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                 /* During recovery, we don't want to send too many early
 755                  * replies, but on the other hand we want to make sure the
 756                  * client has enough time to resend if the rpc is lost. So
 757                  * during the recovery period send at least 4 early replies,
 758                  * spacing them every at_extra if we can. at_estimate should
 759                  * always equal this fixed value during recovery. */
 760                 at_measured(&svc->srv_at_estimate, min(at_extra,
 761                             req->rq_export->exp_obd->obd_recovery_timeout / 4));
 762         } else {
 763                 /* Fake our processing time into the future to ask the
 764                  * clients for some extra amount of time */
 765                 at_measured(&svc->srv_at_estimate, at_extra +
 766                             cfs_time_current_sec() -
 767                             req->rq_arrival_time.tv_sec);
 768
 769                 /* Check to see if we've actually increased the deadline -
 770                  * we may be past adaptive_max */
 771                 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
 772                     at_get(&svc->srv_at_estimate)) {
 773                         DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
 774                                   "(%ld/%ld), not sending early reply\n",
 775                                   olddl, req->rq_arrival_time.tv_sec +
 776                                   at_get(&svc->srv_at_estimate) -
 777                                   cfs_time_current_sec());
 778                         RETURN(-ETIMEDOUT);
 779                 }
 780         }
 781         newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
 782
 783         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 784         if (reqcopy == NULL)
 785                 RETURN(-ENOMEM);
 786         OBD_ALLOC(reqmsg, req->rq_reqlen);
 787         if (!reqmsg) {
 788                 OBD_FREE(reqcopy, sizeof *reqcopy);
 789                 RETURN(-ENOMEM);
 790         }
 791
 792         *reqcopy = *req;
 793         reqcopy->rq_reply_state = NULL;
 794         reqcopy->rq_rep_swab_mask = 0;
 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         LASSERT(atomic_read(&req->rq_refcount));
 800         /** if it is last refcount then early reply isn't needed */
 801         if (atomic_read(&req->rq_refcount) == 1) {
 802                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
 803                           "abort sending early reply\n");
 804                 GOTO(out, rc = -EINVAL);
 805         }
 806
 807         /* Connection ref */
 808         reqcopy->rq_export = class_conn2export(
 809                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 810         if (reqcopy->rq_export == NULL)
 811                 GOTO(out, rc = -ENODEV);
 812
 813         /* RPC ref */
 814         class_export_rpc_get(reqcopy->rq_export);
 815         if (reqcopy->rq_export->exp_obd &&
 816             reqcopy->rq_export->exp_obd->obd_fail)
 817                 GOTO(out_put, rc = -ENODEV);
 818
 819         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 820         if (rc)
 821                 GOTO(out_put, rc);
 822
 823         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 824
 825         if (!rc) {
 826                 /* Adjust our own deadline to what we told the client */
 827                 req->rq_deadline = newdl;
 828                 req->rq_early_count++; /* number sent, server side */
 829         } else {
 830                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 831         }
 832
 833         /* Free the (early) reply state from lustre_pack_reply.
 834            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 835         ptlrpc_req_drop_rs(reqcopy);
 836
 837 out_put:
 838         class_export_rpc_put(reqcopy->rq_export);
 839         class_export_put(reqcopy->rq_export);
 840 out:
 841         OBD_FREE(reqmsg, req->rq_reqlen);
 842         OBD_FREE(reqcopy, sizeof *reqcopy);
 843         RETURN(rc);
 844 }
 845
 846 /* Send early replies to everybody expiring within at_early_margin
 847    asking for at_extra time */
 848 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 849 {
 850         struct ptlrpc_request *rq, *n;
 851         struct list_head work_list;
 852         struct ptlrpc_at_array *array = &svc->srv_at_array;
 853         __u32  index, count;
 854         time_t deadline;
 855         time_t now = cfs_time_current_sec();
 856         cfs_duration_t delay;
 857         int first, counter = 0;
 858         ENTRY;
 859
 860         spin_lock(&svc->srv_at_lock);
 861         if (svc->srv_at_check == 0) {
 862                 spin_unlock(&svc->srv_at_lock);
 863                 RETURN(0);
 864         }
 865         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 866         svc->srv_at_check = 0;
 867
 868         if (array->paa_count == 0) {
 869                 spin_unlock(&svc->srv_at_lock);
 870                 RETURN(0);
 871         }
 872
 873         /* The timer went off, but maybe the nearest rpc already completed. */
 874         first = array->paa_deadline - now;
 875         if (first > at_early_margin) {
 876                 /* We've still got plenty of time.  Reset the timer. */
 877                 spin_unlock(&svc->srv_at_lock);
 878                 ptlrpc_at_set_timer(svc);
 879                 RETURN(0);
 880         }
 881
 882         /* We're close to a timeout, and we don't know how much longer the
 883            server will take. Send early replies to everyone expiring soon. */
 884         CFS_INIT_LIST_HEAD(&work_list);
 885         deadline = -1;
 886         index = (unsigned long)array->paa_deadline % array->paa_size;
 887         count = array->paa_count;
 888         while (count > 0) {
 889                 count -= array->paa_reqs_count[index];
 890                 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
 891                                          rq_timed_list) {
 892                         if (rq->rq_deadline <= now + at_early_margin) {
 893                                 list_del_init(&rq->rq_timed_list);
 894                                 /**
 895                                  * ptlrpc_server_drop_request() may drop
 896                                  * refcount to 0 already. Let's check this and
 897                                  * don't add entry to work_list
 898                                  */
 899                                 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
 900                                         list_add(&rq->rq_timed_list, &work_list);
