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                                             char *threadname)
 260 {
 261         return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
 262                                c->psc_max_req_size, c->psc_max_reply_size,
 263                                c->psc_req_portal, c->psc_rep_portal,
 264                                c->psc_watchdog_timeout,
 265                                h, name, proc_entry,
 266                                prntfn, c->psc_min_threads, c->psc_max_threads,
 267                                threadname, c->psc_ctx_tags);
 268 }
 269 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
 270
 271 /* @threadname should be 11 characters or less - 3 will be added on */
 272 struct ptlrpc_service *
 273 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
 274                 int req_portal, int rep_portal, int watchdog_timeout,
 275                 svc_handler_t handler, char *name,
 276                 cfs_proc_dir_entry_t *proc_entry,
 277                 svcreq_printfn_t svcreq_printfn,
 278                 int min_threads, int max_threads,
 279                 char *threadname, __u32 ctx_tags)
 280 {
 281         int                    rc;
 282         struct ptlrpc_service *service;
 283         ENTRY;
 284
 285         LASSERT (nbufs > 0);
 286         LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
 287         LASSERT (ctx_tags != 0);
 288
 289         OBD_ALLOC_PTR(service);
 290         if (service == NULL)
 291                 RETURN(NULL);
 292
 293         /* First initialise enough for early teardown */
 294
 295         service->srv_name = name;
 296         spin_lock_init(&service->srv_lock);
 297         CFS_INIT_LIST_HEAD(&service->srv_threads);
 298         cfs_waitq_init(&service->srv_waitq);
 299
 300         service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
 301         service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
 302         service->srv_buf_size = bufsize;
 303         service->srv_rep_portal = rep_portal;
 304         service->srv_req_portal = req_portal;
 305         service->srv_watchdog_timeout = watchdog_timeout;
 306         service->srv_handler = handler;
 307         service->srv_request_history_print_fn = svcreq_printfn;
 308         service->srv_request_seq = 1;           /* valid seq #s start at 1 */
 309         service->srv_request_max_cull_seq = 0;
 310         service->srv_threads_min = min_threads;
 311         service->srv_threads_max = max_threads;
 312         service->srv_thread_name = threadname;
 313         service->srv_ctx_tags = ctx_tags;
 314
 315         rc = LNetSetLazyPortal(service->srv_req_portal);
 316         LASSERT (rc == 0);
 317
 318         CFS_INIT_LIST_HEAD(&service->srv_request_queue);
 319         CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
 320         CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
 321         CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
 322         CFS_INIT_LIST_HEAD(&service->srv_request_history);
 323         CFS_INIT_LIST_HEAD(&service->srv_active_replies);
 324         CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
 325         CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
 326         cfs_waitq_init(&service->srv_free_rs_waitq);
 327
 328         spin_lock (&ptlrpc_all_services_lock);
 329         list_add (&service->srv_list, &ptlrpc_all_services);
 330         spin_unlock (&ptlrpc_all_services_lock);
 331
 332         /* Now allocate the request buffers */
 333         rc = ptlrpc_grow_req_bufs(service);
 334         /* We shouldn't be under memory pressure at startup, so
 335          * fail if we can't post all our buffers at this time. */
 336         if (rc != 0)
 337                 GOTO(failed, NULL);
 338
 339         /* Now allocate pool of reply buffers */
 340         /* Increase max reply size to next power of two */
 341         service->srv_max_reply_size = 1;
 342         while (service->srv_max_reply_size <
 343                max_reply_size + SPTLRPC_MAX_PAYLOAD)
 344                 service->srv_max_reply_size <<= 1;
 345
 346         if (proc_entry != NULL)
 347                 ptlrpc_lprocfs_register_service(proc_entry, service);
 348
 349         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 350                service->srv_name, service->srv_req_portal);
 351
 352         RETURN(service);
 353 failed:
 354         ptlrpc_unregister_service(service);
 355         return NULL;
 356 }
 357
 358 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
 359 {
 360         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 361
 362         list_del(&req->rq_list);
 363
 364         if (req->rq_reply_state != NULL) {
 365                 ptlrpc_rs_decref(req->rq_reply_state);
 366                 req->rq_reply_state = NULL;
 367         }
 368
 369         sptlrpc_svc_ctx_decref(req);
 370
 371         if (req != &rqbd->rqbd_req) {
 372                 /* NB request buffers use an embedded
 373                  * req if the incoming req unlinked the
 374                  * MD; this isn't one of them! */
 375                 OBD_FREE(req, sizeof(*req));
 376         }
 377 }
 378
 379 static void
 380 ptlrpc_server_free_request(struct ptlrpc_request *req)
 381 {
 382         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 383         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 384         int                                refcount;
 385         struct list_head                  *tmp;
 386         struct list_head                  *nxt;
 387
 388         spin_lock(&svc->srv_lock);
 389
 390         svc->srv_n_active_reqs--;
 391         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 392
 393         refcount = --(rqbd->rqbd_refcount);
 394         if (refcount == 0) {
 395                 /* request buffer is now idle: add to history */
 396                 list_del(&rqbd->rqbd_list);
 397                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 398                 svc->srv_n_history_rqbds++;
 399
 400                 /* cull some history?
