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 #endif
  30 #include <obd_support.h>
  31 #include <obd_class.h>
  32 #include <lustre_net.h>
  33 #include <lu_object.h>
  34 #include <lnet/types.h>
  35 #include "ptlrpc_internal.h"
  36
  37 int test_req_buffer_pressure = 0;
  38 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
  39                 "set non-zero to put pressure on request buffer pools");
  40
  41 /* forward ref */
  42 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
  43
  44 static CFS_LIST_HEAD(ptlrpc_all_services);
  45 spinlock_t ptlrpc_all_services_lock;
  46
  47 static char *
  48 ptlrpc_alloc_request_buffer (int size)
  49 {
  50         char *ptr;
  51
  52         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  53                 OBD_VMALLOC(ptr, size);
  54         else
  55                 OBD_ALLOC(ptr, size);
  56
  57         return (ptr);
  58 }
  59
  60 static void
  61 ptlrpc_free_request_buffer (char *ptr, int size)
  62 {
  63         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  64                 OBD_VFREE(ptr, size);
  65         else
  66                 OBD_FREE(ptr, size);
  67 }
  68
  69 struct ptlrpc_request_buffer_desc *
  70 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
  71 {
  72         struct ptlrpc_request_buffer_desc *rqbd;
  73
  74         OBD_ALLOC_PTR(rqbd);
  75         if (rqbd == NULL)
  76                 return (NULL);
  77
  78         rqbd->rqbd_service = svc;
  79         rqbd->rqbd_refcount = 0;
  80         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
  81         rqbd->rqbd_cbid.cbid_arg = rqbd;
  82         CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
  83         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
  84
  85         if (rqbd->rqbd_buffer == NULL) {
  86                 OBD_FREE_PTR(rqbd);
  87                 return (NULL);
  88         }
  89
  90         spin_lock(&svc->srv_lock);
  91         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
  92         svc->srv_nbufs++;
  93         spin_unlock(&svc->srv_lock);
  94
  95         return (rqbd);
  96 }
  97
  98 void
  99 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 100 {
 101         struct ptlrpc_service *svc = rqbd->rqbd_service;
 102
 103         LASSERT (rqbd->rqbd_refcount == 0);
 104         LASSERT (list_empty(&rqbd->rqbd_reqs));
 105
 106         spin_lock(&svc->srv_lock);
 107         list_del(&rqbd->rqbd_list);
 108         svc->srv_nbufs--;
 109         spin_unlock(&svc->srv_lock);
 110
 111         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 112         OBD_FREE_PTR(rqbd);
 113 }
 114
 115 int
 116 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
 117 {
 118         struct ptlrpc_request_buffer_desc *rqbd;
 119         int                                i;
 120
 121         CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
 122                svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
 123                svc->srv_nrqbd_receiving, svc->srv_nbufs);
 124         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 125                 rqbd = ptlrpc_alloc_rqbd(svc);
 126
 127                 if (rqbd == NULL) {
 128                         CERROR ("%s: Can't allocate request buffer\n",
 129                                 svc->srv_name);
 130                         return (-ENOMEM);
 131                 }
 132
 133                 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
 134                         return (-EAGAIN);
 135         }
 136
 137         return (0);
 138 }
 139
 140 void
 141 ptlrpc_save_lock (struct ptlrpc_request *req,
 142                   struct lustre_handle *lock, int mode)
 143 {
 144         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 145         int                        idx;
 146
 147         LASSERT(rs != NULL);
 148         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
 149
 150         idx = rs->rs_nlocks++;
 151         rs->rs_locks[idx] = *lock;
 152         rs->rs_modes[idx] = mode;
 153         rs->rs_difficult = 1;
 154 }
 155
 156 void
 157 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 158 {
 159         struct ptlrpc_service *svc = rs->rs_service;
 160
 161 #ifdef CONFIG_SMP
 162         LASSERT (spin_is_locked (&svc->srv_lock));
 163 #endif
 164         LASSERT (rs->rs_difficult);
 165         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 166
 167         if (rs->rs_scheduled)                   /* being set up or already notified */
 168                 return;
 169
 170         rs->rs_scheduled = 1;
 171         list_del (&rs->rs_list);
 172         list_add (&rs->rs_list, &svc->srv_reply_queue);
 173         cfs_waitq_signal (&svc->srv_waitq);
 174 }
 175
 176 void
 177 ptlrpc_commit_replies (struct obd_device *obd)
 178 {
 179         struct list_head   *tmp;
 180         struct list_head   *nxt;
 181
 182         /* Find any replies that have been committed and get their service
 183          * to attend to complete them. */
 184
 185         /* CAVEAT EMPTOR: spinlock ordering!!! */
 186         spin_lock(&obd->obd_uncommitted_replies_lock);
 187
 188         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 189                 struct ptlrpc_reply_state *rs =
 190                         list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
 191
 192                 LASSERT (rs->rs_difficult);
 193
 194                 if (rs->rs_transno <= obd->obd_last_committed) {
 195                         struct ptlrpc_service *svc = rs->rs_service;
 196
 197                         spin_lock (&svc->srv_lock);
 198                         list_del_init (&rs->rs_obd_list);
 199                         ptlrpc_schedule_difficult_reply (rs);
 200                         spin_unlock (&svc->srv_lock);
 201                 }
 202         }
 203
 204         spin_unlock(&obd->obd_uncommitted_replies_lock);
 205 }
 206
 207 static int
 208 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 209 {
 210         struct ptlrpc_request_buffer_desc *rqbd;
 211         int                                rc;
 212         int                                posted = 0;
 213
 214         for (;;) {
 215                 spin_lock(&svc->srv_lock);
 216
 217                 if (list_empty (&svc->srv_idle_rqbds)) {
 218                         spin_unlock(&svc->srv_lock);
 219                         return (posted);
 220                 }
 221
 222                 rqbd = list_entry(svc->srv_idle_rqbds.next,
 223                                   struct ptlrpc_request_buffer_desc,
 224                                   rqbd_list);
 225                 list_del (&rqbd->rqbd_list);
 226
 227                 /* assume we will post successfully */
 228                 svc->srv_nrqbd_receiving++;
