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 #include <libcfs/kp30.h>
  41 #endif
  42 #include <obd_support.h>
  43 #include <obd_class.h>
  44 #include <lustre_net.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
  53 CFS_MODULE_PARM(at_min, "i", int, 0644,
  54                 "Adaptive timeout minimum (sec)");
  55 CFS_MODULE_PARM(at_max, "i", int, 0644,
  56                 "Adaptive timeout maximum (sec)");
  57 CFS_MODULE_PARM(at_history, "i", int, 0644,
  58                 "Adaptive timeouts remember the slowest event that took place "
  59                 "within this period (sec)");
  60 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
  61                 "How soon before an RPC deadline to send an early reply");
  62 CFS_MODULE_PARM(at_extra, "i", int, 0644,
  63                 "How much extra time to give with each early reply");
  64
  65 /* forward ref */
  66 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
  67
  68 static CFS_LIST_HEAD (ptlrpc_all_services);
  69 spinlock_t ptlrpc_all_services_lock;
  70
  71 static char *
  72 ptlrpc_alloc_request_buffer (int size)
  73 {
  74         char *ptr;
  75
  76         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  77                 OBD_VMALLOC(ptr, size);
  78         else
  79                 OBD_ALLOC(ptr, size);
  80
  81         return (ptr);
  82 }
  83
  84 static void
  85 ptlrpc_free_request_buffer (char *ptr, int size)
  86 {
  87         if (size > SVC_BUF_VMALLOC_THRESHOLD)
  88                 OBD_VFREE(ptr, size);
  89         else
  90                 OBD_FREE(ptr, size);
  91 }
  92
  93 struct ptlrpc_request_buffer_desc *
  94 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
  95 {
  96         struct ptlrpc_request_buffer_desc *rqbd;
  97
  98         OBD_ALLOC(rqbd, sizeof (*rqbd));
  99         if (rqbd == NULL)
 100                 return (NULL);
 101
 102         rqbd->rqbd_service = svc;
 103         rqbd->rqbd_refcount = 0;
 104         rqbd->rqbd_cbid.cbid_fn = request_in_callback;
 105         rqbd->rqbd_cbid.cbid_arg = rqbd;
 106         CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
 107         rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
 108
 109         if (rqbd->rqbd_buffer == NULL) {
 110                 OBD_FREE(rqbd, sizeof (*rqbd));
 111                 return (NULL);
 112         }
 113
 114         spin_lock(&svc->srv_lock);
 115         list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 116         svc->srv_nbufs++;
 117         spin_unlock(&svc->srv_lock);
 118
 119         return (rqbd);
 120 }
 121
 122 void
 123 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
 124 {
 125         struct ptlrpc_service *svc = rqbd->rqbd_service;
 126
 127         LASSERT (rqbd->rqbd_refcount == 0);
 128         LASSERT (list_empty(&rqbd->rqbd_reqs));
 129
 130         spin_lock(&svc->srv_lock);
 131         list_del(&rqbd->rqbd_list);
 132         svc->srv_nbufs--;
 133         spin_unlock(&svc->srv_lock);
 134
 135         ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
 136         OBD_FREE (rqbd, sizeof (*rqbd));
 137 }
 138
 139 int
 140 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
 141 {
 142         struct ptlrpc_request_buffer_desc *rqbd;
 143         int                                i;
 144
 145         CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
 146                svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
 147                svc->srv_nrqbd_receiving, svc->srv_nbufs);
 148         for (i = 0; i < svc->srv_nbuf_per_group; i++) {
 149                 rqbd = ptlrpc_alloc_rqbd(svc);
 150
 151                 if (rqbd == NULL) {
 152                         CERROR ("%s: Can't allocate request buffer\n",
 153                                 svc->srv_name);
 154                         return (-ENOMEM);
 155                 }
 156
 157                 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
 158                         return (-EAGAIN);
 159         }
 160
 161         return (0);
 162 }
 163
 164 void
 165 ptlrpc_save_lock(struct ptlrpc_request *req,
 166                  struct lustre_handle *lock, int mode)
 167 {
 168         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 169         int                        idx;
 170
 171         LASSERT(rs != NULL);
 172         LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
 173
 174         if (req->rq_export->exp_disconnected) {
 175                 ldlm_lock_decref(lock, mode);
 176         } else {
 177                 idx = rs->rs_nlocks++;
 178                 rs->rs_locks[idx] = *lock;
 179                 rs->rs_modes[idx] = mode;
 180                 rs->rs_difficult = 1;
 181         }
 182 }
 183
 184 void
 185 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
 186 {
 187         struct ptlrpc_service *svc = rs->rs_service;
 188
 189         LASSERT_SPIN_LOCKED (&svc->srv_lock);
 190         LASSERT (rs->rs_difficult);
 191         rs->rs_scheduled_ever = 1;              /* flag any notification attempt */
 192
 193         if (rs->rs_scheduled)                   /* being set up or already notified */
 194                 return;
 195
 196         rs->rs_scheduled = 1;
 197         list_del (&rs->rs_list);
 198         list_add (&rs->rs_list, &svc->srv_reply_queue);
 199         cfs_waitq_signal (&svc->srv_waitq);
 200 }
 201
 202 void
 203 ptlrpc_commit_replies (struct obd_export *exp)
 204 {
 205         struct list_head   *tmp;
 206         struct list_head   *nxt;
 207
 208         /* Find any replies that have been committed and get their service
 209          * to attend to complete them. */
 210
 211         /* CAVEAT EMPTOR: spinlock ordering!!! */
 212         spin_lock(&exp->exp_uncommitted_replies_lock);
 213
 214         list_for_each_safe(tmp, nxt, &exp->exp_uncommitted_replies) {
 215                 struct ptlrpc_reply_state *rs =
 216                         list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
 217
 218                 LASSERT(rs->rs_difficult);
 219                 /* VBR: per-export last_committed */
 220                 LASSERT(rs->rs_export);
 221                 if (rs->rs_transno <= rs->rs_export->exp_last_committed) {
 222                         struct ptlrpc_service *svc = rs->rs_service;
 223
 224                         spin_lock (&svc->srv_lock);
 225                         list_del_init (&rs->rs_obd_list);
 226                         ptlrpc_schedule_difficult_reply (rs);
 227                         spin_unlock (&svc->srv_lock);
 228                 }
 229         }
 230
 231         spin_unlock(&exp->exp_uncommitted_replies_lock);
 232 }
 233
 234 static int
 235 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
 236 {
 237         struct ptlrpc_request_buffer_desc *rqbd;
 238         int                                rc;
 239         int                                posted = 0;
 240
 241         for (;;) {
 242                 spin_lock(&svc->srv_lock);
 243
 244                 if (list_empty (&svc->srv_idle_rqbds)) {
 245                         spin_unlock(&svc->srv_lock);
 246                         return (posted);
 247                 }
 248
 249                 rqbd = list_entry(svc->srv_idle_rqbds.next,
 250                                   struct ptlrpc_request_buffer_desc,
 251                                   rqbd_list);
 252                 list_del (&rqbd->rqbd_list);
 253
 254                 /* assume we will post successfully */
 255                 svc->srv_nrqbd_receiving++;
 256                 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
 257
 258                 spin_unlock(&svc->srv_lock);
 259
 260                 rc = ptlrpc_register_rqbd(rqbd);
 261                 if (rc != 0)
 262                         break;
 263
 264                 posted = 1;
 265         }
 266
 267         spin_lock(&svc->srv_lock);
 268
 269         svc->srv_nrqbd_receiving--;
 270         list_del(&rqbd->rqbd_list);
 271         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 272
 273         /* Don't complain if no request buffers are posted right now; LNET
 274          * won't drop requests because we set the portal lazy! */
 275
 276         spin_unlock(&svc->srv_lock);
 277
 278         return (-1);
 279 }
 280
 281 static void ptlrpc_at_timer(unsigned long castmeharder)
 282 {
 283         struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
 284         svc->srv_at_check = 1;
 285         svc->srv_at_checktime = cfs_time_current();
 286         cfs_waitq_signal(&svc->srv_waitq);
 287 }
 288
 289 /* @threadname should be 11 characters or less - 3 will be added on */
 290 struct ptlrpc_service *
 291 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
 292                 int req_portal, int rep_portal, int watchdog_factor,
 293                 svc_handler_t handler, char *name,
 294                 cfs_proc_dir_entry_t *proc_entry,
 295                 svcreq_printfn_t svcreq_printfn,
 296                 int min_threads, int max_threads, char *threadname,
 297                 svc_hpreq_handler_t hp_handler)
 298 {
 299         int                     rc;
 300         struct ptlrpc_at_array *array;
 301         struct ptlrpc_service  *service;
 302         unsigned int            size, index;
 303         ENTRY;
 304
 305         LASSERT (nbufs > 0);
 306         LASSERT (bufsize >= max_req_size);
 307
 308         OBD_ALLOC(service, sizeof(*service));
 309         if (service == NULL)
 310                 RETURN(NULL);
 311
 312         /* First initialise enough for early teardown */
 313
 314         service->srv_name = name;
 315         spin_lock_init(&service->srv_lock);
 316         CFS_INIT_LIST_HEAD(&service->srv_threads);
 317         cfs_waitq_init(&service->srv_waitq);
 318
