4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; if not, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2012, 2017, Intel Corporation.
25 * Use is subject to license terms.
27 * Author: Johann Lombardi <johann.lombardi@intel.com>
28 * Author: Niu Yawei <yawei.niu@intel.com>
31 #define DEBUG_SUBSYSTEM S_LQUOTA
33 #include <linux/kthread.h>
35 #include <lustre_dlm.h>
36 #include <lustre_swab.h>
37 #include <obd_class.h>
39 #include "qmt_internal.h"
41 /* intent policy function called from mdt_intent_opc() when the intent is of
43 int qmt_intent_policy(const struct lu_env *env, struct lu_device *ld,
44 struct ptlrpc_request *req, struct ldlm_lock **lockp,
47 struct qmt_device *qmt = lu2qmt_dev(ld);
48 struct ldlm_intent *it;
49 struct quota_body *reqbody;
50 struct quota_body *repbody;
51 struct obd_uuid *uuid;
52 struct lquota_lvb *lvb;
53 struct ldlm_resource *res = (*lockp)->l_resource;
54 struct ldlm_reply *ldlm_rep;
58 req_capsule_extend(&req->rq_pill, &RQF_LDLM_INTENT_QUOTA);
59 req_capsule_set_size(&req->rq_pill, &RMF_DLM_LVB, RCL_SERVER,
60 ldlm_lvbo_size(*lockp));
62 /* extract quota body and intent opc */
63 it = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
65 RETURN(err_serious(-EFAULT));
67 reqbody = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_BODY);
69 RETURN(err_serious(-EFAULT));
72 rc = req_capsule_server_pack(&req->rq_pill);
74 CERROR("Can't pack response, rc %d\n", rc);
75 RETURN(err_serious(rc));
78 repbody = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_BODY);
80 RETURN(err_serious(-EFAULT));
82 ldlm_rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
84 RETURN(err_serious(-EFAULT));
86 uuid = &(*lockp)->l_export->exp_client_uuid;
89 case IT_QUOTA_DQACQ: {
90 struct lquota_entry *lqe;
91 struct ldlm_lock *lock;
94 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
95 /* acquire on global lock? something is wrong ... */
96 GOTO(out, rc = -EPROTO);
98 /* verify global lock isn't stale */
99 if (!lustre_handle_is_used(&reqbody->qb_glb_lockh))
100 GOTO(out, rc = -ENOLCK);
102 lock = ldlm_handle2lock(&reqbody->qb_glb_lockh);
104 GOTO(out, rc = -ENOLCK);
107 rc = qmt_uuid2idx(uuid, &idx);
109 GOTO(out, rc = -EINVAL);
111 /* TODO: it seems we don't need to get lqe from
112 * lq_lvb_data anymore ... And do extra get
114 lqe = res->lr_lvb_data;
115 LASSERT(lqe != NULL);
118 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), rc,
119 lqe_qtype(lqe), &reqbody->qb_id,
126 /* acquire quota space */
127 rc = qmt_dqacq0(env, qmt, uuid,
128 reqbody->qb_flags, reqbody->qb_count,
129 reqbody->qb_usage, repbody);
138 /* new connection from slave */
140 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0)
141 /* connection on per-ID lock? something is wrong ... */
142 GOTO(out, rc = -EPROTO);
144 rc = qmt_pool_new_conn(env, qmt, &reqbody->qb_fid,
145 &repbody->qb_slv_fid,
146 &repbody->qb_slv_ver, uuid);
152 CERROR("%s: invalid intent opcode: %llu\n", qmt->qmt_svname,
154 GOTO(out, rc = -EINVAL);
157 /* on success, pack lvb in reply */
158 lvb = req_capsule_server_get(&req->rq_pill, &RMF_DLM_LVB);
159 lvb_len = ldlm_lvbo_size(*lockp);
160 lvb_len = ldlm_lvbo_fill(*lockp, lvb, &lvb_len);
162 GOTO(out, rc = lvb_len);
164 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB, lvb_len, RCL_SERVER);
166 ldlm_rep->lock_policy_res2 = clear_serious(rc);
172 * Initialize quota LVB associated with quota indexes.
