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 2008 Sun Microsystems, Inc. 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 #ifdef HAVE_QUOTA_SUPPORT
65 static cfs_hash_ops_t lqs_hash_ops;
67 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
68 unsigned long default_btune_ratio = 50; /* 50 percentage */
69 unsigned long default_iunit_sz = 5120; /* 5120 inodes */
70 unsigned long default_itune_ratio = 50; /* 50 percentage */
72 cfs_mem_cache_t *qunit_cachep = NULL;
73 cfs_list_t qunit_hash[NR_DQHASH];
74 cfs_spinlock_t qunit_hash_lock = CFS_SPIN_LOCK_UNLOCKED;
76 /* please sync qunit_state with qunit_state_names */
83 * a qunit is added into qunit hash, that means
84 * a quota req will be sent or is flying
88 * a qunit is removed from qunit hash, that
89 * means a quota req is handled and comes back
91 QUNIT_RM_FROM_HASH = 2,
93 * qunit can wake up all threads waiting for it
98 static const char *qunit_state_names[] = {
99 [QUNIT_CREATED] = "CREATED",
100 [QUNIT_IN_HASH] = "IN_HASH",
101 [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
102 [QUNIT_FINISHED] = "FINISHED",
105 struct lustre_qunit {
106 cfs_list_t lq_hash; /** Hash list in memory */
107 cfs_atomic_t lq_refcnt; /** Use count */
108 struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
109 struct qunit_data lq_data; /** See qunit_data */
110 unsigned int lq_opc; /** QUOTA_DQACQ, QUOTA_DQREL */
111 cfs_waitq_t lq_waitq; /** Threads waiting for this qunit */
112 cfs_spinlock_t lq_lock; /** Protect the whole structure */
113 enum qunit_state lq_state; /** Present the status of qunit */
114 int lq_rc; /** The rc of lq_data */
118 #define QUNIT_SET_STATE(qunit, state) \
120 cfs_spin_lock(&qunit->lq_lock); \
121 QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), " \
122 "lq_rc(%d), lq_owner(%d)\n", \
123 qunit, qunit_state_names[qunit->lq_state], \
124 qunit_state_names[state], qunit->lq_rc, \
126 qunit->lq_state = state; \
127 cfs_spin_unlock(&qunit->lq_lock); \
130 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc) \
132 cfs_spin_lock(&qunit->lq_lock); \
134 QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), " \
135 "lq_rc(%d), lq_owner(%d)\n", \
136 qunit, qunit_state_names[qunit->lq_state], \
137 qunit_state_names[state], qunit->lq_rc, \
139 qunit->lq_state = state; \
140 cfs_spin_unlock(&qunit->lq_lock); \
143 int should_translate_quota (struct obd_import *imp)
148 if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
154 void qunit_cache_cleanup(void)
159 cfs_spin_lock(&qunit_hash_lock);
160 for (i = 0; i < NR_DQHASH; i++)
161 LASSERT(cfs_list_empty(qunit_hash + i));
162 cfs_spin_unlock(&qunit_hash_lock);
166 rc = cfs_mem_cache_destroy(qunit_cachep);
167 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
173 int qunit_cache_init(void)
178 LASSERT(qunit_cachep == NULL);
179 qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
180 sizeof(struct lustre_qunit),
185 cfs_spin_lock(&qunit_hash_lock);
186 for (i = 0; i < NR_DQHASH; i++)
187 CFS_INIT_LIST_HEAD(qunit_hash + i);
188 cfs_spin_unlock(&qunit_hash_lock);
193 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
194 __attribute__((__const__));
197 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
199 unsigned int id = qdata->qd_id;
200 unsigned int type = QDATA_IS_GRP(qdata);
202 unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
203 tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
207 /* caller must hold qunit_hash_lock */
208 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
209 struct lustre_quota_ctxt *qctxt,
210 struct qunit_data *qdata)
212 struct lustre_qunit *qunit = NULL;
213 struct qunit_data *tmp;
215 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
216 cfs_list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
217 tmp = &qunit->lq_data;
218 if (qunit->lq_ctxt == qctxt &&
219 qdata->qd_id == tmp->qd_id &&
220 (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
221 (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
227 /* check_cur_qunit - check the current usage of qunit.
228 * @qctxt: quota context
229 * @qdata: the type of quota unit to be checked
231 * return: 1 - need acquire qunit;
232 * 2 - need release qunit;
233 * 0 - need do nothing.
