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 <class_hash.h>
61 #include <lprocfs_status.h>
62 #include "quota_internal.h"
64 #ifdef HAVE_QUOTA_SUPPORT
66 static lustre_hash_ops_t lqs_hash_ops;
68 unsigned long default_bunit_sz = 128 * 1024 * 1024; /* 128M bytes */
69 unsigned long default_btune_ratio = 50; /* 50 percentage */
70 unsigned long default_iunit_sz = 5120; /* 5120 inodes */
71 unsigned long default_itune_ratio = 50; /* 50 percentage */
73 cfs_mem_cache_t *qunit_cachep = NULL;
74 struct list_head qunit_hash[NR_DQHASH];
75 spinlock_t qunit_hash_lock = SPIN_LOCK_UNLOCKED;
77 /* please sync qunit_state with qunit_state_names */
79 QUNIT_CREATED = 0, /* a qunit is created */
80 QUNIT_IN_HASH = 1, /* a qunit is added into qunit hash, that means
81 * a quota req will be sent or is flying */
82 QUNIT_RM_FROM_HASH = 2, /* a qunit is removed from qunit hash, that
83 * means a quota req is handled and comes
85 QUNIT_FINISHED = 3, /* qunit can wake up all threads waiting
89 static const char *qunit_state_names[] = {
90 [QUNIT_CREATED] = "CREATED",
91 [QUNIT_IN_HASH] = "IN_HASH",
92 [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
93 [QUNIT_FINISHED] = "FINISHED",
97 struct list_head lq_hash; /* Hash list in memory */
98 atomic_t lq_refcnt; /* Use count */
99 struct lustre_quota_ctxt *lq_ctxt; /* Quota context this applies to */
100 struct qunit_data lq_data; /* See qunit_data */
101 unsigned int lq_opc; /* QUOTA_DQACQ, QUOTA_DQREL */
102 cfs_waitq_t lq_waitq; /* Threads waiting for this qunit */
103 spinlock_t lq_lock; /* Protect the whole structure */
104 enum qunit_state lq_state; /* Present the status of qunit */
105 int lq_rc; /* The rc of lq_data */
108 #define QUNIT_SET_STATE(qunit, state) \
110 spin_lock(&qunit->lq_lock); \
111 QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), " \
113 qunit, qunit_state_names[qunit->lq_state], \
114 qunit_state_names[state], qunit->lq_rc); \
115 qunit->lq_state = state; \
116 spin_unlock(&qunit->lq_lock); \
119 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc) \
121 spin_lock(&qunit->lq_lock); \
123 QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), " \
125 qunit, qunit_state_names[qunit->lq_state], \
126 qunit_state_names[state], qunit->lq_rc); \
127 qunit->lq_state = state; \
128 spin_unlock(&qunit->lq_lock); \
132 int should_translate_quota (struct obd_import *imp)
137 if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
143 void qunit_cache_cleanup(void)
148 spin_lock(&qunit_hash_lock);
149 for (i = 0; i < NR_DQHASH; i++)
150 LASSERT(list_empty(qunit_hash + i));
151 spin_unlock(&qunit_hash_lock);
155 rc = cfs_mem_cache_destroy(qunit_cachep);
156 LASSERTF(rc == 0, "couldn't destory qunit_cache slab\n");
162 int qunit_cache_init(void)
167 LASSERT(qunit_cachep == NULL);
168 qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
169 sizeof(struct lustre_qunit),
174 spin_lock(&qunit_hash_lock);
175 for (i = 0; i < NR_DQHASH; i++)
176 INIT_LIST_HEAD(qunit_hash + i);
177 spin_unlock(&qunit_hash_lock);
182 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
183 __attribute__((__const__));
186 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
188 unsigned int id = qdata->qd_id;
189 unsigned int type = QDATA_IS_GRP(qdata);
191 unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
192 tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
196 /* caller must hold qunit_hash_lock */
197 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
198 struct lustre_quota_ctxt *qctxt,
199 struct qunit_data *qdata)
201 struct lustre_qunit *qunit = NULL;
202 struct qunit_data *tmp;
204 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
205 list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
206 tmp = &qunit->lq_data;
207 if (qunit->lq_ctxt == qctxt &&
208 qdata->qd_id == tmp->qd_id &&
209 (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
210 (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
216 /* check_cur_qunit - check the current usage of qunit.
