lustre/ptlrpc/service.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of the Lustre file system, http://www.lustre.org
   7  *   Lustre is a trademark of Cluster File Systems, Inc.
   8  *
   9  *   You may have signed or agreed to another license before downloading
  10  *   this software.  If so, you are bound by the terms and conditions
  11  *   of that agreement, and the following does not apply to you.  See the
  12  *   LICENSE file included with this distribution for more information.
  13  *
  14  *   If you did not agree to a different license, then this copy of Lustre
  15  *   is open source software; you can redistribute it and/or modify it
  16  *   under the terms of version 2 of the GNU General Public License as
  17  *   published by the Free Software Foundation.
  18  *
  19  *   In either case, Lustre is distributed in the hope that it will be
  20  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  21  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  22  *   license text for more details.
  23  *
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 #include <liblustre.h>
  29 #endif
  30 #include <obd_support.h>
  31 #include <obd_class.h>
  32 #include <lustre_net.h>
  33 #include <lu_object.h>
  34 #include <lnet/types.h>
  35 #include "ptlrpc_internal.h"
  36
  37 /* The following are visible and mutable through /sys/module/ptlrpc */
  38 int test_req_buffer_pressure = 0;
  39 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
  40                 "set non-zero to put pressure on request buffer pools");
  41 unsigned int at_min = 0;
  42 CFS_MODULE_PARM(at_min, "i", int, 0644,
  43                 "Adaptive timeout minimum (sec)");
  44 unsigned int at_max = 600;
  45 EXPORT_SYMBOL(at_max);
  46 CFS_MODULE_PARM(at_max, "i", int, 0644,
  47                 "Adaptive timeout maximum (sec)");
  48 unsigned int at_history = 600;
  49 CFS_MODULE_PARM(at_history, "i", int, 0644,
  50                 "Adaptive timeouts remember the slowest event that took place "
  51                 "within this period (sec)");
  52 static int at_early_margin = 5;
  53 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
  54                 "How soon before an RPC deadline to send an early reply");
  55 static int at_extra = 30;
  56 CFS_MODULE_PARM(at_extra, "i", int, 0644,
  57                 "How much extra time to give with each early reply");
  58
  59
  60 /* forward ref */
  61 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
  62
  63 static CFS_LIST_HEAD(ptlrpc_all_services);
  64 spinlock_t ptlrpc_all_services_lock;
  65
  66 static char *
  67 ptlrpc_alloc_request_buffer (int size)
  68 {
  69         char *ptr;
  70
  71         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  72                 OBD_VMALLOC(ptr, size);
  73         else
  74                 OBD_ALLOC(ptr, size);
  75
  76         return (ptr);
  77 }
  78
  79 static void
  80 ptlrpc_free_request_buffer (char *ptr, int size)
  81 {
  82         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  83                 OBD_VFREE(ptr, size);
  84         else
  85                 OBD_FREE(ptr, size);
  86 }
  87
  88 struct ptlrpc_request_buffer_desc *
  89 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
  90 {
  91         struct ptlrpc_request_buffer_desc *rqbd;
  92
  93         OBD_ALLOC_PTR(rqbd);
  94         if (rqbd == NULL)
  95                 return (NULL);
  96
  97         rqbd->rqbd_service = svc;
  98         rqbd->rqbd_refcount = 0;
  99         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
 100         rqbd->rqbd_cbid.cbid_arg = rqbd;
 101         CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
 102         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
 103
 104         if (rqbd->rqbd_buffer == NULL) {
 105                 OBD_FREE_PTR(rqbd);
 106                 return (NULL);
 107         }
 108
 109         spin_lock(&svc->srv_lock);
 110         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 111         svc->srv_nbufs++;
 112         spin_unlock(&svc->srv_lock);
 113
 114         return (rqbd);
 115 }
 116
 117 void
 118 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 119 {
 120         struct ptlrpc_service *svc = rqbd->rqbd_service;
 121
 122         LASSERT (rqbd->rqbd_refcount == 0);
 123         LASSERT (list_empty(&rqbd->rqbd_reqs));
 124
 125         spin_lock(&svc->srv_lock);
 126         list_del(&rqbd->rqbd_list);
 127         svc->srv_nbufs--;
 128         spin_unlock(&svc->srv_lock);
 129
 130         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 131         OBD_FREE_PTR(rqbd);
 132 }
 133
 134 int
 135 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
 136 {
 137         struct ptlrpc_request_buffer_desc *rqbd;
 138         int                                i;
 139
 140         CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
 141                svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
 142                svc->srv_nrqbd_receiving, svc->srv_nbufs);
 143         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 144                 rqbd = ptlrpc_alloc_rqbd(svc);
 145
 146                 if (rqbd == NULL) {
 147                         CERROR ("%s: Can't allocate request buffer\n",
 148                                 svc->srv_name);
 149                         return (-ENOMEM);
 150                 }
 151
 152                 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
 153                         return (-EAGAIN);
 154         }
 155
 156         return (0);
 157 }
 158
 159 void
 160 ptlrpc_save_lock (struct ptlrpc_request *req,
 161                   struct lustre_handle *lock, int mode)
 162 {
 163         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 164         int                        idx;
 165
 166         LASSERT(rs != NULL);
 167         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
 168
 169         idx = rs->rs_nlocks++;
 170         rs->rs_locks[idx] = *lock;
 171         rs->rs_modes[idx] = mode;
 172         rs->rs_difficult = 1;
 173 }
 174
 175 void
 176 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 177 {
 178         struct ptlrpc_service *svc = rs->rs_service;
 179
 180 #ifdef CONFIG_SMP
 181         LASSERT (spin_is_locked (&svc->srv_lock));
 182 #endif
 183         LASSERT (rs->rs_difficult);
 184         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 185
 186         if (rs->rs_scheduled)                   /* being set up or already notified */
 187                 return;
 188
 189         rs->rs_scheduled = 1;
 190         list_del (&rs->rs_list);
 191         list_add (&rs->rs_list, &svc->srv_reply_queue);
 192         cfs_waitq_signal (&svc->srv_waitq);
 193 }
 194
 195 void
 196 ptlrpc_commit_replies (struct obd_device *obd)
 197 {
 198         struct list_head   *tmp;
 199         struct list_head   *nxt;
 200
 201         /* Find any replies that have been committed and get their service
 202          * to attend to complete them. */
 203
 204         /* CAVEAT EMPTOR: spinlock ordering!!! */
 205         spin_lock(&obd->obd_uncommitted_replies_lock);
 206
 207         list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
 208                 struct ptlrpc_reply_state *rs =
 209                         list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
 210
 211                 LASSERT (rs->rs_difficult);
 212
 213                 if (rs->rs_transno <= obd->obd_last_committed) {
 214                         struct ptlrpc_service *svc = rs->rs_service;
 215
 216                         spin_lock (&svc->srv_lock);
 217                         list_del_init (&rs->rs_obd_list);
 218                         ptlrpc_schedule_difficult_reply (rs);
 219                         spin_unlock (&svc->srv_lock);
 220                 }
 221         }
 222
 223         spin_unlock(&obd->obd_uncommitted_replies_lock);
 224 }
 225
 226 static int
 227 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 228 {
 229         struct ptlrpc_request_buffer_desc *rqbd;
 230         int                                rc;
 231         int                                posted = 0;
 232
 233         for (;;) {
 234                 spin_lock(&svc->srv_lock);
 235
 236                 if (list_empty (&svc->srv_idle_rqbds)) {
 237                         spin_unlock(&svc->srv_lock);
 238                         return (posted);
 239                 }
 240
 241                 rqbd = list_entry(svc->srv_idle_rqbds.next,
 242                                   struct ptlrpc_request_buffer_desc,
 243                                   rqbd_list);
 244                 list_del (&rqbd->rqbd_list);
 245
 246                 /* assume we will post successfully */
 247                 svc->srv_nrqbd_receiving++;
 248                 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
 249