 401                  * I expect only about 1 or 2 rqbds need to be recycled here */
 402                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 403                         rqbd = list_entry(svc->srv_history_rqbds.next,
 404                                           struct ptlrpc_request_buffer_desc,
 405                                           rqbd_list);
 406
 407                         list_del(&rqbd->rqbd_list);
 408                         svc->srv_n_history_rqbds--;
 409
 410                         /* remove rqbd's reqs from svc's req history while
 411                          * I've got the service lock */
 412                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 413                                 req = list_entry(tmp, struct ptlrpc_request,
 414                                                  rq_list);
 415                                 /* Track the highest culled req seq */
 416                                 if (req->rq_history_seq >
 417                                     svc->srv_request_max_cull_seq)
 418                                         svc->srv_request_max_cull_seq =
 419                                                 req->rq_history_seq;
 420                                 list_del(&req->rq_history_list);
 421                         }
 422
 423                         spin_unlock(&svc->srv_lock);
 424
 425                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 426                                 req = list_entry(rqbd->rqbd_reqs.next,
 427                                                  struct ptlrpc_request,
 428                                                  rq_list);
 429                                 __ptlrpc_server_free_request(req);
 430                         }
 431
 432                         spin_lock(&svc->srv_lock);
 433
 434                         /* schedule request buffer for re-use.
 435                          * NB I can only do this after I've disposed of their
 436                          * reqs; particularly the embedded req */
 437                         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 438                 }
 439         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 440                  /* If we are low on memory, we are not interested in
 441                     history */
 442                 list_del(&req->rq_history_list);
 443                 __ptlrpc_server_free_request(req);
 444         }
 445
 446         spin_unlock(&svc->srv_lock);
 447
 448 }
 449
 450 /* This function makes sure dead exports are evicted in a timely manner.
 451    This function is only called when some export receives a message (i.e.,
 452    the network is up.) */
 453 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 454 {
 455         struct obd_export *oldest_exp;
 456         time_t oldest_time;
 457
 458         ENTRY;
 459
 460         LASSERT(exp);
 461
 462         /* Compensate for slow machines, etc, by faking our request time
 463            into the future.  Although this can break the strict time-ordering
 464            of the list, we can be really lazy here - we don't have to evict
 465            at the exact right moment.  Eventually, all silent exports
 466            will make it to the top of the list. */
 467         exp->exp_last_request_time = max(exp->exp_last_request_time,
 468                                          (time_t)CURRENT_SECONDS + extra_delay);
 469
 470         CDEBUG(D_HA, "updating export %s at %ld exp %p\n",
 471                exp->exp_client_uuid.uuid,
 472                exp->exp_last_request_time, exp);
 473
 474         /* exports may get disconnected from the chain even though the
 475            export has references, so we must keep the spin lock while
 476            manipulating the lists */
 477         spin_lock(&exp->exp_obd->obd_dev_lock);
 478
 479         if (list_empty(&exp->exp_obd_chain_timed)) {
 480                 /* this one is not timed */
 481                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 482                 EXIT;
 483                 return;
 484         }
 485
 486         list_move_tail(&exp->exp_obd_chain_timed,
 487                        &exp->exp_obd->obd_exports_timed);
 488
 489         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 490                                 struct obd_export, exp_obd_chain_timed);
 491         oldest_time = oldest_exp->exp_last_request_time;
 492         spin_unlock(&exp->exp_obd->obd_dev_lock);
 493
 494         if (exp->exp_obd->obd_recovering) {
 495                 /* be nice to everyone during recovery */
 496                 EXIT;
 497                 return;
 498         }
 499
 500         /* Note - racing to start/reset the obd_eviction timer is safe */
 501         if (exp->exp_obd->obd_eviction_timer == 0) {
 502                 /* Check if the oldest entry is expired. */
 503                 if (CURRENT_SECONDS > (oldest_time +
 504                                        (3 * obd_timeout / 2) + extra_delay)) {
 505                         /* We need a second timer, in case the net was down and
 506                          * it just came back. Since the pinger may skip every
 507                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 508                          * we better wait for 3. */
 509                         exp->exp_obd->obd_eviction_timer = CURRENT_SECONDS +
 510                                 3 * PING_INTERVAL;
 511                         CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
 512                                exp->exp_obd->obd_name, obd_export_nid2str(exp),
 513                                oldest_time);
 514                 }
 515         } else {
 516                 if (CURRENT_SECONDS > (exp->exp_obd->obd_eviction_timer +
 517                                        extra_delay)) {
 518                         /* The evictor won't evict anyone who we've heard from
 519                          * recently, so we don't have to check before we start
 520                          * it. */
 521                         if (!ping_evictor_wake(exp))
 522                                 exp->exp_obd->obd_eviction_timer = 0;
 523                 }
 524         }
 525
 526         EXIT;
 527 }
 528
 529 #ifndef __KERNEL__
 530 int lu_context_init(struct lu_context *ctx, __u32 tags)
 531 {
 532         return 0;
 533 }
 534
 535 void lu_context_fini(struct lu_context *ctx)
 536 {
 537 }
 538
 539 void lu_context_enter(struct lu_context *ctx)
 540 {
 541 }
 542
 543 void lu_context_exit(struct lu_context *ctx)
 544 {
 545 }
 546
 547 #endif
 548
 549 static int
 550 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
 551                              struct ptlrpc_thread *thread)
 552 {
 553         struct obd_export     *export = NULL;
