lustre/ptlrpc/events.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  */
  30 /*
  31  * This file is part of Lustre, http://www.lustre.org/
  32  * Lustre is a trademark of Sun Microsystems, Inc.
  33  */
  34
  35 #define DEBUG_SUBSYSTEM S_RPC
  36
  37 #ifndef __KERNEL__
  38 # include <liblustre.h>
  39 #else
  40 # include <libcfs/libcfs.h>
  41 # ifdef __mips64__
  42 #  include <linux/kernel.h>
  43 # endif
  44 #endif
  45
  46 #include <obd_class.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ptlrpc_internal.h"
  50
  51 lnet_handle_eq_t   ptlrpc_eq_h;
  52
  53 /*
  54  *  Client's outgoing request callback
  55  */
  56 void request_out_callback(lnet_event_t *ev)
  57 {
  58         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  59         struct ptlrpc_request *req = cbid->cbid_arg;
  60         ENTRY;
  61
  62         LASSERT (ev->type == LNET_EVENT_SEND ||
  63                  ev->type == LNET_EVENT_UNLINK);
  64         LASSERT (ev->unlinked);
  65
  66         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  67
  68         sptlrpc_request_out_callback(req);
  69         req->rq_real_sent = cfs_time_current_sec();
  70
  71         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
  72
  73                 /* Failed send: make it seem like the reply timed out, just
  74                  * like failing sends in client.c does currently...  */
  75
  76                 cfs_spin_lock(&req->rq_lock);
  77                 req->rq_net_err = 1;
  78                 cfs_spin_unlock(&req->rq_lock);
  79
  80                 ptlrpc_client_wake_req(req);
  81         }
  82
  83         ptlrpc_req_finished(req);
  84
  85         EXIT;
  86 }
  87
  88 /*
  89  * Client's incoming reply callback
  90  */
  91 void reply_in_callback(lnet_event_t *ev)
  92 {
  93         struct ptlrpc_cb_id   *cbid = ev->md.user_ptr;
  94         struct ptlrpc_request *req = cbid->cbid_arg;
  95         ENTRY;
  96
  97         DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
  98
  99         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
 100         LASSERT (ev->md.start == req->rq_repbuf);
 101         LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
 102         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
 103            for adaptive timeouts' early reply. */
 104         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 105
 106         cfs_spin_lock(&req->rq_lock);
 107
 108         req->rq_receiving_reply = 0;
 109         req->rq_early = 0;
 110         if (ev->unlinked)
 111                 req->rq_must_unlink = 0;
 112
 113         if (ev->status)
 114                 goto out_wake;
 115
 116         if (ev->type == LNET_EVENT_UNLINK) {
 117                 LASSERT(ev->unlinked);
 118                 DEBUG_REQ(D_NET, req, "unlink");
 119                 goto out_wake;
 120         }
 121
 122         if (ev->mlength < ev->rlength ) {
 123                 CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
 124                        req->rq_replen, ev->rlength, ev->offset);
 125                 req->rq_reply_truncate = 1;
 126                 req->rq_replied = 1;
 127                 req->rq_status = -EOVERFLOW;
 128                 req->rq_nob_received = ev->rlength + ev->offset;
 129                 goto out_wake;
 130         }
 131
 132         if ((ev->offset == 0) &&
 133             ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT))) {
 134                 /* Early reply */
 135                 DEBUG_REQ(D_ADAPTTO, req,
 136                           "Early reply received: mlen=%u offset=%d replen=%d "
 137                           "replied=%d unlinked=%d", ev->mlength, ev->offset,
 138                           req->rq_replen, req->rq_replied, ev->unlinked);
 139
 140                 req->rq_early_count++; /* number received, client side */
 141
 142                 if (req->rq_replied)   /* already got the real reply */