 250                 spin_unlock(&svc->srv_lock);
 251
 252                 rc = ptlrpc_register_rqbd(rqbd);
 253                 if (rc != 0)
 254                         break;
 255
 256                 posted = 1;
 257         }
 258
 259         spin_lock(&svc->srv_lock);
 260
 261         svc->srv_nrqbd_receiving--;
 262         list_del(&rqbd->rqbd_list);
 263         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 264
 265         /* Don't complain if no request buffers are posted right now; LNET
 266          * won't drop requests because we set the portal lazy! */
 267
 268         spin_unlock(&svc->srv_lock);
 269
 270         return (-1);
 271 }
 272
 273 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
 274                                             svc_handler_t h, char *name,
 275                                             struct proc_dir_entry *proc_entry,
 276                                             svcreq_printfn_t prntfn,
 277                                             char *threadname)
 278 {
 279         return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
 280                                c->psc_max_req_size, c->psc_max_reply_size,
 281                                c->psc_req_portal, c->psc_rep_portal,
 282                                c->psc_watchdog_factor,
 283                                h, name, proc_entry,
 284                                prntfn, c->psc_min_threads, c->psc_max_threads,
 285                                threadname, c->psc_ctx_tags);
 286 }
 287 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
 288
 289 static void ptlrpc_at_timer(unsigned long castmeharder)
 290 {
 291         struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
 292         CDEBUG(D_INFO, "at timer %s hit at %ld%s\n",
 293                svc->srv_name, cfs_time_current_sec(),
 294                list_empty(&svc->srv_at_list) ? ", empty" : "");
 295         svc->srv_at_check = 1;
 296         svc->srv_at_checktime = cfs_time_current();
 297         cfs_waitq_signal(&svc->srv_waitq);
 298 }
 299
 300 /* @threadname should be 11 characters or less - 3 will be added on */
 301 struct ptlrpc_service *
 302 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
 303                 int req_portal, int rep_portal, int watchdog_factor,
 304                 svc_handler_t handler, char *name,
 305                 cfs_proc_dir_entry_t *proc_entry,
 306                 svcreq_printfn_t svcreq_printfn,
 307                 int min_threads, int max_threads,
 308                 char *threadname, __u32 ctx_tags)
 309 {
 310         int                    rc;
 311         struct ptlrpc_service *service;
 312         ENTRY;
 313
 314         LASSERT (nbufs > 0);
 315         LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
 316         LASSERT (ctx_tags != 0);
 317
 318         OBD_ALLOC_PTR(service);
 319         if (service == NULL)
 320                 RETURN(NULL);
 321
 322         /* First initialise enough for early teardown */
 323
 324         service->srv_name = name;
 325         spin_lock_init(&service->srv_lock);
 326         CFS_INIT_LIST_HEAD(&service->srv_threads);
 327         cfs_waitq_init(&service->srv_waitq);
 328
 329         service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
 330         service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
 331         service->srv_buf_size = bufsize;
 332         service->srv_rep_portal = rep_portal;
 333         service->srv_req_portal = req_portal;
 334         service->srv_watchdog_factor = watchdog_factor;
 335         service->srv_handler = handler;
 336         service->srv_request_history_print_fn = svcreq_printfn;
 337         service->srv_request_seq = 1;           /* valid seq #s start at 1 */
 338         service->srv_request_max_cull_seq = 0;
 339         service->srv_threads_min = min_threads;
 340         service->srv_threads_max = max_threads;
 341         service->srv_thread_name = threadname;
 342         service->srv_ctx_tags = ctx_tags;
 343
 344         rc = LNetSetLazyPortal(service->srv_req_portal);
 345         LASSERT (rc == 0);
 346
 347         CFS_INIT_LIST_HEAD(&service->srv_request_queue);
 348         CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
 349         CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
 350         CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
 351         CFS_INIT_LIST_HEAD(&service->srv_request_history);
 352         CFS_INIT_LIST_HEAD(&service->srv_active_replies);
 353         CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
 354         CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
 355         cfs_waitq_init(&service->srv_free_rs_waitq);
 356
 357         spin_lock_init(&service->srv_at_lock);
 358         CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
 359         CFS_INIT_LIST_HEAD(&service->srv_at_list);
 360         cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
 361         /* At SOW, service time should be quick; 10s seems generous. If client
 362            timeout is less than this, we'll be sending an early reply. */
 363         at_init(&service->srv_at_estimate, 10, 0);
 364
 365         spin_lock (&ptlrpc_all_services_lock);
 366         list_add (&service->srv_list, &ptlrpc_all_services);
 367         spin_unlock (&ptlrpc_all_services_lock);
 368
 369         /* Now allocate the request buffers */
 370         rc = ptlrpc_grow_req_bufs(service);
 371         /* We shouldn't be under memory pressure at startup, so
 372          * fail if we can't post all our buffers at this time. */
 373         if (rc != 0)
 374                 GOTO(failed, NULL);
 375
 376         /* Now allocate pool of reply buffers */
 377         /* Increase max reply size to next power of two */
 378         service->srv_max_reply_size = 1;
 379         while (service->srv_max_reply_size <
 380                max_reply_size + SPTLRPC_MAX_PAYLOAD)
 381                 service->srv_max_reply_size <<= 1;
 382
 383         if (proc_entry != NULL)
 384                 ptlrpc_lprocfs_register_service(proc_entry, service);
 385
 386         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 387                service->srv_name, service->srv_req_portal);
 388
 389         RETURN(service);
 390 failed:
 391         ptlrpc_unregister_service(service);
 392         return NULL;
 393 }
 394
 395 static void ptlrpc_server_req_decref(struct ptlrpc_request *req)
 396 {
 397         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 398
 399         if (!atomic_dec_and_test(&req->rq_refcount))
 400                 return;
 401
 402         sptlrpc_svc_ctx_decref(req);
 403
 404         LASSERT(list_empty(&req->rq_timed_list));
 405         if (req != &rqbd->rqbd_req) {
 406                 /* NB request buffers use an embedded
 407                  * req if the incoming req unlinked the
 408                  * MD; this isn't one of them! */
 409                 OBD_FREE(req, sizeof(*req));
 410         } else {
 411                 struct ptlrpc_service *svc = rqbd->rqbd_service;
 412                 /* schedule request buffer for re-use.
 413                  * NB I can only do this after I've disposed of their
 414                  * reqs; particularly the embedded req */
 415                 spin_lock(&svc->srv_lock);
 416                 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 417                 spin_unlock(&svc->srv_lock);
 418         }
 419 }
 420
 421 static void __ptlrpc_server_free_request(struct ptlrpc_request *req)
 422 {
 423         list_del(&req->rq_list);
 424         ptlrpc_req_drop_rs(req);
 425         ptlrpc_server_req_decref(req);
 426 }
 427
 428 static void
 429 ptlrpc_server_free_request(struct ptlrpc_request *req)
 430 {
 431         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 432         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 433         int                                refcount;
 434         struct list_head                  *tmp;
 435         struct list_head                  *nxt;
 436
 437         if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
 438                 DEBUG_REQ(D_INFO, req, "free req");
 439         spin_lock(&svc->srv_at_lock);
 440         req->rq_sent_final = 1;
 441         list_del_init(&req->rq_timed_list);
 442         spin_unlock(&svc->srv_at_lock);
 443
 444         spin_lock(&svc->srv_lock);
 445
 446         svc->srv_n_active_reqs--;
 447         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 448
 449         refcount = --(rqbd->rqbd_refcount);
 450         if (refcount == 0) {
 451                 /* request buffer is now idle: add to history */
 452                 list_del(&rqbd->rqbd_list);
 453                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 454                 svc->srv_n_history_rqbds++;
 455
 456                 /* cull some history?
