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