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 #ifndef __KERNEL__
 607 int lu_context_init(struct lu_context *ctx, __u32 tags)
 608 {
 609         return 0;
 610 }
 611
 612 void lu_context_fini(struct lu_context *ctx)
 613 {
 614 }
 615
 616 void lu_context_enter(struct lu_context *ctx)
 617 {
 618 }
 619
 620 void lu_context_exit(struct lu_context *ctx)
 621 {
 622 }
 623
 624 #endif
 625
 626 static int ptlrpc_check_req(struct ptlrpc_request *req)
 627 {
 628         if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 629                      req->rq_export->exp_conn_cnt)) {
 630                 DEBUG_REQ(D_ERROR, req,
 631                           "DROPPING req from old connection %d < %d",
 632                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 633                           req->rq_export->exp_conn_cnt);
 634                 return -EEXIST;
 635         }
 636         if (unlikely(req->rq_export->exp_obd &&
 637                      req->rq_export->exp_obd->obd_fail)) {
 638              /* Failing over, don't handle any more reqs, send
 639                 error response instead. */
 640                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 641                        req, req->rq_export->exp_obd->obd_name);
 642                 req->rq_status = -ENODEV;
 643                 ptlrpc_error(req);
 644                 return -ENODEV;
 645         }
 646
 647         return 0;
 648 }
 649
 650 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
 651 {
 652         struct ptlrpc_request *rq;
 653         __s32 next;
 654
 655         spin_lock(&svc->srv_at_lock);
 656         if (list_empty(&svc->srv_at_list)) {
 657                 cfs_timer_disarm(&svc->srv_at_timer);
 658                 spin_unlock(&svc->srv_at_lock);
 659                 return;
 660         }
 661
 662         /* Set timer for closest deadline */
 663         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 664                         rq_timed_list);
 665         next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
 666                        at_early_margin);
 667         if (next <= 0)
 668                 ptlrpc_at_timer((unsigned long)svc);
 669         else
 670                 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
 671         spin_unlock(&svc->srv_at_lock);
 672         CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
 673 }
 674
 675 /* Add rpc to early reply check list */
 676 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 677 {
 678         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 679         struct ptlrpc_request *rq;
 680         int found = 0;
 681
 682         if (AT_OFF)
 683                 return(0);
 684
 685         if (req->rq_no_reply)
 686                 return 0;
 687
 688         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 689                 return(-ENOSYS);
 690
 691         spin_lock(&svc->srv_at_lock);
 692
 693         if (unlikely(req->rq_sent_final)) {
 694                 spin_unlock(&svc->srv_at_lock);
 695                 return 0;
 696         }
 697
 698         LASSERT(list_empty(&req->rq_timed_list));
 699         /* Add to sorted list.  Presumably latest rpcs will have the latest
 700            deadlines, so search backward. */
 701         list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
 702                 if (req->rq_deadline > rq->rq_deadline) {
 703                         list_add(&req->rq_timed_list, &rq->rq_timed_list);
 704                         found++;
 705                         break;
 706                 }
 707         }
 708         if (!found)
 709                 /* Add to front if shortest deadline or list empty */
 710                 list_add(&req->rq_timed_list, &svc->srv_at_list);
 711
 712         /* Check if we're the head of the list */
 713         found = (svc->srv_at_list.next == &req->rq_timed_list);
 714
 715         spin_unlock(&svc->srv_at_lock);
 716
 717         if (found)
 718                 ptlrpc_at_set_timer(svc);
 719
 720         return 0;
 721 }
 722
 723 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
 724                                       int extra_time)
 725 {
 726         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 727         struct ptlrpc_request *reqcopy;
 728         struct lustre_msg *reqmsg;
 729         cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
 730         time_t newdl;
 731         int rc;
 732         ENTRY;
 733
 734         /* deadline is when the client expects us to reply, margin is the
 735            difference between clients' and servers' expectations */
 736         DEBUG_REQ(D_ADAPTTO, req,
 737                   "%ssending early reply (deadline %+lds, margin %+lds) for "
 738                   "%d+%d", AT_OFF ? "AT off - not " : "",
 739                   olddl, olddl - at_get(&svc->srv_at_estimate),
 740                   at_get(&svc->srv_at_estimate), extra_time);
 741
 742         if (AT_OFF)
 743                 RETURN(0);
 744
 745         if (olddl < 0) {
 746                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
 747                           "not sending early reply. Consider increasing "
 748                           "at_early_margin (%d)?", olddl, at_early_margin);
 749
 750                 /* Return an error so we're not re-added to the timed list. */
 751                 RETURN(-ETIMEDOUT);
 752         }
 753
 754         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
 755                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
 756                           "but no AT support");
 757                 RETURN(-ENOSYS);
 758         }
 759
 760         if (extra_time) {
 761                 /* Fake our processing time into the future to ask the
 762                    clients for some extra amount of time */
 763                 extra_time += cfs_time_current_sec() -
 764                         req->rq_arrival_time.tv_sec;
 765                 at_add(&svc->srv_at_estimate, extra_time);
 766         }
 767
 768         newdl = req->rq_arrival_time.tv_sec + at_get(&svc->srv_at_estimate);
 769         if (req->rq_deadline >= newdl) {
 770                 /* We're not adding any time, no need to send an early reply
 771                    (e.g. maybe at adaptive_max) */