 143                         goto out_wake;
 144
 145                 req->rq_early = 1;
 146                 req->rq_reply_off = ev->offset;
 147                 req->rq_nob_received = ev->mlength;
 148                 /* And we're still receiving */
 149                 req->rq_receiving_reply = 1;
 150         } else {
 151                 /* Real reply */
 152                 req->rq_rep_swab_mask = 0;
 153                 req->rq_replied = 1;
 154                 req->rq_reply_off = ev->offset;
 155                 req->rq_nob_received = ev->mlength;
 156                 /* LNetMDUnlink can't be called under the LNET_LOCK,
 157                    so we must unlink in ptlrpc_unregister_reply */
 158                 DEBUG_REQ(D_INFO, req,
 159                           "reply in flags=%x mlen=%u offset=%d replen=%d",
 160                           lustre_msg_get_flags(req->rq_reqmsg),
 161                           ev->mlength, ev->offset, req->rq_replen);
 162         }
 163
 164         req->rq_import->imp_last_reply_time = cfs_time_current_sec();
 165
 166 out_wake:
 167         /* NB don't unlock till after wakeup; req can disappear under us
 168          * since we don't have our own ref */
 169         ptlrpc_client_wake_req(req);
 170         cfs_spin_unlock(&req->rq_lock);
 171         EXIT;
 172 }
 173
 174 /*
 175  * Client's bulk has been written/read
 176  */
 177 void client_bulk_callback (lnet_event_t *ev)
 178 {
 179         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 180         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 181         ENTRY;
 182
 183         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
 184                   ev->type == LNET_EVENT_PUT) ||
 185                  (desc->bd_type == BULK_GET_SOURCE &&
 186                   ev->type == LNET_EVENT_GET) ||
 187                  ev->type == LNET_EVENT_UNLINK);
 188         LASSERT (ev->unlinked);
 189
 190         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
 191                 ev->status = -EIO;
 192
 193         if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
 194                 ev->status = -EIO;
 195
 196         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 197                "event type %d, status %d, desc %p\n",
 198                ev->type, ev->status, desc);
 199
 200         cfs_spin_lock(&desc->bd_lock);
 201
 202         LASSERT(desc->bd_network_rw);
 203         desc->bd_network_rw = 0;
 204
 205         if (ev->type != LNET_EVENT_UNLINK && ev->status == 0) {
 206                 desc->bd_success = 1;
 207                 desc->bd_nob_transferred = ev->mlength;
 208                 desc->bd_sender = ev->sender;
 209         }
 210
 211         /* release the encrypted pages for write */
 212         if (desc->bd_req->rq_bulk_write)
 213                 sptlrpc_enc_pool_put_pages(desc);
 214
 215         /* NB don't unlock till after wakeup; desc can disappear under us
 216          * otherwise */
 217         ptlrpc_client_wake_req(desc->bd_req);
 218
 219         cfs_spin_unlock(&desc->bd_lock);
 220         EXIT;
 221 }
 222
 223 /*
 224  * We will have percpt request history list for ptlrpc service in upcoming
 225  * patches because we don't want to be serialized by current per-service
 226  * history operations. So we require history ID can (somehow) show arriving
 227  * order w/o grabbing global lock, and user can sort them in userspace.
 228  *
 229  * This is how we generate history ID for ptlrpc_request:
 230  * ----------------------------------------------------
 231  * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
 232  * ----------------------------------------------------
 233  * |  seconds  | usec / 16 |   sequence    | CPT id   |
 234  * ----------------------------------------------------
 235  *
 236  * it might not be precise but should be good enough.
