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"
35 static void qmt_work_lvbo_free(struct work_struct *work)
37 struct lqe_glbl_data *lgd;
38 struct lquota_entry *lqe;
40 lqe = container_of(work, struct lquota_entry, lqe_work);
41 mutex_lock(&lqe->lqe_glbl_data_lock);
42 lgd = lqe->lqe_glbl_data;
43 lqe->lqe_glbl_data = NULL;
44 mutex_unlock(&lqe->lqe_glbl_data_lock);
47 if (unlikely(lgd == NULL)) {
48 struct qmt_pool_info *pool;
50 pool = (struct qmt_pool_info *)lqe->lqe_site->lqs_parent;
51 CWARN("%s: lvbo for (id=%llx) not fully inited\n",
52 pool->qpi_qmt->qmt_svname,
56 /* release lqe reference */
61 * Initialize qmt-specific fields of quota entry.
63 * \param lqe - is the quota entry to initialize
64 * \param arg - is the pointer to the qmt_pool_info structure
66 static void qmt_lqe_init(struct lquota_entry *lqe, void *arg)
68 LASSERT(lqe_is_master(lqe));
70 lqe->lqe_revoke_time = 0;
71 init_rwsem(&lqe->lqe_sem);
72 mutex_init(&lqe->lqe_glbl_data_lock);
73 INIT_WORK(&lqe->lqe_work, qmt_work_lvbo_free);
76 /* Apply the default quota setting to the specified quota entry
78 * \param env - is the environment passed by the caller
79 * \param pool - is the quota pool of the quota entry
80 * \param lqe - is the lquota_entry object to apply default quota on
81 * \param create_record - if true, an global quota record will be created and
85 * \retval -ve : other appropriate errors
87 int qmt_lqe_set_default(const struct lu_env *env, struct qmt_pool_info *pool,
88 struct lquota_entry *lqe, bool create_record)
90 struct lquota_entry *lqe_def;
95 if (lqe->lqe_id.qid_uid == 0)
98 lqe_def = pool->qpi_grace_lqe[lqe_qtype(lqe)];
100 LQUOTA_DEBUG(lqe, "inherit default quota");
102 lqe->lqe_is_default = true;
103 lqe->lqe_hardlimit = lqe_def->lqe_hardlimit;
104 lqe->lqe_softlimit = lqe_def->lqe_softlimit;
107 lqe->lqe_uptodate = true;
108 rc = qmt_set_with_lqe(env, pool->qpi_qmt, lqe, 0, 0,
109 LQUOTA_GRACE_FLAG(0, LQUOTA_FLAG_DEFAULT),
110 QIF_TIMES, true, false);
113 LQUOTA_ERROR(lqe, "failed to create the global quota"
117 if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0)
118 lqe->lqe_enforced = false;
120 lqe->lqe_enforced = true;
126 * Update a lquota entry. This is done by reading quota settings from the global
127 * index. The lquota entry must be write locked.
129 * \param env - the environment passed by the caller
130 * \param lqe - is the quota entry to refresh
131 * \param arg - is the pointer to the qmt_pool_info structure
132 * \param find - don't create lqe on disk in case of ENOENT if true
134 static int qmt_lqe_read(const struct lu_env *env, struct lquota_entry *lqe,
135 void *arg, bool find)
137 struct qmt_thread_info *qti = qmt_info(env);
138 struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
142 LASSERT(lqe_is_master(lqe));
144 if (qmt_pool_global(pool))
145 lqe->lqe_is_global = 1;
147 /* read record from disk */
148 rc = lquota_disk_read(env, pool->qpi_glb_obj[lqe->lqe_site->lqs_qtype],
149 &lqe->lqe_id, (struct dt_rec *)&qti->qti_glb_rec);
155 qmt_lqe_set_default(env, pool, lqe, true);
158 /* copy quota settings from on-disk record */
159 lqe->lqe_granted = qti->qti_glb_rec.qbr_granted;
160 lqe->lqe_hardlimit = qti->qti_glb_rec.qbr_hardlimit;
161 lqe->lqe_softlimit = qti->qti_glb_rec.qbr_softlimit;
162 lqe->lqe_gracetime = LQUOTA_GRACE(qti->qti_glb_rec.qbr_time);
164 if (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0 &&
165 (LQUOTA_FLAG(qti->qti_glb_rec.qbr_time) &
166 LQUOTA_FLAG_DEFAULT))
167 qmt_lqe_set_default(env, pool, lqe, false);
168 else if (LQUOTA_FLAG(qti->qti_glb_rec.qbr_time) &
170 lqe->lqe_is_reset = true;
173 LQUOTA_ERROR(lqe, "failed to read quota entry from disk, rc:%d",
178 if (lqe->lqe_id.qid_uid == 0 ||
179 (lqe->lqe_hardlimit == 0 && lqe->lqe_softlimit == 0))
180 /* {hard,soft}limit=0 means no quota enforced */
181 lqe->lqe_enforced = false;
183 lqe->lqe_enforced = true;
185 LQUOTA_DEBUG(lqe, "read");
190 * Print lqe information for debugging.