237 check_cur_qunit(struct obd_device *obd,
238 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
240 struct super_block *sb = qctxt->lqc_sb;
241 unsigned long qunit_sz, tune_sz;
242 __u64 usage, limit, limit_org, pending_write = 0;
243 long long record = 0;
244 struct obd_quotactl *qctl;
245 struct lustre_qunit_size *lqs = NULL;
249 if (!ll_sb_any_quota_active(sb))
252 cfs_spin_lock(&qctxt->lqc_lock);
253 if (!qctxt->lqc_valid){
254 cfs_spin_unlock(&qctxt->lqc_lock);
257 cfs_spin_unlock(&qctxt->lqc_lock);
263 /* get fs quota usage & limit */
264 qctl->qc_cmd = Q_GETQUOTA;
265 qctl->qc_id = qdata->qd_id;
266 qctl->qc_type = QDATA_IS_GRP(qdata);
267 ret = fsfilt_quotactl(obd, sb, qctl);
269 if (ret == -ESRCH) /* no limit */
272 CERROR("can't get fs quota usage! (rc:%d)\n", ret);
276 if (QDATA_IS_BLK(qdata)) {
277 usage = qctl->qc_dqblk.dqb_curspace;
278 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
280 usage = qctl->qc_dqblk.dqb_curinodes;
281 limit = qctl->qc_dqblk.dqb_ihardlimit;
284 /* ignore the no quota limit case; and it can avoid creating
285 * unnecessary lqs for uid/gid */
289 lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
291 if (IS_ERR(lqs) || lqs == NULL) {
292 CERROR("fail to find a lqs for %sid: %u)!\n",
293 QDATA_IS_GRP(qdata) ? "g" : "u", qdata->qd_id);
296 cfs_spin_lock(&lqs->lqs_lock);
298 if (QDATA_IS_BLK(qdata)) {
299 qunit_sz = lqs->lqs_bunit_sz;
300 tune_sz = lqs->lqs_btune_sz;
301 pending_write = lqs->lqs_bwrite_pending;
302 record = lqs->lqs_blk_rec;
303 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
305 /* we didn't need change inode qunit size now */
306 qunit_sz = lqs->lqs_iunit_sz;
307 tune_sz = lqs->lqs_itune_sz;
308 pending_write = lqs->lqs_iwrite_pending;
309 record = lqs->lqs_ino_rec;
312 /* we don't count the MIN_QLIMIT */
313 if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
314 (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
317 usage += pending_write;
319 /* when a releasing quota req is sent, before it returned
320 limit is assigned a small value. limit will overflow */
321 if (limit + record < 0)
326 LASSERT(qdata->qd_count == 0);
327 if (limit <= usage + tune_sz) {
328 while (qdata->qd_count + limit <=
330 qdata->qd_count += qunit_sz;
332 } else if (limit > usage + qunit_sz + tune_sz &&
333 limit_org > qdata->qd_count + qunit_sz) {
334 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
335 limit_org > qdata->qd_count + qunit_sz)
336 qdata->qd_count += qunit_sz;
338 /* if there are other pending writes for this uid/gid, releasing
339 * quota is put off until the last pending write b=16645 */
340 /* if there is an ongoing quota request, a releasing request is aborted.
341 * That ongoing quota request will call this function again when
342 * it returned b=18630 */
343 if (pending_write || record) {
344 CDEBUG(D_QUOTA, "delay quota release\n");
349 quota_compute_lqs(qdata, lqs, 1, (ret == 1) ? 1 : 0);
351 CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
352 ", pending_write: "LPU64", record: %lld"
353 ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
354 QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
355 record, qunit_sz, tune_sz, ret);
356 LASSERT(ret == 0 || qdata->qd_count);
358 cfs_spin_unlock(&lqs->lqs_lock);
368 * Compute the remaining quota for certain gid or uid b=11693
370 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
371 struct qunit_data *qdata, int isblk)
373 struct super_block *sb = qctxt->lqc_sb;
375 struct obd_quotactl *qctl;
376 int ret = QUOTA_RET_OK;
379 if (!ll_sb_any_quota_active(sb))
380 RETURN(QUOTA_RET_NOQUOTA);
382 /* ignore root user */
383 if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
384 RETURN(QUOTA_RET_NOLIMIT);
390 /* get fs quota usage & limit */
391 qctl->qc_cmd = Q_GETQUOTA;
392 qctl->qc_id = qdata->qd_id;
393 qctl->qc_type = QDATA_IS_GRP(qdata);
394 ret = fsfilt_quotactl(obd, sb, qctl);
396 if (ret == -ESRCH) /* no limit */
397 ret = QUOTA_RET_NOLIMIT;
399 CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
404 usage = isblk ? qctl->qc_dqblk.dqb_curspace :
405 qctl->qc_dqblk.dqb_curinodes;
406 limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
407 qctl->qc_dqblk.dqb_ihardlimit;
408 if (!limit){ /* no limit */
409 ret = QUOTA_RET_NOLIMIT;
414 qdata->qd_count = limit - usage;
423 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
424 struct qunit_data *qdata, int opc)
426 struct lustre_qunit *qunit = NULL;
429 OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
433 CFS_INIT_LIST_HEAD(&qunit->lq_hash);
434 cfs_waitq_init(&qunit->lq_waitq);
435 cfs_atomic_set(&qunit->lq_refcnt, 1);
436 qunit->lq_ctxt = qctxt;
437 memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
439 qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
440 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
441 qunit->lq_owner = cfs_curproc_pid();
445 static inline void free_qunit(struct lustre_qunit *qunit)
447 OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
450 static inline void qunit_get(struct lustre_qunit *qunit)
452 cfs_atomic_inc(&qunit->lq_refcnt);
455 static void qunit_put(struct lustre_qunit *qunit)
457 LASSERT(cfs_atomic_read(&qunit->lq_refcnt));
458 if (cfs_atomic_dec_and_test(&qunit->lq_refcnt))