 457                  * I expect only about 1 or 2 rqbds need to be recycled here */
 458                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 459                         rqbd = list_entry(svc->srv_history_rqbds.next,
 460                                           struct ptlrpc_request_buffer_desc,
 461                                           rqbd_list);
 462
 463                         list_del(&rqbd->rqbd_list);
 464                         svc->srv_n_history_rqbds--;
 465
 466                         /* remove rqbd's reqs from svc's req history while
 467                          * I've got the service lock */
 468                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 469                                 req = list_entry(tmp, struct ptlrpc_request,
 470                                                  rq_list);
 471                                 /* Track the highest culled req seq */
 472                                 if (req->rq_history_seq >
 473                                     svc->srv_request_max_cull_seq)
 474                                         svc->srv_request_max_cull_seq =
 475                                                 req->rq_history_seq;
 476                                 list_del(&req->rq_history_list);
 477                         }
 478
 479                         spin_unlock(&svc->srv_lock);
 480
 481                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 482                                 req = list_entry(rqbd->rqbd_reqs.next,
 483                                                  struct ptlrpc_request,
 484                                                  rq_list);
 485                                 __ptlrpc_server_free_request(req);
 486                         }
 487
 488                         spin_lock(&svc->srv_lock);
 489                 }
 490         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 491                  /* If we are low on memory, we are not interested in
 492                     history */
 493                 list_del(&req->rq_history_list);
 494                 __ptlrpc_server_free_request(req);
 495         }
 496
 497         spin_unlock(&svc->srv_lock);
 498
 499 }
 500
 501 /* This function makes sure dead exports are evicted in a timely manner.
 502    This function is only called when some export receives a message (i.e.,
 503    the network is up.) */
 504 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 505 {
 506         struct obd_export *oldest_exp;
 507         time_t oldest_time;
 508
 509         ENTRY;
 510
 511         LASSERT(exp);
 512
 513         /* Compensate for slow machines, etc, by faking our request time
 514            into the future.  Although this can break the strict time-ordering
 515            of the list, we can be really lazy here - we don't have to evict
 516            at the exact right moment.  Eventually, all silent exports
 517            will make it to the top of the list. */
 518         exp->exp_last_request_time = max(exp->exp_last_request_time,
 519                                          cfs_time_current_sec() + extra_delay);
 520
 521         CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
 522                exp->exp_client_uuid.uuid,
 523                exp->exp_last_request_time, exp);
 524
 525         /* exports may get disconnected from the chain even though the
 526            export has references, so we must keep the spin lock while
 527            manipulating the lists */
 528         spin_lock(&exp->exp_obd->obd_dev_lock);
 529
 530         if (list_empty(&exp->exp_obd_chain_timed)) {
 531                 /* this one is not timed */
 532                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 533                 EXIT;
 534                 return;
 535         }
 536
 537         list_move_tail(&exp->exp_obd_chain_timed,
 538                        &exp->exp_obd->obd_exports_timed);
 539
 540         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 541                                 struct obd_export, exp_obd_chain_timed);
 542         oldest_time = oldest_exp->exp_last_request_time;
 543         spin_unlock(&exp->exp_obd->obd_dev_lock);
 544
 545         if (exp->exp_obd->obd_recovering) {
 546                 /* be nice to everyone during recovery */
 547                 EXIT;
 548                 return;
 549         }
 550
 551         /* Note - racing to start/reset the obd_eviction timer is safe */
 552         if (exp->exp_obd->obd_eviction_timer == 0) {
 553                 /* Check if the oldest entry is expired. */
 554                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
 555                                               extra_delay)) {
 556                         /* We need a second timer, in case the net was down and
 557                          * it just came back. Since the pinger may skip every
 558                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 559                          * we better wait for 3. */
 560                         exp->exp_obd->obd_eviction_timer =
 561                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
 562                         CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
 563                                exp->exp_obd->obd_name, obd_export_nid2str(exp),
 564                                oldest_time);
 565                 }
 566         } else {
 567                 if (cfs_time_current_sec() >
 568                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
 569                         /* The evictor won't evict anyone who we've heard from
 570                          * recently, so we don't have to check before we start
 571                          * it. */
 572                         if (!ping_evictor_wake(exp))
 573                                 exp->exp_obd->obd_eviction_timer = 0;
 574                 }
 575         }
 576
 577         EXIT;
 578 }
 579
 580 #ifndef __KERNEL__
 581 int lu_context_init(struct lu_context *ctx, __u32 tags)
 582 {
 583         return 0;
 584 }
 585
 586 void lu_context_fini(struct lu_context *ctx)
 587 {
 588 }
 589
 590 void lu_context_enter(struct lu_context *ctx)
 591 {
 592 }
 593
 594 void lu_context_exit(struct lu_context *ctx)
 595 {
 596 }
 597
 598 #endif
 599
 600 static int ptlrpc_check_req(struct ptlrpc_request *req)
 601 {
 602         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 603                      req->rq_export->exp_conn_cnt)) {
 604                 DEBUG_REQ(D_ERROR, req,
 605                           "DROPPING req from old connection %d < %d",
 606                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 607                           req->rq_export->exp_conn_cnt);
 608                 return -EEXIST;
 609         }
 610         if (unlikely(req->rq_export->exp_obd &&
 611                      req->rq_export->exp_obd->obd_fail)) {
 612              /* Failing over, don't handle any more reqs, send
 613                 error response instead. */
 614                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 615                        req, req->rq_export->exp_obd->obd_name);
 616                 req->rq_status = -ENODEV;
 617                 ptlrpc_error(req);
 618                 return -ENODEV;
 619         }
 620
 621         return 0;
 622 }
 623
 624 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
 625 {
 626         struct ptlrpc_request *rq;
 627         __s32 next;
 628
 629         spin_lock(&svc->srv_at_lock);
 630         if (list_empty(&svc->srv_at_list)) {
 631                 cfs_timer_disarm(&svc->srv_at_timer);
 632                 spin_unlock(&svc->srv_at_lock);
 633                 return;
 634         }
 635
 636         /* Set timer for closest deadline */
 637         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 638                         rq_timed_list);
 639         next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
 640                        at_early_margin);
 641         if (next <= 0)
 642                 ptlrpc_at_timer((unsigned long)svc);
 643         else
 644                 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
 645         spin_unlock(&svc->srv_at_lock);
 646         CDEBUG(D_INFO, "armed %s at %+lds\n", svc->srv_name, next);
 647 }
 648
 649 /* Add rpc to early reply check list */
 650 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 651 {
 652         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 653         struct ptlrpc_request *rq;
 654         int found = 0;
 655
 656         if (AT_OFF)
 657                 return(0);
 658
 659         if (req->rq_no_reply)
 660                 return 0;
 661
 662         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 663                 return(-ENOSYS);
 664
 665         DEBUG_REQ(D_ADAPTTO, req, "add timed %lds",
 666                   req->rq_deadline - cfs_time_current_sec());
 667
 668         spin_lock(&svc->srv_at_lock);
 669
 670         if (unlikely(req->rq_sent_final)) {
 671                 spin_unlock(&svc->srv_at_lock);
 672                 return 0;
 673         }
 674
 675         LASSERT(list_empty(&req->rq_timed_list));
 676         /* Add to sorted list.  Presumably latest rpcs will have the latest
 677            deadlines, so search backward. */
 678         list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
 679                 if (req->rq_deadline > rq->rq_deadline) {
 680                         list_add(&req->rq_timed_list, &rq->rq_timed_list);
 681                         found++;
 682                         break;