173 * Called with res->lr_lvb_sem held
175 int qmt_lvbo_init(struct lu_device *ld, struct ldlm_resource *res)
178 struct qmt_thread_info *qti;
179 struct qmt_device *qmt = lu2qmt_dev(ld);
180 int pool_type, qtype;
184 LASSERT(res != NULL);
186 if (res->lr_type != LDLM_PLAIN)
189 if (res->lr_lvb_data ||
190 res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
197 /* extract global index FID and quota identifier */
198 fid_extract_from_quota_res(&qti->qti_fid, &qti->qti_id, &res->lr_name);
200 /* sanity check the global index FID */
201 rc = lquota_extract_fid(&qti->qti_fid, &pool_type, &qtype);
203 CERROR("can't extract glb index information from FID "DFID"\n",
204 PFID(&qti->qti_fid));
208 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
209 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
210 * we are thus dealing with an ID lock. */
211 struct qmt_pool_info *pool;
212 struct lquota_entry *lqe;
213 struct lqe_glbl_data *lgd;
215 pool = qmt_pool_lookup_glb(env, qmt, pool_type);
217 GOTO(out, rc = -ENOMEM);
219 /* Find the quota entry associated with the quota id */
220 lqe = qmt_pool_lqe_lookup(env, qmt, pool_type, qtype,
223 qpi_putref(env, pool);
224 GOTO(out, rc = PTR_ERR(lqe));
227 /* TODO: need something like qmt_extend_lqe_gd that has
228 * to be calledeach time when qpi_slv_nr is incremented */
229 lgd = qmt_alloc_lqe_gd(pool, qtype);
232 qpi_putref(env, pool);
233 GOTO(out, rc = -ENOMEM);
236 qmt_setup_lqe_gd(env, qmt, lqe, lgd, pool_type);
238 /* store reference to lqe in lr_lvb_data */
239 res->lr_lvb_data = lqe;
240 qpi_putref(env, pool);
241 LQUOTA_DEBUG(lqe, "initialized res lvb");
243 struct dt_object *obj;
245 /* lookup global index */
246 obj = dt_locate(env, qmt->qmt_child, &qti->qti_fid);
248 GOTO(out, rc = PTR_ERR(obj));
249 if (!dt_object_exists(obj)) {
250 dt_object_put(env, obj);
251 GOTO(out, rc = -ENOENT);
254 /* store reference to global index object in lr_lvb_data */
255 res->lr_lvb_data = obj;
256 CDEBUG(D_QUOTA, DFID" initialized lvb\n", PFID(&qti->qti_fid));
259 res->lr_lvb_len = sizeof(struct lquota_lvb);
265 /* clear lge_qunit/edquot_nu flags -
266 * slave recieved new qunit and edquot.
268 * \retval true if revoke is needed - qunit
269 * for this slave reaches least_qunit
271 static bool qmt_clear_lgeg_arr_nu(struct lquota_entry *lqe, int stype, int idx)
273 unsigned long least_qunit = lqe2qpi(lqe)->qpi_least_qunit;
274 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
276 /* There is no array to store lge for the case of DOM.