 229                 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
 230
 231                 spin_unlock(&svc->srv_lock);
 232
 233                 rc = ptlrpc_register_rqbd(rqbd);
 234                 if (rc != 0)
 235                         break;
 236
 237                 posted = 1;
 238         }
 239
 240         spin_lock(&svc->srv_lock);
 241
 242         svc->srv_nrqbd_receiving--;
 243         list_del(&rqbd->rqbd_list);
 244         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 245
 246         /* Don't complain if no request buffers are posted right now; LNET
 247          * won't drop requests because we set the portal lazy! */
 248
 249         spin_unlock(&svc->srv_lock);
 250
 251         return (-1);
 252 }
 253
 254 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
 255                                             svc_handler_t h, char *name,
 256                                             struct proc_dir_entry *proc_entry,
 257                                             svcreq_printfn_t prntfn,
 258                                             char *threadname)
 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_min_threads, c->psc_max_threads,
 266                                threadname, c->psc_ctx_tags);
 267 }
 268 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
 269
 270 /* @threadname should be 11 characters or less - 3 will be added on */
 271 struct ptlrpc_service *
 272 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
 273                 int req_portal, int rep_portal, int watchdog_timeout,
 274                 svc_handler_t handler, char *name,
 275                 cfs_proc_dir_entry_t *proc_entry,
 276                 svcreq_printfn_t svcreq_printfn,
 277                 int min_threads, int max_threads,
 278                 char *threadname, __u32 ctx_tags)
 279 {
 280         int                    rc;
 281         struct ptlrpc_service *service;
 282         ENTRY;
 283
 284         LASSERT (nbufs > 0);
 285         LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
 286         LASSERT (ctx_tags != 0);
 287
 288         OBD_ALLOC_PTR(service);
 289         if (service == NULL)
 290                 RETURN(NULL);
 291
 292         /* First initialise enough for early teardown */
 293
 294         service->srv_name = name;
 295         spin_lock_init(&service->srv_lock);
 296         CFS_INIT_LIST_HEAD(&service->srv_threads);
 297         cfs_waitq_init(&service->srv_waitq);
 298
 299         service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
 300         service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
 301         service->srv_buf_size = bufsize;
 302         service->srv_rep_portal = rep_portal;
 303         service->srv_req_portal = req_portal;
 304         service->srv_watchdog_timeout = watchdog_timeout;
 305         service->srv_handler = handler;
 306         service->srv_request_history_print_fn = svcreq_printfn;
 307         service->srv_request_seq = 1;           /* valid seq #s start at 1 */
 308         service->srv_request_max_cull_seq = 0;
 309         service->srv_threads_min = min_threads;
 310         service->srv_threads_max = max_threads;
 311         service->srv_thread_name = threadname;
 312         service->srv_ctx_tags = ctx_tags;
 313
 314         rc = LNetSetLazyPortal(service->srv_req_portal);
 315         LASSERT (rc == 0);
 316
 317         CFS_INIT_LIST_HEAD(&service->srv_request_queue);
 318         CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
 319         CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
 320         CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
 321         CFS_INIT_LIST_HEAD(&service->srv_request_history);
 322         CFS_INIT_LIST_HEAD(&service->srv_active_replies);
 323         CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
 324         CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
 325         cfs_waitq_init(&service->srv_free_rs_waitq);
 326
 327         spin_lock (&ptlrpc_all_services_lock);
 328         list_add (&service->srv_list, &ptlrpc_all_services);
 329         spin_unlock (&ptlrpc_all_services_lock);
 330
 331         /* Now allocate the request buffers */
 332         rc = ptlrpc_grow_req_bufs(service);
 333         /* We shouldn't be under memory pressure at startup, so
 334          * fail if we can't post all our buffers at this time. */
 335         if (rc != 0)
 336                 GOTO(failed, NULL);
 337
 338         /* Now allocate pool of reply buffers */
 339         /* Increase max reply size to next power of two */
 340         service->srv_max_reply_size = 1;
 341         while (service->srv_max_reply_size <
 342                max_reply_size + SPTLRPC_MAX_PAYLOAD)
 343                 service->srv_max_reply_size <<= 1;
 344
 345         if (proc_entry != NULL)
 346                 ptlrpc_lprocfs_register_service(proc_entry, service);
 347
 348         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 349                service->srv_name, service->srv_req_portal);
 350
 351         RETURN(service);
 352 failed:
 353         ptlrpc_unregister_service(service);
 354         return NULL;
 355 }
 356
 357 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
 358 {
 359         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 360
 361         list_del(&req->rq_list);
 362
 363         if (req->rq_reply_state != NULL) {
 364                 ptlrpc_rs_decref(req->rq_reply_state);
 365                 req->rq_reply_state = NULL;
 366         }
 367
 368         sptlrpc_svc_ctx_decref(req);
 369
 370         if (req != &rqbd->rqbd_req) {
 371                 /* NB request buffers use an embedded
 372                  * req if the incoming req unlinked the
 373                  * MD; this isn't one of them! */
 374                 OBD_FREE(req, sizeof(*req));
 375         }
 376 }
 377
 378 static void
 379 ptlrpc_server_free_request(struct ptlrpc_request *req)
 380 {
 381         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 382         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 383         int                                refcount;
 384         struct list_head                  *tmp;
 385         struct list_head                  *nxt;
 386
 387         spin_lock(&svc->srv_lock);
 388
 389         svc->srv_n_active_reqs--;
 390         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 391
 392         refcount = --(rqbd->rqbd_refcount);
 393         if (refcount == 0) {
 394                 /* request buffer is now idle: add to history */
 395                 list_del(&rqbd->rqbd_list);
 396                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 397                 svc->srv_n_history_rqbds++;
 398
 399                 /* cull some history?
