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 /**
 404  * to actually free the request, must be called without holding svc_lock.
 405  * note it's caller's responsibility to unlink req->rq_list.
 406  */
 407 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
 408 {
 409         LASSERT(atomic_read(&req->rq_refcount) == 0);
 410         LASSERT(list_empty(&req->rq_timed_list));
 411
 412          /* DEBUG_REQ() assumes the reply state of a request with a valid
 413           * ref will not be destroyed until that reference is dropped. */
 414         ptlrpc_req_drop_rs(req);
 415
 416         sptlrpc_svc_ctx_decref(req);
 417
 418         if (req != &req->rq_rqbd->rqbd_req) {
 419                 /* NB request buffers use an embedded
 420                  * req if the incoming req unlinked the
 421                  * MD; this isn't one of them! */
 422                 OBD_FREE(req, sizeof(*req));
 423         }
 424 }
 425
 426 /**
 427  * drop a reference count of the request. if it reaches 0, we either
 428  * put it into history list, or free it immediately.
 429  */
 430 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
 431 {
 432         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 433         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 434         int                                refcount;
 435         struct list_head                  *tmp;
 436         struct list_head                  *nxt;
 437
 438         if (!atomic_dec_and_test(&req->rq_refcount))
 439                 return;
 440
 441         spin_lock(&svc->srv_lock);
 442
 443         svc->srv_n_active_reqs--;
 444         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 445
 446         refcount = --(rqbd->rqbd_refcount);
 447         if (refcount == 0) {
 448                 /* request buffer is now idle: add to history */
 449                 list_del(&rqbd->rqbd_list);
 450                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 451                 svc->srv_n_history_rqbds++;
 452
 453                 /* cull some history?
 454                  * I expect only about 1 or 2 rqbds need to be recycled here */
 455                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 456                         rqbd = list_entry(svc->srv_history_rqbds.next,
 457                                           struct ptlrpc_request_buffer_desc,
 458                                           rqbd_list);
 459
 460                         list_del(&rqbd->rqbd_list);
 461                         svc->srv_n_history_rqbds--;
 462
 463                         /* remove rqbd's reqs from svc's req history while
 464                          * I've got the service lock */
 465                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 466                                 req = list_entry(tmp, struct ptlrpc_request,
 467                                                  rq_list);
 468                                 /* Track the highest culled req seq */
 469                                 if (req->rq_history_seq >
 470                                     svc->srv_request_max_cull_seq)
 471                                         svc->srv_request_max_cull_seq =
 472                                                 req->rq_history_seq;
 473                                 list_del(&req->rq_history_list);
 474                         }
 475
 476                         spin_unlock(&svc->srv_lock);
 477
 478                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 479                                 req = list_entry(rqbd->rqbd_reqs.next,
 480                                                  struct ptlrpc_request,
 481                                                  rq_list);
 482                                 list_del(&req->rq_list);
 483                                 ptlrpc_server_free_request(req);
 484                         }
 485
 486                         spin_lock(&svc->srv_lock);
 487                         /*
 488                          * now all reqs including the embedded req has been
 489                          * disposed, schedule request buffer for re-use.
 490                          */
 491                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
 492                         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 493                 }
 494
 495                 spin_unlock(&svc->srv_lock);
 496         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 497                 /* If we are low on memory, we are not interested in history */
 498                 list_del(&req->rq_list);
 499                 list_del_init(&req->rq_history_list);
 500                 spin_unlock(&svc->srv_lock);
 501
 502                 ptlrpc_server_free_request(req);
 503         } else {
 504                 spin_unlock(&svc->srv_lock);
 505         }
 506 }
 507
 508 /**
 509  * to finish a request: stop sending more early replies, and release
 510  * the request. should be called after we finished handling the request.
 511  */
 512 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
 513 {
 514         struct ptlrpc_service  *svc = req->rq_rqbd->rqbd_service;
 515
 516         if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
 517                 DEBUG_REQ(D_INFO, req, "free req");
 518
 519         spin_lock(&svc->srv_at_lock);
 520         req->rq_sent_final = 1;
 521         list_del_init(&req->rq_timed_list);
 522         spin_unlock(&svc->srv_at_lock);
 523
 524         ptlrpc_server_drop_request(req);
 525 }
 526
 527 /* This function makes sure dead exports are evicted in a timely manner.
 528    This function is only called when some export receives a message (i.e.,
 529    the network is up.) */
 530 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 531 {
 532         struct obd_export *oldest_exp;
 533         time_t oldest_time;
 534
 535         ENTRY;
 536
 537         LASSERT(exp);
 538
 539         /* Compensate for slow machines, etc, by faking our request time
 540            into the future.  Although this can break the strict time-ordering
 541            of the list, we can be really lazy here - we don't have to evict
 542            at the exact right moment.  Eventually, all silent exports
 543            will make it to the top of the list. */
 544         exp->exp_last_request_time = max(exp->exp_last_request_time,
 545                                          cfs_time_current_sec() + extra_delay);
 546
 547         CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
 548                exp->exp_client_uuid.uuid,
 549                exp->exp_last_request_time, exp);
 550
 551         /* exports may get disconnected from the chain even though the
 552            export has references, so we must keep the spin lock while
 553            manipulating the lists */
 554         spin_lock(&exp->exp_obd->obd_dev_lock);
 555
 556         if (list_empty(&exp->exp_obd_chain_timed)) {
 557                 /* this one is not timed */
 558                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 559                 EXIT;
 560                 return;
 561         }
 562
 563         list_move_tail(&exp->exp_obd_chain_timed,
 564                        &exp->exp_obd->obd_exports_timed);
 565
 566         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 567                                 struct obd_export, exp_obd_chain_timed);
 568         oldest_time = oldest_exp->exp_last_request_time;
 569         spin_unlock(&exp->exp_obd->obd_dev_lock);
 570
 571         if (exp->exp_obd->obd_recovering) {
 572                 /* be nice to everyone during recovery */
 573                 EXIT;
 574                 return;
 575         }
 576
 577         /* Note - racing to start/reset the obd_eviction timer is safe */
 578         if (exp->exp_obd->obd_eviction_timer == 0) {
 579                 /* Check if the oldest entry is expired. */
 580                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
 581                                               extra_delay)) {
 582                         /* We need a second timer, in case the net was down and
 583                          * it just came back. Since the pinger may skip every
 584                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 585                          * we better wait for 3. */
 586                         exp->exp_obd->obd_eviction_timer =
 587                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
 588                         CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
 589                                exp->exp_obd->obd_name, obd_export_nid2str(exp),
 590                                oldest_time);
 591                 }
 592         } else {
 593                 if (cfs_time_current_sec() >