462 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
463 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
464 struct qunit_data *qdata)
466 unsigned int hashent = qunit_hashfn(qctxt, qdata);
467 struct lustre_qunit *qunit;
470 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
471 qunit = find_qunit(hashent, qctxt, qdata);
478 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
482 LASSERT(cfs_list_empty(&qunit->lq_hash));
484 head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
485 cfs_list_add(&qunit->lq_hash, head);
486 QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
489 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
491 struct lustre_qunit_size *lqs;
493 lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(&qunit->lq_data),
494 qunit->lq_data.qd_id),
496 if (lqs && !IS_ERR(lqs)) {
497 cfs_spin_lock(&lqs->lqs_lock);
498 if (qunit->lq_opc == QUOTA_DQACQ)
499 quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
500 if (qunit->lq_opc == QUOTA_DQREL)
501 quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
502 cfs_spin_unlock(&lqs->lqs_lock);
503 /* this is for quota_search_lqs */
505 /* this is for schedule_dqacq */
510 static void remove_qunit_nolock(struct lustre_qunit *qunit)
512 LASSERT(!cfs_list_empty(&qunit->lq_hash));
513 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
515 cfs_list_del_init(&qunit->lq_hash);
516 QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
520 void* quota_barrier(struct lustre_quota_ctxt *qctxt,
521 struct obd_quotactl *oqctl, int isblk)
523 struct lustre_qunit *qunit, *find_qunit;
526 OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
528 CERROR("locating %sunit failed for %sid %u\n",
529 isblk ? "b" : "i", oqctl->qc_type ? "g" : "u",
531 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
532 oqctl->qc_type, isblk);
536 CFS_INIT_LIST_HEAD(&qunit->lq_hash);
537 qunit->lq_lock = CFS_SPIN_LOCK_UNLOCKED;
538 cfs_waitq_init(&qunit->lq_waitq);
539 cfs_atomic_set(&qunit->lq_refcnt, 1);
540 qunit->lq_ctxt = qctxt;
541 qunit->lq_data.qd_id = oqctl->qc_id;
542 qunit->lq_data.qd_flags = oqctl->qc_type;
544 QDATA_SET_BLK(&qunit->lq_data);
545 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
546 /* it means it is only an invalid qunit for barrier */
547 qunit->lq_opc = QUOTA_LAST_OPC;
550 cfs_spin_lock(&qunit_hash_lock);
551 find_qunit = dqacq_in_flight(qctxt, &qunit->lq_data);
553 cfs_spin_unlock(&qunit_hash_lock);
554 qunit_put(find_qunit);
555 qctxt_wait_pending_dqacq(qctxt, oqctl->qc_id,
556 oqctl->qc_type, isblk);
557 CDEBUG(D_QUOTA, "cycle=%d\n", cycle++);
562 insert_qunit_nolock(qctxt, qunit);
563 cfs_spin_unlock(&qunit_hash_lock);
567 void quota_unbarrier(void *handle)
569 struct lustre_qunit *qunit = (struct lustre_qunit *)handle;
572 CERROR("handle is NULL\n");
576 LASSERT(qunit->lq_opc == QUOTA_LAST_OPC);
577 cfs_spin_lock(&qunit_hash_lock);
578 remove_qunit_nolock(qunit);
579 cfs_spin_unlock(&qunit_hash_lock);
580 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, QUOTA_REQ_RETURNED);
581 cfs_waitq_signal(&qunit->lq_waitq);
585 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
586 (limit = count) : (limit += count)
589 static inline int is_master(struct lustre_quota_ctxt *qctxt)
591 return qctxt->lqc_handler ? 1 : 0;
595 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
596 struct qunit_data *qdata, int opc, int wait,
597 struct obd_trans_info *oti);
599 static inline void qdata_to_oqaq(struct qunit_data *qdata,
600 struct quota_adjust_qunit *oqaq)
605 oqaq->qaq_flags = qdata->qd_flags;
606 oqaq->qaq_id = qdata->qd_id;
607 if (QDATA_IS_ADJBLK(qdata))
608 oqaq->qaq_bunit_sz = qdata->qd_qunit;
609 if (QDATA_IS_ADJINO(qdata))
610 oqaq->qaq_iunit_sz = qdata->qd_qunit;
614 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
615 struct qunit_data *qdata, int rc, int opc)
617 struct lustre_qunit *qunit = NULL;
618 struct super_block *sb = qctxt->lqc_sb;
620 struct quota_adjust_qunit *oqaq = NULL;
625 QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
626 obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
628 /* do it only when a releasing quota req more than 5MB b=18491 */
629 if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
630 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
632 /* update local operational quota file */
634 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
635 struct obd_quotactl *qctl;
640 GOTO(out, err = -ENOMEM);
642 /* acq/rel qunit for specified uid/gid is serialized,
643 * so there is no race between get fs quota limit and
644 * set fs quota limit */
645 qctl->qc_cmd = Q_GETQUOTA;
646 qctl->qc_id = qdata->qd_id;
647 qctl->qc_type = QDATA_IS_GRP(qdata);
648 err = fsfilt_quotactl(obd, sb, qctl);
650 CERROR("error get quota fs limit! (rc:%d)\n", err);
654 if (QDATA_IS_BLK(qdata)) {
655 qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
656 hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
658 qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
659 hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
662 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
669 INC_QLIMIT(*hardlimit, count);
672 LASSERTF(count < *hardlimit,
673 "id(%u) flag(%u) type(%c) isblk(%c) "
674 "count("LPU64") qd_qunit("LPU64") "
