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, 2016, 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 "qmt_internal.h"
36 * Initialize qmt-specific fields of quota entry.
38 * \param lqe - is the quota entry to initialize
39 * \param arg - is the pointer to the qmt_pool_info structure
41 static void qmt_lqe_init(struct lquota_entry *lqe, void *arg)
43 LASSERT(lqe_is_master(lqe));
45 lqe->lqe_revoke_time = 0;
46 init_rwsem(&lqe->lqe_sem);
47 mutex_init(&lqe->lqe_glbl_data_lock);
50 /* Apply the default quota setting to the specified quota entry
52 * \param env - is the environment passed by the caller
53 * \param pool - is the quota pool of the quota entry
54 * \param lqe - is the lquota_entry object to apply default quota on
55 * \param create_record - if true, an global quota record will be created and
59 * \retval -ve : other appropriate errors
61 int qmt_lqe_set_default(const struct lu_env *env, struct qmt_pool_info *pool,
62 struct lquota_entry *lqe, bool create_record)
64 struct lquota_entry *lqe_def;
69 if (lqe->lqe_id.qid_uid == 0)
72 lqe_def = pool->qpi_grace_lqe[lqe_qtype(lqe)];
74 LQUOTA_DEBUG(lqe, "inherit default quota");
76 lqe->lqe_is_default = true;
77 lqe->lqe_hardlimit = lqe_def->lqe_hardlimit;
78 lqe->lqe_softlimit = lqe_def->lqe_softlimit;
81 lqe->lqe_uptodate = true;
82 rc = qmt_set_with_lqe(env, pool->qpi_qmt, lqe, 0, 0,
83 LQUOTA_GRACE_FLAG(0, LQUOTA_FLAG_DEFAULT),
84 QIF_TIMES, true, false);
87 LQUOTA_ERROR(lqe, "failed to create the global quota"
91 if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0)
92 lqe->lqe_enforced = false;
94 lqe->lqe_enforced = true;
100 * Update a lquota entry. This is done by reading quota settings from the global
101 * index. The lquota entry must be write locked.
103 * \param env - the environment passed by the caller
104 * \param lqe - is the quota entry to refresh
105 * \param arg - is the pointer to the qmt_pool_info structure
106 * \param find - don't create lqe on disk in case of ENOENT if true
108 static int qmt_lqe_read(const struct lu_env *env, struct lquota_entry *lqe,
109 void *arg, bool find)
111 struct qmt_thread_info *qti = qmt_info(env);
112 struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
116 LASSERT(lqe_is_master(lqe));
118 /* read record from disk */
119 rc = lquota_disk_read(env, pool->qpi_glb_obj[lqe->lqe_site->lqs_qtype],
120 &lqe->lqe_id, (struct dt_rec *)&qti->qti_glb_rec);
126 qmt_lqe_set_default(env, pool, lqe, true);
129 /* copy quota settings from on-disk record */
130 lqe->lqe_granted = qti->qti_glb_rec.qbr_granted;
131 lqe->lqe_hardlimit = qti->qti_glb_rec.qbr_hardlimit;
132 lqe->lqe_softlimit = qti->qti_glb_rec.qbr_softlimit;
133 lqe->lqe_gracetime = LQUOTA_GRACE(qti->qti_glb_rec.qbr_time);
135 if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0 &&
136 (LQUOTA_FLAG(qti->qti_glb_rec.qbr_time) &
137 LQUOTA_FLAG_DEFAULT))
138 qmt_lqe_set_default(env, pool, lqe, false);
141 LQUOTA_ERROR(lqe, "failed to read quota entry from disk, rc:%d",
146 if (lqe->lqe_id.qid_uid == 0 ||
147 (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0))
148 /* {hard,soft}limit=0 means no quota enforced */
149 lqe->lqe_enforced = false;
151 lqe->lqe_enforced = true;
153 if (qmt_pool_global(pool))
154 lqe->lqe_is_global = 1;
156 LQUOTA_DEBUG(lqe, "read");
161 * Print lqe information for debugging.
163 * \param lqe - is the quota entry to debug
164 * \param arg - is the pointer to the qmt_pool_info structure
165 * \param msgdata - debug message
166 * \param fmt - format of debug message
168 static void qmt_lqe_debug(struct lquota_entry *lqe, void *arg,
169 struct libcfs_debug_msg_data *msgdata,
170 struct va_format *vaf)
172 struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
174 libcfs_debug_msg(msgdata,
175 "%pV qmt:%s pool:%s-%s id:%llu enforced:%d hard:%llu soft:%llu granted:%llu time:%llu qunit: %llu edquot:%d may_rel:%llu revoke:%lld default:%s\n",
176 vaf, pool->qpi_qmt->qmt_svname,
177 RES_NAME(pool->qpi_rtype),
179 lqe->lqe_id.qid_uid, lqe->lqe_enforced,
180 lqe->lqe_hardlimit, lqe->lqe_softlimit,
181 lqe->lqe_granted, lqe->lqe_gracetime,
182 lqe->lqe_qunit, lqe->lqe_edquot, lqe->lqe_may_rel,
183 lqe->lqe_revoke_time,
184 lqe->lqe_is_default ? "yes" : "no");
188 * Vector of quota entry operations supported on the master
190 const struct lquota_entry_operations qmt_lqe_ops = {
191 .lqe_init = qmt_lqe_init,
192 .lqe_read = qmt_lqe_read,
193 .lqe_debug = qmt_lqe_debug,
197 * Reserve enough credits to update records in both the global index and
198 * the slave index identified by \slv_obj
200 * \param env - is the environment passed by the caller
201 * \param lqe - is the quota entry associated with the identifier
202 * subject to the change. If it is NULL lqes array is
203 * taken from env with qti_lqes_env(env).
