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 bool 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;
297 if (lgd->lqeg_arr[idx].lge_qunit == least_qunit) {
300 qti_lqes_write_lock(env);
301 for (i = 0; i < qti_lqes_cnt(env); i++) {
302 LQUOTA_DEBUG(qti_lqes(env)[i],
303 "idx %d lge_qunit %llu least_qunit %lu\n",
304 idx, lgd->lqeg_arr[idx].lge_qunit,
306 if (qti_lqes(env)[i]->lqe_qunit == least_qunit) {
307 qti_lqes(env)[i]->lqe_revoke_time =
309 notify |= qmt_adjust_edquot(qti_lqes(env)[i],
310 ktime_get_real_seconds());
313 qti_lqes_write_unlock(env);
319 * Update LVB associated with the global quota index.
320 * This function is called from the DLM itself after a glimpse callback, in this
321 * case valid ptlrpc request is passed.
323 int qmt_lvbo_update(struct lu_device *ld, struct ldlm_resource *res,
324 struct ptlrpc_request *req, int increase_only)
327 struct qmt_thread_info *qti;
328 struct qmt_device *qmt = lu2qmt_dev(ld);
329 struct lquota_entry *lqe;
330 struct lquota_lvb *lvb;
331 struct ldlm_lock *lock;
332 struct obd_export *exp;
337 LASSERT(res != NULL);
342 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] == 0)
343 /* no need to update lvb for global quota locks */
346 lvb = req_capsule_server_swab_get(&req->rq_pill, &RMF_DLM_LVB,
347 lustre_swab_lquota_lvb);
349 CERROR("%s: failed to extract lvb from request\n",
354 lqe = res->lr_lvb_data;
355 LASSERT(lqe != NULL);
358 /* allocate environement */
363 /* The request is a glimpse callback which was sent via the
364 * reverse import to the slave. What we care about here is the
365 * export associated with the slave and req->rq_export is
366 * definitely not what we are looking for (it is actually set to
368 * Therefore we extract the lock from the request argument
369 * and use lock->l_export. */
370 lock = ldlm_request_lock(req);
372 CERROR("%s: failed to get lock from request!\n",
374 GOTO(out, rc = PTR_ERR(lock));
377 exp = class_export_get(lock->l_export);
379 CERROR("%s: failed to get export from lock!\n",
381 GOTO(out, rc = -EFAULT);
384 rc = qmt_uuid2idx(&exp->exp_client_uuid, &idx);
388 need_revoke = qmt_clear_lgeg_arr_nu(lqe, rc, idx);
389 if (lvb->lvb_id_rel == 0) {
390 /* nothing to release */
391 if (lvb->lvb_id_may_rel != 0)
392 /* but might still release later ... */
393 lqe->lqe_may_rel += lvb->lvb_id_may_rel;
396 if (!need_revoke && lvb->lvb_id_rel == 0)
397 GOTO(out_exp, rc = 0);
399 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), rc, lqe_qtype(lqe),
400 &lqe->lqe_id, NULL, idx);
404 if (need_revoke && qmt_set_revoke(env, lqe, rc, idx) &&
405 lqe->lqe_glbl_data) {
406 qmt_seed_glbe_edquot(env, lqe->lqe_glbl_data);
407 qmt_id_lock_notify(qmt, lqe);
410 if (lvb->lvb_id_rel) {
411 LQUOTA_DEBUG(lqe, "releasing:%llu may release:%llu",
412 lvb->lvb_id_rel, lvb->lvb_id_may_rel);
414 /* release quota space */
415 rc = qmt_dqacq0(env, qmt, &exp->exp_client_uuid,
416 QUOTA_DQACQ_FL_REL, lvb->lvb_id_rel,
418 if (rc || qti->qti_body.qb_count != lvb->lvb_id_rel)
420 "failed to release quota space on glimpse %llu!=%llu : rc = %d\n",
421 qti->qti_body.qb_count,
422 lvb->lvb_id_rel, rc);
429 class_export_put(exp);
436 * Report size of lvb to ldlm layer in order to allocate lvb buffer
437 * As far as quota locks are concerned, the size is static and is the same
438 * for both global and per-ID locks which shares the same lvb format.