277 * Ignore it until MDT pools will be ready. */
278 if (!(lqe_rtype(lqe) == LQUOTA_RES_DT && stype == QMT_STYPE_MDT)) {
279 lqe->lqe_glbl_data->lqeg_arr[idx].lge_qunit_nu = 0;
280 lqe->lqe_glbl_data->lqeg_arr[idx].lge_edquot_nu = 0;
282 /* We shouldn't call revoke for DOM case, it will be updated
283 * at qmt_id_lock_glimpse. */
284 return (lgd->lqeg_arr[idx].lge_qunit == least_qunit);
290 static void qmt_set_revoke(struct lu_env *env, struct lquota_entry *lqe,
293 unsigned long least_qunit = lqe2qpi(lqe)->qpi_least_qunit;
294 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
296 if (lgd->lqeg_arr[idx].lge_qunit == least_qunit) {
299 qti_lqes_write_lock(env);
300 for (i = 0; i < qti_lqes_cnt(env); i++) {
301 LQUOTA_DEBUG(qti_lqes(env)[i],
302 "idx %d lge_qunit %llu least_qunit %lu\n",
303 idx, lgd->lqeg_arr[idx].lge_qunit,
305 if (qti_lqes(env)[i]->lqe_qunit == least_qunit) {
306 qti_lqes(env)[i]->lqe_revoke_time =
308 qmt_adjust_edquot(qti_lqes(env)[i],
309 ktime_get_real_seconds());
312 qti_lqes_write_unlock(env);
317 * Update LVB associated with the global quota index.
318 * This function is called from the DLM itself after a glimpse callback, in this
319 * case valid ptlrpc request is passed.
321 int qmt_lvbo_update(struct lu_device *ld, struct ldlm_resource *res,
322 struct ptlrpc_request *req, int increase_only)
325 struct qmt_thread_info *qti;
326 struct qmt_device *qmt = lu2qmt_dev(ld);
327 struct lquota_entry *lqe;
328 struct lquota_lvb *lvb;
329 struct ldlm_lock *lock;
330 struct obd_export *exp;
335 LASSERT(res != NULL);
340 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
341 /* no need to update lvb for global quota locks */
344 lvb = req_capsule_server_swab_get(&req->rq_pill, &RMF_DLM_LVB,
345 lustre_swab_lquota_lvb);
347 CERROR("%s: failed to extract lvb from request\n",
352 lqe = res->lr_lvb_data;
353 LASSERT(lqe != NULL);
356 /* allocate environement */
361 /* The request is a glimpse callback which was sent via the
362 * reverse import to the slave. What we care about here is the
363 * export associated with the slave and req->rq_export is
364 * definitely not what we are looking for (it is actually set to
366 * Therefore we extract the lock from the request argument
367 * and use lock->l_export. */
368 lock = ldlm_request_lock(req);
370 CERROR("%s: failed to get lock from request!\n",
372 GOTO(out, rc = PTR_ERR(lock));
375 exp = class_export_get(lock->l_export);
377 CERROR("%s: failed to get export from lock!\n",
379 GOTO(out, rc = -EFAULT);
382 rc = qmt_uuid2idx(&exp->exp_client_uuid, &idx);
386 need_revoke = qmt_clear_lgeg_arr_nu(lqe, rc, idx);
387 if (lvb->lvb_id_rel == 0) {
388 /* nothing to release */
389 if (lvb->lvb_id_may_rel != 0)
390 /* but might still release later ... */
391 lqe->lqe_may_rel += lvb->lvb_id_may_rel;
394 if (!need_revoke && lvb->lvb_id_rel == 0)
395 GOTO(out_exp, rc = 0);
397 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), rc, lqe_qtype(lqe),
398 &lqe->lqe_id, NULL, idx);
403 qmt_set_revoke(env, lqe, rc, idx);
405 if (lvb->lvb_id_rel) {
406 LQUOTA_DEBUG(lqe, "releasing:%llu may release:%llu",
407 lvb->lvb_id_rel, lvb->lvb_id_may_rel);
409 /* release quota space */
410 rc = qmt_dqacq0(env, qmt, &exp->exp_client_uuid,
411 QUOTA_DQACQ_FL_REL, lvb->lvb_id_rel,
413 if (rc || qti->qti_body.qb_count != lvb->lvb_id_rel)
415 "failed to release quota space on glimpse %llu!=%llu : rc = %d\n",
416 qti->qti_body.qb_count,
417 lvb->lvb_id_rel, rc);
424 class_export_put(exp);
431 * Report size of lvb to ldlm layer in order to allocate lvb buffer
432 * As far as quota locks are concerned, the size is static and is the same
433 * for both global and per-ID locks which shares the same lvb format.