 319         service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
 320         service->srv_max_req_size = max_req_size;
 321         service->srv_buf_size = bufsize;
 322         service->srv_rep_portal = rep_portal;
 323         service->srv_req_portal = req_portal;
 324         service->srv_watchdog_factor = watchdog_factor;
 325         service->srv_handler = handler;
 326         service->srv_request_history_print_fn = svcreq_printfn;
 327         service->srv_request_seq = 1;           /* valid seq #s start at 1 */
 328         service->srv_request_max_cull_seq = 0;
 329         service->srv_threads_min = min_threads;
 330         service->srv_threads_max = max_threads;
 331         service->srv_thread_name = threadname;
 332         service->srv_hpreq_handler = hp_handler;
 333         service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
 334         service->srv_hpreq_count = 0;
 335         service->srv_n_hpreq = 0;
 336
 337         rc = LNetSetLazyPortal(service->srv_req_portal);
 338         LASSERT (rc == 0);
 339
 340         CFS_INIT_LIST_HEAD(&service->srv_request_queue);
 341         CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
 342         CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
 343         CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
 344         CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
 345         CFS_INIT_LIST_HEAD(&service->srv_request_history);
 346         CFS_INIT_LIST_HEAD(&service->srv_active_replies);
 347         CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
 348         CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
 349         cfs_waitq_init(&service->srv_free_rs_waitq);
 350
 351         spin_lock_init(&service->srv_at_lock);
 352         CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
 353
 354         array = &service->srv_at_array;
 355         size = at_est2timeout(at_max);
 356         array->paa_size = size;
 357         array->paa_count = 0;
 358         array->paa_deadline = -1;
 359
 360         /* allocate memory for srv_at_array (ptlrpc_at_array) */
 361         OBD_ALLOC(array->paa_reqs_array, sizeof(struct list_head) * size);
 362         if (array->paa_reqs_array == NULL)
 363                 GOTO(failed, NULL);
 364
 365         for (index = 0; index < size; index++)
 366                 CFS_INIT_LIST_HEAD(&array->paa_reqs_array[index]);
 367
 368         OBD_ALLOC(array->paa_reqs_count, sizeof(__u32) * size);
 369         if (array->paa_reqs_count == NULL)
 370                 GOTO(failed, NULL);
 371
 372         cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
 373         /* At SOW, service time should be quick; 10s seems generous. If client
 374            timeout is less than this, we'll be sending an early reply. */
 375         at_init(&service->srv_at_estimate, 10, 0);
 376
 377         spin_lock (&ptlrpc_all_services_lock);
 378         list_add (&service->srv_list, &ptlrpc_all_services);
 379         spin_unlock (&ptlrpc_all_services_lock);
 380
 381         /* Now allocate the request buffers */
 382         rc = ptlrpc_grow_req_bufs(service);
 383         /* We shouldn't be under memory pressure at startup, so
 384          * fail if we can't post all our buffers at this time. */
 385         if (rc != 0)
 386                 GOTO(failed, NULL);
 387
 388         /* Now allocate pool of reply buffers */
 389         /* Increase max reply size to next power of two */
 390         service->srv_max_reply_size = 1;
 391         while (service->srv_max_reply_size < max_reply_size)
 392                 service->srv_max_reply_size <<= 1;
 393
 394         if (proc_entry != NULL)
 395                 ptlrpc_lprocfs_register_service(proc_entry, service);
 396
 397         CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
 398                service->srv_name, service->srv_req_portal);
 399
 400         RETURN(service);
 401 failed:
 402         ptlrpc_unregister_service(service);
 403         return NULL;
 404 }
 405
 406 /**
 407  * to actually free the request, must be called without holding svc_lock.
 408  * note it's caller's responsibility to unlink req->rq_list.
 409  */
 410 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
 411 {
 412         LASSERT(atomic_read(&req->rq_refcount) == 0);
 413         LASSERT(list_empty(&req->rq_timed_list));
 414
 415         /* DEBUG_REQ() assumes the reply state of a request with a valid
 416          * ref will not be destroyed until that reference is dropped. */
 417         ptlrpc_req_drop_rs(req);
 418
 419         if (req != &req->rq_rqbd->rqbd_req) {
 420                 /* NB request buffers use an embedded
 421                  * req if the incoming req unlinked the
 422                  * MD; this isn't one of them! */
 423                 OBD_FREE(req, sizeof(*req));
 424         }
 425 }
 426
 427 /**
 428  * drop a reference count of the request. if it reaches 0, we either
 429  * put it into history list, or free it immediately.
 430  */
 431 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
 432 {
 433         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 434         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 435         int                                refcount;
 436         struct list_head                  *tmp;
 437         struct list_head                  *nxt;
 438
 439         if (!atomic_dec_and_test(&req->rq_refcount))
 440                 return;
 441
 442         spin_lock(&svc->srv_at_lock);
 443         list_del_init(&req->rq_timed_list);
 444         if (req->rq_at_linked) {
 445                 struct ptlrpc_at_array *array = &svc->srv_at_array;
 446                 __u32 index = req->rq_at_index;
 447
 448                 req->rq_at_linked = 0;
 449                 array->paa_reqs_count[index]--;
 450                 array->paa_count--;
 451         }
 452         spin_unlock(&svc->srv_at_lock);
 453
 454         /* finalize request */
 455         if (req->rq_export) {
 456                 class_export_put(req->rq_export);
 457                 req->rq_export = NULL;
 458         }
 459
 460         spin_lock(&svc->srv_lock);
 461
 462         svc->srv_n_active_reqs--;
 463         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 464
 465         refcount = --(rqbd->rqbd_refcount);
 466         if (refcount == 0) {
 467                 /* request buffer is now idle: add to history */
 468                 list_del(&rqbd->rqbd_list);
 469                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 470                 svc->srv_n_history_rqbds++;
 471
 472                 /* cull some history?
 473                  * I expect only about 1 or 2 rqbds need to be recycled here */
 474                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 475                         rqbd = list_entry(svc->srv_history_rqbds.next,
 476                                           struct ptlrpc_request_buffer_desc,
 477                                           rqbd_list);
 478
 479                         list_del(&rqbd->rqbd_list);
 480                         svc->srv_n_history_rqbds--;
 481
 482                         /* remove rqbd's reqs from svc's req history while
 483                          * I've got the service lock */
 484                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 485                                 req = list_entry(tmp, struct ptlrpc_request,
 486                                                  rq_list);
 487                                 /* Track the highest culled req seq */
 488                                 if (req->rq_history_seq >
 489                                     svc->srv_request_max_cull_seq)
 490                                         svc->srv_request_max_cull_seq =
 491                                                 req->rq_history_seq;
 492                                 list_del(&req->rq_history_list);
 493                         }
 494
 495                         spin_unlock(&svc->srv_lock);
 496
 497                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 498                                 req = list_entry(rqbd->rqbd_reqs.next,
 499                                                  struct ptlrpc_request,
 500                                                  rq_list);
 501                                 list_del(&req->rq_list);
 502                                 ptlrpc_server_free_request(req);
 503                         }
 504
 505                         spin_lock(&svc->srv_lock);
 506                         /*
 507                          * now all reqs including the embedded req has been
 508                          * disposed, schedule request buffer for re-use.
 509                          */
 510                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
 511                         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 512                 }
 513
 514                 spin_unlock(&svc->srv_lock);
 515         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 516                  /* If we are low on memory, we are not interested in history */
 517                 list_del(&req->rq_list);
 518                 list_del_init(&req->rq_history_list);
 519                 spin_unlock(&svc->srv_lock);
 520
 521                 ptlrpc_server_free_request(req);
 522         } else {
 523                 spin_unlock(&svc->srv_lock);
 524         }
 525 }
 526
 527 /**
 528  * to finish a request: stop sending more early replies, and release
 529  * the request. should be called after we finished handling the request.
 530  */
 531 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
 532 {
 533         ptlrpc_server_drop_request(req);
 534 }
 535
 536 /* This function makes sure dead exports are evicted in a timely manner.