675 "hardlimit("LPU64").\n",
676 qdata->qd_id, qdata->qd_flags,
677 QDATA_IS_GRP(qdata) ? 'g' : 'u',
678 QDATA_IS_BLK(qdata) ? 'b': 'i',
679 qdata->qd_count, qdata->qd_qunit, *hardlimit);
686 /* clear quota limit */
690 qctl->qc_cmd = Q_SETQUOTA;
691 err = fsfilt_quotactl(obd, sb, qctl);
693 CERROR("error set quota fs limit! (rc:%d)\n", err);
695 QDATA_DEBUG(qdata, "%s completion\n",
696 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
699 } else if (rc == -EDQUOT) {
700 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
701 } else if (rc == -EBUSY) {
702 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
704 CERROR("acquire qunit got error! (rc:%d)\n", rc);
707 /* remove the qunit from hash */
708 cfs_spin_lock(&qunit_hash_lock);
710 qunit = dqacq_in_flight(qctxt, qdata);
711 /* this qunit has been removed by qctxt_cleanup() */
713 cfs_spin_unlock(&qunit_hash_lock);
714 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
715 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
719 LASSERT(opc == qunit->lq_opc);
720 /* remove this qunit from lq_hash so that new processes cannot be added
721 * to qunit->lq_waiters */
722 remove_qunit_nolock(qunit);
723 cfs_spin_unlock(&qunit_hash_lock);
725 compute_lqs_after_removing_qunit(qunit);
728 rc = QUOTA_REQ_RETURNED;
729 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
730 /* wake up all waiters */
731 cfs_waitq_broadcast(&qunit->lq_waitq);
733 /* this is for dqacq_in_flight() */
735 if (rc < 0 && rc != -EDQUOT)
738 /* don't reschedule in such cases:
739 * - acq/rel failure and qunit isn't changed,
740 * but not for quota recovery.
741 * - local dqacq/dqrel.
742 * - local disk io failure.
746 GOTO(out1, err = -ENOMEM);
747 qdata_to_oqaq(qdata, oqaq);
748 /* adjust the qunit size in slaves */
749 rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
752 CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
753 GOTO(out1, err = rc1);
755 if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
758 if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
759 OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
762 /* reschedule another dqacq/dqrel if needed */
764 qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
765 rc1 = check_cur_qunit(obd, qctxt, qdata);
768 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
769 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
770 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
773 /* this is for alloc_qunit() */
778 struct dqacq_async_args {
779 struct lustre_quota_ctxt *aa_ctxt;
780 struct lustre_qunit *aa_qunit;
783 static int dqacq_interpret(const struct lu_env *env,
784 struct ptlrpc_request *req, void *data, int rc)
786 struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
787 struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
788 struct obd_device_target *obt = qctxt->lqc_obt;
789 struct lustre_qunit *qunit = aa->aa_qunit;
790 struct obd_device *obd = req->rq_import->imp_obd;
791 struct qunit_data *qdata = NULL;
795 LASSERT(req->rq_import);
797 cfs_down_read(&obt->obt_rwsem);
798 /* if a quota req timeouts or is dropped, we should update quota
799 * statistics which will be handled in dqacq_completion. And in
800 * this situation we should get qdata from request instead of
802 qdata = quota_get_qdata(req, (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
806 DEBUG_REQ(D_ERROR, req,
807 "error unpacking qunit_data(rc: %ld)\n",
809 qdata = &qunit->lq_data;
812 QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
813 QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
815 if (qdata->qd_id != qunit->lq_data.qd_id ||
816 OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
817 CERROR("the returned qd_id isn't expected!"
818 "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
819 qunit->lq_data.qd_id);
820 qdata->qd_id = qunit->lq_data.qd_id;
823 if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
824 CERROR("the returned grp/usr isn't expected!"
825 "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
826 qunit->lq_data.qd_flags);
827 if (QDATA_IS_GRP(&qunit->lq_data))
828 QDATA_SET_GRP(qdata);
830 QDATA_CLR_GRP(qdata);
833 if (qdata->qd_count > qunit->lq_data.qd_count) {
834 CERROR("the returned qd_count isn't expected!"
835 "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
836 qunit->lq_data.qd_count);
840 if (unlikely(rc == -ESRCH))
841 CERROR("quota for %s has been enabled by master, but disabled "
842 "by slave.\n", QDATA_IS_GRP(qdata) ? "group" : "user");
844 rc = dqacq_completion(obd, qctxt, qdata, rc,
845 lustre_msg_get_opc(req->rq_reqmsg));
847 cfs_up_read(&obt->obt_rwsem);
852 * check if quota master is online
854 int check_qm(struct lustre_quota_ctxt *qctxt)
859 cfs_spin_lock(&qctxt->lqc_lock);
860 /* quit waiting when mds is back or qctxt is cleaned up */
861 rc = qctxt->lqc_import || !qctxt->lqc_valid;
862 cfs_spin_unlock(&qctxt->lqc_lock);
867 /* wake up all waiting threads when lqc_import is NULL */
868 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
870 struct lustre_qunit *qunit, *tmp;
874 cfs_spin_lock(&qunit_hash_lock);
875 for (i = 0; i < NR_DQHASH; i++) {
876 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
878 if (qunit->lq_ctxt != qctxt)
881 /* Wake up all waiters. Do not change lq_state.