435 int qmt_lvbo_size(struct lu_device *ld, struct ldlm_lock *lock)
437 return sizeof(struct lquota_lvb);
441 * Fill request buffer with quota lvb
443 int qmt_lvbo_fill(struct lu_device *ld, struct ldlm_lock *lock, void *lvb,
446 struct ldlm_resource *res = lock->l_resource;
447 struct lquota_lvb *qlvb = lvb;
452 LASSERT(res != NULL);
455 if (res->lr_type != LDLM_PLAIN || res->lr_lvb_data == NULL ||
456 res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
462 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
463 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
464 * we are thus dealing with an ID lock. */
465 struct lquota_entry *lqe = res->lr_lvb_data;
466 struct qmt_device *qmt;
467 struct obd_uuid *uuid;
470 uuid = &(lock)->l_export->exp_client_uuid;
471 rc = qmt_uuid2idx(uuid, &idx);
474 qmt = lu2qmt_dev(ld);
475 /* return current qunit value & edquot flags in lvb */
477 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), rc,
478 lqe_qtype(lqe), &lqe->lqe_id,
481 qlvb->lvb_id_qunit = qti_lqes_min_qunit(env);
483 if (qti_lqes_edquot(env))
484 qlvb->lvb_flags = LQUOTA_FL_EDQUOT;
487 CDEBUG(D_QUOTA, "uuid %s lqe_id %lu, edquot %llu qunit %llu\n",
488 (char *)uuid, (unsigned long)lqe->lqe_id.qid_uid,
489 qlvb->lvb_flags, qlvb->lvb_id_qunit);
492 /* global quota lock */
493 struct dt_object *obj = res->lr_lvb_data;
495 /* return current version of global index */
496 qlvb->lvb_glb_ver = dt_version_get(env, obj);
499 RETURN(rc = rc ?: sizeof(struct lquota_lvb));
503 * Free lvb associated with a given ldlm resource
504 * we don't really allocate a lvb, lr_lvb_data just points to
505 * the appropriate backend structures.
507 int qmt_lvbo_free(struct lu_device *ld, struct ldlm_resource *res)
511 if (res->lr_lvb_data == NULL)
514 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
515 struct lquota_entry *lqe = res->lr_lvb_data;
516 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
518 /* release lqe reference */
519 lqe->lqe_glbl_data = NULL;
521 qmt_free_lqe_gd(lgd);
523 struct dt_object *obj = res->lr_lvb_data;
524 /* release object reference */
525 dt_object_put(lu_env_find(), obj);
528 res->lr_lvb_data = NULL;
534 typedef int (*qmt_glimpse_cb_t)(struct ldlm_lock *, struct lquota_entry *);
536 struct qmt_gl_lock_array {
539 struct ldlm_lock **q_locks;
542 static void qmt_free_lock_array(struct qmt_gl_lock_array *array)
546 if (array->q_max == 0) {
547 LASSERT(array->q_locks == NULL);
551 for (i = 0; i < array->q_cnt; i++) {
552 LASSERT(array->q_locks[i]);
553 LDLM_LOCK_RELEASE(array->q_locks[i]);
554 array->q_locks[i] = NULL;
557 OBD_FREE_PTR_ARRAY(array->q_locks, array->q_max);
558 array->q_locks = NULL;
562 static int qmt_alloc_lock_array(struct ldlm_resource *res,
563 struct qmt_gl_lock_array *array,
564 qmt_glimpse_cb_t cb, void *arg)
566 struct list_head *pos;
567 unsigned long count = 0;
571 LASSERT(!array->q_max && !array->q_cnt && !array->q_locks);
574 /* scan list of granted locks */
575 list_for_each(pos, &res->lr_granted) {
576 struct ldlm_lock *lock;
579 lock = list_entry(pos, struct ldlm_lock, l_res_link);
