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);
49 /* Apply the default quota setting to the specified quota entry
51 * \param env - is the environment passed by the caller
52 * \param pool - is the quota pool of the quota entry
53 * \param lqe - is the lquota_entry object to apply default quota on
54 * \param create_record - if true, an global quota record will be created and
58 * \retval -ve : other appropriate errors
60 int qmt_lqe_set_default(const struct lu_env *env, struct qmt_pool_info *pool,
61 struct lquota_entry *lqe, bool create_record)
63 struct lquota_entry *lqe_def;
68 if (lqe->lqe_id.qid_uid == 0)
71 lqe_def = pool->qpi_grace_lqe[lqe_qtype(lqe)];
73 LQUOTA_DEBUG(lqe, "inherit default quota");
75 lqe->lqe_is_default = true;
76 lqe->lqe_hardlimit = lqe_def->lqe_hardlimit;
77 lqe->lqe_softlimit = lqe_def->lqe_softlimit;
80 lqe->lqe_uptodate = true;
81 rc = qmt_set_with_lqe(env, pool->qpi_qmt, lqe, 0, 0,
82 LQUOTA_GRACE_FLAG(0, LQUOTA_FLAG_DEFAULT),
83 QIF_TIMES, true, false);
86 LQUOTA_ERROR(lqe, "failed to create the global quota"
90 if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0)
91 lqe->lqe_enforced = false;
93 lqe->lqe_enforced = true;
99 * Update a lquota entry. This is done by reading quota settings from the global
100 * index. The lquota entry must be write locked.
102 * \param env - the environment passed by the caller
103 * \param lqe - is the quota entry to refresh
104 * \param arg - is the pointer to the qmt_pool_info structure
105 * \param find - don't create lqe on disk in case of ENOENT if true
107 static int qmt_lqe_read(const struct lu_env *env, struct lquota_entry *lqe,
108 void *arg, bool find)
110 struct qmt_thread_info *qti = qmt_info(env);
111 struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
115 LASSERT(lqe_is_master(lqe));
117 /* read record from disk */
118 rc = lquota_disk_read(env, pool->qpi_glb_obj[lqe->lqe_site->lqs_qtype],
119 &lqe->lqe_id, (struct dt_rec *)&qti->qti_glb_rec);
125 qmt_lqe_set_default(env, pool, lqe, true);
128 /* copy quota settings from on-disk record */
129 lqe->lqe_granted = qti->qti_glb_rec.qbr_granted;
130 lqe->lqe_hardlimit = qti->qti_glb_rec.qbr_hardlimit;
131 lqe->lqe_softlimit = qti->qti_glb_rec.qbr_softlimit;
132 lqe->lqe_gracetime = LQUOTA_GRACE(qti->qti_glb_rec.qbr_time);
134 if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0 &&
135 (LQUOTA_FLAG(qti->qti_glb_rec.qbr_time) &
136 LQUOTA_FLAG_DEFAULT))
137 qmt_lqe_set_default(env, pool, lqe, false);
140 LQUOTA_ERROR(lqe, "failed to read quota entry from disk, rc:%d",
145 if (lqe->lqe_id.qid_uid == 0 ||
146 (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0))
147 /* {hard,soft}limit=0 means no quota enforced */
148 lqe->lqe_enforced = false;
150 lqe->lqe_enforced = true;
152 if (qmt_pool_global(pool))
153 lqe->lqe_is_global = 1;
155 LQUOTA_DEBUG(lqe, "read");
160 * Print lqe information for debugging.
162 * \param lqe - is the quota entry to debug
163 * \param arg - is the pointer to the qmt_pool_info structure
164 * \param msgdata - debug message
165 * \param fmt - format of debug message
167 static void qmt_lqe_debug(struct lquota_entry *lqe, void *arg,
168 struct libcfs_debug_msg_data *msgdata,
169 struct va_format *vaf)
171 struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
173 libcfs_debug_msg(msgdata,
174 "%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",
175 vaf, pool->qpi_qmt->qmt_svname,
176 RES_NAME(pool->qpi_rtype),
178 lqe->lqe_id.qid_uid, lqe->lqe_enforced,
179 lqe->lqe_hardlimit, lqe->lqe_softlimit,
180 lqe->lqe_granted, lqe->lqe_gracetime,
181 lqe->lqe_qunit, lqe->lqe_edquot, lqe->lqe_may_rel,
182 lqe->lqe_revoke_time,
183 lqe->lqe_is_default ? "yes" : "no");
187 * Vector of quota entry operations supported on the master
189 const struct lquota_entry_operations qmt_lqe_ops = {
190 .lqe_init = qmt_lqe_init,
191 .lqe_read = qmt_lqe_read,
192 .lqe_debug = qmt_lqe_debug,
196 * Reserve enough credits to update records in both the global index and
197 * the slave index identified by \slv_obj
199 * \param env - is the environment passed by the caller
200 * \param lqe - is the quota entry associated with the identifier
201 * subject to the change. If it is NULL lqes array is
202 * taken from env with qti_lqes_env(env).
