Whamcloud - gitweb
LU-9611 lod: allow -1 for default stripe count/offset
[fs/lustre-release.git] / lustre / ptlrpc / nrs_tbf.c
index 64e5f97..c858b6c 100644 (file)
@@ -254,6 +254,7 @@ nrs_tbf_cli_init(struct nrs_tbf_head *head,
 {
        struct nrs_tbf_rule *rule;
 
+       memset(cli, 0, sizeof(*cli));
        cli->tc_in_heap = false;
        head->th_ops->o_cli_init(cli, req);
        INIT_LIST_HEAD(&cli->tc_list);
@@ -347,6 +348,9 @@ nrs_tbf_rule_start(struct ptlrpc_nrs_policy *policy,
                head->th_rule = rule;
        }
 
+       CDEBUG(D_RPCTRACE, "TBF starts rule@%p rate %llu gen %llu\n",
+              rule, rule->tr_rpc_rate, rule->tr_generation);
+
        return 0;
 }
 
@@ -624,13 +628,10 @@ nrs_tbf_jobid_hash_lookup(struct cfs_hash *hs,
        struct hlist_node *hnode;
        struct nrs_tbf_client *cli;
 
-       /* cfs_hash_bd_peek_locked is a somehow "internal" function
-        * of cfs_hash, it doesn't add refcount on object. */
-       hnode = cfs_hash_bd_peek_locked(hs, bd, (void *)jobid);
+       hnode = cfs_hash_bd_lookup_locked(hs, bd, (void *)jobid);
        if (hnode == NULL)
                return NULL;
 
-       cfs_hash_get(hs, hnode);
        cli = container_of0(hnode, struct nrs_tbf_client, tc_hnode);
        if (!list_empty(&cli->tc_lru))
                list_del_init(&cli->tc_lru);
@@ -786,6 +787,10 @@ nrs_tbf_jobid_startup(struct ptlrpc_nrs_policy *policy,
        start.tc_name = NRS_TBF_DEFAULT_RULE;
        INIT_LIST_HEAD(&start.u.tc_start.ts_jobids);
        rc = nrs_tbf_rule_start(policy, head, &start);
+       if (rc) {
+               cfs_hash_putref(head->th_cli_hash);
+               head->th_cli_hash = NULL;
+       }
 
        return rc;
 }
@@ -807,9 +812,11 @@ nrs_tbf_jobid_list_free(struct list_head *jobid_list)
 }
 
 static int
-nrs_tbf_jobid_list_add(const struct cfs_lstr *id, struct list_head *jobid_list)
+nrs_tbf_jobid_list_add(struct cfs_lstr *id, struct list_head *jobid_list)
 {
        struct nrs_tbf_jobid *jobid;
+       struct cfs_lstr res;
+       int rc;
 
        OBD_ALLOC(jobid, sizeof(struct nrs_tbf_jobid));
        if (jobid == NULL)
@@ -822,17 +829,62 @@ nrs_tbf_jobid_list_add(const struct cfs_lstr *id, struct list_head *jobid_list)
        }
 
        memcpy(jobid->tj_id, id->ls_str, id->ls_len);
+       rc = cfs_gettok(id, '*', &res);
+       if (rc == 0)
+               jobid->tj_match_flag = NRS_TBF_MATCH_FULL;
+       else
+               jobid->tj_match_flag = NRS_TBF_MATCH_WILDCARD;
+
        list_add_tail(&jobid->tj_linkage, jobid_list);
        return 0;
 }
 
+static bool
+cfs_match_wildcard(const char *pattern, const char *content)
+{
+       if (*pattern == '\0' && *content == '\0')
+               return true;
+
+       if (*pattern == '*' && *(pattern + 1) != '\0' && *content == '\0')
+               return false;
+
+       while (*pattern == *content) {
+               pattern++;
+               content++;
+               if (*pattern == '\0' && *content == '\0')
+                       return true;
+
+               if (*pattern == '*' && *(pattern + 1) != '\0' &&
+                   *content == '\0')
+                       return false;
+       }
+
+       if (*pattern == '*')
+               return (cfs_match_wildcard(pattern + 1, content) ||
+                       cfs_match_wildcard(pattern, content + 1));
+
+       return false;
+}
+
+static inline bool
+nrs_tbf_jobid_match(const struct nrs_tbf_jobid *jobid, const char *id)
+{
+       if (jobid->tj_match_flag == NRS_TBF_MATCH_FULL)
+               return strcmp(jobid->tj_id, id) == 0;
+
+       if (jobid->tj_match_flag == NRS_TBF_MATCH_WILDCARD)
+               return cfs_match_wildcard(jobid->tj_id, id);
+
+       return false;
+}
+
 static int
 nrs_tbf_jobid_list_match(struct list_head *jobid_list, char *id)
 {
        struct nrs_tbf_jobid *jobid;
 
        list_for_each_entry(jobid, jobid_list, tj_linkage) {
-               if (strcmp(id, jobid->tj_id) == 0)
+               if (nrs_tbf_jobid_match(jobid, id))
                        return 1;
        }
        return 0;
@@ -992,8 +1044,8 @@ static struct nrs_tbf_ops nrs_tbf_jobid_ops = {
  * This uses ptlrpc_request::rq_peer.nid as its key, in order to hash
  * nrs_tbf_client objects.
  */
-#define NRS_TBF_NID_BKT_BITS    8
-#define NRS_TBF_NID_BITS        16
+#define NRS_TBF_NID_BKT_BITS   8
+#define NRS_TBF_NID_BITS       16
 
 static unsigned nrs_tbf_nid_hop_hash(struct cfs_hash *hs, const void *key,
                                  unsigned mask)
@@ -1115,6 +1167,10 @@ nrs_tbf_nid_startup(struct ptlrpc_nrs_policy *policy,
        start.tc_name = NRS_TBF_DEFAULT_RULE;
        INIT_LIST_HEAD(&start.u.tc_start.ts_nids);
        rc = nrs_tbf_rule_start(policy, head, &start);
+       if (rc) {
+               cfs_hash_putref(head->th_cli_hash);
+               head->th_cli_hash = NULL;
+       }
 
        return rc;
 }
@@ -1229,220 +1285,1105 @@ static struct nrs_tbf_ops nrs_tbf_nid_ops = {
        .o_rule_fini = nrs_tbf_nid_rule_fini,
 };
 
-/**
- * Is called before the policy transitions into
- * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED; allocates and initializes a
- * policy-specific private data structure.
- *
- * \param[in] policy The policy to start
- *
- * \retval -ENOMEM OOM error
- * \retval  0     success
- *
- * \see nrs_policy_register()
- * \see nrs_policy_ctl()
- */
-static int nrs_tbf_start(struct ptlrpc_nrs_policy *policy, char *arg)
+static unsigned nrs_tbf_hop_hash(struct cfs_hash *hs, const void *key,
+                                unsigned mask)
 {
-       struct nrs_tbf_head     *head;
-       struct nrs_tbf_ops      *ops;
-       __u32                    type;
-       int rc = 0;
-
-       if (arg == NULL || strlen(arg) > NRS_TBF_TYPE_MAX_LEN)
-               GOTO(out, rc = -EINVAL);
+       return cfs_hash_djb2_hash(key, strlen(key), mask);
+}
 
-       if (strcmp(arg, NRS_TBF_TYPE_NID) == 0) {
-               ops = &nrs_tbf_nid_ops;
-               type = NRS_TBF_FLAG_NID;
-       } else if (strcmp(arg, NRS_TBF_TYPE_JOBID) == 0) {
-               ops = &nrs_tbf_jobid_ops;
-               type = NRS_TBF_FLAG_JOBID;
-       } else
-               GOTO(out, rc = -ENOTSUPP);
+static int nrs_tbf_hop_keycmp(const void *key, struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
 