 594                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
 595                         /* The evictor won't evict anyone who we've heard from
 596                          * recently, so we don't have to check before we start
 597                          * it. */
 598                         if (!ping_evictor_wake(exp))
 599                                 exp->exp_obd->obd_eviction_timer = 0;
 600                 }
 601         }
 602
 603         EXIT;
 604 }
 605
 606 static int ptlrpc_check_req(struct ptlrpc_request *req)
 607 {
 608         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 609                      req->rq_export->exp_conn_cnt)) {
 610                 DEBUG_REQ(D_ERROR, req,
 611                           "DROPPING req from old connection %d < %d",
 612                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 613                           req->rq_export->exp_conn_cnt);
 614                 return -EEXIST;
 615         }
 616         if (unlikely(req->rq_export->exp_obd &&
 617                      req->rq_export->exp_obd->obd_fail)) {
 618              /* Failing over, don't handle any more reqs, send
 619                 error response instead. */
 620                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 621                        req, req->rq_export->exp_obd->obd_name);
 622                 req->rq_status = -ENODEV;
 623                 ptlrpc_error(req);
 624                 return -ENODEV;
 625         }
 626
 627         return 0;
 628 }
 629
 630 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
 631 {
 632         struct ptlrpc_request *rq;
 633         __s32 next;
 634
 635         spin_lock(&svc->srv_at_lock);
 636         if (list_empty(&svc->srv_at_list)) {
 637                 cfs_timer_disarm(&svc->srv_at_timer);
 638                 spin_unlock(&svc->srv_at_lock);
 639                 return;
 640         }
 641
 642         /* Set timer for closest deadline */
 643         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 644                         rq_timed_list);
 645         next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
 646                        at_early_margin);
 647         if (next <= 0)
 648                 ptlrpc_at_timer((unsigned long)svc);
 649         else
 650                 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
 651         spin_unlock(&svc->srv_at_lock);
 652         CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
 653 }
 654
 655 /* Add rpc to early reply check list */
 656 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 657 {
 658         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 659         struct ptlrpc_request *rq;
 660         int found = 0;
 661
 662         if (AT_OFF)
 663                 return(0);
 664
 665         if (req->rq_no_reply)
 666                 return 0;
 667
 668         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 669                 return(-ENOSYS);
 670
 671         spin_lock(&svc->srv_at_lock);
 672
 673         if (unlikely(req->rq_sent_final)) {
 674                 spin_unlock(&svc->srv_at_lock);
 675                 return 0;
 676         }
 677
 678         LASSERT(list_empty(&req->rq_timed_list));
 679         /* Add to sorted list.  Presumably latest rpcs will have the latest
 680            deadlines, so search backward. */
 681         list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
 682                 if (req->rq_deadline >= rq->rq_deadline) {
 683                         list_add(&req->rq_timed_list, &rq->rq_timed_list);
 684                         found++;
 685                         break;
 686                 }
 687         }
 688         if (!found)
 689                 /* Add to front if shortest deadline or list empty */
 690                 list_add(&req->rq_timed_list, &svc->srv_at_list);
 691
 692         /* Check if we're the head of the list */
 693         found = (svc->srv_at_list.next == &req->rq_timed_list);
 694
 695         spin_unlock(&svc->srv_at_lock);
 696
 697         if (found)
 698                 ptlrpc_at_set_timer(svc);
 699
 700         return 0;
 701 }
 702
 703 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
 704                                       int extra_time)
 705 {
 706         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 707         struct ptlrpc_request *reqcopy;
 708         struct lustre_msg *reqmsg;
 709         cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
 710         time_t newdl;
 711         int rc;
 712         ENTRY;
 713
 714         /* deadline is when the client expects us to reply, margin is the
 715            difference between clients' and servers' expectations */
 716         DEBUG_REQ(D_ADAPTTO, req,
 717                   "%ssending early reply (deadline %+lds, margin %+lds) for "
 718                   "%d+%d", AT_OFF ? "AT off - not " : "",
 719                   olddl, olddl - at_get(&svc->srv_at_estimate),
 720                   at_get(&svc->srv_at_estimate), extra_time);
 721
 722         if (AT_OFF)
 723                 RETURN(0);
 724
 725         if (olddl < 0) {
 726                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
 727                           "not sending early reply. Consider increasing "
 728                           "at_early_margin (%d)?", olddl, at_early_margin);
 729
 730                 /* Return an error so we're not re-added to the timed list. */
 731                 RETURN(-ETIMEDOUT);
 732         }
 733
 734         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
 735                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
 736                           "but no AT support");
 737                 RETURN(-ENOSYS);
 738         }
 739
 740         if (extra_time) {
 741                 /* Fake our processing time into the future to ask the
 742                    clients for some extra amount of time */
 743                 extra_time += cfs_time_current_sec() -
 744                         req->rq_arrival_time.tv_sec;
 745                 at_add(&svc->srv_at_estimate, extra_time);
 746         }
 747
 748         newdl = req->rq_arrival_time.tv_sec + at_get(&svc->srv_at_estimate);
 749         if (req->rq_deadline >= newdl) {
 750                 /* We're not adding any time, no need to send an early reply
 751                    (e.g. maybe at adaptive_max) */
 752                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
 753                           CFS_DURATION_T"/"CFS_DURATION_T"), "
 754                           "not sending early reply\n", olddl,
 755                           cfs_time_sub(newdl, cfs_time_current_sec()));
 756                 RETURN(-ETIMEDOUT);
 757         }
 758
 759         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 760         if (reqcopy == NULL)
 761                 RETURN(-ENOMEM);
 762         OBD_ALLOC(reqmsg, req->rq_reqlen);
 763         if (!reqmsg) {
 764                 OBD_FREE(reqcopy, sizeof *reqcopy);
 765                 RETURN(-ENOMEM);
 766         }
 767
 768         *reqcopy = *req;
 769         reqcopy->rq_reply_state = NULL;
 770         reqcopy->rq_rep_swab_mask = 0;
 771         reqcopy->rq_pack_bulk = 0;
 772         reqcopy->rq_pack_udesc = 0;
 773         reqcopy->rq_packed_final = 0;
 774         sptlrpc_svc_ctx_addref(reqcopy);
 775         /* We only need the reqmsg for the magic */
 776         reqcopy->rq_reqmsg = reqmsg;
 777         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 778
 779         if (req->rq_sent_final) {
 780                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
 781                           "abort sending early reply\n");
 782                 GOTO(out, rc = 0);
 783         }
 784
 785         /* Connection ref */
 786         reqcopy->rq_export = class_conn2export(
 787                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 788         if (reqcopy->rq_export == NULL)
 789                 GOTO(out, rc = -ENODEV);
 790
 791         /* RPC ref */
 792         class_export_rpc_get(reqcopy->rq_export);
 793         if (reqcopy->rq_export->exp_obd &&
 794             reqcopy->rq_export->exp_obd->obd_fail)
 795                 GOTO(out_put, rc = -ENODEV);
 796
 797         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 798         if (rc)
 799                 GOTO(out_put, rc);
 800
 801         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 802
 803         if (!rc) {
 804                 /* Adjust our own deadline to what we told the client */
 805                 req->rq_deadline = newdl;
 806                 req->rq_early_count++; /* number sent, server side */
 807         } else {
 808                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 809         }
 810
 811         /* Free the (early) reply state from lustre_pack_reply.