204 * \param slv_obj - is the dt_object associated with the index file
205 * \param sync - make transaction sync if true
207 struct thandle *qmt_trans_start_with_slv(const struct lu_env *env,
208 struct lquota_entry *lqe,
209 struct dt_object *slv_obj,
212 struct qmt_device *qmt;
214 struct lquota_entry **lqes;
215 struct qmt_lqe_restore *restore;
219 restore = qti_lqes_rstr(env);
221 lqes_cnt = qti_lqes_cnt(env);
222 lqes = qti_lqes(env);
228 /* qmt is the same for all lqes, so take it from the 1st */
229 qmt = lqe2qpi(lqes[0])->qpi_qmt;
232 LQUOTA_DEBUG(lqes[0], "declare write for slv "DFID,
233 PFID(lu_object_fid(&slv_obj->do_lu)));
235 /* start transaction */
236 th = dt_trans_create(env, qmt->qmt_child);
241 /* quota settings on master are updated synchronously for the
245 /* reserve credits for global index update */
246 for (i = 0; i < lqes_cnt; i++) {
247 rc = lquota_disk_declare_write(env, th,
248 LQE_GLB_OBJ(lqes[i]),
254 if (slv_obj != NULL) {
255 /* reserve credits for slave index update */
256 rc = lquota_disk_declare_write(env, th, slv_obj,
262 /* start transaction */
263 rc = dt_trans_start_local(env, qmt->qmt_child, th);
270 dt_trans_stop(env, qmt->qmt_child, th);
272 LQUOTA_ERROR(lqes[0], "failed to slv declare write for "DFID
273 ", rc:%d", PFID(lu_object_fid(&slv_obj->do_lu)),
276 for (i = 0; i < lqes_cnt; i++) {
277 restore[i].qlr_hardlimit = lqes[i]->lqe_hardlimit;
278 restore[i].qlr_softlimit = lqes[i]->lqe_softlimit;
279 restore[i].qlr_gracetime = lqes[i]->lqe_gracetime;
280 restore[i].qlr_granted = lqes[i]->lqe_granted;
281 restore[i].qlr_qunit = lqes[i]->lqe_qunit;
288 * Reserve enough credits to update a record in the global index
290 * \param env - is the environment passed by the caller
291 * \param lqe - is the quota entry to be modified in the global index
292 * \param restore - is a temporary storage for current quota settings which will
293 * be restored if something goes wrong at index update time.
295 struct thandle *qmt_trans_start(const struct lu_env *env,
296 struct lquota_entry *lqe)
298 LQUOTA_DEBUG(lqe, "declare write");
299 return qmt_trans_start_with_slv(env, lqe, NULL, true);
302 int qmt_glb_write_lqes(const struct lu_env *env, struct thandle *th,
303 __u32 flags, __u64 *ver)
307 for (i = 0; i < qti_lqes_cnt(env); i++) {
308 rc = qmt_glb_write(env, th, qti_lqes(env)[i], flags, ver);
316 * Update record associated with a quota entry in the global index.
317 * If LQUOTA_BUMP_VER is set, then the global index version must also be
319 * The entry must be at least read locked, dirty and up-to-date.
321 * \param env - the environment passed by the caller
322 * \param th - is the transaction handle to be used for the disk writes
323 * \param lqe - is the quota entry to udpate
324 * \param obj - is the dt_object associated with the index file
325 * \param flags - can be LQUOTA_BUMP_VER or LQUOTA_SET_VER.
326 * \param ver - is used to return the new version of the index.
328 * \retval - 0 on success and lqe dirty flag cleared,
329 * appropriate error on failure and uptodate flag cleared.