203 * \param slv_obj - is the dt_object associated with the index file
204 * \param sync - make transaction sync if true
206 struct thandle *qmt_trans_start_with_slv(const struct lu_env *env,
207 struct lquota_entry *lqe,
208 struct dt_object *slv_obj,
211 struct qmt_device *qmt;
213 struct lquota_entry **lqes;
214 struct qmt_lqe_restore *restore;
218 restore = qti_lqes_rstr(env);
220 lqes_cnt = qti_lqes_cnt(env);
221 lqes = qti_lqes(env);
227 /* qmt is the same for all lqes, so take it from the 1st */
228 qmt = lqe2qpi(lqes[0])->qpi_qmt;
231 LQUOTA_DEBUG(lqes[0], "declare write for slv "DFID,
232 PFID(lu_object_fid(&slv_obj->do_lu)));
234 /* start transaction */
235 th = dt_trans_create(env, qmt->qmt_child);
240 /* quota settings on master are updated synchronously for the
244 /* reserve credits for global index update */
245 for (i = 0; i < lqes_cnt; i++) {
246 rc = lquota_disk_declare_write(env, th,
247 LQE_GLB_OBJ(lqes[i]),
253 if (slv_obj != NULL) {
254 /* reserve credits for slave index update */
255 rc = lquota_disk_declare_write(env, th, slv_obj,
261 /* start transaction */
262 rc = dt_trans_start_local(env, qmt->qmt_child, th);
269 dt_trans_stop(env, qmt->qmt_child, th);
271 LQUOTA_ERROR(lqes[0], "failed to slv declare write for "DFID
272 ", rc:%d", PFID(lu_object_fid(&slv_obj->do_lu)),
275 for (i = 0; i < lqes_cnt; i++) {
276 restore[i].qlr_hardlimit = lqes[i]->lqe_hardlimit;
277 restore[i].qlr_softlimit = lqes[i]->lqe_softlimit;
278 restore[i].qlr_gracetime = lqes[i]->lqe_gracetime;
279 restore[i].qlr_granted = lqes[i]->lqe_granted;
280 restore[i].qlr_qunit = lqes[i]->lqe_qunit;
287 * Reserve enough credits to update a record in the global index
289 * \param env - is the environment passed by the caller
290 * \param lqe - is the quota entry to be modified in the global index
291 * \param restore - is a temporary storage for current quota settings which will
292 * be restored if something goes wrong at index update time.
294 struct thandle *qmt_trans_start(const struct lu_env *env,
295 struct lquota_entry *lqe)
297 LQUOTA_DEBUG(lqe, "declare write");
298 return qmt_trans_start_with_slv(env, lqe, NULL, true);
301 int qmt_glb_write_lqes(const struct lu_env *env, struct thandle *th,
302 __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(0, LQUOTA_FLAG_DEFAULT);
359 rec->qbr_hardlimit = lqe->lqe_hardlimit;
360 rec->qbr_softlimit = lqe->lqe_softlimit;
361 rec->qbr_time = lqe->lqe_gracetime;
364 /* write new quota settings */
365 rc = lquota_disk_write(env, th, LQE_GLB_OBJ(lqe), &lqe->lqe_id,
366 (struct dt_rec *)rec, flags, ver);
368 /* we failed to write the new quota settings to disk, report
369 * error to caller who will restore the initial value */
370 LQUOTA_ERROR(lqe, "failed to update global index, rc:%d", rc);
376 * Read from disk how much quota space is allocated to a slave.
377 * This is done by reading records from the dedicated slave index file.
378 * Return in \granted how much quota space is currently allocated to the
380 * The entry must be at least read locked.
382 * \param env - the environment passed by the caller
383 * \param lqe_id - is the quota id associated with the identifier to look-up
385 * \param slv_obj - is the dt_object associated with the slave index
386 * \param granted - is the output parameter where to return how much space
387 * is granted to the slave.
