lustre/ptlrpc/nrs_tbf.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (C) 2013 DataDirect Networks, Inc.
  24  *
  25  * Copyright (c) 2014, Intel Corporation.
  26  */
  27 /*
  28  * lustre/ptlrpc/nrs_tbf.c
  29  *
  30  * Network Request Scheduler (NRS) Token Bucket Filter(TBF) policy
  31  *
  32  */
  33
  34 #ifdef HAVE_SERVER_SUPPORT
  35
  36 /**
  37  * \addtogoup nrs
  38  * @{
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42 #include <obd_support.h>
  43 #include <obd_class.h>
  44 #include <libcfs/libcfs.h>
  45 #include "ptlrpc_internal.h"
  46
  47 /**
  48  * \name tbf
  49  *
  50  * Token Bucket Filter over client NIDs
  51  *
  52  * @{
  53  */
  54
  55 #define NRS_POL_NAME_TBF        "tbf"
  56
  57 static int tbf_jobid_cache_size = 8192;
  58 CFS_MODULE_PARM(tbf_jobid_cache_size, "i", int, 0644,
  59                 "The size of jobid cache");
  60
  61 static int tbf_rate = 10000;
  62 CFS_MODULE_PARM(tbf_rate, "i", int, 0644,
  63                 "Default rate limit in RPCs/s");
  64
  65 static int tbf_depth = 3;
  66 CFS_MODULE_PARM(tbf_depth, "i", int, 0644,
  67                 "How many tokens that a client can save up");
  68
  69 static enum hrtimer_restart nrs_tbf_timer_cb(struct hrtimer *timer)
  70 {
  71         struct nrs_tbf_head *head = container_of(timer, struct nrs_tbf_head,
  72                                                  th_timer);
  73         struct ptlrpc_nrs   *nrs = head->th_res.res_policy->pol_nrs;
  74         struct ptlrpc_service_part *svcpt = nrs->nrs_svcpt;
  75
  76         nrs->nrs_throttling = 0;
  77         wake_up(&svcpt->scp_waitq);
  78
  79         return HRTIMER_NORESTART;
  80 }
  81
  82 #define NRS_TBF_DEFAULT_RULE "default"
  83
  84 static void nrs_tbf_rule_fini(struct nrs_tbf_rule *rule)
  85 {
  86         LASSERT(atomic_read(&rule->tr_ref) == 0);
  87         LASSERT(list_empty(&rule->tr_cli_list));
  88         LASSERT(list_empty(&rule->tr_linkage));
  89
  90         rule->tr_head->th_ops->o_rule_fini(rule);
  91         OBD_FREE_PTR(rule);
  92 }
  93
  94 /**
  95  * Decreases the rule's usage reference count, and stops the rule in case it
  96  * was already stopping and have no more outstanding usage references (which
  97  * indicates it has no more queued or started requests, and can be safely
  98  * stopped).
  99  */
 100 static void nrs_tbf_rule_put(struct nrs_tbf_rule *rule)
 101 {
 102         if (atomic_dec_and_test(&rule->tr_ref))
 103                 nrs_tbf_rule_fini(rule);
 104 }
 105
 106 /**
 107  * Increases the rule's usage reference count.
 108  */
 109 static inline void nrs_tbf_rule_get(struct nrs_tbf_rule *rule)
 110 {
 111         atomic_inc(&rule->tr_ref);
 112 }
 113
 114 static void
 115 nrs_tbf_cli_rule_put(struct nrs_tbf_client *cli)
 116 {
 117         LASSERT(!list_empty(&cli->tc_linkage));
 118         LASSERT(cli->tc_rule);
 119         spin_lock(&cli->tc_rule->tr_rule_lock);
 120         list_del_init(&cli->tc_linkage);
 121         spin_unlock(&cli->tc_rule->tr_rule_lock);
 122         nrs_tbf_rule_put(cli->tc_rule);
 123         cli->tc_rule = NULL;
 124 }
 125
 126 static void
 127 nrs_tbf_cli_reset_value(struct nrs_tbf_head *head,
 128                         struct nrs_tbf_client *cli)
 129
 130 {
 131         struct nrs_tbf_rule *rule = cli->tc_rule;
 132
 133         cli->tc_rpc_rate = rule->tr_rpc_rate;
 134         cli->tc_nsecs = rule->tr_nsecs;
 135         cli->tc_depth = rule->tr_depth;
 136         cli->tc_ntoken = rule->tr_depth;
 137         cli->tc_check_time = ktime_to_ns(ktime_get());
 138         cli->tc_rule_sequence = atomic_read(&head->th_rule_sequence);
 139         cli->tc_rule_generation = rule->tr_generation;
 140
 141         if (cli->tc_in_heap)
 142                 cfs_binheap_relocate(head->th_binheap,
 143                                      &cli->tc_node);
 144 }
 145
 146 static void
 147 nrs_tbf_cli_reset(struct nrs_tbf_head *head,
 148                   struct nrs_tbf_rule *rule,
 149                   struct nrs_tbf_client *cli)
 150 {
 151         spin_lock(&cli->tc_rule_lock);
 152         if (cli->tc_rule != NULL && !list_empty(&cli->tc_linkage)) {
 153                 LASSERT(rule != cli->tc_rule);
 154                 nrs_tbf_cli_rule_put(cli);
 155         }
 156         LASSERT(cli->tc_rule == NULL);
 157         LASSERT(list_empty(&cli->tc_linkage));
 158         /* Rule's ref is added before called */
 159         cli->tc_rule = rule;
 160         spin_lock(&rule->tr_rule_lock);
 161         list_add_tail(&cli->tc_linkage, &rule->tr_cli_list);
 162         spin_unlock(&rule->tr_rule_lock);
 163         spin_unlock(&cli->tc_rule_lock);
 164         nrs_tbf_cli_reset_value(head, cli);
 165 }
 166
 167 static int
 168 nrs_tbf_rule_dump(struct nrs_tbf_rule *rule, struct seq_file *m)
 169 {
 170         return rule->tr_head->th_ops->o_rule_dump(rule, m);
 171 }
 172
 173 static int
 174 nrs_tbf_rule_dump_all(struct nrs_tbf_head *head, struct seq_file *m)
 175 {
 176         struct nrs_tbf_rule *rule;
 177         int rc = 0;
 178
 179         LASSERT(head != NULL);
 180         spin_lock(&head->th_rule_lock);
 181         /* List the rules from newest to oldest */
 182         list_for_each_entry(rule, &head->th_list, tr_linkage) {
 183                 LASSERT((rule->tr_flags & NTRS_STOPPING) == 0);
 184                 rc = nrs_tbf_rule_dump(rule, m);
 185                 if (rc) {
 186                         rc = -ENOSPC;
 187                         break;
 188                 }
 189         }
 190         spin_unlock(&head->th_rule_lock);
 191
 192         return rc;
 193 }
 194
 195 static struct nrs_tbf_rule *
 196 nrs_tbf_rule_find_nolock(struct nrs_tbf_head *head,
 197                          const char *name)
 198 {
 199         struct nrs_tbf_rule *rule;
 200
 201         LASSERT(head != NULL);
 202         list_for_each_entry(rule, &head->th_list, tr_linkage) {
 203                 LASSERT((rule->tr_flags & NTRS_STOPPING) == 0);
 204                 if (strcmp(rule->tr_name, name) == 0) {
 205                         nrs_tbf_rule_get(rule);
 206                         return rule;
 207                 }
 208         }
 209         return NULL;
 210 }
 211
 212 static struct nrs_tbf_rule *
 213 nrs_tbf_rule_find(struct nrs_tbf_head *head,
 214                   const char *name)
 215 {
 216         struct nrs_tbf_rule *rule;
 217
 218         LASSERT(head != NULL);
 219         spin_lock(&head->th_rule_lock);
 220         rule = nrs_tbf_rule_find_nolock(head, name);
 221         spin_unlock(&head->th_rule_lock);
 222         return rule;
 223 }
 224
 225 static struct nrs_tbf_rule *
 226 nrs_tbf_rule_match(struct nrs_tbf_head *head,
 227                    struct nrs_tbf_client *cli)
 228 {
 229         struct nrs_tbf_rule *rule = NULL;
 230         struct nrs_tbf_rule *tmp_rule;
 231
 232         spin_lock(&head->th_rule_lock);
 233         /* Match the newest rule in the list */
 234         list_for_each_entry(tmp_rule, &head->th_list, tr_linkage) {
 235                 LASSERT((tmp_rule->tr_flags & NTRS_STOPPING) == 0);
 236                 if (head->th_ops->o_rule_match(tmp_rule, cli)) {
 237                         rule = tmp_rule;
 238                         break;
 239                 }
 240         }
 241
 242         if (rule == NULL)
 243                 rule = head->th_rule;
 244
 245         nrs_tbf_rule_get(rule);
 246         spin_unlock(&head->th_rule_lock);
 247         return rule;
 248 }
 249
 250 static void
 251 nrs_tbf_cli_init(struct nrs_tbf_head *head,
 252                  struct nrs_tbf_client *cli,
 253                  struct ptlrpc_request *req)
 254 {
 255         struct nrs_tbf_rule *rule;
 256
 257         cli->tc_in_heap = false;
 258         head->th_ops->o_cli_init(cli, req);
 259         INIT_LIST_HEAD(&cli->tc_list);
 260         INIT_LIST_HEAD(&cli->tc_linkage);
 261         spin_lock_init(&cli->tc_rule_lock);
 262         atomic_set(&cli->tc_ref, 1);
 263         rule = nrs_tbf_rule_match(head, cli);