 400                  * I expect only about 1 or 2 rqbds need to be recycled here */
 401                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 402                         rqbd = list_entry(svc->srv_history_rqbds.next,
 403                                           struct ptlrpc_request_buffer_desc,
 404                                           rqbd_list);
 405
 406                         list_del(&rqbd->rqbd_list);
 407                         svc->srv_n_history_rqbds--;
 408
 409                         /* remove rqbd's reqs from svc's req history while
 410                          * I've got the service lock */
 411                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 412                                 req = list_entry(tmp, struct ptlrpc_request,
 413                                                  rq_list);
 414                                 /* Track the highest culled req seq */
 415                                 if (req->rq_history_seq >
 416                                     svc->srv_request_max_cull_seq)
 417                                         svc->srv_request_max_cull_seq =
 418                                                 req->rq_history_seq;
 419                                 list_del(&req->rq_history_list);
 420                         }
 421
 422                         spin_unlock(&svc->srv_lock);
 423
 424                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 425                                 req = list_entry(rqbd->rqbd_reqs.next,
 426                                                  struct ptlrpc_request,
 427                                                  rq_list);
 428                                 __ptlrpc_server_free_request(req);
 429                         }
 430
 431                         spin_lock(&svc->srv_lock);
 432
 433                         /* schedule request buffer for re-use.
 434                          * NB I can only do this after I've disposed of their
 435                          * reqs; particularly the embedded req */
 436                         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 437                 }
 438         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 439                  /* If we are low on memory, we are not interested in
 440                     history */
 441                 list_del(&req->rq_history_list);
 442                 __ptlrpc_server_free_request(req);
 443         }
 444
 445         spin_unlock(&svc->srv_lock);
 446
 447 }
 448
 449 /* This function makes sure dead exports are evicted in a timely manner.
 450    This function is only called when some export receives a message (i.e.,
 451    the network is up.) */
 452 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 453 {
 454         struct obd_export *oldest_exp;
 455         time_t oldest_time;
 456
 457         ENTRY;
 458
 459         LASSERT(exp);
 460
 461         /* Compensate for slow machines, etc, by faking our request time
 462            into the future.  Although this can break the strict time-ordering
 463            of the list, we can be really lazy here - we don't have to evict
 464            at the exact right moment.  Eventually, all silent exports
 465            will make it to the top of the list. */
 466         exp->exp_last_request_time = max(exp->exp_last_request_time,
 467                                          cfs_time_current_sec() + extra_delay);
 468
 469         CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
 470                exp->exp_client_uuid.uuid,
 471                exp->exp_last_request_time, exp);
 472
 473         /* exports may get disconnected from the chain even though the
 474            export has references, so we must keep the spin lock while
 475            manipulating the lists */
 476         spin_lock(&exp->exp_obd->obd_dev_lock);
 477
 478         if (list_empty(&exp->exp_obd_chain_timed)) {
 479                 /* this one is not timed */
 480                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 481                 EXIT;
 482                 return;
 483         }
 484
 485         list_move_tail(&exp->exp_obd_chain_timed,
 486                        &exp->exp_obd->obd_exports_timed);
 487
 488         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 489                                 struct obd_export, exp_obd_chain_timed);
 490         oldest_time = oldest_exp->exp_last_request_time;
 491         spin_unlock(&exp->exp_obd->obd_dev_lock);
 492
 493         if (exp->exp_obd->obd_recovering) {
 494                 /* be nice to everyone during recovery */
 495                 EXIT;
 496                 return;
 497         }
 498
 499         /* Note - racing to start/reset the obd_eviction timer is safe */
 500         if (exp->exp_obd->obd_eviction_timer == 0) {
 501                 /* Check if the oldest entry is expired. */
 502                 if (cfs_time_current_sec() > (oldest_time +
 503                                        (3 * obd_timeout / 2) + extra_delay)) {
 504                         /* We need a second timer, in case the net was down and
 505                          * it just came back. Since the pinger may skip every
 506                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 507                          * we better wait for 3. */
 508                         exp->exp_obd->obd_eviction_timer = cfs_time_current_sec() +
 509                                 3 * PING_INTERVAL;
 510                         CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
 511                                exp->exp_obd->obd_name, obd_export_nid2str(exp),
 512                                oldest_time);
 513                 }
 514         } else {
 515                 if (cfs_time_current_sec() > (exp->exp_obd->obd_eviction_timer +
 516                                        extra_delay)) {
 517                         /* The evictor won't evict anyone who we've heard from
 518                          * recently, so we don't have to check before we start
 519                          * it. */
 520                         if (!ping_evictor_wake(exp))
 521                                 exp->exp_obd->obd_eviction_timer = 0;
 522                 }
 523         }
 524
 525         EXIT;
 526 }
 527
 528 #ifndef __KERNEL__
 529 int lu_context_init(struct lu_context *ctx, __u32 tags)
 530 {
 531         return 0;
 532 }
 533
 534 void lu_context_fini(struct lu_context *ctx)
 535 {
 536 }
 537
 538 void lu_context_enter(struct lu_context *ctx)
 539 {
 540 }
 541
 542 void lu_context_exit(struct lu_context *ctx)
 543 {
 544 }
 545
 546 #endif
 547
 548 static int
 549 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
 550                              struct ptlrpc_thread *thread)
 551 {
 552         struct obd_export     *export = NULL;