217 * @qctxt: quota context
218 * @qdata: the type of quota unit to be checked
220 * return: 1 - need acquire qunit;
221 * 2 - need release qunit;
222 * 0 - need do nothing.
226 check_cur_qunit(struct obd_device *obd,
227 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
229 struct super_block *sb = qctxt->lqc_sb;
230 unsigned long qunit_sz, tune_sz;
231 __u64 usage, limit, limit_org, pending_write = 0;
232 long long record = 0;
233 struct obd_quotactl *qctl;
234 struct lustre_qunit_size *lqs = NULL;
238 if (!sb_any_quota_enabled(sb))
241 spin_lock(&qctxt->lqc_lock);
242 if (!qctxt->lqc_valid){
243 spin_unlock(&qctxt->lqc_lock);
246 spin_unlock(&qctxt->lqc_lock);
252 /* get fs quota usage & limit */
253 qctl->qc_cmd = Q_GETQUOTA;
254 qctl->qc_id = qdata->qd_id;
255 qctl->qc_type = QDATA_IS_GRP(qdata);
256 ret = fsfilt_quotactl(obd, sb, qctl);
258 if (ret == -ESRCH) /* no limit */
261 CERROR("can't get fs quota usage! (rc:%d)\n", ret);
265 if (QDATA_IS_BLK(qdata)) {
266 usage = qctl->qc_dqblk.dqb_curspace;
267 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
269 usage = qctl->qc_dqblk.dqb_curinodes;
270 limit = qctl->qc_dqblk.dqb_ihardlimit;
273 /* ignore the no quota limit case; and it can avoid creating
274 * unnecessary lqs for uid/gid */
279 quota_search_lqs(qdata, NULL, qctxt, &lqs);
281 CDEBUG(D_QUOTA, "Can't find the lustre qunit size!\n");
282 ret = quota_create_lqs(qdata, NULL, qctxt, &lqs);
283 if (ret == -EALREADY) {
290 spin_lock(&lqs->lqs_lock);
292 if (QDATA_IS_BLK(qdata)) {
293 qunit_sz = lqs->lqs_bunit_sz;
294 tune_sz = lqs->lqs_btune_sz;
295 pending_write = lqs->lqs_bwrite_pending;
296 record = lqs->lqs_blk_rec;
297 LASSERT(!(qunit_sz % QUOTABLOCK_SIZE));
299 /* we didn't need change inode qunit size now */
300 qunit_sz = lqs->lqs_iunit_sz;
301 tune_sz = lqs->lqs_itune_sz;
302 pending_write = lqs->lqs_iwrite_pending;
303 record = lqs->lqs_ino_rec;
306 /* we don't count the MIN_QLIMIT */
307 if ((limit == MIN_QLIMIT && !QDATA_IS_BLK(qdata)) ||
308 (toqb(limit) == MIN_QLIMIT && QDATA_IS_BLK(qdata)))
311 usage += pending_write;
313 /* when a releasing quota req is sent, before it returned
314 limit is assigned a small value. limit will overflow */
315 if (limit + record < 0)
320 LASSERT(qdata->qd_count == 0);
321 if (limit <= usage + tune_sz) {
322 while (qdata->qd_count + limit <=
324 qdata->qd_count += qunit_sz;
326 } else if (limit > usage + qunit_sz + tune_sz &&
327 limit_org > qdata->qd_count + qunit_sz) {
328 while (limit - qdata->qd_count > usage + qunit_sz + tune_sz &&
329 limit_org > qdata->qd_count + qunit_sz)
330 qdata->qd_count += qunit_sz;
332 /* if there are other pending writes for this uid/gid, releasing
333 * quota is put off until the last pending write b=16645 */
334 if (ret == 2 && pending_write) {
335 CDEBUG(D_QUOTA, "delay quota release\n");
339 CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
340 ", pending_write: "LPU64", record: "LPD64
341 ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
342 QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
343 record, qunit_sz, tune_sz, ret);
344 LASSERT(ret == 0 || qdata->qd_count);
346 spin_unlock(&lqs->lqs_lock);
354 /* compute the remaining quota for certain gid or uid b=11693 */
355 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
356 struct qunit_data *qdata, int isblk)
358 struct super_block *sb = qctxt->lqc_sb;
360 struct obd_quotactl *qctl;
361 int ret = QUOTA_RET_OK;
364 if (!sb_any_quota_enabled(sb))
365 RETURN(QUOTA_RET_NOQUOTA);
367 /* ignore root user */
368 if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
369 RETURN(QUOTA_RET_NOLIMIT);
375 /* get fs quota usage & limit */
376 qctl->qc_cmd = Q_GETQUOTA;
377 qctl->qc_id = qdata->qd_id;
378 qctl->qc_type = QDATA_IS_GRP(qdata);
379 ret = fsfilt_quotactl(obd, sb, qctl);
381 if (ret == -ESRCH) /* no limit */