192 * \param lqe - is the quota entry to debug
193 * \param arg - is the pointer to the qmt_pool_info structure
194 * \param msgdata - debug message
195 * \param fmt - format of debug message
197 static void qmt_lqe_debug(struct lquota_entry *lqe, void *arg,
198 struct libcfs_debug_msg_data *msgdata,
199 struct va_format *vaf)
201 struct qmt_pool_info *pool = (struct qmt_pool_info *)arg;
203 libcfs_debug_msg(msgdata,
204 "%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",
205 vaf, pool->qpi_qmt->qmt_svname,
206 RES_NAME(pool->qpi_rtype),
208 lqe->lqe_id.qid_uid, lqe->lqe_enforced,
209 lqe->lqe_hardlimit, lqe->lqe_softlimit,
210 lqe->lqe_granted, lqe->lqe_gracetime,
211 lqe->lqe_qunit, lqe->lqe_edquot, lqe->lqe_may_rel,
212 lqe->lqe_revoke_time,
213 lqe->lqe_is_default ? "yes" : "no");
217 * Vector of quota entry operations supported on the master
219 const struct lquota_entry_operations qmt_lqe_ops = {
220 .lqe_init = qmt_lqe_init,
221 .lqe_read = qmt_lqe_read,
222 .lqe_debug = qmt_lqe_debug,
226 * Reserve enough credits to update records in both the global index and
227 * the slave index identified by \slv_obj
229 * \param env - is the environment passed by the caller
230 * \param lqe - is the quota entry associated with the identifier
231 * subject to the change. If it is NULL lqes array is
232 * taken from env with qti_lqes_env(env).
233 * \param slv_obj - is the dt_object associated with the index file
234 * \param sync - make transaction sync if true
236 struct thandle *qmt_trans_start_with_slv(const struct lu_env *env,
237 struct lquota_entry *lqe,
238 struct dt_object *slv_obj,
241 struct qmt_device *qmt;
243 struct lquota_entry **lqes;
244 struct qmt_lqe_restore *restore;
248 restore = qti_lqes_rstr(env);
250 lqes_cnt = qti_lqes_cnt(env);
251 lqes = qti_lqes(env);
257 /* qmt is the same for all lqes, so take it from the 1st */
258 qmt = lqe2qpi(lqes[0])->qpi_qmt;
261 LQUOTA_DEBUG(lqes[0], "declare write for slv "DFID,
262 PFID(lu_object_fid(&slv_obj->do_lu)));
264 /* start transaction */
265 th = dt_trans_create(env, qmt->qmt_child);
270 /* quota settings on master are updated synchronously for the
274 /* reserve credits for global index update */
275 for (i = 0; i < lqes_cnt; i++) {
276 rc = lquota_disk_declare_write(env, th,
277 LQE_GLB_OBJ(lqes[i]),
283 if (slv_obj != NULL) {
284 /* reserve credits for slave index update */
285 rc = lquota_disk_declare_write(env, th, slv_obj,
291 /* start transaction */
292 rc = dt_trans_start_local(env, qmt->qmt_child, th);
299 dt_trans_stop(env, qmt->qmt_child, th);
301 LQUOTA_ERROR(lqes[0], "failed to slv declare write for "DFID
302 ", rc:%d", PFID(lu_object_fid(&slv_obj->do_lu)),
305 for (i = 0; i < lqes_cnt; i++) {
306 restore[i].qlr_hardlimit = lqes[i]->lqe_hardlimit;
307 restore[i].qlr_softlimit = lqes[i]->lqe_softlimit;
308 restore[i].qlr_gracetime = lqes[i]->lqe_gracetime;
309 restore[i].qlr_granted = lqes[i]->lqe_granted;
310 restore[i].qlr_qunit = lqes[i]->lqe_qunit;
317 * Reserve enough credits to update a record in the global index
319 * \param env - is the environment passed by the caller
320 * \param lqe - is the quota entry to be modified in the global index
321 * \param restore - is a temporary storage for current quota settings which will
322 * be restored if something goes wrong at index update time.
324 struct thandle *qmt_trans_start(const struct lu_env *env,
325 struct lquota_entry *lqe)
327 LQUOTA_DEBUG(lqe, "declare write");
328 return qmt_trans_start_with_slv(env, lqe, NULL, true);
331 int qmt_glb_write_lqes(const struct lu_env *env, struct thandle *th,
332 __u32 flags, __u64 *ver)
336 for (i = 0; i < qti_lqes_cnt(env); i++) {
337 rc = qmt_glb_write(env, th, qti_lqes(env)[i], flags, ver);
345 * Update record associated with a quota entry in the global index.
346 * If LQUOTA_BUMP_VER is set, then the global index version must also be
348 * The entry must be at least read locked, dirty and up-to-date.
350 * \param env - the environment passed by the caller
351 * \param th - is the transaction handle to be used for the disk writes
352 * \param lqe - is the quota entry to udpate
353 * \param obj - is the dt_object associated with the index file
354 * \param flags - can be LQUOTA_BUMP_VER or LQUOTA_SET_VER.
355 * \param ver - is used to return the new version of the index.
357 * \retval - 0 on success and lqe dirty flag cleared,
358 * appropriate error on failure and uptodate flag cleared.