 772                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
 773                           CFS_DURATION_T"/"CFS_DURATION_T"), "
 774                           "not sending early reply\n", olddl,
 775                           cfs_time_sub(newdl, cfs_time_current_sec()));
 776                 RETURN(-ETIMEDOUT);
 777         }
 778
 779         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 780         if (reqcopy == NULL)
 781                 RETURN(-ENOMEM);
 782         OBD_ALLOC(reqmsg, req->rq_reqlen);
 783         if (!reqmsg) {
 784                 OBD_FREE(reqcopy, sizeof *reqcopy);
 785                 RETURN(-ENOMEM);
 786         }
 787
 788         *reqcopy = *req;
 789         reqcopy->rq_reply_state = NULL;
 790         reqcopy->rq_rep_swab_mask = 0;
 791         reqcopy->rq_pack_bulk = 0;
 792         reqcopy->rq_pack_udesc = 0;
 793         reqcopy->rq_packed_final = 0;
 794         sptlrpc_svc_ctx_addref(reqcopy);
 795         /* We only need the reqmsg for the magic */
 796         reqcopy->rq_reqmsg = reqmsg;
 797         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 798
 799         if (req->rq_sent_final) {
 800                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
 801                           "abort sending early reply\n");
 802                 GOTO(out, rc = 0);
 803         }
 804
 805         /* Connection ref */
 806         reqcopy->rq_export = class_conn2export(
 807                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 808         if (reqcopy->rq_export == NULL)
 809                 GOTO(out, rc = -ENODEV);
 810
 811         /* RPC ref */
 812         class_export_rpc_get(reqcopy->rq_export);
 813         if (reqcopy->rq_export->exp_obd &&
 814             reqcopy->rq_export->exp_obd->obd_fail)
 815                 GOTO(out_put, rc = -ENODEV);
 816
 817         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 818         if (rc)
 819                 GOTO(out_put, rc);
 820
 821         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 822
 823         if (!rc) {
 824                 /* Adjust our own deadline to what we told the client */
 825                 req->rq_deadline = newdl;
 826                 req->rq_early_count++; /* number sent, server side */
 827         } else {
 828                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 829         }
 830
 831         /* Free the (early) reply state from lustre_pack_reply.
 832            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 833         ptlrpc_req_drop_rs(reqcopy);
 834
 835 out_put:
 836         class_export_rpc_put(reqcopy->rq_export);
 837         class_export_put(reqcopy->rq_export);
 838 out:
 839         sptlrpc_svc_ctx_decref(reqcopy);
 840         OBD_FREE(reqmsg, req->rq_reqlen);
 841         OBD_FREE(reqcopy, sizeof *reqcopy);
 842         RETURN(rc);
 843 }
 844
 845 /* Send early replies to everybody expiring within at_early_margin
 846    asking for at_extra time */
 847 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 848 {
 849         struct ptlrpc_request *rq, *n;
 850         struct list_head work_list;
 851         time_t now = cfs_time_current_sec();
 852         cfs_duration_t delay;
 853         int first, counter = 0;
 854         ENTRY;
 855
 856         spin_lock(&svc->srv_at_lock);
 857         if (svc->srv_at_check == 0) {
 858                 spin_unlock(&svc->srv_at_lock);
 859                 RETURN(0);
 860         }
 861         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 862         svc->srv_at_check = 0;
 863
 864         if (list_empty(&svc->srv_at_list)) {
 865                 spin_unlock(&svc->srv_at_lock);
 866                 RETURN(0);
 867         }
 868
 869         /* The timer went off, but maybe the nearest rpc already completed. */
 870         rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
 871                         rq_timed_list);
 872         first = (int)(rq->rq_deadline - now);
 873         if (first > at_early_margin) {
 874                 /* We've still got plenty of time.  Reset the timer. */
 875                 spin_unlock(&svc->srv_at_lock);
 876                 ptlrpc_at_set_timer(svc);
 877                 RETURN(0);
 878         }
 879
 880         /* We're close to a timeout, and we don't know how much longer the
 881            server will take. Send early replies to everyone expiring soon. */
 882         CFS_INIT_LIST_HEAD(&work_list);
 883         list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
 884                 if (rq->rq_deadline <= now + at_early_margin) {
 885                         list_move_tail(&rq->rq_timed_list, &work_list);
 886                         counter++;
 887                 } else {
 888                         break;
 889                 }
 890         }
 891
 892         spin_unlock(&svc->srv_at_lock);
 893
 894         /* we have a new earliest deadline, restart the timer */
 895         ptlrpc_at_set_timer(svc);
 896
 897         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 898                "replies\n", first, at_extra, counter);
 899         if (first < 0) {
 900                 /* We're already past request deadlines before we even get a
 901                    chance to send early replies */
 902                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 903                               "request traffic (cpu-bound).\n", svc->srv_name);
 904                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 905                       "delay="CFS_DURATION_T"(jiff)\n",
 906                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 907                       at_get(&svc->srv_at_estimate), delay);
 908         }
 909
 910         /* ptlrpc_server_finish_request may delete an entry out of
 911          * the work list */
 912         spin_lock(&svc->srv_at_lock);
 913         while (!list_empty(&work_list)) {
 914                 rq = list_entry(work_list.next, struct ptlrpc_request,
 915                                 rq_timed_list);
 916                 list_del_init(&rq->rq_timed_list);
 917                 /* if the entry is still in the worklist, it hasn't been
 918                    deleted, and is safe to take a ref to keep the req around */
 919                 atomic_inc(&rq->rq_refcount);