 812            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 813         ptlrpc_req_drop_rs(reqcopy);
 814
 815 out_put:
 816         class_export_rpc_put(reqcopy->rq_export);
 817         class_export_put(reqcopy->rq_export);
 818 out:
 819         sptlrpc_svc_ctx_decref(reqcopy);
 820         OBD_FREE(reqmsg, req->rq_reqlen);
 821         OBD_FREE(reqcopy, sizeof *reqcopy);
 822         RETURN(rc);
 823 }
 824
 825 /* Send early replies to everybody expiring within at_early_margin
 826    asking for at_extra time */
 827 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 828 {
 829         struct ptlrpc_request *rq, *n;
 830         struct list_head work_list;
 831         time_t now = cfs_time_current_sec();
 832         cfs_duration_t delay;
 833         int first, counter = 0;
 834         ENTRY;
 835
 836         spin_lock(&svc->srv_at_lock);
 837         if (svc->srv_at_check == 0) {
 838                 spin_unlock(&svc->srv_at_lock);
 839                 RETURN(0);
 840         }
 841         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 842         svc->srv_at_check = 0;
 843
 844         if (list_empty(&svc->srv_at_list)) {
 845                 spin_unlock(&svc->srv_at_lock);
 846                 RETURN(0);
 847         }
 848
 849         /* The timer went off, but maybe the nearest rpc already completed. */
 850         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 851                         rq_timed_list);
 852         first = (int)(rq->rq_deadline - now);
 853         if (first > at_early_margin) {
 854                 /* We've still got plenty of time.  Reset the timer. */
 855                 spin_unlock(&svc->srv_at_lock);
 856                 ptlrpc_at_set_timer(svc);
 857                 RETURN(0);
 858         }
 859
 860         /* We're close to a timeout, and we don't know how much longer the
 861            server will take. Send early replies to everyone expiring soon. */
 862         CFS_INIT_LIST_HEAD(&work_list);
 863         list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
 864                 if (rq->rq_deadline <= now + at_early_margin) {
 865                         list_move_tail(&rq->rq_timed_list, &work_list);
 866                         counter++;
 867                 } else {
 868                         break;
 869                 }
 870         }
 871
 872         spin_unlock(&svc->srv_at_lock);
 873
 874         /* we have a new earliest deadline, restart the timer */
 875         ptlrpc_at_set_timer(svc);
 876
 877         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 878                "replies\n", first, at_extra, counter);
 879         if (first < 0) {
 880                 /* We're already past request deadlines before we even get a
 881                    chance to send early replies */
 882                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 883                               "request traffic (cpu-bound).\n", svc->srv_name);
 884                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 885                       "delay="CFS_DURATION_T"(jiff)\n",
 886                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 887                       at_get(&svc->srv_at_estimate), delay);
 888         }
 889
 890         /* ptlrpc_server_finish_request may delete an entry out of
 891          * the work list */
 892         spin_lock(&svc->srv_at_lock);
 893         while (!list_empty(&work_list)) {
 894                 rq = list_entry(work_list.next, struct ptlrpc_request,
 895                                 rq_timed_list);
 896                 list_del_init(&rq->rq_timed_list);
 897                 /* if the entry is still in the worklist, it hasn't been
 898                    deleted, and is safe to take a ref to keep the req around */
 899                 atomic_inc(&rq->rq_refcount);
 900                 spin_unlock(&svc->srv_at_lock);
 901
 902                 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
 903                         ptlrpc_at_add_timed(rq);
 904
 905                 ptlrpc_server_drop_request(rq);
 906                 spin_lock(&svc->srv_at_lock);
 907         }
 908         spin_unlock(&svc->srv_at_lock);
 909
 910         RETURN(0);
 911 }
 912
 913 /* Handle freshly incoming reqs, add to timed early reply list,
 914    pass on to regular request queue */
 915 static int
 916 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
 917 {
 918         struct ptlrpc_request *req;
 919         __u32                  deadline;
 920         int                    rc;
 921         ENTRY;
 922
 923         LASSERT(svc);
 924
 925         spin_lock(&svc->srv_lock);
 926         if (list_empty(&svc->srv_req_in_queue)) {
 927                 spin_unlock(&svc->srv_lock);
 928                 RETURN(0);
 929         }
 930
 931         req = list_entry(svc->srv_req_in_queue.next,
 932                          struct ptlrpc_request, rq_list);
 933         list_del_init (&req->rq_list);
 934         /* Consider this still a "queued" request as far as stats are
 935            concerned */
 936         spin_unlock(&svc->srv_lock);
 937
 938         /* go through security check/transform */
 939         rc = sptlrpc_svc_unwrap_request(req);
 940         switch (rc) {
 941         case SECSVC_OK:
 942                 break;
 943         case SECSVC_COMPLETE:
 944                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 945                 goto err_req;
 946         case SECSVC_DROP:
 947                 goto err_req;
 948         default:
 949                 LBUG();
 950         }
 951
 952         /* Clear request swab mask; this is a new request */
 953         req->rq_req_swab_mask = 0;
 954
 955         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 956         if (rc != 0) {
 957                 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
 958                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
 959                        req->rq_xid);
 960                 goto err_req;
 961         }
 962
 963         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 964         if (rc) {
 965                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
 966                         LPU64"\n", svc->srv_req_portal,
 967                         libcfs_id2str(req->rq_peer), req->rq_xid);
 968                 goto err_req;
 969         }
 970
 971         rc = -EINVAL;
 972         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 973                 CERROR("wrong packet type received (type=%u) from %s\n",
 974                        lustre_msg_get_type(req->rq_reqmsg),
 975                        libcfs_id2str(req->rq_peer));
 976                 goto err_req;