 553         struct ptlrpc_request *request;
 554         struct timeval         work_start;
 555         struct timeval         work_end;
 556         long                   timediff;
 557         int                    rc, reply;
 558         ENTRY;
 559
 560         LASSERT(svc);
 561
 562         spin_lock(&svc->srv_lock);
 563         if (unlikely(list_empty (&svc->srv_request_queue) ||
 564                      (svc->srv_n_difficult_replies != 0 &&
 565                       svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
 566                 /* If all the other threads are handling requests, I must
 567                  * remain free to handle any 'difficult' reply that might
 568                  * block them */
 569                 spin_unlock(&svc->srv_lock);
 570                 RETURN(0);
 571         }
 572
 573         request = list_entry (svc->srv_request_queue.next,
 574                               struct ptlrpc_request, rq_list);
 575         list_del_init (&request->rq_list);
 576         svc->srv_n_queued_reqs--;
 577         svc->srv_n_active_reqs++;
 578
 579         spin_unlock(&svc->srv_lock);
 580
 581         do_gettimeofday(&work_start);
 582         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
 583         if (likely(svc->srv_stats != NULL)) {
 584                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 585                                     timediff);
 586                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 587                                     svc->srv_n_queued_reqs);
 588                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 589                                     svc->srv_n_active_reqs);
 590         }
 591
 592         /* go through security check/transform */
 593         rc = sptlrpc_svc_unwrap_request(request);
 594         switch (rc) {
 595         case SECSVC_OK:
 596                 break;
 597         case SECSVC_COMPLETE:
 598                 target_send_reply(request, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 599                 goto out_stat;
 600         case SECSVC_DROP:
 601                 goto out_req;
 602         default:
 603                 LBUG();
 604         }
 605
 606         /* Clear request swab mask; this is a new request */
 607         request->rq_req_swab_mask = 0;
 608
 609         rc = lustre_unpack_msg(request->rq_reqmsg, request->rq_reqlen);
 610         if (rc != 0) {
 611                 CERROR ("error unpacking request: ptl %d from %s"
 612                         " xid "LPU64"\n", svc->srv_req_portal,
 613                         libcfs_id2str(request->rq_peer), request->rq_xid);
 614                 goto out_req;
 615         }
 616
 617         rc = lustre_unpack_req_ptlrpc_body(request, MSG_PTLRPC_BODY_OFF);
 618         if (rc) {
 619                 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
 620                         " xid "LPU64"\n", svc->srv_req_portal,
 621                         libcfs_id2str(request->rq_peer), request->rq_xid);
 622                 goto out_req;
 623         }
 624
 625         rc = -EINVAL;
 626         if (lustre_msg_get_type(request->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 627                 CERROR("wrong packet type received (type=%u) from %s\n",
 628                        lustre_msg_get_type(request->rq_reqmsg),
 629                        libcfs_id2str(request->rq_peer));
 630                 goto out_req;
 631         }
 632
 633         rc = lu_context_init(&request->rq_session, LCT_SESSION);
 634         if (rc) {
 635                 CERROR("Failure to initialize session: %d\n", rc);
 636                 goto out_req;
 637         }
 638         request->rq_session.lc_thread = thread;
 639         lu_context_enter(&request->rq_session);
 640
 641         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 642
 643         request->rq_svc_thread = thread;
 644         if (thread)
 645                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
 646
 647         request->rq_export = class_conn2export(
 648                                      lustre_msg_get_handle(request->rq_reqmsg));
 649
 650         if (likely(request->rq_export)) {
 651                 if (unlikely(lustre_msg_get_conn_cnt(request->rq_reqmsg) <
 652                              request->rq_export->exp_conn_cnt)) {
 653                         DEBUG_REQ(D_ERROR, request,
 654                                   "DROPPING req from old connection %d < %d",
 655                                   lustre_msg_get_conn_cnt(request->rq_reqmsg),
 656                                   request->rq_export->exp_conn_cnt);
 657                         goto put_conn;
 658                 }
 659                 if (unlikely(request->rq_export->exp_obd &&
 660                              request->rq_export->exp_obd->obd_fail)) {
 661                         /* Failing over, don't handle any more reqs, send
 662                            error response instead. */
 663                         CDEBUG(D_RPCTRACE,"Dropping req %p for failed obd %s\n",
 664                                request, request->rq_export->exp_obd->obd_name);
 665                         request->rq_status = -ENODEV;
 666                         ptlrpc_error(request);
 667                         goto put_conn;
 668                 }
 669
 670                 rc = sptlrpc_target_export_check(request->rq_export, request);
 671                 if (unlikely(rc)) {
 672                         DEBUG_REQ(D_ERROR, request,
 673                                   "DROPPING req with illegal security flavor,");
 674                         goto put_conn;
 675                 }
 676
 677                 ptlrpc_update_export_timer(request->rq_export, timediff/500000);
 678                 export = class_export_rpc_get(request->rq_export);
 679         }
 680
 681         /* Discard requests queued for longer than my timeout.  If the
 682          * client's timeout is similar to mine, she'll be timing out this
 683          * REQ anyway (bug 1502) */
 684         if (unlikely(timediff / 1000000 > (long)obd_timeout)) {
 685                 CERROR("Dropping timed-out opc %d request from %s"
 686                        ": %ld seconds old\n",
 687                        lustre_msg_get_opc(request->rq_reqmsg),
 688                        libcfs_id2str(request->rq_peer),
 689                        timediff / 1000000);
 690                 goto put_rpc_export;
 691         }
 692
 693         request->rq_phase = RQ_PHASE_INTERPRET;
 694
 695         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
 696                "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 697                (request->rq_export ?
 698                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 699                (request->rq_export ?