 901                                 counter++;
 902                                 array->paa_reqs_count[index]--;
 903                                 array->paa_count--;
 904                                 rq->rq_at_linked = 0;
 905                                 continue;
 906                         }
 907
 908                         /* update the earliest deadline */
 909                         if (deadline == -1 || rq->rq_deadline < deadline)
 910                                 deadline = rq->rq_deadline;
 911
 912                         break;
 913                 }
 914
 915                 if (++index >= array->paa_size)
 916                         index = 0;
 917         }
 918         array->paa_deadline = deadline;
 919         spin_unlock(&svc->srv_at_lock);
 920
 921         /* we have a new earliest deadline, restart the timer */
 922         ptlrpc_at_set_timer(svc);
 923
 924         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 925                "replies\n", first, at_extra, counter);
 926
 927         if (first < 0) {
 928                 /* We're already past request deadlines before we even get a
 929                    chance to send early replies */
 930                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 931                               "request traffic (cpu-bound).\n",  svc->srv_name);
 932                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 933                       "delay="CFS_DURATION_T"(jiff)\n",
 934                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 935                       at_get(&svc->srv_at_estimate), delay);
 936         }
 937
 938         /* we took additional refcount so entries can't be deleted from list, no
 939          * locking is needed */
 940         while (!list_empty(&work_list)) {
 941                 rq = list_entry(work_list.next, struct ptlrpc_request,
 942                                 rq_timed_list);
 943                 list_del_init(&rq->rq_timed_list);
 944
 945                 if (ptlrpc_at_send_early_reply(rq) == 0)
 946                         ptlrpc_at_add_timed(rq);
 947
 948                 ptlrpc_server_drop_request(rq);
 949         }
 950
 951         RETURN(0);
 952 }
 953
 954 /**
 955  * Put the request to the export list if the request may become
 956  * a high priority one.
 957  */
 958 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
 959                              struct ptlrpc_request *req)
 960 {
 961         int rc;
 962         ENTRY;
 963
 964         if (svc->srv_hpreq_handler) {
 965                 rc = svc->srv_hpreq_handler(req);
 966                 if (rc)
 967                         RETURN(rc);
 968         }
 969         if (req->rq_export && req->rq_ops) {
 970                 spin_lock(&req->rq_export->exp_lock);
 971                 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
 972                 spin_unlock(&req->rq_export->exp_lock);
 973         }
 974
 975         RETURN(0);
 976 }
 977
 978 /** Remove the request from the export list. */
 979 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
 980 {
 981         ENTRY;
 982         if (req->rq_export && req->rq_ops) {
 983                 spin_lock(&req->rq_export->exp_lock);
 984                 list_del_init(&req->rq_exp_list);
 985                 spin_unlock(&req->rq_export->exp_lock);
 986         }
 987         EXIT;
 988 }
 989
 990 /**
 991  * Make the request a high priority one.
 992  *
 993  * All the high priority requests are queued in a separate FIFO
 994  * ptlrpc_service::srv_request_hpq list which is parallel to
 995  * ptlrpc_service::srv_request_queue list but has a higher priority
 996  * for handling.
 997  *
 998  * \see ptlrpc_server_handle_request().
 999  */
1000 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1001                                         struct ptlrpc_request *req)
1002 {
1003         ENTRY;
1004         LASSERT(svc != NULL);
1005         spin_lock(&req->rq_lock);
1006         if (req->rq_hp == 0) {
1007                 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1008
1009                 /* Add to the high priority queue. */
1010                 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1011                 req->rq_hp = 1;
1012                 if (opc != OBD_PING)
1013                         DEBUG_REQ(D_NET, req, "high priority req");
1014         }
1015         spin_unlock(&req->rq_lock);
1016         EXIT;
1017 }
1018
1019 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1020 {
1021         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1022         ENTRY;
1023
1024         spin_lock(&svc->srv_lock);
1025         /* It may happen that the request is already taken for the processing
1026          * but still in the export list, do not re-add it into the HP list. */
1027         if (req->rq_phase == RQ_PHASE_NEW)
1028                 ptlrpc_hpreq_reorder_nolock(svc, req);
1029         spin_unlock(&svc->srv_lock);
1030         EXIT;
1031 }
1032
1033 /** Check if the request if a high priority one. */
1034 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1035 {
1036         int opc, rc = 0;
1037         ENTRY;
1038
1039         /* Check by request opc. */
1040         opc = lustre_msg_get_opc(req->rq_reqmsg);
1041         if (opc == OBD_PING)
1042                 RETURN(1);
1043
1044         /* Perform request specific check. */
1045         if (req->rq_ops && req->rq_ops->hpreq_check)
1046                 rc = req->rq_ops->hpreq_check(req);
1047         RETURN(rc);
1048 }
1049
1050 /** Check if a request is a high priority one. */
1051 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1052                                      struct ptlrpc_request *req)
1053 {
1054         int rc;
1055         ENTRY;
1056
1057         rc = ptlrpc_server_hpreq_check(req);
1058         if (rc < 0)
1059                 RETURN(rc);
1060
1061         spin_lock(&svc->srv_lock);
1062         /* Before inserting the request into the queue, check if it is not
1063          * inserted yet, or even already handled -- it may happen due to
1064          * a racing ldlm_server_blocking_ast(). */
1065         if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1066                 if (rc)
1067                         ptlrpc_hpreq_reorder_nolock(svc, req);
1068                 else
1069                         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1070         }
1071         spin_unlock(&svc->srv_lock);
1072
1073         RETURN(0);
1074 }
1075
1076 /* Only allow normal priority requests on a service that has a high-priority
1077  * queue if forced (i.e. cleanup), if there are other high priority requests
1078  * already being processed (i.e. those threads can service more high-priority