 554         struct ptlrpc_request *request;
 555         struct timeval         work_start;
 556         struct timeval         work_end;
 557         long                   timediff;
 558         int                    rc, reply;
 559         ENTRY;
 560
 561         LASSERT(svc);
 562
 563         spin_lock(&svc->srv_lock);
 564         if (unlikely(list_empty (&svc->srv_request_queue) ||
 565                      (svc->srv_n_difficult_replies != 0 &&
 566                       svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
 567                 /* If all the other threads are handling requests, I must
 568                  * remain free to handle any 'difficult' reply that might
 569                  * block them */
 570                 spin_unlock(&svc->srv_lock);
 571                 RETURN(0);
 572         }
 573
 574         request = list_entry (svc->srv_request_queue.next,
 575                               struct ptlrpc_request, rq_list);
 576         list_del_init (&request->rq_list);
 577         svc->srv_n_queued_reqs--;
 578         svc->srv_n_active_reqs++;
 579
 580         spin_unlock(&svc->srv_lock);
 581
 582         do_gettimeofday(&work_start);
 583         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
 584         if (likely(svc->srv_stats != NULL)) {
 585                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
 586                                     timediff);
 587                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
 588                                     svc->srv_n_queued_reqs);
 589                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
 590                                     svc->srv_n_active_reqs);
 591         }
 592
 593         /* go through security check/transform */
 594         request->rq_auth_uid = INVALID_UID;
 595         request->rq_auth_mapped_uid = INVALID_UID;
 596
 597         rc = sptlrpc_svc_unwrap_request(request);
 598         switch (rc) {
 599         case SECSVC_OK:
 600                 break;
 601         case SECSVC_COMPLETE:
 602                 target_send_reply(request, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 603                 goto put_conn;
 604         case SECSVC_DROP:
 605                 goto out_req;
 606         default:
 607                 LBUG();
 608         }
 609
 610 #if SWAB_PARANOIA
 611         /* Clear request swab mask; this is a new request */
 612         request->rq_req_swab_mask = 0;
 613 #endif
 614         rc = lustre_unpack_msg(request->rq_reqmsg, request->rq_reqlen);
 615         if (rc != 0) {
 616                 CERROR ("error unpacking request: ptl %d from %s"
 617                         " xid "LPU64"\n", svc->srv_req_portal,
 618                         libcfs_id2str(request->rq_peer), request->rq_xid);
 619                 goto out_req;
 620         }
 621
 622         rc = lustre_unpack_req_ptlrpc_body(request, MSG_PTLRPC_BODY_OFF);
 623         if (rc) {
 624                 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
 625                         " xid "LPU64"\n", svc->srv_req_portal,
 626                         libcfs_id2str(request->rq_peer), request->rq_xid);
 627                 goto out_req;
 628         }
 629
 630         rc = -EINVAL;
 631         if (lustre_msg_get_type(request->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 632                 CERROR("wrong packet type received (type=%u) from %s\n",
 633                        lustre_msg_get_type(request->rq_reqmsg),
 634                        libcfs_id2str(request->rq_peer));
 635                 goto out_req;
 636         }
 637
 638         rc = lu_context_init(&request->rq_session, LCT_SESSION);
 639         if (rc) {
 640                 CERROR("Failure to initialize session: %d\n", rc);
 641                 goto out_req;
 642         }
 643         request->rq_session.lc_thread = thread;
 644         lu_context_enter(&request->rq_session);
 645
 646         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
 647
 648         request->rq_svc_thread = thread;
 649         if (thread)
 650                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
 651
 652         request->rq_export = class_conn2export(
 653                                      lustre_msg_get_handle(request->rq_reqmsg));
 654
 655         if (likely(request->rq_export)) {
 656                 if (unlikely(lustre_msg_get_conn_cnt(request->rq_reqmsg) <
 657                              request->rq_export->exp_conn_cnt)) {
 658                         DEBUG_REQ(D_ERROR, request,
 659                                   "DROPPING req from old connection %d < %d",
 660                                   lustre_msg_get_conn_cnt(request->rq_reqmsg),
 661                                   request->rq_export->exp_conn_cnt);
 662                         goto put_conn;
 663                 }
 664                 if (unlikely(request->rq_export->exp_obd &&
 665                              request->rq_export->exp_obd->obd_fail)) {
 666                         /* Failing over, don't handle any more reqs, send
 667                            error response instead. */
 668                         CDEBUG(D_HA, "Dropping req %p for failed obd %s\n",
 669                                request, request->rq_export->exp_obd->obd_name);
 670                         request->rq_status = -ENODEV;
 671                         ptlrpc_error(request);
 672                         goto put_conn;
 673                 }
 674
 675                 rc = sptlrpc_target_export_check(request->rq_export, request);
 676                 if (unlikely(rc)) {
 677                         DEBUG_REQ(D_ERROR, request,
 678                                   "DROPPING req with illeagle security flavor");
 679                         goto put_conn;
 680                 }
 681
 682                 ptlrpc_update_export_timer(request->rq_export, timediff/500000);
 683                 export = class_export_rpc_get(request->rq_export);
 684         }
 685
 686         /* Discard requests queued for longer than my timeout.  If the
 687          * client's timeout is similar to mine, she'll be timing out this
 688          * REQ anyway (bug 1502) */
 689         if (unlikely(timediff / 1000000 > (long)obd_timeout)) {
 690                 CERROR("Dropping timed-out opc %d request from %s"
 691                        ": %ld seconds old\n",
 692                        lustre_msg_get_opc(request->rq_reqmsg),
 693                        libcfs_id2str(request->rq_peer),
 694                        timediff / 1000000);
 695                 goto put_rpc_export;
 696         }
 697
 698         request->rq_phase = RQ_PHASE_INTERPRET;
 699
 700         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
 701                "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 702                (request->rq_export ?