 264         nrs_tbf_cli_reset(head, rule, cli);
 265 }
 266
 267 static void
 268 nrs_tbf_cli_fini(struct nrs_tbf_client *cli)
 269 {
 270         LASSERT(list_empty(&cli->tc_list));
 271         LASSERT(!cli->tc_in_heap);
 272         LASSERT(atomic_read(&cli->tc_ref) == 0);
 273         spin_lock(&cli->tc_rule_lock);
 274         nrs_tbf_cli_rule_put(cli);
 275         spin_unlock(&cli->tc_rule_lock);
 276         OBD_FREE_PTR(cli);
 277 }
 278
 279 static int
 280 nrs_tbf_rule_start(struct ptlrpc_nrs_policy *policy,
 281                    struct nrs_tbf_head *head,
 282                    struct nrs_tbf_cmd *start)
 283 {
 284         struct nrs_tbf_rule *rule, *tmp_rule;
 285         int rc;
 286
 287         rule = nrs_tbf_rule_find(head, start->tc_name);
 288         if (rule) {
 289                 nrs_tbf_rule_put(rule);
 290                 return -EEXIST;
 291         }
 292
 293         OBD_CPT_ALLOC_PTR(rule, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
 294         if (rule == NULL)
 295                 return -ENOMEM;
 296
 297         memcpy(rule->tr_name, start->tc_name, strlen(start->tc_name));
 298         rule->tr_rpc_rate = start->tc_rpc_rate;
 299         rule->tr_nsecs = NSEC_PER_SEC / rule->tr_rpc_rate;
 300         rule->tr_depth = tbf_depth;
 301         atomic_set(&rule->tr_ref, 1);
 302         INIT_LIST_HEAD(&rule->tr_cli_list);
 303         INIT_LIST_HEAD(&rule->tr_nids);
 304         INIT_LIST_HEAD(&rule->tr_linkage);
 305         spin_lock_init(&rule->tr_rule_lock);
 306         rule->tr_head = head;
 307
 308         rc = head->th_ops->o_rule_init(policy, rule, start);
 309         if (rc) {
 310                 OBD_FREE_PTR(rule);
 311                 return rc;
 312         }
 313
 314         /* Add as the newest rule */
 315         spin_lock(&head->th_rule_lock);
 316         tmp_rule = nrs_tbf_rule_find_nolock(head, start->tc_name);
 317         if (tmp_rule) {
 318                 spin_unlock(&head->th_rule_lock);
 319                 nrs_tbf_rule_put(tmp_rule);
 320                 nrs_tbf_rule_put(rule);
 321                 return -EEXIST;
 322         }
 323         list_add(&rule->tr_linkage, &head->th_list);
 324         spin_unlock(&head->th_rule_lock);
 325         atomic_inc(&head->th_rule_sequence);
 326         if (start->tc_rule_flags & NTRS_DEFAULT) {
 327                 rule->tr_flags |= NTRS_DEFAULT;
 328                 LASSERT(head->th_rule == NULL);
 329                 head->th_rule = rule;
 330         }
 331
 332         return 0;
 333 }
 334
 335 static int
 336 nrs_tbf_rule_change(struct ptlrpc_nrs_policy *policy,
 337                     struct nrs_tbf_head *head,
 338                     struct nrs_tbf_cmd *change)
 339 {
 340         struct nrs_tbf_rule *rule;
 341
 342         assert_spin_locked(&policy->pol_nrs->nrs_lock);
 343
 344         rule = nrs_tbf_rule_find(head, change->tc_name);
 345         if (rule == NULL)
 346                 return -ENOENT;
 347
 348         rule->tr_rpc_rate = change->tc_rpc_rate;
 349         rule->tr_nsecs = NSEC_PER_SEC / rule->tr_rpc_rate;
 350         rule->tr_generation++;
 351         nrs_tbf_rule_put(rule);
 352
 353         return 0;
 354 }
 355
 356 static int
 357 nrs_tbf_rule_stop(struct ptlrpc_nrs_policy *policy,
 358                   struct nrs_tbf_head *head,
 359                   struct nrs_tbf_cmd *stop)
 360 {
 361         struct nrs_tbf_rule *rule;
 362
 363         assert_spin_locked(&policy->pol_nrs->nrs_lock);
 364
 365         if (strcmp(stop->tc_name, NRS_TBF_DEFAULT_RULE) == 0)
 366                 return -EPERM;
 367
 368         rule = nrs_tbf_rule_find(head, stop->tc_name);
 369         if (rule == NULL)
 370                 return -ENOENT;
 371
 372         list_del_init(&rule->tr_linkage);
 373         rule->tr_flags |= NTRS_STOPPING;
 374         nrs_tbf_rule_put(rule);
 375         nrs_tbf_rule_put(rule);
 376
 377         return 0;
 378 }
 379
 380 static int
 381 nrs_tbf_command(struct ptlrpc_nrs_policy *policy,
 382                 struct nrs_tbf_head *head,
 383                 struct nrs_tbf_cmd *cmd)
 384 {
 385         int rc;
 386
 387         assert_spin_locked(&policy->pol_nrs->nrs_lock);
 388
 389         switch (cmd->tc_cmd) {
 390         case NRS_CTL_TBF_START_RULE:
 391                 if (!(cmd->tc_valid_types & head->th_type_flag))
 392                         return -EINVAL;
 393
 394                 spin_unlock(&policy->pol_nrs->nrs_lock);
 395                 rc = nrs_tbf_rule_start(policy, head, cmd);
 396                 spin_lock(&policy->pol_nrs->nrs_lock);
 397                 return rc;
 398         case NRS_CTL_TBF_CHANGE_RATE:
 399                 rc = nrs_tbf_rule_change(policy, head, cmd);
 400                 return rc;
 401         case NRS_CTL_TBF_STOP_RULE:
 402                 rc = nrs_tbf_rule_stop(policy, head, cmd);
 403                 /* Take it as a success, if not exists at all */
 404                 return rc == -ENOENT ? 0 : rc;
 405         default:
 406                 return -EFAULT;
 407         }
 408 }
 409
 410 /**
 411  * Binary heap predicate.
 412  *
 413  * \param[in] e1 the first binheap node to compare
 414  * \param[in] e2 the second binheap node to compare
 415  *
 416  * \retval 0 e1 > e2
 417  * \retval 1 e1 < e2
 418  */
 419 static int
 420 tbf_cli_compare(struct cfs_binheap_node *e1, struct cfs_binheap_node *e2)
 421 {
 422         struct nrs_tbf_client *cli1;
 423         struct nrs_tbf_client *cli2;
 424
 425         cli1 = container_of(e1, struct nrs_tbf_client, tc_node);
 426         cli2 = container_of(e2, struct nrs_tbf_client, tc_node);
 427
 428         if (cli1->tc_check_time + cli1->tc_nsecs <
 429             cli2->tc_check_time + cli2->tc_nsecs)
 430                 return 1;
 431         else if (cli1->tc_check_time + cli1->tc_nsecs >
 432                  cli2->tc_check_time + cli2->tc_nsecs)
 433                 return 0;
 434
 435         if (cli1->tc_check_time < cli2->tc_check_time)
 436                 return 1;
 437         else if (cli1->tc_check_time > cli2->tc_check_time)
 438                 return 0;
 439
 440         /* Maybe need more comparasion, e.g. request number in the rules */
 441         return 1;
 442 }
 443
 444 /**
 445  * TBF binary heap operations
 446  */
 447 static struct cfs_binheap_ops nrs_tbf_heap_ops = {
 448         .hop_enter      = NULL,
 449         .hop_exit       = NULL,
 450         .hop_compare    = tbf_cli_compare,
 451 };
 452
 453 static unsigned nrs_tbf_jobid_hop_hash(struct cfs_hash *hs, const void *key,
 454                                   unsigned mask)
 455 {
 456         return cfs_hash_djb2_hash(key, strlen(key), mask);
 457 }
 458
 459 static int nrs_tbf_jobid_hop_keycmp(const void *key, struct hlist_node *hnode)
 460 {
 461         struct nrs_tbf_client *cli = hlist_entry(hnode,
 462                                                      struct nrs_tbf_client,
 463                                                      tc_hnode);
 464
 465         return (strcmp(cli->tc_jobid, key) == 0);
 466 }
 467
 468 static void *nrs_tbf_jobid_hop_key(struct hlist_node *hnode)
 469 {
 470         struct nrs_tbf_client *cli = hlist_entry(hnode,
 471                                                      struct nrs_tbf_client,
 472                                                      tc_hnode);
 473
 474         return cli->tc_jobid;
 475 }
 476
 477 static void *nrs_tbf_jobid_hop_object(struct hlist_node *hnode)
 478 {
 479         return hlist_entry(hnode, struct nrs_tbf_client, tc_hnode);
 480 }
 481
 482 static void nrs_tbf_jobid_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
 483 {
 484         struct nrs_tbf_client *cli = hlist_entry(hnode,
 485                                                      struct nrs_tbf_client,
 486                                                      tc_hnode);
 487
 488         atomic_inc(&cli->tc_ref);
 489 }
 490
 491 static void nrs_tbf_jobid_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 492 {
