lustre/ptlrpc/service.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  *
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 #include <liblustre.h>
  29 #include <libcfs/kp30.h>
  30 #endif
  31 #include <obd_support.h>
  32 #include <obd_class.h>
  33 #include <lustre_net.h>
  34 #include <lu_object.h>
  35 #include <lnet/types.h>
  36 #include "ptlrpc_internal.h"
  37
  38 int test_req_buffer_pressure = 0;
  39 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
  40                 "set non-zero to put pressure on request buffer pools");
  41
  42 /* forward ref */
  43 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
  44
  45 static CFS_LIST_HEAD (ptlrpc_all_services);
  46 spinlock_t ptlrpc_all_services_lock;
  47
  48 static char *
  49 ptlrpc_alloc_request_buffer (int size)
  50 {
  51         char *ptr;
  52
  53         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  54                 OBD_VMALLOC(ptr, size);
  55         else
  56                 OBD_ALLOC(ptr, size);
  57
  58         return (ptr);
  59 }
  60
  61 static void
  62 ptlrpc_free_request_buffer (char *ptr, int size)
  63 {
  64         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  65                 OBD_VFREE(ptr, size);
  66         else
  67                 OBD_FREE(ptr, size);
  68 }
  69
  70 struct ptlrpc_request_buffer_desc *
  71 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
  72 {
  73         struct ptlrpc_request_buffer_desc *rqbd;
  74
  75         OBD_ALLOC_PTR(rqbd);
  76         if (rqbd == NULL)
  77                 return (NULL);
  78
  79         rqbd->rqbd_service = svc;
  80         rqbd->rqbd_refcount = 0;
  81         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  82         rqbd->rqbd_cbid.cbid_arg = rqbd;
  83         CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
  84         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
  85
  86         if (rqbd->rqbd_buffer == NULL) {
  87                 OBD_FREE_PTR(rqbd);
  88                 return (NULL);
  89         }
  90
  91         spin_lock(&svc->srv_lock);
  92         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
  93         svc->srv_nbufs++;
  94         spin_unlock(&svc->srv_lock);
  95
  96         return (rqbd);
  97 }
  98
  99 void
 100 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 101 {
 102         struct ptlrpc_service *svc = rqbd->rqbd_service;
 103
 104         LASSERT (rqbd->rqbd_refcount == 0);
 105         LASSERT (list_empty(&rqbd->rqbd_reqs));
 106
 107         spin_lock(&svc->srv_lock);
 108         list_del(&rqbd->rqbd_list);
 109         svc->srv_nbufs--;
 110         spin_unlock(&svc->srv_lock);
 111
 112         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 113         OBD_FREE_PTR(rqbd);
 114 }
 115
 116 int
 117 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
 118 {
 119         struct ptlrpc_request_buffer_desc *rqbd;
 120         int                                i;
 121
 122         CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
 123                svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
 124                svc->srv_nrqbd_receiving, svc->srv_nbufs);
 125         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 126                 rqbd = ptlrpc_alloc_rqbd(svc);
 127
 128                 if (rqbd == NULL) {
 129                         CERROR ("%s: Can't allocate request buffer\n",
 130                                 svc->srv_name);
 131                         return (-ENOMEM);
 132                 }
 133
 134                 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
 135                         return (-EAGAIN);
 136         }
 137
 138         return (0);
 139 }
 140
 141 void
 142 ptlrpc_save_lock (struct ptlrpc_request *req,
 143                   struct lustre_handle *lock, int mode)
 144 {
 145         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 146         int                        idx;
 147
 148         LASSERT(rs != NULL);
 149         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
 150
 151         idx = rs->rs_nlocks++;
 152         rs->rs_locks[idx] = *lock;
 153         rs->rs_modes[idx] = mode;
 154         rs->rs_difficult = 1;
 155 }
 156
 157 void
 158 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 159 {
 160         struct ptlrpc_service *svc = rs->rs_service;
 161
 162 #ifdef CONFIG_SMP
 163         LASSERT (spin_is_locked (&svc->srv_lock));
 164 #endif
 165         LASSERT (rs->rs_difficult);
 166         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 167
 168         if (rs->rs_scheduled)                   /* being set up or already notified */
 169                 return;
 170
 171         rs->rs_scheduled = 1;
 172         list_del (&rs->rs_list);
 173         list_add (&rs->rs_list, &svc->srv_reply_queue);
 174         cfs_waitq_signal (&svc->srv_waitq);
 175 }
 176
 177 void
 178 ptlrpc_commit_replies (struct obd_device *obd)
 179 {
 180         struct list_head   *tmp;
 181         struct list_head   *nxt;
 182
 183         /* Find any replies that have been committed and get their service
 184          * to attend to complete them. */
 185
 186         /* CAVEAT EMPTOR: spinlock ordering!!! */
 187         spin_lock(&obd->obd_uncommitted_replies_lock);
 188
 189         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 190                 struct ptlrpc_reply_state *rs =
 191                         list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
 192
 193                 LASSERT (rs->rs_difficult);
 194
 195                 if (rs->rs_transno <= obd->obd_last_committed) {
 196                         struct ptlrpc_service *svc = rs->rs_service;
 197
 198                         spin_lock (&svc->srv_lock);
 199                         list_del_init (&rs->rs_obd_list);
 200                         ptlrpc_schedule_difficult_reply (rs);
 201                         spin_unlock (&svc->srv_lock);
 202                 }
 203         }
 204
 205         spin_unlock(&obd->obd_uncommitted_replies_lock);
 206 }
 207
 208 static int
 209 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 210 {
 211         struct ptlrpc_request_buffer_desc *rqbd;
 212         int                                rc;
 213         int                                posted = 0;
 214
 215         for (;;) {
 216                 spin_lock(&svc->srv_lock);
 217
 218                 if (list_empty (&svc->srv_idle_rqbds)) {
 219                         spin_unlock(&svc->srv_lock);
 220                         return (posted);
 221                 }
 222
 223                 rqbd = list_entry(svc->srv_idle_rqbds.next,
 224                                   struct ptlrpc_request_buffer_desc,
 225                                   rqbd_list);
 226                 list_del (&rqbd->rqbd_list);
 227