882 * The waiters will check lq_rc which is kept as 0
883 * if no others change it, then the waiters will return
884 * -EAGAIN to caller who can perform related quota
885 * acq/rel if necessary. */
886 cfs_waitq_broadcast(&qunit->lq_waitq);
889 cfs_spin_unlock(&qunit_hash_lock);
893 static int got_qunit(struct lustre_qunit *qunit, int is_master)
895 struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
899 cfs_spin_lock(&qunit->lq_lock);
900 switch (qunit->lq_state) {
902 case QUNIT_RM_FROM_HASH:
908 CERROR("invalid qunit state %d\n", qunit->lq_state);
910 cfs_spin_unlock(&qunit->lq_lock);
913 cfs_spin_lock(&qctxt->lqc_lock);
914 rc = !qctxt->lqc_valid;
916 rc |= !qctxt->lqc_import;
917 cfs_spin_unlock(&qctxt->lqc_lock);
924 revoke_lqs_rec(struct lustre_qunit_size *lqs, struct qunit_data *qdata, int opc)
926 /* revoke lqs_xxx_rec which is computed in check_cur_qunit
928 cfs_spin_lock(&lqs->lqs_lock);
929 quota_compute_lqs(qdata, lqs, 0, (opc == QUOTA_DQACQ) ? 1 : 0);
930 cfs_spin_unlock(&lqs->lqs_lock);
934 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
935 struct qunit_data *qdata, int opc, int wait,
936 struct obd_trans_info *oti)
938 struct lustre_qunit *qunit, *empty;
939 struct l_wait_info lwi = { 0 };
940 struct ptlrpc_request *req;
941 struct dqacq_async_args *aa;
942 struct obd_import *imp = NULL;
943 struct lustre_qunit_size *lqs = NULL;
944 struct timeval work_start;
945 struct timeval work_end;
950 LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
951 cfs_gettimeofday(&work_start);
953 lqs = quota_search_lqs(LQS_KEY(QDATA_IS_GRP(qdata), qdata->qd_id),
955 if (lqs == NULL || IS_ERR(lqs)) {
956 CERROR("Can't find the lustre qunit size!\n");
960 if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL) {
961 revoke_lqs_rec(lqs, qdata, opc);
962 /* this is for quota_search_lqs */
967 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_SD, 5);
969 cfs_spin_lock(&qunit_hash_lock);
970 qunit = dqacq_in_flight(qctxt, qdata);
972 cfs_spin_unlock(&qunit_hash_lock);
975 revoke_lqs_rec(lqs, qdata, opc);
976 /* this is for quota_search_lqs */
978 goto wait_completion;
982 insert_qunit_nolock(qctxt, qunit);
983 cfs_spin_unlock(&qunit_hash_lock);
985 /* From here, the quota request will be sent anyway.
986 * When this qdata request returned or is cancelled,
987 * lqs_putref will be called at that time */
989 /* this is for quota_search_lqs */
992 QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
993 obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
994 /* master is going to dqacq/dqrel from itself */
995 if (is_master(qctxt)) {
997 QDATA_DEBUG(qdata, "local %s.\n",
998 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
999 QDATA_SET_CHANGE_QS(qdata);
1000 rc = qctxt->lqc_handler(obd, qdata, opc);
1001 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
1002 /* this is for qunit_get() */
1005 cfs_gettimeofday(&work_end);
1006 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1007 if (opc == QUOTA_DQACQ)
1008 lprocfs_counter_add(qctxt->lqc_stats,
1009 wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1012 lprocfs_counter_add(qctxt->lqc_stats,
1013 wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1015 RETURN(rc ? rc : rc2);
1018 cfs_spin_lock(&qctxt->lqc_lock);
1019 if (!qctxt->lqc_import) {
1020 cfs_spin_unlock(&qctxt->lqc_lock);
1021 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
1023 cfs_spin_lock(&qunit_hash_lock);
1024 remove_qunit_nolock(qunit);
1025 cfs_spin_unlock(&qunit_hash_lock);
1027 compute_lqs_after_removing_qunit(qunit);
1029 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
1030 cfs_waitq_broadcast(&qunit->lq_waitq);
1032 /* this is for qunit_get() */
1034 /* this for alloc_qunit() */
1036 cfs_spin_lock(&qctxt->lqc_lock);
1037 if (wait && !qctxt->lqc_import) {
1038 cfs_spin_unlock(&qctxt->lqc_lock);
1040 LASSERT(oti && oti->oti_thread &&
1041 oti->oti_thread->t_watchdog);
1043 lc_watchdog_disable(oti->oti_thread->t_watchdog);
1044 CDEBUG(D_QUOTA, "sleep for quota master\n");
1045 l_wait_event(qctxt->lqc_wait_for_qmaster,
1046 check_qm(qctxt), &lwi);
1047 CDEBUG(D_QUOTA, "wake up when quota master is back\n");
1048 lc_watchdog_touch(oti->oti_thread->t_watchdog,
1049 CFS_GET_TIMEOUT(oti->oti_thread->t_svc));
1051 cfs_spin_unlock(&qctxt->lqc_lock);
1056 imp = class_import_get(qctxt->lqc_import);
1057 cfs_spin_unlock(&qctxt->lqc_lock);
1059 /* build dqacq/dqrel request */
1062 req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
