lustre/ptlrpc/service.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #ifndef __KERNEL__
  39 #include <liblustre.h>
  40 #endif
  41 #include <obd_support.h>
  42 #include <obd_class.h>
  43 #include <lustre_net.h>
  44 #include <lu_object.h>
  45 #include <lnet/types.h>
  46 #include "ptlrpc_internal.h"
  47
  48 /* The following are visible and mutable through /sys/module/ptlrpc */
  49 int test_req_buffer_pressure = 0;
  50 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
  51                 "set non-zero to put pressure on request buffer pools");
  52 unsigned int at_min = 0;
  53 CFS_MODULE_PARM(at_min, "i", int, 0644,
  54                 "Adaptive timeout minimum (sec)");
  55 unsigned int at_max = 600;
  56 EXPORT_SYMBOL(at_max);
  57 CFS_MODULE_PARM(at_max, "i", int, 0644,
  58                 "Adaptive timeout maximum (sec)");
  59 unsigned int at_history = 600;
  60 CFS_MODULE_PARM(at_history, "i", int, 0644,
  61                 "Adaptive timeouts remember the slowest event that took place "
  62                 "within this period (sec)");
  63 static int at_early_margin = 5;
  64 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
  65                 "How soon before an RPC deadline to send an early reply");
  66 static int at_extra = 30;
  67 CFS_MODULE_PARM(at_extra, "i", int, 0644,
  68                 "How much extra time to give with each early reply");
  69
  70
  71 /* forward ref */
  72 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
  73
  74 static CFS_LIST_HEAD(ptlrpc_all_services);
  75 spinlock_t ptlrpc_all_services_lock;
  76
  77 static char *
  78 ptlrpc_alloc_request_buffer (int size)
  79 {
  80         char *ptr;
  81
  82         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  83                 OBD_VMALLOC(ptr, size);
  84         else
  85                 OBD_ALLOC(ptr, size);
  86
  87         return (ptr);
  88 }
  89
  90 static void
  91 ptlrpc_free_request_buffer (char *ptr, int size)
  92 {
  93         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  94                 OBD_VFREE(ptr, size);
  95         else
  96                 OBD_FREE(ptr, size);
  97 }
  98
  99 struct ptlrpc_request_buffer_desc *
 100 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
 101 {
 102         struct ptlrpc_request_buffer_desc *rqbd;
 103
 104         OBD_ALLOC_PTR(rqbd);
 105         if (rqbd == NULL)
 106                 return (NULL);
 107
 108         rqbd->rqbd_service = svc;
 109         rqbd->rqbd_refcount = 0;
 110         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
 111         rqbd->rqbd_cbid.cbid_arg = rqbd;
 112         CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
 113         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
 114
 115         if (rqbd->rqbd_buffer == NULL) {
 116                 OBD_FREE_PTR(rqbd);
 117                 return (NULL);
 118         }
 119
 120         spin_lock(&svc->srv_lock);
 121         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 122         svc->srv_nbufs++;
 123         spin_unlock(&svc->srv_lock);
 124
 125         return (rqbd);
 126 }
 127
 128 void
 129 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 130 {
 131         struct ptlrpc_service *svc = rqbd->rqbd_service;
 132
 133         LASSERT (rqbd->rqbd_refcount == 0);
 134         LASSERT (list_empty(&rqbd->rqbd_reqs));
 135
 136         spin_lock(&svc->srv_lock);
 137         list_del(&rqbd->rqbd_list);
 138         svc->srv_nbufs--;
 139         spin_unlock(&svc->srv_lock);
 140
 141         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 142         OBD_FREE_PTR(rqbd);
 143 }
 144
 145 int
 146 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
 147 {
 148         struct ptlrpc_request_buffer_desc *rqbd;
 149         int                                i;
 150
 151         CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
 152                svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
 153                svc->srv_nrqbd_receiving, svc->srv_nbufs);
 154         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 155                 rqbd = ptlrpc_alloc_rqbd(svc);
 156
 157                 if (rqbd == NULL) {
 158                         CERROR ("%s: Can't allocate request buffer\n",
 159                                 svc->srv_name);
 160                         return (-ENOMEM);
 161                 }
 162
 163                 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
 164                         return (-EAGAIN);
 165         }
 166
 167         return (0);
 168 }
 169
 170 void
 171 ptlrpc_save_lock (struct ptlrpc_request *req,
 172                   struct lustre_handle *lock, int mode)
 173 {
 174         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 175         int                        idx;
 176
 177         LASSERT(rs != NULL);
 178         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
 179
 180         idx = rs->rs_nlocks++;
 181         rs->rs_locks[idx] = *lock;
 182         rs->rs_modes[idx] = mode;
 183         rs->rs_difficult = 1;
 184 }
 185
 186 void
 187 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 188 {
 189         struct ptlrpc_service *svc = rs->rs_service;
 190
 191 #ifdef CONFIG_SMP
 192         LASSERT (spin_is_locked (&svc->srv_lock));
 193 #endif
 194         LASSERT (rs->rs_difficult);
 195         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 196
 197         if (rs->rs_scheduled)                   /* being set up or already notified */
 198                 return;
 199
 200         rs->rs_scheduled = 1;
 201         list_del (&rs->rs_list);
 202         list_add (&rs->rs_list, &svc->srv_reply_queue);
 203         cfs_waitq_signal (&svc->srv_waitq);
 204 }
 205
 206 void
 207 ptlrpc_commit_replies (struct obd_device *obd)
 208 {
 209         struct list_head   *tmp;
 210         struct list_head   *nxt;
 211
 212         /* Find any replies that have been committed and get their service
 213          * to attend to complete them. */
 214
 215         /* CAVEAT EMPTOR: spinlock ordering!!! */
 216         spin_lock(&obd->obd_uncommitted_replies_lock);
 217
 218         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 219                 struct ptlrpc_reply_state *rs =
 220                         list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
 221
 222                 LASSERT (rs->rs_difficult);
 223
 224                 if (rs->rs_transno <= obd->obd_last_committed) {
 225                         struct ptlrpc_service *svc = rs->rs_service;
 226
 227                         spin_lock (&svc->srv_lock);
 228                         list_del_init (&rs->rs_obd_list);
 229                         ptlrpc_schedule_difficult_reply (rs);
 230                         spin_unlock (&svc->srv_lock);
 231                 }
 232         }
 233
 234         spin_unlock(&obd->obd_uncommitted_replies_lock);
 235 }
 236
 237 static int
 238 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 239 {
 240         struct ptlrpc_request_buffer_desc *rqbd;
 241         int                                rc;
 242         int                                posted = 0;
 243
 244         for (;;) {
 245                 spin_lock(&svc->srv_lock);
 246
 247                 if (list_empty (&svc->srv_idle_rqbds)) {
 248                         spin_unlock(&svc->srv_lock);
 249                         return (posted);
 250                 }
 251
 252                 rqbd = list_entry(svc->srv_idle_rqbds.next,
 253                                   struct ptlrpc_request_buffer_desc,
 254                                   rqbd_list);
 255                 list_del (&rqbd->rqbd_list);
 256
 257                 /* assume we will post successfully */
 258                 svc->srv_nrqbd_receiving++;
 259                 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
 260
 261                 spin_unlock(&svc->srv_lock);
 262
 263                 rc = ptlrpc_register_rqbd(rqbd);
 264                 if (rc != 0)
 265                         break;
 266
 267                 posted = 1;
 268         }
 269
 270         spin_lock(&svc->srv_lock);
 271
 272         svc->srv_nrqbd_receiving--;
 273         list_del(&rqbd->rqbd_list);
 274         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 275