 703                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 704                (request->rq_export ?
 705                 atomic_read(&request->rq_export->exp_refcount) : -99),
 706                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
 707                libcfs_id2str(request->rq_peer),
 708                lustre_msg_get_opc(request->rq_reqmsg));
 709
 710         rc = svc->srv_handler(request);
 711
 712         request->rq_phase = RQ_PHASE_COMPLETE;
 713
 714         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
 715                "%s:%s+%d:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
 716                (request->rq_export ?
 717                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
 718                (request->rq_export ?
 719                 atomic_read(&request->rq_export->exp_refcount) : -99),
 720                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
 721                libcfs_id2str(request->rq_peer),
 722                lustre_msg_get_opc(request->rq_reqmsg));
 723
 724 put_rpc_export:
 725         if (export != NULL)
 726                 class_export_rpc_put(export);
 727 put_conn:
 728         if (likely(request->rq_export != NULL))
 729                 class_export_put(request->rq_export);
 730
 731         lu_context_exit(&request->rq_session);
 732         lu_context_fini(&request->rq_session);
 733
 734         reply = request->rq_reply_state && request->rq_repmsg;  /* bug 11169 */
 735
 736         do_gettimeofday(&work_end);
 737
 738         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 739
 740         if (unlikely(timediff / 1000000 > (long)obd_timeout))
 741                 CERROR("request "LPU64" opc %u from %s processed in %lds "
 742                        "trans "LPU64" rc %d/%d\n",
 743                        request->rq_xid,
 744                        request->rq_reqmsg ?
 745                                 lustre_msg_get_opc(request->rq_reqmsg) : 0,
 746                        libcfs_id2str(request->rq_peer),
 747                        cfs_timeval_sub(&work_end, &request->rq_arrival_time,
 748                                        NULL) / 1000000,
 749                        reply ? lustre_msg_get_transno(request->rq_repmsg) :
 750                                request->rq_transno, request->rq_status,
 751                        reply ? lustre_msg_get_status(request->rq_repmsg) : -999);
 752         else
 753                 CDEBUG(D_HA, "request "LPU64" opc %u from %s processed in "
 754                        "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
 755                        request->rq_xid,
 756                        request->rq_reqmsg ?