1079  * requests), or if there are enough idle threads that a later thread can do
1080  * a high priority request. */
1081 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1082 {
1083         return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1084                svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1085 }
1086
1087 static struct ptlrpc_request *
1088 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1089 {
1090         struct ptlrpc_request *req = NULL;
1091         ENTRY;
1092
1093         if (ptlrpc_server_allow_normal(svc, force) &&
1094             !list_empty(&svc->srv_request_queue) &&
1095             (list_empty(&svc->srv_request_hpq) ||
1096              svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1097                 req = list_entry(svc->srv_request_queue.next,
1098                                  struct ptlrpc_request, rq_list);
1099                 svc->srv_hpreq_count = 0;
1100         } else if (!list_empty(&svc->srv_request_hpq)) {
1101                 req = list_entry(svc->srv_request_hpq.next,
1102                                  struct ptlrpc_request, rq_list);
1103                 svc->srv_hpreq_count++;
1104         }
1105         RETURN(req);
1106 }
1107
1108 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1109 {
1110         return ((ptlrpc_server_allow_normal(svc, force) &&
1111                  !list_empty(&svc->srv_request_queue)) ||
1112                 !list_empty(&svc->srv_request_hpq));
1113 }
1114
1115 /* Handle freshly incoming reqs, add to timed early reply list,
1116    pass on to regular request queue */
1117 static int
1118 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1119 {
1120         struct ptlrpc_request *req;
1121         __u32                  deadline;
1122         int                    rc;
1123         ENTRY;
1124
1125         LASSERT(svc);
1126
1127         spin_lock(&svc->srv_lock);
1128         if (list_empty(&svc->srv_req_in_queue)) {
1129                 spin_unlock(&svc->srv_lock);
1130                 RETURN(0);
1131         }
1132
1133         req = list_entry(svc->srv_req_in_queue.next,
1134                          struct ptlrpc_request, rq_list);
1135         list_del_init (&req->rq_list);
1136         /* Consider this still a "queued" request as far as stats are
1137            concerned */
1138         spin_unlock(&svc->srv_lock);
1139
1140         /* Clear request swab mask; this is a new request */
1141         req->rq_req_swab_mask = 0;
1142
1143         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1144         if (rc < 0) {
1145                 CERROR ("error unpacking request: ptl %d from %s"
1146                         " xid "LPU64"\n", svc->srv_req_portal,
1147                         libcfs_id2str(req->rq_peer), req->rq_xid);
1148                 goto err_req;
1149         }
1150
1151         if (rc > 0)
1152                 lustre_set_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1153
1154         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1155         if (rc) {
1156                 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
1157                         " xid "LPU64"\n", svc->srv_req_portal,
1158                         libcfs_id2str(req->rq_peer), req->rq_xid);
1159                 goto err_req;
1160         }
1161
1162         rc = -EINVAL;
1163         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1164                 CERROR("wrong packet type received (type=%u) from %s\n",
1165                        lustre_msg_get_type(req->rq_reqmsg),
1166                        libcfs_id2str(req->rq_peer));
1167                 goto err_req;
1168         }
1169
1170         CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1171
1172         req->rq_export = class_conn2export(
1173                 lustre_msg_get_handle(req->rq_reqmsg));
1174         if (req->rq_export) {
1175                 rc = ptlrpc_check_req(req);
1176                 if (rc)
1177                         goto err_req;
1178                 ptlrpc_update_export_timer(req->rq_export, 0);
1179         }
1180
1181         /* req_in handling should/must be fast */
1182         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1183                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus",
1184                           cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
1185
1186         /* Set rpc server deadline and add it to the timed list */
1187         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1188                     MSGHDR_AT_SUPPORT) ?
1189                    /* The max time the client expects us to take */
1190                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1191         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1192         if (unlikely(deadline == 0)) {
1193                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1194                 goto err_req;
1195         }
1196
1197         ptlrpc_at_add_timed(req);
1198         rc = ptlrpc_hpreq_init(svc, req);
1199         if (rc)
1200                 GOTO(err_req, rc);
1201
1202         /* Move it over to the request processing queue */
1203         rc = ptlrpc_server_request_add(svc, req);
1204         if (rc)
1205                 GOTO(err_req, rc);
1206         cfs_waitq_signal(&svc->srv_waitq);
1207         RETURN(1);
1208
1209 err_req:
1210         spin_lock(&svc->srv_lock);
1211         svc->srv_n_queued_reqs--;
1212         svc->srv_n_active_reqs++;
1213         spin_unlock(&svc->srv_lock);
1214         ptlrpc_server_finish_request(req);
1215
1216         RETURN(1);
1217 }
1218
1219 static int
1220 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1221                              struct ptlrpc_thread *thread)
1222 {
1223         struct obd_export     *export = NULL;
1224         struct ptlrpc_request *request;
1225         struct timeval         work_start;
1226         struct timeval         work_end;
1227         long                   timediff;
1228         int                    opc, rc;
1229         int                    fail_opc = 0;
1230         ENTRY;
1231
1232         LASSERT(svc);
1233
1234         spin_lock(&svc->srv_lock);
1235         if (!ptlrpc_server_request_pending(svc, 0) ||
1236             (
1237 #ifndef __KERNEL__
1238              /* !@%$# liblustre only has 1 thread */
1239              svc->srv_n_difficult_replies != 0 &&
1240 #endif
1241              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1))) {
1242                 /* Don't handle regular requests in the last thread, in order
1243                  * to handle difficult replies (which might block other threads)
1244                  * as well as handle any incoming reqs, early replies, etc.