 276         /* Don't complain if no request buffers are posted right now; LNET
 277          * won't drop requests because we set the portal lazy! */
 278
 279         spin_unlock(&svc->srv_lock);
 280
 281         return (-1);
 282 }
 283
 284 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
 285                                             svc_handler_t h, char *name,
 286                                             struct proc_dir_entry *proc_entry,
 287                                             svcreq_printfn_t prntfn,
 288                                             char *threadname)
 289 {
 290         return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
 291                                c->psc_max_req_size, c->psc_max_reply_size,
 292                                c->psc_req_portal, c->psc_rep_portal,
 293                                c->psc_watchdog_factor,
 294                                h, name, proc_entry,
 295                                prntfn, c->psc_min_threads, c->psc_max_threads,
 296                                threadname, c->psc_ctx_tags);
 297 }
 298 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
 299
 300 static void ptlrpc_at_timer(unsigned long castmeharder)
 301 {
 302         struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
 303         svc->srv_at_check = 1;
 304         svc->srv_at_checktime = cfs_time_current();
 305         cfs_waitq_signal(&svc->srv_waitq);
 306 }
 307
 308 /* @threadname should be 11 characters or less - 3 will be added on */
 309 struct ptlrpc_service *
 310 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
 311                 int req_portal, int rep_portal, int watchdog_factor,
 312                 svc_handler_t handler, char *name,
 313                 cfs_proc_dir_entry_t *proc_entry,
 314                 svcreq_printfn_t svcreq_printfn,
 315                 int min_threads, int max_threads,
 316                 char *threadname, __u32 ctx_tags)
 317 {
 318         int                    rc;
 319         struct ptlrpc_service *service;
 320         ENTRY;
 321
 322         LASSERT (nbufs > 0);
 323         LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
 324         LASSERT (ctx_tags != 0);
 325
 326         OBD_ALLOC_PTR(service);
 327         if (service == NULL)
 328                 RETURN(NULL);
 329
 330         /* First initialise enough for early teardown */
 331
 332         service->srv_name = name;
 333         spin_lock_init(&service->srv_lock);
 334         CFS_INIT_LIST_HEAD(&service->srv_threads);
 335         cfs_waitq_init(&service->srv_waitq);
 336
 337         service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
 338         service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
 339         service->srv_buf_size = bufsize;
 340         service->srv_rep_portal = rep_portal;
 341         service->srv_req_portal = req_portal;
 342         service->srv_watchdog_factor = watchdog_factor;
 343         service->srv_handler = handler;
 344         service->srv_request_history_print_fn = svcreq_printfn;
 345         service->srv_request_seq = 1;           /* valid seq #s start at 1 */
 346         service->srv_request_max_cull_seq = 0;
 347         service->srv_threads_min = min_threads;
 348         service->srv_threads_max = max_threads;
 349         service->srv_thread_name = threadname;
 350         service->srv_ctx_tags = ctx_tags;
 351
 352         rc = LNetSetLazyPortal(service->srv_req_portal);
 353         LASSERT (rc == 0);
 354
 355         CFS_INIT_LIST_HEAD(&service->srv_request_queue);
 356         CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
 357         CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
 358         CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
 359         CFS_INIT_LIST_HEAD(&service->srv_request_history);
 360         CFS_INIT_LIST_HEAD(&service->srv_active_replies);
 361         CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
 362         CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
 363         cfs_waitq_init(&service->srv_free_rs_waitq);
 364
 365         spin_lock_init(&service->srv_at_lock);
 366         CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
 367         CFS_INIT_LIST_HEAD(&service->srv_at_list);
 368         cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
 369         /* At SOW, service time should be quick; 10s seems generous. If client
 370            timeout is less than this, we'll be sending an early reply. */
 371         at_init(&service->srv_at_estimate, 10, 0);
 372
 373         spin_lock (&ptlrpc_all_services_lock);
 374         list_add (&service->srv_list, &ptlrpc_all_services);
 375         spin_unlock (&ptlrpc_all_services_lock);
 376
 377         /* Now allocate the request buffers */
 378         rc = ptlrpc_grow_req_bufs(service);
 379         /* We shouldn't be under memory pressure at startup, so
 380          * fail if we can't post all our buffers at this time. */
 381         if (rc != 0)
 382                 GOTO(failed, NULL);
 383
 384         /* Now allocate pool of reply buffers */
 385         /* Increase max reply size to next power of two */
 386         service->srv_max_reply_size = 1;
 387         while (service->srv_max_reply_size <
 388                max_reply_size + SPTLRPC_MAX_PAYLOAD)
 389                 service->srv_max_reply_size <<= 1;
 390
 391         if (proc_entry != NULL)
 392                 ptlrpc_lprocfs_register_service(proc_entry, service);
 393
 394         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 395                service->srv_name, service->srv_req_portal);
 396
 397         RETURN(service);
 398 failed:
 399         ptlrpc_unregister_service(service);
 400         return NULL;
 401 }
 402
 403 static void ptlrpc_server_req_decref(struct ptlrpc_request *req)
 404 {
 405         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 406
 407         if (!atomic_dec_and_test(&req->rq_refcount))
 408                 return;
 409
 410         sptlrpc_svc_ctx_decref(req);
 411
 412         LASSERT(list_empty(&req->rq_timed_list));
 413         if (req != &rqbd->rqbd_req) {
 414                 /* NB request buffers use an embedded
 415                  * req if the incoming req unlinked the
 416                  * MD; this isn't one of them! */
 417                 OBD_FREE(req, sizeof(*req));
 418         } else {
 419                 struct ptlrpc_service *svc = rqbd->rqbd_service;
 420                 /* schedule request buffer for re-use.
 421                  * NB I can only do this after I've disposed of their
 422                  * reqs; particularly the embedded req */
 423                 spin_lock(&svc->srv_lock);
 424                 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 425                 spin_unlock(&svc->srv_lock);
 426         }
 427 }
 428
 429 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
 430 {
 431         list_del(&req->rq_list);
 432         ptlrpc_req_drop_rs(req);
 433         ptlrpc_server_req_decref(req);
 434 }
 435
 436 static void
 437 ptlrpc_server_free_request(struct ptlrpc_request *req)
 438 {
 439         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 440         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 441         int                                refcount;
 442         struct list_head                  *tmp;
 443         struct list_head                  *nxt;
 444
 445         if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
 446                 DEBUG_REQ(D_INFO, req, "free req");
 447         spin_lock(&svc->srv_at_lock);
 448         req->rq_sent_final = 1;
 449         list_del_init(&req->rq_timed_list);
 450         spin_unlock(&svc->srv_at_lock);
 451
 452         spin_lock(&svc->srv_lock);
 453
 454         svc->srv_n_active_reqs--;
 455         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 456
 457         refcount = --(rqbd->rqbd_refcount);
 458         if (refcount == 0) {
 459                 /* request buffer is now idle: add to history */
 460                 list_del(&rqbd->rqbd_list);
 461                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 462                 svc->srv_n_history_rqbds++;
 463
 464                 /* cull some history?