 537    This function is only called when some export receives a message (i.e.,
 538    the network is up.) */
 539 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 540 {
 541         struct obd_export *oldest_exp;
 542         time_t oldest_time, new_time;
 543
 544         ENTRY;
 545
 546         LASSERT(exp);
 547
 548         /* Compensate for slow machines, etc, by faking our request time
 549            into the future.  Although this can break the strict time-ordering
 550            of the list, we can be really lazy here - we don't have to evict
 551            at the exact right moment.  Eventually, all silent exports
 552            will make it to the top of the list. */
 553
 554         /* Do not pay attention on 1sec or smaller renewals. */
 555         new_time = cfs_time_current_sec() + extra_delay;
 556         if (exp->exp_last_request_time + 1 /*second */ >= new_time)
 557                 RETURN_EXIT;
 558
 559         exp->exp_last_request_time = new_time;
 560         CDEBUG(D_INFO, "updating export %s at %ld\n",
 561                exp->exp_client_uuid.uuid,
 562                exp->exp_last_request_time);
 563
 564         /* exports may get disconnected from the chain even though the
 565            export has references, so we must keep the spin lock while
 566            manipulating the lists */
 567         spin_lock(&exp->exp_obd->obd_dev_lock);
 568
 569         if (list_empty(&exp->exp_obd_chain_timed)) {
 570                 /* this one is not timed */
 571                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 572                 RETURN_EXIT;
 573         }
 574
 575         list_move_tail(&exp->exp_obd_chain_timed,
 576                        &exp->exp_obd->obd_exports_timed);
 577
 578         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 579                                 struct obd_export, exp_obd_chain_timed);
 580         oldest_time = oldest_exp->exp_last_request_time;
 581         spin_unlock(&exp->exp_obd->obd_dev_lock);
 582
 583         if (exp->exp_obd->obd_recovering) {
 584                 /* be nice to everyone during recovery */
 585                 EXIT;
 586                 return;
 587         }
 588
 589         /* Note - racing to start/reset the obd_eviction timer is safe */
 590         if (exp->exp_obd->obd_eviction_timer == 0) {
 591                 /* Check if the oldest entry is expired. */
 592                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
 593                                               extra_delay)) {
 594                         /* We need a second timer, in case the net was down and
 595                          * it just came back. Since the pinger may skip every
 596                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 597                          * we better wait for 3. */
 598                         exp->exp_obd->obd_eviction_timer =
 599                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
 600                         CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
 601                                exp->exp_obd->obd_name,
 602                                obd_export_nid2str(oldest_exp), oldest_time);
 603                 }
 604         } else {
 605                 if (cfs_time_current_sec() >
 606                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
 607                         /* The evictor won't evict anyone who we've heard from
 608                          * recently, so we don't have to check before we start
 609                          * it. */
 610                         if (!ping_evictor_wake(exp))
 611                                 exp->exp_obd->obd_eviction_timer = 0;
 612                 }
 613         }
 614
 615         EXIT;
 616 }
 617
 618 static int ptlrpc_check_req(struct ptlrpc_request *req)
 619 {
 620         if (lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 621             req->rq_export->exp_conn_cnt) {
 622                 DEBUG_REQ(D_ERROR, req,
 623                           "DROPPING req from old connection %d < %d",
 624                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 625                           req->rq_export->exp_conn_cnt);
 626                 return -EEXIST;
 627         }
 628         if (req->rq_export->exp_obd && req->rq_export->exp_obd->obd_fail) {
 629              /* Failing over, don't handle any more reqs, send
 630                 error response instead. */
 631                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 632                        req, req->rq_export->exp_obd->obd_name);
 633                 req->rq_status = -ENODEV;
 634                 ptlrpc_error(req);
 635                 return -ENODEV;
 636         }
 637         return 0;
 638 }
 639
 640 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
 641 {
 642         struct ptlrpc_at_array *array = &svc->srv_at_array;
 643         __s32 next;
 644
 645         spin_lock(&svc->srv_at_lock);
 646         if (array->paa_count == 0) {
 647                 cfs_timer_disarm(&svc->srv_at_timer);
 648                 spin_unlock(&svc->srv_at_lock);
 649                 return;
 650         }
 651
 652         /* Set timer for closest deadline */
 653         next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
 654                        at_early_margin);
 655         if (next <= 0)
 656                 ptlrpc_at_timer((unsigned long)svc);
 657         else
 658                 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
 659         spin_unlock(&svc->srv_at_lock);
 660         CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
 661 }
 662
 663 /* Add rpc to early reply check list */
 664 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 665 {
 666         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 667         struct ptlrpc_request *rq = NULL;
 668         struct ptlrpc_at_array *array = &svc->srv_at_array;
 669         __u32 index;
 670         int found = 0;
 671
 672         if (AT_OFF)
 673                 return(0);
 674
 675         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 676                 return(-ENOSYS);
 677
 678         spin_lock(&svc->srv_at_lock);
 679         LASSERT(list_empty(&req->rq_timed_list));
 680
 681         index = (unsigned long)req->rq_deadline % array->paa_size;
 682         if (array->paa_reqs_count[index] > 0) {
 683                 /* latest rpcs will have the latest deadlines in the list,
 684                  * so search backward. */
 685                 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
 686                                             rq_timed_list) {
 687                         if (req->rq_deadline >= rq->rq_deadline) {
 688                                 list_add(&req->rq_timed_list,
 689                                          &rq->rq_timed_list);
 690                                 break;
 691                         }
 692                 }
 693         }
 694
 695         /* Add the request at the head of the list */
 696         if (list_empty(&req->rq_timed_list))
 697                 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
 698
 699         req->rq_at_linked = 1;
 700         req->rq_at_index = index;
 701         array->paa_reqs_count[index]++;
 702         array->paa_count++;
 703         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
 704                 array->paa_deadline = req->rq_deadline;
 705                 found = 1;
 706         }
 707         spin_unlock(&svc->srv_at_lock);
 708
 709         if (found)
 710                 ptlrpc_at_set_timer(svc);
 711
 712         return 0;
 713 }
 714
 715 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
 716 {
 717         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 718         struct ptlrpc_request *reqcopy;
 719         struct lustre_msg *reqmsg;
 720         long olddl = req->rq_deadline - cfs_time_current_sec();
 721         time_t newdl;
 722         int rc;
 723         ENTRY;
 724
 725         /* deadline is when the client expects us to reply, margin is the
 726            difference between clients' and servers' expectations */
 727         DEBUG_REQ(D_ADAPTTO, req,
 728                   "%ssending early reply (deadline %+lds, margin %+lds) for "
 729                   "%d+%d", AT_OFF ? "AT off - not " : "",
 730                   olddl, olddl - at_get(&svc->srv_at_estimate),
 731                   at_get(&svc->srv_at_estimate), at_extra);
 732
 733         if (AT_OFF)
 734                 RETURN(0);
 735
 736         if (olddl < 0) {
 737                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
 738                           "not sending early reply. Consider increasing "
 739                           "at_early_margin (%d)?", olddl, at_early_margin);
 740
 741                 /* Return an error so we're not re-added to the timed list. */
 742                 RETURN(-ETIMEDOUT);
 743         }
 744
 745         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
 746                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
 747                           "but no AT support");
 748                 RETURN(-ENOSYS);
 749         }
 750
 751         if (req->rq_export &&
 752             lustre_msg_get_flags(req->rq_reqmsg) &
 753             (MSG_REPLAY | MSG_LAST_REPLAY)) {
 754                 /* Use at_extra as early reply period for recovery requests
 755                  * but keep it not longer than recovery time / 4 */
 756                 at_add(&svc->srv_at_estimate,
 757                        min(at_extra,
 758                            req->rq_export->exp_obd->obd_recovery_timeout / 4));
 759                 newdl = cfs_time_current_sec();
 760         } else {
 761                 /* Fake our processing time into the future to ask the
 762                  * clients for some extra amount of time */
 763                 at_add(&svc->srv_at_estimate, at_extra);
 764                 newdl = req->rq_arrival_time.tv_sec;
 765         }
 766         newdl += at_get(&svc->srv_at_estimate);
 767         if (req->rq_deadline >= newdl) {
 768                 /* We're not adding any time, no need to send an early reply
 769                    (e.g. maybe at adaptive_max) */
 770                 DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
 771                           "(%ld/%ld), not sending early reply\n",
 772                           olddl, newdl - cfs_time_current_sec());
 773                 RETURN(-ETIMEDOUT);
 774         }
 775
 776         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 777         if (reqcopy == NULL)
 778                 RETURN(-ENOMEM);
 779         OBD_ALLOC(reqmsg, req->rq_reqlen);
 780         if (!reqmsg) {
 781                 OBD_FREE(reqcopy, sizeof *reqcopy);
 782                 RETURN(-ENOMEM);
 783         }
 784
 785         *reqcopy = *req;
 786         reqcopy->rq_reply_state = NULL;
 787         reqcopy->rq_rep_swab_mask = 0;
 788         /* We only need the reqmsg for the magic */
 789         reqcopy->rq_reqmsg = reqmsg;
 790         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 791
 792         LASSERT(atomic_read(&req->rq_refcount));
 793         /** if it is last refcount then early reply isn't needed */
 794         if (atomic_read(&req->rq_refcount) == 1) {
 795                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
 796                           "abort sending early reply\n");
 797                 GOTO(out, rc = -EINVAL);
 798         }
 799
 800         /* Connection ref */
 801         reqcopy->rq_export = class_conn2export(
 802                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 803         if (reqcopy->rq_export == NULL)
 804                 GOTO(out, rc = -ENODEV);
 805
 806         /* RPC ref */
 807         class_export_rpc_get(reqcopy->rq_export);
 808         if (reqcopy->rq_export->exp_obd &&
 809             reqcopy->rq_export->exp_obd->obd_fail)
 810                 GOTO(out_put, rc = -ENODEV);
 811
 812         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 813         if (rc)
 814                 GOTO(out_put, rc);
 815
 816         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 817
 818         if (!rc) {
 819                 /* Adjust our own deadline to what we told the client */
 820                 req->rq_deadline = newdl;
 821                 req->rq_early_count++; /* number sent, server side */
 822         } else {
 823                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 824         }
 825
 826         /* Free the (early) reply state from lustre_pack_reply.