 977         }
 978
 979         CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
 980
 981         req->rq_export = class_conn2export(
 982                 lustre_msg_get_handle(req->rq_reqmsg));
 983         if (req->rq_export) {
 984                 rc = ptlrpc_check_req(req);
 985                 if (rc == 0) {
 986                         rc = sptlrpc_target_export_check(req->rq_export, req);
 987                         if (rc)
 988                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
 989                                           "illegal security flavor,");
 990                 }
 991
 992                 class_export_put(req->rq_export);
 993                 req->rq_export = NULL;
 994                 if (rc)
 995                         goto err_req;
 996         }
 997
 998         /* req_in handling should/must be fast */
 999         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1000                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1001                           cfs_time_sub(cfs_time_current_sec(),
1002                                        req->rq_arrival_time.tv_sec));
1003
1004         /* Set rpc server deadline and add it to the timed list */
1005         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1006                     MSGHDR_AT_SUPPORT) ?
1007                    /* The max time the client expects us to take */
1008                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1009         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1010         if (unlikely(deadline == 0)) {
1011                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1012                 goto err_req;
1013         }
1014
1015         ptlrpc_at_add_timed(req);
1016
1017         /* Move it over to the request processing queue */
1018         spin_lock(&svc->srv_lock);
1019         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1020         cfs_waitq_signal(&svc->srv_waitq);
1021         spin_unlock(&svc->srv_lock);
1022         RETURN(1);
1023
1024 err_req:
1025         spin_lock(&svc->srv_lock);
1026         svc->srv_n_queued_reqs--;
1027         svc->srv_n_active_reqs++;
1028         spin_unlock(&svc->srv_lock);
1029         ptlrpc_server_finish_request(req);
1030
1031         RETURN(1);
1032 }
1033
1034 static int
1035 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1036                              struct ptlrpc_thread *thread)
1037 {
1038         struct obd_export     *export = NULL;
1039         struct ptlrpc_request *request;
1040         struct timeval         work_start;
1041         struct timeval         work_end;
1042         long                   timediff;
1043         int                    rc;
1044         ENTRY;
1045
1046         LASSERT(svc);
1047
1048         spin_lock(&svc->srv_lock);
1049         if (unlikely(list_empty (&svc->srv_request_queue) ||
1050             (
1051 #ifndef __KERNEL__
1052              /* !@%$# liblustre only has 1 thread */
1053              svc->srv_n_difficult_replies != 0 &&
1054 #endif
1055              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1056                  /* Don't handle regular requests in the last thread, in order               * re
1057                   * to handle difficult replies (which might block other threads)
1058                   * as well as handle any incoming reqs, early replies, etc.
1059                   * That means we always need at least 2 service threads. */
1060                 spin_unlock(&svc->srv_lock);
1061                 RETURN(0);
1062         }
1063
1064         request = list_entry (svc->srv_request_queue.next,
1065                               struct ptlrpc_request, rq_list);
1066         list_del_init (&request->rq_list);
1067         svc->srv_n_queued_reqs--;
1068         svc->srv_n_active_reqs++;
1069
1070         spin_unlock(&svc->srv_lock);
1071
1072         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1073                 libcfs_debug_dumplog();
1074
1075         do_gettimeofday(&work_start);
1076         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1077         if (likely(svc->srv_stats != NULL)) {
1078                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1079                                     timediff);
1080                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1081                                     svc->srv_n_queued_reqs);
1082                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1083                                     svc->srv_n_active_reqs);
1084                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1085                                     at_get(&svc->srv_at_estimate));
1086         }
1087
1088         rc = lu_context_init(&request->rq_session, LCT_SESSION);
1089         if (rc) {
1090                 CERROR("Failure to initialize session: %d\n", rc);
1091                 goto out_req;
1092         }
1093         request->rq_session.lc_thread = thread;
1094         lu_context_enter(&request->rq_session);
1095
1096         CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1097
1098         request->rq_svc_thread = thread;
1099         if (thread)
1100                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1101
1102         request->rq_export = class_conn2export(
1103                                      lustre_msg_get_handle(request->rq_reqmsg));
1104
1105         if (likely(request->rq_export)) {
1106                 if (unlikely(ptlrpc_check_req(request)))
1107                         goto put_conn;
1108                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1109                 export = class_export_rpc_get(request->rq_export);
1110         }
1111
1112         /* Discard requests queued for longer than the deadline.
1113            The deadline is increased if we send an early reply. */
1114         if (cfs_time_current_sec() > request->rq_deadline) {
1115                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1116                           ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1117                           libcfs_id2str(request->rq_peer),
1118                           cfs_time_sub(request->rq_deadline,
1119                           request->rq_arrival_time.tv_sec),
1120                           cfs_time_sub(cfs_time_current_sec(),
1121                           request->rq_deadline));
1122                 goto put_rpc_export;
1123         }
1124
1125         request->rq_phase = RQ_PHASE_INTERPRET;
1126
1127         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1128                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1129                (request->rq_export ?