360 int qmt_glb_write(const struct lu_env *env, struct thandle *th,
361 struct lquota_entry *lqe, __u32 flags, __u64 *ver)
363 struct qmt_thread_info *qti = qmt_info(env);
364 struct lquota_glb_rec *rec;
368 LASSERT(lqe != NULL);
369 LASSERT(lqe_is_master(lqe));
370 LASSERT(lqe_is_locked(lqe));
371 LASSERT(lqe->lqe_uptodate);
372 LASSERT((flags & ~(LQUOTA_BUMP_VER | LQUOTA_SET_VER)) == 0);
374 LQUOTA_DEBUG(lqe, "write glb");
376 /* never delete the entry even when the id isn't enforced and
377 * no any guota granted, otherwise, this entry will not be
378 * synced to slave during the reintegration. */
379 rec = &qti->qti_glb_rec;
381 /* fill global index with updated quota settings */
382 rec->qbr_granted = lqe->lqe_granted;
383 if (lqe->lqe_is_default) {
384 rec->qbr_hardlimit = 0;
385 rec->qbr_softlimit = 0;
386 rec->qbr_time = LQUOTA_GRACE_FLAG(lqe->lqe_gracetime,
387 LQUOTA_FLAG_DEFAULT);
388 } else if (lqe->lqe_is_reset) {
389 rec->qbr_hardlimit = 0;
390 rec->qbr_softlimit = 0;
391 rec->qbr_granted = 0;
392 rec->qbr_time = LQUOTA_GRACE_FLAG(lqe->lqe_gracetime,
395 rec->qbr_hardlimit = lqe->lqe_hardlimit;
396 rec->qbr_softlimit = lqe->lqe_softlimit;
397 rec->qbr_time = lqe->lqe_gracetime;
400 /* write new quota settings */
401 rc = lquota_disk_write(env, th, LQE_GLB_OBJ(lqe), &lqe->lqe_id,
402 (struct dt_rec *)rec, flags, ver);
404 /* we failed to write the new quota settings to disk, report
405 * error to caller who will restore the initial value */
406 LQUOTA_ERROR(lqe, "failed to update global index, rc:%d", rc);
412 * Read from disk how much quota space is allocated to a slave.
413 * This is done by reading records from the dedicated slave index file.
414 * Return in \granted how much quota space is currently allocated to the
416 * The entry must be at least read locked.
418 * \param env - the environment passed by the caller
419 * \param lqe_id - is the quota id associated with the identifier to look-up
421 * \param slv_obj - is the dt_object associated with the slave index
422 * \param granted - is the output parameter where to return how much space
423 * is granted to the slave.
425 * \retval - 0 on success, appropriate error on failure
427 int qmt_slv_read(const struct lu_env *env, union lquota_id *qid,
428 struct dt_object *slv_obj, __u64 *granted)
430 struct qmt_thread_info *qti = qmt_info(env);
431 struct lquota_slv_rec *slv_rec = &qti->qti_slv_rec;
435 CDEBUG(D_QUOTA, "read id:%llu form slv "DFID"\n",
436 qid->qid_uid, PFID(lu_object_fid(&slv_obj->do_lu)));
438 /* read slave record from disk */
439 rc = lquota_disk_read(env, slv_obj, qid,
440 (struct dt_rec *)slv_rec);
446 /* extract granted from on-disk record */
447 *granted = slv_rec->qsr_granted;
450 CERROR("Failed to read slave record for %llu from "DFID"\n",
451 qid->qid_uid, PFID(lu_object_fid(&slv_obj->do_lu)));
455 CDEBUG(D_QUOTA, "Successful slv read %llu\n", *granted);
461 * Update record in slave index file.
462 * The entry must be at least read locked.
464 * \param env - the environment passed by the caller
465 * \param th - is the transaction handle to be used for the disk writes
466 * \param lqe - is the dirty quota entry which will be updated at the same time
468 * \param slv_obj - is the dt_object associated with the slave index
469 * \param flags - can be LQUOTA_BUMP_VER or LQUOTA_SET_VER.
470 * \param ver - is used to return the new version of the index.
471 * \param granted - is the new amount of quota space owned by the slave
473 * \retval - 0 on success, appropriate error on failure
475 int qmt_slv_write(const struct lu_env *env, struct thandle *th,
476 struct lquota_entry *lqe, struct dt_object *slv_obj,
477 __u32 flags, __u64 *ver, __u64 granted)
479 struct qmt_thread_info *qti = qmt_info(env);
480 struct lquota_slv_rec *rec;
484 LASSERT(lqe != NULL);
485 LASSERT(lqe_is_master(lqe));
486 LASSERT(lqe_is_locked(lqe));
488 LQUOTA_DEBUG(lqe, "write slv "DFID" granted:%llu",
489 PFID(lu_object_fid(&slv_obj->do_lu)), granted);
491 /* never delete the entry, otherwise, it'll not be transferred
492 * to slave during reintegration. */
493 rec = &qti->qti_slv_rec;
495 /* updated space granted to this slave */
496 rec->qsr_granted = granted;
498 /* write new granted space */
499 rc = lquota_disk_write(env, th, slv_obj, &lqe->lqe_id,
500 (struct dt_rec *)rec, flags, ver);
503 "failed to update slave index "DFID" granted:%llu",
504 PFID(lu_object_fid(&slv_obj->do_lu)),
513 * Check whether new limits are valid for this pool
515 * \param lqe - is the quota entry subject to the setquota
516 * \param hard - is the new hard limit
517 * \param soft - is the new soft limit