440 int qmt_lvbo_size(struct lu_device *ld, struct ldlm_lock *lock)
442 return sizeof(struct lquota_lvb);
446 * Fill request buffer with quota lvb
448 int qmt_lvbo_fill(struct lu_device *ld, struct ldlm_lock *lock, void *lvb,
451 struct ldlm_resource *res = lock->l_resource;
452 struct lquota_lvb *qlvb = lvb;
457 LASSERT(res != NULL);
460 if (res->lr_type != LDLM_PLAIN || res->lr_lvb_data == NULL ||
461 res->lr_name.name[LUSTRE_RES_ID_SEQ_OFF] != FID_SEQ_QUOTA_GLB)
467 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
468 /* no ID quota lock associated with UID/GID 0 or with a seq 0,
469 * we are thus dealing with an ID lock. */
470 struct lquota_entry *lqe = res->lr_lvb_data;
471 struct qmt_device *qmt;
472 struct obd_uuid *uuid;
475 uuid = &(lock)->l_export->exp_client_uuid;
476 rc = qmt_uuid2idx(uuid, &idx);
479 qmt = lu2qmt_dev(ld);
480 /* return current qunit value & edquot flags in lvb */
482 rc = qmt_pool_lqes_lookup(env, qmt, lqe_rtype(lqe), rc,
483 lqe_qtype(lqe), &lqe->lqe_id,
486 qlvb->lvb_id_qunit = qti_lqes_min_qunit(env);
488 if (qti_lqes_edquot(env))
489 qlvb->lvb_flags = LQUOTA_FL_EDQUOT;
492 CDEBUG(D_QUOTA, "uuid %s lqe_id %lu, edquot %llu qunit %llu\n",
493 (char *)uuid, (unsigned long)lqe->lqe_id.qid_uid,
494 qlvb->lvb_flags, qlvb->lvb_id_qunit);
497 /* global quota lock */
498 struct dt_object *obj = res->lr_lvb_data;
500 /* return current version of global index */
501 qlvb->lvb_glb_ver = dt_version_get(env, obj);
504 RETURN(rc = rc ?: sizeof(struct lquota_lvb));
508 * Free lvb associated with a given ldlm resource
509 * we don't really allocate a lvb, lr_lvb_data just points to
510 * the appropriate backend structures.
512 int qmt_lvbo_free(struct lu_device *ld, struct ldlm_resource *res)
516 if (res->lr_lvb_data == NULL)
519 if (res->lr_name.name[LUSTRE_RES_ID_QUOTA_SEQ_OFF] != 0) {
520 struct lquota_entry *lqe = res->lr_lvb_data;
521 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
523 /* release lqe reference */
524 lqe->lqe_glbl_data = NULL;
526 qmt_free_lqe_gd(lgd);
528 struct dt_object *obj = res->lr_lvb_data;
529 /* release object reference */
530 dt_object_put(lu_env_find(), obj);
533 res->lr_lvb_data = NULL;
539 typedef int (*qmt_glimpse_cb_t)(struct ldlm_lock *, struct lquota_entry *);
541 struct qmt_gl_lock_array {
544 struct ldlm_lock **q_locks;
547 static void qmt_free_lock_array(struct qmt_gl_lock_array *array)
551 if (array->q_max == 0) {
552 LASSERT(array->q_locks == NULL);
556 for (i = 0; i < array->q_cnt; i++) {
557 LASSERT(array->q_locks[i]);
558 LDLM_LOCK_RELEASE(array->q_locks[i]);
559 array->q_locks[i] = NULL;
562 OBD_FREE_PTR_ARRAY(array->q_locks, array->q_max);
563 array->q_locks = NULL;
567 static int qmt_alloc_lock_array(struct ldlm_resource *res,
568 struct qmt_gl_lock_array *array,
569 qmt_glimpse_cb_t cb, void *arg)
571 struct list_head *pos;
572 unsigned long count = 0;
576 LASSERT(!array->q_max && !array->q_cnt && !array->q_locks);
579 /* scan list of granted locks */
580 list_for_each(pos, &res->lr_granted) {
581 struct ldlm_lock *lock;
584 lock = list_entry(pos, struct ldlm_lock, l_res_link);