 683                 }
 684         }
 685         if (!found)
 686                 /* Add to front if shortest deadline or list empty */
 687                 list_add(&req->rq_timed_list, &svc->srv_at_list);
 688
 689         /* Check if we're the head of the list */
 690         found = (svc->srv_at_list.next == &req->rq_timed_list);
 691
 692         spin_unlock(&svc->srv_at_lock);
 693
 694         if (found)
 695                 ptlrpc_at_set_timer(svc);
 696
 697         return 0;
 698 }
 699
 700 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
 701                                       int extra_time)
 702 {
 703         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 704         struct ptlrpc_request *reqcopy;
 705         struct lustre_msg *reqmsg;
 706         long olddl = req->rq_deadline - cfs_time_current_sec();
 707         time_t newdl;
 708         int rc;
 709         ENTRY;
 710
 711         /* deadline is when the client expects us to reply, margin is the
 712            difference between clients' and servers' expectations */
 713         DEBUG_REQ(D_ADAPTTO, req,
 714                   "%ssending early reply (deadline %+lds, margin %+lds) for "
 715                   "%d+%d", AT_OFF ? "AT off - not " : "",
 716                   olddl, olddl - at_get(&svc->srv_at_estimate),
 717                   at_get(&svc->srv_at_estimate), extra_time);
 718
 719         if (AT_OFF)
 720                 RETURN(0);
 721
 722         if (olddl < 0) {
 723                 CDEBUG(D_WARNING, "x"LPU64": Already past deadline (%+lds), not"
 724                        " sending early reply. Increase at_early_margin (%d)?\n",
 725                        req->rq_xid, olddl, at_early_margin);
 726                 /* Return an error so we're not re-added to the timed list. */
 727                 RETURN(-ETIMEDOUT);
 728         }
 729
 730         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
 731                 CDEBUG(D_INFO, "Wanted to ask client for more time, but no AT "
 732                       "support\n");
 733                 RETURN(-ENOSYS);
 734         }
 735
 736         if (extra_time) {
 737                 /* Fake our processing time into the future to ask the
 738                    clients for some extra amount of time */
 739                 extra_time += cfs_time_current_sec() -
 740                         req->rq_arrival_time.tv_sec;
 741                 at_add(&svc->srv_at_estimate, extra_time);
 742         }
 743
 744         newdl = req->rq_arrival_time.tv_sec + at_get(&svc->srv_at_estimate);
 745         if (req->rq_deadline >= newdl) {
 746                 /* We're not adding any time, no need to send an early reply
 747                    (e.g. maybe at adaptive_max) */
 748                 CDEBUG(D_ADAPTTO, "x"LPU64": Couldn't add any time (%ld/%ld), "
 749                        "not sending early reply\n", req->rq_xid, olddl,
 750                        newdl - cfs_time_current_sec());
 751                 RETURN(-ETIMEDOUT);
 752         }
 753
 754         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 755         if (reqcopy == NULL)
 756                 RETURN(-ENOMEM);
 757         OBD_ALLOC(reqmsg, req->rq_reqlen);
 758         if (!reqmsg) {
 759                 OBD_FREE(reqcopy, sizeof *reqcopy);
 760                 RETURN(-ENOMEM);
 761         }
 762
 763         *reqcopy = *req;
 764         reqcopy->rq_reply_state = NULL;
 765         reqcopy->rq_rep_swab_mask = 0;
 766         reqcopy->rq_pack_bulk = 0;
 767         reqcopy->rq_pack_udesc = 0;
 768         reqcopy->rq_packed_final = 0;
 769         sptlrpc_svc_ctx_addref(reqcopy);
 770         /* We only need the reqmsg for the magic */
 771         reqcopy->rq_reqmsg = reqmsg;
 772         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 773
 774         if (req->rq_sent_final) {
 775                 CDEBUG(D_ADAPTTO, "x"LPU64": normal reply already sent out, "
 776                        "abort sending early reply\n", req->rq_xid);
 777                 GOTO(out, rc = 0);
 778         }
 779
 780         /* Connection ref */
 781         reqcopy->rq_export = class_conn2export(
 782                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 783         if (reqcopy->rq_export == NULL)
 784                 GOTO(out, rc = -ENODEV);
 785
 786         /* RPC ref */
 787         class_export_rpc_get(reqcopy->rq_export);
 788         if (reqcopy->rq_export->exp_obd &&
 789             reqcopy->rq_export->exp_obd->obd_fail)
 790                 GOTO(out_put, rc = -ENODEV);
 791
 792         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 793         if (rc)
 794                 GOTO(out_put, rc);
 795
 796         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 797
 798         if (!rc) {
 799                 /* Adjust our own deadline to what we told the client */
 800                 req->rq_deadline = newdl;
 801                 req->rq_early_count++; /* number sent, server side */
 802         } else {
 803                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 804         }
 805
 806         /* Free the (early) reply state from lustre_pack_reply.
 807            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 808         ptlrpc_req_drop_rs(reqcopy);
 809
 810 out_put:
 811         class_export_rpc_put(reqcopy->rq_export);
 812         class_export_put(reqcopy->rq_export);
 813 out:
 814         sptlrpc_svc_ctx_decref(reqcopy);
 815         OBD_FREE(reqmsg, req->rq_reqlen);
 816         OBD_FREE(reqcopy, sizeof *reqcopy);
 817         RETURN(rc);
 818 }
 819
 820 /* Send early replies to everybody expiring within at_early_margin
 821    asking for at_extra time */
 822 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 823 {
 824         struct ptlrpc_request *rq, *n;
 825         struct list_head work_list;
 826         time_t now = cfs_time_current_sec();
 827         cfs_duration_t delay;
 828         int first, counter = 0;
 829         ENTRY;
 830
 831         spin_lock(&svc->srv_at_lock);
 832         if (svc->srv_at_check == 0) {
 833                 spin_unlock(&svc->srv_at_lock);
 834                 RETURN(0);
 835         }
 836         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 837         svc->srv_at_check = 0;
 838
 839         if (list_empty(&svc->srv_at_list)) {
 840                 spin_unlock(&svc->srv_at_lock);
 841                 RETURN(0);
 842         }
 843
 844         /* The timer went off, but maybe the nearest rpc already completed. */
 845         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 846                         rq_timed_list);
 847         first = (int)(rq->rq_deadline - now);
 848         if (first > at_early_margin) {
 849                 /* We've still got plenty of time.  Reset the timer. */
 850                 spin_unlock(&svc->srv_at_lock);
 851                 ptlrpc_at_set_timer(svc);
 852                 RETURN(0);
 853         }
 854
 855         /* We're close to a timeout, and we don't know how much longer the
 856            server will take. Send early replies to everyone expiring soon. */
 857         CFS_INIT_LIST_HEAD(&work_list);
 858         list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
 859                 if (rq->rq_deadline <= now + at_early_margin) {
 860                         list_move_tail(&rq->rq_timed_list, &work_list);
 861                         counter++;
 862                 } else {
 863                         break;
 864                 }
 865         }
 866
 867         spin_unlock(&svc->srv_at_lock);
 868
 869         /* we have a new earliest deadline, restart the timer */
 870         ptlrpc_at_set_timer(svc);
 871
 872         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 873                "replies\n", first, at_extra, counter);
 874         if (first < 0) {
 875                 /* We're already past request deadlines before we even get a
 876                    chance to send early replies */
 877                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 878                               "request traffic (cpu-bound).\n", svc->srv_name);
 879                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 880                       "delay="CFS_DURATION_T"(jiff)\n",
 881                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 882                       at_get(&svc->srv_at_estimate), delay);
 883         }
 884
 885         /* ptlrpc_server_free_request may delete an entry out of the work
 886            list */
 887         spin_lock(&svc->srv_at_lock);
 888         while (!list_empty(&work_list)) {
 889                 rq = list_entry(work_list.next, struct ptlrpc_request,
 890                                 rq_timed_list);
 891                 list_del_init(&rq->rq_timed_list);
 892                 /* if the entry is still in the worklist, it hasn't been
 893                    deleted, and is safe to take a ref to keep the req around */
 894                 atomic_inc(&rq->rq_refcount);
 895                 spin_unlock(&svc->srv_at_lock);
 896
 897                 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
 898                         ptlrpc_at_add_timed(rq);
 899
 900                 ptlrpc_server_req_decref(rq);
 901                 spin_lock(&svc->srv_at_lock);
 902         }
 903         spin_unlock(&svc->srv_at_lock);
 904
 905         RETURN(0);
 906 }
 907
 908 /* Handle freshly incoming reqs, add to timed early reply list,
 909    pass on to regular request queue */
 910 static int
 911 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