 493         struct nrs_tbf_client *cli = hlist_entry(hnode,
 494                                                      struct nrs_tbf_client,
 495                                                      tc_hnode);
 496
 497         atomic_dec(&cli->tc_ref);
 498 }
 499
 500 static void
 501 nrs_tbf_jobid_hop_exit(struct cfs_hash *hs, struct hlist_node *hnode)
 502
 503 {
 504         struct nrs_tbf_client *cli = hlist_entry(hnode,
 505                                                  struct nrs_tbf_client,
 506                                                  tc_hnode);
 507
 508         LASSERT(atomic_read(&cli->tc_ref) == 0);
 509         nrs_tbf_cli_fini(cli);
 510 }
 511
 512 static struct cfs_hash_ops nrs_tbf_jobid_hash_ops = {
 513         .hs_hash        = nrs_tbf_jobid_hop_hash,
 514         .hs_keycmp      = nrs_tbf_jobid_hop_keycmp,
 515         .hs_key         = nrs_tbf_jobid_hop_key,
 516         .hs_object      = nrs_tbf_jobid_hop_object,
 517         .hs_get         = nrs_tbf_jobid_hop_get,
 518         .hs_put         = nrs_tbf_jobid_hop_put,
 519         .hs_put_locked  = nrs_tbf_jobid_hop_put,
 520         .hs_exit        = nrs_tbf_jobid_hop_exit,
 521 };
 522
 523 #define NRS_TBF_JOBID_HASH_FLAGS (CFS_HASH_SPIN_BKTLOCK | \
 524                                   CFS_HASH_NO_ITEMREF | \
 525                                   CFS_HASH_DEPTH)
 526
 527 static struct nrs_tbf_client *
 528 nrs_tbf_jobid_hash_lookup(struct cfs_hash *hs,
 529                           struct cfs_hash_bd *bd,
 530                           const char *jobid)
 531 {
 532         struct hlist_node *hnode;
 533         struct nrs_tbf_client *cli;
 534
 535         /* cfs_hash_bd_peek_locked is a somehow "internal" function
 536          * of cfs_hash, it doesn't add refcount on object. */
 537         hnode = cfs_hash_bd_peek_locked(hs, bd, (void *)jobid);
 538         if (hnode == NULL)
 539                 return NULL;
 540
 541         cfs_hash_get(hs, hnode);
 542         cli = container_of0(hnode, struct nrs_tbf_client, tc_hnode);
 543         if (!list_empty(&cli->tc_lru))
 544                 list_del_init(&cli->tc_lru);
 545         return cli;
 546 }
 547
 548 #define NRS_TBF_JOBID_NULL ""
 549
 550 static struct nrs_tbf_client *
 551 nrs_tbf_jobid_cli_find(struct nrs_tbf_head *head,
 552                        struct ptlrpc_request *req)
 553 {
 554         const char              *jobid;
 555         struct nrs_tbf_client   *cli;
 556         struct cfs_hash         *hs = head->th_cli_hash;
 557         struct cfs_hash_bd               bd;
 558
 559         jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 560         if (jobid == NULL)
 561                 jobid = NRS_TBF_JOBID_NULL;
 562         cfs_hash_bd_get_and_lock(hs, (void *)jobid, &bd, 1);
 563         cli = nrs_tbf_jobid_hash_lookup(hs, &bd, jobid);
 564         cfs_hash_bd_unlock(hs, &bd, 1);
 565
 566         return cli;
 567 }
 568
 569 static struct nrs_tbf_client *
 570 nrs_tbf_jobid_cli_findadd(struct nrs_tbf_head *head,
 571                           struct nrs_tbf_client *cli)
 572 {
 573         const char              *jobid;
 574         struct nrs_tbf_client   *ret;
 575         struct cfs_hash         *hs = head->th_cli_hash;
 576         struct cfs_hash_bd               bd;
 577
 578         jobid = cli->tc_jobid;
 579         cfs_hash_bd_get_and_lock(hs, (void *)jobid, &bd, 1);
 580         ret = nrs_tbf_jobid_hash_lookup(hs, &bd, jobid);
 581         if (ret == NULL) {
 582                 cfs_hash_bd_add_locked(hs, &bd, &cli->tc_hnode);
 583                 ret = cli;
 584         }
 585         cfs_hash_bd_unlock(hs, &bd, 1);
 586
 587         return ret;
 588 }
 589
 590 static void
 591 nrs_tbf_jobid_cli_put(struct nrs_tbf_head *head,
 592                       struct nrs_tbf_client *cli)
 593 {
 594         struct cfs_hash_bd               bd;
 595         struct cfs_hash         *hs = head->th_cli_hash;
 596         struct nrs_tbf_bucket   *bkt;
 597         int                      hw;
 598         struct list_head        zombies;
 599
 600         INIT_LIST_HEAD(&zombies);
 601         cfs_hash_bd_get(hs, &cli->tc_jobid, &bd);
 602         bkt = cfs_hash_bd_extra_get(hs, &bd);
 603         if (!cfs_hash_bd_dec_and_lock(hs, &bd, &cli->tc_ref))
 604                 return;
 605         LASSERT(list_empty(&cli->tc_lru));
 606         list_add_tail(&cli->tc_lru, &bkt->ntb_lru);
 607
 608         /*
 609          * Check and purge the LRU, there is at least one client in the LRU.
 610          */
 611         hw = tbf_jobid_cache_size >>
 612              (hs->hs_cur_bits - hs->hs_bkt_bits);
 613         while (cfs_hash_bd_count_get(&bd) > hw) {
 614                 if (unlikely(list_empty(&bkt->ntb_lru)))
 615                         break;
 616                 cli = list_entry(bkt->ntb_lru.next,
 617                                      struct nrs_tbf_client,
 618                                      tc_lru);
 619                 LASSERT(atomic_read(&cli->tc_ref) == 0);
 620                 cfs_hash_bd_del_locked(hs, &bd, &cli->tc_hnode);
 621                 list_move(&cli->tc_lru, &zombies);
 622         }
 623         cfs_hash_bd_unlock(head->th_cli_hash, &bd, 1);
 624
 625         while (!list_empty(&zombies)) {
 626                 cli = container_of0(zombies.next,
 627                                     struct nrs_tbf_client, tc_lru);
 628                 list_del_init(&cli->tc_lru);
 629                 nrs_tbf_cli_fini(cli);
 630         }
 631 }
 632
 633 static void
 634 nrs_tbf_jobid_cli_init(struct nrs_tbf_client *cli,
 635                        struct ptlrpc_request *req)
 636 {
 637         char *jobid = lustre_msg_get_jobid(req->rq_reqmsg);
 638
 639         if (jobid == NULL)
 640                 jobid = NRS_TBF_JOBID_NULL;
 641         LASSERT(strlen(jobid) < LUSTRE_JOBID_SIZE);
 642         INIT_LIST_HEAD(&cli->tc_lru);
 643         memcpy(cli->tc_jobid, jobid, strlen(jobid));
 644 }
 645
 646 static int nrs_tbf_jobid_hash_order(void)
 647 {
 648         int bits;
 649
 650         for (bits = 1; (1 << bits) < tbf_jobid_cache_size; ++bits)
 651                 ;
 652
 653         return bits;
 654 }
 655
 656 #define NRS_TBF_JOBID_BKT_BITS 10
 657
 658 static int
 659 nrs_tbf_jobid_startup(struct ptlrpc_nrs_policy *policy,
 660                       struct nrs_tbf_head *head)
 661 {
 662         struct nrs_tbf_cmd       start;
 663         struct nrs_tbf_bucket   *bkt;
 664         int                      bits;
 665         int                      i;
 666         int                      rc;
 667         struct cfs_hash_bd       bd;
 668
 669         bits = nrs_tbf_jobid_hash_order();
 670         if (bits < NRS_TBF_JOBID_BKT_BITS)
 671                 bits = NRS_TBF_JOBID_BKT_BITS;
 672         head->th_cli_hash = cfs_hash_create("nrs_tbf_hash",
 673                                             bits,
 674                                             bits,
 675                                             NRS_TBF_JOBID_BKT_BITS,
 676                                             sizeof(*bkt),
 677                                             0,
 678                                             0,
 679                                             &nrs_tbf_jobid_hash_ops,
 680                                             NRS_TBF_JOBID_HASH_FLAGS);
 681         if (head->th_cli_hash == NULL)
 682                 return -ENOMEM;
 683
 684         cfs_hash_for_each_bucket(head->th_cli_hash, &bd, i) {
 685                 bkt = cfs_hash_bd_extra_get(head->th_cli_hash, &bd);
 686                 INIT_LIST_HEAD(&bkt->ntb_lru);
 687         }
 688
 689         memset(&start, 0, sizeof(start));
 690         start.tc_jobids_str = "*";
 691
 692         start.tc_rpc_rate = tbf_rate;
 693         start.tc_rule_flags = NTRS_DEFAULT;
 694         start.tc_name = NRS_TBF_DEFAULT_RULE;
 695         INIT_LIST_HEAD(&start.tc_jobids);
 696         rc = nrs_tbf_rule_start(policy, head, &start);
 697
 698         return rc;
 699 }
 700
 701 /**
 702  * Frees jobid of \a list.