331 int qmt_glb_write(const struct lu_env *env, struct thandle *th,
332 struct lquota_entry *lqe, __u32 flags, __u64 *ver)
334 struct qmt_thread_info *qti = qmt_info(env);
335 struct lquota_glb_rec *rec;
339 LASSERT(lqe != NULL);
340 LASSERT(lqe_is_master(lqe));
341 LASSERT(lqe_is_locked(lqe));
342 LASSERT(lqe->lqe_uptodate);
343 LASSERT((flags & ~(LQUOTA_BUMP_VER | LQUOTA_SET_VER)) == 0);
345 LQUOTA_DEBUG(lqe, "write glb");
347 /* never delete the entry even when the id isn't enforced and
348 * no any guota granted, otherwise, this entry will not be
349 * synced to slave during the reintegration. */
350 rec = &qti->qti_glb_rec;
352 /* fill global index with updated quota settings */
353 rec->qbr_granted = lqe->lqe_granted;
354 if (lqe->lqe_is_default) {
355 rec->qbr_hardlimit = 0;
356 rec->qbr_softlimit = 0;
357 rec->qbr_time = LQUOTA_GRACE_FLAG(lqe->lqe_gracetime,
358 LQUOTA_FLAG_DEFAULT);
360 rec->qbr_hardlimit = lqe->lqe_hardlimit;
361 rec->qbr_softlimit = lqe->lqe_softlimit;
362 rec->qbr_time = lqe->lqe_gracetime;
365 /* write new quota settings */
366 rc = lquota_disk_write(env, th, LQE_GLB_OBJ(lqe), &lqe->lqe_id,
367 (struct dt_rec *)rec, flags, ver);
369 /* we failed to write the new quota settings to disk, report
370 * error to caller who will restore the initial value */
371 LQUOTA_ERROR(lqe, "failed to update global index, rc:%d", rc);
377 * Read from disk how much quota space is allocated to a slave.
378 * This is done by reading records from the dedicated slave index file.
379 * Return in \granted how much quota space is currently allocated to the
381 * The entry must be at least read locked.
383 * \param env - the environment passed by the caller
384 * \param lqe_id - is the quota id associated with the identifier to look-up
386 * \param slv_obj - is the dt_object associated with the slave index
387 * \param granted - is the output parameter where to return how much space
388 * is granted to the slave.
390 * \retval - 0 on success, appropriate error on failure
392 int qmt_slv_read(const struct lu_env *env, union lquota_id *qid,
393 struct dt_object *slv_obj, __u64 *granted)
395 struct qmt_thread_info *qti = qmt_info(env);
396 struct lquota_slv_rec *slv_rec = &qti->qti_slv_rec;
400 CDEBUG(D_QUOTA, "read id:%llu form slv "DFID"\n",
401 qid->qid_uid, PFID(lu_object_fid(&slv_obj->do_lu)));
403 /* read slave record from disk */
404 rc = lquota_disk_read(env, slv_obj, qid,
405 (struct dt_rec *)slv_rec);
411 /* extract granted from on-disk record */
412 *granted = slv_rec->qsr_granted;
415 CERROR("Failed to read slave record for %llu from "DFID"\n",
416 qid->qid_uid, PFID(lu_object_fid(&slv_obj->do_lu)));
420 CDEBUG(D_QUOTA, "Successful slv read %llu\n", *granted);
426 * Update record in slave index file.
427 * The entry must be at least read locked.
429 * \param env - the environment passed by the caller
430 * \param th - is the transaction handle to be used for the disk writes
431 * \param lqe - is the dirty quota entry which will be updated at the same time
433 * \param slv_obj - is the dt_object associated with the slave index
434 * \param flags - can be LQUOTA_BUMP_VER or LQUOTA_SET_VER.
435 * \param ver - is used to return the new version of the index.
436 * \param granted - is the new amount of quota space owned by the slave
438 * \retval - 0 on success, appropriate error on failure
440 int qmt_slv_write(const struct lu_env *env, struct thandle *th,
441 struct lquota_entry *lqe, struct dt_object *slv_obj,
442 __u32 flags, __u64 *ver, __u64 granted)
444 struct qmt_thread_info *qti = qmt_info(env);
445 struct lquota_slv_rec *rec;
449 LASSERT(lqe != NULL);
450 LASSERT(lqe_is_master(lqe));
451 LASSERT(lqe_is_locked(lqe));
453 LQUOTA_DEBUG(lqe, "write slv "DFID" granted:%llu",
454 PFID(lu_object_fid(&slv_obj->do_lu)), granted);
456 /* never delete the entry, otherwise, it'll not be transferred
457 * to slave during reintegration. */
458 rec = &qti->qti_slv_rec;
460 /* updated space granted to this slave */
461 rec->qsr_granted = granted;
463 /* write new granted space */
464 rc = lquota_disk_write(env, th, slv_obj, &lqe->lqe_id,
465 (struct dt_rec *)rec, flags, ver);
468 "failed to update slave index "DFID" granted:%llu",
469 PFID(lu_object_fid(&slv_obj->do_lu)),
478 * Check whether new limits are valid for this pool
480 * \param lqe - is the quota entry subject to the setquota
481 * \param hard - is the new hard limit
482 * \param soft - is the new soft limit