-       OBD_CPT_ALLOC_PTR(head, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
-       if (head == NULL)
-               GOTO(out, rc = -ENOMEM);
+       return (strcmp(cli->tc_key, key) == 0);
+}
 
-       memcpy(head->th_type, arg, strlen(arg));
-       head->th_type[strlen(arg)] = '\0';
-       head->th_ops = ops;
-       head->th_type_flag = type;
+static void *nrs_tbf_hop_key(struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
+       return cli->tc_key;
+}
 
-       head->th_binheap = cfs_binheap_create(&nrs_tbf_heap_ops,
-                                             CBH_FLAG_ATOMIC_GROW, 4096, NULL,
-                                             nrs_pol2cptab(policy),
-                                             nrs_pol2cptid(policy));
-       if (head->th_binheap == NULL)
-               GOTO(out_free_head, rc = -ENOMEM);
+static void *nrs_tbf_hop_object(struct hlist_node *hnode)
+{
+       return hlist_entry(hnode, struct nrs_tbf_client, tc_hnode);
+}
 
-       atomic_set(&head->th_rule_sequence, 0);
-       spin_lock_init(&head->th_rule_lock);
-       INIT_LIST_HEAD(&head->th_list);
-       hrtimer_init(&head->th_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
-       head->th_timer.function = nrs_tbf_timer_cb;
-       rc = head->th_ops->o_startup(policy, head);
-       if (rc)
-               GOTO(out_free_heap, rc);
+static void nrs_tbf_hop_get(struct cfs_hash *hs, struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
 
-       policy->pol_private = head;
-       return 0;
-out_free_heap:
-       cfs_binheap_destroy(head->th_binheap);
-out_free_head:
-       OBD_FREE_PTR(head);
-out:
-       return rc;
+       atomic_inc(&cli->tc_ref);
 }
 
-/**
- * Is called before the policy transitions into
- * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED; deallocates the policy-specific
- * private data structure.
- *
- * \param[in] policy The policy to stop
- *
- * \see nrs_policy_stop0()
- */
-static void nrs_tbf_stop(struct ptlrpc_nrs_policy *policy)
+static void nrs_tbf_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
 {
-       struct nrs_tbf_head *head = policy->pol_private;
-       struct ptlrpc_nrs *nrs = policy->pol_nrs;
-       struct nrs_tbf_rule *rule, *n;
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
 
-       LASSERT(head != NULL);
-       LASSERT(head->th_cli_hash != NULL);
-       hrtimer_cancel(&head->th_timer);
-       /* Should cleanup hash first before free rules */
-       cfs_hash_putref(head->th_cli_hash);
-       list_for_each_entry_safe(rule, n, &head->th_list, tr_linkage) {
-               list_del_init(&rule->tr_linkage);
-               nrs_tbf_rule_put(rule);
-       }
-       LASSERT(list_empty(&head->th_list));
-       LASSERT(head->th_binheap != NULL);
-       LASSERT(cfs_binheap_is_empty(head->th_binheap));
-       cfs_binheap_destroy(head->th_binheap);
-       OBD_FREE_PTR(head);
-       nrs->nrs_throttling = 0;
-       wake_up(&policy->pol_nrs->nrs_svcpt->scp_waitq);
+       atomic_dec(&cli->tc_ref);
 }
 
-/**
- * Performs a policy-specific ctl function on TBF policy instances; similar
- * to ioctl.
- *
- * \param[in]    policy the policy instance
- * \param[in]    opc    the opcode
- * \param[in,out] arg   used for passing parameters and information
- *
- * \pre assert_spin_locked(&policy->pol_nrs->->nrs_lock)
- * \post assert_spin_locked(&policy->pol_nrs->->nrs_lock)
- *
- * \retval 0   operation carried out successfully
- * \retval -ve error
- */
-static int nrs_tbf_ctl(struct ptlrpc_nrs_policy *policy,
-                      enum ptlrpc_nrs_ctl opc,
-                      void *arg)
+static void nrs_tbf_hop_exit(struct cfs_hash *hs, struct hlist_node *hnode)
+
 {
-       int rc = 0;
-       ENTRY;
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
 
-       assert_spin_locked(&policy->pol_nrs->nrs_lock);
+       LASSERT(atomic_read(&cli->tc_ref) == 0);
+       nrs_tbf_cli_fini(cli);
+}
 
-       switch ((enum nrs_ctl_tbf)opc) {
-       default:
-               RETURN(-EINVAL);
+static struct cfs_hash_ops nrs_tbf_hash_ops = {
+       .hs_hash        = nrs_tbf_hop_hash,
+       .hs_keycmp      = nrs_tbf_hop_keycmp,
+       .hs_key         = nrs_tbf_hop_key,
+       .hs_object      = nrs_tbf_hop_object,
+       .hs_get         = nrs_tbf_hop_get,
+       .hs_put         = nrs_tbf_hop_put,
+       .hs_put_locked  = nrs_tbf_hop_put,
+       .hs_exit        = nrs_tbf_hop_exit,
+};
 
-       /**
-        * Read RPC rate size of a policy instance.
-        */
-       case NRS_CTL_TBF_RD_RULE: {
-               struct nrs_tbf_head *head = policy->pol_private;
-               struct seq_file *m = (struct seq_file *) arg;
-               struct ptlrpc_service_part *svcpt;
+#define NRS_TBF_GENERIC_BKT_BITS       10
+#define NRS_TBF_GENERIC_HASH_FLAGS     (CFS_HASH_SPIN_BKTLOCK | \
+                                       CFS_HASH_NO_ITEMREF | \
+                                       CFS_HASH_DEPTH)
 
-               svcpt = policy->pol_nrs->nrs_svcpt;
-               seq_printf(m, "CPT %d:\n", svcpt->scp_cpt);
+static int
+nrs_tbf_startup(struct ptlrpc_nrs_policy *policy, struct nrs_tbf_head *head)
+{
+       struct nrs_tbf_cmd       start;
+       struct nrs_tbf_bucket   *bkt;
+       int                      bits;
+       int                      i;
+       int                      rc;
+       struct cfs_hash_bd       bd;
 
-               rc = nrs_tbf_rule_dump_all(head, m);
-               }
-               break;
+       bits = nrs_tbf_jobid_hash_order();
+       if (bits < NRS_TBF_GENERIC_BKT_BITS)
+               bits = NRS_TBF_GENERIC_BKT_BITS;
+       head->th_cli_hash = cfs_hash_create("nrs_tbf_hash",
+                                           bits, bits,
+                                           NRS_TBF_GENERIC_BKT_BITS,
+                                           sizeof(*bkt), 0, 0,
+                                           &nrs_tbf_hash_ops,
+                                           NRS_TBF_GENERIC_HASH_FLAGS);
+       if (head->th_cli_hash == NULL)
+               return -ENOMEM;
 
-       /**
-        * Write RPC rate of a policy instance.
-        */
-       case NRS_CTL_TBF_WR_RULE: {
-               struct nrs_tbf_head *head = policy->pol_private;
-               struct nrs_tbf_cmd *cmd;
+       cfs_hash_for_each_bucket(head->th_cli_hash, &bd, i) {
+               bkt = cfs_hash_bd_extra_get(head->th_cli_hash, &bd);
+               INIT_LIST_HEAD(&bkt->ntb_lru);
+       }
 
