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 (c) 2002, 2010, Oracle and/or its affiliates. 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  * increment the number of active requests consuming service threads.
 429  */
 430 void ptlrpc_server_active_request_inc(struct ptlrpc_request *req)
 431 {
 432         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 433         struct ptlrpc_service *svc = rqbd->rqbd_service;
 434
 435         spin_lock(&svc->srv_lock);
 436         svc->srv_n_active_reqs++;
 437         spin_unlock(&svc->srv_lock);
 438 }
 439
 440 /**
 441  * decrement the number of active requests consuming service threads.
 442  */
 443 void ptlrpc_server_active_request_dec(struct ptlrpc_request *req)
 444 {
 445         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 446         struct ptlrpc_service *svc = rqbd->rqbd_service;
 447
 448         spin_lock(&svc->srv_lock);
 449         svc->srv_n_active_reqs--;
 450         spin_unlock(&svc->srv_lock);
 451 }
 452
 453 /**
 454  * drop a reference count of the request. if it reaches 0, we either
 455  * put it into history list, or free it immediately.
 456  */
 457 void ptlrpc_server_drop_request(struct ptlrpc_request *req)
 458 {
 459         struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
 460         struct ptlrpc_service             *svc = rqbd->rqbd_service;
 461         int                                refcount;
 462         struct list_head                  *tmp;
 463         struct list_head                  *nxt;
 464
 465         if (!atomic_dec_and_test(&req->rq_refcount))
 466                 return;
 467
 468         spin_lock(&svc->srv_at_lock);
 469         list_del_init(&req->rq_timed_list);
 470         if (req->rq_at_linked) {
 471                 struct ptlrpc_at_array *array = &svc->srv_at_array;
 472                 __u32 index = req->rq_at_index;
 473
 474                 spin_lock(&req->rq_lock);
 475                 req->rq_at_linked = 0;
 476                 spin_unlock(&req->rq_lock);
 477                 array->paa_reqs_count[index]--;
 478                 array->paa_count--;
 479         }
 480         spin_unlock(&svc->srv_at_lock);
 481
 482         /* finalize request */
 483         if (req->rq_export) {
 484                 class_export_put(req->rq_export);
 485                 req->rq_export = NULL;
 486         }
 487
 488         spin_lock(&svc->srv_lock);
 489
 490         svc->srv_n_active_reqs--;
 491         list_add(&req->rq_list, &rqbd->rqbd_reqs);
 492
 493         refcount = --(rqbd->rqbd_refcount);
 494         if (refcount == 0) {
 495                 /* request buffer is now idle: add to history */
 496                 list_del(&rqbd->rqbd_list);
 497                 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
 498                 svc->srv_n_history_rqbds++;
 499
 500                 /* cull some history?
 501                  * I expect only about 1 or 2 rqbds need to be recycled here */
 502                 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
 503                         rqbd = list_entry(svc->srv_history_rqbds.next,
 504                                           struct ptlrpc_request_buffer_desc,
 505                                           rqbd_list);
 506
 507                         list_del(&rqbd->rqbd_list);
 508                         svc->srv_n_history_rqbds--;
 509
 510                         /* remove rqbd's reqs from svc's req history while
 511                          * I've got the service lock */
 512                         list_for_each(tmp, &rqbd->rqbd_reqs) {
 513                                 req = list_entry(tmp, struct ptlrpc_request,
 514                                                  rq_list);
 515                                 /* Track the highest culled req seq */
 516                                 if (req->rq_history_seq >
 517                                     svc->srv_request_max_cull_seq)
 518                                         svc->srv_request_max_cull_seq =
 519                                                 req->rq_history_seq;
 520                                 list_del(&req->rq_history_list);
 521                         }
 522
 523                         spin_unlock(&svc->srv_lock);
 524
 525                         list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
 526                                 req = list_entry(rqbd->rqbd_reqs.next,
 527                                                  struct ptlrpc_request,
 528                                                  rq_list);
 529                                 list_del(&req->rq_list);
 530                                 ptlrpc_server_free_request(req);
 531                         }
 532
 533                         spin_lock(&svc->srv_lock);
 534                         /*
 535                          * now all reqs including the embedded req has been
 536                          * disposed, schedule request buffer for re-use.
 537                          */
 538                         LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
 539                         list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
 540                 }
 541
 542                 spin_unlock(&svc->srv_lock);
 543         } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
 544                  /* If we are low on memory, we are not interested in history */
 545                 list_del(&req->rq_list);
 546                 list_del_init(&req->rq_history_list);
 547                 spin_unlock(&svc->srv_lock);
 548
 549                 ptlrpc_server_free_request(req);
 550         } else {
 551                 spin_unlock(&svc->srv_lock);
 552         }
 553 }
 554
 555 /**
 556  * to finish a request: stop sending more early replies, and release
 557  * the request. should be called after we finished handling the request.
 558  */
 559 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
 560 {
 561         ptlrpc_server_drop_request(req);
 562 }
 563
 564 /* This function makes sure dead exports are evicted in a timely manner.
 565    This function is only called when some export receives a message (i.e.,
 566    the network is up.) */
 567 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
 568 {
 569         struct obd_export *oldest_exp;
 570         time_t oldest_time, new_time;
 571
 572         ENTRY;
 573
 574         LASSERT(exp);
 575
 576         /* Compensate for slow machines, etc, by faking our request time
 577            into the future.  Although this can break the strict time-ordering
 578            of the list, we can be really lazy here - we don't have to evict
 579            at the exact right moment.  Eventually, all silent exports
 580            will make it to the top of the list. */
 581
 582         /* Do not pay attention on 1sec or smaller renewals. */
 583         new_time = cfs_time_current_sec() + extra_delay;
 584         if (exp->exp_last_request_time + 1 /*second */ >= new_time)
 585                 RETURN_EXIT;
 586
 587         exp->exp_last_request_time = new_time;
 588         CDEBUG(D_INFO, "updating export %s at %ld\n",
 589                exp->exp_client_uuid.uuid,
 590                exp->exp_last_request_time);
 591
 592         /* exports may get disconnected from the chain even though the
 593            export has references, so we must keep the spin lock while
 594            manipulating the lists */
 595         spin_lock(&exp->exp_obd->obd_dev_lock);
 596
 597         if (list_empty(&exp->exp_obd_chain_timed)) {
 598                 /* this one is not timed */
 599                 spin_unlock(&exp->exp_obd->obd_dev_lock);
 600                 RETURN_EXIT;
 601         }
 602
 603         list_move_tail(&exp->exp_obd_chain_timed,
 604                        &exp->exp_obd->obd_exports_timed);
 605
 606         oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
 607                                 struct obd_export, exp_obd_chain_timed);
 608         oldest_time = oldest_exp->exp_last_request_time;
 609         spin_unlock(&exp->exp_obd->obd_dev_lock);
 610
 611         if (exp->exp_obd->obd_recovering) {
 612                 /* be nice to everyone during recovery */
 613                 EXIT;
 614                 return;
 615         }
 616
 617         /* Note - racing to start/reset the obd_eviction timer is safe */
 618         if (exp->exp_obd->obd_eviction_timer == 0) {
 619                 /* Check if the oldest entry is expired. */
 620                 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
 621                                               extra_delay)) {
 622                         /* We need a second timer, in case the net was down and
 623                          * it just came back. Since the pinger may skip every
 624                          * other PING_INTERVAL (see note in ptlrpc_pinger_main),
 625                          * we better wait for 3. */
 626                         exp->exp_obd->obd_eviction_timer =
 627                                 cfs_time_current_sec() + 3 * PING_INTERVAL;
 628                         CDEBUG(D_HA, "%s: Think about evicting %s from %ld\n",
 629                                exp->exp_obd->obd_name,
 630                                obd_export_nid2str(oldest_exp), oldest_time);
 631                 }
 632         } else {
 633                 if (cfs_time_current_sec() >
 634                     (exp->exp_obd->obd_eviction_timer + extra_delay)) {
 635                         /* The evictor won't evict anyone who we've heard from
 636                          * recently, so we don't have to check before we start
 637                          * it. */
 638                         if (!ping_evictor_wake(exp))