1130                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1131                (request->rq_export ?
1132                 atomic_read(&request->rq_export->exp_refcount) : -99),
1133                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1134                libcfs_id2str(request->rq_peer),
1135                lustre_msg_get_opc(request->rq_reqmsg));
1136
1137         OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1138
1139         rc = svc->srv_handler(request);
1140
1141         request->rq_phase = RQ_PHASE_COMPLETE;
1142
1143         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1144                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1145                (request->rq_export ?
1146                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1147                (request->rq_export ?
1148                 atomic_read(&request->rq_export->exp_refcount) : -99),
1149                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1150                libcfs_id2str(request->rq_peer),
1151                lustre_msg_get_opc(request->rq_reqmsg));
1152
1153 put_rpc_export:
1154         if (export != NULL)
1155                 class_export_rpc_put(export);
1156 put_conn:
1157         if (likely(request->rq_export != NULL))
1158                 class_export_put(request->rq_export);
1159
1160         lu_context_exit(&request->rq_session);
1161         lu_context_fini(&request->rq_session);
1162
1163         if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1164                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1165                           "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1166                           " client may timeout.",
1167                           request->rq_xid, cfs_time_sub(request->rq_deadline,
1168                           request->rq_arrival_time.tv_sec),
1169                           cfs_time_sub(cfs_time_current_sec(),
1170                           request->rq_deadline));
1171         }
1172
1173         do_gettimeofday(&work_end);
1174         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1175         CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1176                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1177                request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1178                libcfs_id2str(request->rq_peer), timediff,
1179                cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1180                request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1181                request->rq_transno, request->rq_status,
1182                request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1183                -999);
1184         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1185                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1186                 int opc = opcode_offset(op);
1187                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1188                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1189                         lprocfs_counter_add(svc->srv_stats,
1190                                             opc + EXTRA_MAX_OPCODES,
1191                                             timediff);
1192                 }
1193         }
1194         if (unlikely(request->rq_early_count)) {
1195                 DEBUG_REQ(D_ADAPTTO, request,
1196                           "sent %d early replies before finishing in "
1197                           CFS_DURATION_T"s",
1198                           request->rq_early_count,
1199                           cfs_time_sub(work_end.tv_sec,
1200                           request->rq_arrival_time.tv_sec));
1201         }
1202
1203 out_req:
1204         ptlrpc_server_finish_request(request);
1205
1206         RETURN(1);
1207 }
1208
1209 static int
1210 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1211 {
1212         struct ptlrpc_reply_state *rs;
1213         struct obd_export         *exp;
1214         struct obd_device         *obd;
1215         int                        nlocks;
1216         int                        been_handled;
1217         ENTRY;
1218
1219         spin_lock(&svc->srv_lock);
1220         if (list_empty (&svc->srv_reply_queue)) {
1221                 spin_unlock(&svc->srv_lock);
1222                 RETURN(0);
1223         }
1224
1225         rs = list_entry (svc->srv_reply_queue.next,
1226                          struct ptlrpc_reply_state, rs_list);
1227
1228         exp = rs->rs_export;
1229         obd = exp->exp_obd;
1230
1231         LASSERT (rs->rs_difficult);
1232         LASSERT (rs->rs_scheduled);
1233
1234         list_del_init (&rs->rs_list);
1235
1236         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1237         spin_unlock(&svc->srv_lock);
1238
1239         spin_lock (&obd->obd_uncommitted_replies_lock);
1240         /* Noop if removed already */
1241         list_del_init (&rs->rs_obd_list);
1242         spin_unlock (&obd->obd_uncommitted_replies_lock);
1243
1244         spin_lock (&exp->exp_lock);
1245         /* Noop if removed already */
1246         list_del_init (&rs->rs_exp_list);
1247         spin_unlock (&exp->exp_lock);
1248
1249         spin_lock(&svc->srv_lock);
1250
1251         been_handled = rs->rs_handled;
1252         rs->rs_handled = 1;
1253
1254         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1255         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1256
1257         if (nlocks == 0 && !been_handled) {
1258                 /* If we see this, we should already have seen the warning
1259                  * in mds_steal_ack_locks()  */
1260                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1261                       " o%d NID %s\n",
1262                       rs,
1263                       rs->rs_xid, rs->rs_transno,
1264                       lustre_msg_get_opc(rs->rs_msg),
1265                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1266         }
1267
1268         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1269                 spin_unlock(&svc->srv_lock);
1270
1271                 if (!been_handled && rs->rs_on_net) {
1272                         LNetMDUnlink(rs->rs_md_h);
1273                         /* Ignore return code; we're racing with
1274                          * completion... */
1275                 }
1276
1277                 while (nlocks-- > 0)
1278                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1279                                          rs->rs_modes[nlocks]);
1280
1281                 spin_lock(&svc->srv_lock);
1282         }
1283
1284         rs->rs_scheduled = 0;
1285
1286         if (!rs->rs_on_net) {
1287                 /* Off the net */
1288                 svc->srv_n_difficult_replies--;
1289                 spin_unlock(&svc->srv_lock);
1290
1291                 class_export_put (exp);
1292                 rs->rs_export = NULL;
1293                 ptlrpc_rs_decref (rs);
1294                 atomic_dec (&svc->srv_outstanding_replies);
1295                 RETURN(1);
1296         }
1297
1298         /* still on the net; callback will schedule */
1299         spin_unlock(&svc->srv_lock);
1300         RETURN(1);
1301 }
1302
1303 #ifndef __KERNEL__
1304 /* FIXME make use of timeout later */
1305 int
1306 liblustre_check_services (void *arg)
1307 {
1308         int  did_something = 0;
1309         int  rc;
1310         struct list_head *tmp, *nxt;
1311         ENTRY;
1312
1313         /* I'm relying on being single threaded, not to have to lock
1314          * ptlrpc_all_services etc */
1315         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1316                 struct ptlrpc_service *svc =
1317                         list_entry (tmp, struct ptlrpc_service, srv_list);
1318
1319                 if (svc->srv_threads_running != 0)     /* I've recursed */
1320                         continue;
1321
1322                 /* service threads can block for bulk, so this limits us
1323                  * (arbitrarily) to recursing 1 stack frame per service.
