lustre/quota/quota_context.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   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.
  11  *
  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).
  17  *
  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
  21  *
  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
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/quota/quota_context.c
  37  *
  38  * Lustre Quota Context
  39  *
  40  * Author: Niu YaWei <niu@clusterfs.com>
  41  */
  42
  43 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46
  47 #define DEBUG_SUBSYSTEM S_LQUOTA
  48
  49 #include <linux/version.h>
  50 #include <linux/fs.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>
  56
  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"
  63
  64 #ifdef HAVE_QUOTA_SUPPORT
  65
  66 static lustre_hash_ops_t lqs_hash_ops;
  67
  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 */
  72
  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;
  76
  77 /* please sync qunit_state with qunit_state_names */
  78 enum qunit_state {
  79         /**
  80          * a qunit is created
  81          */
  82         QUNIT_CREATED      = 0,
  83         /**
  84          * a qunit is added into qunit hash, that means
  85          * a quota req will be sent or is flying
  86          */
  87         QUNIT_IN_HASH      = 1,
  88         /**
  89          * a qunit is removed from qunit hash, that
  90          * means a quota req is handled and comes back
  91          */
  92         QUNIT_RM_FROM_HASH = 2,
  93         /**
  94          * qunit can wake up all threads waiting for it
  95          */
  96         QUNIT_FINISHED     = 3,
  97 };
  98
  99 static const char *qunit_state_names[] = {
 100         [QUNIT_CREATED]      = "CREATED",
 101         [QUNIT_IN_HASH]      = "IN_HASH",
 102         [QUNIT_RM_FROM_HASH] = "RM_FROM_HASH",
 103         [QUNIT_FINISHED]     = "FINISHED",
 104 };
 105
 106 struct lustre_qunit {
 107         struct list_head lq_hash;          /** Hash list in memory */
 108         atomic_t lq_refcnt;                /** Use count */
 109         struct lustre_quota_ctxt *lq_ctxt; /** Quota context this applies to */
 110         struct qunit_data lq_data;         /** See qunit_data */
 111         unsigned int lq_opc;               /** QUOTA_DQACQ, QUOTA_DQREL */
 112         cfs_waitq_t lq_waitq;              /** Threads waiting for this qunit */
 113         spinlock_t lq_lock;                /** Protect the whole structure */
 114         enum qunit_state lq_state;         /** Present the status of qunit */
 115         int lq_rc;                         /** The rc of lq_data */
 116 };
 117
 118 #define QUNIT_SET_STATE(qunit, state)                                   \
 119 do {                                                                    \
 120         spin_lock(&qunit->lq_lock);                                     \
 121         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 122                     "lq_rc(%d)\n",                                      \
 123                     qunit, qunit_state_names[qunit->lq_state],          \
 124                     qunit_state_names[state], qunit->lq_rc);            \
 125         qunit->lq_state = state;                                        \
 126         spin_unlock(&qunit->lq_lock);                                   \
 127 } while(0)
 128
 129 #define QUNIT_SET_STATE_AND_RC(qunit, state, rc)                        \
 130 do {                                                                    \
 131         spin_lock(&qunit->lq_lock);                                     \
 132         qunit->lq_rc = rc;                                              \
 133         QDATA_DEBUG((&qunit->lq_data), "qunit(%p) lq_state(%s->%s), "   \
 134                     "lq_rc(%d)\n",                                      \
 135                     qunit, qunit_state_names[qunit->lq_state],          \
 136                     qunit_state_names[state], qunit->lq_rc);            \
 137         qunit->lq_state = state;                                        \
 138         spin_unlock(&qunit->lq_lock);                                   \
 139 } while(0)
 140
 141
 142 int should_translate_quota (struct obd_import *imp)
 143 {
 144         ENTRY;
 145
 146         LASSERT(imp);
 147         if (imp->imp_connect_data.ocd_connect_flags & OBD_CONNECT_QUOTA64)
 148                 RETURN(0);
 149         else
 150                 RETURN(1);
 151 }
 152
 153 void qunit_cache_cleanup(void)
 154 {
 155         int i;
 156         ENTRY;
 157
 158         spin_lock(&qunit_hash_lock);
 159         for (i = 0; i < NR_DQHASH; i++)
 160                 LASSERT(list_empty(qunit_hash + i));
 161         spin_unlock(&qunit_hash_lock);
 162
 163         if (qunit_cachep) {
 164                 int rc;
 165                 rc = cfs_mem_cache_destroy(qunit_cachep);
 166                 LASSERTF(rc == 0, "couldn't destroy qunit_cache slab\n");
 167                 qunit_cachep = NULL;
 168         }
 169         EXIT;
 170 }
 171
 172 int qunit_cache_init(void)
 173 {
 174         int i;
 175         ENTRY;
 176
 177         LASSERT(qunit_cachep == NULL);
 178         qunit_cachep = cfs_mem_cache_create("ll_qunit_cache",
 179                                             sizeof(struct lustre_qunit),
 180                                             0, 0);
 181         if (!qunit_cachep)
 182                 RETURN(-ENOMEM);
 183
 184         spin_lock(&qunit_hash_lock);
 185         for (i = 0; i < NR_DQHASH; i++)
 186                 CFS_INIT_LIST_HEAD(qunit_hash + i);
 187         spin_unlock(&qunit_hash_lock);
 188         RETURN(0);
 189 }
 190
 191 static inline int
 192 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 193              __attribute__((__const__));
 194
 195 static inline int
 196 qunit_hashfn(struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 197 {
 198         unsigned int id = qdata->qd_id;
 199         unsigned int type = QDATA_IS_GRP(qdata);
 200
 201         unsigned long tmp = ((unsigned long)qctxt >> L1_CACHE_SHIFT) ^ id;
 202         tmp = (tmp * (MAXQUOTAS - type)) % NR_DQHASH;
 203         return tmp;
 204 }
 205
 206 /* caller must hold qunit_hash_lock */
 207 static inline struct lustre_qunit *find_qunit(unsigned int hashent,
 208                                               struct lustre_quota_ctxt *qctxt,
 209                                               struct qunit_data *qdata)
 210 {
 211         struct lustre_qunit *qunit = NULL;
 212         struct qunit_data *tmp;
 213
 214         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 215         list_for_each_entry(qunit, qunit_hash + hashent, lq_hash) {
 216                 tmp = &qunit->lq_data;
 217                 if (qunit->lq_ctxt == qctxt &&
 218                     qdata->qd_id == tmp->qd_id &&
 219                     (qdata->qd_flags & LQUOTA_QUNIT_FLAGS) ==
 220                     (tmp->qd_flags & LQUOTA_QUNIT_FLAGS))
 221                         return qunit;
 222         }
 223         return NULL;
 224 }
 225
 226 /* check_cur_qunit - check the current usage of qunit.
 227  * @qctxt: quota context
 228  * @qdata: the type of quota unit to be checked
 229  *
 230  * return: 1 - need acquire qunit;
 231  *         2 - need release qunit;
 232  *         0 - need do nothing.
 233  *       < 0 - error.