-               cmd = (struct nrs_tbf_cmd *)arg;
-               rc = nrs_tbf_command(policy,
-                                    head,
-                                    cmd);
-               }
-               break;
-       /**
-        * Read the TBF policy type of a policy instance.
-        */
-       case NRS_CTL_TBF_RD_TYPE_FLAG: {
-               struct nrs_tbf_head *head = policy->pol_private;
+       memset(&start, 0, sizeof(start));
+       start.u.tc_start.ts_conds_str = "*";
 
-               *(__u32 *)arg = head->th_type_flag;
-               }
-               break;
-       }
+       start.u.tc_start.ts_rpc_rate = tbf_rate;
+       start.u.tc_start.ts_rule_flags = NTRS_DEFAULT;
+       start.tc_name = NRS_TBF_DEFAULT_RULE;
+       INIT_LIST_HEAD(&start.u.tc_start.ts_conds);
+       rc = nrs_tbf_rule_start(policy, head, &start);
+       if (rc)
+               cfs_hash_putref(head->th_cli_hash);
 
-       RETURN(rc);
+       return rc;
 }
 
-/**
- * Is called for obtaining a TBF policy resource.
- *
- * \param[in]  policy    The policy on which the request is being asked for
- * \param[in]  nrq       The request for which resources are being taken
- * \param[in]  parent    Parent resource, unused in this policy
- * \param[out] resp      Resources references are placed in this array
- * \param[in]  moving_req Signifies limited caller context; unused in this
- *                       policy
- *
- *
- * \see nrs_resource_get_safe()
- */
-static int nrs_tbf_res_get(struct ptlrpc_nrs_policy *policy,
-                          struct ptlrpc_nrs_request *nrq,
-                          const struct ptlrpc_nrs_resource *parent,
-                          struct ptlrpc_nrs_resource **resp,
-                          bool moving_req)
+static struct nrs_tbf_client *
+nrs_tbf_cli_hash_lookup(struct cfs_hash *hs, struct cfs_hash_bd *bd,
+                       const char *key)
 {
-       struct nrs_tbf_head   *head;
+       struct hlist_node *hnode;
        struct nrs_tbf_client *cli;
-       struct nrs_tbf_client *tmp;
-       struct ptlrpc_request *req;
-
-       if (parent == NULL) {
-               *resp = &((struct nrs_tbf_head *)policy->pol_private)->th_res;
-               return 0;
-       }
 
-       head = container_of(parent, struct nrs_tbf_head, th_res);
-       req = container_of(nrq, struct ptlrpc_request, rq_nrq);
-       cli = head->th_ops->o_cli_find(head, req);
-       if (cli != NULL) {
-               spin_lock(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
-               LASSERT(cli->tc_rule);
-               if (cli->tc_rule_sequence !=
-                   atomic_read(&head->th_rule_sequence) ||
-                   cli->tc_rule->tr_flags & NTRS_STOPPING) {
-                       struct nrs_tbf_rule *rule;
+       hnode = cfs_hash_bd_lookup_locked(hs, bd, (void *)key);
+       if (hnode == NULL)
+               return NULL;
 
