
 758                 /* Nothing so far, but I'm allowed to block... */
 759                 found_something = liblustre_check_events(timeout);
 760                 if (!found_something)           /* still nothing */
 761                         break;                  /* I timed out */
 762         }
 763
 764         liblustre_waiting = 0;
 765
 766         return found_something;
 767 }
 768
 769 void
 770 liblustre_wait_idle(void)
 771 {
 772         static int recursed = 0;
 773
 774         cfs_list_t                     *tmp;
 775         struct liblustre_wait_callback *llwc;
 776         int                             idle = 0;
 777
 778         LASSERT(!recursed);
 779         recursed = 1;
 780
 781         do {
 782                 liblustre_wait_event(0);
 783
 784                 idle = 1;
 785
 786                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 787                         llwc = cfs_list_entry(tmp,
 788                                               struct liblustre_wait_callback,
 789                                               llwc_list);
 790
 791                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 792                                 idle = 0;
 793                                 break;
 794                         }
 795                 }
 796
 797         } while (!idle);
 798
 799         recursed = 0;
 800 }
 801
 802 #endif /* __KERNEL__ */
 803
 804 int ptlrpc_init_portals(void)
 805 {
 806         int   rc = ptlrpc_ni_init();
 807
 808         if (rc != 0) {
 809                 CERROR("network initialisation failed\n");
 810                 return -EIO;
 811         }
 812 #ifndef __KERNEL__
 813         liblustre_services_callback =
 814                 liblustre_register_wait_callback("liblustre_check_services",
 815                                                  &liblustre_check_services,
 816                                                  NULL);
 817         init_completion_module(liblustre_wait_event);
 818 #endif
 819         rc = ptlrpcd_addref();
 820         if (rc == 0)
 821                 return 0;
 822
 823         CERROR("rpcd initialisation failed\n");
 824 #ifndef __KERNEL__
 825         liblustre_deregister_wait_callback(liblustre_services_callback);
 826 #endif
 827         ptlrpc_ni_fini();
 828         return rc;
 829 }
 830
 831 void ptlrpc_exit_portals(void)
 832 {
 833 #ifndef __KERNEL__
 834         liblustre_deregister_wait_callback(liblustre_services_callback);
 835 #endif
 836         ptlrpcd_decref();
 837         ptlrpc_ni_fini();
 838 }