 234  */
 235 static int
 236 check_cur_qunit(struct obd_device *obd,
 237                 struct lustre_quota_ctxt *qctxt, struct qunit_data *qdata)
 238 {
 239         struct super_block *sb = qctxt->lqc_sb;
 240         unsigned long qunit_sz, tune_sz;
 241         __u64 usage, limit, limit_org, pending_write = 0;
 242         long long record = 0;
 243         struct obd_quotactl *qctl;
 244         struct lustre_qunit_size *lqs = NULL;
 245         int ret = 0;
 246         ENTRY;
 247
 248         if (!sb_any_quota_enabled(sb))
 249                 RETURN(0);
 250
 251         spin_lock(&qctxt->lqc_lock);
 252         if (!qctxt->lqc_valid){
 253                 spin_unlock(&qctxt->lqc_lock);
 254                 RETURN(0);
 255         }
 256         spin_unlock(&qctxt->lqc_lock);
 257
 258         OBD_ALLOC_PTR(qctl);
 259         if (qctl == NULL)
 260                 RETURN(-ENOMEM);
 261
 262         /* get fs quota usage & limit */
 263         qctl->qc_cmd = Q_GETQUOTA;
 264         qctl->qc_id = qdata->qd_id;
 265         qctl->qc_type = QDATA_IS_GRP(qdata);
 266         ret = fsfilt_quotactl(obd, sb, qctl);
 267         if (ret) {
 268                 if (ret == -ESRCH)      /* no limit */
 269                         ret = 0;
 270                 else
 271                         CERROR("can't get fs quota usage! (rc:%d)\n", ret);
 272                 GOTO(out, ret);
 273         }
 274
 275         if (QDATA_IS_BLK(qdata)) {
 276                 usage = qctl->qc_dqblk.dqb_curspace;
 277                 limit = qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS;
 278         } else {
 279                 usage = qctl->qc_dqblk.dqb_curinodes;
 280                 limit = qctl->qc_dqblk.dqb_ihardlimit;
 281         }
 282
 283         /* ignore the no quota limit case; and it can avoid creating
 284          * unnecessary lqs for uid/gid */
 285         if (!limit)
 286                 GOTO(out, ret = 0);
 287
 288  search_lqs:
 289         quota_search_lqs(qdata, NULL, qctxt, &lqs);
 290         if (!lqs) {
 291                 CDEBUG(D_QUOTA, "Can't find the lustre qunit size!\n");
 292                 ret = quota_create_lqs(qdata, NULL, qctxt, &lqs);
 293                 if (ret == -EALREADY) {
 294                         ret = 0;
 295                         goto search_lqs;
 296                 }
 297                 if (ret < 0)
 298                         GOTO (out, ret);
 299         }
 300         spin_lock(&lqs->lqs_lock);
 301
 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));
 308         } else {
 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;
 314         }
 315
 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)))
 319                 limit = 0;
 320
 321         usage += pending_write;
 322         limit_org = limit;
 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)
 326                 usage -= record;
 327         else
 328                 limit += record;
 329
 330         LASSERT(qdata->qd_count == 0);
 331         if (limit <= usage + tune_sz) {
 332                 while (qdata->qd_count + limit <=
 333                        usage + tune_sz)
 334                         qdata->qd_count += qunit_sz;
 335                 ret = 1;
 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;
 341                 ret = 2;
 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 (ret == 2 && pending_write) {
 345                         CDEBUG(D_QUOTA, "delay quota release\n");
 346                         ret = 0;
 347                 }
 348         }
 349         CDEBUG(D_QUOTA, "type: %c, limit: "LPU64", usage: "LPU64
 350                ", pending_write: "LPU64", record: "LPD64
 351                ", qunit_sz: %lu, tune_sz: %lu, ret: %d.\n",
 352                QDATA_IS_BLK(qdata) ? 'b' : 'i', limit, usage, pending_write,
 353                record, qunit_sz, tune_sz, ret);
 354         LASSERT(ret == 0 || qdata->qd_count);
 355
 356         spin_unlock(&lqs->lqs_lock);
 357         lqs_putref(lqs);
 358         EXIT;
 359  out:
 360         OBD_FREE_PTR(qctl);
 361         return ret;
 362 }
 363
 364 /**
 365  * Compute the remaining quota for certain gid or uid b=11693
 366  */
 367 int compute_remquota(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 368                      struct qunit_data *qdata, int isblk)
 369 {
 370         struct super_block *sb = qctxt->lqc_sb;
 371         __u64 usage, limit;
 372         struct obd_quotactl *qctl;
 373         int ret = QUOTA_RET_OK;
 374         ENTRY;
 375
 376         if (!sb_any_quota_enabled(sb))
 377                 RETURN(QUOTA_RET_NOQUOTA);
 378
 379         /* ignore root user */
 380         if (qdata->qd_id == 0 && QDATA_IS_GRP(qdata) == USRQUOTA)
 381                 RETURN(QUOTA_RET_NOLIMIT);
 382
 383         OBD_ALLOC_PTR(qctl);
 384         if (qctl == NULL)
 385                 RETURN(-ENOMEM);
 386
 387         /* get fs quota usage & limit */
 388         qctl->qc_cmd = Q_GETQUOTA;
 389         qctl->qc_id = qdata->qd_id;
 390         qctl->qc_type = QDATA_IS_GRP(qdata);
 391         ret = fsfilt_quotactl(obd, sb, qctl);
 392         if (ret) {
 393                 if (ret == -ESRCH)      /* no limit */
 394                         ret = QUOTA_RET_NOLIMIT;
 395                 else
 396                         CDEBUG(D_QUOTA, "can't get fs quota usage! (rc:%d)",
 397                                ret);
 398                 GOTO(out, ret);
 399         }
 400
 401         usage = isblk ? qctl->qc_dqblk.dqb_curspace :
 402                 qctl->qc_dqblk.dqb_curinodes;
 403         limit = isblk ? qctl->qc_dqblk.dqb_bhardlimit << QUOTABLOCK_BITS :
 404                 qctl->qc_dqblk.dqb_ihardlimit;
 405         if (!limit){            /* no limit */
 406                 ret = QUOTA_RET_NOLIMIT;
 407                 GOTO(out, ret);
 408         }
 409
 410         if (limit >= usage)
 411                 qdata->qd_count = limit - usage;
 412         else
 413                 qdata->qd_count = 0;
 414         EXIT;
 415 out:
 416         OBD_FREE_PTR(qctl);
 417         return ret;
 418 }
 419
 420 static struct lustre_qunit *alloc_qunit(struct lustre_quota_ctxt *qctxt,
 421                                         struct qunit_data *qdata, int opc)
 422 {
 423         struct lustre_qunit *qunit = NULL;
 424         ENTRY;
 425
 426         OBD_SLAB_ALLOC_PTR_GFP(qunit, qunit_cachep, CFS_ALLOC_IO);
 427         if (qunit == NULL)