1324                  * Note that the problem with recursion is that we have to
1325                  * unwind completely before our caller can resume. */
1326
1327                 svc->srv_threads_running++;
1328
1329                 do {
1330                         rc = ptlrpc_server_handle_req_in(svc);
1331                         rc |= ptlrpc_server_handle_reply(svc);
1332                         rc |= ptlrpc_at_check_timed(svc);
1333                         rc |= ptlrpc_server_handle_request(svc, NULL);
1334                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1335                         did_something |= rc;
1336                 } while (rc);
1337
1338                 svc->srv_threads_running--;
1339         }
1340
1341         RETURN(did_something);
1342 }
1343 #define ptlrpc_stop_all_threads(s) do {} while (0)
1344
1345 #else /* __KERNEL__ */
1346
1347 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1348 void ptlrpc_daemonize(char *name)
1349 {
1350         struct fs_struct *fs = current->fs;
1351
1352         atomic_inc(&fs->count);
1353         cfs_daemonize(name);
1354         exit_fs(cfs_current());
1355         current->fs = fs;
1356         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1357 }
1358
1359 static void
1360 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1361 {
1362         int avail = svc->srv_nrqbd_receiving;
1363         int low_water = test_req_buffer_pressure ? 0 :
1364                         svc->srv_nbuf_per_group/2;
1365
1366         /* NB I'm not locking; just looking. */
1367
1368         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1369          * allowed the request history to grow out of control.  We could put a
1370          * sanity check on that here and cull some history if we need the
1371          * space. */
1372
1373         if (avail <= low_water)
1374                 ptlrpc_grow_req_bufs(svc);
1375
1376         if (svc->srv_stats)
1377                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1378                                     avail);
1379 }
1380
1381 static int
1382 ptlrpc_retry_rqbds(void *arg)
1383 {
1384         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1385
1386         svc->srv_rqbd_timeout = 0;
1387         return (-ETIMEDOUT);
1388 }
1389
1390 static int ptlrpc_main(void *arg)
1391 {
1392         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1393         struct ptlrpc_service  *svc = data->svc;
1394         struct ptlrpc_thread   *thread = data->thread;
1395         struct obd_device      *dev = data->dev;
1396         struct ptlrpc_reply_state *rs;
1397         struct lc_watchdog     *watchdog;
1398 #ifdef WITH_GROUP_INFO
1399         struct group_info *ginfo = NULL;
1400 #endif
1401         struct lu_env env;
1402         int counter = 0, rc = 0;
1403         ENTRY;
1404
1405         ptlrpc_daemonize(data->name);
1406
1407 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1408         /* we need to do this before any per-thread allocation is done so that
1409          * we get the per-thread allocations on local node.  bug 7342 */
1410         if (svc->srv_cpu_affinity) {
1411                 int cpu, num_cpu;
1412
1413                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1414                         if (!cpu_online(cpu))
1415                                 continue;
1416                         if (num_cpu == thread->t_id % num_online_cpus())
1417                                 break;
1418                         num_cpu++;
1419                 }
1420                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1421         }
1422 #endif
1423
1424 #ifdef WITH_GROUP_INFO
1425         ginfo = groups_alloc(0);
1426         if (!ginfo) {
1427                 rc = -ENOMEM;
1428                 goto out;
1429         }
1430
1431         set_current_groups(ginfo);
1432         put_group_info(ginfo);
1433 #endif
1434
1435         if (svc->srv_init != NULL) {
1436                 rc = svc->srv_init(thread);
1437                 if (rc)
1438                         goto out;
1439         }
1440
1441         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1442         if (rc)
1443                 goto out_srv_fini;
1444
1445         thread->t_env = &env;
1446         env.le_ctx.lc_thread = thread;
1447
1448         /* Alloc reply state structure for this one */
1449         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1450         if (!rs) {
1451                 rc = -ENOMEM;
1452                 goto out_srv_fini;
1453         }
1454
1455         /* Record that the thread is running */
1456         thread->t_flags = SVC_RUNNING;
1457         /*
1458          * wake up our creator. Note: @data is invalid after this point,
1459          * because it's allocated on ptlrpc_start_thread() stack.
1460          */
1461         cfs_waitq_signal(&thread->t_ctl_waitq);
1462
1463         watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1464                                    at_get(&svc->srv_at_estimate)) *
1465                                    svc->srv_watchdog_factor, NULL, NULL);
1466
1467         spin_lock(&svc->srv_lock);
1468         svc->srv_threads_running++;
1469         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1470         spin_unlock(&svc->srv_lock);
1471         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1472
1473         CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1474               svc->srv_threads_running);
1475
1476         /* XXX maintain a list of all managed devices: insert here */
1477
1478         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1479                svc->srv_n_difficult_replies != 0) {
1480                 /* Don't exit while there are replies to be handled */
1481                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1482                                                      ptlrpc_retry_rqbds, svc);
1483
1484                 lc_watchdog_disable(watchdog);
1485
1486                 cond_resched();
1487
1488                 l_wait_event_exclusive (svc->srv_waitq,
1489                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1490                                svc->srv_n_difficult_replies == 0) ||
1491                               (!list_empty(&svc->srv_idle_rqbds) &&
1492                                svc->srv_rqbd_timeout == 0) ||
1493                               !list_empty(&svc->srv_req_in_queue) ||
1494                               !list_empty(&svc->srv_reply_queue) ||
1495                               (!list_empty(&svc->srv_request_queue) &&
1496                                (svc->srv_n_active_reqs <
1497                                 (svc->srv_threads_running - 1))) ||
1498                               svc->srv_at_check,
1499                               &lwi);
1500
1501                 lc_watchdog_touch_ms(watchdog, max_t(int, obd_timeout,
1502                                      AT_OFF ? 0 :
1503                                      at_get(&svc->srv_at_estimate)) *
1504                                      svc->srv_watchdog_factor);
1505
1506                 ptlrpc_check_rqbd_pool(svc);
1507
1508                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1509                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1510                         /* Ignore return code - we tried... */