484 int qmt_validate_limits(struct lquota_entry *lqe, __u64 hard, __u64 soft)
488 if (hard != 0 && soft > hard)
489 /* soft limit must be less than hard limit */
495 * Set/clear edquot flag after quota space allocation/release or settings
496 * change. Slaves will be notified of changes via glimpse on per-ID lock
498 * \param lqe - is the quota entry to check
499 * \param now - is the current time in second used for grace time managment
501 bool qmt_adjust_edquot(struct lquota_entry *lqe, __u64 now)
503 struct qmt_pool_info *pool = lqe2qpi(lqe);
506 if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
509 if (!lqe->lqe_edquot) {
510 /* space exhausted flag not set, let's check whether it is time
513 if (!qmt_space_exhausted(lqe, now))
514 /* the qmt still has available space */
517 /* See comment in qmt_adjust_qunit(). LU-4139 */
518 if (qmt_hard_exhausted(lqe) ||
519 pool->qpi_rtype != LQUOTA_RES_DT) {
522 /* we haven't reached the minimal qunit yet so there is
523 * still hope that the rebalancing process might free
524 * up some quota space */
525 if (lqe->lqe_qunit != pool->qpi_least_qunit)
528 /* least qunit value not sent to all slaves yet */
529 if (lqe->lqe_revoke_time == 0)
532 /* Let's give more time to slave to release space */
533 lapse = ktime_get_seconds() - QMT_REBA_TIMEOUT;
534 if (lqe->lqe_may_rel != 0 && lqe->lqe_revoke_time > lapse)
537 if (lqe->lqe_qunit > pool->qpi_soft_least_qunit)
541 /* set edquot flag */
542 lqe->lqe_edquot = true;
544 /* space exhausted flag set, let's check whether it is time to
547 if (qmt_space_exhausted(lqe, now))
548 /* the qmt still has not space */
551 if (lqe->lqe_hardlimit != 0 &&
552 lqe->lqe_granted + pool->qpi_least_qunit >
554 /* we clear the flag only once at least one least qunit
558 /* clear edquot flag */
559 lqe->lqe_edquot = false;
562 LQUOTA_DEBUG(lqe, "changing edquot flag");
564 /* let's notify slave by issuing glimpse on per-ID lock.
565 * the rebalance thread will take care of this */
569 /* Using least_qunit when over block softlimit will seriously impact the
570 * write performance, we need to do some special tweaking on that. */
571 static __u64 qmt_calc_softlimit(struct lquota_entry *lqe, bool *oversoft)
573 struct qmt_pool_info *pool = lqe2qpi(lqe);
575 LASSERT(lqe->lqe_softlimit != 0);
577 /* No need to do special tweaking for inode limit */
578 if (pool->qpi_rtype != LQUOTA_RES_DT)
579 return lqe->lqe_softlimit;
581 if (lqe->lqe_granted <= lqe->lqe_softlimit +
582 pool->qpi_soft_least_qunit) {
583 return lqe->lqe_softlimit;
584 } else if (lqe->lqe_hardlimit != 0) {
586 return lqe->lqe_hardlimit;
594 * Try to grant more quota space back to slave.
596 * \param lqe - is the quota entry for which we would like to allocate more
598 * \param granted - is how much was already granted as part of the request
600 * \param spare - is how much unused quota space the slave already owns
602 * \retval return how additional space can be granted to the slave
604 __u64 qmt_alloc_expand(struct lquota_entry *lqe, __u64 granted, __u64 spare)
606 struct qmt_pool_info *pool = lqe2qpi(lqe);
607 __u64 remaining, qunit;
610 LASSERT(lqe->lqe_enforced && lqe->lqe_qunit != 0);
612 slv_cnt = qpi_slv_nr(lqe2qpi(lqe), lqe_qtype(lqe));
613 qunit = lqe->lqe_qunit;
615 /* See comment in qmt_adjust_qunit(). LU-4139. */
616 if (lqe->lqe_softlimit != 0) {
618 remaining = qmt_calc_softlimit(lqe, &oversoft);
620 remaining = lqe->lqe_granted +
621 pool->qpi_soft_least_qunit;
623 remaining = lqe->lqe_hardlimit;
626 if (lqe->lqe_granted >= remaining)
629 remaining -= lqe->lqe_granted;
635 granted &= (qunit - 1);
637 if (remaining > (slv_cnt * qunit) >> 1) {
638 /* enough room to grant more space w/o additional
639 * shrinking ... at least for now */
640 remaining -= (slv_cnt * qunit) >> 1;
641 } else if (qunit != pool->qpi_least_qunit) {
646 granted &= (qunit - 1);
648 RETURN(min_t(__u64, qunit - spare, remaining));
650 RETURN(min_t(__u64, qunit - granted, remaining));
651 } while (qunit >= pool->qpi_least_qunit);
657 qmt_adjust_qunit_set_revoke(const struct lu_env *env, struct lquota_entry *lqe,
658 unsigned long least_qunit)
660 struct lquota_entry *lqe2;
664 if (qti_lqes_cnt(env) <= 1)
667 for (i = 0; i < qti_lqes_cnt(env); i++) {
668 lqe2 = qti_lqes(env)[i];
669 if ((lqe2->lqe_qunit == least_qunit) && lqe2->lqe_revoke_time) {
671 min = lqe2->lqe_revoke_time;
674 min = lqe2->lqe_revoke_time < min ?