 700                 atomic_read(&request->rq_export->exp_refcount) : -99),
 701                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
 702                libcfs_id2str(request->rq_peer),
 703                lustre_msg_get_opc(request->rq_reqmsg));
 704
 705         rc = svc->srv_handler(request);
 706
 707         request->rq_phase = RQ_PHASE_COMPLETE;
 708
 709         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
 710                "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 711                (request->rq_export ?
 712                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 713                (request->rq_export ?
 714                 atomic_read(&request->rq_export->exp_refcount) : -99),
 715                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
 716                libcfs_id2str(request->rq_peer),
 717                lustre_msg_get_opc(request->rq_reqmsg));
 718
 719 put_rpc_export:
 720         if (export != NULL)
 721                 class_export_rpc_put(export);
 722 put_conn:
 723         if (likely(request->rq_export != NULL))
 724                 class_export_put(request->rq_export);
 725
 726         lu_context_exit(&request->rq_session);
 727         lu_context_fini(&request->rq_session);
 728 out_stat:
 729         reply = request->rq_reply_state && request->rq_repmsg;  /* bug 11169 */
 730
 731         do_gettimeofday(&work_end);
 732
 733         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 734
 735         if (unlikely(timediff / 1000000 > (long)obd_timeout))
 736                 CERROR("request "LPU64" opc %u from %s processed in %lds "
 737                        "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),
 742                        cfs_timeval_sub(&work_end, &request->rq_arrival_time,
 743                                        NULL) / 1000000,
 744                        reply ? lustre_msg_get_transno(request->rq_repmsg) :
 745                                request->rq_transno, request->rq_status,
 746                        reply ? lustre_msg_get_status(request->rq_repmsg) : -999);
 747         else
 748                 CDEBUG(D_RPCTRACE,"request "LPU64" opc %u from %s processed in "
 749                        "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
 750                        request->rq_xid,
 751                        request->rq_reqmsg ?
 752                                 lustre_msg_get_opc(request->rq_reqmsg) : 0,
 753                        libcfs_id2str(request->rq_peer), timediff,
 754                        cfs_timeval_sub(&work_end, &request->rq_arrival_time,
 755                                        NULL),
 756                        request->rq_transno, request->rq_status,
 757                        reply ? lustre_msg_get_status(request->rq_repmsg) : -999);
 758
 759         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
 760                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
 761                 int opc = opcode_offset(op);
 762                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
 763                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 764                         lprocfs_counter_add(svc->srv_stats,
 765                                             opc + EXTRA_MAX_OPCODES,
 766                                             timediff);
 767                 }
 768         }
 769
 770 out_req:
 771         ptlrpc_server_free_request(request);
 772
 773         RETURN(1);
 774 }
 775
 776 static int
 777 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 778 {
 779         struct ptlrpc_reply_state *rs;
 780         struct obd_export         *exp;
 781         struct obd_device         *obd;
 782         int                        nlocks;
 783         int                        been_handled;
 784         ENTRY;
 785
 786         spin_lock(&svc->srv_lock);
 787         if (list_empty (&svc->srv_reply_queue)) {
 788                 spin_unlock(&svc->srv_lock);
 789                 RETURN(0);
 790         }
 791
 792         rs = list_entry (svc->srv_reply_queue.next,
 793                          struct ptlrpc_reply_state, rs_list);
 794
 795         exp = rs->rs_export;
 796         obd = exp->exp_obd;
 797
 798         LASSERT (rs->rs_difficult);
 799         LASSERT (rs->rs_scheduled);
 800
 801         list_del_init (&rs->rs_list);
 802
 803         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
 804         spin_unlock(&svc->srv_lock);
 805
 806         spin_lock (&obd->obd_uncommitted_replies_lock);
 807         /* Noop if removed already */
 808         list_del_init (&rs->rs_obd_list);
 809         spin_unlock (&obd->obd_uncommitted_replies_lock);
 810
 811         spin_lock (&exp->exp_lock);
 812         /* Noop if removed already */
 813         list_del_init (&rs->rs_exp_list);
 814         spin_unlock (&exp->exp_lock);
 815
 816         spin_lock(&svc->srv_lock);
 817
 818         been_handled = rs->rs_handled;
 819         rs->rs_handled = 1;
 820
 821         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 822         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 823
 824         if (nlocks == 0 && !been_handled) {
 825                 /* If we see this, we should already have seen the warning
 826                  * in mds_steal_ack_locks()  */
 827                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 828                       " o%d NID %s\n",
 829                       rs,
 830                       rs->rs_xid, rs->rs_transno,
 831                       lustre_msg_get_opc(rs->rs_msg),
 832                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 833         }
 834
 835         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
 836                 spin_unlock(&svc->srv_lock);
 837
 838                 if (!been_handled && rs->rs_on_net) {
 839                         LNetMDUnlink(rs->rs_md_h);
 840                         /* Ignore return code; we're racing with
 841                          * completion... */
 842                 }
 843
 844                 while (nlocks-- > 0)
 845                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 846                                          rs->rs_modes[nlocks]);
 847
 848                 spin_lock(&svc->srv_lock);
 849         }
 850
 851         rs->rs_scheduled = 0;
 852
 853         if (!rs->rs_on_net) {
 854                 /* Off the net */
 855                 svc->srv_n_difficult_replies--;
 856                 spin_unlock(&svc->srv_lock);
 857
 858                 class_export_put (exp);
 859                 rs->rs_export = NULL;
 860                 ptlrpc_rs_decref (rs);
 861                 atomic_dec (&svc->srv_outstanding_replies);
 862                 RETURN(1);
 863         }
 864
 865         /* still on the net; callback will schedule */
 866         spin_unlock(&svc->srv_lock);
 867         RETURN(1);
 868 }
 869
 870 #ifndef __KERNEL__
 871 /* FIXME make use of timeout later */
 872 int
 873 liblustre_check_services (void *arg)
 874 {
 875         int  did_something = 0;
 876         int  rc;
 877         struct list_head *tmp, *nxt;
 878         ENTRY;
 879
 880         /* I'm relying on being single threaded, not to have to lock
 881          * ptlrpc_all_services etc */
 882         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 883                 struct ptlrpc_service *svc =
 884                         list_entry (tmp, struct ptlrpc_service, srv_list);
 885
 886                 if (svc->srv_threads_running != 0)     /* I've recursed */
 887                         continue;
 888
 889                 /* service threads can block for bulk, so this limits us
 890                  * (arbitrarily) to recursing 1 stack frame per service.