 465                  * I expect only about 1 or 2 rqbds need to be recycled here */
 466                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 467                         rqbd = list_entry(svc->srv_history_rqbds.next,
 468                                           struct ptlrpc_request_buffer_desc,
 469                                           rqbd_list);
 470
 471                         list_del(&rqbd->rqbd_list);
 472                         svc->srv_n_history_rqbds--;
 473
 474                         /* remove rqbd's reqs from svc's req history while
 475                          * I've got the service lock */
 476                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 477                                 req = list_entry(tmp, struct ptlrpc_request,
 478                                                  rq_list);
 479                                 /* Track the highest culled req seq */
 480                                 if (req->rq_history_seq >
 481                                     svc->srv_request_max_cull_seq)
 482                                         svc->srv_request_max_cull_seq =
 483                                                 req->rq_history_seq;
 484                                 list_del(&req->rq_history_list);
 485                         }
 486
 487                         spin_unlock(&svc->srv_lock);
 488
 489                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 490                                 req = list_entry(rqbd->rqbd_reqs.next,
 491                                                  struct ptlrpc_request,
 492                                                  rq_list);
 493                                 __ptlrpc_server_free_request(req);
 494                         }
 495
 496                         spin_lock(&svc->srv_lock);
 497                 }
 498         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 499                  /* If we are low on memory, we are not interested in
 500                     history */
 501                 list_del(&req->rq_history_list);
 502                 __ptlrpc_server_free_request(req);
 503         }
 504
 505         spin_unlock(&svc->srv_lock);
 506
 507 }
 508
 509 /* This function makes sure dead exports are evicted in a timely manner.
 510    This function is only called when some export receives a message (i.e.,
 511    the network is up.) */
 512 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 513 {
 514         struct obd_export *oldest_exp;
 515         time_t oldest_time;
 516
 517         ENTRY;
 518
 519         LASSERT(exp);
 520
 521         /* Compensate for slow machines, etc, by faking our request time
 522            into the future.  Although this can break the strict time-ordering
 523            of the list, we can be really lazy here - we don't have to evict
 524            at the exact right moment.  Eventually, all silent exports
 525            will make it to the top of the list. */
 526         exp->exp_last_request_time = max(exp->exp_last_request_time,
 527                                          cfs_time_current_sec() + extra_delay);
 528
 529         CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
 530                exp->exp_client_uuid.uuid,
 531                exp->exp_last_request_time, exp);
 532
 533         /* exports may get disconnected from the chain even though the
 534            export has references, so we must keep the spin lock while
 535            manipulating the lists */
 536         spin_lock(&exp->exp_obd->obd_dev_lock);
 537
 538         if (list_empty(&exp->exp_obd_chain_timed)) {
 539                 /* this one is not timed */
 540                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 541                 EXIT;
 542                 return;
 543         }
 544
 545         list_move_tail(&exp->exp_obd_chain_timed,
 546                        &exp->exp_obd->obd_exports_timed);
 547
 548         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 549                                 struct obd_export, exp_obd_chain_timed);
 550         oldest_time = oldest_exp->exp_last_request_time;
 551         spin_unlock(&exp->exp_obd->obd_dev_lock);
 552
 553         if (exp->exp_obd->obd_recovering) {
 554                 /* be nice to everyone during recovery */
 555                 EXIT;
 556                 return;
 557         }
 558
 559         /* Note - racing to start/reset the obd_eviction timer is safe */
 560         if (exp->exp_obd->obd_eviction_timer == 0) {
 561                 /* Check if the oldest entry is expired. */
 562                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
 563                                               extra_delay)) {
 564                         /* We need a second timer, in case the net was down and
 565                          * it just came back. Since the pinger may skip every
 566                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 567                          * we better wait for 3. */
 568                         exp->exp_obd->obd_eviction_timer =
 569                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
 570                         CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
 571                                exp->exp_obd->obd_name, obd_export_nid2str(exp),
 572                                oldest_time);
 573                 }
 574         } else {
 575                 if (cfs_time_current_sec() >
 576                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
 577                         /* The evictor won't evict anyone who we've heard from
 578                          * recently, so we don't have to check before we start
 579                          * it. */
 580                         if (!ping_evictor_wake(exp))
 581                                 exp->exp_obd->obd_eviction_timer = 0;
 582                 }
 583         }
 584
 585         EXIT;
 586 }
 587
 588 #ifndef __KERNEL__
 589 int lu_context_init(struct lu_context *ctx, __u32 tags)
 590 {
 591         return 0;
 592 }
 593
 594 void lu_context_fini(struct lu_context *ctx)
 595 {
 596 }
 597
 598 void lu_context_enter(struct lu_context *ctx)
 599 {
 600 }
 601
 602 void lu_context_exit(struct lu_context *ctx)
 603 {
 604 }
 605
 606 #endif
 607
 608 static int ptlrpc_check_req(struct ptlrpc_request *req)
 609 {
 610         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 611                      req->rq_export->exp_conn_cnt)) {
 612                 DEBUG_REQ(D_ERROR, req,
 613                           "DROPPING req from old connection %d < %d",
 614                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 615                           req->rq_export->exp_conn_cnt);
 616                 return -EEXIST;
 617         }
 618         if (unlikely(req->rq_export->exp_obd &&
 619                      req->rq_export->exp_obd->obd_fail)) {
 620              /* Failing over, don't handle any more reqs, send
 621                 error response instead. */
 622                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 623                        req, req->rq_export->exp_obd->obd_name);
 624                 req->rq_status = -ENODEV;
 625                 ptlrpc_error(req);
 626                 return -ENODEV;
 627         }
 628
 629         return 0;
 630 }
 631
 632 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
 633 {
 634         struct ptlrpc_request *rq;
 635         __s32 next;
 636
 637         spin_lock(&svc->srv_at_lock);
 638         if (list_empty(&svc->srv_at_list)) {
 639                 cfs_timer_disarm(&svc->srv_at_timer);
 640                 spin_unlock(&svc->srv_at_lock);
 641                 return;
 642         }
 643
 644         /* Set timer for closest deadline */
 645         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 646                         rq_timed_list);
 647         next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
 648                        at_early_margin);
 649         if (next <= 0)
 650                 ptlrpc_at_timer((unsigned long)svc);
 651         else
 652                 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
 653         spin_unlock(&svc->srv_at_lock);
 654         CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
 655 }
 656
 657 /* Add rpc to early reply check list */
 658 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 659 {
 660         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 661         struct ptlrpc_request *rq;
 662         int found = 0;
 663
 664         if (AT_OFF)
 665                 return(0);
 666
 667         if (req->rq_no_reply)
 668                 return 0;
 669
 670         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 671                 return(-ENOSYS);
 672
 673         DEBUG_REQ(D_ADAPTTO, req, "add timed %lds",
 674                   req->rq_deadline - cfs_time_current_sec());
 675
 676         spin_lock(&svc->srv_at_lock);
 677
 678         if (unlikely(req->rq_sent_final)) {
 679                 spin_unlock(&svc->srv_at_lock);
 680                 return 0;
 681         }
 682
 683         LASSERT(list_empty(&req->rq_timed_list));
 684         /* Add to sorted list.  Presumably latest rpcs will have the latest
 685            deadlines, so search backward. */
 686         list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
 687                 if (req->rq_deadline > rq->rq_deadline) {
 688                         list_add(&req->rq_timed_list, &rq->rq_timed_list);
 689                         found++;
 690                         break;
