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