 639                                 exp->exp_obd->obd_eviction_timer = 0;
 640                 }
 641         }
 642
 643         EXIT;
 644 }
 645
 646 static int ptlrpc_check_req(struct ptlrpc_request *req)
 647 {
 648         if (lustre_msg_get_conn_cnt(req->rq_reqmsg) <
 649             req->rq_export->exp_conn_cnt) {
 650                 DEBUG_REQ(D_ERROR, req,
 651                           "DROPPING req from old connection %d < %d",
 652                           lustre_msg_get_conn_cnt(req->rq_reqmsg),
 653                           req->rq_export->exp_conn_cnt);
 654                 return -EEXIST;
 655         }
 656         if (req->rq_export->exp_obd && req->rq_export->exp_obd->obd_fail) {
 657              /* Failing over, don't handle any more reqs, send
 658                 error response instead. */
 659                 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
 660                        req, req->rq_export->exp_obd->obd_name);
 661                 req->rq_status = -ENODEV;
 662                 ptlrpc_error(req);
 663                 return -ENODEV;
 664         }
 665         return 0;
 666 }
 667
 668 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
 669 {
 670         struct ptlrpc_at_array *array = &svc->srv_at_array;
 671         __s32 next;
 672
 673         spin_lock(&svc->srv_at_lock);
 674         if (array->paa_count == 0) {
 675                 cfs_timer_disarm(&svc->srv_at_timer);
 676                 spin_unlock(&svc->srv_at_lock);
 677                 return;
 678         }
 679
 680         /* Set timer for closest deadline */
 681         next = (__s32)(array->paa_deadline - cfs_time_current_sec() -
 682                        at_early_margin);
 683         if (next <= 0)
 684                 ptlrpc_at_timer((unsigned long)svc);
 685         else
 686                 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
 687         spin_unlock(&svc->srv_at_lock);
 688         CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
 689 }
 690
 691 /* Add rpc to early reply check list */
 692 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
 693 {
 694         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 695         struct ptlrpc_request *rq = NULL;
 696         struct ptlrpc_at_array *array = &svc->srv_at_array;
 697         __u32 index;
 698         int found = 0;
 699
 700         if (AT_OFF)
 701                 return(0);
 702
 703         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
 704                 return(-ENOSYS);
 705
 706         spin_lock(&svc->srv_at_lock);
 707         LASSERT(list_empty(&req->rq_timed_list));
 708
 709         index = (unsigned long)req->rq_deadline % array->paa_size;
 710         if (array->paa_reqs_count[index] > 0) {
 711                 /* latest rpcs will have the latest deadlines in the list,
 712                  * so search backward. */
 713                 list_for_each_entry_reverse(rq, &array->paa_reqs_array[index],
 714                                             rq_timed_list) {
 715                         if (req->rq_deadline >= rq->rq_deadline) {
 716                                 list_add(&req->rq_timed_list,
 717                                          &rq->rq_timed_list);
 718                                 break;
 719                         }
 720                 }
 721         }
 722
 723         /* Add the request at the head of the list */
 724         if (list_empty(&req->rq_timed_list))
 725                 list_add(&req->rq_timed_list, &array->paa_reqs_array[index]);
 726
 727         spin_lock(&req->rq_lock);
 728         req->rq_at_linked = 1;
 729         spin_unlock(&req->rq_lock);
 730
 731         req->rq_at_index = index;
 732         array->paa_reqs_count[index]++;
 733         array->paa_count++;
 734         if (array->paa_count == 1 || array->paa_deadline > req->rq_deadline) {
 735                 array->paa_deadline = req->rq_deadline;
 736                 found = 1;
 737         }
 738         spin_unlock(&svc->srv_at_lock);
 739
 740         if (found)
 741                 ptlrpc_at_set_timer(svc);
 742
 743         return 0;
 744 }
 745
 746 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req)
 747 {
 748         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
 749         struct ptlrpc_request *reqcopy;
 750         struct lustre_msg *reqmsg;
 751         long olddl = req->rq_deadline - cfs_time_current_sec();
 752         time_t newdl;
 753         int rc;
 754         ENTRY;
 755
 756         /* deadline is when the client expects us to reply, margin is the
 757            difference between clients' and servers' expectations */
 758         DEBUG_REQ(D_ADAPTTO, req,
 759                   "%ssending early reply (deadline %+lds, margin %+lds) for "
 760                   "%d+%d", AT_OFF ? "AT off - not " : "",
 761                   olddl, olddl - at_get(&svc->srv_at_estimate),
 762                   at_get(&svc->srv_at_estimate), at_extra);
 763
 764         if (AT_OFF)
 765                 RETURN(0);
 766
 767         if (olddl < 0) {
 768                 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
 769                           "not sending early reply. Consider increasing "
 770                           "at_early_margin (%d)?", olddl, at_early_margin);
 771
 772                 /* Return an error so we're not re-added to the timed list. */
 773                 RETURN(-ETIMEDOUT);
 774         }
 775
 776         if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
 777                 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
 778                           "but no AT support");
 779                 RETURN(-ENOSYS);
 780         }
 781
 782         if (req->rq_export &&
 783             lustre_msg_get_flags(req->rq_reqmsg) &
 784             (MSG_REPLAY | MSG_LAST_REPLAY)) {
 785                 /* During recovery, we don't want to send too many early
 786                  * replies, but on the other hand we want to make sure the
 787                  * client has enough time to resend if the rpc is lost. So
 788                  * during the recovery period send at least 4 early replies,
 789                  * spacing them every at_extra if we can. at_estimate should
 790                  * always equal this fixed value during recovery. */
 791                 at_measured(&svc->srv_at_estimate, min(at_extra,
 792                             req->rq_export->exp_obd->obd_recovery_timeout / 4));
 793         } else {
 794                 /* Fake our processing time into the future to ask the
 795                  * clients for some extra amount of time */
 796                 at_measured(&svc->srv_at_estimate, at_extra +
 797                             cfs_time_current_sec() -
 798                             req->rq_arrival_time.tv_sec);
 799
 800                 /* Check to see if we've actually increased the deadline -
 801                  * we may be past adaptive_max */
 802                 if (req->rq_deadline >= req->rq_arrival_time.tv_sec +
 803                     at_get(&svc->srv_at_estimate)) {
 804                         DEBUG_REQ(D_WARNING, req, "Couldn't add any time "
 805                                   "(%ld/%ld), not sending early reply\n",
 806                                   olddl, req->rq_arrival_time.tv_sec +
 807                                   at_get(&svc->srv_at_estimate) -
 808                                   cfs_time_current_sec());
 809                         RETURN(-ETIMEDOUT);
 810                 }
 811         }
 812         newdl = cfs_time_current_sec() + at_get(&svc->srv_at_estimate);
 813
 814         OBD_ALLOC(reqcopy, sizeof *reqcopy);
 815         if (reqcopy == NULL)
 816                 RETURN(-ENOMEM);
 817         OBD_ALLOC(reqmsg, req->rq_reqlen);
 818         if (!reqmsg) {
 819                 OBD_FREE(reqcopy, sizeof *reqcopy);
 820                 RETURN(-ENOMEM);
 821         }
 822
 823         *reqcopy = *req;
 824         reqcopy->rq_reply_state = NULL;
 825         reqcopy->rq_rep_swab_mask = 0;
 826         /* We only need the reqmsg for the magic */
 827         reqcopy->rq_reqmsg = reqmsg;
 828         memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
 829
 830         LASSERT(atomic_read(&req->rq_refcount));
 831         /** if it is last refcount then early reply isn't needed */
 832         if (atomic_read(&req->rq_refcount) == 1) {
 833                 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
 834                           "abort sending early reply\n");
 835                 GOTO(out, rc = -EINVAL);
 836         }
 837
 838         /* Connection ref */
 839         reqcopy->rq_export = class_conn2export(
 840                                      lustre_msg_get_handle(reqcopy->rq_reqmsg));
 841         if (reqcopy->rq_export == NULL)
 842                 GOTO(out, rc = -ENODEV);
 843
 844         /* RPC ref */
 845         class_export_rpc_get(reqcopy->rq_export);
 846         if (reqcopy->rq_export->exp_obd &&
 847             reqcopy->rq_export->exp_obd->obd_fail)
 848                 GOTO(out_put, rc = -ENODEV);
 849
 850         rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
 851         if (rc)
 852                 GOTO(out_put, rc);
 853
 854         rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
 855
 856         if (!rc) {
 857                 /* Adjust our own deadline to what we told the client */
 858                 req->rq_deadline = newdl;
 859                 req->rq_early_count++; /* number sent, server side */
 860         } else {
 861                 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
 862         }
 863
 864         /* Free the (early) reply state from lustre_pack_reply.
