1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/quota/quota_context.c
38 * Lustre Quota Context
40 * Author: Niu YaWei <niu@clusterfs.com>
44 # define EXPORT_SYMTAB
47 #define DEBUG_SUBSYSTEM S_LQUOTA
49 #include <linux/version.h>
51 #include <asm/unistd.h>
52 #include <linux/slab.h>
53 #include <linux/quotaops.h>
54 #include <linux/module.h>
55 #include <linux/init.h>
57 #include <obd_class.h>
58 #include <lustre_quota.h>
59 #include <lustre_fsfilt.h>
60 #include <lprocfs_status.h>
61 #include "quota_internal.h"
63 static int hash_lqs_cur_bits = HASH_LQS_CUR_BITS;
64 CFS_MODULE_PARM(hash_lqs_cur_bits, "i", int, 0444,
65 "the current bits of lqs hash");
67 #ifdef HAVE_QUOTA_SUPPORT
69 static cfs_hash_ops_t lqs_hash_ops;
71 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
72 unsigned long default_btune_ratio = 50; /* 50 percentage */
73 unsigned long default_iunit_sz = 5120; /* 5120 inodes */
74 unsigned long default_itune_ratio = 50; /* 50 percentage */
76 cfs_mem_cache_t *qunit_cachep = NULL;
77 cfs_list_t qunit_hash[NR_DQHASH];
78 cfs_spinlock_t qunit_hash_lock = CFS_SPIN_LOCK_UNLOCKED;
80 /* please sync qunit_state with qunit_state_names */
87 * a qunit is added into qunit hash, that means
88 * a quota req will be sent or is flying
92 * a qunit is removed from qunit hash, that
93 * means a quota req is handled and comes back
95 QUNIT_RM_FROM_HASH = 2,
97 * qunit can wake up all threads waiting for it
102 static const char *qunit_state_names[] = {
103 [QUNIT_CREATED] = "CREATED",
104 [QUNIT_IN_HASH] = "IN_HASH",
105 [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
106 [QUNIT_FINISHED] = "FINISHED",
109 struct lustre_qunit {
110 cfs_list_t lq_hash; /** Hash list in memory */
111 cfs_atomic_t lq_refcnt; /** Use count */
112 struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
113 struct qunit_data lq_data; /** See qunit_data */
114 unsigned int lq_opc; /** QUOTA_DQACQ, QUOTA_DQREL */
115 cfs_waitq_t lq_waitq; /** Threads waiting for this qunit */
116 cfs_spinlock_t lq_lock; /** Protect the whole structure */
117 enum qunit_state lq_state; /** Present the status of qunit */
118 int lq_rc; /** The rc of lq_data */
122 #define QUNIT_SET_STATE(qunit, state) \
124 cfs_spin_lock(&qunit->lq_lock); \
125 QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), " \
126 "lq_rc(%d), lq_owner(%d)\n", \
127 qunit, qunit_state_names[qunit->lq_state], \
128 qunit_state_names[state], qunit->lq_rc, \
130 qunit->lq_state = state; \
131 cfs_spin_unlock(&qunit->lq_lock); \
134 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc) \
136 cfs_spin_lock(&qunit->lq_lock); \
138 QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), " \
139 "lq_rc(%d), lq_owner(%d)\n", \
140 qunit, qunit_state_names[qunit->lq_state], \
141 qunit_state_names[state], qunit->lq_rc, \
143 qunit->lq_state = state; \
144 cfs_spin_unlock(&qunit->lq_lock); \
147 int should_translate_quota (struct obd_import *imp)
152 if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
158 void qunit_cache_cleanup(void)
163 cfs_spin_lock(&qunit_hash_lock);
164 for (i = 0; i < NR_DQHASH; i++)
165 LASSERT(cfs_list_empty(qunit_hash + i));
166 cfs_spin_unlock(&qunit_hash_lock);
170 rc = cfs_mem_cache_destroy(qunit_cachep);
171 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
177 int qunit_cache_init(void)
182 LASSERT(qunit_cachep == NULL);
183 qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
184 sizeof(struct lustre_qunit),
189 cfs_spin_lock(&qunit_hash_lock);
190 for (i = 0; i < NR_DQHASH; i++)
191 CFS_INIT_LIST_HEAD(qunit_hash + i);
192 cfs_spin_unlock(&qunit_hash_lock);
197 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
198 __attribute__((__const__));
201 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
203 unsigned int id = qdata->qd_id;
204 unsigned int type = QDATA_IS_GRP(qdata);
206 unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
207 tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
211 /* caller must hold qunit_hash_lock */
212 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
213 struct lustre_quota_ctxt *qctxt,
214 struct qunit_data *qdata)
216 struct lustre_qunit *qunit = NULL;
217 struct qunit_data *tmp;
219 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
220 cfs_list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
221 tmp = &qunit->lq_data;
222 if (qunit->lq_ctxt == qctxt &&
223 qdata->qd_id == tmp->qd_id &&
224 (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
225 (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
231 /* check_cur_qunit - check the current usage of qunit.