 237  */
 238 #define REQS_ALL_BITS(svcpt)    ((int)(sizeof((svcpt)->scp_hist_seq) * 8))
 239 #define REQS_SEC_BITS           32
 240 #define REQS_USEC_BITS          16
 241 /* will be replaced by bits for total service partition number soon */
 242 #define REQS_CPT_BITS(svcpt)    0
 243 #define REQS_SEQ_BITS(svcpt)    (REQS_ALL_BITS(svcpt) - REQS_CPT_BITS(svcpt) -\
 244                                  REQS_SEC_BITS - REQS_USEC_BITS)
 245
 246 #define REQS_SEQ_SHIFT(svcpt)   (REQS_CPT_BITS(svcpt))
 247 #define REQS_USEC_SHIFT(svcpt)  (REQS_SEQ_SHIFT(svcpt) + REQS_SEQ_BITS(svcpt))
 248 #define REQS_SEC_SHIFT(svcpt)   (REQS_USEC_SHIFT(svcpt) + REQS_USEC_BITS)
 249
 250 static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
 251                                    struct ptlrpc_request *req)
 252 {
 253         __u64   sec = req->rq_arrival_time.tv_sec;
 254         __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
 255         __u64   new_seq;
 256
 257         /* set sequence ID for request and add it to history list,
 258          * it must be called with hold svcpt::scp_lock */
 259
 260         LASSERT(REQS_SEQ_BITS(svcpt) > 0);
 261
 262         new_seq = (sec << REQS_SEC_SHIFT(svcpt)) |
 263                   (usec << REQS_USEC_SHIFT(svcpt)) | svcpt->scp_cpt;
 264         if (new_seq > svcpt->scp_hist_seq) {
 265                 /* This handles the initial case of scp_hist_seq == 0 or
 266                  * we just jumped into a new time window */
 267                 svcpt->scp_hist_seq = new_seq;
 268         } else {
 269                 /* NB: increase sequence number in current usec bucket,
 270                  * however, it's possible that we used up all bits for
 271                  * sequence and jumped into the next usec bucket (future time),
 272                  * then we hope there will be less RPCs per bucket at some
 273                  * point, and sequence will catch up again */
 274                 svcpt->scp_hist_seq += (1U << REQS_CPT_BITS(svcpt));
 275                 new_seq = svcpt->scp_hist_seq;
 276         }
 277
 278         req->rq_history_seq = new_seq;
 279
 280         cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
 281 }
 282
 283 /*
 284  * Server's incoming request callback
 285  */
 286 void request_in_callback(lnet_event_t *ev)
 287 {
 288         struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
 289         struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
 290         struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
 291         struct ptlrpc_service             *service = svcpt->scp_service;
 292         struct ptlrpc_request             *req;
 293         ENTRY;
 294
 295         LASSERT (ev->type == LNET_EVENT_PUT ||
 296                  ev->type == LNET_EVENT_UNLINK);
 297         LASSERT ((char *)ev->md.start >= rqbd->rqbd_buffer);
 298         LASSERT ((char *)ev->md.start + ev->offset + ev->mlength <=
 299                  rqbd->rqbd_buffer + service->srv_buf_size);
 300
 301         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 302                "event type %d, status %d, service %s\n",
 303                ev->type, ev->status, service->srv_name);
 304
 305         if (ev->unlinked) {
 306                 /* If this is the last request message to fit in the
 307                  * request buffer we can use the request object embedded in
 308                  * rqbd.  Note that if we failed to allocate a request,
 309                  * we'd have to re-post the rqbd, which we can't do in this
 310                  * context. */
 311                 req = &rqbd->rqbd_req;
 312                 memset(req, 0, sizeof (*req));
 313         } else {
 314                 LASSERT (ev->type == LNET_EVENT_PUT);
 315                 if (ev->status != 0) {
 316                         /* We moaned above already... */
 317                         return;
 318                 }
 319                 OBD_ALLOC_GFP(req, sizeof(*req), CFS_ALLOC_ATOMIC_TRY);
 320                 if (req == NULL) {
 321                         CERROR("Can't allocate incoming request descriptor: "
 322                                "Dropping %s RPC from %s\n",
 323                                service->srv_name,
 324                                libcfs_id2str(ev->initiator));
 325                         return;
 326                 }
 327         }
 328
 329         /* NB we ABSOLUTELY RELY on req being zeroed, so pointers are NULL,
 330          * flags are reset and scalars are zero.  We only set the message