1511                         ptlrpc_start_thread(dev, svc);
1512                 }
1513
1514                 if (!list_empty(&svc->srv_reply_queue))
1515                         ptlrpc_server_handle_reply(svc);
1516
1517                 if (!list_empty(&svc->srv_req_in_queue)) {
1518                         /* Process all incoming reqs before handling any */
1519                         ptlrpc_server_handle_req_in(svc);
1520                         /* but limit ourselves in case of flood */
1521                         if (counter++ < 1000)
1522                                 continue;
1523                         counter = 0;
1524                 }
1525
1526                 if (svc->srv_at_check)
1527                         ptlrpc_at_check_timed(svc);
1528
1529                 /* don't handle requests in the last thread */
1530                 if (!list_empty (&svc->srv_request_queue) &&
1531                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1532                         lu_context_enter(&env.le_ctx);
1533                         ptlrpc_server_handle_request(svc, thread);
1534                         lu_context_exit(&env.le_ctx);
1535                 }
1536
1537                 if (!list_empty(&svc->srv_idle_rqbds) &&
1538                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1539                         /* I just failed to repost request buffers.  Wait
1540                          * for a timeout (unless something else happens)
1541                          * before I try again */
1542                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1543                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1544                                svc->srv_nrqbd_receiving);
1545                 }
1546         }
1547
1548         lc_watchdog_delete(watchdog);
1549
1550 out_srv_fini:
1551         /*
1552          * deconstruct service specific state created by ptlrpc_start_thread()
1553          */
1554         if (svc->srv_done != NULL)
1555                 svc->srv_done(thread);
1556
1557         lu_context_fini(&env.le_ctx);
1558 out:
1559         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1560
1561         spin_lock(&svc->srv_lock);
1562         svc->srv_threads_running--; /* must know immediately */
1563         thread->t_id = rc;
1564         thread->t_flags = SVC_STOPPED;
1565
1566         cfs_waitq_signal(&thread->t_ctl_waitq);
1567         spin_unlock(&svc->srv_lock);
1568
1569         return rc;
1570 }
1571
1572 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1573                                struct ptlrpc_thread *thread)
1574 {
1575         struct l_wait_info lwi = { 0 };
1576
1577         spin_lock(&svc->srv_lock);
1578         thread->t_flags = SVC_STOPPING;
1579         spin_unlock(&svc->srv_lock);
1580
1581         cfs_waitq_broadcast(&svc->srv_waitq);
1582         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1583                      &lwi);
1584
1585         spin_lock(&svc->srv_lock);
1586         list_del(&thread->t_link);
1587         spin_unlock(&svc->srv_lock);
1588
1589         OBD_FREE_PTR(thread);
1590 }
1591
1592 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1593 {
1594         struct ptlrpc_thread *thread;
1595
1596         spin_lock(&svc->srv_lock);
1597         while (!list_empty(&svc->srv_threads)) {
1598                 thread = list_entry(svc->srv_threads.next,
1599                                     struct ptlrpc_thread, t_link);
1600
1601                 spin_unlock(&svc->srv_lock);
1602                 ptlrpc_stop_thread(svc, thread);
1603                 spin_lock(&svc->srv_lock);
1604         }
1605
1606         spin_unlock(&svc->srv_lock);
1607 }
1608
1609 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1610 {
1611         int i, rc = 0;
1612         ENTRY;
1613
1614         /* We require 2 threads min - see note in
1615            ptlrpc_server_handle_request */
1616         LASSERT(svc->srv_threads_min >= 2);
1617         for (i = 0; i < svc->srv_threads_min; i++) {
1618                 rc = ptlrpc_start_thread(dev, svc);
1619                 /* We have enough threads, don't start more.  b=15759 */
1620                 if (rc == -EMFILE)
1621                         break;
1622                 if (rc) {
1623                         CERROR("cannot start %s thread #%d: rc %d\n",
1624                                svc->srv_thread_name, i, rc);
1625                         ptlrpc_stop_all_threads(svc);
1626                 }
1627         }
1628         RETURN(rc);
1629 }
1630
1631 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1632 {
1633         struct l_wait_info lwi = { 0 };
1634         struct ptlrpc_svc_data d;
1635         struct ptlrpc_thread *thread;
1636         char name[32];
1637         int id, rc;
1638         ENTRY;
1639
1640         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1641                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1642                svc->srv_threads_max, svc->srv_threads_running);
1643         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1644             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1645              svc->srv_threads_started == svc->srv_threads_min - 1))
1646                 RETURN(-EMFILE);
1647
1648         OBD_ALLOC_PTR(thread);
1649         if (thread == NULL)
1650                 RETURN(-ENOMEM);
1651         cfs_waitq_init(&thread->t_ctl_waitq);
1652
1653         spin_lock(&svc->srv_lock);
1654         if (svc->srv_threads_started >= svc->srv_threads_max) {
1655                 spin_unlock(&svc->srv_lock);
1656                 OBD_FREE_PTR(thread);
1657                 RETURN(-EMFILE);
1658         }
1659         list_add(&thread->t_link, &svc->srv_threads);
1660         id = svc->srv_threads_started++;
1661         spin_unlock(&svc->srv_lock);
1662
1663         thread->t_id = id;
1664         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1665         d.dev = dev;
1666         d.svc = svc;
1667         d.name = name;
1668         d.thread = thread;
1669
1670         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1671
1672           /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1673          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1674          */
1675         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1676         if (rc < 0) {
1677                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1678
1679                 spin_lock(&svc->srv_lock);
1680                 list_del(&thread->t_link);
1681                 --svc->srv_threads_started;
1682                 spin_unlock(&svc->srv_lock);
1683
1684                 OBD_FREE(thread, sizeof(*thread));
1685                 RETURN(rc);
1686         }
1687         l_wait_event(thread->t_ctl_waitq,
1688                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1689
1690         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1691         RETURN(rc);
1692 }
1693 #endif
1694
1695 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1696 {
1697         int                   rc;
1698         struct l_wait_info    lwi;
1699         struct list_head     *tmp;
1700         struct ptlrpc_reply_state *rs, *t;