232 * @qctxt: quota context
233 * @qdata: the type of quota unit to be checked
235 * return: 1 - need acquire qunit;
236 * 2 - need release qunit;
237 * 0 - need do nothing.
241 check_cur_qunit(struct obd_device *obd,
242 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
244 struct super_block *sb = qctxt->lqc_sb;
245 unsigned long qunit_sz, tune_sz;
246 __u64 usage, limit, limit_org, pending_write = 0;
247 long long record = 0;
248 struct obd_quotactl *qctl;
249 struct lustre_qunit_size *lqs = NULL;
253 if (!ll_sb_any_quota_active(sb))
256 cfs_spin_lock(&qctxt->lqc_lock);
257 if (!qctxt->lqc_valid){
258 cfs_spin_unlock(&qctxt->lqc_lock);
261 cfs_spin_unlock(&qctxt->lqc_lock);
267 /* get fs quota usage & limit */
268 qctl->qc_cmd = Q_GETQUOTA;
269 qctl->qc_id = qdata->qd_id;
270 qctl->qc_type = QDATA_IS_GRP(qdata);
271 ret = fsfilt_quotactl(obd, sb, qctl);
273 if (ret == -ESRCH) /* no limit */
276 CERROR("can't get fs quota usage! (rc:%d)\n", ret);
280 if (QDATA_IS_BLK(qdata)) {
281 usage = qctl->qc_dqblk.dqb_curspace;
282 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
284 usage = qctl->qc_dqblk.dqb_curinodes;
285 limit = qctl->qc_dqblk.dqb_ihardlimit;
288 /* ignore the no quota limit case; and it can avoid creating
289 * unnecessary lqs for uid/gid */
293 lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
295 if (IS_ERR(lqs) || lqs == NULL) {
296 CERROR("fail to find a lqs for %sid: %u)!\n",
297 QDATA_IS_GRP(qdata) ? "g" : "u", qdata->qd_id);
300 cfs_spin_lock(&lqs->lqs_lock);
302 if (QDATA_IS_BLK(qdata)) {
303 qunit_sz = lqs->lqs_bunit_sz;
304 tune_sz = lqs->lqs_btune_sz;
305 pending_write = lqs->lqs_bwrite_pending;
306 record = lqs->lqs_blk_rec;
307 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
309 /* we didn't need change inode qunit size now */
310 qunit_sz = lqs->lqs_iunit_sz;
311 tune_sz = lqs->lqs_itune_sz;
312 pending_write = lqs->lqs_iwrite_pending;
313 record = lqs->lqs_ino_rec;
316 /* we don't count the MIN_QLIMIT */
317 if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
318 (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
321 usage += pending_write;
323 /* when a releasing quota req is sent, before it returned
324 limit is assigned a small value. limit will overflow */
325 if (limit + record < 0)
330 LASSERT(qdata->qd_count == 0);
331 if (limit <= usage + tune_sz) {
332 while (qdata->qd_count + limit <=
334 qdata->qd_count += qunit_sz;
336 } else if (limit > usage + qunit_sz + tune_sz &&
337 limit_org > qdata->qd_count + qunit_sz) {
338 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
339 limit_org > qdata->qd_count + qunit_sz)
340 qdata->qd_count += qunit_sz;
342 /* if there are other pending writes for this uid/gid, releasing
343 * quota is put off until the last pending write b=16645 */
344 /* if there is an ongoing quota request, a releasing request is aborted.
345 * That ongoing quota request will call this function again when
346 * it returned b=18630 */
347 if (pending_write || record) {
348 CDEBUG(D_QUOTA, "delay quota release\n");
353 quota_compute_lqs(qdata, lqs, 1, (ret == 1) ? 1 : 0);
355 CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
356 ", pending_write: "LPU64", record: %lld"
357 ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
358 QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
359 record, qunit_sz, tune_sz, ret);
360 LASSERT(ret == 0 || qdata->qd_count);
362 cfs_spin_unlock(&lqs->lqs_lock);
372 * Compute the remaining quota for certain gid or uid b=11693
374 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
375 struct qunit_data *qdata, int isblk)
377 struct super_block *sb = qctxt->lqc_sb;
379 struct obd_quotactl *qctl;
380 int ret = QUOTA_RET_OK;
383 if (!ll_sb_any_quota_active(sb))
384 RETURN(QUOTA_RET_NOQUOTA);
386 /* ignore root user */
387 if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
388 RETURN(QUOTA_RET_NOLIMIT);
394 /* get fs quota usage & limit */
395 qctl->qc_cmd = Q_GETQUOTA;
396 qctl->qc_id = qdata->qd_id;
397 qctl->qc_type = QDATA_IS_GRP(qdata);
398 ret = fsfilt_quotactl(obd, sb, qctl);
400 if (ret == -ESRCH) /* no limit */
401 ret = QUOTA_RET_NOLIMIT;
403 CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
408 usage = isblk ? qctl->qc_dqblk.dqb_curspace :
409 qctl->qc_dqblk.dqb_curinodes;
410 limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
411 qctl->qc_dqblk.dqb_ihardlimit;
412 if (!limit){ /* no limit */
413 ret = QUOTA_RET_NOLIMIT;
418 qdata->qd_count = limit - usage;
427 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
428 struct qunit_data *qdata, int opc)
430 struct lustre_qunit *qunit = NULL;
433 OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
437 CFS_INIT_LIST_HEAD(&qunit->lq_hash);
438 cfs_waitq_init(&qunit->lq_waitq);
439 cfs_atomic_set(&qunit->lq_refcnt, 1);
440 qunit->lq_ctxt = qctxt;