 865            (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
 866         ptlrpc_req_drop_rs(reqcopy);
 867
 868 out_put:
 869         class_export_rpc_put(reqcopy->rq_export);
 870         class_export_put(reqcopy->rq_export);
 871 out:
 872         OBD_FREE(reqmsg, req->rq_reqlen);
 873         OBD_FREE(reqcopy, sizeof *reqcopy);
 874         RETURN(rc);
 875 }
 876
 877 /* Send early replies to everybody expiring within at_early_margin
 878    asking for at_extra time */
 879 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
 880 {
 881         struct ptlrpc_request *rq, *n;
 882         struct list_head work_list;
 883         struct ptlrpc_at_array *array = &svc->srv_at_array;
 884         __u32  index, count;
 885         time_t deadline;
 886         time_t now = cfs_time_current_sec();
 887         cfs_duration_t delay;
 888         int first, counter = 0;
 889         ENTRY;
 890
 891         spin_lock(&svc->srv_at_lock);
 892         if (svc->srv_at_check == 0) {
 893                 spin_unlock(&svc->srv_at_lock);
 894                 RETURN(0);
 895         }
 896         delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
 897         svc->srv_at_check = 0;
 898
 899         if (array->paa_count == 0) {
 900                 spin_unlock(&svc->srv_at_lock);
 901                 RETURN(0);
 902         }
 903
 904         /* The timer went off, but maybe the nearest rpc already completed. */
 905         first = array->paa_deadline - now;
 906         if (first > at_early_margin) {
 907                 /* We've still got plenty of time.  Reset the timer. */
 908                 spin_unlock(&svc->srv_at_lock);
 909                 ptlrpc_at_set_timer(svc);
 910                 RETURN(0);
 911         }
 912
 913         /* We're close to a timeout, and we don't know how much longer the
 914            server will take. Send early replies to everyone expiring soon. */
 915         CFS_INIT_LIST_HEAD(&work_list);
 916         deadline = -1;
 917         index = (unsigned long)array->paa_deadline % array->paa_size;
 918         count = array->paa_count;
 919         while (count > 0) {
 920                 count -= array->paa_reqs_count[index];
 921                 list_for_each_entry_safe(rq, n, &array->paa_reqs_array[index],
 922                                          rq_timed_list) {
 923                         if (rq->rq_deadline <= now + at_early_margin) {
 924                                 list_del_init(&rq->rq_timed_list);
 925                                 /**
 926                                  * ptlrpc_server_drop_request() may drop
 927                                  * refcount to 0 already. Let's check this and
 928                                  * don't add entry to work_list
 929                                  */
 930                                 if (likely(atomic_inc_not_zero(&rq->rq_refcount)))
 931                                         list_add(&rq->rq_timed_list, &work_list);
 932                                 counter++;
 933                                 array->paa_reqs_count[index]--;
 934                                 array->paa_count--;
 935                                 spin_lock(&rq->rq_lock);
 936                                 rq->rq_at_linked = 0;
 937                                 spin_unlock(&rq->rq_lock);
 938                                 continue;
 939                         }
 940
 941                         /* update the earliest deadline */
 942                         if (deadline == -1 || rq->rq_deadline < deadline)
 943                                 deadline = rq->rq_deadline;
 944
 945                         break;
 946                 }
 947
 948                 if (++index >= array->paa_size)
 949                         index = 0;
 950         }
 951         array->paa_deadline = deadline;
 952         spin_unlock(&svc->srv_at_lock);
 953
 954         /* we have a new earliest deadline, restart the timer */
 955         ptlrpc_at_set_timer(svc);
 956
 957         CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
 958                "replies\n", first, at_extra, counter);
 959
 960         if (first < 0) {
 961                 /* We're already past request deadlines before we even get a
 962                    chance to send early replies */
 963                 LCONSOLE_WARN("%s: This server is not able to keep up with "
 964                               "request traffic (cpu-bound).\n",  svc->srv_name);
 965                 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
 966                       "delay="CFS_DURATION_T"(jiff)\n",
 967                       counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
 968                       at_get(&svc->srv_at_estimate), delay);
 969         }
 970
 971         /* we took additional refcount so entries can't be deleted from list, no
 972          * locking is needed */
 973         while (!list_empty(&work_list)) {
 974                 rq = list_entry(work_list.next, struct ptlrpc_request,
 975                                 rq_timed_list);
 976                 list_del_init(&rq->rq_timed_list);
 977
 978                 if (ptlrpc_at_send_early_reply(rq) == 0)
 979                         ptlrpc_at_add_timed(rq);
 980
 981                 ptlrpc_server_drop_request(rq);
 982         }
 983
 984         RETURN(0);
 985 }
 986
 987 /**
 988  * Put the request to the export list if the request may become
 989  * a high priority one.
 990  */
 991 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
 992                              struct ptlrpc_request *req)
 993 {
 994         int rc;
 995         ENTRY;
 996
 997         if (svc->srv_hpreq_handler) {
 998                 rc = svc->srv_hpreq_handler(req);
 999                 if (rc)
1000                         RETURN(rc);
1001         }
1002         if (req->rq_export && req->rq_ops) {
1003                 spin_lock(&req->rq_export->exp_lock);
1004                 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
1005                 spin_unlock(&req->rq_export->exp_lock);
1006         }
1007
1008         RETURN(0);
1009 }
1010
1011 /** Remove the request from the export list. */
1012 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
1013 {
1014         ENTRY;
1015         if (req->rq_export && req->rq_ops) {
1016                 spin_lock(&req->rq_export->exp_lock);
1017                 list_del_init(&req->rq_exp_list);
1018                 spin_unlock(&req->rq_export->exp_lock);
1019         }
1020         EXIT;
1021 }
1022
1023 /**
1024  * Make the request a high priority one.
1025  *
1026  * All the high priority requests are queued in a separate FIFO
1027  * ptlrpc_service::srv_request_hpq list which is parallel to
1028  * ptlrpc_service::srv_request_queue list but has a higher priority
1029  * for handling.
1030  *
1031  * \see ptlrpc_server_handle_request().