382 ret = QUOTA_RET_NOLIMIT;
384 CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
389 usage = isblk ? qctl->qc_dqblk.dqb_curspace :
390 qctl->qc_dqblk.dqb_curinodes;
391 limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
392 qctl->qc_dqblk.dqb_ihardlimit;
393 if (!limit){ /* no limit */
394 ret = QUOTA_RET_NOLIMIT;
399 qdata->qd_count = limit - usage;
408 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
409 struct qunit_data *qdata, int opc)
411 struct lustre_qunit *qunit = NULL;
414 OBD_SLAB_ALLOC(qunit, qunit_cachep, CFS_ALLOC_IO, sizeof(*qunit));
418 INIT_LIST_HEAD(&qunit->lq_hash);
419 init_waitqueue_head(&qunit->lq_waitq);
420 atomic_set(&qunit->lq_refcnt, 1);
421 qunit->lq_ctxt = qctxt;
422 memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
424 qunit->lq_lock = SPIN_LOCK_UNLOCKED;
425 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
429 static inline void free_qunit(struct lustre_qunit *qunit)
431 OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
434 static inline void qunit_get(struct lustre_qunit *qunit)
436 atomic_inc(&qunit->lq_refcnt);
439 static void qunit_put(struct lustre_qunit *qunit)
441 LASSERT(atomic_read(&qunit->lq_refcnt));
442 if (atomic_dec_and_test(&qunit->lq_refcnt))
446 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
447 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
448 struct qunit_data *qdata)
450 unsigned int hashent = qunit_hashfn(qctxt, qdata);
451 struct lustre_qunit *qunit;
454 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
455 qunit = find_qunit(hashent, qctxt, qdata);
462 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
464 struct list_head *head;
466 LASSERT(list_empty(&qunit->lq_hash));
468 head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
469 list_add(&qunit->lq_hash, head);
470 QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
473 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
475 struct lustre_qunit_size *lqs = NULL;
477 quota_search_lqs(&qunit->lq_data, NULL, qunit->lq_ctxt, &lqs);
479 spin_lock(&lqs->lqs_lock);
480 if (qunit->lq_opc == QUOTA_DQACQ)
481 quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
482 if (qunit->lq_opc == QUOTA_DQREL)
483 quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
484 spin_unlock(&lqs->lqs_lock);
485 /* this is for quota_search_lqs */
487 /* this is for schedule_dqacq */
493 static void remove_qunit_nolock(struct lustre_qunit *qunit)
495 LASSERT(!list_empty(&qunit->lq_hash));
496 LASSERT_SPIN_LOCKED(&qunit_hash_lock);
498 list_del_init(&qunit->lq_hash);
499 QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
503 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
504 (limit = count) : (limit += count)
507 /* FIXME check if this mds is the master of specified id */
509 is_master(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
510 unsigned int id, int type)
512 return qctxt->lqc_handler ? 1 : 0;
516 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
517 struct qunit_data *qdata, int opc, int wait,
518 struct obd_trans_info *oti);
521 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
522 struct qunit_data *qdata, int rc, int opc)
524 struct lustre_qunit *qunit = NULL;
525 struct super_block *sb = qctxt->lqc_sb;
527 struct quota_adjust_qunit *oqaq = NULL;
532 QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
533 obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
535 /* update local operational quota file */
537 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
538 struct obd_quotactl *qctl;
543 GOTO(out, err = -ENOMEM);
545 /* acq/rel qunit for specified uid/gid is serialized,
546 * so there is no race between get fs quota limit and
547 * set fs quota limit */
548 qctl->qc_cmd = Q_GETQUOTA;
549 qctl->qc_id = qdata->qd_id;
550 qctl->qc_type = QDATA_IS_GRP(qdata);
551 err = fsfilt_quotactl(obd, sb, qctl);