 691                 }
 692         }
 693         if (!found)
 694                 /* Add to front if shortest deadline or list empty */
 695                 list_add(&req->rq_timed_list, &svc->srv_at_list);
 696
 697         /* Check if we're the head of the list */
 698         found = (svc->srv_at_list.next == &req->rq_timed_list);
 699
 700         spin_unlock(&svc->srv_at_lock);
 701
 702         if (found)
 703                 ptlrpc_at_set_timer(svc);
 704
 705         return 0;
 706 }
 707
 708 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
 709                                       int extra_time)
 710 {
 711         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 712         struct ptlrpc_request *reqcopy;
 713         struct lustre_msg *reqmsg;
 714         long olddl = req->rq_deadline - cfs_time_current_sec();
 715         time_t newdl;
 716         int rc;
 717         ENTRY;
 718
 719         /* deadline is when the client expects us to reply, margin is the
 720            difference between clients' and servers' expectations */
 721         DEBUG_REQ(D_ADAPTTO, req,
 722                   "%ssending early reply (deadline %+lds, margin %+lds) for "
 723                   "%d+%d", AT_OFF ? "AT off - not " : "",
 724                   olddl, olddl - at_get(&svc->srv_at_estimate),
 725                   at_get(&svc->srv_at_estimate), extra_time);
 726
 727         if (AT_OFF)
 728                 RETURN(0);
 729
 730         if (olddl < 0) {
 731                 CDEBUG(D_WARNING, "x"LPU64": Already past deadline (%+lds), not"
 732                        " sending early reply. Increase at_early_margin (%d)?\n",
 733                        req->rq_xid, olddl, at_early_margin);
 734                 /* Return an error so we're not re-added to the timed list. */
 735                 RETURN(-ETIMEDOUT);
 736         }
 737
 738         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
 739                 CDEBUG(D_INFO, "Wanted to ask client for more time, but no AT "
 740                       "support\n");
 741                 RETURN(-ENOSYS);
 742         }
 743
 744         if (extra_time) {
 745                 /* Fake our processing time into the future to ask the
 746                    clients for some extra amount of time */
 747                 extra_time += cfs_time_current_sec() -
 748                         req->rq_arrival_time.tv_sec;
 749                 at_add(&svc->srv_at_estimate, extra_time);
 750         }
 751
 752         newdl = req->rq_arrival_time.tv_sec + at_get(&svc->srv_at_estimate);
 753         if (req->rq_deadline >= newdl) {
 754                 /* We're not adding any time, no need to send an early reply
 755                    (e.g. maybe at adaptive_max) */
 756                 CDEBUG(D_ADAPTTO, "x"LPU64": Couldn't add any time (%ld/%ld), "
 757                        "not sending early reply\n", req->rq_xid, olddl,
 758                        newdl - cfs_time_current_sec());
 759                 RETURN(-ETIMEDOUT);
 760         }
 761
 762         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 763         if (reqcopy == NULL)
 764                 RETURN(-ENOMEM);
 765         OBD_ALLOC(reqmsg, req->rq_reqlen);
 766         if (!reqmsg) {
 767                 OBD_FREE(reqcopy, sizeof *reqcopy);
 768                 RETURN(-ENOMEM);
 769         }
 770
 771         *reqcopy = *req;
 772         reqcopy->rq_reply_state = NULL;
 773         reqcopy->rq_rep_swab_mask = 0;
 774         reqcopy->rq_pack_bulk = 0;
 775         reqcopy->rq_pack_udesc = 0;
 776         reqcopy->rq_packed_final = 0;
 777         sptlrpc_svc_ctx_addref(reqcopy);
 778         /* We only need the reqmsg for the magic */
 779         reqcopy->rq_reqmsg = reqmsg;
 780         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 781
 782         if (req->rq_sent_final) {
 783                 CDEBUG(D_ADAPTTO, "x"LPU64": normal reply already sent out, "
 784                        "abort sending early reply\n", req->rq_xid);
 785                 GOTO(out, rc = 0);
 786         }
 787
 788         /* Connection ref */
 789         reqcopy->rq_export = class_conn2export(
 790                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 791         if (reqcopy->rq_export == NULL)
 792                 GOTO(out, rc = -ENODEV);
 793
 794         /* RPC ref */
 795         class_export_rpc_get(reqcopy->rq_export);
 796         if (reqcopy->rq_export->exp_obd &&
 797             reqcopy->rq_export->exp_obd->obd_fail)
 798                 GOTO(out_put, rc = -ENODEV);
 799
 800         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 801         if (rc)
 802                 GOTO(out_put, rc);
 803
 804         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 805
 806         if (!rc) {
 807                 /* Adjust our own deadline to what we told the client */
 808                 req->rq_deadline = newdl;
 809                 req->rq_early_count++; /* number sent, server side */
 810         } else {
 811                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 812         }
 813
 814         /* Free the (early) reply state from lustre_pack_reply.
 815            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 816         ptlrpc_req_drop_rs(reqcopy);
 817
 818 out_put:
 819         class_export_rpc_put(reqcopy->rq_export);
 820         class_export_put(reqcopy->rq_export);
 821 out:
 822         sptlrpc_svc_ctx_decref(reqcopy);
 823         OBD_FREE(reqmsg, req->rq_reqlen);
 824         OBD_FREE(reqcopy, sizeof *reqcopy);
 825         RETURN(rc);
 826 }
 827
 828 /* Send early replies to everybody expiring within at_early_margin
 829    asking for at_extra time */
 830 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 831 {
 832         struct ptlrpc_request *rq, *n;
 833         struct list_head work_list;
 834         time_t now = cfs_time_current_sec();
 835         cfs_duration_t delay;
 836         int first, counter = 0;
 837         ENTRY;
 838
 839         spin_lock(&svc->srv_at_lock);
 840         if (svc->srv_at_check == 0) {
 841                 spin_unlock(&svc->srv_at_lock);
 842                 RETURN(0);
 843         }
 844         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 845         svc->srv_at_check = 0;
 846
 847         if (list_empty(&svc->srv_at_list)) {
 848                 spin_unlock(&svc->srv_at_lock);
 849                 RETURN(0);
 850         }
 851
 852         /* The timer went off, but maybe the nearest rpc already completed. */
 853         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 854                         rq_timed_list);
 855         first = (int)(rq->rq_deadline - now);
 856         if (first > at_early_margin) {
 857                 /* We've still got plenty of time.  Reset the timer. */
 858                 spin_unlock(&svc->srv_at_lock);
 859                 ptlrpc_at_set_timer(svc);
 860                 RETURN(0);
 861         }
 862
 863         /* We're close to a timeout, and we don't know how much longer the
 864            server will take. Send early replies to everyone expiring soon. */
 865         CFS_INIT_LIST_HEAD(&work_list);
 866         list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
 867                 if (rq->rq_deadline <= now + at_early_margin) {
 868                         list_move_tail(&rq->rq_timed_list, &work_list);
 869                         counter++;
 870                 } else {
 871                         break;
 872                 }
 873         }
 874
 875         spin_unlock(&svc->srv_at_lock);
 876
 877         /* we have a new earliest deadline, restart the timer */
 878         ptlrpc_at_set_timer(svc);
 879
 880         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 881                "replies\n", first, at_extra, counter);
 882         if (first < 0) {
 883                 /* We're already past request deadlines before we even get a
 884                    chance to send early replies */
 885                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 886                               "request traffic (cpu-bound).\n", svc->srv_name);
 887                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 888                       "delay="CFS_DURATION_T"(jiff)\n",
 889                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 890                       at_get(&svc->srv_at_estimate), delay);
 891         }
 892
 893         /* ptlrpc_server_free_request may delete an entry out of the work
 894            list */
 895         spin_lock(&svc->srv_at_lock);
 896         while (!list_empty(&work_list)) {
 897                 rq = list_entry(work_list.next, struct ptlrpc_request,
 898                                 rq_timed_list);
 899                 list_del_init(&rq->rq_timed_list);
 900                 /* if the entry is still in the worklist, it hasn't been
 901                    deleted, and is safe to take a ref to keep the req around */
 902                 atomic_inc(&rq->rq_refcount);
 903                 spin_unlock(&svc->srv_at_lock);
 904
 905                 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
 906                         ptlrpc_at_add_timed(rq);
 907
 908                 ptlrpc_server_req_decref(rq);
 909                 spin_lock(&svc->srv_at_lock);
 910         }
 911         spin_unlock(&svc->srv_at_lock);
 912
 913         RETURN(0);
 914 }
 915
 916 /* Handle freshly incoming reqs, add to timed early reply list,
 917    pass on to regular request queue */
 918 static int
 919 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