1245                  * That means we always need at least 2 service threads. */
1246                 spin_unlock(&svc->srv_lock);
1247                 RETURN(0);
1248         }
1249
1250         request = ptlrpc_server_request_get(svc, 0);
1251         if  (request == NULL) {
1252                 spin_unlock(&svc->srv_lock);
1253                 RETURN(0);
1254         }
1255
1256         opc = lustre_msg_get_opc(request->rq_reqmsg);
1257         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1258                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1259         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1260                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1261
1262         if (unlikely(fail_opc)) {
1263                 if (request->rq_export && request->rq_ops) {
1264                         spin_unlock(&svc->srv_lock);
1265                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1266                         spin_lock(&svc->srv_lock);
1267                         request = ptlrpc_server_request_get(svc, 0);
1268                         if  (request == NULL) {
1269                                 spin_unlock(&svc->srv_lock);
1270                                 RETURN(0);
1271                         }
1272                         LASSERT(ptlrpc_server_request_pending(svc, 0));
1273                 }
1274         }
1275
1276         list_del_init(&request->rq_list);
1277         svc->srv_n_queued_reqs--;
1278         svc->srv_n_active_reqs++;
1279
1280         if (request->rq_hp)
1281                 svc->srv_n_hpreq++;
1282
1283         /* The phase is changed under the lock here because we need to know
1284          * the request is under processing (see ptlrpc_hpreq_reorder()). */
1285         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1286         spin_unlock(&svc->srv_lock);
1287
1288         ptlrpc_hpreq_fini(request);
1289
1290         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1291                 libcfs_debug_dumplog();
1292
1293         do_gettimeofday(&work_start);
1294         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1295         if (svc->srv_stats != NULL) {
1296                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1297                                     timediff);
1298                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1299                                     svc->srv_n_queued_reqs);
1300                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1301                                     svc->srv_n_active_reqs);
1302                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1303                                     at_get(&svc->srv_at_estimate));
1304         }
1305
1306         CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1307
1308         request->rq_svc_thread = thread;
1309         if (request->rq_export) {
1310                 if (ptlrpc_check_req(request))
1311                         goto put_conn;
1312                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1313                 export = class_export_rpc_get(request->rq_export);
1314         }
1315
1316         /* Discard requests queued for longer than the deadline.
1317            The deadline is increased if we send an early reply. */
1318         if (cfs_time_current_sec() > request->rq_deadline) {
1319                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1320                           ": deadline %ld%+lds ago\n",
1321                           libcfs_id2str(request->rq_peer),
1322                           request->rq_deadline -
1323                           request->rq_arrival_time.tv_sec,
1324                           cfs_time_current_sec() - request->rq_deadline);
1325                 goto put_rpc_export;
1326         }
1327
1328         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1329                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1330                (request->rq_export ?
1331                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1332                (request->rq_export ?
1333                 atomic_read(&request->rq_export->exp_refcount) : -99),
1334                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1335                libcfs_id2str(request->rq_peer),
1336                lustre_msg_get_opc(request->rq_reqmsg));
1337
1338         OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1339
1340         rc = svc->srv_handler(request);
1341
1342         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1343
1344 put_rpc_export:
1345         if (export != NULL && !request->rq_copy_queued)
1346                 class_export_rpc_put(export);
1347
1348 put_conn:
1349         if (cfs_time_current_sec() > request->rq_deadline) {
1350                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1351                           "than estimated (%ld%+lds); client may timeout.",
1352                           request->rq_xid, request->rq_deadline -
1353                           request->rq_arrival_time.tv_sec,
1354                           cfs_time_current_sec() - request->rq_deadline);
1355         }
1356
1357         do_gettimeofday(&work_end);
1358         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1359
1360         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1361                "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1362                "%lds (%lds total) trans "LPU64" rc %d/%d\n",
1363                cfs_curproc_comm(),
1364                (request->rq_export ?
1365                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1366                (request->rq_export ?
1367                 atomic_read(&request->rq_export->exp_refcount) : -99),
1368                lustre_msg_get_status(request->rq_reqmsg),
1369                request->rq_xid,
1370                libcfs_id2str(request->rq_peer),
1371                lustre_msg_get_opc(request->rq_reqmsg),
1372                timediff,
1373                cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1374                (request->rq_repmsg ?
1375                 lustre_msg_get_transno(request->rq_repmsg) :
1376                 request->rq_transno),
1377                request->rq_status,
1378                (request->rq_repmsg ?