389 * \retval - 0 on success, appropriate error on failure
391 int qmt_slv_read(const struct lu_env *env, union lquota_id *qid,
392 struct dt_object *slv_obj, __u64 *granted)
394 struct qmt_thread_info *qti = qmt_info(env);
395 struct lquota_slv_rec *slv_rec = &qti->qti_slv_rec;
399 CDEBUG(D_QUOTA, "read id:%llu form slv "DFID"\n",
400 qid->qid_uid, PFID(lu_object_fid(&slv_obj->do_lu)));
402 /* read slave record from disk */
403 rc = lquota_disk_read(env, slv_obj, qid,
404 (struct dt_rec *)slv_rec);
410 /* extract granted from on-disk record */
411 *granted = slv_rec->qsr_granted;
414 CERROR("Failed to read slave record for %llu from "DFID"\n",
415 qid->qid_uid, PFID(lu_object_fid(&slv_obj->do_lu)));
419 CDEBUG(D_QUOTA, "Successful slv read %llu\n", *granted);
425 * Update record in slave index file.
426 * The entry must be at least read locked.
428 * \param env - the environment passed by the caller
429 * \param th - is the transaction handle to be used for the disk writes
430 * \param lqe - is the dirty quota entry which will be updated at the same time
432 * \param slv_obj - is the dt_object associated with the slave index
433 * \param flags - can be LQUOTA_BUMP_VER or LQUOTA_SET_VER.
434 * \param ver - is used to return the new version of the index.
435 * \param granted - is the new amount of quota space owned by the slave
437 * \retval - 0 on success, appropriate error on failure
439 int qmt_slv_write(const struct lu_env *env, struct thandle *th,
440 struct lquota_entry *lqe, struct dt_object *slv_obj,
441 __u32 flags, __u64 *ver, __u64 granted)
443 struct qmt_thread_info *qti = qmt_info(env);
444 struct lquota_slv_rec *rec;
448 LASSERT(lqe != NULL);
449 LASSERT(lqe_is_master(lqe));
450 LASSERT(lqe_is_locked(lqe));
452 LQUOTA_DEBUG(lqe, "write slv "DFID" granted:%llu",
453 PFID(lu_object_fid(&slv_obj->do_lu)), granted);
455 /* never delete the entry, otherwise, it'll not be transferred
456 * to slave during reintegration. */
457 rec = &qti->qti_slv_rec;
459 /* updated space granted to this slave */
460 rec->qsr_granted = granted;
462 /* write new granted space */
463 rc = lquota_disk_write(env, th, slv_obj, &lqe->lqe_id,
464 (struct dt_rec *)rec, flags, ver);
467 "failed to update slave index "DFID" granted:%llu",
468 PFID(lu_object_fid(&slv_obj->do_lu)),
477 * Check whether new limits are valid for this pool
479 * \param lqe - is the quota entry subject to the setquota
480 * \param hard - is the new hard limit
481 * \param soft - is the new soft limit
483 int qmt_validate_limits(struct lquota_entry *lqe, __u64 hard, __u64 soft)
487 if (hard != 0 && soft > hard)
488 /* soft limit must be less than hard limit */
494 * Set/clear edquot flag after quota space allocation/release or settings
495 * change. Slaves will be notified of changes via glimpse on per-ID lock
497 * \param lqe - is the quota entry to check
498 * \param now - is the current time in second used for grace time managment
500 bool qmt_adjust_edquot(struct lquota_entry *lqe, __u64 now)
502 struct qmt_pool_info *pool = lqe2qpi(lqe);
505 if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
508 if (!lqe->lqe_edquot) {
509 /* space exhausted flag not set, let's check whether it is time
512 if (!qmt_space_exhausted(lqe, now))
513 /* the qmt still has available space */
516 /* See comment in qmt_adjust_qunit(). LU-4139 */
517 if (qmt_hard_exhausted(lqe) ||
518 pool->qpi_rtype != LQUOTA_RES_DT) {
521 /* we haven't reached the minimal qunit yet so there is
522 * still hope that the rebalancing process might free
523 * up some quota space */
524 if (lqe->lqe_qunit != pool->qpi_least_qunit)
527 /* least qunit value not sent to all slaves yet */
528 if (lqe->lqe_revoke_time == 0)
531 /* Let's give more time to slave to release space */
532 lapse = ktime_get_seconds() - QMT_REBA_TIMEOUT;
533 if (lqe->lqe_may_rel != 0 && lqe->lqe_revoke_time > lapse)
536 if (lqe->lqe_qunit > pool->qpi_soft_least_qunit)
540 /* set edquot flag */
541 lqe->lqe_edquot = true;
543 /* space exhausted flag set, let's check whether it is time to
546 if (qmt_space_exhausted(lqe, now))
547 /* the qmt still has not space */
550 if (lqe->lqe_hardlimit != 0 &&
551 lqe->lqe_granted + pool->qpi_least_qunit >
553 /* we clear the flag only once at least one least qunit
557 /* clear edquot flag */
558 lqe->lqe_edquot = false;
561 LQUOTA_DEBUG(lqe, "changing edquot flag");
563 /* let's notify slave by issuing glimpse on per-ID lock.