 920 {
 921         struct ptlrpc_request *req;
 922         __u32                  deadline;
 923         int                    rc;
 924         ENTRY;
 925
 926         LASSERT(svc);
 927
 928         spin_lock(&svc->srv_lock);
 929         if (list_empty(&svc->srv_req_in_queue)) {
 930                 spin_unlock(&svc->srv_lock);
 931                 RETURN(0);
 932         }
 933
 934         req = list_entry(svc->srv_req_in_queue.next,
 935                          struct ptlrpc_request, rq_list);
 936         list_del_init (&req->rq_list);
 937         /* Consider this still a "queued" request as far as stats are
 938            concerned */
 939         spin_unlock(&svc->srv_lock);
 940
 941         /* go through security check/transform */
 942         rc = sptlrpc_svc_unwrap_request(req);
 943         switch (rc) {
 944         case SECSVC_OK:
 945                 break;
 946         case SECSVC_COMPLETE:
 947                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 948                 goto err_req;
 949         case SECSVC_DROP:
 950                 goto err_req;
 951         default:
 952                 LBUG();
 953         }
 954
 955         /* Clear request swab mask; this is a new request */
 956         req->rq_req_swab_mask = 0;
 957
 958         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 959         if (rc != 0) {
 960                 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
 961                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
 962                        req->rq_xid);
 963                 goto err_req;
 964         }
 965
 966         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 967         if (rc) {
 968                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
 969                         LPU64"\n", svc->srv_req_portal,
 970                         libcfs_id2str(req->rq_peer), req->rq_xid);
 971                 goto err_req;
 972         }
 973
 974         rc = -EINVAL;
 975         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 976                 CERROR("wrong packet type received (type=%u) from %s\n",
 977                        lustre_msg_get_type(req->rq_reqmsg),
 978                        libcfs_id2str(req->rq_peer));
 979                 goto err_req;
 980         }
 981
 982         CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
 983
 984         req->rq_export = class_conn2export(
 985                 lustre_msg_get_handle(req->rq_reqmsg));
 986         if (req->rq_export) {
 987                 rc = ptlrpc_check_req(req);
 988                 if (rc == 0) {
 989                         rc = sptlrpc_target_export_check(req->rq_export, req);
 990                         if (rc)
 991                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
 992                                           "illegal security flavor,");
 993                 }
 994
 995                 class_export_put(req->rq_export);
 996                 req->rq_export = NULL;
 997                 if (rc)
 998                         goto err_req;
 999         }
1000
1001         /* req_in handling should/must be fast */
1002         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1003                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus",
1004                           cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
1005
1006         /* Set rpc server deadline and add it to the timed list */
1007         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1008                     MSGHDR_AT_SUPPORT) ?
1009                    /* The max time the client expects us to take */
1010                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1011         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1012         if (unlikely(deadline == 0)) {
1013                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1014                 goto err_req;
1015         }
1016
1017         ptlrpc_at_add_timed(req);
1018
1019         /* Move it over to the request processing queue */
1020         spin_lock(&svc->srv_lock);
1021         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1022         cfs_waitq_signal(&svc->srv_waitq);
1023         spin_unlock(&svc->srv_lock);
1024         RETURN(1);
1025
1026 err_req:
1027         spin_lock(&svc->srv_lock);
1028         svc->srv_n_queued_reqs--;
1029         svc->srv_n_active_reqs++;
1030         spin_unlock(&svc->srv_lock);
1031         ptlrpc_server_free_request(req);
1032
1033         RETURN(1);
1034 }
1035
1036 static int
1037 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1038                              struct ptlrpc_thread *thread)
1039 {
1040         struct obd_export     *export = NULL;
1041         struct ptlrpc_request *request;
1042         struct timeval         work_start;
1043         struct timeval         work_end;
1044         long                   timediff;
1045         int                    rc;
1046         ENTRY;
1047
1048         LASSERT(svc);
1049
1050         spin_lock(&svc->srv_lock);
1051         if (unlikely(list_empty (&svc->srv_request_queue) ||
1052             (
1053 #ifndef __KERNEL__
1054              /* !@%$# liblustre only has 1 thread */
1055              svc->srv_n_difficult_replies != 0 &&
1056 #endif
1057              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1058                  /* Don't handle regular requests in the last thread, in order               * re
1059                   * to handle difficult replies (which might block other threads)
1060                   * as well as handle any incoming reqs, early replies, etc.
1061                   * That means we always need at least 2 service threads. */
1062                 spin_unlock(&svc->srv_lock);
1063                 RETURN(0);
1064         }
1065
1066         request = list_entry (svc->srv_request_queue.next,
1067                               struct ptlrpc_request, rq_list);
1068         list_del_init (&request->rq_list);
1069         svc->srv_n_queued_reqs--;
1070         svc->srv_n_active_reqs++;
1071
1072         spin_unlock(&svc->srv_lock);
1073
1074         do_gettimeofday(&work_start);
1075         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1076         if (likely(svc->srv_stats != NULL)) {
1077                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1078                                     timediff);
1079                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1080                                     svc->srv_n_queued_reqs);
1081                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1082                                     svc->srv_n_active_reqs);
1083                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1084                                     at_get(&svc->srv_at_estimate));
1085         }
1086
1087         rc = lu_context_init(&request->rq_session, LCT_SESSION);
1088         if (rc) {
1089                 CERROR("Failure to initialize session: %d\n", rc);
1090                 goto out_req;
1091         }
1092         request->rq_session.lc_thread = thread;
1093         lu_context_enter(&request->rq_session);
1094
1095         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
1096
1097         request->rq_svc_thread = thread;
1098         if (thread)
1099                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1100
1101         request->rq_export = class_conn2export(
1102                                      lustre_msg_get_handle(request->rq_reqmsg));
1103
1104         if (likely(request->rq_export)) {
1105                 if (unlikely(ptlrpc_check_req(request)))
1106                         goto put_conn;
1107                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1108                 export = class_export_rpc_get(request->rq_export);
1109         }
1110
1111         /* Discard requests queued for longer than the deadline.
1112            The deadline is increased if we send an early reply. */
1113         if (cfs_time_current_sec() > request->rq_deadline) {
1114                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1115                           ": deadline %ld%+lds ago\n",
1116                           libcfs_id2str(request->rq_peer),
1117                           request->rq_deadline -
1118                           request->rq_arrival_time.tv_sec,
1119                           cfs_time_current_sec() - request->rq_deadline);
1120                 goto put_rpc_export;
1121         }
1122
1123         request->rq_phase = RQ_PHASE_INTERPRET;
1124
1125         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1126                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1127                (request->rq_export ?
1128                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1129                (request->rq_export ?
1130                 atomic_read(&request->rq_export->exp_refcount) : -99),
1131                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1132                libcfs_id2str(request->rq_peer),
1133                lustre_msg_get_opc(request->rq_reqmsg));
1134
1135         OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1136
1137         rc = svc->srv_handler(request);
1138
1139         request->rq_phase = RQ_PHASE_COMPLETE;
1140
1141         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1142                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1143                (request->rq_export ?
1144                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1145                (request->rq_export ?