675 lqe2->lqe_revoke_time : min;
679 lqe->lqe_revoke_time = min;
684 * Adjust qunit size according to quota limits and total granted count.
685 * The caller must have locked the lqe.
687 * \param env - the environment passed by the caller
688 * \param lqe - is the qid entry to be adjusted
689 * \retval true - need reseed glbe array
691 bool qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
693 struct qmt_pool_info *pool = lqe2qpi(lqe);
694 bool need_reseed = false;
696 __u64 qunit, limit, qunit2 = 0;
699 LASSERT(lqe_is_locked(lqe));
701 if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
702 /* no quota limits */
705 /* record how many slaves have already registered */
706 slv_cnt = qpi_slv_nr(pool, lqe_qtype(lqe));
708 /* Pool hasn't slaves anymore. Qunit will be adjusted
709 * again when new slaves would be added. */
710 if (lqe->lqe_qunit) {
714 /* wait for at least one slave to join */
718 /* Qunit calculation is based on soft limit, if any, hard limit
719 * otherwise. This means that qunit is shrunk to the minimum when
720 * beyond the soft limit. This will impact performance, but that's the
721 * price of an accurate grace time management. */
722 if (lqe->lqe_softlimit != 0) {
724 /* As a compromise of write performance and the grace time
725 * accuracy, the block qunit size will be shrunk to
726 * qpi_soft_least_qunit when over softlimit. LU-4139. */
727 limit = qmt_calc_softlimit(lqe, &oversoft);
729 qunit2 = pool->qpi_soft_least_qunit;
731 GOTO(done, qunit = qunit2);
732 } else if (lqe->lqe_hardlimit != 0) {
733 limit = lqe->lqe_hardlimit;
735 LQUOTA_ERROR(lqe, "enforced bit set, but neither hard nor soft "
740 qunit = lqe->lqe_qunit == 0 ? pool->qpi_least_qunit : lqe->lqe_qunit;
742 /* The qunit value is computed as follows: limit / (2 * slv_cnt).
743 * Then 75% of the quota space can be granted with current qunit value.
744 * The remaining 25% are then used with reduced qunit size (by a factor
745 * of 4) which is then divided in a similar manner.
747 * |---------------------limit---------------------|
748 * |-------limit / 2-------|-limit / 4-|-limit / 4-|
749 * |qunit|qunit|qunit|qunit| | |
750 * |----slv_cnt * qunit----| | |
751 * |-grow limit-| | | |
752 * |--------------shrink limit---------| |
753 * |---space granted in qunit chunks---|-remaining-|
759 * qunit >>= 2; |qunit*slv_cnt|qunit*slv_cnt|
760 * |---space in qunit---|remain|
762 if (qunit == pool->qpi_least_qunit ||
763 limit >= lqe->lqe_granted + ((slv_cnt * qunit) >> 1)) {
764 /* current qunit value still fits, let's see if we can afford to
765 * increase qunit now ...
766 * To increase qunit again, we have to be under 25% */
767 while (qunit && limit >= lqe->lqe_granted + 6 * qunit * slv_cnt)
772 do_div(qunit, 2 * slv_cnt);
776 /* shrink qunit until we find a suitable value */
777 while (qunit > pool->qpi_least_qunit &&
778 limit < lqe->lqe_granted + ((slv_cnt * qunit) >> 1))
782 if (qunit2 && qunit > qunit2)
785 if (lqe->lqe_qunit == qunit)
786 /* keep current qunit */
789 LQUOTA_DEBUG(lqe, "%s qunit to %llu",
790 lqe->lqe_qunit < qunit ? "increasing" : "decreasing",
793 /* store new qunit value */
794 swap(lqe->lqe_qunit, qunit);
796 /* reseed glbe array and notify
797 * slave if qunit was shrinked */
799 /* reset revoke time */
800 lqe->lqe_revoke_time = 0;
802 if (lqe->lqe_qunit == pool->qpi_least_qunit) {
803 if (lqe->lqe_qunit >= qunit)
804 /* initial qunit value is the smallest one */
805 lqe->lqe_revoke_time = ktime_get_seconds();
806 /* If there are several lqes and lqe_revoke_time is set for
807 * some of them, it means appropriate OSTs have been already
808 * notified with the least qunit and there is no chance to
809 * free more space. Find an lqe with the minimum(earliest)
810 * revoke_time and set this time to the current one.