585 LASSERT(lock->l_export);
589 /* slave should not be notified */
595 if (array->q_max != 0 && array->q_cnt < array->q_max) {
596 array->q_locks[array->q_cnt] = LDLM_LOCK_GET(lock);
602 if (count > array->q_max) {
603 qmt_free_lock_array(array);
607 * allocate more slots in case of more qualified locks are
608 * found during next loop
610 array->q_max = count + count / 2 + 10;
612 LASSERT(array->q_locks == NULL && array->q_cnt == 0);
613 OBD_ALLOC_PTR_ARRAY(array->q_locks, array->q_max);
614 if (array->q_locks == NULL) {
624 void qmt_setup_id_desc(struct ldlm_lock *lock, union ldlm_gl_desc *desc,
625 struct lquota_entry *lqe)
627 struct obd_uuid *uuid = &(lock)->l_export->exp_client_uuid;
628 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
633 stype = qmt_uuid2idx(uuid, &idx);
636 /* DOM case - set global lqe settings */
637 if (lqe_rtype(lqe) == LQUOTA_RES_DT && stype == QMT_STYPE_MDT) {
638 edquot = lqe->lqe_edquot;
639 qunit = lqe->lqe_qunit;
641 edquot = lgd->lqeg_arr[idx].lge_edquot;
642 qunit = lgd->lqeg_arr[idx].lge_qunit;
645 /* fill glimpse descriptor with lqe settings */
646 desc->lquota_desc.gl_flags = edquot ? LQUOTA_FL_EDQUOT : 0;
647 desc->lquota_desc.gl_qunit = qunit;
648 CDEBUG(D_QUOTA, "setup desc: stype %d idx %d, edquot %llu qunit %llu\n",
649 stype, idx, desc->lquota_desc.gl_flags,
650 desc->lquota_desc.gl_qunit);
654 * Send glimpse callback to slaves holding a lock on resource \res.
655 * This is used to notify slaves of new quota settings or to claim quota space
658 * \param env - is the environment passed by the caller
659 * \param qmt - is the quota master target
660 * \param res - is the dlm resource associated with the quota object
661 * \param desc - is the glimpse descriptor to pack in glimpse callback
662 * \param cb - is the callback function called on every lock and determine
663 * whether a glimpse should be issued
664 * \param arg - is an opaq parameter passed to the callback function
666 static int qmt_glimpse_lock(const struct lu_env *env, struct qmt_device *qmt,
667 struct ldlm_resource *res, union ldlm_gl_desc *desc,
668 qmt_glimpse_cb_t cb, struct lquota_entry *lqe)
670 union ldlm_gl_desc *descs = NULL;
671 struct lqe_glbl_data *gld;
672 struct list_head *tmp, *pos;
674 struct qmt_gl_lock_array locks;
675 unsigned long i, locks_count;
679 gld = lqe ? lqe->lqe_glbl_data : NULL;
680 memset(&locks, 0, sizeof(locks));
681 rc = qmt_alloc_lock_array(res, &locks, cb, lqe);
683 CERROR("%s: failed to allocate glimpse lock array (%d)\n",
684 qmt->qmt_svname, rc);
688 CDEBUG(D_QUOTA, "%s: no granted locks to send glimpse\n",
692 CDEBUG(D_QUOTA, "found granted locks %lu\n", locks.q_cnt);
693 locks_count = locks.q_cnt;
695 /* Use one desc for all works, when called from qmt_glb_lock_notify */
696 if (gld && locks.q_cnt > 1) {
697 /* TODO: think about to store this preallocated descs
698 * in lqe_global in lqeg_arr as a part of lqe_glbl_entry.
699 * The benefit is that we don't need to allocate/free
700 * and setup this descs each time. But the drawback is
701 * memory use (sizeof ldlm_gl_desc * OST_COUNT * user_number).