 757                                 lustre_msg_get_opc(request->rq_reqmsg) : 0,
 758                        libcfs_id2str(request->rq_peer), timediff,
 759                        cfs_timeval_sub(&work_end, &request->rq_arrival_time,
 760                                        NULL),
 761                        request->rq_transno, request->rq_status,
 762                        reply ? lustre_msg_get_status(request->rq_repmsg) : -999);
 763
 764         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
 765                 int opc = opcode_offset(lustre_msg_get_opc(request->rq_reqmsg));
 766                 if (opc > 0) {
 767                         LASSERT(opc < LUSTRE_MAX_OPCODES);
 768                         lprocfs_counter_add(svc->srv_stats,
 769                                             opc + PTLRPC_LAST_CNTR,
 770                                             timediff);
 771                 }
 772         }
 773
 774 out_req:
 775         ptlrpc_server_free_request(request);
 776
 777         RETURN(1);
 778 }
 779
 780 static int
 781 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
 782 {
 783         struct ptlrpc_reply_state *rs;
 784         struct obd_export         *exp;
 785         struct obd_device         *obd;
 786         int                        nlocks;
 787         int                        been_handled;
 788         ENTRY;
 789
 790         spin_lock(&svc->srv_lock);
 791         if (list_empty (&svc->srv_reply_queue)) {
 792                 spin_unlock(&svc->srv_lock);
 793                 RETURN(0);
 794         }
 795
 796         rs = list_entry (svc->srv_reply_queue.next,
 797                          struct ptlrpc_reply_state, rs_list);
 798
 799         exp = rs->rs_export;
 800         obd = exp->exp_obd;
 801
 802         LASSERT (rs->rs_difficult);
 803         LASSERT (rs->rs_scheduled);
 804
 805         list_del_init (&rs->rs_list);
 806
 807         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
 808         spin_unlock(&svc->srv_lock);
 809
 810         spin_lock (&obd->obd_uncommitted_replies_lock);
 811         /* Noop if removed already */
 812         list_del_init (&rs->rs_obd_list);
 813         spin_unlock (&obd->obd_uncommitted_replies_lock);
 814
 815         spin_lock (&exp->exp_lock);
 816         /* Noop if removed already */
 817         list_del_init (&rs->rs_exp_list);
 818         spin_unlock (&exp->exp_lock);
 819
 820         spin_lock(&svc->srv_lock);
 821
 822         been_handled = rs->rs_handled;
 823         rs->rs_handled = 1;
 824
 825         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
 826         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
 827
 828         if (nlocks == 0 && !been_handled) {
 829                 /* If we see this, we should already have seen the warning
 830                  * in mds_steal_ack_locks()  */
 831                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
 832                       " o%d NID %s\n",
 833                       rs,
 834                       rs->rs_xid, rs->rs_transno,
 835                       lustre_msg_get_opc(rs->rs_msg),
 836                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 837         }
 838
 839         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
 840                 spin_unlock(&svc->srv_lock);
 841
 842                 if (!been_handled && rs->rs_on_net) {
 843                         LNetMDUnlink(rs->rs_md_h);
 844                         /* Ignore return code; we're racing with
 845                          * completion... */
 846                 }
 847
 848                 while (nlocks-- > 0)
 849                         ldlm_lock_decref(&rs->rs_locks[nlocks],
 850                                          rs->rs_modes[nlocks]);
 851
 852                 spin_lock(&svc->srv_lock);
 853         }
 854
 855         rs->rs_scheduled = 0;
 856
 857         if (!rs->rs_on_net) {
 858                 /* Off the net */
 859                 svc->srv_n_difficult_replies--;
 860                 spin_unlock(&svc->srv_lock);
 861
 862                 class_export_put (exp);
 863                 rs->rs_export = NULL;
 864                 ptlrpc_rs_decref (rs);
 865                 atomic_dec (&svc->srv_outstanding_replies);
 866                 RETURN(1);
 867         }
 868
 869         /* still on the net; callback will schedule */
 870         spin_unlock(&svc->srv_lock);
 871         RETURN(1);
 872 }
 873
 874 #ifndef __KERNEL__
 875 /* FIXME make use of timeout later */
 876 int
 877 liblustre_check_services (void *arg)
 878 {
 879         int  did_something = 0;
 880         int  rc;
 881         struct list_head *tmp, *nxt;
 882         ENTRY;
 883
 884         /* I'm relying on being single threaded, not to have to lock
 885          * ptlrpc_all_services etc */
 886         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
 887                 struct ptlrpc_service *svc =
 888                         list_entry (tmp, struct ptlrpc_service, srv_list);
 889
 890                 if (svc->srv_threads_running != 0)     /* I've recursed */
 891                         continue;
 892
 893                 /* service threads can block for bulk, so this limits us
 894                  * (arbitrarily) to recursing 1 stack frame per service.
 895                  * Note that the problem with recursion is that we have to
 896                  * unwind completely before our caller can resume. */
 897
 898                 svc->srv_threads_running++;
 899
 900                 do {
 901                         rc = ptlrpc_server_handle_reply(svc);
 902                         rc |= ptlrpc_server_handle_request(svc, NULL);
 903                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
 904                         did_something |= rc;
 905                 } while (rc);
 906
 907                 svc->srv_threads_running--;
 908         }
 909
 910         RETURN(did_something);
 911 }
 912 #define ptlrpc_stop_all_threads(s) do {} while (0)
 913
 914 #else /* __KERNEL__ */
 915
 916 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
 917 void ptlrpc_daemonize(char *name)
 918 {
 919         struct fs_struct *fs = current->fs;
 920
 921         atomic_inc(&fs->count);