 827            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 828         ptlrpc_req_drop_rs(reqcopy);
 829
 830 out_put:
 831         class_export_rpc_put(reqcopy->rq_export);
 832         class_export_put(reqcopy->rq_export);
 833 out:
 834         OBD_FREE(reqmsg, req->rq_reqlen);
 835         OBD_FREE(reqcopy, sizeof *reqcopy);
 836         RETURN(rc);
 837 }
 838
 839 /* Send early replies to everybody expiring within at_early_margin
 840    asking for at_extra time */
 841 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 842 {
 843         struct ptlrpc_request *rq, *n;
 844         struct list_head work_list;
 845         struct ptlrpc_at_array *array = &svc->srv_at_array;
 846         __u32  index, count;
 847         time_t deadline;
 848         time_t now = cfs_time_current_sec();
 849         cfs_duration_t delay;
 850         int first, counter = 0;
 851         ENTRY;
 852
 853         spin_lock(&svc->srv_at_lock);
 854         if (svc->srv_at_check == 0) {
 855                 spin_unlock(&svc->srv_at_lock);
 856                 RETURN(0);
 857         }
 858         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 859         svc->srv_at_check = 0;
 860
 861         if (array->paa_count == 0) {
 862                 spin_unlock(&svc->srv_at_lock);
 863                 RETURN(0);
 864         }
 865
 866         /* The timer went off, but maybe the nearest rpc already completed. */
 867         first = array->paa_deadline - now;
 868         if (first > at_early_margin) {
 869                 /* We've still got plenty of time.  Reset the timer. */
 870                 spin_unlock(&svc->srv_at_lock);
 871                 ptlrpc_at_set_timer(svc);
 872                 RETURN(0);
 873         }
 874
 875         /* We're close to a timeout, and we don't know how much longer the
 876            server will take. Send early replies to everyone expiring soon. */
 877         CFS_INIT_LIST_HEAD(&work_list);
 878         deadline = -1;
 879         index = (unsigned long)array->paa_deadline % array->paa_size;
 880         count = array->paa_count;
 881         while (count > 0) {
 882                 count -= array->paa_reqs_count[index];
 883                 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
 884                                          rq_timed_list) {
 885                         if (rq->rq_deadline <= now + at_early_margin) {
 886                                 list_del_init(&rq->rq_timed_list);
 887                                 /**
 888                                  * ptlrpc_server_drop_request() may drop
 889                                  * refcount to 0 already. Let's check this and
 890                                  * don't add entry to work_list
 891                                  */
 892                                 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
 893                                         list_add(&rq->rq_timed_list, &work_list);
 894                                 counter++;
 895                                 array->paa_reqs_count[index]--;
 896                                 array->paa_count--;
 897                                 rq->rq_at_linked = 0;
 898                                 continue;
 899                         }
 900
 901                         /* update the earliest deadline */
 902                         if (deadline == -1 || rq->rq_deadline < deadline)
 903                                 deadline = rq->rq_deadline;
 904
 905                         break;
 906                 }
 907
 908                 if (++index >= array->paa_size)
 909                         index = 0;
 910         }
 911         array->paa_deadline = deadline;
 912         spin_unlock(&svc->srv_at_lock);
 913
 914         /* we have a new earliest deadline, restart the timer */
 915         ptlrpc_at_set_timer(svc);
 916
 917         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 918                "replies\n", first, at_extra, counter);
 919
 920         if (first < 0) {
 921                 /* We're already past request deadlines before we even get a
 922                    chance to send early replies */
 923                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 924                               "request traffic (cpu-bound).\n",  svc->srv_name);
 925                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 926                       "delay="CFS_DURATION_T"(jiff)\n",
 927                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 928                       at_get(&svc->srv_at_estimate), delay);
 929         }
 930
 931         /* we took additional refcount so entries can't be deleted from list, no
 932          * locking is needed */
 933         while (!list_empty(&work_list)) {
 934                 rq = list_entry(work_list.next, struct ptlrpc_request,
 935                                 rq_timed_list);
 936                 list_del_init(&rq->rq_timed_list);
 937
 938                 if (ptlrpc_at_send_early_reply(rq) == 0)
 939                         ptlrpc_at_add_timed(rq);
 940
 941                 ptlrpc_server_drop_request(rq);
 942         }
 943
 944         RETURN(0);
 945 }
 946
 947 /**
 948  * Put the request to the export list if the request may become
 949  * a high priority one.
 950  */
 951 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
 952                              struct ptlrpc_request *req)
 953 {
 954         int rc;
 955         ENTRY;
 956
 957         if (svc->srv_hpreq_handler) {
 958                 rc = svc->srv_hpreq_handler(req);
 959                 if (rc)
 960                         RETURN(rc);
 961         }
 962         if (req->rq_export && req->rq_ops) {
 963                 spin_lock(&req->rq_export->exp_lock);
 964                 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
 965                 spin_unlock(&req->rq_export->exp_lock);
 966         }
 967
 968         RETURN(0);
 969 }
 970
 971 /** Remove the request from the export list. */
 972 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
 973 {
 974         ENTRY;
 975         if (req->rq_export && req->rq_ops) {
 976                 spin_lock(&req->rq_export->exp_lock);
 977                 list_del_init(&req->rq_exp_list);
 978                 spin_unlock(&req->rq_export->exp_lock);
 979         }
 980         EXIT;
 981 }
 982
 983 /**
 984  * Make the request a high priority one.
 985  *
 986  * All the high priority requests are queued in a separate FIFO
 987  * ptlrpc_service::srv_request_hpq list which is parallel to
 988  * ptlrpc_service::srv_request_queue list but has a higher priority
 989  * for handling.
 990  *
 991  * \see ptlrpc_server_handle_request().