 912 {
 913         struct ptlrpc_request *req;
 914         __u32                  deadline;
 915         int                    rc;
 916         ENTRY;
 917
 918         LASSERT(svc);
 919
 920         spin_lock(&svc->srv_lock);
 921         if (list_empty(&svc->srv_req_in_queue)) {
 922                 spin_unlock(&svc->srv_lock);
 923                 RETURN(0);
 924         }
 925
 926         req = list_entry(svc->srv_req_in_queue.next,
 927                          struct ptlrpc_request, rq_list);
 928         list_del_init (&req->rq_list);
 929         /* Consider this still a "queued" request as far as stats are
 930            concerned */
 931         spin_unlock(&svc->srv_lock);
 932
 933         /* go through security check/transform */
 934         rc = sptlrpc_svc_unwrap_request(req);
 935         switch (rc) {
 936         case SECSVC_OK:
 937                 break;
 938         case SECSVC_COMPLETE:
 939                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 940                 goto err_req;
 941         case SECSVC_DROP:
 942                 goto err_req;
 943         default:
 944                 LBUG();
 945         }
 946
 947         /* Clear request swab mask; this is a new request */
 948         req->rq_req_swab_mask = 0;
 949
 950         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 951         if (rc != 0) {
 952                 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
 953                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
 954                        req->rq_xid);
 955                 goto err_req;
 956         }
 957
 958         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 959         if (rc) {
 960                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
 961                         LPU64"\n", svc->srv_req_portal,
 962                         libcfs_id2str(req->rq_peer), req->rq_xid);
 963                 goto err_req;
 964         }
 965
 966         rc = -EINVAL;
 967         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 968                 CERROR("wrong packet type received (type=%u) from %s\n",
 969                        lustre_msg_get_type(req->rq_reqmsg),
 970                        libcfs_id2str(req->rq_peer));
 971                 goto err_req;
 972         }
 973
 974         CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
 975
 976         req->rq_export = class_conn2export(
 977                 lustre_msg_get_handle(req->rq_reqmsg));
 978         if (req->rq_export) {
 979                 rc = ptlrpc_check_req(req);
 980                 if (rc == 0) {
 981                         rc = sptlrpc_target_export_check(req->rq_export, req);
 982                         if (rc)
 983                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
 984                                           "illegal security flavor,");
 985                 }
 986
 987                 class_export_put(req->rq_export);
 988                 req->rq_export = NULL;
 989                 if (rc)
 990                         goto err_req;
 991         }
 992
 993         /* req_in handling should/must be fast */
 994         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
 995                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus",
 996                           cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
 997
 998         /* Set rpc server deadline and add it to the timed list */
 999         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1000                     MSGHDR_AT_SUPPORT) ?
1001                    /* The max time the client expects us to take */
1002                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1003         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1004         if (unlikely(deadline == 0)) {
1005                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1006                 goto err_req;
1007         }
1008
1009         ptlrpc_at_add_timed(req);
1010
1011         /* Move it over to the request processing queue */
1012         spin_lock(&svc->srv_lock);
1013         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1014         cfs_waitq_signal(&svc->srv_waitq);
1015         spin_unlock(&svc->srv_lock);
1016         RETURN(1);
1017
1018 err_req:
1019         spin_lock(&svc->srv_lock);
1020         svc->srv_n_queued_reqs--;
1021         svc->srv_n_active_reqs++;
1022         spin_unlock(&svc->srv_lock);
1023         ptlrpc_server_free_request(req);
1024
1025         RETURN(1);
1026 }
1027
1028 static int
1029 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1030                              struct ptlrpc_thread *thread)
1031 {
1032         struct obd_export     *export = NULL;
1033         struct ptlrpc_request *request;
1034         struct timeval         work_start;
1035         struct timeval         work_end;
1036         long                   timediff;
1037         int                    rc;
1038         ENTRY;
1039
1040         LASSERT(svc);
1041
1042         spin_lock(&svc->srv_lock);
1043         if (unlikely(list_empty (&svc->srv_request_queue) ||
1044             (
1045 #ifndef __KERNEL__
1046              /* !@%$# liblustre only has 1 thread */
1047              svc->srv_n_difficult_replies != 0 &&
1048 #endif
1049              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1050                  /* Don't handle regular requests in the last thread, in order               * re
1051                   * to handle difficult replies (which might block other threads)
1052                   * as well as handle any incoming reqs, early replies, etc.
1053                   * That means we always need at least 2 service threads. */
1054                 spin_unlock(&svc->srv_lock);
1055                 RETURN(0);
1056         }
1057
1058         request = list_entry (svc->srv_request_queue.next,
1059                               struct ptlrpc_request, rq_list);
1060         list_del_init (&request->rq_list);
1061         svc->srv_n_queued_reqs--;
1062         svc->srv_n_active_reqs++;
1063
1064         spin_unlock(&svc->srv_lock);
1065
1066         do_gettimeofday(&work_start);
1067         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1068         if (likely(svc->srv_stats != NULL)) {
1069                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1070                                     timediff);
1071                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1072                                     svc->srv_n_queued_reqs);
1073                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1074                                     svc->srv_n_active_reqs);
1075                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1076                                     at_get(&svc->srv_at_estimate));
1077         }
1078
1079         rc = lu_context_init(&request->rq_session, LCT_SESSION);
1080         if (rc) {
1081                 CERROR("Failure to initialize session: %d\n", rc);
1082                 goto out_req;
1083         }
1084         request->rq_session.lc_thread = thread;
1085         lu_context_enter(&request->rq_session);
1086
1087         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
1088
1089         request->rq_svc_thread = thread;
1090         if (thread)
1091                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1092
1093         request->rq_export = class_conn2export(
1094                                      lustre_msg_get_handle(request->rq_reqmsg));
1095
1096         if (likely(request->rq_export)) {
1097                 if (unlikely(ptlrpc_check_req(request)))
1098                         goto put_conn;
1099                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1100                 export = class_export_rpc_get(request->rq_export);
1101         }
1102
1103         /* Discard requests queued for longer than the deadline.
1104            The deadline is increased if we send an early reply. */
1105         if (cfs_time_current_sec() > request->rq_deadline) {
1106                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1107                           ": deadline %ld%+lds ago\n",
1108                           libcfs_id2str(request->rq_peer),
1109                           request->rq_deadline -
1110                           request->rq_arrival_time.tv_sec,
1111                           cfs_time_current_sec() - request->rq_deadline);
1112                 goto put_rpc_export;
1113         }
1114
1115         request->rq_phase = RQ_PHASE_INTERPRET;
1116
1117         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1118                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1119                (request->rq_export ?
1120                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1121                (request->rq_export ?
1122                 atomic_read(&request->rq_export->exp_refcount) : -99),
1123                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1124                libcfs_id2str(request->rq_peer),
1125                lustre_msg_get_opc(request->rq_reqmsg));
1126
1127         OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1128
1129         rc = svc->srv_handler(request);
1130
1131         request->rq_phase = RQ_PHASE_COMPLETE;
1132
1133         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1134                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1135                (request->rq_export ?
1136                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1137                (request->rq_export ?