441 memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
443 qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
444 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
445 qunit->lq_owner = cfs_curproc_pid();
449 static inline void free_qunit(struct lustre_qunit *qunit)
451 OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
454 static inline void qunit_get(struct lustre_qunit *qunit)
456 cfs_atomic_inc(&qunit->lq_refcnt);
459 static void qunit_put(struct lustre_qunit *qunit)
461 LASSERT(cfs_atomic_read(&qunit->lq_refcnt));
462 if (cfs_atomic_dec_and_test(&qunit->lq_refcnt))
466 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
467 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
468 struct qunit_data *qdata)
470 unsigned int hashent = qunit_hashfn(qctxt, qdata);
471 struct lustre_qunit *qunit;
474 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
475 qunit = find_qunit(hashent, qctxt, qdata);
482 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
486 LASSERT(cfs_list_empty(&qunit->lq_hash));
488 head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
489 cfs_list_add(&qunit->lq_hash, head);
490 QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
493 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
495 struct lustre_qunit_size *lqs;
497 lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
498 qunit->lq_data.qd_id),
500 if (lqs && !IS_ERR(lqs)) {
501 cfs_spin_lock(&lqs->lqs_lock);
502 if (qunit->lq_opc == QUOTA_DQACQ)
503 quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
504 if (qunit->lq_opc == QUOTA_DQREL)
505 quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
506 cfs_spin_unlock(&lqs->lqs_lock);
507 /* this is for quota_search_lqs */
509 /* this is for schedule_dqacq */
514 static void remove_qunit_nolock(struct lustre_qunit *qunit)
516 LASSERT(!cfs_list_empty(&qunit->lq_hash));
517 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
519 cfs_list_del_init(&qunit->lq_hash);
520 QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
524 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
525 struct obd_quotactl *oqctl, int isblk)
527 struct lustre_qunit *qunit, *find_qunit;
530 OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
532 CERROR("locating %sunit failed for %sid %u\n",
533 isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
535 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
536 oqctl->qc_type, isblk);
540 CFS_INIT_LIST_HEAD(&qunit->lq_hash);
541 qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
542 cfs_waitq_init(&qunit->lq_waitq);
543 cfs_atomic_set(&qunit->lq_refcnt, 1);
544 qunit->lq_ctxt = qctxt;
545 qunit->lq_data.qd_id = oqctl->qc_id;
546 qunit->lq_data.qd_flags = oqctl->qc_type;
548 QDATA_SET_BLK(&qunit->lq_data);
549 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
550 /* it means it is only an invalid qunit for barrier */
551 qunit->lq_opc = QUOTA_LAST_OPC;
554 cfs_spin_lock(&qunit_hash_lock);
555 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
557 cfs_spin_unlock(&qunit_hash_lock);
558 qunit_put(find_qunit);
559 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
560 oqctl->qc_type, isblk);
561 CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
566 insert_qunit_nolock(qctxt, qunit);
567 cfs_spin_unlock(&qunit_hash_lock);
571 void quota_unbarrier(void *handle)
573 struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
576 CERROR("handle is NULL\n");
580 LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
581 cfs_spin_lock(&qunit_hash_lock);
582 remove_qunit_nolock(qunit);
583 cfs_spin_unlock(&qunit_hash_lock);
584 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
585 cfs_waitq_signal(&qunit->lq_waitq);
589 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
590 (limit = count) : (limit += count)
593 static inline int is_master(struct lustre_quota_ctxt *qctxt)
595 return qctxt->lqc_handler ? 1 : 0;
599 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
600 struct qunit_data *qdata, int opc, int wait,
601 struct obd_trans_info *oti);
603 static inline void qdata_to_oqaq(struct qunit_data *qdata,
604 struct quota_adjust_qunit *oqaq)
609 oqaq->qaq_flags = qdata->qd_flags;
610 oqaq->qaq_id = qdata->qd_id;
611 if (QDATA_IS_ADJBLK(qdata))
612 oqaq->qaq_bunit_sz = qdata->qd_qunit;
613 if (QDATA_IS_ADJINO(qdata))
614 oqaq->qaq_iunit_sz = qdata->qd_qunit;
618 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
619 struct qunit_data *qdata, int rc, int opc)
621 struct lustre_qunit *qunit = NULL;
622 struct super_block *sb = qctxt->lqc_sb;
624 struct quota_adjust_qunit *oqaq = NULL;
629 QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
630 obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
632 /* do it only when a releasing quota req more than 5MB b=18491 */
633 if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