1701
1702         cfs_timer_disarm(&service->srv_at_timer);
1703
1704         ptlrpc_stop_all_threads(service);
1705         LASSERT(list_empty(&service->srv_threads));
1706
1707         spin_lock (&ptlrpc_all_services_lock);
1708         list_del_init (&service->srv_list);
1709         spin_unlock (&ptlrpc_all_services_lock);
1710
1711         ptlrpc_lprocfs_unregister_service(service);
1712
1713         /* All history will be culled when the next request buffer is
1714          * freed */
1715         service->srv_max_history_rqbds = 0;
1716
1717         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1718
1719         rc = LNetClearLazyPortal(service->srv_req_portal);
1720         LASSERT (rc == 0);
1721
1722         /* Unlink all the request buffers.  This forces a 'final' event with
1723          * its 'unlink' flag set for each posted rqbd */
1724         list_for_each(tmp, &service->srv_active_rqbds) {
1725                 struct ptlrpc_request_buffer_desc *rqbd =
1726                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1727                                    rqbd_list);
1728
1729                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1730                 LASSERT (rc == 0 || rc == -ENOENT);
1731         }
1732
1733         /* Wait for the network to release any buffers it's currently
1734          * filling */
1735         for (;;) {
1736                 spin_lock(&service->srv_lock);
1737                 rc = service->srv_nrqbd_receiving;
1738                 spin_unlock(&service->srv_lock);
1739
1740                 if (rc == 0)
1741                         break;
1742
1743                 /* Network access will complete in finite time but the HUGE
1744                  * timeout lets us CWARN for visibility of sluggish NALs */
1745                 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1746                 rc = l_wait_event(service->srv_waitq,
1747                                   service->srv_nrqbd_receiving == 0,
1748                                   &lwi);
1749                 if (rc == -ETIMEDOUT)
1750                         CWARN("Service %s waiting for request buffers\n",
1751                               service->srv_name);
1752         }
1753
1754         /* schedule all outstanding replies to terminate them */
1755         spin_lock(&service->srv_lock);
1756         while (!list_empty(&service->srv_active_replies)) {
1757                 struct ptlrpc_reply_state *rs =
1758                         list_entry(service->srv_active_replies.next,
1759                                    struct ptlrpc_reply_state, rs_list);
1760                 ptlrpc_schedule_difficult_reply(rs);
1761         }
1762         spin_unlock(&service->srv_lock);
1763
1764         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1765          * and no service threads, so I'm the only thread noodling the
1766          * request queue now */
1767         while (!list_empty(&service->srv_req_in_queue)) {
1768                 struct ptlrpc_request *req =
1769                         list_entry(service->srv_req_in_queue.next,
1770                                    struct ptlrpc_request,
1771                                    rq_list);
1772
1773                 list_del(&req->rq_list);
1774                 service->srv_n_queued_reqs--;
1775                 service->srv_n_active_reqs++;
1776                 ptlrpc_server_finish_request(req);
1777         }
1778         while (!list_empty(&service->srv_request_queue)) {
1779                 struct ptlrpc_request *req =
1780                         list_entry(service->srv_request_queue.next,
1781                                    struct ptlrpc_request,
1782                                    rq_list);
1783
1784                 list_del(&req->rq_list);
1785                 service->srv_n_queued_reqs--;
1786                 service->srv_n_active_reqs++;
1787
1788                 ptlrpc_server_finish_request(req);
1789         }
1790         LASSERT(service->srv_n_queued_reqs == 0);
1791         LASSERT(service->srv_n_active_reqs == 0);
1792         LASSERT(service->srv_n_history_rqbds == 0);
1793         LASSERT(list_empty(&service->srv_active_rqbds));
1794
1795         /* Now free all the request buffers since nothing references them
1796          * any more... */
1797         while (!list_empty(&service->srv_idle_rqbds)) {
1798                 struct ptlrpc_request_buffer_desc *rqbd =
1799                         list_entry(service->srv_idle_rqbds.next,
1800                                    struct ptlrpc_request_buffer_desc,
1801                                    rqbd_list);
1802
1803                 ptlrpc_free_rqbd(rqbd);
1804         }
1805
1806         /* wait for all outstanding replies to complete (they were
1807          * scheduled having been flagged to abort above) */
1808         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1809                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1810
1811                 rc = l_wait_event(service->srv_waitq,
1812                                   !list_empty(&service->srv_reply_queue), &lwi);
1813                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1814
1815                 if (rc == 0) {
1816                         ptlrpc_server_handle_reply(service);
1817                         continue;
1818                 }
1819                 CWARN("Unexpectedly long timeout %p\n", service);
1820         }
1821
1822         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1823                 list_del(&rs->rs_list);
1824                 OBD_FREE(rs, service->srv_max_reply_size);
1825         }
1826
1827         /* In case somebody rearmed this in the meantime */
1828         cfs_timer_disarm(&service->srv_at_timer);
1829
1830         OBD_FREE_PTR(service);
1831         return 0;
1832 }
1833
1834 /* Returns 0 if the service is healthy.
1835  *
1836  * Right now, it just checks to make sure that requests aren't languishing
1837  * in the queue.  We'll use this health check to govern whether a node needs
1838  * to be shot, so it's intentionally non-aggressive. */
1839 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1840 {
1841         struct ptlrpc_request *request;
1842         struct timeval         right_now;
1843         long                   timediff;
1844
1845         if (svc == NULL)
1846                 return 0;
1847
1848         do_gettimeofday(&right_now);
1849
1850         spin_lock(&svc->srv_lock);
1851         if (list_empty(&svc->srv_request_queue)) {
1852                 spin_unlock(&svc->srv_lock);
1853                 return 0;
1854         }
1855
1856         /* How long has the next entry been waiting? */
1857         request = list_entry(svc->srv_request_queue.next,
1858                              struct ptlrpc_request, rq_list);
1859         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1860         spin_unlock(&svc->srv_lock);
1861
1862         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
1863                                         at_max)) {
1864                 CERROR("%s: unhealthy - request has been waiting %lds\n",
1865                        svc->srv_name, timediff / ONE_MILLION);
1866                 return (-1);
1867         }
1868
1869         return 0;
1870 }