 331          * size to non-zero if this was a successful receive. */
 332         req->rq_xid = ev->match_bits;
 333         req->rq_reqbuf = ev->md.start + ev->offset;
 334         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
 335                 req->rq_reqdata_len = ev->mlength;
 336         cfs_gettimeofday(&req->rq_arrival_time);
 337         req->rq_peer = ev->initiator;
 338         req->rq_self = ev->target.nid;
 339         req->rq_rqbd = rqbd;
 340         req->rq_phase = RQ_PHASE_NEW;
 341 #ifdef CRAY_XT3
 342         req->rq_uid = ev->uid;
 343 #endif
 344         cfs_spin_lock_init(&req->rq_lock);
 345         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
 346         cfs_atomic_set(&req->rq_refcount, 1);
 347         if (ev->type == LNET_EVENT_PUT)
 348                 CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
 349                        req, req->rq_xid, ev->mlength);
 350
 351         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 352
 353         cfs_spin_lock(&svcpt->scp_lock);
 354
 355         ptlrpc_req_add_history(svcpt, req);
 356
 357         if (ev->unlinked) {
 358                 svcpt->scp_nrqbds_posted--;
 359                 CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
 360                        svcpt->scp_nrqbds_posted);
 361
 362                 /* Normally, don't complain about 0 buffers posted; LNET won't
 363                  * drop incoming reqs since we set the portal lazy */
 364                 if (test_req_buffer_pressure &&
 365                     ev->type != LNET_EVENT_UNLINK &&
 366                     svcpt->scp_nrqbds_posted == 0)
 367                         CWARN("All %s request buffers busy\n",
 368                               service->srv_name);
 369
 370                 /* req takes over the network's ref on rqbd */
 371         } else {
 372                 /* req takes a ref on rqbd */
 373                 rqbd->rqbd_refcount++;
 374         }
 375
 376         cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
 377         svcpt->scp_nreqs_incoming++;
 378
 379         /* NB everything can disappear under us once the request
 380          * has been queued and we unlock, so do the wake now... */
 381         cfs_waitq_signal(&svcpt->scp_waitq);
 382
 383         cfs_spin_unlock(&svcpt->scp_lock);
 384         EXIT;
 385 }
 386
 387 /*
 388  *  Server's outgoing reply callback
 389  */
 390 void reply_out_callback(lnet_event_t *ev)
 391 {
 392         struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
 393         struct ptlrpc_reply_state *rs = cbid->cbid_arg;
 394         struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 395         ENTRY;
 396
 397         LASSERT (ev->type == LNET_EVENT_SEND ||
 398                  ev->type == LNET_EVENT_ACK ||
 399                  ev->type == LNET_EVENT_UNLINK);
 400
 401         if (!rs->rs_difficult) {
 402                 /* 'Easy' replies have no further processing so I drop the
 403                  * net's ref on 'rs' */
 404                 LASSERT (ev->unlinked);
 405                 ptlrpc_rs_decref(rs);
 406                 EXIT;
 407                 return;
 408         }
 409
 410         LASSERT (rs->rs_on_net);
 411
 412         if (ev->unlinked) {
 413                 /* Last network callback. The net's ref on 'rs' stays put
 414                  * until ptlrpc_handle_rs() is done with it */
 415                 cfs_spin_lock(&svcpt->scp_rep_lock);
 416                 cfs_spin_lock(&rs->rs_lock);
 417
 418                 rs->rs_on_net = 0;
 419                 if (!rs->rs_no_ack ||
 420                     rs->rs_transno <=
 421                     rs->rs_export->exp_obd->obd_last_committed)
 422                         ptlrpc_schedule_difficult_reply(rs);
 423
 424                 cfs_spin_unlock(&rs->rs_lock);
 425                 cfs_spin_unlock(&svcpt->scp_rep_lock);
 426         }
 427         EXIT;
 428 }
 429
 430 #ifdef HAVE_SERVER_SUPPORT
 431 /*
 432  * Server's bulk completion callback
 433  */
 434 void server_bulk_callback (lnet_event_t *ev)
 435 {
 436         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
 437         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
 438         ENTRY;
 439
 440         LASSERT (ev->type == LNET_EVENT_SEND ||
 441                  ev->type == LNET_EVENT_UNLINK ||
 442                  (desc->bd_type == BULK_PUT_SOURCE &&
 443                   ev->type == LNET_EVENT_ACK) ||