519 int qmt_validate_limits(struct lquota_entry *lqe, __u64 hard, __u64 soft)
523 if (hard != 0 && soft > hard)
524 /* soft limit must be less than hard limit */
530 * Set/clear edquot flag after quota space allocation/release or settings
531 * change. Slaves will be notified of changes via glimpse on per-ID lock
533 * \param lqe - is the quota entry to check
534 * \param now - is the current time in second used for grace time managment
536 bool qmt_adjust_edquot(struct lquota_entry *lqe, __u64 now)
538 struct qmt_pool_info *pool = lqe2qpi(lqe);
541 if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
544 if (!lqe->lqe_edquot) {
545 /* space exhausted flag not set, let's check whether it is time
548 if (!qmt_space_exhausted(lqe, now))
549 /* the qmt still has available space */
552 /* See comment in qmt_adjust_qunit(). LU-4139 */
553 if (qmt_hard_exhausted(lqe) ||
554 pool->qpi_rtype != LQUOTA_RES_DT) {
557 /* we haven't reached the minimal qunit yet so there is
558 * still hope that the rebalancing process might free
559 * up some quota space */
560 if (lqe->lqe_qunit != pool->qpi_least_qunit)
563 /* least qunit value not sent to all slaves yet */
564 if (lqe->lqe_revoke_time == 0)
567 /* Let's give more time to slave to release space */
568 lapse = ktime_get_seconds() - QMT_REBA_TIMEOUT;
569 if (lqe->lqe_may_rel != 0 && lqe->lqe_revoke_time > lapse)
572 if (lqe->lqe_qunit > pool->qpi_soft_least_qunit)
576 /* set edquot flag */
577 lqe->lqe_edquot = true;
579 /* space exhausted flag set, let's check whether it is time to
582 if (qmt_space_exhausted(lqe, now))
583 /* the qmt still has not space */
586 if (lqe->lqe_hardlimit != 0 &&
587 lqe->lqe_granted + pool->qpi_least_qunit >
589 /* we clear the flag only once at least one least qunit
593 /* clear edquot flag */
594 lqe->lqe_edquot = false;
597 LQUOTA_DEBUG(lqe, "changing edquot flag");
599 /* let's notify slave by issuing glimpse on per-ID lock.
600 * the rebalance thread will take care of this */
604 /* Using least_qunit when over block softlimit will seriously impact the
605 * write performance, we need to do some special tweaking on that. */
606 static __u64 qmt_calc_softlimit(struct lquota_entry *lqe, bool *oversoft)
608 struct qmt_pool_info *pool = lqe2qpi(lqe);
610 LASSERT(lqe->lqe_softlimit != 0);
612 /* No need to do special tweaking for inode limit */
613 if (pool->qpi_rtype != LQUOTA_RES_DT)
614 return lqe->lqe_softlimit;
616 if (lqe->lqe_granted <= lqe->lqe_softlimit +
617 pool->qpi_soft_least_qunit) {
618 return lqe->lqe_softlimit;
619 } else if (lqe->lqe_hardlimit != 0) {
621 return lqe->lqe_hardlimit;
629 * Try to grant more quota space back to slave.
631 * \param lqe - is the quota entry for which we would like to allocate more
633 * \param granted - is how much was already granted as part of the request
635 * \param spare - is how much unused quota space the slave already owns
637 * \retval return how additional space can be granted to the slave
639 __u64 qmt_alloc_expand(struct lquota_entry *lqe, __u64 granted, __u64 spare)
641 struct qmt_pool_info *pool = lqe2qpi(lqe);
642 __u64 remaining, qunit;
645 LASSERT(lqe->lqe_enforced && lqe->lqe_qunit != 0);
647 slv_cnt = qpi_slv_nr(lqe2qpi(lqe), lqe_qtype(lqe));
648 qunit = lqe->lqe_qunit;
650 /* See comment in qmt_adjust_qunit(). LU-4139. */
651 if (lqe->lqe_softlimit != 0) {
653 remaining = qmt_calc_softlimit(lqe, &oversoft);
655 remaining = lqe->lqe_granted +
656 pool->qpi_soft_least_qunit;
658 remaining = lqe->lqe_hardlimit;
661 if (lqe->lqe_granted >= remaining)
664 remaining -= lqe->lqe_granted;
670 granted &= (qunit - 1);
672 if (remaining > (slv_cnt * qunit) >> 1) {
673 /* enough room to grant more space w/o additional
674 * shrinking ... at least for now */
675 remaining -= (slv_cnt * qunit) >> 1;
676 } else if (qunit != pool->qpi_least_qunit) {
681 granted &= (qunit - 1);
683 RETURN(min_t(__u64, qunit - spare, remaining));
685 RETURN(min_t(__u64, qunit - granted, remaining));
686 } while (qunit >= pool->qpi_least_qunit);
692 qmt_adjust_qunit_set_revoke(const struct lu_env *env, struct lquota_entry *lqe,
693 unsigned long least_qunit)
695 struct lquota_entry *lqe2;
699 if (qti_lqes_cnt(env) <= 1)
702 for (i = 0; i < qti_lqes_cnt(env); i++) {
703 lqe2 = qti_lqes(env)[i];
704 if ((lqe2->lqe_qunit == least_qunit) && lqe2->lqe_revoke_time) {
706 min = lqe2->lqe_revoke_time;
709 min = lqe2->lqe_revoke_time < min ?
710 lqe2->lqe_revoke_time : min;
714 lqe->lqe_revoke_time = min;
719 * Adjust qunit size according to quota limits and total granted count.
720 * The caller must have locked the lqe.