1146                 atomic_read(&request->rq_export->exp_refcount) : -99),
1147                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1148                libcfs_id2str(request->rq_peer),
1149                lustre_msg_get_opc(request->rq_reqmsg));
1150
1151 put_rpc_export:
1152         if (export != NULL)
1153                 class_export_rpc_put(export);
1154 put_conn:
1155         if (likely(request->rq_export != NULL))
1156                 class_export_put(request->rq_export);
1157
1158         lu_context_exit(&request->rq_session);
1159         lu_context_fini(&request->rq_session);
1160
1161         if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1162                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1163                           "than estimated (%ld%+lds); client may timeout.",
1164                           request->rq_xid, request->rq_deadline -
1165                           request->rq_arrival_time.tv_sec,
1166                           cfs_time_current_sec() - request->rq_deadline);
1167         }
1168
1169         do_gettimeofday(&work_end);
1170         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1171         CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1172                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1173                request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1174                libcfs_id2str(request->rq_peer), timediff,
1175                cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1176                request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1177                request->rq_transno, request->rq_status,
1178                request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1179                -999);
1180         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1181                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1182                 int opc = opcode_offset(op);
1183                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1184                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1185                         lprocfs_counter_add(svc->srv_stats,
1186                                             opc + EXTRA_MAX_OPCODES,
1187                                             timediff);
1188                 }
1189         }
1190         if (unlikely(request->rq_early_count)) {
1191                 DEBUG_REQ(D_ADAPTTO, request,
1192                           "sent %d early replies before finishing in %lds",
1193                           request->rq_early_count,
1194                           work_end.tv_sec - request->rq_arrival_time.tv_sec);
1195         }
1196
1197 out_req:
1198         ptlrpc_server_free_request(request);
1199
1200         RETURN(1);
1201 }
1202
1203 static int
1204 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1205 {
1206         struct ptlrpc_reply_state *rs;
1207         struct obd_export         *exp;
1208         struct obd_device         *obd;
1209         int                        nlocks;
1210         int                        been_handled;
1211         ENTRY;
1212
1213         spin_lock(&svc->srv_lock);
1214         if (list_empty (&svc->srv_reply_queue)) {
1215                 spin_unlock(&svc->srv_lock);
1216                 RETURN(0);
1217         }
1218
1219         rs = list_entry (svc->srv_reply_queue.next,
1220                          struct ptlrpc_reply_state, rs_list);
1221
1222         exp = rs->rs_export;
1223         obd = exp->exp_obd;
1224
1225         LASSERT (rs->rs_difficult);
1226         LASSERT (rs->rs_scheduled);
1227
1228         list_del_init (&rs->rs_list);
1229
1230         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1231         spin_unlock(&svc->srv_lock);
1232
1233         spin_lock (&obd->obd_uncommitted_replies_lock);
1234         /* Noop if removed already */
1235         list_del_init (&rs->rs_obd_list);
1236         spin_unlock (&obd->obd_uncommitted_replies_lock);
1237
1238         spin_lock (&exp->exp_lock);
1239         /* Noop if removed already */
1240         list_del_init (&rs->rs_exp_list);
1241         spin_unlock (&exp->exp_lock);
1242
1243         spin_lock(&svc->srv_lock);
1244
1245         been_handled = rs->rs_handled;
1246         rs->rs_handled = 1;
1247
1248         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1249         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1250
1251         if (nlocks == 0 && !been_handled) {
1252                 /* If we see this, we should already have seen the warning
1253                  * in mds_steal_ack_locks()  */
1254                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1255                       " o%d NID %s\n",
1256                       rs,
1257                       rs->rs_xid, rs->rs_transno,
1258                       lustre_msg_get_opc(rs->rs_msg),
1259                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1260         }
1261
1262         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1263                 spin_unlock(&svc->srv_lock);
1264
1265                 if (!been_handled && rs->rs_on_net) {
1266                         LNetMDUnlink(rs->rs_md_h);
1267                         /* Ignore return code; we're racing with
1268                          * completion... */
1269                 }
1270
1271                 while (nlocks-- > 0)
1272                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1273                                          rs->rs_modes[nlocks]);
1274
1275                 spin_lock(&svc->srv_lock);
1276         }
1277
1278         rs->rs_scheduled = 0;
1279
1280         if (!rs->rs_on_net) {
1281                 /* Off the net */
1282                 svc->srv_n_difficult_replies--;
1283                 spin_unlock(&svc->srv_lock);
1284
1285                 class_export_put (exp);
1286                 rs->rs_export = NULL;
1287                 ptlrpc_rs_decref (rs);
1288                 atomic_dec (&svc->srv_outstanding_replies);
1289                 RETURN(1);
1290         }
1291
1292         /* still on the net; callback will schedule */
1293         spin_unlock(&svc->srv_lock);
1294         RETURN(1);
1295 }
1296
1297 #ifndef __KERNEL__
1298 /* FIXME make use of timeout later */
1299 int
1300 liblustre_check_services (void *arg)
1301 {
1302         int  did_something = 0;
1303         int  rc;
1304         struct list_head *tmp, *nxt;
1305         ENTRY;
1306
1307         /* I'm relying on being single threaded, not to have to lock
1308          * ptlrpc_all_services etc */
1309         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1310                 struct ptlrpc_service *svc =
1311                         list_entry (tmp, struct ptlrpc_service, srv_list);
1312
1313                 if (svc->srv_threads_running != 0)     /* I've recursed */
1314                         continue;
1315
1316                 /* service threads can block for bulk, so this limits us
1317                  * (arbitrarily) to recursing 1 stack frame per service.
1318                  * Note that the problem with recursion is that we have to
1319                  * unwind completely before our caller can resume. */
1320
1321                 svc->srv_threads_running++;
1322
1323                 do {
1324                         rc = ptlrpc_server_handle_req_in(svc);
1325                         rc |= ptlrpc_server_handle_reply(svc);
1326                         rc |= ptlrpc_at_check_timed(svc);
1327                         rc |= ptlrpc_server_handle_request(svc, NULL);
1328                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1329                         did_something |= rc;
1330                 } while (rc);
1331
1332                 svc->srv_threads_running--;
1333         }
1334
1335         RETURN(did_something);
1336 }
1337 #define ptlrpc_stop_all_threads(s) do {} while (0)
1338
1339 #else /* __KERNEL__ */
1340
1341 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1342 void ptlrpc_daemonize(char *name)
1343 {
1344         struct fs_struct *fs = current->fs;
1345
1346         atomic_inc(&fs->count);
1347         cfs_daemonize(name);
1348         exit_fs(cfs_current());
1349         current->fs = fs;
1350         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1351 }
1352
1353 static void
1354 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1355 {
1356         int avail = svc->srv_nrqbd_receiving;
1357         int low_water = test_req_buffer_pressure ? 0 :
1358                         svc->srv_nbuf_per_group/2;
1359
1360         /* NB I'm not locking; just looking. */
1361
1362         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1363          * allowed the request history to grow out of control.  We could put a
1364          * sanity check on that here and cull some history if we need the
1365          * space. */
1366
1367         if (avail <= low_water)
1368                 ptlrpc_grow_req_bufs(svc);
1369
1370         if (svc->srv_stats)
1371                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1372                                     avail);
1373 }
1374
1375 static int
1376 ptlrpc_retry_rqbds(void *arg)
1377 {
1378         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1379
1380         svc->srv_rqbd_timeout = 0;
1381         return (-ETIMEDOUT);
1382 }
1383
1384 static int ptlrpc_main(void *arg)
1385 {
1386         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1387         struct ptlrpc_service  *svc = data->svc;
1388         struct ptlrpc_thread   *thread = data->thread;
1389         struct obd_device      *dev = data->dev;
1390         struct ptlrpc_reply_state *rs;
1391         struct lc_watchdog     *watchdog;
1392 #ifdef WITH_GROUP_INFO
1393         struct group_info *ginfo = NULL;
1394 #endif
1395         struct lu_env env;
1396         int counter = 0, rc = 0;
1397         ENTRY;
1398
1399         ptlrpc_daemonize(data->name);
1400
1401 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1402         /* we need to do this before any per-thread allocation is done so that
1403          * we get the per-thread allocations on local node.  bug 7342 */
1404         if (svc->srv_cpu_affinity) {
1405                 int cpu, num_cpu;
1406
1407                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1408                         if (!cpu_online(cpu))
1409                                 continue;
1410                         if (num_cpu == thread->t_id % num_online_cpus())
1411                                 break;
1412                         num_cpu++;
1413                 }
1414                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1415         }
1416 #endif
1417
1418 #ifdef WITH_GROUP_INFO
1419         ginfo = groups_alloc(0);
1420         if (!ginfo) {
1421                 rc = -ENOMEM;
1422                 goto out;
1423         }
1424
1425         set_current_groups(ginfo);
1426         put_group_info(ginfo);
1427 #endif
1428
1429         if (svc->srv_init != NULL) {
1430                 rc = svc->srv_init(thread);
1431                 if (rc)
1432                         goto out;
1433         }
1434
1435         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1436         if (rc)
1437                 goto out_srv_fini;
1438
1439         thread->t_env = &env;
1440         env.le_ctx.lc_thread = thread;
1441
1442         /* Alloc reply state structure for this one */
1443         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1444         if (!rs) {
1445                 rc = -ENOMEM;
1446                 goto out_srv_fini;
1447         }
1448
1449         /* Record that the thread is running */
1450         thread->t_flags = SVC_RUNNING;
1451         /*
1452          * wake up our creator. Note: @data is invalid after this point,
1453          * because it's allocated on ptlrpc_start_thread() stack.