 428                 RETURN(NULL);
 429
 430         CFS_INIT_LIST_HEAD(&qunit->lq_hash);
 431         init_waitqueue_head(&qunit->lq_waitq);
 432         atomic_set(&qunit->lq_refcnt, 1);
 433         qunit->lq_ctxt = qctxt;
 434         memcpy(&qunit->lq_data, qdata, sizeof(*qdata));
 435         qunit->lq_opc = opc;
 436         qunit->lq_lock = SPIN_LOCK_UNLOCKED;
 437         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_CREATED, 0);
 438         RETURN(qunit);
 439 }
 440
 441 static inline void free_qunit(struct lustre_qunit *qunit)
 442 {
 443         OBD_SLAB_FREE(qunit, qunit_cachep, sizeof(*qunit));
 444 }
 445
 446 static inline void qunit_get(struct lustre_qunit *qunit)
 447 {
 448         atomic_inc(&qunit->lq_refcnt);
 449 }
 450
 451 static void qunit_put(struct lustre_qunit *qunit)
 452 {
 453         LASSERT(atomic_read(&qunit->lq_refcnt));
 454         if (atomic_dec_and_test(&qunit->lq_refcnt))
 455                 free_qunit(qunit);
 456 }
 457
 458 /* caller must hold qunit_hash_lock and release ref of qunit after using it */
 459 static struct lustre_qunit *dqacq_in_flight(struct lustre_quota_ctxt *qctxt,
 460                                             struct qunit_data *qdata)
 461 {
 462         unsigned int hashent = qunit_hashfn(qctxt, qdata);
 463         struct lustre_qunit *qunit;
 464         ENTRY;
 465
 466         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 467         qunit = find_qunit(hashent, qctxt, qdata);
 468         if (qunit)
 469                 qunit_get(qunit);
 470         RETURN(qunit);
 471 }
 472
 473 static void
 474 insert_qunit_nolock(struct lustre_quota_ctxt *qctxt, struct lustre_qunit *qunit)
 475 {
 476         struct list_head *head;
 477
 478         LASSERT(list_empty(&qunit->lq_hash));
 479         qunit_get(qunit);
 480         head = qunit_hash + qunit_hashfn(qctxt, &qunit->lq_data);
 481         list_add(&qunit->lq_hash, head);
 482         QUNIT_SET_STATE(qunit, QUNIT_IN_HASH);
 483 }
 484
 485 static void compute_lqs_after_removing_qunit(struct lustre_qunit *qunit)
 486 {
 487         struct lustre_qunit_size *lqs = NULL;
 488
 489         quota_search_lqs(&qunit->lq_data, NULL, qunit->lq_ctxt, &lqs);
 490         if (lqs) {
 491                 spin_lock(&lqs->lqs_lock);
 492                 if (qunit->lq_opc == QUOTA_DQACQ)
 493                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 1);
 494                 if (qunit->lq_opc == QUOTA_DQREL)
 495                         quota_compute_lqs(&qunit->lq_data, lqs, 0, 0);
 496                 spin_unlock(&lqs->lqs_lock);
 497                 /* this is for quota_search_lqs */
 498                 lqs_putref(lqs);
 499                 /* this is for schedule_dqacq */
 500                 lqs_putref(lqs);
 501         }
 502
 503 }
 504
 505 static void remove_qunit_nolock(struct lustre_qunit *qunit)
 506 {
 507         LASSERT(!list_empty(&qunit->lq_hash));
 508         LASSERT_SPIN_LOCKED(&qunit_hash_lock);
 509
 510         list_del_init(&qunit->lq_hash);
 511         QUNIT_SET_STATE(qunit, QUNIT_RM_FROM_HASH);
 512         qunit_put(qunit);
 513 }
 514
 515 #define INC_QLIMIT(limit, count) (limit == MIN_QLIMIT) ? \
 516                                  (limit = count) : (limit += count)
 517
 518
 519 static inline int is_master(struct lustre_quota_ctxt *qctxt)
 520 {
 521         return qctxt->lqc_handler ? 1 : 0;
 522 }
 523
 524 static int
 525 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 526                struct qunit_data *qdata, int opc, int wait,
 527                struct obd_trans_info *oti);
 528
 529 static int
 530 dqacq_completion(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 531                  struct qunit_data *qdata, int rc, int opc)
 532 {
 533         struct lustre_qunit *qunit = NULL;
 534         struct super_block *sb = qctxt->lqc_sb;
 535         int err = 0;
 536         struct quota_adjust_qunit *oqaq = NULL;
 537         int rc1 = 0;
 538         ENTRY;
 539
 540         LASSERT(qdata);
 541         QDATA_DEBUG(qdata, "obd(%s): complete %s quota req\n",
 542                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 543
 544         /* do it only when a releasing quota req more than 5MB b=18491 */
 545         if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880)
 546                 OBD_FAIL_TIMEOUT(OBD_FAIL_QUOTA_DELAY_REL, 5);
 547
 548         /* update local operational quota file */
 549         if (rc == 0) {
 550                 __u64 count = QUSG(qdata->qd_count, QDATA_IS_BLK(qdata));
 551                 struct obd_quotactl *qctl;
 552                 __u64 *hardlimit;
 553
 554                 OBD_ALLOC_PTR(qctl);
 555                 if (qctl == NULL)
 556                         GOTO(out, err = -ENOMEM);
 557
 558                 /* acq/rel qunit for specified uid/gid is serialized,
 559                  * so there is no race between get fs quota limit and
 560                  * set fs quota limit */
 561                 qctl->qc_cmd = Q_GETQUOTA;
 562                 qctl->qc_id = qdata->qd_id;
 563                 qctl->qc_type = QDATA_IS_GRP(qdata);
 564                 err = fsfilt_quotactl(obd, sb, qctl);
 565                 if (err) {
 566                         CERROR("error get quota fs limit! (rc:%d)\n", err);
 567                         GOTO(out_mem, err);
 568                 }
 569
 570                 if (QDATA_IS_BLK(qdata)) {
 571                         qctl->qc_dqblk.dqb_valid = QIF_BLIMITS;
 572                         hardlimit = &qctl->qc_dqblk.dqb_bhardlimit;
 573                 } else {
 574                         qctl->qc_dqblk.dqb_valid = QIF_ILIMITS;
 575                         hardlimit = &qctl->qc_dqblk.dqb_ihardlimit;
 576                 }
 577
 578                 CDEBUG(D_QUOTA, "hardlimt: "LPU64"\n", *hardlimit);
 579
 580                 if (*hardlimit == 0)
 581                         goto out_mem;
 582
 583                 switch (opc) {
 584                 case QUOTA_DQACQ:
 585                         INC_QLIMIT(*hardlimit, count);
 586                         break;
 587                 case QUOTA_DQREL:
 588                         LASSERTF(count < *hardlimit,
 589                                  "id(%u) flag(%u) type(%c) isblk(%c) "
 590                                  "count("LPU64") qd_qunit("LPU64") "
 591                                  "hardlimit("LPU64").\n",