 228                 /* assume we will post successfully */
 229                 svc->srv_nrqbd_receiving++;
 230                 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
 231
 232                 spin_unlock(&svc->srv_lock);
 233
 234                 rc = ptlrpc_register_rqbd(rqbd);
 235                 if (rc != 0)
 236                         break;
 237
 238                 posted = 1;
 239         }
 240
 241         spin_lock(&svc->srv_lock);
 242
 243         svc->srv_nrqbd_receiving--;
 244         list_del(&rqbd->rqbd_list);
 245         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 246
 247         /* Don't complain if no request buffers are posted right now; LNET
 248          * won't drop requests because we set the portal lazy! */
 249
 250         spin_unlock(&svc->srv_lock);
 251
 252         return (-1);
 253 }
 254
 255 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
 256                                             svc_handler_t h, char *name,
 257                                             struct proc_dir_entry *proc_entry,
 258                                             svcreq_printfn_t prntfn)
 259 {
 260         return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
 261                                c->psc_max_req_size, c->psc_max_reply_size,
 262                                c->psc_req_portal, c->psc_rep_portal,
 263                                c->psc_watchdog_timeout,
 264                                h, name, proc_entry,
 265                                prntfn, c->psc_num_threads, c->psc_ctx_tags);
 266 }
 267 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
 268
 269 struct ptlrpc_service *
 270 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
 271                 int req_portal, int rep_portal, int watchdog_timeout,
 272                 svc_handler_t handler, char *name,
 273                 cfs_proc_dir_entry_t *proc_entry,
 274                 svcreq_printfn_t svcreq_printfn, int num_threads,
 275                 __u32 ctx_tags)
 276 {
 277         int                    rc;
 278         struct ptlrpc_service *service;
 279         ENTRY;
 280
 281         LASSERT (nbufs > 0);
 282         LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
 283         LASSERT (ctx_tags != 0);
 284
 285         OBD_ALLOC_PTR(service);
 286         if (service == NULL)
 287                 RETURN(NULL);
 288
 289         /* First initialise enough for early teardown */
 290
 291         service->srv_name = name;
 292         spin_lock_init(&service->srv_lock);
 293         CFS_INIT_LIST_HEAD(&service->srv_threads);
 294         cfs_waitq_init(&service->srv_waitq);
 295
 296         service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
 297         service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
 298         service->srv_buf_size = bufsize;
 299         service->srv_rep_portal = rep_portal;
 300         service->srv_req_portal = req_portal;
 301         service->srv_watchdog_timeout = watchdog_timeout;
 302         service->srv_handler = handler;
 303         service->srv_request_history_print_fn = svcreq_printfn;
 304         service->srv_request_seq = 1;           /* valid seq #s start at 1 */
 305         service->srv_request_max_cull_seq = 0;
 306         service->srv_num_threads = num_threads;
 307         service->srv_ctx_tags = ctx_tags;
 308
 309         rc = LNetSetLazyPortal(service->srv_req_portal);
 310         LASSERT (rc == 0);
 311
 312         CFS_INIT_LIST_HEAD(&service->srv_request_queue);
 313         CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
 314         CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
 315         CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
 316         CFS_INIT_LIST_HEAD(&service->srv_request_history);
 317         CFS_INIT_LIST_HEAD(&service->srv_active_replies);
 318         CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
 319         CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
 320         cfs_waitq_init(&service->srv_free_rs_waitq);
 321
 322         spin_lock (&ptlrpc_all_services_lock);
 323         list_add (&service->srv_list, &ptlrpc_all_services);
 324         spin_unlock (&ptlrpc_all_services_lock);
 325
 326         /* Now allocate the request buffers */
 327         rc = ptlrpc_grow_req_bufs(service);
 328         /* We shouldn't be under memory pressure at startup, so
 329          * fail if we can't post all our buffers at this time. */
 330         if (rc != 0)
 331                 GOTO(failed, NULL);
 332
 333         /* Now allocate pool of reply buffers */
 334         /* Increase max reply size to next power of two */
 335         service->srv_max_reply_size = 1;
 336         while (service->srv_max_reply_size <
 337                max_reply_size + SPTLRPC_MAX_PAYLOAD)
 338                 service->srv_max_reply_size <<= 1;
 339
 340         if (proc_entry != NULL)
 341                 ptlrpc_lprocfs_register_service(proc_entry, service);
 342
 343         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 344                service->srv_name, service->srv_req_portal);
 345
 346         RETURN(service);
 347 failed:
 348         ptlrpc_unregister_service(service);
 349         return NULL;
 350 }
 351
 352 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
 353 {
 354         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 355
 356         list_del(&req->rq_list);
 357
 358         if (req->rq_reply_state != NULL) {
 359                 ptlrpc_rs_decref(req->rq_reply_state);
 360                 req->rq_reply_state = NULL;
 361         }
 362
 363         sptlrpc_svc_ctx_decref(req);
 364
 365         if (req != &rqbd->rqbd_req) {
 366                 /* NB request buffers use an embedded
 367                  * req if the incoming req unlinked the
 368                  * MD; this isn't one of them! */
 369                 OBD_FREE(req, sizeof(*req));
 370         }
 371 }
 372
 373 static void
 374 ptlrpc_server_free_request(struct ptlrpc_request *req)
 375 {
 376         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 377         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 378         int                                refcount;
 379         struct list_head                  *tmp;
 380         struct list_head                  *nxt;
 381
 382         spin_lock(&svc->srv_lock);
 383
 384         svc->srv_n_active_reqs--;
 385         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 386
 387         refcount = --(rqbd->rqbd_refcount);
 388         if (refcount == 0) {
 389                 /* request buffer is now idle: add to history */
 390                 list_del(&rqbd->rqbd_list);
 391                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 392                 svc->srv_n_history_rqbds++;
 393
 394                 /* cull some history?