1032  */
1033 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
1034                                         struct ptlrpc_request *req)
1035 {
1036         ENTRY;
1037         LASSERT(svc != NULL);
1038         spin_lock(&req->rq_lock);
1039         if (req->rq_hp == 0) {
1040                 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1041
1042                 /* Add to the high priority queue. */
1043                 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
1044                 req->rq_hp = 1;
1045                 if (opc != OBD_PING)
1046                         DEBUG_REQ(D_NET, req, "high priority req");
1047         }
1048         spin_unlock(&req->rq_lock);
1049         EXIT;
1050 }
1051
1052 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1053 {
1054         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1055         ENTRY;
1056
1057         spin_lock(&svc->srv_lock);
1058         /* It may happen that the request is already taken for the processing
1059          * but still in the export list, do not re-add it into the HP list. */
1060         if (req->rq_phase == RQ_PHASE_NEW)
1061                 ptlrpc_hpreq_reorder_nolock(svc, req);
1062         spin_unlock(&svc->srv_lock);
1063         EXIT;
1064 }
1065
1066 /** Check if the request if a high priority one. */
1067 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1068 {
1069         int opc, rc = 0;
1070         ENTRY;
1071
1072         /* Check by request opc. */
1073         opc = lustre_msg_get_opc(req->rq_reqmsg);
1074         if (opc == OBD_PING)
1075                 RETURN(1);
1076
1077         /* Perform request specific check. */
1078         if (req->rq_ops && req->rq_ops->hpreq_check)
1079                 rc = req->rq_ops->hpreq_check(req);
1080         RETURN(rc);
1081 }
1082
1083 /** Check if a request is a high priority one. */
1084 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1085                                      struct ptlrpc_request *req)
1086 {
1087         int rc;
1088         ENTRY;
1089
1090         rc = ptlrpc_server_hpreq_check(req);
1091         if (rc < 0)
1092                 RETURN(rc);
1093
1094         spin_lock(&svc->srv_lock);
1095         /* Before inserting the request into the queue, check if it is not
1096          * inserted yet, or even already handled -- it may happen due to
1097          * a racing ldlm_server_blocking_ast(). */
1098         if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1099                 if (rc)
1100                         ptlrpc_hpreq_reorder_nolock(svc, req);
1101                 else
1102                         list_add_tail(&req->rq_list, &svc->srv_request_queue);
1103         }
1104         spin_unlock(&svc->srv_lock);
1105
1106         RETURN(0);
1107 }
1108
1109 /* Only allow normal priority requests on a service that has a high-priority
1110  * queue if forced (i.e. cleanup), if there are other high priority requests
1111  * already being processed (i.e. those threads can service more high-priority
1112  * requests), or if there are enough idle threads that a later thread can do
1113  * a high priority request. */
1114 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1115 {
1116         return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1117                svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1118 }
1119
1120 static struct ptlrpc_request *
1121 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1122 {
1123         struct ptlrpc_request *req = NULL;
1124         ENTRY;
1125
1126         if (ptlrpc_server_allow_normal(svc, force) &&
1127             !list_empty(&svc->srv_request_queue) &&
1128             (list_empty(&svc->srv_request_hpq) ||
1129              svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1130                 req = list_entry(svc->srv_request_queue.next,
1131                                  struct ptlrpc_request, rq_list);
1132                 svc->srv_hpreq_count = 0;
1133         } else if (!list_empty(&svc->srv_request_hpq)) {
1134                 req = list_entry(svc->srv_request_hpq.next,
1135                                  struct ptlrpc_request, rq_list);
1136                 svc->srv_hpreq_count++;
1137         }
1138         RETURN(req);
1139 }
1140
1141 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1142 {
1143         return ((ptlrpc_server_allow_normal(svc, force) &&
1144                  !list_empty(&svc->srv_request_queue)) ||
1145                 !list_empty(&svc->srv_request_hpq));
1146 }
1147
1148 /* Handle freshly incoming reqs, add to timed early reply list,
1149    pass on to regular request queue */
1150 static int
1151 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1152 {
1153         struct ptlrpc_request *req;
1154         __u32                  deadline;
1155         int                    rc;
1156         ENTRY;
1157
1158         LASSERT(svc);
1159
1160         spin_lock(&svc->srv_lock);
1161         if (list_empty(&svc->srv_req_in_queue)) {
1162                 spin_unlock(&svc->srv_lock);
1163                 RETURN(0);
1164         }
1165
1166         req = list_entry(svc->srv_req_in_queue.next,
1167                          struct ptlrpc_request, rq_list);
1168         list_del_init (&req->rq_list);
1169         /* Consider this still a "queued" request as far as stats are
1170            concerned */
1171         spin_unlock(&svc->srv_lock);
1172
1173         /* Clear request swab mask; this is a new request */
1174         req->rq_req_swab_mask = 0;
1175
1176         rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1177         if (rc < 0) {
1178                 CERROR ("error unpacking request: ptl %d from %s"
1179                         " xid "LPU64"\n", svc->srv_req_portal,
1180                         libcfs_id2str(req->rq_peer), req->rq_xid);
1181                 goto err_req;
1182         }
1183
1184         if (rc > 0)
1185                 lustre_set_req_swabbed(req, MSG_PTLRPC_HEADER_OFF);
1186
1187         rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1188         if (rc) {
1189                 CERROR ("error unpacking ptlrpc body: ptl %d from %s"
1190                         " xid "LPU64"\n", svc->srv_req_portal,
1191                         libcfs_id2str(req->rq_peer), req->rq_xid);
1192                 goto err_req;
1193         }
1194
1195         rc = -EINVAL;
1196         if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1197                 CERROR("wrong packet type received (type=%u) from %s\n",
1198                        lustre_msg_get_type(req->rq_reqmsg),
1199                        libcfs_id2str(req->rq_peer));
1200                 goto err_req;
1201         }
1202
1203         CDEBUG(D_NET, "got req "LPU64"\n", req->rq_xid);
1204
1205         req->rq_export = class_conn2export(
1206                 lustre_msg_get_handle(req->rq_reqmsg));
1207         if (req->rq_export) {
1208                 rc = ptlrpc_check_req(req);
1209                 if (rc)
1210                         goto err_req;
1211                 ptlrpc_update_export_timer(req->rq_export, 0);
1212         }
1213
1214         /* req_in handling should/must be fast */
1215         if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1216                 DEBUG_REQ(D_WARNING, req, "Slow req_in handling %lus",
1217                           cfs_time_current_sec() - req->rq_arrival_time.tv_sec);
1218
1219         /* Set rpc server deadline and add it to the timed list */
1220         deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1221                     MSGHDR_AT_SUPPORT) ?
1222                    /* The max time the client expects us to take */
1223                    lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1224         req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1225         if (unlikely(deadline == 0)) {
1226                 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1227                 goto err_req;
1228         }
1229
1230         ptlrpc_at_add_timed(req);
1231         rc = ptlrpc_hpreq_init(svc, req);
1232         if (rc)
1233                 GOTO(err_req, rc);
1234
1235         /* Move it over to the request processing queue */
1236         rc = ptlrpc_server_request_add(svc, req);
1237         if (rc)
1238                 GOTO(err_req, rc);
1239         cfs_waitq_signal(&svc->srv_waitq);
1240         RETURN(1);
1241
1242 err_req:
1243         spin_lock(&svc->srv_lock);
1244         svc->srv_n_queued_reqs--;
1245         svc->srv_n_active_reqs++;
1246         spin_unlock(&svc->srv_lock);
1247         ptlrpc_server_finish_request(req);
1248
1249         RETURN(1);
1250 }
1251
1252 #ifndef noinline
1253 #define noinline __attribute__((noinline))
1254 #endif
1255
1256 /*
1257  * The sole purpose of these functions is to avoid unreasonable stack frame
1258  * sizes such as assigned by the gcc compiler. Should NOT be inlined.