 592                                  qdata->qd_id, qdata->qd_flags,
 593                                  QDATA_IS_GRP(qdata) ? 'g' : 'u',
 594                                  QDATA_IS_BLK(qdata) ? 'b': 'i',
 595                                  qdata->qd_count, qdata->qd_qunit, *hardlimit);
 596                         *hardlimit -= count;
 597                         break;
 598                 default:
 599                         LBUG();
 600                 }
 601
 602                 /* clear quota limit */
 603                 if (count == 0)
 604                         *hardlimit = 0;
 605
 606                 qctl->qc_cmd = Q_SETQUOTA;
 607                 err = fsfilt_quotactl(obd, sb, qctl);
 608                 if (err)
 609                         CERROR("error set quota fs limit! (rc:%d)\n", err);
 610
 611                 QDATA_DEBUG(qdata, "%s completion\n",
 612                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 613 out_mem:
 614                 OBD_FREE_PTR(qctl);
 615         } else if (rc == -EDQUOT) {
 616                 QDATA_DEBUG(qdata, "acquire qunit got EDQUOT.\n");
 617         } else if (rc == -EBUSY) {
 618                 QDATA_DEBUG(qdata, "it's is recovering, got EBUSY.\n");
 619         } else {
 620                 CERROR("acquire qunit got error! (rc:%d)\n", rc);
 621         }
 622 out:
 623         /* remove the qunit from hash */
 624         spin_lock(&qunit_hash_lock);
 625
 626         qunit = dqacq_in_flight(qctxt, qdata);
 627         /* this qunit has been removed by qctxt_cleanup() */
 628         if (!qunit) {
 629                 spin_unlock(&qunit_hash_lock);
 630                 QDATA_DEBUG(qdata, "%s is discarded because qunit isn't found\n",
 631                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 632                 RETURN(err);
 633         }
 634
 635         LASSERT(opc == qunit->lq_opc);
 636         /* remove this qunit from lq_hash so that new processes cannot be added
 637          * to qunit->lq_waiters */
 638         remove_qunit_nolock(qunit);
 639         spin_unlock(&qunit_hash_lock);
 640
 641         compute_lqs_after_removing_qunit(qunit);
 642
 643         if (rc == 0)
 644                 rc = QUOTA_REQ_RETURNED;
 645         QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, rc);
 646         /* wake up all waiters */
 647         wake_up_all(&qunit->lq_waitq);
 648
 649         /* this is for dqacq_in_flight() */
 650         qunit_put(qunit);
 651         /* this is for alloc_qunit() */
 652         qunit_put(qunit);
 653         if (rc < 0 && rc != -EDQUOT)
 654                  RETURN(err);
 655
 656         /* don't reschedule in such cases:
 657          *   - acq/rel failure and qunit isn't changed,
 658          *     but not for quota recovery.
 659          *   - local dqacq/dqrel.
 660          *   - local disk io failure.
 661          */
 662          OBD_ALLOC_PTR(oqaq);
 663          if (!oqaq)
 664                  RETURN(-ENOMEM);
 665          qdata_to_oqaq(qdata, oqaq);
 666          /* adjust the qunit size in slaves */
 667          rc1 = quota_adjust_slave_lqs(oqaq, qctxt);
 668          OBD_FREE_PTR(oqaq);
 669          if (rc1 < 0) {
 670                  CERROR("adjust slave's qunit size failed!(rc:%d)\n", rc1);
 671                  RETURN(rc1);
 672          }
 673          if (err || (rc < 0 && rc != -EBUSY && rc1 == 0) || is_master(qctxt))
 674                  RETURN(err);
 675
 676          if (opc == QUOTA_DQREL && qdata->qd_count >= 5242880 &&
 677              OBD_FAIL_CHECK(OBD_FAIL_QUOTA_DELAY_REL))
 678                  RETURN(err);
 679
 680         /* reschedule another dqacq/dqrel if needed */
 681         qdata->qd_count = 0;
 682         qdata->qd_flags &= LQUOTA_QUNIT_FLAGS;
 683         rc1 = check_cur_qunit(obd, qctxt, qdata);
 684         if (rc1 > 0) {
 685                 int opc;
 686                 opc = rc1 == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
 687                 rc1 = schedule_dqacq(obd, qctxt, qdata, opc, 0, NULL);
 688                 QDATA_DEBUG(qdata, "reschedudle opc(%d) rc(%d)\n", opc, rc1);
 689         }
 690         RETURN(err);
 691 }
 692
 693 struct dqacq_async_args {
 694         struct lustre_quota_ctxt *aa_ctxt;
 695         struct lustre_qunit *aa_qunit;
 696 };
 697
 698 static int dqacq_interpret(const struct lu_env *env,
 699                            struct ptlrpc_request *req, void *data, int rc)
 700 {
 701         struct dqacq_async_args *aa = (struct dqacq_async_args *)data;
 702         struct lustre_quota_ctxt *qctxt = aa->aa_ctxt;
 703         struct obd_device_target *obt = qctxt->lqc_obt;
 704         struct lustre_qunit *qunit = aa->aa_qunit;
 705         struct obd_device *obd = req->rq_import->imp_obd;
 706         struct qunit_data *qdata = NULL;
 707         int rc1 = 0;
 708         ENTRY;
 709
 710         LASSERT(req);
 711         LASSERT(req->rq_import);
 712
 713         /* there are several forms of qunit(historic causes), so we need to
 714          * adjust qunit from slaves to the same form here */
 715         OBD_ALLOC(qdata, sizeof(struct qunit_data));
 716         if (!qdata)
 717                 RETURN(-ENOMEM);
 718
 719         down_read(&obt->obt_rwsem);
 720         /* if a quota req timeouts or is dropped, we should update quota
 721          * statistics which will be handled in dqacq_completion. And in
 722          * this situation we should get qdata from request instead of
 723          * reply */
 724         rc1 = quota_get_qdata(req, qdata,
 725                               (rc != 0) ? QUOTA_REQUEST : QUOTA_REPLY,
 726                               QUOTA_IMPORT);
 727         if (rc1 < 0) {
 728                 DEBUG_REQ(D_ERROR, req,
 729                           "error unpacking qunit_data(rc: %d)\n", rc1);
 730                 GOTO(exit, rc = rc1);
 731         }
 732
 733         QDATA_DEBUG(qdata, "qdata: interpret rc(%d).\n", rc);
 734         QDATA_DEBUG((&qunit->lq_data), "lq_data: \n");
 735
 736         if (qdata->qd_id != qunit->lq_data.qd_id ||
 737             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_RET_QDATA)) {
 738                 CDEBUG(D_ERROR, "the returned qd_id isn't expected!"
 739                        "(qdata: %u, lq_data: %u)\n", qdata->qd_id,
 740                        qunit->lq_data.qd_id);
 741                 qdata->qd_id = qunit->lq_data.qd_id;
 742                 rc = -EPROTO;