634 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
636 /* update local operational quota file */
638 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
639 struct obd_quotactl *qctl;
644 GOTO(out, err = -ENOMEM);
646 /* acq/rel qunit for specified uid/gid is serialized,
647 * so there is no race between get fs quota limit and
648 * set fs quota limit */
649 qctl->qc_cmd = Q_GETQUOTA;
650 qctl->qc_id = qdata->qd_id;
651 qctl->qc_type = QDATA_IS_GRP(qdata);
652 err = fsfilt_quotactl(obd, sb, qctl);
654 CERROR("error get quota fs limit! (rc:%d)\n", err);
658 if (QDATA_IS_BLK(qdata)) {
659 qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
660 hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
662 qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
663 hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
666 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
673 INC_QLIMIT(*hardlimit, count);
676 LASSERTF(count < *hardlimit,
677 "id(%u) flag(%u) type(%c) isblk(%c) "
678 "count("LPU64") qd_qunit("LPU64") "
679 "hardlimit("LPU64").\n",
680 qdata->qd_id, qdata->qd_flags,
681 QDATA_IS_GRP(qdata) ? 'g' : 'u',
682 QDATA_IS_BLK(qdata) ? 'b': 'i',
683 qdata->qd_count, qdata->qd_qunit, *hardlimit);
690 /* clear quota limit */
694 qctl->qc_cmd = Q_SETQUOTA;
695 err = fsfilt_quotactl(obd, sb, qctl);
697 CERROR("error set quota fs limit! (rc:%d)\n", err);
699 QDATA_DEBUG(qdata, "%s completion\n",
700 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
703 } else if (rc == -EDQUOT) {
704 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
705 } else if (rc == -EBUSY) {
706 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
708 CERROR("acquire qunit got error! (rc:%d)\n", rc);
711 /* remove the qunit from hash */
712 cfs_spin_lock(&qunit_hash_lock);
714 qunit = dqacq_in_flight(qctxt, qdata);
715 /* this qunit has been removed by qctxt_cleanup() */
717 cfs_spin_unlock(&qunit_hash_lock);
718 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
719 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
723 LASSERT(opc == qunit->lq_opc);
724 /* remove this qunit from lq_hash so that new processes cannot be added
725 * to qunit->lq_waiters */
726 remove_qunit_nolock(qunit);
727 cfs_spin_unlock(&qunit_hash_lock);
729 compute_lqs_after_removing_qunit(qunit);
732 rc = QUOTA_REQ_RETURNED;
733 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
734 /* wake up all waiters */
735 cfs_waitq_broadcast(&qunit->lq_waitq);
737 /* this is for dqacq_in_flight() */
739 if (rc < 0 && rc != -EDQUOT)
742 /* don't reschedule in such cases:
743 * - acq/rel failure and qunit isn't changed,
744 * but not for quota recovery.
745 * - local dqacq/dqrel.
746 * - local disk io failure.
750 GOTO(out1, err = -ENOMEM);
751 qdata_to_oqaq(qdata, oqaq);
752 /* adjust the qunit size in slaves */
753 rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
756 CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
757 GOTO(out1, err = rc1);
759 if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
762 if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
763 OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
766 /* reschedule another dqacq/dqrel if needed */
768 qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
769 rc1 = check_cur_qunit(obd, qctxt, qdata);
772 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
773 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
774 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
777 /* this is for alloc_qunit() */
782 struct dqacq_async_args {
783 struct lustre_quota_ctxt *aa_ctxt;
784 struct lustre_qunit *aa_qunit;
787 static int dqacq_interpret(const struct lu_env *env,
788 struct ptlrpc_request *req, void *data, int rc)
790 struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
791 struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
792 struct obd_device_target *obt = qctxt->lqc_obt;
793 struct lustre_qunit *qunit = aa->aa_qunit;
794 struct obd_device *obd = req->rq_import->imp_obd;
795 struct qunit_data *qdata = NULL;
799 LASSERT(req->rq_import);
801 cfs_down_read(&obt->obt_rwsem);
802 /* if a quota req timeouts or is dropped, we should update quota
803 * statistics which will be handled in dqacq_completion. And in
804 * this situation we should get qdata from request instead of
806 qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
810 DEBUG_REQ(D_ERROR, req,
811 "error unpacking qunit_data(rc: %ld)\n",
813 qdata = &qunit->lq_data;
816 QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
817 QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
819 if (qdata->qd_id != qunit->lq_data.qd_id ||
820 OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
821 CERROR("the returned qd_id isn't expected!"
822 "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
823 qunit->lq_data.qd_id);
824 qdata->qd_id = qunit->lq_data.qd_id;