1259  */
1260 static void noinline
1261 ptlrpc_server_log_handling_request(struct ptlrpc_request *request)
1262 {
1263         CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1264                "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1265                (request->rq_export ?
1266                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1267                (request->rq_export ?
1268                 atomic_read(&request->rq_export->exp_refcount) : -99),
1269                lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1270                libcfs_id2str(request->rq_peer),
1271                lustre_msg_get_opc(request->rq_reqmsg));
1272 }
1273
1274 static void noinline
1275 ptlrpc_server_log_handled_request(struct ptlrpc_request *request,
1276                                   long timediff,
1277                                   struct timeval *work_end)
1278 {
1279         CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1280                "%s:%s+%d:%d:x"LPU64":%s:%d Request procesed in "
1281                "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1282                cfs_curproc_comm(),
1283                (request->rq_export ?
1284                 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1285                (request->rq_export ?
1286                 atomic_read(&request->rq_export->exp_refcount) : -99),
1287                lustre_msg_get_status(request->rq_reqmsg),
1288                request->rq_xid,
1289                libcfs_id2str(request->rq_peer),
1290                lustre_msg_get_opc(request->rq_reqmsg),
1291                timediff,
1292                cfs_timeval_sub(work_end, &request->rq_arrival_time, NULL),
1293                (request->rq_repmsg ?
1294                 lustre_msg_get_transno(request->rq_repmsg) :
1295                 request->rq_transno),
1296                request->rq_status,
1297                (request->rq_repmsg ?
1298                 lustre_msg_get_status(request->rq_repmsg) : -999));
1299 }
1300
1301 static int
1302 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1303                              struct ptlrpc_thread *thread)
1304 {
1305         struct obd_export     *export = NULL;
1306         struct ptlrpc_request *request;
1307         struct timeval         work_start;
1308         struct timeval         work_end;
1309         long                   timediff;
1310         int                    opc, rc;
1311         int                    fail_opc = 0;
1312         ENTRY;
1313
1314         LASSERT(svc);
1315
1316         spin_lock(&svc->srv_lock);
1317         if (!ptlrpc_server_request_pending(svc, 0) ||
1318             (
1319 #ifndef __KERNEL__
1320              /* !@%$# liblustre only has 1 thread */
1321              svc->srv_n_difficult_replies != 0 &&
1322 #endif
1323              svc->srv_n_active_reqs >= (svc->srv_threads_running - 1))) {
1324                 /* Don't handle regular requests in the last thread, in order
1325                  * to handle difficult replies (which might block other threads)
1326                  * as well as handle any incoming reqs, early replies, etc.
1327                  * That means we always need at least 2 service threads. */
1328                 spin_unlock(&svc->srv_lock);
1329                 RETURN(0);
1330         }
1331
1332         request = ptlrpc_server_request_get(svc, 0);
1333         if  (request == NULL) {
1334                 spin_unlock(&svc->srv_lock);
1335                 RETURN(0);
1336         }
1337
1338         opc = lustre_msg_get_opc(request->rq_reqmsg);
1339         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1340                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1341         else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1342                 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1343
1344         if (unlikely(fail_opc)) {
1345                 if (request->rq_export && request->rq_ops) {
1346                         spin_unlock(&svc->srv_lock);
1347                         OBD_FAIL_TIMEOUT(fail_opc, 4);
1348                         spin_lock(&svc->srv_lock);
1349                         request = ptlrpc_server_request_get(svc, 0);
1350                         if  (request == NULL) {
1351                                 spin_unlock(&svc->srv_lock);
1352                                 RETURN(0);
1353                         }
1354                         LASSERT(ptlrpc_server_request_pending(svc, 0));
1355                 }
1356         }
1357
1358         list_del_init(&request->rq_list);
1359         svc->srv_n_queued_reqs--;
1360         svc->srv_n_active_reqs++;
1361
1362         if (request->rq_hp)
1363                 svc->srv_n_hpreq++;
1364
1365         /* The phase is changed under the lock here because we need to know
1366          * the request is under processing (see ptlrpc_hpreq_reorder()). */
1367         ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1368         spin_unlock(&svc->srv_lock);
1369
1370         ptlrpc_hpreq_fini(request);
1371
1372         if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1373                 libcfs_debug_dumplog();
1374
1375         do_gettimeofday(&work_start);
1376         timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1377         if (svc->srv_stats != NULL) {
1378                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1379                                     timediff);
1380                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1381                                     svc->srv_n_queued_reqs);
1382                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1383                                     svc->srv_n_active_reqs);
1384                 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1385                                     at_get(&svc->srv_at_estimate));
1386         }
1387
1388         CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1389
1390         request->rq_svc_thread = thread;
1391         if (request->rq_export) {
1392                 if (ptlrpc_check_req(request))
1393                         goto put_conn;
1394                 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1395                 export = class_export_rpc_get(request->rq_export);
1396         }
1397
1398         /* Discard requests queued for longer than the deadline.
1399            The deadline is increased if we send an early reply. */
1400         if (cfs_time_current_sec() > request->rq_deadline) {
1401                 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1402                           ": deadline %ld%+lds ago\n",
1403                           libcfs_id2str(request->rq_peer),
1404                           request->rq_deadline -
1405                           request->rq_arrival_time.tv_sec,
1406                           cfs_time_current_sec() - request->rq_deadline);
1407                 goto put_rpc_export;
1408         }
1409
1410         ptlrpc_server_log_handling_request(request);
1411
1412         if (lustre_msg_get_opc(request->rq_reqmsg) != OBD_PING)
1413                 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1414
1415         rc = svc->srv_handler(request);
1416
1417         ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1418
1419 put_rpc_export:
1420         if (export != NULL && !request->rq_copy_queued)
1421                 class_export_rpc_put(export);
1422
1423 put_conn:
1424         if (cfs_time_current_sec() > request->rq_deadline) {
1425                 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1426                           "than estimated (%ld%+lds); client may timeout.",
1427                           request->rq_xid, request->rq_deadline -
1428                           request->rq_arrival_time.tv_sec,
1429                           cfs_time_current_sec() - request->rq_deadline);
1430         }
1431
1432         do_gettimeofday(&work_end);
1433         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1434
1435         ptlrpc_server_log_handled_request(request, timediff, &work_end);
1436
1437         if (svc->srv_stats != NULL) {
1438                 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1439                 int opc = opcode_offset(op);
1440                 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1441                         LASSERT(opc < LUSTRE_MAX_OPCODES);
1442                         lprocfs_counter_add(svc->srv_stats,
1443                                             opc + EXTRA_MAX_OPCODES,
1444                                             timediff);
1445                 }
1446         }
1447         if (request->rq_early_count) {
1448                 DEBUG_REQ(D_ADAPTTO, request,
1449                           "sent %d early replies before finishing in %lds",
1450                           request->rq_early_count,
1451                           work_end.tv_sec - request->rq_arrival_time.tv_sec);
1452         }
1453