-                       rule = nrs_tbf_rule_match(head, cli);
-                       if (rule != cli->tc_rule)
-                               nrs_tbf_cli_reset(head, rule, cli);
-                       else
+       cli = container_of0(hnode, struct nrs_tbf_client, tc_hnode);
+       if (!list_empty(&cli->tc_lru))
+               list_del_init(&cli->tc_lru);
+       return cli;
+}
+
+static struct nrs_tbf_client *
+nrs_tbf_cli_find(struct nrs_tbf_head *head, struct ptlrpc_request *req)
+{
+       struct nrs_tbf_client *cli;
+       struct cfs_hash *hs = head->th_cli_hash;
+       struct cfs_hash_bd bd;
+       char keystr[NRS_TBF_KEY_LEN] = { '\0' };
+       const char *jobid;
+       __u32 opc;
+
+       jobid = lustre_msg_get_jobid(req->rq_reqmsg);
+       if (jobid == NULL)
+               jobid = NRS_TBF_JOBID_NULL;
+       opc = lustre_msg_get_opc(req->rq_reqmsg);
+       snprintf(keystr, sizeof(keystr), "%s_%s_%d", jobid,
+                libcfs_nid2str(req->rq_peer.nid), opc);
+       LASSERT(strlen(keystr) < NRS_TBF_KEY_LEN);
+       cfs_hash_bd_get_and_lock(hs, (void *)keystr, &bd, 1);
+       cli = nrs_tbf_cli_hash_lookup(hs, &bd, keystr);
+       cfs_hash_bd_unlock(hs, &bd, 1);
+
+       return cli;
+}
+
+static struct nrs_tbf_client *
+nrs_tbf_cli_findadd(struct nrs_tbf_head *head,
+                   struct nrs_tbf_client *cli)
+{
+       const char              *key;
+       struct nrs_tbf_client   *ret;
+       struct cfs_hash         *hs = head->th_cli_hash;
+       struct cfs_hash_bd       bd;
+
+       key = cli->tc_key;
+       cfs_hash_bd_get_and_lock(hs, (void *)key, &bd, 1);
+       ret = nrs_tbf_cli_hash_lookup(hs, &bd, key);
+       if (ret == NULL) {
+               cfs_hash_bd_add_locked(hs, &bd, &cli->tc_hnode);
+               ret = cli;
+       }
+       cfs_hash_bd_unlock(hs, &bd, 1);
+
+       return ret;
+}
+
+static void
+nrs_tbf_cli_put(struct nrs_tbf_head *head, struct nrs_tbf_client *cli)
+{
+       struct cfs_hash_bd       bd;
+       struct cfs_hash         *hs = head->th_cli_hash;
+       struct nrs_tbf_bucket   *bkt;
+       int                      hw;
+       struct list_head         zombies;
+
+       INIT_LIST_HEAD(&zombies);
+       cfs_hash_bd_get(hs, &cli->tc_key, &bd);
+       bkt = cfs_hash_bd_extra_get(hs, &bd);
+       if (!cfs_hash_bd_dec_and_lock(hs, &bd, &cli->tc_ref))
+               return;
+       LASSERT(list_empty(&cli->tc_lru));
+       list_add_tail(&cli->tc_lru, &bkt->ntb_lru);
+
+       /**
+        * Check and purge the LRU, there is at least one client in the LRU.
+        */
+       hw = tbf_jobid_cache_size >> (hs->hs_cur_bits - hs->hs_bkt_bits);
+       while (cfs_hash_bd_count_get(&bd) > hw) {
+               if (unlikely(list_empty(&bkt->ntb_lru)))
+                       break;
+               cli = list_entry(bkt->ntb_lru.next,
+                                struct nrs_tbf_client,
+                                tc_lru);
+               LASSERT(atomic_read(&cli->tc_ref) == 0);
+               cfs_hash_bd_del_locked(hs, &bd, &cli->tc_hnode);
+               list_move(&cli->tc_lru, &zombies);
+       }
+       cfs_hash_bd_unlock(head->th_cli_hash, &bd, 1);
+
+       while (!list_empty(&zombies)) {
+               cli = container_of0(zombies.next,
+                                   struct nrs_tbf_client, tc_lru);
+               list_del_init(&cli->tc_lru);
+               nrs_tbf_cli_fini(cli);
+       }
+}
+
+static void
+nrs_tbf_generic_cli_init(struct nrs_tbf_client *cli,
+                        struct ptlrpc_request *req)
+{
+       char keystr[NRS_TBF_KEY_LEN];
+       const char *jobid;
+       __u32 opc;
+
+       jobid = lustre_msg_get_jobid(req->rq_reqmsg);
+       if (jobid == NULL)
+               jobid = NRS_TBF_JOBID_NULL;
+       opc = lustre_msg_get_opc(req->rq_reqmsg);
+       snprintf(keystr, sizeof(keystr), "%s_%s_%d", jobid,
+                libcfs_nid2str(req->rq_peer.nid), opc);
+
+       LASSERT(strlen(keystr) < NRS_TBF_KEY_LEN);
+       INIT_LIST_HEAD(&cli->tc_lru);
+       memcpy(cli->tc_key, keystr, strlen(keystr));
+       memcpy(cli->tc_jobid, jobid, strlen(jobid));
+       cli->tc_nid = req->rq_peer.nid;
+       cli->tc_opcode = opc;
+}
+
+static void
+nrs_tbf_expression_free(struct nrs_tbf_expression *expr)
+{
+       LASSERT(expr->te_field >= NRS_TBF_FIELD_NID &&
+               expr->te_field < NRS_TBF_FIELD_MAX);
+       switch (expr->te_field) {
+       case NRS_TBF_FIELD_NID:
+               cfs_free_nidlist(&expr->te_cond);
+               break;
+       case NRS_TBF_FIELD_JOBID:
+               nrs_tbf_jobid_list_free(&expr->te_cond);
+               break;
+       case NRS_TBF_FIELD_OPCODE:
+               CFS_FREE_BITMAP(expr->te_opcodes);
+               break;
+       default:
+               LBUG();
+       }
+       OBD_FREE_PTR(expr);
+}
+
+static void
+nrs_tbf_conjunction_free(struct nrs_tbf_conjunction *conjunction)
+{
+       struct nrs_tbf_expression *expression;
+       struct nrs_tbf_expression *n;
+
+       LASSERT(list_empty(&conjunction->tc_linkage));
+       list_for_each_entry_safe(expression, n,
+                                &conjunction->tc_expressions,
+                                te_linkage) {
+               list_del_init(&expression->te_linkage);
+               nrs_tbf_expression_free(expression);
+       }
+       OBD_FREE_PTR(conjunction);
+}
+
+static void
+nrs_tbf_conds_free(struct list_head *cond_list)
+{
+       struct nrs_tbf_conjunction *conjunction;
+       struct nrs_tbf_conjunction *n;
+
+       list_for_each_entry_safe(conjunction, n, cond_list, tc_linkage) {
+               list_del_init(&conjunction->tc_linkage);
+               nrs_tbf_conjunction_free(conjunction);
+       }
+}
+
+static void
+nrs_tbf_generic_cmd_fini(struct nrs_tbf_cmd *cmd)
+{
+       if (!list_empty(&cmd->u.tc_start.ts_conds))
+               nrs_tbf_conds_free(&cmd->u.tc_start.ts_conds);
+       if (cmd->u.tc_start.ts_conds_str)
+               OBD_FREE(cmd->u.tc_start.ts_conds_str,
+                        strlen(cmd->u.tc_start.ts_conds_str) + 1);
+}
+
+#define NRS_TBF_DISJUNCTION_DELIM      (',')
+#define NRS_TBF_CONJUNCTION_DELIM      ('&')
+#define NRS_TBF_EXPRESSION_DELIM       ('=')
+
+static inline bool
+nrs_tbf_check_field(struct cfs_lstr *field, char *str)
+{
+       int len = strlen(str);
+
+       return (field->ls_len == len &&
+               strncmp(field->ls_str, str, len) == 0);
+}
+
+static int
+nrs_tbf_opcode_list_parse(char *str, int len, struct cfs_bitmap **bitmaptr);
+
+static int
+nrs_tbf_expression_parse(struct cfs_lstr *src, struct list_head *cond_list)
+{
+       struct nrs_tbf_expression *expr;
+       struct cfs_lstr field;
+       int rc = 0;
+
+       OBD_ALLOC(expr, sizeof(struct nrs_tbf_expression));
+       if (expr == NULL)
+               return -ENOMEM;
+
+       rc = cfs_gettok(src, NRS_TBF_EXPRESSION_DELIM, &field);
+       if (rc == 0 || src->ls_len <= 2 || src->ls_str[0] != '{' ||
+           src->ls_str[src->ls_len - 1] != '}')
+               GOTO(out, rc = -EINVAL);
+
+       /* Skip '{' and '}' */
+       src->ls_str++;
+       src->ls_len -= 2;
+
+       if (nrs_tbf_check_field(&field, "nid")) {
+               if (cfs_parse_nidlist(src->ls_str,
+                                     src->ls_len,
+                                     &expr->te_cond) <= 0)
+                       GOTO(out, rc = -EINVAL);
+               expr->te_field = NRS_TBF_FIELD_NID;
+       } else if (nrs_tbf_check_field(&field, "jobid")) {
+               if (nrs_tbf_jobid_list_parse(src->ls_str,
+                                            src->ls_len,
+                                            &expr->te_cond) < 0)
+                       GOTO(out, rc = -EINVAL);
+               expr->te_field = NRS_TBF_FIELD_JOBID;
+       } else if (nrs_tbf_check_field(&field, "opcode")) {
+               if (nrs_tbf_opcode_list_parse(src->ls_str,
+                                             src->ls_len,
+                                             &expr->te_opcodes) < 0)
+                       GOTO(out, rc = -EINVAL);
+               expr->te_field = NRS_TBF_FIELD_OPCODE;
+       } else
+               GOTO(out, rc = -EINVAL);
+
+       list_add_tail(&expr->te_linkage, cond_list);
+       return 0;
+out:
+       OBD_FREE_PTR(expr);
+       return rc;
+}
+
+static int