564 * the rebalance thread will take care of this */
568 /* Using least_qunit when over block softlimit will seriously impact the
569 * write performance, we need to do some special tweaking on that. */
570 static __u64 qmt_calc_softlimit(struct lquota_entry *lqe, bool *oversoft)
572 struct qmt_pool_info *pool = lqe2qpi(lqe);
574 LASSERT(lqe->lqe_softlimit != 0);
576 /* No need to do special tweaking for inode limit */
577 if (pool->qpi_rtype != LQUOTA_RES_DT)
578 return lqe->lqe_softlimit;
580 if (lqe->lqe_granted <= lqe->lqe_softlimit +
581 pool->qpi_soft_least_qunit) {
582 return lqe->lqe_softlimit;
583 } else if (lqe->lqe_hardlimit != 0) {
585 return lqe->lqe_hardlimit;
593 * Try to grant more quota space back to slave.
595 * \param lqe - is the quota entry for which we would like to allocate more
597 * \param granted - is how much was already granted as part of the request
599 * \param spare - is how much unused quota space the slave already owns
601 * \retval return how additional space can be granted to the slave
603 __u64 qmt_alloc_expand(struct lquota_entry *lqe, __u64 granted, __u64 spare)
605 struct qmt_pool_info *pool = lqe2qpi(lqe);
606 __u64 remaining, qunit;
609 LASSERT(lqe->lqe_enforced && lqe->lqe_qunit != 0);
611 slv_cnt = qpi_slv_nr(lqe2qpi(lqe), lqe_qtype(lqe));
612 qunit = lqe->lqe_qunit;
614 /* See comment in qmt_adjust_qunit(). LU-4139. */
615 if (lqe->lqe_softlimit != 0) {
617 remaining = qmt_calc_softlimit(lqe, &oversoft);
619 remaining = lqe->lqe_granted +
620 pool->qpi_soft_least_qunit;
622 remaining = lqe->lqe_hardlimit;
625 if (lqe->lqe_granted >= remaining)
628 remaining -= lqe->lqe_granted;
634 granted &= (qunit - 1);
636 if (remaining > (slv_cnt * qunit) >> 1) {
637 /* enough room to grant more space w/o additional
638 * shrinking ... at least for now */
639 remaining -= (slv_cnt * qunit) >> 1;
640 } else if (qunit != pool->qpi_least_qunit) {
645 granted &= (qunit - 1);
647 RETURN(min_t(__u64, qunit - spare, remaining));
649 RETURN(min_t(__u64, qunit - granted, remaining));
650 } while (qunit >= pool->qpi_least_qunit);
656 qmt_adjust_qunit_set_revoke(const struct lu_env *env, struct lquota_entry *lqe,
657 unsigned long least_qunit)
659 struct lquota_entry *lqe2;
663 if (qti_lqes_cnt(env) <= 1)
666 for (i = 0; i < qti_lqes_cnt(env); i++) {
667 lqe2 = qti_lqes(env)[i];
668 if ((lqe2->lqe_qunit == least_qunit) && lqe2->lqe_revoke_time) {
670 min = lqe2->lqe_revoke_time;
673 min = lqe2->lqe_revoke_time < min ?
674 lqe2->lqe_revoke_time : min;
678 lqe->lqe_revoke_time = min;
683 * Adjust qunit size according to quota limits and total granted count.
684 * The caller must have locked the lqe.
686 * \param env - the environment passed by the caller
687 * \param lqe - is the qid entry to be adjusted
688 * \retval true - need reseed glbe array
690 bool qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
692 struct qmt_pool_info *pool = lqe2qpi(lqe);
693 bool need_reseed = false;
695 __u64 qunit, limit, qunit2 = 0;
698 LASSERT(lqe_is_locked(lqe));
700 if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
701 /* no quota limits */
704 /* record how many slaves have already registered */
705 slv_cnt = qpi_slv_nr(pool, lqe_qtype(lqe));
707 /* Pool hasn't slaves anymore. Qunit will be adjusted
708 * again when new slaves would be added. */
709 if (lqe->lqe_qunit) {
713 /* wait for at least one slave to join */
717 /* Qunit calculation is based on soft limit, if any, hard limit
718 * otherwise. This means that qunit is shrunk to the minimum when
719 * beyond the soft limit. This will impact performance, but that's the
720 * price of an accurate grace time management. */
721 if (lqe->lqe_softlimit != 0) {
723 /* As a compromise of write performance and the grace time
724 * accuracy, the block qunit size will be shrunk to
725 * qpi_soft_least_qunit when over softlimit. LU-4139. */
726 limit = qmt_calc_softlimit(lqe, &oversoft);
728 qunit2 = pool->qpi_soft_least_qunit;
730 GOTO(done, qunit = qunit2);
731 } else if (lqe->lqe_hardlimit != 0) {
732 limit = lqe->lqe_hardlimit;
734 LQUOTA_ERROR(lqe, "enforced bit set, but neither hard nor soft "
739 qunit = lqe->lqe_qunit == 0 ? pool->qpi_least_qunit : lqe->lqe_qunit;
741 /* The qunit value is computed as follows: limit / (2 * slv_cnt).