 922         cfs_daemonize(name);
 923         exit_fs(cfs_current());
 924         current->fs = fs;
 925         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
 926 }
 927
 928 static void
 929 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
 930 {
 931         int avail = svc->srv_nrqbd_receiving;
 932         int low_water = test_req_buffer_pressure ? 0 :
 933                         svc->srv_nbuf_per_group/2;
 934
 935         /* NB I'm not locking; just looking. */
 936
 937         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
 938          * allowed the request history to grow out of control.  We could put a
 939          * sanity check on that here and cull some history if we need the
 940          * space. */
 941
 942         if (avail <= low_water)
 943                 ptlrpc_grow_req_bufs(svc);
 944
 945         if (svc->srv_stats)
 946                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
 947                                     avail);
 948 }
 949
 950 static int
 951 ptlrpc_retry_rqbds(void *arg)
 952 {
 953         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
 954
 955         svc->srv_rqbd_timeout = 0;
 956         return (-ETIMEDOUT);
 957 }
 958
 959 static int ptlrpc_main(void *arg)
 960 {
 961         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
 962         struct ptlrpc_service  *svc = data->svc;
 963         struct ptlrpc_thread   *thread = data->thread;
 964         struct obd_device      *dev = data->dev;
 965         struct ptlrpc_reply_state *rs;
 966         struct lc_watchdog     *watchdog;
 967 #ifdef WITH_GROUP_INFO
 968         struct group_info *ginfo = NULL;
 969 #endif
 970         struct lu_env env;
 971         int rc = 0;
 972         ENTRY;
 973
 974         ptlrpc_daemonize(data->name);
 975
 976 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
 977         /* we need to do this before any per-thread allocation is done so that
 978          * we get the per-thread allocations on local node.  bug 7342 */
 979         if (svc->srv_cpu_affinity) {
 980                 int cpu, num_cpu;
 981
 982                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
 983                         if (!cpu_online(cpu))
 984                                 continue;
 985                         if (num_cpu == thread->t_id % num_online_cpus())
 986                                 break;
 987                         num_cpu++;
 988                 }
 989                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
 990         }
 991 #endif
 992
 993 #ifdef WITH_GROUP_INFO
 994         ginfo = groups_alloc(0);
 995         if (!ginfo) {
 996                 rc = -ENOMEM;
 997                 goto out;
 998         }
 999
1000         set_current_groups(ginfo);
1001         put_group_info(ginfo);
1002 #endif
1003
1004         if (svc->srv_init != NULL) {
1005                 rc = svc->srv_init(thread);
1006                 if (rc)
1007                         goto out;
1008         }
1009
1010         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1011         if (rc)
1012                 goto out_srv_fini;
1013
1014         thread->t_env = &env;
1015         env.le_ctx.lc_thread = thread;
1016
1017         /* Alloc reply state structure for this one */
1018         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1019         if (!rs) {
1020                 rc = -ENOMEM;
1021                 goto out_srv_fini;
1022         }
1023
1024         /* Record that the thread is running */
1025         thread->t_flags = SVC_RUNNING;
1026         /*
1027          * wake up our creator. Note: @data is invalid after this point,
1028          * because it's allocated on ptlrpc_start_thread() stack.
1029          */
1030         cfs_waitq_signal(&thread->t_ctl_waitq);
1031
1032         watchdog = lc_watchdog_add(svc->srv_watchdog_timeout, NULL, NULL);
1033
1034         spin_lock(&svc->srv_lock);
1035         svc->srv_threads_running++;
1036         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1037         spin_unlock(&svc->srv_lock);
1038         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1039
1040         CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1041               svc->srv_threads_running);
1042
1043         /* XXX maintain a list of all managed devices: insert here */
1044
1045         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1046                svc->srv_n_difficult_replies != 0) {
1047                 /* Don't exit while there are replies to be handled */
1048                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1049                                                      ptlrpc_retry_rqbds, svc);
1050
1051                 lc_watchdog_disable(watchdog);
1052
1053                 cond_resched();
1054
1055                 l_wait_event_exclusive (svc->srv_waitq,
1056                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1057                                svc->srv_n_difficult_replies == 0) ||
1058                               (!list_empty(&svc->srv_idle_rqbds) &&
1059                                svc->srv_rqbd_timeout == 0) ||
1060                               !list_empty (&svc->srv_reply_queue) ||
1061                               (!list_empty (&svc->srv_request_queue) &&
1062                                (svc->srv_n_difficult_replies == 0 ||
1063                                 svc->srv_n_active_reqs <
1064                                 (svc->srv_threads_running - 1))),
1065                               &lwi);
1066
1067                 lc_watchdog_touch(watchdog);
1068
1069                 ptlrpc_check_rqbd_pool(svc);
1070
1071                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1072                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1073                         /* Ignore return code - we tried... */
1074                         ptlrpc_start_thread(dev, svc);
1075                 }
1076
1077                 if (!list_empty (&svc->srv_reply_queue))
1078                         ptlrpc_server_handle_reply (svc);
1079
1080                 /* only handle requests if there are no difficult replies
1081                  * outstanding, or I'm not the last thread handling