+nrs_tbf_conjunction_parse(struct cfs_lstr *src, struct list_head *cond_list)
+{
+       struct nrs_tbf_conjunction *conjunction;
+       struct cfs_lstr expr;
+       int rc = 0;
+
+       OBD_ALLOC(conjunction, sizeof(struct nrs_tbf_conjunction));
+       if (conjunction == NULL)
+               return -ENOMEM;
+
+       INIT_LIST_HEAD(&conjunction->tc_expressions);
+       list_add_tail(&conjunction->tc_linkage, cond_list);
+
+       while (src->ls_str) {
+               rc = cfs_gettok(src, NRS_TBF_CONJUNCTION_DELIM, &expr);
+               if (rc == 0) {
+                       rc = -EINVAL;
+                       break;
+               }
+               rc = nrs_tbf_expression_parse(&expr,
+                                             &conjunction->tc_expressions);
+               if (rc)
+                       break;
+       }
+       return rc;
+}
+
+static int
+nrs_tbf_conds_parse(char *str, int len, struct list_head *cond_list)
+{
+       struct cfs_lstr src;
+       struct cfs_lstr res;
+       int rc = 0;
+
+       src.ls_str = str;
+       src.ls_len = len;
+       INIT_LIST_HEAD(cond_list);
+       while (src.ls_str) {
+               rc = cfs_gettok(&src, NRS_TBF_DISJUNCTION_DELIM, &res);
+               if (rc == 0) {
+                       rc = -EINVAL;
+                       break;
+               }
+               rc = nrs_tbf_conjunction_parse(&res, cond_list);
+               if (rc)
+                       break;
+       }
+       return rc;
+}
+
+static int
+nrs_tbf_generic_parse(struct nrs_tbf_cmd *cmd, const char *id)
+{
+       int rc;
+
+       OBD_ALLOC(cmd->u.tc_start.ts_conds_str, strlen(id) + 1);
+       if (cmd->u.tc_start.ts_conds_str == NULL)
+               return -ENOMEM;
+
+       memcpy(cmd->u.tc_start.ts_conds_str, id, strlen(id));
+
+       /* Parse hybird NID and JOBID conditions */
+       rc = nrs_tbf_conds_parse(cmd->u.tc_start.ts_conds_str,
+                                strlen(cmd->u.tc_start.ts_conds_str),
+                                &cmd->u.tc_start.ts_conds);
+       if (rc)
+               nrs_tbf_generic_cmd_fini(cmd);
+
+       return rc;
+}
+
+static int
+nrs_tbf_expression_match(struct nrs_tbf_expression *expr,
+                        struct nrs_tbf_rule *rule,
+                        struct nrs_tbf_client *cli)
+{
+       switch (expr->te_field) {
+       case NRS_TBF_FIELD_NID:
+               return cfs_match_nid(cli->tc_nid, &expr->te_cond);
+       case NRS_TBF_FIELD_JOBID:
+               return nrs_tbf_jobid_list_match(&expr->te_cond, cli->tc_jobid);
+       case NRS_TBF_FIELD_OPCODE:
+               return cfs_bitmap_check(expr->te_opcodes, cli->tc_opcode);
+       default:
+               return 0;
+       }
+}
+
+static int
+nrs_tbf_conjunction_match(struct nrs_tbf_conjunction *conjunction,
+                         struct nrs_tbf_rule *rule,
+                         struct nrs_tbf_client *cli)
+{
+       struct nrs_tbf_expression *expr;
+       int matched;
+
+       list_for_each_entry(expr, &conjunction->tc_expressions, te_linkage) {
+               matched = nrs_tbf_expression_match(expr, rule, cli);
+               if (!matched)
+                       return 0;
+       }
+
+       return 1;
+}
+
+static int
+nrs_tbf_cond_match(struct nrs_tbf_rule *rule, struct nrs_tbf_client *cli)
+{
+       struct nrs_tbf_conjunction *conjunction;
+       int matched;
+
+       list_for_each_entry(conjunction, &rule->tr_conds, tc_linkage) {
+               matched = nrs_tbf_conjunction_match(conjunction, rule, cli);
+               if (matched)
+                       return 1;
+       }
+
+       return 0;
+}
+
+static void
+nrs_tbf_generic_rule_fini(struct nrs_tbf_rule *rule)
+{
+       if (!list_empty(&rule->tr_conds))
+               nrs_tbf_conds_free(&rule->tr_conds);
+       LASSERT(rule->tr_conds_str != NULL);
+       OBD_FREE(rule->tr_conds_str, strlen(rule->tr_conds_str) + 1);
+}
+
+static int
+nrs_tbf_rule_init(struct ptlrpc_nrs_policy *policy,
+                 struct nrs_tbf_rule *rule, struct nrs_tbf_cmd *start)
+{
+       int rc = 0;
+
+       LASSERT(start->u.tc_start.ts_conds_str);
+       OBD_ALLOC(rule->tr_conds_str,
+                 strlen(start->u.tc_start.ts_conds_str) + 1);
+       if (rule->tr_conds_str == NULL)
+               return -ENOMEM;
+
+       memcpy(rule->tr_conds_str,
+              start->u.tc_start.ts_conds_str,
+              strlen(start->u.tc_start.ts_conds_str));
+
+       INIT_LIST_HEAD(&rule->tr_conds);
+       if (!list_empty(&start->u.tc_start.ts_conds)) {
+               rc = nrs_tbf_conds_parse(rule->tr_conds_str,
+                                        strlen(rule->tr_conds_str),
+                                        &rule->tr_conds);
+       }
+       if (rc)
+               nrs_tbf_generic_rule_fini(rule);
+
+       return rc;
+}
+
+static int
+nrs_tbf_generic_rule_dump(struct nrs_tbf_rule *rule, struct seq_file *m)
+{
+       seq_printf(m, "%s %s %llu, ref %d\n", rule->tr_name,
+                  rule->tr_conds_str, rule->tr_rpc_rate,
+                  atomic_read(&rule->tr_ref) - 1);
+       return 0;
+}
+
+static int
+nrs_tbf_generic_rule_match(struct nrs_tbf_rule *rule,
+                          struct nrs_tbf_client *cli)
+{
+       return nrs_tbf_cond_match(rule, cli);
+}
+
+static struct nrs_tbf_ops nrs_tbf_generic_ops = {
+       .o_name = NRS_TBF_TYPE_GENERIC,
+       .o_startup = nrs_tbf_startup,
+       .o_cli_find = nrs_tbf_cli_find,
+       .o_cli_findadd = nrs_tbf_cli_findadd,
+       .o_cli_put = nrs_tbf_cli_put,
+       .o_cli_init = nrs_tbf_generic_cli_init,
+       .o_rule_init = nrs_tbf_rule_init,
+       .o_rule_dump = nrs_tbf_generic_rule_dump,
+       .o_rule_match = nrs_tbf_generic_rule_match,
+       .o_rule_fini = nrs_tbf_generic_rule_fini,
+};
+
+static void nrs_tbf_opcode_rule_fini(struct nrs_tbf_rule *rule)
+{
+       if (rule->tr_opcodes != NULL)
+               CFS_FREE_BITMAP(rule->tr_opcodes);
+
+       LASSERT(rule->tr_opcodes_str != NULL);
+       OBD_FREE(rule->tr_opcodes_str, strlen(rule->tr_opcodes_str) + 1);
+}
+
+static unsigned nrs_tbf_opcode_hop_hash(struct cfs_hash *hs, const void *key,
+                                       unsigned mask)
+{
+       return cfs_hash_djb2_hash(key, sizeof(__u32), mask);
+}
+
+static int nrs_tbf_opcode_hop_keycmp(const void *key, struct hlist_node *hnode)
+{
+       const __u32     *opc = key;
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
+
+       return *opc == cli->tc_opcode;
+}
+
+static void *nrs_tbf_opcode_hop_key(struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
+
+       return &cli->tc_opcode;
+}
+
+static void *nrs_tbf_opcode_hop_object(struct hlist_node *hnode)
+{
+       return hlist_entry(hnode, struct nrs_tbf_client, tc_hnode);
+}
+
+static void nrs_tbf_opcode_hop_get(struct cfs_hash *hs,
+                                  struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
+
+       atomic_inc(&cli->tc_ref);
+}
+
+static void nrs_tbf_opcode_hop_put(struct cfs_hash *hs,
+                                  struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
+
+       atomic_dec(&cli->tc_ref);
+}
+
+static void nrs_tbf_opcode_hop_exit(struct cfs_hash *hs,
+                                   struct hlist_node *hnode)
+{
+       struct nrs_tbf_client *cli = hlist_entry(hnode,
+                                                struct nrs_tbf_client,
+                                                tc_hnode);
+
+       LASSERTF(atomic_read(&cli->tc_ref) == 0,
+                "Busy TBF object from client with opcode %s, with %d refs\n",
+                ll_opcode2str(cli->tc_opcode),
+                atomic_read(&cli->tc_ref));
+
+       nrs_tbf_cli_fini(cli);
+}
+static struct cfs_hash_ops nrs_tbf_opcode_hash_ops = {
+       .hs_hash        = nrs_tbf_opcode_hop_hash,
+       .hs_keycmp      = nrs_tbf_opcode_hop_keycmp,
+       .hs_key         = nrs_tbf_opcode_hop_key,
+       .hs_object      = nrs_tbf_opcode_hop_object,
+       .hs_get         = nrs_tbf_opcode_hop_get,
+       .hs_put         = nrs_tbf_opcode_hop_put,
+       .hs_put_locked  = nrs_tbf_opcode_hop_put,
+       .hs_exit        = nrs_tbf_opcode_hop_exit,
+};
+
+static int
+nrs_tbf_opcode_startup(struct ptlrpc_nrs_policy *policy,
+                   struct nrs_tbf_head *head)
+{
+       struct nrs_tbf_cmd      start = { 0 };
+       int rc;
+