 992  */
 993 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
 994                                         struct ptlrpc_request *req)
 995 {
 996         ENTRY;
 997         LASSERT(svc != NULL);
 998         spin_lock(&req->rq_lock);
 999         if (req->rq_hp == 0) {
1000                 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1001
1002                 /* Add to the high priority queue. */
1003                 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1004                 req->rq_hp = 1;
1005                 if (opc != OBD_PING)
1006                         DEBUG_REQ(D_NET, req, "high priority req");
1007         }
1008         spin_unlock(&req->rq_lock);
1009         EXIT;
1010 }
1011
1012 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1013 {
1014         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1015         ENTRY;
1016
1017         spin_lock(&svc->srv_lock);
1018         /* It may happen that the request is already taken for the processing
1019          * but still in the export list, do not re-add it into the HP list. */
1020         if (req->rq_phase == RQ_PHASE_NEW)
1021                 ptlrpc_hpreq_reorder_nolock(svc, req);
1022         spin_unlock(&svc->srv_lock);
1023         EXIT;
1024 }
1025
1026 /** Check if the request if a high priority one. */
1027 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1028 {
1029         int opc, rc = 0;
1030         ENTRY;
1031
1032         /* Check by request opc. */
1033         opc = lustre_msg_get_opc(req->rq_reqmsg);
1034         if (opc == OBD_PING)
1035                 RETURN(1);
1036
1037         /* Perform request specific check. */
1038         if (req->rq_ops && req->rq_ops->hpreq_check)
1039                 rc = req->rq_ops->hpreq_check(req);
1040         RETURN(rc);
1041 }
1042
1043 /** Check if a request is a high priority one. */
1044 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1045                                      struct ptlrpc_request *req)
1046 {
1047         int rc;
1048         ENTRY;
1049
1050         rc = ptlrpc_server_hpreq_check(req);
1051         if (rc < 0)
1052                 RETURN(rc);
1053
1054         spin_lock(&svc->srv_lock);
1055         /* Before inserting the request into the queue, check if it is not
1056          * inserted yet, or even already handled -- it may happen due to
1057          * a racing ldlm_server_blocking_ast(). */
1058         if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1059                 if (rc)
1060                         ptlrpc_hpreq_reorder_nolock(svc, req);
1061                 else
1062                         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1063         }
1064         spin_unlock(&svc->srv_lock);
1065
1066         RETURN(0);
1067 }
1068
1069 /* Only allow normal priority requests on a service that has a high-priority
1070  * queue if forced (i.e. cleanup), if there are other high priority requests
1071  * already being processed (i.e. those threads can service more high-priority
1072  * requests), or if there are enough idle threads that a later thread can do
1073  * a high priority request. */
1074 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1075 {
1076         return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1077                svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1078 }
1079
1080 static struct ptlrpc_request *
1081 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1082 {
1083         struct ptlrpc_request *req = NULL;
1084         ENTRY;
1085
1086         if (ptlrpc_server_allow_normal(svc, force) &&
1087             !list_empty(&svc->srv_request_queue) &&
1088             (list_empty(&svc->srv_request_hpq) ||
1089              svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1090                 req = list_entry(svc->srv_request_queue.next,
1091                                  struct ptlrpc_request, rq_list);
1092                 svc->srv_hpreq_count = 0;
1093         } else if (!list_empty(&svc->srv_request_hpq)) {
1094                 req = list_entry(svc->srv_request_hpq.next,
1095                                  struct ptlrpc_request, rq_list);
1096                 svc->srv_hpreq_count++;
1097         }
1098         RETURN(req);
1099 }
1100
1101 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1102 {
1103         return ((ptlrpc_server_allow_normal(svc, force) &&
1104                  !list_empty(&svc->srv_request_queue)) ||
1105                 !list_empty(&svc->srv_request_hpq));
1106 }
1107
1108 /* Handle freshly incoming reqs, add to timed early reply list,
1109    pass on to regular request queue */
1110 static int
1111 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1112 {
1113         struct ptlrpc_request *req;
1114         __u32                  deadline;
1115         int                    rc;
1116         ENTRY;
1117
1118         LASSERT(svc);
1119
1120         spin_lock(&svc->srv_lock);
1121         if (list_empty(&svc->srv_req_in_queue)) {
1122                 spin_unlock(&svc->srv_lock);
1123                 RETURN(0);
1124         }
1125
1126         req = list_entry(svc->srv_req_in_queue.next,
1127                          struct ptlrpc_request, rq_list);
1128         list_del_init (&req->rq_list);
1129         /* Consider this still a "queued" request as far as stats are
1130            concerned */
1131         spin_unlock(&svc->srv_lock);
1132
1133         /* Clear request swab mask; this is a new request */
1134         req->rq_req_swab_mask = 0;
1135
1136         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1137         if (rc < 0) {
1138                 CERROR ("error unpacking request: ptl %d from %s"
1139                         " xid "LPU64"\n", svc->srv_req_portal,
1140                         libcfs_id2str(req->rq_peer), req->rq_xid);
1141                 goto err_req;
1142         }
1143
1144         if (rc > 0)
1145                 lustre_set_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1146
1147         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1148         if (rc) {
1149                 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
1150                         " xid "LPU64"\n", svc->srv_req_portal,
1151                         libcfs_id2str(req->rq_peer), req->rq_xid);
1152                 goto err_req;
1153         }
1154
1155         rc = -EINVAL;
1156         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1157                 CERROR("wrong packet type received (type=%u) from %s\n",
1158                        lustre_msg_get_type(req->rq_reqmsg),
1159                        libcfs_id2str(req->rq_peer));
1160                 goto err_req;
1161         }
1162
1163         CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1164
1165         req->rq_export = class_conn2export(
1166                 lustre_msg_get_handle(req->rq_reqmsg));
1167         if (req->rq_export) {
1168                 rc = ptlrpc_check_req(req);
1169                 if (rc)
1170                         goto err_req;
1171                 ptlrpc_update_export_timer(req->rq_export, 0);
1172         }
1173
1174         /* req_in handling should/must be fast */
1175         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1176                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus",
1177                           cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
1178
1179         /* Set rpc server deadline and add it to the timed list */
1180         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1181                     MSGHDR_AT_SUPPORT) ?
1182                    /* The max time the client expects us to take */
1183                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1184         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1185         if (unlikely(deadline == 0)) {
1186                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1187                 goto err_req;
1188         }
1189
1190         ptlrpc_at_add_timed(req);
1191         rc = ptlrpc_hpreq_init(svc, req);
1192         if (rc)
1193                 GOTO(err_req, rc);
1194
1195         /* Move it over to the request processing queue */
1196         rc = ptlrpc_server_request_add(svc, req);
1197         if (rc)
1198                 GOTO(err_req, rc);
1199         cfs_waitq_signal(&svc->srv_waitq);
1200         RETURN(1);
1201
1202 err_req:
1203         spin_lock(&svc->srv_lock);
1204         svc->srv_n_queued_reqs--;
1205         svc->srv_n_active_reqs++;
1206         spin_unlock(&svc->srv_lock);
1207         ptlrpc_server_finish_request(req);
1208
1209         RETURN(1);
1210 }
1211
1212 static int
1213 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1214                              struct ptlrpc_thread *thread)
1215 {
1216         struct obd_export     *export = NULL;
1217         struct ptlrpc_request *request;
1218         struct timeval         work_start;
1219         struct timeval         work_end;
1220         long                   timediff;
1221         int                    opc, rc;
1222         int                    fail_opc = 0;
1223         ENTRY;
1224
1225         LASSERT(svc);
1226
1227         spin_lock(&svc->srv_lock);
1228         if (!ptlrpc_server_request_pending(svc, 0) ||
1229             (
1230 #ifndef __KERNEL__
1231              /* !@%$# liblustre only has 1 thread */
1232              svc->srv_n_difficult_replies != 0 &&
1233 #endif
1234              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1))) {
1235                 /* Don't handle regular requests in the last thread, in order
1236                  * to handle difficult replies (which might block other threads)
1237                  * as well as handle any incoming reqs, early replies, etc.
1238                  * That means we always need at least 2 service threads. */
1239                 spin_unlock(&svc->srv_lock);
1240                 RETURN(0);
1241         }
1242
1243         request = ptlrpc_server_request_get(svc, 0);
1244         if  (request == NULL) {
1245                 spin_unlock(&svc->srv_lock);
1246                 RETURN(0);
1247         }
1248
1249         opc = lustre_msg_get_opc(request->rq_reqmsg);
1250         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1251                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1252         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1253                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1254
1255         if (unlikely(fail_opc)) {
1256                 if (request->rq_export && request->rq_ops) {
1257                         spin_unlock(&svc->srv_lock);
1258                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1259                         spin_lock(&svc->srv_lock);
1260                         request = ptlrpc_server_request_get(svc, 0);
1261                         if  (request == NULL) {
1262                                 spin_unlock(&svc->srv_lock);
1263                                 RETURN(0);
1264                         }
1265                         LASSERT(ptlrpc_server_request_pending(svc, 0));
1266                 }
1267         }
1268
1269         list_del_init(&request->rq_list);
1270         svc->srv_n_queued_reqs--;
1271         svc->srv_n_active_reqs++;
1272
1273         if (request->rq_hp)
1274                 svc->srv_n_hpreq++;
1275
1276         /* The phase is changed under the lock here because we need to know
1277          * the request is under processing (see ptlrpc_hpreq_reorder()). */
1278         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1279         spin_unlock(&svc->srv_lock);
1280
1281         ptlrpc_hpreq_fini(request);
1282
1283         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1284                 libcfs_debug_dumplog();
1285
1286         do_gettimeofday(&work_start);
1287         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1288         if (svc->srv_stats != NULL) {
1289                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1290                                     timediff);
1291                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1292                                     svc->srv_n_queued_reqs);
1293                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1294                                     svc->srv_n_active_reqs);
1295                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1296                                     at_get(&svc->srv_at_estimate));
1297         }
1298
1299         CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1300
1301         request->rq_svc_thread = thread;
1302         if (request->rq_export) {
1303                 if (ptlrpc_check_req(request))
1304                         goto put_conn;
1305                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1306                 export = class_export_rpc_get(request->rq_export);
1307         }
1308
1309         /* Discard requests queued for longer than the deadline.