553 CERROR("error get quota fs limit! (rc:%d)\n", err);
557 if (QDATA_IS_BLK(qdata)) {
558 qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
559 hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
561 qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
562 hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
565 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
572 INC_QLIMIT(*hardlimit, count);
575 LASSERTF(count < *hardlimit,
576 "id(%u) flag(%u) type(%c) isblk(%c) "
577 "count("LPU64") qd_qunit("LPU64") "
578 "hardlimit("LPU64").\n",
579 qdata->qd_id, qdata->qd_flags,
580 QDATA_IS_GRP(qdata) ? 'g' : 'u',
581 QDATA_IS_BLK(qdata) ? 'b': 'i',
582 qdata->qd_count, qdata->qd_qunit, *hardlimit);
589 /* clear quota limit */
593 qctl->qc_cmd = Q_SETQUOTA;
594 err = fsfilt_quotactl(obd, sb, qctl);
596 CERROR("error set quota fs limit! (rc:%d)\n", err);
598 QDATA_DEBUG(qdata, "%s completion\n",
599 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
602 } else if (rc == -EDQUOT) {
603 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
604 } else if (rc == -EBUSY) {
605 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
607 CERROR("acquire qunit got error! (rc:%d)\n", rc);
610 /* remove the qunit from hash */
611 spin_lock(&qunit_hash_lock);
613 qunit = dqacq_in_flight(qctxt, qdata);
614 /* this qunit has been removed by qctxt_cleanup() */
616 spin_unlock(&qunit_hash_lock);
617 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
618 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
622 LASSERT(opc == qunit->lq_opc);
623 /* remove this qunit from lq_hash so that new processes cannot be added
624 * to qunit->lq_waiters */
625 remove_qunit_nolock(qunit);
626 spin_unlock(&qunit_hash_lock);
628 compute_lqs_after_removing_qunit(qunit);
630 /* wake up all waiters */
631 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
632 wake_up(&qunit->lq_waitq);
634 /* this is for dqacq_in_flight() */
636 /* this is for alloc_qunit() */
638 if (rc < 0 && rc != -EDQUOT)
641 /* don't reschedule in such cases:
642 * - acq/rel failure and qunit isn't changed,
643 * but not for quota recovery.
644 * - local dqacq/dqrel.
645 * - local disk io failure.
650 qdata_to_oqaq(qdata, oqaq);
651 /* adjust the qunit size in slaves */
652 rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
655 CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
658 if (err || (rc && rc != -EBUSY && rc1 == 0) ||
659 is_master(obd, qctxt, qdata->qd_id, QDATA_IS_GRP(qdata)))
662 /* reschedule another dqacq/dqrel if needed */
664 qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
665 rc1 = check_cur_qunit(obd, qctxt, qdata);
668 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
669 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
670 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
675 struct dqacq_async_args {
676 struct lustre_quota_ctxt *aa_ctxt;
677 struct lustre_qunit *aa_qunit;
680 static int dqacq_interpret(struct ptlrpc_request *req, void *data, int rc)
682 struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
683 struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
684 struct lustre_qunit *qunit = aa->aa_qunit;
685 struct obd_device *obd = req->rq_import->imp_obd;
686 struct qunit_data *qdata = NULL;
691 LASSERT(req->rq_import);
693 /* there are several forms of qunit(historic causes), so we need to
694 * adjust qunit from slaves to the same form here */
695 OBD_ALLOC(qdata, sizeof(struct qunit_data));
699 /* if a quota req timeouts or is dropped, we should update quota
700 * statistics which will be handled in dqacq_completion. And in
701 * this situation we should get qdata from request instead of
703 rc1 = quota_get_qdata(req, qdata,
704 (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
707 DEBUG_REQ(D_ERROR, req,
708 "error unpacking qunit_data(rc: %d)\n", rc1);