 743         }
 744         if (QDATA_IS_GRP(qdata) != QDATA_IS_GRP(&qunit->lq_data)) {
 745                 CDEBUG(D_ERROR, "the returned grp/usr isn't expected!"
 746                        "(qdata: %u, lq_data: %u)\n", qdata->qd_flags,
 747                        qunit->lq_data.qd_flags);
 748                 if (QDATA_IS_GRP(&qunit->lq_data))
 749                         QDATA_SET_GRP(qdata);
 750                 else
 751                         QDATA_CLR_GRP(qdata);
 752                 rc = -EPROTO;
 753         }
 754         if (qdata->qd_count > qunit->lq_data.qd_count) {
 755                 CDEBUG(D_ERROR, "the returned qd_count isn't expected!"
 756                        "(qdata: "LPU64", lq_data: "LPU64")\n", qdata->qd_count,
 757                        qunit->lq_data.qd_count);
 758                 rc = -EPROTO;
 759         }
 760
 761         rc = dqacq_completion(obd, qctxt, qdata, rc,
 762                               lustre_msg_get_opc(req->rq_reqmsg));
 763
 764 exit:
 765         up_read(&obt->obt_rwsem);
 766         OBD_FREE(qdata, sizeof(struct qunit_data));
 767
 768         RETURN(rc);
 769 }
 770
 771 /**
 772  * check if quota master is online
 773  */
 774 int check_qm(struct lustre_quota_ctxt *qctxt)
 775 {
 776         int rc;
 777         ENTRY;
 778
 779         spin_lock(&qctxt->lqc_lock);
 780         /* quit waiting when mds is back or qctxt is cleaned up */
 781         rc = qctxt->lqc_import || !qctxt->lqc_valid;
 782         spin_unlock(&qctxt->lqc_lock);
 783
 784         RETURN(rc);
 785 }
 786
 787 /* wake up all waiting threads when lqc_import is NULL */
 788 void dqacq_interrupt(struct lustre_quota_ctxt *qctxt)
 789 {
 790         struct lustre_qunit *qunit, *tmp;
 791         int i;
 792         ENTRY;
 793
 794         spin_lock(&qunit_hash_lock);
 795         for (i = 0; i < NR_DQHASH; i++) {
 796                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
 797                         if (qunit->lq_ctxt != qctxt)
 798                                 continue;
 799
 800                         /* Wake up all waiters. Do not change lq_state.
 801                          * The waiters will check lq_rc which is kept as 0
 802                          * if no others change it, then the waiters will return
 803                          * -EAGAIN to caller who can perform related quota
 804                          * acq/rel if necessary. */
 805                         wake_up_all(&qunit->lq_waitq);
 806                 }
 807         }
 808         spin_unlock(&qunit_hash_lock);
 809         EXIT;
 810 }
 811
 812 static int got_qunit(struct lustre_qunit *qunit)
 813 {
 814         struct lustre_quota_ctxt *qctxt = qunit->lq_ctxt;
 815         int rc = 0;
 816         ENTRY;
 817
 818         spin_lock(&qunit->lq_lock);
 819         switch (qunit->lq_state) {
 820         case QUNIT_IN_HASH:
 821         case QUNIT_RM_FROM_HASH:
 822                 break;
 823         case QUNIT_FINISHED:
 824                 rc = 1;
 825                 break;
 826         default:
 827                 CERROR("invalid qunit state %d\n", qunit->lq_state);
 828         }
 829         spin_unlock(&qunit->lq_lock);
 830
 831         if (!rc) {
 832                 spin_lock(&qctxt->lqc_lock);
 833                 rc = !qctxt->lqc_import || !qctxt->lqc_valid;
 834                 spin_unlock(&qctxt->lqc_lock);
 835         }
 836
 837         RETURN(rc);
 838 }
 839
 840 static int
 841 schedule_dqacq(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
 842                struct qunit_data *qdata, int opc, int wait,
 843                struct obd_trans_info *oti)
 844 {
 845         struct lustre_qunit *qunit, *empty;
 846         struct l_wait_info lwi = { 0 };
 847         struct ptlrpc_request *req;
 848         struct dqacq_async_args *aa;
 849         struct obd_import *imp = NULL;
 850         struct lustre_qunit_size *lqs = NULL;
 851         struct timeval work_start;
 852         struct timeval work_end;
 853         long timediff;
 854         int rc = 0;
 855         ENTRY;
 856
 857         LASSERT(opc == QUOTA_DQACQ || opc == QUOTA_DQREL);
 858         do_gettimeofday(&work_start);
 859         if ((empty = alloc_qunit(qctxt, qdata, opc)) == NULL)
 860                 RETURN(-ENOMEM);
 861
 862         spin_lock(&qunit_hash_lock);
 863         qunit = dqacq_in_flight(qctxt, qdata);
 864         if (qunit) {
 865                 spin_unlock(&qunit_hash_lock);
 866                 qunit_put(empty);
 867
 868                 goto wait_completion;
 869         }
 870         qunit = empty;
 871         qunit_get(qunit);
 872         insert_qunit_nolock(qctxt, qunit);
 873         spin_unlock(&qunit_hash_lock);
 874
 875         quota_search_lqs(qdata, NULL, qctxt, &lqs);
 876         if (lqs) {
 877                 spin_lock(&lqs->lqs_lock);
 878                 quota_compute_lqs(qdata, lqs, 1, (opc == QUOTA_DQACQ) ? 1 : 0);
 879                 /* when this qdata returned from mds, it will call lqs_putref */
 880                 lqs_getref(lqs);
 881                 spin_unlock(&lqs->lqs_lock);
 882                 /* this is for quota_search_lqs */
 883                 lqs_putref(lqs);
 884         } else {
 885                 CDEBUG(D_ERROR, "Can't find the lustre qunit size!\n");
 886         }
 887
 888         QDATA_DEBUG(qdata, "obd(%s): send %s quota req\n",
 889                     obd->obd_name, (opc == QUOTA_DQACQ) ? "acq" : "rel");
 890         /* master is going to dqacq/dqrel from itself */
 891         if (is_master(qctxt)) {
 892                 int rc2;
 893                 QDATA_DEBUG(qdata, "local %s.\n",
 894                             opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 895                 QDATA_SET_CHANGE_QS(qdata);
 896                 rc = qctxt->lqc_handler(obd, qdata, opc);
 897                 rc2 = dqacq_completion(obd, qctxt, qdata, rc, opc);
 898                 /* this is for qunit_get() */
 899                 qunit_put(qunit);
 900
 901                 do_gettimeofday(&work_end);
 902                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
 903                 if (opc == QUOTA_DQACQ)
 904                         lprocfs_counter_add(qctxt->lqc_stats,
 905                                             wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
 906                                             timediff);
 907                 else
 908                         lprocfs_counter_add(qctxt->lqc_stats,
 909                                             wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
 910                                             timediff);
 911                 RETURN(rc ? rc : rc2);
 912         }
 913
 914         spin_lock(&qctxt->lqc_lock);
 915         if (!qctxt->lqc_import) {