702 * for examfple it could be 88 * 256 * 10 000 about 225 MB. */
704 sizeof(struct ldlm_gl_lquota_desc) * locks.q_cnt);
706 CERROR("%s: alloc glimpse lock array failed: rc = %d\n",
707 qmt->qmt_svname, rc);
708 qmt_free_lock_array(&locks);
713 for (i = locks.q_cnt; i > 0; i--) {
714 struct ldlm_glimpse_work *work;
718 CERROR("%s: failed to notify a lock.\n",
725 desc = &descs[i - 1];
726 qmt_setup_id_desc(locks.q_locks[i - 1], desc, lqe);
727 work->gl_interpret_data = lqe;
730 list_add_tail(&work->gl_list, &gl_list);
731 work->gl_lock = locks.q_locks[i - 1];
733 work->gl_desc = desc;
735 locks.q_locks[i - 1] = NULL;
739 qmt_free_lock_array(&locks);
741 if (list_empty(&gl_list)) {
742 CDEBUG(D_QUOTA, "%s: nobody to notify\n", qmt->qmt_svname);
746 /* issue glimpse callbacks to all connected slaves */
747 rc = ldlm_glimpse_locks(res, &gl_list);
749 list_for_each_safe(pos, tmp, &gl_list) {
750 struct ldlm_glimpse_work *work;
752 work = list_entry(pos, struct ldlm_glimpse_work, gl_list);
754 list_del(&work->gl_list);
755 CERROR("%s: failed to notify %s of new quota settings\n",
757 obd_uuid2str(&work->gl_lock->l_export->exp_client_uuid));
758 LDLM_LOCK_RELEASE(work->gl_lock);
764 sizeof(struct ldlm_gl_lquota_desc) * locks_count);
770 * Send glimpse request to all global quota locks to push new quota setting to
773 * \param env - is the environment passed by the caller
774 * \param lqe - is the lquota entry which has new settings
775 * \param ver - is the version associated with the setting change
777 void qmt_glb_lock_notify(const struct lu_env *env, struct lquota_entry *lqe,
780 struct qmt_thread_info *qti = qmt_info(env);
781 struct qmt_pool_info *pool = lqe2qpi(lqe);
782 struct ldlm_resource *res = NULL;
785 lquota_generate_fid(&qti->qti_fid, pool->qpi_rtype, lqe_qtype(lqe));
787 /* send glimpse callback to notify slaves of new quota settings */
788 qti->qti_gl_desc.lquota_desc.gl_id = lqe->lqe_id;
789 qti->qti_gl_desc.lquota_desc.gl_flags = 0;
790 if (lqe->lqe_is_default) {
791 qti->qti_gl_desc.lquota_desc.gl_hardlimit = 0;
792 qti->qti_gl_desc.lquota_desc.gl_softlimit = 0;
793 qti->qti_gl_desc.lquota_desc.gl_time = LQUOTA_GRACE_FLAG(0,
794 LQUOTA_FLAG_DEFAULT);
797 qti->qti_gl_desc.lquota_desc.gl_hardlimit = lqe->lqe_hardlimit;
798 qti->qti_gl_desc.lquota_desc.gl_softlimit = lqe->lqe_softlimit;
799 qti->qti_gl_desc.lquota_desc.gl_time = lqe->lqe_gracetime;
801 qti->qti_gl_desc.lquota_desc.gl_ver = ver;
803 /* look up ldlm resource associated with global index */
804 fid_build_reg_res_name(&qti->qti_fid, &qti->qti_resid);
805 res = ldlm_resource_get(pool->qpi_qmt->qmt_ns, NULL, &qti->qti_resid,
808 /* this might happen if no slaves have enqueued global quota
810 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource associated "
811 "with "DFID, PFID(&qti->qti_fid));
815 qmt_glimpse_lock(env, pool->qpi_qmt, res, &qti->qti_gl_desc,
817 ldlm_resource_putref(res);
821 /* Callback function used to select locks that should be glimpsed when
822 * broadcasting the new qunit value */
823 static int qmt_id_lock_cb(struct ldlm_lock *lock, struct lquota_entry *lqe)
825 struct obd_uuid *uuid = &(lock)->l_export->exp_client_uuid;
826 struct lqe_glbl_data *lgd = lqe->lqe_glbl_data;
828 int stype = qmt_uuid2idx(uuid, &idx);
830 LASSERT(stype == QMT_STYPE_OST || stype == QMT_STYPE_MDT);
832 /* Quota pools support only OSTs, despite MDTs also could be registered
833 * as LQUOTA_RES_DT devices(DOM). */
834 if (lqe_rtype(lqe) == LQUOTA_RES_DT && stype == QMT_STYPE_MDT)
837 return lgd->lqeg_arr[idx].lge_edquot_nu ||
838 lgd->lqeg_arr[idx].lge_qunit_nu;
843 * Send glimpse request on per-ID lock to push new qunit value to slave.