709 GOTO(exit, rc = rc1);
712 QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
713 QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
715 if (qdata->qd_id != qunit->lq_data.qd_id ||
716 OBD_FAIL_CHECK_ONCE(OBD_FAIL_QUOTA_RET_QDATA)) {
717 CDEBUG(D_ERROR, "the returned qd_id isn't expected!"
718 "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
719 qunit->lq_data.qd_id);
720 qdata->qd_id = qunit->lq_data.qd_id;
723 if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
724 CDEBUG(D_ERROR, "the returned grp/usr isn't expected!"
725 "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
726 qunit->lq_data.qd_flags);
727 if (QDATA_IS_GRP(&qunit->lq_data))
728 QDATA_SET_GRP(qdata);
730 QDATA_CLR_GRP(qdata);
733 if (qdata->qd_count > qunit->lq_data.qd_count) {
734 CDEBUG(D_ERROR, "the returned qd_count isn't expected!"
735 "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
736 qunit->lq_data.qd_count);
740 rc = dqacq_completion(obd, qctxt, qdata, rc,
741 lustre_msg_get_opc(req->rq_reqmsg));
744 OBD_FREE(qdata, sizeof(struct qunit_data));
749 /* check if quota master is online */
750 int check_qm(struct lustre_quota_ctxt *qctxt)
755 spin_lock(&qctxt->lqc_lock);
756 /* quit waiting when mds is back or qctxt is cleaned up */
757 rc = qctxt->lqc_import || !qctxt->lqc_valid;
758 spin_unlock(&qctxt->lqc_lock);
763 static int got_qunit(struct lustre_qunit *qunit)
768 spin_lock(&qunit->lq_lock);
769 switch (qunit->lq_state) {
771 case QUNIT_RM_FROM_HASH:
779 CERROR("invalid qunit state %d\n", qunit->lq_state);
781 spin_unlock(&qunit->lq_lock);
786 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
787 struct qunit_data *qdata, int opc, int wait,
788 struct obd_trans_info *oti)
790 struct lustre_qunit *qunit, *empty;
791 struct l_wait_info lwi = { 0 };
792 struct ptlrpc_request *req;
793 struct dqacq_async_args *aa;
794 int size[2] = { sizeof(struct ptlrpc_body), 0 };
795 struct obd_import *imp = NULL;
796 struct lustre_qunit_size *lqs = NULL;
797 struct timeval work_start;
798 struct timeval work_end;
803 LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
804 do_gettimeofday(&work_start);
805 if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
808 spin_lock(&qunit_hash_lock);
809 qunit = dqacq_in_flight(qctxt, qdata);
811 spin_unlock(&qunit_hash_lock);
814 goto wait_completion;
818 insert_qunit_nolock(qctxt, qunit);
819 spin_unlock(&qunit_hash_lock);
821 quota_search_lqs(qdata, NULL, qctxt, &lqs);
823 spin_lock(&lqs->lqs_lock);
824 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
825 /* when this qdata returned from mds, it will call lqs_putref */
827 spin_unlock(&lqs->lqs_lock);
828 /* this is for quota_search_lqs */
831 CDEBUG(D_ERROR, "Can't find the lustre qunit size!\n");
834 QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
835 obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
836 /* master is going to dqacq/dqrel from itself */
837 if (is_master(obd, qctxt, qdata->qd_id, QDATA_IS_GRP(qdata))) {
839 QDATA_DEBUG(qdata, "local %s.\n",
840 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
841 QDATA_SET_CHANGE_QS(qdata);
842 rc = qctxt->lqc_handler(obd, qdata, opc);
843 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
844 /* this is for qunit_get() */
847 do_gettimeofday(&work_end);
848 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
849 if (opc == QUOTA_DQACQ)
850 lprocfs_counter_add(qctxt->lqc_stats,
851 wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
854 lprocfs_counter_add(qctxt->lqc_stats,
855 wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
857 RETURN(rc ? rc : rc2);
860 spin_lock(&qctxt->lqc_lock);
861 if (!qctxt->lqc_import) {
862 spin_unlock(&qctxt->lqc_lock);
863 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
865 spin_lock(&qunit_hash_lock);
866 remove_qunit_nolock(qunit);
867 spin_unlock(&qunit_hash_lock);