 891                  * Note that the problem with recursion is that we have to
 892                  * unwind completely before our caller can resume. */
 893
 894                 svc->srv_threads_running++;
 895
 896                 do {
 897                         rc = ptlrpc_server_handle_reply(svc);
 898                         rc |= ptlrpc_server_handle_request(svc, NULL);
 899                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 900                         did_something |= rc;
 901                 } while (rc);
 902
 903                 svc->srv_threads_running--;
 904         }
 905
 906         RETURN(did_something);
 907 }
 908 #define ptlrpc_stop_all_threads(s) do {} while (0)
 909
 910 #else /* __KERNEL__ */
 911
 912 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 913 void ptlrpc_daemonize(char *name)
 914 {
 915         struct fs_struct *fs = current->fs;
 916
 917         atomic_inc(&fs->count);
 918         cfs_daemonize(name);
 919         exit_fs(cfs_current());
 920         current->fs = fs;
 921         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
 922 }
 923
 924 static void
 925 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
 926 {
 927         int avail = svc->srv_nrqbd_receiving;
 928         int low_water = test_req_buffer_pressure ? 0 :
 929                         svc->srv_nbuf_per_group/2;
 930
 931         /* NB I'm not locking; just looking. */
 932
 933         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
 934          * allowed the request history to grow out of control.  We could put a
 935          * sanity check on that here and cull some history if we need the
 936          * space. */
 937
 938         if (avail <= low_water)
 939                 ptlrpc_grow_req_bufs(svc);
 940
 941         if (svc->srv_stats)
 942                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
 943                                     avail);
 944 }
 945
 946 static int
 947 ptlrpc_retry_rqbds(void *arg)
 948 {
 949         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 950
 951         svc->srv_rqbd_timeout = 0;
 952         return (-ETIMEDOUT);
 953 }
 954
 955 static int ptlrpc_main(void *arg)
 956 {
 957         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 958         struct ptlrpc_service  *svc = data->svc;
 959         struct ptlrpc_thread   *thread = data->thread;
 960         struct obd_device      *dev = data->dev;
 961         struct ptlrpc_reply_state *rs;
 962         struct lc_watchdog     *watchdog;
 963 #ifdef WITH_GROUP_INFO
 964         struct group_info *ginfo = NULL;
 965 #endif
 966         struct lu_env env;
 967         int rc = 0;
 968         ENTRY;
 969
 970         ptlrpc_daemonize(data->name);
 971
 972 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
 973         /* we need to do this before any per-thread allocation is done so that
 974          * we get the per-thread allocations on local node.  bug 7342 */
 975         if (svc->srv_cpu_affinity) {
 976                 int cpu, num_cpu;
 977
 978                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
 979                         if (!cpu_online(cpu))
 980                                 continue;
 981                         if (num_cpu == thread->t_id % num_online_cpus())
 982                                 break;
 983                         num_cpu++;
 984                 }
 985                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
 986         }
 987 #endif
 988
 989 #ifdef WITH_GROUP_INFO
 990         ginfo = groups_alloc(0);
 991         if (!ginfo) {
 992                 rc = -ENOMEM;
 993                 goto out;
 994         }
 995
 996         set_current_groups(ginfo);
 997         put_group_info(ginfo);
 998 #endif
 999
1000         if (svc->srv_init != NULL) {
1001                 rc = svc->srv_init(thread);
1002                 if (rc)
1003                         goto out;
1004         }
1005
1006         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1007         if (rc)
1008                 goto out_srv_fini;
1009
1010         thread->t_env = &env;
1011         env.le_ctx.lc_thread = thread;
1012
1013         /* Alloc reply state structure for this one */
1014         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1015         if (!rs) {
1016                 rc = -ENOMEM;
1017                 goto out_srv_fini;
1018         }
1019
1020         /* Record that the thread is running */
1021         thread->t_flags = SVC_RUNNING;
1022         /*
1023          * wake up our creator. Note: @data is invalid after this point,
1024          * because it's allocated on ptlrpc_start_thread() stack.