1138                 atomic_read(&request->rq_export->exp_refcount) : -99),
1139                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1140                libcfs_id2str(request->rq_peer),
1141                lustre_msg_get_opc(request->rq_reqmsg));
1142
1143 put_rpc_export:
1144         if (export != NULL)
1145                 class_export_rpc_put(export);
1146 put_conn:
1147         if (likely(request->rq_export != NULL))
1148                 class_export_put(request->rq_export);
1149
1150         lu_context_exit(&request->rq_session);
1151         lu_context_fini(&request->rq_session);
1152
1153         if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1154                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1155                           "than estimated (%ld%+lds); client may timeout.",
1156                           request->rq_xid, request->rq_deadline -
1157                           request->rq_arrival_time.tv_sec,
1158                           cfs_time_current_sec() - request->rq_deadline);
1159         }
1160
1161         do_gettimeofday(&work_end);
1162         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1163         CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1164                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1165                request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1166                libcfs_id2str(request->rq_peer), timediff,
1167                cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1168                request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1169                request->rq_transno, request->rq_status,
1170                request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1171                -999);
1172         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1173                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1174                 int opc = opcode_offset(op);
1175                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1176                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1177                         lprocfs_counter_add(svc->srv_stats,
1178                                             opc + EXTRA_MAX_OPCODES,
1179                                             timediff);
1180                 }
1181         }
1182         if (unlikely(request->rq_early_count)) {
1183                 DEBUG_REQ(D_ADAPTTO, request,
1184                           "sent %d early replies before finishing in %lds",
1185                           request->rq_early_count,
1186                           work_end.tv_sec - request->rq_arrival_time.tv_sec);
1187         }
1188
1189 out_req:
1190         ptlrpc_server_free_request(request);
1191
1192         RETURN(1);
1193 }
1194
1195 static int
1196 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1197 {
1198         struct ptlrpc_reply_state *rs;
1199         struct obd_export         *exp;
1200         struct obd_device         *obd;
1201         int                        nlocks;
1202         int                        been_handled;
1203         ENTRY;
1204
1205         spin_lock(&svc->srv_lock);
1206         if (list_empty (&svc->srv_reply_queue)) {
1207                 spin_unlock(&svc->srv_lock);
1208                 RETURN(0);
1209         }
1210
1211         rs = list_entry (svc->srv_reply_queue.next,
1212                          struct ptlrpc_reply_state, rs_list);
1213
1214         exp = rs->rs_export;
1215         obd = exp->exp_obd;
1216
1217         LASSERT (rs->rs_difficult);
1218         LASSERT (rs->rs_scheduled);
1219
1220         list_del_init (&rs->rs_list);
1221
1222         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1223         spin_unlock(&svc->srv_lock);
1224
1225         spin_lock (&obd->obd_uncommitted_replies_lock);
1226         /* Noop if removed already */
1227         list_del_init (&rs->rs_obd_list);
1228         spin_unlock (&obd->obd_uncommitted_replies_lock);
1229
1230         spin_lock (&exp->exp_lock);
1231         /* Noop if removed already */
1232         list_del_init (&rs->rs_exp_list);
1233         spin_unlock (&exp->exp_lock);
1234
1235         spin_lock(&svc->srv_lock);
1236
1237         been_handled = rs->rs_handled;
1238         rs->rs_handled = 1;
1239
1240         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1241         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1242
1243         if (nlocks == 0 && !been_handled) {
1244                 /* If we see this, we should already have seen the warning
1245                  * in mds_steal_ack_locks()  */
1246                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1247                       " o%d NID %s\n",
1248                       rs,
1249                       rs->rs_xid, rs->rs_transno,
1250                       lustre_msg_get_opc(rs->rs_msg),
1251                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1252         }
1253
1254         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1255                 spin_unlock(&svc->srv_lock);
1256
1257                 if (!been_handled && rs->rs_on_net) {
1258                         LNetMDUnlink(rs->rs_md_h);
1259                         /* Ignore return code; we're racing with
1260                          * completion... */
1261                 }
1262
1263                 while (nlocks-- > 0)
1264                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1265                                          rs->rs_modes[nlocks]);
1266
1267                 spin_lock(&svc->srv_lock);
1268         }
1269
1270         rs->rs_scheduled = 0;
1271
1272         if (!rs->rs_on_net) {
1273                 /* Off the net */
1274                 svc->srv_n_difficult_replies--;
1275                 spin_unlock(&svc->srv_lock);
1276
1277                 class_export_put (exp);
1278                 rs->rs_export = NULL;
1279                 ptlrpc_rs_decref (rs);
1280                 atomic_dec (&svc->srv_outstanding_replies);
1281                 RETURN(1);
1282         }
1283
1284         /* still on the net; callback will schedule */
1285         spin_unlock(&svc->srv_lock);
1286         RETURN(1);
1287 }
1288
1289 #ifndef __KERNEL__
1290 /* FIXME make use of timeout later */
1291 int
1292 liblustre_check_services (void *arg)
1293 {
1294         int  did_something = 0;
1295         int  rc;
1296         struct list_head *tmp, *nxt;
1297         ENTRY;
1298
1299         /* I'm relying on being single threaded, not to have to lock
1300          * ptlrpc_all_services etc */
1301         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1302                 struct ptlrpc_service *svc =
1303                         list_entry (tmp, struct ptlrpc_service, srv_list);
1304
1305                 if (svc->srv_threads_running != 0)     /* I've recursed */
1306                         continue;
1307
1308                 /* service threads can block for bulk, so this limits us
1309                  * (arbitrarily) to recursing 1 stack frame per service.
1310                  * Note that the problem with recursion is that we have to
1311                  * unwind completely before our caller can resume. */
1312
1313                 svc->srv_threads_running++;
1314
1315                 do {
1316                         rc = ptlrpc_server_handle_req_in(svc);
1317                         rc |= ptlrpc_server_handle_reply(svc);
1318                         rc |= ptlrpc_at_check_timed(svc);
1319                         rc |= ptlrpc_server_handle_request(svc, NULL);
1320                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1321                         did_something |= rc;
1322                 } while (rc);
1323
1324                 svc->srv_threads_running--;
1325         }
1326
1327         RETURN(did_something);
1328 }
1329 #define ptlrpc_stop_all_threads(s) do {} while (0)
1330
1331 #else /* __KERNEL__ */
1332
1333 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1334 void ptlrpc_daemonize(char *name)
1335 {
1336         struct fs_struct *fs = current->fs;
1337
1338         atomic_inc(&fs->count);
1339         cfs_daemonize(name);
1340         exit_fs(cfs_current());
1341         current->fs = fs;
1342         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1343 }
1344
1345 static void
1346 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1347 {
1348         int avail = svc->srv_nrqbd_receiving;
1349         int low_water = test_req_buffer_pressure ? 0 :
1350                         svc->srv_nbuf_per_group/2;
1351
1352         /* NB I'm not locking; just looking. */
1353
1354         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1355          * allowed the request history to grow out of control.  We could put a
1356          * sanity check on that here and cull some history if we need the
1357          * space. */
1358
1359         if (avail <= low_water)
1360                 ptlrpc_grow_req_bufs(svc);
1361
1362         if (svc->srv_stats)
1363                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1364                                     avail);
1365 }
1366
1367 static int
1368 ptlrpc_retry_rqbds(void *arg)
1369 {
1370         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1371
1372         svc->srv_rqbd_timeout = 0;
1373         return (-ETIMEDOUT);
1374 }
1375
1376 static int ptlrpc_main(void *arg)
1377 {
1378         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1379         struct ptlrpc_service  *svc = data->svc;
1380         struct ptlrpc_thread   *thread = data->thread;
1381         struct obd_device      *dev = data->dev;
1382         struct ptlrpc_reply_state *rs;
1383         struct lc_watchdog     *watchdog;
1384 #ifdef WITH_GROUP_INFO
1385         struct group_info *ginfo = NULL;
1386 #endif
1387         struct lu_env env;
1388         int counter = 0, rc = 0;
1389         ENTRY;
1390
1391         ptlrpc_daemonize(data->name);
1392
1393 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1394         /* we need to do this before any per-thread allocation is done so that
1395          * we get the per-thread allocations on local node.  bug 7342 */
1396         if (svc->srv_cpu_affinity) {
1397                 int cpu, num_cpu;
1398
1399                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1400                         if (!cpu_online(cpu))
1401                                 continue;
1402                         if (num_cpu == thread->t_id % num_online_cpus())
1403                                 break;
1404                         num_cpu++;
1405                 }
1406                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1407         }
1408 #endif
1409
1410 #ifdef WITH_GROUP_INFO
1411         ginfo = groups_alloc(0);
1412         if (!ginfo) {
1413                 rc = -ENOMEM;
1414                 goto out;
1415         }
1416
1417         set_current_groups(ginfo);
1418         put_group_info(ginfo);
1419 #endif
1420
1421         if (svc->srv_init != NULL) {
1422                 rc = svc->srv_init(thread);
1423                 if (rc)
1424                         goto out;
1425         }
1426
1427         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1428         if (rc)
1429                 goto out_srv_fini;
1430
1431         thread->t_env = &env;
1432         env.le_ctx.lc_thread = thread;
1433
1434         /* Alloc reply state structure for this one */
1435         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1436         if (!rs) {
1437                 rc = -ENOMEM;
1438                 goto out_srv_fini;
1439         }
1440
1441         /* Record that the thread is running */
1442         thread->t_flags = SVC_RUNNING;
1443         /*
1444          * wake up our creator. Note: @data is invalid after this point,
1445          * because it's allocated on ptlrpc_start_thread() stack.