869 compute_lqs_after_removing_qunit(qunit);
871 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
872 wake_up(&qunit->lq_waitq);
874 /* this is for qunit_get() */
876 /* this for alloc_qunit() */
878 spin_lock(&qctxt->lqc_lock);
879 if (wait && !qctxt->lqc_import) {
880 spin_unlock(&qctxt->lqc_lock);
882 LASSERT(oti && oti->oti_thread &&
883 oti->oti_thread->t_watchdog);
885 lc_watchdog_disable(oti->oti_thread->t_watchdog);
886 CDEBUG(D_QUOTA, "sleep for quota master\n");
887 l_wait_event(qctxt->lqc_wait_for_qmaster,
888 check_qm(qctxt), &lwi);
889 CDEBUG(D_QUOTA, "wake up when quota master is back\n");
890 lc_watchdog_touch(oti->oti_thread->t_watchdog);
892 spin_unlock(&qctxt->lqc_lock);
897 imp = class_import_get(qctxt->lqc_import);
898 spin_unlock(&qctxt->lqc_lock);
900 /* build dqacq/dqrel request */
902 size[1] = quota_get_qunit_data_size(imp->
903 imp_connect_data.ocd_connect_flags);
905 req = ptlrpc_prep_req(imp, LUSTRE_MDS_VERSION, opc, 2,
908 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
909 class_import_put(imp);
910 /* this is for qunit_get() */
915 rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
917 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
918 dqacq_completion(obd, qctxt, qdata, rc, opc);
919 class_import_put(imp);
920 /* this is for qunit_get() */
924 ptlrpc_req_set_repsize(req, 2, size);
925 req->rq_no_resend = req->rq_no_delay = 1;
926 class_import_put(imp);
928 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
929 aa = ptlrpc_req_async_args(req);
931 aa->aa_qunit = qunit;
933 req->rq_interpret_reply = dqacq_interpret;
934 ptlrpcd_add_req(req);
936 QDATA_DEBUG(qdata, "%s scheduled.\n",
937 opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
940 struct qunit_data *p = &qunit->lq_data;
942 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
943 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
944 /* rc = -EAGAIN, it means a quota req is finished;
945 * rc = -EDQUOT, it means out of quota
946 * rc = -EBUSY, it means recovery is happening
947 * other rc < 0, it means real errors, functions who call
948 * schedule_dqacq should take care of this */
949 spin_lock(&qunit->lq_lock);
950 if (qunit->lq_rc == 0)
954 spin_unlock(&qunit->lq_lock);
955 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
960 do_gettimeofday(&work_end);
961 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
962 if (opc == QUOTA_DQACQ)
963 lprocfs_counter_add(qctxt->lqc_stats,
964 wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
967 lprocfs_counter_add(qctxt->lqc_stats,
968 wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
975 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
976 uid_t uid, gid_t gid, __u32 isblk, int wait,
977 struct obd_trans_info *oti)
979 int rc = 0, i = USRQUOTA;
980 __u32 id[MAXQUOTAS] = { uid, gid };
981 struct qunit_data qdata[MAXQUOTAS];
984 CLASSERT(MAXQUOTAS < 4);
985 if (!sb_any_quota_enabled(qctxt->lqc_sb))
988 for (i = 0; i < MAXQUOTAS; i++) {
989 qdata[i].qd_id = id[i];
990 qdata[i].qd_flags = i;
992 QDATA_SET_BLK(&qdata[i]);
993 qdata[i].qd_count = 0;
995 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
998 /* need acquire or release */
999 opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1000 rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1004 } else if (wait == 1) {
1005 /* when wait equates 1, that means mds_quota_acquire
1006 * or filter_quota_acquire is calling it. */
1007 rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1017 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1018 unsigned short type, int isblk)
1020 struct lustre_qunit *qunit = NULL;
1021 struct qunit_data qdata;
1022 struct timeval work_start;
1023 struct timeval work_end;
1025 struct l_wait_info lwi = { 0 };
1029 do_gettimeofday(&work_start);
1031 qdata.qd_flags = type;