 444                  (desc->bd_type == BULK_GET_SINK &&
 445                   ev->type == LNET_EVENT_REPLY));
 446
 447         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
 448                "event type %d, status %d, desc %p\n",
 449                ev->type, ev->status, desc);
 450
 451         cfs_spin_lock(&desc->bd_lock);
 452
 453         if ((ev->type == LNET_EVENT_ACK ||
 454              ev->type == LNET_EVENT_REPLY) &&
 455             ev->status == 0) {
 456                 /* We heard back from the peer, so even if we get this
 457                  * before the SENT event (oh yes we can), we know we
 458                  * read/wrote the peer buffer and how much... */
 459                 desc->bd_success = 1;
 460                 desc->bd_nob_transferred = ev->mlength;
 461                 desc->bd_sender = ev->sender;
 462         }
 463
 464         if (ev->unlinked) {
 465                 /* This is the last callback no matter what... */
 466                 desc->bd_network_rw = 0;
 467                 cfs_waitq_signal(&desc->bd_waitq);
 468         }
 469
 470         cfs_spin_unlock(&desc->bd_lock);
 471         EXIT;
 472 }
 473 #endif
 474
 475 static void ptlrpc_master_callback(lnet_event_t *ev)
 476 {
 477         struct ptlrpc_cb_id *cbid = ev->md.user_ptr;
 478         void (*callback)(lnet_event_t *ev) = cbid->cbid_fn;
 479
 480         /* Honestly, it's best to find out early. */
 481         LASSERT (cbid->cbid_arg != LP_POISON);
 482         LASSERT (callback == request_out_callback ||
 483                  callback == reply_in_callback ||
 484                  callback == client_bulk_callback ||
 485                  callback == request_in_callback ||
 486                  callback == reply_out_callback
 487 #ifdef HAVE_SERVER_SUPPORT
 488                  || callback == server_bulk_callback
 489 #endif
 490                  );
 491
 492         callback (ev);
 493 }
 494
 495 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
 496                          lnet_process_id_t *peer, lnet_nid_t *self)
 497 {
 498         int               best_dist = 0;
 499         __u32             best_order = 0;
 500         int               count = 0;
 501         int               rc = -ENOENT;
 502         int               portals_compatibility;
 503         int               dist;
 504         __u32             order;
 505         lnet_nid_t        dst_nid;
 506         lnet_nid_t        src_nid;
 507
 508         portals_compatibility = LNetCtl(IOC_LIBCFS_PORTALS_COMPATIBILITY, NULL);
 509
 510         peer->pid = LUSTRE_SRV_LNET_PID;
 511
 512         /* Choose the matching UUID that's closest */
 513         while (lustre_uuid_to_peer(uuid->uuid, &dst_nid, count++) == 0) {
 514                 dist = LNetDist(dst_nid, &src_nid, &order);
 515                 if (dist < 0)
 516                         continue;
 517
 518                 if (dist == 0) {                /* local! use loopback LND */
 519                         peer->nid = *self = LNET_MKNID(LNET_MKNET(LOLND, 0), 0);
 520                         rc = 0;
 521                         break;
 522                 }
 523
 524                 if (rc < 0 ||
 525                     dist < best_dist ||
 526                     (dist == best_dist && order < best_order)) {
 527                         best_dist = dist;
 528                         best_order = order;
 529
 530                         if (portals_compatibility > 1) {
 531                                 /* Strong portals compatibility: Zero the nid's
 532                                  * NET, so if I'm reading new config logs, or
 533                                  * getting configured by (new) lconf I can
 534                                  * still talk to old servers. */
 535                                 dst_nid = LNET_MKNID(0, LNET_NIDADDR(dst_nid));
 536                                 src_nid = LNET_MKNID(0, LNET_NIDADDR(src_nid));
 537                         }
 538                         peer->nid = dst_nid;
 539                         *self = src_nid;
 540                         rc = 0;
 541                 }
 542         }
 543
 544         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
 545         return rc;
 546 }
 547
 548 void ptlrpc_ni_fini(void)
 549 {
 550         cfs_waitq_t         waitq;
 551         struct l_wait_info  lwi;
 552         int                 rc;
 553         int                 retries;
 554
 555         /* Wait for the event queue to become idle since there may still be