1310            The deadline is increased if we send an early reply. */
1311         if (cfs_time_current_sec() > request->rq_deadline) {
1312                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1313                           ": deadline %ld%+lds ago\n",
1314                           libcfs_id2str(request->rq_peer),
1315                           request->rq_deadline -
1316                           request->rq_arrival_time.tv_sec,
1317                           cfs_time_current_sec() - request->rq_deadline);
1318                 goto put_rpc_export;
1319         }
1320
1321         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1322                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1323                (request->rq_export ?
1324                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1325                (request->rq_export ?
1326                 atomic_read(&request->rq_export->exp_refcount) : -99),
1327                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1328                libcfs_id2str(request->rq_peer),
1329                lustre_msg_get_opc(request->rq_reqmsg));
1330
1331         OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1332
1333         rc = svc->srv_handler(request);
1334
1335         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1336
1337 put_rpc_export:
1338         if (export != NULL && !request->rq_copy_queued)
1339                 class_export_rpc_put(export);
1340
1341 put_conn:
1342         if (cfs_time_current_sec() > request->rq_deadline) {
1343                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1344                           "than estimated (%ld%+lds); client may timeout.",
1345                           request->rq_xid, request->rq_deadline -
1346                           request->rq_arrival_time.tv_sec,
1347                           cfs_time_current_sec() - request->rq_deadline);
1348         }
1349
1350         do_gettimeofday(&work_end);
1351         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1352
1353         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1354                "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1355                "%lds (%lds total) trans "LPU64" rc %d/%d\n",
1356                cfs_curproc_comm(),
1357                (request->rq_export ?
1358                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1359                (request->rq_export ?
1360                 atomic_read(&request->rq_export->exp_refcount) : -99),
1361                lustre_msg_get_status(request->rq_reqmsg),
1362                request->rq_xid,
1363                libcfs_id2str(request->rq_peer),
1364                lustre_msg_get_opc(request->rq_reqmsg),
1365                timediff,
1366                cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1367                (request->rq_repmsg ?
1368                 lustre_msg_get_transno(request->rq_repmsg) :
1369                 request->rq_transno),
1370                request->rq_status,
1371                (request->rq_repmsg ?
1372                 lustre_msg_get_status(request->rq_repmsg) : -999));
1373
1374         if (svc->srv_stats != NULL) {
1375                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1376                 int opc = opcode_offset(op);
1377                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1378                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1379                         lprocfs_counter_add(svc->srv_stats,
1380                                             opc + EXTRA_MAX_OPCODES,
1381                                             timediff);
1382                 }
1383         }
1384         if (request->rq_early_count) {
1385                 DEBUG_REQ(D_ADAPTTO, request,
1386                           "sent %d early replies before finishing in %lds",
1387                           request->rq_early_count,
1388                           work_end.tv_sec - request->rq_arrival_time.tv_sec);
1389         }
1390
1391         spin_lock(&svc->srv_lock);
1392         if (request->rq_hp)
1393                 svc->srv_n_hpreq--;
1394         spin_unlock(&svc->srv_lock);
1395         ptlrpc_server_finish_request(request);
1396
1397         RETURN(1);
1398 }
1399
1400 static int
1401 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1402 {
1403         struct ptlrpc_reply_state *rs;
1404         struct obd_export         *exp;
1405         struct obd_device         *obd;
1406         int                        nlocks;
1407         int                        been_handled;
1408         ENTRY;
1409
1410         spin_lock(&svc->srv_lock);
1411         if (list_empty (&svc->srv_reply_queue)) {
1412                 spin_unlock(&svc->srv_lock);
1413                 RETURN(0);
1414         }
1415
1416         rs = list_entry (svc->srv_reply_queue.next,
1417                          struct ptlrpc_reply_state, rs_list);
1418
1419         exp = rs->rs_export;
1420         obd = exp->exp_obd;
1421
1422         LASSERT (rs->rs_difficult);
1423         LASSERT (rs->rs_scheduled);
1424
1425         list_del_init (&rs->rs_list);
1426
1427         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1428         spin_unlock(&svc->srv_lock);
1429
1430         spin_lock (&exp->exp_uncommitted_replies_lock);
1431         /* Noop if removed already */
1432         list_del_init (&rs->rs_obd_list);
1433         spin_unlock (&exp->exp_uncommitted_replies_lock);
1434
1435         spin_lock (&exp->exp_lock);
1436         /* Noop if removed already */
1437         list_del_init (&rs->rs_exp_list);
1438         spin_unlock (&exp->exp_lock);
1439
1440         spin_lock(&svc->srv_lock);
1441
1442         been_handled = rs->rs_handled;
1443         rs->rs_handled = 1;
1444
1445         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1446         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1447
1448         if (nlocks == 0 && !been_handled) {
1449                 /* If we see this, we should already have seen the warning
1450                  * in mds_steal_ack_locks()  */
1451                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1452                       " o%d NID %s\n", rs, rs->rs_xid, rs->rs_transno,
1453                       lustre_msg_get_opc(rs->rs_msg),
1454                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1455         }
1456
1457         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1458                 spin_unlock(&svc->srv_lock);
1459
1460                 if (!been_handled && rs->rs_on_net) {
1461                         LNetMDUnlink(rs->rs_md_h);
1462                         /* Ignore return code; we're racing with
1463                          * completion... */
1464                 }
1465
1466                 while (nlocks-- > 0)
1467                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1468                                          rs->rs_modes[nlocks]);
1469
1470                 spin_lock(&svc->srv_lock);
1471         }
1472
1473         rs->rs_scheduled = 0;
1474
1475         if (!rs->rs_on_net) {
1476                 /* Off the net */
1477                 svc->srv_n_difficult_replies--;
1478                 spin_unlock(&svc->srv_lock);
1479
1480                 class_export_put (exp);
1481                 rs->rs_export = NULL;
1482                 ptlrpc_rs_decref (rs);
1483                 atomic_dec (&svc->srv_outstanding_replies);
1484                 RETURN(1);
1485         }
1486
1487         /* still on the net; callback will schedule */
1488         spin_unlock(&svc->srv_lock);
1489         RETURN(1);
1490 }
1491
1492 #ifndef __KERNEL__
1493 /* FIXME make use of timeout later */
1494 int
1495 liblustre_check_services (void *arg)
1496 {
1497         int  did_something = 0;
1498         int  rc;
1499         struct list_head *tmp, *nxt;
1500         ENTRY;
1501
1502         /* I'm relying on being single threaded, not to have to lock
1503          * ptlrpc_all_services etc */
1504         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1505                 struct ptlrpc_service *svc =
1506                         list_entry (tmp, struct ptlrpc_service, srv_list);
1507
1508                 if (svc->srv_threads_running != 0)     /* I've recursed */
1509                         continue;
1510
1511                 /* service threads can block for bulk, so this limits us
1512                  * (arbitrarily) to recursing 1 stack frame per service.