742 * Then 75% of the quota space can be granted with current qunit value.
743 * The remaining 25% are then used with reduced qunit size (by a factor
744 * of 4) which is then divided in a similar manner.
746 * |---------------------limit---------------------|
747 * |-------limit / 2-------|-limit / 4-|-limit / 4-|
748 * |qunit|qunit|qunit|qunit| | |
749 * |----slv_cnt * qunit----| | |
750 * |-grow limit-| | | |
751 * |--------------shrink limit---------| |
752 * |---space granted in qunit chunks---|-remaining-|
758 * qunit >>= 2; |qunit*slv_cnt|qunit*slv_cnt|
759 * |---space in qunit---|remain|
761 if (qunit == pool->qpi_least_qunit ||
762 limit >= lqe->lqe_granted + ((slv_cnt * qunit) >> 1)) {
763 /* current qunit value still fits, let's see if we can afford to
764 * increase qunit now ...
765 * To increase qunit again, we have to be under 25% */
766 while (qunit && limit >= lqe->lqe_granted + 6 * qunit * slv_cnt)
771 do_div(qunit, 2 * slv_cnt);
775 /* shrink qunit until we find a suitable value */
776 while (qunit > pool->qpi_least_qunit &&
777 limit < lqe->lqe_granted + ((slv_cnt * qunit) >> 1))
781 if (qunit2 && qunit > qunit2)
784 if (lqe->lqe_qunit == qunit)
785 /* keep current qunit */
788 LQUOTA_DEBUG(lqe, "%s qunit to %llu",
789 lqe->lqe_qunit < qunit ? "increasing" : "decreasing",
792 /* store new qunit value */
793 swap(lqe->lqe_qunit, qunit);
795 /* reseed glbe array and notify
796 * slave if qunit was shrinked */
798 /* reset revoke time */
799 lqe->lqe_revoke_time = 0;
801 if (lqe->lqe_qunit == pool->qpi_least_qunit) {
802 if (lqe->lqe_qunit >= qunit)
803 /* initial qunit value is the smallest one */
804 lqe->lqe_revoke_time = ktime_get_seconds();
805 /* If there are several lqes and lqe_revoke_time is set for
806 * some of them, it means appropriate OSTs have been already
807 * notified with the least qunit and there is no chance to
808 * free more space. Find an lqe with the minimum(earliest)
809 * revoke_time and set this time to the current one.
811 qmt_adjust_qunit_set_revoke(env, lqe, pool->qpi_least_qunit);
816 bool qmt_adjust_edquot_qunit_notify(const struct lu_env *env,
817 struct qmt_device *qmt,
818 __u64 now, bool edquot,
819 bool qunit, __u32 qb_flags)
821 struct lquota_entry *lqe_gl, *lqe;
822 bool need_reseed = false;
825 lqe_gl = qti_lqes_glbl(env);
827 for (i = 0; i < qti_lqes_cnt(env); i++) {
828 lqe = qti_lqes(env)[i];
830 need_reseed |= qmt_adjust_qunit(env, lqe);
832 need_reseed |= qmt_adjust_edquot(lqe, now);
836 if (!lqe_gl->lqe_glbl_data &&
837 (req_has_rep(qb_flags) || req_is_rel(qb_flags))) {
840 "%s: can not notify - lge_glbl_data is not set\n",
845 if (lqe_gl->lqe_glbl_data && need_reseed) {
846 qmt_seed_glbe_all(env, lqe_gl->lqe_glbl_data, qunit, edquot);
847 qmt_id_lock_notify(qmt, lqe_gl);
854 * Adjust qunit & edquot flag in case it wasn't initialized already (e.g.