1379                 lustre_msg_get_status(request->rq_repmsg) : -999));
1380
1381         if (svc->srv_stats != NULL) {
1382                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1383                 int opc = opcode_offset(op);
1384                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1385                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1386                         lprocfs_counter_add(svc->srv_stats,
1387                                             opc + EXTRA_MAX_OPCODES,
1388                                             timediff);
1389                 }
1390         }
1391         if (request->rq_early_count) {
1392                 DEBUG_REQ(D_ADAPTTO, request,
1393                           "sent %d early replies before finishing in %lds",
1394                           request->rq_early_count,
1395                           work_end.tv_sec - request->rq_arrival_time.tv_sec);
1396         }
1397
1398         spin_lock(&svc->srv_lock);
1399         if (request->rq_hp)
1400                 svc->srv_n_hpreq--;
1401         spin_unlock(&svc->srv_lock);
1402         ptlrpc_server_finish_request(request);
1403
1404         RETURN(1);
1405 }
1406
1407 static int
1408 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1409 {
1410         struct ptlrpc_reply_state *rs;
1411         struct obd_export         *exp;
1412         struct obd_device         *obd;
1413         int                        nlocks;
1414         int                        been_handled;
1415         ENTRY;
1416
1417         spin_lock(&svc->srv_lock);
1418         if (list_empty (&svc->srv_reply_queue)) {
1419                 spin_unlock(&svc->srv_lock);
1420                 RETURN(0);
1421         }
1422
1423         rs = list_entry (svc->srv_reply_queue.next,
1424                          struct ptlrpc_reply_state, rs_list);
1425
1426         exp = rs->rs_export;
1427         obd = exp->exp_obd;
1428
1429         LASSERT (rs->rs_difficult);
1430         LASSERT (rs->rs_scheduled);
1431
1432         list_del_init (&rs->rs_list);
1433
1434         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1435         spin_unlock(&svc->srv_lock);
1436
1437         spin_lock (&exp->exp_uncommitted_replies_lock);
1438         /* Noop if removed already */
1439         list_del_init (&rs->rs_obd_list);
1440         spin_unlock (&exp->exp_uncommitted_replies_lock);
1441
1442         spin_lock (&exp->exp_lock);
1443         /* Noop if removed already */
1444         list_del_init (&rs->rs_exp_list);
1445         spin_unlock (&exp->exp_lock);
1446
1447         spin_lock(&svc->srv_lock);
1448
1449         been_handled = rs->rs_handled;
1450         rs->rs_handled = 1;
1451
1452         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1453         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1454
1455         if (nlocks == 0 && !been_handled) {
1456                 /* If we see this, we should already have seen the warning
1457                  * in mds_steal_ack_locks()  */
1458                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1459                       " o%d NID %s\n", rs, rs->rs_xid, rs->rs_transno,
1460                       lustre_msg_get_opc(rs->rs_msg),
1461                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1462         }
1463
1464         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1465                 spin_unlock(&svc->srv_lock);
1466
1467                 if (!been_handled && rs->rs_on_net) {
1468                         LNetMDUnlink(rs->rs_md_h);
1469                         /* Ignore return code; we're racing with
1470                          * completion... */
1471                 }
1472
1473                 while (nlocks-- > 0)
1474                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1475                                          rs->rs_modes[nlocks]);
1476
1477                 spin_lock(&svc->srv_lock);
1478         }
1479
1480         rs->rs_scheduled = 0;
1481
1482         if (!rs->rs_on_net) {
1483                 /* Off the net */
1484                 svc->srv_n_difficult_replies--;
1485                 spin_unlock(&svc->srv_lock);
1486
1487                 class_export_put (exp);
1488                 rs->rs_export = NULL;
1489                 ptlrpc_rs_decref (rs);
1490                 atomic_dec (&svc->srv_outstanding_replies);
1491                 RETURN(1);
1492         }
1493
1494         /* still on the net; callback will schedule */
1495         spin_unlock(&svc->srv_lock);
1496         RETURN(1);
1497 }
1498
1499 #ifndef __KERNEL__
1500 /* FIXME make use of timeout later */
1501 int
1502 liblustre_check_services (void *arg)
1503 {
1504         int  did_something = 0;
1505         int  rc;
1506         struct list_head *tmp, *nxt;
1507         ENTRY;
1508
1509         /* I'm relying on being single threaded, not to have to lock
1510          * ptlrpc_all_services etc */
1511         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1512                 struct ptlrpc_service *svc =
1513                         list_entry (tmp, struct ptlrpc_service, srv_list);
1514
1515                 if (svc->srv_threads_running != 0)     /* I've recursed */
1516                         continue;
1517
1518                 /* service threads can block for bulk, so this limits us
1519                  * (arbitrarily) to recursing 1 stack frame per service.
1520                  * Note that the problem with recursion is that we have to
1521                  * unwind completely before our caller can resume. */
1522
1523                 svc->srv_threads_running++;
1524
1525                 do {
1526                         rc = ptlrpc_server_handle_req_in(svc);
1527                         rc |= ptlrpc_server_handle_reply(svc);
1528                         rc |= ptlrpc_at_check_timed(svc);
1529                         rc |= ptlrpc_server_handle_request(svc, NULL);
1530                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1531                         did_something |= rc;
1532                 } while (rc);
1533
1534                 svc->srv_threads_running--;
1535         }
1536
1537         RETURN(did_something);
1538 }
1539 #define ptlrpc_stop_all_threads(s) do {} while (0)
1540
1541 #else /* __KERNEL__ */
1542
1543 static void
1544 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1545 {
1546         int avail = svc->srv_nrqbd_receiving;
1547         int low_water = test_req_buffer_pressure ? 0 :
1548                         svc->srv_nbuf_per_group/2;
1549
1550         /* NB I'm not locking; just looking. */
1551
1552         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1553          * allowed the request history to grow out of control.  We could put a
1554          * sanity check on that here and cull some history if we need the
1555          * space. */
1556