 556          * messages in flight with pending events (i.e. the fire-and-forget
 557          * messages == client requests and "non-difficult" server
 558          * replies */
 559
 560         for (retries = 0;; retries++) {
 561                 rc = LNetEQFree(ptlrpc_eq_h);
 562                 switch (rc) {
 563                 default:
 564                         LBUG();
 565
 566                 case 0:
 567                         LNetNIFini();
 568                         return;
 569
 570                 case -EBUSY:
 571                         if (retries != 0)
 572                                 CWARN("Event queue still busy\n");
 573
 574                         /* Wait for a bit */
 575                         cfs_waitq_init(&waitq);
 576                         lwi = LWI_TIMEOUT(cfs_time_seconds(2), NULL, NULL);
 577                         l_wait_event(waitq, 0, &lwi);
 578                         break;
 579                 }
 580         }
 581         /* notreached */
 582 }
 583
 584 lnet_pid_t ptl_get_pid(void)
 585 {
 586         lnet_pid_t        pid;
 587
 588 #ifndef  __KERNEL__
 589         pid = getpid();
 590 #else
 591         pid = LUSTRE_SRV_LNET_PID;
 592 #endif
 593         return pid;
 594 }
 595
 596 int ptlrpc_ni_init(void)
 597 {
 598         int              rc;
 599         lnet_pid_t       pid;
 600
 601         pid = ptl_get_pid();
 602         CDEBUG(D_NET, "My pid is: %x\n", pid);
 603
 604         /* We're not passing any limits yet... */
 605         rc = LNetNIInit(pid);
 606         if (rc < 0) {
 607                 CDEBUG (D_NET, "Can't init network interface: %d\n", rc);
 608                 return (-ENOENT);
 609         }
 610
 611         /* CAVEAT EMPTOR: how we process portals events is _radically_
 612          * different depending on... */
 613 #ifdef __KERNEL__
 614         /* kernel LNet calls our master callback when there are new event,
 615          * because we are guaranteed to get every event via callback,
 616          * so we just set EQ size to 0 to avoid overhread of serializing
 617          * enqueue/dequeue operations in LNet. */
 618         rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 619 #else
 620         /* liblustre calls the master callback when it removes events from the
 621          * event queue.  The event queue has to be big enough not to drop
 622          * anything */
 623         rc = LNetEQAlloc(10240, LNET_EQ_HANDLER_NONE, &ptlrpc_eq_h);
 624 #endif
 625         if (rc == 0)
 626                 return 0;
 627
 628         CERROR ("Failed to allocate event queue: %d\n", rc);
 629         LNetNIFini();
 630
 631         return (-ENOMEM);
 632 }
 633
 634 #ifndef __KERNEL__
 635 CFS_LIST_HEAD(liblustre_wait_callbacks);
 636 CFS_LIST_HEAD(liblustre_idle_callbacks);
 637 void *liblustre_services_callback;
 638
 639 void *
 640 liblustre_register_waitidle_callback (cfs_list_t *callback_list,
 641                                       const char *name,
 642                                       int (*fn)(void *arg), void *arg)
 643 {
 644         struct liblustre_wait_callback *llwc;
 645
 646         OBD_ALLOC(llwc, sizeof(*llwc));
 647         LASSERT (llwc != NULL);
 648
 649         llwc->llwc_name = name;
 650         llwc->llwc_fn = fn;
 651         llwc->llwc_arg = arg;
 652         cfs_list_add_tail(&llwc->llwc_list, callback_list);
 653
 654         return (llwc);
 655 }
 656
 657 void
 658 liblustre_deregister_waitidle_callback (void *opaque)
 659 {
 660         struct liblustre_wait_callback *llwc = opaque;
 661
 662         cfs_list_del(&llwc->llwc_list);
 663         OBD_FREE(llwc, sizeof(*llwc));
 664 }
 665
 666 void *
 667 liblustre_register_wait_callback (const char *name,
 668                                   int (*fn)(void *arg), void *arg)
 669 {
 670         return liblustre_register_waitidle_callback(&liblustre_wait_callbacks,
 671                                                     name, fn, arg);
 672 }
 673
 674 void
 675 liblustre_deregister_wait_callback (void *opaque)
 676 {
 677         liblustre_deregister_waitidle_callback(opaque);
 678 }
 679
 680 void *
 681 liblustre_register_idle_callback (const char *name,
 682                                   int (*fn)(void *arg), void *arg)
 683 {
 684         return liblustre_register_waitidle_callback(&liblustre_idle_callbacks,
 685                                                     name, fn, arg);
 686 }
 687
 688 void
 689 liblustre_deregister_idle_callback (void *opaque)