 395                  * I expect only about 1 or 2 rqbds need to be recycled here */
 396                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 397                         rqbd = list_entry(svc->srv_history_rqbds.next,
 398                                           struct ptlrpc_request_buffer_desc,
 399                                           rqbd_list);
 400
 401                         list_del(&rqbd->rqbd_list);
 402                         svc->srv_n_history_rqbds--;
 403
 404                         /* remove rqbd's reqs from svc's req history while
 405                          * I've got the service lock */
 406                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 407                                 req = list_entry(tmp, struct ptlrpc_request,
 408                                                  rq_list);
 409                                 /* Track the highest culled req seq */
 410                                 if (req->rq_history_seq >
 411                                     svc->srv_request_max_cull_seq)
 412                                         svc->srv_request_max_cull_seq =
 413                                                 req->rq_history_seq;
 414                                 list_del(&req->rq_history_list);
 415                         }
 416
 417                         spin_unlock(&svc->srv_lock);
 418
 419                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 420                                 req = list_entry(rqbd->rqbd_reqs.next,
 421                                                  struct ptlrpc_request,
 422                                                  rq_list);
 423                                 __ptlrpc_server_free_request(req);
 424                         }
 425
 426                         spin_lock(&svc->srv_lock);
 427
 428                         /* schedule request buffer for re-use.
 429                          * NB I can only do this after I've disposed of their
 430                          * reqs; particularly the embedded req */
 431                         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 432                 }
 433         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 434                  /* If we are low on memory, we are not interested in
 435                     history */
 436                 list_del(&req->rq_history_list);
 437                 __ptlrpc_server_free_request(req);
 438         }
 439
 440         spin_unlock(&svc->srv_lock);
 441
 442 }
 443
 444 /* This function makes sure dead exports are evicted in a timely manner.
 445    This function is only called when some export receives a message (i.e.,
 446    the network is up.) */
 447 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 448 {
 449         struct obd_export *oldest_exp;
 450         time_t oldest_time;
 451
 452         ENTRY;
 453
 454         LASSERT(exp);
 455
 456         /* Compensate for slow machines, etc, by faking our request time
 457            into the future.  Although this can break the strict time-ordering
 458            of the list, we can be really lazy here - we don't have to evict
 459            at the exact right moment.  Eventually, all silent exports
 460            will make it to the top of the list. */
 461         exp->exp_last_request_time = max(exp->exp_last_request_time,
 462                                          (time_t)CURRENT_SECONDS + extra_delay);
 463
 464         CDEBUG(D_INFO, "updating export %s at %ld\n",
 465                exp->exp_client_uuid.uuid,
 466                exp->exp_last_request_time);
 467
 468         /* exports may get disconnected from the chain even though the
 469            export has references, so we must keep the spin lock while
 470            manipulating the lists */
 471         spin_lock(&exp->exp_obd->obd_dev_lock);
 472
 473         if (list_empty(&exp->exp_obd_chain_timed)) {
 474                 /* this one is not timed */
 475                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 476                 EXIT;
 477                 return;
 478         }
 479
 480         list_move_tail(&exp->exp_obd_chain_timed,
 481                        &exp->exp_obd->obd_exports_timed);
 482
 483         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 484                                 struct obd_export, exp_obd_chain_timed);
 485         oldest_time = oldest_exp->exp_last_request_time;
 486         spin_unlock(&exp->exp_obd->obd_dev_lock);
 487
 488         if (exp->exp_obd->obd_recovering) {
 489                 /* be nice to everyone during recovery */
 490                 EXIT;
 491                 return;
 492         }
 493
 494         /* Note - racing to start/reset the obd_eviction timer is safe */
 495         if (exp->exp_obd->obd_eviction_timer == 0) {
 496                 /* Check if the oldest entry is expired. */
 497                 if (CURRENT_SECONDS > (oldest_time +
 498                                        (3 * obd_timeout / 2) + extra_delay)) {
 499                         /* We need a second timer, in case the net was down and
 500                          * it just came back. Since the pinger may skip every
 501                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 502                          * we better wait for 3. */
 503                         exp->exp_obd->obd_eviction_timer = CURRENT_SECONDS +
 504                                 3 * PING_INTERVAL;
 505                         CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
 506                                exp->exp_obd->obd_name, obd_export_nid2str(exp),
 507                                oldest_time);
 508                 }
 509         } else {
 510                 if (CURRENT_SECONDS > (exp->exp_obd->obd_eviction_timer +
 511                                        extra_delay)) {
 512                         /* The evictor won't evict anyone who we've heard from
 513                          * recently, so we don't have to check before we start
 514                          * it. */
 515                         if (!ping_evictor_wake(exp))
 516                                 exp->exp_obd->obd_eviction_timer = 0;
 517                 }
 518         }
 519
 520         EXIT;
 521 }
 522
 523 #ifndef __KERNEL__
 524 int lu_context_init(struct lu_context *ctx, __u32 tags)
 525 {
 526         return 0;
 527 }
 528
 529 void lu_context_fini(struct lu_context *ctx)
 530 {
 531 }
 532
 533 void lu_context_enter(struct lu_context *ctx)
 534 {
 535 }
 536
 537 void lu_context_exit(struct lu_context *ctx)