845 * \param env - is the environment passed by the caller
846 * \param qmt - is the quota master target device
847 * \param lqe - is the lquota entry with the new qunit value
848 * \param uuid - is the uuid of the slave acquiring space, if any
850 static void qmt_id_lock_glimpse(const struct lu_env *env,
851 struct qmt_device *qmt,
852 struct lquota_entry *lqe, struct obd_uuid *uuid)
854 struct qmt_thread_info *qti = qmt_info(env);
855 struct qmt_pool_info *pool = lqe2qpi(lqe);
856 struct ldlm_resource *res = NULL;
859 if (!lqe->lqe_enforced)
862 lquota_generate_fid(&qti->qti_fid, pool->qpi_rtype, lqe_qtype(lqe));
863 fid_build_quota_res_name(&qti->qti_fid, &lqe->lqe_id, &qti->qti_resid);
864 res = ldlm_resource_get(qmt->qmt_ns, NULL, &qti->qti_resid, LDLM_PLAIN,
867 /* this might legitimately happens if slaves haven't had the
868 * opportunity to enqueue quota lock yet. */
869 LQUOTA_DEBUG(lqe, "failed to lookup ldlm resource for per-ID "
870 "lock "DFID, PFID(&qti->qti_fid));
872 if (lqe->lqe_revoke_time == 0 &&
873 lqe->lqe_qunit == pool->qpi_least_qunit)
874 lqe->lqe_revoke_time = ktime_get_seconds();
875 lqe_write_unlock(lqe);
880 /* The purpose of glimpse callback on per-ID lock is twofold:
881 * - notify slaves of new qunit value and hope they will release some
882 * spare quota space in return
883 * - notify slaves that master ran out of quota space and there is no
884 * need to send acquire request any more until further notice */
886 /* TODO: it is not clear how to implement below case for all lqes
887 * from where slaves will be notified in qmt_glimpse_lock. Because
888 * here we have just global lqe with an array of OSTs that should
889 * be notified. Theoretically we can find all lqes that includes
890 * these OSTs, but it is not trivial. So I would propose to move
891 * this case to another place ... */
892 if (lqe->lqe_revoke_time == 0 &&
893 lqe->lqe_qunit == pool->qpi_least_qunit)
894 /* reset lqe_may_rel, it will be updated on glimpse callback
895 * replies if needed */
896 lqe->lqe_may_rel = 0;
898 /* The rebalance thread is the only thread which can issue glimpses */
899 LASSERT(!lqe->lqe_gl);
901 lqe_write_unlock(lqe);
903 /* issue glimpse callback to slaves */
904 qmt_glimpse_lock(env, qmt, res, &qti->qti_gl_desc,
905 qmt_id_lock_cb, lqe);
908 if (lqe->lqe_revoke_time == 0 &&
909 lqe->lqe_qunit == pool->qpi_least_qunit) {
910 lqe->lqe_revoke_time = ktime_get_seconds();
911 qmt_adjust_edquot(lqe, ktime_get_real_seconds());
913 LASSERT(lqe->lqe_gl);
915 lqe_write_unlock(lqe);
917 ldlm_resource_putref(res);
922 * Schedule a glimpse request on per-ID locks to push new qunit value or
923 * edquot flag to quota slaves.