1025          */
1026         cfs_waitq_signal(&thread->t_ctl_waitq);
1027
1028         watchdog = lc_watchdog_add(svc->srv_watchdog_timeout, NULL, NULL);
1029
1030         spin_lock(&svc->srv_lock);
1031         svc->srv_threads_running++;
1032         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1033         spin_unlock(&svc->srv_lock);
1034         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1035
1036         CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1037               svc->srv_threads_running);
1038
1039         /* XXX maintain a list of all managed devices: insert here */
1040
1041         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1042                svc->srv_n_difficult_replies != 0) {
1043                 /* Don't exit while there are replies to be handled */
1044                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1045                                                      ptlrpc_retry_rqbds, svc);
1046
1047                 lc_watchdog_disable(watchdog);
1048
1049                 cond_resched();
1050
1051                 l_wait_event_exclusive (svc->srv_waitq,
1052                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1053                                svc->srv_n_difficult_replies == 0) ||
1054                               (!list_empty(&svc->srv_idle_rqbds) &&
1055                                svc->srv_rqbd_timeout == 0) ||
1056                               !list_empty (&svc->srv_reply_queue) ||
1057                               (!list_empty (&svc->srv_request_queue) &&
1058                                (svc->srv_n_difficult_replies == 0 ||
1059                                 svc->srv_n_active_reqs <
1060                                 (svc->srv_threads_running - 1))),
1061                               &lwi);
1062
1063                 lc_watchdog_touch(watchdog);
1064
1065                 ptlrpc_check_rqbd_pool(svc);
1066
1067                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1068                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1069                         /* Ignore return code - we tried... */
1070                         ptlrpc_start_thread(dev, svc);
1071                 }
1072
1073                 if (!list_empty (&svc->srv_reply_queue))
1074                         ptlrpc_server_handle_reply (svc);
1075
1076                 /* only handle requests if there are no difficult replies
1077                  * outstanding, or I'm not the last thread handling
1078                  * requests */
1079                 if (!list_empty (&svc->srv_request_queue) &&
1080                     (svc->srv_n_difficult_replies == 0 ||
1081                      svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1082                         lu_context_enter(&env.le_ctx);
1083                         ptlrpc_server_handle_request(svc, thread);
1084                         lu_context_exit(&env.le_ctx);
1085                 }
1086
1087                 if (!list_empty(&svc->srv_idle_rqbds) &&
1088                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1089                         /* I just failed to repost request buffers.  Wait
1090                          * for a timeout (unless something else happens)
1091                          * before I try again */
1092                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1093                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1094                                svc->srv_nrqbd_receiving);
1095                 }
1096         }
1097
1098         lc_watchdog_delete(watchdog);
1099
1100 out_srv_fini:
1101         /*
1102          * deconstruct service specific state created by ptlrpc_start_thread()
1103          */
1104         if (svc->srv_done != NULL)
1105                 svc->srv_done(thread);
1106
1107         lu_context_fini(&env.le_ctx);
1108 out:
1109         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1110
1111         spin_lock(&svc->srv_lock);
1112         svc->srv_threads_running--; /* must know immediately */
1113         thread->t_id = rc;
1114         thread->t_flags = SVC_STOPPED;
1115
1116         cfs_waitq_signal(&thread->t_ctl_waitq);
1117         spin_unlock(&svc->srv_lock);
1118
1119         return rc;
1120 }
1121
1122 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1123                                struct ptlrpc_thread *thread)
1124 {
1125         struct l_wait_info lwi = { 0 };
1126
1127         spin_lock(&svc->srv_lock);
1128         thread->t_flags = SVC_STOPPING;
1129         spin_unlock(&svc->srv_lock);
1130
1131         cfs_waitq_broadcast(&svc->srv_waitq);
1132         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1133                      &lwi);
1134
1135         spin_lock(&svc->srv_lock);
1136         list_del(&thread->t_link);
1137         spin_unlock(&svc->srv_lock);
1138
1139         OBD_FREE_PTR(thread);
1140 }
1141
1142 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1143 {
1144         struct ptlrpc_thread *thread;
1145
1146         spin_lock(&svc->srv_lock);
1147         while (!list_empty(&svc->srv_threads)) {
1148                 thread = list_entry(svc->srv_threads.next,
1149                                     struct ptlrpc_thread, t_link);
1150
1151                 spin_unlock(&svc->srv_lock);
1152                 ptlrpc_stop_thread(svc, thread);
1153                 spin_lock(&svc->srv_lock);
1154         }
1155
1156         spin_unlock(&svc->srv_lock);
1157 }
1158
1159 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1160 {
1161         int i, rc = 0;
1162         ENTRY;
1163
1164         LASSERT(svc->srv_threads_min > 0);
1165         for (i = 0; i < svc->srv_threads_min; i++) {
1166                 rc = ptlrpc_start_thread(dev, svc);
1167                 /* We have enough threads, don't start more.  b=15759 */
1168                 if (rc == -EMFILE)
1169                         break;
1170                 if (rc) {
1171                         CERROR("cannot start %s thread #%d: rc %d\n",
1172                                svc->srv_thread_name, i, rc);
1173                         ptlrpc_stop_all_threads(svc);
1174                 }
1175         }
1176         RETURN(rc);
1177 }
1178
1179 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1180 {
1181         struct l_wait_info lwi = { 0 };
1182         struct ptlrpc_svc_data d;
1183         struct ptlrpc_thread *thread;
1184         char name[32];
1185         int id, rc;
1186         ENTRY;
1187
1188         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1189                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1190                svc->srv_threads_max, svc->srv_threads_running);
1191         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1192             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1193              svc->srv_threads_started == svc->srv_threads_min - 1))
1194                 RETURN(-EMFILE);
1195
1196         OBD_ALLOC_PTR(thread);
1197         if (thread == NULL)
1198                 RETURN(-ENOMEM);
1199         cfs_waitq_init(&thread->t_ctl_waitq);
1200
1201         spin_lock(&svc->srv_lock);
1202         if (svc->srv_threads_started >= svc->srv_threads_max) {
1203                 spin_unlock(&svc->srv_lock);
1204                 OBD_FREE_PTR(thread);
1205                 RETURN(-EMFILE);
1206         }
1207         list_add(&thread->t_link, &svc->srv_threads);
1208         id = svc->srv_threads_started++;
1209         spin_unlock(&svc->srv_lock);
1210
1211         thread->t_id = id;
1212         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1213         d.dev = dev;