 538 {
 539 }
 540
 541 #endif
 542
 543 static int
 544 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
 545                              struct ptlrpc_thread *thread)
 546 {
 547         struct ptlrpc_request *request;
 548         struct timeval         work_start;
 549         struct timeval         work_end;
 550         long                   timediff;
 551         int                    rc;
 552         ENTRY;
 553
 554         LASSERT(svc);
 555
 556         spin_lock(&svc->srv_lock);
 557         if (list_empty (&svc->srv_request_queue) ||
 558             (svc->srv_n_difficult_replies != 0 &&
 559              svc->srv_n_active_reqs >= (svc->srv_nthreads - 1))) {
 560                 /* If all the other threads are handling requests, I must
 561                  * remain free to handle any 'difficult' reply that might
 562                  * block them */
 563                 spin_unlock(&svc->srv_lock);
 564                 RETURN(0);
 565         }
 566
 567         request = list_entry (svc->srv_request_queue.next,
 568                               struct ptlrpc_request, rq_list);
 569         list_del_init (&request->rq_list);
 570         svc->srv_n_queued_reqs--;
 571         svc->srv_n_active_reqs++;
 572
 573         spin_unlock(&svc->srv_lock);
 574
 575         do_gettimeofday(&work_start);
 576         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
 577         if (svc->srv_stats != NULL) {
 578                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 579                                     timediff);
 580                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 581                                     svc->srv_n_queued_reqs);
 582                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 583                                     svc->srv_n_active_reqs);
 584         }
 585
 586         /* go through security check/transform */
 587         request->rq_auth_uid = INVALID_UID;
 588         request->rq_auth_mapped_uid = INVALID_UID;
 589
 590         rc = sptlrpc_svc_unwrap_request(request);
 591         switch (rc) {
 592         case SECSVC_OK:
 593                 break;
 594         case SECSVC_COMPLETE:
 595                 target_send_reply(request, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 596                 goto put_conn;
 597         case SECSVC_DROP:
 598                 goto out;
 599         default:
 600                 LBUG();
 601         }
 602
 603 #if SWAB_PARANOIA
 604         /* Clear request swab mask; this is a new request */
 605         request->rq_req_swab_mask = 0;
 606 #endif
 607         rc = lustre_unpack_msg(request->rq_reqmsg, request->rq_reqlen);
 608         if (rc != 0) {
 609                 CERROR ("error unpacking request: ptl %d from %s"
 610                         " xid "LPU64"\n", svc->srv_req_portal,
 611                         libcfs_id2str(request->rq_peer), request->rq_xid);
 612                 goto out;
 613         }
 614
 615         rc = lustre_unpack_ptlrpc_body(request->rq_reqmsg);
 616         if (rc) {
 617                 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
 618                         " xid "LPU64"\n", svc->srv_req_portal,
 619                         libcfs_id2str(request->rq_peer), request->rq_xid);
 620                 goto out;
 621         }
 622
 623         rc = -EINVAL;
 624         if (lustre_msg_get_type(request->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 625                 CERROR("wrong packet type received (type=%u) from %s\n",
 626                        lustre_msg_get_type(request->rq_reqmsg),
 627                        libcfs_id2str(request->rq_peer));
 628                 goto out;
 629         }
 630
 631         rc = lu_context_init(&request->rq_session, LCT_SESSION);
 632         if (rc) {
 633                 CERROR("Failure to initialize session: %d\n", rc);
 634                 goto out;
 635         }
 636         request->rq_session.lc_thread = thread;
 637         lu_context_enter(&request->rq_session);
 638
 639         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 640
 641         request->rq_svc_thread = thread;
 642         request->rq_svc_thread->t_env->le_ses = &request->rq_session;
 643
 644         request->rq_export = class_conn2export(
 645                                      lustre_msg_get_handle(request->rq_reqmsg));
 646
 647         if (request->rq_export) {
 648                 if (lustre_msg_get_conn_cnt(request->rq_reqmsg) <
 649                     request->rq_export->exp_conn_cnt) {
 650                         DEBUG_REQ(D_ERROR, request,
 651                                   "DROPPING req from old connection %d < %d",
 652                                   lustre_msg_get_conn_cnt(request->rq_reqmsg),
 653                                   request->rq_export->exp_conn_cnt);
 654                         goto put_conn;
 655                 }
 656                 if (request->rq_export->exp_obd &&
 657                     request->rq_export->exp_obd->obd_fail) {
 658                         /* Failing over, don't handle any more reqs, send
 659                            error response instead. */
 660                         CDEBUG(D_HA, "Dropping req %p for failed obd %s\n",
 661                                request, request->rq_export->exp_obd->obd_name);
 662                         request->rq_status = -ENODEV;
 663                         ptlrpc_error(request);
 664                         goto put_conn;
 665                 }
 666
 667                 ptlrpc_update_export_timer(request->rq_export, timediff/500000);
 668         }
 669
 670         /* Discard requests queued for longer than my timeout.  If the
 671          * client's timeout is similar to mine, she'll be timing out this
 672          * REQ anyway (bug 1502) */
 673         if (timediff / 1000000 > (long)obd_timeout) {
 674                 CERROR("Dropping timed-out opc %d request from %s"
 675                        ": %ld seconds old\n",
 676                        lustre_msg_get_opc(request->rq_reqmsg),
 677                        libcfs_id2str(request->rq_peer),
 678                        timediff / 1000000);
 679                 goto put_conn;
 680         }
 681
 682         request->rq_phase = RQ_PHASE_INTERPRET;
 683
 684         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
 685                "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 686                (request->rq_export ?
 687                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 688                (request->rq_export ?