1033 QDATA_SET_BLK(&qdata);
1036 spin_lock(&qunit_hash_lock);
1037 qunit = dqacq_in_flight(qctxt, &qdata);
1038 spin_unlock(&qunit_hash_lock);
1041 struct qunit_data *p = &qunit->lq_data;
1043 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1044 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
1045 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1046 qunit, qunit->lq_rc);
1047 /* keep same as schedule_dqacq() b=17030 */
1048 spin_lock(&qunit->lq_lock);
1049 if (qunit->lq_rc == 0)
1053 spin_unlock(&qunit->lq_lock);
1054 /* this is for dqacq_in_flight() */
1056 do_gettimeofday(&work_end);
1057 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1058 lprocfs_counter_add(qctxt->lqc_stats,
1059 isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1060 LQUOTA_WAIT_PENDING_INO_QUOTA,
1063 do_gettimeofday(&work_end);
1064 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1065 lprocfs_counter_add(qctxt->lqc_stats,
1066 isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1067 LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1075 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1077 struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1078 struct super_block *sb = obd->u.obt.obt_sb;
1084 rc = ptlrpcd_addref();
1088 cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1089 spin_lock_init(&qctxt->lqc_lock);
1090 spin_lock(&qctxt->lqc_lock);
1091 qctxt->lqc_handler = handler;
1093 qctxt->lqc_import = NULL;
1094 qctxt->lqc_recovery = 0;
1095 qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1096 qctxt->lqc_valid = 1;
1097 qctxt->lqc_cqs_boundary_factor = 4;
1098 qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1099 qctxt->lqc_cqs_least_iunit = 2;
1100 qctxt->lqc_cqs_qs_factor = 2;
1101 qctxt->lqc_flags = 0;
1102 QUOTA_MASTER_UNREADY(qctxt);
1103 qctxt->lqc_bunit_sz = default_bunit_sz;
1104 qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1105 qctxt->lqc_iunit_sz = default_iunit_sz;
1106 qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1107 qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1108 * after the last shrinking */
1109 qctxt->lqc_sync_blk = 0;
1110 spin_unlock(&qctxt->lqc_lock);
1112 qctxt->lqc_lqs_hash = lustre_hash_init("LQS_HASH", 7, 7,
1114 if (!qctxt->lqc_lqs_hash)
1115 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1118 if (lquota_proc_setup(obd, is_master(obd, qctxt, 0, 0)))
1119 CERROR("initialize proc for %s error!\n", obd->obd_name);
1125 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1127 struct lustre_qunit *qunit, *tmp;
1128 struct list_head tmp_list;
1132 INIT_LIST_HEAD(&tmp_list);
1134 spin_lock(&qctxt->lqc_lock);
1135 qctxt->lqc_valid = 0;
1136 spin_unlock(&qctxt->lqc_lock);
1138 spin_lock(&qunit_hash_lock);
1139 for (i = 0; i < NR_DQHASH; i++) {
1140 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1141 if (qunit->lq_ctxt != qctxt)
1143 remove_qunit_nolock(qunit);
1144 list_add(&qunit->lq_hash, &tmp_list);
1147 spin_unlock(&qunit_hash_lock);
1149 list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1150 list_del_init(&qunit->lq_hash);
1151 compute_lqs_after_removing_qunit(qunit);
1153 /* wake up all waiters */
1154 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1155 wake_up(&qunit->lq_waitq);
1159 lustre_hash_exit(qctxt->lqc_lqs_hash);
1161 /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1162 * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1163 while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1164 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1165 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1166 cfs_time_seconds(1));
1172 if (lquota_proc_cleanup(qctxt))
1173 CERROR("cleanup proc error!\n");
1179 struct qslave_recov_thread_data {
1180 struct obd_device *obd;
1181 struct lustre_quota_ctxt *qctxt;
1182 struct completion comp;