827 if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
828 CERROR("the returned grp/usr isn't expected!"
829 "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
830 qunit->lq_data.qd_flags);
831 if (QDATA_IS_GRP(&qunit->lq_data))
832 QDATA_SET_GRP(qdata);
834 QDATA_CLR_GRP(qdata);
837 if (qdata->qd_count > qunit->lq_data.qd_count) {
838 CERROR("the returned qd_count isn't expected!"
839 "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
840 qunit->lq_data.qd_count);
844 if (unlikely(rc == -ESRCH))
845 CERROR("quota for %s has been enabled by master, but disabled "
846 "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
848 rc = dqacq_completion(obd, qctxt, qdata, rc,
849 lustre_msg_get_opc(req->rq_reqmsg));
851 cfs_up_read(&obt->obt_rwsem);
856 * check if quota master is online
858 int check_qm(struct lustre_quota_ctxt *qctxt)
863 cfs_spin_lock(&qctxt->lqc_lock);
864 /* quit waiting when mds is back or qctxt is cleaned up */
865 rc = qctxt->lqc_import || !qctxt->lqc_valid;
866 cfs_spin_unlock(&qctxt->lqc_lock);
871 /* wake up all waiting threads when lqc_import is NULL */
872 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
874 struct lustre_qunit *qunit, *tmp;
878 cfs_spin_lock(&qunit_hash_lock);
879 for (i = 0; i < NR_DQHASH; i++) {
880 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
882 if (qunit->lq_ctxt != qctxt)
885 /* Wake up all waiters. Do not change lq_state.
886 * The waiters will check lq_rc which is kept as 0
887 * if no others change it, then the waiters will return
888 * -EAGAIN to caller who can perform related quota
889 * acq/rel if necessary. */
890 cfs_waitq_broadcast(&qunit->lq_waitq);
893 cfs_spin_unlock(&qunit_hash_lock);
897 static int got_qunit(struct lustre_qunit *qunit, int is_master)
899 struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
903 cfs_spin_lock(&qunit->lq_lock);
904 switch (qunit->lq_state) {
906 case QUNIT_RM_FROM_HASH:
912 CERROR("invalid qunit state %d\n", qunit->lq_state);
914 cfs_spin_unlock(&qunit->lq_lock);
917 cfs_spin_lock(&qctxt->lqc_lock);
918 rc = !qctxt->lqc_valid;
920 rc |= !qctxt->lqc_import;
921 cfs_spin_unlock(&qctxt->lqc_lock);
928 revoke_lqs_rec(struct lustre_qunit_size *lqs, struct qunit_data *qdata, int opc)
930 /* revoke lqs_xxx_rec which is computed in check_cur_qunit
932 cfs_spin_lock(&lqs->lqs_lock);
933 quota_compute_lqs(qdata, lqs, 0, (opc == QUOTA_DQACQ) ? 1 : 0);
934 cfs_spin_unlock(&lqs->lqs_lock);
938 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
939 struct qunit_data *qdata, int opc, int wait,
940 struct obd_trans_info *oti)
942 struct lustre_qunit *qunit, *empty;
943 struct l_wait_info lwi = { 0 };
944 struct ptlrpc_request *req;
945 struct dqacq_async_args *aa;
946 struct obd_import *imp = NULL;
947 struct lustre_qunit_size *lqs = NULL;
948 struct timeval work_start;
949 struct timeval work_end;
954 LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
955 cfs_gettimeofday(&work_start);
957 lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
959 if (lqs == NULL || IS_ERR(lqs)) {
960 CERROR("Can't find the lustre qunit size!\n");
964 if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL) {
965 revoke_lqs_rec(lqs, qdata, opc);
966 /* this is for quota_search_lqs */
971 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_SD, 5);
973 cfs_spin_lock(&qunit_hash_lock);
974 qunit = dqacq_in_flight(qctxt, qdata);
976 cfs_spin_unlock(&qunit_hash_lock);
979 revoke_lqs_rec(lqs, qdata, opc);
980 /* this is for quota_search_lqs */
982 goto wait_completion;
986 insert_qunit_nolock(qctxt, qunit);
987 cfs_spin_unlock(&qunit_hash_lock);
989 /* From here, the quota request will be sent anyway.