812 qmt_adjust_qunit_set_revoke(env, lqe, pool->qpi_least_qunit);
817 bool qmt_adjust_edquot_qunit_notify(const struct lu_env *env,
818 struct qmt_device *qmt,
819 __u64 now, bool edquot,
820 bool qunit, __u32 qb_flags)
822 struct lquota_entry *lqe_gl, *lqe;
823 bool need_reseed = false;
826 lqe_gl = qti_lqes_glbl(env);
828 for (i = 0; i < qti_lqes_cnt(env); i++) {
829 lqe = qti_lqes(env)[i];
831 need_reseed |= qmt_adjust_qunit(env, lqe);
833 need_reseed |= qmt_adjust_edquot(lqe, now);
837 if (!lqe_gl->lqe_glbl_data &&
838 (req_has_rep(qb_flags) || req_is_rel(qb_flags))) {
841 "%s: can not notify - lge_glbl_data is not set\n",
847 mutex_lock(&lqe_gl->lqe_glbl_data_lock);
848 if (lqe_gl->lqe_glbl_data)
849 qmt_seed_glbe_all(env, lqe_gl->lqe_glbl_data, qunit,
851 mutex_unlock(&lqe_gl->lqe_glbl_data_lock);
853 qmt_id_lock_notify(qmt, lqe_gl);
860 * Adjust qunit & edquot flag in case it wasn't initialized already (e.g.
861 * limit set while no slaves were connected yet)
863 bool qmt_revalidate(const struct lu_env *env, struct lquota_entry *lqe)
865 bool need_notify = false;
867 if (lqe->lqe_qunit == 0) {
868 /* lqe was read from disk, but neither qunit, nor edquot flag
869 * were initialized */
870 need_notify = qmt_adjust_qunit(env, lqe);
871 if (lqe->lqe_qunit != 0)
872 need_notify |= qmt_adjust_edquot(lqe,
873 ktime_get_real_seconds());
879 void qmt_revalidate_lqes(const struct lu_env *env,
880 struct qmt_device *qmt, __u32 qb_flags)
882 struct lquota_entry *lqe_gl = qti_lqes_glbl(env);
883 bool need_notify = false;
886 for (i = 0; i < qti_lqes_cnt(env); i++)
887 need_notify |= qmt_revalidate(env, qti_lqes(env)[i]);
892 /* There could be no ID lock to the moment of reconciliation.
893 * As a result lqe global data is not initialised yet. It is ok
894 * for release and report requests. */
895 if (!lqe_gl->lqe_glbl_data &&
896 (req_is_rel(qb_flags) || req_has_rep(qb_flags))) {
900 mutex_lock(&lqe_gl->lqe_glbl_data_lock);
901 if (lqe_gl->lqe_glbl_data)
902 qmt_seed_glbe(env, lqe_gl->lqe_glbl_data);
903 mutex_unlock(&lqe_gl->lqe_glbl_data_lock);
905 qmt_id_lock_notify(qmt, lqe_gl);
908 void qti_lqes_init(const struct lu_env *env)
910 struct qmt_thread_info *qti = qmt_info(env);
912 qti->qti_lqes_cnt = 0;
913 qti->qti_glbl_lqe_idx = 0;
914 qti->qti_lqes_num = QMT_MAX_POOL_NUM;
917 int qti_lqes_add(const struct lu_env *env, struct lquota_entry *lqe)
919 struct qmt_thread_info *qti = qmt_info(env);
921 if (qti->qti_lqes_cnt >= qti->qti_lqes_num) {
922 struct lquota_entry **lqes;
923 lqes = qti->qti_lqes;
924 OBD_ALLOC(lqes, sizeof(lqe) * qti->qti_lqes_num * 2);
927 memcpy(lqes, qti_lqes(env), qti->qti_lqes_cnt * sizeof(lqe));
928 /* Don't need to free, if it is the very 1st allocation */
929 if (qti->qti_lqes_num > QMT_MAX_POOL_NUM)
930 OBD_FREE(qti->qti_lqes,
931 qti->qti_lqes_num * sizeof(lqe));
932 qti->qti_lqes = lqes;
933 qti->qti_lqes_num *= 2;
936 if (lqe->lqe_is_global)
937 qti->qti_glbl_lqe_idx = qti->qti_lqes_cnt;
938 qti_lqes(env)[qti->qti_lqes_cnt++] = lqe;
940 /* The pool could be accessed directly from lqe, so take
941 * extra reference that is put in qti_lqes_fini */
942 qpi_getref(lqe2qpi(lqe));
944 CDEBUG(D_QUOTA, "LQE %p %lu is added, lqe_cnt %d lqes_num %d\n",
945 lqe, (long unsigned)lqe->lqe_id.qid_uid,
946 qti->qti_lqes_cnt, qti->qti_lqes_num);
947 LASSERT(qti->qti_lqes_num != 0);
952 void qti_lqes_del(const struct lu_env *env, int index)
954 struct lquota_entry **lqes;
955 int lqes_cnt = qti_lqes_cnt(env);
956 int lqep_size = sizeof(struct lquota_entry *);