1446          */
1447         cfs_waitq_signal(&thread->t_ctl_waitq);
1448
1449         watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1450                                    at_get(&svc->srv_at_estimate)) *
1451                                    svc->srv_watchdog_factor, NULL, NULL);
1452
1453         spin_lock(&svc->srv_lock);
1454         svc->srv_threads_running++;
1455         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1456         spin_unlock(&svc->srv_lock);
1457         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1458
1459         CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1460               svc->srv_threads_running);
1461
1462         /* XXX maintain a list of all managed devices: insert here */
1463
1464         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1465                svc->srv_n_difficult_replies != 0) {
1466                 /* Don't exit while there are replies to be handled */
1467                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1468                                                      ptlrpc_retry_rqbds, svc);
1469
1470                 lc_watchdog_disable(watchdog);
1471
1472                 cond_resched();
1473
1474                 l_wait_event_exclusive (svc->srv_waitq,
1475                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1476                                svc->srv_n_difficult_replies == 0) ||
1477                               (!list_empty(&svc->srv_idle_rqbds) &&
1478                                svc->srv_rqbd_timeout == 0) ||
1479                               !list_empty(&svc->srv_req_in_queue) ||
1480                               !list_empty(&svc->srv_reply_queue) ||
1481                               (!list_empty(&svc->srv_request_queue) &&
1482                                (svc->srv_n_active_reqs <
1483                                 (svc->srv_threads_running - 1))) ||
1484                               svc->srv_at_check,
1485                               &lwi);
1486
1487                 lc_watchdog_touch_ms(watchdog, max_t(int, obd_timeout,
1488                                      AT_OFF ? 0 :
1489                                      at_get(&svc->srv_at_estimate)) *
1490                                      svc->srv_watchdog_factor);
1491
1492                 ptlrpc_check_rqbd_pool(svc);
1493
1494                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1495                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1496                         /* Ignore return code - we tried... */
1497                         ptlrpc_start_thread(dev, svc);
1498                 }
1499
1500                 if (!list_empty(&svc->srv_reply_queue))
1501                         ptlrpc_server_handle_reply(svc);
1502
1503                 if (!list_empty(&svc->srv_req_in_queue)) {
1504                         /* Process all incoming reqs before handling any */
1505                         ptlrpc_server_handle_req_in(svc);
1506                         /* but limit ourselves in case of flood */
1507                         if (counter++ < 1000)
1508                                 continue;
1509                         counter = 0;
1510                 }
1511
1512                 if (svc->srv_at_check)
1513                         ptlrpc_at_check_timed(svc);
1514
1515                 /* don't handle requests in the last thread */
1516                 if (!list_empty (&svc->srv_request_queue) &&
1517                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1518                         lu_context_enter(&env.le_ctx);
1519                         ptlrpc_server_handle_request(svc, thread);
1520                         lu_context_exit(&env.le_ctx);
1521                 }
1522
1523                 if (!list_empty(&svc->srv_idle_rqbds) &&
1524                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1525                         /* I just failed to repost request buffers.  Wait
1526                          * for a timeout (unless something else happens)
1527                          * before I try again */
1528                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1529                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1530                                svc->srv_nrqbd_receiving);
1531                 }
1532         }
1533
1534         lc_watchdog_delete(watchdog);
1535
1536 out_srv_fini:
1537         /*
1538          * deconstruct service specific state created by ptlrpc_start_thread()
1539          */
1540         if (svc->srv_done != NULL)
1541                 svc->srv_done(thread);
1542
1543         lu_context_fini(&env.le_ctx);
1544 out:
1545         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1546
1547         spin_lock(&svc->srv_lock);
1548         svc->srv_threads_running--; /* must know immediately */
1549         thread->t_id = rc;
1550         thread->t_flags = SVC_STOPPED;
1551
1552         cfs_waitq_signal(&thread->t_ctl_waitq);
1553         spin_unlock(&svc->srv_lock);
1554
1555         return rc;
1556 }
1557
1558 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1559                                struct ptlrpc_thread *thread)
1560 {
1561         struct l_wait_info lwi = { 0 };
1562
1563         spin_lock(&svc->srv_lock);
1564         thread->t_flags = SVC_STOPPING;
1565         spin_unlock(&svc->srv_lock);
1566
1567         cfs_waitq_broadcast(&svc->srv_waitq);
1568         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1569                      &lwi);
1570
1571         spin_lock(&svc->srv_lock);
1572         list_del(&thread->t_link);
1573         spin_unlock(&svc->srv_lock);
1574
1575         OBD_FREE_PTR(thread);
1576 }
1577
1578 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1579 {
1580         struct ptlrpc_thread *thread;
1581
1582         spin_lock(&svc->srv_lock);
1583         while (!list_empty(&svc->srv_threads)) {
1584                 thread = list_entry(svc->srv_threads.next,
1585                                     struct ptlrpc_thread, t_link);
1586
1587                 spin_unlock(&svc->srv_lock);
1588                 ptlrpc_stop_thread(svc, thread);
1589                 spin_lock(&svc->srv_lock);
1590         }
1591
1592         spin_unlock(&svc->srv_lock);
1593 }
1594
1595 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1596 {
1597         int i, rc = 0;
1598         ENTRY;
1599
1600         /* We require 2 threads min - see note in
1601            ptlrpc_server_handle_request */
1602         LASSERT(svc->srv_threads_min >= 2);
1603         for (i = 0; i < svc->srv_threads_min; i++) {
1604                 rc = ptlrpc_start_thread(dev, svc);
1605                 /* We have enough threads, don't start more.  b=15759 */
1606                 if (rc == -EMFILE)
1607                         break;
1608                 if (rc) {
1609                         CERROR("cannot start %s thread #%d: rc %d\n",
1610                                svc->srv_thread_name, i, rc);
1611                         ptlrpc_stop_all_threads(svc);
1612                 }
1613         }
1614         RETURN(rc);
1615 }
1616
1617 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1618 {
1619         struct l_wait_info lwi = { 0 };
1620         struct ptlrpc_svc_data d;
1621         struct ptlrpc_thread *thread;
1622         char name[32];
1623         int id, rc;
1624         ENTRY;
1625
1626         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1627                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1628                svc->srv_threads_max, svc->srv_threads_running);
1629         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1630             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1631              svc->srv_threads_started == svc->srv_threads_min - 1))
1632                 RETURN(-EMFILE);
1633
1634         OBD_ALLOC_PTR(thread);
1635         if (thread == NULL)
1636                 RETURN(-ENOMEM);
1637         cfs_waitq_init(&thread->t_ctl_waitq);
1638
1639         spin_lock(&svc->srv_lock);
1640         if (svc->srv_threads_started >= svc->srv_threads_max) {
1641                 spin_unlock(&svc->srv_lock);
1642                 OBD_FREE_PTR(thread);
1643                 RETURN(-EMFILE);
1644         }
1645         list_add(&thread->t_link, &svc->srv_threads);
1646         id = svc->srv_threads_started++;
1647         spin_unlock(&svc->srv_lock);
1648
1649         thread->t_id = id;
1650         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1651         d.dev = dev;