855 * limit set while no slaves were connected yet)
857 bool qmt_revalidate(const struct lu_env *env, struct lquota_entry *lqe)
859 bool need_notify = false;
861 if (lqe->lqe_qunit == 0) {
862 /* lqe was read from disk, but neither qunit, nor edquot flag
863 * were initialized */
864 need_notify = qmt_adjust_qunit(env, lqe);
865 if (lqe->lqe_qunit != 0)
866 need_notify |= qmt_adjust_edquot(lqe,
867 ktime_get_real_seconds());
873 void qmt_revalidate_lqes(const struct lu_env *env,
874 struct qmt_device *qmt, __u32 qb_flags)
876 struct lquota_entry *lqe_gl = qti_lqes_glbl(env);
877 bool need_notify = false;
880 for (i = 0; i < qti_lqes_cnt(env); i++)
881 need_notify |= qmt_revalidate(env, qti_lqes(env)[i]);
883 /* There could be no ID lock to the moment of reconciliation.
884 * As a result lqe global data is not initialised yet. It is ok
885 * for release and report requests. */
886 if (!lqe_gl->lqe_glbl_data &&
887 (req_is_rel(qb_flags) || req_has_rep(qb_flags)))
891 qmt_seed_glbe(env, lqe_gl->lqe_glbl_data);
892 qmt_id_lock_notify(qmt, lqe_gl);
896 void qti_lqes_init(const struct lu_env *env)
898 struct qmt_thread_info *qti = qmt_info(env);
900 qti->qti_lqes_cnt = 0;
901 qti->qti_glbl_lqe_idx = 0;
902 qti->qti_lqes_num = QMT_MAX_POOL_NUM;
905 int qti_lqes_add(const struct lu_env *env, struct lquota_entry *lqe)
907 struct qmt_thread_info *qti = qmt_info(env);
909 if (qti->qti_lqes_cnt >= qti->qti_lqes_num) {
910 struct lquota_entry **lqes;
911 lqes = qti->qti_lqes;
912 OBD_ALLOC(lqes, sizeof(lqe) * qti->qti_lqes_num * 2);
915 memcpy(lqes, qti_lqes(env), qti->qti_lqes_cnt * sizeof(lqe));
916 /* Don't need to free, if it is the very 1st allocation */
917 if (qti->qti_lqes_num > QMT_MAX_POOL_NUM)
918 OBD_FREE(qti->qti_lqes,
919 qti->qti_lqes_num * sizeof(lqe));
920 qti->qti_lqes = lqes;
921 qti->qti_lqes_num *= 2;
924 if (lqe->lqe_is_global)
925 qti->qti_glbl_lqe_idx = qti->qti_lqes_cnt;
926 qti_lqes(env)[qti->qti_lqes_cnt++] = lqe;
928 /* The pool could be accessed directly from lqe, so take
929 * extra reference that is put in qti_lqes_fini */
930 qpi_getref(lqe2qpi(lqe));
932 CDEBUG(D_QUOTA, "LQE %p %lu is added, lqe_cnt %d lqes_num %d\n",
933 lqe, (long unsigned)lqe->lqe_id.qid_uid,
934 qti->qti_lqes_cnt, qti->qti_lqes_num);
935 LASSERT(qti->qti_lqes_num != 0);
940 void qti_lqes_del(const struct lu_env *env, int index)
942 struct lquota_entry **lqes;
943 int lqes_cnt = qti_lqes_cnt(env);
944 int lqep_size = sizeof(struct lquota_entry *);
947 /* We can't handle non global lqes correctly without
948 * global lqe located at index 0. If we try to do so,
949 * something goes wrong. */
950 LQUOTA_ERROR(qti_lqes_glbl(env),
951 "quota: cannot remove lqe at index 0 as it is global");
952 LASSERT(qti_lqes_glbl(env)->lqe_is_global);
955 lqes = qti_lqes(env);
956 qpi_putref(env, lqe2qpi(lqes[index]));
957 lqe_putref(lqes[index]);
958 memcpy((unsigned char *)lqes + index * lqep_size,
959 (unsigned char *)lqes + (index + 1) * lqep_size,