1082                  * requests */
1083                 if (!list_empty (&svc->srv_request_queue) &&
1084                     (svc->srv_n_difficult_replies == 0 ||
1085                      svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1086                         lu_context_enter(&env.le_ctx);
1087                         ptlrpc_server_handle_request(svc, thread);
1088                         lu_context_exit(&env.le_ctx);
1089                 }
1090
1091                 if (!list_empty(&svc->srv_idle_rqbds) &&
1092                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1093                         /* I just failed to repost request buffers.  Wait
1094                          * for a timeout (unless something else happens)
1095                          * before I try again */
1096                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1097                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1098                                svc->srv_nrqbd_receiving);
1099                 }
1100         }
1101
1102         lc_watchdog_delete(watchdog);
1103
1104 out_srv_fini:
1105         /*
1106          * deconstruct service specific state created by ptlrpc_start_thread()
1107          */
1108         if (svc->srv_done != NULL)
1109                 svc->srv_done(thread);
1110
1111         lu_env_fini(&env);
1112 out:
1113         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1114
1115         spin_lock(&svc->srv_lock);
1116         svc->srv_threads_running--; /* must know immediately */
1117         thread->t_id = rc;
1118         thread->t_flags = SVC_STOPPED;
1119
1120         cfs_waitq_signal(&thread->t_ctl_waitq);
1121         spin_unlock(&svc->srv_lock);
1122
1123         return rc;
1124 }
1125
1126 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1127                                struct ptlrpc_thread *thread)
1128 {
1129         struct l_wait_info lwi = { 0 };
1130
1131         spin_lock(&svc->srv_lock);
1132         thread->t_flags = SVC_STOPPING;
1133         spin_unlock(&svc->srv_lock);
1134
1135         cfs_waitq_broadcast(&svc->srv_waitq);
1136         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1137                      &lwi);
1138
1139         spin_lock(&svc->srv_lock);
1140         list_del(&thread->t_link);
1141         spin_unlock(&svc->srv_lock);
1142
1143         OBD_FREE_PTR(thread);
1144 }
1145
1146 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1147 {
1148         struct ptlrpc_thread *thread;
1149
1150         spin_lock(&svc->srv_lock);
1151         while (!list_empty(&svc->srv_threads)) {
1152                 thread = list_entry(svc->srv_threads.next,
1153                                     struct ptlrpc_thread, t_link);
1154
1155                 spin_unlock(&svc->srv_lock);
1156                 ptlrpc_stop_thread(svc, thread);
1157                 spin_lock(&svc->srv_lock);
1158         }
1159
1160         spin_unlock(&svc->srv_lock);
1161 }
1162
1163 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1164 {
1165         int i, rc = 0;
1166         ENTRY;
1167
1168         LASSERT(svc->srv_threads_min > 0);
1169         for (i = 0; i < svc->srv_threads_min; i++) {
1170                 rc = ptlrpc_start_thread(dev, svc);
1171                 if (rc) {
1172                         CERROR("cannot start %s thread #%d: rc %d\n",
1173                                svc->srv_thread_name, i, rc);
1174                         ptlrpc_stop_all_threads(svc);
1175                 }
1176         }
1177         RETURN(rc);
1178 }
1179
1180 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1181 {
1182         struct l_wait_info lwi = { 0 };
1183         struct ptlrpc_svc_data d;
1184         struct ptlrpc_thread *thread;
1185         char name[32];
1186         int id, rc;
1187         ENTRY;
1188
1189         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1190                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1191                svc->srv_threads_max, svc->srv_threads_running);
1192         if (svc->srv_threads_started >= svc->srv_threads_max)
1193                 RETURN(-EMFILE);
1194
1195         OBD_ALLOC_PTR(thread);
1196         if (thread == NULL)
1197                 RETURN(-ENOMEM);
1198         cfs_waitq_init(&thread->t_ctl_waitq);
1199
1200         spin_lock(&svc->srv_lock);
1201         if (svc->srv_threads_started >= svc->srv_threads_max) {
1202                 spin_unlock(&svc->srv_lock);
1203                 OBD_FREE_PTR(thread);
1204                 RETURN(-EMFILE);
1205         }
1206         list_add(&thread->t_link, &svc->srv_threads);
1207         id = svc->srv_threads_started++;
1208         spin_unlock(&svc->srv_lock);
1209
1210         thread->t_id = id;
1211         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1212         d.dev = dev;
1213         d.svc = svc;
1214         d.name = name;
1215         d.thread = thread;
1216
1217         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1218
1219           /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1220          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1221          */
1222         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1223         if (rc < 0) {
1224                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1225
1226                 spin_lock(&svc->srv_lock);
1227                 list_del(&thread->t_link);
1228                 --svc->srv_threads_started;
1229                 spin_unlock(&svc->srv_lock);
1230
1231                 OBD_FREE(thread, sizeof(*thread));
1232                 RETURN(rc);
1233         }
1234         l_wait_event(thread->t_ctl_waitq,
1235                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1236
1237         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1238         RETURN(rc);
1239 }
1240 #endif
1241
1242 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1243 {
1244         int                   rc;
1245         struct l_wait_info    lwi;
1246         struct list_head     *tmp;
1247         struct ptlrpc_reply_state *rs, *t;
1248
1249         ptlrpc_stop_all_threads(service);
1250         LASSERT(list_empty(&service->srv_threads));
1251