580 LASSERT(lock->l_export);
584 /* slave should not be notified */
590 if (array->q_max != 0 && array->q_cnt < array->q_max) {
591 array->q_locks[array->q_cnt] = LDLM_LOCK_GET(lock);
597 if (count > array->q_max) {
598 qmt_free_lock_array(array);
602 * allocate more slots in case of more qualified locks are
603 * found during next loop
605 array->q_max = count + count / 2 + 10;
607 LASSERT(array->q_locks == NULL && array->q_cnt == 0);
608 OBD_ALLOC_PTR_ARRAY(array->q_locks, array->q_max);
609 if (array->q_locks == NULL) {
619 void qmt_setup_id_desc(struct ldlm_lock *lock, union ldlm_gl_desc *desc,
620 struct lquota_entry *lqe)
622 struct obd_uuid *uuid = &(lock)->l_export->exp_client_uuid;
623 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
628 stype = qmt_uuid2idx(uuid, &idx);
631 /* DOM case - set global lqe settings */
632 if (lqe_rtype(lqe) == LQUOTA_RES_DT && stype == QMT_STYPE_MDT) {
633 edquot = lqe->lqe_edquot;
634 qunit = lqe->lqe_qunit;
636 edquot = lgd->lqeg_arr[idx].lge_edquot;
637 qunit = lgd->lqeg_arr[idx].lge_qunit;
640 /* fill glimpse descriptor with lqe settings */
641 desc->lquota_desc.gl_flags = edquot ? LQUOTA_FL_EDQUOT : 0;
642 desc->lquota_desc.gl_qunit = qunit;
643 CDEBUG(D_QUOTA, "setup desc: stype %d idx %d, edquot %llu qunit %llu\n",
644 stype, idx, desc->lquota_desc.gl_flags,
645 desc->lquota_desc.gl_qunit);
649 * Send glimpse callback to slaves holding a lock on resource \res.
650 * This is used to notify slaves of new quota settings or to claim quota space
653 * \param env - is the environment passed by the caller
654 * \param qmt - is the quota master target
655 * \param res - is the dlm resource associated with the quota object
656 * \param desc - is the glimpse descriptor to pack in glimpse callback
657 * \param cb - is the callback function called on every lock and determine
658 * whether a glimpse should be issued
659 * \param arg - is an opaq parameter passed to the callback function
661 static int qmt_glimpse_lock(const struct lu_env *env, struct qmt_device *qmt,
662 struct ldlm_resource *res, union ldlm_gl_desc *desc,
663 qmt_glimpse_cb_t cb, struct lquota_entry *lqe)
665 union ldlm_gl_desc *descs = NULL;
666 struct lqe_glbl_data *gld;
667 struct list_head *tmp, *pos;
669 struct qmt_gl_lock_array locks;
670 unsigned long i, locks_count;
674 gld = lqe ? lqe->lqe_glbl_data : NULL;
675 memset(&locks, 0, sizeof(locks));
676 rc = qmt_alloc_lock_array(res, &locks, cb, lqe);
678 CERROR("%s: failed to allocate glimpse lock array (%d)\n",
679 qmt->qmt_svname, rc);
683 CDEBUG(D_QUOTA, "%s: no granted locks to send glimpse\n",
687 CDEBUG(D_QUOTA, "found granted locks %lu\n", locks.q_cnt);
688 locks_count = locks.q_cnt;
690 /* Use one desc for all works, when called from qmt_glb_lock_notify */
691 if (gld && locks.q_cnt > 1) {
692 /* TODO: think about to store this preallocated descs
693 * in lqe_global in lqeg_arr as a part of lqe_glbl_entry.
694 * The benefit is that we don't need to allocate/free
695 * and setup this descs each time. But the drawback is
696 * memory use (sizeof ldlm_gl_desc * OST_COUNT * user_number).