 916                 spin_unlock(&qctxt->lqc_lock);
 917                 QDATA_DEBUG(qdata, "lqc_import is invalid.\n");
 918
 919                 spin_lock(&qunit_hash_lock);
 920                 remove_qunit_nolock(qunit);
 921                 spin_unlock(&qunit_hash_lock);
 922
 923                 compute_lqs_after_removing_qunit(qunit);
 924
 925                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, -EAGAIN);
 926                 wake_up_all(&qunit->lq_waitq);
 927
 928                 /* this is for qunit_get() */
 929                 qunit_put(qunit);
 930                 /* this for alloc_qunit() */
 931                 qunit_put(qunit);
 932                 spin_lock(&qctxt->lqc_lock);
 933                 if (wait && !qctxt->lqc_import) {
 934                         spin_unlock(&qctxt->lqc_lock);
 935
 936                         LASSERT(oti && oti->oti_thread &&
 937                                 oti->oti_thread->t_watchdog);
 938
 939                         lc_watchdog_disable(oti->oti_thread->t_watchdog);
 940                         CDEBUG(D_QUOTA, "sleep for quota master\n");
 941                         l_wait_event(qctxt->lqc_wait_for_qmaster,
 942                                      check_qm(qctxt), &lwi);
 943                         CDEBUG(D_QUOTA, "wake up when quota master is back\n");
 944                         lc_watchdog_touch(oti->oti_thread->t_watchdog);
 945                 } else {
 946                         spin_unlock(&qctxt->lqc_lock);
 947                 }
 948
 949                 RETURN(-EAGAIN);
 950         }
 951         imp = class_import_get(qctxt->lqc_import);
 952         spin_unlock(&qctxt->lqc_lock);
 953
 954         /* build dqacq/dqrel request */
 955         LASSERT(imp);
 956
 957         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_QUOTA_DQACQ,
 958                                         LUSTRE_MDS_VERSION, opc);
 959         class_import_put(imp);
 960         if (req == NULL) {
 961                 CDEBUG(D_ERROR, "Can't alloc request\n");
 962                 dqacq_completion(obd, qctxt, qdata, -ENOMEM, opc);
 963                 /* this is for qunit_get() */
 964                 qunit_put(qunit);
 965                 RETURN(-ENOMEM);
 966         }
 967
 968         ptlrpc_request_set_replen(req);
 969         req->rq_no_resend = req->rq_no_delay = 1;
 970         rc = quota_copy_qdata(req, qdata, QUOTA_REQUEST, QUOTA_IMPORT);
 971         if (rc < 0) {
 972                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
 973                 ptlrpc_req_finished(req);
 974                 dqacq_completion(obd, qctxt, qdata, -EPROTO, opc);
 975                 /* this is for qunit_get() */
 976                 qunit_put(qunit);
 977                 RETURN(rc);
 978         }
 979
 980         CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
 981         aa = ptlrpc_req_async_args(req);
 982         aa->aa_ctxt = qctxt;
 983         aa->aa_qunit = qunit;
 984
 985         req->rq_interpret_reply = dqacq_interpret;
 986         ptlrpcd_add_req(req, PSCOPE_OTHER);
 987
 988         QDATA_DEBUG(qdata, "%s scheduled.\n",
 989                     opc == QUOTA_DQACQ ? "DQACQ" : "DQREL");
 990 wait_completion:
 991         if (wait && qunit) {
 992                 struct qunit_data *p = &qunit->lq_data;
 993
 994                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
 995                 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
 996                 /* rc = -EAGAIN, it means the quota master isn't ready yet
 997                  * rc = QUOTA_REQ_RETURNED, it means a quota req is finished;
 998                  * rc = -EDQUOT, it means out of quota
 999                  * rc = -EBUSY, it means recovery is happening
1000                  * other rc < 0, it means real errors, functions who call
1001                  * schedule_dqacq should take care of this */
1002                 spin_lock(&qunit->lq_lock);
1003                 rc = qunit->lq_rc;
1004                 spin_unlock(&qunit->lq_lock);
1005                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1006                        qunit, rc);
1007         }
1008
1009         qunit_put(qunit);
1010         do_gettimeofday(&work_end);
1011         timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1012         if (opc == QUOTA_DQACQ)
1013                 lprocfs_counter_add(qctxt->lqc_stats,
1014                                     wait ? LQUOTA_SYNC_ACQ : LQUOTA_ASYNC_ACQ,
1015                                     timediff);
1016         else
1017                 lprocfs_counter_add(qctxt->lqc_stats,
1018                                     wait ? LQUOTA_SYNC_REL : LQUOTA_ASYNC_REL,
1019                                     timediff);
1020
1021         RETURN(rc);
1022 }
1023
1024 int
1025 qctxt_adjust_qunit(struct obd_device *obd, struct lustre_quota_ctxt *qctxt,
1026                    uid_t uid, gid_t gid, __u32 isblk, int wait,
1027                    struct obd_trans_info *oti)
1028 {
1029         int rc = 0, i = USRQUOTA;
1030         __u32 id[MAXQUOTAS] = { uid, gid };
1031         struct qunit_data qdata[MAXQUOTAS];
1032         ENTRY;
1033
1034         CLASSERT(MAXQUOTAS < 4);
1035         if (!sb_any_quota_enabled(qctxt->lqc_sb))
1036                 RETURN(0);
1037
1038         for (i = 0; i < MAXQUOTAS; i++) {
1039                 qdata[i].qd_id = id[i];
1040                 qdata[i].qd_flags = i;
1041                 if (isblk)
1042                         QDATA_SET_BLK(&qdata[i]);
1043                 qdata[i].qd_count = 0;
1044
1045                 rc = check_cur_qunit(obd, qctxt, &qdata[i]);
1046                 if (rc > 0) {
1047                         int opc;
1048                         /* need acquire or release */
1049                         opc = rc == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1050                         rc = schedule_dqacq(obd, qctxt, &qdata[i], opc,
1051                                             wait,oti);
1052                         if (rc < 0)
1053                                 RETURN(rc);
1054                 } else if (wait == 1) {
1055                         /* when wait equates 1, that means mds_quota_acquire
1056                          * or filter_quota_acquire is calling it. */
1057                         rc = qctxt_wait_pending_dqacq(qctxt, id[i], i, isblk);
1058                         if (rc < 0)
1059                                 RETURN(rc);
1060                 }
1061         }
1062
1063         RETURN(rc);
1064 }
1065
1066 int
1067 qctxt_wait_pending_dqacq(struct lustre_quota_ctxt *qctxt, unsigned int id,
1068                          unsigned short type, int isblk)
1069 {
1070         struct lustre_qunit *qunit = NULL;
1071         struct qunit_data qdata;