 703  *
 704  */
 705 static void
 706 nrs_tbf_jobid_list_free(struct list_head *jobid_list)
 707 {
 708         struct nrs_tbf_jobid *jobid, *n;
 709
 710         list_for_each_entry_safe(jobid, n, jobid_list, tj_linkage) {
 711                 OBD_FREE(jobid->tj_id, strlen(jobid->tj_id) + 1);
 712                 list_del(&jobid->tj_linkage);
 713                 OBD_FREE(jobid, sizeof(struct nrs_tbf_jobid));
 714         }
 715 }
 716
 717 static int
 718 nrs_tbf_jobid_list_add(const struct cfs_lstr *id, struct list_head *jobid_list)
 719 {
 720         struct nrs_tbf_jobid *jobid;
 721
 722         OBD_ALLOC(jobid, sizeof(struct nrs_tbf_jobid));
 723         if (jobid == NULL)
 724                 return -ENOMEM;
 725
 726         OBD_ALLOC(jobid->tj_id, id->ls_len + 1);
 727         if (jobid->tj_id == NULL) {
 728                 OBD_FREE(jobid, sizeof(struct nrs_tbf_jobid));
 729                 return -ENOMEM;
 730         }
 731
 732         memcpy(jobid->tj_id, id->ls_str, id->ls_len);
 733         list_add_tail(&jobid->tj_linkage, jobid_list);
 734         return 0;
 735 }
 736
 737 static int
 738 nrs_tbf_jobid_list_match(struct list_head *jobid_list, char *id)
 739 {
 740         struct nrs_tbf_jobid *jobid;
 741
 742         list_for_each_entry(jobid, jobid_list, tj_linkage) {
 743                 if (strcmp(id, jobid->tj_id) == 0)
 744                         return 1;
 745         }
 746         return 0;
 747 }
 748
 749 static int
 750 nrs_tbf_jobid_list_parse(char *str, int len, struct list_head *jobid_list)
 751 {
 752         struct cfs_lstr src;
 753         struct cfs_lstr res;
 754         int rc = 0;
 755         ENTRY;
 756
 757         src.ls_str = str;
 758         src.ls_len = len;
 759         INIT_LIST_HEAD(jobid_list);
 760         while (src.ls_str) {
 761                 rc = cfs_gettok(&src, ' ', &res);
 762                 if (rc == 0) {
 763                         rc = -EINVAL;
 764                         break;
 765                 }
 766                 rc = nrs_tbf_jobid_list_add(&res, jobid_list);
 767                 if (rc)
 768                         break;
 769         }
 770         if (rc)
 771                 nrs_tbf_jobid_list_free(jobid_list);
 772         RETURN(rc);
 773 }
 774
 775 static void nrs_tbf_jobid_cmd_fini(struct nrs_tbf_cmd *cmd)
 776 {
 777         if (!list_empty(&cmd->tc_jobids))
 778                 nrs_tbf_jobid_list_free(&cmd->tc_jobids);
 779         if (cmd->tc_jobids_str)
 780                 OBD_FREE(cmd->tc_jobids_str, strlen(cmd->tc_jobids_str) + 1);
 781 }
 782
 783 static int nrs_tbf_jobid_parse(struct nrs_tbf_cmd *cmd, const char *id)
 784 {
 785         int rc;
 786
 787         OBD_ALLOC(cmd->tc_jobids_str, strlen(id) + 1);
 788         if (cmd->tc_jobids_str == NULL)
 789                 return -ENOMEM;
 790
 791         memcpy(cmd->tc_jobids_str, id, strlen(id));
 792
 793         /* parse jobid list */
 794         rc = nrs_tbf_jobid_list_parse(cmd->tc_jobids_str,
 795                                       strlen(cmd->tc_jobids_str),
 796                                       &cmd->tc_jobids);
 797         if (rc)
 798                 nrs_tbf_jobid_cmd_fini(cmd);
 799
 800         return rc;
 801 }
 802
 803 static int nrs_tbf_jobid_rule_init(struct ptlrpc_nrs_policy *policy,
 804                                    struct nrs_tbf_rule *rule,
 805                                    struct nrs_tbf_cmd *start)
 806 {
 807         int rc = 0;
 808
 809         LASSERT(start->tc_jobids_str);
 810         OBD_ALLOC(rule->tr_jobids_str,
 811                   strlen(start->tc_jobids_str) + 1);
 812         if (rule->tr_jobids_str == NULL)
 813                 return -ENOMEM;
 814
 815         memcpy(rule->tr_jobids_str,
 816                start->tc_jobids_str,
 817                strlen(start->tc_jobids_str));
 818
 819         INIT_LIST_HEAD(&rule->tr_jobids);
 820         if (!list_empty(&start->tc_jobids)) {
 821                 rc = nrs_tbf_jobid_list_parse(rule->tr_jobids_str,
 822                                               strlen(rule->tr_jobids_str),
 823                                               &rule->tr_jobids);
 824                 if (rc)
 825                         CERROR("jobids {%s} illegal\n", rule->tr_jobids_str);
 826         }
 827         if (rc)
 828                 OBD_FREE(rule->tr_jobids_str,
 829                          strlen(start->tc_jobids_str) + 1);
 830         return rc;
 831 }
 832
 833 static int
 834 nrs_tbf_jobid_rule_dump(struct nrs_tbf_rule *rule, struct seq_file *m)
 835 {
 836         return seq_printf(m, "%s {%s} %llu, ref %d\n", rule->tr_name,
 837                           rule->tr_jobids_str, rule->tr_rpc_rate,
 838                           atomic_read(&rule->tr_ref) - 1);
 839 }
 840
 841 static int
 842 nrs_tbf_jobid_rule_match(struct nrs_tbf_rule *rule,
 843                          struct nrs_tbf_client *cli)
 844 {
 845         return nrs_tbf_jobid_list_match(&rule->tr_jobids, cli->tc_jobid);
 846 }
 847
 848 static void nrs_tbf_jobid_rule_fini(struct nrs_tbf_rule *rule)
 849 {
 850         if (!list_empty(&rule->tr_jobids))
 851                 nrs_tbf_jobid_list_free(&rule->tr_jobids);
 852         LASSERT(rule->tr_jobids_str != NULL);
 853         OBD_FREE(rule->tr_jobids_str, strlen(rule->tr_jobids_str) + 1);
 854 }
 855
 856 static struct nrs_tbf_ops nrs_tbf_jobid_ops = {
 857         .o_name = NRS_TBF_TYPE_JOBID,
 858         .o_startup = nrs_tbf_jobid_startup,
 859         .o_cli_find = nrs_tbf_jobid_cli_find,
 860         .o_cli_findadd = nrs_tbf_jobid_cli_findadd,
 861         .o_cli_put = nrs_tbf_jobid_cli_put,
 862         .o_cli_init = nrs_tbf_jobid_cli_init,
 863         .o_rule_init = nrs_tbf_jobid_rule_init,
 864         .o_rule_dump = nrs_tbf_jobid_rule_dump,
 865         .o_rule_match = nrs_tbf_jobid_rule_match,
 866         .o_rule_fini = nrs_tbf_jobid_rule_fini,
 867 };
 868
 869 /**
 870  * libcfs_hash operations for nrs_tbf_net::cn_cli_hash
 871  *
 872  * This uses ptlrpc_request::rq_peer.nid as its key, in order to hash
 873  * nrs_tbf_client objects.
 874  */
 875 #define NRS_TBF_NID_BKT_BITS    8
 876 #define NRS_TBF_NID_BITS        16
 877
 878 static unsigned nrs_tbf_nid_hop_hash(struct cfs_hash *hs, const void *key,
 879                                   unsigned mask)
 880 {
 881         return cfs_hash_djb2_hash(key, sizeof(lnet_nid_t), mask);
 882 }
 883
 884 static int nrs_tbf_nid_hop_keycmp(const void *key, struct hlist_node *hnode)
 885 {
 886         lnet_nid_t            *nid = (lnet_nid_t *)key;
 887         struct nrs_tbf_client *cli = hlist_entry(hnode,
 888                                                      struct nrs_tbf_client,
 889                                                      tc_hnode);
 890
 891         return *nid == cli->tc_nid;
 892 }
 893
 894 static void *nrs_tbf_nid_hop_key(struct hlist_node *hnode)
 895 {
 896         struct nrs_tbf_client *cli = hlist_entry(hnode,
 897                                                      struct nrs_tbf_client,
 898                                                      tc_hnode);
 899
 900         return &cli->tc_nid;
 901 }
 902
 903 static void *nrs_tbf_nid_hop_object(struct hlist_node *hnode)
 904 {
 905         return hlist_entry(hnode, struct nrs_tbf_client, tc_hnode);
 906 }
 907
 908 static void nrs_tbf_nid_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
 909 {
 910         struct nrs_tbf_client *cli = hlist_entry(hnode,
 911                                                      struct nrs_tbf_client,
 912                                                      tc_hnode);
 913
 914         atomic_inc(&cli->tc_ref);
 915 }
 916
 917 static void nrs_tbf_nid_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 918 {
 919         struct nrs_tbf_client *cli = hlist_entry(hnode,
 920                                                      struct nrs_tbf_client,
 921                                                      tc_hnode);
 922
 923         atomic_dec(&cli->tc_ref);
 924 }
 925
 926 static void nrs_tbf_nid_hop_exit(struct cfs_hash *hs, struct hlist_node *hnode)
 927 {
 928         struct nrs_tbf_client *cli = hlist_entry(hnode,
 929                                                      struct nrs_tbf_client,
 930                                                      tc_hnode);
 931
 932         LASSERTF(atomic_read(&cli->tc_ref) == 0,
 933                  "Busy TBF object from client with NID %s, with %d refs\n",
 934                  libcfs_nid2str(cli->tc_nid), atomic_read(&cli->tc_ref));
 935
 936         nrs_tbf_cli_fini(cli);
 937 }
 938
 939 static struct cfs_hash_ops nrs_tbf_nid_hash_ops = {
 940         .hs_hash        = nrs_tbf_nid_hop_hash,
 941         .hs_keycmp      = nrs_tbf_nid_hop_keycmp,