 689                 atomic_read(&request->rq_export->exp_refcount) : -99),
 690                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
 691                libcfs_id2str(request->rq_peer),
 692                lustre_msg_get_opc(request->rq_reqmsg));
 693
 694         rc = svc->srv_handler(request);
 695
 696         request->rq_phase = RQ_PHASE_COMPLETE;
 697
 698         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
 699                "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 700                (request->rq_export ?
 701                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 702                (request->rq_export ?
 703                 atomic_read(&request->rq_export->exp_refcount) : -99),
 704                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
 705                libcfs_id2str(request->rq_peer),
 706                lustre_msg_get_opc(request->rq_reqmsg));
 707
 708 put_conn:
 709         if (request->rq_export != NULL)
 710                 class_export_put(request->rq_export);
 711
 712  out:
 713         lu_context_exit(&request->rq_session);
 714         lu_context_fini(&request->rq_session);
 715
 716         do_gettimeofday(&work_end);
 717
 718         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 719
 720         if (timediff / 1000000 > (long)obd_timeout)
 721                 CERROR("request "LPU64" opc %u from %s processed in %lds "
 722                        "trans "LPU64" rc %d/%d\n",
 723                        request->rq_xid,
 724                        request->rq_reqmsg ?
 725                                 lustre_msg_get_opc(request->rq_reqmsg) : 0,
 726                        libcfs_id2str(request->rq_peer),
 727                        cfs_timeval_sub(&work_end, &request->rq_arrival_time,
 728                                        NULL) / 1000000,
 729                        request->rq_repmsg ?
 730                                 lustre_msg_get_transno(request->rq_repmsg) :
 731                                 request->rq_transno, request->rq_status,
 732                        request->rq_repmsg ?
 733                                 lustre_msg_get_status(request->rq_repmsg) :
 734                                  -999);
 735         else
 736                 CDEBUG(D_HA, "request "LPU64" opc %u from %s processed in "
 737                        "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
 738                        request->rq_xid,
 739                        request->rq_reqmsg ?
 740                                 lustre_msg_get_opc(request->rq_reqmsg) : 0,
 741                        libcfs_id2str(request->rq_peer), timediff,
 742                        cfs_timeval_sub(&work_end, &request->rq_arrival_time,
 743                                        NULL),
 744                        request->rq_transno, request->rq_status,
 745                        request->rq_repmsg ?
 746                                 lustre_msg_get_status(request->rq_repmsg) :
 747                                 -999);
 748
 749         if (svc->srv_stats != NULL && request->rq_reqmsg != NULL) {
 750                 int opc = opcode_offset(lustre_msg_get_opc(request->rq_reqmsg));
 751                 if (opc > 0) {
 752                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 753                         lprocfs_counter_add(svc->srv_stats,
 754                                             opc + PTLRPC_LAST_CNTR,
 755                                             timediff);
 756                 }
 757         }
 758
 759         ptlrpc_server_free_request(request);
 760
 761         RETURN(1);
 762 }
 763
 764 static int
 765 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 766 {
 767         struct ptlrpc_reply_state *rs;
 768         struct obd_export         *exp;
 769         struct obd_device         *obd;
 770         int                        nlocks;
 771         int                        been_handled;
 772         ENTRY;
 773
 774         spin_lock(&svc->srv_lock);
 775         if (list_empty (&svc->srv_reply_queue)) {
 776                 spin_unlock(&svc->srv_lock);
 777                 RETURN(0);
 778         }
 779
 780         rs = list_entry (svc->srv_reply_queue.next,
 781                          struct ptlrpc_reply_state, rs_list);
 782
 783         exp = rs->rs_export;
 784         obd = exp->exp_obd;
 785
 786         LASSERT (rs->rs_difficult);
 787         LASSERT (rs->rs_scheduled);
 788
 789         list_del_init (&rs->rs_list);
 790
 791         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
 792         spin_unlock(&svc->srv_lock);
 793
 794         spin_lock (&obd->obd_uncommitted_replies_lock);
 795         /* Noop if removed already */
 796         list_del_init (&rs->rs_obd_list);
 797         spin_unlock (&obd->obd_uncommitted_replies_lock);
 798
 799         spin_lock (&exp->exp_lock);
 800         /* Noop if removed already */
 801         list_del_init (&rs->rs_exp_list);
 802         spin_unlock (&exp->exp_lock);
 803
 804         spin_lock(&svc->srv_lock);
 805
 806         been_handled = rs->rs_handled;
 807         rs->rs_handled = 1;
 808
 809         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 810         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 811
 812         if (nlocks == 0 && !been_handled) {
 813                 /* If we see this, we should already have seen the warning
 814                  * in mds_steal_ack_locks()  */
 815                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 816                       " o%d NID %s\n",
 817                       rs,
 818                       rs->rs_xid, rs->rs_transno,
 819                       lustre_msg_get_opc(rs->rs_msg),
 820                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 821         }
 822
 823         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
 824                 spin_unlock(&svc->srv_lock);
 825
 826                 if (!been_handled && rs->rs_on_net) {
 827                         LNetMDUnlink(rs->rs_md_h);
 828                         /* Ignore return code; we're racing with
 829                          * completion... */
 830                 }
 831
 832                 while (nlocks-- > 0)
 833                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 834                                          rs->rs_modes[nlocks]);
 835
 836                 spin_lock(&svc->srv_lock);
 837         }
 838
 839         rs->rs_scheduled = 0;
 840
 841         if (!rs->rs_on_net) {
 842                 /* Off the net */
 843                 svc->srv_n_difficult_replies--;
 844                 spin_unlock(&svc->srv_lock);
 845
 846                 class_export_put (exp);
 847                 rs->rs_export = NULL;
 848                 ptlrpc_rs_decref (rs);
 849                 atomic_dec (&svc->srv_outstanding_replies);
 850                 RETURN(1);
 851         }
 852
 853         /* still on the net; callback will schedule */
 854         spin_unlock(&svc->srv_lock);
 855         RETURN(1);
 856 }
 857
 858 #ifndef __KERNEL__
 859 /* FIXME make use of timeout later */
 860 int
 861 liblustre_check_services (void *arg)
 862 {
 863         int  did_something = 0;
 864         int  rc;
 865         struct list_head *tmp, *nxt;
 866         ENTRY;
 867
 868         /* I'm relying on being single threaded, not to have to lock
 869          * ptlrpc_all_services etc */
 870         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 871                 struct ptlrpc_service *svc =
 872                         list_entry (tmp, struct ptlrpc_service, srv_list);
 873
 874                 if (svc->srv_nthreads != 0)     /* I've recursed */
 875                         continue;
 876
 877                 /* service threads can block for bulk, so this limits us
 878                  * (arbitrarily) to recursing 1 stack frame per service.