1072         struct timeval work_start;
1073         struct timeval work_end;
1074         long timediff;
1075         struct l_wait_info lwi = { 0 };
1076         int rc = 0;
1077         ENTRY;
1078
1079         do_gettimeofday(&work_start);
1080         qdata.qd_id = id;
1081         qdata.qd_flags = type;
1082         if (isblk)
1083                 QDATA_SET_BLK(&qdata);
1084         qdata.qd_count = 0;
1085
1086         spin_lock(&qunit_hash_lock);
1087         qunit = dqacq_in_flight(qctxt, &qdata);
1088         spin_unlock(&qunit_hash_lock);
1089
1090         if (qunit) {
1091                 struct qunit_data *p = &qunit->lq_data;
1092
1093                 QDATA_DEBUG(p, "qunit(%p) is waiting for dqacq.\n", qunit);
1094                 l_wait_event(qunit->lq_waitq, got_qunit(qunit), &lwi);
1095                 CDEBUG(D_QUOTA, "qunit(%p) finishes waiting. (rc:%d)\n",
1096                        qunit, qunit->lq_rc);
1097                 /* keep same as schedule_dqacq() b=17030 */
1098                 spin_lock(&qunit->lq_lock);
1099                 rc = qunit->lq_rc;
1100                 spin_unlock(&qunit->lq_lock);
1101                 /* this is for dqacq_in_flight() */
1102                 qunit_put(qunit);
1103                 do_gettimeofday(&work_end);
1104                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1105                 lprocfs_counter_add(qctxt->lqc_stats,
1106                                     isblk ? LQUOTA_WAIT_PENDING_BLK_QUOTA :
1107                                             LQUOTA_WAIT_PENDING_INO_QUOTA,
1108                                     timediff);
1109         } else {
1110                 do_gettimeofday(&work_end);
1111                 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1112                 lprocfs_counter_add(qctxt->lqc_stats,
1113                                     isblk ? LQUOTA_NOWAIT_PENDING_BLK_QUOTA :
1114                                             LQUOTA_NOWAIT_PENDING_INO_QUOTA,
1115                                     timediff);
1116         }
1117
1118         RETURN(rc);
1119 }
1120
1121 int
1122 qctxt_init(struct obd_device *obd, dqacq_handler_t handler)
1123 {
1124         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
1125         struct obd_device_target *obt = &obd->u.obt;
1126         struct super_block *sb = obt->obt_sb;
1127         int rc = 0;
1128         ENTRY;
1129
1130         LASSERT(qctxt);
1131
1132         rc = ptlrpcd_addref();
1133         if (rc)
1134                 RETURN(rc);
1135
1136         cfs_waitq_init(&qctxt->lqc_wait_for_qmaster);
1137         spin_lock_init(&qctxt->lqc_lock);
1138         spin_lock(&qctxt->lqc_lock);
1139         qctxt->lqc_handler = handler;
1140         qctxt->lqc_sb = sb;
1141         qctxt->lqc_obt = obt;
1142         qctxt->lqc_import = NULL;
1143         qctxt->lqc_recovery = 0;
1144         qctxt->lqc_switch_qs = 1; /* Change qunit size in default setting */
1145         qctxt->lqc_valid = 1;
1146         qctxt->lqc_cqs_boundary_factor = 4;
1147         qctxt->lqc_cqs_least_bunit = PTLRPC_MAX_BRW_SIZE;
1148         qctxt->lqc_cqs_least_iunit = 2;
1149         qctxt->lqc_cqs_qs_factor = 2;
1150         qctxt->lqc_flags = 0;
1151         QUOTA_MASTER_UNREADY(qctxt);
1152         qctxt->lqc_bunit_sz = default_bunit_sz;
1153         qctxt->lqc_btune_sz = default_bunit_sz / 100 * default_btune_ratio;
1154         qctxt->lqc_iunit_sz = default_iunit_sz;
1155         qctxt->lqc_itune_sz = default_iunit_sz * default_itune_ratio / 100;
1156         qctxt->lqc_switch_seconds = 300; /* enlarging will wait 5 minutes
1157                                           * after the last shrinking */
1158         qctxt->lqc_sync_blk = 0;
1159         spin_unlock(&qctxt->lqc_lock);
1160
1161         qctxt->lqc_lqs_hash = lustre_hash_init("LQS_HASH", 7, 7,
1162                                                &lqs_hash_ops, 0);
1163         if (!qctxt->lqc_lqs_hash) {
1164                 CERROR("initialize hash lqs for %s error!\n", obd->obd_name);
1165                 RETURN(-ENOMEM);
1166         }
1167
1168 #ifdef LPROCFS
1169         rc = lquota_proc_setup(obd, is_master(qctxt));
1170         if (rc)
1171                 CERROR("initialize proc for %s error!\n", obd->obd_name);
1172 #endif
1173
1174         RETURN(rc);
1175 }
1176
1177 void qctxt_cleanup(struct lustre_quota_ctxt *qctxt, int force)
1178 {
1179         struct lustre_qunit *qunit, *tmp;
1180         struct list_head tmp_list;
1181         struct obd_device_target *obt = qctxt->lqc_obt;
1182         int i;
1183         ENTRY;
1184
1185         CFS_INIT_LIST_HEAD(&tmp_list);
1186
1187         spin_lock(&qctxt->lqc_lock);
1188         qctxt->lqc_valid = 0;
1189         spin_unlock(&qctxt->lqc_lock);
1190
1191         spin_lock(&qunit_hash_lock);
1192         for (i = 0; i < NR_DQHASH; i++) {
1193                 list_for_each_entry_safe(qunit, tmp, &qunit_hash[i], lq_hash) {
1194                         if (qunit->lq_ctxt != qctxt)
1195                                 continue;
1196                         remove_qunit_nolock(qunit);
1197                         list_add(&qunit->lq_hash, &tmp_list);
1198                 }
1199         }
1200         spin_unlock(&qunit_hash_lock);
1201
1202         list_for_each_entry_safe(qunit, tmp, &tmp_list, lq_hash) {
1203                 list_del_init(&qunit->lq_hash);
1204                 compute_lqs_after_removing_qunit(qunit);
1205
1206                 /* wake up all waiters */
1207                 QUNIT_SET_STATE_AND_RC(qunit, QUNIT_FINISHED, 0);
1208                 wake_up_all(&qunit->lq_waitq);
1209                 qunit_put(qunit);
1210         }
1211
1212         down_write(&obt->obt_rwsem);
1213         lustre_hash_exit(qctxt->lqc_lqs_hash);
1214         qctxt->lqc_lqs_hash = NULL;
1215         up_write(&obt->obt_rwsem);
1216
1217         /* after qctxt_cleanup, qctxt might be freed, then check_qm() is
1218          * unpredicted. So we must wait until lqc_wait_for_qmaster is empty */
1219         while (cfs_waitq_active(&qctxt->lqc_wait_for_qmaster)) {
1220                 cfs_waitq_signal(&qctxt->lqc_wait_for_qmaster);
1221                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
1222                                      cfs_time_seconds(1));
1223         }
1224
1225         ptlrpcd_decref();
1226
1227 #ifdef LPROCFS
1228         if (lquota_proc_cleanup(qctxt))
1229                 CERROR("cleanup proc error!\n");
1230 #endif
1231
1232         EXIT;
1233 }
1234
1235 struct qslave_recov_thread_data {
1236         struct obd_device *obd;
1237         struct lustre_quota_ctxt *qctxt;
1238         struct completion comp;
1239 };
1240
1241 /* FIXME only recovery block quota by now */
1242 static int qslave_recovery_main(void *arg)
1243 {