1513                  * Note that the problem with recursion is that we have to
1514                  * unwind completely before our caller can resume. */
1515
1516                 svc->srv_threads_running++;
1517
1518                 do {
1519                         rc = ptlrpc_server_handle_req_in(svc);
1520                         rc |= ptlrpc_server_handle_reply(svc);
1521                         rc |= ptlrpc_at_check_timed(svc);
1522                         rc |= ptlrpc_server_handle_request(svc, NULL);
1523                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1524                         did_something |= rc;
1525                 } while (rc);
1526
1527                 svc->srv_threads_running--;
1528         }
1529
1530         RETURN(did_something);
1531 }
1532 #define ptlrpc_stop_all_threads(s) do {} while (0)
1533
1534 #else /* __KERNEL__ */
1535
1536 static void
1537 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1538 {
1539         int avail = svc->srv_nrqbd_receiving;
1540         int low_water = test_req_buffer_pressure ? 0 :
1541                         svc->srv_nbuf_per_group/2;
1542
1543         /* NB I'm not locking; just looking. */
1544
1545         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1546          * allowed the request history to grow out of control.  We could put a
1547          * sanity check on that here and cull some history if we need the
1548          * space. */
1549
1550         if (avail <= low_water)
1551                 ptlrpc_grow_req_bufs(svc);
1552
1553         if (svc->srv_stats)
1554                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1555                                     avail);
1556 }
1557
1558 static int
1559 ptlrpc_retry_rqbds(void *arg)
1560 {
1561         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1562
1563         svc->srv_rqbd_timeout = 0;
1564         return (-ETIMEDOUT);
1565 }
1566
1567 static int ptlrpc_main(void *arg)
1568 {
1569         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1570         struct ptlrpc_service  *svc = data->svc;
1571         struct ptlrpc_thread   *thread = data->thread;
1572         struct obd_device      *dev = data->dev;
1573         struct ptlrpc_reply_state *rs;
1574 #ifdef WITH_GROUP_INFO
1575         struct group_info *ginfo = NULL;
1576 #endif
1577         int counter = 0, rc = 0;
1578         ENTRY;
1579
1580         cfs_daemonize_ctxt(data->name);
1581
1582 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1583         /* we need to do this before any per-thread allocation is done so that
1584          * we get the per-thread allocations on local node.  bug 7342 */
1585         if (svc->srv_cpu_affinity) {
1586                 int cpu, num_cpu;
1587
1588                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1589                         if (!cpu_online(cpu))
1590                                 continue;
1591                         if (num_cpu == thread->t_id % num_online_cpus())
1592                                 break;
1593                         num_cpu++;
1594                 }
1595                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1596         }
1597 #endif
1598
1599 #ifdef WITH_GROUP_INFO
1600         ginfo = groups_alloc(0);
1601         if (!ginfo) {
1602                 rc = -ENOMEM;
1603                 goto out;
1604         }
1605
1606         set_current_groups(ginfo);
1607         put_group_info(ginfo);
1608 #endif
1609
1610         if (svc->srv_init != NULL) {
1611                 rc = svc->srv_init(thread);
1612                 if (rc)
1613                         goto out;
1614         }
1615
1616         /* Alloc reply state structure for this one */
1617         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1618         if (!rs) {
1619                 rc = -ENOMEM;
1620                 goto out_srv_init;
1621         }
1622
1623         spin_lock(&svc->srv_lock);
1624         /* SVC_STOPPING may already be set here if someone else is trying
1625          * to stop the service while this new thread has been dynamically
1626          * forked. We still set SVC_RUNNING to let our creator know that
1627          * we are now running, however we will exit as soon as possible */
1628         thread->t_flags |= SVC_RUNNING;
1629         spin_unlock(&svc->srv_lock);
1630
1631         /*
1632          * wake up our creator. Note: @data is invalid after this point,
1633          * because it's allocated on ptlrpc_start_thread() stack.
1634          */
1635         cfs_waitq_signal(&thread->t_ctl_waitq);
1636
1637         thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL);
1638
1639         spin_lock(&svc->srv_lock);
1640         svc->srv_threads_running++;
1641         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1642         spin_unlock(&svc->srv_lock);
1643         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1644
1645         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1646                svc->srv_threads_running);
1647
1648         /* XXX maintain a list of all managed devices: insert here */
1649
1650         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1651                svc->srv_n_difficult_replies != 0) {
1652                 /* Don't exit while there are replies to be handled */
1653                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1654                                                      ptlrpc_retry_rqbds, svc);
1655
1656                 lc_watchdog_disable(thread->t_watchdog);
1657
1658                 cfs_cond_resched();
1659
1660                 l_wait_event_exclusive (svc->srv_waitq,
1661                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1662                                svc->srv_n_difficult_replies == 0) ||
1663                               (!list_empty(&svc->srv_idle_rqbds) &&
1664                                svc->srv_rqbd_timeout == 0) ||
1665                               !list_empty(&svc->srv_req_in_queue) ||
1666                               !list_empty(&svc->srv_reply_queue) ||
1667                               (ptlrpc_server_request_pending(svc, 0) &&
1668                                (svc->srv_n_active_reqs <
1669                                 (svc->srv_threads_running - 1))) ||
1670                               svc->srv_at_check,
1671                               &lwi);
1672
1673                 lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc));
1674
1675                 ptlrpc_check_rqbd_pool(svc);
1676
1677                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1678                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1679                         /* Ignore return code - we tried... */
1680                         ptlrpc_start_thread(dev, svc);
1681                 }
1682
1683                 if (!list_empty(&svc->srv_reply_queue))
1684                         ptlrpc_server_handle_reply (svc);
1685
1686                 if (!list_empty(&svc->srv_req_in_queue)) {
1687                         /* Process all incoming reqs before handling any */
1688                         ptlrpc_server_handle_req_in(svc);
1689                         /* but limit ourselves in case of flood */
1690                         if (counter++ < 1000)
1691                                 continue;
1692                         counter = 0;
1693                 }
1694
1695                 if (svc->srv_at_check)
1696                         ptlrpc_at_check_timed(svc);
1697
1698                 /* don't handle requests in the last thread */
1699                 if (ptlrpc_server_request_pending(svc, 0) &&
1700                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1)))
1701                         ptlrpc_server_handle_request(svc, thread);
1702
1703                 if (!list_empty(&svc->srv_idle_rqbds) &&
1704                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1705                         /* I just failed to repost request buffers.  Wait
1706                          * for a timeout (unless something else happens)
1707                          * before I try again */
1708                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1709                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1710                                svc->srv_nrqbd_receiving);
1711                 }
1712         }
1713
1714         lc_watchdog_delete(thread->t_watchdog);
1715         thread->t_watchdog = NULL;
1716
1717 out_srv_init:
1718         /*
1719          * deconstruct service specific state created by ptlrpc_start_thread()
1720          */
1721         if (svc->srv_done != NULL)
1722                 svc->srv_done(thread);
1723
1724 out:
1725         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1726
1727         spin_lock(&svc->srv_lock);
1728         svc->srv_threads_running--;              /* must know immediately */
1729         thread->t_id = rc;
1730         thread->t_flags = SVC_STOPPED;
1731
1732         cfs_waitq_signal(&thread->t_ctl_waitq);
1733         spin_unlock(&svc->srv_lock);
1734
1735         return rc;
1736 }
1737
1738 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1739                                struct ptlrpc_thread *thread)
1740 {
1741         struct l_wait_info lwi = { 0 };
1742
1743         spin_lock(&svc->srv_lock);
1744         /* let the thread know that we would like it to stop asap */
1745         thread->t_flags |= SVC_STOPPING;
1746         spin_unlock(&svc->srv_lock);
1747
1748         cfs_waitq_broadcast(&svc->srv_waitq);
1749         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1750                      &lwi);
1751
1752         spin_lock(&svc->srv_lock);
1753         list_del(&thread->t_link);
1754         spin_unlock(&svc->srv_lock);
1755
1756         OBD_FREE(thread, sizeof(*thread));
1757 }
1758
1759 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1760 {
1761         struct ptlrpc_thread *thread;
1762
1763         spin_lock(&svc->srv_lock);
1764         while (!list_empty(&svc->srv_threads)) {
1765                 thread = list_entry(svc->srv_threads.next,
1766                                     struct ptlrpc_thread, t_link);
1767
1768                 spin_unlock(&svc->srv_lock);
1769                 ptlrpc_stop_thread(svc, thread);
1770                 spin_lock(&svc->srv_lock);
1771         }
1772
1773         spin_unlock(&svc->srv_lock);
1774 }
1775
1776 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1777 {
1778         int i, rc = 0;
1779         ENTRY;
1780
1781         /* We require 2 threads min - see note in
1782          * ptlrpc_server_handle_request() */
1783
1784         LASSERT(svc->srv_threads_min >= 2);
1785         for (i = 0; i < svc->srv_threads_min; i++) {
1786                 rc = ptlrpc_start_thread(dev, svc);
1787                 /* We have enough threads, don't start more.  b=15759 */
1788                 if (rc == -EMFILE)
1789                         break;
1790                 if (rc) {
1791                         CERROR("cannot start %s thread #%d: rc %d\n",