 879                  * Note that the problem with recursion is that we have to
 880                  * unwind completely before our caller can resume. */
 881
 882                 svc->srv_nthreads++;
 883
 884                 do {
 885                         rc = ptlrpc_server_handle_reply(svc);
 886                         rc |= ptlrpc_server_handle_request(svc, NULL);
 887                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 888                         did_something |= rc;
 889                 } while (rc);
 890
 891                 svc->srv_nthreads--;
 892         }
 893
 894         RETURN(did_something);
 895 }
 896 #define ptlrpc_stop_all_threads(s) do {} while (0)
 897
 898 #else /* __KERNEL__ */
 899
 900 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 901 void ptlrpc_daemonize(char *name)
 902 {
 903         struct fs_struct *fs = current->fs;
 904
 905         atomic_inc(&fs->count);
 906         cfs_daemonize(name);
 907         exit_fs(cfs_current());
 908         current->fs = fs;
 909         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
 910 }
 911
 912 static void
 913 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
 914 {
 915         int avail = svc->srv_nrqbd_receiving;
 916         int low_water = test_req_buffer_pressure ? 0 :
 917                         svc->srv_nbuf_per_group/2;
 918
 919         /* NB I'm not locking; just looking. */
 920
 921         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
 922          * allowed the request history to grow out of control.  We could put a
 923          * sanity check on that here and cull some history if we need the
 924          * space. */
 925
 926         if (avail <= low_water)
 927                 ptlrpc_grow_req_bufs(svc);
 928
 929         if (svc->srv_stats)
 930                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
 931                                     avail);
 932 }
 933
 934 static int
 935 ptlrpc_retry_rqbds(void *arg)
 936 {
 937         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 938
 939         svc->srv_rqbd_timeout = 0;
 940         return (-ETIMEDOUT);
 941 }
 942
 943 static int ptlrpc_main(void *arg)
 944 {
 945         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 946         struct ptlrpc_service  *svc = data->svc;
 947         struct ptlrpc_thread   *thread = data->thread;
 948         struct ptlrpc_reply_state *rs;
 949         struct lc_watchdog     *watchdog;
 950 #ifdef WITH_GROUP_INFO
 951         struct group_info *ginfo = NULL;
 952 #endif
 953         struct lu_env env;
 954         int rc = 0;
 955         ENTRY;
 956
 957         ptlrpc_daemonize(data->name);
 958
 959 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,9) && CONFIG_NUMA
 960         /* we need to do this before any per-thread allocation is done so that
 961          * we get the per-thread allocations on local node.  bug 7342 */
 962         if (svc->srv_cpu_affinity) {
 963                 int cpu, num_cpu;
 964
 965                 for (cpu = 0, num_cpu = 0; cpu < NR_CPUS; cpu++) {
 966                         if (!cpu_online(cpu))
 967                                 continue;
 968                         if (num_cpu == thread->t_id % num_online_cpus())
 969                                 break;
 970                         num_cpu++;
 971                 }
 972                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
 973         }
 974 #endif
 975
 976 #ifdef WITH_GROUP_INFO
 977         ginfo = groups_alloc(0);
 978         if (!ginfo) {
 979                 rc = -ENOMEM;
 980                 goto out;
 981         }
 982
 983         set_current_groups(ginfo);
 984         put_group_info(ginfo);
 985 #endif
 986
 987         if (svc->srv_init != NULL) {
 988                 rc = svc->srv_init(thread);
 989                 if (rc)
 990                         goto out;
 991         }
 992
 993         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
 994         if (rc)
 995                 goto out_srv_init;
 996
 997         thread->t_env = &env;
 998         env.le_ctx.lc_thread = thread;
 999
1000         /* Alloc reply state structure for this one */
1001         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1002         if (!rs) {
1003                 rc = -ENOMEM;
1004                 goto out_srv_init;
1005         }
1006
1007         /* Record that the thread is running */
1008         thread->t_flags = SVC_RUNNING;
1009         /*
1010          * wake up our creator. Note: @data is invalid after this point,
1011          * because it's allocated on ptlrpc_start_thread() stack.