1454         spin_lock(&svc->srv_lock);
1455         if (request->rq_hp)
1456                 svc->srv_n_hpreq--;
1457         spin_unlock(&svc->srv_lock);
1458         ptlrpc_server_finish_request(request);
1459
1460         RETURN(1);
1461 }
1462
1463 static int
1464 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1465 {
1466         struct ptlrpc_reply_state *rs;
1467         struct obd_export         *exp;
1468         struct obd_device         *obd;
1469         int                        nlocks;
1470         int                        been_handled;
1471         ENTRY;
1472
1473         spin_lock(&svc->srv_lock);
1474         if (list_empty (&svc->srv_reply_queue)) {
1475                 spin_unlock(&svc->srv_lock);
1476                 RETURN(0);
1477         }
1478
1479         rs = list_entry (svc->srv_reply_queue.next,
1480                          struct ptlrpc_reply_state, rs_list);
1481
1482         exp = rs->rs_export;
1483         obd = exp->exp_obd;
1484
1485         LASSERT (rs->rs_difficult);
1486         LASSERT (rs->rs_scheduled);
1487
1488         list_del_init (&rs->rs_list);
1489
1490         /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1491         spin_unlock(&svc->srv_lock);
1492
1493         spin_lock (&exp->exp_uncommitted_replies_lock);
1494         /* Noop if removed already */
1495         list_del_init (&rs->rs_obd_list);
1496         spin_unlock (&exp->exp_uncommitted_replies_lock);
1497
1498         spin_lock (&exp->exp_lock);
1499         /* Noop if removed already */
1500         list_del_init (&rs->rs_exp_list);
1501         spin_unlock (&exp->exp_lock);
1502
1503         spin_lock(&svc->srv_lock);
1504
1505         been_handled = rs->rs_handled;
1506         rs->rs_handled = 1;
1507
1508         nlocks = rs->rs_nlocks;                 /* atomic "steal", but */
1509         rs->rs_nlocks = 0;                      /* locks still on rs_locks! */
1510
1511         if (nlocks == 0 && !been_handled) {
1512                 /* If we see this, we should already have seen the warning
1513                  * in mds_steal_ack_locks()  */
1514                 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1515                       " o%d NID %s\n", rs, rs->rs_xid, rs->rs_transno,
1516                       lustre_msg_get_opc(rs->rs_msg),
1517                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1518         }
1519
1520         if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1521                 spin_unlock(&svc->srv_lock);
1522
1523                 if (!been_handled && rs->rs_on_net) {
1524                         LNetMDUnlink(rs->rs_md_h);
1525                         /* Ignore return code; we're racing with
1526                          * completion... */
1527                 }
1528
1529                 while (nlocks-- > 0)
1530                         ldlm_lock_decref(&rs->rs_locks[nlocks],
1531                                          rs->rs_modes[nlocks]);
1532
1533                 spin_lock(&svc->srv_lock);
1534         }
1535
1536         rs->rs_scheduled = 0;
1537
1538         if (!rs->rs_on_net) {
1539                 /* Off the net */
1540                 svc->srv_n_difficult_replies--;
1541                 spin_unlock(&svc->srv_lock);
1542
1543                 class_export_put (exp);
1544                 rs->rs_export = NULL;
1545                 ptlrpc_rs_decref (rs);
1546                 atomic_dec (&svc->srv_outstanding_replies);
1547                 RETURN(1);
1548         }
1549
1550         /* still on the net; callback will schedule */
1551         spin_unlock(&svc->srv_lock);
1552         RETURN(1);
1553 }
1554
1555 #ifndef __KERNEL__
1556 /* FIXME make use of timeout later */
1557 int
1558 liblustre_check_services (void *arg)
1559 {
1560         int  did_something = 0;
1561         int  rc;
1562         struct list_head *tmp, *nxt;
1563         ENTRY;
1564
1565         /* I'm relying on being single threaded, not to have to lock
1566          * ptlrpc_all_services etc */
1567         list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1568                 struct ptlrpc_service *svc =
1569                         list_entry (tmp, struct ptlrpc_service, srv_list);
1570
1571                 if (svc->srv_threads_running != 0)     /* I've recursed */
1572                         continue;
1573
1574                 /* service threads can block for bulk, so this limits us
1575                  * (arbitrarily) to recursing 1 stack frame per service.
1576                  * Note that the problem with recursion is that we have to
1577                  * unwind completely before our caller can resume. */
1578
1579                 svc->srv_threads_running++;
1580
1581                 do {
1582                         rc = ptlrpc_server_handle_req_in(svc);
1583                         rc |= ptlrpc_server_handle_reply(svc);
1584                         rc |= ptlrpc_at_check_timed(svc);
1585                         rc |= ptlrpc_server_handle_request(svc, NULL);
1586                         rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1587                         did_something |= rc;
1588                 } while (rc);
1589
1590                 svc->srv_threads_running--;
1591         }
1592
1593         RETURN(did_something);
1594 }
1595 #define ptlrpc_stop_all_threads(s) do {} while (0)
1596
1597 #else /* __KERNEL__ */
1598
1599 static void
1600 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1601 {
1602         int avail = svc->srv_nrqbd_receiving;
1603         int low_water = test_req_buffer_pressure ? 0 :
1604                         svc->srv_nbuf_per_group/2;
1605
1606         /* NB I'm not locking; just looking. */
1607
1608         /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1609          * allowed the request history to grow out of control.  We could put a
1610          * sanity check on that here and cull some history if we need the
1611          * space. */
1612
1613         if (avail <= low_water)
1614                 ptlrpc_grow_req_bufs(svc);
1615
1616         if (svc->srv_stats)
1617                 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1618                                     avail);
1619 }
1620
1621 static int
1622 ptlrpc_retry_rqbds(void *arg)
1623 {
1624         struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1625
1626         svc->srv_rqbd_timeout = 0;
1627         return (-ETIMEDOUT);
1628 }
1629
1630 static int ptlrpc_main(void *arg)
1631 {
1632         struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1633         struct ptlrpc_service  *svc = data->svc;
1634         struct ptlrpc_thread   *thread = data->thread;
1635         struct obd_device      *dev = data->dev;
1636         struct ptlrpc_reply_state *rs;
1637 #ifdef WITH_GROUP_INFO
1638         struct group_info *ginfo = NULL;
1639 #endif
1640         int counter = 0, rc = 0;
1641         ENTRY;
1642
1643         cfs_daemonize_ctxt(data->name);
1644
1645 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1646         /* we need to do this before any per-thread allocation is done so that
1647          * we get the per-thread allocations on local node.  bug 7342 */
1648         if (svc->srv_cpu_affinity) {
1649                 int cpu, num_cpu;
1650
1651                 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1652                         if (!cpu_online(cpu))
1653                                 continue;
1654                         if (num_cpu == thread->t_id % num_online_cpus())
1655                                 break;
1656                         num_cpu++;
1657                 }
1658                 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1659         }
1660 #endif
1661
1662 #ifdef WITH_GROUP_INFO
1663         ginfo = groups_alloc(0);
1664         if (!ginfo) {
1665                 rc = -ENOMEM;
1666                 goto out;
1667         }
1668
1669         set_current_groups(ginfo);
1670         put_group_info(ginfo);
1671 #endif
1672
1673         if (svc->srv_init != NULL) {
1674                 rc = svc->srv_init(thread);
1675                 if (rc)
1676                         goto out;
1677         }
1678
1679         /* Alloc reply state structure for this one */
1680         OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1681         if (!rs) {
1682                 rc = -ENOMEM;
1683                 goto out_srv_init;
1684         }
1685
1686         spin_lock(&svc->srv_lock);
1687         /* SVC_STOPPING may already be set here if someone else is trying
1688          * to stop the service while this new thread has been dynamically
1689          * forked. We still set SVC_RUNNING to let our creator know that
1690          * we are now running, however we will exit as soon as possible */
1691         thread->t_flags |= SVC_RUNNING;
1692         spin_unlock(&svc->srv_lock);
1693
1694         /*
1695          * wake up our creator. Note: @data is invalid after this point,
1696          * because it's allocated on ptlrpc_start_thread() stack.