697 * for examfple it could be 88 * 256 * 10 000 about 225 MB. */
699 sizeof(struct ldlm_gl_lquota_desc) * locks.q_cnt);
701 CERROR("%s: alloc glimpse lock array failed: rc = %d\n",
702 qmt->qmt_svname, rc);
703 qmt_free_lock_array(&locks);
708 for (i = locks.q_cnt; i > 0; i--) {
709 struct ldlm_glimpse_work *work;
713 CERROR("%s: failed to notify a lock.\n",
720 desc = &descs[i - 1];
721 qmt_setup_id_desc(locks.q_locks[i - 1], desc, lqe);
722 work->gl_interpret_data = lqe;
725 list_add_tail(&work->gl_list, &gl_list);
726 work->gl_lock = locks.q_locks[i - 1];
728 work->gl_desc = desc;
730 locks.q_locks[i - 1] = NULL;
734 qmt_free_lock_array(&locks);
736 if (list_empty(&gl_list)) {
737 CDEBUG(D_QUOTA, "%s: nobody to notify\n", qmt->qmt_svname);
741 /* issue glimpse callbacks to all connected slaves */
742 rc = ldlm_glimpse_locks(res, &gl_list);
744 list_for_each_safe(pos, tmp, &gl_list) {
745 struct ldlm_glimpse_work *work;
747 work = list_entry(pos, struct ldlm_glimpse_work, gl_list);
749 list_del(&work->gl_list);
750 CERROR("%s: failed to notify %s of new quota settings\n",
752 obd_uuid2str(&work->gl_lock->l_export->exp_client_uuid));
753 LDLM_LOCK_RELEASE(work->gl_lock);
759 sizeof(struct ldlm_gl_lquota_desc) * locks_count);
765 * Send glimpse request to all global quota locks to push new quota setting to
768 * \param env - is the environment passed by the caller
769 * \param lqe - is the lquota entry which has new settings
770 * \param ver - is the version associated with the setting change
772 void qmt_glb_lock_notify(const struct lu_env *env, struct lquota_entry *lqe,
775 struct qmt_thread_info *qti = qmt_info(env);
776 struct qmt_pool_info *pool = lqe2qpi(lqe);
777 struct ldlm_resource *res = NULL;
780 lquota_generate_fid(&qti->qti_fid, pool->qpi_rtype, lqe_qtype(lqe));
782 /* send glimpse callback to notify slaves of new quota settings */
783 qti->qti_gl_desc.lquota_desc.gl_id = lqe->lqe_id;
784 qti->qti_gl_desc.lquota_desc.gl_flags = 0;
785 if (lqe->lqe_is_default) {
786 qti->qti_gl_desc.lquota_desc.gl_hardlimit = 0;
787 qti->qti_gl_desc.lquota_desc.gl_softlimit = 0;
788 qti->qti_gl_desc.lquota_desc.gl_time = LQUOTA_GRACE_FLAG(0,
789 LQUOTA_FLAG_DEFAULT);
792 qti->qti_gl_desc.lquota_desc.gl_hardlimit = lqe->lqe_hardlimit;
793 qti->qti_gl_desc.lquota_desc.gl_softlimit = lqe->lqe_softlimit;
794 qti->qti_gl_desc.lquota_desc.gl_time = lqe->lqe_gracetime;
796 qti->qti_gl_desc.lquota_desc.gl_ver = ver;
798 /* look up ldlm resource associated with global index */
799 fid_build_reg_res_name(&qti->qti_fid, &qti->qti_resid);
800 res = ldlm_resource_get(pool->qpi_qmt->qmt_ns, NULL, &qti->qti_resid,
803 /* this might happen if no slaves have enqueued global quota
805 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource associated "
806 "with "DFID, PFID(&qti->qti_fid));
810 qmt_glimpse_lock(env, pool->qpi_qmt, res, &qti->qti_gl_desc,
812 ldlm_resource_putref(res);
816 /* Callback function used to select locks that should be glimpsed when
817 * broadcasting the new qunit value */
818 static int qmt_id_lock_cb(struct ldlm_lock *lock, struct lquota_entry *lqe)
820 struct obd_uuid *uuid = &(lock)->l_export->exp_client_uuid;
821 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
823 int stype = qmt_uuid2idx(uuid, &idx);
825 LASSERT(stype == QMT_STYPE_OST || stype == QMT_STYPE_MDT);
827 /* Quota pools support only OSTs, despite MDTs also could be registered
828 * as LQUOTA_RES_DT devices(DOM). */
829 if (lqe_rtype(lqe) == LQUOTA_RES_DT && stype == QMT_STYPE_MDT)
832 return lgd->lqeg_arr[idx].lge_edquot_nu ||
833 lgd->lqeg_arr[idx].lge_qunit_nu;
838 * Send glimpse request on per-ID lock to push new qunit value to slave.