 942         .hs_key         = nrs_tbf_nid_hop_key,
 943         .hs_object      = nrs_tbf_nid_hop_object,
 944         .hs_get         = nrs_tbf_nid_hop_get,
 945         .hs_put         = nrs_tbf_nid_hop_put,
 946         .hs_put_locked  = nrs_tbf_nid_hop_put,
 947         .hs_exit        = nrs_tbf_nid_hop_exit,
 948 };
 949
 950 static struct nrs_tbf_client *
 951 nrs_tbf_nid_cli_find(struct nrs_tbf_head *head,
 952                      struct ptlrpc_request *req)
 953 {
 954         return cfs_hash_lookup(head->th_cli_hash, &req->rq_peer.nid);
 955 }
 956
 957 static struct nrs_tbf_client *
 958 nrs_tbf_nid_cli_findadd(struct nrs_tbf_head *head,
 959                         struct nrs_tbf_client *cli)
 960 {
 961         return cfs_hash_findadd_unique(head->th_cli_hash, &cli->tc_nid,
 962                                        &cli->tc_hnode);
 963 }
 964
 965 static void
 966 nrs_tbf_nid_cli_put(struct nrs_tbf_head *head,
 967                       struct nrs_tbf_client *cli)
 968 {
 969         cfs_hash_put(head->th_cli_hash, &cli->tc_hnode);
 970 }
 971
 972 static int
 973 nrs_tbf_nid_startup(struct ptlrpc_nrs_policy *policy,
 974                     struct nrs_tbf_head *head)
 975 {
 976         struct nrs_tbf_cmd      start;
 977         int rc;
 978
 979         head->th_cli_hash = cfs_hash_create("nrs_tbf_hash",
 980                                             NRS_TBF_NID_BITS,
 981                                             NRS_TBF_NID_BITS,
 982                                             NRS_TBF_NID_BKT_BITS, 0,
 983                                             CFS_HASH_MIN_THETA,
 984                                             CFS_HASH_MAX_THETA,
 985                                             &nrs_tbf_nid_hash_ops,
 986                                             CFS_HASH_RW_BKTLOCK);
 987         if (head->th_cli_hash == NULL)
 988                 return -ENOMEM;
 989
 990         memset(&start, 0, sizeof(start));
 991         start.tc_nids_str = "*";
 992
 993         start.tc_rpc_rate = tbf_rate;
 994         start.tc_rule_flags = NTRS_DEFAULT;
 995         start.tc_name = NRS_TBF_DEFAULT_RULE;
 996         INIT_LIST_HEAD(&start.tc_nids);
 997         rc = nrs_tbf_rule_start(policy, head, &start);
 998
 999         return rc;
1000 }
1001
1002 static void
1003 nrs_tbf_nid_cli_init(struct nrs_tbf_client *cli,
1004                              struct ptlrpc_request *req)
1005 {
1006         cli->tc_nid = req->rq_peer.nid;
1007 }
1008
1009 static int nrs_tbf_nid_rule_init(struct ptlrpc_nrs_policy *policy,
1010                                  struct nrs_tbf_rule *rule,
1011                                  struct nrs_tbf_cmd *start)
1012 {
1013         LASSERT(start->tc_nids_str);
1014         OBD_ALLOC(rule->tr_nids_str,
1015                   strlen(start->tc_nids_str) + 1);
1016         if (rule->tr_nids_str == NULL)
1017                 return -ENOMEM;
1018
1019         memcpy(rule->tr_nids_str,
1020                start->tc_nids_str,
1021                strlen(start->tc_nids_str));
1022
1023         INIT_LIST_HEAD(&rule->tr_nids);
1024         if (!list_empty(&start->tc_nids)) {
1025                 if (cfs_parse_nidlist(rule->tr_nids_str,
1026                                       strlen(rule->tr_nids_str),
1027                                       &rule->tr_nids) <= 0) {
1028                         CERROR("nids {%s} illegal\n",
1029                                rule->tr_nids_str);
1030                         OBD_FREE(rule->tr_nids_str,
1031                                  strlen(start->tc_nids_str) + 1);
1032                         return -EINVAL;
1033                 }
1034         }
1035         return 0;
1036 }
1037
1038 static int
1039 nrs_tbf_nid_rule_dump(struct nrs_tbf_rule *rule, struct seq_file *m)
1040 {
1041         return seq_printf(m, "%s {%s} %llu, ref %d\n", rule->tr_name,
1042                           rule->tr_nids_str, rule->tr_rpc_rate,
1043                           atomic_read(&rule->tr_ref) - 1);
1044 }
1045
1046 static int
1047 nrs_tbf_nid_rule_match(struct nrs_tbf_rule *rule,
1048                        struct nrs_tbf_client *cli)
1049 {
1050         return cfs_match_nid(cli->tc_nid, &rule->tr_nids);
1051 }
1052
1053 static void nrs_tbf_nid_rule_fini(struct nrs_tbf_rule *rule)
1054 {
1055         if (!list_empty(&rule->tr_nids))
1056                 cfs_free_nidlist(&rule->tr_nids);
1057         LASSERT(rule->tr_nids_str != NULL);
1058         OBD_FREE(rule->tr_nids_str, strlen(rule->tr_nids_str) + 1);
1059 }
1060
1061 static void nrs_tbf_nid_cmd_fini(struct nrs_tbf_cmd *cmd)
1062 {
1063         if (!list_empty(&cmd->tc_nids))
1064                 cfs_free_nidlist(&cmd->tc_nids);
1065         if (cmd->tc_nids_str)
1066                 OBD_FREE(cmd->tc_nids_str, strlen(cmd->tc_nids_str) + 1);
1067 }
1068
1069 static int nrs_tbf_nid_parse(struct nrs_tbf_cmd *cmd, const char *id)
1070 {
1071         OBD_ALLOC(cmd->tc_nids_str, strlen(id) + 1);
1072         if (cmd->tc_nids_str == NULL)
1073                 return -ENOMEM;
1074
1075         memcpy(cmd->tc_nids_str, id, strlen(id));
1076
1077         /* parse NID list */
1078         if (cfs_parse_nidlist(cmd->tc_nids_str,
1079                               strlen(cmd->tc_nids_str),
1080                               &cmd->tc_nids) <= 0) {
1081                 nrs_tbf_nid_cmd_fini(cmd);
1082                 return -EINVAL;
1083         }
1084
1085         return 0;
1086 }
1087
1088 static struct nrs_tbf_ops nrs_tbf_nid_ops = {
1089         .o_name = NRS_TBF_TYPE_NID,
1090         .o_startup = nrs_tbf_nid_startup,
1091         .o_cli_find = nrs_tbf_nid_cli_find,
1092         .o_cli_findadd = nrs_tbf_nid_cli_findadd,
1093         .o_cli_put = nrs_tbf_nid_cli_put,
1094         .o_cli_init = nrs_tbf_nid_cli_init,
1095         .o_rule_init = nrs_tbf_nid_rule_init,
1096         .o_rule_dump = nrs_tbf_nid_rule_dump,
1097         .o_rule_match = nrs_tbf_nid_rule_match,
1098         .o_rule_fini = nrs_tbf_nid_rule_fini,
1099 };
1100
1101 /**
1102  * Is called before the policy transitions into
1103  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED; allocates and initializes a
1104  * policy-specific private data structure.
1105  *
1106  * \param[in] policy The policy to start
1107  *
1108  * \retval -ENOMEM OOM error
1109  * \retval  0      success
1110  *
1111  * \see nrs_policy_register()
1112  * \see nrs_policy_ctl()
1113  */
1114 static int nrs_tbf_start(struct ptlrpc_nrs_policy *policy, char *arg)
1115 {
1116         struct nrs_tbf_head     *head;
1117         struct nrs_tbf_ops      *ops;
1118         __u32                    type;
1119         int rc = 0;
1120
1121         if (arg == NULL || strlen(arg) > NRS_TBF_TYPE_MAX_LEN)
1122                 GOTO(out, rc = -EINVAL);
1123
1124         if (strcmp(arg, NRS_TBF_TYPE_NID) == 0) {
1125                 ops = &nrs_tbf_nid_ops;
1126                 type = NRS_TBF_FLAG_NID;
1127         } else if (strcmp(arg, NRS_TBF_TYPE_JOBID) == 0) {
1128                 ops = &nrs_tbf_jobid_ops;
1129                 type = NRS_TBF_FLAG_JOBID;
1130         } else
1131                 GOTO(out, rc = -ENOTSUPP);
1132
1133         OBD_CPT_ALLOC_PTR(head, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
1134         if (head == NULL)
1135                 GOTO(out, rc = -ENOMEM);
1136
1137         memcpy(head->th_type, arg, strlen(arg));
1138         head->th_type[strlen(arg)] = '\0';
1139         head->th_ops = ops;
1140         head->th_type_flag = type;
1141
1142         head->th_binheap = cfs_binheap_create(&nrs_tbf_heap_ops,
1143                                               CBH_FLAG_ATOMIC_GROW, 4096, NULL,
1144                                               nrs_pol2cptab(policy),
1145                                               nrs_pol2cptid(policy));
1146         if (head->th_binheap == NULL)
1147                 GOTO(out_free_head, rc = -ENOMEM);
1148
1149         atomic_set(&head->th_rule_sequence, 0);
1150         spin_lock_init(&head->th_rule_lock);
1151         INIT_LIST_HEAD(&head->th_list);
1152         hrtimer_init(&head->th_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
1153         head->th_timer.function = nrs_tbf_timer_cb;
1154         rc = head->th_ops->o_startup(policy, head);
1155         if (rc)
1156                 GOTO(out_free_heap, rc);
1157
1158         policy->pol_private = head;
1159         return 0;
1160 out_free_heap:
1161         cfs_binheap_destroy(head->th_binheap);
1162 out_free_head:
1163         OBD_FREE_PTR(head);
1164 out:
1165         return rc;
1166 }
1167
1168 /**
1169  * Is called before the policy transitions into
1170  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED; deallocates the policy-specific
1171  * private data structure.