1697          */
1698         cfs_waitq_signal(&thread->t_ctl_waitq);
1699
1700         thread->t_watchdog = lc_watchdog_add(GET_TIMEOUT(svc), NULL, NULL);
1701
1702         spin_lock(&svc->srv_lock);
1703         svc->srv_threads_running++;
1704         list_add(&rs->rs_list, &svc->srv_free_rs_list);
1705         spin_unlock(&svc->srv_lock);
1706         cfs_waitq_signal(&svc->srv_free_rs_waitq);
1707
1708         CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1709                svc->srv_threads_running);
1710
1711         /* XXX maintain a list of all managed devices: insert here */
1712
1713         while ((thread->t_flags & SVC_STOPPING) == 0 ||
1714                svc->srv_n_difficult_replies != 0) {
1715                 /* Don't exit while there are replies to be handled */
1716                 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1717                                                      ptlrpc_retry_rqbds, svc);
1718
1719                 lc_watchdog_disable(thread->t_watchdog);
1720
1721                 cfs_cond_resched();
1722
1723                 l_wait_event_exclusive (svc->srv_waitq,
1724                               ((thread->t_flags & SVC_STOPPING) != 0 &&
1725                                svc->srv_n_difficult_replies == 0) ||
1726                               (!list_empty(&svc->srv_idle_rqbds) &&
1727                                svc->srv_rqbd_timeout == 0) ||
1728                               !list_empty(&svc->srv_req_in_queue) ||
1729                               !list_empty(&svc->srv_reply_queue) ||
1730                               (ptlrpc_server_request_pending(svc, 0) &&
1731                                (svc->srv_n_active_reqs <
1732                                 (svc->srv_threads_running - 1))) ||
1733                               svc->srv_at_check,
1734                               &lwi);
1735
1736                 lc_watchdog_touch(thread->t_watchdog, GET_TIMEOUT(svc));
1737
1738                 ptlrpc_check_rqbd_pool(svc);
1739
1740                 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1741                     (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1742                         /* Ignore return code - we tried... */
1743                         ptlrpc_start_thread(dev, svc);
1744                 }
1745
1746                 if (!list_empty(&svc->srv_reply_queue))
1747                         ptlrpc_server_handle_reply (svc);
1748
1749                 if (!list_empty(&svc->srv_req_in_queue)) {
1750                         /* Process all incoming reqs before handling any */
1751                         ptlrpc_server_handle_req_in(svc);
1752                         /* but limit ourselves in case of flood */
1753                         if (counter++ < 1000)
1754                                 continue;
1755                         counter = 0;
1756                 }
1757
1758                 if (svc->srv_at_check)
1759                         ptlrpc_at_check_timed(svc);
1760
1761                 /* don't handle requests in the last thread */
1762                 if (ptlrpc_server_request_pending(svc, 0) &&
1763                     (svc->srv_n_active_reqs < (svc->srv_threads_running - 1)))
1764                         ptlrpc_server_handle_request(svc, thread);
1765
1766                 if (!list_empty(&svc->srv_idle_rqbds) &&
1767                     ptlrpc_server_post_idle_rqbds(svc) < 0) {
1768                         /* I just failed to repost request buffers.  Wait
1769                          * for a timeout (unless something else happens)
1770                          * before I try again */
1771                         svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1772                         CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1773                                svc->srv_nrqbd_receiving);
1774                 }
1775         }
1776
1777         lc_watchdog_delete(thread->t_watchdog);
1778         thread->t_watchdog = NULL;
1779
1780 out_srv_init:
1781         /*
1782          * deconstruct service specific state created by ptlrpc_start_thread()
1783          */
1784         if (svc->srv_done != NULL)
1785                 svc->srv_done(thread);
1786
1787 out:
1788         CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1789
1790         spin_lock(&svc->srv_lock);
1791         svc->srv_threads_running--;              /* must know immediately */
1792         thread->t_id = rc;
1793         thread->t_flags = SVC_STOPPED;
1794
1795         cfs_waitq_signal(&thread->t_ctl_waitq);
1796         spin_unlock(&svc->srv_lock);
1797
1798         return rc;
1799 }
1800
1801 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1802                                struct ptlrpc_thread *thread)
1803 {
1804         struct l_wait_info lwi = { 0 };
1805
1806         spin_lock(&svc->srv_lock);
1807         /* let the thread know that we would like it to stop asap */
1808         thread->t_flags |= SVC_STOPPING;
1809         spin_unlock(&svc->srv_lock);
1810
1811         cfs_waitq_broadcast(&svc->srv_waitq);
1812         l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1813                      &lwi);
1814
1815         spin_lock(&svc->srv_lock);
1816         list_del(&thread->t_link);
1817         spin_unlock(&svc->srv_lock);
1818
1819         OBD_FREE(thread, sizeof(*thread));
1820 }
1821
1822 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1823 {
1824         struct ptlrpc_thread *thread;
1825
1826         spin_lock(&svc->srv_lock);
1827         while (!list_empty(&svc->srv_threads)) {
1828                 thread = list_entry(svc->srv_threads.next,
1829                                     struct ptlrpc_thread, t_link);
1830
1831                 spin_unlock(&svc->srv_lock);
1832                 ptlrpc_stop_thread(svc, thread);
1833                 spin_lock(&svc->srv_lock);
1834         }
1835
1836         spin_unlock(&svc->srv_lock);
1837 }
1838
1839 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1840 {
1841         int i, rc = 0;
1842         ENTRY;
1843
1844         /* We require 2 threads min - see note in
1845          * ptlrpc_server_handle_request() */
1846
1847         LASSERT(svc->srv_threads_min >= 2);
1848         for (i = 0; i < svc->srv_threads_min; i++) {
1849                 rc = ptlrpc_start_thread(dev, svc);
1850                 /* We have enough threads, don't start more.  b=15759 */
1851                 if (rc == -EMFILE)
1852                         break;
1853                 if (rc) {
1854                         CERROR("cannot start %s thread #%d: rc %d\n",
1855                                svc->srv_thread_name, i, rc);
1856                         ptlrpc_stop_all_threads(svc);
1857                 }
1858         }
1859         RETURN(rc);
1860 }
1861
1862 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1863 {
1864         struct l_wait_info lwi = { 0 };
1865         struct ptlrpc_svc_data d;
1866         struct ptlrpc_thread *thread;
1867         char name[32];
1868         int id, rc;
1869         ENTRY;
1870
1871         CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1872                svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1873                svc->srv_threads_max, svc->srv_threads_running);
1874         if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1875             (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1876              svc->srv_threads_started == svc->srv_threads_min - 1))
1877                 RETURN(-EMFILE);
1878
1879         OBD_ALLOC(thread, sizeof(*thread));
1880         if (thread == NULL)
1881                 RETURN(-ENOMEM);
1882         cfs_waitq_init(&thread->t_ctl_waitq);
1883
1884         spin_lock(&svc->srv_lock);
1885         if (svc->srv_threads_started >= svc->srv_threads_max) {
1886                 spin_unlock(&svc->srv_lock);
1887                 OBD_FREE(thread, sizeof(*thread));
1888                 RETURN(-EMFILE);
1889         }
1890         list_add(&thread->t_link, &svc->srv_threads);
1891         id = svc->srv_threads_started++;
1892         spin_unlock(&svc->srv_lock);
1893
1894         thread->t_svc = svc;
1895         thread->t_id = id;
1896         sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1897         d.dev = dev;
1898         d.svc = svc;
1899         d.name = name;
1900         d.thread = thread;
1901
1902         CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1903
1904         /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1905          * just drop the VM and FILES in cfs_daemonize_ctxt() right away.