840 * \param env - is the environment passed by the caller
841 * \param qmt - is the quota master target device
842 * \param lqe - is the lquota entry with the new qunit value
843 * \param uuid - is the uuid of the slave acquiring space, if any
845 static void qmt_id_lock_glimpse(const struct lu_env *env,
846 struct qmt_device *qmt,
847 struct lquota_entry *lqe, struct obd_uuid *uuid)
849 struct qmt_thread_info *qti = qmt_info(env);
850 struct qmt_pool_info *pool = lqe2qpi(lqe);
851 struct ldlm_resource *res = NULL;
854 if (!lqe->lqe_enforced)
857 lquota_generate_fid(&qti->qti_fid, pool->qpi_rtype, lqe_qtype(lqe));
858 fid_build_quota_res_name(&qti->qti_fid, &lqe->lqe_id, &qti->qti_resid);
859 res = ldlm_resource_get(qmt->qmt_ns, NULL, &qti->qti_resid, LDLM_PLAIN,
862 /* this might legitimately happens if slaves haven't had the
863 * opportunity to enqueue quota lock yet. */
864 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource for per-ID "
865 "lock "DFID, PFID(&qti->qti_fid));
867 if (lqe->lqe_revoke_time == 0 &&
868 lqe->lqe_qunit == pool->qpi_least_qunit)
869 lqe->lqe_revoke_time = ktime_get_seconds();
870 lqe_write_unlock(lqe);
875 /* The purpose of glimpse callback on per-ID lock is twofold:
876 * - notify slaves of new qunit value and hope they will release some
877 * spare quota space in return
878 * - notify slaves that master ran out of quota space and there is no
879 * need to send acquire request any more until further notice */
881 /* TODO: it is not clear how to implement below case for all lqes
882 * from where slaves will be notified in qmt_glimpse_lock. Because
883 * here we have just global lqe with an array of OSTs that should
884 * be notified. Theoretically we can find all lqes that includes
885 * these OSTs, but it is not trivial. So I would propose to move
886 * this case to another place ... */
887 if (lqe->lqe_revoke_time == 0 &&
888 lqe->lqe_qunit == pool->qpi_least_qunit)
889 /* reset lqe_may_rel, it will be updated on glimpse callback
890 * replies if needed */
891 lqe->lqe_may_rel = 0;
893 /* The rebalance thread is the only thread which can issue glimpses */
894 LASSERT(!lqe->lqe_gl);
896 lqe_write_unlock(lqe);
898 /* issue glimpse callback to slaves */
899 qmt_glimpse_lock(env, qmt, res, &qti->qti_gl_desc,
900 qmt_id_lock_cb, lqe);
903 if (lqe->lqe_revoke_time == 0 &&
904 lqe->lqe_qunit == pool->qpi_least_qunit) {
905 lqe->lqe_revoke_time = ktime_get_seconds();
906 qmt_adjust_edquot(lqe, ktime_get_real_seconds());
908 LASSERT(lqe->lqe_gl);
910 lqe_write_unlock(lqe);
912 ldlm_resource_putref(res);
917 * Schedule a glimpse request on per-ID locks to push new qunit value or
918 * edquot flag to quota slaves.