+       head->th_cli_hash = cfs_hash_create("nrs_tbf_hash",
+                                           NRS_TBF_NID_BITS,
+                                           NRS_TBF_NID_BITS,
+                                           NRS_TBF_NID_BKT_BITS, 0,
+                                           CFS_HASH_MIN_THETA,
+                                           CFS_HASH_MAX_THETA,
+                                           &nrs_tbf_opcode_hash_ops,
+                                           CFS_HASH_RW_BKTLOCK);
+       if (head->th_cli_hash == NULL)
+               return -ENOMEM;
+
+       start.u.tc_start.ts_opcodes = NULL;
+       start.u.tc_start.ts_opcodes_str = "*";
+
+       start.u.tc_start.ts_rpc_rate = tbf_rate;
+       start.u.tc_start.ts_rule_flags = NTRS_DEFAULT;
+       start.tc_name = NRS_TBF_DEFAULT_RULE;
+       rc = nrs_tbf_rule_start(policy, head, &start);
+
+       return rc;
+}
+
+static struct nrs_tbf_client *
+nrs_tbf_opcode_cli_find(struct nrs_tbf_head *head,
+                       struct ptlrpc_request *req)
+{
+       __u32 opc;
+
+       opc = lustre_msg_get_opc(req->rq_reqmsg);
+       return cfs_hash_lookup(head->th_cli_hash, &opc);
+}
+
+static struct nrs_tbf_client *
+nrs_tbf_opcode_cli_findadd(struct nrs_tbf_head *head,
+                          struct nrs_tbf_client *cli)
+{
+       return cfs_hash_findadd_unique(head->th_cli_hash, &cli->tc_opcode,
+                                      &cli->tc_hnode);
+}
+
+static void
+nrs_tbf_opcode_cli_init(struct nrs_tbf_client *cli,
+                       struct ptlrpc_request *req)
+{
+       cli->tc_opcode = lustre_msg_get_opc(req->rq_reqmsg);
+}
+
+#define MAX_OPCODE_LEN 32
+static int
+nrs_tbf_opcode_set_bit(const struct cfs_lstr *id, struct cfs_bitmap *opcodes)
+{
+       int     op = 0;
+       char    opcode_str[MAX_OPCODE_LEN];
+
+       if (id->ls_len + 1 > MAX_OPCODE_LEN)
+               return -EINVAL;
+
+       memcpy(opcode_str, id->ls_str, id->ls_len);
+       opcode_str[id->ls_len] = '\0';
+
+       op = ll_str2opcode(opcode_str);
+       if (op < 0)
+               return -EINVAL;
+
+       cfs_bitmap_set(opcodes, op);
+       return 0;
+}
+
+static int
+nrs_tbf_opcode_list_parse(char *str, int len, struct cfs_bitmap **bitmaptr)
+{
+       struct cfs_bitmap *opcodes;
+       struct cfs_lstr src;
+       struct cfs_lstr res;
+       int rc = 0;
+       ENTRY;
+
+       opcodes = CFS_ALLOCATE_BITMAP(LUSTRE_MAX_OPCODES);
+       if (opcodes == NULL)
+               return -ENOMEM;
+
+       src.ls_str = str;
+       src.ls_len = len;
+       while (src.ls_str) {
+               rc = cfs_gettok(&src, ' ', &res);
+               if (rc == 0) {
+                       rc = -EINVAL;
+                       break;
+               }
+               rc = nrs_tbf_opcode_set_bit(&res, opcodes);
+               if (rc)
+                       break;
+       }
+
+       if (rc == 0)
+               *bitmaptr = opcodes;
+       else
+               CFS_FREE_BITMAP(opcodes);
+
+       RETURN(rc);
+}
+
+static void nrs_tbf_opcode_cmd_fini(struct nrs_tbf_cmd *cmd)
+{
+       if (cmd->u.tc_start.ts_opcodes)
+               CFS_FREE_BITMAP(cmd->u.tc_start.ts_opcodes);
+
+       if (cmd->u.tc_start.ts_opcodes_str)
+               OBD_FREE(cmd->u.tc_start.ts_opcodes_str,
+                        strlen(cmd->u.tc_start.ts_opcodes_str) + 1);
+
+}
+
+static int nrs_tbf_opcode_parse(struct nrs_tbf_cmd *cmd, char *id)
+{
+       struct cfs_lstr src;
+       int rc;
+
+       src.ls_str = id;
+       src.ls_len = strlen(id);
+       rc = nrs_tbf_check_id_value(&src, "opcode");
+       if (rc)
+               return rc;
+
+       OBD_ALLOC(cmd->u.tc_start.ts_opcodes_str, src.ls_len + 1);
+       if (cmd->u.tc_start.ts_opcodes_str == NULL)
+               return -ENOMEM;
+
+       memcpy(cmd->u.tc_start.ts_opcodes_str, src.ls_str, src.ls_len);
+
+       /* parse opcode list */
+       rc = nrs_tbf_opcode_list_parse(cmd->u.tc_start.ts_opcodes_str,
+                                      strlen(cmd->u.tc_start.ts_opcodes_str),
+                                      &cmd->u.tc_start.ts_opcodes);
+       if (rc)
+               nrs_tbf_opcode_cmd_fini(cmd);
+
+       return rc;
+}
+
+static int
+nrs_tbf_opcode_rule_match(struct nrs_tbf_rule *rule,
+                         struct nrs_tbf_client *cli)
+{
+       if (rule->tr_opcodes == NULL)
+               return 0;
+
+       return cfs_bitmap_check(rule->tr_opcodes, cli->tc_opcode);
+}
+
+static int nrs_tbf_opcode_rule_init(struct ptlrpc_nrs_policy *policy,
+                                   struct nrs_tbf_rule *rule,
+                                   struct nrs_tbf_cmd *start)
+{
+       int rc = 0;
+
+       LASSERT(start->u.tc_start.ts_opcodes_str != NULL);
+       OBD_ALLOC(rule->tr_opcodes_str,
+                 strlen(start->u.tc_start.ts_opcodes_str) + 1);
+       if (rule->tr_opcodes_str == NULL)
+               return -ENOMEM;
+
+       strncpy(rule->tr_opcodes_str, start->u.tc_start.ts_opcodes_str,
+               strlen(start->u.tc_start.ts_opcodes_str) + 1);
+
+       /* Default rule '*' */
+       if (start->u.tc_start.ts_opcodes == NULL)
+               return 0;
+
+       rc = nrs_tbf_opcode_list_parse(rule->tr_opcodes_str,
+                                      strlen(rule->tr_opcodes_str),
+                                      &rule->tr_opcodes);
+       if (rc)
+               OBD_FREE(rule->tr_opcodes_str,
+                        strlen(start->u.tc_start.ts_opcodes_str) + 1);
+
+       return rc;
+}
+
+static int
+nrs_tbf_opcode_rule_dump(struct nrs_tbf_rule *rule, struct seq_file *m)
+{
+       seq_printf(m, "%s {%s} %llu, ref %d\n", rule->tr_name,
+                  rule->tr_opcodes_str, rule->tr_rpc_rate,
+                  atomic_read(&rule->tr_ref) - 1);
+       return 0;
+}
+
+
+struct nrs_tbf_ops nrs_tbf_opcode_ops = {
+       .o_name = NRS_TBF_TYPE_OPCODE,
+       .o_startup = nrs_tbf_opcode_startup,
+       .o_cli_find = nrs_tbf_opcode_cli_find,
+       .o_cli_findadd = nrs_tbf_opcode_cli_findadd,
+       .o_cli_put = nrs_tbf_nid_cli_put,
+       .o_cli_init = nrs_tbf_opcode_cli_init,
+       .o_rule_init = nrs_tbf_opcode_rule_init,
+       .o_rule_dump = nrs_tbf_opcode_rule_dump,
+       .o_rule_match = nrs_tbf_opcode_rule_match,
+       .o_rule_fini = nrs_tbf_opcode_rule_fini,
+};
+
+static struct nrs_tbf_type nrs_tbf_types[] = {
+       {
+               .ntt_name = NRS_TBF_TYPE_JOBID,
+               .ntt_flag = NRS_TBF_FLAG_JOBID,
+               .ntt_ops = &nrs_tbf_jobid_ops,
+       },
+       {
+               .ntt_name = NRS_TBF_TYPE_NID,
+               .ntt_flag = NRS_TBF_FLAG_NID,
+               .ntt_ops = &nrs_tbf_nid_ops,
+       },
+       {
+               .ntt_name = NRS_TBF_TYPE_OPCODE,
+               .ntt_flag = NRS_TBF_FLAG_OPCODE,
+               .ntt_ops = &nrs_tbf_opcode_ops,
+       },
+       {
+               .ntt_name = NRS_TBF_TYPE_GENERIC,
+               .ntt_flag = NRS_TBF_FLAG_GENERIC,
+               .ntt_ops = &nrs_tbf_generic_ops,
+       },
+};
+
+/**
+ * Is called before the policy transitions into
+ * ptlrpc_nrs_pol_state::NRS_POL_STATE_STARTED; allocates and initializes a
+ * policy-specific private data structure.
+ *
+ * \param[in] policy The policy to start
+ *
+ * \retval -ENOMEM OOM error
+ * \retval  0     success
+ *
+ * \see nrs_policy_register()
+ * \see nrs_policy_ctl()
+ */
+static int nrs_tbf_start(struct ptlrpc_nrs_policy *policy, char *arg)
+{
+       struct nrs_tbf_head     *head;
+       struct nrs_tbf_ops      *ops;
+       __u32                    type;
+       char                    *name;
+       int found = 0;
+       int i;
+       int rc = 0;
+
+       if (arg == NULL)
+               name = NRS_TBF_TYPE_GENERIC;
+       else if (strlen(arg) < NRS_TBF_TYPE_MAX_LEN)
+               name = arg;
+       else
+               GOTO(out, rc = -EINVAL);
+
+       for (i = 0; i < ARRAY_SIZE(nrs_tbf_types); i++) {
+               if (strcmp(name, nrs_tbf_types[i].ntt_name) == 0) {
+                       ops = nrs_tbf_types[i].ntt_ops;
+                       type = nrs_tbf_types[i].ntt_flag;
+                       found = 1;
+                       break;
+               }
+       }
+       if (found == 0)
+               GOTO(out, rc = -ENOTSUPP);
+
+       OBD_CPT_ALLOC_PTR(head, nrs_pol2cptab(policy), nrs_pol2cptid(policy));
+       if (head == NULL)