960 (lqes_cnt - index - 1) * lqep_size);
964 void qti_lqes_fini(const struct lu_env *env)
966 struct qmt_thread_info *qti = qmt_info(env);
967 struct lquota_entry **lqes = qti->qti_lqes;
970 lqes = qti_lqes(env);
971 for (i = 0; i < qti->qti_lqes_cnt; i++) {
972 qpi_putref(env, lqe2qpi(lqes[i]));
976 if (qti->qti_lqes_num > QMT_MAX_POOL_NUM)
977 OBD_FREE(qti->qti_lqes,
978 qti->qti_lqes_num * sizeof(struct lquota_entry *));
980 qti->qti_lqes_num = 0;
981 qti->qti_lqes_cnt = 0;
984 __u64 qti_lqes_min_qunit(const struct lu_env *env)
989 for (i = 1, min = qti_lqe_qunit(env, 0); i < qti_lqes_cnt(env); i++) {
990 qunit = qti_lqe_qunit(env, i);
991 /* if qunit is 0, lqe is not enforced and we can ignore it */
992 if (qunit && qunit < min)
999 int qti_lqes_edquot(const struct lu_env *env)
1003 for (i = 0; i < qti_lqes_cnt(env); i++) {
1004 if (qti_lqes(env)[i]->lqe_edquot)
1011 int qti_lqes_restore_init(const struct lu_env *env)
1015 if (qti_lqes_inited(env) && qti_lqes_cnt(env) > QMT_MAX_POOL_NUM) {
1016 OBD_ALLOC(qmt_info(env)->qti_lqes_rstr,
1017 qti_lqes_cnt(env) * sizeof(struct qmt_lqe_restore));
1018 if (!qti_lqes_rstr(env))
1025 void qti_lqes_restore_fini(const struct lu_env *env)
1027 if (qti_lqes_inited(env) && qti_lqes_cnt(env) > QMT_MAX_POOL_NUM)
1028 OBD_FREE(qmt_info(env)->qti_lqes_rstr,
1029 qti_lqes_cnt(env) * sizeof(struct qmt_lqe_restore));
1032 void qti_lqes_write_lock(const struct lu_env *env)
1036 for (i = 0; i < qti_lqes_cnt(env); i++)
1037 lqe_write_lock(qti_lqes(env)[i]);
1040 void qti_lqes_write_unlock(const struct lu_env *env)
1044 for (i = 0; i < qti_lqes_cnt(env); i++)
1045 lqe_write_unlock(qti_lqes(env)[i]);
1048 #define QMT_INIT_SLV_CNT 64
1049 struct lqe_glbl_data *qmt_alloc_lqe_gd(struct qmt_pool_info *pool, int qtype)
1051 struct lqe_glbl_data *lgd;
1052 struct lqe_glbl_entry *lqeg_arr;
1053 int slv_cnt, glbe_num;
1055 OBD_ALLOC(lgd, sizeof(struct lqe_glbl_data));
1059 slv_cnt = qpi_slv_nr_by_rtype(pool, qtype);
1061 glbe_num = slv_cnt < QMT_INIT_SLV_CNT ? QMT_INIT_SLV_CNT : slv_cnt;
1062 OBD_ALLOC(lqeg_arr, sizeof(struct lqe_glbl_entry) * glbe_num);
1064 OBD_FREE(lgd, sizeof(struct lqe_glbl_data));
1068 CDEBUG(D_QUOTA, "slv_cnt %d glbe_num %d\n", slv_cnt, glbe_num);
1070 lgd->lqeg_num_used = slv_cnt;
1071 lgd->lqeg_num_alloc = glbe_num;
1072 lgd->lqeg_arr = lqeg_arr;
1077 void qmt_free_lqe_gd(struct lqe_glbl_data *lgd)
1079 OBD_FREE(lgd->lqeg_arr,
1080 sizeof(struct lqe_glbl_entry) * lgd->lqeg_num_alloc);
1081 OBD_FREE(lgd, sizeof(struct lqe_glbl_data));
1084 void qmt_seed_glbe_all(const struct lu_env *env, struct lqe_glbl_data *lgd,
1085 bool qunit, bool edquot)
1087 struct rw_semaphore *sem = NULL;
1088 struct qmt_pool_info *qpi;
1092 if (!qti_lqes_cnt(env))
1094 /* lqes array is sorted by qunit - the first entry has minimum qunit.
1095 * Thus start seeding global qunit's array beginning from the 1st lqe
1096 * and appropriate pool. If pools overlapped, slaves from this
1097 * overlapping get minimum qunit value.