1454          */
1455         cfs_waitq_signal(&thread->t_ctl_waitq);
1456
1457         watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1458                                    at_get(&svc->srv_at_estimate)) *
1459                                    svc->srv_watchdog_factor, NULL, NULL);
1460
1461         spin_lock(&svc->srv_lock);
1462         svc->srv_threads_running++;
1463         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1464         spin_unlock(&svc->srv_lock);
1465         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1466
1467         CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1468               svc->srv_threads_running);
1469
1470         /* XXX maintain a list of all managed devices: insert here */
1471
1472         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1473                svc->srv_n_difficult_replies != 0) {
1474                 /* Don't exit while there are replies to be handled */
1475                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1476                                                      ptlrpc_retry_rqbds, svc);
1477
1478                 lc_watchdog_disable(watchdog);
1479
1480                 cond_resched();
1481
1482                 l_wait_event_exclusive (svc->srv_waitq,
1483                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1484                                svc->srv_n_difficult_replies == 0) ||
1485                               (!list_empty(&svc->srv_idle_rqbds) &&
1486                                svc->srv_rqbd_timeout == 0) ||
1487                               !list_empty(&svc->srv_req_in_queue) ||
1488                               !list_empty(&svc->srv_reply_queue) ||
1489                               (!list_empty(&svc->srv_request_queue) &&
1490                                (svc->srv_n_active_reqs <
1491                                 (svc->srv_threads_running - 1))) ||
1492                               svc->srv_at_check,
1493                               &lwi);
1494
1495                 lc_watchdog_touch_ms(watchdog, max_t(int, obd_timeout,
1496                                      AT_OFF ? 0 :
1497                                      at_get(&svc->srv_at_estimate)) *
1498                                      svc->srv_watchdog_factor);
1499
1500                 ptlrpc_check_rqbd_pool(svc);
1501
1502                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1503                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1504                         /* Ignore return code - we tried... */
1505                         ptlrpc_start_thread(dev, svc);
1506                 }
1507
1508                 if (!list_empty(&svc->srv_reply_queue))
1509                         ptlrpc_server_handle_reply(svc);
1510
1511                 if (!list_empty(&svc->srv_req_in_queue)) {
1512                         /* Process all incoming reqs before handling any */
1513                         ptlrpc_server_handle_req_in(svc);
1514                         /* but limit ourselves in case of flood */
1515                         if (counter++ < 1000)
1516                                 continue;
1517                         counter = 0;
1518                 }
1519
1520                 if (svc->srv_at_check)
1521                         ptlrpc_at_check_timed(svc);
1522
1523                 /* don't handle requests in the last thread */
1524                 if (!list_empty (&svc->srv_request_queue) &&
1525                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1526                         lu_context_enter(&env.le_ctx);
1527                         ptlrpc_server_handle_request(svc, thread);
1528                         lu_context_exit(&env.le_ctx);
1529                 }
1530
1531                 if (!list_empty(&svc->srv_idle_rqbds) &&
1532                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1533                         /* I just failed to repost request buffers.  Wait
1534                          * for a timeout (unless something else happens)
1535                          * before I try again */
1536                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1537                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1538                                svc->srv_nrqbd_receiving);
1539                 }
1540         }
1541
1542         lc_watchdog_delete(watchdog);
1543
1544 out_srv_fini:
1545         /*
1546          * deconstruct service specific state created by ptlrpc_start_thread()
1547          */
1548         if (svc->srv_done != NULL)
1549                 svc->srv_done(thread);
1550
1551         lu_context_fini(&env.le_ctx);
1552 out:
1553         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1554
1555         spin_lock(&svc->srv_lock);
1556         svc->srv_threads_running--; /* must know immediately */
1557         thread->t_id = rc;
1558         thread->t_flags = SVC_STOPPED;
1559
1560         cfs_waitq_signal(&thread->t_ctl_waitq);
1561         spin_unlock(&svc->srv_lock);
1562
1563         return rc;
1564 }
1565
1566 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1567                                struct ptlrpc_thread *thread)
1568 {
1569         struct l_wait_info lwi = { 0 };
1570
1571         spin_lock(&svc->srv_lock);
1572         thread->t_flags = SVC_STOPPING;
1573         spin_unlock(&svc->srv_lock);
1574
1575         cfs_waitq_broadcast(&svc->srv_waitq);
1576         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1577                      &lwi);
1578
1579         spin_lock(&svc->srv_lock);
1580         list_del(&thread->t_link);
1581         spin_unlock(&svc->srv_lock);
1582
1583         OBD_FREE_PTR(thread);
1584 }
1585
1586 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1587 {
1588         struct ptlrpc_thread *thread;
1589
1590         spin_lock(&svc->srv_lock);
1591         while (!list_empty(&svc->srv_threads)) {
1592                 thread = list_entry(svc->srv_threads.next,
1593                                     struct ptlrpc_thread, t_link);
1594
1595                 spin_unlock(&svc->srv_lock);
1596                 ptlrpc_stop_thread(svc, thread);
1597                 spin_lock(&svc->srv_lock);
1598         }
1599
1600         spin_unlock(&svc->srv_lock);
1601 }
1602
1603 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1604 {
1605         int i, rc = 0;
1606         ENTRY;
1607
1608         /* We require 2 threads min - see note in
1609            ptlrpc_server_handle_request */
1610         LASSERT(svc->srv_threads_min >= 2);
1611         for (i = 0; i < svc->srv_threads_min; i++) {
1612                 rc = ptlrpc_start_thread(dev, svc);
1613                 /* We have enough threads, don't start more.  b=15759 */
1614                 if (rc == -EMFILE)
1615                         break;
1616                 if (rc) {
1617                         CERROR("cannot start %s thread #%d: rc %d\n",
1618                                svc->srv_thread_name, i, rc);
1619                         ptlrpc_stop_all_threads(svc);
1620                 }
1621         }
1622         RETURN(rc);
1623 }
1624
1625 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1626 {
1627         struct l_wait_info lwi = { 0 };
1628         struct ptlrpc_svc_data d;
1629         struct ptlrpc_thread *thread;
1630         char name[32];
1631         int id, rc;
1632         ENTRY;
1633
1634         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1635                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1636                svc->srv_threads_max, svc->srv_threads_running);
1637         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1638             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1639              svc->srv_threads_started == svc->srv_threads_min - 1))
1640                 RETURN(-EMFILE);
1641
1642         OBD_ALLOC_PTR(thread);
1643         if (thread == NULL)
1644                 RETURN(-ENOMEM);
1645         cfs_waitq_init(&thread->t_ctl_waitq);
1646
1647         spin_lock(&svc->srv_lock);
1648         if (svc->srv_threads_started >= svc->srv_threads_max) {
1649                 spin_unlock(&svc->srv_lock);
1650                 OBD_FREE_PTR(thread);
1651                 RETURN(-EMFILE);
1652         }
1653         list_add(&thread->t_link, &svc->srv_threads);
1654         id = svc->srv_threads_started++;
1655         spin_unlock(&svc->srv_lock);
1656
1657         thread->t_id = id;
1658         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1659         d.dev = dev;