722 * \param env - the environment passed by the caller
723 * \param lqe - is the qid entry to be adjusted
724 * \retval true - need reseed glbe array
726 bool qmt_adjust_qunit(const struct lu_env *env, struct lquota_entry *lqe)
728 struct qmt_pool_info *pool = lqe2qpi(lqe);
729 bool need_reseed = false;
731 __u64 qunit, limit, qunit2 = 0;
734 LASSERT(lqe_is_locked(lqe));
736 if (!lqe->lqe_enforced || lqe->lqe_id.qid_uid == 0)
737 /* no quota limits */
740 /* record how many slaves have already registered */
741 slv_cnt = qpi_slv_nr(pool, lqe_qtype(lqe));
743 /* Pool hasn't slaves anymore. Qunit will be adjusted
744 * again when new slaves would be added. */
745 if (lqe->lqe_qunit) {
749 /* wait for at least one slave to join */
753 /* Qunit calculation is based on soft limit, if any, hard limit
754 * otherwise. This means that qunit is shrunk to the minimum when
755 * beyond the soft limit. This will impact performance, but that's the
756 * price of an accurate grace time management. */
757 if (lqe->lqe_softlimit != 0) {
759 /* As a compromise of write performance and the grace time
760 * accuracy, the block qunit size will be shrunk to
761 * qpi_soft_least_qunit when over softlimit. LU-4139. */
762 limit = qmt_calc_softlimit(lqe, &oversoft);
764 qunit2 = pool->qpi_soft_least_qunit;
766 GOTO(done, qunit = qunit2);
767 } else if (lqe->lqe_hardlimit != 0) {
768 limit = lqe->lqe_hardlimit;
770 LQUOTA_ERROR(lqe, "enforced bit set, but neither hard nor soft "
775 qunit = lqe->lqe_qunit == 0 ? pool->qpi_least_qunit : lqe->lqe_qunit;
777 /* The qunit value is computed as follows: limit / (2 * slv_cnt).
778 * Then 75% of the quota space can be granted with current qunit value.
779 * The remaining 25% are then used with reduced qunit size (by a factor
780 * of 4) which is then divided in a similar manner.
782 * |---------------------limit---------------------|
783 * |-------limit / 2-------|-limit / 4-|-limit / 4-|
784 * |qunit|qunit|qunit|qunit| | |
785 * |----slv_cnt * qunit----| | |
786 * |-grow limit-| | | |
787 * |--------------shrink limit---------| |
788 * |---space granted in qunit chunks---|-remaining-|
794 * qunit >>= 2; |qunit*slv_cnt|qunit*slv_cnt|
795 * |---space in qunit---|remain|
797 if (qunit == pool->qpi_least_qunit ||
798 limit >= lqe->lqe_granted + ((slv_cnt * qunit) >> 1)) {
799 /* current qunit value still fits, let's see if we can afford to
800 * increase qunit now ...
801 * To increase qunit again, we have to be under 25% */
802 while (qunit && limit >= lqe->lqe_granted + 6 * qunit * slv_cnt)
807 do_div(qunit, 2 * slv_cnt);
811 /* shrink qunit until we find a suitable value */
812 while (qunit > pool->qpi_least_qunit &&
813 limit < lqe->lqe_granted + ((slv_cnt * qunit) >> 1))
817 if (qunit2 && qunit > qunit2)
820 if (lqe->lqe_qunit == qunit)
821 /* keep current qunit */
824 LQUOTA_DEBUG(lqe, "%s qunit to %llu",
825 lqe->lqe_qunit < qunit ? "increasing" : "decreasing",
828 /* store new qunit value */
829 swap(lqe->lqe_qunit, qunit);
831 /* reseed glbe array and notify
832 * slave if qunit was shrinked */
834 /* reset revoke time */
835 lqe->lqe_revoke_time = 0;
837 if (lqe->lqe_qunit == pool->qpi_least_qunit) {
838 if (lqe->lqe_qunit >= qunit)
839 /* initial qunit value is the smallest one */
840 lqe->lqe_revoke_time = ktime_get_seconds();
841 /* If there are several lqes and lqe_revoke_time is set for
842 * some of them, it means appropriate OSTs have been already
843 * notified with the least qunit and there is no chance to
844 * free more space. Find an lqe with the minimum(earliest)
845 * revoke_time and set this time to the current one.
847 qmt_adjust_qunit_set_revoke(env, lqe, pool->qpi_least_qunit);
852 bool qmt_adjust_edquot_qunit_notify(const struct lu_env *env,
853 struct qmt_device *qmt,
854 __u64 now, bool edquot,
855 bool qunit, __u32 qb_flags,
858 struct lquota_entry *lqe_gl, *lqe;
863 lqe_gl = qti_lqes_glbl(env);
865 for (i = 0; i < qti_lqes_cnt(env); i++) {
866 lqe = qti_lqes(env)[i];
868 reseed |= qmt_adjust_qunit(env, lqe);
870 reseed |= qmt_adjust_edquot(lqe, now);
874 if (!lqe_gl->lqe_glbl_data &&
875 (req_has_rep(qb_flags) || req_is_rel(qb_flags))) {
878 "%s: can not notify - lge_glbl_data is not set\n",
883 if (reseed || idx >= 0) {
884 mutex_lock(&lqe_gl->lqe_glbl_data_lock);
885 if (lqe_gl->lqe_glbl_data) {
886 struct lqe_glbl_data *lgd = lqe_gl->lqe_glbl_data;
889 qmt_seed_glbe_all(env, lgd, qunit, edquot,
891 } else if (idx >= 0) {
892 int lge_idx = qmt_map_lge_idx(lgd, idx);
894 /* If there are no locks yet when
895 * lge_qunit/edquot_nu is set, slaves
896 * are still not notified with new
897 * qunit/edquot value. In a such case
898 * we need to notify them with new values to
899 * avoid endless EINPROGRESS if qunit is equal
900 * to the least qunit, but lqe_revoke_time is
903 notify = lgd->lqeg_arr[lge_idx].lge_qunit_nu ||
904 lgd->lqeg_arr[lge_idx].lge_edquot_nu;
907 mutex_unlock(&lqe_gl->lqe_glbl_data_lock);
910 if (reseed || notify)
911 qmt_id_lock_notify(qmt, lqe_gl);
918 * Adjust qunit & edquot flag in case it wasn't initialized already (e.g.