1244         struct qslave_recov_thread_data *data = arg;
1245         struct obd_device *obd = data->obd;
1246         struct lustre_quota_ctxt *qctxt = data->qctxt;
1247         unsigned int type;
1248         int rc = 0;
1249         ENTRY;
1250
1251         ptlrpc_daemonize("qslave_recovd");
1252
1253         complete(&data->comp);
1254
1255         if (qctxt->lqc_recovery)
1256                 RETURN(0);
1257         qctxt->lqc_recovery = 1;
1258
1259         for (type = USRQUOTA; type < MAXQUOTAS; type++) {
1260                 struct qunit_data qdata;
1261                 struct quota_info *dqopt = sb_dqopt(qctxt->lqc_sb);
1262                 struct list_head id_list;
1263                 struct dquot_id *dqid, *tmp;
1264                 int ret;
1265
1266                 LOCK_DQONOFF_MUTEX(dqopt);
1267                 if (!sb_has_quota_enabled(qctxt->lqc_sb, type)) {
1268                         UNLOCK_DQONOFF_MUTEX(dqopt);
1269                         break;
1270                 }
1271
1272                 LASSERT(dqopt->files[type] != NULL);
1273                 CFS_INIT_LIST_HEAD(&id_list);
1274 #ifndef KERNEL_SUPPORTS_QUOTA_READ
1275                 rc = fsfilt_qids(obd, dqopt->files[type], NULL, type, &id_list);
1276 #else
1277                 rc = fsfilt_qids(obd, NULL, dqopt->files[type], type, &id_list);
1278 #endif
1279                 UNLOCK_DQONOFF_MUTEX(dqopt);
1280                 if (rc)
1281                         CERROR("Get ids from quota file failed. (rc:%d)\n", rc);
1282
1283                 list_for_each_entry_safe(dqid, tmp, &id_list, di_link) {
1284                         list_del_init(&dqid->di_link);
1285                         /* skip slave recovery on itself */
1286                         if (is_master(qctxt))
1287                                 goto free;
1288                         if (rc && rc != -EBUSY)
1289                                 goto free;
1290
1291                         qdata.qd_id = dqid->di_id;
1292                         qdata.qd_flags = type;
1293                         QDATA_SET_BLK(&qdata);
1294                         qdata.qd_count = 0;
1295
1296                         ret = check_cur_qunit(obd, qctxt, &qdata);
1297                         if (ret > 0) {
1298                                 int opc;
1299                                 opc = ret == 1 ? QUOTA_DQACQ : QUOTA_DQREL;
1300                                 rc = schedule_dqacq(obd, qctxt, &qdata, opc,
1301                                                     0, NULL);
1302                                 if (rc == -EDQUOT)
1303                                         rc = 0;
1304                         } else {
1305                                 rc = 0;
1306                         }
1307
1308                         if (rc)
1309                                 CDEBUG(rc == -EBUSY ? D_QUOTA : D_ERROR,
1310                                        "qslave recovery failed! (id:%d type:%d "
1311                                        " rc:%d)\n", dqid->di_id, type, rc);
1312 free:
1313                         kfree(dqid);
1314                 }
1315         }
1316
1317         qctxt->lqc_recovery = 0;
1318         RETURN(rc);
1319 }
1320
1321 void
1322 qslave_start_recovery(struct obd_device *obd, struct lustre_quota_ctxt *qctxt)
1323 {
1324         struct qslave_recov_thread_data data;
1325         int rc;
1326         ENTRY;
1327
1328         if (!sb_any_quota_enabled(qctxt->lqc_sb))
1329                 goto exit;
1330
1331         data.obd = obd;
1332         data.qctxt = qctxt;
1333         init_completion(&data.comp);
1334
1335         rc = kernel_thread(qslave_recovery_main, &data, CLONE_VM|CLONE_FILES);
1336         if (rc < 0) {
1337                 CERROR("Cannot start quota recovery thread: rc %d\n", rc);
1338                 goto exit;
1339         }
1340         wait_for_completion(&data.comp);
1341 exit:
1342         EXIT;
1343 }
1344
1345
1346 /**
1347  * lqs<->qctxt hash operations
1348  */
1349
1350 /**
1351  * string hashing using djb2 hash algorithm
1352  */
1353 static unsigned
1354 lqs_hash(lustre_hash_t *lh, void *key, unsigned mask)
1355 {
1356         struct quota_adjust_qunit *lqs_key;
1357         unsigned hash;
1358         ENTRY;
1359
1360         LASSERT(key);
1361         lqs_key = (struct quota_adjust_qunit *)key;
1362         hash = (QAQ_IS_GRP(lqs_key) ? 5381 : 5387) * lqs_key->qaq_id;
1363
1364         RETURN(hash & mask);
1365 }
1366
1367 static int
1368 lqs_compare(void *key, struct hlist_node *hnode)
1369 {
1370         struct quota_adjust_qunit *lqs_key;
1371         struct lustre_qunit_size *q;
1372         int rc;
1373         ENTRY;
1374
1375         LASSERT(key);
1376         lqs_key = (struct quota_adjust_qunit *)key;
1377         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1378
1379         spin_lock(&q->lqs_lock);
1380         rc = ((lqs_key->qaq_id == q->lqs_id) &&
1381               (QAQ_IS_GRP(lqs_key) == LQS_IS_GRP(q)));
1382         spin_unlock(&q->lqs_lock);
1383
1384         RETURN(rc);
1385 }
1386
1387 static void *
1388 lqs_get(struct hlist_node *hnode)
1389 {
1390         struct lustre_qunit_size *q =
1391             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1392         ENTRY;
1393
1394         atomic_inc(&q->lqs_refcount);
1395         CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1396                q, atomic_read(&q->lqs_refcount));
1397
1398         RETURN(q);
1399 }
1400
1401 static void *
1402 lqs_put(struct hlist_node *hnode)
1403 {
1404         struct lustre_qunit_size *q =
1405             hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1406         ENTRY;
1407
1408         LASSERT(atomic_read(&q->lqs_refcount) > 0);
1409         atomic_dec(&q->lqs_refcount);
1410         CDEBUG(D_QUOTA, "lqs=%p refcount %d\n",
1411                q, atomic_read(&q->lqs_refcount));
1412
1413         RETURN(q);
1414 }
1415
1416 static void
1417 lqs_exit(struct hlist_node *hnode)
1418 {
1419         struct lustre_qunit_size *q;
1420         ENTRY;
1421
1422         q = hlist_entry(hnode, struct lustre_qunit_size, lqs_hash);
1423         /*
1424          * Nothing should be left. User of lqs put it and
1425          * lqs also was deleted from table by this time
1426          * so we should have 0 refs.
1427          */
1428         LASSERTF(atomic_read(&q->lqs_refcount) == 0,
1429                  "Busy lqs %p with %d refs\n", q,
1430                  atomic_read(&q->lqs_refcount));
1431         OBD_FREE_PTR(q);
1432         EXIT;
1433 }
1434
1435 static lustre_hash_ops_t lqs_hash_ops = {
1436         .lh_hash    = lqs_hash,
1437         .lh_compare = lqs_compare,
1438         .lh_get     = lqs_get,
1439         .lh_put     = lqs_put,
1440         .lh_exit    = lqs_exit
1441 };
1442 #endif /* HAVE_QUOTA_SUPPORT */