1660         d.svc = svc;
1661         d.name = name;
1662         d.thread = thread;
1663
1664         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1665
1666           /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1667          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1668          */
1669         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1670         if (rc < 0) {
1671                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1672
1673                 spin_lock(&svc->srv_lock);
1674                 list_del(&thread->t_link);
1675                 --svc->srv_threads_started;
1676                 spin_unlock(&svc->srv_lock);
1677
1678                 OBD_FREE(thread, sizeof(*thread));
1679                 RETURN(rc);
1680         }
1681         l_wait_event(thread->t_ctl_waitq,
1682                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1683
1684         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1685         RETURN(rc);
1686 }
1687 #endif
1688
1689 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1690 {
1691         int                   rc;
1692         struct l_wait_info    lwi;
1693         struct list_head     *tmp;
1694         struct ptlrpc_reply_state *rs, *t;
1695
1696         cfs_timer_disarm(&service->srv_at_timer);
1697
1698         ptlrpc_stop_all_threads(service);
1699         LASSERT(list_empty(&service->srv_threads));
1700
1701         spin_lock (&ptlrpc_all_services_lock);
1702         list_del_init (&service->srv_list);
1703         spin_unlock (&ptlrpc_all_services_lock);
1704
1705         ptlrpc_lprocfs_unregister_service(service);
1706
1707         /* All history will be culled when the next request buffer is
1708          * freed */
1709         service->srv_max_history_rqbds = 0;
1710
1711         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1712
1713         rc = LNetClearLazyPortal(service->srv_req_portal);
1714         LASSERT (rc == 0);
1715
1716         /* Unlink all the request buffers.  This forces a 'final' event with
1717          * its 'unlink' flag set for each posted rqbd */
1718         list_for_each(tmp, &service->srv_active_rqbds) {
1719                 struct ptlrpc_request_buffer_desc *rqbd =
1720                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1721                                    rqbd_list);
1722
1723                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1724                 LASSERT (rc == 0 || rc == -ENOENT);
1725         }
1726
1727         /* Wait for the network to release any buffers it's currently
1728          * filling */
1729         for (;;) {
1730                 spin_lock(&service->srv_lock);
1731                 rc = service->srv_nrqbd_receiving;
1732                 spin_unlock(&service->srv_lock);
1733
1734                 if (rc == 0)
1735                         break;
1736
1737                 /* Network access will complete in finite time but the HUGE
1738                  * timeout lets us CWARN for visibility of sluggish NALs */
1739                 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1740                 rc = l_wait_event(service->srv_waitq,
1741                                   service->srv_nrqbd_receiving == 0,
1742                                   &lwi);
1743                 if (rc == -ETIMEDOUT)
1744                         CWARN("Service %s waiting for request buffers\n",
1745                               service->srv_name);
1746         }
1747
1748         /* schedule all outstanding replies to terminate them */
1749         spin_lock(&service->srv_lock);
1750         while (!list_empty(&service->srv_active_replies)) {
1751                 struct ptlrpc_reply_state *rs =
1752                         list_entry(service->srv_active_replies.next,
1753                                    struct ptlrpc_reply_state, rs_list);
1754                 ptlrpc_schedule_difficult_reply(rs);
1755         }
1756         spin_unlock(&service->srv_lock);
1757
1758         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1759          * and no service threads, so I'm the only thread noodling the
1760          * request queue now */
1761         while (!list_empty(&service->srv_req_in_queue)) {
1762                 struct ptlrpc_request *req =
1763                         list_entry(service->srv_req_in_queue.next,
1764                                    struct ptlrpc_request,
1765                                    rq_list);
1766
1767                 list_del(&req->rq_list);
1768                 service->srv_n_queued_reqs--;
1769                 service->srv_n_active_reqs++;
1770                 ptlrpc_server_free_request(req);
1771         }
1772         while (!list_empty(&service->srv_request_queue)) {
1773                 struct ptlrpc_request *req =
1774                         list_entry(service->srv_request_queue.next,
1775                                    struct ptlrpc_request,
1776                                    rq_list);
1777
1778                 list_del(&req->rq_list);
1779                 service->srv_n_queued_reqs--;
1780                 service->srv_n_active_reqs++;
1781
1782                 ptlrpc_server_free_request(req);
1783         }
1784         LASSERT(service->srv_n_queued_reqs == 0);
1785         LASSERT(service->srv_n_active_reqs == 0);
1786         LASSERT(service->srv_n_history_rqbds == 0);
1787         LASSERT(list_empty(&service->srv_active_rqbds));
1788
1789         /* Now free all the request buffers since nothing references them
1790          * any more... */
1791         while (!list_empty(&service->srv_idle_rqbds)) {
1792                 struct ptlrpc_request_buffer_desc *rqbd =
1793                         list_entry(service->srv_idle_rqbds.next,
1794                                    struct ptlrpc_request_buffer_desc,
1795                                    rqbd_list);
1796
1797                 ptlrpc_free_rqbd(rqbd);
1798         }
1799
1800         /* wait for all outstanding replies to complete (they were
1801          * scheduled having been flagged to abort above) */
1802         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1803                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1804
1805                 rc = l_wait_event(service->srv_waitq,
1806                                   !list_empty(&service->srv_reply_queue), &lwi);
1807                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1808
1809                 if (rc == 0) {
1810                         ptlrpc_server_handle_reply(service);
1811                         continue;
1812                 }
1813                 CWARN("Unexpectedly long timeout %p\n", service);
1814         }
1815
1816         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1817                 list_del(&rs->rs_list);
1818                 OBD_FREE(rs, service->srv_max_reply_size);
1819         }
1820
1821         /* In case somebody rearmed this in the meantime */
1822         cfs_timer_disarm(&service->srv_at_timer);
1823
1824         OBD_FREE_PTR(service);
1825         return 0;
1826 }
1827
1828 /* Returns 0 if the service is healthy.
1829  *
1830  * Right now, it just checks to make sure that requests aren't languishing
1831  * in the queue.  We'll use this health check to govern whether a node needs
1832  * to be shot, so it's intentionally non-aggressive. */
1833 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1834 {
1835         struct ptlrpc_request *request;
1836         struct timeval         right_now;
1837         long                   timediff;
1838
1839         if (svc == NULL)
1840                 return 0;
1841
1842         do_gettimeofday(&right_now);
1843
1844         spin_lock(&svc->srv_lock);
1845         if (list_empty(&svc->srv_request_queue)) {
1846                 spin_unlock(&svc->srv_lock);
1847                 return 0;
1848         }
1849
1850         /* How long has the next entry been waiting? */
1851         request = list_entry(svc->srv_request_queue.next,
1852                              struct ptlrpc_request, rq_list);
1853         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1854         spin_unlock(&svc->srv_lock);
1855
1856         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
1857                                         at_max)) {
1858                 CERROR("%s: unhealthy - request has been waiting %lds\n",
1859                        svc->srv_name, timediff / ONE_MILLION);
1860                 return (-1);
1861         }
1862
1863         return 0;
1864 }