 920                 spin_unlock(&svc->srv_at_lock);
 921
 922                 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
 923                         ptlrpc_at_add_timed(rq);
 924
 925                 ptlrpc_server_drop_request(rq);
 926                 spin_lock(&svc->srv_at_lock);
 927         }
 928         spin_unlock(&svc->srv_at_lock);
 929
 930         RETURN(0);
 931 }
 932
 933 /* Handle freshly incoming reqs, add to timed early reply list,
 934    pass on to regular request queue */
 935 static int
 936 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
 937 {
 938         struct ptlrpc_request *req;
 939         __u32                  deadline;
 940         int                    rc;
 941         ENTRY;
 942
 943         LASSERT(svc);
 944
 945         spin_lock(&svc->srv_lock);
 946         if (list_empty(&svc->srv_req_in_queue)) {
 947                 spin_unlock(&svc->srv_lock);
 948                 RETURN(0);
 949         }
 950
 951         req = list_entry(svc->srv_req_in_queue.next,
 952                          struct ptlrpc_request, rq_list);
 953         list_del_init (&req->rq_list);
 954         /* Consider this still a "queued" request as far as stats are
 955            concerned */
 956         spin_unlock(&svc->srv_lock);
 957
 958         /* go through security check/transform */
 959         rc = sptlrpc_svc_unwrap_request(req);
 960         switch (rc) {
 961         case SECSVC_OK:
 962                 break;
 963         case SECSVC_COMPLETE:
 964                 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
 965                 goto err_req;
 966         case SECSVC_DROP:
 967                 goto err_req;
 968         default:
 969                 LBUG();
 970         }
 971
 972         /* Clear request swab mask; this is a new request */
 973         req->rq_req_swab_mask = 0;
 974
 975         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
 976         if (rc != 0) {
 977                 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
 978                        svc->srv_req_portal, libcfs_id2str(req->rq_peer),
 979                        req->rq_xid);
 980                 goto err_req;
 981         }
 982
 983         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
 984         if (rc) {
 985                 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
 986                         LPU64"\n", svc->srv_req_portal,
 987                         libcfs_id2str(req->rq_peer), req->rq_xid);
 988                 goto err_req;
 989         }
 990
 991         rc = -EINVAL;
 992         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
 993                 CERROR("wrong packet type received (type=%u) from %s\n",
 994                        lustre_msg_get_type(req->rq_reqmsg),
 995                        libcfs_id2str(req->rq_peer));
 996                 goto err_req;
 997         }
 998
 999         CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
1000
1001         req->rq_export = class_conn2export(
1002                 lustre_msg_get_handle(req->rq_reqmsg));
1003         if (req->rq_export) {
1004                 rc = ptlrpc_check_req(req);
1005                 if (rc == 0) {
1006                         rc = sptlrpc_target_export_check(req->rq_export, req);
1007                         if (rc)
1008                                 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1009                                           "illegal security flavor,");
1010                 }
1011
1012                 class_export_put(req->rq_export);
1013                 req->rq_export = NULL;
1014                 if (rc)
1015                         goto err_req;
1016         }
1017
1018         /* req_in handling should/must be fast */
1019         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1020                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1021                           cfs_time_sub(cfs_time_current_sec(),
1022                                        req->rq_arrival_time.tv_sec));
1023
1024         /* Set rpc server deadline and add it to the timed list */
1025         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1026                     MSGHDR_AT_SUPPORT) ?
1027                    /* The max time the client expects us to take */
1028                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1029         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1030         if (unlikely(deadline == 0)) {
1031                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1032                 goto err_req;
1033         }
1034
1035         ptlrpc_at_add_timed(req);
1036
1037         /* Move it over to the request processing queue */
1038         spin_lock(&svc->srv_lock);
1039         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1040         cfs_waitq_signal(&svc->srv_waitq);
1041         spin_unlock(&svc->srv_lock);
1042         RETURN(1);
1043
1044 err_req:
1045         spin_lock(&svc->srv_lock);
1046         svc->srv_n_queued_reqs--;
1047         svc->srv_n_active_reqs++;
1048         spin_unlock(&svc->srv_lock);
1049         ptlrpc_server_finish_request(req);
1050
1051         RETURN(1);
1052 }
1053
1054 static int
1055 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1056                              struct ptlrpc_thread *thread)
1057 {
1058         struct obd_export     *export = NULL;
1059         struct ptlrpc_request *request;
1060         struct timeval         work_start;
1061         struct timeval         work_end;
1062         long                   timediff;
1063         int                    rc;
1064         ENTRY;
1065
1066         LASSERT(svc);
1067
1068         spin_lock(&svc->srv_lock);
1069         if (unlikely(list_empty (&svc->srv_request_queue) ||
1070             (
1071 #ifndef __KERNEL__
1072              /* !@%$# liblustre only has 1 thread */
1073              svc->srv_n_difficult_replies != 0 &&
1074 #endif
1075              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1076                  /* Don't handle regular requests in the last thread, in order               * re
1077                   * to handle difficult replies (which might block other threads)
1078                   * as well as handle any incoming reqs, early replies, etc.