1012          */
1013         cfs_waitq_signal(&thread->t_ctl_waitq);
1014
1015         watchdog = lc_watchdog_add(svc->srv_watchdog_timeout, NULL, NULL);
1016
1017         spin_lock(&svc->srv_lock);
1018         svc->srv_nthreads++;
1019         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1020         spin_unlock(&svc->srv_lock);
1021         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1022
1023         CDEBUG(D_NET, "service thread %d started\n", thread->t_id);
1024
1025         /* XXX maintain a list of all managed devices: insert here */
1026
1027         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1028                svc->srv_n_difficult_replies != 0) {
1029                 /* Don't exit while there are replies to be handled */
1030                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1031                                                      ptlrpc_retry_rqbds, svc);
1032
1033                 lc_watchdog_disable(watchdog);
1034
1035                 l_wait_event_exclusive (svc->srv_waitq,
1036                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1037                                svc->srv_n_difficult_replies == 0) ||
1038                               (!list_empty(&svc->srv_idle_rqbds) &&
1039                                svc->srv_rqbd_timeout == 0) ||
1040                               !list_empty (&svc->srv_reply_queue) ||
1041                               (!list_empty (&svc->srv_request_queue) &&
1042                                (svc->srv_n_difficult_replies == 0 ||
1043                                 svc->srv_n_active_reqs <
1044                                 (svc->srv_nthreads - 1))),
1045                               &lwi);
1046
1047                 lc_watchdog_touch(watchdog);
1048
1049                 ptlrpc_check_rqbd_pool(svc);
1050
1051                 if (!list_empty (&svc->srv_reply_queue))
1052                         ptlrpc_server_handle_reply (svc);
1053
1054                 /* only handle requests if there are no difficult replies
1055                  * outstanding, or I'm not the last thread handling
1056                  * requests */
1057                 if (!list_empty (&svc->srv_request_queue) &&
1058                     (svc->srv_n_difficult_replies == 0 ||
1059                      svc->srv_n_active_reqs < (svc->srv_nthreads - 1))) {
1060                         lu_context_enter(&env.le_ctx);
1061                         ptlrpc_server_handle_request(svc, thread);
1062                         lu_context_exit(&env.le_ctx);
1063                 }
1064
1065                 if (!list_empty(&svc->srv_idle_rqbds) &&
1066                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1067                         /* I just failed to repost request buffers.  Wait
1068                          * for a timeout (unless something else happens)
1069                          * before I try again */
1070                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1071                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1072                                svc->srv_nrqbd_receiving);
1073                 }
1074         }
1075
1076         lc_watchdog_delete(watchdog);
1077
1078 out_srv_init:
1079         /*
1080          * deconstruct service specific state created by ptlrpc_start_thread()
1081          */
1082         if (svc->srv_done != NULL)
1083                 svc->srv_done(thread);
1084
1085 out:
1086         lu_env_fini(&env);
1087
1088         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1089
1090         spin_lock(&svc->srv_lock);
1091         svc->srv_nthreads--;                    /* must know immediately */
1092         thread->t_id = rc;
1093         thread->t_flags = SVC_STOPPED;
1094
1095         cfs_waitq_signal(&thread->t_ctl_waitq);
1096         spin_unlock(&svc->srv_lock);
1097
1098         return rc;
1099 }
1100
1101 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1102                                struct ptlrpc_thread *thread)
1103 {
1104         struct l_wait_info lwi = { 0 };
1105
1106         spin_lock(&svc->srv_lock);
1107         thread->t_flags = SVC_STOPPING;
1108         spin_unlock(&svc->srv_lock);
1109
1110         cfs_waitq_broadcast(&svc->srv_waitq);
1111         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1112                      &lwi);
1113
1114         spin_lock(&svc->srv_lock);
1115         list_del(&thread->t_link);
1116         spin_unlock(&svc->srv_lock);
1117
1118         OBD_FREE_PTR(thread);
1119 }
1120
1121 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1122 {
1123         struct ptlrpc_thread *thread;
1124
1125         spin_lock(&svc->srv_lock);
1126         while (!list_empty(&svc->srv_threads)) {
1127                 thread = list_entry(svc->srv_threads.next,
1128                                     struct ptlrpc_thread, t_link);
1129
1130                 spin_unlock(&svc->srv_lock);
1131                 ptlrpc_stop_thread(svc, thread);
1132                 spin_lock(&svc->srv_lock);
1133         }
1134
1135         spin_unlock(&svc->srv_lock);
1136 }
1137
1138 /* @base_name should be 11 characters or less - 3 will be added on */
1139 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc,
1140                          char *base_name)
1141 {
1142         int i, rc = 0;
1143         ENTRY;
1144
1145         LASSERT(svc->srv_num_threads > 0);
1146         for (i = 0; i < svc->srv_num_threads; i++) {
1147                 char name[32];
1148                 sprintf(name, "%s_%02d", base_name, i);
1149                 rc = ptlrpc_start_thread(dev, svc, name, i);
1150                 if (rc) {
1151                         CERROR("cannot start %s thread #%d: rc %d\n", base_name,
1152                                i, rc);
1153                         ptlrpc_stop_all_threads(svc);
1154                 }
1155         }
1156         RETURN(rc);
1157 }
1158
1159 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc,
1160                         char *name, int id)
1161 {
1162         struct l_wait_info lwi = { 0 };
1163         struct ptlrpc_svc_data d;
1164         struct ptlrpc_thread *thread;
1165         int rc;
1166         ENTRY;
1167
1168         OBD_ALLOC_PTR(thread);
1169         if (thread == NULL)
1170                 RETURN(-ENOMEM);
1171         cfs_waitq_init(&thread->t_ctl_waitq);
1172         thread->t_id = id;
1173
1174         spin_lock(&svc->srv_lock);
1175         list_add(&thread->t_link, &svc->srv_threads);
1176         spin_unlock(&svc->srv_lock);
1177
1178         d.dev = dev;
1179         d.svc = svc;
1180         d.name = name;
1181         d.thread = thread;
1182
1183         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1184          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1185          */
1186         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1187         if (rc < 0) {
1188                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1189
1190                 spin_lock(&svc->srv_lock);