920 * \param qmt - is the quota master target device
921 * \param lqe - is the lquota entry with the new qunit value
923 void qmt_id_lock_notify(struct qmt_device *qmt, struct lquota_entry *lqe)
928 LASSERT(lqe->lqe_is_global);
930 spin_lock(&qmt->qmt_reba_lock);
931 if (!qmt->qmt_stopping && list_empty(&lqe->lqe_link)) {
932 list_add_tail(&lqe->lqe_link, &qmt->qmt_reba_list);
934 if (qmt->qmt_reba_task)
935 wake_up_process(qmt->qmt_reba_task);
937 spin_unlock(&qmt->qmt_reba_lock);
944 struct qmt_reba_args {
945 struct qmt_device *qra_dev;
946 struct lu_env qra_env;
947 struct completion *qra_started;
951 #define TASK_IDLE TASK_INTERRUPTIBLE
955 * The rebalance thread is in charge of sending glimpse callbacks on per-ID
956 * quota locks owned by slaves in order to notify them of:
957 * - a qunit shrink in which case slaves might release quota space back in
959 * - set/clear edquot flag used to cache the "quota exhausted" state of the
960 * master. When the flag is set, slaves know that there is no need to
961 * try to acquire quota from the master since this latter has already
962 * distributed all the space.
964 static int qmt_reba_thread(void *_args)
966 struct qmt_reba_args *args = _args;
967 struct qmt_device *qmt = args->qra_dev;
968 struct lu_env *env = &args->qra_env;
969 struct lquota_entry *lqe, *tmp;
972 complete(args->qra_started);
973 while (({set_current_state(TASK_IDLE);
974 !kthread_should_stop(); })) {
976 spin_lock(&qmt->qmt_reba_lock);
977 list_for_each_entry_safe(lqe, tmp, &qmt->qmt_reba_list,
979 __set_current_state(TASK_RUNNING);
980 list_del_init(&lqe->lqe_link);
981 spin_unlock(&qmt->qmt_reba_lock);
983 if (!kthread_should_stop())
984 qmt_id_lock_glimpse(env, qmt, lqe, NULL);
987 spin_lock(&qmt->qmt_reba_lock);
989 spin_unlock(&qmt->qmt_reba_lock);
992 __set_current_state(TASK_RUNNING);
1001 * Start rebalance thread. Called when the QMT is being setup
1003 int qmt_start_reba_thread(struct qmt_device *qmt)
1005 struct task_struct *task;
1006 struct qmt_reba_args *args;
1007 DECLARE_COMPLETION_ONSTACK(started);
1011 OBD_ALLOC_PTR(args);
1014 args->qra_dev = qmt;
1015 args->qra_started = &started;
1017 rc = lu_env_init(&args->qra_env, LCT_MD_THREAD);
1019 CERROR("%s: failed to init env.", qmt->qmt_svname);
1023 task = kthread_create(qmt_reba_thread, args,
1024 "qmt_reba_%s", qmt->qmt_svname);
1026 CERROR("%s: failed to start rebalance thread (%ld)\n",
1027 qmt->qmt_svname, PTR_ERR(task));
1028 GOTO(out_env_fini, rc = PTR_ERR(task));
1031 rc = lu_env_add_task(&args->qra_env, task);
1034 GOTO(out_env_fini, rc);
1036 qmt->qmt_reba_task = task;
1037 wake_up_process(task);
1038 wait_for_completion(&started);
1042 lu_env_fini(&args->qra_env);
1049 * Stop rebalance thread. Called when the QMT is about to shutdown.
1051 void qmt_stop_reba_thread(struct qmt_device *qmt)
1053 struct task_struct *task;
1055 spin_lock(&qmt->qmt_reba_lock);
1056 task = qmt->qmt_reba_task;
1057 qmt->qmt_reba_task = NULL;
1058 spin_unlock(&qmt->qmt_reba_lock);
1063 LASSERT(list_empty(&qmt->qmt_reba_list));