1557         if (avail <= low_water)
1558                 ptlrpc_grow_req_bufs(svc);
1559
1560         if (svc->srv_stats)
1561                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1562                                     avail);
1563 }
1564
1565 static int
1566 ptlrpc_retry_rqbds(void *arg)
1567 {
1568         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1569
1570         svc->srv_rqbd_timeout = 0;
1571         return (-ETIMEDOUT);
1572 }
1573
1574 static int ptlrpc_main(void *arg)
1575 {
1576         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1577         struct ptlrpc_service  *svc = data->svc;
1578         struct ptlrpc_thread   *thread = data->thread;
1579         struct obd_device      *dev = data->dev;
1580         struct ptlrpc_reply_state *rs;
1581 #ifdef WITH_GROUP_INFO
1582         struct group_info *ginfo = NULL;
1583 #endif
1584         int counter = 0, rc = 0;
1585         ENTRY;
1586
1587         cfs_daemonize_ctxt(data->name);
1588
1589 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1590         /* we need to do this before any per-thread allocation is done so that
1591          * we get the per-thread allocations on local node.  bug 7342 */
1592         if (svc->srv_cpu_affinity) {
1593                 int cpu, num_cpu;
1594
1595                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1596                         if (!cpu_online(cpu))
1597                                 continue;
1598                         if (num_cpu == thread->t_id % num_online_cpus())
1599                                 break;
1600                         num_cpu++;
1601                 }
1602                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1603         }
1604 #endif
1605
1606 #ifdef WITH_GROUP_INFO
1607         ginfo = groups_alloc(0);
1608         if (!ginfo) {
1609                 rc = -ENOMEM;
1610                 goto out;
1611         }
1612
1613         set_current_groups(ginfo);
1614         put_group_info(ginfo);
1615 #endif
1616
1617         if (svc->srv_init != NULL) {
1618                 rc = svc->srv_init(thread);
1619                 if (rc)
1620                         goto out;
1621         }
1622
1623         /* Alloc reply state structure for this one */
1624         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1625         if (!rs) {
1626                 rc = -ENOMEM;
1627                 goto out_srv_init;
1628         }
1629
1630         spin_lock(&svc->srv_lock);
1631         /* SVC_STOPPING may already be set here if someone else is trying
1632          * to stop the service while this new thread has been dynamically
1633          * forked. We still set SVC_RUNNING to let our creator know that
1634          * we are now running, however we will exit as soon as possible */
1635         thread->t_flags |= SVC_RUNNING;
1636         spin_unlock(&svc->srv_lock);
1637
1638         /*
1639          * wake up our creator. Note: @data is invalid after this point,
1640          * because it's allocated on ptlrpc_start_thread() stack.
1641          */
1642         cfs_waitq_signal(&thread->t_ctl_waitq);
1643
1644         thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL);
1645
1646         spin_lock(&svc->srv_lock);
1647         svc->srv_threads_running++;
1648         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1649         spin_unlock(&svc->srv_lock);
1650         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1651
1652         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1653                svc->srv_threads_running);
1654
1655         /* XXX maintain a list of all managed devices: insert here */
1656
1657         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1658                svc->srv_n_difficult_replies != 0) {
1659                 /* Don't exit while there are replies to be handled */
1660                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1661                                                      ptlrpc_retry_rqbds, svc);
1662
1663                 lc_watchdog_disable(thread->t_watchdog);
1664
1665                 cfs_cond_resched();
1666
1667                 l_wait_event_exclusive (svc->srv_waitq,
1668                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1669                                svc->srv_n_difficult_replies == 0) ||
1670                               (!list_empty(&svc->srv_idle_rqbds) &&
1671                                svc->srv_rqbd_timeout == 0) ||
1672                               !list_empty(&svc->srv_req_in_queue) ||
1673                               !list_empty(&svc->srv_reply_queue) ||
1674                               (ptlrpc_server_request_pending(svc, 0) &&
1675                                (svc->srv_n_active_reqs <
1676                                 (svc->srv_threads_running - 1))) ||
1677                               svc->srv_at_check,
1678                               &lwi);
1679
1680                 lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc));
1681
1682                 ptlrpc_check_rqbd_pool(svc);
1683
1684                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1685                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1686                         /* Ignore return code - we tried... */
1687                         ptlrpc_start_thread(dev, svc);
1688                 }
1689
1690                 if (!list_empty(&svc->srv_reply_queue))
1691                         ptlrpc_server_handle_reply (svc);
1692
1693                 if (!list_empty(&svc->srv_req_in_queue)) {
1694                         /* Process all incoming reqs before handling any */
1695                         ptlrpc_server_handle_req_in(svc);
1696                         /* but limit ourselves in case of flood */
1697                         if (counter++ < 1000)
1698                                 continue;
1699                         counter = 0;
1700                 }
1701
1702                 if (svc->srv_at_check)
1703                         ptlrpc_at_check_timed(svc);
1704
1705                 /* don't handle requests in the last thread */
1706                 if (ptlrpc_server_request_pending(svc, 0) &&
1707                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1)))
1708                         ptlrpc_server_handle_request(svc, thread);
1709
1710                 if (!list_empty(&svc->srv_idle_rqbds) &&
1711                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1712                         /* I just failed to repost request buffers.  Wait
1713                          * for a timeout (unless something else happens)
1714                          * before I try again */
1715                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1716                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1717                                svc->srv_nrqbd_receiving);