1191                 list_del(&thread->t_link);
1192                 spin_unlock(&svc->srv_lock);
1193
1194                 OBD_FREE(thread, sizeof(*thread));
1195                 RETURN(rc);
1196         }
1197         l_wait_event(thread->t_ctl_waitq,
1198                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1199
1200         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1201         RETURN(rc);
1202 }
1203 #endif
1204
1205 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1206 {
1207         int                   rc;
1208         struct l_wait_info    lwi;
1209         struct list_head     *tmp;
1210         struct ptlrpc_reply_state *rs, *t;
1211
1212         ptlrpc_stop_all_threads(service);
1213         LASSERT(list_empty(&service->srv_threads));
1214
1215         spin_lock (&ptlrpc_all_services_lock);
1216         list_del_init (&service->srv_list);
1217         spin_unlock (&ptlrpc_all_services_lock);
1218
1219         ptlrpc_lprocfs_unregister_service(service);
1220
1221         /* All history will be culled when the next request buffer is
1222          * freed */
1223         service->srv_max_history_rqbds = 0;
1224
1225         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1226
1227         rc = LNetClearLazyPortal(service->srv_req_portal);
1228         LASSERT (rc == 0);
1229
1230         /* Unlink all the request buffers.  This forces a 'final' event with
1231          * its 'unlink' flag set for each posted rqbd */
1232         list_for_each(tmp, &service->srv_active_rqbds) {
1233                 struct ptlrpc_request_buffer_desc *rqbd =
1234                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1235                                    rqbd_list);
1236
1237                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1238                 LASSERT (rc == 0 || rc == -ENOENT);
1239         }
1240
1241         /* Wait for the network to release any buffers it's currently
1242          * filling */
1243         for (;;) {
1244                 spin_lock(&service->srv_lock);
1245                 rc = service->srv_nrqbd_receiving;
1246                 spin_unlock(&service->srv_lock);
1247
1248                 if (rc == 0)
1249                         break;
1250
1251                 /* Network access will complete in finite time but the HUGE
1252                  * timeout lets us CWARN for visibility of sluggish NALs */
1253                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1254                 rc = l_wait_event(service->srv_waitq,
1255                                   service->srv_nrqbd_receiving == 0,
1256                                   &lwi);
1257                 if (rc == -ETIMEDOUT)
1258                         CWARN("Service %s waiting for request buffers\n",
1259                               service->srv_name);
1260         }
1261
1262         /* schedule all outstanding replies to terminate them */
1263         spin_lock(&service->srv_lock);
1264         while (!list_empty(&service->srv_active_replies)) {
1265                 struct ptlrpc_reply_state *rs =
1266                         list_entry(service->srv_active_replies.next,
1267                                    struct ptlrpc_reply_state, rs_list);
1268                 ptlrpc_schedule_difficult_reply(rs);
1269         }
1270         spin_unlock(&service->srv_lock);
1271
1272         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1273          * and no service threads, so I'm the only thread noodling the
1274          * request queue now */
1275         while (!list_empty(&service->srv_request_queue)) {
1276                 struct ptlrpc_request *req =
1277                         list_entry(service->srv_request_queue.next,
1278                                    struct ptlrpc_request,
1279                                    rq_list);
1280
1281                 list_del(&req->rq_list);
1282                 service->srv_n_queued_reqs--;
1283                 service->srv_n_active_reqs++;
1284
1285                 ptlrpc_server_free_request(req);
1286         }
1287         LASSERT(service->srv_n_queued_reqs == 0);
1288         LASSERT(service->srv_n_active_reqs == 0);
1289         LASSERT(service->srv_n_history_rqbds == 0);
1290         LASSERT(list_empty(&service->srv_active_rqbds));
1291
1292         /* Now free all the request buffers since nothing references them
1293          * any more... */
1294         while (!list_empty(&service->srv_idle_rqbds)) {
1295                 struct ptlrpc_request_buffer_desc *rqbd =
1296                         list_entry(service->srv_idle_rqbds.next,
1297                                    struct ptlrpc_request_buffer_desc,
1298                                    rqbd_list);
1299
1300                 ptlrpc_free_rqbd(rqbd);
1301         }
1302
1303         /* wait for all outstanding replies to complete (they were
1304          * scheduled having been flagged to abort above) */
1305         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1306                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1307
1308                 rc = l_wait_event(service->srv_waitq,
1309                                   !list_empty(&service->srv_reply_queue), &lwi);
1310                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1311
1312                 if (rc == 0) {
1313                         ptlrpc_server_handle_reply(service);
1314                         continue;
1315                 }
1316                 CWARN("Unexpectedly long timeout %p\n", service);
1317         }
1318
1319         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1320                 list_del(&rs->rs_list);
1321                 OBD_FREE(rs, service->srv_max_reply_size);
1322         }
1323
1324         OBD_FREE_PTR(service);
1325         return 0;
1326 }
1327
1328 /* Returns 0 if the service is healthy.
1329  *
1330  * Right now, it just checks to make sure that requests aren't languishing
1331  * in the queue.  We'll use this health check to govern whether a node needs
1332  * to be shot, so it's intentionally non-aggressive. */
1333 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1334 {
1335         struct ptlrpc_request *request;
1336         struct timeval         right_now;
1337         long                   timediff, cutoff;
1338         int                    rc = 0;
1339
1340         if (svc == NULL)
1341                 return 0;
1342
1343         spin_lock(&svc->srv_lock);
1344
1345         if (list_empty(&svc->srv_request_queue))
1346                 goto out;
1347
1348         request = list_entry(svc->srv_request_queue.next,
1349                              struct ptlrpc_request, rq_list);
1350
1351         do_gettimeofday(&right_now);
1352         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1353
1354         cutoff = obd_health_check_timeout;
1355
1356         if (timediff / 1000000 > cutoff) {
1357                 rc = -1;
1358                 goto out;
1359         }
1360
1361  out:
1362         spin_unlock(&svc->srv_lock);
1363         return rc;
1364 }