1063 LUSTRE_MDS_VERSION, opc);
1064 class_import_put(imp);
1066 CERROR("Can't alloc request\n");
1067 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
1068 /* this is for qunit_get() */
1073 ptlrpc_request_set_replen(req);
1074 req->rq_no_resend = req->rq_no_delay = 1;
1075 rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
1077 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
1078 ptlrpc_req_finished(req);
1079 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
1080 /* this is for qunit_get() */
1085 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1086 aa = ptlrpc_req_async_args(req);
1087 aa->aa_ctxt = qctxt;
1088 aa->aa_qunit = qunit;
1090 req->rq_interpret_reply = dqacq_interpret;
1091 ptlrpcd_add_req(req, PSCOPE_OTHER);
1093 QDATA_DEBUG(qdata, "%s scheduled.\n",
1094 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
1096 if (wait && qunit) {
1097 struct qunit_data *p = &qunit->lq_data;
1099 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1100 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1102 /* rc = -EAGAIN, it means the quota master isn't ready yet
1103 * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
1104 * rc = -EDQUOT, it means out of quota
1105 * rc = -EBUSY, it means recovery is happening
1106 * other rc < 0, it means real errors, functions who call
1107 * schedule_dqacq should take care of this */
1108 cfs_spin_lock(&qunit->lq_lock);
1110 cfs_spin_unlock(&qunit->lq_lock);
1111 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: id(%u) flag(%u) "
1112 "rc(%d) owner(%d)\n", qunit, qunit->lq_data.qd_id,
1113 qunit->lq_data.qd_flags, rc, qunit->lq_owner);
1117 cfs_gettimeofday(&work_end);
1118 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1119 if (opc == QUOTA_DQACQ)
1120 lprocfs_counter_add(qctxt->lqc_stats,
1121 wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1124 lprocfs_counter_add(qctxt->lqc_stats,
1125 wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1132 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1133 const unsigned int id[], __u32 isblk, int wait,
1134 struct obd_trans_info *oti)
1136 int rc = 0, i = USRQUOTA;
1137 struct qunit_data qdata[MAXQUOTAS];
1140 if (quota_is_set(obd, id, isblk ? QB_SET : QI_SET) == 0)
1143 for (i = 0; i < MAXQUOTAS; i++) {
1144 qdata[i].qd_id = id[i];
1145 qdata[i].qd_flags = i;
1147 QDATA_SET_BLK(&qdata[i]);
1148 qdata[i].qd_count = 0;
1150 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1153 /* need acquire or release */
1154 opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1155 rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1159 } else if (wait == 1) {
1160 /* when wait equates 1, that means mds_quota_acquire
1161 * or filter_quota_acquire is calling it. */
1162 rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1172 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1173 unsigned short type, int isblk)
1175 struct lustre_qunit *qunit = NULL;
1176 struct qunit_data qdata;
1177 struct timeval work_start;
1178 struct timeval work_end;
1180 struct l_wait_info lwi = { 0 };
1184 cfs_gettimeofday(&work_start);
1186 qdata.qd_flags = type;
1188 QDATA_SET_BLK(&qdata);
1191 cfs_spin_lock(&qunit_hash_lock);
1192 qunit = dqacq_in_flight(qctxt, &qdata);
1193 cfs_spin_unlock(&qunit_hash_lock);
1196 struct qunit_data *p = &qunit->lq_data;
1198 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1199 l_wait_event(qunit->lq_waitq, got_qunit(qunit, is_master(qctxt)),
1201 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting: rc(%d) "
1202 "owner(%d)\n", qunit, qunit->lq_rc, qunit->lq_owner);
1203 /* keep same as schedule_dqacq() b=17030 */
1204 cfs_spin_lock(&qunit->lq_lock);
1206 cfs_spin_unlock(&qunit->lq_lock);
1207 /* this is for dqacq_in_flight() */
1209 cfs_gettimeofday(&work_end);
1210 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1211 lprocfs_counter_add(qctxt->lqc_stats,
1212 isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1213 LQUOTA_WAIT_PENDING_INO_QUOTA,
1216 cfs_gettimeofday(&work_end);
1217 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1218 lprocfs_counter_add(qctxt->lqc_stats,
1219 isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1220 LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1228 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1230 struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1231 struct obd_device_target *obt = &obd->u.obt;
1232 struct super_block *sb = obt->obt_sb;
1238 rc = ptlrpcd_addref();