1718                 }
1719         }
1720
1721         lc_watchdog_delete(thread->t_watchdog);
1722         thread->t_watchdog = NULL;
1723
1724 out_srv_init:
1725         /*
1726          * deconstruct service specific state created by ptlrpc_start_thread()
1727          */
1728         if (svc->srv_done != NULL)
1729                 svc->srv_done(thread);
1730
1731 out:
1732         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1733
1734         spin_lock(&svc->srv_lock);
1735         svc->srv_threads_running--;              /* must know immediately */
1736         thread->t_id = rc;
1737         thread->t_flags = SVC_STOPPED;
1738
1739         cfs_waitq_signal(&thread->t_ctl_waitq);
1740         spin_unlock(&svc->srv_lock);
1741
1742         return rc;
1743 }
1744
1745 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1746                                struct ptlrpc_thread *thread)
1747 {
1748         struct l_wait_info lwi = { 0 };
1749
1750         spin_lock(&svc->srv_lock);
1751         /* let the thread know that we would like it to stop asap */
1752         thread->t_flags |= SVC_STOPPING;
1753         spin_unlock(&svc->srv_lock);
1754
1755         cfs_waitq_broadcast(&svc->srv_waitq);
1756         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1757                      &lwi);
1758
1759         spin_lock(&svc->srv_lock);
1760         list_del(&thread->t_link);
1761         spin_unlock(&svc->srv_lock);
1762
1763         OBD_FREE(thread, sizeof(*thread));
1764 }
1765
1766 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1767 {
1768         struct ptlrpc_thread *thread;
1769
1770         spin_lock(&svc->srv_lock);
1771         while (!list_empty(&svc->srv_threads)) {
1772                 thread = list_entry(svc->srv_threads.next,
1773                                     struct ptlrpc_thread, t_link);
1774
1775                 spin_unlock(&svc->srv_lock);
1776                 ptlrpc_stop_thread(svc, thread);
1777                 spin_lock(&svc->srv_lock);
1778         }
1779
1780         spin_unlock(&svc->srv_lock);
1781 }
1782
1783 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1784 {
1785         int i, rc = 0;
1786         ENTRY;
1787
1788         /* We require 2 threads min - see note in
1789          * ptlrpc_server_handle_request() */
1790
1791         LASSERT(svc->srv_threads_min >= 2);
1792         for (i = 0; i < svc->srv_threads_min; i++) {
1793                 rc = ptlrpc_start_thread(dev, svc);
1794                 /* We have enough threads, don't start more.  b=15759 */
1795                 if (rc == -EMFILE)
1796                         break;
1797                 if (rc) {
1798                         CERROR("cannot start %s thread #%d: rc %d\n",
1799                                svc->srv_thread_name, i, rc);
1800                         ptlrpc_stop_all_threads(svc);
1801                 }
1802         }
1803         RETURN(rc);
1804 }
1805
1806 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1807 {
1808         struct l_wait_info lwi = { 0 };
1809         struct ptlrpc_svc_data d;
1810         struct ptlrpc_thread *thread;
1811         char name[32];
1812         int id, rc;
1813         ENTRY;
1814
1815         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1816                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1817                svc->srv_threads_max, svc->srv_threads_running);
1818         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1819             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1820              svc->srv_threads_started == svc->srv_threads_min - 1))
1821                 RETURN(-EMFILE);
1822
1823         OBD_ALLOC(thread, sizeof(*thread));
1824         if (thread == NULL)
1825                 RETURN(-ENOMEM);
1826         cfs_waitq_init(&thread->t_ctl_waitq);
1827
1828         spin_lock(&svc->srv_lock);
1829         if (svc->srv_threads_started >= svc->srv_threads_max) {
1830                 spin_unlock(&svc->srv_lock);
1831                 OBD_FREE(thread, sizeof(*thread));
1832                 RETURN(-EMFILE);
1833         }
1834         list_add(&thread->t_link, &svc->srv_threads);
1835         id = svc->srv_threads_started++;
1836         spin_unlock(&svc->srv_lock);
1837
1838         thread->t_svc = svc;
1839         thread->t_id = id;
1840         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1841         d.dev = dev;
1842         d.svc = svc;
1843         d.name = name;
1844         d.thread = thread;
1845
1846         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1847
1848         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1849          * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
1850          */
1851         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1852         if (rc < 0) {
1853                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1854
1855                 spin_lock(&svc->srv_lock);
1856                 list_del(&thread->t_link);
1857                 --svc->srv_threads_started;
1858                 spin_unlock(&svc->srv_lock);
1859
1860                 OBD_FREE(thread, sizeof(*thread));
1861                 RETURN(rc);
1862         }
1863         l_wait_event(thread->t_ctl_waitq,
1864                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1865
1866         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1867         RETURN(rc);
1868 }
1869 #endif
1870
1871 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1872 {
1873         int                   rc;
1874         struct l_wait_info    lwi;
1875         struct list_head     *tmp;
1876         struct ptlrpc_reply_state *rs, *t;
1877         struct ptlrpc_at_array *array = &service->srv_at_array;
1878
1879         cfs_timer_disarm(&service->srv_at_timer);
1880
1881         ptlrpc_stop_all_threads(service);
1882         LASSERT(list_empty(&service->srv_threads));
1883
1884         spin_lock (&ptlrpc_all_services_lock);
1885         list_del_init (&service->srv_list);
1886         spin_unlock (&ptlrpc_all_services_lock);
1887
1888         ptlrpc_lprocfs_unregister_service(service);
1889
1890         /* All history will be culled when the next request buffer is
1891          * freed */
1892         service->srv_max_history_rqbds = 0;
1893
1894         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1895
1896         rc = LNetClearLazyPortal(service->srv_req_portal);
1897         LASSERT (rc == 0);
1898
1899         /* Unlink all the request buffers.  This forces a 'final' event with
1900          * its 'unlink' flag set for each posted rqbd */
1901         list_for_each(tmp, &service->srv_active_rqbds) {
1902                 struct ptlrpc_request_buffer_desc *rqbd =
1903                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1904                                    rqbd_list);
1905