1792                                svc->srv_thread_name, i, rc);
1793                         ptlrpc_stop_all_threads(svc);
1794                 }
1795         }
1796         RETURN(rc);
1797 }
1798
1799 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1800 {
1801         struct l_wait_info lwi = { 0 };
1802         struct ptlrpc_svc_data d;
1803         struct ptlrpc_thread *thread;
1804         char name[32];
1805         int id, rc;
1806         ENTRY;
1807
1808         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1809                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1810                svc->srv_threads_max, svc->srv_threads_running);
1811         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1812             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1813              svc->srv_threads_started == svc->srv_threads_min - 1))
1814                 RETURN(-EMFILE);
1815
1816         OBD_ALLOC(thread, sizeof(*thread));
1817         if (thread == NULL)
1818                 RETURN(-ENOMEM);
1819         cfs_waitq_init(&thread->t_ctl_waitq);
1820
1821         spin_lock(&svc->srv_lock);
1822         if (svc->srv_threads_started >= svc->srv_threads_max) {
1823                 spin_unlock(&svc->srv_lock);
1824                 OBD_FREE(thread, sizeof(*thread));
1825                 RETURN(-EMFILE);
1826         }
1827         list_add(&thread->t_link, &svc->srv_threads);
1828         id = svc->srv_threads_started++;
1829         spin_unlock(&svc->srv_lock);
1830
1831         thread->t_svc = svc;
1832         thread->t_id = id;
1833         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1834         d.dev = dev;
1835         d.svc = svc;
1836         d.name = name;
1837         d.thread = thread;
1838
1839         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1840
1841         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1842          * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
1843          */
1844         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1845         if (rc < 0) {
1846                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1847
1848                 spin_lock(&svc->srv_lock);
1849                 list_del(&thread->t_link);
1850                 --svc->srv_threads_started;
1851                 spin_unlock(&svc->srv_lock);
1852
1853                 OBD_FREE(thread, sizeof(*thread));
1854                 RETURN(rc);
1855         }
1856         l_wait_event(thread->t_ctl_waitq,
1857                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1858
1859         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1860         RETURN(rc);
1861 }
1862 #endif
1863
1864 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1865 {
1866         int                   rc;
1867         struct l_wait_info    lwi;
1868         struct list_head     *tmp;
1869         struct ptlrpc_reply_state *rs, *t;
1870         struct ptlrpc_at_array *array = &service->srv_at_array;
1871
1872         cfs_timer_disarm(&service->srv_at_timer);
1873
1874         ptlrpc_stop_all_threads(service);
1875         LASSERT(list_empty(&service->srv_threads));
1876
1877         spin_lock (&ptlrpc_all_services_lock);
1878         list_del_init (&service->srv_list);
1879         spin_unlock (&ptlrpc_all_services_lock);
1880
1881         ptlrpc_lprocfs_unregister_service(service);
1882
1883         /* All history will be culled when the next request buffer is
1884          * freed */
1885         service->srv_max_history_rqbds = 0;
1886
1887         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1888
1889         rc = LNetClearLazyPortal(service->srv_req_portal);
1890         LASSERT (rc == 0);
1891
1892         /* Unlink all the request buffers.  This forces a 'final' event with
1893          * its 'unlink' flag set for each posted rqbd */
1894         list_for_each(tmp, &service->srv_active_rqbds) {
1895                 struct ptlrpc_request_buffer_desc *rqbd =
1896                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1897                                    rqbd_list);
1898
1899                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1900                 LASSERT (rc == 0 || rc == -ENOENT);
1901         }
1902
1903         /* Wait for the network to release any buffers it's currently
1904          * filling */
1905         for (;;) {
1906                 spin_lock(&service->srv_lock);
1907                 rc = service->srv_nrqbd_receiving;
1908                 spin_unlock(&service->srv_lock);
1909
1910                 if (rc == 0)
1911                         break;
1912
1913                 /* Network access will complete in finite time but the HUGE
1914                  * timeout lets us CWARN for visibility of sluggish NALs */
1915                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1916                                            cfs_time_seconds(1), NULL, NULL);
1917                 rc = l_wait_event(service->srv_waitq,
1918                                   service->srv_nrqbd_receiving == 0,
1919                                   &lwi);
1920                 if (rc == -ETIMEDOUT)
1921                         CWARN("Service %s waiting for request buffers\n",
1922                               service->srv_name);
1923         }
1924
1925         /* schedule all outstanding replies to terminate them */
1926         spin_lock(&service->srv_lock);
1927         while (!list_empty(&service->srv_active_replies)) {
1928                 struct ptlrpc_reply_state *rs =
1929                         list_entry(service->srv_active_replies.next,
1930                                    struct ptlrpc_reply_state, rs_list);
1931                 ptlrpc_schedule_difficult_reply(rs);
1932         }
1933         spin_unlock(&service->srv_lock);
1934
1935         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1936          * and no service threads, so I'm the only thread noodling the
1937          * request queue now */
1938         while (!list_empty(&service->srv_req_in_queue)) {
1939                 struct ptlrpc_request *req =
1940                         list_entry(service->srv_req_in_queue.next,
1941                                    struct ptlrpc_request,
1942                                    rq_list);
1943
1944                 list_del(&req->rq_list);
1945                 service->srv_n_queued_reqs--;
1946                 service->srv_n_active_reqs++;
1947                 ptlrpc_server_finish_request(req);
1948         }
1949         while (ptlrpc_server_request_pending(service, 1)) {
1950                 struct ptlrpc_request *req;
1951
1952                 req = ptlrpc_server_request_get(service, 1);
1953                 list_del(&req->rq_list);
1954                 service->srv_n_queued_reqs--;
1955                 service->srv_n_active_reqs++;
1956                 ptlrpc_hpreq_fini(req);
1957                 ptlrpc_server_finish_request(req);
1958         }
1959         LASSERT(service->srv_n_queued_reqs == 0);
1960         LASSERT(service->srv_n_active_reqs == 0);
1961         LASSERT(service->srv_n_history_rqbds == 0);
1962         LASSERT(list_empty(&service->srv_active_rqbds));
1963
1964         /* Now free all the request buffers since nothing references them
1965          * any more... */
1966         while (!list_empty(&service->srv_idle_rqbds)) {
1967                 struct ptlrpc_request_buffer_desc *rqbd =
1968                         list_entry(service->srv_idle_rqbds.next,
1969                                    struct ptlrpc_request_buffer_desc,
1970                                    rqbd_list);
1971
1972                 ptlrpc_free_rqbd(rqbd);
1973         }
1974
1975         /* wait for all outstanding replies to complete (they were
1976          * scheduled having been flagged to abort above) */
1977         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1978                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1979
1980                 rc = l_wait_event(service->srv_waitq,
1981                                   !list_empty(&service->srv_reply_queue), &lwi);
1982                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1983
1984                 if (rc == 0) {
1985                         ptlrpc_server_handle_reply(service);
1986                         continue;
1987                 }
1988                 CWARN("Unexpectedly long timeout %p\n", service);
1989         }
1990
1991         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1992                 list_del(&rs->rs_list);
1993                 OBD_FREE(rs, service->srv_max_reply_size);
1994         }
1995
1996         /* In case somebody rearmed this in the meantime */
1997         cfs_timer_disarm(&service->srv_at_timer);
1998
1999         if (array->paa_reqs_array != NULL) {
2000                 OBD_FREE(array->paa_reqs_array,
2001                          sizeof(struct list_head) * array->paa_size);
2002                 array->paa_reqs_array = NULL;
2003         }
2004
2005         if (array->paa_reqs_count != NULL) {
2006                 OBD_FREE(array->paa_reqs_count,
2007                          sizeof(__u32) * array->paa_size);
2008                 array->paa_reqs_count= NULL;
2009         }
2010
2011         OBD_FREE(service, sizeof(*service));
2012         return 0;
2013 }
2014
2015 /* Returns 0 if the service is healthy.
2016  *
2017  * Right now, it just checks to make sure that requests aren't languishing
2018  * in the queue.  We'll use this health check to govern whether a node needs
2019  * to be shot, so it's intentionally non-aggressive. */
2020 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2021 {
2022         struct ptlrpc_request *request;
2023         struct timeval         right_now;
2024         long                   timediff;
2025
2026         if (svc == NULL)
2027                 return 0;
2028
2029         do_gettimeofday(&right_now);
2030
2031         spin_lock(&svc->srv_lock);
2032         if (!ptlrpc_server_request_pending(svc, 1)) {
2033                 spin_unlock(&svc->srv_lock);
2034                 return 0;
2035         }
2036
2037         /* How long has the next entry been waiting? */
2038         if (list_empty(&svc->srv_request_queue))
2039                 request = list_entry(svc->srv_request_hpq.next,
2040                                      struct ptlrpc_request, rq_list);
2041         else
2042                 request = list_entry(svc->srv_request_queue.next,
2043                                      struct ptlrpc_request, rq_list);
2044         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2045         spin_unlock(&svc->srv_lock);
2046
2047         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2048                                         at_max)) {
2049                 CERROR("%s: unhealthy - request has been waiting %lds\n",
2050                        svc->srv_name, timediff / ONE_MILLION);
2051                 return (-1);
2052         }
2053
2054         return 0;
2055 }