1906          */
1907         rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1908         if (rc < 0) {
1909                 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1910
1911                 spin_lock(&svc->srv_lock);
1912                 list_del(&thread->t_link);
1913                 --svc->srv_threads_started;
1914                 spin_unlock(&svc->srv_lock);
1915
1916                 OBD_FREE(thread, sizeof(*thread));
1917                 RETURN(rc);
1918         }
1919         l_wait_event(thread->t_ctl_waitq,
1920                      thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1921
1922         rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1923         RETURN(rc);
1924 }
1925 #endif
1926
1927 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1928 {
1929         int                   rc;
1930         struct l_wait_info    lwi;
1931         struct list_head     *tmp;
1932         struct ptlrpc_reply_state *rs, *t;
1933         struct ptlrpc_at_array *array = &service->srv_at_array;
1934
1935         cfs_timer_disarm(&service->srv_at_timer);
1936
1937         ptlrpc_stop_all_threads(service);
1938         LASSERT(list_empty(&service->srv_threads));
1939
1940         spin_lock (&ptlrpc_all_services_lock);
1941         list_del_init (&service->srv_list);
1942         spin_unlock (&ptlrpc_all_services_lock);
1943
1944         ptlrpc_lprocfs_unregister_service(service);
1945
1946         /* All history will be culled when the next request buffer is
1947          * freed */
1948         service->srv_max_history_rqbds = 0;
1949
1950         CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1951
1952         rc = LNetClearLazyPortal(service->srv_req_portal);
1953         LASSERT (rc == 0);
1954
1955         /* Unlink all the request buffers.  This forces a 'final' event with
1956          * its 'unlink' flag set for each posted rqbd */
1957         list_for_each(tmp, &service->srv_active_rqbds) {
1958                 struct ptlrpc_request_buffer_desc *rqbd =
1959                         list_entry(tmp, struct ptlrpc_request_buffer_desc,
1960                                    rqbd_list);
1961
1962                 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1963                 LASSERT (rc == 0 || rc == -ENOENT);
1964         }
1965
1966         /* Wait for the network to release any buffers it's currently
1967          * filling */
1968         for (;;) {
1969                 spin_lock(&service->srv_lock);
1970                 rc = service->srv_nrqbd_receiving;
1971                 spin_unlock(&service->srv_lock);
1972
1973                 if (rc == 0)
1974                         break;
1975
1976                 /* Network access will complete in finite time but the HUGE
1977                  * timeout lets us CWARN for visibility of sluggish NALs */
1978                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1979                                            cfs_time_seconds(1), NULL, NULL);
1980                 rc = l_wait_event(service->srv_waitq,
1981                                   service->srv_nrqbd_receiving == 0,
1982                                   &lwi);
1983                 if (rc == -ETIMEDOUT)
1984                         CWARN("Service %s waiting for request buffers\n",
1985                               service->srv_name);
1986         }
1987
1988         /* schedule all outstanding replies to terminate them */
1989         spin_lock(&service->srv_lock);
1990         while (!list_empty(&service->srv_active_replies)) {
1991                 struct ptlrpc_reply_state *rs =
1992                         list_entry(service->srv_active_replies.next,
1993                                    struct ptlrpc_reply_state, rs_list);
1994                 ptlrpc_schedule_difficult_reply(rs);
1995         }
1996         spin_unlock(&service->srv_lock);
1997
1998         /* purge the request queue.  NB No new replies (rqbds all unlinked)
1999          * and no service threads, so I'm the only thread noodling the
2000          * request queue now */
2001         while (!list_empty(&service->srv_req_in_queue)) {
2002                 struct ptlrpc_request *req =
2003                         list_entry(service->srv_req_in_queue.next,
2004                                    struct ptlrpc_request,
2005                                    rq_list);
2006
2007                 list_del(&req->rq_list);
2008                 service->srv_n_queued_reqs--;
2009                 service->srv_n_active_reqs++;
2010                 ptlrpc_server_finish_request(req);
2011         }
2012         while (ptlrpc_server_request_pending(service, 1)) {
2013                 struct ptlrpc_request *req;
2014
2015                 req = ptlrpc_server_request_get(service, 1);
2016                 list_del(&req->rq_list);
2017                 service->srv_n_queued_reqs--;
2018                 service->srv_n_active_reqs++;
2019                 ptlrpc_hpreq_fini(req);
2020                 ptlrpc_server_finish_request(req);
2021         }
2022         LASSERT(service->srv_n_queued_reqs == 0);
2023         LASSERT(service->srv_n_active_reqs == 0);
2024         LASSERT(service->srv_n_history_rqbds == 0);
2025         LASSERT(list_empty(&service->srv_active_rqbds));
2026
2027         /* Now free all the request buffers since nothing references them
2028          * any more... */
2029         while (!list_empty(&service->srv_idle_rqbds)) {
2030                 struct ptlrpc_request_buffer_desc *rqbd =
2031                         list_entry(service->srv_idle_rqbds.next,
2032                                    struct ptlrpc_request_buffer_desc,
2033                                    rqbd_list);
2034
2035                 ptlrpc_free_rqbd(rqbd);
2036         }
2037
2038         /* wait for all outstanding replies to complete (they were
2039          * scheduled having been flagged to abort above) */
2040         while (atomic_read(&service->srv_outstanding_replies) != 0) {
2041                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
2042
2043                 rc = l_wait_event(service->srv_waitq,
2044                                   !list_empty(&service->srv_reply_queue), &lwi);
2045                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2046
2047                 if (rc == 0) {
2048                         ptlrpc_server_handle_reply(service);
2049                         continue;
2050                 }
2051                 CWARN("Unexpectedly long timeout %p\n", service);
2052         }
2053
2054         list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
2055                 list_del(&rs->rs_list);
2056                 OBD_FREE(rs, service->srv_max_reply_size);
2057         }
2058
2059         /* In case somebody rearmed this in the meantime */
2060         cfs_timer_disarm(&service->srv_at_timer);
2061
2062         if (array->paa_reqs_array != NULL) {
2063                 OBD_FREE(array->paa_reqs_array,
2064                          sizeof(struct list_head) * array->paa_size);
2065                 array->paa_reqs_array = NULL;
2066         }
2067
2068         if (array->paa_reqs_count != NULL) {
2069                 OBD_FREE(array->paa_reqs_count,
2070                          sizeof(__u32) * array->paa_size);
2071                 array->paa_reqs_count= NULL;
2072         }
2073
2074         OBD_FREE(service, sizeof(*service));
2075         return 0;
2076 }
2077
2078 /* Returns 0 if the service is healthy.
2079  *
2080  * Right now, it just checks to make sure that requests aren't languishing
2081  * in the queue.  We'll use this health check to govern whether a node needs
2082  * to be shot, so it's intentionally non-aggressive. */
2083 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2084 {
2085         struct ptlrpc_request *request;
2086         struct timeval         right_now;
2087         long                   timediff;
2088
2089         if (svc == NULL)
2090                 return 0;
2091
2092         do_gettimeofday(&right_now);
2093
2094         spin_lock(&svc->srv_lock);
2095         if (!ptlrpc_server_request_pending(svc, 1)) {
2096                 spin_unlock(&svc->srv_lock);
2097                 return 0;
2098         }
2099
2100         /* How long has the next entry been waiting? */
2101         if (list_empty(&svc->srv_request_queue))
2102                 request = list_entry(svc->srv_request_hpq.next,
2103                                      struct ptlrpc_request, rq_list);
2104         else
2105                 request = list_entry(svc->srv_request_queue.next,
2106                                      struct ptlrpc_request, rq_list);
2107         timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2108         spin_unlock(&svc->srv_lock);
2109
2110         if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2111                                         at_max)) {
2112                 CERROR("%s: unhealthy - request has been waiting %lds\n",
2113                        svc->srv_name, timediff / ONE_MILLION);
2114                 return (-1);
2115         }
2116
2117         return 0;
2118 }