1242 cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1243 cfs_waitq_init(&qctxt->lqc_lqs_waitq);
1244 cfs_atomic_set(&qctxt->lqc_lqs, 0);
1245 cfs_spin_lock_init(&qctxt->lqc_lock);
1246 cfs_spin_lock(&qctxt->lqc_lock);
1247 qctxt->lqc_handler = handler;
1249 qctxt->lqc_obt = obt;
1250 qctxt->lqc_import = NULL;
1251 qctxt->lqc_recovery = 0;
1252 qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1253 qctxt->lqc_valid = 1;
1254 qctxt->lqc_cqs_boundary_factor = 4;
1255 qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1256 qctxt->lqc_cqs_least_iunit = 2;
1257 qctxt->lqc_cqs_qs_factor = 2;
1258 qctxt->lqc_flags = 0;
1259 QUOTA_MASTER_UNREADY(qctxt);
1260 qctxt->lqc_bunit_sz = default_bunit_sz;
1261 qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1262 qctxt->lqc_iunit_sz = default_iunit_sz;
1263 qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1264 qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1265 * after the last shrinking */
1266 qctxt->lqc_sync_blk = 0;
1267 cfs_spin_unlock(&qctxt->lqc_lock);
1269 qctxt->lqc_lqs_hash = cfs_hash_create("LQS_HASH",
1272 &lqs_hash_ops, CFS_HASH_REHASH);
1273 if (!qctxt->lqc_lqs_hash) {
1274 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1279 rc = lquota_proc_setup(obd, is_master(qctxt));
1281 CERROR("initialize proc for %s error!\n", obd->obd_name);
1287 static int check_lqs(struct lustre_quota_ctxt *qctxt)
1292 rc = !cfs_atomic_read(&qctxt->lqc_lqs);
1298 void hash_put_lqs(void *obj, void *data)
1300 lqs_putref((struct lustre_qunit_size *)obj);
1303 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1305 struct lustre_qunit *qunit, *tmp;
1306 cfs_list_t tmp_list;
1307 struct l_wait_info lwi = { 0 };
1308 struct obd_device_target *obt = qctxt->lqc_obt;
1312 CFS_INIT_LIST_HEAD(&tmp_list);
1314 cfs_spin_lock(&qctxt->lqc_lock);
1315 qctxt->lqc_valid = 0;
1316 cfs_spin_unlock(&qctxt->lqc_lock);
1318 cfs_spin_lock(&qunit_hash_lock);
1319 for (i = 0; i < NR_DQHASH; i++) {
1320 cfs_list_for_each_entry_safe(qunit, tmp, &qunit_hash[i],
1322 if (qunit->lq_ctxt != qctxt)
1324 remove_qunit_nolock(qunit);
1325 cfs_list_add(&qunit->lq_hash, &tmp_list);
1328 cfs_spin_unlock(&qunit_hash_lock);
1330 cfs_list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1331 cfs_list_del_init(&qunit->lq_hash);
1332 compute_lqs_after_removing_qunit(qunit);
1334 /* wake up all waiters */
1335 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1336 cfs_waitq_broadcast(&qunit->lq_waitq);
1340 /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1341 * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1342 while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1343 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1344 cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
1345 cfs_time_seconds(1));
1348 cfs_hash_for_each_safe(qctxt->lqc_lqs_hash, hash_put_lqs, NULL);
1349 l_wait_event(qctxt->lqc_lqs_waitq, check_lqs(qctxt), &lwi);
1350 cfs_down_write(&obt->obt_rwsem);
1351 cfs_hash_destroy(qctxt->lqc_lqs_hash);
1352 qctxt->lqc_lqs_hash = NULL;
1353 cfs_up_write(&obt->obt_rwsem);
1358 if (lquota_proc_cleanup(qctxt))
1359 CERROR("cleanup proc error!\n");
1365 struct qslave_recov_thread_data {
1366 struct obd_device *obd;
1367 struct lustre_quota_ctxt *qctxt;
1368 cfs_completion_t comp;
1371 /* FIXME only recovery block quota by now */
1372 static int qslave_recovery_main(void *arg)
1374 struct qslave_recov_thread_data *data = arg;
1375 struct obd_device *obd = data->obd;
1376 struct lustre_quota_ctxt *qctxt = data->qctxt;
1381 cfs_daemonize_ctxt("qslave_recovd");
1384 class_incref(obd, "qslave_recovd_filter", obd);
1386 cfs_complete(&data->comp);
1388 cfs_spin_lock(&qctxt->lqc_lock);
1389 if (qctxt->lqc_recovery) {
1390 cfs_spin_unlock(&qctxt->lqc_lock);
1391 class_decref(obd, "qslave_recovd_filter", obd);
1394 qctxt->lqc_recovery = 1;
1395 cfs_spin_unlock(&qctxt->lqc_lock);
1398 for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1399 struct qunit_data qdata;
1400 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1402 struct dquot_id *dqid, *tmp;
1405 LOCK_DQONOFF_MUTEX(dqopt);
1406 if (!ll_sb_has_quota_active(qctxt->lqc_sb, type)) {
1407 UNLOCK_DQONOFF_MUTEX(dqopt);
1411 LASSERT(dqopt->files[type] != NULL);
1412 CFS_INIT_LIST_HEAD(&id_list);
1413 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1414 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1416 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1418 UNLOCK_DQONOFF_MUTEX(dqopt);