 690 {
 691         liblustre_deregister_waitidle_callback(opaque);
 692 }
 693
 694 int
 695 liblustre_check_events (int timeout)
 696 {
 697         lnet_event_t ev;
 698         int         rc;
 699         int         i;
 700         ENTRY;
 701
 702         rc = LNetEQPoll(&ptlrpc_eq_h, 1, timeout * 1000, &ev, &i);
 703         if (rc == 0)
 704                 RETURN(0);
 705
 706         LASSERT (rc == -EOVERFLOW || rc == 1);
 707
 708         /* liblustre: no asynch callback so we can't afford to miss any
 709          * events... */
 710         if (rc == -EOVERFLOW) {
 711                 CERROR ("Dropped an event!!!\n");
 712                 abort();
 713         }
 714
 715         ptlrpc_master_callback (&ev);
 716         RETURN(1);
 717 }
 718
 719 int liblustre_waiting = 0;
 720
 721 int
 722 liblustre_wait_event (int timeout)
 723 {
 724         cfs_list_t                     *tmp;
 725         struct liblustre_wait_callback *llwc;
 726         int                             found_something = 0;
 727
 728         /* single threaded recursion check... */
 729         liblustre_waiting = 1;
 730
 731         for (;;) {
 732                 /* Deal with all pending events */
 733                 while (liblustre_check_events(0))
 734                         found_something = 1;
 735
 736                 /* Give all registered callbacks a bite at the cherry */
 737                 cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
 738                         llwc = cfs_list_entry(tmp,
 739                                               struct liblustre_wait_callback,
 740                                               llwc_list);
 741
 742                         if (llwc->llwc_fn(llwc->llwc_arg))
 743                                 found_something = 1;
 744                 }
 745
 746                 if (found_something || timeout == 0)
 747                         break;
 748
 749                 /* Nothing so far, but I'm allowed to block... */
 750                 found_something = liblustre_check_events(timeout);
 751                 if (!found_something)           /* still nothing */
 752                         break;                  /* I timed out */
 753         }
 754
 755         liblustre_waiting = 0;
 756
 757         return found_something;
 758 }
 759
 760 void
 761 liblustre_wait_idle(void)
 762 {
 763         static int recursed = 0;
 764
 765         cfs_list_t                     *tmp;
 766         struct liblustre_wait_callback *llwc;
 767         int                             idle = 0;
 768
 769         LASSERT(!recursed);
 770         recursed = 1;
 771
 772         do {
 773                 liblustre_wait_event(0);
 774
 775                 idle = 1;
 776
 777                 cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
 778                         llwc = cfs_list_entry(tmp,
 779                                               struct liblustre_wait_callback,
 780                                               llwc_list);
 781
 782                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
 783                                 idle = 0;
 784                                 break;
 785                         }
 786                 }
 787
 788         } while (!idle);
 789
 790         recursed = 0;
 791 }
 792
 793 #endif /* __KERNEL__ */
 794
 795 int ptlrpc_init_portals(void)
 796 {
 797         int   rc = ptlrpc_ni_init();
 798
 799         if (rc != 0) {
 800                 CERROR("network initialisation failed\n");
 801                 return -EIO;
 802         }
 803 #ifndef __KERNEL__
 804         liblustre_services_callback =
 805                 liblustre_register_wait_callback("liblustre_check_services",
 806                                                  &liblustre_check_services,
 807                                                  NULL);
 808         cfs_init_completion_module(liblustre_wait_event);
 809 #endif
 810         rc = ptlrpcd_addref();
 811         if (rc == 0)
 812                 return 0;
 813
 814         CERROR("rpcd initialisation failed\n");
 815 #ifndef __KERNEL__
 816         liblustre_deregister_wait_callback(liblustre_services_callback);
 817 #endif
 818         ptlrpc_ni_fini();
 819         return rc;
 820 }
 821
 822 void ptlrpc_exit_portals(void)
 823 {
 824 #ifndef __KERNEL__
 825         liblustre_deregister_wait_callback(liblustre_services_callback);
 826 #endif
 827         ptlrpcd_decref();
 828         ptlrpc_ni_fini();
 829 }