919 * limit set while no slaves were connected yet)
921 bool qmt_revalidate(const struct lu_env *env, struct lquota_entry *lqe)
923 bool need_notify = false;
925 if (lqe->lqe_qunit == 0) {
926 /* lqe was read from disk, but neither qunit, nor edquot flag
927 * were initialized */
928 need_notify = qmt_adjust_qunit(env, lqe);
929 if (lqe->lqe_qunit != 0)
930 need_notify |= qmt_adjust_edquot(lqe,
931 ktime_get_real_seconds());
937 void qmt_revalidate_lqes(const struct lu_env *env,
938 struct qmt_device *qmt, __u32 qb_flags)
940 struct lquota_entry *lqe_gl = qti_lqes_glbl(env);
941 bool need_notify = false;
944 for (i = 0; i < qti_lqes_cnt(env); i++)
945 need_notify |= qmt_revalidate(env, qti_lqes(env)[i]);
950 /* There could be no ID lock to the moment of reconciliation.
951 * As a result lqe global data is not initialised yet. It is ok
952 * for release and report requests. */
953 if (!lqe_gl->lqe_glbl_data &&
954 (req_is_rel(qb_flags) || req_has_rep(qb_flags))) {
958 mutex_lock(&lqe_gl->lqe_glbl_data_lock);
959 if (lqe_gl->lqe_glbl_data)
960 qmt_seed_glbe(env, lqe_gl->lqe_glbl_data, false);
961 mutex_unlock(&lqe_gl->lqe_glbl_data_lock);
963 qmt_id_lock_notify(qmt, lqe_gl);
966 void qti_lqes_init(const struct lu_env *env)
968 struct qmt_thread_info *qti = qmt_info(env);
970 qti->qti_lqes_cnt = 0;
971 qti->qti_glbl_lqe_idx = 0;
972 qti->qti_lqes_num = QMT_MAX_POOL_NUM;
975 int qti_lqes_add(const struct lu_env *env, struct lquota_entry *lqe)
977 struct qmt_thread_info *qti = qmt_info(env);
979 if (qti->qti_lqes_cnt >= qti->qti_lqes_num) {
980 struct lquota_entry **lqes;
981 lqes = qti->qti_lqes;
982 OBD_ALLOC(lqes, sizeof(lqe) * qti->qti_lqes_num * 2);
985 memcpy(lqes, qti_lqes(env), qti->qti_lqes_cnt * sizeof(lqe));
986 /* Don't need to free, if it is the very 1st allocation */
987 if (qti->qti_lqes_num > QMT_MAX_POOL_NUM)
988 OBD_FREE(qti->qti_lqes,
989 qti->qti_lqes_num * sizeof(lqe));
990 qti->qti_lqes = lqes;
991 qti->qti_lqes_num *= 2;
994 if (lqe->lqe_is_global)
995 qti->qti_glbl_lqe_idx = qti->qti_lqes_cnt;
996 qti_lqes(env)[qti->qti_lqes_cnt++] = lqe;
998 /* The pool could be accessed directly from lqe, so take
999 * extra reference that is put in qti_lqes_fini */
1000 qpi_getref(lqe2qpi(lqe));
1002 CDEBUG(D_QUOTA, "LQE %px %lu is added, lqe_cnt %d lqes_num %d\n",
1003 lqe, (long unsigned)lqe->lqe_id.qid_uid,
1004 qti->qti_lqes_cnt, qti->qti_lqes_num);
1005 LASSERT(qti->qti_lqes_num != 0);
1010 void qti_lqes_del(const struct lu_env *env, int index)
1012 struct lquota_entry **lqes;
1013 int lqes_cnt = qti_lqes_cnt(env);
1014 int lqep_size = sizeof(struct lquota_entry *);
1017 /* We can't handle non global lqes correctly without
1018 * global lqe located at index 0. If we try to do so,
1019 * something goes wrong. */
1020 LQUOTA_ERROR(qti_lqes_glbl(env),
1021 "quota: cannot remove lqe at index 0 as it is global");
1022 LASSERT(qti_lqes_glbl(env)->lqe_is_global);
1025 lqes = qti_lqes(env);
1026 qpi_putref(env, lqe2qpi(lqes[index]));
1027 lqe_putref(lqes[index]);
1028 memcpy((unsigned char *)lqes + index * lqep_size,
1029 (unsigned char *)lqes + (index + 1) * lqep_size,
1030 (lqes_cnt - index - 1) * lqep_size);
1031 qti_lqes_cnt(env)--;
1034 void qti_lqes_fini(const struct lu_env *env)
1036 struct qmt_thread_info *qti = qmt_info(env);