1079                   * That means we always need at least 2 service threads. */
1080                 spin_unlock(&svc->srv_lock);
1081                 RETURN(0);
1082         }
1083
1084         request = list_entry (svc->srv_request_queue.next,
1085                               struct ptlrpc_request, rq_list);
1086         list_del_init (&request->rq_list);
1087         svc->srv_n_queued_reqs--;
1088         svc->srv_n_active_reqs++;
1089
1090         spin_unlock(&svc->srv_lock);
1091
1092         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1093                 libcfs_debug_dumplog();
1094
1095         do_gettimeofday(&work_start);
1096         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1097         if (likely(svc->srv_stats != NULL)) {
1098                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1099                                     timediff);
1100                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1101                                     svc->srv_n_queued_reqs);
1102                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1103                                     svc->srv_n_active_reqs);
1104                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1105                                     at_get(&svc->srv_at_estimate));
1106         }
1107
1108         rc = lu_context_init(&request->rq_session, LCT_SESSION);
1109         if (rc) {
1110                 CERROR("Failure to initialize session: %d\n", rc);
1111                 goto out_req;
1112         }
1113         request->rq_session.lc_thread = thread;
1114         lu_context_enter(&request->rq_session);
1115
1116         CDEBUG(D_NET, "got req "LPD64"\n", request->rq_xid);
1117
1118         request->rq_svc_thread = thread;
1119         if (thread)
1120                 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1121
1122         request->rq_export = class_conn2export(
1123                                      lustre_msg_get_handle(request->rq_reqmsg));
1124
1125         if (likely(request->rq_export)) {
1126                 if (unlikely(ptlrpc_check_req(request)))
1127                         goto put_conn;
1128                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1129                 export = class_export_rpc_get(request->rq_export);
1130         }
1131
1132         /* Discard requests queued for longer than the deadline.
1133            The deadline is increased if we send an early reply. */
1134         if (cfs_time_current_sec() > request->rq_deadline) {
1135                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1136                           ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1137                           libcfs_id2str(request->rq_peer),
1138                           cfs_time_sub(request->rq_deadline,
1139                           request->rq_arrival_time.tv_sec),
1140                           cfs_time_sub(cfs_time_current_sec(),
1141                           request->rq_deadline));
1142                 goto put_rpc_export;
1143         }
1144
1145         request->rq_phase = RQ_PHASE_INTERPRET;
1146
1147         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1148                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1149                (request->rq_export ?
1150                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1151                (request->rq_export ?
1152                 atomic_read(&request->rq_export->exp_refcount) : -99),
1153                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1154                libcfs_id2str(request->rq_peer),
1155                lustre_msg_get_opc(request->rq_reqmsg));
1156
1157         OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1158
1159         rc = svc->srv_handler(request);
1160
1161         request->rq_phase = RQ_PHASE_COMPLETE;
1162
1163         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1164                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1165                (request->rq_export ?
1166                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1167                (request->rq_export ?
1168                 atomic_read(&request->rq_export->exp_refcount) : -99),
1169                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1170                libcfs_id2str(request->rq_peer),
1171                lustre_msg_get_opc(request->rq_reqmsg));
1172
1173 put_rpc_export:
1174         if (export != NULL)
1175                 class_export_rpc_put(export);
1176 put_conn:
1177         if (likely(request->rq_export != NULL))
1178                 class_export_put(request->rq_export);
1179
1180         lu_context_exit(&request->rq_session);
1181         lu_context_fini(&request->rq_session);
1182
1183         if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1184                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1185                           "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1186                           " client may timeout.",
1187                           request->rq_xid, cfs_time_sub(request->rq_deadline,
1188                           request->rq_arrival_time.tv_sec),
1189                           cfs_time_sub(cfs_time_current_sec(),
1190                           request->rq_deadline));
1191         }
1192
1193         do_gettimeofday(&work_end);
1194         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1195         CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1196                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1197                request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1198                libcfs_id2str(request->rq_peer), timediff,
1199                cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1200                request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1201                request->rq_transno, request->rq_status,
1202                request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1203                -999);
1204         if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1205                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1206                 int opc = opcode_offset(op);