1420 CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1422 cfs_list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1423 cfs_list_del_init(&dqid->di_link);
1424 /* skip slave recovery on itself */
1425 if (is_master(qctxt))
1427 if (rc && rc != -EBUSY)
1430 qdata.qd_id = dqid->di_id;
1431 qdata.qd_flags = type;
1432 QDATA_SET_BLK(&qdata);
1435 ret = check_cur_qunit(obd, qctxt, &qdata);
1438 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1439 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1447 if (rc && rc != -EBUSY)
1448 CERROR("qslave recovery failed! (id:%d type:%d "
1449 " rc:%d)\n", dqid->di_id, type, rc);
1455 cfs_spin_lock(&qctxt->lqc_lock);
1456 qctxt->lqc_recovery = 0;
1457 cfs_spin_unlock(&qctxt->lqc_lock);
1458 class_decref(obd, "qslave_recovd_filter", obd);
1463 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1465 struct qslave_recov_thread_data data;
1469 if (!ll_sb_any_quota_active(qctxt->lqc_sb))
1474 cfs_init_completion(&data.comp);
1476 rc = cfs_kernel_thread(qslave_recovery_main, &data,
1477 CLONE_VM|CLONE_FILES);
1479 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1482 cfs_wait_for_completion(&data.comp);
1487 int quota_is_on(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1491 for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1492 if (!Q_TYPESET(oqctl, type))
1494 if (!(qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type)))
1500 int quota_is_off(struct lustre_quota_ctxt *qctxt, struct obd_quotactl *oqctl)
1504 for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1505 if (!Q_TYPESET(oqctl, type))
1507 if (qctxt->lqc_flags & UGQUOTA2LQC(oqctl->qc_type))
1514 * When quotaon, build a lqs for every uid/gid who has been set limitation
1515 * for quota. After quota_search_lqs, it will hold one ref for the lqs.
1516 * It will be released when qctxt_cleanup() is executed b=18574
1518 * Should be called with obt->obt_quotachecking held. b=20152
1520 void build_lqs(struct obd_device *obd)
1522 struct obd_device_target *obt = &obd->u.obt;
1523 struct lustre_quota_ctxt *qctxt = &obt->obt_qctxt;
1527 LASSERT_SEM_LOCKED(&obt->obt_quotachecking);
1528 CFS_INIT_LIST_HEAD(&id_list);
1529 for (i = 0; i < MAXQUOTAS; i++) {
1530 struct dquot_id *dqid, *tmp;
1532 if (sb_dqopt(qctxt->lqc_sb)->files[i] == NULL)
1535 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1536 rc = fsfilt_qids(obd, sb_dqopt(qctxt->lqc_sb)->files[i], NULL,
1539 rc = fsfilt_qids(obd, NULL, sb_dqopt(qctxt->lqc_sb)->files[i],
1543 CERROR("%s: failed to get %s qids!\n", obd->obd_name,
1544 i ? "group" : "user");
1548 cfs_list_for_each_entry_safe(dqid, tmp, &id_list,
1550 struct lustre_qunit_size *lqs;
1552 cfs_list_del_init(&dqid->di_link);
1553 lqs = quota_search_lqs(LQS_KEY(i, dqid->di_id),
1555 if (lqs && !IS_ERR(lqs)) {
1556 lqs->lqs_flags |= dqid->di_flag;
1559 CERROR("%s: failed to create a lqs for %sid %u"
1560 "\n", obd->obd_name, i ? "g" : "u",
1570 * lqs<->qctxt hash operations
1574 * string hashing using djb2 hash algorithm
1577 lqs_hash(cfs_hash_t *hs, void *key, unsigned mask)
1579 struct quota_adjust_qunit *lqs_key;
1584 lqs_key = (struct quota_adjust_qunit *)key;
1585 hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1587 RETURN(hash & mask);
1591 lqs_key(cfs_hlist_node_t *hnode)
1593 struct lustre_qunit_size *lqs;
1596 lqs = cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1597 RETURN(&lqs->lqs_key);
1601 lqs_compare(void *key, cfs_hlist_node_t *hnode)
1603 struct lustre_qunit_size *q;
1608 q = cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1610 cfs_spin_lock(&q->lqs_lock);
1611 rc = (q->lqs_key == *((unsigned long long *)key));
1612 cfs_spin_unlock(&q->lqs_lock);
1618 lqs_get(cfs_hlist_node_t *hnode)
1620 struct lustre_qunit_size *q =
1621 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1630 lqs_put(cfs_hlist_node_t *hnode)
1632 struct lustre_qunit_size *q =
1633 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1642 lqs_exit(cfs_hlist_node_t *hnode)
1644 struct lustre_qunit_size *q =
1645 cfs_hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1649 * Nothing should be left. User of lqs put it and
1650 * lqs also was deleted from table by this time
1651 * so we should have 0 refs.
1653 LASSERTF(cfs_atomic_read(&q->lqs_refcount) == 0,
1654 "Busy lqs %p with %d refs\n", q,
1655 cfs_atomic_read(&q->lqs_refcount));
1660 static cfs_hash_ops_t lqs_hash_ops = {
1661 .hs_hash = lqs_hash,
1663 .hs_compare = lqs_compare,
1668 #endif /* HAVE_QUOTA_SUPPORT */