+               GOTO(out, rc = -ENOMEM);
+
+       memcpy(head->th_type, name, strlen(name));
+       head->th_type[strlen(name)] = '\0';
+       head->th_ops = ops;
+       head->th_type_flag = type;
+
+       head->th_binheap = cfs_binheap_create(&nrs_tbf_heap_ops,
+                                             CBH_FLAG_ATOMIC_GROW, 4096, NULL,
+                                             nrs_pol2cptab(policy),
+                                             nrs_pol2cptid(policy));
+       if (head->th_binheap == NULL)
+               GOTO(out_free_head, rc = -ENOMEM);
+
+       atomic_set(&head->th_rule_sequence, 0);
+       spin_lock_init(&head->th_rule_lock);
+       INIT_LIST_HEAD(&head->th_list);
+       hrtimer_init(&head->th_timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
+       head->th_timer.function = nrs_tbf_timer_cb;
+       rc = head->th_ops->o_startup(policy, head);
+       if (rc)
+               GOTO(out_free_heap, rc);
+
+       policy->pol_private = head;
+       return 0;
+out_free_heap:
+       cfs_binheap_destroy(head->th_binheap);
+out_free_head:
+       OBD_FREE_PTR(head);
+out:
+       return rc;
+}
+
+/**
+ * Is called before the policy transitions into
+ * ptlrpc_nrs_pol_state::NRS_POL_STATE_STOPPED; deallocates the policy-specific
+ * private data structure.
+ *
+ * \param[in] policy The policy to stop
+ *
+ * \see nrs_policy_stop0()
+ */
+static void nrs_tbf_stop(struct ptlrpc_nrs_policy *policy)
+{
+       struct nrs_tbf_head *head = policy->pol_private;
+       struct ptlrpc_nrs *nrs = policy->pol_nrs;
+       struct nrs_tbf_rule *rule, *n;
+
+       LASSERT(head != NULL);
+       LASSERT(head->th_cli_hash != NULL);
+       hrtimer_cancel(&head->th_timer);
+       /* Should cleanup hash first before free rules */
+       cfs_hash_putref(head->th_cli_hash);
+       list_for_each_entry_safe(rule, n, &head->th_list, tr_linkage) {
+               list_del_init(&rule->tr_linkage);
+               nrs_tbf_rule_put(rule);
+       }
+       LASSERT(list_empty(&head->th_list));
+       LASSERT(head->th_binheap != NULL);
+       LASSERT(cfs_binheap_is_empty(head->th_binheap));
+       cfs_binheap_destroy(head->th_binheap);
+       OBD_FREE_PTR(head);
+       nrs->nrs_throttling = 0;
+       wake_up(&policy->pol_nrs->nrs_svcpt->scp_waitq);
+}
+
+/**
+ * Performs a policy-specific ctl function on TBF policy instances; similar
+ * to ioctl.
+ *
+ * \param[in]    policy the policy instance
+ * \param[in]    opc    the opcode
+ * \param[in,out] arg   used for passing parameters and information
+ *
+ * \pre assert_spin_locked(&policy->pol_nrs->->nrs_lock)
+ * \post assert_spin_locked(&policy->pol_nrs->->nrs_lock)
+ *
+ * \retval 0   operation carried out successfully
+ * \retval -ve error
+ */
+static int nrs_tbf_ctl(struct ptlrpc_nrs_policy *policy,
+                      enum ptlrpc_nrs_ctl opc,
+                      void *arg)
+{
+       int rc = 0;
+       ENTRY;
+
+       assert_spin_locked(&policy->pol_nrs->nrs_lock);
+
+       switch ((enum nrs_ctl_tbf)opc) {
+       default:
+               RETURN(-EINVAL);
+
+       /**
+        * Read RPC rate size of a policy instance.
+        */
+       case NRS_CTL_TBF_RD_RULE: {
+               struct nrs_tbf_head *head = policy->pol_private;
+               struct seq_file *m = (struct seq_file *) arg;
+               struct ptlrpc_service_part *svcpt;
+
+               svcpt = policy->pol_nrs->nrs_svcpt;
+               seq_printf(m, "CPT %d:\n", svcpt->scp_cpt);
+
+               rc = nrs_tbf_rule_dump_all(head, m);
+               }
+               break;
+
+       /**
+        * Write RPC rate of a policy instance.
+        */
+       case NRS_CTL_TBF_WR_RULE: {
+               struct nrs_tbf_head *head = policy->pol_private;
+               struct nrs_tbf_cmd *cmd;
+
+               cmd = (struct nrs_tbf_cmd *)arg;
+               rc = nrs_tbf_command(policy,
+                                    head,
+                                    cmd);
+               }
+               break;
+       /**
+        * Read the TBF policy type of a policy instance.
+        */
+       case NRS_CTL_TBF_RD_TYPE_FLAG: {
+               struct nrs_tbf_head *head = policy->pol_private;
+
+               *(__u32 *)arg = head->th_type_flag;
+               }
+               break;
+       }
+
+       RETURN(rc);
+}
+
+/**
+ * Is called for obtaining a TBF policy resource.
+ *
+ * \param[in]  policy    The policy on which the request is being asked for
+ * \param[in]  nrq       The request for which resources are being taken
+ * \param[in]  parent    Parent resource, unused in this policy
+ * \param[out] resp      Resources references are placed in this array
+ * \param[in]  moving_req Signifies limited caller context; unused in this
+ *                       policy
+ *
+ *
+ * \see nrs_resource_get_safe()
+ */
+static int nrs_tbf_res_get(struct ptlrpc_nrs_policy *policy,
+                          struct ptlrpc_nrs_request *nrq,
+                          const struct ptlrpc_nrs_resource *parent,
+                          struct ptlrpc_nrs_resource **resp,
+                          bool moving_req)
+{
+       struct nrs_tbf_head   *head;
+       struct nrs_tbf_client *cli;
+       struct nrs_tbf_client *tmp;
+       struct ptlrpc_request *req;
+
+       if (parent == NULL) {
+               *resp = &((struct nrs_tbf_head *)policy->pol_private)->th_res;
+               return 0;
+       }
+
+       head = container_of(parent, struct nrs_tbf_head, th_res);
+       req = container_of(nrq, struct ptlrpc_request, rq_nrq);
+       cli = head->th_ops->o_cli_find(head, req);
+       if (cli != NULL) {
+               spin_lock(&policy->pol_nrs->nrs_svcpt->scp_req_lock);
+               LASSERT(cli->tc_rule);
+               if (cli->tc_rule_sequence !=
+                   atomic_read(&head->th_rule_sequence) ||
+                   cli->tc_rule->tr_flags & NTRS_STOPPING) {
+                       struct nrs_tbf_rule *rule;
+
+                       CDEBUG(D_RPCTRACE,
+                              "TBF class@%p rate %llu sequence %d, "
+                              "rule flags %d, head sequence %d\n",
+                              cli, cli->tc_rpc_rate,
+                              cli->tc_rule_sequence,
+                              cli->tc_rule->tr_flags,
+                              atomic_read(&head->th_rule_sequence));
+                       rule = nrs_tbf_rule_match(head, cli);
+                       if (rule != cli->tc_rule) {
+                               nrs_tbf_cli_reset(head, rule, cli);
+                       } else {
+                               if (cli->tc_rule_generation != rule->tr_generation)
+                                       nrs_tbf_cli_reset_value(head, cli);
                                nrs_tbf_rule_put(rule);
+                       }
                } else if (cli->tc_rule_generation !=
                           cli->tc_rule->tr_generation) {
                        nrs_tbf_cli_reset_value(head, cli);
@@ -1455,6 +2396,7 @@ static int nrs_tbf_res_get(struct ptlrpc_nrs_policy *policy,
                          sizeof(*cli), moving_req ? GFP_ATOMIC : __GFP_IO);
        if (cli == NULL)
                return -ENOMEM;
+
        nrs_tbf_cli_init(head, cli, req);
        tmp = head->th_ops->o_cli_findadd(head, cli);
        if (tmp != cli) {
@@ -1570,11 +2512,12 @@ struct ptlrpc_nrs_request *nrs_tbf_req_get(struct ptlrpc_nrs_policy *policy,
                                                     &cli->tc_node);
                        }
                        CDEBUG(D_RPCTRACE,
-                              "NRS start %s request from %s, "
-                              "seq: %llu\n",
-                              policy->pol_desc->pd_name,
-                              libcfs_id2str(req->rq_peer),
-                              nrq->nr_u.tbf.tr_sequence);
+                              "TBF dequeues: class@%p rate %llu gen %llu "
+                              "token %llu, rule@%p rate %llu gen %llu\n",
+                              cli, cli->tc_rpc_rate,
+                              cli->tc_rule_generation, cli->tc_ntoken,
+                              cli->tc_rule, cli->tc_rule->tr_rpc_rate,
+                              cli->tc_rule->tr_generation);
                } else {
                        ktime_t time;
 