1652         d.svc = svc;
1653         d.name = name;
1654         d.thread = thread;
1655
1656         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1657
1658           /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1659          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1660          */
1661         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1662         if (rc < 0) {
1663                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1664
1665                 spin_lock(&svc->srv_lock);
1666                 list_del(&thread->t_link);
1667                 --svc->srv_threads_started;
1668                 spin_unlock(&svc->srv_lock);
1669
1670                 OBD_FREE(thread, sizeof(*thread));
1671                 RETURN(rc);
1672         }
1673         l_wait_event(thread->t_ctl_waitq,
1674                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1675
1676         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1677         RETURN(rc);
1678 }
1679 #endif
1680
1681 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1682 {
1683         int                   rc;
1684         struct l_wait_info    lwi;
1685         struct list_head     *tmp;
1686         struct ptlrpc_reply_state *rs, *t;
1687
1688         cfs_timer_disarm(&service->srv_at_timer);
1689
1690         ptlrpc_stop_all_threads(service);
1691         LASSERT(list_empty(&service->srv_threads));
1692
1693         spin_lock (&ptlrpc_all_services_lock);
1694         list_del_init (&service->srv_list);
1695         spin_unlock (&ptlrpc_all_services_lock);
1696
1697         ptlrpc_lprocfs_unregister_service(service);
1698
1699         /* All history will be culled when the next request buffer is
1700          * freed */
1701         service->srv_max_history_rqbds = 0;
1702
1703         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1704
1705         rc = LNetClearLazyPortal(service->srv_req_portal);
1706         LASSERT (rc == 0);
1707
1708         /* Unlink all the request buffers.  This forces a 'final' event with
1709          * its 'unlink' flag set for each posted rqbd */
1710         list_for_each(tmp, &service->srv_active_rqbds) {
1711                 struct ptlrpc_request_buffer_desc *rqbd =
1712                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1713                                    rqbd_list);
1714
1715                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1716                 LASSERT (rc == 0 || rc == -ENOENT);
1717         }
1718
1719         /* Wait for the network to release any buffers it's currently
1720          * filling */
1721         for (;;) {
1722                 spin_lock(&service->srv_lock);
1723                 rc = service->srv_nrqbd_receiving;
1724                 spin_unlock(&service->srv_lock);
1725
1726                 if (rc == 0)
1727                         break;
1728
1729                 /* Network access will complete in finite time but the HUGE
1730                  * timeout lets us CWARN for visibility of sluggish NALs */
1731                 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1732                 rc = l_wait_event(service->srv_waitq,
1733                                   service->srv_nrqbd_receiving == 0,
1734                                   &lwi);
1735                 if (rc == -ETIMEDOUT)
1736                         CWARN("Service %s waiting for request buffers\n",
1737                               service->srv_name);
1738         }
1739
1740         /* schedule all outstanding replies to terminate them */
1741         spin_lock(&service->srv_lock);
1742         while (!list_empty(&service->srv_active_replies)) {
1743                 struct ptlrpc_reply_state *rs =
1744                         list_entry(service->srv_active_replies.next,
1745                                    struct ptlrpc_reply_state, rs_list);
1746                 ptlrpc_schedule_difficult_reply(rs);
1747         }
1748         spin_unlock(&service->srv_lock);
1749
1750         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1751          * and no service threads, so I'm the only thread noodling the
1752          * request queue now */
1753         while (!list_empty(&service->srv_req_in_queue)) {
1754                 struct ptlrpc_request *req =
1755                         list_entry(service->srv_req_in_queue.next,
1756                                    struct ptlrpc_request,
1757                                    rq_list);
1758
1759                 list_del(&req->rq_list);
1760                 service->srv_n_queued_reqs--;
1761                 service->srv_n_active_reqs++;
1762                 ptlrpc_server_free_request(req);
1763         }
1764         while (!list_empty(&service->srv_request_queue)) {
1765                 struct ptlrpc_request *req =
1766                         list_entry(service->srv_request_queue.next,
1767                                    struct ptlrpc_request,
1768                                    rq_list);
1769
1770                 list_del(&req->rq_list);
1771                 service->srv_n_queued_reqs--;
1772                 service->srv_n_active_reqs++;
1773
1774                 ptlrpc_server_free_request(req);
1775         }
1776         LASSERT(service->srv_n_queued_reqs == 0);
1777         LASSERT(service->srv_n_active_reqs == 0);
1778         LASSERT(service->srv_n_history_rqbds == 0);
1779         LASSERT(list_empty(&service->srv_active_rqbds));
1780
1781         /* Now free all the request buffers since nothing references them
1782          * any more... */
1783         while (!list_empty(&service->srv_idle_rqbds)) {
1784                 struct ptlrpc_request_buffer_desc *rqbd =
1785                         list_entry(service->srv_idle_rqbds.next,
1786                                    struct ptlrpc_request_buffer_desc,
1787                                    rqbd_list);
1788
1789                 ptlrpc_free_rqbd(rqbd);
1790         }
1791
1792         /* wait for all outstanding replies to complete (they were
1793          * scheduled having been flagged to abort above) */
1794         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1795                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1796
1797                 rc = l_wait_event(service->srv_waitq,
1798                                   !list_empty(&service->srv_reply_queue), &lwi);
1799                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1800
1801                 if (rc == 0) {
1802                         ptlrpc_server_handle_reply(service);
1803                         continue;
1804                 }
1805                 CWARN("Unexpectedly long timeout %p\n", service);
1806         }
1807
1808         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1809                 list_del(&rs->rs_list);
1810                 OBD_FREE(rs, service->srv_max_reply_size);
1811         }
1812
1813         /* In case somebody rearmed this in the meantime */
1814         cfs_timer_disarm(&service->srv_at_timer);
1815
1816         OBD_FREE_PTR(service);
1817         return 0;
1818 }
1819
1820 /* Returns 0 if the service is healthy.
1821  *
1822  * Right now, it just checks to make sure that requests aren't languishing
1823  * in the queue.  We'll use this health check to govern whether a node needs
1824  * to be shot, so it's intentionally non-aggressive. */
1825 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1826 {
1827         struct ptlrpc_request *request;
1828         struct timeval         right_now;
1829         long                   timediff;
1830
1831         if (svc == NULL)
1832                 return 0;
1833
1834         do_gettimeofday(&right_now);
1835
1836         spin_lock(&svc->srv_lock);
1837         if (list_empty(&svc->srv_request_queue)) {
1838                 spin_unlock(&svc->srv_lock);
1839                 return 0;
1840         }
1841
1842         /* How long has the next entry been waiting? */
1843         request = list_entry(svc->srv_request_queue.next,
1844                              struct ptlrpc_request, rq_list);
1845         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1846         spin_unlock(&svc->srv_lock);
1847
1848         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
1849                                         at_max)) {
1850                 CERROR("%s: unhealthy - request has been waiting %lds\n",
1851                        svc->srv_name, timediff / ONE_MILLION);
1852                 return (-1);
1853         }
1854
1855         return 0;
1856 }