990 * When this qdata request returned or is cancelled,
991 * lqs_putref will be called at that time */
993 /* this is for quota_search_lqs */
996 QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
997 obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
998 /* master is going to dqacq/dqrel from itself */
999 if (is_master(qctxt)) {
1001 QDATA_DEBUG(qdata, "local %s.\n",
1002 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1003 QDATA_SET_CHANGE_QS(qdata);
1004 rc = qctxt->lqc_handler(obd, qdata, opc);
1005 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
1006 /* this is for qunit_get() */
1009 cfs_gettimeofday(&work_end);
1010 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1011 if (opc == QUOTA_DQACQ)
1012 lprocfs_counter_add(qctxt->lqc_stats,
1013 wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1016 lprocfs_counter_add(qctxt->lqc_stats,
1017 wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1019 RETURN(rc ? rc : rc2);
1022 cfs_spin_lock(&qctxt->lqc_lock);
1023 if (!qctxt->lqc_import) {
1024 cfs_spin_unlock(&qctxt->lqc_lock);
1025 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
1027 cfs_spin_lock(&qunit_hash_lock);
1028 remove_qunit_nolock(qunit);
1029 cfs_spin_unlock(&qunit_hash_lock);
1031 compute_lqs_after_removing_qunit(qunit);
1033 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1034 cfs_waitq_broadcast(&qunit->lq_waitq);
1036 /* this is for qunit_get() */
1038 /* this for alloc_qunit() */
1040 cfs_spin_lock(&qctxt->lqc_lock);
1041 if (wait && !qctxt->lqc_import) {
1042 cfs_spin_unlock(&qctxt->lqc_lock);
1043 LASSERT(oti && oti->oti_thread);
1044 /* The recovery thread doesn't have watchdog
1045 * attached. LU-369 */
1046 if (oti->oti_thread->t_watchdog)
1047 lc_watchdog_disable(oti->oti_thread->\
1049 CDEBUG(D_QUOTA, "sleep for quota master\n");
1050 l_wait_event(qctxt->lqc_wait_for_qmaster,
1051 check_qm(qctxt), &lwi);
1052 CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1053 if (oti->oti_thread->t_watchdog)
1054 lc_watchdog_touch(oti->oti_thread->t_watchdog,
1055 CFS_GET_TIMEOUT(oti->oti_thread->t_svc));
1057 cfs_spin_unlock(&qctxt->lqc_lock);
1062 imp = class_import_get(qctxt->lqc_import);
1063 cfs_spin_unlock(&qctxt->lqc_lock);
1065 /* build dqacq/dqrel request */
1068 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1069 LUSTRE_MDS_VERSION, opc);
1070 class_import_put(imp);
1072 CERROR("Can't alloc request\n");
1073 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1074 /* this is for qunit_get() */
1079 ptlrpc_request_set_replen(req);
1080 req->rq_no_resend = req->rq_no_delay = 1;
1081 rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1083 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1084 ptlrpc_req_finished(req);
1085 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1086 /* this is for qunit_get() */
1091 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1092 aa = ptlrpc_req_async_args(req);
1093 aa->aa_ctxt = qctxt;
1094 aa->aa_qunit = qunit;
1096 req->rq_interpret_reply = dqacq_interpret;
1097 ptlrpcd_add_req(req, PSCOPE_OTHER);
1099 QDATA_DEBUG(qdata, "%s scheduled.\n",
1100 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1102 if (wait && qunit) {
1103 struct qunit_data *p = &qunit->lq_data;
1105 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1106 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1108 /* rc = -EAGAIN, it means the quota master isn't ready yet
1109 * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1110 * rc = -EDQUOT, it means out of quota
1111 * rc = -EBUSY, it means recovery is happening
1112 * other rc < 0, it means real errors, functions who call
1113 * schedule_dqacq should take care of this */
1114 cfs_spin_lock(&qunit->lq_lock);
1116 cfs_spin_unlock(&qunit->lq_lock);
1117 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1118 "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1119 qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1123 cfs_gettimeofday(&work_end);
1124 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1125 if (opc == QUOTA_DQACQ)
1126 lprocfs_counter_add(qctxt->lqc_stats,
1127 wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1130 lprocfs_counter_add(qctxt->lqc_stats,
1131 wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1138 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1139 const unsigned int id[], __u32 isblk, int wait,
1140 struct obd_trans_info *oti)
1142 int rc = 0, i = USRQUOTA;
1143 struct qunit_data qdata[MAXQUOTAS];
1146 if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1149 for (i = 0; i < MAXQUOTAS; i++) {
1150 qdata[i].qd_id = id[i];
1151 qdata[i].qd_flags = i;
1153 QDATA_SET_BLK(&qdata[i]);
1154 qdata[i].qd_count = 0;
1156 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1159 /* need acquire or release */
1160 opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1161 rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1165 } else if (wait == 1) {
1166 /* when wait equates 1, that means mds_quota_acquire
1167 * or filter_quota_acquire is calling it. */
1168 rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1178 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1179 unsigned short type, int isblk)
1181 struct lustre_qunit *qunit = NULL;