1906                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1907                 LASSERT (rc == 0 || rc == -ENOENT);
1908         }
1909
1910         /* Wait for the network to release any buffers it's currently
1911          * filling */
1912         for (;;) {
1913                 spin_lock(&service->srv_lock);
1914                 rc = service->srv_nrqbd_receiving;
1915                 spin_unlock(&service->srv_lock);
1916
1917                 if (rc == 0)
1918                         break;
1919
1920                 /* Network access will complete in finite time but the HUGE
1921                  * timeout lets us CWARN for visibility of sluggish NALs */
1922                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1923                                            cfs_time_seconds(1), NULL, NULL);
1924                 rc = l_wait_event(service->srv_waitq,
1925                                   service->srv_nrqbd_receiving == 0,
1926                                   &lwi);
1927                 if (rc == -ETIMEDOUT)
1928                         CWARN("Service %s waiting for request buffers\n",
1929                               service->srv_name);
1930         }
1931
1932         /* schedule all outstanding replies to terminate them */
1933         spin_lock(&service->srv_lock);
1934         while (!list_empty(&service->srv_active_replies)) {
1935                 struct ptlrpc_reply_state *rs =
1936                         list_entry(service->srv_active_replies.next,
1937                                    struct ptlrpc_reply_state, rs_list);
1938                 ptlrpc_schedule_difficult_reply(rs);
1939         }
1940         spin_unlock(&service->srv_lock);
1941
1942         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1943          * and no service threads, so I'm the only thread noodling the
1944          * request queue now */
1945         while (!list_empty(&service->srv_req_in_queue)) {
1946                 struct ptlrpc_request *req =
1947                         list_entry(service->srv_req_in_queue.next,
1948                                    struct ptlrpc_request,
1949                                    rq_list);
1950
1951                 list_del(&req->rq_list);
1952                 service->srv_n_queued_reqs--;
1953                 service->srv_n_active_reqs++;
1954                 ptlrpc_server_finish_request(req);
1955         }
1956         while (ptlrpc_server_request_pending(service, 1)) {
1957                 struct ptlrpc_request *req;
1958
1959                 req = ptlrpc_server_request_get(service, 1);
1960                 list_del(&req->rq_list);
1961                 service->srv_n_queued_reqs--;
1962                 service->srv_n_active_reqs++;
1963                 ptlrpc_hpreq_fini(req);
1964                 ptlrpc_server_finish_request(req);
1965         }
1966         LASSERT(service->srv_n_queued_reqs == 0);
1967         LASSERT(service->srv_n_active_reqs == 0);
1968         LASSERT(service->srv_n_history_rqbds == 0);
1969         LASSERT(list_empty(&service->srv_active_rqbds));
1970
1971         /* Now free all the request buffers since nothing references them
1972          * any more... */
1973         while (!list_empty(&service->srv_idle_rqbds)) {
1974                 struct ptlrpc_request_buffer_desc *rqbd =
1975                         list_entry(service->srv_idle_rqbds.next,
1976                                    struct ptlrpc_request_buffer_desc,
1977                                    rqbd_list);
1978
1979                 ptlrpc_free_rqbd(rqbd);
1980         }
1981
1982         /* wait for all outstanding replies to complete (they were
1983          * scheduled having been flagged to abort above) */
1984         while (atomic_read(&service->srv_outstanding_replies) != 0) {
1985                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
1986
1987                 rc = l_wait_event(service->srv_waitq,
1988                                   !list_empty(&service->srv_reply_queue), &lwi);
1989                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1990
1991                 if (rc == 0) {
1992                         ptlrpc_server_handle_reply(service);
1993                         continue;
1994                 }
1995                 CWARN("Unexpectedly long timeout %p\n", service);
1996         }
1997
1998         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
1999                 list_del(&rs->rs_list);
2000                 OBD_FREE(rs, service->srv_max_reply_size);
2001         }
2002
2003         /* In case somebody rearmed this in the meantime */
2004         cfs_timer_disarm(&service->srv_at_timer);
2005
2006         if (array->paa_reqs_array != NULL) {
2007                 OBD_FREE(array->paa_reqs_array,
2008                          sizeof(struct list_head) * array->paa_size);
2009                 array->paa_reqs_array = NULL;
2010         }
2011
2012         if (array->paa_reqs_count != NULL) {
2013                 OBD_FREE(array->paa_reqs_count,
2014                          sizeof(__u32) * array->paa_size);
2015                 array->paa_reqs_count= NULL;
2016         }
2017
2018         OBD_FREE(service, sizeof(*service));
2019         return 0;
2020 }
2021
2022 /* Returns 0 if the service is healthy.
2023  *
2024  * Right now, it just checks to make sure that requests aren't languishing
2025  * in the queue.  We'll use this health check to govern whether a node needs
2026  * to be shot, so it's intentionally non-aggressive. */
2027 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2028 {
2029         struct ptlrpc_request *request;
2030         struct timeval         right_now;
2031         long                   timediff;
2032
2033         if (svc == NULL)
2034                 return 0;
2035
2036         do_gettimeofday(&right_now);
2037
2038         spin_lock(&svc->srv_lock);
2039         if (!ptlrpc_server_request_pending(svc, 1)) {
2040                 spin_unlock(&svc->srv_lock);
2041                 return 0;
2042         }
2043
2044         /* How long has the next entry been waiting? */
2045         if (list_empty(&svc->srv_request_queue))
2046                 request = list_entry(svc->srv_request_hpq.next,
2047                                      struct ptlrpc_request, rq_list);
2048         else
2049                 request = list_entry(svc->srv_request_queue.next,
2050                                      struct ptlrpc_request, rq_list);
2051         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2052         spin_unlock(&svc->srv_lock);
2053
2054         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2055                                         at_max)) {
2056                 CERROR("%s: unhealthy - request has been waiting %lds\n",
2057                        svc->srv_name, timediff / ONE_MILLION);
2058                 return (-1);
2059         }
2060
2061         return 0;
2062 }