1037 struct lquota_entry **lqes = qti->qti_lqes;
1040 lqes = qti_lqes(env);
1041 for (i = 0; i < qti->qti_lqes_cnt; i++) {
1042 qpi_putref(env, lqe2qpi(lqes[i]));
1043 lqe_putref(lqes[i]);
1046 if (qti->qti_lqes_num > QMT_MAX_POOL_NUM)
1047 OBD_FREE(qti->qti_lqes,
1048 qti->qti_lqes_num * sizeof(struct lquota_entry *));
1050 qti->qti_lqes_num = 0;
1051 qti->qti_lqes_cnt = 0;
1054 __u64 qti_lqes_min_qunit(const struct lu_env *env)
1059 for (i = 1, min = qti_lqe_qunit(env, 0); i < qti_lqes_cnt(env); i++) {
1060 qunit = qti_lqe_qunit(env, i);
1061 /* if qunit is 0, lqe is not enforced and we can ignore it */
1062 if (qunit && qunit < min)
1069 int qti_lqes_edquot(const struct lu_env *env)
1073 for (i = 0; i < qti_lqes_cnt(env); i++) {
1074 if (qti_lqes(env)[i]->lqe_edquot)
1081 int qti_lqes_restore_init(const struct lu_env *env)
1085 if (qti_lqes_inited(env) && qti_lqes_cnt(env) > QMT_MAX_POOL_NUM) {
1086 OBD_ALLOC(qmt_info(env)->qti_lqes_rstr,
1087 qti_lqes_cnt(env) * sizeof(struct qmt_lqe_restore));
1088 if (!qmt_info(env)->qti_lqes_rstr)
1095 void qti_lqes_restore_fini(const struct lu_env *env)
1097 if (qti_lqes_inited(env) && qti_lqes_cnt(env) > QMT_MAX_POOL_NUM)
1098 OBD_FREE(qmt_info(env)->qti_lqes_rstr,
1099 qti_lqes_cnt(env) * sizeof(struct qmt_lqe_restore));
1102 void qti_lqes_write_lock(const struct lu_env *env)
1106 for (i = 0; i < qti_lqes_cnt(env); i++)
1107 lqe_write_lock(qti_lqes(env)[i]);
1110 void qti_lqes_write_unlock(const struct lu_env *env)
1114 for (i = 0; i < qti_lqes_cnt(env); i++)
1115 lqe_write_unlock(qti_lqes(env)[i]);
1118 #define QMT_INIT_SLV_CNT 64
1119 struct lqe_glbl_data *qmt_alloc_lqe_gd(struct qmt_pool_info *pool, int qtype)
1121 struct lqe_glbl_data *lgd;
1122 struct lqe_glbl_entry *lqeg_arr;
1123 int slv_cnt, glbe_num;
1125 OBD_ALLOC(lgd, sizeof(struct lqe_glbl_data));
1129 slv_cnt = qpi_slv_nr_by_rtype(pool, qtype);
1131 glbe_num = slv_cnt < QMT_INIT_SLV_CNT ? QMT_INIT_SLV_CNT : slv_cnt;
1132 OBD_ALLOC(lqeg_arr, sizeof(struct lqe_glbl_entry) * glbe_num);
1134 OBD_FREE(lgd, sizeof(struct lqe_glbl_data));
1138 CDEBUG(D_QUOTA, "slv_cnt %d glbe_num %d\n", slv_cnt, glbe_num);
1140 lgd->lqeg_num_used = slv_cnt;
1141 lgd->lqeg_num_alloc = glbe_num;
1142 lgd->lqeg_arr = lqeg_arr;
1147 void qmt_free_lqe_gd(struct lqe_glbl_data *lgd)
1152 OBD_FREE(lgd->lqeg_arr,
1153 sizeof(struct lqe_glbl_entry) * lgd->lqeg_num_alloc);
1154 OBD_FREE(lgd, sizeof(struct lqe_glbl_data));
1157 int qmt_map_lge_idx(struct lqe_glbl_data *lgd, int ostidx)
1161 /* check common case of sequential OST numbers first */
1162 if (ostidx < lgd->lqeg_num_used &&
1163 lgd->lqeg_arr[ostidx].lge_idx == ostidx)
1166 for (k = 0; k < lgd->lqeg_num_used; k++)
1167 if (lgd->lqeg_arr[k].lge_idx == ostidx)
1170 LASSERTF(k < lgd->lqeg_num_used, "Cannot map ostidx %d for %p\n",
1175 void qmt_seed_glbe_all(const struct lu_env *env, struct lqe_glbl_data *lgd,
1176 bool qunit, bool edquot, bool pool_locked)
1178 struct qmt_pool_info *qpi;
1182 if (!qti_lqes_cnt(env))
1184 /* lqes array is sorted by qunit - the first entry has minimum qunit.
1185 * Thus start seeding global qunit's array beginning from the 1st lqe
1186 * and appropriate pool. If pools overlapped, slaves from this
1187 * overlapping get minimum qunit value.