1207                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1208                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1209                         lprocfs_counter_add(svc->srv_stats,
1210                                             opc + EXTRA_MAX_OPCODES,
1211                                             timediff);
1212                 }
1213         }
1214         if (unlikely(request->rq_early_count)) {
1215                 DEBUG_REQ(D_ADAPTTO, request,
1216                           "sent %d early replies before finishing in "
1217                           CFS_DURATION_T"s",
1218                           request->rq_early_count,
1219                           cfs_time_sub(work_end.tv_sec,
1220                           request->rq_arrival_time.tv_sec));
1221         }
1222
1223 out_req:
1224         ptlrpc_server_finish_request(request);
1225
1226         RETURN(1);
1227 }
1228
1229 static int
1230 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1231 {
1232         struct ptlrpc_reply_state *rs;
1233         struct obd_export         *exp;
1234         struct obd_device         *obd;
1235         int                        nlocks;
1236         int                        been_handled;
1237         ENTRY;
1238
1239         spin_lock(&svc->srv_lock);
1240         if (list_empty (&svc->srv_reply_queue)) {
1241                 spin_unlock(&svc->srv_lock);
1242                 RETURN(0);
1243         }
1244
1245         rs = list_entry (svc->srv_reply_queue.next,
1246                          struct ptlrpc_reply_state, rs_list);
1247
1248         exp = rs->rs_export;
1249         obd = exp->exp_obd;
1250
1251         LASSERT (rs->rs_difficult);
1252         LASSERT (rs->rs_scheduled);
1253
1254         list_del_init (&rs->rs_list);
1255
1256         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1257         spin_unlock(&svc->srv_lock);
1258
1259         spin_lock (&obd->obd_uncommitted_replies_lock);
1260         /* Noop if removed already */
1261         list_del_init (&rs->rs_obd_list);
1262         spin_unlock (&obd->obd_uncommitted_replies_lock);
1263
1264         spin_lock (&exp->exp_lock);
1265         /* Noop if removed already */
1266         list_del_init (&rs->rs_exp_list);
1267         spin_unlock (&exp->exp_lock);
1268
1269         spin_lock(&svc->srv_lock);
1270
1271         been_handled = rs->rs_handled;
1272         rs->rs_handled = 1;
1273
1274         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1275         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1276
1277         if (nlocks == 0 && !been_handled) {
1278                 /* If we see this, we should already have seen the warning
1279                  * in mds_steal_ack_locks()  */
1280                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1281                       " o%d NID %s\n",
1282                       rs,
1283                       rs->rs_xid, rs->rs_transno,
1284                       lustre_msg_get_opc(rs->rs_msg),
1285                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1286         }
1287
1288         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1289                 spin_unlock(&svc->srv_lock);
1290
1291                 if (!been_handled && rs->rs_on_net) {
1292                         LNetMDUnlink(rs->rs_md_h);
1293                         /* Ignore return code; we're racing with
1294                          * completion... */
1295                 }
1296
1297                 while (nlocks-- > 0)
1298                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1299                                          rs->rs_modes[nlocks]);
1300
1301                 spin_lock(&svc->srv_lock);
1302         }
1303
1304         rs->rs_scheduled = 0;
1305
1306         if (!rs->rs_on_net) {
1307                 /* Off the net */
1308                 svc->srv_n_difficult_replies--;
1309                 spin_unlock(&svc->srv_lock);
1310
1311                 class_export_put (exp);
1312                 rs->rs_export = NULL;
1313                 ptlrpc_rs_decref (rs);
1314                 atomic_dec (&svc->srv_outstanding_replies);
1315                 RETURN(1);
1316         }
1317
1318         /* still on the net; callback will schedule */
1319         spin_unlock(&svc->srv_lock);
1320         RETURN(1);
1321 }
1322
1323 #ifndef __KERNEL__
1324 /* FIXME make use of timeout later */
1325 int
1326 liblustre_check_services (void *arg)
1327 {
1328         int  did_something = 0;
1329         int  rc;
1330         struct list_head *tmp, *nxt;
1331         ENTRY;
1332
1333         /* I'm relying on being single threaded, not to have to lock
1334          * ptlrpc_all_services etc */
1335         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1336                 struct ptlrpc_service *svc =
1337                         list_entry (tmp, struct ptlrpc_service, srv_list);
1338
1339                 if (svc->srv_threads_running != 0)     /* I've recursed */
1340                         continue;
1341
1342                 /* service threads can block for bulk, so this limits us
1343                  * (arbitrarily) to recursing 1 stack frame per service.
1344                  * Note that the problem with recursion is that we have to
1345                  * unwind completely before our caller can resume. */
1346
1347                 svc->srv_threads_running++;
1348
1349                 do {
1350                         rc = ptlrpc_server_handle_req_in(svc);
1351                         rc |= ptlrpc_server_handle_reply(svc);
1352                         rc |= ptlrpc_at_check_timed(svc);
1353                         rc |= ptlrpc_server_handle_request(svc, NULL);
1354                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1355                         did_something |= rc;
1356                 } while (rc);
1357
1358                 svc->srv_threads_running--;
1359         }
1360
1361         RETURN(did_something);
1362 }
1363 #define ptlrpc_stop_all_threads(s) do {} while (0)
1364
1365 #else /* __KERNEL__ */
1366
1367 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1368 void ptlrpc_daemonize(char *name)
1369 {
1370         struct fs_struct *fs = current->fs;
1371
1372         atomic_inc(&fs->count);
1373         cfs_daemonize(name);