1182 struct qunit_data qdata;
1183 struct timeval work_start;
1184 struct timeval work_end;
1186 struct l_wait_info lwi = { 0 };
1190 cfs_gettimeofday(&work_start);
1192 qdata.qd_flags = type;
1194 QDATA_SET_BLK(&qdata);
1197 cfs_spin_lock(&qunit_hash_lock);
1198 qunit = dqacq_in_flight(qctxt, &qdata);
1199 cfs_spin_unlock(&qunit_hash_lock);
1202 struct qunit_data *p = &qunit->lq_data;
1204 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1205 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1207 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1208 "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1209 /* keep same as schedule_dqacq() b=17030 */
1210 cfs_spin_lock(&qunit->lq_lock);
1212 cfs_spin_unlock(&qunit->lq_lock);
1213 /* this is for dqacq_in_flight() */
1215 cfs_gettimeofday(&work_end);
1216 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1217 lprocfs_counter_add(qctxt->lqc_stats,
1218 isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1219 LQUOTA_WAIT_PENDING_INO_QUOTA,
1222 cfs_gettimeofday(&work_end);
1223 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1224 lprocfs_counter_add(qctxt->lqc_stats,
1225 isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1226 LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1234 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1236 struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1237 struct obd_device_target *obt = &obd->u.obt;
1238 struct super_block *sb = obt->obt_sb;
1244 rc = ptlrpcd_addref();
1248 cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1249 cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1250 cfs_atomic_set(&qctxt->lqc_lqs, 0);
1251 cfs_spin_lock_init(&qctxt->lqc_lock);
1252 cfs_spin_lock(&qctxt->lqc_lock);
1253 qctxt->lqc_handler = handler;
1255 qctxt->lqc_obt = obt;
1256 qctxt->lqc_import = NULL;
1257 qctxt->lqc_recovery = 0;
1258 qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1259 qctxt->lqc_valid = 1;
1260 qctxt->lqc_cqs_boundary_factor = 4;
1261 qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1262 qctxt->lqc_cqs_least_iunit = 2;
1263 qctxt->lqc_cqs_qs_factor = 2;
1264 qctxt->lqc_flags = 0;
1265 QUOTA_MASTER_UNREADY(qctxt);
1266 qctxt->lqc_bunit_sz = default_bunit_sz;
1267 qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1268 qctxt->lqc_iunit_sz = default_iunit_sz;
1269 qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1270 qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1271 * after the last shrinking */
1272 qctxt->lqc_sync_blk = 0;
1273 cfs_spin_unlock(&qctxt->lqc_lock);
1275 qctxt->lqc_lqs_hash = cfs_hash_create("LQS_HASH",
1278 min(hash_lqs_cur_bits,
1280 0, CFS_HASH_MIN_THETA,
1282 &lqs_hash_ops, CFS_HASH_DEFAULT);
1283 if (!qctxt->lqc_lqs_hash) {
1284 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1289 rc = lquota_proc_setup(obd, is_master(qctxt));
1291 CERROR("initialize proc for %s error!\n", obd->obd_name);
1297 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1302 rc = !cfs_atomic_read(&qctxt->lqc_lqs);
1307 int qctxt_del_lqs(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1308 cfs_hlist_node_t *hnode, void *data)
1310 /* remove from hash and -1 refcount */
1311 cfs_hash_bd_del_locked(hs, bd, hnode);
1315 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1317 struct lustre_qunit *qunit, *tmp;
1318 cfs_list_t tmp_list;
1319 struct l_wait_info lwi = { 0 };
1320 struct obd_device_target *obt = qctxt->lqc_obt;
1324 CFS_INIT_LIST_HEAD(&tmp_list);
1326 cfs_spin_lock(&qctxt->lqc_lock);
1327 qctxt->lqc_valid = 0;
1328 cfs_spin_unlock(&qctxt->lqc_lock);
1330 cfs_spin_lock(&qunit_hash_lock);
1331 for (i = 0; i < NR_DQHASH; i++) {
1332 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
1334 if (qunit->lq_ctxt != qctxt)
1336 remove_qunit_nolock(qunit);
1337 cfs_list_add(&qunit->lq_hash, &tmp_list);
1340 cfs_spin_unlock(&qunit_hash_lock);
1342 cfs_list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1343 cfs_list_del_init(&qunit->lq_hash);
1344 compute_lqs_after_removing_qunit(qunit);
1346 /* wake up all waiters */
1347 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1348 cfs_waitq_broadcast(&qunit->lq_waitq);
1352 /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1353 * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1354 while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1355 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1356 cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
1357 cfs_time_seconds(1));
1360 /* release refcount on lustre_qunit_size holding by lqs_hash */
1361 cfs_hash_for_each_safe(qctxt->lqc_lqs_hash, qctxt_del_lqs, NULL);
1363 l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1364 cfs_down_write(&obt->obt_rwsem);
1365 cfs_hash_putref(qctxt->lqc_lqs_hash);
1366 qctxt->lqc_lqs_hash = NULL;
1367 cfs_up_write(&obt->obt_rwsem);
1372 if (lquota_proc_cleanup(qctxt))
1373 CERROR("cleanup proc error!\n");
1379 struct qslave_recov_thread_data {
1380 struct obd_device *obd;
1381 struct lustre_quota_ctxt *qctxt;
1382 cfs_completion_t comp;
1385 /* FIXME only recovery block quota by now */
1386 static int qslave_recovery_main(void *arg)
1388 struct qslave_recov_thread_data *data = arg;
1389 struct obd_device *obd = data->obd;