1188 * user1: pool1, pool2, pool_glbl;
1189 * pool1: OST1; user1_qunit = 10M;
1190 * pool2: OST0, OST1, OST2; user1_qunit = 30M;
1191 * pool_glbl: OST0, OST1, OST2, OST3; user1_qunit = 160M;
1192 * qunit array after seeding should be:
1193 * OST0: 30M; OST1: 10M; OST2: 30M; OST3: 160M; */
1195 /* edquot resetup algorythm works fine
1196 * with not sorted lqes */
1200 for (i = 0; i < lgd->lqeg_num_used; i++) {
1201 lgd->lqeg_arr[i].lge_qunit_set = 0;
1202 lgd->lqeg_arr[i].lge_qunit_nu = 0;
1203 lgd->lqeg_arr[i].lge_edquot_nu = 0;
1206 for (i = 0; i < qti_lqes_cnt(env); i++) {
1207 struct lquota_entry *lqe = qti_lqes(env)[i];
1210 CDEBUG(D_QUOTA, "lqes_cnt %d, i %d\n", qti_lqes_cnt(env), i);
1213 qmt_sarr_read_down(qpi);
1215 slaves_cnt = qmt_sarr_count(qpi);
1217 for (j = 0; j < slaves_cnt; j++) {
1220 tgt_idx = qmt_sarr_get_idx(qpi, j);
1221 LASSERT(tgt_idx >= 0);
1222 idx = qmt_map_lge_idx(lgd, tgt_idx);
1225 int lge_edquot, new_edquot, edquot_nu;
1227 lge_edquot = lgd->lqeg_arr[idx].lge_edquot;
1228 edquot_nu = lgd->lqeg_arr[idx].lge_edquot_nu;
1229 new_edquot = lqe->lqe_edquot;
1231 if (lge_edquot == new_edquot ||
1232 (edquot_nu && lge_edquot == 1))
1234 lgd->lqeg_arr[idx].lge_edquot = new_edquot;
1235 /* it is needed for the following case:
1236 * initial values for idx i -
1237 * lqe_edquot = 1, lqe_edquot_nu == 0;
1238 * 1: new_edquot == 0 ->
1239 * lqe_edquot = 0, lqe_edquot_nu = 1;
1240 * 2: new_edquot == 1 ->
1241 * lqe_edquot = 1, lqe_edquot_nu = 0;
1242 * At the 2nd iteration lge_edquot comes back
1243 * to 1, so no changes and we don't need
1244 * to notify slave. */
1245 lgd->lqeg_arr[idx].lge_edquot_nu = !edquot_nu;
1249 __u64 lge_qunit, new_qunit;
1252 "tgt_idx %d idx %d lge_qunit_set %d lge_qunit %llu new_qunit %llu\n",
1254 lgd->lqeg_arr[idx].lge_qunit_set,
1255 lgd->lqeg_arr[idx].lge_qunit,
1257 /* lge for this idx is already set
1258 * on previous iteration */
1259 if (lgd->lqeg_arr[idx].lge_qunit_set)
1261 lge_qunit = lgd->lqeg_arr[idx].lge_qunit;
1262 new_qunit = lqe->lqe_qunit;
1263 /* qunit could be not set,
1264 * so use global lqe's qunit */
1268 if (lge_qunit != new_qunit)
1269 lgd->lqeg_arr[idx].lge_qunit =
1272 /* TODO: initially slaves notification was done
1273 * only for qunit shrinking. Should we always
1274 * notify slaves with new qunit ? */
1275 if (lge_qunit > new_qunit)
1276 lgd->lqeg_arr[idx].lge_qunit_nu = 1;
1277 lgd->lqeg_arr[idx].lge_qunit_set = 1;
1282 qmt_sarr_read_up(qpi);
1284 /* TODO: only for debug purposes - remove it later */
1285 for (i = 0; i < lgd->lqeg_num_used; i++)
1287 "lgd i %d tgt_idx %d qunit %lu nu %d; edquot %d nu %d\n",
1288 i, lgd->lqeg_arr[i].lge_idx,
1289 (unsigned long)lgd->lqeg_arr[i].lge_qunit,
1290 lgd->lqeg_arr[i].lge_qunit_nu,
1291 lgd->lqeg_arr[i].lge_edquot,
1292 lgd->lqeg_arr[i].lge_edquot_nu);
1297 void qmt_setup_lqe_gd(const struct lu_env *env, struct qmt_device *qmt,
1298 struct lquota_entry *lqe, struct lqe_glbl_data *lgd,
1305 qunit = lqe->lqe_qunit;
1306 edquot = lqe->lqe_edquot;
1308 /* Firstly set all elements in array with
1309 * qunit and edquot of global pool */
1310 qmt_sarr_read_down(lqe2qpi(lqe));
1311 for (i = 0; i < lgd->lqeg_num_used; i++) {
1312 lgd->lqeg_arr[i].lge_qunit = qunit;
1313 lgd->lqeg_arr[i].lge_edquot = edquot;
1314 /* It is the very first lvb setup - qunit and other flags
1315 * will be sent to slaves during qmt_lvbo_fill. */
1316 lgd->lqeg_arr[i].lge_qunit_nu = 0;
1317 lgd->lqeg_arr[i].lge_edquot_nu = 0;
1318 lgd->lqeg_arr[i].lge_idx = qmt_sarr_get_idx(lqe2qpi(lqe), i);
1320 qmt_sarr_read_up(lqe2qpi(lqe));
1322 qmt_pool_lqes_lookup_spec(env, qmt, pool_type,
1323 lqe_qtype(lqe), &lqe->lqe_id);
1324 qmt_seed_glbe(env, lgd, false);
1326 mutex_lock(&lqe->lqe_glbl_data_lock);
1327 if (lqe->lqe_glbl_data == NULL) {
1328 lqe->lqe_glbl_data = lgd;
1331 mutex_unlock(&lqe->lqe_glbl_data_lock);
1333 qmt_free_lqe_gd(lgd);
1335 qmt_id_lock_notify(qmt, lqe);