1374         exit_fs(cfs_current());
1375         current->fs = fs;
1376         ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1377 }
1378
1379 static void
1380 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1381 {
1382         int avail = svc->srv_nrqbd_receiving;
1383         int low_water = test_req_buffer_pressure ? 0 :
1384                         svc->srv_nbuf_per_group/2;
1385
1386         /* NB I'm not locking; just looking. */
1387
1388         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1389          * allowed the request history to grow out of control.  We could put a
1390          * sanity check on that here and cull some history if we need the
1391          * space. */
1392
1393         if (avail <= low_water)
1394                 ptlrpc_grow_req_bufs(svc);
1395
1396         if (svc->srv_stats)
1397                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1398                                     avail);
1399 }
1400
1401 static int
1402 ptlrpc_retry_rqbds(void *arg)
1403 {
1404         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1405
1406         svc->srv_rqbd_timeout = 0;
1407         return (-ETIMEDOUT);
1408 }
1409
1410 static int ptlrpc_main(void *arg)
1411 {
1412         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1413         struct ptlrpc_service  *svc = data->svc;
1414         struct ptlrpc_thread   *thread = data->thread;
1415         struct obd_device      *dev = data->dev;
1416         struct ptlrpc_reply_state *rs;
1417         struct lc_watchdog     *watchdog;
1418 #ifdef WITH_GROUP_INFO
1419         struct group_info *ginfo = NULL;
1420 #endif
1421         struct lu_env env;
1422         int counter = 0, rc = 0;
1423         ENTRY;
1424
1425         ptlrpc_daemonize(data->name);
1426
1427 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1428         /* we need to do this before any per-thread allocation is done so that
1429          * we get the per-thread allocations on local node.  bug 7342 */
1430         if (svc->srv_cpu_affinity) {
1431                 int cpu, num_cpu;
1432
1433                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1434                         if (!cpu_online(cpu))
1435                                 continue;
1436                         if (num_cpu == thread->t_id % num_online_cpus())
1437                                 break;
1438                         num_cpu++;
1439                 }
1440                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1441         }
1442 #endif
1443
1444 #ifdef WITH_GROUP_INFO
1445         ginfo = groups_alloc(0);
1446         if (!ginfo) {
1447                 rc = -ENOMEM;
1448                 goto out;
1449         }
1450
1451         set_current_groups(ginfo);
1452         put_group_info(ginfo);
1453 #endif
1454
1455         if (svc->srv_init != NULL) {
1456                 rc = svc->srv_init(thread);
1457                 if (rc)
1458                         goto out;
1459         }
1460
1461         rc = lu_context_init(&env.le_ctx, svc->srv_ctx_tags);
1462         if (rc)
1463                 goto out_srv_fini;
1464
1465         thread->t_env = &env;
1466         env.le_ctx.lc_thread = thread;
1467
1468         /* Alloc reply state structure for this one */
1469         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1470         if (!rs) {
1471                 rc = -ENOMEM;
1472                 goto out_srv_fini;
1473         }
1474
1475         /* Record that the thread is running */
1476         thread->t_flags = SVC_RUNNING;
1477         /*
1478          * wake up our creator. Note: @data is invalid after this point,
1479          * because it's allocated on ptlrpc_start_thread() stack.
1480          */
1481         cfs_waitq_signal(&thread->t_ctl_waitq);
1482
1483         watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1484                                    at_get(&svc->srv_at_estimate)) *
1485                                    svc->srv_watchdog_factor, NULL, NULL);
1486
1487         spin_lock(&svc->srv_lock);
1488         svc->srv_threads_running++;
1489         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1490         spin_unlock(&svc->srv_lock);
1491         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1492
1493         CDEBUG(D_NET, "service thread %d (#%d)started\n", thread->t_id,
1494               svc->srv_threads_running);
1495
1496         /* XXX maintain a list of all managed devices: insert here */
1497
1498         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1499                svc->srv_n_difficult_replies != 0) {
1500                 /* Don't exit while there are replies to be handled */
1501                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1502                                                      ptlrpc_retry_rqbds, svc);
1503
1504                 lc_watchdog_disable(watchdog);
1505
1506                 cond_resched();
1507
1508                 l_wait_event_exclusive (svc->srv_waitq,
1509                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1510                                svc->srv_n_difficult_replies == 0) ||
1511                               (!list_empty(&svc->srv_idle_rqbds) &&
1512                                svc->srv_rqbd_timeout == 0) ||
1513                               !list_empty(&svc->srv_req_in_queue) ||
1514                               !list_empty(&svc->srv_reply_queue) ||
1515                               (!list_empty(&svc->srv_request_queue) &&
1516                                (svc->srv_n_active_reqs <
1517                                 (svc->srv_threads_running - 1))) ||
1518                               svc->srv_at_check,
1519                               &lwi);
1520
1521                 lc_watchdog_touch_ms(watchdog, max_t(int, obd_timeout,
1522                                      AT_OFF ? 0 :
1523                                      at_get(&svc->srv_at_estimate)) *
1524                                      svc->srv_watchdog_factor);
1525
1526                 ptlrpc_check_rqbd_pool(svc);
1527
1528                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1529                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1530                         /* Ignore return code - we tried... */