1185 /* FIXME only recovery block quota by now */
1186 static int qslave_recovery_main(void *arg)
1188 struct qslave_recov_thread_data *data = arg;
1189 struct obd_device *obd = data->obd;
1190 struct lustre_quota_ctxt *qctxt = data->qctxt;
1195 ptlrpc_daemonize("qslave_recovd");
1197 complete(&data->comp);
1199 if (qctxt->lqc_recovery)
1201 qctxt->lqc_recovery = 1;
1203 for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1204 struct qunit_data qdata;
1205 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1206 struct list_head id_list;
1207 struct dquot_id *dqid, *tmp;
1210 LOCK_DQONOFF_MUTEX(dqopt);
1211 if (!sb_has_quota_enabled(qctxt->lqc_sb, type)) {
1212 UNLOCK_DQONOFF_MUTEX(dqopt);
1216 LASSERT(dqopt->files[type] != NULL);
1217 INIT_LIST_HEAD(&id_list);
1218 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1219 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1221 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1223 UNLOCK_DQONOFF_MUTEX(dqopt);
1225 CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1227 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1228 list_del_init(&dqid->di_link);
1229 /* skip slave recovery on itself */
1230 if (is_master(obd, qctxt, dqid->di_id, type))
1232 if (rc && rc != -EBUSY)
1235 qdata.qd_id = dqid->di_id;
1236 qdata.qd_flags = type;
1237 QDATA_SET_BLK(&qdata);
1240 ret = check_cur_qunit(obd, qctxt, &qdata);
1243 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1244 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1253 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
1254 "qslave recovery failed! (id:%d type:%d "
1255 " rc:%d)\n", dqid->di_id, type, rc);
1261 qctxt->lqc_recovery = 0;
1266 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1268 struct qslave_recov_thread_data data;
1272 if (!sb_any_quota_enabled(qctxt->lqc_sb))
1277 init_completion(&data.comp);
1279 rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1281 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1284 wait_for_completion(&data.comp);
1291 * lqs<->qctxt hash operations
1294 /* string hashing using djb2 hash algorithm */
1296 lqs_hash(lustre_hash_t *lh, void *key, unsigned mask)
1298 struct quota_adjust_qunit *lqs_key;
1303 lqs_key = (struct quota_adjust_qunit *)key;
1304 hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1306 RETURN(hash & mask);
1310 lqs_compare(void *key, struct hlist_node *hnode)
1312 struct quota_adjust_qunit *lqs_key;
1313 struct lustre_qunit_size *q;
1318 lqs_key = (struct quota_adjust_qunit *)key;
1319 q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1321 spin_lock(&q->lqs_lock);
1322 rc = ((lqs_key->qaq_id == q->lqs_id) &&
1323 (QAQ_IS_GRP(lqs_key) == LQS_IS_GRP(q)));
1324 spin_unlock(&q->lqs_lock);
1330 lqs_get(struct hlist_node *hnode)
1332 struct lustre_qunit_size *q =
1333 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1336 atomic_inc(&q->lqs_refcount);
1337 CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1338 q, atomic_read(&q->lqs_refcount));
1344 lqs_put(struct hlist_node *hnode)
1346 struct lustre_qunit_size *q =
1347 hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1350 LASSERT(atomic_read(&q->lqs_refcount) > 0);
1351 atomic_dec(&q->lqs_refcount);
1352 CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1353 q, atomic_read(&q->lqs_refcount));
1359 lqs_exit(struct hlist_node *hnode)
1361 struct lustre_qunit_size *q;
1364 q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1366 * Nothing should be left. User of lqs put it and
1367 * lqs also was deleted from table by this time
1368 * so we should have 0 refs.
1370 LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1371 "Busy lqs %p with %d refs\n", q,
1372 atomic_read(&q->lqs_refcount));
1377 static lustre_hash_ops_t lqs_hash_ops = {
1378 .lh_hash = lqs_hash,
1379 .lh_compare = lqs_compare,
1384 #endif /* HAVE_QUOTA_SUPPORT */