925 * \param qmt - is the quota master target device
926 * \param lqe - is the lquota entry with the new qunit value
928 void qmt_id_lock_notify(struct qmt_device *qmt, struct lquota_entry *lqe)
933 LASSERT(lqe->lqe_is_global);
935 spin_lock(&qmt->qmt_reba_lock);
936 if (!qmt->qmt_stopping && list_empty(&lqe->lqe_link)) {
937 list_add_tail(&lqe->lqe_link, &qmt->qmt_reba_list);
939 if (qmt->qmt_reba_task)
940 wake_up_process(qmt->qmt_reba_task);
942 spin_unlock(&qmt->qmt_reba_lock);
949 struct qmt_reba_args {
950 struct qmt_device *qra_dev;
951 struct lu_env qra_env;
952 struct completion *qra_started;
956 #define TASK_IDLE TASK_INTERRUPTIBLE
960 * The rebalance thread is in charge of sending glimpse callbacks on per-ID
961 * quota locks owned by slaves in order to notify them of:
962 * - a qunit shrink in which case slaves might release quota space back in
964 * - set/clear edquot flag used to cache the "quota exhausted" state of the
965 * master. When the flag is set, slaves know that there is no need to
966 * try to acquire quota from the master since this latter has already
967 * distributed all the space.
969 static int qmt_reba_thread(void *_args)
971 struct qmt_reba_args *args = _args;
972 struct qmt_device *qmt = args->qra_dev;
973 struct lu_env *env = &args->qra_env;
974 struct lquota_entry *lqe, *tmp;
977 complete(args->qra_started);
978 while (({set_current_state(TASK_IDLE);
979 !kthread_should_stop(); })) {
981 spin_lock(&qmt->qmt_reba_lock);
982 list_for_each_entry_safe(lqe, tmp, &qmt->qmt_reba_list,
984 __set_current_state(TASK_RUNNING);
985 list_del_init(&lqe->lqe_link);
986 spin_unlock(&qmt->qmt_reba_lock);
988 if (!kthread_should_stop())
989 qmt_id_lock_glimpse(env, qmt, lqe, NULL);
992 spin_lock(&qmt->qmt_reba_lock);
994 spin_unlock(&qmt->qmt_reba_lock);
997 __set_current_state(TASK_RUNNING);
1006 * Start rebalance thread. Called when the QMT is being setup
1008 int qmt_start_reba_thread(struct qmt_device *qmt)
1010 struct task_struct *task;
1011 struct qmt_reba_args *args;
1012 DECLARE_COMPLETION_ONSTACK(started);
1016 OBD_ALLOC_PTR(args);
1019 args->qra_dev = qmt;
1020 args->qra_started = &started;
1022 rc = lu_env_init(&args->qra_env, LCT_MD_THREAD);
1024 CERROR("%s: failed to init env.\n", qmt->qmt_svname);
1028 task = kthread_create(qmt_reba_thread, args,
1029 "qmt_reba_%s", qmt->qmt_svname);
1031 CERROR("%s: failed to start rebalance thread (%ld)\n",
1032 qmt->qmt_svname, PTR_ERR(task));
1033 GOTO(out_env_fini, rc = PTR_ERR(task));
1036 rc = lu_env_add_task(&args->qra_env, task);
1039 GOTO(out_env_fini, rc);
1041 qmt->qmt_reba_task = task;
1042 wake_up_process(task);
1043 wait_for_completion(&started);
1047 lu_env_fini(&args->qra_env);
1054 * Stop rebalance thread. Called when the QMT is about to shutdown.
1056 void qmt_stop_reba_thread(struct qmt_device *qmt)
1058 struct task_struct *task;
1060 spin_lock(&qmt->qmt_reba_lock);
1061 task = qmt->qmt_reba_task;
1062 qmt->qmt_reba_task = NULL;
1063 spin_unlock(&qmt->qmt_reba_lock);
1068 LASSERT(list_empty(&qmt->qmt_reba_list));