1252         spin_lock (&ptlrpc_all_services_lock);
1253         list_del_init (&service->srv_list);
1254         spin_unlock (&ptlrpc_all_services_lock);
1255
1256         ptlrpc_lprocfs_unregister_service(service);
1257
1258         /* All history will be culled when the next request buffer is
1259          * freed */
1260         service->srv_max_history_rqbds = 0;
1261
1262         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1263
1264         rc = LNetClearLazyPortal(service->srv_req_portal);
1265         LASSERT (rc == 0);
1266
1267         /* Unlink all the request buffers.  This forces a 'final' event with
1268          * its 'unlink' flag set for each posted rqbd */
1269         list_for_each(tmp, &service->srv_active_rqbds) {
1270                 struct ptlrpc_request_buffer_desc *rqbd =
1271                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1272                                    rqbd_list);
1273
1274                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1275                 LASSERT (rc == 0 || rc == -ENOENT);
1276         }
1277
1278         /* Wait for the network to release any buffers it's currently
1279          * filling */
1280         for (;;) {
1281                 spin_lock(&service->srv_lock);
1282                 rc = service->srv_nrqbd_receiving;
1283                 spin_unlock(&service->srv_lock);
1284
1285                 if (rc == 0)
1286                         break;
1287
1288                 /* Network access will complete in finite time but the HUGE
1289                  * timeout lets us CWARN for visibility of sluggish NALs */
1290                 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1291                 rc = l_wait_event(service->srv_waitq,
1292                                   service->srv_nrqbd_receiving == 0,
1293                                   &lwi);
1294                 if (rc == -ETIMEDOUT)
1295                         CWARN("Service %s waiting for request buffers\n",
1296                               service->srv_name);
1297         }
1298
1299         /* schedule all outstanding replies to terminate them */
1300         spin_lock(&service->srv_lock);
1301         while (!list_empty(&service->srv_active_replies)) {
1302                 struct ptlrpc_reply_state *rs =
1303                         list_entry(service->srv_active_replies.next,
1304                                    struct ptlrpc_reply_state, rs_list);
1305                 ptlrpc_schedule_difficult_reply(rs);
1306         }
1307         spin_unlock(&service->srv_lock);
1308
1309         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1310          * and no service threads, so I'm the only thread noodling the
1311          * request queue now */
1312         while (!list_empty(&service->srv_request_queue)) {
1313                 struct ptlrpc_request *req =
1314                         list_entry(service->srv_request_queue.next,
1315                                    struct ptlrpc_request,
1316                                    rq_list);
1317
1318                 list_del(&req->rq_list);
1319                 service->srv_n_queued_reqs--;
1320                 service->srv_n_active_reqs++;
1321
1322                 ptlrpc_server_free_request(req);
1323         }
1324         LASSERT(service->srv_n_queued_reqs == 0);
1325         LASSERT(service->srv_n_active_reqs == 0);
1326         LASSERT(service->srv_n_history_rqbds == 0);
1327         LASSERT(list_empty(&service->srv_active_rqbds));
1328
1329         /* Now free all the request buffers since nothing references them
1330          * any more... */
1331         while (!list_empty(&service->srv_idle_rqbds)) {
1332                 struct ptlrpc_request_buffer_desc *rqbd =
1333                         list_entry(service->srv_idle_rqbds.next,
1334                                    struct ptlrpc_request_buffer_desc,
1335                                    rqbd_list);
1336
1337                 ptlrpc_free_rqbd(rqbd);
1338         }
1339
1340         /* wait for all outstanding replies to complete (they were
1341          * scheduled having been flagged to abort above) */
1342         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1343                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1344
1345                 rc = l_wait_event(service->srv_waitq,
1346                                   !list_empty(&service->srv_reply_queue), &lwi);
1347                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1348
1349                 if (rc == 0) {
1350                         ptlrpc_server_handle_reply(service);
1351                         continue;
1352                 }
1353                 CWARN("Unexpectedly long timeout %p\n", service);
1354         }
1355
1356         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1357                 list_del(&rs->rs_list);
1358                 OBD_FREE(rs, service->srv_max_reply_size);
1359         }
1360
1361         OBD_FREE_PTR(service);
1362         return 0;
1363 }
1364
1365 /* Returns 0 if the service is healthy.
1366  *
1367  * Right now, it just checks to make sure that requests aren't languishing
1368  * in the queue.  We'll use this health check to govern whether a node needs
1369  * to be shot, so it's intentionally non-aggressive. */
1370 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1371 {
1372         struct ptlrpc_request *request;
1373         struct timeval         right_now;
1374         long                   timediff, cutoff;
1375         int                    rc = 0;
1376
1377         if (svc == NULL)
1378                 return 0;
1379
1380         spin_lock(&svc->srv_lock);
1381
1382         if (list_empty(&svc->srv_request_queue))
1383                 goto out;
1384
1385         request = list_entry(svc->srv_request_queue.next,
1386                              struct ptlrpc_request, rq_list);
1387
1388         do_gettimeofday(&right_now);
1389         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1390
1391         cutoff = obd_health_check_timeout;
1392
1393         if (timediff / 1000000 > cutoff) {
1394                 rc = -1;
1395                 goto out;
1396         }
1397
1398  out:
1399         spin_unlock(&svc->srv_lock);
1400         return rc;
1401 }