1390 struct lustre_quota_ctxt *qctxt = data->qctxt;
1395 cfs_daemonize_ctxt("qslave_recovd");
1398 class_incref(obd, "qslave_recovd_filter", obd);
1400 cfs_complete(&data->comp);
1402 cfs_spin_lock(&qctxt->lqc_lock);
1403 if (qctxt->lqc_recovery) {
1404 cfs_spin_unlock(&qctxt->lqc_lock);
1405 class_decref(obd, "qslave_recovd_filter", obd);
1408 qctxt->lqc_recovery = 1;
1409 cfs_spin_unlock(&qctxt->lqc_lock);
1412 for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1413 struct qunit_data qdata;
1414 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1416 struct dquot_id *dqid, *tmp;
1419 LOCK_DQONOFF_MUTEX(dqopt);
1420 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1421 UNLOCK_DQONOFF_MUTEX(dqopt);
1425 LASSERT(dqopt->files[type] != NULL);
1426 CFS_INIT_LIST_HEAD(&id_list);
1427 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1428 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1430 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1432 UNLOCK_DQONOFF_MUTEX(dqopt);
1434 CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1436 cfs_list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1437 cfs_list_del_init(&dqid->di_link);
1438 /* skip slave recovery on itself */
1439 if (is_master(qctxt))
1441 if (rc && rc != -EBUSY)
1444 qdata.qd_id = dqid->di_id;
1445 qdata.qd_flags = type;
1446 QDATA_SET_BLK(&qdata);
1449 ret = check_cur_qunit(obd, qctxt, &qdata);
1452 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1453 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1461 if (rc && rc != -EBUSY)
1462 CERROR("qslave recovery failed! (id:%d type:%d "
1463 " rc:%d)\n", dqid->di_id, type, rc);
1469 cfs_spin_lock(&qctxt->lqc_lock);
1470 qctxt->lqc_recovery = 0;
1471 cfs_spin_unlock(&qctxt->lqc_lock);
1472 class_decref(obd, "qslave_recovd_filter", obd);
1477 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1479 struct qslave_recov_thread_data data;
1483 if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1488 cfs_init_completion(&data.comp);
1490 rc = cfs_create_thread(qslave_recovery_main, &data,
1493 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1496 cfs_wait_for_completion(&data.comp);
1501 inline int quota_is_on(struct lustre_quota_ctxt *qctxt,
1502 struct obd_quotactl *oqctl)
1504 return ((qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)) ==
1505 UGQUOTA2LQC(oqctl->qc_type));
1508 inline int quota_is_off(struct lustre_quota_ctxt *qctxt,
1509 struct obd_quotactl *oqctl)
1511 return !(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type));
1515 * When quotaon, build a lqs for every uid/gid who has been set limitation
1516 * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1517 * It will be released when qctxt_cleanup() is executed b=18574
1519 * Should be called with obt->obt_quotachecking held. b=20152
1521 void build_lqs(struct obd_device *obd)
1523 struct obd_device_target *obt = &obd->u.obt;
1524 struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1528 LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1529 CFS_INIT_LIST_HEAD(&id_list);
1530 for (i = 0; i < MAXQUOTAS; i++) {
1531 struct dquot_id *dqid, *tmp;
1533 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1536 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1537 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1540 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1544 CERROR("%s: failed to get %s qids!\n", obd->obd_name,
1545 i ? "group" : "user");
1549 cfs_list_for_each_entry_safe(dqid, tmp, &id_list,
1551 struct lustre_qunit_size *lqs;
1553 cfs_list_del_init(&dqid->di_link);
1554 lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1556 if (lqs && !IS_ERR(lqs)) {
1557 lqs->lqs_flags |= dqid->di_flag;
1560 CERROR("%s: failed to create a lqs for %sid %u"
1561 "\n", obd->obd_name, i ? "g" : "u",
1571 * lqs<->qctxt hash operations
1575 * string hashing using djb2 hash algorithm
1578 lqs_hash(cfs_hash_t *hs, const void *key, unsigned mask)
1580 struct quota_adjust_qunit *lqs_key;
1585 lqs_key = (struct quota_adjust_qunit *)key;
1586 hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1588 RETURN(hash & mask);
1592 lqs_key(cfs_hlist_node_t *hnode)
1594 struct lustre_qunit_size *lqs;
1597 lqs = cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1598 RETURN(&lqs->lqs_key);
1602 lqs_keycmp(const void *key, cfs_hlist_node_t *hnode)
1604 struct lustre_qunit_size *q =
1605 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1607 RETURN(q->lqs_key == *((unsigned long long *)key));
1611 lqs_object(cfs_hlist_node_t *hnode)
1613 return cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1617 lqs_get(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1619 struct lustre_qunit_size *q =
1620 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1626 lqs_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1628 struct lustre_qunit_size *q =
1629 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1635 lqs_exit(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
1637 CERROR("It should not have any item left to be handled by this!");
1640 static cfs_hash_ops_t lqs_hash_ops = {
1641 .hs_hash = lqs_hash,
1643 .hs_keycmp = lqs_keycmp,
1644 .hs_object = lqs_object,
1646 .hs_put_locked = lqs_put_locked,
1649 #endif /* HAVE_QUOTA_SUPPORT */