959 /* We can't handle non global lqes correctly without
960 * global lqe located at index 0. If we try to do so,
961 * something goes wrong. */
962 LQUOTA_ERROR(qti_lqes_glbl(env),
963 "quota: cannot remove lqe at index 0 as it is global");
964 LASSERT(qti_lqes_glbl(env)->lqe_is_global);
967 lqes = qti_lqes(env);
968 qpi_putref(env, lqe2qpi(lqes[index]));
969 lqe_putref(lqes[index]);
970 memcpy((unsigned char *)lqes + index * lqep_size,
971 (unsigned char *)lqes + (index + 1) * lqep_size,
972 (lqes_cnt - index - 1) * lqep_size);
976 void qti_lqes_fini(const struct lu_env *env)
978 struct qmt_thread_info *qti = qmt_info(env);
979 struct lquota_entry **lqes = qti->qti_lqes;
982 lqes = qti_lqes(env);
983 for (i = 0; i < qti->qti_lqes_cnt; i++) {
984 qpi_putref(env, lqe2qpi(lqes[i]));
988 if (qti->qti_lqes_num > QMT_MAX_POOL_NUM)
989 OBD_FREE(qti->qti_lqes,
990 qti->qti_lqes_num * sizeof(struct lquota_entry *));
992 qti->qti_lqes_num = 0;
993 qti->qti_lqes_cnt = 0;
996 __u64 qti_lqes_min_qunit(const struct lu_env *env)
1001 for (i = 1, min = qti_lqe_qunit(env, 0); i < qti_lqes_cnt(env); i++) {
1002 qunit = qti_lqe_qunit(env, i);
1003 /* if qunit is 0, lqe is not enforced and we can ignore it */
1004 if (qunit && qunit < min)
1011 int qti_lqes_edquot(const struct lu_env *env)
1015 for (i = 0; i < qti_lqes_cnt(env); i++) {
1016 if (qti_lqes(env)[i]->lqe_edquot)
1023 int qti_lqes_restore_init(const struct lu_env *env)
1027 if (qti_lqes_inited(env) && qti_lqes_cnt(env) > QMT_MAX_POOL_NUM) {
1028 OBD_ALLOC(qmt_info(env)->qti_lqes_rstr,
1029 qti_lqes_cnt(env) * sizeof(struct qmt_lqe_restore));
1030 if (!qti_lqes_rstr(env))
1037 void qti_lqes_restore_fini(const struct lu_env *env)
1039 if (qti_lqes_inited(env) && qti_lqes_cnt(env) > QMT_MAX_POOL_NUM)
1040 OBD_FREE(qmt_info(env)->qti_lqes_rstr,
1041 qti_lqes_cnt(env) * sizeof(struct qmt_lqe_restore));
1044 void qti_lqes_write_lock(const struct lu_env *env)
1048 for (i = 0; i < qti_lqes_cnt(env); i++)
1049 lqe_write_lock(qti_lqes(env)[i]);
1052 void qti_lqes_write_unlock(const struct lu_env *env)
1056 for (i = 0; i < qti_lqes_cnt(env); i++)
1057 lqe_write_unlock(qti_lqes(env)[i]);
1060 #define QMT_INIT_SLV_CNT 64
1061 struct lqe_glbl_data *qmt_alloc_lqe_gd(struct qmt_pool_info *pool, int qtype)
1063 struct lqe_glbl_data *lgd;
1064 struct lqe_glbl_entry *lqeg_arr;
1065 int slv_cnt, glbe_num;
1067 OBD_ALLOC(lgd, sizeof(struct lqe_glbl_data));
1071 slv_cnt = qpi_slv_nr_by_rtype(pool, qtype);
1073 glbe_num = slv_cnt < QMT_INIT_SLV_CNT ? QMT_INIT_SLV_CNT : slv_cnt;
1074 OBD_ALLOC(lqeg_arr, sizeof(struct lqe_glbl_entry) * glbe_num);
1076 OBD_FREE(lgd, sizeof(struct lqe_glbl_data));
1080 CDEBUG(D_QUOTA, "slv_cnt %d glbe_num %d\n", slv_cnt, glbe_num);
1082 lgd->lqeg_num_used = slv_cnt;
1083 lgd->lqeg_num_alloc = glbe_num;
1084 lgd->lqeg_arr = lqeg_arr;
1089 void qmt_free_lqe_gd(struct lqe_glbl_data *lgd)
1091 OBD_FREE(lgd->lqeg_arr,
1092 sizeof(struct lqe_glbl_entry) * lgd->lqeg_num_alloc);
1093 OBD_FREE(lgd, sizeof(struct lqe_glbl_data));
1096 void qmt_seed_glbe_all(const struct lu_env *env, struct lqe_glbl_data *lgd,
1097 bool qunit, bool edquot)
1099 struct rw_semaphore *sem = NULL;
1100 struct qmt_pool_info *qpi;
1104 if (!qti_lqes_cnt(env))
1106 /* lqes array is sorted by qunit - the first entry has minimum qunit.
1107 * Thus start seeding global qunit's array beginning from the 1st lqe
1108 * and appropriate pool. If pools overlapped, slaves from this
1109 * overlapping get minimum qunit value.