1172  *
1173  * \param[in] policy The policy to stop
1174  *
1175  * \see nrs_policy_stop0()
1176  */
1177 static void nrs_tbf_stop(struct ptlrpc_nrs_policy *policy)
1178 {
1179         struct nrs_tbf_head *head = policy->pol_private;
1180         struct ptlrpc_nrs *nrs = policy->pol_nrs;
1181         struct nrs_tbf_rule *rule, *n;
1182
1183         LASSERT(head != NULL);
1184         LASSERT(head->th_cli_hash != NULL);
1185         hrtimer_cancel(&head->th_timer);
1186         /* Should cleanup hash first before free rules */
1187         cfs_hash_putref(head->th_cli_hash);
1188         list_for_each_entry_safe(rule, n, &head->th_list, tr_linkage) {
1189                 list_del_init(&rule->tr_linkage);
1190                 nrs_tbf_rule_put(rule);
1191         }
1192         LASSERT(list_empty(&head->th_list));
1193         LASSERT(head->th_binheap != NULL);
1194         LASSERT(cfs_binheap_is_empty(head->th_binheap));
1195         cfs_binheap_destroy(head->th_binheap);
1196         OBD_FREE_PTR(head);
1197         nrs->nrs_throttling = 0;
1198         wake_up(&policy->pol_nrs->nrs_svcpt->scp_waitq);
1199 }
1200
1201 /**
1202  * Performs a policy-specific ctl function on TBF policy instances; similar
1203  * to ioctl.
1204  *
1205  * \param[in]     policy the policy instance
1206  * \param[in]     opc    the opcode
1207  * \param[in,out] arg    used for passing parameters and information
1208  *
1209  * \pre assert_spin_locked(&policy->pol_nrs->->nrs_lock)
1210  * \post assert_spin_locked(&policy->pol_nrs->->nrs_lock)
1211  *
1212  * \retval 0   operation carried out successfully
1213  * \retval -ve error
1214  */
1215 static int nrs_tbf_ctl(struct ptlrpc_nrs_policy *policy,
1216                        enum ptlrpc_nrs_ctl opc,
1217                        void *arg)
1218 {
1219         int rc = 0;
1220         ENTRY;
1221
1222         assert_spin_locked(&policy->pol_nrs->nrs_lock);
1223
1224         switch ((enum nrs_ctl_tbf)opc) {
1225         default:
1226                 RETURN(-EINVAL);
1227
1228         /**
1229          * Read RPC rate size of a policy instance.
1230          */
1231         case NRS_CTL_TBF_RD_RULE: {
1232                 struct nrs_tbf_head *head = policy->pol_private;
1233                 struct seq_file *m = (struct seq_file *) arg;
1234                 struct ptlrpc_service_part *svcpt;
1235
1236                 svcpt = policy->pol_nrs->nrs_svcpt;
1237                 seq_printf(m, "CPT %d:\n", svcpt->scp_cpt);
1238
1239                 rc = nrs_tbf_rule_dump_all(head, m);
1240                 }
1241                 break;
1242
1243         /**
1244          * Write RPC rate of a policy instance.
1245          */
1246         case NRS_CTL_TBF_WR_RULE: {
1247                 struct nrs_tbf_head *head = policy->pol_private;
1248                 struct nrs_tbf_cmd *cmd;
1249
1250                 cmd = (struct nrs_tbf_cmd *)arg;
1251                 rc = nrs_tbf_command(policy,
1252                                      head,
1253                                      cmd);
1254                 }
1255                 break;
1256         }
1257
1258         RETURN(rc);
1259 }
1260
1261 /**
1262  * Is called for obtaining a TBF policy resource.
1263  *
1264  * \param[in]  policy     The policy on which the request is being asked for
1265  * \param[in]  nrq        The request for which resources are being taken
1266  * \param[in]  parent     Parent resource, unused in this policy
1267  * \param[out] resp       Resources references are placed in this array
1268  * \param[in]  moving_req Signifies limited caller context; unused in this
1269  *                        policy
1270  *
1271  *
1272  * \see nrs_resource_get_safe()
1273  */
1274 static int nrs_tbf_res_get(struct ptlrpc_nrs_policy *policy,
1275                            struct ptlrpc_nrs_request *nrq,
1276                            const struct ptlrpc_nrs_resource *parent,
1277                            struct ptlrpc_nrs_resource **resp,
1278                            bool moving_req)
1279 {
1280         struct nrs_tbf_head   *head;
1281         struct nrs_tbf_client *cli;
1282         struct nrs_tbf_client *tmp;
1283         struct ptlrpc_request *req;
1284
1285         if (parent == NULL) {
1286                 *resp = &((struct nrs_tbf_head *)policy->pol_private)->th_res;
1287                 return 0;
1288         }
1289
1290         head = container_of(parent, struct nrs_tbf_head, th_res);
1291         req = container_of(nrq, struct ptlrpc_request, rq_nrq);
1292         cli = head->th_ops->o_cli_find(head, req);
1293         if (cli != NULL) {
1294                 spin_lock(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
1295                 LASSERT(cli->tc_rule);
1296                 if (cli->tc_rule_sequence !=
1297                     atomic_read(&head->th_rule_sequence) ||
1298                     cli->tc_rule->tr_flags & NTRS_STOPPING) {
1299                         struct nrs_tbf_rule *rule;
1300
1301                         rule = nrs_tbf_rule_match(head, cli);
1302                         if (rule != cli->tc_rule)
1303                                 nrs_tbf_cli_reset(head, rule, cli);
1304                         else
1305                                 nrs_tbf_rule_put(rule);
1306                 } else if (cli->tc_rule_generation !=
1307                            cli->tc_rule->tr_generation) {
1308                         nrs_tbf_cli_reset_value(head, cli);
1309                 }
1310                 spin_unlock(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
1311                 goto out;
1312         }
1313
1314         OBD_CPT_ALLOC_GFP(cli, nrs_pol2cptab(policy), nrs_pol2cptid(policy),
1315                           sizeof(*cli), moving_req ? GFP_ATOMIC : __GFP_IO);
1316         if (cli == NULL)
1317                 return -ENOMEM;
1318         nrs_tbf_cli_init(head, cli, req);
1319         tmp = head->th_ops->o_cli_findadd(head, cli);
1320         if (tmp != cli) {
1321                 atomic_dec(&cli->tc_ref);
1322                 nrs_tbf_cli_fini(cli);
1323                 cli = tmp;
1324         }
1325 out:
1326         *resp = &cli->tc_res;
1327
1328         return 1;
1329 }
1330
1331 /**
1332  * Called when releasing references to the resource hierachy obtained for a
1333  * request for scheduling using the TBF policy.
1334  *
1335  * \param[in] policy   the policy the resource belongs to
1336  * \param[in] res      the resource to be released
1337  */
1338 static void nrs_tbf_res_put(struct ptlrpc_nrs_policy *policy,
1339                             const struct ptlrpc_nrs_resource *res)
1340 {
1341         struct nrs_tbf_head   *head;
1342         struct nrs_tbf_client *cli;
1343
1344         /**
1345          * Do nothing for freeing parent, nrs_tbf_net resources
1346          */
1347         if (res->res_parent == NULL)
1348                 return;
1349
1350         cli = container_of(res, struct nrs_tbf_client, tc_res);
1351         head = container_of(res->res_parent, struct nrs_tbf_head, th_res);
1352
1353         head->th_ops->o_cli_put(head, cli);
1354 }
1355
1356 /**
1357  * Called when getting a request from the TBF policy for handling, or just
1358  * peeking; removes the request from the policy when it is to be handled.
1359  *
1360  * \param[in] policy The policy
1361  * \param[in] peek   When set, signifies that we just want to examine the
1362  *                   request, and not handle it, so the request is not removed
1363  *                   from the policy.