@@ -1640,6 +2583,16 @@ static int nrs_tbf_req_add(struct ptlrpc_nrs_policy *policy,
                list_add_tail(&nrq->nr_u.tbf.tr_list,
                                  &cli->tc_list);
        }
+
+       if (rc == 0)
+               CDEBUG(D_RPCTRACE,
+                      "TBF enqueues: class@%p rate %llu gen %llu "
+                      "token %llu, rule@%p rate %llu gen %llu\n",
+                      cli, cli->tc_rpc_rate,
+                      cli->tc_rule_generation, cli->tc_ntoken,
+                      cli->tc_rule, cli->tc_rule->tr_rpc_rate,
+                      cli->tc_rule->tr_generation);
+
        return rc;
 }
 
@@ -1768,26 +2721,37 @@ static int nrs_tbf_id_parse(struct nrs_tbf_cmd *cmd, char *token)
 {
        int rc;
 
-       if (cmd->u.tc_start.ts_valid_type & NRS_TBF_FLAG_JOBID)
+       switch (cmd->u.tc_start.ts_valid_type) {
+       case NRS_TBF_FLAG_JOBID:
                rc = nrs_tbf_jobid_parse(cmd, token);
-       else if (cmd->u.tc_start.ts_valid_type & NRS_TBF_FLAG_NID)
+               break;
+       case NRS_TBF_FLAG_NID:
                rc = nrs_tbf_nid_parse(cmd, token);
-       else if (cmd->u.tc_start.ts_valid_type == NRS_TBF_FLAG_INVALID)
-               rc = -EINVAL;
-       else
-               rc = 0;
+               break;
+       case NRS_TBF_FLAG_OPCODE:
+               rc = nrs_tbf_opcode_parse(cmd, token);
+               break;
+       case NRS_TBF_FLAG_GENERIC:
+               rc = nrs_tbf_generic_parse(cmd, token);
+               break;
+       default:
+               RETURN(-EINVAL);
+       }
 
        return rc;
 }
 
-
 static void nrs_tbf_cmd_fini(struct nrs_tbf_cmd *cmd)
 {
        if (cmd->tc_cmd == NRS_CTL_TBF_START_RULE) {
-               if (cmd->u.tc_start.ts_valid_type & NRS_TBF_FLAG_JOBID)
+               if (cmd->u.tc_start.ts_valid_type == NRS_TBF_FLAG_JOBID)
                        nrs_tbf_jobid_cmd_fini(cmd);
-               else if (cmd->u.tc_start.ts_valid_type & NRS_TBF_FLAG_NID)
+               else if (cmd->u.tc_start.ts_valid_type == NRS_TBF_FLAG_NID)
                        nrs_tbf_nid_cmd_fini(cmd);
+               else if (cmd->u.tc_start.ts_valid_type == NRS_TBF_FLAG_OPCODE)
+                       nrs_tbf_opcode_cmd_fini(cmd);
+               else if (cmd->u.tc_start.ts_valid_type == NRS_TBF_FLAG_GENERIC)
+                       nrs_tbf_generic_cmd_fini(cmd);
        }
 }