1110 * user1: pool1, pool2, pool_glbl;
1111 * pool1: OST1; user1_qunit = 10M;
1112 * pool2: OST0, OST1, OST2; user1_qunit = 30M;
1113 * pool_glbl: OST0, OST1, OST2, OST3; user1_qunit = 160M;
1114 * qunit array after seeding should be:
1115 * OST0: 30M; OST1: 10M; OST2: 30M; OST3: 160M; */
1117 /* edquot resetup algorythm works fine
1118 * with not sorted lqes */
1122 for (i = 0; i < lgd->lqeg_num_used; i++) {
1123 lgd->lqeg_arr[i].lge_qunit_set = 0;
1124 lgd->lqeg_arr[i].lge_qunit_nu = 0;
1125 lgd->lqeg_arr[i].lge_edquot_nu = 0;
1128 for (i = 0; i < qti_lqes_cnt(env); i++) {
1129 struct lquota_entry *lqe = qti_lqes(env)[i];
1132 CDEBUG(D_QUOTA, "lqes_cnt %d, i %d\n", qti_lqes_cnt(env), i);
1134 if (qmt_pool_global(qpi)) {
1135 slaves_cnt = qpi_slv_nr_by_rtype(lqe2qpi(lqe),
1138 sem = qmt_sarr_rwsem(qpi);
1140 slaves_cnt = qmt_sarr_count(qpi);
1143 for (j = 0; j < slaves_cnt; j++) {
1144 idx = qmt_sarr_get_idx(qpi, j);
1148 int lge_edquot, new_edquot, edquot_nu;
1150 lge_edquot = lgd->lqeg_arr[idx].lge_edquot;
1151 edquot_nu = lgd->lqeg_arr[idx].lge_edquot_nu;
1152 new_edquot = lqe->lqe_edquot;
1154 if (lge_edquot == new_edquot ||
1155 (edquot_nu && lge_edquot == 1))
1157 lgd->lqeg_arr[idx].lge_edquot = new_edquot;
1158 /* it is needed for the following case:
1159 * initial values for idx i -
1160 * lqe_edquot = 1, lqe_edquot_nu == 0;
1161 * 1: new_edquot == 0 ->
1162 * lqe_edquot = 0, lqe_edquot_nu = 1;
1163 * 2: new_edquot == 1 ->
1164 * lqe_edquot = 1, lqe_edquot_nu = 0;
1165 * At the 2nd iteration lge_edquot comes back
1166 * to 1, so no changes and we don't need
1167 * to notify slave. */
1168 lgd->lqeg_arr[idx].lge_edquot_nu = !edquot_nu;
1172 __u64 lge_qunit, new_qunit;
1175 "idx %d lge_qunit_set %d lge_qunit %llu new_qunit %llu\n",
1176 idx, lgd->lqeg_arr[idx].lge_qunit_set,
1177 lgd->lqeg_arr[idx].lge_qunit,
1179 /* lge for this idx is already set
1180 * on previous iteration */
1181 if (lgd->lqeg_arr[idx].lge_qunit_set)
1183 lge_qunit = lgd->lqeg_arr[idx].lge_qunit;
1184 new_qunit = lqe->lqe_qunit;
1185 /* qunit could be not set,
1186 * so use global lqe's qunit */
1190 if (lge_qunit != new_qunit)
1191 lgd->lqeg_arr[idx].lge_qunit =
1194 /* TODO: initially slaves notification was done
1195 * only for qunit shrinking. Should we always
1196 * notify slaves with new qunit ? */
1197 if (lge_qunit > new_qunit)
1198 lgd->lqeg_arr[idx].lge_qunit_nu = 1;
1199 lgd->lqeg_arr[idx].lge_qunit_set = 1;
1203 if (!qmt_pool_global(qpi))
1206 /* TODO: only for debug purposes - remove it later */
1207 for (i = 0; i < lgd->lqeg_num_used; i++)
1209 "lgd ost %d, qunit %lu nu %d; edquot %d nu %d\n",
1210 i, (long unsigned)lgd->lqeg_arr[i].lge_qunit,
1211 lgd->lqeg_arr[i].lge_qunit_nu,
1212 lgd->lqeg_arr[i].lge_edquot,
1213 lgd->lqeg_arr[i].lge_edquot_nu);
1218 void qmt_setup_lqe_gd(const struct lu_env *env, struct qmt_device *qmt,
1219 struct lquota_entry *lqe, struct lqe_glbl_data *lgd,
1226 qunit = lqe->lqe_qunit;
1227 edquot = lqe->lqe_edquot;
1229 /* Firstly set all elements in array with
1230 * qunit and edquot of global pool */
1231 for (i = 0; i < lgd->lqeg_num_used; i++) {
1232 lgd->lqeg_arr[i].lge_qunit = qunit;
1233 lgd->lqeg_arr[i].lge_edquot = edquot;
1234 /* It is the very first lvb setup - qunit and other flags
1235 * will be sent to slaves during qmt_lvbo_fill. */
1236 lgd->lqeg_arr[i].lge_qunit_nu = 0;
1237 lgd->lqeg_arr[i].lge_edquot_nu = 0;
1240 qmt_pool_lqes_lookup_spec(env, qmt, pool_type,
1241 lqe_qtype(lqe), &lqe->lqe_id);
1242 qmt_seed_glbe(env, lgd);
1244 lqe->lqe_glbl_data = lgd;
1245 qmt_id_lock_notify(qmt, lqe);