1364  * \param[in] force  Force the policy to return a request; unused in this
1365  *                   policy
1366  *
1367  * \retval The request to be handled; this is the next request in the TBF
1368  *         rule
1369  *
1370  * \see ptlrpc_nrs_req_get_nolock()
1371  * \see nrs_request_get()
1372  */
1373 static
1374 struct ptlrpc_nrs_request *nrs_tbf_req_get(struct ptlrpc_nrs_policy *policy,
1375                                            bool peek, bool force)
1376 {
1377         struct nrs_tbf_head       *head = policy->pol_private;
1378         struct ptlrpc_nrs_request *nrq = NULL;
1379         struct nrs_tbf_client     *cli;
1380         struct cfs_binheap_node   *node;
1381
1382         assert_spin_locked(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
1383
1384         if (!peek && policy->pol_nrs->nrs_throttling)
1385                 return NULL;
1386
1387         node = cfs_binheap_root(head->th_binheap);
1388         if (unlikely(node == NULL))
1389                 return NULL;
1390
1391         cli = container_of(node, struct nrs_tbf_client, tc_node);
1392         LASSERT(cli->tc_in_heap);
1393         if (peek) {
1394                 nrq = list_entry(cli->tc_list.next,
1395                                      struct ptlrpc_nrs_request,
1396                                      nr_u.tbf.tr_list);
1397         } else {
1398                 __u64 now = ktime_to_ns(ktime_get());
1399                 __u64 passed;
1400                 long  ntoken;
1401                 __u64 deadline;
1402
1403                 deadline = cli->tc_check_time +
1404                           cli->tc_nsecs;
1405                 LASSERT(now >= cli->tc_check_time);
1406                 passed = now - cli->tc_check_time;
1407                 ntoken = (passed * cli->tc_rpc_rate) / NSEC_PER_SEC;
1408                 ntoken += cli->tc_ntoken;
1409                 if (ntoken > cli->tc_depth)
1410                         ntoken = cli->tc_depth;
1411                 if (ntoken > 0) {
1412                         struct ptlrpc_request *req;
1413                         nrq = list_entry(cli->tc_list.next,
1414                                              struct ptlrpc_nrs_request,
1415                                              nr_u.tbf.tr_list);
1416                         req = container_of(nrq,
1417                                            struct ptlrpc_request,
1418                                            rq_nrq);
1419                         ntoken--;
1420                         cli->tc_ntoken = ntoken;
1421                         cli->tc_check_time = now;
1422                         list_del_init(&nrq->nr_u.tbf.tr_list);
1423                         if (list_empty(&cli->tc_list)) {
1424                                 cfs_binheap_remove(head->th_binheap,
1425                                                    &cli->tc_node);
1426                                 cli->tc_in_heap = false;
1427                         } else {
1428                                 cfs_binheap_relocate(head->th_binheap,
1429                                                      &cli->tc_node);
1430                         }
1431                         CDEBUG(D_RPCTRACE,
1432                                "NRS start %s request from %s, "
1433                                "seq: "LPU64"\n",
1434                                policy->pol_desc->pd_name,
1435                                libcfs_id2str(req->rq_peer),
1436                                nrq->nr_u.tbf.tr_sequence);
1437                 } else {
1438                         ktime_t time;
1439
1440                         policy->pol_nrs->nrs_throttling = 1;
1441                         head->th_deadline = deadline;
1442                         time = ktime_set(0, 0);
1443                         time = ktime_add_ns(time, deadline);
1444                         hrtimer_start(&head->th_timer, time, HRTIMER_MODE_ABS);
1445                 }
1446         }
1447
1448         return nrq;
1449 }
1450
1451 /**
1452  * Adds request \a nrq to \a policy's list of queued requests
1453  *
1454  * \param[in] policy The policy
1455  * \param[in] nrq    The request to add
1456  *
1457  * \retval 0 success; nrs_request_enqueue() assumes this function will always
1458  *                    succeed
1459  */
1460 static int nrs_tbf_req_add(struct ptlrpc_nrs_policy *policy,
1461                            struct ptlrpc_nrs_request *nrq)
1462 {
1463         struct nrs_tbf_head   *head;
1464         struct nrs_tbf_client *cli;
1465         int                    rc = 0;
1466
1467         assert_spin_locked(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
1468
1469         cli = container_of(nrs_request_resource(nrq),
1470                            struct nrs_tbf_client, tc_res);
1471         head = container_of(nrs_request_resource(nrq)->res_parent,
1472                             struct nrs_tbf_head, th_res);
1473         if (list_empty(&cli->tc_list)) {
1474                 LASSERT(!cli->tc_in_heap);
1475                 rc = cfs_binheap_insert(head->th_binheap, &cli->tc_node);
1476                 if (rc == 0) {
1477                         cli->tc_in_heap = true;
1478                         nrq->nr_u.tbf.tr_sequence = head->th_sequence++;
1479                         list_add_tail(&nrq->nr_u.tbf.tr_list,
1480                                           &cli->tc_list);
1481                         if (policy->pol_nrs->nrs_throttling) {
1482                                 __u64 deadline = cli->tc_check_time +
1483                                                  cli->tc_nsecs;
1484                                 if ((head->th_deadline > deadline) &&
1485                                     (hrtimer_try_to_cancel(&head->th_timer)
1486                                      >= 0)) {
1487                                         ktime_t time;
1488                                         head->th_deadline = deadline;
1489                                         time = ktime_set(0, 0);
1490                                         time = ktime_add_ns(time, deadline);
1491                                         hrtimer_start(&head->th_timer, time,
1492                                                       HRTIMER_MODE_ABS);
1493                                 }
1494                         }
1495                 }
1496         } else {
1497                 LASSERT(cli->tc_in_heap);
1498                 nrq->nr_u.tbf.tr_sequence = head->th_sequence++;
1499                 list_add_tail(&nrq->nr_u.tbf.tr_list,
1500                                   &cli->tc_list);
1501         }
1502         return rc;
1503 }
1504
1505 /**
1506  * Removes request \a nrq from \a policy's list of queued requests.
1507  *
1508  * \param[in] policy The policy
1509  * \param[in] nrq    The request to remove
1510  */
1511 static void nrs_tbf_req_del(struct ptlrpc_nrs_policy *policy,
1512                              struct ptlrpc_nrs_request *nrq)
1513 {
1514         struct nrs_tbf_head   *head;
1515         struct nrs_tbf_client *cli;
1516
1517         assert_spin_locked(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
1518
1519         cli = container_of(nrs_request_resource(nrq),
1520                            struct nrs_tbf_client, tc_res);
1521         head = container_of(nrs_request_resource(nrq)->res_parent,
1522                             struct nrs_tbf_head, th_res);
1523
1524         LASSERT(!list_empty(&nrq->nr_u.tbf.tr_list));
1525         list_del_init(&nrq->nr_u.tbf.tr_list);
1526         if (list_empty(&cli->tc_list)) {
1527                 cfs_binheap_remove(head->th_binheap,
1528                                    &cli->tc_node);
1529                 cli->tc_in_heap = false;
1530         } else {
1531                 cfs_binheap_relocate(head->th_binheap,
1532                                      &cli->tc_node);
1533         }
1534 }
1535
1536 /**
1537  * Prints a debug statement right before the request \a nrq stops being
1538  * handled.
1539  *
1540  * \param[in] policy The policy handling the request
1541  * \param[in] nrq    The request being handled
1542  *
1543  * \see ptlrpc_server_finish_request()
1544  * \see ptlrpc_nrs_req_stop_nolock()
1545  */
1546 static void nrs_tbf_req_stop(struct ptlrpc_nrs_policy *policy,
1547                               struct ptlrpc_nrs_request *nrq)
1548 {
1549         struct ptlrpc_request *req = container_of(nrq, struct ptlrpc_request,
1550                                                   rq_nrq);
1551
1552         assert_spin_locked(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
1553
1554         CDEBUG(D_RPCTRACE, "NRS stop %s request from %s, seq: "LPU64"\n",
1555                policy->pol_desc->pd_name, libcfs_id2str(req->rq_peer),
1556                nrq->nr_u.tbf.tr_sequence);
1557 }
1558
1559 #ifdef CONFIG_PROC_FS
1560
1561 /**
1562  * lprocfs interface
1563  */
1564
1565 /**
1566  * The maximum RPC rate.
1567  */
1568 #define LPROCFS_NRS_RATE_MAX            65535
1569
1570 static int
1571 ptlrpc_lprocfs_nrs_tbf_rule_seq_show(struct seq_file *m, void *data)
1572 {
1573         struct ptlrpc_service       *svc = m->private;
1574         int                          rc;
1575
1576         seq_printf(m, "regular_requests:\n");
1577         /**
1578          * Perform two separate calls to this as only one of the NRS heads'
1579          * policies may be in the ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED or
1580          * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPING state.
1581          */
1582         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_REG,
1583                                        NRS_POL_NAME_TBF,
1584                                        NRS_CTL_TBF_RD_RULE,
1585                                        false, m);
1586         if (rc == 0) {
1587                 /**
1588                  * -ENOSPC means buf in the parameter m is overflow, return 0
1589                  * here to let upper layer function seq_read alloc a larger
1590                  * memory area and do this process again.
1591                  */
1592         } else if (rc == -ENOSPC) {
1593                 return 0;
1594
1595                 /**
1596                  * Ignore -ENODEV as the regular NRS head's policy may be in the
1597                  * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED state.
1598                  */
1599         } else if (rc != -ENODEV) {
1600                 return rc;
1601         }
1602
1603         if (!nrs_svc_has_hp(svc))
1604                 goto no_hp;
1605
1606         seq_printf(m, "high_priority_requests:\n");
1607         rc = ptlrpc_nrs_policy_control(svc, PTLRPC_NRS_QUEUE_HP,
1608                                        NRS_POL_NAME_TBF,
1609                                        NRS_CTL_TBF_RD_RULE,
1610                                        false, m);
1611         if (rc == 0) {
1612                 /**
1613                  * -ENOSPC means buf in the parameter m is overflow, return 0
1614                  * here to let upper layer function seq_read alloc a larger
1615                  * memory area and do this process again.
1616                  */
1617         } else if (rc == -ENOSPC) {
1618                 return 0;
1619         }
1620
1621 no_hp:
1622
1623         return rc;
1624 }
1625
1626 static int nrs_tbf_id_parse(struct nrs_tbf_cmd *cmd, char **val)
1627 {
1628         int rc;
1629         char *token;
1630
1631         token = strsep(val, "}");
1632         if (*val == NULL)
1633                 GOTO(out, rc = -EINVAL);
1634
1635         if (strlen(token) <= 1 ||
1636             token[0] != '{')
1637                 GOTO(out, rc = -EINVAL);
1638         /* Skip '{' */
1639         token++;
1640
1641         /* Should be followed by ' ' or nothing */
1642         if ((*val)[0] == '\0')
1643                 *val = NULL;
1644         else if ((*val)[0] == ' ')
1645                 (*val)++;
1646         else
1647                 GOTO(out, rc = -EINVAL);
1648
1649         rc = nrs_tbf_jobid_parse(cmd, token);
1650         if (!rc)
1651                 cmd->tc_valid_types |= NRS_TBF_FLAG_JOBID;
1652
1653         rc = nrs_tbf_nid_parse(cmd, token);
1654         if (!rc)
1655                 cmd->tc_valid_types |= NRS_TBF_FLAG_NID;
1656
1657         if (!cmd->tc_valid_types)
1658                 rc = -EINVAL;
1659         else
1660                 rc = 0;
1661 out:
1662         return rc;
1663 }
1664
1665
1666 static void nrs_tbf_cmd_fini(struct nrs_tbf_cmd *cmd)
1667 {
1668         if (cmd->tc_valid_types & NRS_TBF_FLAG_JOBID)
1669                 nrs_tbf_jobid_cmd_fini(cmd);
1670         if (cmd->tc_valid_types & NRS_TBF_FLAG_NID)
1671                 nrs_tbf_nid_cmd_fini(cmd);
1672 }
1673
1674 static struct nrs_tbf_cmd *
1675 nrs_tbf_parse_cmd(char *buffer, unsigned long count)
1676 {
1677         static struct nrs_tbf_cmd *cmd;
1678         char                      *token;
1679         char                      *val;
1680         int                        i;
1681         int                        rc = 0;
1682
1683         OBD_ALLOC_PTR(cmd);
1684         if (cmd == NULL)
1685                 GOTO(out, rc = -ENOMEM);
1686
1687         val = buffer;
1688         token = strsep(&val, " ");
1689         if (val == NULL || strlen(val) == 0)
1690                 GOTO(out_free_cmd, rc = -EINVAL);
1691
1692         /* Type of the command */
1693         if (strcmp(token, "start") == 0)
1694                 cmd->tc_cmd = NRS_CTL_TBF_START_RULE;
1695         else if (strcmp(token, "stop") == 0)
1696                 cmd->tc_cmd = NRS_CTL_TBF_STOP_RULE;
1697         else if (strcmp(token, "change") == 0)
1698                 cmd->tc_cmd = NRS_CTL_TBF_CHANGE_RATE;
1699         else
1700                 GOTO(out_free_cmd, rc = -EINVAL);
1701
1702         /* Name of the rule */
1703         token = strsep(&val, " ");
1704         if (val == NULL) {
1705                 /**
1706                  * Stop comand only need name argument,
1707                  * But other commands need ID or rate argument.