1531                         ptlrpc_start_thread(dev, svc);
1532                 }
1533
1534                 if (!list_empty(&svc->srv_reply_queue))
1535                         ptlrpc_server_handle_reply(svc);
1536
1537                 if (!list_empty(&svc->srv_req_in_queue)) {
1538                         /* Process all incoming reqs before handling any */
1539                         ptlrpc_server_handle_req_in(svc);
1540                         /* but limit ourselves in case of flood */
1541                         if (counter++ < 1000)
1542                                 continue;
1543                         counter = 0;
1544                 }
1545
1546                 if (svc->srv_at_check)
1547                         ptlrpc_at_check_timed(svc);
1548
1549                 /* don't handle requests in the last thread */
1550                 if (!list_empty (&svc->srv_request_queue) &&
1551                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1552                         lu_context_enter(&env.le_ctx);
1553                         ptlrpc_server_handle_request(svc, thread);
1554                         lu_context_exit(&env.le_ctx);
1555                 }
1556
1557                 if (!list_empty(&svc->srv_idle_rqbds) &&
1558                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1559                         /* I just failed to repost request buffers.  Wait
1560                          * for a timeout (unless something else happens)
1561                          * before I try again */
1562                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1563                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1564                                svc->srv_nrqbd_receiving);
1565                 }
1566         }
1567
1568         lc_watchdog_delete(watchdog);
1569
1570 out_srv_fini:
1571         /*
1572          * deconstruct service specific state created by ptlrpc_start_thread()
1573          */
1574         if (svc->srv_done != NULL)
1575                 svc->srv_done(thread);
1576
1577         lu_context_fini(&env.le_ctx);
1578 out:
1579         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1580
1581         spin_lock(&svc->srv_lock);
1582         svc->srv_threads_running--; /* must know immediately */
1583         thread->t_id = rc;
1584         thread->t_flags = SVC_STOPPED;
1585
1586         cfs_waitq_signal(&thread->t_ctl_waitq);
1587         spin_unlock(&svc->srv_lock);
1588
1589         return rc;
1590 }
1591
1592 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1593                                struct ptlrpc_thread *thread)
1594 {
1595         struct l_wait_info lwi = { 0 };
1596
1597         spin_lock(&svc->srv_lock);
1598         thread->t_flags = SVC_STOPPING;
1599         spin_unlock(&svc->srv_lock);
1600
1601         cfs_waitq_broadcast(&svc->srv_waitq);
1602         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1603                      &lwi);
1604
1605         spin_lock(&svc->srv_lock);
1606         list_del(&thread->t_link);
1607         spin_unlock(&svc->srv_lock);
1608
1609         OBD_FREE_PTR(thread);
1610 }
1611
1612 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1613 {
1614         struct ptlrpc_thread *thread;
1615
1616         spin_lock(&svc->srv_lock);
1617         while (!list_empty(&svc->srv_threads)) {
1618                 thread = list_entry(svc->srv_threads.next,
1619                                     struct ptlrpc_thread, t_link);
1620
1621                 spin_unlock(&svc->srv_lock);
1622                 ptlrpc_stop_thread(svc, thread);
1623                 spin_lock(&svc->srv_lock);
1624         }
1625
1626         spin_unlock(&svc->srv_lock);
1627 }
1628
1629 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1630 {
1631         int i, rc = 0;
1632         ENTRY;
1633
1634         /* We require 2 threads min - see note in
1635            ptlrpc_server_handle_request */
1636         LASSERT(svc->srv_threads_min >= 2);
1637         for (i = 0; i < svc->srv_threads_min; i++) {
1638                 rc = ptlrpc_start_thread(dev, svc);
1639                 /* We have enough threads, don't start more.  b=15759 */
1640                 if (rc == -EMFILE)
1641                         break;
1642                 if (rc) {
1643                         CERROR("cannot start %s thread #%d: rc %d\n",
1644                                svc->srv_thread_name, i, rc);
1645                         ptlrpc_stop_all_threads(svc);
1646                 }
1647         }
1648         RETURN(rc);
1649 }
1650
1651 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1652 {
1653         struct l_wait_info lwi = { 0 };
1654         struct ptlrpc_svc_data d;
1655         struct ptlrpc_thread *thread;
1656         char name[32];
1657         int id, rc;
1658         ENTRY;
1659
1660         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1661                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1662                svc->srv_threads_max, svc->srv_threads_running);
1663         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1664             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1665              svc->srv_threads_started == svc->srv_threads_min - 1))
1666                 RETURN(-EMFILE);
1667
1668         OBD_ALLOC_PTR(thread);
1669         if (thread == NULL)
1670                 RETURN(-ENOMEM);
1671         cfs_waitq_init(&thread->t_ctl_waitq);
1672
1673         spin_lock(&svc->srv_lock);
1674         if (svc->srv_threads_started >= svc->srv_threads_max) {
1675                 spin_unlock(&svc->srv_lock);
1676                 OBD_FREE_PTR(thread);
1677                 RETURN(-EMFILE);
1678         }
1679         list_add(&thread->t_link, &svc->srv_threads);
1680         id = svc->srv_threads_started++;
1681         spin_unlock(&svc->srv_lock);
1682
1683         thread->t_id = id;
1684         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1685         d.dev = dev;
1686         d.svc = svc;
1687         d.name = name;
1688         d.thread = thread;
1689
1690         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1691
1692           /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1693          * just drop the VM and FILES in ptlrpc_daemonize() right away.