1098 * user1: pool1, pool2, pool_glbl;
1099 * pool1: OST1; user1_qunit = 10M;
1100 * pool2: OST0, OST1, OST2; user1_qunit = 30M;
1101 * pool_glbl: OST0, OST1, OST2, OST3; user1_qunit = 160M;
1102 * qunit array after seeding should be:
1103 * OST0: 30M; OST1: 10M; OST2: 30M; OST3: 160M; */
1105 /* edquot resetup algorythm works fine
1106 * with not sorted lqes */
1110 for (i = 0; i < lgd->lqeg_num_used; i++) {
1111 lgd->lqeg_arr[i].lge_qunit_set = 0;
1112 lgd->lqeg_arr[i].lge_qunit_nu = 0;
1113 lgd->lqeg_arr[i].lge_edquot_nu = 0;
1116 for (i = 0; i < qti_lqes_cnt(env); i++) {
1117 struct lquota_entry *lqe = qti_lqes(env)[i];
1120 CDEBUG(D_QUOTA, "lqes_cnt %d, i %d\n", qti_lqes_cnt(env), i);
1122 if (qmt_pool_global(qpi)) {
1123 slaves_cnt = qpi_slv_nr_by_rtype(lqe2qpi(lqe),
1126 sem = qmt_sarr_rwsem(qpi);
1128 slaves_cnt = qmt_sarr_count(qpi);
1131 for (j = 0; j < slaves_cnt; j++) {
1132 idx = qmt_sarr_get_idx(qpi, j);
1136 int lge_edquot, new_edquot, edquot_nu;
1138 lge_edquot = lgd->lqeg_arr[idx].lge_edquot;
1139 edquot_nu = lgd->lqeg_arr[idx].lge_edquot_nu;
1140 new_edquot = lqe->lqe_edquot;
1142 if (lge_edquot == new_edquot ||
1143 (edquot_nu && lge_edquot == 1))
1145 lgd->lqeg_arr[idx].lge_edquot = new_edquot;
1146 /* it is needed for the following case:
1147 * initial values for idx i -
1148 * lqe_edquot = 1, lqe_edquot_nu == 0;
1149 * 1: new_edquot == 0 ->
1150 * lqe_edquot = 0, lqe_edquot_nu = 1;
1151 * 2: new_edquot == 1 ->
1152 * lqe_edquot = 1, lqe_edquot_nu = 0;
1153 * At the 2nd iteration lge_edquot comes back
1154 * to 1, so no changes and we don't need
1155 * to notify slave. */
1156 lgd->lqeg_arr[idx].lge_edquot_nu = !edquot_nu;
1160 __u64 lge_qunit, new_qunit;
1163 "idx %d lge_qunit_set %d lge_qunit %llu new_qunit %llu\n",
1164 idx, lgd->lqeg_arr[idx].lge_qunit_set,
1165 lgd->lqeg_arr[idx].lge_qunit,
1167 /* lge for this idx is already set
1168 * on previous iteration */
1169 if (lgd->lqeg_arr[idx].lge_qunit_set)
1171 lge_qunit = lgd->lqeg_arr[idx].lge_qunit;
1172 new_qunit = lqe->lqe_qunit;
1173 /* qunit could be not set,
1174 * so use global lqe's qunit */
1178 if (lge_qunit != new_qunit)
1179 lgd->lqeg_arr[idx].lge_qunit =
1182 /* TODO: initially slaves notification was done
1183 * only for qunit shrinking. Should we always
1184 * notify slaves with new qunit ? */
1185 if (lge_qunit > new_qunit)
1186 lgd->lqeg_arr[idx].lge_qunit_nu = 1;
1187 lgd->lqeg_arr[idx].lge_qunit_set = 1;
1191 if (!qmt_pool_global(qpi))
1194 /* TODO: only for debug purposes - remove it later */
1195 for (i = 0; i < lgd->lqeg_num_used; i++)
1197 "lgd ost %d, qunit %lu nu %d; edquot %d nu %d\n",
1198 i, (long unsigned)lgd->lqeg_arr[i].lge_qunit,
1199 lgd->lqeg_arr[i].lge_qunit_nu,
1200 lgd->lqeg_arr[i].lge_edquot,
1201 lgd->lqeg_arr[i].lge_edquot_nu);
1206 void qmt_setup_lqe_gd(const struct lu_env *env, struct qmt_device *qmt,
1207 struct lquota_entry *lqe, struct lqe_glbl_data *lgd,
1214 qunit = lqe->lqe_qunit;
1215 edquot = lqe->lqe_edquot;
1217 /* Firstly set all elements in array with
1218 * qunit and edquot of global pool */
1219 for (i = 0; i < lgd->lqeg_num_used; i++) {
1220 lgd->lqeg_arr[i].lge_qunit = qunit;
1221 lgd->lqeg_arr[i].lge_edquot = edquot;
1222 /* It is the very first lvb setup - qunit and other flags
1223 * will be sent to slaves during qmt_lvbo_fill. */
1224 lgd->lqeg_arr[i].lge_qunit_nu = 0;
1225 lgd->lqeg_arr[i].lge_edquot_nu = 0;
1228 qmt_pool_lqes_lookup_spec(env, qmt, pool_type,
1229 lqe_qtype(lqe), &lqe->lqe_id);
1230 qmt_seed_glbe(env, lgd);
1232 lqe->lqe_glbl_data = lgd;
1233 qmt_id_lock_notify(qmt, lqe);