1708                  */
1709                 if (cmd->tc_cmd != NRS_CTL_TBF_STOP_RULE)
1710                         GOTO(out_free_cmd, rc = -EINVAL);
1711         }
1712
1713         for (i = 0; i < strlen(token); i++) {
1714                 if ((!isalnum(token[i])) &&
1715                     (token[i] != '_'))
1716                         GOTO(out_free_cmd, rc = -EINVAL);
1717         }
1718         cmd->tc_name = token;
1719
1720         if (cmd->tc_cmd == NRS_CTL_TBF_START_RULE) {
1721                 /* List of ID */
1722                 LASSERT(val);
1723                 rc = nrs_tbf_id_parse(cmd, &val);
1724                 if (rc)
1725                         GOTO(out_free_cmd, rc);
1726         }
1727
1728         if (val != NULL) {
1729                 if (cmd->tc_cmd == NRS_CTL_TBF_STOP_RULE ||
1730                     strlen(val) == 0 || !isdigit(val[0]))
1731                         GOTO(out_free_nid, rc = -EINVAL);
1732
1733                 cmd->tc_rpc_rate = simple_strtoull(val, NULL, 10);
1734                 if (cmd->tc_rpc_rate <= 0 ||
1735                     cmd->tc_rpc_rate >= LPROCFS_NRS_RATE_MAX)
1736                         GOTO(out_free_nid, rc = -EINVAL);
1737         } else {
1738                 if (cmd->tc_cmd == NRS_CTL_TBF_CHANGE_RATE)
1739                         GOTO(out_free_nid, rc = -EINVAL);
1740                 /* No RPC rate given */
1741                 cmd->tc_rpc_rate = tbf_rate;
1742         }
1743         goto out;
1744 out_free_nid:
1745         nrs_tbf_cmd_fini(cmd);
1746 out_free_cmd:
1747         OBD_FREE_PTR(cmd);
1748 out:
1749         if (rc)
1750                 cmd = ERR_PTR(rc);
1751         return cmd;
1752 }
1753
1754 extern struct nrs_core nrs_core;
1755 #define LPROCFS_WR_NRS_TBF_MAX_CMD (4096)
1756 static ssize_t
1757 ptlrpc_lprocfs_nrs_tbf_rule_seq_write(struct file *file,
1758                                       const char __user *buffer,
1759                                       size_t count, loff_t *off)
1760 {
1761         struct seq_file           *m = file->private_data;
1762         struct ptlrpc_service     *svc = m->private;
1763         char                      *kernbuf;
1764         char                      *val;
1765         int                        rc;
1766         static struct nrs_tbf_cmd *cmd;
1767         enum ptlrpc_nrs_queue_type queue = PTLRPC_NRS_QUEUE_BOTH;
1768         unsigned long              length;
1769         char                      *token;
1770
1771         OBD_ALLOC(kernbuf, LPROCFS_WR_NRS_TBF_MAX_CMD);
1772         if (kernbuf == NULL)
1773                 GOTO(out, rc = -ENOMEM);
1774
1775         if (count > LPROCFS_WR_NRS_TBF_MAX_CMD - 1)
1776                 GOTO(out_free_kernbuff, rc = -EINVAL);
1777
1778         if (copy_from_user(kernbuf, buffer, count))
1779                 GOTO(out_free_kernbuff, rc = -EFAULT);
1780
1781         val = kernbuf;
1782         token = strsep(&val, " ");
1783         if (val == NULL)
1784                 GOTO(out_free_kernbuff, rc = -EINVAL);
1785
1786         if (strcmp(token, "reg") == 0) {
1787                 queue = PTLRPC_NRS_QUEUE_REG;
1788         } else if (strcmp(token, "hp") == 0) {
1789                 queue = PTLRPC_NRS_QUEUE_HP;
1790         } else {
1791                 kernbuf[strlen(token)] = ' ';
1792                 val = kernbuf;
1793         }
1794         length = strlen(val);
1795
1796         if (length == 0)
1797                 GOTO(out_free_kernbuff, rc = -EINVAL);
1798
1799         if (queue == PTLRPC_NRS_QUEUE_HP && !nrs_svc_has_hp(svc))
1800                 GOTO(out_free_kernbuff, rc = -ENODEV);
1801         else if (queue == PTLRPC_NRS_QUEUE_BOTH && !nrs_svc_has_hp(svc))
1802                 queue = PTLRPC_NRS_QUEUE_REG;
1803
1804         cmd = nrs_tbf_parse_cmd(val, length);
1805         if (IS_ERR(cmd))
1806                 GOTO(out_free_kernbuff, rc = PTR_ERR(cmd));
1807
1808         /**
1809          * Serialize NRS core lprocfs operations with policy registration/
1810          * unregistration.
1811          */
1812         mutex_lock(&nrs_core.nrs_mutex);
1813         rc = ptlrpc_nrs_policy_control(svc, queue,
1814                                        NRS_POL_NAME_TBF,
1815                                        NRS_CTL_TBF_WR_RULE,
1816                                        false, cmd);
1817         mutex_unlock(&nrs_core.nrs_mutex);
1818
1819         nrs_tbf_cmd_fini(cmd);
1820         OBD_FREE_PTR(cmd);
1821 out_free_kernbuff:
1822         OBD_FREE(kernbuf, LPROCFS_WR_NRS_TBF_MAX_CMD);
1823 out:
1824         return rc ? rc : count;
1825 }
1826 LPROC_SEQ_FOPS(ptlrpc_lprocfs_nrs_tbf_rule);
1827
1828 /**
1829  * Initializes a TBF policy's lprocfs interface for service \a svc
1830  *
1831  * \param[in] svc the service
1832  *
1833  * \retval 0    success
1834  * \retval != 0 error
1835  */
1836 static int nrs_tbf_lprocfs_init(struct ptlrpc_service *svc)
1837 {
1838         struct lprocfs_vars nrs_tbf_lprocfs_vars[] = {
1839                 { .name         = "nrs_tbf_rule",
1840                   .fops         = &ptlrpc_lprocfs_nrs_tbf_rule_fops,
1841                   .data = svc },
1842                 { NULL }
1843         };
1844
1845         if (svc->srv_procroot == NULL)
1846                 return 0;
1847
1848         return lprocfs_add_vars(svc->srv_procroot, nrs_tbf_lprocfs_vars, NULL);
1849 }
1850
1851 /**
1852  * Cleans up a TBF policy's lprocfs interface for service \a svc
1853  *
1854  * \param[in] svc the service
1855  */
1856 static void nrs_tbf_lprocfs_fini(struct ptlrpc_service *svc)
1857 {
1858         if (svc->srv_procroot == NULL)
1859                 return;
1860
1861         lprocfs_remove_proc_entry("nrs_tbf_rule", svc->srv_procroot);
1862 }
1863
1864 #endif /* CONFIG_PROC_FS */
1865
1866 /**
1867  * TBF policy operations
1868  */
1869 static const struct ptlrpc_nrs_pol_ops nrs_tbf_ops = {
1870         .op_policy_start        = nrs_tbf_start,
1871         .op_policy_stop         = nrs_tbf_stop,
1872         .op_policy_ctl          = nrs_tbf_ctl,
1873         .op_res_get             = nrs_tbf_res_get,
1874         .op_res_put             = nrs_tbf_res_put,
1875         .op_req_get             = nrs_tbf_req_get,
1876         .op_req_enqueue         = nrs_tbf_req_add,
1877         .op_req_dequeue         = nrs_tbf_req_del,
1878         .op_req_stop            = nrs_tbf_req_stop,
1879 #ifdef CONFIG_PROC_FS
1880         .op_lprocfs_init        = nrs_tbf_lprocfs_init,
1881         .op_lprocfs_fini        = nrs_tbf_lprocfs_fini,
1882 #endif
1883 };
1884
1885 /**
1886  * TBF policy configuration
1887  */
1888 struct ptlrpc_nrs_pol_conf nrs_conf_tbf = {
1889         .nc_name                = NRS_POL_NAME_TBF,
1890         .nc_ops                 = &nrs_tbf_ops,
1891         .nc_compat              = nrs_policy_compat_all,
1892 };
1893
1894 /** @} tbf */
1895
1896 /** @} nrs */
1897
1898 #endif /* HAVE_SERVER_SUPPORT */