1694          */
1695         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1696         if (rc < 0) {
1697                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1698
1699                 spin_lock(&svc->srv_lock);
1700                 list_del(&thread->t_link);
1701                 --svc->srv_threads_started;
1702                 spin_unlock(&svc->srv_lock);
1703
1704                 OBD_FREE(thread, sizeof(*thread));
1705                 RETURN(rc);
1706         }
1707         l_wait_event(thread->t_ctl_waitq,
1708                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1709
1710         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1711         RETURN(rc);
1712 }
1713 #endif
1714
1715 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1716 {
1717         int                   rc;
1718         struct l_wait_info    lwi;
1719         struct list_head     *tmp;
1720         struct ptlrpc_reply_state *rs, *t;
1721
1722         cfs_timer_disarm(&service->srv_at_timer);
1723
1724         ptlrpc_stop_all_threads(service);
1725         LASSERT(list_empty(&service->srv_threads));
1726
1727         spin_lock (&ptlrpc_all_services_lock);
1728         list_del_init (&service->srv_list);
1729         spin_unlock (&ptlrpc_all_services_lock);
1730
1731         ptlrpc_lprocfs_unregister_service(service);
1732
1733         /* All history will be culled when the next request buffer is
1734          * freed */
1735         service->srv_max_history_rqbds = 0;
1736
1737         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1738
1739         rc = LNetClearLazyPortal(service->srv_req_portal);
1740         LASSERT (rc == 0);
1741
1742         /* Unlink all the request buffers.  This forces a 'final' event with
1743          * its 'unlink' flag set for each posted rqbd */
1744         list_for_each(tmp, &service->srv_active_rqbds) {
1745                 struct ptlrpc_request_buffer_desc *rqbd =
1746                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1747                                    rqbd_list);
1748
1749                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1750                 LASSERT (rc == 0 || rc == -ENOENT);
1751         }
1752
1753         /* Wait for the network to release any buffers it's currently
1754          * filling */
1755         for (;;) {
1756                 spin_lock(&service->srv_lock);
1757                 rc = service->srv_nrqbd_receiving;
1758                 spin_unlock(&service->srv_lock);
1759
1760                 if (rc == 0)
1761                         break;
1762
1763                 /* Network access will complete in finite time but the HUGE
1764                  * timeout lets us CWARN for visibility of sluggish NALs */
1765                 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1766                 rc = l_wait_event(service->srv_waitq,
1767                                   service->srv_nrqbd_receiving == 0,
1768                                   &lwi);
1769                 if (rc == -ETIMEDOUT)
1770                         CWARN("Service %s waiting for request buffers\n",
1771                               service->srv_name);
1772         }
1773
1774         /* schedule all outstanding replies to terminate them */
1775         spin_lock(&service->srv_lock);
1776         while (!list_empty(&service->srv_active_replies)) {
1777                 struct ptlrpc_reply_state *rs =
1778                         list_entry(service->srv_active_replies.next,
1779                                    struct ptlrpc_reply_state, rs_list);
1780                 ptlrpc_schedule_difficult_reply(rs);
1781         }
1782         spin_unlock(&service->srv_lock);
1783
1784         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1785          * and no service threads, so I'm the only thread noodling the
1786          * request queue now */
1787         while (!list_empty(&service->srv_req_in_queue)) {
1788                 struct ptlrpc_request *req =
1789                         list_entry(service->srv_req_in_queue.next,
1790                                    struct ptlrpc_request,
1791                                    rq_list);
1792
1793                 list_del(&req->rq_list);
1794                 service->srv_n_queued_reqs--;
1795                 service->srv_n_active_reqs++;
1796                 ptlrpc_server_finish_request(req);
1797         }
1798         while (!list_empty(&service->srv_request_queue)) {
1799                 struct ptlrpc_request *req =
1800                         list_entry(service->srv_request_queue.next,
1801                                    struct ptlrpc_request,
1802                                    rq_list);
1803
1804                 list_del(&req->rq_list);
1805                 service->srv_n_queued_reqs--;
1806                 service->srv_n_active_reqs++;
1807
1808                 ptlrpc_server_finish_request(req);
1809         }
1810         LASSERT(service->srv_n_queued_reqs == 0);
1811         LASSERT(service->srv_n_active_reqs == 0);
1812         LASSERT(service->srv_n_history_rqbds == 0);
1813         LASSERT(list_empty(&service->srv_active_rqbds));
1814
1815         /* Now free all the request buffers since nothing references them
1816          * any more... */
1817         while (!list_empty(&service->srv_idle_rqbds)) {
1818                 struct ptlrpc_request_buffer_desc *rqbd =
1819                         list_entry(service->srv_idle_rqbds.next,
1820                                    struct ptlrpc_request_buffer_desc,
1821                                    rqbd_list);
1822
1823                 ptlrpc_free_rqbd(rqbd);
1824         }
1825
1826         /* wait for all outstanding replies to complete (they were
1827          * scheduled having been flagged to abort above) */
1828         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1829                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1830
1831                 rc = l_wait_event(service->srv_waitq,
1832                                   !list_empty(&service->srv_reply_queue), &lwi);
1833                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1834
1835                 if (rc == 0) {
1836                         ptlrpc_server_handle_reply(service);
1837                         continue;
1838                 }
1839                 CWARN("Unexpectedly long timeout %p\n", service);
1840         }
1841
1842         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1843                 list_del(&rs->rs_list);
1844                 OBD_FREE(rs, service->srv_max_reply_size);
1845         }
1846
1847         /* In case somebody rearmed this in the meantime */
1848         cfs_timer_disarm(&service->srv_at_timer);
1849
1850         OBD_FREE_PTR(service);
1851         return 0;
1852 }
1853
1854 /* Returns 0 if the service is healthy.
1855  *
1856  * Right now, it just checks to make sure that requests aren't languishing
1857  * in the queue.  We'll use this health check to govern whether a node needs
1858  * to be shot, so it's intentionally non-aggressive. */
1859 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
1860 {
1861         struct ptlrpc_request *request;
1862         struct timeval         right_now;
1863         long                   timediff;
1864
1865         if (svc == NULL)
1866                 return 0;
1867
1868         do_gettimeofday(&right_now);
1869
1870         spin_lock(&svc->srv_lock);
1871         if (list_empty(&svc->srv_request_queue)) {
1872                 spin_unlock(&svc->srv_lock);
1873                 return 0;
1874         }
1875
1876         /* How long has the next entry been waiting? */
1877         request = list_entry(svc->srv_request_queue.next,
1878                              struct ptlrpc_request, rq_list);
1879         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
1880         spin_unlock(&svc->srv_lock);
1881
1882         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
1883                                         at_max)) {
1884                 CERROR("%s: unhealthy - request has been waiting %lds\n",
1885                        svc->srv_name, timediff / ONE_MILLION);
1886                 return (-1);
1887         }
1888
1889         return 0;
1890 }