1214         d.svc = svc;
1215         d.name = name;
1216         d.thread = thread;
1217
1218         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1219
1220           /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1221          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1222          */
1223         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1224         if (rc < 0) {
1225                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1226
1227                 spin_lock(&svc->srv_lock);
1228                 list_del(&thread->t_link);
1229                 --svc->srv_threads_started;
1230                 spin_unlock(&svc->srv_lock);
1231
1232                 OBD_FREE(thread, sizeof(*thread));
1233                 RETURN(rc);
1234         }
1235         l_wait_event(thread->t_ctl_waitq,
1236                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1237
1238         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1239         RETURN(rc);
1240 }
1241 #endif
1242
1243 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1244 {
1245         int                   rc;
1246         struct l_wait_info    lwi;
1247         struct list_head     *tmp;
1248         struct ptlrpc_reply_state *rs, *t;
1249
1250         ptlrpc_stop_all_threads(service);
1251         LASSERT(list_empty(&service->srv_threads));
1252
1253         spin_lock (&ptlrpc_all_services_lock);
1254         list_del_init (&service->srv_list);
1255         spin_unlock (&ptlrpc_all_services_lock);
1256
1257         ptlrpc_lprocfs_unregister_service(service);
1258
1259         /* All history will be culled when the next request buffer is
1260          * freed */
1261         service->srv_max_history_rqbds = 0;
1262
1263         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1264
1265         rc = LNetClearLazyPortal(service->srv_req_portal);
1266         LASSERT (rc == 0);
1267
1268         /* Unlink all the request buffers.  This forces a 'final' event with
1269          * its 'unlink' flag set for each posted rqbd */
1270         list_for_each(tmp, &service->srv_active_rqbds) {
1271                 struct ptlrpc_request_buffer_desc *rqbd =
1272                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1273                                    rqbd_list);
1274
1275                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1276                 LASSERT (rc == 0 || rc == -ENOENT);
1277         }
1278
1279         /* Wait for the network to release any buffers it's currently
1280          * filling */
1281         for (;;) {
1282                 spin_lock(&service->srv_lock);
1283                 rc = service->srv_nrqbd_receiving;
1284                 spin_unlock(&service->srv_lock);
1285
1286                 if (rc == 0)
1287                         break;
1288
1289                 /* Network access will complete in finite time but the HUGE
1290                  * timeout lets us CWARN for visibility of sluggish NALs */
1291                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1292                 rc = l_wait_event(service->srv_waitq,
1293                                   service->srv_nrqbd_receiving == 0,
1294                                   &lwi);
1295                 if (rc == -ETIMEDOUT)
1296                         CWARN("Service %s waiting for request buffers\n",
1297                               service->srv_name);
1298         }
1299
1300         /* schedule all outstanding replies to terminate them */
1301         spin_lock(&service->srv_lock);
1302         while (!list_empty(&service->srv_active_replies)) {
1303                 struct ptlrpc_reply_state *rs =
1304                         list_entry(service->srv_active_replies.next,
1305                                    struct ptlrpc_reply_state, rs_list);
1306                 ptlrpc_schedule_difficult_reply(rs);
1307         }
1308         spin_unlock(&service->srv_lock);
1309
1310         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1311          * and no service threads, so I'm the only thread noodling the
1312          * request queue now */
1313         while (!list_empty(&service->srv_request_queue)) {
1314                 struct ptlrpc_request *req =
1315                         list_entry(service->srv_request_queue.next,
1316                                    struct ptlrpc_request,
1317                                    rq_list);
1318
1319                 list_del(&req->rq_list);
1320                 service->srv_n_queued_reqs--;
1321                 service->srv_n_active_reqs++;
1322
1323                 ptlrpc_server_free_request(req);
1324         }
1325         LASSERT(service->srv_n_queued_reqs == 0);
1326         LASSERT(service->srv_n_active_reqs == 0);
1327         LASSERT(service->srv_n_history_rqbds == 0);
1328         LASSERT(list_empty(&service->srv_active_rqbds));
1329
1330         /* Now free all the request buffers since nothing references them
1331          * any more... */
1332         while (!list_empty(&service->srv_idle_rqbds)) {
1333                 struct ptlrpc_request_buffer_desc *rqbd =
1334                         list_entry(service->srv_idle_rqbds.next,
1335                                    struct ptlrpc_request_buffer_desc,
1336                                    rqbd_list);
1337
1338                 ptlrpc_free_rqbd(rqbd);
1339         }
1340
1341         /* wait for all outstanding replies to complete (they were
1342          * scheduled having been flagged to abort above) */
1343         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1344                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1345
1346                 rc = l_wait_event(service->srv_waitq,
1347                                   !list_empty(&service->srv_reply_queue), &lwi);
1348                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1349
1350                 if (rc == 0) {
1351                         ptlrpc_server_handle_reply(service);
1352                         continue;
1353                 }
1354                 CWARN("Unexpectedly long timeout %p\n", service);
1355         }
1356
1357         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1358                 list_del(&rs->rs_list);
1359                 OBD_FREE(rs, service->srv_max_reply_size);
1360         }
1361
1362         OBD_FREE_PTR(service);
1363         return 0;
1364 }
1365
1366 /* Returns 0 if the service is healthy.
1367  *
1368  * Right now, it just checks to make sure that requests aren't languishing
1369  * in the queue.  We'll use this health check to govern whether a node needs
1370  * to be shot, so it's intentionally non-aggressive. */
1371 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1372 {
1373         struct ptlrpc_request *request;
1374         struct timeval         right_now;
1375         long                   timediff, cutoff;
1376         int                    rc = 0;
1377
1378         if (svc == NULL)
1379                 return 0;
1380
1381         spin_lock(&svc->srv_lock);
1382
1383         if (list_empty(&svc->srv_request_queue))
1384                 goto out;
1385
1386         request = list_entry(svc->srv_request_queue.next,
1387                              struct ptlrpc_request, rq_list);
1388
1389         do_gettimeofday(&right_now);
1390         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1391
1392         cutoff = obd_health_check_timeout;
1393
1394         if (timediff / 1000000 > cutoff) {
1395                 rc = -1;
1396                 goto out;
1397         }
1398
1399  out:
1400         spin_unlock(&svc->srv_lock);
1401         return rc;
1402 }